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 semaphore 'cache_chain_sem'.
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>
107 #include <asm/uaccess.h>
108 #include <asm/cacheflush.h>
109 #include <asm/tlbflush.h>
110 #include <asm/page.h>
113 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
114 * SLAB_RED_ZONE & SLAB_POISON.
115 * 0 for faster, smaller code (especially in the critical paths).
117 * STATS - 1 to collect stats for /proc/slabinfo.
118 * 0 for faster, smaller code (especially in the critical paths).
120 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
123 #ifdef CONFIG_DEBUG_SLAB
126 #define FORCED_DEBUG 1
130 #define FORCED_DEBUG 0
133 /* Shouldn't this be in a header file somewhere? */
134 #define BYTES_PER_WORD sizeof(void *)
136 #ifndef cache_line_size
137 #define cache_line_size() L1_CACHE_BYTES
140 #ifndef ARCH_KMALLOC_MINALIGN
142 * Enforce a minimum alignment for the kmalloc caches.
143 * Usually, the kmalloc caches are cache_line_size() aligned, except when
144 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
145 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
146 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
147 * Note that this flag disables some debug features.
149 #define ARCH_KMALLOC_MINALIGN 0
152 #ifndef ARCH_SLAB_MINALIGN
154 * Enforce a minimum alignment for all caches.
155 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
156 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
157 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
158 * some debug features.
160 #define ARCH_SLAB_MINALIGN 0
163 #ifndef ARCH_KMALLOC_FLAGS
164 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
167 /* Legal flag mask for kmem_cache_create(). */
169 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
170 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
171 SLAB_NO_REAP | SLAB_CACHE_DMA | \
172 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
173 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
176 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
177 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 * Bufctl's are used for linking objs within a slab
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 typedef unsigned int kmem_bufctl_t;
202 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
204 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
206 /* Max number of objs-per-slab for caches which use off-slab slabs.
207 * Needed to avoid a possible looping condition in cache_grow().
209 static unsigned long offslab_limit;
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 kmem_cache_t *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 [].
275 /* bootstrap: The caches do not work without cpuarrays anymore,
276 * but the cpuarrays are allocated from the generic caches...
278 #define BOOT_CPUCACHE_ENTRIES 1
279 struct arraycache_init {
280 struct array_cache cache;
281 void *entries[BOOT_CPUCACHE_ENTRIES];
285 * The slab lists for all objects.
288 struct list_head slabs_partial; /* partial list first, better asm code */
289 struct list_head slabs_full;
290 struct list_head slabs_free;
291 unsigned long free_objects;
292 unsigned long next_reap;
294 unsigned int free_limit;
295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
301 * Need this for bootstrapping a per node allocator.
303 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
304 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
305 #define CACHE_CACHE 0
307 #define SIZE_L3 (1 + MAX_NUMNODES)
310 * This function must be completely optimized away if
311 * a constant is passed to it. Mostly the same as
312 * what is in linux/slab.h except it returns an
315 static __always_inline int index_of(const size_t size)
317 if (__builtin_constant_p(size)) {
325 #include "linux/kmalloc_sizes.h"
328 extern void __bad_size(void);
336 #define INDEX_AC index_of(sizeof(struct arraycache_init))
337 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
339 static inline void kmem_list3_init(struct kmem_list3 *parent)
341 INIT_LIST_HEAD(&parent->slabs_full);
342 INIT_LIST_HEAD(&parent->slabs_partial);
343 INIT_LIST_HEAD(&parent->slabs_free);
344 parent->shared = NULL;
345 parent->alien = NULL;
346 spin_lock_init(&parent->list_lock);
347 parent->free_objects = 0;
348 parent->free_touched = 0;
351 #define MAKE_LIST(cachep, listp, slab, nodeid) \
353 INIT_LIST_HEAD(listp); \
354 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
357 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
359 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
360 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
371 /* 1) per-cpu data, touched during every alloc/free */
372 struct array_cache *array[NR_CPUS];
373 unsigned int batchcount;
376 unsigned int objsize;
377 /* 2) touched by every alloc & free from the backend */
378 struct kmem_list3 *nodelists[MAX_NUMNODES];
379 unsigned int flags; /* constant flags */
380 unsigned int num; /* # of objs per slab */
383 /* 3) cache_grow/shrink */
384 /* order of pgs per slab (2^n) */
385 unsigned int gfporder;
387 /* force GFP flags, e.g. GFP_DMA */
390 size_t colour; /* cache colouring range */
391 unsigned int colour_off; /* colour offset */
392 unsigned int colour_next; /* cache colouring */
393 kmem_cache_t *slabp_cache;
394 unsigned int slab_size;
395 unsigned int dflags; /* dynamic flags */
397 /* constructor func */
398 void (*ctor) (void *, kmem_cache_t *, unsigned long);
400 /* de-constructor func */
401 void (*dtor) (void *, kmem_cache_t *, unsigned long);
403 /* 4) cache creation/removal */
405 struct list_head next;
409 unsigned long num_active;
410 unsigned long num_allocations;
411 unsigned long high_mark;
413 unsigned long reaped;
414 unsigned long errors;
415 unsigned long max_freeable;
416 unsigned long node_allocs;
417 unsigned long node_frees;
429 #define CFLGS_OFF_SLAB (0x80000000UL)
430 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
432 #define BATCHREFILL_LIMIT 16
433 /* Optimization question: fewer reaps means less
434 * probability for unnessary cpucache drain/refill cycles.
436 * OTOH the cpuarrays can contain lots of objects,
437 * which could lock up otherwise freeable slabs.
439 #define REAPTIMEOUT_CPUC (2*HZ)
440 #define REAPTIMEOUT_LIST3 (4*HZ)
443 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
444 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
445 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
446 #define STATS_INC_GROWN(x) ((x)->grown++)
447 #define STATS_INC_REAPED(x) ((x)->reaped++)
448 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
449 (x)->high_mark = (x)->num_active; \
451 #define STATS_INC_ERR(x) ((x)->errors++)
452 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
453 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
454 #define STATS_SET_FREEABLE(x, i) \
455 do { if ((x)->max_freeable < i) \
456 (x)->max_freeable = i; \
459 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
460 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
461 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
462 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
464 #define STATS_INC_ACTIVE(x) do { } while (0)
465 #define STATS_DEC_ACTIVE(x) do { } while (0)
466 #define STATS_INC_ALLOCED(x) do { } while (0)
467 #define STATS_INC_GROWN(x) do { } while (0)
468 #define STATS_INC_REAPED(x) do { } while (0)
469 #define STATS_SET_HIGH(x) do { } while (0)
470 #define STATS_INC_ERR(x) do { } while (0)
471 #define STATS_INC_NODEALLOCS(x) do { } while (0)
472 #define STATS_INC_NODEFREES(x) do { } while (0)
473 #define STATS_SET_FREEABLE(x, i) \
476 #define STATS_INC_ALLOCHIT(x) do { } while (0)
477 #define STATS_INC_ALLOCMISS(x) do { } while (0)
478 #define STATS_INC_FREEHIT(x) do { } while (0)
479 #define STATS_INC_FREEMISS(x) do { } while (0)
483 /* Magic nums for obj red zoning.
484 * Placed in the first word before and the first word after an obj.
486 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
487 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
489 /* ...and for poisoning */
490 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
491 #define POISON_FREE 0x6b /* for use-after-free poisoning */
492 #define POISON_END 0xa5 /* end-byte of poisoning */
494 /* memory layout of objects:
496 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
497 * the end of an object is aligned with the end of the real
498 * allocation. Catches writes behind the end of the allocation.
499 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
501 * cachep->dbghead: The real object.
502 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
503 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
505 static int obj_dbghead(kmem_cache_t *cachep)
507 return cachep->dbghead;
510 static int obj_reallen(kmem_cache_t *cachep)
512 return cachep->reallen;
515 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
517 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
518 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
521 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
523 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
524 if (cachep->flags & SLAB_STORE_USER)
525 return (unsigned long *)(objp + cachep->objsize -
527 return (unsigned long *)(objp + cachep->objsize - BYTES_PER_WORD);
530 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
532 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
533 return (void **)(objp + cachep->objsize - BYTES_PER_WORD);
538 #define obj_dbghead(x) 0
539 #define obj_reallen(cachep) (cachep->objsize)
540 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
541 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
542 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
547 * Maximum size of an obj (in 2^order pages)
548 * and absolute limit for the gfp order.
550 #if defined(CONFIG_LARGE_ALLOCS)
551 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
552 #define MAX_GFP_ORDER 13 /* up to 32Mb */
553 #elif defined(CONFIG_MMU)
554 #define MAX_OBJ_ORDER 5 /* 32 pages */
555 #define MAX_GFP_ORDER 5 /* 32 pages */
557 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
558 #define MAX_GFP_ORDER 8 /* up to 1Mb */
562 * Do not go above this order unless 0 objects fit into the slab.
564 #define BREAK_GFP_ORDER_HI 1
565 #define BREAK_GFP_ORDER_LO 0
566 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
568 /* Functions for storing/retrieving the cachep and or slab from the
569 * global 'mem_map'. These are used to find the slab an obj belongs to.
570 * With kfree(), these are used to find the cache which an obj belongs to.
572 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
574 page->lru.next = (struct list_head *)cache;
577 static inline struct kmem_cache *page_get_cache(struct page *page)
579 return (struct kmem_cache *)page->lru.next;
582 static inline void page_set_slab(struct page *page, struct slab *slab)
584 page->lru.prev = (struct list_head *)slab;
587 static inline struct slab *page_get_slab(struct page *page)
589 return (struct slab *)page->lru.prev;
592 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
593 struct cache_sizes malloc_sizes[] = {
594 #define CACHE(x) { .cs_size = (x) },
595 #include <linux/kmalloc_sizes.h>
599 EXPORT_SYMBOL(malloc_sizes);
601 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
607 static struct cache_names __initdata cache_names[] = {
608 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
609 #include <linux/kmalloc_sizes.h>
614 static struct arraycache_init initarray_cache __initdata =
615 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
616 static struct arraycache_init initarray_generic =
617 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
619 /* internal cache of cache description objs */
620 static kmem_cache_t cache_cache = {
622 .limit = BOOT_CPUCACHE_ENTRIES,
624 .objsize = sizeof(kmem_cache_t),
625 .flags = SLAB_NO_REAP,
626 .spinlock = SPIN_LOCK_UNLOCKED,
627 .name = "kmem_cache",
629 .reallen = sizeof(kmem_cache_t),
633 /* Guard access to the cache-chain. */
634 static struct semaphore cache_chain_sem;
635 static struct list_head cache_chain;
638 * vm_enough_memory() looks at this to determine how many
639 * slab-allocated pages are possibly freeable under pressure
641 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
643 atomic_t slab_reclaim_pages;
646 * chicken and egg problem: delay the per-cpu array allocation
647 * until the general caches are up.
