3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/nodemask.h>
107 #include <linux/mempolicy.h>
108 #include <linux/mutex.h>
109 #include <linux/rtmutex.h>
111 #include <asm/uaccess.h>
112 #include <asm/cacheflush.h>
113 #include <asm/tlbflush.h>
114 #include <asm/page.h>
117 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
118 * SLAB_RED_ZONE & SLAB_POISON.
119 * 0 for faster, smaller code (especially in the critical paths).
121 * STATS - 1 to collect stats for /proc/slabinfo.
122 * 0 for faster, smaller code (especially in the critical paths).
124 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
127 #ifdef CONFIG_DEBUG_SLAB
130 #define FORCED_DEBUG 1
134 #define FORCED_DEBUG 0
137 /* Shouldn't this be in a header file somewhere? */
138 #define BYTES_PER_WORD sizeof(void *)
140 #ifndef cache_line_size
141 #define cache_line_size() L1_CACHE_BYTES
144 #ifndef ARCH_KMALLOC_MINALIGN
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
150 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
151 * Note that this flag disables some debug features.
153 #define ARCH_KMALLOC_MINALIGN 0
156 #ifndef ARCH_SLAB_MINALIGN
158 * Enforce a minimum alignment for all caches.
159 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
160 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
161 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
162 * some debug features.
164 #define ARCH_SLAB_MINALIGN 0
167 #ifndef ARCH_KMALLOC_FLAGS
168 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171 /* Legal flag mask for kmem_cache_create(). */
173 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
176 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
180 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
181 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
182 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
183 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
189 * Bufctl's are used for linking objs within a slab
192 * This implementation relies on "struct page" for locating the cache &
193 * slab an object belongs to.
194 * This allows the bufctl structure to be small (one int), but limits
195 * the number of objects a slab (not a cache) can contain when off-slab
196 * bufctls are used. The limit is the size of the largest general cache
197 * that does not use off-slab slabs.
198 * For 32bit archs with 4 kB pages, is this 56.
199 * This is not serious, as it is only for large objects, when it is unwise
200 * to have too many per slab.
201 * Note: This limit can be raised by introducing a general cache whose size
202 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 typedef unsigned int kmem_bufctl_t;
206 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
207 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
208 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
209 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 struct list_head list;
220 unsigned long colouroff;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
224 unsigned short nodeid;
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
241 * We assume struct slab_rcu can overlay struct slab when destroying.
244 struct rcu_head head;
245 struct kmem_cache *cachep;
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
257 * The limit is stored in the per-cpu structure to reduce the data cache
264 unsigned int batchcount;
265 unsigned int touched;
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
271 * [0] is for gcc 2.95. It should really be [].
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
279 #define BOOT_CPUCACHE_ENTRIES 1
280 struct arraycache_init {
281 struct array_cache cache;
282 void *entries[BOOT_CPUCACHE_ENTRIES];
286 * The slab lists for all objects.
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
293 unsigned int free_limit;
294 unsigned int colour_next; /* Per-node cache coloring */
295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
298 unsigned long next_reap; /* updated without locking */
299 int free_touched; /* updated without locking */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
311 static int drain_freelist(struct kmem_cache *cache,
312 struct kmem_list3 *l3, int tofree);
313 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
315 static int enable_cpucache(struct kmem_cache *cachep);
316 static void cache_reap(void *unused);
319 * This function must be completely optimized away if a constant is passed to
320 * it. Mostly the same as what is in linux/slab.h except it returns an index.
322 static __always_inline int index_of(const size_t size)
324 extern void __bad_size(void);
326 if (__builtin_constant_p(size)) {
334 #include "linux/kmalloc_sizes.h"
342 static int slab_early_init = 1;
344 #define INDEX_AC index_of(sizeof(struct arraycache_init))
345 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
347 static void kmem_list3_init(struct kmem_list3 *parent)
349 INIT_LIST_HEAD(&parent->slabs_full);
350 INIT_LIST_HEAD(&parent->slabs_partial);
351 INIT_LIST_HEAD(&parent->slabs_free);
352 parent->shared = NULL;
353 parent->alien = NULL;
354 parent->colour_next = 0;
355 spin_lock_init(&parent->list_lock);
356 parent->free_objects = 0;
357 parent->free_touched = 0;
360 #define MAKE_LIST(cachep, listp, slab, nodeid) \
362 INIT_LIST_HEAD(listp); \
363 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
366 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
368 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
369 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
380 /* 1) per-cpu data, touched during every alloc/free */
381 struct array_cache *array[NR_CPUS];
382 /* 2) Cache tunables. Protected by cache_chain_mutex */
383 unsigned int batchcount;
387 unsigned int buffer_size;
388 /* 3) touched by every alloc & free from the backend */
389 struct kmem_list3 *nodelists[MAX_NUMNODES];
391 unsigned int flags; /* constant flags */
392 unsigned int num; /* # of objs per slab */
394 /* 4) cache_grow/shrink */
395 /* order of pgs per slab (2^n) */
396 unsigned int gfporder;
398 /* force GFP flags, e.g. GFP_DMA */
401 size_t colour; /* cache colouring range */
402 unsigned int colour_off; /* colour offset */
403 struct kmem_cache *slabp_cache;
404 unsigned int slab_size;
405 unsigned int dflags; /* dynamic flags */
407 /* constructor func */
408 void (*ctor) (void *, struct kmem_cache *, unsigned long);
410 /* de-constructor func */
411 void (*dtor) (void *, struct kmem_cache *, unsigned long);
413 /* 5) cache creation/removal */
415 struct list_head next;
419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
428 unsigned long node_overflow;
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
446 #define CFLGS_OFF_SLAB (0x80000000UL)
447 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
449 #define BATCHREFILL_LIMIT 16
451 * Optimization question: fewer reaps means less probability for unnessary
452 * cpucache drain/refill cycles.
454 * OTOH the cpuarrays can contain lots of objects,
455 * which could lock up otherwise freeable slabs.
457 #define REAPTIMEOUT_CPUC (2*HZ)
458 #define REAPTIMEOUT_LIST3 (4*HZ)
461 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
462 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
463 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
464 #define STATS_INC_GROWN(x) ((x)->grown++)
465 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
466 #define STATS_SET_HIGH(x) \
468 if ((x)->num_active > (x)->high_mark) \
469 (x)->high_mark = (x)->num_active; \
471 #define STATS_INC_ERR(x) ((x)->errors++)
472 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
473 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
474 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
475 #define STATS_SET_FREEABLE(x, i) \
477 if ((x)->max_freeable < i) \
478 (x)->max_freeable = i; \
480 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
481 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
482 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
483 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
485 #define STATS_INC_ACTIVE(x) do { } while (0)
486 #define STATS_DEC_ACTIVE(x) do { } while (0)
487 #define STATS_INC_ALLOCED(x) do { } while (0)
488 #define STATS_INC_GROWN(x) do { } while (0)
489 #define STATS_ADD_REAPED(x,y) do { } while (0)
490 #define STATS_SET_HIGH(x) do { } while (0)
491 #define STATS_INC_ERR(x) do { } while (0)
492 #define STATS_INC_NODEALLOCS(x) do { } while (0)
493 #define STATS_INC_NODEFREES(x) do { } while (0)
494 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
495 #define STATS_SET_FREEABLE(x, i) do { } while (0)
496 #define STATS_INC_ALLOCHIT(x) do { } while (0)
497 #define STATS_INC_ALLOCMISS(x) do { } while (0)
498 #define STATS_INC_FREEHIT(x) do { } while (0)
499 #define STATS_INC_FREEMISS(x) do { } while (0)
505 * memory layout of objects:
507 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
508 * the end of an object is aligned with the end of the real
509 * allocation. Catches writes behind the end of the allocation.
510 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
512 * cachep->obj_offset: The real object.
513 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
514 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
515 * [BYTES_PER_WORD long]
517 static int obj_offset(struct kmem_cache *cachep)
519 return cachep->obj_offset;
522 static int obj_size(struct kmem_cache *cachep)
524 return cachep->obj_size;
527 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
529 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
530 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
533 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
535 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
536 if (cachep->flags & SLAB_STORE_USER)
537 return (unsigned long *)(objp + cachep->buffer_size -
539 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
542 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
544 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
545 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
550 #define obj_offset(x) 0
551 #define obj_size(cachep) (cachep->buffer_size)
552 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
553 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
559 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
562 #if defined(CONFIG_LARGE_ALLOCS)
563 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
564 #define MAX_GFP_ORDER 13 /* up to 32Mb */
565 #elif defined(CONFIG_MMU)
566 #define MAX_OBJ_ORDER 5 /* 32 pages */
567 #define MAX_GFP_ORDER 5 /* 32 pages */
569 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
570 #define MAX_GFP_ORDER 8 /* up to 1Mb */
574 * Do not go above this order unless 0 objects fit into the slab.
576 #define BREAK_GFP_ORDER_HI 1
577 #define BREAK_GFP_ORDER_LO 0
578 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
581 * Functions for storing/retrieving the cachep and or slab from the page
582 * allocator. These are used to find the slab an obj belongs to. With kfree(),
583 * these are used to find the cache which an obj belongs to.
585 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
587 page->lru.next = (struct list_head *)cache;
590 static inline struct kmem_cache *page_get_cache(struct page *page)
592 if (unlikely(PageCompound(page)))
593 page = (struct page *)page_private(page);
594 BUG_ON(!PageSlab(page));
595 return (struct kmem_cache *)page->lru.next;
598 static inline void page_set_slab(struct page *page, struct slab *slab)
600 page->lru.prev = (struct list_head *)slab;
603 static inline struct slab *page_get_slab(struct page *page)
605 if (unlikely(PageCompound(page)))
606 page = (struct page *)page_private(page);
607 BUG_ON(!PageSlab(page));
608 return (struct slab *)page->lru.prev;
611 static inline struct kmem_cache *virt_to_cache(const void *obj)
613 struct page *page = virt_to_page(obj);
614 return page_get_cache(page);
617 static inline struct slab *virt_to_slab(const void *obj)
619 struct page *page = virt_to_page(obj);
620 return page_get_slab(page);
623 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
626 return slab->s_mem + cache->buffer_size * idx;
629 static inline unsigned int obj_to_index(struct kmem_cache *cache,
630 struct slab *slab, void *obj)
632 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
636 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 struct cache_sizes malloc_sizes[] = {
639 #define CACHE(x) { .cs_size = (x) },
640 #include <linux/kmalloc_sizes.h>
644 EXPORT_SYMBOL(malloc_sizes);
646 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
652 static struct cache_names __initdata cache_names[] = {
653 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
654 #include <linux/kmalloc_sizes.h>
659 static struct arraycache_init initarray_cache __initdata =
660 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
661 static struct arraycache_init initarray_generic =
662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
664 /* internal cache of cache description objs */
665 static struct kmem_cache cache_cache = {
667 .limit = BOOT_CPUCACHE_ENTRIES,
669 .buffer_size = sizeof(struct kmem_cache),
670 .name = "kmem_cache",
672 .obj_size = sizeof(struct kmem_cache),
676 #define BAD_ALIEN_MAGIC 0x01020304ul
678 #ifdef CONFIG_LOCKDEP
681 * Slab sometimes uses the kmalloc slabs to store the slab headers
682 * for other slabs "off slab".
683 * The locking for this is tricky in that it nests within the locks
684 * of all other slabs in a few places; to deal with this special
685 * locking we put on-slab caches into a separate lock-class.
687 * We set lock class for alien array caches which are up during init.
