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
134 /* Shouldn't this be in a header file somewhere? */
135 #define BYTES_PER_WORD sizeof(void *)
137 #ifndef cache_line_size
138 #define cache_line_size() L1_CACHE_BYTES
141 #ifndef ARCH_KMALLOC_MINALIGN
143 * Enforce a minimum alignment for the kmalloc caches.
144 * Usually, the kmalloc caches are cache_line_size() aligned, except when
145 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
146 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
147 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
148 * Note that this flag disables some debug features.
150 #define ARCH_KMALLOC_MINALIGN 0
153 #ifndef ARCH_SLAB_MINALIGN
155 * Enforce a minimum alignment for all caches.
156 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
157 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
158 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
159 * some debug features.
161 #define ARCH_SLAB_MINALIGN 0
164 #ifndef ARCH_KMALLOC_FLAGS
165 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
168 /* Legal flag mask for kmem_cache_create(). */
170 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
172 SLAB_NO_REAP | SLAB_CACHE_DMA | \
173 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
174 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
177 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
178 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
186 * Bufctl's are used for linking objs within a slab
189 * This implementation relies on "struct page" for locating the cache &
190 * slab an object belongs to.
191 * This allows the bufctl structure to be small (one int), but limits
192 * the number of objects a slab (not a cache) can contain when off-slab
193 * bufctls are used. The limit is the size of the largest general cache
194 * that does not use off-slab slabs.
195 * For 32bit archs with 4 kB pages, is this 56.
196 * This is not serious, as it is only for large objects, when it is unwise
197 * to have too many per slab.
198 * Note: This limit can be raised by introducing a general cache whose size
199 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
202 typedef unsigned int kmem_bufctl_t;
203 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
204 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
205 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
207 /* Max number of objs-per-slab for caches which use off-slab slabs.
208 * Needed to avoid a possible looping condition in cache_grow().
210 static unsigned long offslab_limit;
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
225 unsigned short nodeid;
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
242 * We assume struct slab_rcu can overlay struct slab when destroying.
245 struct rcu_head head;
246 kmem_cache_t *cachep;
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
258 * The limit is stored in the per-cpu structure to reduce the data cache
265 unsigned int batchcount;
266 unsigned int touched;
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
272 * [0] is for gcc 2.95. It should really be [].
276 /* bootstrap: The caches do not work without cpuarrays anymore,
277 * but the cpuarrays are allocated from the generic caches...
279 #define BOOT_CPUCACHE_ENTRIES 1
280 struct arraycache_init {
281 struct array_cache cache;
282 void * entries[BOOT_CPUCACHE_ENTRIES];
286 * The slab lists for all objects.
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
293 unsigned long next_reap;
295 unsigned int free_limit;
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
302 * Need this for bootstrapping a per node allocator.
304 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
305 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
306 #define CACHE_CACHE 0
308 #define SIZE_L3 (1 + MAX_NUMNODES)
311 * This function must be completely optimized away if
312 * a constant is passed to it. Mostly the same as
313 * what is in linux/slab.h except it returns an
316 static __always_inline int index_of(const size_t size)
318 if (__builtin_constant_p(size)) {
326 #include "linux/kmalloc_sizes.h"
329 extern void __bad_size(void);
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 objsize;
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;
430 #define CFLGS_OFF_SLAB (0x80000000UL)
431 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
433 #define BATCHREFILL_LIMIT 16
434 /* Optimization question: fewer reaps means less
435 * probability for unnessary cpucache drain/refill cycles.
437 * OTOH the cpuarrays can contain lots of objects,
438 * which could lock up otherwise freeable slabs.
440 #define REAPTIMEOUT_CPUC (2*HZ)
441 #define REAPTIMEOUT_LIST3 (4*HZ)
444 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
445 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
446 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
447 #define STATS_INC_GROWN(x) ((x)->grown++)
448 #define STATS_INC_REAPED(x) ((x)->reaped++)
449 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
450 (x)->high_mark = (x)->num_active; \
452 #define STATS_INC_ERR(x) ((x)->errors++)
453 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
454 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
455 #define STATS_SET_FREEABLE(x, i) \
456 do { if ((x)->max_freeable < i) \
457 (x)->max_freeable = i; \
460 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
461 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
462 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
463 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
465 #define STATS_INC_ACTIVE(x) do { } while (0)
466 #define STATS_DEC_ACTIVE(x) do { } while (0)
467 #define STATS_INC_ALLOCED(x) do { } while (0)
468 #define STATS_INC_GROWN(x) do { } while (0)
469 #define STATS_INC_REAPED(x) do { } while (0)
470 #define STATS_SET_HIGH(x) do { } while (0)
471 #define STATS_INC_ERR(x) do { } while (0)
472 #define STATS_INC_NODEALLOCS(x) do { } while (0)
473 #define STATS_INC_NODEFREES(x) do { } while (0)
474 #define STATS_SET_FREEABLE(x, i) \
477 #define STATS_INC_ALLOCHIT(x) do { } while (0)
478 #define STATS_INC_ALLOCMISS(x) do { } while (0)
479 #define STATS_INC_FREEHIT(x) do { } while (0)
480 #define STATS_INC_FREEMISS(x) do { } while (0)
484 /* Magic nums for obj red zoning.
485 * Placed in the first word before and the first word after an obj.
487 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
488 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
490 /* ...and for poisoning */
491 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
492 #define POISON_FREE 0x6b /* for use-after-free poisoning */
493 #define POISON_END 0xa5 /* end-byte of poisoning */
495 /* memory layout of objects:
497 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
498 * the end of an object is aligned with the end of the real
499 * allocation. Catches writes behind the end of the allocation.
500 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
502 * cachep->dbghead: The real object.
503 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
504 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
506 static int obj_dbghead(kmem_cache_t *cachep)
508 return cachep->dbghead;
511 static int obj_reallen(kmem_cache_t *cachep)
513 return cachep->reallen;
516 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
518 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
519 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
522 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
525 if (cachep->flags & SLAB_STORE_USER)
526 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
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, struct array_cache *ac, int node)
817 struct kmem_list3 *rl3 = cachep->nodelists[node];
820 spin_lock(&rl3->list_lock);
821 free_block(cachep, ac->entry, ac->avail, node);
823 spin_unlock(&rl3->list_lock);
827 static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
830 struct array_cache *ac;
833 for_each_online_node(i) {
836 spin_lock_irqsave(&ac->lock, flags);
837 __drain_alien_cache(cachep, ac, i);
838 spin_unlock_irqrestore(&ac->lock, flags);
843 #define alloc_alien_cache(node, limit) do { } while (0)
844 #define free_alien_cache(ac_ptr) do { } while (0)
845 #define drain_alien_cache(cachep, l3) do { } while (0)
848 static int __devinit cpuup_callback(struct notifier_block *nfb,
849 unsigned long action, void *hcpu)
851 long cpu = (long)hcpu;
852 kmem_cache_t* cachep;
853 struct kmem_list3 *l3 = NULL;
854 int node = cpu_to_node(cpu);
855 int memsize = sizeof(struct kmem_list3);
856 struct array_cache *nc = NULL;
860 down(&cache_chain_sem);
861 /* we need to do this right in the beginning since
862 * alloc_arraycache's are going to use this list.
863 * kmalloc_node allows us to add the slab to the right
864 * kmem_list3 and not this cpu's kmem_list3
867 list_for_each_entry(cachep, &cache_chain, next) {
868 /* setup the size64 kmemlist for cpu before we can
869 * begin anything. Make sure some other cpu on this
870 * node has not already allocated this
872 if (!cachep->nodelists[node]) {
873 if (!(l3 = kmalloc_node(memsize,
877 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
878 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
880 cachep->nodelists[node] = l3;
883 spin_lock_irq(&cachep->nodelists[node]->list_lock);
884 cachep->nodelists[node]->free_limit =
885 (1 + nr_cpus_node(node)) *
886 cachep->batchcount + cachep->num;
887 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
890 /* Now we can go ahead with allocating the shared array's
892 list_for_each_entry(cachep, &cache_chain, next) {
893 nc = alloc_arraycache(node, cachep->limit,
897 cachep->array[cpu] = nc;
899 l3 = cachep->nodelists[node];
902 if (!(nc = alloc_arraycache(node,
903 cachep->shared*cachep->batchcount,
907 /* we are serialised from CPU_DEAD or
908 CPU_UP_CANCELLED by the cpucontrol lock */
912 up(&cache_chain_sem);
915 start_cpu_timer(cpu);
917 #ifdef CONFIG_HOTPLUG_CPU
920 case CPU_UP_CANCELED:
921 down(&cache_chain_sem);
923 list_for_each_entry(cachep, &cache_chain, next) {
924 struct array_cache *nc;
927 mask = node_to_cpumask(node);
928 spin_lock_irq(&cachep->spinlock);
929 /* cpu is dead; no one can alloc from it. */
930 nc = cachep->array[cpu];
931 cachep->array[cpu] = NULL;
932 l3 = cachep->nodelists[node];
937 spin_lock(&l3->list_lock);
939 /* Free limit for this kmem_list3 */
940 l3->free_limit -= cachep->batchcount;
942 free_block(cachep, nc->entry, nc->avail, node);
944 if (!cpus_empty(mask)) {
945 spin_unlock(&l3->list_lock);
950 free_block(cachep, l3->shared->entry,
951 l3->shared->avail, node);
956 drain_alien_cache(cachep, l3);
957 free_alien_cache(l3->alien);
961 /* free slabs belonging to this node */
962 if (__node_shrink(cachep, node)) {
963 cachep->nodelists[node] = NULL;
964 spin_unlock(&l3->list_lock);
967 spin_unlock(&l3->list_lock);
970 spin_unlock_irq(&cachep->spinlock);
973 up(&cache_chain_sem);
979 up(&cache_chain_sem);
983 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
986 * swap the static kmem_list3 with kmalloced memory
988 static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list,
991 struct kmem_list3 *ptr;
993 BUG_ON(cachep->nodelists[nodeid] != list);
994 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
998 memcpy(ptr, list, sizeof(struct kmem_list3));
999 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1000 cachep->nodelists[nodeid] = ptr;
1005 * Called after the gfp() functions have been enabled, and before smp_init().
