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 initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <trace/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
117 #include <asm/cacheflush.h>
118 #include <asm/tlbflush.h>
119 #include <asm/page.h>
122 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * STATS - 1 to collect stats for /proc/slabinfo.
126 * 0 for faster, smaller code (especially in the critical paths).
128 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 #ifdef CONFIG_DEBUG_SLAB
134 #define FORCED_DEBUG 1
138 #define FORCED_DEBUG 0
141 /* Shouldn't this be in a header file somewhere? */
142 #define BYTES_PER_WORD sizeof(void *)
143 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
183 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
193 * Bufctl's are used for linking objs within a slab
196 * This implementation relies on "struct page" for locating the cache &
197 * slab an object belongs to.
198 * This allows the bufctl structure to be small (one int), but limits
199 * the number of objects a slab (not a cache) can contain when off-slab
200 * bufctls are used. The limit is the size of the largest general cache
201 * that does not use off-slab slabs.
202 * For 32bit archs with 4 kB pages, is this 56.
203 * This is not serious, as it is only for large objects, when it is unwise
204 * to have too many per slab.
205 * Note: This limit can be raised by introducing a general cache whose size
206 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
209 typedef unsigned int kmem_bufctl_t;
210 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
211 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
212 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
213 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
228 unsigned short nodeid;
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
245 * We assume struct slab_rcu can overlay struct slab when destroying.
248 struct rcu_head head;
249 struct kmem_cache *cachep;
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
261 * The limit is stored in the per-cpu structure to reduce the data cache
268 unsigned int batchcount;
269 unsigned int touched;
272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
289 * The slab lists for all objects.
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
309 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
311 #define SIZE_AC MAX_NUMNODES
312 #define SIZE_L3 (2 * MAX_NUMNODES)
314 static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
318 static int enable_cpucache(struct kmem_cache *cachep);
319 static void cache_reap(struct work_struct *unused);
322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
325 static __always_inline int index_of(const size_t size)
327 extern void __bad_size(void);
329 if (__builtin_constant_p(size)) {
337 #include <linux/kmalloc_sizes.h>
345 static int slab_early_init = 1;
347 #define INDEX_AC index_of(sizeof(struct arraycache_init))
348 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
350 static void kmem_list3_init(struct kmem_list3 *parent)
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
357 parent->colour_next = 0;
358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
363 #define MAKE_LIST(cachep, listp, slab, nodeid) \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
369 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
383 /* 1) per-cpu data, touched during every alloc/free */
384 struct array_cache *array[NR_CPUS];
385 /* 2) Cache tunables. Protected by cache_chain_mutex */
386 unsigned int batchcount;
390 unsigned int buffer_size;
391 u32 reciprocal_buffer_size;
392 /* 3) touched by every alloc & free from the backend */
394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
397 /* 4) cache_grow/shrink */
398 /* order of pgs per slab (2^n) */
399 unsigned int gfporder;
401 /* force GFP flags, e.g. GFP_DMA */
404 size_t colour; /* cache colouring range */
405 unsigned int colour_off; /* colour offset */
406 struct kmem_cache *slabp_cache;
407 unsigned int slab_size;
408 unsigned int dflags; /* dynamic flags */
410 /* constructor func */
411 void (*ctor)(void *obj);
413 /* 5) cache creation/removal */
415 struct list_head next;
419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
428 unsigned long node_overflow;
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
445 * We put nodelists[] at the end of kmem_cache, because we want to size
446 * this array to nr_node_ids slots instead of MAX_NUMNODES
447 * (see kmem_cache_init())
448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
449 * is statically defined, so we reserve the max number of nodes.
451 struct kmem_list3 *nodelists[MAX_NUMNODES];
453 * Do not add fields after nodelists[]
457 #define CFLGS_OFF_SLAB (0x80000000UL)
458 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
460 #define BATCHREFILL_LIMIT 16
462 * Optimization question: fewer reaps means less probability for unnessary
463 * cpucache drain/refill cycles.
465 * OTOH the cpuarrays can contain lots of objects,
466 * which could lock up otherwise freeable slabs.
468 #define REAPTIMEOUT_CPUC (2*HZ)
469 #define REAPTIMEOUT_LIST3 (4*HZ)
472 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
473 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
474 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
475 #define STATS_INC_GROWN(x) ((x)->grown++)
476 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
477 #define STATS_SET_HIGH(x) \
479 if ((x)->num_active > (x)->high_mark) \
480 (x)->high_mark = (x)->num_active; \
482 #define STATS_INC_ERR(x) ((x)->errors++)
483 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
484 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
485 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
486 #define STATS_SET_FREEABLE(x, i) \
488 if ((x)->max_freeable < i) \
489 (x)->max_freeable = i; \
491 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
492 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
493 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
494 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
496 #define STATS_INC_ACTIVE(x) do { } while (0)
497 #define STATS_DEC_ACTIVE(x) do { } while (0)
498 #define STATS_INC_ALLOCED(x) do { } while (0)
499 #define STATS_INC_GROWN(x) do { } while (0)
500 #define STATS_ADD_REAPED(x,y) do { } while (0)
501 #define STATS_SET_HIGH(x) do { } while (0)
502 #define STATS_INC_ERR(x) do { } while (0)
503 #define STATS_INC_NODEALLOCS(x) do { } while (0)
504 #define STATS_INC_NODEFREES(x) do { } while (0)
505 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
506 #define STATS_SET_FREEABLE(x, i) do { } while (0)
507 #define STATS_INC_ALLOCHIT(x) do { } while (0)
508 #define STATS_INC_ALLOCMISS(x) do { } while (0)
509 #define STATS_INC_FREEHIT(x) do { } while (0)
510 #define STATS_INC_FREEMISS(x) do { } while (0)
516 * memory layout of objects:
518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
519 * the end of an object is aligned with the end of the real
520 * allocation. Catches writes behind the end of the allocation.
521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
523 * cachep->obj_offset: The real object.
524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
526 * [BYTES_PER_WORD long]
528 static int obj_offset(struct kmem_cache *cachep)
530 return cachep->obj_offset;
533 static int obj_size(struct kmem_cache *cachep)
535 return cachep->obj_size;
538 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
541 return (unsigned long long*) (objp + obj_offset(cachep) -
542 sizeof(unsigned long long));
545 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 if (cachep->flags & SLAB_STORE_USER)
549 return (unsigned long long *)(objp + cachep->buffer_size -
550 sizeof(unsigned long long) -
552 return (unsigned long long *) (objp + cachep->buffer_size -
553 sizeof(unsigned long long));
556 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
558 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
564 #define obj_offset(x) 0
565 #define obj_size(cachep) (cachep->buffer_size)
566 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
568 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
572 #ifdef CONFIG_KMEMTRACE
573 size_t slab_buffer_size(struct kmem_cache *cachep)
575 return cachep->buffer_size;
577 EXPORT_SYMBOL(slab_buffer_size);
581 * Do not go above this order unless 0 objects fit into the slab.
583 #define BREAK_GFP_ORDER_HI 1
584 #define BREAK_GFP_ORDER_LO 0
585 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
588 * Functions for storing/retrieving the cachep and or slab from the page
589 * allocator. These are used to find the slab an obj belongs to. With kfree(),
590 * these are used to find the cache which an obj belongs to.
592 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
594 page->lru.next = (struct list_head *)cache;
597 static inline struct kmem_cache *page_get_cache(struct page *page)
599 page = compound_head(page);
600 BUG_ON(!PageSlab(page));
601 return (struct kmem_cache *)page->lru.next;
604 static inline void page_set_slab(struct page *page, struct slab *slab)
606 page->lru.prev = (struct list_head *)slab;
609 static inline struct slab *page_get_slab(struct page *page)
611 BUG_ON(!PageSlab(page));
612 return (struct slab *)page->lru.prev;
615 static inline struct kmem_cache *virt_to_cache(const void *obj)
617 struct page *page = virt_to_head_page(obj);
618 return page_get_cache(page);
621 static inline struct slab *virt_to_slab(const void *obj)
623 struct page *page = virt_to_head_page(obj);
624 return page_get_slab(page);
627 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
630 return slab->s_mem + cache->buffer_size * idx;
634 * We want to avoid an expensive divide : (offset / cache->buffer_size)
635 * Using the fact that buffer_size is a constant for a particular cache,
636 * we can replace (offset / cache->buffer_size) by
637 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
639 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
640 const struct slab *slab, void *obj)
642 u32 offset = (obj - slab->s_mem);
643 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
647 * These are the default caches for kmalloc. Custom caches can have other sizes.
649 struct cache_sizes malloc_sizes[] = {
650 #define CACHE(x) { .cs_size = (x) },
651 #include <linux/kmalloc_sizes.h>
655 EXPORT_SYMBOL(malloc_sizes);
657 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
663 static struct cache_names __initdata cache_names[] = {
664 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
665 #include <linux/kmalloc_sizes.h>
670 static struct arraycache_init initarray_cache __initdata =
671 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
672 static struct arraycache_init initarray_generic =
673 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
675 /* internal cache of cache description objs */
676 static struct kmem_cache cache_cache = {
678 .limit = BOOT_CPUCACHE_ENTRIES,
680 .buffer_size = sizeof(struct kmem_cache),
681 .name = "kmem_cache",
684 #define BAD_ALIEN_MAGIC 0x01020304ul
686 #ifdef CONFIG_LOCKDEP
689 * Slab sometimes uses the kmalloc slabs to store the slab headers
690 * for other slabs "off slab".
691 * The locking for this is tricky in that it nests within the locks
692 * of all other slabs in a few places; to deal with this special
693 * locking we put on-slab caches into a separate lock-class.
695 * We set lock class for alien array caches which are up during init.
696 * The lock annotation will be lost if all cpus of a node goes down and
697 * then comes back up during hotplug
699 static struct lock_class_key on_slab_l3_key;
700 static struct lock_class_key on_slab_alc_key;
702 static inline void init_lock_keys(void)
706 struct cache_sizes *s = malloc_sizes;
708 while (s->cs_size != ULONG_MAX) {
710 struct array_cache **alc;
712 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
713 if (!l3 || OFF_SLAB(s->cs_cachep))
715 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
718 * FIXME: This check for BAD_ALIEN_MAGIC
719 * should go away when common slab code is taught to
720 * work even without alien caches.
721 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
722 * for alloc_alien_cache,
724 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
728 lockdep_set_class(&alc[r]->lock,
736 static inline void init_lock_keys(void)
742 * Guard access to the cache-chain.
744 static DEFINE_MUTEX(cache_chain_mutex);
745 static struct list_head cache_chain;
748 * chicken and egg problem: delay the per-cpu array allocation
749 * until the general caches are up.
759 * used by boot code to determine if it can use slab based allocator
761 int slab_is_available(void)
763 return g_cpucache_up == FULL;
766 static DEFINE_PER_CPU(struct delayed_work, reap_work);
768 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
770 return cachep->array[smp_processor_id()];
773 static inline struct kmem_cache *__find_general_cachep(size_t size,
776 struct cache_sizes *csizep = malloc_sizes;
779 /* This happens if someone tries to call
780 * kmem_cache_create(), or __kmalloc(), before
781 * the generic caches are initialized.
783 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
786 return ZERO_SIZE_PTR;
788 while (size > csizep->cs_size)
792 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
793 * has cs_{dma,}cachep==NULL. Thus no special case
794 * for large kmalloc calls required.
796 #ifdef CONFIG_ZONE_DMA
797 if (unlikely(gfpflags & GFP_DMA))
798 return csizep->cs_dmacachep;
800 return csizep->cs_cachep;
803 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
805 return __find_general_cachep(size, gfpflags);
808 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
810 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
814 * Calculate the number of objects and left-over bytes for a given buffer size.