656 static DEFINE_PER_CPU(struct work_struct, reap_work);
658 static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node);
659 static void enable_cpucache(kmem_cache_t *cachep);
660 static void cache_reap(void *unused);
661 static int __node_shrink(kmem_cache_t *cachep, int node);
663 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
665 return cachep->array[smp_processor_id()];
668 static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags)
670 struct cache_sizes *csizep = malloc_sizes;
673 /* This happens if someone tries to call
674 * kmem_cache_create(), or __kmalloc(), before
675 * the generic caches are initialized.
677 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
679 while (size > csizep->cs_size)
683 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
684 * has cs_{dma,}cachep==NULL. Thus no special case
685 * for large kmalloc calls required.
687 if (unlikely(gfpflags & GFP_DMA))
688 return csizep->cs_dmacachep;
689 return csizep->cs_cachep;
692 kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
694 return __find_general_cachep(size, gfpflags);
696 EXPORT_SYMBOL(kmem_find_general_cachep);
698 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
699 static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
700 int flags, size_t *left_over, unsigned int *num)
703 size_t wastage = PAGE_SIZE << gfporder;
707 if (!(flags & CFLGS_OFF_SLAB)) {
708 base = sizeof(struct slab);
709 extra = sizeof(kmem_bufctl_t);
712 while (i * size + ALIGN(base + i * extra, align) <= wastage)
722 wastage -= ALIGN(base + i * extra, align);
723 *left_over = wastage;
726 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
728 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
730 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
731 function, cachep->name, msg);
736 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
737 * via the workqueue/eventd.
738 * Add the CPU number into the expiration time to minimize the possibility of
739 * the CPUs getting into lockstep and contending for the global cache chain
742 static void __devinit start_cpu_timer(int cpu)
744 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
747 * When this gets called from do_initcalls via cpucache_init(),
748 * init_workqueues() has already run, so keventd will be setup
751 if (keventd_up() && reap_work->func == NULL) {
752 INIT_WORK(reap_work, cache_reap, NULL);
753 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
757 static struct array_cache *alloc_arraycache(int node, int entries,
760 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
761 struct array_cache *nc = NULL;
763 nc = kmalloc_node(memsize, GFP_KERNEL, node);
767 nc->batchcount = batchcount;
769 spin_lock_init(&nc->lock);
775 static inline struct array_cache **alloc_alien_cache(int node, int limit)
777 struct array_cache **ac_ptr;
778 int memsize = sizeof(void *) * MAX_NUMNODES;
783 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
786 if (i == node || !node_online(i)) {
790 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
792 for (i--; i <= 0; i--)
802 static inline void free_alien_cache(struct array_cache **ac_ptr)
815 static inline void __drain_alien_cache(kmem_cache_t *cachep,
816 struct array_cache *ac, int node)
818 struct kmem_list3 *rl3 = cachep->nodelists[node];
821 spin_lock(&rl3->list_lock);
822 free_block(cachep, ac->entry, ac->avail, node);
824 spin_unlock(&rl3->list_lock);
828 static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
831 struct array_cache *ac;
834 for_each_online_node(i) {
837 spin_lock_irqsave(&ac->lock, flags);
838 __drain_alien_cache(cachep, ac, i);
839 spin_unlock_irqrestore(&ac->lock, flags);
844 #define alloc_alien_cache(node, limit) do { } while (0)
845 #define free_alien_cache(ac_ptr) do { } while (0)
846 #define drain_alien_cache(cachep, l3) do { } while (0)
849 static int __devinit cpuup_callback(struct notifier_block *nfb,
850 unsigned long action, void *hcpu)
852 long cpu = (long)hcpu;
853 kmem_cache_t *cachep;
854 struct kmem_list3 *l3 = NULL;
855 int node = cpu_to_node(cpu);
856 int memsize = sizeof(struct kmem_list3);
857 struct array_cache *nc = NULL;
861 down(&cache_chain_sem);
862 /* we need to do this right in the beginning since
863 * alloc_arraycache's are going to use this list.
864 * kmalloc_node allows us to add the slab to the right
865 * kmem_list3 and not this cpu's kmem_list3
868 list_for_each_entry(cachep, &cache_chain, next) {
869 /* setup the size64 kmemlist for cpu before we can
870 * begin anything. Make sure some other cpu on this
871 * node has not already allocated this
873 if (!cachep->nodelists[node]) {
874 if (!(l3 = kmalloc_node(memsize,
878 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
879 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
881 cachep->nodelists[node] = l3;
884 spin_lock_irq(&cachep->nodelists[node]->list_lock);
885 cachep->nodelists[node]->free_limit =
886 (1 + nr_cpus_node(node)) *
887 cachep->batchcount + cachep->num;
888 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
891 /* Now we can go ahead with allocating the shared array's
893 list_for_each_entry(cachep, &cache_chain, next) {
894 nc = alloc_arraycache(node, cachep->limit,
898 cachep->array[cpu] = nc;
900 l3 = cachep->nodelists[node];
903 if (!(nc = alloc_arraycache(node,
909 /* we are serialised from CPU_DEAD or
910 CPU_UP_CANCELLED by the cpucontrol lock */
914 up(&cache_chain_sem);
917 start_cpu_timer(cpu);
919 #ifdef CONFIG_HOTPLUG_CPU
922 case CPU_UP_CANCELED:
923 down(&cache_chain_sem);
925 list_for_each_entry(cachep, &cache_chain, next) {
926 struct array_cache *nc;
929 mask = node_to_cpumask(node);
930 spin_lock_irq(&cachep->spinlock);
931 /* cpu is dead; no one can alloc from it. */
932 nc = cachep->array[cpu];
933 cachep->array[cpu] = NULL;
934 l3 = cachep->nodelists[node];
939 spin_lock(&l3->list_lock);
941 /* Free limit for this kmem_list3 */
942 l3->free_limit -= cachep->batchcount;
944 free_block(cachep, nc->entry, nc->avail, node);
946 if (!cpus_empty(mask)) {
947 spin_unlock(&l3->list_lock);
952 free_block(cachep, l3->shared->entry,
953 l3->shared->avail, node);
958 drain_alien_cache(cachep, l3);
959 free_alien_cache(l3->alien);
963 /* free slabs belonging to this node */
964 if (__node_shrink(cachep, node)) {
965 cachep->nodelists[node] = NULL;
966 spin_unlock(&l3->list_lock);
969 spin_unlock(&l3->list_lock);
972 spin_unlock_irq(&cachep->spinlock);
975 up(&cache_chain_sem);
981 up(&cache_chain_sem);
985 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
988 * swap the static kmem_list3 with kmalloced memory
990 static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid)
992 struct kmem_list3 *ptr;
994 BUG_ON(cachep->nodelists[nodeid] != list);
995 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
999 memcpy(ptr, list, sizeof(struct kmem_list3));
1000 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1001 cachep->nodelists[nodeid] = ptr;
1006 * Called after the gfp() functions have been enabled, and before smp_init().
1008 void __init kmem_cache_init(void)
1011 struct cache_sizes *sizes;
1012 struct cache_names *names;
1015 for (i = 0; i < NUM_INIT_LISTS; i++) {
1016 kmem_list3_init(&initkmem_list3[i]);
1017 if (i < MAX_NUMNODES)
1018 cache_cache.nodelists[i] = NULL;
1022 * Fragmentation resistance on low memory - only use bigger
1023 * page orders on machines with more than 32MB of memory.
1025 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1026 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1028 /* Bootstrap is tricky, because several objects are allocated
1029 * from caches that do not exist yet:
1030 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1031 * structures of all caches, except cache_cache itself: cache_cache
1032 * is statically allocated.
1033 * Initially an __init data area is used for the head array and the
1034 * kmem_list3 structures, it's replaced with a kmalloc allocated
1035 * array at the end of the bootstrap.
1036 * 2) Create the first kmalloc cache.
1037 * The kmem_cache_t for the new cache is allocated normally.
1038 * An __init data area is used for the head array.
1039 * 3) Create the remaining kmalloc caches, with minimally sized
1041 * 4) Replace the __init data head arrays for cache_cache and the first
1042 * kmalloc cache with kmalloc allocated arrays.
1043 * 5) Replace the __init data for kmem_list3 for cache_cache and
1044 * the other cache's with kmalloc allocated memory.
1045 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1048 /* 1) create the cache_cache */
1049 init_MUTEX(&cache_chain_sem);
1050 INIT_LIST_HEAD(&cache_chain);
1051 list_add(&cache_cache.next, &cache_chain);
1052 cache_cache.colour_off = cache_line_size();
1053 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1054 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1056 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1058 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
1059 &left_over, &cache_cache.num);
1060 if (!cache_cache.num)
1063 cache_cache.colour = left_over / cache_cache.colour_off;
1064 cache_cache.colour_next = 0;
1065 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1066 sizeof(struct slab), cache_line_size());
1068 /* 2+3) create the kmalloc caches */
1069 sizes = malloc_sizes;
1070 names = cache_names;
1072 /* Initialize the caches that provide memory for the array cache
1073 * and the kmem_list3 structures first.
1074 * Without this, further allocations will bug
1077 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1078 sizes[INDEX_AC].cs_size,
1079 ARCH_KMALLOC_MINALIGN,
1080 (ARCH_KMALLOC_FLAGS |
1081 SLAB_PANIC), NULL, NULL);
1083 if (INDEX_AC != INDEX_L3)
1084 sizes[INDEX_L3].cs_cachep =
1085 kmem_cache_create(names[INDEX_L3].name,
1086 sizes[INDEX_L3].cs_size,
1087 ARCH_KMALLOC_MINALIGN,
1088 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1091 while (sizes->cs_size != ULONG_MAX) {
1093 * For performance, all the general caches are L1 aligned.