1007 void __init kmem_cache_init(void)
1010 struct cache_sizes *sizes;
1011 struct cache_names *names;
1014 for (i = 0; i < NUM_INIT_LISTS; i++) {
1015 kmem_list3_init(&initkmem_list3[i]);
1016 if (i < MAX_NUMNODES)
1017 cache_cache.nodelists[i] = NULL;
1021 * Fragmentation resistance on low memory - only use bigger
1022 * page orders on machines with more than 32MB of memory.
1024 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1025 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1027 /* Bootstrap is tricky, because several objects are allocated
1028 * from caches that do not exist yet:
1029 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1030 * structures of all caches, except cache_cache itself: cache_cache
1031 * is statically allocated.
1032 * Initially an __init data area is used for the head array and the
1033 * kmem_list3 structures, it's replaced with a kmalloc allocated
1034 * array at the end of the bootstrap.
1035 * 2) Create the first kmalloc cache.
1036 * The kmem_cache_t for the new cache is allocated normally.
1037 * An __init data area is used for the head array.
1038 * 3) Create the remaining kmalloc caches, with minimally sized
1040 * 4) Replace the __init data head arrays for cache_cache and the first
1041 * kmalloc cache with kmalloc allocated arrays.
1042 * 5) Replace the __init data for kmem_list3 for cache_cache and
1043 * the other cache's with kmalloc allocated memory.
1044 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1047 /* 1) create the cache_cache */
1048 init_MUTEX(&cache_chain_sem);
1049 INIT_LIST_HEAD(&cache_chain);
1050 list_add(&cache_cache.next, &cache_chain);
1051 cache_cache.colour_off = cache_line_size();
1052 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1053 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1055 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1057 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
1058 &left_over, &cache_cache.num);
1059 if (!cache_cache.num)
1062 cache_cache.colour = left_over/cache_cache.colour_off;
1063 cache_cache.colour_next = 0;
1064 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
1065 sizeof(struct slab), cache_line_size());
1067 /* 2+3) create the kmalloc caches */
1068 sizes = malloc_sizes;
1069 names = cache_names;
1071 /* Initialize the caches that provide memory for the array cache
1072 * and the kmem_list3 structures first.
1073 * Without this, further allocations will bug
1076 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1077 sizes[INDEX_AC].cs_size, ARCH_KMALLOC_MINALIGN,
1078 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1080 if (INDEX_AC != INDEX_L3)
1081 sizes[INDEX_L3].cs_cachep =
1082 kmem_cache_create(names[INDEX_L3].name,
1083 sizes[INDEX_L3].cs_size, ARCH_KMALLOC_MINALIGN,
1084 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1086 while (sizes->cs_size != ULONG_MAX) {
1088 * For performance, all the general caches are L1 aligned.
1089 * This should be particularly beneficial on SMP boxes, as it
1090 * eliminates "false sharing".
1091 * Note for systems short on memory removing the alignment will
1092 * allow tighter packing of the smaller caches.
1094 if(!sizes->cs_cachep)
1095 sizes->cs_cachep = kmem_cache_create(names->name,
1096 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1097 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1099 /* Inc off-slab bufctl limit until the ceiling is hit. */
1100 if (!(OFF_SLAB(sizes->cs_cachep))) {
1101 offslab_limit = sizes->cs_size-sizeof(struct slab);
1102 offslab_limit /= sizeof(kmem_bufctl_t);
1105 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1106 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1107 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
1113 /* 4) Replace the bootstrap head arrays */
1117 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1119 local_irq_disable();
1120 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
1121 memcpy(ptr, ac_data(&cache_cache),
1122 sizeof(struct arraycache_init));
1123 cache_cache.array[smp_processor_id()] = ptr;
1126 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1128 local_irq_disable();
1129 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
1130 != &initarray_generic.cache);
1131 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
1132 sizeof(struct arraycache_init));
1133 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1137 /* 5) Replace the bootstrap kmem_list3's */
1140 /* Replace the static kmem_list3 structures for the boot cpu */
1141 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1144 for_each_online_node(node) {
1145 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1146 &initkmem_list3[SIZE_AC+node], node);
1148 if (INDEX_AC != INDEX_L3) {
1149 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1150 &initkmem_list3[SIZE_L3+node],
1156 /* 6) resize the head arrays to their final sizes */
1158 kmem_cache_t *cachep;
1159 down(&cache_chain_sem);
1160 list_for_each_entry(cachep, &cache_chain, next)
1161 enable_cpucache(cachep);
1162 up(&cache_chain_sem);
1166 g_cpucache_up = FULL;
1168 /* Register a cpu startup notifier callback
1169 * that initializes ac_data for all new cpus
1171 register_cpu_notifier(&cpucache_notifier);
1173 /* The reap timers are started later, with a module init call:
1174 * That part of the kernel is not yet operational.
1178 static int __init cpucache_init(void)
1183 * Register the timers that return unneeded
1186 for_each_online_cpu(cpu)
1187 start_cpu_timer(cpu);
1192 __initcall(cpucache_init);
1195 * Interface to system's page allocator. No need to hold the cache-lock.
1197 * If we requested dmaable memory, we will get it. Even if we
1198 * did not request dmaable memory, we might get it, but that
1199 * would be relatively rare and ignorable.
1201 static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
1207 flags |= cachep->gfpflags;
1208 if (likely(nodeid == -1)) {
1209 page = alloc_pages(flags, cachep->gfporder);
1211 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1215 addr = page_address(page);
1217 i = (1 << cachep->gfporder);
1218 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1219 atomic_add(i, &slab_reclaim_pages);
1220 add_page_state(nr_slab, i);
1229 * Interface to system's page release.
1231 static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1233 unsigned long i = (1<<cachep->gfporder);
1234 struct page *page = virt_to_page(addr);
1235 const unsigned long nr_freed = i;
1238 if (!TestClearPageSlab(page))
1242 sub_page_state(nr_slab, nr_freed);
1243 if (current->reclaim_state)
1244 current->reclaim_state->reclaimed_slab += nr_freed;
1245 free_pages((unsigned long)addr, cachep->gfporder);
1246 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1247 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
1250 static void kmem_rcu_free(struct rcu_head *head)
1252 struct slab_rcu *slab_rcu = (struct slab_rcu *) head;
1253 kmem_cache_t *cachep = slab_rcu->cachep;
1255 kmem_freepages(cachep, slab_rcu->addr);
1256 if (OFF_SLAB(cachep))
1257 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1262 #ifdef CONFIG_DEBUG_PAGEALLOC
1263 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1264 unsigned long caller)
1266 int size = obj_reallen(cachep);
1268 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
1270 if (size < 5*sizeof(unsigned long))
1275 *addr++=smp_processor_id();
1276 size -= 3*sizeof(unsigned long);
1278 unsigned long *sptr = &caller;
1279 unsigned long svalue;
1281 while (!kstack_end(sptr)) {
1283 if (kernel_text_address(svalue)) {
1285 size -= sizeof(unsigned long);
1286 if (size <= sizeof(unsigned long))
1296 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1298 int size = obj_reallen(cachep);
1299 addr = &((char*)addr)[obj_dbghead(cachep)];
1301 memset(addr, val, size);
1302 *(unsigned char *)(addr+size-1) = POISON_END;
1305 static void dump_line(char *data, int offset, int limit)
1308 printk(KERN_ERR "%03x:", offset);
1309 for (i=0;i<limit;i++) {
1310 printk(" %02x", (unsigned char)data[offset+i]);
1318 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1323 if (cachep->flags & SLAB_RED_ZONE) {
1324 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1325 *dbg_redzone1(cachep, objp),
1326 *dbg_redzone2(cachep, objp));
1329 if (cachep->flags & SLAB_STORE_USER) {
1330 printk(KERN_ERR "Last user: [<%p>]",
1331 *dbg_userword(cachep, objp));
1332 print_symbol("(%s)",
1333 (unsigned long)*dbg_userword(cachep, objp));
1336 realobj = (char*)objp+obj_dbghead(cachep);
1337 size = obj_reallen(cachep);
1338 for (i=0; i<size && lines;i+=16, lines--) {
1343 dump_line(realobj, i, limit);
1347 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1353 realobj = (char*)objp+obj_dbghead(cachep);
1354 size = obj_reallen(cachep);
1356 for (i=0;i<size;i++) {
1357 char exp = POISON_FREE;
1360 if (realobj[i] != exp) {
1365 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1367 print_objinfo(cachep, objp, 0);
1369 /* Hexdump the affected line */
1374 dump_line(realobj, i, limit);
1377 /* Limit to 5 lines */
1383 /* Print some data about the neighboring objects, if they
1386 struct slab *slabp = page_get_slab(virt_to_page(objp));
1389 objnr = (objp-slabp->s_mem)/cachep->objsize;
1391 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1392 realobj = (char*)objp+obj_dbghead(cachep);
1393 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1395 print_objinfo(cachep, objp, 2);
1397 if (objnr+1 < cachep->num) {
1398 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1399 realobj = (char*)objp+obj_dbghead(cachep);
1400 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1402 print_objinfo(cachep, objp, 2);
1408 /* Destroy all the objs in a slab, and release the mem back to the system.