816 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
817 size_t align, int flags, size_t *left_over,
822 size_t slab_size = PAGE_SIZE << gfporder;
825 * The slab management structure can be either off the slab or
826 * on it. For the latter case, the memory allocated for a
830 * - One kmem_bufctl_t for each object
831 * - Padding to respect alignment of @align
832 * - @buffer_size bytes for each object
834 * If the slab management structure is off the slab, then the
835 * alignment will already be calculated into the size. Because
836 * the slabs are all pages aligned, the objects will be at the
837 * correct alignment when allocated.
839 if (flags & CFLGS_OFF_SLAB) {
841 nr_objs = slab_size / buffer_size;
843 if (nr_objs > SLAB_LIMIT)
844 nr_objs = SLAB_LIMIT;
847 * Ignore padding for the initial guess. The padding
848 * is at most @align-1 bytes, and @buffer_size is at
849 * least @align. In the worst case, this result will
850 * be one greater than the number of objects that fit
851 * into the memory allocation when taking the padding
854 nr_objs = (slab_size - sizeof(struct slab)) /
855 (buffer_size + sizeof(kmem_bufctl_t));
858 * This calculated number will be either the right
859 * amount, or one greater than what we want.
861 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
865 if (nr_objs > SLAB_LIMIT)
866 nr_objs = SLAB_LIMIT;
868 mgmt_size = slab_mgmt_size(nr_objs, align);
871 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
874 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
876 static void __slab_error(const char *function, struct kmem_cache *cachep,
879 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
880 function, cachep->name, msg);
885 * By default on NUMA we use alien caches to stage the freeing of
886 * objects allocated from other nodes. This causes massive memory
887 * inefficiencies when using fake NUMA setup to split memory into a
888 * large number of small nodes, so it can be disabled on the command
892 static int use_alien_caches __read_mostly = 1;
893 static int numa_platform __read_mostly = 1;
894 static int __init noaliencache_setup(char *s)
896 use_alien_caches = 0;
899 __setup("noaliencache", noaliencache_setup);
903 * Special reaping functions for NUMA systems called from cache_reap().
904 * These take care of doing round robin flushing of alien caches (containing
905 * objects freed on different nodes from which they were allocated) and the
906 * flushing of remote pcps by calling drain_node_pages.
908 static DEFINE_PER_CPU(unsigned long, reap_node);
910 static void init_reap_node(int cpu)
914 node = next_node(cpu_to_node(cpu), node_online_map);
915 if (node == MAX_NUMNODES)
916 node = first_node(node_online_map);
918 per_cpu(reap_node, cpu) = node;
921 static void next_reap_node(void)
923 int node = __get_cpu_var(reap_node);
925 node = next_node(node, node_online_map);
926 if (unlikely(node >= MAX_NUMNODES))
927 node = first_node(node_online_map);
928 __get_cpu_var(reap_node) = node;
932 #define init_reap_node(cpu) do { } while (0)
933 #define next_reap_node(void) do { } while (0)
937 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
938 * via the workqueue/eventd.
939 * Add the CPU number into the expiration time to minimize the possibility of
940 * the CPUs getting into lockstep and contending for the global cache chain
943 static void __cpuinit start_cpu_timer(int cpu)
945 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
948 * When this gets called from do_initcalls via cpucache_init(),
949 * init_workqueues() has already run, so keventd will be setup
952 if (keventd_up() && reap_work->work.func == NULL) {
954 INIT_DELAYED_WORK(reap_work, cache_reap);
955 schedule_delayed_work_on(cpu, reap_work,
956 __round_jiffies_relative(HZ, cpu));
960 static struct array_cache *alloc_arraycache(int node, int entries,
963 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
964 struct array_cache *nc = NULL;
966 nc = kmalloc_node(memsize, GFP_KERNEL, node);
970 nc->batchcount = batchcount;
972 spin_lock_init(&nc->lock);
978 * Transfer objects in one arraycache to another.
979 * Locking must be handled by the caller.
981 * Return the number of entries transferred.
983 static int transfer_objects(struct array_cache *to,
984 struct array_cache *from, unsigned int max)
986 /* Figure out how many entries to transfer */
987 int nr = min(min(from->avail, max), to->limit - to->avail);
992 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1003 #define drain_alien_cache(cachep, alien) do { } while (0)
1004 #define reap_alien(cachep, l3) do { } while (0)
1006 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1008 return (struct array_cache **)BAD_ALIEN_MAGIC;
1011 static inline void free_alien_cache(struct array_cache **ac_ptr)
1015 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1020 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1026 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1027 gfp_t flags, int nodeid)
1032 #else /* CONFIG_NUMA */
1034 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1035 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1037 static struct array_cache **alloc_alien_cache(int node, int limit)
1039 struct array_cache **ac_ptr;
1040 int memsize = sizeof(void *) * nr_node_ids;
1045 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1048 if (i == node || !node_online(i)) {
1052 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1054 for (i--; i >= 0; i--)
1064 static void free_alien_cache(struct array_cache **ac_ptr)
1075 static void __drain_alien_cache(struct kmem_cache *cachep,
1076 struct array_cache *ac, int node)
1078 struct kmem_list3 *rl3 = cachep->nodelists[node];
1081 spin_lock(&rl3->list_lock);
1083 * Stuff objects into the remote nodes shared array first.
1084 * That way we could avoid the overhead of putting the objects
1085 * into the free lists and getting them back later.
1088 transfer_objects(rl3->shared, ac, ac->limit);
1090 free_block(cachep, ac->entry, ac->avail, node);
1092 spin_unlock(&rl3->list_lock);
1097 * Called from cache_reap() to regularly drain alien caches round robin.
1099 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1101 int node = __get_cpu_var(reap_node);
1104 struct array_cache *ac = l3->alien[node];
1106 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1107 __drain_alien_cache(cachep, ac, node);
1108 spin_unlock_irq(&ac->lock);
1113 static void drain_alien_cache(struct kmem_cache *cachep,
1114 struct array_cache **alien)
1117 struct array_cache *ac;
1118 unsigned long flags;
1120 for_each_online_node(i) {
1123 spin_lock_irqsave(&ac->lock, flags);
1124 __drain_alien_cache(cachep, ac, i);
1125 spin_unlock_irqrestore(&ac->lock, flags);
1130 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1132 struct slab *slabp = virt_to_slab(objp);
1133 int nodeid = slabp->nodeid;
1134 struct kmem_list3 *l3;
1135 struct array_cache *alien = NULL;
1138 node = numa_node_id();
1141 * Make sure we are not freeing a object from another node to the array
1142 * cache on this cpu.
1144 if (likely(slabp->nodeid == node))
1147 l3 = cachep->nodelists[node];
1148 STATS_INC_NODEFREES(cachep);
1149 if (l3->alien && l3->alien[nodeid]) {
1150 alien = l3->alien[nodeid];
1151 spin_lock(&alien->lock);
1152 if (unlikely(alien->avail == alien->limit)) {
1153 STATS_INC_ACOVERFLOW(cachep);
1154 __drain_alien_cache(cachep, alien, nodeid);
1156 alien->entry[alien->avail++] = objp;
1157 spin_unlock(&alien->lock);
1159 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1160 free_block(cachep, &objp, 1, nodeid);
1161 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1167 static void __cpuinit cpuup_canceled(long cpu)
1169 struct kmem_cache *cachep;
1170 struct kmem_list3 *l3 = NULL;
1171 int node = cpu_to_node(cpu);
1172 const struct cpumask *mask = cpumask_of_node(node);
1174 list_for_each_entry(cachep, &cache_chain, next) {
1175 struct array_cache *nc;
1176 struct array_cache *shared;
1177 struct array_cache **alien;
1179 /* cpu is dead; no one can alloc from it. */
1180 nc = cachep->array[cpu];
1181 cachep->array[cpu] = NULL;
1182 l3 = cachep->nodelists[node];
1185 goto free_array_cache;
1187 spin_lock_irq(&l3->list_lock);
1189 /* Free limit for this kmem_list3 */
1190 l3->free_limit -= cachep->batchcount;
1192 free_block(cachep, nc->entry, nc->avail, node);
1194 if (!cpus_empty(*mask)) {
1195 spin_unlock_irq(&l3->list_lock);
1196 goto free_array_cache;
1199 shared = l3->shared;
1201 free_block(cachep, shared->entry,
1202 shared->avail, node);
1209 spin_unlock_irq(&l3->list_lock);
1213 drain_alien_cache(cachep, alien);
1214 free_alien_cache(alien);
1220 * In the previous loop, all the objects were freed to
1221 * the respective cache's slabs, now we can go ahead and
1222 * shrink each nodelist to its limit.
1224 list_for_each_entry(cachep, &cache_chain, next) {
1225 l3 = cachep->nodelists[node];
1228 drain_freelist(cachep, l3, l3->free_objects);
1232 static int __cpuinit cpuup_prepare(long cpu)
1234 struct kmem_cache *cachep;
1235 struct kmem_list3 *l3 = NULL;
1236 int node = cpu_to_node(cpu);
1237 const int memsize = sizeof(struct kmem_list3);
1240 * We need to do this right in the beginning since
1241 * alloc_arraycache's are going to use this list.
1242 * kmalloc_node allows us to add the slab to the right
1243 * kmem_list3 and not this cpu's kmem_list3
1246 list_for_each_entry(cachep, &cache_chain, next) {
1248 * Set up the size64 kmemlist for cpu before we can
1249 * begin anything. Make sure some other cpu on this
1250 * node has not already allocated this
1252 if (!cachep->nodelists[node]) {
1253 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1256 kmem_list3_init(l3);
1257 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1258 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1261 * The l3s don't come and go as CPUs come and
1262 * go. cache_chain_mutex is sufficient
1265 cachep->nodelists[node] = l3;
1268 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1269 cachep->nodelists[node]->free_limit =
1270 (1 + nr_cpus_node(node)) *
1271 cachep->batchcount + cachep->num;
1272 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1276 * Now we can go ahead with allocating the shared arrays and
1279 list_for_each_entry(cachep, &cache_chain, next) {
1280 struct array_cache *nc;
1281 struct array_cache *shared = NULL;
1282 struct array_cache **alien = NULL;
1284 nc = alloc_arraycache(node, cachep->limit,
1285 cachep->batchcount);
1288 if (cachep->shared) {
1289 shared = alloc_arraycache(node,
1290 cachep->shared * cachep->batchcount,
1297 if (use_alien_caches) {
1298 alien = alloc_alien_cache(node, cachep->limit);
1305 cachep->array[cpu] = nc;
1306 l3 = cachep->nodelists[node];
1309 spin_lock_irq(&l3->list_lock);
1312 * We are serialised from CPU_DEAD or
1313 * CPU_UP_CANCELLED by the cpucontrol lock
1315 l3->shared = shared;
1324 spin_unlock_irq(&l3->list_lock);
1326 free_alien_cache(alien);
1330 cpuup_canceled(cpu);
1334 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1335 unsigned long action, void *hcpu)
1337 long cpu = (long)hcpu;
1341 case CPU_UP_PREPARE:
1342 case CPU_UP_PREPARE_FROZEN:
1343 mutex_lock(&cache_chain_mutex);
1344 err = cpuup_prepare(cpu);
1345 mutex_unlock(&cache_chain_mutex);
1348 case CPU_ONLINE_FROZEN:
1349 start_cpu_timer(cpu);
1351 #ifdef CONFIG_HOTPLUG_CPU
1352 case CPU_DOWN_PREPARE:
1353 case CPU_DOWN_PREPARE_FROZEN:
1355 * Shutdown cache reaper. Note that the cache_chain_mutex is
1356 * held so that if cache_reap() is invoked it cannot do
1357 * anything expensive but will only modify reap_work
1358 * and reschedule the timer.
1360 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1361 /* Now the cache_reaper is guaranteed to be not running. */
1362 per_cpu(reap_work, cpu).work.func = NULL;
1364 case CPU_DOWN_FAILED:
1365 case CPU_DOWN_FAILED_FROZEN:
1366 start_cpu_timer(cpu);
1369 case CPU_DEAD_FROZEN:
1371 * Even if all the cpus of a node are down, we don't free the
1372 * kmem_list3 of any cache. This to avoid a race between
1373 * cpu_down, and a kmalloc allocation from another cpu for
1374 * memory from the node of the cpu going down. The list3
1375 * structure is usually allocated from kmem_cache_create() and
1376 * gets destroyed at kmem_cache_destroy().