1094 * This should be particularly beneficial on SMP boxes, as it
1095 * eliminates "false sharing".
1096 * Note for systems short on memory removing the alignment will
1097 * allow tighter packing of the smaller caches.
1099 if (!sizes->cs_cachep)
1100 sizes->cs_cachep = kmem_cache_create(names->name,
1102 ARCH_KMALLOC_MINALIGN,
1107 /* Inc off-slab bufctl limit until the ceiling is hit. */
1108 if (!(OFF_SLAB(sizes->cs_cachep))) {
1109 offslab_limit = sizes->cs_size - sizeof(struct slab);
1110 offslab_limit /= sizeof(kmem_bufctl_t);
1113 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1115 ARCH_KMALLOC_MINALIGN,
1116 (ARCH_KMALLOC_FLAGS |
1124 /* 4) Replace the bootstrap head arrays */
1128 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1130 local_irq_disable();
1131 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
1132 memcpy(ptr, ac_data(&cache_cache),
1133 sizeof(struct arraycache_init));
1134 cache_cache.array[smp_processor_id()] = ptr;
1137 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1139 local_irq_disable();
1140 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
1141 != &initarray_generic.cache);
1142 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
1143 sizeof(struct arraycache_init));
1144 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1148 /* 5) Replace the bootstrap kmem_list3's */
1151 /* Replace the static kmem_list3 structures for the boot cpu */
1152 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1155 for_each_online_node(node) {
1156 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1157 &initkmem_list3[SIZE_AC + node], node);
1159 if (INDEX_AC != INDEX_L3) {
1160 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1161 &initkmem_list3[SIZE_L3 + node],
1167 /* 6) resize the head arrays to their final sizes */
1169 kmem_cache_t *cachep;
1170 down(&cache_chain_sem);
1171 list_for_each_entry(cachep, &cache_chain, next)
1172 enable_cpucache(cachep);
1173 up(&cache_chain_sem);
1177 g_cpucache_up = FULL;
1179 /* Register a cpu startup notifier callback
1180 * that initializes ac_data for all new cpus
1182 register_cpu_notifier(&cpucache_notifier);
1184 /* The reap timers are started later, with a module init call:
1185 * That part of the kernel is not yet operational.
1189 static int __init cpucache_init(void)
1194 * Register the timers that return unneeded
1197 for_each_online_cpu(cpu)
1198 start_cpu_timer(cpu);
1203 __initcall(cpucache_init);
1206 * Interface to system's page allocator. No need to hold the cache-lock.
1208 * If we requested dmaable memory, we will get it. Even if we
1209 * did not request dmaable memory, we might get it, but that
1210 * would be relatively rare and ignorable.
1212 static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
1218 flags |= cachep->gfpflags;
1219 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1222 addr = page_address(page);
1224 i = (1 << cachep->gfporder);
1225 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1226 atomic_add(i, &slab_reclaim_pages);
1227 add_page_state(nr_slab, i);
1236 * Interface to system's page release.
1238 static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1240 unsigned long i = (1 << cachep->gfporder);
1241 struct page *page = virt_to_page(addr);
1242 const unsigned long nr_freed = i;
1245 if (!TestClearPageSlab(page))
1249 sub_page_state(nr_slab, nr_freed);
1250 if (current->reclaim_state)
1251 current->reclaim_state->reclaimed_slab += nr_freed;
1252 free_pages((unsigned long)addr, cachep->gfporder);
1253 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1254 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1257 static void kmem_rcu_free(struct rcu_head *head)
1259 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1260 kmem_cache_t *cachep = slab_rcu->cachep;
1262 kmem_freepages(cachep, slab_rcu->addr);
1263 if (OFF_SLAB(cachep))
1264 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1269 #ifdef CONFIG_DEBUG_PAGEALLOC
1270 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1271 unsigned long caller)
1273 int size = obj_reallen(cachep);
1275 addr = (unsigned long *)&((char *)addr)[obj_dbghead(cachep)];
1277 if (size < 5 * sizeof(unsigned long))
1280 *addr++ = 0x12345678;
1282 *addr++ = smp_processor_id();
1283 size -= 3 * sizeof(unsigned long);
1285 unsigned long *sptr = &caller;
1286 unsigned long svalue;
1288 while (!kstack_end(sptr)) {
1290 if (kernel_text_address(svalue)) {
1292 size -= sizeof(unsigned long);
1293 if (size <= sizeof(unsigned long))
1299 *addr++ = 0x87654321;
1303 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1305 int size = obj_reallen(cachep);
1306 addr = &((char *)addr)[obj_dbghead(cachep)];
1308 memset(addr, val, size);
1309 *(unsigned char *)(addr + size - 1) = POISON_END;
1312 static void dump_line(char *data, int offset, int limit)
1315 printk(KERN_ERR "%03x:", offset);
1316 for (i = 0; i < limit; i++) {
1317 printk(" %02x", (unsigned char)data[offset + i]);
1325 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1330 if (cachep->flags & SLAB_RED_ZONE) {
1331 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1332 *dbg_redzone1(cachep, objp),
1333 *dbg_redzone2(cachep, objp));
1336 if (cachep->flags & SLAB_STORE_USER) {
1337 printk(KERN_ERR "Last user: [<%p>]",
1338 *dbg_userword(cachep, objp));
1339 print_symbol("(%s)",
1340 (unsigned long)*dbg_userword(cachep, objp));
1343 realobj = (char *)objp + obj_dbghead(cachep);
1344 size = obj_reallen(cachep);
1345 for (i = 0; i < size && lines; i += 16, lines--) {
1348 if (i + limit > size)
1350 dump_line(realobj, i, limit);
1354 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1360 realobj = (char *)objp + obj_dbghead(cachep);
1361 size = obj_reallen(cachep);
1363 for (i = 0; i < size; i++) {
1364 char exp = POISON_FREE;
1367 if (realobj[i] != exp) {
1373 "Slab corruption: start=%p, len=%d\n",
1375 print_objinfo(cachep, objp, 0);
1377 /* Hexdump the affected line */
1380 if (i + limit > size)
1382 dump_line(realobj, i, limit);
1385 /* Limit to 5 lines */
1391 /* Print some data about the neighboring objects, if they
1394 struct slab *slabp = page_get_slab(virt_to_page(objp));
1397 objnr = (objp - slabp->s_mem) / cachep->objsize;
1399 objp = slabp->s_mem + (objnr - 1) * cachep->objsize;
1400 realobj = (char *)objp + obj_dbghead(cachep);
1401 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1403 print_objinfo(cachep, objp, 2);
1405 if (objnr + 1 < cachep->num) {
1406 objp = slabp->s_mem + (objnr + 1) * cachep->objsize;
1407 realobj = (char *)objp + obj_dbghead(cachep);
1408 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1410 print_objinfo(cachep, objp, 2);
1416 /* Destroy all the objs in a slab, and release the mem back to the system.
1417 * Before calling the slab must have been unlinked from the cache.
1418 * The cache-lock is not held/needed.
1420 static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp)
1422 void *addr = slabp->s_mem - slabp->colouroff;
1426 for (i = 0; i < cachep->num; i++) {
1427 void *objp = slabp->s_mem + cachep->objsize * i;
1429 if (cachep->flags & SLAB_POISON) {
1430 #ifdef CONFIG_DEBUG_PAGEALLOC
1431 if ((cachep->objsize % PAGE_SIZE) == 0
1432 && OFF_SLAB(cachep))
1433 kernel_map_pages(virt_to_page(objp),
1434 cachep->objsize / PAGE_SIZE,
1437 check_poison_obj(cachep, objp);
1439 check_poison_obj(cachep, objp);
1442 if (cachep->flags & SLAB_RED_ZONE) {
1443 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1444 slab_error(cachep, "start of a freed object "
1446 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1447 slab_error(cachep, "end of a freed object "
1450 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1451 (cachep->dtor) (objp + obj_dbghead(cachep), cachep, 0);
1456 for (i = 0; i < cachep->num; i++) {
1457 void *objp = slabp->s_mem + cachep->objsize * i;
1458 (cachep->dtor) (objp, cachep, 0);
1463 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1464 struct slab_rcu *slab_rcu;
1466 slab_rcu = (struct slab_rcu *)slabp;
1467 slab_rcu->cachep = cachep;
1468 slab_rcu->addr = addr;
1469 call_rcu(&slab_rcu->head, kmem_rcu_free);
1471 kmem_freepages(cachep, addr);
1472 if (OFF_SLAB(cachep))
1473 kmem_cache_free(cachep->slabp_cache, slabp);
1477 /* For setting up all the kmem_list3s for cache whose objsize is same
1478 as size of kmem_list3. */
1479 static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1483 for_each_online_node(node) {
1484 cachep->nodelists[node] = &initkmem_list3[index + node];
1485 cachep->nodelists[node]->next_reap = jiffies +
1487 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1492 * calculate_slab_order - calculate size (page order) of slabs and the number
1493 * of objects per slab.
1495 * This could be made much more intelligent. For now, try to avoid using
1496 * high order pages for slabs. When the gfp() functions are more friendly
1497 * towards high-order requests, this should be changed.
1499 static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size,
1500 size_t align, gfp_t flags)
1502 size_t left_over = 0;
1504 for (;; cachep->gfporder++) {
1508 if (cachep->gfporder > MAX_GFP_ORDER) {
1513 cache_estimate(cachep->gfporder, size, align, flags,
1517 /* More than offslab_limit objects will cause problems */
1518 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1522 left_over = remainder;
1525 * Large number of objects is good, but very large slabs are
1526 * currently bad for the gfp()s.
1528 if (cachep->gfporder >= slab_break_gfp_order)
1531 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1532 /* Acceptable internal fragmentation */
1539 * kmem_cache_create - Create a cache.
1540 * @name: A string which is used in /proc/slabinfo to identify this cache.
1541 * @size: The size of objects to be created in this cache.
1542 * @align: The required alignment for the objects.
