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 kmem_cache_t and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/config.h>
90 #include <linux/slab.h>
92 #include <linux/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/seq_file.h>
98 #include <linux/notifier.h>
99 #include <linux/kallsyms.h>
100 #include <linux/cpu.h>
101 #include <linux/sysctl.h>
102 #include <linux/module.h>
103 #include <linux/rcupdate.h>
104 #include <linux/string.h>
105 #include <linux/nodemask.h>
106 #include <linux/mempolicy.h>
107 #include <linux/mutex.h>
109 #include <asm/uaccess.h>
110 #include <asm/cacheflush.h>
111 #include <asm/tlbflush.h>
112 #include <asm/page.h>
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
125 #ifdef CONFIG_DEBUG_SLAB
128 #define FORCED_DEBUG 1
132 #define FORCED_DEBUG 0
135 /* Shouldn't this be in a header file somewhere? */
136 #define BYTES_PER_WORD sizeof(void *)
138 #ifndef cache_line_size
139 #define cache_line_size() L1_CACHE_BYTES
142 #ifndef ARCH_KMALLOC_MINALIGN
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
151 #define ARCH_KMALLOC_MINALIGN 0
154 #ifndef ARCH_SLAB_MINALIGN
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
162 #define ARCH_SLAB_MINALIGN 0
165 #ifndef ARCH_KMALLOC_FLAGS
166 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
169 /* Legal flag mask for kmem_cache_create(). */
171 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 * Bufctl's are used for linking objs within a slab
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
203 typedef unsigned int kmem_bufctl_t;
204 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
208 /* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
211 static unsigned long offslab_limit;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 kmem_cache_t *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
277 /* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
287 * The slab lists for all objects.
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned long next_reap;
296 unsigned int free_limit;
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
312 * This function must be completely optimized away if
313 * a constant is passed to it. Mostly the same as
314 * what is in linux/slab.h except it returns an
317 static __always_inline int index_of(const size_t size)
319 extern void __bad_size(void);
321 if (__builtin_constant_p(size)) {
329 #include "linux/kmalloc_sizes.h"
337 #define INDEX_AC index_of(sizeof(struct arraycache_init))
338 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
340 static inline void kmem_list3_init(struct kmem_list3 *parent)
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 spin_lock_init(&parent->list_lock);
348 parent->free_objects = 0;
349 parent->free_touched = 0;
352 #define MAKE_LIST(cachep, listp, slab, nodeid) \
354 INIT_LIST_HEAD(listp); \
355 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
358 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
360 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
372 /* 1) per-cpu data, touched during every alloc/free */
373 struct array_cache *array[NR_CPUS];
374 unsigned int batchcount;
377 unsigned int buffer_size;
378 /* 2) touched by every alloc & free from the backend */
379 struct kmem_list3 *nodelists[MAX_NUMNODES];
380 unsigned int flags; /* constant flags */
381 unsigned int num; /* # of objs per slab */
384 /* 3) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
386 unsigned int gfporder;
388 /* force GFP flags, e.g. GFP_DMA */
391 size_t colour; /* cache colouring range */
392 unsigned int colour_off; /* colour offset */
393 unsigned int colour_next; /* cache colouring */
394 kmem_cache_t *slabp_cache;
395 unsigned int slab_size;
396 unsigned int dflags; /* dynamic flags */
398 /* constructor func */
399 void (*ctor) (void *, kmem_cache_t *, unsigned long);
401 /* de-constructor func */
402 void (*dtor) (void *, kmem_cache_t *, unsigned long);
404 /* 4) cache creation/removal */
406 struct list_head next;
410 unsigned long num_active;
411 unsigned long num_allocations;
412 unsigned long high_mark;
414 unsigned long reaped;
415 unsigned long errors;
416 unsigned long max_freeable;
417 unsigned long node_allocs;
418 unsigned long node_frees;
426 * If debugging is enabled, then the allocator can add additional
427 * fields and/or padding to every object. buffer_size contains the total
428 * object size including these internal fields, the following two
429 * variables contain the offset to the user object and its size.
436 #define CFLGS_OFF_SLAB (0x80000000UL)
437 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
439 #define BATCHREFILL_LIMIT 16
440 /* Optimization question: fewer reaps means less
441 * probability for unnessary cpucache drain/refill cycles.
443 * OTOH the cpuarrays can contain lots of objects,
444 * which could lock up otherwise freeable slabs.
446 #define REAPTIMEOUT_CPUC (2*HZ)
447 #define REAPTIMEOUT_LIST3 (4*HZ)
450 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
451 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
452 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
453 #define STATS_INC_GROWN(x) ((x)->grown++)
454 #define STATS_INC_REAPED(x) ((x)->reaped++)
455 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
456 (x)->high_mark = (x)->num_active; \
458 #define STATS_INC_ERR(x) ((x)->errors++)
459 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
460 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
461 #define STATS_SET_FREEABLE(x, i) \
462 do { if ((x)->max_freeable < i) \
463 (x)->max_freeable = i; \
466 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
467 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
468 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
469 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
471 #define STATS_INC_ACTIVE(x) do { } while (0)
472 #define STATS_DEC_ACTIVE(x) do { } while (0)
473 #define STATS_INC_ALLOCED(x) do { } while (0)
474 #define STATS_INC_GROWN(x) do { } while (0)
475 #define STATS_INC_REAPED(x) do { } while (0)
476 #define STATS_SET_HIGH(x) do { } while (0)
477 #define STATS_INC_ERR(x) do { } while (0)
478 #define STATS_INC_NODEALLOCS(x) do { } while (0)
479 #define STATS_INC_NODEFREES(x) do { } while (0)
480 #define STATS_SET_FREEABLE(x, i) \
483 #define STATS_INC_ALLOCHIT(x) do { } while (0)
484 #define STATS_INC_ALLOCMISS(x) do { } while (0)
485 #define STATS_INC_FREEHIT(x) do { } while (0)
486 #define STATS_INC_FREEMISS(x) do { } while (0)
490 /* Magic nums for obj red zoning.
491 * Placed in the first word before and the first word after an obj.
493 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
494 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
496 /* ...and for poisoning */
497 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
498 #define POISON_FREE 0x6b /* for use-after-free poisoning */
499 #define POISON_END 0xa5 /* end-byte of poisoning */
501 /* memory layout of objects:
503 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
504 * the end of an object is aligned with the end of the real
505 * allocation. Catches writes behind the end of the allocation.
506 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
508 * cachep->obj_offset: The real object.
509 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
510 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
512 static int obj_offset(kmem_cache_t *cachep)
514 return cachep->obj_offset;
517 static int obj_size(kmem_cache_t *cachep)
519 return cachep->obj_size;
522 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
525 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
528 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 if (cachep->flags & SLAB_STORE_USER)
532 return (unsigned long *)(objp + cachep->buffer_size -
534 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
537 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
539 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
540 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
545 #define obj_offset(x) 0
546 #define obj_size(cachep) (cachep->buffer_size)
547 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
548 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
554 * Maximum size of an obj (in 2^order pages)
555 * and absolute limit for the gfp order.
557 #if defined(CONFIG_LARGE_ALLOCS)
558 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
559 #define MAX_GFP_ORDER 13 /* up to 32Mb */
560 #elif defined(CONFIG_MMU)
561 #define MAX_OBJ_ORDER 5 /* 32 pages */
562 #define MAX_GFP_ORDER 5 /* 32 pages */
564 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
565 #define MAX_GFP_ORDER 8 /* up to 1Mb */
569 * Do not go above this order unless 0 objects fit into the slab.
571 #define BREAK_GFP_ORDER_HI 1
572 #define BREAK_GFP_ORDER_LO 0
573 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
575 /* Functions for storing/retrieving the cachep and or slab from the
576 * global 'mem_map'. These are used to find the slab an obj belongs to.
577 * With kfree(), these are used to find the cache which an obj belongs to.
579 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
581 page->lru.next = (struct list_head *)cache;
584 static inline struct kmem_cache *page_get_cache(struct page *page)
586 return (struct kmem_cache *)page->lru.next;
589 static inline void page_set_slab(struct page *page, struct slab *slab)
591 page->lru.prev = (struct list_head *)slab;
594 static inline struct slab *page_get_slab(struct page *page)
596 return (struct slab *)page->lru.prev;
599 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
600 struct cache_sizes malloc_sizes[] = {
601 #define CACHE(x) { .cs_size = (x) },
602 #include <linux/kmalloc_sizes.h>
606 EXPORT_SYMBOL(malloc_sizes);
608 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
614 static struct cache_names __initdata cache_names[] = {
615 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
616 #include <linux/kmalloc_sizes.h>
621 static struct arraycache_init initarray_cache __initdata =
622 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
623 static struct arraycache_init initarray_generic =
624 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
626 /* internal cache of cache description objs */
627 static kmem_cache_t cache_cache = {
629 .limit = BOOT_CPUCACHE_ENTRIES,
631 .buffer_size = sizeof(kmem_cache_t),
632 .flags = SLAB_NO_REAP,
633 .spinlock = SPIN_LOCK_UNLOCKED,
634 .name = "kmem_cache",
636 .obj_size = sizeof(kmem_cache_t),
640 /* Guard access to the cache-chain. */
641 static DEFINE_MUTEX(cache_chain_mutex);
642 static struct list_head cache_chain;
645 * vm_enough_memory() looks at this to determine how many
646 * slab-allocated pages are possibly freeable under pressure
648 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
650 atomic_t slab_reclaim_pages;
653 * chicken and egg problem: delay the per-cpu array allocation
654 * until the general caches are up.
663 static DEFINE_PER_CPU(struct work_struct, reap_work);
665 static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node);
666 static void enable_cpucache(kmem_cache_t *cachep);
667 static void cache_reap(void *unused);
668 static int __node_shrink(kmem_cache_t *cachep, int node);
670 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
672 return cachep->array[smp_processor_id()];
675 static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags)
677 struct cache_sizes *csizep = malloc_sizes;
680 /* This happens if someone tries to call
681 * kmem_cache_create(), or __kmalloc(), before
682 * the generic caches are initialized.