1380 case CPU_UP_CANCELED:
1381 case CPU_UP_CANCELED_FROZEN:
1382 mutex_lock(&cache_chain_mutex);
1383 cpuup_canceled(cpu);
1384 mutex_unlock(&cache_chain_mutex);
1387 return err ? NOTIFY_BAD : NOTIFY_OK;
1390 static struct notifier_block __cpuinitdata cpucache_notifier = {
1391 &cpuup_callback, NULL, 0
1395 * swap the static kmem_list3 with kmalloced memory
1397 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1400 struct kmem_list3 *ptr;
1402 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1405 local_irq_disable();
1406 memcpy(ptr, list, sizeof(struct kmem_list3));
1408 * Do not assume that spinlocks can be initialized via memcpy:
1410 spin_lock_init(&ptr->list_lock);
1412 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1413 cachep->nodelists[nodeid] = ptr;
1418 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1419 * size of kmem_list3.
1421 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1425 for_each_online_node(node) {
1426 cachep->nodelists[node] = &initkmem_list3[index + node];
1427 cachep->nodelists[node]->next_reap = jiffies +
1429 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1434 * Initialisation. Called after the page allocator have been initialised and
1435 * before smp_init().
1437 void __init kmem_cache_init(void)
1440 struct cache_sizes *sizes;
1441 struct cache_names *names;
1446 if (num_possible_nodes() == 1) {
1447 use_alien_caches = 0;
1451 for (i = 0; i < NUM_INIT_LISTS; i++) {
1452 kmem_list3_init(&initkmem_list3[i]);
1453 if (i < MAX_NUMNODES)
1454 cache_cache.nodelists[i] = NULL;
1456 set_up_list3s(&cache_cache, CACHE_CACHE);
1459 * Fragmentation resistance on low memory - only use bigger
1460 * page orders on machines with more than 32MB of memory.
1462 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1463 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1465 /* Bootstrap is tricky, because several objects are allocated
1466 * from caches that do not exist yet:
1467 * 1) initialize the cache_cache cache: it contains the struct
1468 * kmem_cache structures of all caches, except cache_cache itself:
1469 * cache_cache is statically allocated.
1470 * Initially an __init data area is used for the head array and the
1471 * kmem_list3 structures, it's replaced with a kmalloc allocated
1472 * array at the end of the bootstrap.
1473 * 2) Create the first kmalloc cache.
1474 * The struct kmem_cache for the new cache is allocated normally.
1475 * An __init data area is used for the head array.
1476 * 3) Create the remaining kmalloc caches, with minimally sized
1478 * 4) Replace the __init data head arrays for cache_cache and the first
1479 * kmalloc cache with kmalloc allocated arrays.
1480 * 5) Replace the __init data for kmem_list3 for cache_cache and
1481 * the other cache's with kmalloc allocated memory.
1482 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1485 node = numa_node_id();
1487 /* 1) create the cache_cache */
1488 INIT_LIST_HEAD(&cache_chain);
1489 list_add(&cache_cache.next, &cache_chain);
1490 cache_cache.colour_off = cache_line_size();
1491 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1492 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1495 * struct kmem_cache size depends on nr_node_ids, which
1496 * can be less than MAX_NUMNODES.
1498 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1499 nr_node_ids * sizeof(struct kmem_list3 *);
1501 cache_cache.obj_size = cache_cache.buffer_size;
1503 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1505 cache_cache.reciprocal_buffer_size =
1506 reciprocal_value(cache_cache.buffer_size);
1508 for (order = 0; order < MAX_ORDER; order++) {
1509 cache_estimate(order, cache_cache.buffer_size,
1510 cache_line_size(), 0, &left_over, &cache_cache.num);
1511 if (cache_cache.num)
1514 BUG_ON(!cache_cache.num);
1515 cache_cache.gfporder = order;
1516 cache_cache.colour = left_over / cache_cache.colour_off;
1517 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1518 sizeof(struct slab), cache_line_size());
1520 /* 2+3) create the kmalloc caches */
1521 sizes = malloc_sizes;
1522 names = cache_names;
1525 * Initialize the caches that provide memory for the array cache and the
1526 * kmem_list3 structures first. Without this, further allocations will
1530 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1531 sizes[INDEX_AC].cs_size,
1532 ARCH_KMALLOC_MINALIGN,
1533 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1536 if (INDEX_AC != INDEX_L3) {
1537 sizes[INDEX_L3].cs_cachep =
1538 kmem_cache_create(names[INDEX_L3].name,
1539 sizes[INDEX_L3].cs_size,
1540 ARCH_KMALLOC_MINALIGN,
1541 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1545 slab_early_init = 0;
1547 while (sizes->cs_size != ULONG_MAX) {
1549 * For performance, all the general caches are L1 aligned.
1550 * This should be particularly beneficial on SMP boxes, as it
1551 * eliminates "false sharing".
1552 * Note for systems short on memory removing the alignment will
1553 * allow tighter packing of the smaller caches.
1555 if (!sizes->cs_cachep) {
1556 sizes->cs_cachep = kmem_cache_create(names->name,
1558 ARCH_KMALLOC_MINALIGN,
1559 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1562 #ifdef CONFIG_ZONE_DMA
1563 sizes->cs_dmacachep = kmem_cache_create(
1566 ARCH_KMALLOC_MINALIGN,
1567 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1574 /* 4) Replace the bootstrap head arrays */
1576 struct array_cache *ptr;
1578 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1580 local_irq_disable();
1581 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1582 memcpy(ptr, cpu_cache_get(&cache_cache),
1583 sizeof(struct arraycache_init));
1585 * Do not assume that spinlocks can be initialized via memcpy:
1587 spin_lock_init(&ptr->lock);
1589 cache_cache.array[smp_processor_id()] = ptr;
1592 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1594 local_irq_disable();
1595 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1596 != &initarray_generic.cache);
1597 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1598 sizeof(struct arraycache_init));
1600 * Do not assume that spinlocks can be initialized via memcpy:
1602 spin_lock_init(&ptr->lock);
1604 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1608 /* 5) Replace the bootstrap kmem_list3's */
1612 for_each_online_node(nid) {
1613 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1615 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1616 &initkmem_list3[SIZE_AC + nid], nid);
1618 if (INDEX_AC != INDEX_L3) {
1619 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1620 &initkmem_list3[SIZE_L3 + nid], nid);
1625 /* 6) resize the head arrays to their final sizes */
1627 struct kmem_cache *cachep;
1628 mutex_lock(&cache_chain_mutex);
1629 list_for_each_entry(cachep, &cache_chain, next)
1630 if (enable_cpucache(cachep))
1632 mutex_unlock(&cache_chain_mutex);
1635 /* Annotate slab for lockdep -- annotate the malloc caches */
1640 g_cpucache_up = FULL;
1643 * Register a cpu startup notifier callback that initializes
1644 * cpu_cache_get for all new cpus
1646 register_cpu_notifier(&cpucache_notifier);
1649 * The reap timers are started later, with a module init call: That part
1650 * of the kernel is not yet operational.
1654 static int __init cpucache_init(void)
1659 * Register the timers that return unneeded pages to the page allocator
1661 for_each_online_cpu(cpu)
1662 start_cpu_timer(cpu);
1665 __initcall(cpucache_init);
1668 * Interface to system's page allocator. No need to hold the cache-lock.
1670 * If we requested dmaable memory, we will get it. Even if we
1671 * did not request dmaable memory, we might get it, but that
1672 * would be relatively rare and ignorable.
1674 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1682 * Nommu uses slab's for process anonymous memory allocations, and thus
1683 * requires __GFP_COMP to properly refcount higher order allocations
1685 flags |= __GFP_COMP;
1688 flags |= cachep->gfpflags;
1689 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1690 flags |= __GFP_RECLAIMABLE;
1692 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1696 nr_pages = (1 << cachep->gfporder);
1697 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1698 add_zone_page_state(page_zone(page),
1699 NR_SLAB_RECLAIMABLE, nr_pages);
1701 add_zone_page_state(page_zone(page),
1702 NR_SLAB_UNRECLAIMABLE, nr_pages);
1703 for (i = 0; i < nr_pages; i++)
1704 __SetPageSlab(page + i);
1705 return page_address(page);
1709 * Interface to system's page release.
1711 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1713 unsigned long i = (1 << cachep->gfporder);
1714 struct page *page = virt_to_page(addr);
1715 const unsigned long nr_freed = i;
1717 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1718 sub_zone_page_state(page_zone(page),
1719 NR_SLAB_RECLAIMABLE, nr_freed);
1721 sub_zone_page_state(page_zone(page),
1722 NR_SLAB_UNRECLAIMABLE, nr_freed);
1724 BUG_ON(!PageSlab(page));
1725 __ClearPageSlab(page);
1728 if (current->reclaim_state)
1729 current->reclaim_state->reclaimed_slab += nr_freed;
1730 free_pages((unsigned long)addr, cachep->gfporder);
1733 static void kmem_rcu_free(struct rcu_head *head)
1735 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1736 struct kmem_cache *cachep = slab_rcu->cachep;
1738 kmem_freepages(cachep, slab_rcu->addr);
1739 if (OFF_SLAB(cachep))
1740 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1745 #ifdef CONFIG_DEBUG_PAGEALLOC
1746 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1747 unsigned long caller)
1749 int size = obj_size(cachep);
1751 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1753 if (size < 5 * sizeof(unsigned long))
1756 *addr++ = 0x12345678;
1758 *addr++ = smp_processor_id();
1759 size -= 3 * sizeof(unsigned long);
1761 unsigned long *sptr = &caller;
1762 unsigned long svalue;
1764 while (!kstack_end(sptr)) {
1766 if (kernel_text_address(svalue)) {
1768 size -= sizeof(unsigned long);
1769 if (size <= sizeof(unsigned long))
1775 *addr++ = 0x87654321;
1779 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1781 int size = obj_size(cachep);
1782 addr = &((char *)addr)[obj_offset(cachep)];
1784 memset(addr, val, size);
1785 *(unsigned char *)(addr + size - 1) = POISON_END;
1788 static void dump_line(char *data, int offset, int limit)
1791 unsigned char error = 0;
1794 printk(KERN_ERR "%03x:", offset);
1795 for (i = 0; i < limit; i++) {
1796 if (data[offset + i] != POISON_FREE) {
1797 error = data[offset + i];
1800 printk(" %02x", (unsigned char)data[offset + i]);
1804 if (bad_count == 1) {
1805 error ^= POISON_FREE;
1806 if (!(error & (error - 1))) {
1807 printk(KERN_ERR "Single bit error detected. Probably "
1810 printk(KERN_ERR "Run memtest86+ or a similar memory "
1813 printk(KERN_ERR "Run a memory test tool.\n");
1822 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1827 if (cachep->flags & SLAB_RED_ZONE) {
1828 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1829 *dbg_redzone1(cachep, objp),
1830 *dbg_redzone2(cachep, objp));
1833 if (cachep->flags & SLAB_STORE_USER) {
1834 printk(KERN_ERR "Last user: [<%p>]",
1835 *dbg_userword(cachep, objp));
1836 print_symbol("(%s)",
1837 (unsigned long)*dbg_userword(cachep, objp));
1840 realobj = (char *)objp + obj_offset(cachep);
1841 size = obj_size(cachep);
1842 for (i = 0; i < size && lines; i += 16, lines--) {
1845 if (i + limit > size)
1847 dump_line(realobj, i, limit);
1851 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1857 realobj = (char *)objp + obj_offset(cachep);
1858 size = obj_size(cachep);
1860 for (i = 0; i < size; i++) {
1861 char exp = POISON_FREE;
1864 if (realobj[i] != exp) {
1870 "Slab corruption: %s start=%p, len=%d\n",
1871 cachep->name, realobj, size);
1872 print_objinfo(cachep, objp, 0);
1874 /* Hexdump the affected line */
1877 if (i + limit > size)
1879 dump_line(realobj, i, limit);
1882 /* Limit to 5 lines */
1888 /* Print some data about the neighboring objects, if they
1891 struct slab *slabp = virt_to_slab(objp);
1894 objnr = obj_to_index(cachep, slabp, objp);
1896 objp = index_to_obj(cachep, slabp, objnr - 1);
1897 realobj = (char *)objp + obj_offset(cachep);
1898 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1900 print_objinfo(cachep, objp, 2);
1902 if (objnr + 1 < cachep->num) {
1903 objp = index_to_obj(cachep, slabp, objnr + 1);
1904 realobj = (char *)objp + obj_offset(cachep);
1905 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1907 print_objinfo(cachep, objp, 2);
1914 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1917 for (i = 0; i < cachep->num; i++) {
1918 void *objp = index_to_obj(cachep, slabp, i);
1920 if (cachep->flags & SLAB_POISON) {
1921 #ifdef CONFIG_DEBUG_PAGEALLOC
1922 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1924 kernel_map_pages(virt_to_page(objp),
1925 cachep->buffer_size / PAGE_SIZE, 1);
1927 check_poison_obj(cachep, objp);
1929 check_poison_obj(cachep, objp);
1932 if (cachep->flags & SLAB_RED_ZONE) {
1933 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1934 slab_error(cachep, "start of a freed object "
1936 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1937 slab_error(cachep, "end of a freed object "
1943 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1949 * slab_destroy - destroy and release all objects in a slab
1950 * @cachep: cache pointer being destroyed
1951 * @slabp: slab pointer being destroyed
1953 * Destroy all the objs in a slab, and release the mem back to the system.