1543 * @flags: SLAB flags
1544 * @ctor: A constructor for the objects.
1545 * @dtor: A destructor for the objects.
1547 * Returns a ptr to the cache on success, NULL on failure.
1548 * Cannot be called within a int, but can be interrupted.
1549 * The @ctor is run when new pages are allocated by the cache
1550 * and the @dtor is run before the pages are handed back.
1552 * @name must be valid until the cache is destroyed. This implies that
1553 * the module calling this has to destroy the cache before getting
1558 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1559 * to catch references to uninitialised memory.
1561 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1562 * for buffer overruns.
1564 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1567 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1568 * cacheline. This can be beneficial if you're counting cycles as closely
1572 kmem_cache_create (const char *name, size_t size, size_t align,
1573 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1574 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1576 size_t left_over, slab_size, ralign;
1577 kmem_cache_t *cachep = NULL;
1578 struct list_head *p;
1581 * Sanity checks... these are all serious usage bugs.
1585 (size < BYTES_PER_WORD) ||
1586 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1587 printk(KERN_ERR "%s: Early error in slab %s\n",
1588 __FUNCTION__, name);
1592 down(&cache_chain_sem);
1594 list_for_each(p, &cache_chain) {
1595 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1596 mm_segment_t old_fs = get_fs();
1601 * This happens when the module gets unloaded and doesn't
1602 * destroy its slab cache and no-one else reuses the vmalloc
1603 * area of the module. Print a warning.
1606 res = __get_user(tmp, pc->name);
1609 printk("SLAB: cache with size %d has lost its name\n",
1614 if (!strcmp(pc->name, name)) {
1615 printk("kmem_cache_create: duplicate cache %s\n", name);
1622 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1623 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1624 /* No constructor, but inital state check requested */
1625 printk(KERN_ERR "%s: No con, but init state check "
1626 "requested - %s\n", __FUNCTION__, name);
1627 flags &= ~SLAB_DEBUG_INITIAL;
1631 * Enable redzoning and last user accounting, except for caches with
1632 * large objects, if the increased size would increase the object size
1633 * above the next power of two: caches with object sizes just above a
1634 * power of two have a significant amount of internal fragmentation.
1637 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1638 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1639 if (!(flags & SLAB_DESTROY_BY_RCU))
1640 flags |= SLAB_POISON;
1642 if (flags & SLAB_DESTROY_BY_RCU)
1643 BUG_ON(flags & SLAB_POISON);
1645 if (flags & SLAB_DESTROY_BY_RCU)
1649 * Always checks flags, a caller might be expecting debug
1650 * support which isn't available.
1652 if (flags & ~CREATE_MASK)
1655 /* Check that size is in terms of words. This is needed to avoid
1656 * unaligned accesses for some archs when redzoning is used, and makes
1657 * sure any on-slab bufctl's are also correctly aligned.
1659 if (size & (BYTES_PER_WORD - 1)) {
1660 size += (BYTES_PER_WORD - 1);
1661 size &= ~(BYTES_PER_WORD - 1);
1664 /* calculate out the final buffer alignment: */
1665 /* 1) arch recommendation: can be overridden for debug */
1666 if (flags & SLAB_HWCACHE_ALIGN) {
1667 /* Default alignment: as specified by the arch code.
1668 * Except if an object is really small, then squeeze multiple
1669 * objects into one cacheline.
1671 ralign = cache_line_size();
1672 while (size <= ralign / 2)
1675 ralign = BYTES_PER_WORD;
1677 /* 2) arch mandated alignment: disables debug if necessary */
1678 if (ralign < ARCH_SLAB_MINALIGN) {
1679 ralign = ARCH_SLAB_MINALIGN;
1680 if (ralign > BYTES_PER_WORD)
1681 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1683 /* 3) caller mandated alignment: disables debug if necessary */
1684 if (ralign < align) {
1686 if (ralign > BYTES_PER_WORD)
1687 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1689 /* 4) Store it. Note that the debug code below can reduce
1690 * the alignment to BYTES_PER_WORD.
1694 /* Get cache's description obj. */
1695 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1698 memset(cachep, 0, sizeof(kmem_cache_t));
1701 cachep->reallen = size;
1703 if (flags & SLAB_RED_ZONE) {
1704 /* redzoning only works with word aligned caches */
1705 align = BYTES_PER_WORD;
1707 /* add space for red zone words */
1708 cachep->dbghead += BYTES_PER_WORD;
1709 size += 2 * BYTES_PER_WORD;
1711 if (flags & SLAB_STORE_USER) {
1712 /* user store requires word alignment and
1713 * one word storage behind the end of the real
1716 align = BYTES_PER_WORD;
1717 size += BYTES_PER_WORD;
1719 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1720 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1721 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1722 cachep->dbghead += PAGE_SIZE - size;
1728 /* Determine if the slab management is 'on' or 'off' slab. */
1729 if (size >= (PAGE_SIZE >> 3))
1731 * Size is large, assume best to place the slab management obj
1732 * off-slab (should allow better packing of objs).
1734 flags |= CFLGS_OFF_SLAB;
1736 size = ALIGN(size, align);
1738 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1740 * A VFS-reclaimable slab tends to have most allocations
1741 * as GFP_NOFS and we really don't want to have to be allocating
1742 * higher-order pages when we are unable to shrink dcache.
1744 cachep->gfporder = 0;
1745 cache_estimate(cachep->gfporder, size, align, flags,
1746 &left_over, &cachep->num);
1748 left_over = calculate_slab_order(cachep, size, align, flags);
1751 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1752 kmem_cache_free(&cache_cache, cachep);
1756 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1757 + sizeof(struct slab), align);
1760 * If the slab has been placed off-slab, and we have enough space then
1761 * move it on-slab. This is at the expense of any extra colouring.
1763 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1764 flags &= ~CFLGS_OFF_SLAB;
1765 left_over -= slab_size;
1768 if (flags & CFLGS_OFF_SLAB) {
1769 /* really off slab. No need for manual alignment */
1771 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
1774 cachep->colour_off = cache_line_size();
1775 /* Offset must be a multiple of the alignment. */
1776 if (cachep->colour_off < align)
1777 cachep->colour_off = align;
1778 cachep->colour = left_over / cachep->colour_off;
1779 cachep->slab_size = slab_size;
1780 cachep->flags = flags;
1781 cachep->gfpflags = 0;
1782 if (flags & SLAB_CACHE_DMA)
1783 cachep->gfpflags |= GFP_DMA;
1784 spin_lock_init(&cachep->spinlock);
1785 cachep->objsize = size;
1787 if (flags & CFLGS_OFF_SLAB)
1788 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
1789 cachep->ctor = ctor;
1790 cachep->dtor = dtor;
1791 cachep->name = name;
1793 /* Don't let CPUs to come and go */
1796 if (g_cpucache_up == FULL) {
1797 enable_cpucache(cachep);
1799 if (g_cpucache_up == NONE) {
1800 /* Note: the first kmem_cache_create must create
1801 * the cache that's used by kmalloc(24), otherwise
1802 * the creation of further caches will BUG().
1804 cachep->array[smp_processor_id()] =
1805 &initarray_generic.cache;
1807 /* If the cache that's used by
1808 * kmalloc(sizeof(kmem_list3)) is the first cache,
1809 * then we need to set up all its list3s, otherwise
1810 * the creation of further caches will BUG().
1812 set_up_list3s(cachep, SIZE_AC);
1813 if (INDEX_AC == INDEX_L3)
1814 g_cpucache_up = PARTIAL_L3;
1816 g_cpucache_up = PARTIAL_AC;
1818 cachep->array[smp_processor_id()] =
1819 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1821 if (g_cpucache_up == PARTIAL_AC) {
1822 set_up_list3s(cachep, SIZE_L3);
1823 g_cpucache_up = PARTIAL_L3;
1826 for_each_online_node(node) {
1828 cachep->nodelists[node] =
1830 (struct kmem_list3),
1832 BUG_ON(!cachep->nodelists[node]);
1833 kmem_list3_init(cachep->
1838 cachep->nodelists[numa_node_id()]->next_reap =
1839 jiffies + REAPTIMEOUT_LIST3 +
1840 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1842 BUG_ON(!ac_data(cachep));
1843 ac_data(cachep)->avail = 0;
1844 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1845 ac_data(cachep)->batchcount = 1;
1846 ac_data(cachep)->touched = 0;
1847 cachep->batchcount = 1;
1848 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1851 /* cache setup completed, link it into the list */
1852 list_add(&cachep->next, &cache_chain);
1853 unlock_cpu_hotplug();
1855 if (!cachep && (flags & SLAB_PANIC))
1856 panic("kmem_cache_create(): failed to create slab `%s'\n",
1858 up(&cache_chain_sem);
1861 EXPORT_SYMBOL(kmem_cache_create);
1864 static void check_irq_off(void)
1866 BUG_ON(!irqs_disabled());
1869 static void check_irq_on(void)
1871 BUG_ON(irqs_disabled());
1874 static void check_spinlock_acquired(kmem_cache_t *cachep)
1878 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
1882 static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1886 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1891 #define check_irq_off() do { } while(0)
1892 #define check_irq_on() do { } while(0)
1893 #define check_spinlock_acquired(x) do { } while(0)
1894 #define check_spinlock_acquired_node(x, y) do { } while(0)
1898 * Waits for all CPUs to execute func().