684 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
686 while (size > csizep->cs_size)
690 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
691 * has cs_{dma,}cachep==NULL. Thus no special case
692 * for large kmalloc calls required.
694 if (unlikely(gfpflags & GFP_DMA))
695 return csizep->cs_dmacachep;
696 return csizep->cs_cachep;
699 kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
701 return __find_general_cachep(size, gfpflags);
703 EXPORT_SYMBOL(kmem_find_general_cachep);
705 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
706 static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
707 int flags, size_t *left_over, unsigned int *num)
710 size_t wastage = PAGE_SIZE << gfporder;
714 if (!(flags & CFLGS_OFF_SLAB)) {
715 base = sizeof(struct slab);
716 extra = sizeof(kmem_bufctl_t);
719 while (i * size + ALIGN(base + i * extra, align) <= wastage)
729 wastage -= ALIGN(base + i * extra, align);
730 *left_over = wastage;
733 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
735 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
737 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
738 function, cachep->name, msg);
743 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
744 * via the workqueue/eventd.
745 * Add the CPU number into the expiration time to minimize the possibility of
746 * the CPUs getting into lockstep and contending for the global cache chain
749 static void __devinit start_cpu_timer(int cpu)
751 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
754 * When this gets called from do_initcalls via cpucache_init(),
755 * init_workqueues() has already run, so keventd will be setup
758 if (keventd_up() && reap_work->func == NULL) {
759 INIT_WORK(reap_work, cache_reap, NULL);
760 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
764 static struct array_cache *alloc_arraycache(int node, int entries,
767 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
768 struct array_cache *nc = NULL;
770 nc = kmalloc_node(memsize, GFP_KERNEL, node);
774 nc->batchcount = batchcount;
776 spin_lock_init(&nc->lock);
782 static void *__cache_alloc_node(kmem_cache_t *, gfp_t, int);
784 static inline struct array_cache **alloc_alien_cache(int node, int limit)
786 struct array_cache **ac_ptr;
787 int memsize = sizeof(void *) * MAX_NUMNODES;
792 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
795 if (i == node || !node_online(i)) {
799 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
801 for (i--; i <= 0; i--)
811 static inline void free_alien_cache(struct array_cache **ac_ptr)
824 static inline void __drain_alien_cache(kmem_cache_t *cachep,
825 struct array_cache *ac, int node)
827 struct kmem_list3 *rl3 = cachep->nodelists[node];
830 spin_lock(&rl3->list_lock);
831 free_block(cachep, ac->entry, ac->avail, node);
833 spin_unlock(&rl3->list_lock);
837 static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
840 struct array_cache *ac;
843 for_each_online_node(i) {
846 spin_lock_irqsave(&ac->lock, flags);
847 __drain_alien_cache(cachep, ac, i);
848 spin_unlock_irqrestore(&ac->lock, flags);
853 #define alloc_alien_cache(node, limit) do { } while (0)
854 #define free_alien_cache(ac_ptr) do { } while (0)
855 #define drain_alien_cache(cachep, l3) do { } while (0)
858 static int __devinit cpuup_callback(struct notifier_block *nfb,
859 unsigned long action, void *hcpu)
861 long cpu = (long)hcpu;
862 kmem_cache_t *cachep;
863 struct kmem_list3 *l3 = NULL;
864 int node = cpu_to_node(cpu);
865 int memsize = sizeof(struct kmem_list3);
869 mutex_lock(&cache_chain_mutex);
870 /* we need to do this right in the beginning since
871 * alloc_arraycache's are going to use this list.
872 * kmalloc_node allows us to add the slab to the right
873 * kmem_list3 and not this cpu's kmem_list3
876 list_for_each_entry(cachep, &cache_chain, next) {
877 /* setup the size64 kmemlist for cpu before we can
878 * begin anything. Make sure some other cpu on this
879 * node has not already allocated this
881 if (!cachep->nodelists[node]) {
882 if (!(l3 = kmalloc_node(memsize,
886 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
887 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
889 cachep->nodelists[node] = l3;
892 spin_lock_irq(&cachep->nodelists[node]->list_lock);
893 cachep->nodelists[node]->free_limit =
894 (1 + nr_cpus_node(node)) *
895 cachep->batchcount + cachep->num;
896 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
899 /* Now we can go ahead with allocating the shared array's
901 list_for_each_entry(cachep, &cache_chain, next) {
902 struct array_cache *nc;
904 nc = alloc_arraycache(node, cachep->limit,
908 cachep->array[cpu] = nc;
910 l3 = cachep->nodelists[node];
913 if (!(nc = alloc_arraycache(node,
919 /* we are serialised from CPU_DEAD or
920 CPU_UP_CANCELLED by the cpucontrol lock */
924 mutex_unlock(&cache_chain_mutex);
927 start_cpu_timer(cpu);
929 #ifdef CONFIG_HOTPLUG_CPU
932 case CPU_UP_CANCELED:
933 mutex_lock(&cache_chain_mutex);
935 list_for_each_entry(cachep, &cache_chain, next) {
936 struct array_cache *nc;
939 mask = node_to_cpumask(node);
940 spin_lock_irq(&cachep->spinlock);
941 /* cpu is dead; no one can alloc from it. */
942 nc = cachep->array[cpu];
943 cachep->array[cpu] = NULL;
944 l3 = cachep->nodelists[node];
949 spin_lock(&l3->list_lock);
951 /* Free limit for this kmem_list3 */
952 l3->free_limit -= cachep->batchcount;
954 free_block(cachep, nc->entry, nc->avail, node);
956 if (!cpus_empty(mask)) {
957 spin_unlock(&l3->list_lock);
962 free_block(cachep, l3->shared->entry,
963 l3->shared->avail, node);
968 drain_alien_cache(cachep, l3);
969 free_alien_cache(l3->alien);
973 /* free slabs belonging to this node */
974 if (__node_shrink(cachep, node)) {
975 cachep->nodelists[node] = NULL;
976 spin_unlock(&l3->list_lock);
979 spin_unlock(&l3->list_lock);
982 spin_unlock_irq(&cachep->spinlock);
985 mutex_unlock(&cache_chain_mutex);
991 mutex_unlock(&cache_chain_mutex);
995 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
998 * swap the static kmem_list3 with kmalloced memory
1000 static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid)
1002 struct kmem_list3 *ptr;
1004 BUG_ON(cachep->nodelists[nodeid] != list);
1005 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1008 local_irq_disable();
1009 memcpy(ptr, list, sizeof(struct kmem_list3));
1010 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1011 cachep->nodelists[nodeid] = ptr;
1016 * Called after the gfp() functions have been enabled, and before smp_init().
1018 void __init kmem_cache_init(void)
1021 struct cache_sizes *sizes;
1022 struct cache_names *names;
1025 for (i = 0; i < NUM_INIT_LISTS; i++) {
1026 kmem_list3_init(&initkmem_list3[i]);
1027 if (i < MAX_NUMNODES)
1028 cache_cache.nodelists[i] = NULL;
1032 * Fragmentation resistance on low memory - only use bigger
1033 * page orders on machines with more than 32MB of memory.
1035 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1036 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1038 /* Bootstrap is tricky, because several objects are allocated
1039 * from caches that do not exist yet:
1040 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1041 * structures of all caches, except cache_cache itself: cache_cache
1042 * is statically allocated.
1043 * Initially an __init data area is used for the head array and the
1044 * kmem_list3 structures, it's replaced with a kmalloc allocated
1045 * array at the end of the bootstrap.
1046 * 2) Create the first kmalloc cache.
1047 * The kmem_cache_t for the new cache is allocated normally.
1048 * An __init data area is used for the head array.
1049 * 3) Create the remaining kmalloc caches, with minimally sized
1051 * 4) Replace the __init data head arrays for cache_cache and the first
1052 * kmalloc cache with kmalloc allocated arrays.
1053 * 5) Replace the __init data for kmem_list3 for cache_cache and
1054 * the other cache's with kmalloc allocated memory.
1055 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1058 /* 1) create the cache_cache */
1059 INIT_LIST_HEAD(&cache_chain);
1060 list_add(&cache_cache.next, &cache_chain);
1061 cache_cache.colour_off = cache_line_size();
1062 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1063 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1065 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
1067 cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0,
1068 &left_over, &cache_cache.num);
1069 if (!cache_cache.num)
1072 cache_cache.colour = left_over / cache_cache.colour_off;
1073 cache_cache.colour_next = 0;
1074 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1075 sizeof(struct slab), cache_line_size());
1077 /* 2+3) create the kmalloc caches */
1078 sizes = malloc_sizes;
1079 names = cache_names;
1081 /* Initialize the caches that provide memory for the array cache
1082 * and the kmem_list3 structures first.
1083 * Without this, further allocations will bug
1086 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1087 sizes[INDEX_AC].cs_size,
1088 ARCH_KMALLOC_MINALIGN,
1089 (ARCH_KMALLOC_FLAGS |
1090 SLAB_PANIC), NULL, NULL);
1092 if (INDEX_AC != INDEX_L3)
1093 sizes[INDEX_L3].cs_cachep =
1094 kmem_cache_create(names[INDEX_L3].name,
1095 sizes[INDEX_L3].cs_size,
1096 ARCH_KMALLOC_MINALIGN,
1097 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1100 while (sizes->cs_size != ULONG_MAX) {
1102 * For performance, all the general caches are L1 aligned.
1103 * This should be particularly beneficial on SMP boxes, as it
1104 * eliminates "false sharing".
1105 * Note for systems short on memory removing the alignment will
1106 * allow tighter packing of the smaller caches.
