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>
110 #include <linux/rtmutex.h>
112 #include <asm/uaccess.h>
113 #include <asm/cacheflush.h>
114 #include <asm/tlbflush.h>
115 #include <asm/page.h>
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
154 #define ARCH_KMALLOC_MINALIGN 0
157 #ifndef ARCH_SLAB_MINALIGN
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
165 #define ARCH_SLAB_MINALIGN 0
168 #ifndef ARCH_KMALLOC_FLAGS
169 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 /* Legal flag mask for kmem_cache_create(). */
174 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
181 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
190 * Bufctl's are used for linking objs within a slab
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 typedef unsigned int kmem_bufctl_t;
207 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
209 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
225 unsigned short nodeid;
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
242 * We assume struct slab_rcu can overlay struct slab when destroying.
245 struct rcu_head head;
246 struct kmem_cache *cachep;
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
258 * The limit is stored in the per-cpu structure to reduce the data cache
265 unsigned int batchcount;
266 unsigned int touched;
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
272 * [0] is for gcc 2.95. It should really be [].
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
287 * The slab lists for all objects.
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned int free_limit;
295 unsigned int colour_next; /* Per-node cache coloring */
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
312 static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 static void enable_cpucache(struct kmem_cache *cachep);
317 static void cache_reap(void *unused);
320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
323 static __always_inline int index_of(const size_t size)
325 extern void __bad_size(void);
327 if (__builtin_constant_p(size)) {
335 #include "linux/kmalloc_sizes.h"
343 static int slab_early_init = 1;
345 #define INDEX_AC index_of(sizeof(struct arraycache_init))
346 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
348 static void kmem_list3_init(struct kmem_list3 *parent)
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
355 parent->colour_next = 0;
356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
361 #define MAKE_LIST(cachep, listp, slab, nodeid) \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
367 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 /* 1) per-cpu data, touched during every alloc/free */
382 struct array_cache *array[NR_CPUS];
383 /* 2) Cache tunables. Protected by cache_chain_mutex */
384 unsigned int batchcount;
388 unsigned int buffer_size;
389 /* 3) touched by every alloc & free from the backend */
390 struct kmem_list3 *nodelists[MAX_NUMNODES];
392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
395 /* 4) cache_grow/shrink */
396 /* order of pgs per slab (2^n) */
397 unsigned int gfporder;
399 /* force GFP flags, e.g. GFP_DMA */
402 size_t colour; /* cache colouring range */
403 unsigned int colour_off; /* colour offset */
404 struct kmem_cache *slabp_cache;
405 unsigned int slab_size;
406 unsigned int dflags; /* dynamic flags */
408 /* constructor func */
409 void (*ctor) (void *, struct kmem_cache *, unsigned long);
411 /* de-constructor func */
412 void (*dtor) (void *, struct kmem_cache *, unsigned long);
414 /* 5) cache creation/removal */
416 struct list_head next;
420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
429 unsigned long node_overflow;
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
447 #define CFLGS_OFF_SLAB (0x80000000UL)
448 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
450 #define BATCHREFILL_LIMIT 16
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
455 * OTOH the cpuarrays can contain lots of objects,
456 * which could lock up otherwise freeable slabs.
458 #define REAPTIMEOUT_CPUC (2*HZ)
459 #define REAPTIMEOUT_LIST3 (4*HZ)
462 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
463 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465 #define STATS_INC_GROWN(x) ((x)->grown++)
466 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
467 #define STATS_SET_HIGH(x) \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
472 #define STATS_INC_ERR(x) ((x)->errors++)
473 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
474 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
475 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
476 #define STATS_SET_FREEABLE(x, i) \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
481 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
486 #define STATS_INC_ACTIVE(x) do { } while (0)
487 #define STATS_DEC_ACTIVE(x) do { } while (0)
488 #define STATS_INC_ALLOCED(x) do { } while (0)
489 #define STATS_INC_GROWN(x) do { } while (0)
490 #define STATS_ADD_REAPED(x,y) do { } while (0)
491 #define STATS_SET_HIGH(x) do { } while (0)
492 #define STATS_INC_ERR(x) do { } while (0)
493 #define STATS_INC_NODEALLOCS(x) do { } while (0)
494 #define STATS_INC_NODEFREES(x) do { } while (0)
495 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
496 #define STATS_SET_FREEABLE(x, i) do { } while (0)
497 #define STATS_INC_ALLOCHIT(x) do { } while (0)
498 #define STATS_INC_ALLOCMISS(x) do { } while (0)
499 #define STATS_INC_FREEHIT(x) do { } while (0)
500 #define STATS_INC_FREEMISS(x) do { } while (0)
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache *cachep)
520 return cachep->obj_offset;
523 static int obj_size(struct kmem_cache *cachep)
525 return cachep->obj_size;
528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
538 return (unsigned long *)(objp + cachep->buffer_size -
540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
543 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
588 page->lru.next = (struct list_head *)cache;
591 static inline struct kmem_cache *page_get_cache(struct page *page)
593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
595 BUG_ON(!PageSlab(page));
596 return (struct kmem_cache *)page->lru.next;
599 static inline void page_set_slab(struct page *page, struct slab *slab)
601 page->lru.prev = (struct list_head *)slab;
604 static inline struct slab *page_get_slab(struct page *page)
606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
608 BUG_ON(!PageSlab(page));
609 return (struct slab *)page->lru.prev;
612 static inline struct kmem_cache *virt_to_cache(const void *obj)
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
618 static inline struct slab *virt_to_slab(const void *obj)
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
624 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
627 return slab->s_mem + cache->buffer_size * idx;
630 static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 struct cache_sizes malloc_sizes[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
645 EXPORT_SYMBOL(malloc_sizes);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
653 static struct cache_names __initdata cache_names[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
660 static struct arraycache_init initarray_cache __initdata =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
662 static struct arraycache_init initarray_generic =
663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache = {
668 .limit = BOOT_CPUCACHE_ENTRIES,
670 .buffer_size = sizeof(struct kmem_cache),
671 .name = "kmem_cache",
673 .obj_size = sizeof(struct kmem_cache),
677 #ifdef CONFIG_LOCKDEP
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
686 static struct lock_class_key on_slab_key;
688 static inline void init_lock_keys(struct cache_sizes *s)
692 for (q = 0; q < MAX_NUMNODES; q++) {
693 if (!s->cs_cachep->nodelists[q] || OFF_SLAB(s->cs_cachep))
695 lockdep_set_class(&s->cs_cachep->nodelists[q]->list_lock,
701 static inline void init_lock_keys(struct cache_sizes *s)
708 /* Guard access to the cache-chain. */
709 static DEFINE_MUTEX(cache_chain_mutex);
710 static struct list_head cache_chain;
713 * vm_enough_memory() looks at this to determine how many slab-allocated pages
714 * are possibly freeable under pressure
716 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
718 atomic_t slab_reclaim_pages;
721 * chicken and egg problem: delay the per-cpu array allocation
722 * until the general caches are up.
732 * used by boot code to determine if it can use slab based allocator
734 int slab_is_available(void)
736 return g_cpucache_up == FULL;
739 static DEFINE_PER_CPU(struct work_struct, reap_work);
741 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
743 return cachep->array[smp_processor_id()];
746 static inline struct kmem_cache *__find_general_cachep(size_t size,
749 struct cache_sizes *csizep = malloc_sizes;
752 /* This happens if someone tries to call
753 * kmem_cache_create(), or __kmalloc(), before
754 * the generic caches are initialized.
756 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
758 while (size > csizep->cs_size)
762 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
763 * has cs_{dma,}cachep==NULL. Thus no special case
764 * for large kmalloc calls required.
766 if (unlikely(gfpflags & GFP_DMA))
767 return csizep->cs_dmacachep;
768 return csizep->cs_cachep;
771 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
773 return __find_general_cachep(size, gfpflags);
776 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
778 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
782 * Calculate the number of objects and left-over bytes for a given buffer size.
784 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
785 size_t align, int flags, size_t *left_over,
790 size_t slab_size = PAGE_SIZE << gfporder;
793 * The slab management structure can be either off the slab or
794 * on it. For the latter case, the memory allocated for a
798 * - One kmem_bufctl_t for each object
799 * - Padding to respect alignment of @align
800 * - @buffer_size bytes for each object
802 * If the slab management structure is off the slab, then the
803 * alignment will already be calculated into the size. Because
804 * the slabs are all pages aligned, the objects will be at the
805 * correct alignment when allocated.
807 if (flags & CFLGS_OFF_SLAB) {
809 nr_objs = slab_size / buffer_size;
811 if (nr_objs > SLAB_LIMIT)
812 nr_objs = SLAB_LIMIT;
815 * Ignore padding for the initial guess. The padding
816 * is at most @align-1 bytes, and @buffer_size is at
817 * least @align. In the worst case, this result will
818 * be one greater than the number of objects that fit
819 * into the memory allocation when taking the padding
822 nr_objs = (slab_size - sizeof(struct slab)) /
823 (buffer_size + sizeof(kmem_bufctl_t));
826 * This calculated number will be either the right
827 * amount, or one greater than what we want.
829 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
833 if (nr_objs > SLAB_LIMIT)
834 nr_objs = SLAB_LIMIT;
836 mgmt_size = slab_mgmt_size(nr_objs, align);
839 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
842 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
844 static void __slab_error(const char *function, struct kmem_cache *cachep,
847 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
848 function, cachep->name, msg);
854 * Special reaping functions for NUMA systems called from cache_reap().
855 * These take care of doing round robin flushing of alien caches (containing
856 * objects freed on different nodes from which they were allocated) and the
857 * flushing of remote pcps by calling drain_node_pages.
859 static DEFINE_PER_CPU(unsigned long, reap_node);
861 static void init_reap_node(int cpu)
865 node = next_node(cpu_to_node(cpu), node_online_map);
866 if (node == MAX_NUMNODES)
867 node = first_node(node_online_map);
869 __get_cpu_var(reap_node) = node;
872 static void next_reap_node(void)
874 int node = __get_cpu_var(reap_node);
877 * Also drain per cpu pages on remote zones
879 if (node != numa_node_id())
880 drain_node_pages(node);
882 node = next_node(node, node_online_map);
883 if (unlikely(node >= MAX_NUMNODES))
884 node = first_node(node_online_map);
885 __get_cpu_var(reap_node) = node;
889 #define init_reap_node(cpu) do { } while (0)
890 #define next_reap_node(void) do { } while (0)
894 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
895 * via the workqueue/eventd.
896 * Add the CPU number into the expiration time to minimize the possibility of
897 * the CPUs getting into lockstep and contending for the global cache chain
900 static void __devinit start_cpu_timer(int cpu)
902 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
905 * When this gets called from do_initcalls via cpucache_init(),
906 * init_workqueues() has already run, so keventd will be setup
909 if (keventd_up() && reap_work->func == NULL) {
911 INIT_WORK(reap_work, cache_reap, NULL);
912 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
916 static struct array_cache *alloc_arraycache(int node, int entries,
919 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
920 struct array_cache *nc = NULL;
922 nc = kmalloc_node(memsize, GFP_KERNEL, node);
926 nc->batchcount = batchcount;
928 spin_lock_init(&nc->lock);
934 * Transfer objects in one arraycache to another.