1900 static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
1905 local_irq_disable();
1909 if (smp_call_function(func, arg, 1, 1))
1915 static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
1916 int force, int node);
1918 static void do_drain(void *arg)
1920 kmem_cache_t *cachep = (kmem_cache_t *) arg;
1921 struct array_cache *ac;
1922 int node = numa_node_id();
1925 ac = ac_data(cachep);
1926 spin_lock(&cachep->nodelists[node]->list_lock);
1927 free_block(cachep, ac->entry, ac->avail, node);
1928 spin_unlock(&cachep->nodelists[node]->list_lock);
1932 static void drain_cpu_caches(kmem_cache_t *cachep)
1934 struct kmem_list3 *l3;
1937 smp_call_function_all_cpus(do_drain, cachep);
1939 spin_lock_irq(&cachep->spinlock);
1940 for_each_online_node(node) {
1941 l3 = cachep->nodelists[node];
1943 spin_lock(&l3->list_lock);
1944 drain_array_locked(cachep, l3->shared, 1, node);
1945 spin_unlock(&l3->list_lock);
1947 drain_alien_cache(cachep, l3);
1950 spin_unlock_irq(&cachep->spinlock);
1953 static int __node_shrink(kmem_cache_t *cachep, int node)
1956 struct kmem_list3 *l3 = cachep->nodelists[node];
1960 struct list_head *p;
1962 p = l3->slabs_free.prev;
1963 if (p == &l3->slabs_free)
1966 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
1971 list_del(&slabp->list);
1973 l3->free_objects -= cachep->num;
1974 spin_unlock_irq(&l3->list_lock);
1975 slab_destroy(cachep, slabp);
1976 spin_lock_irq(&l3->list_lock);
1978 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
1982 static int __cache_shrink(kmem_cache_t *cachep)
1985 struct kmem_list3 *l3;
1987 drain_cpu_caches(cachep);
1990 for_each_online_node(i) {
1991 l3 = cachep->nodelists[i];
1993 spin_lock_irq(&l3->list_lock);
1994 ret += __node_shrink(cachep, i);
1995 spin_unlock_irq(&l3->list_lock);
1998 return (ret ? 1 : 0);
2002 * kmem_cache_shrink - Shrink a cache.
2003 * @cachep: The cache to shrink.
2005 * Releases as many slabs as possible for a cache.
2006 * To help debugging, a zero exit status indicates all slabs were released.
2008 int kmem_cache_shrink(kmem_cache_t *cachep)
2010 if (!cachep || in_interrupt())
2013 return __cache_shrink(cachep);
2015 EXPORT_SYMBOL(kmem_cache_shrink);
2018 * kmem_cache_destroy - delete a cache
2019 * @cachep: the cache to destroy
2021 * Remove a kmem_cache_t object from the slab cache.
2022 * Returns 0 on success.
2024 * It is expected this function will be called by a module when it is
2025 * unloaded. This will remove the cache completely, and avoid a duplicate
2026 * cache being allocated each time a module is loaded and unloaded, if the
2027 * module doesn't have persistent in-kernel storage across loads and unloads.
2029 * The cache must be empty before calling this function.
2031 * The caller must guarantee that noone will allocate memory from the cache
2032 * during the kmem_cache_destroy().
2034 int kmem_cache_destroy(kmem_cache_t *cachep)
2037 struct kmem_list3 *l3;
2039 if (!cachep || in_interrupt())
2042 /* Don't let CPUs to come and go */
2045 /* Find the cache in the chain of caches. */
2046 down(&cache_chain_sem);
2048 * the chain is never empty, cache_cache is never destroyed
2050 list_del(&cachep->next);
2051 up(&cache_chain_sem);
2053 if (__cache_shrink(cachep)) {
2054 slab_error(cachep, "Can't free all objects");
2055 down(&cache_chain_sem);
2056 list_add(&cachep->next, &cache_chain);
2057 up(&cache_chain_sem);
2058 unlock_cpu_hotplug();
2062 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2065 for_each_online_cpu(i)
2066 kfree(cachep->array[i]);
2068 /* NUMA: free the list3 structures */
2069 for_each_online_node(i) {
2070 if ((l3 = cachep->nodelists[i])) {
2072 free_alien_cache(l3->alien);
2076 kmem_cache_free(&cache_cache, cachep);
2078 unlock_cpu_hotplug();
2082 EXPORT_SYMBOL(kmem_cache_destroy);
2084 /* Get the memory for a slab management obj. */
2085 static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2086 int colour_off, gfp_t local_flags)
2090 if (OFF_SLAB(cachep)) {
2091 /* Slab management obj is off-slab. */
2092 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2096 slabp = objp + colour_off;
2097 colour_off += cachep->slab_size;
2100 slabp->colouroff = colour_off;
2101 slabp->s_mem = objp + colour_off;
2106 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2108 return (kmem_bufctl_t *) (slabp + 1);
2111 static void cache_init_objs(kmem_cache_t *cachep,
2112 struct slab *slabp, unsigned long ctor_flags)
2116 for (i = 0; i < cachep->num; i++) {
2117 void *objp = slabp->s_mem + cachep->objsize * i;
2119 /* need to poison the objs? */
2120 if (cachep->flags & SLAB_POISON)
2121 poison_obj(cachep, objp, POISON_FREE);
2122 if (cachep->flags & SLAB_STORE_USER)
2123 *dbg_userword(cachep, objp) = NULL;
2125 if (cachep->flags & SLAB_RED_ZONE) {
2126 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2127 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2130 * Constructors are not allowed to allocate memory from
2131 * the same cache which they are a constructor for.
2132 * Otherwise, deadlock. They must also be threaded.
2134 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2135 cachep->ctor(objp + obj_dbghead(cachep), cachep,
2138 if (cachep->flags & SLAB_RED_ZONE) {
2139 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2140 slab_error(cachep, "constructor overwrote the"
2141 " end of an object");
2142 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2143 slab_error(cachep, "constructor overwrote the"
2144 " start of an object");
2146 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2147 && cachep->flags & SLAB_POISON)
2148 kernel_map_pages(virt_to_page(objp),
2149 cachep->objsize / PAGE_SIZE, 0);
2152 cachep->ctor(objp, cachep, ctor_flags);
2154 slab_bufctl(slabp)[i] = i + 1;
2156 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2160 static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
2162 if (flags & SLAB_DMA) {
2163 if (!(cachep->gfpflags & GFP_DMA))
2166 if (cachep->gfpflags & GFP_DMA)
2171 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2176 /* Nasty!!!!!! I hope this is OK. */
2177 i = 1 << cachep->gfporder;
2178 page = virt_to_page(objp);
2180 page_set_cache(page, cachep);
2181 page_set_slab(page, slabp);
2187 * Grow (by 1) the number of slabs within a cache. This is called by
2188 * kmem_cache_alloc() when there are no active objs left in a cache.
2190 static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2196 unsigned long ctor_flags;
2197 struct kmem_list3 *l3;
2199 /* Be lazy and only check for valid flags here,
2200 * keeping it out of the critical path in kmem_cache_alloc().
2202 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
2204 if (flags & SLAB_NO_GROW)
2207 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2208 local_flags = (flags & SLAB_LEVEL_MASK);
2209 if (!(local_flags & __GFP_WAIT))
2211 * Not allowed to sleep. Need to tell a constructor about
2212 * this - it might need to know...
2214 ctor_flags |= SLAB_CTOR_ATOMIC;
2216 /* About to mess with non-constant members - lock. */
2218 spin_lock(&cachep->spinlock);
2220 /* Get colour for the slab, and cal the next value. */
2221 offset = cachep->colour_next;
2222 cachep->colour_next++;
2223 if (cachep->colour_next >= cachep->colour)
2224 cachep->colour_next = 0;
2225 offset *= cachep->colour_off;
2227 spin_unlock(&cachep->spinlock);
2230 if (local_flags & __GFP_WAIT)
2234 * The test for missing atomic flag is performed here, rather than
2235 * the more obvious place, simply to reduce the critical path length
2236 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2237 * will eventually be caught here (where it matters).
2239 kmem_flagcheck(cachep, flags);
2241 /* Get mem for the objs.
2242 * Attempt to allocate a physical page from 'nodeid',
2244 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2247 /* Get slab management. */
2248 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2251 slabp->nodeid = nodeid;
2252 set_slab_attr(cachep, slabp, objp);
2254 cache_init_objs(cachep, slabp, ctor_flags);
2256 if (local_flags & __GFP_WAIT)
2257 local_irq_disable();
2259 l3 = cachep->nodelists[nodeid];
2260 spin_lock(&l3->list_lock);
2262 /* Make slab active. */
2263 list_add_tail(&slabp->list, &(l3->slabs_free));
2264 STATS_INC_GROWN(cachep);
2265 l3->free_objects += cachep->num;
2266 spin_unlock(&l3->list_lock);
2269 kmem_freepages(cachep, objp);
2271 if (local_flags & __GFP_WAIT)
2272 local_irq_disable();
2279 * Perform extra freeing checks:
2280 * - detect bad pointers.
2281 * - POISON/RED_ZONE checking
2282 * - destructor calls, for caches with POISON+dtor
2284 static void kfree_debugcheck(const void *objp)
2288 if (!virt_addr_valid(objp)) {
2289 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2290 (unsigned long)objp);
2293 page = virt_to_page(objp);
2294 if (!PageSlab(page)) {
2295 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2296 (unsigned long)objp);
2301 static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2308 objp -= obj_dbghead(cachep);
2309 kfree_debugcheck(objp);
2310 page = virt_to_page(objp);
2312 if (page_get_cache(page) != cachep) {
2314 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2315 page_get_cache(page), cachep);
2316 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2317 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2318 page_get_cache(page)->name);
2321 slabp = page_get_slab(page);
2323 if (cachep->flags & SLAB_RED_ZONE) {
2324 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2325 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2327 "double free, or memory outside"
2328 " object was overwritten");
2330 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2331 objp, *dbg_redzone1(cachep, objp),
2332 *dbg_redzone2(cachep, objp));
2334 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2335 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2337 if (cachep->flags & SLAB_STORE_USER)
2338 *dbg_userword(cachep, objp) = caller;
2340 objnr = (objp - slabp->s_mem) / cachep->objsize;
2342 BUG_ON(objnr >= cachep->num);
2343 BUG_ON(objp != slabp->s_mem + objnr * cachep->objsize);
2345 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2346 /* Need to call the slab's constructor so the
2347 * caller can perform a verify of its state (debugging).
2348 * Called without the cache-lock held.