1409 * Before calling the slab must have been unlinked from the cache.
1410 * The cache-lock is not held/needed.
1412 static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1414 void *addr = slabp->s_mem - slabp->colouroff;
1418 for (i = 0; i < cachep->num; i++) {
1419 void *objp = slabp->s_mem + cachep->objsize * i;
1421 if (cachep->flags & SLAB_POISON) {
1422 #ifdef CONFIG_DEBUG_PAGEALLOC
1423 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1424 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1426 check_poison_obj(cachep, objp);
1428 check_poison_obj(cachep, objp);
1431 if (cachep->flags & SLAB_RED_ZONE) {
1432 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1433 slab_error(cachep, "start of a freed object "
1435 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1436 slab_error(cachep, "end of a freed object "
1439 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1440 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1445 for (i = 0; i < cachep->num; i++) {
1446 void* objp = slabp->s_mem+cachep->objsize*i;
1447 (cachep->dtor)(objp, cachep, 0);
1452 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1453 struct slab_rcu *slab_rcu;
1455 slab_rcu = (struct slab_rcu *) slabp;
1456 slab_rcu->cachep = cachep;
1457 slab_rcu->addr = addr;
1458 call_rcu(&slab_rcu->head, kmem_rcu_free);
1460 kmem_freepages(cachep, addr);
1461 if (OFF_SLAB(cachep))
1462 kmem_cache_free(cachep->slabp_cache, slabp);
1466 /* For setting up all the kmem_list3s for cache whose objsize is same
1467 as size of kmem_list3. */
1468 static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1472 for_each_online_node(node) {
1473 cachep->nodelists[node] = &initkmem_list3[index+node];
1474 cachep->nodelists[node]->next_reap = jiffies +
1476 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1481 * kmem_cache_create - Create a cache.
1482 * @name: A string which is used in /proc/slabinfo to identify this cache.
1483 * @size: The size of objects to be created in this cache.
1484 * @align: The required alignment for the objects.
1485 * @flags: SLAB flags
1486 * @ctor: A constructor for the objects.
1487 * @dtor: A destructor for the objects.
1489 * Returns a ptr to the cache on success, NULL on failure.
1490 * Cannot be called within a int, but can be interrupted.
1491 * The @ctor is run when new pages are allocated by the cache
1492 * and the @dtor is run before the pages are handed back.
1494 * @name must be valid until the cache is destroyed. This implies that
1495 * the module calling this has to destroy the cache before getting
1500 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1501 * to catch references to uninitialised memory.
1503 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1504 * for buffer overruns.
1506 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1509 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1510 * cacheline. This can be beneficial if you're counting cycles as closely
1514 kmem_cache_create (const char *name, size_t size, size_t align,
1515 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1516 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1518 size_t left_over, slab_size, ralign;
1519 kmem_cache_t *cachep = NULL;
1520 struct list_head *p;
1523 * Sanity checks... these are all serious usage bugs.
1527 (size < BYTES_PER_WORD) ||
1528 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1530 printk(KERN_ERR "%s: Early error in slab %s\n",
1531 __FUNCTION__, name);
1535 down(&cache_chain_sem);
1537 list_for_each(p, &cache_chain) {
1538 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1539 mm_segment_t old_fs = get_fs();
1544 * This happens when the module gets unloaded and doesn't
1545 * destroy its slab cache and no-one else reuses the vmalloc
1546 * area of the module. Print a warning.
1549 res = __get_user(tmp, pc->name);
1552 printk("SLAB: cache with size %d has lost its name\n",
1557 if (!strcmp(pc->name,name)) {
1558 printk("kmem_cache_create: duplicate cache %s\n", name);
1565 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1566 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1567 /* No constructor, but inital state check requested */
1568 printk(KERN_ERR "%s: No con, but init state check "
1569 "requested - %s\n", __FUNCTION__, name);
1570 flags &= ~SLAB_DEBUG_INITIAL;
1575 * Enable redzoning and last user accounting, except for caches with
1576 * large objects, if the increased size would increase the object size
1577 * above the next power of two: caches with object sizes just above a
1578 * power of two have a significant amount of internal fragmentation.
1580 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1581 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1582 if (!(flags & SLAB_DESTROY_BY_RCU))
1583 flags |= SLAB_POISON;
1585 if (flags & SLAB_DESTROY_BY_RCU)
1586 BUG_ON(flags & SLAB_POISON);
1588 if (flags & SLAB_DESTROY_BY_RCU)
1592 * Always checks flags, a caller might be expecting debug
1593 * support which isn't available.
1595 if (flags & ~CREATE_MASK)
1598 /* Check that size is in terms of words. This is needed to avoid
1599 * unaligned accesses for some archs when redzoning is used, and makes
1600 * sure any on-slab bufctl's are also correctly aligned.
1602 if (size & (BYTES_PER_WORD-1)) {
1603 size += (BYTES_PER_WORD-1);
1604 size &= ~(BYTES_PER_WORD-1);
1607 /* calculate out the final buffer alignment: */
1608 /* 1) arch recommendation: can be overridden for debug */
1609 if (flags & SLAB_HWCACHE_ALIGN) {
1610 /* Default alignment: as specified by the arch code.
1611 * Except if an object is really small, then squeeze multiple
1612 * objects into one cacheline.
1614 ralign = cache_line_size();
1615 while (size <= ralign/2)
1618 ralign = BYTES_PER_WORD;
1620 /* 2) arch mandated alignment: disables debug if necessary */
1621 if (ralign < ARCH_SLAB_MINALIGN) {
1622 ralign = ARCH_SLAB_MINALIGN;
1623 if (ralign > BYTES_PER_WORD)
1624 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1626 /* 3) caller mandated alignment: disables debug if necessary */
1627 if (ralign < align) {
1629 if (ralign > BYTES_PER_WORD)
1630 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1632 /* 4) Store it. Note that the debug code below can reduce
1633 * the alignment to BYTES_PER_WORD.
1637 /* Get cache's description obj. */
1638 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1641 memset(cachep, 0, sizeof(kmem_cache_t));
1644 cachep->reallen = size;
1646 if (flags & SLAB_RED_ZONE) {
1647 /* redzoning only works with word aligned caches */
1648 align = BYTES_PER_WORD;
1650 /* add space for red zone words */
1651 cachep->dbghead += BYTES_PER_WORD;
1652 size += 2*BYTES_PER_WORD;
1654 if (flags & SLAB_STORE_USER) {
1655 /* user store requires word alignment and
1656 * one word storage behind the end of the real
1659 align = BYTES_PER_WORD;
1660 size += BYTES_PER_WORD;
1662 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1663 if (size >= malloc_sizes[INDEX_L3+1].cs_size && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1664 cachep->dbghead += PAGE_SIZE - size;
1670 /* Determine if the slab management is 'on' or 'off' slab. */
1671 if (size >= (PAGE_SIZE>>3))
1673 * Size is large, assume best to place the slab management obj
1674 * off-slab (should allow better packing of objs).
1676 flags |= CFLGS_OFF_SLAB;
1678 size = ALIGN(size, align);
1680 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1682 * A VFS-reclaimable slab tends to have most allocations
1683 * as GFP_NOFS and we really don't want to have to be allocating
1684 * higher-order pages when we are unable to shrink dcache.
1686 cachep->gfporder = 0;
1687 cache_estimate(cachep->gfporder, size, align, flags,
1688 &left_over, &cachep->num);
1691 * Calculate size (in pages) of slabs, and the num of objs per
1692 * slab. This could be made much more intelligent. For now,
1693 * try to avoid using high page-orders for slabs. When the
1694 * gfp() funcs are more friendly towards high-order requests,
1695 * this should be changed.
1698 unsigned int break_flag = 0;
1700 cache_estimate(cachep->gfporder, size, align, flags,
1701 &left_over, &cachep->num);
1704 if (cachep->gfporder >= MAX_GFP_ORDER)
1708 if (flags & CFLGS_OFF_SLAB &&
1709 cachep->num > offslab_limit) {
1710 /* This num of objs will cause problems. */
1717 * Large num of objs is good, but v. large slabs are
1718 * currently bad for the gfp()s.