1108 if (!sizes->cs_cachep)
1109 sizes->cs_cachep = kmem_cache_create(names->name,
1111 ARCH_KMALLOC_MINALIGN,
1116 /* Inc off-slab bufctl limit until the ceiling is hit. */
1117 if (!(OFF_SLAB(sizes->cs_cachep))) {
1118 offslab_limit = sizes->cs_size - sizeof(struct slab);
1119 offslab_limit /= sizeof(kmem_bufctl_t);
1122 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1124 ARCH_KMALLOC_MINALIGN,
1125 (ARCH_KMALLOC_FLAGS |
1133 /* 4) Replace the bootstrap head arrays */
1137 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1139 local_irq_disable();
1140 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
1141 memcpy(ptr, ac_data(&cache_cache),
1142 sizeof(struct arraycache_init));
1143 cache_cache.array[smp_processor_id()] = ptr;
1146 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1148 local_irq_disable();
1149 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
1150 != &initarray_generic.cache);
1151 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
1152 sizeof(struct arraycache_init));
1153 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1157 /* 5) Replace the bootstrap kmem_list3's */
1160 /* Replace the static kmem_list3 structures for the boot cpu */
1161 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1164 for_each_online_node(node) {
1165 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1166 &initkmem_list3[SIZE_AC + node], node);
1168 if (INDEX_AC != INDEX_L3) {
1169 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1170 &initkmem_list3[SIZE_L3 + node],
1176 /* 6) resize the head arrays to their final sizes */
1178 kmem_cache_t *cachep;
1179 mutex_lock(&cache_chain_mutex);
1180 list_for_each_entry(cachep, &cache_chain, next)
1181 enable_cpucache(cachep);
1182 mutex_unlock(&cache_chain_mutex);
1186 g_cpucache_up = FULL;
1188 /* Register a cpu startup notifier callback
1189 * that initializes ac_data for all new cpus
1191 register_cpu_notifier(&cpucache_notifier);
1193 /* The reap timers are started later, with a module init call:
1194 * That part of the kernel is not yet operational.
1198 static int __init cpucache_init(void)
1203 * Register the timers that return unneeded
1206 for_each_online_cpu(cpu)
1207 start_cpu_timer(cpu);
1212 __initcall(cpucache_init);
1215 * Interface to system's page allocator. No need to hold the cache-lock.
1217 * If we requested dmaable memory, we will get it. Even if we
1218 * did not request dmaable memory, we might get it, but that
1219 * would be relatively rare and ignorable.
1221 static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
1227 flags |= cachep->gfpflags;
1228 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1231 addr = page_address(page);
1233 i = (1 << cachep->gfporder);
1234 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1235 atomic_add(i, &slab_reclaim_pages);
1236 add_page_state(nr_slab, i);
1245 * Interface to system's page release.
1247 static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1249 unsigned long i = (1 << cachep->gfporder);
1250 struct page *page = virt_to_page(addr);
1251 const unsigned long nr_freed = i;
1254 if (!TestClearPageSlab(page))
1258 sub_page_state(nr_slab, nr_freed);
1259 if (current->reclaim_state)
1260 current->reclaim_state->reclaimed_slab += nr_freed;
1261 free_pages((unsigned long)addr, cachep->gfporder);
1262 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1263 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1266 static void kmem_rcu_free(struct rcu_head *head)
1268 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1269 kmem_cache_t *cachep = slab_rcu->cachep;
1271 kmem_freepages(cachep, slab_rcu->addr);
1272 if (OFF_SLAB(cachep))
1273 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1278 #ifdef CONFIG_DEBUG_PAGEALLOC
1279 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1280 unsigned long caller)
1282 int size = obj_size(cachep);
1284 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1286 if (size < 5 * sizeof(unsigned long))
1289 *addr++ = 0x12345678;
1291 *addr++ = smp_processor_id();
1292 size -= 3 * sizeof(unsigned long);
1294 unsigned long *sptr = &caller;
1295 unsigned long svalue;
1297 while (!kstack_end(sptr)) {
1299 if (kernel_text_address(svalue)) {
1301 size -= sizeof(unsigned long);
1302 if (size <= sizeof(unsigned long))
1308 *addr++ = 0x87654321;
1312 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1314 int size = obj_size(cachep);
1315 addr = &((char *)addr)[obj_offset(cachep)];
1317 memset(addr, val, size);
1318 *(unsigned char *)(addr + size - 1) = POISON_END;
1321 static void dump_line(char *data, int offset, int limit)
1324 printk(KERN_ERR "%03x:", offset);
1325 for (i = 0; i < limit; i++) {
1326 printk(" %02x", (unsigned char)data[offset + i]);
1334 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1339 if (cachep->flags & SLAB_RED_ZONE) {
1340 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1341 *dbg_redzone1(cachep, objp),
1342 *dbg_redzone2(cachep, objp));
1345 if (cachep->flags & SLAB_STORE_USER) {
1346 printk(KERN_ERR "Last user: [<%p>]",
1347 *dbg_userword(cachep, objp));
1348 print_symbol("(%s)",
1349 (unsigned long)*dbg_userword(cachep, objp));
1352 realobj = (char *)objp + obj_offset(cachep);
1353 size = obj_size(cachep);
1354 for (i = 0; i < size && lines; i += 16, lines--) {
1357 if (i + limit > size)
1359 dump_line(realobj, i, limit);
1363 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1369 realobj = (char *)objp + obj_offset(cachep);
1370 size = obj_size(cachep);
1372 for (i = 0; i < size; i++) {
1373 char exp = POISON_FREE;
1376 if (realobj[i] != exp) {
1382 "Slab corruption: start=%p, len=%d\n",
1384 print_objinfo(cachep, objp, 0);
1386 /* Hexdump the affected line */
1389 if (i + limit > size)
1391 dump_line(realobj, i, limit);
1394 /* Limit to 5 lines */
1400 /* Print some data about the neighboring objects, if they
1403 struct slab *slabp = page_get_slab(virt_to_page(objp));
1406 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
1408 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1409 realobj = (char *)objp + obj_offset(cachep);
1410 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1412 print_objinfo(cachep, objp, 2);
1414 if (objnr + 1 < cachep->num) {
1415 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1416 realobj = (char *)objp + obj_offset(cachep);
1417 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1419 print_objinfo(cachep, objp, 2);
1425 /* Destroy all the objs in a slab, and release the mem back to the system.
1426 * Before calling the slab must have been unlinked from the cache.
1427 * The cache-lock is not held/needed.
1429 static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp)
1431 void *addr = slabp->s_mem - slabp->colouroff;
1435 for (i = 0; i < cachep->num; i++) {
1436 void *objp = slabp->s_mem + cachep->buffer_size * i;
1438 if (cachep->flags & SLAB_POISON) {
1439 #ifdef CONFIG_DEBUG_PAGEALLOC
1440 if ((cachep->buffer_size % PAGE_SIZE) == 0
1441 && OFF_SLAB(cachep))
1442 kernel_map_pages(virt_to_page(objp),
1443 cachep->buffer_size / PAGE_SIZE,
1446 check_poison_obj(cachep, objp);
1448 check_poison_obj(cachep, objp);
1451 if (cachep->flags & SLAB_RED_ZONE) {
1452 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1453 slab_error(cachep, "start of a freed object "
1455 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1456 slab_error(cachep, "end of a freed object "
1459 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1460 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1465 for (i = 0; i < cachep->num; i++) {
1466 void *objp = slabp->s_mem + cachep->buffer_size * i;
1467 (cachep->dtor) (objp, cachep, 0);
1472 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1473 struct slab_rcu *slab_rcu;
1475 slab_rcu = (struct slab_rcu *)slabp;
1476 slab_rcu->cachep = cachep;
1477 slab_rcu->addr = addr;
1478 call_rcu(&slab_rcu->head, kmem_rcu_free);
1480 kmem_freepages(cachep, addr);
1481 if (OFF_SLAB(cachep))
1482 kmem_cache_free(cachep->slabp_cache, slabp);
1486 /* For setting up all the kmem_list3s for cache whose buffer_size is same
1487 as size of kmem_list3. */
1488 static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1492 for_each_online_node(node) {
1493 cachep->nodelists[node] = &initkmem_list3[index + node];
1494 cachep->nodelists[node]->next_reap = jiffies +
1496 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1501 * calculate_slab_order - calculate size (page order) of slabs and the number
1502 * of objects per slab.
1504 * This could be made much more intelligent. For now, try to avoid using
1505 * high order pages for slabs. When the gfp() functions are more friendly
1506 * towards high-order requests, this should be changed.
1508 static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size,
1509 size_t align, gfp_t flags)
1511 size_t left_over = 0;
1513 for (;; cachep->gfporder++) {
1517 if (cachep->gfporder > MAX_GFP_ORDER) {
1522 cache_estimate(cachep->gfporder, size, align, flags,
1526 /* More than offslab_limit objects will cause problems */
1527 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1531 left_over = remainder;
1534 * Large number of objects is good, but very large slabs are
1535 * currently bad for the gfp()s.
1537 if (cachep->gfporder >= slab_break_gfp_order)
1540 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1541 /* Acceptable internal fragmentation */
1548 * kmem_cache_create - Create a cache.
1549 * @name: A string which is used in /proc/slabinfo to identify this cache.
1550 * @size: The size of objects to be created in this cache.
1551 * @align: The required alignment for the objects.
1552 * @flags: SLAB flags
1553 * @ctor: A constructor for the objects.
1554 * @dtor: A destructor for the objects.
1556 * Returns a ptr to the cache on success, NULL on failure.
1557 * Cannot be called within a int, but can be interrupted.
1558 * The @ctor is run when new pages are allocated by the cache
1559 * and the @dtor is run before the pages are handed back.
1561 * @name must be valid until the cache is destroyed. This implies that
1562 * the module calling this has to destroy the cache before getting
1567 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1568 * to catch references to uninitialised memory.
1570 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1571 * for buffer overruns.