688 * The lock annotation will be lost if all cpus of a node goes down and
689 * then comes back up during hotplug
691 static struct lock_class_key on_slab_l3_key;
692 static struct lock_class_key on_slab_alc_key;
694 static inline void init_lock_keys(void)
698 struct cache_sizes *s = malloc_sizes;
700 while (s->cs_size != ULONG_MAX) {
702 struct array_cache **alc;
704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
705 if (!l3 || OFF_SLAB(s->cs_cachep))
707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
710 * FIXME: This check for BAD_ALIEN_MAGIC
711 * should go away when common slab code is taught to
712 * work even without alien caches.
713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
714 * for alloc_alien_cache,
716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
720 lockdep_set_class(&alc[r]->lock,
728 static inline void init_lock_keys(void)
733 /* Guard access to the cache-chain. */
734 static DEFINE_MUTEX(cache_chain_mutex);
735 static struct list_head cache_chain;
738 * chicken and egg problem: delay the per-cpu array allocation
739 * until the general caches are up.
749 * used by boot code to determine if it can use slab based allocator
751 int slab_is_available(void)
753 return g_cpucache_up == FULL;
756 static DEFINE_PER_CPU(struct work_struct, reap_work);
758 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
760 return cachep->array[smp_processor_id()];
763 static inline struct kmem_cache *__find_general_cachep(size_t size,
766 struct cache_sizes *csizep = malloc_sizes;
769 /* This happens if someone tries to call
770 * kmem_cache_create(), or __kmalloc(), before
771 * the generic caches are initialized.
773 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
775 while (size > csizep->cs_size)
779 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
780 * has cs_{dma,}cachep==NULL. Thus no special case
781 * for large kmalloc calls required.
783 if (unlikely(gfpflags & GFP_DMA))
784 return csizep->cs_dmacachep;
785 return csizep->cs_cachep;
788 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
790 return __find_general_cachep(size, gfpflags);
793 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
795 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
799 * Calculate the number of objects and left-over bytes for a given buffer size.
801 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
802 size_t align, int flags, size_t *left_over,
807 size_t slab_size = PAGE_SIZE << gfporder;
810 * The slab management structure can be either off the slab or
811 * on it. For the latter case, the memory allocated for a
815 * - One kmem_bufctl_t for each object
816 * - Padding to respect alignment of @align
817 * - @buffer_size bytes for each object
819 * If the slab management structure is off the slab, then the
820 * alignment will already be calculated into the size. Because
821 * the slabs are all pages aligned, the objects will be at the
822 * correct alignment when allocated.
824 if (flags & CFLGS_OFF_SLAB) {
826 nr_objs = slab_size / buffer_size;
828 if (nr_objs > SLAB_LIMIT)
829 nr_objs = SLAB_LIMIT;
832 * Ignore padding for the initial guess. The padding
833 * is at most @align-1 bytes, and @buffer_size is at
834 * least @align. In the worst case, this result will
835 * be one greater than the number of objects that fit
836 * into the memory allocation when taking the padding
839 nr_objs = (slab_size - sizeof(struct slab)) /
840 (buffer_size + sizeof(kmem_bufctl_t));
843 * This calculated number will be either the right
844 * amount, or one greater than what we want.
846 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
850 if (nr_objs > SLAB_LIMIT)
851 nr_objs = SLAB_LIMIT;
853 mgmt_size = slab_mgmt_size(nr_objs, align);
856 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
859 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
861 static void __slab_error(const char *function, struct kmem_cache *cachep,
864 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
865 function, cachep->name, msg);
871 * Special reaping functions for NUMA systems called from cache_reap().
872 * These take care of doing round robin flushing of alien caches (containing
873 * objects freed on different nodes from which they were allocated) and the
874 * flushing of remote pcps by calling drain_node_pages.
876 static DEFINE_PER_CPU(unsigned long, reap_node);
878 static void init_reap_node(int cpu)
882 node = next_node(cpu_to_node(cpu), node_online_map);
883 if (node == MAX_NUMNODES)
884 node = first_node(node_online_map);
886 __get_cpu_var(reap_node) = node;
889 static void next_reap_node(void)
891 int node = __get_cpu_var(reap_node);
894 * Also drain per cpu pages on remote zones
896 if (node != numa_node_id())
897 drain_node_pages(node);
899 node = next_node(node, node_online_map);
900 if (unlikely(node >= MAX_NUMNODES))
901 node = first_node(node_online_map);
902 __get_cpu_var(reap_node) = node;
906 #define init_reap_node(cpu) do { } while (0)
907 #define next_reap_node(void) do { } while (0)
911 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
912 * via the workqueue/eventd.
913 * Add the CPU number into the expiration time to minimize the possibility of
914 * the CPUs getting into lockstep and contending for the global cache chain
917 static void __devinit start_cpu_timer(int cpu)
919 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
922 * When this gets called from do_initcalls via cpucache_init(),
923 * init_workqueues() has already run, so keventd will be setup
926 if (keventd_up() && reap_work->func == NULL) {
928 INIT_WORK(reap_work, cache_reap, NULL);
929 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
933 static struct array_cache *alloc_arraycache(int node, int entries,
936 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
937 struct array_cache *nc = NULL;
939 nc = kmalloc_node(memsize, GFP_KERNEL, node);
943 nc->batchcount = batchcount;
945 spin_lock_init(&nc->lock);
951 * Transfer objects in one arraycache to another.
952 * Locking must be handled by the caller.
954 * Return the number of entries transferred.
956 static int transfer_objects(struct array_cache *to,
957 struct array_cache *from, unsigned int max)
959 /* Figure out how many entries to transfer */
960 int nr = min(min(from->avail, max), to->limit - to->avail);
965 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
976 #define drain_alien_cache(cachep, alien) do { } while (0)
977 #define reap_alien(cachep, l3) do { } while (0)
979 static inline struct array_cache **alloc_alien_cache(int node, int limit)
981 return (struct array_cache **)BAD_ALIEN_MAGIC;
984 static inline void free_alien_cache(struct array_cache **ac_ptr)
988 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
993 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
999 static inline void *__cache_alloc_node(struct kmem_cache *cachep,
1000 gfp_t flags, int nodeid)
1005 #else /* CONFIG_NUMA */
1007 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
1008 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1010 static struct array_cache **alloc_alien_cache(int node, int limit)
1012 struct array_cache **ac_ptr;
1013 int memsize = sizeof(void *) * MAX_NUMNODES;
1018 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1021 if (i == node || !node_online(i)) {
1025 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1027 for (i--; i <= 0; i--)
1037 static void free_alien_cache(struct array_cache **ac_ptr)
1048 static void __drain_alien_cache(struct kmem_cache *cachep,
1049 struct array_cache *ac, int node)
1051 struct kmem_list3 *rl3 = cachep->nodelists[node];
1054 spin_lock(&rl3->list_lock);
1056 * Stuff objects into the remote nodes shared array first.
1057 * That way we could avoid the overhead of putting the objects
1058 * into the free lists and getting them back later.
1061 transfer_objects(rl3->shared, ac, ac->limit);
1063 free_block(cachep, ac->entry, ac->avail, node);
1065 spin_unlock(&rl3->list_lock);
1070 * Called from cache_reap() to regularly drain alien caches round robin.
1072 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1074 int node = __get_cpu_var(reap_node);
1077 struct array_cache *ac = l3->alien[node];
1079 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1080 __drain_alien_cache(cachep, ac, node);
1081 spin_unlock_irq(&ac->lock);
1086 static void drain_alien_cache(struct kmem_cache *cachep,
1087 struct array_cache **alien)
1090 struct array_cache *ac;
1091 unsigned long flags;
1093 for_each_online_node(i) {
1096 spin_lock_irqsave(&ac->lock, flags);
1097 __drain_alien_cache(cachep, ac, i);
1098 spin_unlock_irqrestore(&ac->lock, flags);
1103 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1105 struct slab *slabp = virt_to_slab(objp);
1106 int nodeid = slabp->nodeid;
1107 struct kmem_list3 *l3;
1108 struct array_cache *alien = NULL;
1111 node = numa_node_id();
1114 * Make sure we are not freeing a object from another node to the array
1115 * cache on this cpu.
1117 if (likely(slabp->nodeid == node))
1120 l3 = cachep->nodelists[node];
1121 STATS_INC_NODEFREES(cachep);
1122 if (l3->alien && l3->alien[nodeid]) {
1123 alien = l3->alien[nodeid];
1124 spin_lock(&alien->lock);
1125 if (unlikely(alien->avail == alien->limit)) {
1126 STATS_INC_ACOVERFLOW(cachep);
1127 __drain_alien_cache(cachep, alien, nodeid);
1129 alien->entry[alien->avail++] = objp;
1130 spin_unlock(&alien->lock);
1132 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1133 free_block(cachep, &objp, 1, nodeid);
1134 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1140 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1141 unsigned long action, void *hcpu)
1143 long cpu = (long)hcpu;
1144 struct kmem_cache *cachep;
1145 struct kmem_list3 *l3 = NULL;
1146 int node = cpu_to_node(cpu);
1147 int memsize = sizeof(struct kmem_list3);
1150 case CPU_UP_PREPARE:
1151 mutex_lock(&cache_chain_mutex);
1153 * We need to do this right in the beginning since
1154 * alloc_arraycache's are going to use this list.
1155 * kmalloc_node allows us to add the slab to the right
1156 * kmem_list3 and not this cpu's kmem_list3
1159 list_for_each_entry(cachep, &cache_chain, next) {
1161 * Set up the size64 kmemlist for cpu before we can
1162 * begin anything. Make sure some other cpu on this
1163 * node has not already allocated this
1165 if (!cachep->nodelists[node]) {
1166 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1169 kmem_list3_init(l3);
1170 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1171 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1174 * The l3s don't come and go as CPUs come and
1175 * go. cache_chain_mutex is sufficient
1178 cachep->nodelists[node] = l3;
1181 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1182 cachep->nodelists[node]->free_limit =
1183 (1 + nr_cpus_node(node)) *
1184 cachep->batchcount + cachep->num;
1185 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1189 * Now we can go ahead with allocating the shared arrays and
1192 list_for_each_entry(cachep, &cache_chain, next) {
1193 struct array_cache *nc;
1194 struct array_cache *shared;
1195 struct array_cache **alien;
1197 nc = alloc_arraycache(node, cachep->limit,
1198 cachep->batchcount);
1201 shared = alloc_arraycache(node,
1202 cachep->shared * cachep->batchcount,
1207 alien = alloc_alien_cache(node, cachep->limit);
1210 cachep->array[cpu] = nc;
1211 l3 = cachep->nodelists[node];
1214 spin_lock_irq(&l3->list_lock);
1217 * We are serialised from CPU_DEAD or
1218 * CPU_UP_CANCELLED by the cpucontrol lock
1220 l3->shared = shared;
1229 spin_unlock_irq(&l3->list_lock);
1231 free_alien_cache(alien);
1233 mutex_unlock(&cache_chain_mutex);
1236 start_cpu_timer(cpu);
1238 #ifdef CONFIG_HOTPLUG_CPU
1241 * Even if all the cpus of a node are down, we don't free the
1242 * kmem_list3 of any cache. This to avoid a race between
1243 * cpu_down, and a kmalloc allocation from another cpu for
1244 * memory from the node of the cpu going down. The list3
1245 * structure is usually allocated from kmem_cache_create() and
1246 * gets destroyed at kmem_cache_destroy().