935 * Locking must be handled by the caller.
937 * Return the number of entries transferred.
939 static int transfer_objects(struct array_cache *to,
940 struct array_cache *from, unsigned int max)
942 /* Figure out how many entries to transfer */
943 int nr = min(min(from->avail, max), to->limit - to->avail);
948 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
958 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
959 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
961 static struct array_cache **alloc_alien_cache(int node, int limit)
963 struct array_cache **ac_ptr;
964 int memsize = sizeof(void *) * MAX_NUMNODES;
969 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
972 if (i == node || !node_online(i)) {
976 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
978 for (i--; i <= 0; i--)
988 static void free_alien_cache(struct array_cache **ac_ptr)
999 static void __drain_alien_cache(struct kmem_cache *cachep,
1000 struct array_cache *ac, int node)
1002 struct kmem_list3 *rl3 = cachep->nodelists[node];
1005 spin_lock(&rl3->list_lock);
1007 * Stuff objects into the remote nodes shared array first.
1008 * That way we could avoid the overhead of putting the objects
1009 * into the free lists and getting them back later.
1012 transfer_objects(rl3->shared, ac, ac->limit);
1014 free_block(cachep, ac->entry, ac->avail, node);
1016 spin_unlock(&rl3->list_lock);
1021 * Called from cache_reap() to regularly drain alien caches round robin.
1023 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1025 int node = __get_cpu_var(reap_node);
1028 struct array_cache *ac = l3->alien[node];
1030 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1031 __drain_alien_cache(cachep, ac, node);
1032 spin_unlock_irq(&ac->lock);
1037 static void drain_alien_cache(struct kmem_cache *cachep,
1038 struct array_cache **alien)
1041 struct array_cache *ac;
1042 unsigned long flags;
1044 for_each_online_node(i) {
1047 spin_lock_irqsave(&ac->lock, flags);
1048 __drain_alien_cache(cachep, ac, i);
1049 spin_unlock_irqrestore(&ac->lock, flags);
1054 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1056 struct slab *slabp = virt_to_slab(objp);
1057 int nodeid = slabp->nodeid;
1058 struct kmem_list3 *l3;
1059 struct array_cache *alien = NULL;
1062 * Make sure we are not freeing a object from another node to the array
1063 * cache on this cpu.
1065 if (likely(slabp->nodeid == numa_node_id()))
1068 l3 = cachep->nodelists[numa_node_id()];
1069 STATS_INC_NODEFREES(cachep);
1070 if (l3->alien && l3->alien[nodeid]) {
1071 alien = l3->alien[nodeid];
1072 spin_lock(&alien->lock);
1073 if (unlikely(alien->avail == alien->limit)) {
1074 STATS_INC_ACOVERFLOW(cachep);
1075 __drain_alien_cache(cachep, alien, nodeid);
1077 alien->entry[alien->avail++] = objp;
1078 spin_unlock(&alien->lock);
1080 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1081 free_block(cachep, &objp, 1, nodeid);
1082 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1089 #define drain_alien_cache(cachep, alien) do { } while (0)
1090 #define reap_alien(cachep, l3) do { } while (0)
1092 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1094 return (struct array_cache **) 0x01020304ul;
1097 static inline void free_alien_cache(struct array_cache **ac_ptr)
1101 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1108 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1109 unsigned long action, void *hcpu)
1111 long cpu = (long)hcpu;
1112 struct kmem_cache *cachep;
1113 struct kmem_list3 *l3 = NULL;
1114 int node = cpu_to_node(cpu);
1115 int memsize = sizeof(struct kmem_list3);
1118 case CPU_UP_PREPARE:
1119 mutex_lock(&cache_chain_mutex);
1121 * We need to do this right in the beginning since
1122 * alloc_arraycache's are going to use this list.
1123 * kmalloc_node allows us to add the slab to the right
1124 * kmem_list3 and not this cpu's kmem_list3
1127 list_for_each_entry(cachep, &cache_chain, next) {
1129 * Set up the size64 kmemlist for cpu before we can
1130 * begin anything. Make sure some other cpu on this
1131 * node has not already allocated this
1133 if (!cachep->nodelists[node]) {
1134 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1137 kmem_list3_init(l3);
1138 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1139 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1142 * The l3s don't come and go as CPUs come and
1143 * go. cache_chain_mutex is sufficient
1146 cachep->nodelists[node] = l3;
1149 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1150 cachep->nodelists[node]->free_limit =
1151 (1 + nr_cpus_node(node)) *
1152 cachep->batchcount + cachep->num;
1153 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1157 * Now we can go ahead with allocating the shared arrays and
1160 list_for_each_entry(cachep, &cache_chain, next) {
1161 struct array_cache *nc;
1162 struct array_cache *shared;
1163 struct array_cache **alien;
1165 nc = alloc_arraycache(node, cachep->limit,
1166 cachep->batchcount);
1169 shared = alloc_arraycache(node,
1170 cachep->shared * cachep->batchcount,
1175 alien = alloc_alien_cache(node, cachep->limit);
1178 cachep->array[cpu] = nc;
1179 l3 = cachep->nodelists[node];
1182 spin_lock_irq(&l3->list_lock);
1185 * We are serialised from CPU_DEAD or
1186 * CPU_UP_CANCELLED by the cpucontrol lock
1188 l3->shared = shared;
1197 spin_unlock_irq(&l3->list_lock);
1199 free_alien_cache(alien);
1201 mutex_unlock(&cache_chain_mutex);
1204 start_cpu_timer(cpu);
1206 #ifdef CONFIG_HOTPLUG_CPU
1209 * Even if all the cpus of a node are down, we don't free the
1210 * kmem_list3 of any cache. This to avoid a race between
1211 * cpu_down, and a kmalloc allocation from another cpu for
1212 * memory from the node of the cpu going down. The list3
1213 * structure is usually allocated from kmem_cache_create() and
1214 * gets destroyed at kmem_cache_destroy().
1217 case CPU_UP_CANCELED:
1218 mutex_lock(&cache_chain_mutex);
1219 list_for_each_entry(cachep, &cache_chain, next) {
1220 struct array_cache *nc;
1221 struct array_cache *shared;
1222 struct array_cache **alien;
1225 mask = node_to_cpumask(node);
1226 /* cpu is dead; no one can alloc from it. */
1227 nc = cachep->array[cpu];
1228 cachep->array[cpu] = NULL;
1229 l3 = cachep->nodelists[node];
1232 goto free_array_cache;
1234 spin_lock_irq(&l3->list_lock);
1236 /* Free limit for this kmem_list3 */
1237 l3->free_limit -= cachep->batchcount;
1239 free_block(cachep, nc->entry, nc->avail, node);
1241 if (!cpus_empty(mask)) {
1242 spin_unlock_irq(&l3->list_lock);
1243 goto free_array_cache;
1246 shared = l3->shared;
1248 free_block(cachep, l3->shared->entry,
1249 l3->shared->avail, node);
1256 spin_unlock_irq(&l3->list_lock);
1260 drain_alien_cache(cachep, alien);
1261 free_alien_cache(alien);
1267 * In the previous loop, all the objects were freed to
1268 * the respective cache's slabs, now we can go ahead and
1269 * shrink each nodelist to its limit.
1271 list_for_each_entry(cachep, &cache_chain, next) {
1272 l3 = cachep->nodelists[node];
1275 drain_freelist(cachep, l3, l3->free_objects);
1277 mutex_unlock(&cache_chain_mutex);
1283 mutex_unlock(&cache_chain_mutex);
1287 static struct notifier_block __cpuinitdata cpucache_notifier = {
1288 &cpuup_callback, NULL, 0
1292 * swap the static kmem_list3 with kmalloced memory
1294 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1297 struct kmem_list3 *ptr;
1299 BUG_ON(cachep->nodelists[nodeid] != list);
1300 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1303 local_irq_disable();
1304 memcpy(ptr, list, sizeof(struct kmem_list3));
1306 * Do not assume that spinlocks can be initialized via memcpy:
1308 spin_lock_init(&ptr->list_lock);
1310 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1311 cachep->nodelists[nodeid] = ptr;
1316 * Initialisation. Called after the page allocator have been initialised and
1317 * before smp_init().
1319 void __init kmem_cache_init(void)
1322 struct cache_sizes *sizes;
1323 struct cache_names *names;
1327 for (i = 0; i < NUM_INIT_LISTS; i++) {
1328 kmem_list3_init(&initkmem_list3[i]);
1329 if (i < MAX_NUMNODES)
1330 cache_cache.nodelists[i] = NULL;
1334 * Fragmentation resistance on low memory - only use bigger
1335 * page orders on machines with more than 32MB of memory.
1337 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1338 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1340 /* Bootstrap is tricky, because several objects are allocated
1341 * from caches that do not exist yet:
1342 * 1) initialize the cache_cache cache: it contains the struct
1343 * kmem_cache structures of all caches, except cache_cache itself:
1344 * cache_cache is statically allocated.
1345 * Initially an __init data area is used for the head array and the
1346 * kmem_list3 structures, it's replaced with a kmalloc allocated
1347 * array at the end of the bootstrap.
1348 * 2) Create the first kmalloc cache.
1349 * The struct kmem_cache for the new cache is allocated normally.
1350 * An __init data area is used for the head array.
1351 * 3) Create the remaining kmalloc caches, with minimally sized
1353 * 4) Replace the __init data head arrays for cache_cache and the first
1354 * kmalloc cache with kmalloc allocated arrays.
1355 * 5) Replace the __init data for kmem_list3 for cache_cache and
1356 * the other cache's with kmalloc allocated memory.
1357 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1360 /* 1) create the cache_cache */
1361 INIT_LIST_HEAD(&cache_chain);
1362 list_add(&cache_cache.next, &cache_chain);
1363 cache_cache.colour_off = cache_line_size();
1364 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1365 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1367 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1370 for (order = 0; order < MAX_ORDER; order++) {
1371 cache_estimate(order, cache_cache.buffer_size,
1372 cache_line_size(), 0, &left_over, &cache_cache.num);
1373 if (cache_cache.num)
1376 BUG_ON(!cache_cache.num);
1377 cache_cache.gfporder = order;
1378 cache_cache.colour = left_over / cache_cache.colour_off;
1379 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1380 sizeof(struct slab), cache_line_size());
1382 /* 2+3) create the kmalloc caches */
1383 sizes = malloc_sizes;
1384 names = cache_names;
1387 * Initialize the caches that provide memory for the array cache and the
1388 * kmem_list3 structures first. Without this, further allocations will
1392 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1393 sizes[INDEX_AC].cs_size,
1394 ARCH_KMALLOC_MINALIGN,
1395 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1398 if (INDEX_AC != INDEX_L3) {
1399 sizes[INDEX_L3].cs_cachep =
1400 kmem_cache_create(names[INDEX_L3].name,
1401 sizes[INDEX_L3].cs_size,
1402 ARCH_KMALLOC_MINALIGN,
1403 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1407 slab_early_init = 0;
1409 while (sizes->cs_size != ULONG_MAX) {
1411 * For performance, all the general caches are L1 aligned.