1573 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1576 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1577 * cacheline. This can be beneficial if you're counting cycles as closely
1581 kmem_cache_create (const char *name, size_t size, size_t align,
1582 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1583 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1585 size_t left_over, slab_size, ralign;
1586 kmem_cache_t *cachep = NULL;
1587 struct list_head *p;
1590 * Sanity checks... these are all serious usage bugs.
1594 (size < BYTES_PER_WORD) ||
1595 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1596 printk(KERN_ERR "%s: Early error in slab %s\n",
1597 __FUNCTION__, name);
1601 mutex_lock(&cache_chain_mutex);
1603 list_for_each(p, &cache_chain) {
1604 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1605 mm_segment_t old_fs = get_fs();
1610 * This happens when the module gets unloaded and doesn't
1611 * destroy its slab cache and no-one else reuses the vmalloc
1612 * area of the module. Print a warning.
1615 res = __get_user(tmp, pc->name);
1618 printk("SLAB: cache with size %d has lost its name\n",
1623 if (!strcmp(pc->name, name)) {
1624 printk("kmem_cache_create: duplicate cache %s\n", name);
1631 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1632 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1633 /* No constructor, but inital state check requested */
1634 printk(KERN_ERR "%s: No con, but init state check "
1635 "requested - %s\n", __FUNCTION__, name);
1636 flags &= ~SLAB_DEBUG_INITIAL;
1640 * Enable redzoning and last user accounting, except for caches with
1641 * large objects, if the increased size would increase the object size
1642 * above the next power of two: caches with object sizes just above a
1643 * power of two have a significant amount of internal fragmentation.
1646 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1647 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1648 if (!(flags & SLAB_DESTROY_BY_RCU))
1649 flags |= SLAB_POISON;
1651 if (flags & SLAB_DESTROY_BY_RCU)
1652 BUG_ON(flags & SLAB_POISON);
1654 if (flags & SLAB_DESTROY_BY_RCU)
1658 * Always checks flags, a caller might be expecting debug
1659 * support which isn't available.
1661 if (flags & ~CREATE_MASK)
1664 /* Check that size is in terms of words. This is needed to avoid
1665 * unaligned accesses for some archs when redzoning is used, and makes
1666 * sure any on-slab bufctl's are also correctly aligned.
1668 if (size & (BYTES_PER_WORD - 1)) {
1669 size += (BYTES_PER_WORD - 1);
1670 size &= ~(BYTES_PER_WORD - 1);
1673 /* calculate out the final buffer alignment: */
1674 /* 1) arch recommendation: can be overridden for debug */
1675 if (flags & SLAB_HWCACHE_ALIGN) {
1676 /* Default alignment: as specified by the arch code.
1677 * Except if an object is really small, then squeeze multiple
1678 * objects into one cacheline.
1680 ralign = cache_line_size();
1681 while (size <= ralign / 2)
1684 ralign = BYTES_PER_WORD;
1686 /* 2) arch mandated alignment: disables debug if necessary */
1687 if (ralign < ARCH_SLAB_MINALIGN) {
1688 ralign = ARCH_SLAB_MINALIGN;
1689 if (ralign > BYTES_PER_WORD)
1690 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1692 /* 3) caller mandated alignment: disables debug if necessary */
1693 if (ralign < align) {
1695 if (ralign > BYTES_PER_WORD)
1696 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1698 /* 4) Store it. Note that the debug code below can reduce
1699 * the alignment to BYTES_PER_WORD.
1703 /* Get cache's description obj. */
1704 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1707 memset(cachep, 0, sizeof(kmem_cache_t));
1710 cachep->obj_size = size;
1712 if (flags & SLAB_RED_ZONE) {
1713 /* redzoning only works with word aligned caches */
1714 align = BYTES_PER_WORD;
1716 /* add space for red zone words */
1717 cachep->obj_offset += BYTES_PER_WORD;
1718 size += 2 * BYTES_PER_WORD;
1720 if (flags & SLAB_STORE_USER) {
1721 /* user store requires word alignment and
1722 * one word storage behind the end of the real
1725 align = BYTES_PER_WORD;
1726 size += BYTES_PER_WORD;
1728 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1729 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1730 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1731 cachep->obj_offset += PAGE_SIZE - size;
1737 /* Determine if the slab management is 'on' or 'off' slab. */
1738 if (size >= (PAGE_SIZE >> 3))
1740 * Size is large, assume best to place the slab management obj
1741 * off-slab (should allow better packing of objs).
1743 flags |= CFLGS_OFF_SLAB;
1745 size = ALIGN(size, align);
1747 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1749 * A VFS-reclaimable slab tends to have most allocations
1750 * as GFP_NOFS and we really don't want to have to be allocating
1751 * higher-order pages when we are unable to shrink dcache.
1753 cachep->gfporder = 0;
1754 cache_estimate(cachep->gfporder, size, align, flags,
1755 &left_over, &cachep->num);
1757 left_over = calculate_slab_order(cachep, size, align, flags);
1760 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1761 kmem_cache_free(&cache_cache, cachep);
1765 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1766 + sizeof(struct slab), align);
1769 * If the slab has been placed off-slab, and we have enough space then
1770 * move it on-slab. This is at the expense of any extra colouring.
1772 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1773 flags &= ~CFLGS_OFF_SLAB;
1774 left_over -= slab_size;
1777 if (flags & CFLGS_OFF_SLAB) {
1778 /* really off slab. No need for manual alignment */
1780 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
1783 cachep->colour_off = cache_line_size();
1784 /* Offset must be a multiple of the alignment. */
1785 if (cachep->colour_off < align)
1786 cachep->colour_off = align;
1787 cachep->colour = left_over / cachep->colour_off;
1788 cachep->slab_size = slab_size;
1789 cachep->flags = flags;
1790 cachep->gfpflags = 0;
1791 if (flags & SLAB_CACHE_DMA)
1792 cachep->gfpflags |= GFP_DMA;
1793 spin_lock_init(&cachep->spinlock);
1794 cachep->buffer_size = size;
1796 if (flags & CFLGS_OFF_SLAB)
1797 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
1798 cachep->ctor = ctor;
1799 cachep->dtor = dtor;
1800 cachep->name = name;
1802 /* Don't let CPUs to come and go */
1805 if (g_cpucache_up == FULL) {
1806 enable_cpucache(cachep);
1808 if (g_cpucache_up == NONE) {
1809 /* Note: the first kmem_cache_create must create
1810 * the cache that's used by kmalloc(24), otherwise
1811 * the creation of further caches will BUG().
1813 cachep->array[smp_processor_id()] =
1814 &initarray_generic.cache;
1816 /* If the cache that's used by
1817 * kmalloc(sizeof(kmem_list3)) is the first cache,
1818 * then we need to set up all its list3s, otherwise
1819 * the creation of further caches will BUG().
1821 set_up_list3s(cachep, SIZE_AC);
1822 if (INDEX_AC == INDEX_L3)
1823 g_cpucache_up = PARTIAL_L3;
1825 g_cpucache_up = PARTIAL_AC;
1827 cachep->array[smp_processor_id()] =
1828 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1830 if (g_cpucache_up == PARTIAL_AC) {
1831 set_up_list3s(cachep, SIZE_L3);
1832 g_cpucache_up = PARTIAL_L3;
1835 for_each_online_node(node) {
1837 cachep->nodelists[node] =
1839 (struct kmem_list3),
1841 BUG_ON(!cachep->nodelists[node]);
1842 kmem_list3_init(cachep->
1847 cachep->nodelists[numa_node_id()]->next_reap =
1848 jiffies + REAPTIMEOUT_LIST3 +
1849 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1851 BUG_ON(!ac_data(cachep));
1852 ac_data(cachep)->avail = 0;
1853 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1854 ac_data(cachep)->batchcount = 1;
1855 ac_data(cachep)->touched = 0;
1856 cachep->batchcount = 1;
1857 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1860 /* cache setup completed, link it into the list */
1861 list_add(&cachep->next, &cache_chain);
1862 unlock_cpu_hotplug();
1864 if (!cachep && (flags & SLAB_PANIC))
1865 panic("kmem_cache_create(): failed to create slab `%s'\n",
1867 mutex_unlock(&cache_chain_mutex);
1870 EXPORT_SYMBOL(kmem_cache_create);
1873 static void check_irq_off(void)
1875 BUG_ON(!irqs_disabled());
1878 static void check_irq_on(void)
1880 BUG_ON(irqs_disabled());
1883 static void check_spinlock_acquired(kmem_cache_t *cachep)
1887 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
1891 static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1895 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1900 #define check_irq_off() do { } while(0)
1901 #define check_irq_on() do { } while(0)
1902 #define check_spinlock_acquired(x) do { } while(0)
1903 #define check_spinlock_acquired_node(x, y) do { } while(0)
1907 * Waits for all CPUs to execute func().
1909 static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
1914 local_irq_disable();
1918 if (smp_call_function(func, arg, 1, 1))
1924 static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
1925 int force, int node);
1927 static void do_drain(void *arg)
1929 kmem_cache_t *cachep = (kmem_cache_t *) arg;
1930 struct array_cache *ac;
1931 int node = numa_node_id();
1934 ac = ac_data(cachep);
1935 spin_lock(&cachep->nodelists[node]->list_lock);
1936 free_block(cachep, ac->entry, ac->avail, node);
1937 spin_unlock(&cachep->nodelists[node]->list_lock);
1941 static void drain_cpu_caches(kmem_cache_t *cachep)
1943 struct kmem_list3 *l3;
1946 smp_call_function_all_cpus(do_drain, cachep);
1948 spin_lock_irq(&cachep->spinlock);
1949 for_each_online_node(node) {
1950 l3 = cachep->nodelists[node];
1952 spin_lock(&l3->list_lock);
1953 drain_array_locked(cachep, l3->shared, 1, node);
1954 spin_unlock(&l3->list_lock);
1956 drain_alien_cache(cachep, l3);
1959 spin_unlock_irq(&cachep->spinlock);
1962 static int __node_shrink(kmem_cache_t *cachep, int node)
1965 struct kmem_list3 *l3 = cachep->nodelists[node];
1969 struct list_head *p;
1971 p = l3->slabs_free.prev;
1972 if (p == &l3->slabs_free)
1975 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
1980 list_del(&slabp->list);
1982 l3->free_objects -= cachep->num;
1983 spin_unlock_irq(&l3->list_lock);
1984 slab_destroy(cachep, slabp);
1985 spin_lock_irq(&l3->list_lock);
1987 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
1991 static int __cache_shrink(kmem_cache_t *cachep)
1994 struct kmem_list3 *l3;
1996 drain_cpu_caches(cachep);
1999 for_each_online_node(i) {
2000 l3 = cachep->nodelists[i];
2002 spin_lock_irq(&l3->list_lock);
2003 ret += __node_shrink(cachep, i);
2004 spin_unlock_irq(&l3->list_lock);
2007 return (ret ? 1 : 0);
2011 * kmem_cache_shrink - Shrink a cache.