1249 case CPU_UP_CANCELED:
1250 mutex_lock(&cache_chain_mutex);
1251 list_for_each_entry(cachep, &cache_chain, next) {
1252 struct array_cache *nc;
1253 struct array_cache *shared;
1254 struct array_cache **alien;
1257 mask = node_to_cpumask(node);
1258 /* cpu is dead; no one can alloc from it. */
1259 nc = cachep->array[cpu];
1260 cachep->array[cpu] = NULL;
1261 l3 = cachep->nodelists[node];
1264 goto free_array_cache;
1266 spin_lock_irq(&l3->list_lock);
1268 /* Free limit for this kmem_list3 */
1269 l3->free_limit -= cachep->batchcount;
1271 free_block(cachep, nc->entry, nc->avail, node);
1273 if (!cpus_empty(mask)) {
1274 spin_unlock_irq(&l3->list_lock);
1275 goto free_array_cache;
1278 shared = l3->shared;
1280 free_block(cachep, l3->shared->entry,
1281 l3->shared->avail, node);
1288 spin_unlock_irq(&l3->list_lock);
1292 drain_alien_cache(cachep, alien);
1293 free_alien_cache(alien);
1299 * In the previous loop, all the objects were freed to
1300 * the respective cache's slabs, now we can go ahead and
1301 * shrink each nodelist to its limit.
1303 list_for_each_entry(cachep, &cache_chain, next) {
1304 l3 = cachep->nodelists[node];
1307 drain_freelist(cachep, l3, l3->free_objects);
1309 mutex_unlock(&cache_chain_mutex);
1315 mutex_unlock(&cache_chain_mutex);
1319 static struct notifier_block __cpuinitdata cpucache_notifier = {
1320 &cpuup_callback, NULL, 0
1324 * swap the static kmem_list3 with kmalloced memory
1326 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1329 struct kmem_list3 *ptr;
1331 BUG_ON(cachep->nodelists[nodeid] != list);
1332 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1335 local_irq_disable();
1336 memcpy(ptr, list, sizeof(struct kmem_list3));
1338 * Do not assume that spinlocks can be initialized via memcpy:
1340 spin_lock_init(&ptr->list_lock);
1342 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1343 cachep->nodelists[nodeid] = ptr;
1348 * Initialisation. Called after the page allocator have been initialised and
1349 * before smp_init().
1351 void __init kmem_cache_init(void)
1354 struct cache_sizes *sizes;
1355 struct cache_names *names;
1360 for (i = 0; i < NUM_INIT_LISTS; i++) {
1361 kmem_list3_init(&initkmem_list3[i]);
1362 if (i < MAX_NUMNODES)
1363 cache_cache.nodelists[i] = NULL;
1367 * Fragmentation resistance on low memory - only use bigger
1368 * page orders on machines with more than 32MB of memory.
1370 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1371 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1373 /* Bootstrap is tricky, because several objects are allocated
1374 * from caches that do not exist yet:
1375 * 1) initialize the cache_cache cache: it contains the struct
1376 * kmem_cache structures of all caches, except cache_cache itself:
1377 * cache_cache is statically allocated.
1378 * Initially an __init data area is used for the head array and the
1379 * kmem_list3 structures, it's replaced with a kmalloc allocated
1380 * array at the end of the bootstrap.
1381 * 2) Create the first kmalloc cache.
1382 * The struct kmem_cache for the new cache is allocated normally.
1383 * An __init data area is used for the head array.
1384 * 3) Create the remaining kmalloc caches, with minimally sized
1386 * 4) Replace the __init data head arrays for cache_cache and the first
1387 * kmalloc cache with kmalloc allocated arrays.
1388 * 5) Replace the __init data for kmem_list3 for cache_cache and
1389 * the other cache's with kmalloc allocated memory.
1390 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1393 node = numa_node_id();
1395 /* 1) create the cache_cache */
1396 INIT_LIST_HEAD(&cache_chain);
1397 list_add(&cache_cache.next, &cache_chain);
1398 cache_cache.colour_off = cache_line_size();
1399 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1400 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1402 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1405 for (order = 0; order < MAX_ORDER; order++) {
1406 cache_estimate(order, cache_cache.buffer_size,
1407 cache_line_size(), 0, &left_over, &cache_cache.num);
1408 if (cache_cache.num)
1411 BUG_ON(!cache_cache.num);
1412 cache_cache.gfporder = order;
1413 cache_cache.colour = left_over / cache_cache.colour_off;
1414 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1415 sizeof(struct slab), cache_line_size());
1417 /* 2+3) create the kmalloc caches */
1418 sizes = malloc_sizes;
1419 names = cache_names;
1422 * Initialize the caches that provide memory for the array cache and the
1423 * kmem_list3 structures first. Without this, further allocations will
1427 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1428 sizes[INDEX_AC].cs_size,
1429 ARCH_KMALLOC_MINALIGN,
1430 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1433 if (INDEX_AC != INDEX_L3) {
1434 sizes[INDEX_L3].cs_cachep =
1435 kmem_cache_create(names[INDEX_L3].name,
1436 sizes[INDEX_L3].cs_size,
1437 ARCH_KMALLOC_MINALIGN,
1438 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1442 slab_early_init = 0;
1444 while (sizes->cs_size != ULONG_MAX) {
1446 * For performance, all the general caches are L1 aligned.
1447 * This should be particularly beneficial on SMP boxes, as it
1448 * eliminates "false sharing".
1449 * Note for systems short on memory removing the alignment will
1450 * allow tighter packing of the smaller caches.
1452 if (!sizes->cs_cachep) {
1453 sizes->cs_cachep = kmem_cache_create(names->name,
1455 ARCH_KMALLOC_MINALIGN,
1456 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1460 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1462 ARCH_KMALLOC_MINALIGN,
1463 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1469 /* 4) Replace the bootstrap head arrays */
1471 struct array_cache *ptr;
1473 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1475 local_irq_disable();
1476 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1477 memcpy(ptr, cpu_cache_get(&cache_cache),
1478 sizeof(struct arraycache_init));
1480 * Do not assume that spinlocks can be initialized via memcpy:
1482 spin_lock_init(&ptr->lock);
1484 cache_cache.array[smp_processor_id()] = ptr;
1487 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1489 local_irq_disable();
1490 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1491 != &initarray_generic.cache);
1492 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1493 sizeof(struct arraycache_init));
1495 * Do not assume that spinlocks can be initialized via memcpy:
1497 spin_lock_init(&ptr->lock);
1499 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1503 /* 5) Replace the bootstrap kmem_list3's */
1507 /* Replace the static kmem_list3 structures for the boot cpu */
1508 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1510 for_each_online_node(nid) {
1511 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1512 &initkmem_list3[SIZE_AC + nid], nid);
1514 if (INDEX_AC != INDEX_L3) {
1515 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1516 &initkmem_list3[SIZE_L3 + nid], nid);
1521 /* 6) resize the head arrays to their final sizes */
1523 struct kmem_cache *cachep;
1524 mutex_lock(&cache_chain_mutex);
1525 list_for_each_entry(cachep, &cache_chain, next)
1526 if (enable_cpucache(cachep))
1528 mutex_unlock(&cache_chain_mutex);
1531 /* Annotate slab for lockdep -- annotate the malloc caches */
1536 g_cpucache_up = FULL;
1539 * Register a cpu startup notifier callback that initializes
1540 * cpu_cache_get for all new cpus
1542 register_cpu_notifier(&cpucache_notifier);
1545 * The reap timers are started later, with a module init call: That part
1546 * of the kernel is not yet operational.
1550 static int __init cpucache_init(void)
1555 * Register the timers that return unneeded pages to the page allocator
1557 for_each_online_cpu(cpu)
1558 start_cpu_timer(cpu);
1561 __initcall(cpucache_init);
1564 * Interface to system's page allocator. No need to hold the cache-lock.
1566 * If we requested dmaable memory, we will get it. Even if we
1567 * did not request dmaable memory, we might get it, but that
1568 * would be relatively rare and ignorable.
1570 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1578 * Nommu uses slab's for process anonymous memory allocations, and thus
1579 * requires __GFP_COMP to properly refcount higher order allocations
1581 flags |= __GFP_COMP;
1585 * Under NUMA we want memory on the indicated node. We will handle
1586 * the needed fallback ourselves since we want to serve from our
1587 * per node object lists first for other nodes.
1589 flags |= cachep->gfpflags | GFP_THISNODE;
1591 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1595 nr_pages = (1 << cachep->gfporder);
1596 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1597 add_zone_page_state(page_zone(page),
1598 NR_SLAB_RECLAIMABLE, nr_pages);
1600 add_zone_page_state(page_zone(page),
1601 NR_SLAB_UNRECLAIMABLE, nr_pages);
1602 for (i = 0; i < nr_pages; i++)
1603 __SetPageSlab(page + i);
1604 return page_address(page);
1608 * Interface to system's page release.
1610 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1612 unsigned long i = (1 << cachep->gfporder);
1613 struct page *page = virt_to_page(addr);
1614 const unsigned long nr_freed = i;
1616 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1617 sub_zone_page_state(page_zone(page),
1618 NR_SLAB_RECLAIMABLE, nr_freed);
1620 sub_zone_page_state(page_zone(page),
1621 NR_SLAB_UNRECLAIMABLE, nr_freed);
1623 BUG_ON(!PageSlab(page));
1624 __ClearPageSlab(page);
1627 if (current->reclaim_state)
1628 current->reclaim_state->reclaimed_slab += nr_freed;
1629 free_pages((unsigned long)addr, cachep->gfporder);
1632 static void kmem_rcu_free(struct rcu_head *head)
1634 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1635 struct kmem_cache *cachep = slab_rcu->cachep;
1637 kmem_freepages(cachep, slab_rcu->addr);
1638 if (OFF_SLAB(cachep))
1639 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1644 #ifdef CONFIG_DEBUG_PAGEALLOC
1645 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1646 unsigned long caller)
1648 int size = obj_size(cachep);
1650 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1652 if (size < 5 * sizeof(unsigned long))
1655 *addr++ = 0x12345678;
1657 *addr++ = smp_processor_id();
1658 size -= 3 * sizeof(unsigned long);
1660 unsigned long *sptr = &caller;
1661 unsigned long svalue;
1663 while (!kstack_end(sptr)) {
1665 if (kernel_text_address(svalue)) {
1667 size -= sizeof(unsigned long);
1668 if (size <= sizeof(unsigned long))
1674 *addr++ = 0x87654321;
1678 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1680 int size = obj_size(cachep);
1681 addr = &((char *)addr)[obj_offset(cachep)];
1683 memset(addr, val, size);
1684 *(unsigned char *)(addr + size - 1) = POISON_END;
1687 static void dump_line(char *data, int offset, int limit)
1690 unsigned char error = 0;
1693 printk(KERN_ERR "%03x:", offset);
1694 for (i = 0; i < limit; i++) {
1695 if (data[offset + i] != POISON_FREE) {
1696 error = data[offset + i];
1699 printk(" %02x", (unsigned char)data[offset + i]);
1703 if (bad_count == 1) {
1704 error ^= POISON_FREE;
1705 if (!(error & (error - 1))) {
1706 printk(KERN_ERR "Single bit error detected. Probably "
1709 printk(KERN_ERR "Run memtest86+ or a similar memory "
1712 printk(KERN_ERR "Run a memory test tool.\n");
1721 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1726 if (cachep->flags & SLAB_RED_ZONE) {
1727 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1728 *dbg_redzone1(cachep, objp),
1729 *dbg_redzone2(cachep, objp));
1732 if (cachep->flags & SLAB_STORE_USER) {
1733 printk(KERN_ERR "Last user: [<%p>]",
1734 *dbg_userword(cachep, objp));
1735 print_symbol("(%s)",
1736 (unsigned long)*dbg_userword(cachep, objp));
1739 realobj = (char *)objp + obj_offset(cachep);
1740 size = obj_size(cachep);
1741 for (i = 0; i < size && lines; i += 16, lines--) {
1744 if (i + limit > size)
1746 dump_line(realobj, i, limit);
1750 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1756 realobj = (char *)objp + obj_offset(cachep);
1757 size = obj_size(cachep);
1759 for (i = 0; i < size; i++) {
1760 char exp = POISON_FREE;
1763 if (realobj[i] != exp) {
1769 "Slab corruption: start=%p, len=%d\n",
1771 print_objinfo(cachep, objp, 0);
1773 /* Hexdump the affected line */
1776 if (i + limit > size)
1778 dump_line(realobj, i, limit);
1781 /* Limit to 5 lines */
1787 /* Print some data about the neighboring objects, if they
1790 struct slab *slabp = virt_to_slab(objp);
1793 objnr = obj_to_index(cachep, slabp, objp);
1795 objp = index_to_obj(cachep, slabp, objnr - 1);
1796 realobj = (char *)objp + obj_offset(cachep);
1797 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1799 print_objinfo(cachep, objp, 2);
1801 if (objnr + 1 < cachep->num) {
1802 objp = index_to_obj(cachep, slabp, objnr + 1);
1803 realobj = (char *)objp + obj_offset(cachep);
1804 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1806 print_objinfo(cachep, objp, 2);
1814 * slab_destroy_objs - destroy a slab and its objects
1815 * @cachep: cache pointer being destroyed
1816 * @slabp: slab pointer being destroyed
1818 * Call the registered destructor for each object in a slab that is being
1821 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1824 for (i = 0; i < cachep->num; i++) {
1825 void *objp = index_to_obj(cachep, slabp, i);
1827 if (cachep->flags & SLAB_POISON) {
1828 #ifdef CONFIG_DEBUG_PAGEALLOC
1829 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1831 kernel_map_pages(virt_to_page(objp),
1832 cachep->buffer_size / PAGE_SIZE, 1);
1834 check_poison_obj(cachep, objp);
1836 check_poison_obj(cachep, objp);
1839 if (cachep->flags & SLAB_RED_ZONE) {
1840 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1841 slab_error(cachep, "start of a freed object "
1843 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1844 slab_error(cachep, "end of a freed object "
1847 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1848 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1852 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1856 for (i = 0; i < cachep->num; i++) {
1857 void *objp = index_to_obj(cachep, slabp, i);
1858 (cachep->dtor) (objp, cachep, 0);
1865 * slab_destroy - destroy and release all objects in a slab
1866 * @cachep: cache pointer being destroyed
1867 * @slabp: slab pointer being destroyed
1869 * Destroy all the objs in a slab, and release the mem back to the system.