1954 * Before calling the slab must have been unlinked from the cache. The
1955 * cache-lock is not held/needed.
1957 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1959 void *addr = slabp->s_mem - slabp->colouroff;
1961 slab_destroy_debugcheck(cachep, slabp);
1962 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1963 struct slab_rcu *slab_rcu;
1965 slab_rcu = (struct slab_rcu *)slabp;
1966 slab_rcu->cachep = cachep;
1967 slab_rcu->addr = addr;
1968 call_rcu(&slab_rcu->head, kmem_rcu_free);
1970 kmem_freepages(cachep, addr);
1971 if (OFF_SLAB(cachep))
1972 kmem_cache_free(cachep->slabp_cache, slabp);
1976 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1979 struct kmem_list3 *l3;
1981 for_each_online_cpu(i)
1982 kfree(cachep->array[i]);
1984 /* NUMA: free the list3 structures */
1985 for_each_online_node(i) {
1986 l3 = cachep->nodelists[i];
1989 free_alien_cache(l3->alien);
1993 kmem_cache_free(&cache_cache, cachep);
1998 * calculate_slab_order - calculate size (page order) of slabs
1999 * @cachep: pointer to the cache that is being created
2000 * @size: size of objects to be created in this cache.
2001 * @align: required alignment for the objects.
2002 * @flags: slab allocation flags
2004 * Also calculates the number of objects per slab.
2006 * This could be made much more intelligent. For now, try to avoid using
2007 * high order pages for slabs. When the gfp() functions are more friendly
2008 * towards high-order requests, this should be changed.
2010 static size_t calculate_slab_order(struct kmem_cache *cachep,
2011 size_t size, size_t align, unsigned long flags)
2013 unsigned long offslab_limit;
2014 size_t left_over = 0;
2017 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2021 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2025 if (flags & CFLGS_OFF_SLAB) {
2027 * Max number of objs-per-slab for caches which
2028 * use off-slab slabs. Needed to avoid a possible
2029 * looping condition in cache_grow().
2031 offslab_limit = size - sizeof(struct slab);
2032 offslab_limit /= sizeof(kmem_bufctl_t);
2034 if (num > offslab_limit)
2038 /* Found something acceptable - save it away */
2040 cachep->gfporder = gfporder;
2041 left_over = remainder;
2044 * A VFS-reclaimable slab tends to have most allocations
2045 * as GFP_NOFS and we really don't want to have to be allocating
2046 * higher-order pages when we are unable to shrink dcache.
2048 if (flags & SLAB_RECLAIM_ACCOUNT)
2052 * Large number of objects is good, but very large slabs are
2053 * currently bad for the gfp()s.
2055 if (gfporder >= slab_break_gfp_order)
2059 * Acceptable internal fragmentation?
2061 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2067 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
2069 if (g_cpucache_up == FULL)
2070 return enable_cpucache(cachep);
2072 if (g_cpucache_up == NONE) {
2074 * Note: the first kmem_cache_create must create the cache
2075 * that's used by kmalloc(24), otherwise the creation of
2076 * further caches will BUG().
2078 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2081 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2082 * the first cache, then we need to set up all its list3s,
2083 * otherwise the creation of further caches will BUG().
2085 set_up_list3s(cachep, SIZE_AC);
2086 if (INDEX_AC == INDEX_L3)
2087 g_cpucache_up = PARTIAL_L3;
2089 g_cpucache_up = PARTIAL_AC;
2091 cachep->array[smp_processor_id()] =
2092 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2094 if (g_cpucache_up == PARTIAL_AC) {
2095 set_up_list3s(cachep, SIZE_L3);
2096 g_cpucache_up = PARTIAL_L3;
2099 for_each_online_node(node) {
2100 cachep->nodelists[node] =
2101 kmalloc_node(sizeof(struct kmem_list3),
2103 BUG_ON(!cachep->nodelists[node]);
2104 kmem_list3_init(cachep->nodelists[node]);
2108 cachep->nodelists[numa_node_id()]->next_reap =
2109 jiffies + REAPTIMEOUT_LIST3 +
2110 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2112 cpu_cache_get(cachep)->avail = 0;
2113 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2114 cpu_cache_get(cachep)->batchcount = 1;
2115 cpu_cache_get(cachep)->touched = 0;
2116 cachep->batchcount = 1;
2117 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2122 * kmem_cache_create - Create a cache.
2123 * @name: A string which is used in /proc/slabinfo to identify this cache.
2124 * @size: The size of objects to be created in this cache.
2125 * @align: The required alignment for the objects.
2126 * @flags: SLAB flags
2127 * @ctor: A constructor for the objects.
2129 * Returns a ptr to the cache on success, NULL on failure.
2130 * Cannot be called within a int, but can be interrupted.
2131 * The @ctor is run when new pages are allocated by the cache.
2133 * @name must be valid until the cache is destroyed. This implies that
2134 * the module calling this has to destroy the cache before getting unloaded.
2135 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2136 * therefore applications must manage it themselves.
2140 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2141 * to catch references to uninitialised memory.
2143 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2144 * for buffer overruns.
2146 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2147 * cacheline. This can be beneficial if you're counting cycles as closely
2151 kmem_cache_create (const char *name, size_t size, size_t align,
2152 unsigned long flags, void (*ctor)(void *))
2154 size_t left_over, slab_size, ralign;
2155 struct kmem_cache *cachep = NULL, *pc;
2158 * Sanity checks... these are all serious usage bugs.
2160 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2161 size > KMALLOC_MAX_SIZE) {
2162 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2168 * We use cache_chain_mutex to ensure a consistent view of
2169 * cpu_online_mask as well. Please see cpuup_callback
2172 mutex_lock(&cache_chain_mutex);
2174 list_for_each_entry(pc, &cache_chain, next) {
2179 * This happens when the module gets unloaded and doesn't
2180 * destroy its slab cache and no-one else reuses the vmalloc
2181 * area of the module. Print a warning.
2183 res = probe_kernel_address(pc->name, tmp);
2186 "SLAB: cache with size %d has lost its name\n",
2191 if (!strcmp(pc->name, name)) {
2193 "kmem_cache_create: duplicate cache %s\n", name);
2200 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2203 * Enable redzoning and last user accounting, except for caches with
2204 * large objects, if the increased size would increase the object size
2205 * above the next power of two: caches with object sizes just above a
2206 * power of two have a significant amount of internal fragmentation.
2208 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2209 2 * sizeof(unsigned long long)))
2210 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2211 if (!(flags & SLAB_DESTROY_BY_RCU))
2212 flags |= SLAB_POISON;
2214 if (flags & SLAB_DESTROY_BY_RCU)
2215 BUG_ON(flags & SLAB_POISON);
2218 * Always checks flags, a caller might be expecting debug support which
2221 BUG_ON(flags & ~CREATE_MASK);
2224 * Check that size is in terms of words. This is needed to avoid
2225 * unaligned accesses for some archs when redzoning is used, and makes
2226 * sure any on-slab bufctl's are also correctly aligned.
2228 if (size & (BYTES_PER_WORD - 1)) {
2229 size += (BYTES_PER_WORD - 1);
2230 size &= ~(BYTES_PER_WORD - 1);
2233 /* calculate the final buffer alignment: */
2235 /* 1) arch recommendation: can be overridden for debug */
2236 if (flags & SLAB_HWCACHE_ALIGN) {
2238 * Default alignment: as specified by the arch code. Except if
2239 * an object is really small, then squeeze multiple objects into
2242 ralign = cache_line_size();
2243 while (size <= ralign / 2)
2246 ralign = BYTES_PER_WORD;
2250 * Redzoning and user store require word alignment or possibly larger.
2251 * Note this will be overridden by architecture or caller mandated
2252 * alignment if either is greater than BYTES_PER_WORD.
2254 if (flags & SLAB_STORE_USER)
2255 ralign = BYTES_PER_WORD;
2257 if (flags & SLAB_RED_ZONE) {
2258 ralign = REDZONE_ALIGN;
2259 /* If redzoning, ensure that the second redzone is suitably
2260 * aligned, by adjusting the object size accordingly. */
2261 size += REDZONE_ALIGN - 1;
2262 size &= ~(REDZONE_ALIGN - 1);
2265 /* 2) arch mandated alignment */
2266 if (ralign < ARCH_SLAB_MINALIGN) {
2267 ralign = ARCH_SLAB_MINALIGN;
2269 /* 3) caller mandated alignment */
2270 if (ralign < align) {
2273 /* disable debug if necessary */
2274 if (ralign > __alignof__(unsigned long long))
2275 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2281 /* Get cache's description obj. */
2282 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2287 cachep->obj_size = size;
2290 * Both debugging options require word-alignment which is calculated
2293 if (flags & SLAB_RED_ZONE) {
2294 /* add space for red zone words */
2295 cachep->obj_offset += sizeof(unsigned long long);
2296 size += 2 * sizeof(unsigned long long);
2298 if (flags & SLAB_STORE_USER) {
2299 /* user store requires one word storage behind the end of
2300 * the real object. But if the second red zone needs to be
2301 * aligned to 64 bits, we must allow that much space.
2303 if (flags & SLAB_RED_ZONE)
2304 size += REDZONE_ALIGN;
2306 size += BYTES_PER_WORD;
2308 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2309 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2310 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2311 cachep->obj_offset += PAGE_SIZE - size;
2318 * Determine if the slab management is 'on' or 'off' slab.
2319 * (bootstrapping cannot cope with offslab caches so don't do
2322 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2324 * Size is large, assume best to place the slab management obj
2325 * off-slab (should allow better packing of objs).
2327 flags |= CFLGS_OFF_SLAB;
2329 size = ALIGN(size, align);
2331 left_over = calculate_slab_order(cachep, size, align, flags);
2335 "kmem_cache_create: couldn't create cache %s.\n", name);
2336 kmem_cache_free(&cache_cache, cachep);
2340 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2341 + sizeof(struct slab), align);
2344 * If the slab has been placed off-slab, and we have enough space then
2345 * move it on-slab. This is at the expense of any extra colouring.