1720 if (cachep->gfporder >= slab_break_gfp_order)
1723 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1724 break; /* Acceptable internal fragmentation. */
1731 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1732 kmem_cache_free(&cache_cache, cachep);
1736 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1737 + sizeof(struct slab), align);
1740 * If the slab has been placed off-slab, and we have enough space then
1741 * move it on-slab. This is at the expense of any extra colouring.
1743 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1744 flags &= ~CFLGS_OFF_SLAB;
1745 left_over -= slab_size;
1748 if (flags & CFLGS_OFF_SLAB) {
1749 /* really off slab. No need for manual alignment */
1750 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1753 cachep->colour_off = cache_line_size();
1754 /* Offset must be a multiple of the alignment. */
1755 if (cachep->colour_off < align)
1756 cachep->colour_off = align;
1757 cachep->colour = left_over/cachep->colour_off;
1758 cachep->slab_size = slab_size;
1759 cachep->flags = flags;
1760 cachep->gfpflags = 0;
1761 if (flags & SLAB_CACHE_DMA)
1762 cachep->gfpflags |= GFP_DMA;
1763 spin_lock_init(&cachep->spinlock);
1764 cachep->objsize = size;
1766 if (flags & CFLGS_OFF_SLAB)
1767 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
1768 cachep->ctor = ctor;
1769 cachep->dtor = dtor;
1770 cachep->name = name;
1772 /* Don't let CPUs to come and go */
1775 if (g_cpucache_up == FULL) {
1776 enable_cpucache(cachep);
1778 if (g_cpucache_up == NONE) {
1779 /* Note: the first kmem_cache_create must create
1780 * the cache that's used by kmalloc(24), otherwise
1781 * the creation of further caches will BUG().
1783 cachep->array[smp_processor_id()] =
1784 &initarray_generic.cache;
1786 /* If the cache that's used by
1787 * kmalloc(sizeof(kmem_list3)) is the first cache,
1788 * then we need to set up all its list3s, otherwise
1789 * the creation of further caches will BUG().
1791 set_up_list3s(cachep, SIZE_AC);
1792 if (INDEX_AC == INDEX_L3)
1793 g_cpucache_up = PARTIAL_L3;
1795 g_cpucache_up = PARTIAL_AC;
1797 cachep->array[smp_processor_id()] =
1798 kmalloc(sizeof(struct arraycache_init),
1801 if (g_cpucache_up == PARTIAL_AC) {
1802 set_up_list3s(cachep, SIZE_L3);
1803 g_cpucache_up = PARTIAL_L3;
1806 for_each_online_node(node) {
1808 cachep->nodelists[node] =
1809 kmalloc_node(sizeof(struct kmem_list3),
1811 BUG_ON(!cachep->nodelists[node]);
1812 kmem_list3_init(cachep->nodelists[node]);
1816 cachep->nodelists[numa_node_id()]->next_reap =
1817 jiffies + REAPTIMEOUT_LIST3 +
1818 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1820 BUG_ON(!ac_data(cachep));
1821 ac_data(cachep)->avail = 0;
1822 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1823 ac_data(cachep)->batchcount = 1;
1824 ac_data(cachep)->touched = 0;
1825 cachep->batchcount = 1;
1826 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1829 /* cache setup completed, link it into the list */
1830 list_add(&cachep->next, &cache_chain);
1831 unlock_cpu_hotplug();
1833 if (!cachep && (flags & SLAB_PANIC))
1834 panic("kmem_cache_create(): failed to create slab `%s'\n",
1836 up(&cache_chain_sem);
1839 EXPORT_SYMBOL(kmem_cache_create);
1842 static void check_irq_off(void)
1844 BUG_ON(!irqs_disabled());
1847 static void check_irq_on(void)
1849 BUG_ON(irqs_disabled());
1852 static void check_spinlock_acquired(kmem_cache_t *cachep)
1856 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
1860 static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1864 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1869 #define check_irq_off() do { } while(0)
1870 #define check_irq_on() do { } while(0)
1871 #define check_spinlock_acquired(x) do { } while(0)
1872 #define check_spinlock_acquired_node(x, y) do { } while(0)
1876 * Waits for all CPUs to execute func().
1878 static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1883 local_irq_disable();
1887 if (smp_call_function(func, arg, 1, 1))
1893 static void drain_array_locked(kmem_cache_t* cachep,
1894 struct array_cache *ac, int force, int node);
1896 static void do_drain(void *arg)
1898 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1899 struct array_cache *ac;
1900 int node = numa_node_id();
1903 ac = ac_data(cachep);
1904 spin_lock(&cachep->nodelists[node]->list_lock);
1905 free_block(cachep, ac->entry, ac->avail, node);
1906 spin_unlock(&cachep->nodelists[node]->list_lock);
1910 static void drain_cpu_caches(kmem_cache_t *cachep)
1912 struct kmem_list3 *l3;
1915 smp_call_function_all_cpus(do_drain, cachep);
1917 spin_lock_irq(&cachep->spinlock);
1918 for_each_online_node(node) {
1919 l3 = cachep->nodelists[node];
1921 spin_lock(&l3->list_lock);
1922 drain_array_locked(cachep, l3->shared, 1, node);
1923 spin_unlock(&l3->list_lock);
1925 drain_alien_cache(cachep, l3);
1928 spin_unlock_irq(&cachep->spinlock);
1931 static int __node_shrink(kmem_cache_t *cachep, int node)
1934 struct kmem_list3 *l3 = cachep->nodelists[node];
1938 struct list_head *p;
1940 p = l3->slabs_free.prev;
1941 if (p == &l3->slabs_free)
1944 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
1949 list_del(&slabp->list);
1951 l3->free_objects -= cachep->num;
1952 spin_unlock_irq(&l3->list_lock);
1953 slab_destroy(cachep, slabp);
1954 spin_lock_irq(&l3->list_lock);
1956 ret = !list_empty(&l3->slabs_full) ||
1957 !list_empty(&l3->slabs_partial);
1961 static int __cache_shrink(kmem_cache_t *cachep)
1964 struct kmem_list3 *l3;
1966 drain_cpu_caches(cachep);
1969 for_each_online_node(i) {
1970 l3 = cachep->nodelists[i];
1972 spin_lock_irq(&l3->list_lock);
1973 ret += __node_shrink(cachep, i);
1974 spin_unlock_irq(&l3->list_lock);
1977 return (ret ? 1 : 0);
1981 * kmem_cache_shrink - Shrink a cache.
1982 * @cachep: The cache to shrink.
1984 * Releases as many slabs as possible for a cache.
1985 * To help debugging, a zero exit status indicates all slabs were released.
1987 int kmem_cache_shrink(kmem_cache_t *cachep)
1989 if (!cachep || in_interrupt())
1992 return __cache_shrink(cachep);
1994 EXPORT_SYMBOL(kmem_cache_shrink);
1997 * kmem_cache_destroy - delete a cache
1998 * @cachep: the cache to destroy
2000 * Remove a kmem_cache_t object from the slab cache.
2001 * Returns 0 on success.
2003 * It is expected this function will be called by a module when it is
2004 * unloaded. This will remove the cache completely, and avoid a duplicate
2005 * cache being allocated each time a module is loaded and unloaded, if the
2006 * module doesn't have persistent in-kernel storage across loads and unloads.
2008 * The cache must be empty before calling this function.
2010 * The caller must guarantee that noone will allocate memory from the cache
2011 * during the kmem_cache_destroy().
2013 int kmem_cache_destroy(kmem_cache_t * cachep)
2016 struct kmem_list3 *l3;
2018 if (!cachep || in_interrupt())
2021 /* Don't let CPUs to come and go */
2024 /* Find the cache in the chain of caches. */
2025 down(&cache_chain_sem);
2027 * the chain is never empty, cache_cache is never destroyed
2029 list_del(&cachep->next);
2030 up(&cache_chain_sem);
2032 if (__cache_shrink(cachep)) {
2033 slab_error(cachep, "Can't free all objects");
2034 down(&cache_chain_sem);
2035 list_add(&cachep->next,&cache_chain);
2036 up(&cache_chain_sem);
2037 unlock_cpu_hotplug();
2041 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2044 for_each_online_cpu(i)
2045 kfree(cachep->array[i]);
2047 /* NUMA: free the list3 structures */
2048 for_each_online_node(i) {
2049 if ((l3 = cachep->nodelists[i])) {
2051 free_alien_cache(l3->alien);
2055 kmem_cache_free(&cache_cache, cachep);
2057 unlock_cpu_hotplug();
2061 EXPORT_SYMBOL(kmem_cache_destroy);
2063 /* Get the memory for a slab management obj. */
2064 static struct slab* alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2065 int colour_off, gfp_t local_flags)
2069 if (OFF_SLAB(cachep)) {
2070 /* Slab management obj is off-slab. */
2071 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2075 slabp = objp+colour_off;
2076 colour_off += cachep->slab_size;
2079 slabp->colouroff = colour_off;
2080 slabp->s_mem = objp+colour_off;
2085 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2087 return (kmem_bufctl_t *)(slabp+1);
2090 static void cache_init_objs(kmem_cache_t *cachep,
2091 struct slab *slabp, unsigned long ctor_flags)
2095 for (i = 0; i < cachep->num; i++) {
2096 void *objp = slabp->s_mem+cachep->objsize*i;
2098 /* need to poison the objs? */
2099 if (cachep->flags & SLAB_POISON)
2100 poison_obj(cachep, objp, POISON_FREE);
2101 if (cachep->flags & SLAB_STORE_USER)
2102 *dbg_userword(cachep, objp) = NULL;
2104 if (cachep->flags & SLAB_RED_ZONE) {
2105 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2106 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2109 * Constructors are not allowed to allocate memory from
2110 * the same cache which they are a constructor for.