1412 * This should be particularly beneficial on SMP boxes, as it
1413 * eliminates "false sharing".
1414 * Note for systems short on memory removing the alignment will
1415 * allow tighter packing of the smaller caches.
1417 if (!sizes->cs_cachep) {
1418 sizes->cs_cachep = kmem_cache_create(names->name,
1420 ARCH_KMALLOC_MINALIGN,
1421 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1424 init_lock_keys(sizes);
1426 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1428 ARCH_KMALLOC_MINALIGN,
1429 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1435 /* 4) Replace the bootstrap head arrays */
1437 struct array_cache *ptr;
1439 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1441 local_irq_disable();
1442 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1443 memcpy(ptr, cpu_cache_get(&cache_cache),
1444 sizeof(struct arraycache_init));
1446 * Do not assume that spinlocks can be initialized via memcpy:
1448 spin_lock_init(&ptr->lock);
1450 cache_cache.array[smp_processor_id()] = ptr;
1453 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1455 local_irq_disable();
1456 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1457 != &initarray_generic.cache);
1458 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1459 sizeof(struct arraycache_init));
1461 * Do not assume that spinlocks can be initialized via memcpy:
1463 spin_lock_init(&ptr->lock);
1465 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1469 /* 5) Replace the bootstrap kmem_list3's */
1472 /* Replace the static kmem_list3 structures for the boot cpu */
1473 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1476 for_each_online_node(node) {
1477 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1478 &initkmem_list3[SIZE_AC + node], node);
1480 if (INDEX_AC != INDEX_L3) {
1481 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1482 &initkmem_list3[SIZE_L3 + node],
1488 /* 6) resize the head arrays to their final sizes */
1490 struct kmem_cache *cachep;
1491 mutex_lock(&cache_chain_mutex);
1492 list_for_each_entry(cachep, &cache_chain, next)
1493 enable_cpucache(cachep);
1494 mutex_unlock(&cache_chain_mutex);
1498 g_cpucache_up = FULL;
1501 * Register a cpu startup notifier callback that initializes
1502 * cpu_cache_get for all new cpus
1504 register_cpu_notifier(&cpucache_notifier);
1507 * The reap timers are started later, with a module init call: That part
1508 * of the kernel is not yet operational.
1512 static int __init cpucache_init(void)
1517 * Register the timers that return unneeded pages to the page allocator
1519 for_each_online_cpu(cpu)
1520 start_cpu_timer(cpu);
1523 __initcall(cpucache_init);
1526 * Interface to system's page allocator. No need to hold the cache-lock.
1528 * If we requested dmaable memory, we will get it. Even if we
1529 * did not request dmaable memory, we might get it, but that
1530 * would be relatively rare and ignorable.
1532 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1540 * Nommu uses slab's for process anonymous memory allocations, and thus
1541 * requires __GFP_COMP to properly refcount higher order allocations
1543 flags |= __GFP_COMP;
1545 flags |= cachep->gfpflags;
1547 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1551 nr_pages = (1 << cachep->gfporder);
1552 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1553 atomic_add(nr_pages, &slab_reclaim_pages);
1554 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages);
1555 for (i = 0; i < nr_pages; i++)
1556 __SetPageSlab(page + i);
1557 return page_address(page);
1561 * Interface to system's page release.
1563 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1565 unsigned long i = (1 << cachep->gfporder);
1566 struct page *page = virt_to_page(addr);
1567 const unsigned long nr_freed = i;
1569 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed);
1571 BUG_ON(!PageSlab(page));
1572 __ClearPageSlab(page);
1575 if (current->reclaim_state)
1576 current->reclaim_state->reclaimed_slab += nr_freed;
1577 free_pages((unsigned long)addr, cachep->gfporder);
1578 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1579 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1582 static void kmem_rcu_free(struct rcu_head *head)
1584 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1585 struct kmem_cache *cachep = slab_rcu->cachep;
1587 kmem_freepages(cachep, slab_rcu->addr);
1588 if (OFF_SLAB(cachep))
1589 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1594 #ifdef CONFIG_DEBUG_PAGEALLOC
1595 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1596 unsigned long caller)
1598 int size = obj_size(cachep);
1600 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1602 if (size < 5 * sizeof(unsigned long))
1605 *addr++ = 0x12345678;
1607 *addr++ = smp_processor_id();
1608 size -= 3 * sizeof(unsigned long);
1610 unsigned long *sptr = &caller;
1611 unsigned long svalue;
1613 while (!kstack_end(sptr)) {
1615 if (kernel_text_address(svalue)) {
1617 size -= sizeof(unsigned long);
1618 if (size <= sizeof(unsigned long))
1624 *addr++ = 0x87654321;
1628 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1630 int size = obj_size(cachep);
1631 addr = &((char *)addr)[obj_offset(cachep)];
1633 memset(addr, val, size);
1634 *(unsigned char *)(addr + size - 1) = POISON_END;
1637 static void dump_line(char *data, int offset, int limit)
1640 printk(KERN_ERR "%03x:", offset);
1641 for (i = 0; i < limit; i++)
1642 printk(" %02x", (unsigned char)data[offset + i]);
1649 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1654 if (cachep->flags & SLAB_RED_ZONE) {
1655 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1656 *dbg_redzone1(cachep, objp),
1657 *dbg_redzone2(cachep, objp));
1660 if (cachep->flags & SLAB_STORE_USER) {
1661 printk(KERN_ERR "Last user: [<%p>]",
1662 *dbg_userword(cachep, objp));
1663 print_symbol("(%s)",
1664 (unsigned long)*dbg_userword(cachep, objp));
1667 realobj = (char *)objp + obj_offset(cachep);
1668 size = obj_size(cachep);
1669 for (i = 0; i < size && lines; i += 16, lines--) {
1672 if (i + limit > size)
1674 dump_line(realobj, i, limit);
1678 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1684 realobj = (char *)objp + obj_offset(cachep);
1685 size = obj_size(cachep);
1687 for (i = 0; i < size; i++) {
1688 char exp = POISON_FREE;
1691 if (realobj[i] != exp) {
1697 "Slab corruption: start=%p, len=%d\n",
1699 print_objinfo(cachep, objp, 0);
1701 /* Hexdump the affected line */
1704 if (i + limit > size)
1706 dump_line(realobj, i, limit);
1709 /* Limit to 5 lines */
1715 /* Print some data about the neighboring objects, if they
1718 struct slab *slabp = virt_to_slab(objp);
1721 objnr = obj_to_index(cachep, slabp, objp);
1723 objp = index_to_obj(cachep, slabp, objnr - 1);
1724 realobj = (char *)objp + obj_offset(cachep);
1725 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1727 print_objinfo(cachep, objp, 2);
1729 if (objnr + 1 < cachep->num) {
1730 objp = index_to_obj(cachep, slabp, objnr + 1);
1731 realobj = (char *)objp + obj_offset(cachep);
1732 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1734 print_objinfo(cachep, objp, 2);
1742 * slab_destroy_objs - destroy a slab and its objects
1743 * @cachep: cache pointer being destroyed
1744 * @slabp: slab pointer being destroyed
1746 * Call the registered destructor for each object in a slab that is being
1749 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1752 for (i = 0; i < cachep->num; i++) {
1753 void *objp = index_to_obj(cachep, slabp, i);
1755 if (cachep->flags & SLAB_POISON) {
1756 #ifdef CONFIG_DEBUG_PAGEALLOC
1757 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1759 kernel_map_pages(virt_to_page(objp),
1760 cachep->buffer_size / PAGE_SIZE, 1);
1762 check_poison_obj(cachep, objp);
1764 check_poison_obj(cachep, objp);
1767 if (cachep->flags & SLAB_RED_ZONE) {
1768 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1769 slab_error(cachep, "start of a freed object "
1771 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1772 slab_error(cachep, "end of a freed object "
1775 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1776 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1780 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1784 for (i = 0; i < cachep->num; i++) {
1785 void *objp = index_to_obj(cachep, slabp, i);
1786 (cachep->dtor) (objp, cachep, 0);
1793 * slab_destroy - destroy and release all objects in a slab
1794 * @cachep: cache pointer being destroyed
1795 * @slabp: slab pointer being destroyed
1797 * Destroy all the objs in a slab, and release the mem back to the system.
1798 * Before calling the slab must have been unlinked from the cache. The
1799 * cache-lock is not held/needed.
1801 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1803 void *addr = slabp->s_mem - slabp->colouroff;
1805 slab_destroy_objs(cachep, slabp);
1806 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1807 struct slab_rcu *slab_rcu;
1809 slab_rcu = (struct slab_rcu *)slabp;
1810 slab_rcu->cachep = cachep;
1811 slab_rcu->addr = addr;
1812 call_rcu(&slab_rcu->head, kmem_rcu_free);
1814 kmem_freepages(cachep, addr);
1815 if (OFF_SLAB(cachep))
1816 kmem_cache_free(cachep->slabp_cache, slabp);
1821 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1822 * size of kmem_list3.
1824 static void set_up_list3s(struct kmem_cache *cachep, int index)
1828 for_each_online_node(node) {
1829 cachep->nodelists[node] = &initkmem_list3[index + node];
1830 cachep->nodelists[node]->next_reap = jiffies +
1832 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1837 * calculate_slab_order - calculate size (page order) of slabs
1838 * @cachep: pointer to the cache that is being created
1839 * @size: size of objects to be created in this cache.
1840 * @align: required alignment for the objects.
1841 * @flags: slab allocation flags
1843 * Also calculates the number of objects per slab.
1845 * This could be made much more intelligent. For now, try to avoid using
1846 * high order pages for slabs. When the gfp() functions are more friendly
1847 * towards high-order requests, this should be changed.
1849 static size_t calculate_slab_order(struct kmem_cache *cachep,
1850 size_t size, size_t align, unsigned long flags)
1852 unsigned long offslab_limit;
1853 size_t left_over = 0;
1856 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1860 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1864 if (flags & CFLGS_OFF_SLAB) {
1866 * Max number of objs-per-slab for caches which
1867 * use off-slab slabs. Needed to avoid a possible
1868 * looping condition in cache_grow().
1870 offslab_limit = size - sizeof(struct slab);
1871 offslab_limit /= sizeof(kmem_bufctl_t);
1873 if (num > offslab_limit)
1877 /* Found something acceptable - save it away */
1879 cachep->gfporder = gfporder;
1880 left_over = remainder;
1883 * A VFS-reclaimable slab tends to have most allocations
1884 * as GFP_NOFS and we really don't want to have to be allocating
1885 * higher-order pages when we are unable to shrink dcache.
1887 if (flags & SLAB_RECLAIM_ACCOUNT)
1891 * Large number of objects is good, but very large slabs are
1892 * currently bad for the gfp()s.
1894 if (gfporder >= slab_break_gfp_order)
1898 * Acceptable internal fragmentation?