2012 * @cachep: The cache to shrink.
2014 * Releases as many slabs as possible for a cache.
2015 * To help debugging, a zero exit status indicates all slabs were released.
2017 int kmem_cache_shrink(kmem_cache_t *cachep)
2019 if (!cachep || in_interrupt())
2022 return __cache_shrink(cachep);
2024 EXPORT_SYMBOL(kmem_cache_shrink);
2027 * kmem_cache_destroy - delete a cache
2028 * @cachep: the cache to destroy
2030 * Remove a kmem_cache_t object from the slab cache.
2031 * Returns 0 on success.
2033 * It is expected this function will be called by a module when it is
2034 * unloaded. This will remove the cache completely, and avoid a duplicate
2035 * cache being allocated each time a module is loaded and unloaded, if the
2036 * module doesn't have persistent in-kernel storage across loads and unloads.
2038 * The cache must be empty before calling this function.
2040 * The caller must guarantee that noone will allocate memory from the cache
2041 * during the kmem_cache_destroy().
2043 int kmem_cache_destroy(kmem_cache_t *cachep)
2046 struct kmem_list3 *l3;
2048 if (!cachep || in_interrupt())
2051 /* Don't let CPUs to come and go */
2054 /* Find the cache in the chain of caches. */
2055 mutex_lock(&cache_chain_mutex);
2057 * the chain is never empty, cache_cache is never destroyed
2059 list_del(&cachep->next);
2060 mutex_unlock(&cache_chain_mutex);
2062 if (__cache_shrink(cachep)) {
2063 slab_error(cachep, "Can't free all objects");
2064 mutex_lock(&cache_chain_mutex);
2065 list_add(&cachep->next, &cache_chain);
2066 mutex_unlock(&cache_chain_mutex);
2067 unlock_cpu_hotplug();
2071 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2074 for_each_online_cpu(i)
2075 kfree(cachep->array[i]);
2077 /* NUMA: free the list3 structures */
2078 for_each_online_node(i) {
2079 if ((l3 = cachep->nodelists[i])) {
2081 free_alien_cache(l3->alien);
2085 kmem_cache_free(&cache_cache, cachep);
2087 unlock_cpu_hotplug();
2091 EXPORT_SYMBOL(kmem_cache_destroy);
2093 /* Get the memory for a slab management obj. */
2094 static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2095 int colour_off, gfp_t local_flags)
2099 if (OFF_SLAB(cachep)) {
2100 /* Slab management obj is off-slab. */
2101 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2105 slabp = objp + colour_off;
2106 colour_off += cachep->slab_size;
2109 slabp->colouroff = colour_off;
2110 slabp->s_mem = objp + colour_off;
2115 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2117 return (kmem_bufctl_t *) (slabp + 1);
2120 static void cache_init_objs(kmem_cache_t *cachep,
2121 struct slab *slabp, unsigned long ctor_flags)
2125 for (i = 0; i < cachep->num; i++) {
2126 void *objp = slabp->s_mem + cachep->buffer_size * i;
2128 /* need to poison the objs? */
2129 if (cachep->flags & SLAB_POISON)
2130 poison_obj(cachep, objp, POISON_FREE);
2131 if (cachep->flags & SLAB_STORE_USER)
2132 *dbg_userword(cachep, objp) = NULL;
2134 if (cachep->flags & SLAB_RED_ZONE) {
2135 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2136 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2139 * Constructors are not allowed to allocate memory from
2140 * the same cache which they are a constructor for.
2141 * Otherwise, deadlock. They must also be threaded.
2143 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2144 cachep->ctor(objp + obj_offset(cachep), cachep,
2147 if (cachep->flags & SLAB_RED_ZONE) {
2148 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2149 slab_error(cachep, "constructor overwrote the"
2150 " end of an object");
2151 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2152 slab_error(cachep, "constructor overwrote the"
2153 " start of an object");
2155 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2156 && cachep->flags & SLAB_POISON)
2157 kernel_map_pages(virt_to_page(objp),
2158 cachep->buffer_size / PAGE_SIZE, 0);
2161 cachep->ctor(objp, cachep, ctor_flags);
2163 slab_bufctl(slabp)[i] = i + 1;
2165 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2169 static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
2171 if (flags & SLAB_DMA) {
2172 if (!(cachep->gfpflags & GFP_DMA))
2175 if (cachep->gfpflags & GFP_DMA)
2180 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2185 /* Nasty!!!!!! I hope this is OK. */
2186 i = 1 << cachep->gfporder;
2187 page = virt_to_page(objp);
2189 page_set_cache(page, cachep);
2190 page_set_slab(page, slabp);
2196 * Grow (by 1) the number of slabs within a cache. This is called by
2197 * kmem_cache_alloc() when there are no active objs left in a cache.
2199 static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2205 unsigned long ctor_flags;
2206 struct kmem_list3 *l3;
2208 /* Be lazy and only check for valid flags here,
2209 * keeping it out of the critical path in kmem_cache_alloc().
2211 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2213 if (flags & SLAB_NO_GROW)
2216 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2217 local_flags = (flags & SLAB_LEVEL_MASK);
2218 if (!(local_flags & __GFP_WAIT))
2220 * Not allowed to sleep. Need to tell a constructor about
2221 * this - it might need to know...
2223 ctor_flags |= SLAB_CTOR_ATOMIC;
2225 /* About to mess with non-constant members - lock. */
2227 spin_lock(&cachep->spinlock);
2229 /* Get colour for the slab, and cal the next value. */
2230 offset = cachep->colour_next;
2231 cachep->colour_next++;
2232 if (cachep->colour_next >= cachep->colour)
2233 cachep->colour_next = 0;
2234 offset *= cachep->colour_off;
2236 spin_unlock(&cachep->spinlock);
2239 if (local_flags & __GFP_WAIT)
2243 * The test for missing atomic flag is performed here, rather than
2244 * the more obvious place, simply to reduce the critical path length
2245 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2246 * will eventually be caught here (where it matters).
2248 kmem_flagcheck(cachep, flags);
2250 /* Get mem for the objs.
2251 * Attempt to allocate a physical page from 'nodeid',
2253 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2256 /* Get slab management. */
2257 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2260 slabp->nodeid = nodeid;
2261 set_slab_attr(cachep, slabp, objp);
2263 cache_init_objs(cachep, slabp, ctor_flags);
2265 if (local_flags & __GFP_WAIT)
2266 local_irq_disable();
2268 l3 = cachep->nodelists[nodeid];
2269 spin_lock(&l3->list_lock);
2271 /* Make slab active. */
2272 list_add_tail(&slabp->list, &(l3->slabs_free));
2273 STATS_INC_GROWN(cachep);
2274 l3->free_objects += cachep->num;
2275 spin_unlock(&l3->list_lock);
2278 kmem_freepages(cachep, objp);
2280 if (local_flags & __GFP_WAIT)
2281 local_irq_disable();
2288 * Perform extra freeing checks:
2289 * - detect bad pointers.
2290 * - POISON/RED_ZONE checking
2291 * - destructor calls, for caches with POISON+dtor
2293 static void kfree_debugcheck(const void *objp)
2297 if (!virt_addr_valid(objp)) {
2298 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2299 (unsigned long)objp);
2302 page = virt_to_page(objp);
2303 if (!PageSlab(page)) {
2304 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2305 (unsigned long)objp);
2310 static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2317 objp -= obj_offset(cachep);
2318 kfree_debugcheck(objp);
2319 page = virt_to_page(objp);
2321 if (page_get_cache(page) != cachep) {
2323 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2324 page_get_cache(page), cachep);
2325 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2326 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2327 page_get_cache(page)->name);
2330 slabp = page_get_slab(page);
2332 if (cachep->flags & SLAB_RED_ZONE) {
2333 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2334 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2336 "double free, or memory outside"
2337 " object was overwritten");
2339 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2340 objp, *dbg_redzone1(cachep, objp),
2341 *dbg_redzone2(cachep, objp));
2343 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2344 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2346 if (cachep->flags & SLAB_STORE_USER)
2347 *dbg_userword(cachep, objp) = caller;
2349 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2351 BUG_ON(objnr >= cachep->num);
2352 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
2354 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2355 /* Need to call the slab's constructor so the
2356 * caller can perform a verify of its state (debugging).
2357 * Called without the cache-lock held.