2350 cachep->ctor(objp + obj_dbghead(cachep),
2351 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2353 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2354 /* we want to cache poison the object,
2355 * call the destruction callback
2357 cachep->dtor(objp + obj_dbghead(cachep), cachep, 0);
2359 if (cachep->flags & SLAB_POISON) {
2360 #ifdef CONFIG_DEBUG_PAGEALLOC
2361 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2362 store_stackinfo(cachep, objp, (unsigned long)caller);
2363 kernel_map_pages(virt_to_page(objp),
2364 cachep->objsize / PAGE_SIZE, 0);
2366 poison_obj(cachep, objp, POISON_FREE);
2369 poison_obj(cachep, objp, POISON_FREE);
2375 static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2380 /* Check slab's freelist to see if this obj is there. */
2381 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2383 if (entries > cachep->num || i >= cachep->num)
2386 if (entries != cachep->num - slabp->inuse) {
2389 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2390 cachep->name, cachep->num, slabp, slabp->inuse);
2392 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2395 printk("\n%03x:", i);
2396 printk(" %02x", ((unsigned char *)slabp)[i]);
2403 #define kfree_debugcheck(x) do { } while(0)
2404 #define cache_free_debugcheck(x,objp,z) (objp)
2405 #define check_slabp(x,y) do { } while(0)
2408 static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
2411 struct kmem_list3 *l3;
2412 struct array_cache *ac;
2415 ac = ac_data(cachep);
2417 batchcount = ac->batchcount;
2418 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2419 /* if there was little recent activity on this
2420 * cache, then perform only a partial refill.
2421 * Otherwise we could generate refill bouncing.
2423 batchcount = BATCHREFILL_LIMIT;
2425 l3 = cachep->nodelists[numa_node_id()];
2427 BUG_ON(ac->avail > 0 || !l3);
2428 spin_lock(&l3->list_lock);
2431 struct array_cache *shared_array = l3->shared;
2432 if (shared_array->avail) {
2433 if (batchcount > shared_array->avail)
2434 batchcount = shared_array->avail;
2435 shared_array->avail -= batchcount;
2436 ac->avail = batchcount;
2438 &(shared_array->entry[shared_array->avail]),
2439 sizeof(void *) * batchcount);
2440 shared_array->touched = 1;
2444 while (batchcount > 0) {
2445 struct list_head *entry;
2447 /* Get slab alloc is to come from. */
2448 entry = l3->slabs_partial.next;
2449 if (entry == &l3->slabs_partial) {
2450 l3->free_touched = 1;
2451 entry = l3->slabs_free.next;
2452 if (entry == &l3->slabs_free)
2456 slabp = list_entry(entry, struct slab, list);
2457 check_slabp(cachep, slabp);
2458 check_spinlock_acquired(cachep);
2459 while (slabp->inuse < cachep->num && batchcount--) {
2461 STATS_INC_ALLOCED(cachep);
2462 STATS_INC_ACTIVE(cachep);
2463 STATS_SET_HIGH(cachep);
2465 /* get obj pointer */
2466 ac->entry[ac->avail++] = slabp->s_mem +
2467 slabp->free * cachep->objsize;
2470 next = slab_bufctl(slabp)[slabp->free];
2472 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2473 WARN_ON(numa_node_id() != slabp->nodeid);
2477 check_slabp(cachep, slabp);
2479 /* move slabp to correct slabp list: */
2480 list_del(&slabp->list);
2481 if (slabp->free == BUFCTL_END)
2482 list_add(&slabp->list, &l3->slabs_full);
2484 list_add(&slabp->list, &l3->slabs_partial);
2488 l3->free_objects -= ac->avail;
2490 spin_unlock(&l3->list_lock);
2492 if (unlikely(!ac->avail)) {
2494 x = cache_grow(cachep, flags, numa_node_id());
2496 // cache_grow can reenable interrupts, then ac could change.
2497 ac = ac_data(cachep);
2498 if (!x && ac->avail == 0) // no objects in sight? abort
2501 if (!ac->avail) // objects refilled by interrupt?
2505 return ac->entry[--ac->avail];
2509 cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
2511 might_sleep_if(flags & __GFP_WAIT);
2513 kmem_flagcheck(cachep, flags);
2518 static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags,
2519 void *objp, void *caller)
2523 if (cachep->flags & SLAB_POISON) {
2524 #ifdef CONFIG_DEBUG_PAGEALLOC
2525 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2526 kernel_map_pages(virt_to_page(objp),
2527 cachep->objsize / PAGE_SIZE, 1);
2529 check_poison_obj(cachep, objp);
2531 check_poison_obj(cachep, objp);
2533 poison_obj(cachep, objp, POISON_INUSE);
2535 if (cachep->flags & SLAB_STORE_USER)
2536 *dbg_userword(cachep, objp) = caller;
2538 if (cachep->flags & SLAB_RED_ZONE) {
2539 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2540 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2542 "double free, or memory outside"
2543 " object was overwritten");
2545 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2546 objp, *dbg_redzone1(cachep, objp),
2547 *dbg_redzone2(cachep, objp));
2549 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2550 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2552 objp += obj_dbghead(cachep);
2553 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2554 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2556 if (!(flags & __GFP_WAIT))
2557 ctor_flags |= SLAB_CTOR_ATOMIC;
2559 cachep->ctor(objp, cachep, ctor_flags);
2564 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2567 static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2570 struct array_cache *ac;
2573 ac = ac_data(cachep);
2574 if (likely(ac->avail)) {
2575 STATS_INC_ALLOCHIT(cachep);
2577 objp = ac->entry[--ac->avail];
2579 STATS_INC_ALLOCMISS(cachep);
2580 objp = cache_alloc_refill(cachep, flags);
2585 static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2587 unsigned long save_flags;
2590 cache_alloc_debugcheck_before(cachep, flags);
2592 local_irq_save(save_flags);
2593 objp = ____cache_alloc(cachep, flags);
2594 local_irq_restore(save_flags);
2595 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2596 __builtin_return_address(0));
2603 * A interface to enable slab creation on nodeid
2605 static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2607 struct list_head *entry;
2609 struct kmem_list3 *l3;
2614 l3 = cachep->nodelists[nodeid];
2618 spin_lock(&l3->list_lock);
2619 entry = l3->slabs_partial.next;
2620 if (entry == &l3->slabs_partial) {
2621 l3->free_touched = 1;
2622 entry = l3->slabs_free.next;
2623 if (entry == &l3->slabs_free)
2627 slabp = list_entry(entry, struct slab, list);
2628 check_spinlock_acquired_node(cachep, nodeid);
2629 check_slabp(cachep, slabp);
2631 STATS_INC_NODEALLOCS(cachep);
2632 STATS_INC_ACTIVE(cachep);
2633 STATS_SET_HIGH(cachep);
2635 BUG_ON(slabp->inuse == cachep->num);
2637 /* get obj pointer */
2638 obj = slabp->s_mem + slabp->free * cachep->objsize;
2640 next = slab_bufctl(slabp)[slabp->free];
2642 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2645 check_slabp(cachep, slabp);
2647 /* move slabp to correct slabp list: */
2648 list_del(&slabp->list);
2650 if (slabp->free == BUFCTL_END) {
2651 list_add(&slabp->list, &l3->slabs_full);
2653 list_add(&slabp->list, &l3->slabs_partial);
2656 spin_unlock(&l3->list_lock);
2660 spin_unlock(&l3->list_lock);
2661 x = cache_grow(cachep, flags, nodeid);
2673 * Caller needs to acquire correct kmem_list's list_lock
2675 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects,
2679 struct kmem_list3 *l3;
2681 for (i = 0; i < nr_objects; i++) {
2682 void *objp = objpp[i];
2686 slabp = page_get_slab(virt_to_page(objp));
2687 l3 = cachep->nodelists[node];
2688 list_del(&slabp->list);
2689 objnr = (objp - slabp->s_mem) / cachep->objsize;
2690 check_spinlock_acquired_node(cachep, node);
2691 check_slabp(cachep, slabp);
2694 /* Verify that the slab belongs to the intended node */
2695 WARN_ON(slabp->nodeid != node);
2697 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2698 printk(KERN_ERR "slab: double free detected in cache "
2699 "'%s', objp %p\n", cachep->name, objp);
2703 slab_bufctl(slabp)[objnr] = slabp->free;
2704 slabp->free = objnr;
2705 STATS_DEC_ACTIVE(cachep);
2708 check_slabp(cachep, slabp);
2710 /* fixup slab chains */
2711 if (slabp->inuse == 0) {
2712 if (l3->free_objects > l3->free_limit) {
2713 l3->free_objects -= cachep->num;
2714 slab_destroy(cachep, slabp);
2716 list_add(&slabp->list, &l3->slabs_free);
2719 /* Unconditionally move a slab to the end of the
2720 * partial list on free - maximum time for the
2721 * other objects to be freed, too.
2723 list_add_tail(&slabp->list, &l3->slabs_partial);
2728 static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2731 struct kmem_list3 *l3;
2732 int node = numa_node_id();
2734 batchcount = ac->batchcount;
2736 BUG_ON(!batchcount || batchcount > ac->avail);
2739 l3 = cachep->nodelists[node];
2740 spin_lock(&l3->list_lock);
2742 struct array_cache *shared_array = l3->shared;
2743 int max = shared_array->limit - shared_array->avail;
2745 if (batchcount > max)
2747 memcpy(&(shared_array->entry[shared_array->avail]),
2748 ac->entry, sizeof(void *) * batchcount);
2749 shared_array->avail += batchcount;
2754 free_block(cachep, ac->entry, batchcount, node);
2759 struct list_head *p;
2761 p = l3->slabs_free.next;
2762 while (p != &(l3->slabs_free)) {
2765 slabp = list_entry(p, struct slab, list);
2766 BUG_ON(slabp->inuse);
2771 STATS_SET_FREEABLE(cachep, i);
2774 spin_unlock(&l3->list_lock);
2775 ac->avail -= batchcount;
2776 memmove(ac->entry, &(ac->entry[batchcount]),
2777 sizeof(void *) * ac->avail);
2782 * Release an obj back to its cache. If the obj has a constructed
2783 * state, it must be in this state _before_ it is released.
2785 * Called with disabled ints.
2787 static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2789 struct array_cache *ac = ac_data(cachep);
2792 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2794 /* Make sure we are not freeing a object from another
2795 * node to the array cache on this cpu.