2111 * Otherwise, deadlock. They must also be threaded.
2113 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2114 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
2116 if (cachep->flags & SLAB_RED_ZONE) {
2117 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2118 slab_error(cachep, "constructor overwrote the"
2119 " end of an object");
2120 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2121 slab_error(cachep, "constructor overwrote the"
2122 " start of an object");
2124 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2125 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2128 cachep->ctor(objp, cachep, ctor_flags);
2130 slab_bufctl(slabp)[i] = i+1;
2132 slab_bufctl(slabp)[i-1] = BUFCTL_END;
2136 static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
2138 if (flags & SLAB_DMA) {
2139 if (!(cachep->gfpflags & GFP_DMA))
2142 if (cachep->gfpflags & GFP_DMA)
2147 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2152 /* Nasty!!!!!! I hope this is OK. */
2153 i = 1 << cachep->gfporder;
2154 page = virt_to_page(objp);
2156 page_set_cache(page, cachep);
2157 page_set_slab(page, slabp);
2163 * Grow (by 1) the number of slabs within a cache. This is called by
2164 * kmem_cache_alloc() when there are no active objs left in a cache.
2166 static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2172 unsigned long ctor_flags;
2173 struct kmem_list3 *l3;
2175 /* Be lazy and only check for valid flags here,
2176 * keeping it out of the critical path in kmem_cache_alloc().
2178 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
2180 if (flags & SLAB_NO_GROW)
2183 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2184 local_flags = (flags & SLAB_LEVEL_MASK);
2185 if (!(local_flags & __GFP_WAIT))
2187 * Not allowed to sleep. Need to tell a constructor about
2188 * this - it might need to know...
2190 ctor_flags |= SLAB_CTOR_ATOMIC;
2192 /* About to mess with non-constant members - lock. */
2194 spin_lock(&cachep->spinlock);
2196 /* Get colour for the slab, and cal the next value. */
2197 offset = cachep->colour_next;
2198 cachep->colour_next++;
2199 if (cachep->colour_next >= cachep->colour)
2200 cachep->colour_next = 0;
2201 offset *= cachep->colour_off;
2203 spin_unlock(&cachep->spinlock);
2206 if (local_flags & __GFP_WAIT)
2210 * The test for missing atomic flag is performed here, rather than
2211 * the more obvious place, simply to reduce the critical path length
2212 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2213 * will eventually be caught here (where it matters).
2215 kmem_flagcheck(cachep, flags);
2217 /* Get mem for the objs.
2218 * Attempt to allocate a physical page from 'nodeid',
2220 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2223 /* Get slab management. */
2224 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2227 slabp->nodeid = nodeid;
2228 set_slab_attr(cachep, slabp, objp);
2230 cache_init_objs(cachep, slabp, ctor_flags);
2232 if (local_flags & __GFP_WAIT)
2233 local_irq_disable();
2235 l3 = cachep->nodelists[nodeid];
2236 spin_lock(&l3->list_lock);
2238 /* Make slab active. */
2239 list_add_tail(&slabp->list, &(l3->slabs_free));
2240 STATS_INC_GROWN(cachep);
2241 l3->free_objects += cachep->num;
2242 spin_unlock(&l3->list_lock);
2245 kmem_freepages(cachep, objp);
2247 if (local_flags & __GFP_WAIT)
2248 local_irq_disable();
2255 * Perform extra freeing checks:
2256 * - detect bad pointers.
2257 * - POISON/RED_ZONE checking
2258 * - destructor calls, for caches with POISON+dtor
2260 static void kfree_debugcheck(const void *objp)
2264 if (!virt_addr_valid(objp)) {
2265 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2266 (unsigned long)objp);
2269 page = virt_to_page(objp);
2270 if (!PageSlab(page)) {
2271 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
2276 static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2283 objp -= obj_dbghead(cachep);
2284 kfree_debugcheck(objp);
2285 page = virt_to_page(objp);
2287 if (page_get_cache(page) != cachep) {
2288 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2289 page_get_cache(page),cachep);
2290 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2291 printk(KERN_ERR "%p is %s.\n", page_get_cache(page), page_get_cache(page)->name);
2294 slabp = page_get_slab(page);
2296 if (cachep->flags & SLAB_RED_ZONE) {
2297 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2298 slab_error(cachep, "double free, or memory outside"
2299 " object was overwritten");
2300 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2301 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2303 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2304 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2306 if (cachep->flags & SLAB_STORE_USER)
2307 *dbg_userword(cachep, objp) = caller;
2309 objnr = (objp-slabp->s_mem)/cachep->objsize;
2311 BUG_ON(objnr >= cachep->num);
2312 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
2314 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2315 /* Need to call the slab's constructor so the
2316 * caller can perform a verify of its state (debugging).
2317 * Called without the cache-lock held.
2319 cachep->ctor(objp+obj_dbghead(cachep),
2320 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
2322 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2323 /* we want to cache poison the object,
2324 * call the destruction callback
2326 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
2328 if (cachep->flags & SLAB_POISON) {
2329 #ifdef CONFIG_DEBUG_PAGEALLOC
2330 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2331 store_stackinfo(cachep, objp, (unsigned long)caller);
2332 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2334 poison_obj(cachep, objp, POISON_FREE);
2337 poison_obj(cachep, objp, POISON_FREE);
2343 static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2348 /* Check slab's freelist to see if this obj is there. */
2349 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2351 if (entries > cachep->num || i >= cachep->num)
2354 if (entries != cachep->num - slabp->inuse) {
2356 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2357 cachep->name, cachep->num, slabp, slabp->inuse);
2358 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
2360 printk("\n%03x:", i);
2361 printk(" %02x", ((unsigned char*)slabp)[i]);
2368 #define kfree_debugcheck(x) do { } while(0)
2369 #define cache_free_debugcheck(x,objp,z) (objp)
2370 #define check_slabp(x,y) do { } while(0)
2373 static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
2376 struct kmem_list3 *l3;
2377 struct array_cache *ac;
2380 ac = ac_data(cachep);
2382 batchcount = ac->batchcount;
2383 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2384 /* if there was little recent activity on this
2385 * cache, then perform only a partial refill.
2386 * Otherwise we could generate refill bouncing.
2388 batchcount = BATCHREFILL_LIMIT;
2390 l3 = cachep->nodelists[numa_node_id()];
2392 BUG_ON(ac->avail > 0 || !l3);
2393 spin_lock(&l3->list_lock);
2396 struct array_cache *shared_array = l3->shared;
2397 if (shared_array->avail) {
2398 if (batchcount > shared_array->avail)
2399 batchcount = shared_array->avail;
2400 shared_array->avail -= batchcount;
2401 ac->avail = batchcount;
2403 &(shared_array->entry[shared_array->avail]),
2404 sizeof(void*)*batchcount);
2405 shared_array->touched = 1;
2409 while (batchcount > 0) {
2410 struct list_head *entry;
2412 /* Get slab alloc is to come from. */
2413 entry = l3->slabs_partial.next;
2414 if (entry == &l3->slabs_partial) {
2415 l3->free_touched = 1;
2416 entry = l3->slabs_free.next;
2417 if (entry == &l3->slabs_free)
2421 slabp = list_entry(entry, struct slab, list);
2422 check_slabp(cachep, slabp);
2423 check_spinlock_acquired(cachep);
2424 while (slabp->inuse < cachep->num && batchcount--) {
2426 STATS_INC_ALLOCED(cachep);
2427 STATS_INC_ACTIVE(cachep);
2428 STATS_SET_HIGH(cachep);
2430 /* get obj pointer */
2431 ac->entry[ac->avail++] = slabp->s_mem +
2432 slabp->free*cachep->objsize;
2435 next = slab_bufctl(slabp)[slabp->free];
2437 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2438 WARN_ON(numa_node_id() != slabp->nodeid);
2442 check_slabp(cachep, slabp);
2444 /* move slabp to correct slabp list: */
2445 list_del(&slabp->list);
2446 if (slabp->free == BUFCTL_END)
2447 list_add(&slabp->list, &l3->slabs_full);
2449 list_add(&slabp->list, &l3->slabs_partial);
2453 l3->free_objects -= ac->avail;
2455 spin_unlock(&l3->list_lock);
2457 if (unlikely(!ac->avail)) {
2459 x = cache_grow(cachep, flags, numa_node_id());
2461 // cache_grow can reenable interrupts, then ac could change.