2359 cachep->ctor(objp + obj_offset(cachep),
2360 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2362 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2363 /* we want to cache poison the object,
2364 * call the destruction callback
2366 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2368 if (cachep->flags & SLAB_POISON) {
2369 #ifdef CONFIG_DEBUG_PAGEALLOC
2370 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2371 store_stackinfo(cachep, objp, (unsigned long)caller);
2372 kernel_map_pages(virt_to_page(objp),
2373 cachep->buffer_size / PAGE_SIZE, 0);
2375 poison_obj(cachep, objp, POISON_FREE);
2378 poison_obj(cachep, objp, POISON_FREE);
2384 static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2389 /* Check slab's freelist to see if this obj is there. */
2390 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2392 if (entries > cachep->num || i >= cachep->num)
2395 if (entries != cachep->num - slabp->inuse) {
2398 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2399 cachep->name, cachep->num, slabp, slabp->inuse);
2401 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2404 printk("\n%03x:", i);
2405 printk(" %02x", ((unsigned char *)slabp)[i]);
2412 #define kfree_debugcheck(x) do { } while(0)
2413 #define cache_free_debugcheck(x,objp,z) (objp)
2414 #define check_slabp(x,y) do { } while(0)
2417 static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
2420 struct kmem_list3 *l3;
2421 struct array_cache *ac;
2424 ac = ac_data(cachep);
2426 batchcount = ac->batchcount;
2427 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2428 /* if there was little recent activity on this
2429 * cache, then perform only a partial refill.
2430 * Otherwise we could generate refill bouncing.
2432 batchcount = BATCHREFILL_LIMIT;
2434 l3 = cachep->nodelists[numa_node_id()];
2436 BUG_ON(ac->avail > 0 || !l3);
2437 spin_lock(&l3->list_lock);
2440 struct array_cache *shared_array = l3->shared;
2441 if (shared_array->avail) {
2442 if (batchcount > shared_array->avail)
2443 batchcount = shared_array->avail;
2444 shared_array->avail -= batchcount;
2445 ac->avail = batchcount;
2447 &(shared_array->entry[shared_array->avail]),
2448 sizeof(void *) * batchcount);
2449 shared_array->touched = 1;
2453 while (batchcount > 0) {
2454 struct list_head *entry;
2456 /* Get slab alloc is to come from. */
2457 entry = l3->slabs_partial.next;
2458 if (entry == &l3->slabs_partial) {
2459 l3->free_touched = 1;
2460 entry = l3->slabs_free.next;
2461 if (entry == &l3->slabs_free)
2465 slabp = list_entry(entry, struct slab, list);
2466 check_slabp(cachep, slabp);
2467 check_spinlock_acquired(cachep);
2468 while (slabp->inuse < cachep->num && batchcount--) {
2470 STATS_INC_ALLOCED(cachep);
2471 STATS_INC_ACTIVE(cachep);
2472 STATS_SET_HIGH(cachep);
2474 /* get obj pointer */
2475 ac->entry[ac->avail++] = slabp->s_mem +
2476 slabp->free * cachep->buffer_size;
2479 next = slab_bufctl(slabp)[slabp->free];
2481 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2482 WARN_ON(numa_node_id() != slabp->nodeid);
2486 check_slabp(cachep, slabp);
2488 /* move slabp to correct slabp list: */
2489 list_del(&slabp->list);
2490 if (slabp->free == BUFCTL_END)
2491 list_add(&slabp->list, &l3->slabs_full);
2493 list_add(&slabp->list, &l3->slabs_partial);
2497 l3->free_objects -= ac->avail;
2499 spin_unlock(&l3->list_lock);
2501 if (unlikely(!ac->avail)) {
2503 x = cache_grow(cachep, flags, numa_node_id());
2505 // cache_grow can reenable interrupts, then ac could change.
2506 ac = ac_data(cachep);
2507 if (!x && ac->avail == 0) // no objects in sight? abort
2510 if (!ac->avail) // objects refilled by interrupt?
2514 return ac->entry[--ac->avail];
2518 cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
2520 might_sleep_if(flags & __GFP_WAIT);
2522 kmem_flagcheck(cachep, flags);
2527 static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags,
2528 void *objp, void *caller)
2532 if (cachep->flags & SLAB_POISON) {
2533 #ifdef CONFIG_DEBUG_PAGEALLOC
2534 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2535 kernel_map_pages(virt_to_page(objp),
2536 cachep->buffer_size / PAGE_SIZE, 1);
2538 check_poison_obj(cachep, objp);
2540 check_poison_obj(cachep, objp);
2542 poison_obj(cachep, objp, POISON_INUSE);
2544 if (cachep->flags & SLAB_STORE_USER)
2545 *dbg_userword(cachep, objp) = caller;
2547 if (cachep->flags & SLAB_RED_ZONE) {
2548 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2549 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2551 "double free, or memory outside"
2552 " object was overwritten");
2554 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2555 objp, *dbg_redzone1(cachep, objp),
2556 *dbg_redzone2(cachep, objp));
2558 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2559 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2561 objp += obj_offset(cachep);
2562 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2563 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2565 if (!(flags & __GFP_WAIT))
2566 ctor_flags |= SLAB_CTOR_ATOMIC;
2568 cachep->ctor(objp, cachep, ctor_flags);
2573 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2576 static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2579 struct array_cache *ac;
2582 if (unlikely(current->mempolicy && !in_interrupt())) {
2583 int nid = slab_node(current->mempolicy);
2585 if (nid != numa_node_id())
2586 return __cache_alloc_node(cachep, flags, nid);
2591 ac = ac_data(cachep);
2592 if (likely(ac->avail)) {
2593 STATS_INC_ALLOCHIT(cachep);
2595 objp = ac->entry[--ac->avail];
2597 STATS_INC_ALLOCMISS(cachep);
2598 objp = cache_alloc_refill(cachep, flags);
2603 static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2605 unsigned long save_flags;
2608 cache_alloc_debugcheck_before(cachep, flags);
2610 local_irq_save(save_flags);
2611 objp = ____cache_alloc(cachep, flags);
2612 local_irq_restore(save_flags);
2613 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2614 __builtin_return_address(0));
2621 * A interface to enable slab creation on nodeid
2623 static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2625 struct list_head *entry;
2627 struct kmem_list3 *l3;
2632 l3 = cachep->nodelists[nodeid];
2636 spin_lock(&l3->list_lock);
2637 entry = l3->slabs_partial.next;
2638 if (entry == &l3->slabs_partial) {
2639 l3->free_touched = 1;
2640 entry = l3->slabs_free.next;
2641 if (entry == &l3->slabs_free)
2645 slabp = list_entry(entry, struct slab, list);
2646 check_spinlock_acquired_node(cachep, nodeid);
2647 check_slabp(cachep, slabp);
2649 STATS_INC_NODEALLOCS(cachep);
2650 STATS_INC_ACTIVE(cachep);
2651 STATS_SET_HIGH(cachep);
2653 BUG_ON(slabp->inuse == cachep->num);
2655 /* get obj pointer */
2656 obj = slabp->s_mem + slabp->free * cachep->buffer_size;
2658 next = slab_bufctl(slabp)[slabp->free];
2660 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2663 check_slabp(cachep, slabp);
2665 /* move slabp to correct slabp list: */
2666 list_del(&slabp->list);
2668 if (slabp->free == BUFCTL_END) {
2669 list_add(&slabp->list, &l3->slabs_full);
2671 list_add(&slabp->list, &l3->slabs_partial);
2674 spin_unlock(&l3->list_lock);
2678 spin_unlock(&l3->list_lock);
2679 x = cache_grow(cachep, flags, nodeid);
2691 * Caller needs to acquire correct kmem_list's list_lock
2693 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects,
2697 struct kmem_list3 *l3;
2699 for (i = 0; i < nr_objects; i++) {
2700 void *objp = objpp[i];
2704 slabp = page_get_slab(virt_to_page(objp));
2705 l3 = cachep->nodelists[node];
2706 list_del(&slabp->list);
2707 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2708 check_spinlock_acquired_node(cachep, node);
2709 check_slabp(cachep, slabp);
2712 /* Verify that the slab belongs to the intended node */
2713 WARN_ON(slabp->nodeid != node);
2715 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2716 printk(KERN_ERR "slab: double free detected in cache "
2717 "'%s', objp %p\n", cachep->name, objp);
2721 slab_bufctl(slabp)[objnr] = slabp->free;
2722 slabp->free = objnr;
2723 STATS_DEC_ACTIVE(cachep);
2726 check_slabp(cachep, slabp);
2728 /* fixup slab chains */
2729 if (slabp->inuse == 0) {
2730 if (l3->free_objects > l3->free_limit) {
2731 l3->free_objects -= cachep->num;
2732 slab_destroy(cachep, slabp);
2734 list_add(&slabp->list, &l3->slabs_free);
2737 /* Unconditionally move a slab to the end of the
2738 * partial list on free - maximum time for the
2739 * other objects to be freed, too.
2741 list_add_tail(&slabp->list, &l3->slabs_partial);
2746 static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2749 struct kmem_list3 *l3;
2750 int node = numa_node_id();
2752 batchcount = ac->batchcount;
2754 BUG_ON(!batchcount || batchcount > ac->avail);
2757 l3 = cachep->nodelists[node];
2758 spin_lock(&l3->list_lock);
2760 struct array_cache *shared_array = l3->shared;
2761 int max = shared_array->limit - shared_array->avail;
2763 if (batchcount > max)
2765 memcpy(&(shared_array->entry[shared_array->avail]),
2766 ac->entry, sizeof(void *) * batchcount);
2767 shared_array->avail += batchcount;
2772 free_block(cachep, ac->entry, batchcount, node);
2777 struct list_head *p;
2779 p = l3->slabs_free.next;
2780 while (p != &(l3->slabs_free)) {
2783 slabp = list_entry(p, struct slab, list);
2784 BUG_ON(slabp->inuse);
2789 STATS_SET_FREEABLE(cachep, i);
2792 spin_unlock(&l3->list_lock);
2793 ac->avail -= batchcount;
2794 memmove(ac->entry, &(ac->entry[batchcount]),
2795 sizeof(void *) * ac->avail);
2800 * Release an obj back to its cache. If the obj has a constructed
2801 * state, it must be in this state _before_ it is released.
2803 * Called with disabled ints.
2805 static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2807 struct array_cache *ac = ac_data(cachep);
2810 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2812 /* Make sure we are not freeing a object from another
2813 * node to the array cache on this cpu.