1870 * Before calling the slab must have been unlinked from the cache. The
1871 * cache-lock is not held/needed.
1873 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1875 void *addr = slabp->s_mem - slabp->colouroff;
1877 slab_destroy_objs(cachep, slabp);
1878 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1879 struct slab_rcu *slab_rcu;
1881 slab_rcu = (struct slab_rcu *)slabp;
1882 slab_rcu->cachep = cachep;
1883 slab_rcu->addr = addr;
1884 call_rcu(&slab_rcu->head, kmem_rcu_free);
1886 kmem_freepages(cachep, addr);
1887 if (OFF_SLAB(cachep))
1888 kmem_cache_free(cachep->slabp_cache, slabp);
1893 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1894 * size of kmem_list3.
1896 static void set_up_list3s(struct kmem_cache *cachep, int index)
1900 for_each_online_node(node) {
1901 cachep->nodelists[node] = &initkmem_list3[index + node];
1902 cachep->nodelists[node]->next_reap = jiffies +
1904 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1908 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1911 struct kmem_list3 *l3;
1913 for_each_online_cpu(i)
1914 kfree(cachep->array[i]);
1916 /* NUMA: free the list3 structures */
1917 for_each_online_node(i) {
1918 l3 = cachep->nodelists[i];
1921 free_alien_cache(l3->alien);
1925 kmem_cache_free(&cache_cache, cachep);
1930 * calculate_slab_order - calculate size (page order) of slabs
1931 * @cachep: pointer to the cache that is being created
1932 * @size: size of objects to be created in this cache.
1933 * @align: required alignment for the objects.
1934 * @flags: slab allocation flags
1936 * Also calculates the number of objects per slab.
1938 * This could be made much more intelligent. For now, try to avoid using
1939 * high order pages for slabs. When the gfp() functions are more friendly
1940 * towards high-order requests, this should be changed.
1942 static size_t calculate_slab_order(struct kmem_cache *cachep,
1943 size_t size, size_t align, unsigned long flags)
1945 unsigned long offslab_limit;
1946 size_t left_over = 0;
1949 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1953 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1957 if (flags & CFLGS_OFF_SLAB) {
1959 * Max number of objs-per-slab for caches which
1960 * use off-slab slabs. Needed to avoid a possible
1961 * looping condition in cache_grow().
1963 offslab_limit = size - sizeof(struct slab);
1964 offslab_limit /= sizeof(kmem_bufctl_t);
1966 if (num > offslab_limit)
1970 /* Found something acceptable - save it away */
1972 cachep->gfporder = gfporder;
1973 left_over = remainder;
1976 * A VFS-reclaimable slab tends to have most allocations
1977 * as GFP_NOFS and we really don't want to have to be allocating
1978 * higher-order pages when we are unable to shrink dcache.
1980 if (flags & SLAB_RECLAIM_ACCOUNT)
1984 * Large number of objects is good, but very large slabs are
1985 * currently bad for the gfp()s.
1987 if (gfporder >= slab_break_gfp_order)
1991 * Acceptable internal fragmentation?
1993 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1999 static int setup_cpu_cache(struct kmem_cache *cachep)
2001 if (g_cpucache_up == FULL)
2002 return enable_cpucache(cachep);
2004 if (g_cpucache_up == NONE) {
2006 * Note: the first kmem_cache_create must create the cache
2007 * that's used by kmalloc(24), otherwise the creation of
2008 * further caches will BUG().
2010 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2013 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2014 * the first cache, then we need to set up all its list3s,
2015 * otherwise the creation of further caches will BUG().
2017 set_up_list3s(cachep, SIZE_AC);
2018 if (INDEX_AC == INDEX_L3)
2019 g_cpucache_up = PARTIAL_L3;
2021 g_cpucache_up = PARTIAL_AC;
2023 cachep->array[smp_processor_id()] =
2024 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2026 if (g_cpucache_up == PARTIAL_AC) {
2027 set_up_list3s(cachep, SIZE_L3);
2028 g_cpucache_up = PARTIAL_L3;
2031 for_each_online_node(node) {
2032 cachep->nodelists[node] =
2033 kmalloc_node(sizeof(struct kmem_list3),
2035 BUG_ON(!cachep->nodelists[node]);
2036 kmem_list3_init(cachep->nodelists[node]);
2040 cachep->nodelists[numa_node_id()]->next_reap =
2041 jiffies + REAPTIMEOUT_LIST3 +
2042 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2044 cpu_cache_get(cachep)->avail = 0;
2045 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2046 cpu_cache_get(cachep)->batchcount = 1;
2047 cpu_cache_get(cachep)->touched = 0;
2048 cachep->batchcount = 1;
2049 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2054 * kmem_cache_create - Create a cache.
2055 * @name: A string which is used in /proc/slabinfo to identify this cache.
2056 * @size: The size of objects to be created in this cache.
2057 * @align: The required alignment for the objects.
2058 * @flags: SLAB flags
2059 * @ctor: A constructor for the objects.
2060 * @dtor: A destructor for the objects.
2062 * Returns a ptr to the cache on success, NULL on failure.
2063 * Cannot be called within a int, but can be interrupted.
2064 * The @ctor is run when new pages are allocated by the cache
2065 * and the @dtor is run before the pages are handed back.
2067 * @name must be valid until the cache is destroyed. This implies that
2068 * the module calling this has to destroy the cache before getting unloaded.
2072 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2073 * to catch references to uninitialised memory.
2075 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2076 * for buffer overruns.
2078 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2079 * cacheline. This can be beneficial if you're counting cycles as closely
2083 kmem_cache_create (const char *name, size_t size, size_t align,
2084 unsigned long flags,
2085 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2086 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2088 size_t left_over, slab_size, ralign;
2089 struct kmem_cache *cachep = NULL, *pc;
2092 * Sanity checks... these are all serious usage bugs.
2094 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2095 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2096 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2102 * Prevent CPUs from coming and going.
2103 * lock_cpu_hotplug() nests outside cache_chain_mutex
2107 mutex_lock(&cache_chain_mutex);
2109 list_for_each_entry(pc, &cache_chain, next) {
2110 mm_segment_t old_fs = get_fs();
2115 * This happens when the module gets unloaded and doesn't
2116 * destroy its slab cache and no-one else reuses the vmalloc
2117 * area of the module. Print a warning.
2120 res = __get_user(tmp, pc->name);
2123 printk("SLAB: cache with size %d has lost its name\n",
2128 if (!strcmp(pc->name, name)) {
2129 printk("kmem_cache_create: duplicate cache %s\n", name);
2136 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2137 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2138 /* No constructor, but inital state check requested */
2139 printk(KERN_ERR "%s: No con, but init state check "
2140 "requested - %s\n", __FUNCTION__, name);
2141 flags &= ~SLAB_DEBUG_INITIAL;
2145 * Enable redzoning and last user accounting, except for caches with
2146 * large objects, if the increased size would increase the object size
2147 * above the next power of two: caches with object sizes just above a
2148 * power of two have a significant amount of internal fragmentation.
2150 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2151 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2152 if (!(flags & SLAB_DESTROY_BY_RCU))
2153 flags |= SLAB_POISON;
2155 if (flags & SLAB_DESTROY_BY_RCU)
2156 BUG_ON(flags & SLAB_POISON);
2158 if (flags & SLAB_DESTROY_BY_RCU)
2162 * Always checks flags, a caller might be expecting debug support which
2165 BUG_ON(flags & ~CREATE_MASK);
2168 * Check that size is in terms of words. This is needed to avoid
2169 * unaligned accesses for some archs when redzoning is used, and makes
2170 * sure any on-slab bufctl's are also correctly aligned.
2172 if (size & (BYTES_PER_WORD - 1)) {
2173 size += (BYTES_PER_WORD - 1);
2174 size &= ~(BYTES_PER_WORD - 1);
2177 /* calculate the final buffer alignment: */
2179 /* 1) arch recommendation: can be overridden for debug */
2180 if (flags & SLAB_HWCACHE_ALIGN) {
2182 * Default alignment: as specified by the arch code. Except if
2183 * an object is really small, then squeeze multiple objects into
2186 ralign = cache_line_size();
2187 while (size <= ralign / 2)
2190 ralign = BYTES_PER_WORD;
2194 * Redzoning and user store require word alignment. Note this will be
2195 * overridden by architecture or caller mandated alignment if either
2196 * is greater than BYTES_PER_WORD.