2462 ac = ac_data(cachep);
2463 if (!x && ac->avail == 0) // no objects in sight? abort
2466 if (!ac->avail) // objects refilled by interrupt?
2470 return ac->entry[--ac->avail];
2474 cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
2476 might_sleep_if(flags & __GFP_WAIT);
2478 kmem_flagcheck(cachep, flags);
2484 cache_alloc_debugcheck_after(kmem_cache_t *cachep,
2485 gfp_t flags, void *objp, void *caller)
2489 if (cachep->flags & SLAB_POISON) {
2490 #ifdef CONFIG_DEBUG_PAGEALLOC
2491 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2492 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2494 check_poison_obj(cachep, objp);
2496 check_poison_obj(cachep, objp);
2498 poison_obj(cachep, objp, POISON_INUSE);
2500 if (cachep->flags & SLAB_STORE_USER)
2501 *dbg_userword(cachep, objp) = caller;
2503 if (cachep->flags & SLAB_RED_ZONE) {
2504 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2505 slab_error(cachep, "double free, or memory outside"
2506 " object was overwritten");
2507 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2508 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2510 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2511 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2513 objp += obj_dbghead(cachep);
2514 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2515 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2517 if (!(flags & __GFP_WAIT))
2518 ctor_flags |= SLAB_CTOR_ATOMIC;
2520 cachep->ctor(objp, cachep, ctor_flags);
2525 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2528 static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2531 struct array_cache *ac;
2534 ac = ac_data(cachep);
2535 if (likely(ac->avail)) {
2536 STATS_INC_ALLOCHIT(cachep);
2538 objp = ac->entry[--ac->avail];
2540 STATS_INC_ALLOCMISS(cachep);
2541 objp = cache_alloc_refill(cachep, flags);
2546 static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2548 unsigned long save_flags;
2551 cache_alloc_debugcheck_before(cachep, flags);
2553 local_irq_save(save_flags);
2554 objp = ____cache_alloc(cachep, flags);
2555 local_irq_restore(save_flags);
2556 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2557 __builtin_return_address(0));
2564 * A interface to enable slab creation on nodeid
2566 static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2568 struct list_head *entry;
2570 struct kmem_list3 *l3;
2575 l3 = cachep->nodelists[nodeid];
2579 spin_lock(&l3->list_lock);
2580 entry = l3->slabs_partial.next;
2581 if (entry == &l3->slabs_partial) {
2582 l3->free_touched = 1;
2583 entry = l3->slabs_free.next;
2584 if (entry == &l3->slabs_free)
2588 slabp = list_entry(entry, struct slab, list);
2589 check_spinlock_acquired_node(cachep, nodeid);
2590 check_slabp(cachep, slabp);
2592 STATS_INC_NODEALLOCS(cachep);
2593 STATS_INC_ACTIVE(cachep);
2594 STATS_SET_HIGH(cachep);
2596 BUG_ON(slabp->inuse == cachep->num);
2598 /* get obj pointer */
2599 obj = slabp->s_mem + slabp->free*cachep->objsize;
2601 next = slab_bufctl(slabp)[slabp->free];
2603 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2606 check_slabp(cachep, slabp);
2608 /* move slabp to correct slabp list: */
2609 list_del(&slabp->list);
2611 if (slabp->free == BUFCTL_END) {
2612 list_add(&slabp->list, &l3->slabs_full);
2614 list_add(&slabp->list, &l3->slabs_partial);
2617 spin_unlock(&l3->list_lock);
2621 spin_unlock(&l3->list_lock);
2622 x = cache_grow(cachep, flags, nodeid);
2634 * Caller needs to acquire correct kmem_list's list_lock
2636 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects, int node)
2639 struct kmem_list3 *l3;
2641 for (i = 0; i < nr_objects; i++) {
2642 void *objp = objpp[i];
2646 slabp = page_get_slab(virt_to_page(objp));
2647 l3 = cachep->nodelists[node];
2648 list_del(&slabp->list);
2649 objnr = (objp - slabp->s_mem) / cachep->objsize;
2650 check_spinlock_acquired_node(cachep, node);
2651 check_slabp(cachep, slabp);
2654 /* Verify that the slab belongs to the intended node */
2655 WARN_ON(slabp->nodeid != node);
2657 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2658 printk(KERN_ERR "slab: double free detected in cache "
2659 "'%s', objp %p\n", cachep->name, objp);
2663 slab_bufctl(slabp)[objnr] = slabp->free;
2664 slabp->free = objnr;
2665 STATS_DEC_ACTIVE(cachep);
2668 check_slabp(cachep, slabp);
2670 /* fixup slab chains */
2671 if (slabp->inuse == 0) {
2672 if (l3->free_objects > l3->free_limit) {
2673 l3->free_objects -= cachep->num;
2674 slab_destroy(cachep, slabp);
2676 list_add(&slabp->list, &l3->slabs_free);
2679 /* Unconditionally move a slab to the end of the
2680 * partial list on free - maximum time for the
2681 * other objects to be freed, too.
2683 list_add_tail(&slabp->list, &l3->slabs_partial);
2688 static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2691 struct kmem_list3 *l3;
2692 int node = numa_node_id();
2694 batchcount = ac->batchcount;
2696 BUG_ON(!batchcount || batchcount > ac->avail);
2699 l3 = cachep->nodelists[node];
2700 spin_lock(&l3->list_lock);
2702 struct array_cache *shared_array = l3->shared;
2703 int max = shared_array->limit-shared_array->avail;
2705 if (batchcount > max)
2707 memcpy(&(shared_array->entry[shared_array->avail]),
2709 sizeof(void*)*batchcount);
2710 shared_array->avail += batchcount;
2715 free_block(cachep, ac->entry, batchcount, node);
2720 struct list_head *p;
2722 p = l3->slabs_free.next;
2723 while (p != &(l3->slabs_free)) {
2726 slabp = list_entry(p, struct slab, list);
2727 BUG_ON(slabp->inuse);
2732 STATS_SET_FREEABLE(cachep, i);
2735 spin_unlock(&l3->list_lock);
2736 ac->avail -= batchcount;
2737 memmove(ac->entry, &(ac->entry[batchcount]),
2738 sizeof(void*)*ac->avail);
2744 * Release an obj back to its cache. If the obj has a constructed
2745 * state, it must be in this state _before_ it is released.
2747 * Called with disabled ints.
2749 static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2751 struct array_cache *ac = ac_data(cachep);
2754 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2756 /* Make sure we are not freeing a object from another
2757 * node to the array cache on this cpu.
2762 slabp = page_get_slab(virt_to_page(objp));
2763 if (unlikely(slabp->nodeid != numa_node_id())) {
2764 struct array_cache *alien = NULL;
2765 int nodeid = slabp->nodeid;
2766 struct kmem_list3 *l3 = cachep->nodelists[numa_node_id()];
2768 STATS_INC_NODEFREES(cachep);
2769 if (l3->alien && l3->alien[nodeid]) {
2770 alien = l3->alien[nodeid];
2771 spin_lock(&alien->lock);
2772 if (unlikely(alien->avail == alien->limit))
2773 __drain_alien_cache(cachep,
2775 alien->entry[alien->avail++] = objp;
2776 spin_unlock(&alien->lock);
2778 spin_lock(&(cachep->nodelists[nodeid])->
2780 free_block(cachep, &objp, 1, nodeid);
2781 spin_unlock(&(cachep->nodelists[nodeid])->
2788 if (likely(ac->avail < ac->limit)) {
2789 STATS_INC_FREEHIT(cachep);
2790 ac->entry[ac->avail++] = objp;
2793 STATS_INC_FREEMISS(cachep);
2794 cache_flusharray(cachep, ac);
2795 ac->entry[ac->avail++] = objp;
2800 * kmem_cache_alloc - Allocate an object
2801 * @cachep: The cache to allocate from.
2802 * @flags: See kmalloc().
2804 * Allocate an object from this cache. The flags are only relevant
2805 * if the cache has no available objects.
2807 void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
2809 return __cache_alloc(cachep, flags);
2811 EXPORT_SYMBOL(kmem_cache_alloc);
2814 * kmem_ptr_validate - check if an untrusted pointer might
2816 * @cachep: the cache we're checking against
2817 * @ptr: pointer to validate
2819 * This verifies that the untrusted pointer looks sane:
2820 * it is _not_ a guarantee that the pointer is actually
2821 * part of the slab cache in question, but it at least
2822 * validates that the pointer can be dereferenced and
2823 * looks half-way sane.
2825 * Currently only used for dentry validation.
2827 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2829 unsigned long addr = (unsigned long) ptr;
2830 unsigned long min_addr = PAGE_OFFSET;
2831 unsigned long align_mask = BYTES_PER_WORD-1;
2832 unsigned long size = cachep->objsize;
2835 if (unlikely(addr < min_addr))
2837 if (unlikely(addr > (unsigned long)high_memory - size))
2839 if (unlikely(addr & align_mask))
2841 if (unlikely(!kern_addr_valid(addr)))
2843 if (unlikely(!kern_addr_valid(addr + size - 1)))
2845 page = virt_to_page(ptr);
2846 if (unlikely(!PageSlab(page)))
2848 if (unlikely(page_get_cache(page) != cachep))
2857 * kmem_cache_alloc_node - Allocate an object on the specified node
2858 * @cachep: The cache to allocate from.