2818 slabp = page_get_slab(virt_to_page(objp));
2819 if (unlikely(slabp->nodeid != numa_node_id())) {
2820 struct array_cache *alien = NULL;
2821 int nodeid = slabp->nodeid;
2822 struct kmem_list3 *l3 =
2823 cachep->nodelists[numa_node_id()];
2825 STATS_INC_NODEFREES(cachep);
2826 if (l3->alien && l3->alien[nodeid]) {
2827 alien = l3->alien[nodeid];
2828 spin_lock(&alien->lock);
2829 if (unlikely(alien->avail == alien->limit))
2830 __drain_alien_cache(cachep,
2832 alien->entry[alien->avail++] = objp;
2833 spin_unlock(&alien->lock);
2835 spin_lock(&(cachep->nodelists[nodeid])->
2837 free_block(cachep, &objp, 1, nodeid);
2838 spin_unlock(&(cachep->nodelists[nodeid])->
2845 if (likely(ac->avail < ac->limit)) {
2846 STATS_INC_FREEHIT(cachep);
2847 ac->entry[ac->avail++] = objp;
2850 STATS_INC_FREEMISS(cachep);
2851 cache_flusharray(cachep, ac);
2852 ac->entry[ac->avail++] = objp;
2857 * kmem_cache_alloc - Allocate an object
2858 * @cachep: The cache to allocate from.
2859 * @flags: See kmalloc().
2861 * Allocate an object from this cache. The flags are only relevant
2862 * if the cache has no available objects.
2864 void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2866 return __cache_alloc(cachep, flags);
2868 EXPORT_SYMBOL(kmem_cache_alloc);
2871 * kmem_ptr_validate - check if an untrusted pointer might
2873 * @cachep: the cache we're checking against
2874 * @ptr: pointer to validate
2876 * This verifies that the untrusted pointer looks sane:
2877 * it is _not_ a guarantee that the pointer is actually
2878 * part of the slab cache in question, but it at least
2879 * validates that the pointer can be dereferenced and
2880 * looks half-way sane.
2882 * Currently only used for dentry validation.
2884 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2886 unsigned long addr = (unsigned long)ptr;
2887 unsigned long min_addr = PAGE_OFFSET;
2888 unsigned long align_mask = BYTES_PER_WORD - 1;
2889 unsigned long size = cachep->buffer_size;
2892 if (unlikely(addr < min_addr))
2894 if (unlikely(addr > (unsigned long)high_memory - size))
2896 if (unlikely(addr & align_mask))
2898 if (unlikely(!kern_addr_valid(addr)))
2900 if (unlikely(!kern_addr_valid(addr + size - 1)))
2902 page = virt_to_page(ptr);
2903 if (unlikely(!PageSlab(page)))
2905 if (unlikely(page_get_cache(page) != cachep))
2914 * kmem_cache_alloc_node - Allocate an object on the specified node
2915 * @cachep: The cache to allocate from.
2916 * @flags: See kmalloc().
2917 * @nodeid: node number of the target node.
2919 * Identical to kmem_cache_alloc, except that this function is slow
2920 * and can sleep. And it will allocate memory on the given node, which
2921 * can improve the performance for cpu bound structures.
2922 * New and improved: it will now make sure that the object gets
2923 * put on the correct node list so that there is no false sharing.
2925 void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2927 unsigned long save_flags;
2930 cache_alloc_debugcheck_before(cachep, flags);
2931 local_irq_save(save_flags);
2933 if (nodeid == -1 || nodeid == numa_node_id() ||
2934 !cachep->nodelists[nodeid])
2935 ptr = ____cache_alloc(cachep, flags);
2937 ptr = __cache_alloc_node(cachep, flags, nodeid);
2938 local_irq_restore(save_flags);
2940 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
2941 __builtin_return_address(0));
2945 EXPORT_SYMBOL(kmem_cache_alloc_node);
2947 void *kmalloc_node(size_t size, gfp_t flags, int node)
2949 kmem_cache_t *cachep;
2951 cachep = kmem_find_general_cachep(size, flags);
2952 if (unlikely(cachep == NULL))
2954 return kmem_cache_alloc_node(cachep, flags, node);
2956 EXPORT_SYMBOL(kmalloc_node);
2960 * kmalloc - allocate memory
2961 * @size: how many bytes of memory are required.
2962 * @flags: the type of memory to allocate.
2964 * kmalloc is the normal method of allocating memory
2967 * The @flags argument may be one of:
2969 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2971 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2973 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2975 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2976 * must be suitable for DMA. This can mean different things on different
2977 * platforms. For example, on i386, it means that the memory must come
2978 * from the first 16MB.
2980 void *__kmalloc(size_t size, gfp_t flags)
2982 kmem_cache_t *cachep;
2984 /* If you want to save a few bytes .text space: replace
2986 * Then kmalloc uses the uninlined functions instead of the inline
2989 cachep = __find_general_cachep(size, flags);
2990 if (unlikely(cachep == NULL))
2992 return __cache_alloc(cachep, flags);
2994 EXPORT_SYMBOL(__kmalloc);
2998 * __alloc_percpu - allocate one copy of the object for every present
2999 * cpu in the system, zeroing them.
3000 * Objects should be dereferenced using the per_cpu_ptr macro only.
3002 * @size: how many bytes of memory are required.
3004 void *__alloc_percpu(size_t size)
3007 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3013 * Cannot use for_each_online_cpu since a cpu may come online
3014 * and we have no way of figuring out how to fix the array
3015 * that we have allocated then....
3018 int node = cpu_to_node(i);
3020 if (node_online(node))
3021 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3023 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3025 if (!pdata->ptrs[i])
3027 memset(pdata->ptrs[i], 0, size);
3030 /* Catch derefs w/o wrappers */
3031 return (void *)(~(unsigned long)pdata);
3035 if (!cpu_possible(i))
3037 kfree(pdata->ptrs[i]);
3042 EXPORT_SYMBOL(__alloc_percpu);
3046 * kmem_cache_free - Deallocate an object
3047 * @cachep: The cache the allocation was from.
3048 * @objp: The previously allocated object.
3050 * Free an object which was previously allocated from this
3053 void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3055 unsigned long flags;
3057 local_irq_save(flags);
3058 __cache_free(cachep, objp);
3059 local_irq_restore(flags);
3061 EXPORT_SYMBOL(kmem_cache_free);
3064 * kfree - free previously allocated memory
3065 * @objp: pointer returned by kmalloc.
3067 * If @objp is NULL, no operation is performed.
3069 * Don't free memory not originally allocated by kmalloc()
3070 * or you will run into trouble.
3072 void kfree(const void *objp)
3075 unsigned long flags;
3077 if (unlikely(!objp))
3079 local_irq_save(flags);
3080 kfree_debugcheck(objp);
3081 c = page_get_cache(virt_to_page(objp));
3082 mutex_debug_check_no_locks_freed(objp, obj_size(c));
3083 __cache_free(c, (void *)objp);
3084 local_irq_restore(flags);
3086 EXPORT_SYMBOL(kfree);
3090 * free_percpu - free previously allocated percpu memory
3091 * @objp: pointer returned by alloc_percpu.
3093 * Don't free memory not originally allocated by alloc_percpu()
3094 * The complemented objp is to check for that.
3096 void free_percpu(const void *objp)
3099 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3102 * We allocate for all cpus so we cannot use for online cpu here.
3108 EXPORT_SYMBOL(free_percpu);
3111 unsigned int kmem_cache_size(kmem_cache_t *cachep)
3113 return obj_size(cachep);
3115 EXPORT_SYMBOL(kmem_cache_size);
3117 const char *kmem_cache_name(kmem_cache_t *cachep)
3119 return cachep->name;
3121 EXPORT_SYMBOL_GPL(kmem_cache_name);
3124 * This initializes kmem_list3 for all nodes.
3126 static int alloc_kmemlist(kmem_cache_t *cachep)
3129 struct kmem_list3 *l3;
3132 for_each_online_node(node) {
3133 struct array_cache *nc = NULL, *new;
3134 struct array_cache **new_alien = NULL;
3136 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3139 if (!(new = alloc_arraycache(node, (cachep->shared *
3140 cachep->batchcount),
3143 if ((l3 = cachep->nodelists[node])) {
3145 spin_lock_irq(&l3->list_lock);
3147 if ((nc = cachep->nodelists[node]->shared))
3148 free_block(cachep, nc->entry, nc->avail, node);
3151 if (!cachep->nodelists[node]->alien) {
3152 l3->alien = new_alien;
3155 l3->free_limit = (1 + nr_cpus_node(node)) *
3156 cachep->batchcount + cachep->num;
3157 spin_unlock_irq(&l3->list_lock);
3159 free_alien_cache(new_alien);
3162 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3166 kmem_list3_init(l3);
3167 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3168 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3170 l3->alien = new_alien;
3171 l3->free_limit = (1 + nr_cpus_node(node)) *
3172 cachep->batchcount + cachep->num;
3173 cachep->nodelists[node] = l3;
3181 struct ccupdate_struct {
3182 kmem_cache_t *cachep;
3183 struct array_cache *new[NR_CPUS];
3186 static void do_ccupdate_local(void *info)
3188 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3189 struct array_cache *old;
3192 old = ac_data(new->cachep);
3194 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3195 new->new[smp_processor_id()] = old;
3198 static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3201 struct ccupdate_struct new;
3204 memset(&new.new, 0, sizeof(new.new));
3205 for_each_online_cpu(i) {
3207 alloc_arraycache(cpu_to_node(i), limit, batchcount);
3209 for (i--; i >= 0; i--)
3214 new.cachep = cachep;
3216 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3219 spin_lock_irq(&cachep->spinlock);
3220 cachep->batchcount = batchcount;
3221 cachep->limit = limit;
3222 cachep->shared = shared;
3223 spin_unlock_irq(&cachep->spinlock);
3225 for_each_online_cpu(i) {
3226 struct array_cache *ccold = new.new[i];
3229 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3230 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3231 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3235 err = alloc_kmemlist(cachep);
3237 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3238 cachep->name, -err);
3244 static void enable_cpucache(kmem_cache_t *cachep)
3249 /* The head array serves three purposes:
3250 * - create a LIFO ordering, i.e. return objects that are cache-warm
3251 * - reduce the number of spinlock operations.