1900 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1906 static void setup_cpu_cache(struct kmem_cache *cachep)
1908 if (g_cpucache_up == FULL) {
1909 enable_cpucache(cachep);
1912 if (g_cpucache_up == NONE) {
1914 * Note: the first kmem_cache_create must create the cache
1915 * that's used by kmalloc(24), otherwise the creation of
1916 * further caches will BUG().
1918 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1921 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1922 * the first cache, then we need to set up all its list3s,
1923 * otherwise the creation of further caches will BUG().
1925 set_up_list3s(cachep, SIZE_AC);
1926 if (INDEX_AC == INDEX_L3)
1927 g_cpucache_up = PARTIAL_L3;
1929 g_cpucache_up = PARTIAL_AC;
1931 cachep->array[smp_processor_id()] =
1932 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1934 if (g_cpucache_up == PARTIAL_AC) {
1935 set_up_list3s(cachep, SIZE_L3);
1936 g_cpucache_up = PARTIAL_L3;
1939 for_each_online_node(node) {
1940 cachep->nodelists[node] =
1941 kmalloc_node(sizeof(struct kmem_list3),
1943 BUG_ON(!cachep->nodelists[node]);
1944 kmem_list3_init(cachep->nodelists[node]);
1948 cachep->nodelists[numa_node_id()]->next_reap =
1949 jiffies + REAPTIMEOUT_LIST3 +
1950 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1952 cpu_cache_get(cachep)->avail = 0;
1953 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1954 cpu_cache_get(cachep)->batchcount = 1;
1955 cpu_cache_get(cachep)->touched = 0;
1956 cachep->batchcount = 1;
1957 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1961 * kmem_cache_create - Create a cache.
1962 * @name: A string which is used in /proc/slabinfo to identify this cache.
1963 * @size: The size of objects to be created in this cache.
1964 * @align: The required alignment for the objects.
1965 * @flags: SLAB flags
1966 * @ctor: A constructor for the objects.
1967 * @dtor: A destructor for the objects.
1969 * Returns a ptr to the cache on success, NULL on failure.
1970 * Cannot be called within a int, but can be interrupted.
1971 * The @ctor is run when new pages are allocated by the cache
1972 * and the @dtor is run before the pages are handed back.
1974 * @name must be valid until the cache is destroyed. This implies that
1975 * the module calling this has to destroy the cache before getting unloaded.
1979 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1980 * to catch references to uninitialised memory.
1982 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1983 * for buffer overruns.
1985 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1986 * cacheline. This can be beneficial if you're counting cycles as closely
1990 kmem_cache_create (const char *name, size_t size, size_t align,
1991 unsigned long flags,
1992 void (*ctor)(void*, struct kmem_cache *, unsigned long),
1993 void (*dtor)(void*, struct kmem_cache *, unsigned long))
1995 size_t left_over, slab_size, ralign;
1996 struct kmem_cache *cachep = NULL, *pc;
1999 * Sanity checks... these are all serious usage bugs.
2001 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2002 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2003 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2009 * Prevent CPUs from coming and going.
2010 * lock_cpu_hotplug() nests outside cache_chain_mutex
2014 mutex_lock(&cache_chain_mutex);
2016 list_for_each_entry(pc, &cache_chain, next) {
2017 mm_segment_t old_fs = get_fs();
2022 * This happens when the module gets unloaded and doesn't
2023 * destroy its slab cache and no-one else reuses the vmalloc
2024 * area of the module. Print a warning.
2027 res = __get_user(tmp, pc->name);
2030 printk("SLAB: cache with size %d has lost its name\n",
2035 if (!strcmp(pc->name, name)) {
2036 printk("kmem_cache_create: duplicate cache %s\n", name);
2043 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2044 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2045 /* No constructor, but inital state check requested */
2046 printk(KERN_ERR "%s: No con, but init state check "
2047 "requested - %s\n", __FUNCTION__, name);
2048 flags &= ~SLAB_DEBUG_INITIAL;
2052 * Enable redzoning and last user accounting, except for caches with
2053 * large objects, if the increased size would increase the object size
2054 * above the next power of two: caches with object sizes just above a
2055 * power of two have a significant amount of internal fragmentation.
2057 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2058 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2059 if (!(flags & SLAB_DESTROY_BY_RCU))
2060 flags |= SLAB_POISON;
2062 if (flags & SLAB_DESTROY_BY_RCU)
2063 BUG_ON(flags & SLAB_POISON);
2065 if (flags & SLAB_DESTROY_BY_RCU)
2069 * Always checks flags, a caller might be expecting debug support which
2072 BUG_ON(flags & ~CREATE_MASK);
2075 * Check that size is in terms of words. This is needed to avoid
2076 * unaligned accesses for some archs when redzoning is used, and makes
2077 * sure any on-slab bufctl's are also correctly aligned.
2079 if (size & (BYTES_PER_WORD - 1)) {
2080 size += (BYTES_PER_WORD - 1);
2081 size &= ~(BYTES_PER_WORD - 1);
2084 /* calculate the final buffer alignment: */
2086 /* 1) arch recommendation: can be overridden for debug */
2087 if (flags & SLAB_HWCACHE_ALIGN) {
2089 * Default alignment: as specified by the arch code. Except if
2090 * an object is really small, then squeeze multiple objects into
2093 ralign = cache_line_size();
2094 while (size <= ralign / 2)
2097 ralign = BYTES_PER_WORD;
2101 * Redzoning and user store require word alignment. Note this will be
2102 * overridden by architecture or caller mandated alignment if either
2103 * is greater than BYTES_PER_WORD.
2105 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2106 ralign = BYTES_PER_WORD;
2108 /* 2) arch mandated alignment: disables debug if necessary */
2109 if (ralign < ARCH_SLAB_MINALIGN) {
2110 ralign = ARCH_SLAB_MINALIGN;
2111 if (ralign > BYTES_PER_WORD)
2112 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2114 /* 3) caller mandated alignment: disables debug if necessary */
2115 if (ralign < align) {
2117 if (ralign > BYTES_PER_WORD)
2118 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2125 /* Get cache's description obj. */
2126 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2131 cachep->obj_size = size;
2134 * Both debugging options require word-alignment which is calculated
2137 if (flags & SLAB_RED_ZONE) {
2138 /* add space for red zone words */
2139 cachep->obj_offset += BYTES_PER_WORD;
2140 size += 2 * BYTES_PER_WORD;
2142 if (flags & SLAB_STORE_USER) {
2143 /* user store requires one word storage behind the end of
2146 size += BYTES_PER_WORD;
2148 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2149 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2150 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2151 cachep->obj_offset += PAGE_SIZE - size;
2158 * Determine if the slab management is 'on' or 'off' slab.
2159 * (bootstrapping cannot cope with offslab caches so don't do
2162 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2164 * Size is large, assume best to place the slab management obj
2165 * off-slab (should allow better packing of objs).
2167 flags |= CFLGS_OFF_SLAB;
2169 size = ALIGN(size, align);
2171 left_over = calculate_slab_order(cachep, size, align, flags);
2174 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2175 kmem_cache_free(&cache_cache, cachep);
2179 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2180 + sizeof(struct slab), align);
2183 * If the slab has been placed off-slab, and we have enough space then
2184 * move it on-slab. This is at the expense of any extra colouring.
2186 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2187 flags &= ~CFLGS_OFF_SLAB;
2188 left_over -= slab_size;
2191 if (flags & CFLGS_OFF_SLAB) {
2192 /* really off slab. No need for manual alignment */
2194 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2197 cachep->colour_off = cache_line_size();
2198 /* Offset must be a multiple of the alignment. */
2199 if (cachep->colour_off < align)
2200 cachep->colour_off = align;
2201 cachep->colour = left_over / cachep->colour_off;
2202 cachep->slab_size = slab_size;
2203 cachep->flags = flags;
2204 cachep->gfpflags = 0;
2205 if (flags & SLAB_CACHE_DMA)
2206 cachep->gfpflags |= GFP_DMA;
2207 cachep->buffer_size = size;
2209 if (flags & CFLGS_OFF_SLAB)
2210 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2211 cachep->ctor = ctor;
2212 cachep->dtor = dtor;
2213 cachep->name = name;
2216 setup_cpu_cache(cachep);
2218 /* cache setup completed, link it into the list */
2219 list_add(&cachep->next, &cache_chain);
2221 if (!cachep && (flags & SLAB_PANIC))
2222 panic("kmem_cache_create(): failed to create slab `%s'\n",
2224 mutex_unlock(&cache_chain_mutex);
2225 unlock_cpu_hotplug();
2228 EXPORT_SYMBOL(kmem_cache_create);
2231 static void check_irq_off(void)
2233 BUG_ON(!irqs_disabled());
2236 static void check_irq_on(void)
2238 BUG_ON(irqs_disabled());
2241 static void check_spinlock_acquired(struct kmem_cache *cachep)
2245 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2249 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2253 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2258 #define check_irq_off() do { } while(0)
2259 #define check_irq_on() do { } while(0)
2260 #define check_spinlock_acquired(x) do { } while(0)
2261 #define check_spinlock_acquired_node(x, y) do { } while(0)
2264 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2265 struct array_cache *ac,
2266 int force, int node);
2268 static void do_drain(void *arg)
2270 struct kmem_cache *cachep = arg;
2271 struct array_cache *ac;
2272 int node = numa_node_id();
2275 ac = cpu_cache_get(cachep);
2276 spin_lock(&cachep->nodelists[node]->list_lock);
2277 free_block(cachep, ac->entry, ac->avail, node);
2278 spin_unlock(&cachep->nodelists[node]->list_lock);
2282 static void drain_cpu_caches(struct kmem_cache *cachep)
2284 struct kmem_list3 *l3;
2287 on_each_cpu(do_drain, cachep, 1, 1);
2289 for_each_online_node(node) {
2290 l3 = cachep->nodelists[node];
2291 if (l3 && l3->alien)
2292 drain_alien_cache(cachep, l3->alien);
2295 for_each_online_node(node) {
2296 l3 = cachep->nodelists[node];
2298 drain_array(cachep, l3, l3->shared, 1, node);
2303 * Remove slabs from the list of free slabs.
2304 * Specify the number of slabs to drain in tofree.
2306 * Returns the actual number of slabs released.
2308 static int drain_freelist(struct kmem_cache *cache,
2309 struct kmem_list3 *l3, int tofree)
2311 struct list_head *p;
2316 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2318 spin_lock_irq(&l3->list_lock);
2319 p = l3->slabs_free.prev;
2320 if (p == &l3->slabs_free) {
2321 spin_unlock_irq(&l3->list_lock);
2325 slabp = list_entry(p, struct slab, list);
2327 BUG_ON(slabp->inuse);
2329 list_del(&slabp->list);
2331 * Safe to drop the lock. The slab is no longer linked
2334 l3->free_objects -= cache->num;
2335 spin_unlock_irq(&l3->list_lock);
2336 slab_destroy(cache, slabp);
2343 static int __cache_shrink(struct kmem_cache *cachep)
2346 struct kmem_list3 *l3;
2348 drain_cpu_caches(cachep);
2351 for_each_online_node(i) {
2352 l3 = cachep->nodelists[i];
2356 drain_freelist(cachep, l3, l3->free_objects);
2358 ret += !list_empty(&l3->slabs_full) ||
2359 !list_empty(&l3->slabs_partial);
2361 return (ret ? 1 : 0);
2365 * kmem_cache_shrink - Shrink a cache.