2347 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2348 flags &= ~CFLGS_OFF_SLAB;
2349 left_over -= slab_size;
2352 if (flags & CFLGS_OFF_SLAB) {
2353 /* really off slab. No need for manual alignment */
2355 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2358 cachep->colour_off = cache_line_size();
2359 /* Offset must be a multiple of the alignment. */
2360 if (cachep->colour_off < align)
2361 cachep->colour_off = align;
2362 cachep->colour = left_over / cachep->colour_off;
2363 cachep->slab_size = slab_size;
2364 cachep->flags = flags;
2365 cachep->gfpflags = 0;
2366 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2367 cachep->gfpflags |= GFP_DMA;
2368 cachep->buffer_size = size;
2369 cachep->reciprocal_buffer_size = reciprocal_value(size);
2371 if (flags & CFLGS_OFF_SLAB) {
2372 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2374 * This is a possibility for one of the malloc_sizes caches.
2375 * But since we go off slab only for object size greater than
2376 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2377 * this should not happen at all.
2378 * But leave a BUG_ON for some lucky dude.
2380 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2382 cachep->ctor = ctor;
2383 cachep->name = name;
2385 if (setup_cpu_cache(cachep)) {
2386 __kmem_cache_destroy(cachep);
2391 /* cache setup completed, link it into the list */
2392 list_add(&cachep->next, &cache_chain);
2394 if (!cachep && (flags & SLAB_PANIC))
2395 panic("kmem_cache_create(): failed to create slab `%s'\n",
2397 mutex_unlock(&cache_chain_mutex);
2401 EXPORT_SYMBOL(kmem_cache_create);
2404 static void check_irq_off(void)
2406 BUG_ON(!irqs_disabled());
2409 static void check_irq_on(void)
2411 BUG_ON(irqs_disabled());
2414 static void check_spinlock_acquired(struct kmem_cache *cachep)
2418 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2422 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2426 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2431 #define check_irq_off() do { } while(0)
2432 #define check_irq_on() do { } while(0)
2433 #define check_spinlock_acquired(x) do { } while(0)
2434 #define check_spinlock_acquired_node(x, y) do { } while(0)
2437 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2438 struct array_cache *ac,
2439 int force, int node);
2441 static void do_drain(void *arg)
2443 struct kmem_cache *cachep = arg;
2444 struct array_cache *ac;
2445 int node = numa_node_id();
2448 ac = cpu_cache_get(cachep);
2449 spin_lock(&cachep->nodelists[node]->list_lock);
2450 free_block(cachep, ac->entry, ac->avail, node);
2451 spin_unlock(&cachep->nodelists[node]->list_lock);
2455 static void drain_cpu_caches(struct kmem_cache *cachep)
2457 struct kmem_list3 *l3;
2460 on_each_cpu(do_drain, cachep, 1);
2462 for_each_online_node(node) {
2463 l3 = cachep->nodelists[node];
2464 if (l3 && l3->alien)
2465 drain_alien_cache(cachep, l3->alien);
2468 for_each_online_node(node) {
2469 l3 = cachep->nodelists[node];
2471 drain_array(cachep, l3, l3->shared, 1, node);
2476 * Remove slabs from the list of free slabs.
2477 * Specify the number of slabs to drain in tofree.
2479 * Returns the actual number of slabs released.
2481 static int drain_freelist(struct kmem_cache *cache,
2482 struct kmem_list3 *l3, int tofree)
2484 struct list_head *p;
2489 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2491 spin_lock_irq(&l3->list_lock);
2492 p = l3->slabs_free.prev;
2493 if (p == &l3->slabs_free) {
2494 spin_unlock_irq(&l3->list_lock);
2498 slabp = list_entry(p, struct slab, list);
2500 BUG_ON(slabp->inuse);
2502 list_del(&slabp->list);
2504 * Safe to drop the lock. The slab is no longer linked
2507 l3->free_objects -= cache->num;
2508 spin_unlock_irq(&l3->list_lock);
2509 slab_destroy(cache, slabp);
2516 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2517 static int __cache_shrink(struct kmem_cache *cachep)
2520 struct kmem_list3 *l3;
2522 drain_cpu_caches(cachep);
2525 for_each_online_node(i) {
2526 l3 = cachep->nodelists[i];
2530 drain_freelist(cachep, l3, l3->free_objects);
2532 ret += !list_empty(&l3->slabs_full) ||
2533 !list_empty(&l3->slabs_partial);
2535 return (ret ? 1 : 0);
2539 * kmem_cache_shrink - Shrink a cache.
2540 * @cachep: The cache to shrink.
2542 * Releases as many slabs as possible for a cache.
2543 * To help debugging, a zero exit status indicates all slabs were released.
2545 int kmem_cache_shrink(struct kmem_cache *cachep)
2548 BUG_ON(!cachep || in_interrupt());
2551 mutex_lock(&cache_chain_mutex);
2552 ret = __cache_shrink(cachep);
2553 mutex_unlock(&cache_chain_mutex);
2557 EXPORT_SYMBOL(kmem_cache_shrink);
2560 * kmem_cache_destroy - delete a cache
2561 * @cachep: the cache to destroy
2563 * Remove a &struct kmem_cache object from the slab cache.
2565 * It is expected this function will be called by a module when it is
2566 * unloaded. This will remove the cache completely, and avoid a duplicate
2567 * cache being allocated each time a module is loaded and unloaded, if the
2568 * module doesn't have persistent in-kernel storage across loads and unloads.
2570 * The cache must be empty before calling this function.
2572 * The caller must guarantee that noone will allocate memory from the cache
2573 * during the kmem_cache_destroy().
2575 void kmem_cache_destroy(struct kmem_cache *cachep)
2577 BUG_ON(!cachep || in_interrupt());
2579 /* Find the cache in the chain of caches. */
2581 mutex_lock(&cache_chain_mutex);
2583 * the chain is never empty, cache_cache is never destroyed
2585 list_del(&cachep->next);
2586 if (__cache_shrink(cachep)) {
2587 slab_error(cachep, "Can't free all objects");
2588 list_add(&cachep->next, &cache_chain);
2589 mutex_unlock(&cache_chain_mutex);
2594 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2597 __kmem_cache_destroy(cachep);
2598 mutex_unlock(&cache_chain_mutex);
2601 EXPORT_SYMBOL(kmem_cache_destroy);
2604 * Get the memory for a slab management obj.
2605 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2606 * always come from malloc_sizes caches. The slab descriptor cannot
2607 * come from the same cache which is getting created because,
2608 * when we are searching for an appropriate cache for these
2609 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2610 * If we are creating a malloc_sizes cache here it would not be visible to
2611 * kmem_find_general_cachep till the initialization is complete.
2612 * Hence we cannot have slabp_cache same as the original cache.
2614 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2615 int colour_off, gfp_t local_flags,
2620 if (OFF_SLAB(cachep)) {
2621 /* Slab management obj is off-slab. */
2622 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2623 local_flags, nodeid);
2627 slabp = objp + colour_off;
2628 colour_off += cachep->slab_size;
2631 slabp->colouroff = colour_off;
2632 slabp->s_mem = objp + colour_off;
2633 slabp->nodeid = nodeid;
2638 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2640 return (kmem_bufctl_t *) (slabp + 1);
2643 static void cache_init_objs(struct kmem_cache *cachep,
2648 for (i = 0; i < cachep->num; i++) {
2649 void *objp = index_to_obj(cachep, slabp, i);
2651 /* need to poison the objs? */
2652 if (cachep->flags & SLAB_POISON)
2653 poison_obj(cachep, objp, POISON_FREE);
2654 if (cachep->flags & SLAB_STORE_USER)
2655 *dbg_userword(cachep, objp) = NULL;
2657 if (cachep->flags & SLAB_RED_ZONE) {
2658 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2659 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2662 * Constructors are not allowed to allocate memory from the same
2663 * cache which they are a constructor for. Otherwise, deadlock.
2664 * They must also be threaded.
2666 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2667 cachep->ctor(objp + obj_offset(cachep));
2669 if (cachep->flags & SLAB_RED_ZONE) {
2670 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2671 slab_error(cachep, "constructor overwrote the"
2672 " end of an object");
2673 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2674 slab_error(cachep, "constructor overwrote the"
2675 " start of an object");
2677 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2678 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2679 kernel_map_pages(virt_to_page(objp),
2680 cachep->buffer_size / PAGE_SIZE, 0);
2685 slab_bufctl(slabp)[i] = i + 1;
2687 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2690 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2692 if (CONFIG_ZONE_DMA_FLAG) {
2693 if (flags & GFP_DMA)
2694 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2696 BUG_ON(cachep->gfpflags & GFP_DMA);
2700 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2703 void *objp = index_to_obj(cachep, slabp, slabp->free);
2707 next = slab_bufctl(slabp)[slabp->free];
2709 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2710 WARN_ON(slabp->nodeid != nodeid);
2717 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2718 void *objp, int nodeid)
2720 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2723 /* Verify that the slab belongs to the intended node */
2724 WARN_ON(slabp->nodeid != nodeid);
2726 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2727 printk(KERN_ERR "slab: double free detected in cache "
2728 "'%s', objp %p\n", cachep->name, objp);
2732 slab_bufctl(slabp)[objnr] = slabp->free;
2733 slabp->free = objnr;
2738 * Map pages beginning at addr to the given cache and slab. This is required
2739 * for the slab allocator to be able to lookup the cache and slab of a
2740 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2742 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2748 page = virt_to_page(addr);
2751 if (likely(!PageCompound(page)))
2752 nr_pages <<= cache->gfporder;
2755 page_set_cache(page, cache);
2756 page_set_slab(page, slab);
2758 } while (--nr_pages);
2762 * Grow (by 1) the number of slabs within a cache. This is called by
2763 * kmem_cache_alloc() when there are no active objs left in a cache.
2765 static int cache_grow(struct kmem_cache *cachep,
2766 gfp_t flags, int nodeid, void *objp)
2771 struct kmem_list3 *l3;
2774 * Be lazy and only check for valid flags here, keeping it out of the
2775 * critical path in kmem_cache_alloc().
2777 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2778 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2780 /* Take the l3 list lock to change the colour_next on this node */
2782 l3 = cachep->nodelists[nodeid];
2783 spin_lock(&l3->list_lock);
2785 /* Get colour for the slab, and cal the next value. */
2786 offset = l3->colour_next;
2788 if (l3->colour_next >= cachep->colour)
2789 l3->colour_next = 0;
2790 spin_unlock(&l3->list_lock);
2792 offset *= cachep->colour_off;
2794 if (local_flags & __GFP_WAIT)
2798 * The test for missing atomic flag is performed here, rather than
2799 * the more obvious place, simply to reduce the critical path length
2800 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2801 * will eventually be caught here (where it matters).
2803 kmem_flagcheck(cachep, flags);
2806 * Get mem for the objs. Attempt to allocate a physical page from
2810 objp = kmem_getpages(cachep, local_flags, nodeid);
2814 /* Get slab management. */
2815 slabp = alloc_slabmgmt(cachep, objp, offset,
2816 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2820 slab_map_pages(cachep, slabp, objp);
2822 cache_init_objs(cachep, slabp);
2824 if (local_flags & __GFP_WAIT)
2825 local_irq_disable();
2827 spin_lock(&l3->list_lock);
2829 /* Make slab active. */
2830 list_add_tail(&slabp->list, &(l3->slabs_free));
2831 STATS_INC_GROWN(cachep);
2832 l3->free_objects += cachep->num;
2833 spin_unlock(&l3->list_lock);
2836 kmem_freepages(cachep, objp);
2838 if (local_flags & __GFP_WAIT)
2839 local_irq_disable();
2846 * Perform extra freeing checks:
2847 * - detect bad pointers.