2198 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2199 ralign = BYTES_PER_WORD;
2201 /* 2) arch mandated alignment: disables debug if necessary */
2202 if (ralign < ARCH_SLAB_MINALIGN) {
2203 ralign = ARCH_SLAB_MINALIGN;
2204 if (ralign > BYTES_PER_WORD)
2205 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2207 /* 3) caller mandated alignment: disables debug if necessary */
2208 if (ralign < align) {
2210 if (ralign > BYTES_PER_WORD)
2211 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2218 /* Get cache's description obj. */
2219 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2224 cachep->obj_size = size;
2227 * Both debugging options require word-alignment which is calculated
2230 if (flags & SLAB_RED_ZONE) {
2231 /* add space for red zone words */
2232 cachep->obj_offset += BYTES_PER_WORD;
2233 size += 2 * BYTES_PER_WORD;
2235 if (flags & SLAB_STORE_USER) {
2236 /* user store requires one word storage behind the end of
2239 size += BYTES_PER_WORD;
2241 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2242 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2243 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2244 cachep->obj_offset += PAGE_SIZE - size;
2251 * Determine if the slab management is 'on' or 'off' slab.
2252 * (bootstrapping cannot cope with offslab caches so don't do
2255 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2257 * Size is large, assume best to place the slab management obj
2258 * off-slab (should allow better packing of objs).
2260 flags |= CFLGS_OFF_SLAB;
2262 size = ALIGN(size, align);
2264 left_over = calculate_slab_order(cachep, size, align, flags);
2267 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2268 kmem_cache_free(&cache_cache, cachep);
2272 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2273 + sizeof(struct slab), align);
2276 * If the slab has been placed off-slab, and we have enough space then
2277 * move it on-slab. This is at the expense of any extra colouring.
2279 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2280 flags &= ~CFLGS_OFF_SLAB;
2281 left_over -= slab_size;
2284 if (flags & CFLGS_OFF_SLAB) {
2285 /* really off slab. No need for manual alignment */
2287 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2290 cachep->colour_off = cache_line_size();
2291 /* Offset must be a multiple of the alignment. */
2292 if (cachep->colour_off < align)
2293 cachep->colour_off = align;
2294 cachep->colour = left_over / cachep->colour_off;
2295 cachep->slab_size = slab_size;
2296 cachep->flags = flags;
2297 cachep->gfpflags = 0;
2298 if (flags & SLAB_CACHE_DMA)
2299 cachep->gfpflags |= GFP_DMA;
2300 cachep->buffer_size = size;
2302 if (flags & CFLGS_OFF_SLAB) {
2303 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2305 * This is a possibility for one of the malloc_sizes caches.
2306 * But since we go off slab only for object size greater than
2307 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2308 * this should not happen at all.
2309 * But leave a BUG_ON for some lucky dude.
2311 BUG_ON(!cachep->slabp_cache);
2313 cachep->ctor = ctor;
2314 cachep->dtor = dtor;
2315 cachep->name = name;
2317 if (setup_cpu_cache(cachep)) {
2318 __kmem_cache_destroy(cachep);
2323 /* cache setup completed, link it into the list */
2324 list_add(&cachep->next, &cache_chain);
2326 if (!cachep && (flags & SLAB_PANIC))
2327 panic("kmem_cache_create(): failed to create slab `%s'\n",
2329 mutex_unlock(&cache_chain_mutex);
2330 unlock_cpu_hotplug();
2333 EXPORT_SYMBOL(kmem_cache_create);
2336 static void check_irq_off(void)
2338 BUG_ON(!irqs_disabled());
2341 static void check_irq_on(void)
2343 BUG_ON(irqs_disabled());
2346 static void check_spinlock_acquired(struct kmem_cache *cachep)
2350 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2354 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2358 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2363 #define check_irq_off() do { } while(0)
2364 #define check_irq_on() do { } while(0)
2365 #define check_spinlock_acquired(x) do { } while(0)
2366 #define check_spinlock_acquired_node(x, y) do { } while(0)
2369 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2370 struct array_cache *ac,
2371 int force, int node);
2373 static void do_drain(void *arg)
2375 struct kmem_cache *cachep = arg;
2376 struct array_cache *ac;
2377 int node = numa_node_id();
2380 ac = cpu_cache_get(cachep);
2381 spin_lock(&cachep->nodelists[node]->list_lock);
2382 free_block(cachep, ac->entry, ac->avail, node);
2383 spin_unlock(&cachep->nodelists[node]->list_lock);
2387 static void drain_cpu_caches(struct kmem_cache *cachep)
2389 struct kmem_list3 *l3;
2392 on_each_cpu(do_drain, cachep, 1, 1);
2394 for_each_online_node(node) {
2395 l3 = cachep->nodelists[node];
2396 if (l3 && l3->alien)
2397 drain_alien_cache(cachep, l3->alien);
2400 for_each_online_node(node) {
2401 l3 = cachep->nodelists[node];
2403 drain_array(cachep, l3, l3->shared, 1, node);
2408 * Remove slabs from the list of free slabs.
2409 * Specify the number of slabs to drain in tofree.
2411 * Returns the actual number of slabs released.
2413 static int drain_freelist(struct kmem_cache *cache,
2414 struct kmem_list3 *l3, int tofree)
2416 struct list_head *p;
2421 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2423 spin_lock_irq(&l3->list_lock);
2424 p = l3->slabs_free.prev;
2425 if (p == &l3->slabs_free) {
2426 spin_unlock_irq(&l3->list_lock);
2430 slabp = list_entry(p, struct slab, list);
2432 BUG_ON(slabp->inuse);
2434 list_del(&slabp->list);
2436 * Safe to drop the lock. The slab is no longer linked
2439 l3->free_objects -= cache->num;
2440 spin_unlock_irq(&l3->list_lock);
2441 slab_destroy(cache, slabp);
2448 static int __cache_shrink(struct kmem_cache *cachep)
2451 struct kmem_list3 *l3;
2453 drain_cpu_caches(cachep);
2456 for_each_online_node(i) {
2457 l3 = cachep->nodelists[i];
2461 drain_freelist(cachep, l3, l3->free_objects);
2463 ret += !list_empty(&l3->slabs_full) ||
2464 !list_empty(&l3->slabs_partial);
2466 return (ret ? 1 : 0);
2470 * kmem_cache_shrink - Shrink a cache.
2471 * @cachep: The cache to shrink.
2473 * Releases as many slabs as possible for a cache.
2474 * To help debugging, a zero exit status indicates all slabs were released.
2476 int kmem_cache_shrink(struct kmem_cache *cachep)
2478 BUG_ON(!cachep || in_interrupt());
2480 return __cache_shrink(cachep);
2482 EXPORT_SYMBOL(kmem_cache_shrink);
2485 * kmem_cache_destroy - delete a cache
2486 * @cachep: the cache to destroy
2488 * Remove a struct kmem_cache object from the slab cache.
2490 * It is expected this function will be called by a module when it is
2491 * unloaded. This will remove the cache completely, and avoid a duplicate
2492 * cache being allocated each time a module is loaded and unloaded, if the
2493 * module doesn't have persistent in-kernel storage across loads and unloads.
2495 * The cache must be empty before calling this function.
2497 * The caller must guarantee that noone will allocate memory from the cache
2498 * during the kmem_cache_destroy().
2500 void kmem_cache_destroy(struct kmem_cache *cachep)
2502 BUG_ON(!cachep || in_interrupt());
2504 /* Don't let CPUs to come and go */
2507 /* Find the cache in the chain of caches. */
2508 mutex_lock(&cache_chain_mutex);
2510 * the chain is never empty, cache_cache is never destroyed
2512 list_del(&cachep->next);
2513 mutex_unlock(&cache_chain_mutex);
2515 if (__cache_shrink(cachep)) {
2516 slab_error(cachep, "Can't free all objects");
2517 mutex_lock(&cache_chain_mutex);
2518 list_add(&cachep->next, &cache_chain);
2519 mutex_unlock(&cache_chain_mutex);
2520 unlock_cpu_hotplug();
2524 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2527 __kmem_cache_destroy(cachep);
2528 unlock_cpu_hotplug();
2530 EXPORT_SYMBOL(kmem_cache_destroy);
2533 * Get the memory for a slab management obj.
2534 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2535 * always come from malloc_sizes caches. The slab descriptor cannot
2536 * come from the same cache which is getting created because,
2537 * when we are searching for an appropriate cache for these
2538 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2539 * If we are creating a malloc_sizes cache here it would not be visible to
2540 * kmem_find_general_cachep till the initialization is complete.
2541 * Hence we cannot have slabp_cache same as the original cache.
2543 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2544 int colour_off, gfp_t local_flags,
2549 if (OFF_SLAB(cachep)) {
2550 /* Slab management obj is off-slab. */
2551 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2552 local_flags, nodeid);
2556 slabp = objp + colour_off;
2557 colour_off += cachep->slab_size;
2560 slabp->colouroff = colour_off;
2561 slabp->s_mem = objp + colour_off;
2562 slabp->nodeid = nodeid;
2566 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2568 return (kmem_bufctl_t *) (slabp + 1);
2571 static void cache_init_objs(struct kmem_cache *cachep,
2572 struct slab *slabp, unsigned long ctor_flags)
2576 for (i = 0; i < cachep->num; i++) {
2577 void *objp = index_to_obj(cachep, slabp, i);
2579 /* need to poison the objs? */
2580 if (cachep->flags & SLAB_POISON)
2581 poison_obj(cachep, objp, POISON_FREE);
2582 if (cachep->flags & SLAB_STORE_USER)
2583 *dbg_userword(cachep, objp) = NULL;
2585 if (cachep->flags & SLAB_RED_ZONE) {
2586 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2587 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2590 * Constructors are not allowed to allocate memory from the same
2591 * cache which they are a constructor for. Otherwise, deadlock.
2592 * They must also be threaded.
2594 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2595 cachep->ctor(objp + obj_offset(cachep), cachep,
2598 if (cachep->flags & SLAB_RED_ZONE) {
2599 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2600 slab_error(cachep, "constructor overwrote the"
2601 " end of an object");
2602 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2603 slab_error(cachep, "constructor overwrote the"
2604 " start of an object");
2606 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2607 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2608 kernel_map_pages(virt_to_page(objp),
2609 cachep->buffer_size / PAGE_SIZE, 0);
2612 cachep->ctor(objp, cachep, ctor_flags);
2614 slab_bufctl(slabp)[i] = i + 1;
2616 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2620 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2622 if (flags & SLAB_DMA)
2623 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2625 BUG_ON(cachep->gfpflags & GFP_DMA);
2628 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2631 void *objp = index_to_obj(cachep, slabp, slabp->free);
2635 next = slab_bufctl(slabp)[slabp->free];
2637 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2638 WARN_ON(slabp->nodeid != nodeid);
2645 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2646 void *objp, int nodeid)
2648 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2651 /* Verify that the slab belongs to the intended node */
2652 WARN_ON(slabp->nodeid != nodeid);
2654 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2655 printk(KERN_ERR "slab: double free detected in cache "
2656 "'%s', objp %p\n", cachep->name, objp);
2660 slab_bufctl(slabp)[objnr] = slabp->free;
2661 slabp->free = objnr;
2666 * Map pages beginning at addr to the given cache and slab. This is required
2667 * for the slab allocator to be able to lookup the cache and slab of a
2668 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2670 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2676 page = virt_to_page(addr);
2679 if (likely(!PageCompound(page)))
2680 nr_pages <<= cache->gfporder;
2683 page_set_cache(page, cache);
2684 page_set_slab(page, slab);
2686 } while (--nr_pages);
2690 * Grow (by 1) the number of slabs within a cache. This is called by
2691 * kmem_cache_alloc() when there are no active objs left in a cache.
2693 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2699 unsigned long ctor_flags;
2700 struct kmem_list3 *l3;
2703 * Be lazy and only check for valid flags here, keeping it out of the
2704 * critical path in kmem_cache_alloc().