3252 * - reduce the number of linked list operations on the slab and
3253 * bufctl chains: array operations are cheaper.
3254 * The numbers are guessed, we should auto-tune as described by
3257 if (cachep->buffer_size > 131072)
3259 else if (cachep->buffer_size > PAGE_SIZE)
3261 else if (cachep->buffer_size > 1024)
3263 else if (cachep->buffer_size > 256)
3268 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3269 * allocation behaviour: Most allocs on one cpu, most free operations
3270 * on another cpu. For these cases, an efficient object passing between
3271 * cpus is necessary. This is provided by a shared array. The array
3272 * replaces Bonwick's magazine layer.
3273 * On uniprocessor, it's functionally equivalent (but less efficient)
3274 * to a larger limit. Thus disabled by default.
3278 if (cachep->buffer_size <= PAGE_SIZE)
3283 /* With debugging enabled, large batchcount lead to excessively
3284 * long periods with disabled local interrupts. Limit the
3290 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3292 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3293 cachep->name, -err);
3296 static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
3297 int force, int node)
3301 check_spinlock_acquired_node(cachep, node);
3302 if (ac->touched && !force) {
3304 } else if (ac->avail) {
3305 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3306 if (tofree > ac->avail) {
3307 tofree = (ac->avail + 1) / 2;
3309 free_block(cachep, ac->entry, tofree, node);
3310 ac->avail -= tofree;
3311 memmove(ac->entry, &(ac->entry[tofree]),
3312 sizeof(void *) * ac->avail);
3317 * cache_reap - Reclaim memory from caches.
3318 * @unused: unused parameter
3320 * Called from workqueue/eventd every few seconds.
3322 * - clear the per-cpu caches for this CPU.
3323 * - return freeable pages to the main free memory pool.
3325 * If we cannot acquire the cache chain mutex then just give up - we'll
3326 * try again on the next iteration.
3328 static void cache_reap(void *unused)
3330 struct list_head *walk;
3331 struct kmem_list3 *l3;
3333 if (!mutex_trylock(&cache_chain_mutex)) {
3334 /* Give up. Setup the next iteration. */
3335 schedule_delayed_work(&__get_cpu_var(reap_work),
3340 list_for_each(walk, &cache_chain) {
3341 kmem_cache_t *searchp;
3342 struct list_head *p;
3346 searchp = list_entry(walk, kmem_cache_t, next);
3348 if (searchp->flags & SLAB_NO_REAP)
3353 l3 = searchp->nodelists[numa_node_id()];
3355 drain_alien_cache(searchp, l3);
3356 spin_lock_irq(&l3->list_lock);
3358 drain_array_locked(searchp, ac_data(searchp), 0,
3361 if (time_after(l3->next_reap, jiffies))
3364 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3367 drain_array_locked(searchp, l3->shared, 0,
3370 if (l3->free_touched) {
3371 l3->free_touched = 0;
3376 (l3->free_limit + 5 * searchp->num -
3377 1) / (5 * searchp->num);
3379 p = l3->slabs_free.next;
3380 if (p == &(l3->slabs_free))
3383 slabp = list_entry(p, struct slab, list);
3384 BUG_ON(slabp->inuse);
3385 list_del(&slabp->list);
3386 STATS_INC_REAPED(searchp);
3388 /* Safe to drop the lock. The slab is no longer
3389 * linked to the cache.
3390 * searchp cannot disappear, we hold
3393 l3->free_objects -= searchp->num;
3394 spin_unlock_irq(&l3->list_lock);
3395 slab_destroy(searchp, slabp);
3396 spin_lock_irq(&l3->list_lock);
3397 } while (--tofree > 0);
3399 spin_unlock_irq(&l3->list_lock);
3404 mutex_unlock(&cache_chain_mutex);
3405 drain_remote_pages();
3406 /* Setup the next iteration */
3407 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3410 #ifdef CONFIG_PROC_FS
3412 static void print_slabinfo_header(struct seq_file *m)
3415 * Output format version, so at least we can change it
3416 * without _too_ many complaints.
3419 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3421 seq_puts(m, "slabinfo - version: 2.1\n");
3423 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3424 "<objperslab> <pagesperslab>");
3425 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3426 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3428 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3429 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3430 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3435 static void *s_start(struct seq_file *m, loff_t *pos)
3438 struct list_head *p;
3440 mutex_lock(&cache_chain_mutex);
3442 print_slabinfo_header(m);
3443 p = cache_chain.next;
3446 if (p == &cache_chain)
3449 return list_entry(p, kmem_cache_t, next);
3452 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3454 kmem_cache_t *cachep = p;
3456 return cachep->next.next == &cache_chain ? NULL
3457 : list_entry(cachep->next.next, kmem_cache_t, next);
3460 static void s_stop(struct seq_file *m, void *p)
3462 mutex_unlock(&cache_chain_mutex);
3465 static int s_show(struct seq_file *m, void *p)
3467 kmem_cache_t *cachep = p;
3468 struct list_head *q;
3470 unsigned long active_objs;
3471 unsigned long num_objs;
3472 unsigned long active_slabs = 0;
3473 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3477 struct kmem_list3 *l3;
3480 spin_lock_irq(&cachep->spinlock);
3483 for_each_online_node(node) {
3484 l3 = cachep->nodelists[node];
3488 spin_lock(&l3->list_lock);
3490 list_for_each(q, &l3->slabs_full) {
3491 slabp = list_entry(q, struct slab, list);
3492 if (slabp->inuse != cachep->num && !error)
3493 error = "slabs_full accounting error";
3494 active_objs += cachep->num;
3497 list_for_each(q, &l3->slabs_partial) {
3498 slabp = list_entry(q, struct slab, list);
3499 if (slabp->inuse == cachep->num && !error)
3500 error = "slabs_partial inuse accounting error";
3501 if (!slabp->inuse && !error)
3502 error = "slabs_partial/inuse accounting error";
3503 active_objs += slabp->inuse;
3506 list_for_each(q, &l3->slabs_free) {
3507 slabp = list_entry(q, struct slab, list);
3508 if (slabp->inuse && !error)
3509 error = "slabs_free/inuse accounting error";
3512 free_objects += l3->free_objects;
3513 shared_avail += l3->shared->avail;
3515 spin_unlock(&l3->list_lock);
3517 num_slabs += active_slabs;
3518 num_objs = num_slabs * cachep->num;
3519 if (num_objs - active_objs != free_objects && !error)
3520 error = "free_objects accounting error";
3522 name = cachep->name;
3524 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3526 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3527 name, active_objs, num_objs, cachep->buffer_size,
3528 cachep->num, (1 << cachep->gfporder));
3529 seq_printf(m, " : tunables %4u %4u %4u",
3530 cachep->limit, cachep->batchcount, cachep->shared);
3531 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3532 active_slabs, num_slabs, shared_avail);
3535 unsigned long high = cachep->high_mark;
3536 unsigned long allocs = cachep->num_allocations;
3537 unsigned long grown = cachep->grown;
3538 unsigned long reaped = cachep->reaped;
3539 unsigned long errors = cachep->errors;
3540 unsigned long max_freeable = cachep->max_freeable;
3541 unsigned long node_allocs = cachep->node_allocs;
3542 unsigned long node_frees = cachep->node_frees;
3544 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3545 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
3549 unsigned long allochit = atomic_read(&cachep->allochit);
3550 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3551 unsigned long freehit = atomic_read(&cachep->freehit);
3552 unsigned long freemiss = atomic_read(&cachep->freemiss);
3554 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3555 allochit, allocmiss, freehit, freemiss);
3559 spin_unlock_irq(&cachep->spinlock);
3564 * slabinfo_op - iterator that generates /proc/slabinfo
3573 * num-pages-per-slab
3574 * + further values on SMP and with statistics enabled
3577 struct seq_operations slabinfo_op = {
3584 #define MAX_SLABINFO_WRITE 128
3586 * slabinfo_write - Tuning for the slab allocator
3588 * @buffer: user buffer
3589 * @count: data length
3592 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3593 size_t count, loff_t *ppos)
3595 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3596 int limit, batchcount, shared, res;
3597 struct list_head *p;
3599 if (count > MAX_SLABINFO_WRITE)
3601 if (copy_from_user(&kbuf, buffer, count))
3603 kbuf[MAX_SLABINFO_WRITE] = '\0';
3605 tmp = strchr(kbuf, ' ');
3610 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3613 /* Find the cache in the chain of caches. */
3614 mutex_lock(&cache_chain_mutex);
3616 list_for_each(p, &cache_chain) {
3617 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3619 if (!strcmp(cachep->name, kbuf)) {
3622 batchcount > limit || shared < 0) {
3625 res = do_tune_cpucache(cachep, limit,
3626 batchcount, shared);
3631 mutex_unlock(&cache_chain_mutex);
3639 * ksize - get the actual amount of memory allocated for a given object
3640 * @objp: Pointer to the object
3642 * kmalloc may internally round up allocations and return more memory
3643 * than requested. ksize() can be used to determine the actual amount of
3644 * memory allocated. The caller may use this additional memory, even though
3645 * a smaller amount of memory was initially specified with the kmalloc call.
3646 * The caller must guarantee that objp points to a valid object previously
3647 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3648 * must not be freed during the duration of the call.
3650 unsigned int ksize(const void *objp)
3652 if (unlikely(objp == NULL))
3655 return obj_size(page_get_cache(virt_to_page(objp)));