2848 * - POISON/RED_ZONE checking
2850 static void kfree_debugcheck(const void *objp)
2852 if (!virt_addr_valid(objp)) {
2853 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2854 (unsigned long)objp);
2859 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2861 unsigned long long redzone1, redzone2;
2863 redzone1 = *dbg_redzone1(cache, obj);
2864 redzone2 = *dbg_redzone2(cache, obj);
2869 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2872 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2873 slab_error(cache, "double free detected");
2875 slab_error(cache, "memory outside object was overwritten");
2877 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2878 obj, redzone1, redzone2);
2881 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2888 BUG_ON(virt_to_cache(objp) != cachep);
2890 objp -= obj_offset(cachep);
2891 kfree_debugcheck(objp);
2892 page = virt_to_head_page(objp);
2894 slabp = page_get_slab(page);
2896 if (cachep->flags & SLAB_RED_ZONE) {
2897 verify_redzone_free(cachep, objp);
2898 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2899 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2901 if (cachep->flags & SLAB_STORE_USER)
2902 *dbg_userword(cachep, objp) = caller;
2904 objnr = obj_to_index(cachep, slabp, objp);
2906 BUG_ON(objnr >= cachep->num);
2907 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2909 #ifdef CONFIG_DEBUG_SLAB_LEAK
2910 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2912 if (cachep->flags & SLAB_POISON) {
2913 #ifdef CONFIG_DEBUG_PAGEALLOC
2914 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2915 store_stackinfo(cachep, objp, (unsigned long)caller);
2916 kernel_map_pages(virt_to_page(objp),
2917 cachep->buffer_size / PAGE_SIZE, 0);
2919 poison_obj(cachep, objp, POISON_FREE);
2922 poison_obj(cachep, objp, POISON_FREE);
2928 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2933 /* Check slab's freelist to see if this obj is there. */
2934 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2936 if (entries > cachep->num || i >= cachep->num)
2939 if (entries != cachep->num - slabp->inuse) {
2941 printk(KERN_ERR "slab: Internal list corruption detected in "
2942 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2943 cachep->name, cachep->num, slabp, slabp->inuse);
2945 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2948 printk("\n%03x:", i);
2949 printk(" %02x", ((unsigned char *)slabp)[i]);
2956 #define kfree_debugcheck(x) do { } while(0)
2957 #define cache_free_debugcheck(x,objp,z) (objp)
2958 #define check_slabp(x,y) do { } while(0)
2961 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2964 struct kmem_list3 *l3;
2965 struct array_cache *ac;
2970 node = numa_node_id();
2971 ac = cpu_cache_get(cachep);
2972 batchcount = ac->batchcount;
2973 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2975 * If there was little recent activity on this cache, then
2976 * perform only a partial refill. Otherwise we could generate
2979 batchcount = BATCHREFILL_LIMIT;
2981 l3 = cachep->nodelists[node];
2983 BUG_ON(ac->avail > 0 || !l3);
2984 spin_lock(&l3->list_lock);
2986 /* See if we can refill from the shared array */
2987 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2990 while (batchcount > 0) {
2991 struct list_head *entry;
2993 /* Get slab alloc is to come from. */
2994 entry = l3->slabs_partial.next;
2995 if (entry == &l3->slabs_partial) {
2996 l3->free_touched = 1;
2997 entry = l3->slabs_free.next;
2998 if (entry == &l3->slabs_free)
3002 slabp = list_entry(entry, struct slab, list);
3003 check_slabp(cachep, slabp);
3004 check_spinlock_acquired(cachep);
3007 * The slab was either on partial or free list so
3008 * there must be at least one object available for
3011 BUG_ON(slabp->inuse >= cachep->num);
3013 while (slabp->inuse < cachep->num && batchcount--) {
3014 STATS_INC_ALLOCED(cachep);
3015 STATS_INC_ACTIVE(cachep);
3016 STATS_SET_HIGH(cachep);
3018 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3021 check_slabp(cachep, slabp);
3023 /* move slabp to correct slabp list: */
3024 list_del(&slabp->list);
3025 if (slabp->free == BUFCTL_END)
3026 list_add(&slabp->list, &l3->slabs_full);
3028 list_add(&slabp->list, &l3->slabs_partial);
3032 l3->free_objects -= ac->avail;
3034 spin_unlock(&l3->list_lock);
3036 if (unlikely(!ac->avail)) {
3038 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3040 /* cache_grow can reenable interrupts, then ac could change. */
3041 ac = cpu_cache_get(cachep);
3042 if (!x && ac->avail == 0) /* no objects in sight? abort */
3045 if (!ac->avail) /* objects refilled by interrupt? */
3049 return ac->entry[--ac->avail];
3052 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3055 might_sleep_if(flags & __GFP_WAIT);
3057 kmem_flagcheck(cachep, flags);
3062 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3063 gfp_t flags, void *objp, void *caller)
3067 if (cachep->flags & SLAB_POISON) {
3068 #ifdef CONFIG_DEBUG_PAGEALLOC
3069 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3070 kernel_map_pages(virt_to_page(objp),
3071 cachep->buffer_size / PAGE_SIZE, 1);
3073 check_poison_obj(cachep, objp);
3075 check_poison_obj(cachep, objp);
3077 poison_obj(cachep, objp, POISON_INUSE);
3079 if (cachep->flags & SLAB_STORE_USER)
3080 *dbg_userword(cachep, objp) = caller;
3082 if (cachep->flags & SLAB_RED_ZONE) {
3083 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3084 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3085 slab_error(cachep, "double free, or memory outside"
3086 " object was overwritten");
3088 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3089 objp, *dbg_redzone1(cachep, objp),
3090 *dbg_redzone2(cachep, objp));
3092 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3093 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3095 #ifdef CONFIG_DEBUG_SLAB_LEAK
3100 slabp = page_get_slab(virt_to_head_page(objp));
3101 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3102 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3105 objp += obj_offset(cachep);
3106 if (cachep->ctor && cachep->flags & SLAB_POISON)
3108 #if ARCH_SLAB_MINALIGN
3109 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3110 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3111 objp, ARCH_SLAB_MINALIGN);
3117 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3120 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3122 if (cachep == &cache_cache)
3125 return should_failslab(obj_size(cachep), flags);
3128 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3131 struct array_cache *ac;
3135 ac = cpu_cache_get(cachep);
3136 if (likely(ac->avail)) {
3137 STATS_INC_ALLOCHIT(cachep);
3139 objp = ac->entry[--ac->avail];
3141 STATS_INC_ALLOCMISS(cachep);
3142 objp = cache_alloc_refill(cachep, flags);
3149 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3151 * If we are in_interrupt, then process context, including cpusets and
3152 * mempolicy, may not apply and should not be used for allocation policy.
3154 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3156 int nid_alloc, nid_here;
3158 if (in_interrupt() || (flags & __GFP_THISNODE))
3160 nid_alloc = nid_here = numa_node_id();
3161 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3162 nid_alloc = cpuset_mem_spread_node();
3163 else if (current->mempolicy)
3164 nid_alloc = slab_node(current->mempolicy);
3165 if (nid_alloc != nid_here)
3166 return ____cache_alloc_node(cachep, flags, nid_alloc);
3171 * Fallback function if there was no memory available and no objects on a
3172 * certain node and fall back is permitted. First we scan all the
3173 * available nodelists for available objects. If that fails then we
3174 * perform an allocation without specifying a node. This allows the page
3175 * allocator to do its reclaim / fallback magic. We then insert the
3176 * slab into the proper nodelist and then allocate from it.
3178 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3180 struct zonelist *zonelist;
3184 enum zone_type high_zoneidx = gfp_zone(flags);
3188 if (flags & __GFP_THISNODE)
3191 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3192 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3196 * Look through allowed nodes for objects available
3197 * from existing per node queues.
3199 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3200 nid = zone_to_nid(zone);
3202 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3203 cache->nodelists[nid] &&
3204 cache->nodelists[nid]->free_objects) {
3205 obj = ____cache_alloc_node(cache,
3206 flags | GFP_THISNODE, nid);
3214 * This allocation will be performed within the constraints
3215 * of the current cpuset / memory policy requirements.
3216 * We may trigger various forms of reclaim on the allowed
3217 * set and go into memory reserves if necessary.
3219 if (local_flags & __GFP_WAIT)
3221 kmem_flagcheck(cache, flags);
3222 obj = kmem_getpages(cache, local_flags, -1);
3223 if (local_flags & __GFP_WAIT)
3224 local_irq_disable();
3227 * Insert into the appropriate per node queues
3229 nid = page_to_nid(virt_to_page(obj));
3230 if (cache_grow(cache, flags, nid, obj)) {
3231 obj = ____cache_alloc_node(cache,
3232 flags | GFP_THISNODE, nid);
3235 * Another processor may allocate the
3236 * objects in the slab since we are
3237 * not holding any locks.
3241 /* cache_grow already freed obj */
3250 * A interface to enable slab creation on nodeid
3252 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3255 struct list_head *entry;
3257 struct kmem_list3 *l3;
3261 l3 = cachep->nodelists[nodeid];
3266 spin_lock(&l3->list_lock);
3267 entry = l3->slabs_partial.next;
3268 if (entry == &l3->slabs_partial) {
3269 l3->free_touched = 1;
3270 entry = l3->slabs_free.next;
3271 if (entry == &l3->slabs_free)
3275 slabp = list_entry(entry, struct slab, list);
3276 check_spinlock_acquired_node(cachep, nodeid);
3277 check_slabp(cachep, slabp);
3279 STATS_INC_NODEALLOCS(cachep);
3280 STATS_INC_ACTIVE(cachep);
3281 STATS_SET_HIGH(cachep);
3283 BUG_ON(slabp->inuse == cachep->num);
3285 obj = slab_get_obj(cachep, slabp, nodeid);
3286 check_slabp(cachep, slabp);
3288 /* move slabp to correct slabp list: */
3289 list_del(&slabp->list);
3291 if (slabp->free == BUFCTL_END)
3292 list_add(&slabp->list, &l3->slabs_full);
3294 list_add(&slabp->list, &l3->slabs_partial);
3296 spin_unlock(&l3->list_lock);
3300 spin_unlock(&l3->list_lock);
3301 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3305 return fallback_alloc(cachep, flags);
3312 * kmem_cache_alloc_node - Allocate an object on the specified node
3313 * @cachep: The cache to allocate from.
3314 * @flags: See kmalloc().
3315 * @nodeid: node number of the target node.
3316 * @caller: return address of caller, used for debug information
3318 * Identical to kmem_cache_alloc but it will allocate memory on the given
3319 * node, which can improve the performance for cpu bound structures.
3321 * Fallback to other node is possible if __GFP_THISNODE is not set.
3323 static __always_inline void *
3324 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3327 unsigned long save_flags;
3330 lockdep_trace_alloc(flags);
3332 if (slab_should_failslab(cachep, flags))
3335 cache_alloc_debugcheck_before(cachep, flags);
3336 local_irq_save(save_flags);
3338 if (unlikely(nodeid == -1))
3339 nodeid = numa_node_id();
3341 if (unlikely(!cachep->nodelists[nodeid])) {
3342 /* Node not bootstrapped yet */
3343 ptr = fallback_alloc(cachep, flags);
3347 if (nodeid == numa_node_id()) {
3349 * Use the locally cached objects if possible.
3350 * However ____cache_alloc does not allow fallback
3351 * to other nodes. It may fail while we still have
3352 * objects on other nodes available.
3354 ptr = ____cache_alloc(cachep, flags);
3358 /* ___cache_alloc_node can fall back to other nodes */
3359 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3361 local_irq_restore(save_flags);
3362 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3364 if (unlikely((flags & __GFP_ZERO) && ptr))
3365 memset(ptr, 0, obj_size(cachep));
3370 static __always_inline void *
3371 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3375 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3376 objp = alternate_node_alloc(cache, flags);
3380 objp = ____cache_alloc(cache, flags);
3383 * We may just have run out of memory on the local node.