2366 * @cachep: The cache to shrink.
2368 * Releases as many slabs as possible for a cache.
2369 * To help debugging, a zero exit status indicates all slabs were released.
2371 int kmem_cache_shrink(struct kmem_cache *cachep)
2373 BUG_ON(!cachep || in_interrupt());
2375 return __cache_shrink(cachep);
2377 EXPORT_SYMBOL(kmem_cache_shrink);
2380 * kmem_cache_destroy - delete a cache
2381 * @cachep: the cache to destroy
2383 * Remove a struct kmem_cache object from the slab cache.
2384 * Returns 0 on success.
2386 * It is expected this function will be called by a module when it is
2387 * unloaded. This will remove the cache completely, and avoid a duplicate
2388 * cache being allocated each time a module is loaded and unloaded, if the
2389 * module doesn't have persistent in-kernel storage across loads and unloads.
2391 * The cache must be empty before calling this function.
2393 * The caller must guarantee that noone will allocate memory from the cache
2394 * during the kmem_cache_destroy().
2396 int kmem_cache_destroy(struct kmem_cache *cachep)
2399 struct kmem_list3 *l3;
2401 BUG_ON(!cachep || in_interrupt());
2403 /* Don't let CPUs to come and go */
2406 /* Find the cache in the chain of caches. */
2407 mutex_lock(&cache_chain_mutex);
2409 * the chain is never empty, cache_cache is never destroyed
2411 list_del(&cachep->next);
2412 mutex_unlock(&cache_chain_mutex);
2414 if (__cache_shrink(cachep)) {
2415 slab_error(cachep, "Can't free all objects");
2416 mutex_lock(&cache_chain_mutex);
2417 list_add(&cachep->next, &cache_chain);
2418 mutex_unlock(&cache_chain_mutex);
2419 unlock_cpu_hotplug();
2423 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2426 for_each_online_cpu(i)
2427 kfree(cachep->array[i]);
2429 /* NUMA: free the list3 structures */
2430 for_each_online_node(i) {
2431 l3 = cachep->nodelists[i];
2434 free_alien_cache(l3->alien);
2438 kmem_cache_free(&cache_cache, cachep);
2439 unlock_cpu_hotplug();
2442 EXPORT_SYMBOL(kmem_cache_destroy);
2444 /* Get the memory for a slab management obj. */
2445 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2446 int colour_off, gfp_t local_flags,
2451 if (OFF_SLAB(cachep)) {
2452 /* Slab management obj is off-slab. */
2453 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2454 local_flags, nodeid);
2458 slabp = objp + colour_off;
2459 colour_off += cachep->slab_size;
2462 slabp->colouroff = colour_off;
2463 slabp->s_mem = objp + colour_off;
2464 slabp->nodeid = nodeid;
2468 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2470 return (kmem_bufctl_t *) (slabp + 1);
2473 static void cache_init_objs(struct kmem_cache *cachep,
2474 struct slab *slabp, unsigned long ctor_flags)
2478 for (i = 0; i < cachep->num; i++) {
2479 void *objp = index_to_obj(cachep, slabp, i);
2481 /* need to poison the objs? */
2482 if (cachep->flags & SLAB_POISON)
2483 poison_obj(cachep, objp, POISON_FREE);
2484 if (cachep->flags & SLAB_STORE_USER)
2485 *dbg_userword(cachep, objp) = NULL;
2487 if (cachep->flags & SLAB_RED_ZONE) {
2488 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2489 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2492 * Constructors are not allowed to allocate memory from the same
2493 * cache which they are a constructor for. Otherwise, deadlock.
2494 * They must also be threaded.
2496 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2497 cachep->ctor(objp + obj_offset(cachep), cachep,
2500 if (cachep->flags & SLAB_RED_ZONE) {
2501 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2502 slab_error(cachep, "constructor overwrote the"
2503 " end of an object");
2504 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2505 slab_error(cachep, "constructor overwrote the"
2506 " start of an object");
2508 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2509 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2510 kernel_map_pages(virt_to_page(objp),
2511 cachep->buffer_size / PAGE_SIZE, 0);
2514 cachep->ctor(objp, cachep, ctor_flags);
2516 slab_bufctl(slabp)[i] = i + 1;
2518 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2522 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2524 if (flags & SLAB_DMA)
2525 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2527 BUG_ON(cachep->gfpflags & GFP_DMA);
2530 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2533 void *objp = index_to_obj(cachep, slabp, slabp->free);
2537 next = slab_bufctl(slabp)[slabp->free];
2539 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2540 WARN_ON(slabp->nodeid != nodeid);
2547 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2548 void *objp, int nodeid)
2550 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2553 /* Verify that the slab belongs to the intended node */
2554 WARN_ON(slabp->nodeid != nodeid);
2556 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2557 printk(KERN_ERR "slab: double free detected in cache "
2558 "'%s', objp %p\n", cachep->name, objp);
2562 slab_bufctl(slabp)[objnr] = slabp->free;
2563 slabp->free = objnr;
2568 * Map pages beginning at addr to the given cache and slab. This is required
2569 * for the slab allocator to be able to lookup the cache and slab of a
2570 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2572 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2578 page = virt_to_page(addr);
2581 if (likely(!PageCompound(page)))
2582 nr_pages <<= cache->gfporder;
2585 page_set_cache(page, cache);
2586 page_set_slab(page, slab);
2588 } while (--nr_pages);
2592 * Grow (by 1) the number of slabs within a cache. This is called by
2593 * kmem_cache_alloc() when there are no active objs left in a cache.
2595 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2601 unsigned long ctor_flags;
2602 struct kmem_list3 *l3;
2605 * Be lazy and only check for valid flags here, keeping it out of the
2606 * critical path in kmem_cache_alloc().
2608 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2609 if (flags & SLAB_NO_GROW)
2612 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2613 local_flags = (flags & SLAB_LEVEL_MASK);
2614 if (!(local_flags & __GFP_WAIT))
2616 * Not allowed to sleep. Need to tell a constructor about
2617 * this - it might need to know...
2619 ctor_flags |= SLAB_CTOR_ATOMIC;
2621 /* Take the l3 list lock to change the colour_next on this node */
2623 l3 = cachep->nodelists[nodeid];
2624 spin_lock(&l3->list_lock);
2626 /* Get colour for the slab, and cal the next value. */
2627 offset = l3->colour_next;
2629 if (l3->colour_next >= cachep->colour)
2630 l3->colour_next = 0;
2631 spin_unlock(&l3->list_lock);
2633 offset *= cachep->colour_off;
2635 if (local_flags & __GFP_WAIT)
2639 * The test for missing atomic flag is performed here, rather than
2640 * the more obvious place, simply to reduce the critical path length
2641 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2642 * will eventually be caught here (where it matters).
2644 kmem_flagcheck(cachep, flags);
2647 * Get mem for the objs. Attempt to allocate a physical page from
2650 objp = kmem_getpages(cachep, flags, nodeid);
2654 /* Get slab management. */
2655 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2659 slabp->nodeid = nodeid;
2660 slab_map_pages(cachep, slabp, objp);
2662 cache_init_objs(cachep, slabp, ctor_flags);
2664 if (local_flags & __GFP_WAIT)
2665 local_irq_disable();
2667 spin_lock(&l3->list_lock);
2669 /* Make slab active. */
2670 list_add_tail(&slabp->list, &(l3->slabs_free));
2671 STATS_INC_GROWN(cachep);
2672 l3->free_objects += cachep->num;
2673 spin_unlock(&l3->list_lock);
2676 kmem_freepages(cachep, objp);
2678 if (local_flags & __GFP_WAIT)
2679 local_irq_disable();
2686 * Perform extra freeing checks:
2687 * - detect bad pointers.
2688 * - POISON/RED_ZONE checking
2689 * - destructor calls, for caches with POISON+dtor
2691 static void kfree_debugcheck(const void *objp)
2695 if (!virt_addr_valid(objp)) {
2696 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2697 (unsigned long)objp);
2700 page = virt_to_page(objp);
2701 if (!PageSlab(page)) {
2702 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2703 (unsigned long)objp);
2708 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2710 unsigned long redzone1, redzone2;
2712 redzone1 = *dbg_redzone1(cache, obj);
2713 redzone2 = *dbg_redzone2(cache, obj);
2718 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2721 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2722 slab_error(cache, "double free detected");
2724 slab_error(cache, "memory outside object was overwritten");
2726 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2727 obj, redzone1, redzone2);
2730 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2737 objp -= obj_offset(cachep);
2738 kfree_debugcheck(objp);
2739 page = virt_to_page(objp);
2741 slabp = page_get_slab(page);
2743 if (cachep->flags & SLAB_RED_ZONE) {
2744 verify_redzone_free(cachep, objp);
2745 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2746 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2748 if (cachep->flags & SLAB_STORE_USER)
2749 *dbg_userword(cachep, objp) = caller;
2751 objnr = obj_to_index(cachep, slabp, objp);
2753 BUG_ON(objnr >= cachep->num);
2754 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2756 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2758 * Need to call the slab's constructor so the caller can
2759 * perform a verify of its state (debugging). Called without
2760 * the cache-lock held.