2706 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2707 if (flags & SLAB_NO_GROW)
2710 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2711 local_flags = (flags & SLAB_LEVEL_MASK);
2712 if (!(local_flags & __GFP_WAIT))
2714 * Not allowed to sleep. Need to tell a constructor about
2715 * this - it might need to know...
2717 ctor_flags |= SLAB_CTOR_ATOMIC;
2719 /* Take the l3 list lock to change the colour_next on this node */
2721 l3 = cachep->nodelists[nodeid];
2722 spin_lock(&l3->list_lock);
2724 /* Get colour for the slab, and cal the next value. */
2725 offset = l3->colour_next;
2727 if (l3->colour_next >= cachep->colour)
2728 l3->colour_next = 0;
2729 spin_unlock(&l3->list_lock);
2731 offset *= cachep->colour_off;
2733 if (local_flags & __GFP_WAIT)
2737 * The test for missing atomic flag is performed here, rather than
2738 * the more obvious place, simply to reduce the critical path length
2739 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2740 * will eventually be caught here (where it matters).
2742 kmem_flagcheck(cachep, flags);
2745 * Get mem for the objs. Attempt to allocate a physical page from
2748 objp = kmem_getpages(cachep, flags, nodeid);
2752 /* Get slab management. */
2753 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2757 slabp->nodeid = nodeid;
2758 slab_map_pages(cachep, slabp, objp);
2760 cache_init_objs(cachep, slabp, ctor_flags);
2762 if (local_flags & __GFP_WAIT)
2763 local_irq_disable();
2765 spin_lock(&l3->list_lock);
2767 /* Make slab active. */
2768 list_add_tail(&slabp->list, &(l3->slabs_free));
2769 STATS_INC_GROWN(cachep);
2770 l3->free_objects += cachep->num;
2771 spin_unlock(&l3->list_lock);
2774 kmem_freepages(cachep, objp);
2776 if (local_flags & __GFP_WAIT)
2777 local_irq_disable();
2784 * Perform extra freeing checks:
2785 * - detect bad pointers.
2786 * - POISON/RED_ZONE checking
2787 * - destructor calls, for caches with POISON+dtor
2789 static void kfree_debugcheck(const void *objp)
2793 if (!virt_addr_valid(objp)) {
2794 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2795 (unsigned long)objp);
2798 page = virt_to_page(objp);
2799 if (!PageSlab(page)) {
2800 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2801 (unsigned long)objp);
2806 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2808 unsigned long redzone1, redzone2;
2810 redzone1 = *dbg_redzone1(cache, obj);
2811 redzone2 = *dbg_redzone2(cache, obj);
2816 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2819 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2820 slab_error(cache, "double free detected");
2822 slab_error(cache, "memory outside object was overwritten");
2824 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2825 obj, redzone1, redzone2);
2828 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2835 objp -= obj_offset(cachep);
2836 kfree_debugcheck(objp);
2837 page = virt_to_page(objp);
2839 slabp = page_get_slab(page);
2841 if (cachep->flags & SLAB_RED_ZONE) {
2842 verify_redzone_free(cachep, objp);
2843 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2844 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2846 if (cachep->flags & SLAB_STORE_USER)
2847 *dbg_userword(cachep, objp) = caller;
2849 objnr = obj_to_index(cachep, slabp, objp);
2851 BUG_ON(objnr >= cachep->num);
2852 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2854 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2856 * Need to call the slab's constructor so the caller can
2857 * perform a verify of its state (debugging). Called without
2858 * the cache-lock held.
2860 cachep->ctor(objp + obj_offset(cachep),
2861 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2863 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2864 /* we want to cache poison the object,
2865 * call the destruction callback
2867 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2869 #ifdef CONFIG_DEBUG_SLAB_LEAK
2870 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2872 if (cachep->flags & SLAB_POISON) {
2873 #ifdef CONFIG_DEBUG_PAGEALLOC
2874 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2875 store_stackinfo(cachep, objp, (unsigned long)caller);
2876 kernel_map_pages(virt_to_page(objp),
2877 cachep->buffer_size / PAGE_SIZE, 0);
2879 poison_obj(cachep, objp, POISON_FREE);
2882 poison_obj(cachep, objp, POISON_FREE);
2888 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2893 /* Check slab's freelist to see if this obj is there. */
2894 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2896 if (entries > cachep->num || i >= cachep->num)
2899 if (entries != cachep->num - slabp->inuse) {
2901 printk(KERN_ERR "slab: Internal list corruption detected in "
2902 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2903 cachep->name, cachep->num, slabp, slabp->inuse);
2905 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2908 printk("\n%03x:", i);
2909 printk(" %02x", ((unsigned char *)slabp)[i]);
2916 #define kfree_debugcheck(x) do { } while(0)
2917 #define cache_free_debugcheck(x,objp,z) (objp)
2918 #define check_slabp(x,y) do { } while(0)
2921 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2924 struct kmem_list3 *l3;
2925 struct array_cache *ac;
2928 node = numa_node_id();
2931 ac = cpu_cache_get(cachep);
2933 batchcount = ac->batchcount;
2934 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2936 * If there was little recent activity on this cache, then
2937 * perform only a partial refill. Otherwise we could generate
2940 batchcount = BATCHREFILL_LIMIT;
2942 l3 = cachep->nodelists[node];
2944 BUG_ON(ac->avail > 0 || !l3);
2945 spin_lock(&l3->list_lock);
2947 /* See if we can refill from the shared array */
2948 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2951 while (batchcount > 0) {
2952 struct list_head *entry;
2954 /* Get slab alloc is to come from. */
2955 entry = l3->slabs_partial.next;
2956 if (entry == &l3->slabs_partial) {
2957 l3->free_touched = 1;
2958 entry = l3->slabs_free.next;
2959 if (entry == &l3->slabs_free)
2963 slabp = list_entry(entry, struct slab, list);
2964 check_slabp(cachep, slabp);
2965 check_spinlock_acquired(cachep);
2966 while (slabp->inuse < cachep->num && batchcount--) {
2967 STATS_INC_ALLOCED(cachep);
2968 STATS_INC_ACTIVE(cachep);
2969 STATS_SET_HIGH(cachep);
2971 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2974 check_slabp(cachep, slabp);
2976 /* move slabp to correct slabp list: */
2977 list_del(&slabp->list);
2978 if (slabp->free == BUFCTL_END)
2979 list_add(&slabp->list, &l3->slabs_full);
2981 list_add(&slabp->list, &l3->slabs_partial);
2985 l3->free_objects -= ac->avail;
2987 spin_unlock(&l3->list_lock);
2989 if (unlikely(!ac->avail)) {
2991 x = cache_grow(cachep, flags, node);
2993 /* cache_grow can reenable interrupts, then ac could change. */
2994 ac = cpu_cache_get(cachep);
2995 if (!x && ac->avail == 0) /* no objects in sight? abort */
2998 if (!ac->avail) /* objects refilled by interrupt? */
3002 return ac->entry[--ac->avail];
3005 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3008 might_sleep_if(flags & __GFP_WAIT);
3010 kmem_flagcheck(cachep, flags);
3015 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3016 gfp_t flags, void *objp, void *caller)
3020 if (cachep->flags & SLAB_POISON) {
3021 #ifdef CONFIG_DEBUG_PAGEALLOC
3022 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3023 kernel_map_pages(virt_to_page(objp),
3024 cachep->buffer_size / PAGE_SIZE, 1);
3026 check_poison_obj(cachep, objp);
3028 check_poison_obj(cachep, objp);
3030 poison_obj(cachep, objp, POISON_INUSE);
3032 if (cachep->flags & SLAB_STORE_USER)
3033 *dbg_userword(cachep, objp) = caller;
3035 if (cachep->flags & SLAB_RED_ZONE) {
3036 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3037 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3038 slab_error(cachep, "double free, or memory outside"
3039 " object was overwritten");
3041 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3042 objp, *dbg_redzone1(cachep, objp),
3043 *dbg_redzone2(cachep, objp));
3045 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3046 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3048 #ifdef CONFIG_DEBUG_SLAB_LEAK
3053 slabp = page_get_slab(virt_to_page(objp));
3054 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3055 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3058 objp += obj_offset(cachep);
3059 if (cachep->ctor && cachep->flags & SLAB_POISON) {
3060 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
3062 if (!(flags & __GFP_WAIT))
3063 ctor_flags |= SLAB_CTOR_ATOMIC;
3065 cachep->ctor(objp, cachep, ctor_flags);
3070 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3073 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3076 struct array_cache *ac;
3079 ac = cpu_cache_get(cachep);
3080 if (likely(ac->avail)) {
3081 STATS_INC_ALLOCHIT(cachep);
3083 objp = ac->entry[--ac->avail];
3085 STATS_INC_ALLOCMISS(cachep);
3086 objp = cache_alloc_refill(cachep, flags);
3091 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3092 gfp_t flags, void *caller)
3094 unsigned long save_flags;
3097 cache_alloc_debugcheck_before(cachep, flags);
3099 local_irq_save(save_flags);
3101 if (unlikely(NUMA_BUILD &&
3102 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
3103 objp = alternate_node_alloc(cachep, flags);
3106 objp = ____cache_alloc(cachep, flags);
3108 * We may just have run out of memory on the local node.
3109 * __cache_alloc_node() knows how to locate memory on other nodes
3111 if (NUMA_BUILD && !objp)
3112 objp = __cache_alloc_node(cachep, flags, numa_node_id());
3113 local_irq_restore(save_flags);
3114 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3122 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3124 * If we are in_interrupt, then process context, including cpusets and
3125 * mempolicy, may not apply and should not be used for allocation policy.
3127 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3129 int nid_alloc, nid_here;
3131 if (in_interrupt() || (flags & __GFP_THISNODE))
3133 nid_alloc = nid_here = numa_node_id();
3134 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3135 nid_alloc = cpuset_mem_spread_node();
3136 else if (current->mempolicy)
3137 nid_alloc = slab_node(current->mempolicy);
3138 if (nid_alloc != nid_here)
3139 return __cache_alloc_node(cachep, flags, nid_alloc);
3144 * Fallback function if there was no memory available and no objects on a
3145 * certain node and we are allowed to fall back. We mimick the behavior of
3146 * the page allocator. We fall back according to a zonelist determined by
3147 * the policy layer while obeying cpuset constraints.