3384 * ____cache_alloc_node() knows how to locate memory on other nodes
3387 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3394 static __always_inline void *
3395 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3397 return ____cache_alloc(cachep, flags);
3400 #endif /* CONFIG_NUMA */
3402 static __always_inline void *
3403 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3405 unsigned long save_flags;
3408 lockdep_trace_alloc(flags);
3410 if (slab_should_failslab(cachep, flags))
3413 cache_alloc_debugcheck_before(cachep, flags);
3414 local_irq_save(save_flags);
3415 objp = __do_cache_alloc(cachep, flags);
3416 local_irq_restore(save_flags);
3417 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3420 if (unlikely((flags & __GFP_ZERO) && objp))
3421 memset(objp, 0, obj_size(cachep));
3427 * Caller needs to acquire correct kmem_list's list_lock
3429 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3433 struct kmem_list3 *l3;
3435 for (i = 0; i < nr_objects; i++) {
3436 void *objp = objpp[i];
3439 slabp = virt_to_slab(objp);
3440 l3 = cachep->nodelists[node];
3441 list_del(&slabp->list);
3442 check_spinlock_acquired_node(cachep, node);
3443 check_slabp(cachep, slabp);
3444 slab_put_obj(cachep, slabp, objp, node);
3445 STATS_DEC_ACTIVE(cachep);
3447 check_slabp(cachep, slabp);
3449 /* fixup slab chains */
3450 if (slabp->inuse == 0) {
3451 if (l3->free_objects > l3->free_limit) {
3452 l3->free_objects -= cachep->num;
3453 /* No need to drop any previously held
3454 * lock here, even if we have a off-slab slab
3455 * descriptor it is guaranteed to come from
3456 * a different cache, refer to comments before
3459 slab_destroy(cachep, slabp);
3461 list_add(&slabp->list, &l3->slabs_free);
3464 /* Unconditionally move a slab to the end of the
3465 * partial list on free - maximum time for the
3466 * other objects to be freed, too.
3468 list_add_tail(&slabp->list, &l3->slabs_partial);
3473 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3476 struct kmem_list3 *l3;
3477 int node = numa_node_id();
3479 batchcount = ac->batchcount;
3481 BUG_ON(!batchcount || batchcount > ac->avail);
3484 l3 = cachep->nodelists[node];
3485 spin_lock(&l3->list_lock);
3487 struct array_cache *shared_array = l3->shared;
3488 int max = shared_array->limit - shared_array->avail;
3490 if (batchcount > max)
3492 memcpy(&(shared_array->entry[shared_array->avail]),
3493 ac->entry, sizeof(void *) * batchcount);
3494 shared_array->avail += batchcount;
3499 free_block(cachep, ac->entry, batchcount, node);
3504 struct list_head *p;
3506 p = l3->slabs_free.next;
3507 while (p != &(l3->slabs_free)) {
3510 slabp = list_entry(p, struct slab, list);
3511 BUG_ON(slabp->inuse);
3516 STATS_SET_FREEABLE(cachep, i);
3519 spin_unlock(&l3->list_lock);
3520 ac->avail -= batchcount;
3521 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3525 * Release an obj back to its cache. If the obj has a constructed state, it must
3526 * be in this state _before_ it is released. Called with disabled ints.
3528 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3530 struct array_cache *ac = cpu_cache_get(cachep);
3533 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3536 * Skip calling cache_free_alien() when the platform is not numa.
3537 * This will avoid cache misses that happen while accessing slabp (which
3538 * is per page memory reference) to get nodeid. Instead use a global
3539 * variable to skip the call, which is mostly likely to be present in
3542 if (numa_platform && cache_free_alien(cachep, objp))
3545 if (likely(ac->avail < ac->limit)) {
3546 STATS_INC_FREEHIT(cachep);
3547 ac->entry[ac->avail++] = objp;
3550 STATS_INC_FREEMISS(cachep);
3551 cache_flusharray(cachep, ac);
3552 ac->entry[ac->avail++] = objp;
3557 * kmem_cache_alloc - Allocate an object
3558 * @cachep: The cache to allocate from.
3559 * @flags: See kmalloc().
3561 * Allocate an object from this cache. The flags are only relevant
3562 * if the cache has no available objects.
3564 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3566 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3568 trace_kmem_cache_alloc(_RET_IP_, ret,
3569 obj_size(cachep), cachep->buffer_size, flags);
3573 EXPORT_SYMBOL(kmem_cache_alloc);
3575 #ifdef CONFIG_KMEMTRACE
3576 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3578 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3580 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3584 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3585 * @cachep: the cache we're checking against
3586 * @ptr: pointer to validate
3588 * This verifies that the untrusted pointer looks sane;
3589 * it is _not_ a guarantee that the pointer is actually
3590 * part of the slab cache in question, but it at least
3591 * validates that the pointer can be dereferenced and
3592 * looks half-way sane.
3594 * Currently only used for dentry validation.
3596 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3598 unsigned long addr = (unsigned long)ptr;
3599 unsigned long min_addr = PAGE_OFFSET;
3600 unsigned long align_mask = BYTES_PER_WORD - 1;
3601 unsigned long size = cachep->buffer_size;
3604 if (unlikely(addr < min_addr))
3606 if (unlikely(addr > (unsigned long)high_memory - size))
3608 if (unlikely(addr & align_mask))
3610 if (unlikely(!kern_addr_valid(addr)))
3612 if (unlikely(!kern_addr_valid(addr + size - 1)))
3614 page = virt_to_page(ptr);
3615 if (unlikely(!PageSlab(page)))
3617 if (unlikely(page_get_cache(page) != cachep))
3625 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3627 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3628 __builtin_return_address(0));
3630 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3631 obj_size(cachep), cachep->buffer_size,
3636 EXPORT_SYMBOL(kmem_cache_alloc_node);
3638 #ifdef CONFIG_KMEMTRACE
3639 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3643 return __cache_alloc_node(cachep, flags, nodeid,
3644 __builtin_return_address(0));
3646 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3649 static __always_inline void *
3650 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3652 struct kmem_cache *cachep;
3655 cachep = kmem_find_general_cachep(size, flags);
3656 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3658 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3660 trace_kmalloc_node((unsigned long) caller, ret,
3661 size, cachep->buffer_size, flags, node);
3666 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3667 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3669 return __do_kmalloc_node(size, flags, node,
3670 __builtin_return_address(0));
3672 EXPORT_SYMBOL(__kmalloc_node);
3674 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3675 int node, unsigned long caller)
3677 return __do_kmalloc_node(size, flags, node, (void *)caller);
3679 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3681 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3683 return __do_kmalloc_node(size, flags, node, NULL);
3685 EXPORT_SYMBOL(__kmalloc_node);
3686 #endif /* CONFIG_DEBUG_SLAB */
3687 #endif /* CONFIG_NUMA */
3690 * __do_kmalloc - allocate memory
3691 * @size: how many bytes of memory are required.
3692 * @flags: the type of memory to allocate (see kmalloc).
3693 * @caller: function caller for debug tracking of the caller
3695 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3698 struct kmem_cache *cachep;
3701 /* If you want to save a few bytes .text space: replace
3703 * Then kmalloc uses the uninlined functions instead of the inline
3706 cachep = __find_general_cachep(size, flags);
3707 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3709 ret = __cache_alloc(cachep, flags, caller);
3711 trace_kmalloc((unsigned long) caller, ret,
3712 size, cachep->buffer_size, flags);
3718 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3719 void *__kmalloc(size_t size, gfp_t flags)
3721 return __do_kmalloc(size, flags, __builtin_return_address(0));
3723 EXPORT_SYMBOL(__kmalloc);
3725 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3727 return __do_kmalloc(size, flags, (void *)caller);
3729 EXPORT_SYMBOL(__kmalloc_track_caller);
3732 void *__kmalloc(size_t size, gfp_t flags)
3734 return __do_kmalloc(size, flags, NULL);
3736 EXPORT_SYMBOL(__kmalloc);
3740 * kmem_cache_free - Deallocate an object
3741 * @cachep: The cache the allocation was from.
3742 * @objp: The previously allocated object.
3744 * Free an object which was previously allocated from this
3747 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3749 unsigned long flags;
3751 local_irq_save(flags);
3752 debug_check_no_locks_freed(objp, obj_size(cachep));
3753 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3754 debug_check_no_obj_freed(objp, obj_size(cachep));
3755 __cache_free(cachep, objp);
3756 local_irq_restore(flags);
3758 trace_kmem_cache_free(_RET_IP_, objp);
3760 EXPORT_SYMBOL(kmem_cache_free);
3763 * kfree - free previously allocated memory
3764 * @objp: pointer returned by kmalloc.
3766 * If @objp is NULL, no operation is performed.
3768 * Don't free memory not originally allocated by kmalloc()
3769 * or you will run into trouble.
3771 void kfree(const void *objp)
3773 struct kmem_cache *c;
3774 unsigned long flags;
3776 trace_kfree(_RET_IP_, objp);
3778 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3780 local_irq_save(flags);
3781 kfree_debugcheck(objp);
3782 c = virt_to_cache(objp);
3783 debug_check_no_locks_freed(objp, obj_size(c));
3784 debug_check_no_obj_freed(objp, obj_size(c));
3785 __cache_free(c, (void *)objp);
3786 local_irq_restore(flags);
3788 EXPORT_SYMBOL(kfree);
3790 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3792 return obj_size(cachep);
3794 EXPORT_SYMBOL(kmem_cache_size);
3796 const char *kmem_cache_name(struct kmem_cache *cachep)
3798 return cachep->name;
3800 EXPORT_SYMBOL_GPL(kmem_cache_name);
3803 * This initializes kmem_list3 or resizes various caches for all nodes.
3805 static int alloc_kmemlist(struct kmem_cache *cachep)
3808 struct kmem_list3 *l3;
3809 struct array_cache *new_shared;
3810 struct array_cache **new_alien = NULL;
3812 for_each_online_node(node) {
3814 if (use_alien_caches) {
3815 new_alien = alloc_alien_cache(node, cachep->limit);
3821 if (cachep->shared) {
3822 new_shared = alloc_arraycache(node,
3823 cachep->shared*cachep->batchcount,
3826 free_alien_cache(new_alien);
3831 l3 = cachep->nodelists[node];
3833 struct array_cache *shared = l3->shared;
3835 spin_lock_irq(&l3->list_lock);
3838 free_block(cachep, shared->entry,
3839 shared->avail, node);
3841 l3->shared = new_shared;
3843 l3->alien = new_alien;
3846 l3->free_limit = (1 + nr_cpus_node(node)) *
3847 cachep->batchcount + cachep->num;
3848 spin_unlock_irq(&l3->list_lock);
3850 free_alien_cache(new_alien);
3853 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3855 free_alien_cache(new_alien);
3860 kmem_list3_init(l3);
3861 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3862 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3863 l3->shared = new_shared;
3864 l3->alien = new_alien;
3865 l3->free_limit = (1 + nr_cpus_node(node)) *
3866 cachep->batchcount + cachep->num;
3867 cachep->nodelists[node] = l3;
3872 if (!cachep->next.next) {
3873 /* Cache is not active yet. Roll back what we did */
3876 if (cachep->nodelists[node]) {
3877 l3 = cachep->nodelists[node];
3880 free_alien_cache(l3->alien);
3882 cachep->nodelists[node] = NULL;
3890 struct ccupdate_struct {
3891 struct kmem_cache *cachep;
3892 struct array_cache *new[NR_CPUS];
3895 static void do_ccupdate_local(void *info)
3897 struct ccupdate_struct *new = info;
3898 struct array_cache *old;
3901 old = cpu_cache_get(new->cachep);
3903 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3904 new->new[smp_processor_id()] = old;
3907 /* Always called with the cache_chain_mutex held */
3908 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3909 int batchcount, int shared)
3911 struct ccupdate_struct *new;
3914 new = kzalloc(sizeof(*new), GFP_KERNEL);
3918 for_each_online_cpu(i) {
3919 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3922 for (i--; i >= 0; i--)
3928 new->cachep = cachep;
3930 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3933 cachep->batchcount = batchcount;
3934 cachep->limit = limit;
3935 cachep->shared = shared;
3937 for_each_online_cpu(i) {
3938 struct array_cache *ccold = new->new[i];
3941 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3942 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3943 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3947 return alloc_kmemlist(cachep);
3950 /* Called with cache_chain_mutex held always */
3951 static int enable_cpucache(struct kmem_cache *cachep)
3957 * The head array serves three purposes:
3958 * - create a LIFO ordering, i.e. return objects that are cache-warm
3959 * - reduce the number of spinlock operations.