2762 cachep->ctor(objp + obj_offset(cachep),
2763 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2765 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2766 /* we want to cache poison the object,
2767 * call the destruction callback
2769 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2771 #ifdef CONFIG_DEBUG_SLAB_LEAK
2772 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2774 if (cachep->flags & SLAB_POISON) {
2775 #ifdef CONFIG_DEBUG_PAGEALLOC
2776 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2777 store_stackinfo(cachep, objp, (unsigned long)caller);
2778 kernel_map_pages(virt_to_page(objp),
2779 cachep->buffer_size / PAGE_SIZE, 0);
2781 poison_obj(cachep, objp, POISON_FREE);
2784 poison_obj(cachep, objp, POISON_FREE);
2790 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2795 /* Check slab's freelist to see if this obj is there. */
2796 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2798 if (entries > cachep->num || i >= cachep->num)
2801 if (entries != cachep->num - slabp->inuse) {
2803 printk(KERN_ERR "slab: Internal list corruption detected in "
2804 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2805 cachep->name, cachep->num, slabp, slabp->inuse);
2807 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2810 printk("\n%03x:", i);
2811 printk(" %02x", ((unsigned char *)slabp)[i]);
2818 #define kfree_debugcheck(x) do { } while(0)
2819 #define cache_free_debugcheck(x,objp,z) (objp)
2820 #define check_slabp(x,y) do { } while(0)
2823 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2826 struct kmem_list3 *l3;
2827 struct array_cache *ac;
2830 ac = cpu_cache_get(cachep);
2832 batchcount = ac->batchcount;
2833 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2835 * If there was little recent activity on this cache, then
2836 * perform only a partial refill. Otherwise we could generate
2839 batchcount = BATCHREFILL_LIMIT;
2841 l3 = cachep->nodelists[numa_node_id()];
2843 BUG_ON(ac->avail > 0 || !l3);
2844 spin_lock(&l3->list_lock);
2846 /* See if we can refill from the shared array */
2847 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2850 while (batchcount > 0) {
2851 struct list_head *entry;
2853 /* Get slab alloc is to come from. */
2854 entry = l3->slabs_partial.next;
2855 if (entry == &l3->slabs_partial) {
2856 l3->free_touched = 1;
2857 entry = l3->slabs_free.next;
2858 if (entry == &l3->slabs_free)
2862 slabp = list_entry(entry, struct slab, list);
2863 check_slabp(cachep, slabp);
2864 check_spinlock_acquired(cachep);
2865 while (slabp->inuse < cachep->num && batchcount--) {
2866 STATS_INC_ALLOCED(cachep);
2867 STATS_INC_ACTIVE(cachep);
2868 STATS_SET_HIGH(cachep);
2870 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2873 check_slabp(cachep, slabp);
2875 /* move slabp to correct slabp list: */
2876 list_del(&slabp->list);
2877 if (slabp->free == BUFCTL_END)
2878 list_add(&slabp->list, &l3->slabs_full);
2880 list_add(&slabp->list, &l3->slabs_partial);
2884 l3->free_objects -= ac->avail;
2886 spin_unlock(&l3->list_lock);
2888 if (unlikely(!ac->avail)) {
2890 x = cache_grow(cachep, flags, numa_node_id());
2892 /* cache_grow can reenable interrupts, then ac could change. */
2893 ac = cpu_cache_get(cachep);
2894 if (!x && ac->avail == 0) /* no objects in sight? abort */
2897 if (!ac->avail) /* objects refilled by interrupt? */
2901 return ac->entry[--ac->avail];
2904 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2907 might_sleep_if(flags & __GFP_WAIT);
2909 kmem_flagcheck(cachep, flags);
2914 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2915 gfp_t flags, void *objp, void *caller)
2919 if (cachep->flags & SLAB_POISON) {
2920 #ifdef CONFIG_DEBUG_PAGEALLOC
2921 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2922 kernel_map_pages(virt_to_page(objp),
2923 cachep->buffer_size / PAGE_SIZE, 1);
2925 check_poison_obj(cachep, objp);
2927 check_poison_obj(cachep, objp);
2929 poison_obj(cachep, objp, POISON_INUSE);
2931 if (cachep->flags & SLAB_STORE_USER)
2932 *dbg_userword(cachep, objp) = caller;
2934 if (cachep->flags & SLAB_RED_ZONE) {
2935 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2936 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2937 slab_error(cachep, "double free, or memory outside"
2938 " object was overwritten");
2940 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2941 objp, *dbg_redzone1(cachep, objp),
2942 *dbg_redzone2(cachep, objp));
2944 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2945 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2947 #ifdef CONFIG_DEBUG_SLAB_LEAK
2952 slabp = page_get_slab(virt_to_page(objp));
2953 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2954 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2957 objp += obj_offset(cachep);
2958 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2959 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2961 if (!(flags & __GFP_WAIT))
2962 ctor_flags |= SLAB_CTOR_ATOMIC;
2964 cachep->ctor(objp, cachep, ctor_flags);
2969 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2972 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2975 struct array_cache *ac;
2978 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
2979 objp = alternate_node_alloc(cachep, flags);
2986 ac = cpu_cache_get(cachep);
2987 if (likely(ac->avail)) {
2988 STATS_INC_ALLOCHIT(cachep);
2990 objp = ac->entry[--ac->avail];
2992 STATS_INC_ALLOCMISS(cachep);
2993 objp = cache_alloc_refill(cachep, flags);
2998 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2999 gfp_t flags, void *caller)
3001 unsigned long save_flags;
3004 cache_alloc_debugcheck_before(cachep, flags);
3006 local_irq_save(save_flags);
3007 objp = ____cache_alloc(cachep, flags);
3008 local_irq_restore(save_flags);
3009 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3017 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3019 * If we are in_interrupt, then process context, including cpusets and
3020 * mempolicy, may not apply and should not be used for allocation policy.
3022 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3024 int nid_alloc, nid_here;
3028 nid_alloc = nid_here = numa_node_id();
3029 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3030 nid_alloc = cpuset_mem_spread_node();
3031 else if (current->mempolicy)
3032 nid_alloc = slab_node(current->mempolicy);
3033 if (nid_alloc != nid_here)
3034 return __cache_alloc_node(cachep, flags, nid_alloc);
3039 * A interface to enable slab creation on nodeid
3041 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3044 struct list_head *entry;
3046 struct kmem_list3 *l3;
3050 l3 = cachep->nodelists[nodeid];
3055 spin_lock(&l3->list_lock);
3056 entry = l3->slabs_partial.next;
3057 if (entry == &l3->slabs_partial) {
3058 l3->free_touched = 1;
3059 entry = l3->slabs_free.next;
3060 if (entry == &l3->slabs_free)
3064 slabp = list_entry(entry, struct slab, list);
3065 check_spinlock_acquired_node(cachep, nodeid);
3066 check_slabp(cachep, slabp);
3068 STATS_INC_NODEALLOCS(cachep);
3069 STATS_INC_ACTIVE(cachep);
3070 STATS_SET_HIGH(cachep);
3072 BUG_ON(slabp->inuse == cachep->num);
3074 obj = slab_get_obj(cachep, slabp, nodeid);
3075 check_slabp(cachep, slabp);
3077 /* move slabp to correct slabp list: */
3078 list_del(&slabp->list);
3080 if (slabp->free == BUFCTL_END)
3081 list_add(&slabp->list, &l3->slabs_full);
3083 list_add(&slabp->list, &l3->slabs_partial);
3085 spin_unlock(&l3->list_lock);
3089 spin_unlock(&l3->list_lock);
3090 x = cache_grow(cachep, flags, nodeid);
3102 * Caller needs to acquire correct kmem_list's list_lock
3104 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3108 struct kmem_list3 *l3;
3110 for (i = 0; i < nr_objects; i++) {
3111 void *objp = objpp[i];
3114 slabp = virt_to_slab(objp);
3115 l3 = cachep->nodelists[node];
3116 list_del(&slabp->list);
3117 check_spinlock_acquired_node(cachep, node);
3118 check_slabp(cachep, slabp);
3119 slab_put_obj(cachep, slabp, objp, node);
3120 STATS_DEC_ACTIVE(cachep);
3122 check_slabp(cachep, slabp);
3124 /* fixup slab chains */
3125 if (slabp->inuse == 0) {
3126 if (l3->free_objects > l3->free_limit) {
3127 l3->free_objects -= cachep->num;
3128 slab_destroy(cachep, slabp);
3130 list_add(&slabp->list, &l3->slabs_free);
3133 /* Unconditionally move a slab to the end of the
3134 * partial list on free - maximum time for the
3135 * other objects to be freed, too.
3137 list_add_tail(&slabp->list, &l3->slabs_partial);
3142 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3145 struct kmem_list3 *l3;
3146 int node = numa_node_id();
3148 batchcount = ac->batchcount;
3150 BUG_ON(!batchcount || batchcount > ac->avail);
3153 l3 = cachep->nodelists[node];
3154 spin_lock(&l3->list_lock);
3156 struct array_cache *shared_array = l3->shared;
3157 int max = shared_array->limit - shared_array->avail;
3159 if (batchcount > max)
3161 memcpy(&(shared_array->entry[shared_array->avail]),
3162 ac->entry, sizeof(void *) * batchcount);
3163 shared_array->avail += batchcount;
3168 free_block(cachep, ac->entry, batchcount, node);
3173 struct list_head *p;
3175 p = l3->slabs_free.next;
3176 while (p != &(l3->slabs_free)) {
3179 slabp = list_entry(p, struct slab, list);
3180 BUG_ON(slabp->inuse);
3185 STATS_SET_FREEABLE(cachep, i);
3188 spin_unlock(&l3->list_lock);
3189 ac->avail -= batchcount;
3190 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3194 * Release an obj back to its cache. If the obj has a constructed state, it must
3195 * be in this state _before_ it is released. Called with disabled ints.
3197 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3199 struct array_cache *ac = cpu_cache_get(cachep);
3202 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3204 if (cache_free_alien(cachep, objp))
3207 if (likely(ac->avail < ac->limit)) {
3208 STATS_INC_FREEHIT(cachep);
3209 ac->entry[ac->avail++] = objp;
3212 STATS_INC_FREEMISS(cachep);
3213 cache_flusharray(cachep, ac);
3214 ac->entry[ac->avail++] = objp;
3219 * kmem_cache_alloc - Allocate an object
3220 * @cachep: The cache to allocate from.
3221 * @flags: See kmalloc().
3223 * Allocate an object from this cache. The flags are only relevant
3224 * if the cache has no available objects.
3226 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3228 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3230 EXPORT_SYMBOL(kmem_cache_alloc);
3233 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3234 * @cache: The cache to allocate from.
3235 * @flags: See kmalloc().
3237 * Allocate an object from this cache and set the allocated memory to zero.
3238 * The flags are only relevant if the cache has no available objects.
3240 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3242 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3244 memset(ret, 0, obj_size(cache));
3247 EXPORT_SYMBOL(kmem_cache_zalloc);
3250 * kmem_ptr_validate - check if an untrusted pointer might
3252 * @cachep: the cache we're checking against
3253 * @ptr: pointer to validate
3255 * This verifies that the untrusted pointer looks sane:
3256 * it is _not_ a guarantee that the pointer is actually
3257 * part of the slab cache in question, but it at least
3258 * validates that the pointer can be dereferenced and
3259 * looks half-way sane.
3261 * Currently only used for dentry validation.
3263 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3265 unsigned long addr = (unsigned long)ptr;
3266 unsigned long min_addr = PAGE_OFFSET;
3267 unsigned long align_mask = BYTES_PER_WORD - 1;
3268 unsigned long size = cachep->buffer_size;
3271 if (unlikely(addr < min_addr))
3273 if (unlikely(addr > (unsigned long)high_memory - size))
3275 if (unlikely(addr & align_mask))
3277 if (unlikely(!kern_addr_valid(addr)))
3279 if (unlikely(!kern_addr_valid(addr + size - 1)))
3281 page = virt_to_page(ptr);
3282 if (unlikely(!PageSlab(page)))
3284 if (unlikely(page_get_cache(page) != cachep))
3293 * kmem_cache_alloc_node - Allocate an object on the specified node
3294 * @cachep: The cache to allocate from.
3295 * @flags: See kmalloc().
3296 * @nodeid: node number of the target node.
3298 * Identical to kmem_cache_alloc, except that this function is slow
3299 * and can sleep. And it will allocate memory on the given node, which
3300 * can improve the performance for cpu bound structures.