3149 void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3151 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3152 ->node_zonelists[gfp_zone(flags)];
3156 for (z = zonelist->zones; *z && !obj; z++)
3157 if (zone_idx(*z) <= ZONE_NORMAL &&
3158 cpuset_zone_allowed(*z, flags))
3159 obj = __cache_alloc_node(cache,
3160 flags | __GFP_THISNODE,
3166 * A interface to enable slab creation on nodeid
3168 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3171 struct list_head *entry;
3173 struct kmem_list3 *l3;
3177 l3 = cachep->nodelists[nodeid];
3182 spin_lock(&l3->list_lock);
3183 entry = l3->slabs_partial.next;
3184 if (entry == &l3->slabs_partial) {
3185 l3->free_touched = 1;
3186 entry = l3->slabs_free.next;
3187 if (entry == &l3->slabs_free)
3191 slabp = list_entry(entry, struct slab, list);
3192 check_spinlock_acquired_node(cachep, nodeid);
3193 check_slabp(cachep, slabp);
3195 STATS_INC_NODEALLOCS(cachep);
3196 STATS_INC_ACTIVE(cachep);
3197 STATS_SET_HIGH(cachep);
3199 BUG_ON(slabp->inuse == cachep->num);
3201 obj = slab_get_obj(cachep, slabp, nodeid);
3202 check_slabp(cachep, slabp);
3204 /* move slabp to correct slabp list: */
3205 list_del(&slabp->list);
3207 if (slabp->free == BUFCTL_END)
3208 list_add(&slabp->list, &l3->slabs_full);
3210 list_add(&slabp->list, &l3->slabs_partial);
3212 spin_unlock(&l3->list_lock);
3216 spin_unlock(&l3->list_lock);
3217 x = cache_grow(cachep, flags, nodeid);
3221 if (!(flags & __GFP_THISNODE))
3222 /* Unable to grow the cache. Fall back to other nodes. */
3223 return fallback_alloc(cachep, flags);
3233 * Caller needs to acquire correct kmem_list's list_lock
3235 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3239 struct kmem_list3 *l3;
3241 for (i = 0; i < nr_objects; i++) {
3242 void *objp = objpp[i];
3245 slabp = virt_to_slab(objp);
3246 l3 = cachep->nodelists[node];
3247 list_del(&slabp->list);
3248 check_spinlock_acquired_node(cachep, node);
3249 check_slabp(cachep, slabp);
3250 slab_put_obj(cachep, slabp, objp, node);
3251 STATS_DEC_ACTIVE(cachep);
3253 check_slabp(cachep, slabp);
3255 /* fixup slab chains */
3256 if (slabp->inuse == 0) {
3257 if (l3->free_objects > l3->free_limit) {
3258 l3->free_objects -= cachep->num;
3259 /* No need to drop any previously held
3260 * lock here, even if we have a off-slab slab
3261 * descriptor it is guaranteed to come from
3262 * a different cache, refer to comments before
3265 slab_destroy(cachep, slabp);
3267 list_add(&slabp->list, &l3->slabs_free);
3270 /* Unconditionally move a slab to the end of the
3271 * partial list on free - maximum time for the
3272 * other objects to be freed, too.
3274 list_add_tail(&slabp->list, &l3->slabs_partial);
3279 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3282 struct kmem_list3 *l3;
3283 int node = numa_node_id();
3285 batchcount = ac->batchcount;
3287 BUG_ON(!batchcount || batchcount > ac->avail);
3290 l3 = cachep->nodelists[node];
3291 spin_lock(&l3->list_lock);
3293 struct array_cache *shared_array = l3->shared;
3294 int max = shared_array->limit - shared_array->avail;
3296 if (batchcount > max)
3298 memcpy(&(shared_array->entry[shared_array->avail]),
3299 ac->entry, sizeof(void *) * batchcount);
3300 shared_array->avail += batchcount;
3305 free_block(cachep, ac->entry, batchcount, node);
3310 struct list_head *p;
3312 p = l3->slabs_free.next;
3313 while (p != &(l3->slabs_free)) {
3316 slabp = list_entry(p, struct slab, list);
3317 BUG_ON(slabp->inuse);
3322 STATS_SET_FREEABLE(cachep, i);
3325 spin_unlock(&l3->list_lock);
3326 ac->avail -= batchcount;
3327 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3331 * Release an obj back to its cache. If the obj has a constructed state, it must
3332 * be in this state _before_ it is released. Called with disabled ints.
3334 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3336 struct array_cache *ac = cpu_cache_get(cachep);
3339 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3341 if (cache_free_alien(cachep, objp))
3344 if (likely(ac->avail < ac->limit)) {
3345 STATS_INC_FREEHIT(cachep);
3346 ac->entry[ac->avail++] = objp;
3349 STATS_INC_FREEMISS(cachep);
3350 cache_flusharray(cachep, ac);
3351 ac->entry[ac->avail++] = objp;
3356 * kmem_cache_alloc - Allocate an object
3357 * @cachep: The cache to allocate from.
3358 * @flags: See kmalloc().
3360 * Allocate an object from this cache. The flags are only relevant
3361 * if the cache has no available objects.
3363 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3365 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3367 EXPORT_SYMBOL(kmem_cache_alloc);
3370 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3371 * @cache: The cache to allocate from.
3372 * @flags: See kmalloc().
3374 * Allocate an object from this cache and set the allocated memory to zero.
3375 * The flags are only relevant if the cache has no available objects.
3377 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3379 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3381 memset(ret, 0, obj_size(cache));
3384 EXPORT_SYMBOL(kmem_cache_zalloc);
3387 * kmem_ptr_validate - check if an untrusted pointer might
3389 * @cachep: the cache we're checking against
3390 * @ptr: pointer to validate
3392 * This verifies that the untrusted pointer looks sane:
3393 * it is _not_ a guarantee that the pointer is actually
3394 * part of the slab cache in question, but it at least
3395 * validates that the pointer can be dereferenced and
3396 * looks half-way sane.
3398 * Currently only used for dentry validation.
3400 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3402 unsigned long addr = (unsigned long)ptr;
3403 unsigned long min_addr = PAGE_OFFSET;
3404 unsigned long align_mask = BYTES_PER_WORD - 1;
3405 unsigned long size = cachep->buffer_size;
3408 if (unlikely(addr < min_addr))
3410 if (unlikely(addr > (unsigned long)high_memory - size))
3412 if (unlikely(addr & align_mask))
3414 if (unlikely(!kern_addr_valid(addr)))
3416 if (unlikely(!kern_addr_valid(addr + size - 1)))
3418 page = virt_to_page(ptr);
3419 if (unlikely(!PageSlab(page)))
3421 if (unlikely(page_get_cache(page) != cachep))
3430 * kmem_cache_alloc_node - Allocate an object on the specified node
3431 * @cachep: The cache to allocate from.
3432 * @flags: See kmalloc().
3433 * @nodeid: node number of the target node.
3435 * Identical to kmem_cache_alloc, except that this function is slow
3436 * and can sleep. And it will allocate memory on the given node, which
3437 * can improve the performance for cpu bound structures.
3438 * New and improved: it will now make sure that the object gets
3439 * put on the correct node list so that there is no false sharing.
3441 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3443 unsigned long save_flags;
3446 cache_alloc_debugcheck_before(cachep, flags);
3447 local_irq_save(save_flags);
3449 if (nodeid == -1 || nodeid == numa_node_id() ||
3450 !cachep->nodelists[nodeid])
3451 ptr = ____cache_alloc(cachep, flags);
3453 ptr = __cache_alloc_node(cachep, flags, nodeid);
3454 local_irq_restore(save_flags);
3456 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3457 __builtin_return_address(0));
3461 EXPORT_SYMBOL(kmem_cache_alloc_node);
3463 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3465 struct kmem_cache *cachep;
3467 cachep = kmem_find_general_cachep(size, flags);
3468 if (unlikely(cachep == NULL))
3470 return kmem_cache_alloc_node(cachep, flags, node);
3472 EXPORT_SYMBOL(__kmalloc_node);
3476 * __do_kmalloc - allocate memory
3477 * @size: how many bytes of memory are required.
3478 * @flags: the type of memory to allocate (see kmalloc).
3479 * @caller: function caller for debug tracking of the caller
3481 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3484 struct kmem_cache *cachep;
3486 /* If you want to save a few bytes .text space: replace
3488 * Then kmalloc uses the uninlined functions instead of the inline
3491 cachep = __find_general_cachep(size, flags);
3492 if (unlikely(cachep == NULL))
3494 return __cache_alloc(cachep, flags, caller);
3498 #ifdef CONFIG_DEBUG_SLAB
3499 void *__kmalloc(size_t size, gfp_t flags)
3501 return __do_kmalloc(size, flags, __builtin_return_address(0));
3503 EXPORT_SYMBOL(__kmalloc);
3505 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3507 return __do_kmalloc(size, flags, caller);
3509 EXPORT_SYMBOL(__kmalloc_track_caller);
3512 void *__kmalloc(size_t size, gfp_t flags)
3514 return __do_kmalloc(size, flags, NULL);
3516 EXPORT_SYMBOL(__kmalloc);
3520 * kmem_cache_free - Deallocate an object
3521 * @cachep: The cache the allocation was from.
3522 * @objp: The previously allocated object.
3524 * Free an object which was previously allocated from this
3527 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3529 unsigned long flags;
3531 BUG_ON(virt_to_cache(objp) != cachep);
3533 local_irq_save(flags);
3534 __cache_free(cachep, objp);
3535 local_irq_restore(flags);
3537 EXPORT_SYMBOL(kmem_cache_free);
3540 * kfree - free previously allocated memory
3541 * @objp: pointer returned by kmalloc.
3543 * If @objp is NULL, no operation is performed.
3545 * Don't free memory not originally allocated by kmalloc()
3546 * or you will run into trouble.
3548 void kfree(const void *objp)
3550 struct kmem_cache *c;
3551 unsigned long flags;
3553 if (unlikely(!objp))
3555 local_irq_save(flags);
3556 kfree_debugcheck(objp);
3557 c = virt_to_cache(objp);
3558 debug_check_no_locks_freed(objp, obj_size(c));
3559 __cache_free(c, (void *)objp);
3560 local_irq_restore(flags);
3562 EXPORT_SYMBOL(kfree);
3564 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3566 return obj_size(cachep);
3568 EXPORT_SYMBOL(kmem_cache_size);
3570 const char *kmem_cache_name(struct kmem_cache *cachep)
3572 return cachep->name;
3574 EXPORT_SYMBOL_GPL(kmem_cache_name);
3577 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3579 static int alloc_kmemlist(struct kmem_cache *cachep)
3582 struct kmem_list3 *l3;
3583 struct array_cache *new_shared;
3584 struct array_cache **new_alien;
3586 for_each_online_node(node) {
3588 new_alien = alloc_alien_cache(node, cachep->limit);
3592 new_shared = alloc_arraycache(node,
3593 cachep->shared*cachep->batchcount,
3596 free_alien_cache(new_alien);
3600 l3 = cachep->nodelists[node];
3602 struct array_cache *shared = l3->shared;
3604 spin_lock_irq(&l3->list_lock);
3607 free_block(cachep, shared->entry,
3608 shared->avail, node);
3610 l3->shared = new_shared;
3612 l3->alien = new_alien;
3615 l3->free_limit = (1 + nr_cpus_node(node)) *
3616 cachep->batchcount + cachep->num;
3617 spin_unlock_irq(&l3->list_lock);
3619 free_alien_cache(new_alien);
3622 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3624 free_alien_cache(new_alien);
3629 kmem_list3_init(l3);
3630 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3631 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3632 l3->shared = new_shared;
3633 l3->alien = new_alien;
3634 l3->free_limit = (1 + nr_cpus_node(node)) *
3635 cachep->batchcount + cachep->num;
3636 cachep->nodelists[node] = l3;
3641 if (!cachep->next.next) {
3642 /* Cache is not active yet. Roll back what we did */
3645 if (cachep->nodelists[node]) {
3646 l3 = cachep->nodelists[node];
3649 free_alien_cache(l3->alien);
3651 cachep->nodelists[node] = NULL;
3659 struct ccupdate_struct {
3660 struct kmem_cache *cachep;
3661 struct array_cache *new[NR_CPUS];
3664 static void do_ccupdate_local(void *info)
3666 struct ccupdate_struct *new = info;
3667 struct array_cache *old;
3670 old = cpu_cache_get(new->cachep);
3672 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3673 new->new[smp_processor_id()] = old;
3676 /* Always called with the cache_chain_mutex held */
3677 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3678 int batchcount, int shared)
3680 struct ccupdate_struct *new;
3683 new = kzalloc(sizeof(*new), GFP_KERNEL);
3687 for_each_online_cpu(i) {
3688 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3691 for (i--; i >= 0; i--)
3697 new->cachep = cachep;
3699 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3702 cachep->batchcount = batchcount;
3703 cachep->limit = limit;
3704 cachep->shared = shared;
3706 for_each_online_cpu(i) {
3707 struct array_cache *ccold = new->new[i];
3710 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3711 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3712 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3716 return alloc_kmemlist(cachep);
3719 /* Called with cache_chain_mutex held always */
3720 static int enable_cpucache(struct kmem_cache *cachep)
3726 * The head array serves three purposes:
3727 * - create a LIFO ordering, i.e. return objects that are cache-warm
3728 * - reduce the number of spinlock operations.