2800 slabp = page_get_slab(virt_to_page(objp));
2801 if (unlikely(slabp->nodeid != numa_node_id())) {
2802 struct array_cache *alien = NULL;
2803 int nodeid = slabp->nodeid;
2804 struct kmem_list3 *l3 =
2805 cachep->nodelists[numa_node_id()];
2807 STATS_INC_NODEFREES(cachep);
2808 if (l3->alien && l3->alien[nodeid]) {
2809 alien = l3->alien[nodeid];
2810 spin_lock(&alien->lock);
2811 if (unlikely(alien->avail == alien->limit))
2812 __drain_alien_cache(cachep,
2814 alien->entry[alien->avail++] = objp;
2815 spin_unlock(&alien->lock);
2817 spin_lock(&(cachep->nodelists[nodeid])->
2819 free_block(cachep, &objp, 1, nodeid);
2820 spin_unlock(&(cachep->nodelists[nodeid])->
2827 if (likely(ac->avail < ac->limit)) {
2828 STATS_INC_FREEHIT(cachep);
2829 ac->entry[ac->avail++] = objp;
2832 STATS_INC_FREEMISS(cachep);
2833 cache_flusharray(cachep, ac);
2834 ac->entry[ac->avail++] = objp;
2839 * kmem_cache_alloc - Allocate an object
2840 * @cachep: The cache to allocate from.
2841 * @flags: See kmalloc().
2843 * Allocate an object from this cache. The flags are only relevant
2844 * if the cache has no available objects.
2846 void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2848 return __cache_alloc(cachep, flags);
2850 EXPORT_SYMBOL(kmem_cache_alloc);
2853 * kmem_ptr_validate - check if an untrusted pointer might
2855 * @cachep: the cache we're checking against
2856 * @ptr: pointer to validate
2858 * This verifies that the untrusted pointer looks sane:
2859 * it is _not_ a guarantee that the pointer is actually
2860 * part of the slab cache in question, but it at least
2861 * validates that the pointer can be dereferenced and
2862 * looks half-way sane.
2864 * Currently only used for dentry validation.
2866 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2868 unsigned long addr = (unsigned long)ptr;
2869 unsigned long min_addr = PAGE_OFFSET;
2870 unsigned long align_mask = BYTES_PER_WORD - 1;
2871 unsigned long size = cachep->objsize;
2874 if (unlikely(addr < min_addr))
2876 if (unlikely(addr > (unsigned long)high_memory - size))
2878 if (unlikely(addr & align_mask))
2880 if (unlikely(!kern_addr_valid(addr)))
2882 if (unlikely(!kern_addr_valid(addr + size - 1)))
2884 page = virt_to_page(ptr);
2885 if (unlikely(!PageSlab(page)))
2887 if (unlikely(page_get_cache(page) != cachep))
2896 * kmem_cache_alloc_node - Allocate an object on the specified node
2897 * @cachep: The cache to allocate from.
2898 * @flags: See kmalloc().
2899 * @nodeid: node number of the target node.
2901 * Identical to kmem_cache_alloc, except that this function is slow
2902 * and can sleep. And it will allocate memory on the given node, which
2903 * can improve the performance for cpu bound structures.
2904 * New and improved: it will now make sure that the object gets
2905 * put on the correct node list so that there is no false sharing.
2907 void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2909 unsigned long save_flags;
2913 return __cache_alloc(cachep, flags);
2915 if (unlikely(!cachep->nodelists[nodeid])) {
2916 /* Fall back to __cache_alloc if we run into trouble */
2918 "slab: not allocating in inactive node %d for cache %s\n",
2919 nodeid, cachep->name);
2920 return __cache_alloc(cachep, flags);
2923 cache_alloc_debugcheck_before(cachep, flags);
2924 local_irq_save(save_flags);
2925 if (nodeid == numa_node_id())
2926 ptr = ____cache_alloc(cachep, flags);
2928 ptr = __cache_alloc_node(cachep, flags, nodeid);
2929 local_irq_restore(save_flags);
2931 cache_alloc_debugcheck_after(cachep, flags, ptr,
2932 __builtin_return_address(0));
2936 EXPORT_SYMBOL(kmem_cache_alloc_node);
2938 void *kmalloc_node(size_t size, gfp_t flags, int node)
2940 kmem_cache_t *cachep;
2942 cachep = kmem_find_general_cachep(size, flags);
2943 if (unlikely(cachep == NULL))
2945 return kmem_cache_alloc_node(cachep, flags, node);
2947 EXPORT_SYMBOL(kmalloc_node);
2951 * kmalloc - allocate memory
2952 * @size: how many bytes of memory are required.
2953 * @flags: the type of memory to allocate.
2955 * kmalloc is the normal method of allocating memory
2958 * The @flags argument may be one of:
2960 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2962 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2964 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2966 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2967 * must be suitable for DMA. This can mean different things on different
2968 * platforms. For example, on i386, it means that the memory must come
2969 * from the first 16MB.
2971 void *__kmalloc(size_t size, gfp_t flags)
2973 kmem_cache_t *cachep;
2975 /* If you want to save a few bytes .text space: replace
2977 * Then kmalloc uses the uninlined functions instead of the inline
2980 cachep = __find_general_cachep(size, flags);
2981 if (unlikely(cachep == NULL))
2983 return __cache_alloc(cachep, flags);
2985 EXPORT_SYMBOL(__kmalloc);
2989 * __alloc_percpu - allocate one copy of the object for every present
2990 * cpu in the system, zeroing them.
2991 * Objects should be dereferenced using the per_cpu_ptr macro only.
2993 * @size: how many bytes of memory are required.
2995 void *__alloc_percpu(size_t size)
2998 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3004 * Cannot use for_each_online_cpu since a cpu may come online
3005 * and we have no way of figuring out how to fix the array
3006 * that we have allocated then....
3009 int node = cpu_to_node(i);
3011 if (node_online(node))
3012 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3014 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3016 if (!pdata->ptrs[i])
3018 memset(pdata->ptrs[i], 0, size);
3021 /* Catch derefs w/o wrappers */
3022 return (void *)(~(unsigned long)pdata);
3026 if (!cpu_possible(i))
3028 kfree(pdata->ptrs[i]);
3033 EXPORT_SYMBOL(__alloc_percpu);
3037 * kmem_cache_free - Deallocate an object
3038 * @cachep: The cache the allocation was from.
3039 * @objp: The previously allocated object.
3041 * Free an object which was previously allocated from this
3044 void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3046 unsigned long flags;
3048 local_irq_save(flags);
3049 __cache_free(cachep, objp);
3050 local_irq_restore(flags);
3052 EXPORT_SYMBOL(kmem_cache_free);
3055 * kzalloc - allocate memory. The memory is set to zero.
3056 * @size: how many bytes of memory are required.
3057 * @flags: the type of memory to allocate.
3059 void *kzalloc(size_t size, gfp_t flags)
3061 void *ret = kmalloc(size, flags);
3063 memset(ret, 0, size);
3066 EXPORT_SYMBOL(kzalloc);
3069 * kfree - free previously allocated memory
3070 * @objp: pointer returned by kmalloc.
3072 * If @objp is NULL, no operation is performed.
3074 * Don't free memory not originally allocated by kmalloc()
3075 * or you will run into trouble.
3077 void kfree(const void *objp)
3080 unsigned long flags;
3082 if (unlikely(!objp))
3084 local_irq_save(flags);
3085 kfree_debugcheck(objp);
3086 c = page_get_cache(virt_to_page(objp));
3087 __cache_free(c, (void *)objp);
3088 local_irq_restore(flags);
3090 EXPORT_SYMBOL(kfree);
3094 * free_percpu - free previously allocated percpu memory
3095 * @objp: pointer returned by alloc_percpu.
3097 * Don't free memory not originally allocated by alloc_percpu()
3098 * The complemented objp is to check for that.
3100 void free_percpu(const void *objp)
3103 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3106 * We allocate for all cpus so we cannot use for online cpu here.
3112 EXPORT_SYMBOL(free_percpu);
3115 unsigned int kmem_cache_size(kmem_cache_t *cachep)
3117 return obj_reallen(cachep);
3119 EXPORT_SYMBOL(kmem_cache_size);
3121 const char *kmem_cache_name(kmem_cache_t *cachep)
3123 return cachep->name;
3125 EXPORT_SYMBOL_GPL(kmem_cache_name);
3128 * This initializes kmem_list3 for all nodes.
3130 static int alloc_kmemlist(kmem_cache_t *cachep)
3133 struct kmem_list3 *l3;
3136 for_each_online_node(node) {
3137 struct array_cache *nc = NULL, *new;
3138 struct array_cache **new_alien = NULL;
3140 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3143 if (!(new = alloc_arraycache(node, (cachep->shared *
3144 cachep->batchcount),
3147 if ((l3 = cachep->nodelists[node])) {
3149 spin_lock_irq(&l3->list_lock);
3151 if ((nc = cachep->nodelists[node]->shared))
3152 free_block(cachep, nc->entry, nc->avail, node);
3155 if (!cachep->nodelists[node]->alien) {
3156 l3->alien = new_alien;
3159 l3->free_limit = (1 + nr_cpus_node(node)) *
3160 cachep->batchcount + cachep->num;
3161 spin_unlock_irq(&l3->list_lock);
3163 free_alien_cache(new_alien);
3166 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3170 kmem_list3_init(l3);
3171 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3172 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3174 l3->alien = new_alien;
3175 l3->free_limit = (1 + nr_cpus_node(node)) *
3176 cachep->batchcount + cachep->num;
3177 cachep->nodelists[node] = l3;
3185 struct ccupdate_struct {
3186 kmem_cache_t *cachep;
3187 struct array_cache *new[NR_CPUS];
3190 static void do_ccupdate_local(void *info)
3192 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3193 struct array_cache *old;
3196 old = ac_data(new->cachep);
3198 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3199 new->new[smp_processor_id()] = old;
3202 static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3205 struct ccupdate_struct new;
3208 memset(&new.new, 0, sizeof(new.new));
3209 for_each_online_cpu(i) {
3211 alloc_arraycache(cpu_to_node(i), limit, batchcount);
3213 for (i--; i >= 0; i--)
3218 new.cachep = cachep;
3220 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3223 spin_lock_irq(&cachep->spinlock);
3224 cachep->batchcount = batchcount;
3225 cachep->limit = limit;
3226 cachep->shared = shared;
3227 spin_unlock_irq(&cachep->spinlock);
3229 for_each_online_cpu(i) {
3230 struct array_cache *ccold = new.new[i];
3233 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3234 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3235 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3239 err = alloc_kmemlist(cachep);
3241 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3242 cachep->name, -err);
3248 static void enable_cpucache(kmem_cache_t *cachep)
3253 /* The head array serves three purposes:
3254 * - create a LIFO ordering, i.e. return objects that are cache-warm
3255 * - reduce the number of spinlock operations.