3960 * - reduce the number of linked list operations on the slab and
3961 * bufctl chains: array operations are cheaper.
3962 * The numbers are guessed, we should auto-tune as described by
3965 if (cachep->buffer_size > 131072)
3967 else if (cachep->buffer_size > PAGE_SIZE)
3969 else if (cachep->buffer_size > 1024)
3971 else if (cachep->buffer_size > 256)
3977 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3978 * allocation behaviour: Most allocs on one cpu, most free operations
3979 * on another cpu. For these cases, an efficient object passing between
3980 * cpus is necessary. This is provided by a shared array. The array
3981 * replaces Bonwick's magazine layer.
3982 * On uniprocessor, it's functionally equivalent (but less efficient)
3983 * to a larger limit. Thus disabled by default.
3986 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3991 * With debugging enabled, large batchcount lead to excessively long
3992 * periods with disabled local interrupts. Limit the batchcount
3997 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3999 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4000 cachep->name, -err);
4005 * Drain an array if it contains any elements taking the l3 lock only if
4006 * necessary. Note that the l3 listlock also protects the array_cache
4007 * if drain_array() is used on the shared array.
4009 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4010 struct array_cache *ac, int force, int node)
4014 if (!ac || !ac->avail)
4016 if (ac->touched && !force) {
4019 spin_lock_irq(&l3->list_lock);
4021 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4022 if (tofree > ac->avail)
4023 tofree = (ac->avail + 1) / 2;
4024 free_block(cachep, ac->entry, tofree, node);
4025 ac->avail -= tofree;
4026 memmove(ac->entry, &(ac->entry[tofree]),
4027 sizeof(void *) * ac->avail);
4029 spin_unlock_irq(&l3->list_lock);
4034 * cache_reap - Reclaim memory from caches.
4035 * @w: work descriptor
4037 * Called from workqueue/eventd every few seconds.
4039 * - clear the per-cpu caches for this CPU.
4040 * - return freeable pages to the main free memory pool.
4042 * If we cannot acquire the cache chain mutex then just give up - we'll try
4043 * again on the next iteration.
4045 static void cache_reap(struct work_struct *w)
4047 struct kmem_cache *searchp;
4048 struct kmem_list3 *l3;
4049 int node = numa_node_id();
4050 struct delayed_work *work = to_delayed_work(w);
4052 if (!mutex_trylock(&cache_chain_mutex))
4053 /* Give up. Setup the next iteration. */
4056 list_for_each_entry(searchp, &cache_chain, next) {
4060 * We only take the l3 lock if absolutely necessary and we
4061 * have established with reasonable certainty that
4062 * we can do some work if the lock was obtained.
4064 l3 = searchp->nodelists[node];
4066 reap_alien(searchp, l3);
4068 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4071 * These are racy checks but it does not matter
4072 * if we skip one check or scan twice.
4074 if (time_after(l3->next_reap, jiffies))
4077 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4079 drain_array(searchp, l3, l3->shared, 0, node);
4081 if (l3->free_touched)
4082 l3->free_touched = 0;
4086 freed = drain_freelist(searchp, l3, (l3->free_limit +
4087 5 * searchp->num - 1) / (5 * searchp->num));
4088 STATS_ADD_REAPED(searchp, freed);
4094 mutex_unlock(&cache_chain_mutex);
4097 /* Set up the next iteration */
4098 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4101 #ifdef CONFIG_SLABINFO
4103 static void print_slabinfo_header(struct seq_file *m)
4106 * Output format version, so at least we can change it
4107 * without _too_ many complaints.
4110 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4112 seq_puts(m, "slabinfo - version: 2.1\n");
4114 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4115 "<objperslab> <pagesperslab>");
4116 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4117 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4119 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4120 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4121 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4126 static void *s_start(struct seq_file *m, loff_t *pos)
4130 mutex_lock(&cache_chain_mutex);
4132 print_slabinfo_header(m);
4134 return seq_list_start(&cache_chain, *pos);
4137 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4139 return seq_list_next(p, &cache_chain, pos);
4142 static void s_stop(struct seq_file *m, void *p)
4144 mutex_unlock(&cache_chain_mutex);
4147 static int s_show(struct seq_file *m, void *p)
4149 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4151 unsigned long active_objs;
4152 unsigned long num_objs;
4153 unsigned long active_slabs = 0;
4154 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4158 struct kmem_list3 *l3;
4162 for_each_online_node(node) {
4163 l3 = cachep->nodelists[node];
4168 spin_lock_irq(&l3->list_lock);
4170 list_for_each_entry(slabp, &l3->slabs_full, list) {
4171 if (slabp->inuse != cachep->num && !error)
4172 error = "slabs_full accounting error";
4173 active_objs += cachep->num;
4176 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4177 if (slabp->inuse == cachep->num && !error)
4178 error = "slabs_partial inuse accounting error";
4179 if (!slabp->inuse && !error)
4180 error = "slabs_partial/inuse accounting error";
4181 active_objs += slabp->inuse;
4184 list_for_each_entry(slabp, &l3->slabs_free, list) {
4185 if (slabp->inuse && !error)
4186 error = "slabs_free/inuse accounting error";
4189 free_objects += l3->free_objects;
4191 shared_avail += l3->shared->avail;
4193 spin_unlock_irq(&l3->list_lock);
4195 num_slabs += active_slabs;
4196 num_objs = num_slabs * cachep->num;
4197 if (num_objs - active_objs != free_objects && !error)
4198 error = "free_objects accounting error";
4200 name = cachep->name;
4202 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4204 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4205 name, active_objs, num_objs, cachep->buffer_size,
4206 cachep->num, (1 << cachep->gfporder));
4207 seq_printf(m, " : tunables %4u %4u %4u",
4208 cachep->limit, cachep->batchcount, cachep->shared);
4209 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4210 active_slabs, num_slabs, shared_avail);
4213 unsigned long high = cachep->high_mark;
4214 unsigned long allocs = cachep->num_allocations;
4215 unsigned long grown = cachep->grown;
4216 unsigned long reaped = cachep->reaped;
4217 unsigned long errors = cachep->errors;
4218 unsigned long max_freeable = cachep->max_freeable;
4219 unsigned long node_allocs = cachep->node_allocs;
4220 unsigned long node_frees = cachep->node_frees;
4221 unsigned long overflows = cachep->node_overflow;
4223 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4224 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4225 reaped, errors, max_freeable, node_allocs,
4226 node_frees, overflows);
4230 unsigned long allochit = atomic_read(&cachep->allochit);
4231 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4232 unsigned long freehit = atomic_read(&cachep->freehit);
4233 unsigned long freemiss = atomic_read(&cachep->freemiss);
4235 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4236 allochit, allocmiss, freehit, freemiss);
4244 * slabinfo_op - iterator that generates /proc/slabinfo
4253 * num-pages-per-slab
4254 * + further values on SMP and with statistics enabled
4257 static const struct seq_operations slabinfo_op = {
4264 #define MAX_SLABINFO_WRITE 128
4266 * slabinfo_write - Tuning for the slab allocator
4268 * @buffer: user buffer
4269 * @count: data length
4272 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4273 size_t count, loff_t *ppos)
4275 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4276 int limit, batchcount, shared, res;
4277 struct kmem_cache *cachep;
4279 if (count > MAX_SLABINFO_WRITE)
4281 if (copy_from_user(&kbuf, buffer, count))
4283 kbuf[MAX_SLABINFO_WRITE] = '\0';
4285 tmp = strchr(kbuf, ' ');
4290 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4293 /* Find the cache in the chain of caches. */
4294 mutex_lock(&cache_chain_mutex);
4296 list_for_each_entry(cachep, &cache_chain, next) {
4297 if (!strcmp(cachep->name, kbuf)) {
4298 if (limit < 1 || batchcount < 1 ||
4299 batchcount > limit || shared < 0) {
4302 res = do_tune_cpucache(cachep, limit,
4303 batchcount, shared);
4308 mutex_unlock(&cache_chain_mutex);
4314 static int slabinfo_open(struct inode *inode, struct file *file)
4316 return seq_open(file, &slabinfo_op);
4319 static const struct file_operations proc_slabinfo_operations = {
4320 .open = slabinfo_open,
4322 .write = slabinfo_write,
4323 .llseek = seq_lseek,
4324 .release = seq_release,
4327 #ifdef CONFIG_DEBUG_SLAB_LEAK
4329 static void *leaks_start(struct seq_file *m, loff_t *pos)
4331 mutex_lock(&cache_chain_mutex);
4332 return seq_list_start(&cache_chain, *pos);
4335 static inline int add_caller(unsigned long *n, unsigned long v)
4345 unsigned long *q = p + 2 * i;
4359 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4365 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4371 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4372 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4374 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4379 static void show_symbol(struct seq_file *m, unsigned long address)
4381 #ifdef CONFIG_KALLSYMS
4382 unsigned long offset, size;
4383 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4385 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4386 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4388 seq_printf(m, " [%s]", modname);
4392 seq_printf(m, "%p", (void *)address);
4395 static int leaks_show(struct seq_file *m, void *p)
4397 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4399 struct kmem_list3 *l3;
4401 unsigned long *n = m->private;
4405 if (!(cachep->flags & SLAB_STORE_USER))
4407 if (!(cachep->flags & SLAB_RED_ZONE))
4410 /* OK, we can do it */
4414 for_each_online_node(node) {
4415 l3 = cachep->nodelists[node];
4420 spin_lock_irq(&l3->list_lock);
4422 list_for_each_entry(slabp, &l3->slabs_full, list)
4423 handle_slab(n, cachep, slabp);
4424 list_for_each_entry(slabp, &l3->slabs_partial, list)
4425 handle_slab(n, cachep, slabp);
4426 spin_unlock_irq(&l3->list_lock);
4428 name = cachep->name;
4430 /* Increase the buffer size */
4431 mutex_unlock(&cache_chain_mutex);
4432 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4434 /* Too bad, we are really out */
4436 mutex_lock(&cache_chain_mutex);
4439 *(unsigned long *)m->private = n[0] * 2;
4441 mutex_lock(&cache_chain_mutex);
4442 /* Now make sure this entry will be retried */
4446 for (i = 0; i < n[1]; i++) {
4447 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4448 show_symbol(m, n[2*i+2]);
4455 static const struct seq_operations slabstats_op = {
4456 .start = leaks_start,
4462 static int slabstats_open(struct inode *inode, struct file *file)
4464 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4467 ret = seq_open(file, &slabstats_op);
4469 struct seq_file *m = file->private_data;
4470 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4479 static const struct file_operations proc_slabstats_operations = {
4480 .open = slabstats_open,
4482 .llseek = seq_lseek,
4483 .release = seq_release_private,
4487 static int __init slab_proc_init(void)
4489 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4490 #ifdef CONFIG_DEBUG_SLAB_LEAK
4491 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4495 module_init(slab_proc_init);
4499 * ksize - get the actual amount of memory allocated for a given object
4500 * @objp: Pointer to the object
4502 * kmalloc may internally round up allocations and return more memory
4503 * than requested. ksize() can be used to determine the actual amount of
4504 * memory allocated. The caller may use this additional memory, even though
4505 * a smaller amount of memory was initially specified with the kmalloc call.
4506 * The caller must guarantee that objp points to a valid object previously
4507 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4508 * must not be freed during the duration of the call.
4510 size_t ksize(const void *objp)
4513 if (unlikely(objp == ZERO_SIZE_PTR))
4516 return obj_size(virt_to_cache(objp));
4518 EXPORT_SYMBOL(ksize);