3301 * New and improved: it will now make sure that the object gets
3302 * put on the correct node list so that there is no false sharing.
3304 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3306 unsigned long save_flags;
3309 cache_alloc_debugcheck_before(cachep, flags);
3310 local_irq_save(save_flags);
3312 if (nodeid == -1 || nodeid == numa_node_id() ||
3313 !cachep->nodelists[nodeid])
3314 ptr = ____cache_alloc(cachep, flags);
3316 ptr = __cache_alloc_node(cachep, flags, nodeid);
3317 local_irq_restore(save_flags);
3319 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3320 __builtin_return_address(0));
3324 EXPORT_SYMBOL(kmem_cache_alloc_node);
3326 void *kmalloc_node(size_t size, gfp_t flags, int node)
3328 struct kmem_cache *cachep;
3330 cachep = kmem_find_general_cachep(size, flags);
3331 if (unlikely(cachep == NULL))
3333 return kmem_cache_alloc_node(cachep, flags, node);
3335 EXPORT_SYMBOL(kmalloc_node);
3339 * __do_kmalloc - allocate memory
3340 * @size: how many bytes of memory are required.
3341 * @flags: the type of memory to allocate (see kmalloc).
3342 * @caller: function caller for debug tracking of the caller
3344 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3347 struct kmem_cache *cachep;
3349 /* If you want to save a few bytes .text space: replace
3351 * Then kmalloc uses the uninlined functions instead of the inline
3354 cachep = __find_general_cachep(size, flags);
3355 if (unlikely(cachep == NULL))
3357 return __cache_alloc(cachep, flags, caller);
3361 void *__kmalloc(size_t size, gfp_t flags)
3363 #ifndef CONFIG_DEBUG_SLAB
3364 return __do_kmalloc(size, flags, NULL);
3366 return __do_kmalloc(size, flags, __builtin_return_address(0));
3369 EXPORT_SYMBOL(__kmalloc);
3371 #ifdef CONFIG_DEBUG_SLAB
3372 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3374 return __do_kmalloc(size, flags, caller);
3376 EXPORT_SYMBOL(__kmalloc_track_caller);
3381 * percpu_depopulate - depopulate per-cpu data for given cpu
3382 * @__pdata: per-cpu data to depopulate
3383 * @cpu: depopulate per-cpu data for this cpu
3385 * Depopulating per-cpu data for a cpu going offline would be a typical
3386 * use case. You need to register a cpu hotplug handler for that purpose.
3388 void percpu_depopulate(void *__pdata, int cpu)
3390 struct percpu_data *pdata = __percpu_disguise(__pdata);
3391 if (pdata->ptrs[cpu]) {
3392 kfree(pdata->ptrs[cpu]);
3393 pdata->ptrs[cpu] = NULL;
3396 EXPORT_SYMBOL_GPL(percpu_depopulate);
3399 * percpu_depopulate_mask - depopulate per-cpu data for some cpu's
3400 * @__pdata: per-cpu data to depopulate
3401 * @mask: depopulate per-cpu data for cpu's selected through mask bits
3403 void __percpu_depopulate_mask(void *__pdata, cpumask_t *mask)
3406 for_each_cpu_mask(cpu, *mask)
3407 percpu_depopulate(__pdata, cpu);
3409 EXPORT_SYMBOL_GPL(__percpu_depopulate_mask);
3412 * percpu_populate - populate per-cpu data for given cpu
3413 * @__pdata: per-cpu data to populate further
3414 * @size: size of per-cpu object
3415 * @gfp: may sleep or not etc.
3416 * @cpu: populate per-data for this cpu
3418 * Populating per-cpu data for a cpu coming online would be a typical
3419 * use case. You need to register a cpu hotplug handler for that purpose.
3420 * Per-cpu object is populated with zeroed buffer.
3422 void *percpu_populate(void *__pdata, size_t size, gfp_t gfp, int cpu)
3424 struct percpu_data *pdata = __percpu_disguise(__pdata);
3425 int node = cpu_to_node(cpu);
3427 BUG_ON(pdata->ptrs[cpu]);
3428 if (node_online(node)) {
3429 /* FIXME: kzalloc_node(size, gfp, node) */
3430 pdata->ptrs[cpu] = kmalloc_node(size, gfp, node);
3431 if (pdata->ptrs[cpu])
3432 memset(pdata->ptrs[cpu], 0, size);
3434 pdata->ptrs[cpu] = kzalloc(size, gfp);
3435 return pdata->ptrs[cpu];
3437 EXPORT_SYMBOL_GPL(percpu_populate);
3440 * percpu_populate_mask - populate per-cpu data for more cpu's
3441 * @__pdata: per-cpu data to populate further
3442 * @size: size of per-cpu object
3443 * @gfp: may sleep or not etc.
3444 * @mask: populate per-cpu data for cpu's selected through mask bits
3446 * Per-cpu objects are populated with zeroed buffers.
3448 int __percpu_populate_mask(void *__pdata, size_t size, gfp_t gfp,
3451 cpumask_t populated = CPU_MASK_NONE;
3454 for_each_cpu_mask(cpu, *mask)
3455 if (unlikely(!percpu_populate(__pdata, size, gfp, cpu))) {
3456 __percpu_depopulate_mask(__pdata, &populated);
3459 cpu_set(cpu, populated);
3462 EXPORT_SYMBOL_GPL(__percpu_populate_mask);
3465 * percpu_alloc_mask - initial setup of per-cpu data
3466 * @size: size of per-cpu object
3467 * @gfp: may sleep or not etc.
3468 * @mask: populate per-data for cpu's selected through mask bits
3470 * Populating per-cpu data for all online cpu's would be a typical use case,
3471 * which is simplified by the percpu_alloc() wrapper.
3472 * Per-cpu objects are populated with zeroed buffers.
3474 void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask)
3476 void *pdata = kzalloc(sizeof(struct percpu_data), gfp);
3477 void *__pdata = __percpu_disguise(pdata);
3479 if (unlikely(!pdata))
3481 if (likely(!__percpu_populate_mask(__pdata, size, gfp, mask)))
3486 EXPORT_SYMBOL_GPL(__percpu_alloc_mask);
3489 * percpu_free - final cleanup of per-cpu data
3490 * @__pdata: object to clean up
3492 * We simply clean up any per-cpu object left. No need for the client to
3493 * track and specify through a bis mask which per-cpu objects are to free.
3495 void percpu_free(void *__pdata)
3497 __percpu_depopulate_mask(__pdata, &cpu_possible_map);
3498 kfree(__percpu_disguise(__pdata));
3500 EXPORT_SYMBOL_GPL(percpu_free);
3501 #endif /* CONFIG_SMP */
3504 * kmem_cache_free - Deallocate an object
3505 * @cachep: The cache the allocation was from.
3506 * @objp: The previously allocated object.
3508 * Free an object which was previously allocated from this
3511 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3513 unsigned long flags;
3515 BUG_ON(virt_to_cache(objp) != cachep);
3517 local_irq_save(flags);
3518 __cache_free(cachep, objp);
3519 local_irq_restore(flags);
3521 EXPORT_SYMBOL(kmem_cache_free);
3524 * kfree - free previously allocated memory
3525 * @objp: pointer returned by kmalloc.
3527 * If @objp is NULL, no operation is performed.
3529 * Don't free memory not originally allocated by kmalloc()
3530 * or you will run into trouble.
3532 void kfree(const void *objp)
3534 struct kmem_cache *c;
3535 unsigned long flags;
3537 if (unlikely(!objp))
3539 local_irq_save(flags);
3540 kfree_debugcheck(objp);
3541 c = virt_to_cache(objp);
3542 debug_check_no_locks_freed(objp, obj_size(c));
3543 __cache_free(c, (void *)objp);
3544 local_irq_restore(flags);
3546 EXPORT_SYMBOL(kfree);
3548 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3550 return obj_size(cachep);
3552 EXPORT_SYMBOL(kmem_cache_size);
3554 const char *kmem_cache_name(struct kmem_cache *cachep)
3556 return cachep->name;
3558 EXPORT_SYMBOL_GPL(kmem_cache_name);
3561 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3563 static int alloc_kmemlist(struct kmem_cache *cachep)
3566 struct kmem_list3 *l3;
3567 struct array_cache *new_shared;
3568 struct array_cache **new_alien;
3570 for_each_online_node(node) {
3572 new_alien = alloc_alien_cache(node, cachep->limit);
3576 new_shared = alloc_arraycache(node,
3577 cachep->shared*cachep->batchcount,
3580 free_alien_cache(new_alien);
3584 l3 = cachep->nodelists[node];
3586 struct array_cache *shared = l3->shared;
3588 spin_lock_irq(&l3->list_lock);
3591 free_block(cachep, shared->entry,
3592 shared->avail, node);
3594 l3->shared = new_shared;
3596 l3->alien = new_alien;
3599 l3->free_limit = (1 + nr_cpus_node(node)) *
3600 cachep->batchcount + cachep->num;
3601 spin_unlock_irq(&l3->list_lock);
3603 free_alien_cache(new_alien);
3606 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3608 free_alien_cache(new_alien);
3613 kmem_list3_init(l3);
3614 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3615 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3616 l3->shared = new_shared;
3617 l3->alien = new_alien;
3618 l3->free_limit = (1 + nr_cpus_node(node)) *
3619 cachep->batchcount + cachep->num;
3620 cachep->nodelists[node] = l3;
3625 if (!cachep->next.next) {
3626 /* Cache is not active yet. Roll back what we did */
3629 if (cachep->nodelists[node]) {
3630 l3 = cachep->nodelists[node];
3633 free_alien_cache(l3->alien);
3635 cachep->nodelists[node] = NULL;
3643 struct ccupdate_struct {
3644 struct kmem_cache *cachep;
3645 struct array_cache *new[NR_CPUS];
3648 static void do_ccupdate_local(void *info)
3650 struct ccupdate_struct *new = info;
3651 struct array_cache *old;
3654 old = cpu_cache_get(new->cachep);
3656 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3657 new->new[smp_processor_id()] = old;
3660 /* Always called with the cache_chain_mutex held */
3661 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3662 int batchcount, int shared)
3664 struct ccupdate_struct new;
3667 memset(&new.new, 0, sizeof(new.new));
3668 for_each_online_cpu(i) {
3669 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3672 for (i--; i >= 0; i--)
3677 new.cachep = cachep;
3679 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3682 cachep->batchcount = batchcount;
3683 cachep->limit = limit;
3684 cachep->shared = shared;
3686 for_each_online_cpu(i) {
3687 struct array_cache *ccold = new.new[i];
3690 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3691 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3692 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3696 err = alloc_kmemlist(cachep);
3698 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3699 cachep->name, -err);
3705 /* Called with cache_chain_mutex held always */
3706 static void enable_cpucache(struct kmem_cache *cachep)
3712 * The head array serves three purposes:
3713 * - create a LIFO ordering, i.e. return objects that are cache-warm
3714 * - reduce the number of spinlock operations.
3715 * - reduce the number of linked list operations on the slab and
3716 * bufctl chains: array operations are cheaper.