2859 * @flags: See kmalloc().
2860 * @nodeid: node number of the target node.
2862 * Identical to kmem_cache_alloc, except that this function is slow
2863 * and can sleep. And it will allocate memory on the given node, which
2864 * can improve the performance for cpu bound structures.
2865 * New and improved: it will now make sure that the object gets
2866 * put on the correct node list so that there is no false sharing.
2868 void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
2870 unsigned long save_flags;
2874 return __cache_alloc(cachep, flags);
2876 if (unlikely(!cachep->nodelists[nodeid])) {
2877 /* Fall back to __cache_alloc if we run into trouble */
2878 printk(KERN_WARNING "slab: not allocating in inactive node %d for cache %s\n", nodeid, cachep->name);
2879 return __cache_alloc(cachep,flags);
2882 cache_alloc_debugcheck_before(cachep, flags);
2883 local_irq_save(save_flags);
2884 if (nodeid == numa_node_id())
2885 ptr = ____cache_alloc(cachep, flags);
2887 ptr = __cache_alloc_node(cachep, flags, nodeid);
2888 local_irq_restore(save_flags);
2889 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, __builtin_return_address(0));
2893 EXPORT_SYMBOL(kmem_cache_alloc_node);
2895 void *kmalloc_node(size_t size, gfp_t flags, int node)
2897 kmem_cache_t *cachep;
2899 cachep = kmem_find_general_cachep(size, flags);
2900 if (unlikely(cachep == NULL))
2902 return kmem_cache_alloc_node(cachep, flags, node);
2904 EXPORT_SYMBOL(kmalloc_node);
2908 * kmalloc - allocate memory
2909 * @size: how many bytes of memory are required.
2910 * @flags: the type of memory to allocate.
2912 * kmalloc is the normal method of allocating memory
2915 * The @flags argument may be one of:
2917 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2919 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2921 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2923 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2924 * must be suitable for DMA. This can mean different things on different
2925 * platforms. For example, on i386, it means that the memory must come
2926 * from the first 16MB.
2928 void *__kmalloc(size_t size, gfp_t flags)
2930 kmem_cache_t *cachep;
2932 /* If you want to save a few bytes .text space: replace
2934 * Then kmalloc uses the uninlined functions instead of the inline
2937 cachep = __find_general_cachep(size, flags);
2938 if (unlikely(cachep == NULL))
2940 return __cache_alloc(cachep, flags);
2942 EXPORT_SYMBOL(__kmalloc);
2946 * __alloc_percpu - allocate one copy of the object for every present
2947 * cpu in the system, zeroing them.
2948 * Objects should be dereferenced using the per_cpu_ptr macro only.
2950 * @size: how many bytes of memory are required.
2951 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2953 void *__alloc_percpu(size_t size, size_t align)
2956 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2962 * Cannot use for_each_online_cpu since a cpu may come online
2963 * and we have no way of figuring out how to fix the array
2964 * that we have allocated then....
2967 int node = cpu_to_node(i);
2969 if (node_online(node))
2970 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
2972 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
2974 if (!pdata->ptrs[i])
2976 memset(pdata->ptrs[i], 0, size);
2979 /* Catch derefs w/o wrappers */
2980 return (void *) (~(unsigned long) pdata);
2984 if (!cpu_possible(i))
2986 kfree(pdata->ptrs[i]);
2991 EXPORT_SYMBOL(__alloc_percpu);
2995 * kmem_cache_free - Deallocate an object
2996 * @cachep: The cache the allocation was from.
2997 * @objp: The previously allocated object.
2999 * Free an object which was previously allocated from this
3002 void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3004 unsigned long flags;
3006 local_irq_save(flags);
3007 __cache_free(cachep, objp);
3008 local_irq_restore(flags);
3010 EXPORT_SYMBOL(kmem_cache_free);
3013 * kzalloc - allocate memory. The memory is set to zero.
3014 * @size: how many bytes of memory are required.
3015 * @flags: the type of memory to allocate.
3017 void *kzalloc(size_t size, gfp_t flags)
3019 void *ret = kmalloc(size, flags);
3021 memset(ret, 0, size);
3024 EXPORT_SYMBOL(kzalloc);
3027 * kfree - free previously allocated memory
3028 * @objp: pointer returned by kmalloc.
3030 * If @objp is NULL, no operation is performed.
3032 * Don't free memory not originally allocated by kmalloc()
3033 * or you will run into trouble.
3035 void kfree(const void *objp)
3038 unsigned long flags;
3040 if (unlikely(!objp))
3042 local_irq_save(flags);
3043 kfree_debugcheck(objp);
3044 c = page_get_cache(virt_to_page(objp));
3045 __cache_free(c, (void*)objp);
3046 local_irq_restore(flags);
3048 EXPORT_SYMBOL(kfree);
3052 * free_percpu - free previously allocated percpu memory
3053 * @objp: pointer returned by alloc_percpu.
3055 * Don't free memory not originally allocated by alloc_percpu()
3056 * The complemented objp is to check for that.
3059 free_percpu(const void *objp)
3062 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
3065 * We allocate for all cpus so we cannot use for online cpu here.
3071 EXPORT_SYMBOL(free_percpu);
3074 unsigned int kmem_cache_size(kmem_cache_t *cachep)
3076 return obj_reallen(cachep);
3078 EXPORT_SYMBOL(kmem_cache_size);
3080 const char *kmem_cache_name(kmem_cache_t *cachep)
3082 return cachep->name;
3084 EXPORT_SYMBOL_GPL(kmem_cache_name);
3087 * This initializes kmem_list3 for all nodes.
3089 static int alloc_kmemlist(kmem_cache_t *cachep)
3092 struct kmem_list3 *l3;
3095 for_each_online_node(node) {
3096 struct array_cache *nc = NULL, *new;
3097 struct array_cache **new_alien = NULL;
3099 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3102 if (!(new = alloc_arraycache(node, (cachep->shared*
3103 cachep->batchcount), 0xbaadf00d)))
3105 if ((l3 = cachep->nodelists[node])) {
3107 spin_lock_irq(&l3->list_lock);
3109 if ((nc = cachep->nodelists[node]->shared))
3110 free_block(cachep, nc->entry,
3114 if (!cachep->nodelists[node]->alien) {
3115 l3->alien = new_alien;
3118 l3->free_limit = (1 + nr_cpus_node(node))*
3119 cachep->batchcount + cachep->num;
3120 spin_unlock_irq(&l3->list_lock);
3122 free_alien_cache(new_alien);
3125 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3129 kmem_list3_init(l3);
3130 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3131 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
3133 l3->alien = new_alien;
3134 l3->free_limit = (1 + nr_cpus_node(node))*
3135 cachep->batchcount + cachep->num;
3136 cachep->nodelists[node] = l3;
3144 struct ccupdate_struct {
3145 kmem_cache_t *cachep;
3146 struct array_cache *new[NR_CPUS];
3149 static void do_ccupdate_local(void *info)
3151 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3152 struct array_cache *old;
3155 old = ac_data(new->cachep);
3157 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3158 new->new[smp_processor_id()] = old;
3162 static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3165 struct ccupdate_struct new;
3168 memset(&new.new,0,sizeof(new.new));
3169 for_each_online_cpu(i) {
3170 new.new[i] = alloc_arraycache(cpu_to_node(i), limit, batchcount);
3172 for (i--; i >= 0; i--) kfree(new.new[i]);
3176 new.cachep = cachep;
3178 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
3181 spin_lock_irq(&cachep->spinlock);
3182 cachep->batchcount = batchcount;
3183 cachep->limit = limit;
3184 cachep->shared = shared;
3185 spin_unlock_irq(&cachep->spinlock);
3187 for_each_online_cpu(i) {
3188 struct array_cache *ccold = new.new[i];
3191 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3192 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3193 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3197 err = alloc_kmemlist(cachep);
3199 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3200 cachep->name, -err);
3207 static void enable_cpucache(kmem_cache_t *cachep)
3212 /* The head array serves three purposes:
3213 * - create a LIFO ordering, i.e. return objects that are cache-warm
3214 * - reduce the number of spinlock operations.
3215 * - reduce the number of linked list operations on the slab and
3216 * bufctl chains: array operations are cheaper.
3217 * The numbers are guessed, we should auto-tune as described by
3220 if (cachep->objsize > 131072)
3222 else if (cachep->objsize > PAGE_SIZE)
3224 else if (cachep->objsize > 1024)
3226 else if (cachep->objsize > 256)
3231 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3232 * allocation behaviour: Most allocs on one cpu, most free operations
3233 * on another cpu. For these cases, an efficient object passing between
3234 * cpus is necessary. This is provided by a shared array. The array
3235 * replaces Bonwick's magazine layer.
3236 * On uniprocessor, it's functionally equivalent (but less efficient)
3237 * to a larger limit. Thus disabled by default.