3729 * - reduce the number of linked list operations on the slab and
3730 * bufctl chains: array operations are cheaper.
3731 * The numbers are guessed, we should auto-tune as described by
3734 if (cachep->buffer_size > 131072)
3736 else if (cachep->buffer_size > PAGE_SIZE)
3738 else if (cachep->buffer_size > 1024)
3740 else if (cachep->buffer_size > 256)
3746 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3747 * allocation behaviour: Most allocs on one cpu, most free operations
3748 * on another cpu. For these cases, an efficient object passing between
3749 * cpus is necessary. This is provided by a shared array. The array
3750 * replaces Bonwick's magazine layer.
3751 * On uniprocessor, it's functionally equivalent (but less efficient)
3752 * to a larger limit. Thus disabled by default.
3756 if (cachep->buffer_size <= PAGE_SIZE)
3762 * With debugging enabled, large batchcount lead to excessively long
3763 * periods with disabled local interrupts. Limit the batchcount
3768 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3770 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3771 cachep->name, -err);
3776 * Drain an array if it contains any elements taking the l3 lock only if
3777 * necessary. Note that the l3 listlock also protects the array_cache
3778 * if drain_array() is used on the shared array.
3780 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3781 struct array_cache *ac, int force, int node)
3785 if (!ac || !ac->avail)
3787 if (ac->touched && !force) {
3790 spin_lock_irq(&l3->list_lock);
3792 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3793 if (tofree > ac->avail)
3794 tofree = (ac->avail + 1) / 2;
3795 free_block(cachep, ac->entry, tofree, node);
3796 ac->avail -= tofree;
3797 memmove(ac->entry, &(ac->entry[tofree]),
3798 sizeof(void *) * ac->avail);
3800 spin_unlock_irq(&l3->list_lock);
3805 * cache_reap - Reclaim memory from caches.
3806 * @unused: unused parameter
3808 * Called from workqueue/eventd every few seconds.
3810 * - clear the per-cpu caches for this CPU.
3811 * - return freeable pages to the main free memory pool.
3813 * If we cannot acquire the cache chain mutex then just give up - we'll try
3814 * again on the next iteration.
3816 static void cache_reap(void *unused)
3818 struct kmem_cache *searchp;
3819 struct kmem_list3 *l3;
3820 int node = numa_node_id();
3822 if (!mutex_trylock(&cache_chain_mutex)) {
3823 /* Give up. Setup the next iteration. */
3824 schedule_delayed_work(&__get_cpu_var(reap_work),
3829 list_for_each_entry(searchp, &cache_chain, next) {
3833 * We only take the l3 lock if absolutely necessary and we
3834 * have established with reasonable certainty that
3835 * we can do some work if the lock was obtained.
3837 l3 = searchp->nodelists[node];
3839 reap_alien(searchp, l3);
3841 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3844 * These are racy checks but it does not matter
3845 * if we skip one check or scan twice.
3847 if (time_after(l3->next_reap, jiffies))
3850 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3852 drain_array(searchp, l3, l3->shared, 0, node);
3854 if (l3->free_touched)
3855 l3->free_touched = 0;
3859 freed = drain_freelist(searchp, l3, (l3->free_limit +
3860 5 * searchp->num - 1) / (5 * searchp->num));
3861 STATS_ADD_REAPED(searchp, freed);
3867 mutex_unlock(&cache_chain_mutex);
3869 refresh_cpu_vm_stats(smp_processor_id());
3870 /* Set up the next iteration */
3871 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3874 #ifdef CONFIG_PROC_FS
3876 static void print_slabinfo_header(struct seq_file *m)
3879 * Output format version, so at least we can change it
3880 * without _too_ many complaints.
3883 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3885 seq_puts(m, "slabinfo - version: 2.1\n");
3887 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3888 "<objperslab> <pagesperslab>");
3889 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3890 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3892 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3893 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3894 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3899 static void *s_start(struct seq_file *m, loff_t *pos)
3902 struct list_head *p;
3904 mutex_lock(&cache_chain_mutex);
3906 print_slabinfo_header(m);
3907 p = cache_chain.next;
3910 if (p == &cache_chain)
3913 return list_entry(p, struct kmem_cache, next);
3916 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3918 struct kmem_cache *cachep = p;
3920 return cachep->next.next == &cache_chain ?
3921 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3924 static void s_stop(struct seq_file *m, void *p)
3926 mutex_unlock(&cache_chain_mutex);
3929 static int s_show(struct seq_file *m, void *p)
3931 struct kmem_cache *cachep = p;
3933 unsigned long active_objs;
3934 unsigned long num_objs;
3935 unsigned long active_slabs = 0;
3936 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3940 struct kmem_list3 *l3;
3944 for_each_online_node(node) {
3945 l3 = cachep->nodelists[node];
3950 spin_lock_irq(&l3->list_lock);
3952 list_for_each_entry(slabp, &l3->slabs_full, list) {
3953 if (slabp->inuse != cachep->num && !error)
3954 error = "slabs_full accounting error";
3955 active_objs += cachep->num;
3958 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3959 if (slabp->inuse == cachep->num && !error)
3960 error = "slabs_partial inuse accounting error";
3961 if (!slabp->inuse && !error)
3962 error = "slabs_partial/inuse accounting error";
3963 active_objs += slabp->inuse;
3966 list_for_each_entry(slabp, &l3->slabs_free, list) {
3967 if (slabp->inuse && !error)
3968 error = "slabs_free/inuse accounting error";
3971 free_objects += l3->free_objects;
3973 shared_avail += l3->shared->avail;
3975 spin_unlock_irq(&l3->list_lock);
3977 num_slabs += active_slabs;
3978 num_objs = num_slabs * cachep->num;
3979 if (num_objs - active_objs != free_objects && !error)
3980 error = "free_objects accounting error";
3982 name = cachep->name;
3984 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3986 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3987 name, active_objs, num_objs, cachep->buffer_size,
3988 cachep->num, (1 << cachep->gfporder));
3989 seq_printf(m, " : tunables %4u %4u %4u",
3990 cachep->limit, cachep->batchcount, cachep->shared);
3991 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3992 active_slabs, num_slabs, shared_avail);
3995 unsigned long high = cachep->high_mark;
3996 unsigned long allocs = cachep->num_allocations;
3997 unsigned long grown = cachep->grown;
3998 unsigned long reaped = cachep->reaped;
3999 unsigned long errors = cachep->errors;
4000 unsigned long max_freeable = cachep->max_freeable;
4001 unsigned long node_allocs = cachep->node_allocs;
4002 unsigned long node_frees = cachep->node_frees;
4003 unsigned long overflows = cachep->node_overflow;
4005 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4006 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4007 reaped, errors, max_freeable, node_allocs,
4008 node_frees, overflows);
4012 unsigned long allochit = atomic_read(&cachep->allochit);
4013 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4014 unsigned long freehit = atomic_read(&cachep->freehit);
4015 unsigned long freemiss = atomic_read(&cachep->freemiss);
4017 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4018 allochit, allocmiss, freehit, freemiss);
4026 * slabinfo_op - iterator that generates /proc/slabinfo
4035 * num-pages-per-slab
4036 * + further values on SMP and with statistics enabled
4039 struct seq_operations slabinfo_op = {
4046 #define MAX_SLABINFO_WRITE 128
4048 * slabinfo_write - Tuning for the slab allocator
4050 * @buffer: user buffer
4051 * @count: data length
4054 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4055 size_t count, loff_t *ppos)
4057 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4058 int limit, batchcount, shared, res;
4059 struct kmem_cache *cachep;
4061 if (count > MAX_SLABINFO_WRITE)
4063 if (copy_from_user(&kbuf, buffer, count))
4065 kbuf[MAX_SLABINFO_WRITE] = '\0';
4067 tmp = strchr(kbuf, ' ');
4072 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4075 /* Find the cache in the chain of caches. */
4076 mutex_lock(&cache_chain_mutex);
4078 list_for_each_entry(cachep, &cache_chain, next) {
4079 if (!strcmp(cachep->name, kbuf)) {
4080 if (limit < 1 || batchcount < 1 ||
4081 batchcount > limit || shared < 0) {
4084 res = do_tune_cpucache(cachep, limit,
4085 batchcount, shared);
4090 mutex_unlock(&cache_chain_mutex);
4096 #ifdef CONFIG_DEBUG_SLAB_LEAK
4098 static void *leaks_start(struct seq_file *m, loff_t *pos)
4101 struct list_head *p;
4103 mutex_lock(&cache_chain_mutex);
4104 p = cache_chain.next;
4107 if (p == &cache_chain)
4110 return list_entry(p, struct kmem_cache, next);
4113 static inline int add_caller(unsigned long *n, unsigned long v)
4123 unsigned long *q = p + 2 * i;
4137 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4143 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4149 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4150 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4152 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4157 static void show_symbol(struct seq_file *m, unsigned long address)
4159 #ifdef CONFIG_KALLSYMS
4162 unsigned long offset, size;
4163 char namebuf[KSYM_NAME_LEN+1];
4165 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4168 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4170 seq_printf(m, " [%s]", modname);
4174 seq_printf(m, "%p", (void *)address);
4177 static int leaks_show(struct seq_file *m, void *p)
4179 struct kmem_cache *cachep = p;
4181 struct kmem_list3 *l3;
4183 unsigned long *n = m->private;
4187 if (!(cachep->flags & SLAB_STORE_USER))
4189 if (!(cachep->flags & SLAB_RED_ZONE))
4192 /* OK, we can do it */
4196 for_each_online_node(node) {
4197 l3 = cachep->nodelists[node];
4202 spin_lock_irq(&l3->list_lock);
4204 list_for_each_entry(slabp, &l3->slabs_full, list)
4205 handle_slab(n, cachep, slabp);
4206 list_for_each_entry(slabp, &l3->slabs_partial, list)
4207 handle_slab(n, cachep, slabp);
4208 spin_unlock_irq(&l3->list_lock);
4210 name = cachep->name;
4212 /* Increase the buffer size */
4213 mutex_unlock(&cache_chain_mutex);
4214 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4216 /* Too bad, we are really out */
4218 mutex_lock(&cache_chain_mutex);
4221 *(unsigned long *)m->private = n[0] * 2;
4223 mutex_lock(&cache_chain_mutex);
4224 /* Now make sure this entry will be retried */
4228 for (i = 0; i < n[1]; i++) {
4229 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4230 show_symbol(m, n[2*i+2]);
4237 struct seq_operations slabstats_op = {
4238 .start = leaks_start,
4247 * ksize - get the actual amount of memory allocated for a given object
4248 * @objp: Pointer to the object
4250 * kmalloc may internally round up allocations and return more memory
4251 * than requested. ksize() can be used to determine the actual amount of
4252 * memory allocated. The caller may use this additional memory, even though
4253 * a smaller amount of memory was initially specified with the kmalloc call.
4254 * The caller must guarantee that objp points to a valid object previously
4255 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4256 * must not be freed during the duration of the call.
4258 unsigned int ksize(const void *objp)
4260 if (unlikely(objp == NULL))
4263 return obj_size(virt_to_cache(objp));