3256 * - reduce the number of linked list operations on the slab and
3257 * bufctl chains: array operations are cheaper.
3258 * The numbers are guessed, we should auto-tune as described by
3261 if (cachep->objsize > 131072)
3263 else if (cachep->objsize > PAGE_SIZE)
3265 else if (cachep->objsize > 1024)
3267 else if (cachep->objsize > 256)
3272 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3273 * allocation behaviour: Most allocs on one cpu, most free operations
3274 * on another cpu. For these cases, an efficient object passing between
3275 * cpus is necessary. This is provided by a shared array. The array
3276 * replaces Bonwick's magazine layer.
3277 * On uniprocessor, it's functionally equivalent (but less efficient)
3278 * to a larger limit. Thus disabled by default.
3282 if (cachep->objsize <= PAGE_SIZE)
3287 /* With debugging enabled, large batchcount lead to excessively
3288 * long periods with disabled local interrupts. Limit the
3294 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3296 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3297 cachep->name, -err);
3300 static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
3301 int force, int node)
3305 check_spinlock_acquired_node(cachep, node);
3306 if (ac->touched && !force) {
3308 } else if (ac->avail) {
3309 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3310 if (tofree > ac->avail) {
3311 tofree = (ac->avail + 1) / 2;
3313 free_block(cachep, ac->entry, tofree, node);
3314 ac->avail -= tofree;
3315 memmove(ac->entry, &(ac->entry[tofree]),
3316 sizeof(void *) * ac->avail);
3321 * cache_reap - Reclaim memory from caches.
3322 * @unused: unused parameter
3324 * Called from workqueue/eventd every few seconds.
3326 * - clear the per-cpu caches for this CPU.
3327 * - return freeable pages to the main free memory pool.
3329 * If we cannot acquire the cache chain semaphore then just give up - we'll
3330 * try again on the next iteration.
3332 static void cache_reap(void *unused)
3334 struct list_head *walk;
3335 struct kmem_list3 *l3;
3337 if (down_trylock(&cache_chain_sem)) {
3338 /* Give up. Setup the next iteration. */
3339 schedule_delayed_work(&__get_cpu_var(reap_work),
3344 list_for_each(walk, &cache_chain) {
3345 kmem_cache_t *searchp;
3346 struct list_head *p;
3350 searchp = list_entry(walk, kmem_cache_t, next);
3352 if (searchp->flags & SLAB_NO_REAP)
3357 l3 = searchp->nodelists[numa_node_id()];
3359 drain_alien_cache(searchp, l3);
3360 spin_lock_irq(&l3->list_lock);
3362 drain_array_locked(searchp, ac_data(searchp), 0,
3365 if (time_after(l3->next_reap, jiffies))
3368 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3371 drain_array_locked(searchp, l3->shared, 0,
3374 if (l3->free_touched) {
3375 l3->free_touched = 0;
3380 (l3->free_limit + 5 * searchp->num -
3381 1) / (5 * searchp->num);
3383 p = l3->slabs_free.next;
3384 if (p == &(l3->slabs_free))
3387 slabp = list_entry(p, struct slab, list);
3388 BUG_ON(slabp->inuse);
3389 list_del(&slabp->list);
3390 STATS_INC_REAPED(searchp);
3392 /* Safe to drop the lock. The slab is no longer
3393 * linked to the cache.
3394 * searchp cannot disappear, we hold
3397 l3->free_objects -= searchp->num;
3398 spin_unlock_irq(&l3->list_lock);
3399 slab_destroy(searchp, slabp);
3400 spin_lock_irq(&l3->list_lock);
3401 } while (--tofree > 0);
3403 spin_unlock_irq(&l3->list_lock);
3408 up(&cache_chain_sem);
3409 drain_remote_pages();
3410 /* Setup the next iteration */
3411 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3414 #ifdef CONFIG_PROC_FS
3416 static void print_slabinfo_header(struct seq_file *m)
3419 * Output format version, so at least we can change it
3420 * without _too_ many complaints.
3423 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3425 seq_puts(m, "slabinfo - version: 2.1\n");
3427 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3428 "<objperslab> <pagesperslab>");
3429 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3430 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3432 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3433 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3434 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3439 static void *s_start(struct seq_file *m, loff_t *pos)
3442 struct list_head *p;
3444 down(&cache_chain_sem);
3446 print_slabinfo_header(m);
3447 p = cache_chain.next;
3450 if (p == &cache_chain)
3453 return list_entry(p, kmem_cache_t, next);
3456 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3458 kmem_cache_t *cachep = p;
3460 return cachep->next.next == &cache_chain ? NULL
3461 : list_entry(cachep->next.next, kmem_cache_t, next);
3464 static void s_stop(struct seq_file *m, void *p)
3466 up(&cache_chain_sem);
3469 static int s_show(struct seq_file *m, void *p)
3471 kmem_cache_t *cachep = p;
3472 struct list_head *q;
3474 unsigned long active_objs;
3475 unsigned long num_objs;
3476 unsigned long active_slabs = 0;
3477 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3481 struct kmem_list3 *l3;
3484 spin_lock_irq(&cachep->spinlock);
3487 for_each_online_node(node) {
3488 l3 = cachep->nodelists[node];
3492 spin_lock(&l3->list_lock);
3494 list_for_each(q, &l3->slabs_full) {
3495 slabp = list_entry(q, struct slab, list);
3496 if (slabp->inuse != cachep->num && !error)
3497 error = "slabs_full accounting error";
3498 active_objs += cachep->num;
3501 list_for_each(q, &l3->slabs_partial) {
3502 slabp = list_entry(q, struct slab, list);
3503 if (slabp->inuse == cachep->num && !error)
3504 error = "slabs_partial inuse accounting error";
3505 if (!slabp->inuse && !error)
3506 error = "slabs_partial/inuse accounting error";
3507 active_objs += slabp->inuse;
3510 list_for_each(q, &l3->slabs_free) {
3511 slabp = list_entry(q, struct slab, list);
3512 if (slabp->inuse && !error)
3513 error = "slabs_free/inuse accounting error";
3516 free_objects += l3->free_objects;
3517 shared_avail += l3->shared->avail;
3519 spin_unlock(&l3->list_lock);
3521 num_slabs += active_slabs;
3522 num_objs = num_slabs * cachep->num;
3523 if (num_objs - active_objs != free_objects && !error)
3524 error = "free_objects accounting error";
3526 name = cachep->name;
3528 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3530 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3531 name, active_objs, num_objs, cachep->objsize,
3532 cachep->num, (1 << cachep->gfporder));
3533 seq_printf(m, " : tunables %4u %4u %4u",
3534 cachep->limit, cachep->batchcount, cachep->shared);
3535 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3536 active_slabs, num_slabs, shared_avail);
3539 unsigned long high = cachep->high_mark;
3540 unsigned long allocs = cachep->num_allocations;
3541 unsigned long grown = cachep->grown;
3542 unsigned long reaped = cachep->reaped;
3543 unsigned long errors = cachep->errors;
3544 unsigned long max_freeable = cachep->max_freeable;
3545 unsigned long node_allocs = cachep->node_allocs;
3546 unsigned long node_frees = cachep->node_frees;
3548 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3549 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
3553 unsigned long allochit = atomic_read(&cachep->allochit);
3554 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3555 unsigned long freehit = atomic_read(&cachep->freehit);
3556 unsigned long freemiss = atomic_read(&cachep->freemiss);
3558 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3559 allochit, allocmiss, freehit, freemiss);
3563 spin_unlock_irq(&cachep->spinlock);
3568 * slabinfo_op - iterator that generates /proc/slabinfo
3577 * num-pages-per-slab
3578 * + further values on SMP and with statistics enabled
3581 struct seq_operations slabinfo_op = {
3588 #define MAX_SLABINFO_WRITE 128
3590 * slabinfo_write - Tuning for the slab allocator
3592 * @buffer: user buffer
3593 * @count: data length
3596 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3597 size_t count, loff_t *ppos)
3599 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3600 int limit, batchcount, shared, res;
3601 struct list_head *p;
3603 if (count > MAX_SLABINFO_WRITE)
3605 if (copy_from_user(&kbuf, buffer, count))
3607 kbuf[MAX_SLABINFO_WRITE] = '\0';
3609 tmp = strchr(kbuf, ' ');
3614 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3617 /* Find the cache in the chain of caches. */
3618 down(&cache_chain_sem);
3620 list_for_each(p, &cache_chain) {
3621 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3623 if (!strcmp(cachep->name, kbuf)) {
3626 batchcount > limit || shared < 0) {
3629 res = do_tune_cpucache(cachep, limit,
3630 batchcount, shared);
3635 up(&cache_chain_sem);
3643 * ksize - get the actual amount of memory allocated for a given object
3644 * @objp: Pointer to the object
3646 * kmalloc may internally round up allocations and return more memory
3647 * than requested. ksize() can be used to determine the actual amount of
3648 * memory allocated. The caller may use this additional memory, even though
3649 * a smaller amount of memory was initially specified with the kmalloc call.
3650 * The caller must guarantee that objp points to a valid object previously
3651 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3652 * must not be freed during the duration of the call.
3654 unsigned int ksize(const void *objp)
3656 if (unlikely(objp == NULL))
3659 return obj_reallen(page_get_cache(virt_to_page(objp)));
3664 * kstrdup - allocate space for and copy an existing string
3666 * @s: the string to duplicate
3667 * @gfp: the GFP mask used in the kmalloc() call when allocating memory
3669 char *kstrdup(const char *s, gfp_t gfp)
3677 len = strlen(s) + 1;
3678 buf = kmalloc(len, gfp);
3680 memcpy(buf, s, len);
3683 EXPORT_SYMBOL(kstrdup);