3241 if (cachep->objsize <= PAGE_SIZE)
3246 /* With debugging enabled, large batchcount lead to excessively
3247 * long periods with disabled local interrupts. Limit the
3253 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
3255 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3256 cachep->name, -err);
3259 static void drain_array_locked(kmem_cache_t *cachep,
3260 struct array_cache *ac, int force, int node)
3264 check_spinlock_acquired_node(cachep, node);
3265 if (ac->touched && !force) {
3267 } else if (ac->avail) {
3268 tofree = force ? ac->avail : (ac->limit+4)/5;
3269 if (tofree > ac->avail) {
3270 tofree = (ac->avail+1)/2;
3272 free_block(cachep, ac->entry, tofree, node);
3273 ac->avail -= tofree;
3274 memmove(ac->entry, &(ac->entry[tofree]),
3275 sizeof(void*)*ac->avail);
3280 * cache_reap - Reclaim memory from caches.
3281 * @unused: unused parameter
3283 * Called from workqueue/eventd every few seconds.
3285 * - clear the per-cpu caches for this CPU.
3286 * - return freeable pages to the main free memory pool.
3288 * If we cannot acquire the cache chain semaphore then just give up - we'll
3289 * try again on the next iteration.
3291 static void cache_reap(void *unused)
3293 struct list_head *walk;
3294 struct kmem_list3 *l3;
3296 if (down_trylock(&cache_chain_sem)) {
3297 /* Give up. Setup the next iteration. */
3298 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3302 list_for_each(walk, &cache_chain) {
3303 kmem_cache_t *searchp;
3304 struct list_head* p;
3308 searchp = list_entry(walk, kmem_cache_t, next);
3310 if (searchp->flags & SLAB_NO_REAP)
3315 l3 = searchp->nodelists[numa_node_id()];
3317 drain_alien_cache(searchp, l3);
3318 spin_lock_irq(&l3->list_lock);
3320 drain_array_locked(searchp, ac_data(searchp), 0,
3323 if (time_after(l3->next_reap, jiffies))
3326 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3329 drain_array_locked(searchp, l3->shared, 0,
3332 if (l3->free_touched) {
3333 l3->free_touched = 0;
3337 tofree = (l3->free_limit+5*searchp->num-1)/(5*searchp->num);
3339 p = l3->slabs_free.next;
3340 if (p == &(l3->slabs_free))
3343 slabp = list_entry(p, struct slab, list);
3344 BUG_ON(slabp->inuse);
3345 list_del(&slabp->list);
3346 STATS_INC_REAPED(searchp);
3348 /* Safe to drop the lock. The slab is no longer
3349 * linked to the cache.
3350 * searchp cannot disappear, we hold
3353 l3->free_objects -= searchp->num;
3354 spin_unlock_irq(&l3->list_lock);
3355 slab_destroy(searchp, slabp);
3356 spin_lock_irq(&l3->list_lock);
3357 } while(--tofree > 0);
3359 spin_unlock_irq(&l3->list_lock);
3364 up(&cache_chain_sem);
3365 drain_remote_pages();
3366 /* Setup the next iteration */
3367 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3370 #ifdef CONFIG_PROC_FS
3372 static void *s_start(struct seq_file *m, loff_t *pos)
3375 struct list_head *p;
3377 down(&cache_chain_sem);
3380 * Output format version, so at least we can change it
3381 * without _too_ many complaints.
3384 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3386 seq_puts(m, "slabinfo - version: 2.1\n");
3388 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
3389 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3390 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3392 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>"
3393 " <error> <maxfreeable> <nodeallocs> <remotefrees>");
3394 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3398 p = cache_chain.next;
3401 if (p == &cache_chain)
3404 return list_entry(p, kmem_cache_t, next);
3407 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3409 kmem_cache_t *cachep = p;
3411 return cachep->next.next == &cache_chain ? NULL
3412 : list_entry(cachep->next.next, kmem_cache_t, next);
3415 static void s_stop(struct seq_file *m, void *p)
3417 up(&cache_chain_sem);
3420 static int s_show(struct seq_file *m, void *p)
3422 kmem_cache_t *cachep = p;
3423 struct list_head *q;
3425 unsigned long active_objs;
3426 unsigned long num_objs;
3427 unsigned long active_slabs = 0;
3428 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3432 struct kmem_list3 *l3;
3435 spin_lock_irq(&cachep->spinlock);
3438 for_each_online_node(node) {
3439 l3 = cachep->nodelists[node];
3443 spin_lock(&l3->list_lock);
3445 list_for_each(q,&l3->slabs_full) {
3446 slabp = list_entry(q, struct slab, list);
3447 if (slabp->inuse != cachep->num && !error)
3448 error = "slabs_full accounting error";
3449 active_objs += cachep->num;
3452 list_for_each(q,&l3->slabs_partial) {
3453 slabp = list_entry(q, struct slab, list);
3454 if (slabp->inuse == cachep->num && !error)
3455 error = "slabs_partial inuse accounting error";
3456 if (!slabp->inuse && !error)
3457 error = "slabs_partial/inuse accounting error";
3458 active_objs += slabp->inuse;
3461 list_for_each(q,&l3->slabs_free) {
3462 slabp = list_entry(q, struct slab, list);
3463 if (slabp->inuse && !error)
3464 error = "slabs_free/inuse accounting error";
3467 free_objects += l3->free_objects;
3468 shared_avail += l3->shared->avail;
3470 spin_unlock(&l3->list_lock);
3472 num_slabs+=active_slabs;
3473 num_objs = num_slabs*cachep->num;
3474 if (num_objs - active_objs != free_objects && !error)
3475 error = "free_objects accounting error";
3477 name = cachep->name;
3479 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3481 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3482 name, active_objs, num_objs, cachep->objsize,
3483 cachep->num, (1<<cachep->gfporder));
3484 seq_printf(m, " : tunables %4u %4u %4u",
3485 cachep->limit, cachep->batchcount,
3487 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3488 active_slabs, num_slabs, shared_avail);
3491 unsigned long high = cachep->high_mark;
3492 unsigned long allocs = cachep->num_allocations;
3493 unsigned long grown = cachep->grown;
3494 unsigned long reaped = cachep->reaped;
3495 unsigned long errors = cachep->errors;
3496 unsigned long max_freeable = cachep->max_freeable;
3497 unsigned long node_allocs = cachep->node_allocs;
3498 unsigned long node_frees = cachep->node_frees;
3500 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3501 %4lu %4lu %4lu %4lu",
3502 allocs, high, grown, reaped, errors,
3503 max_freeable, node_allocs, node_frees);
3507 unsigned long allochit = atomic_read(&cachep->allochit);
3508 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3509 unsigned long freehit = atomic_read(&cachep->freehit);
3510 unsigned long freemiss = atomic_read(&cachep->freemiss);
3512 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3513 allochit, allocmiss, freehit, freemiss);
3517 spin_unlock_irq(&cachep->spinlock);
3522 * slabinfo_op - iterator that generates /proc/slabinfo
3531 * num-pages-per-slab
3532 * + further values on SMP and with statistics enabled
3535 struct seq_operations slabinfo_op = {
3542 #define MAX_SLABINFO_WRITE 128
3544 * slabinfo_write - Tuning for the slab allocator
3546 * @buffer: user buffer
3547 * @count: data length
3550 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3551 size_t count, loff_t *ppos)
3553 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
3554 int limit, batchcount, shared, res;
3555 struct list_head *p;
3557 if (count > MAX_SLABINFO_WRITE)
3559 if (copy_from_user(&kbuf, buffer, count))
3561 kbuf[MAX_SLABINFO_WRITE] = '\0';
3563 tmp = strchr(kbuf, ' ');
3568 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3571 /* Find the cache in the chain of caches. */
3572 down(&cache_chain_sem);
3574 list_for_each(p,&cache_chain) {
3575 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3577 if (!strcmp(cachep->name, kbuf)) {
3580 batchcount > limit ||
3584 res = do_tune_cpucache(cachep, limit,
3585 batchcount, shared);
3590 up(&cache_chain_sem);
3598 * ksize - get the actual amount of memory allocated for a given object
3599 * @objp: Pointer to the object
3601 * kmalloc may internally round up allocations and return more memory
3602 * than requested. ksize() can be used to determine the actual amount of
3603 * memory allocated. The caller may use this additional memory, even though
3604 * a smaller amount of memory was initially specified with the kmalloc call.
3605 * The caller must guarantee that objp points to a valid object previously
3606 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3607 * must not be freed during the duration of the call.
3609 unsigned int ksize(const void *objp)
3611 if (unlikely(objp == NULL))
3614 return obj_reallen(page_get_cache(virt_to_page(objp)));
3619 * kstrdup - allocate space for and copy an existing string
3621 * @s: the string to duplicate
3622 * @gfp: the GFP mask used in the kmalloc() call when allocating memory
3624 char *kstrdup(const char *s, gfp_t gfp)
3632 len = strlen(s) + 1;
3633 buf = kmalloc(len, gfp);
3635 memcpy(buf, s, len);
3638 EXPORT_SYMBOL(kstrdup);