3717 * The numbers are guessed, we should auto-tune as described by
3720 if (cachep->buffer_size > 131072)
3722 else if (cachep->buffer_size > PAGE_SIZE)
3724 else if (cachep->buffer_size > 1024)
3726 else if (cachep->buffer_size > 256)
3732 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3733 * allocation behaviour: Most allocs on one cpu, most free operations
3734 * on another cpu. For these cases, an efficient object passing between
3735 * cpus is necessary. This is provided by a shared array. The array
3736 * replaces Bonwick's magazine layer.
3737 * On uniprocessor, it's functionally equivalent (but less efficient)
3738 * to a larger limit. Thus disabled by default.
3742 if (cachep->buffer_size <= PAGE_SIZE)
3748 * With debugging enabled, large batchcount lead to excessively long
3749 * periods with disabled local interrupts. Limit the batchcount
3754 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3756 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3757 cachep->name, -err);
3761 * Drain an array if it contains any elements taking the l3 lock only if
3762 * necessary. Note that the l3 listlock also protects the array_cache
3763 * if drain_array() is used on the shared array.
3765 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3766 struct array_cache *ac, int force, int node)
3770 if (!ac || !ac->avail)
3772 if (ac->touched && !force) {
3775 spin_lock_irq(&l3->list_lock);
3777 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3778 if (tofree > ac->avail)
3779 tofree = (ac->avail + 1) / 2;
3780 free_block(cachep, ac->entry, tofree, node);
3781 ac->avail -= tofree;
3782 memmove(ac->entry, &(ac->entry[tofree]),
3783 sizeof(void *) * ac->avail);
3785 spin_unlock_irq(&l3->list_lock);
3790 * cache_reap - Reclaim memory from caches.
3791 * @unused: unused parameter
3793 * Called from workqueue/eventd every few seconds.
3795 * - clear the per-cpu caches for this CPU.
3796 * - return freeable pages to the main free memory pool.
3798 * If we cannot acquire the cache chain mutex then just give up - we'll try
3799 * again on the next iteration.
3801 static void cache_reap(void *unused)
3803 struct kmem_cache *searchp;
3804 struct kmem_list3 *l3;
3805 int node = numa_node_id();
3807 if (!mutex_trylock(&cache_chain_mutex)) {
3808 /* Give up. Setup the next iteration. */
3809 schedule_delayed_work(&__get_cpu_var(reap_work),
3814 list_for_each_entry(searchp, &cache_chain, next) {
3818 * We only take the l3 lock if absolutely necessary and we
3819 * have established with reasonable certainty that
3820 * we can do some work if the lock was obtained.
3822 l3 = searchp->nodelists[node];
3824 reap_alien(searchp, l3);
3826 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3829 * These are racy checks but it does not matter
3830 * if we skip one check or scan twice.
3832 if (time_after(l3->next_reap, jiffies))
3835 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3837 drain_array(searchp, l3, l3->shared, 0, node);
3839 if (l3->free_touched)
3840 l3->free_touched = 0;
3844 freed = drain_freelist(searchp, l3, (l3->free_limit +
3845 5 * searchp->num - 1) / (5 * searchp->num));
3846 STATS_ADD_REAPED(searchp, freed);
3852 mutex_unlock(&cache_chain_mutex);
3854 refresh_cpu_vm_stats(smp_processor_id());
3855 /* Set up the next iteration */
3856 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3859 #ifdef CONFIG_PROC_FS
3861 static void print_slabinfo_header(struct seq_file *m)
3864 * Output format version, so at least we can change it
3865 * without _too_ many complaints.
3868 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3870 seq_puts(m, "slabinfo - version: 2.1\n");
3872 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3873 "<objperslab> <pagesperslab>");
3874 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3875 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3877 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3878 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3879 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3884 static void *s_start(struct seq_file *m, loff_t *pos)
3887 struct list_head *p;
3889 mutex_lock(&cache_chain_mutex);
3891 print_slabinfo_header(m);
3892 p = cache_chain.next;
3895 if (p == &cache_chain)
3898 return list_entry(p, struct kmem_cache, next);
3901 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3903 struct kmem_cache *cachep = p;
3905 return cachep->next.next == &cache_chain ?
3906 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3909 static void s_stop(struct seq_file *m, void *p)
3911 mutex_unlock(&cache_chain_mutex);
3914 static int s_show(struct seq_file *m, void *p)
3916 struct kmem_cache *cachep = p;
3918 unsigned long active_objs;
3919 unsigned long num_objs;
3920 unsigned long active_slabs = 0;
3921 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3925 struct kmem_list3 *l3;
3929 for_each_online_node(node) {
3930 l3 = cachep->nodelists[node];
3935 spin_lock_irq(&l3->list_lock);
3937 list_for_each_entry(slabp, &l3->slabs_full, list) {
3938 if (slabp->inuse != cachep->num && !error)
3939 error = "slabs_full accounting error";
3940 active_objs += cachep->num;
3943 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3944 if (slabp->inuse == cachep->num && !error)
3945 error = "slabs_partial inuse accounting error";
3946 if (!slabp->inuse && !error)
3947 error = "slabs_partial/inuse accounting error";
3948 active_objs += slabp->inuse;
3951 list_for_each_entry(slabp, &l3->slabs_free, list) {
3952 if (slabp->inuse && !error)
3953 error = "slabs_free/inuse accounting error";
3956 free_objects += l3->free_objects;
3958 shared_avail += l3->shared->avail;
3960 spin_unlock_irq(&l3->list_lock);
3962 num_slabs += active_slabs;
3963 num_objs = num_slabs * cachep->num;
3964 if (num_objs - active_objs != free_objects && !error)
3965 error = "free_objects accounting error";
3967 name = cachep->name;
3969 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3971 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3972 name, active_objs, num_objs, cachep->buffer_size,
3973 cachep->num, (1 << cachep->gfporder));
3974 seq_printf(m, " : tunables %4u %4u %4u",
3975 cachep->limit, cachep->batchcount, cachep->shared);
3976 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3977 active_slabs, num_slabs, shared_avail);
3980 unsigned long high = cachep->high_mark;
3981 unsigned long allocs = cachep->num_allocations;
3982 unsigned long grown = cachep->grown;
3983 unsigned long reaped = cachep->reaped;
3984 unsigned long errors = cachep->errors;
3985 unsigned long max_freeable = cachep->max_freeable;
3986 unsigned long node_allocs = cachep->node_allocs;
3987 unsigned long node_frees = cachep->node_frees;
3988 unsigned long overflows = cachep->node_overflow;
3990 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3991 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3992 reaped, errors, max_freeable, node_allocs,
3993 node_frees, overflows);
3997 unsigned long allochit = atomic_read(&cachep->allochit);
3998 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3999 unsigned long freehit = atomic_read(&cachep->freehit);
4000 unsigned long freemiss = atomic_read(&cachep->freemiss);
4002 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4003 allochit, allocmiss, freehit, freemiss);
4011 * slabinfo_op - iterator that generates /proc/slabinfo
4020 * num-pages-per-slab
4021 * + further values on SMP and with statistics enabled
4024 struct seq_operations slabinfo_op = {
4031 #define MAX_SLABINFO_WRITE 128
4033 * slabinfo_write - Tuning for the slab allocator
4035 * @buffer: user buffer
4036 * @count: data length
4039 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4040 size_t count, loff_t *ppos)
4042 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4043 int limit, batchcount, shared, res;
4044 struct kmem_cache *cachep;
4046 if (count > MAX_SLABINFO_WRITE)
4048 if (copy_from_user(&kbuf, buffer, count))
4050 kbuf[MAX_SLABINFO_WRITE] = '\0';
4052 tmp = strchr(kbuf, ' ');
4057 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4060 /* Find the cache in the chain of caches. */
4061 mutex_lock(&cache_chain_mutex);
4063 list_for_each_entry(cachep, &cache_chain, next) {
4064 if (!strcmp(cachep->name, kbuf)) {
4065 if (limit < 1 || batchcount < 1 ||
4066 batchcount > limit || shared < 0) {
4069 res = do_tune_cpucache(cachep, limit,
4070 batchcount, shared);
4075 mutex_unlock(&cache_chain_mutex);
4081 #ifdef CONFIG_DEBUG_SLAB_LEAK
4083 static void *leaks_start(struct seq_file *m, loff_t *pos)
4086 struct list_head *p;
4088 mutex_lock(&cache_chain_mutex);
4089 p = cache_chain.next;
4092 if (p == &cache_chain)
4095 return list_entry(p, struct kmem_cache, next);
4098 static inline int add_caller(unsigned long *n, unsigned long v)
4108 unsigned long *q = p + 2 * i;
4122 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4128 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4134 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4135 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4137 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4142 static void show_symbol(struct seq_file *m, unsigned long address)
4144 #ifdef CONFIG_KALLSYMS
4147 unsigned long offset, size;
4148 char namebuf[KSYM_NAME_LEN+1];
4150 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4153 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4155 seq_printf(m, " [%s]", modname);
4159 seq_printf(m, "%p", (void *)address);
4162 static int leaks_show(struct seq_file *m, void *p)
4164 struct kmem_cache *cachep = p;
4166 struct kmem_list3 *l3;
4168 unsigned long *n = m->private;
4172 if (!(cachep->flags & SLAB_STORE_USER))
4174 if (!(cachep->flags & SLAB_RED_ZONE))
4177 /* OK, we can do it */
4181 for_each_online_node(node) {
4182 l3 = cachep->nodelists[node];
4187 spin_lock_irq(&l3->list_lock);
4189 list_for_each_entry(slabp, &l3->slabs_full, list)
4190 handle_slab(n, cachep, slabp);
4191 list_for_each_entry(slabp, &l3->slabs_partial, list)
4192 handle_slab(n, cachep, slabp);
4193 spin_unlock_irq(&l3->list_lock);
4195 name = cachep->name;
4197 /* Increase the buffer size */
4198 mutex_unlock(&cache_chain_mutex);
4199 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4201 /* Too bad, we are really out */
4203 mutex_lock(&cache_chain_mutex);
4206 *(unsigned long *)m->private = n[0] * 2;
4208 mutex_lock(&cache_chain_mutex);
4209 /* Now make sure this entry will be retried */
4213 for (i = 0; i < n[1]; i++) {
4214 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4215 show_symbol(m, n[2*i+2]);
4221 struct seq_operations slabstats_op = {
4222 .start = leaks_start,
4231 * ksize - get the actual amount of memory allocated for a given object
4232 * @objp: Pointer to the object
4234 * kmalloc may internally round up allocations and return more memory
4235 * than requested. ksize() can be used to determine the actual amount of
4236 * memory allocated. The caller may use this additional memory, even though
4237 * a smaller amount of memory was initially specified with the kmalloc call.
4238 * The caller must guarantee that objp points to a valid object previously
4239 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4240 * must not be freed during the duration of the call.
4242 unsigned int ksize(const void *objp)
4244 if (unlikely(objp == NULL))
4247 return obj_size(virt_to_cache(objp));