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 <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.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>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <asm/cacheflush.h>
121 #include <asm/tlbflush.h>
122 #include <asm/page.h>
124 #include <trace/events/kmem.h>
127 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * STATS - 1 to collect stats for /proc/slabinfo.
131 * 0 for faster, smaller code (especially in the critical paths).
133 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
136 #ifdef CONFIG_DEBUG_SLAB
139 #define FORCED_DEBUG 1
143 #define FORCED_DEBUG 0
146 /* Shouldn't this be in a header file somewhere? */
147 #define BYTES_PER_WORD sizeof(void *)
148 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
150 #ifndef ARCH_KMALLOC_FLAGS
151 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
154 /* Legal flag mask for kmem_cache_create(). */
156 # define CREATE_MASK (SLAB_RED_ZONE | \
157 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
160 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
161 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
162 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
164 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
166 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
167 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
168 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
174 * Bufctl's are used for linking objs within a slab
177 * This implementation relies on "struct page" for locating the cache &
178 * slab an object belongs to.
179 * This allows the bufctl structure to be small (one int), but limits
180 * the number of objects a slab (not a cache) can contain when off-slab
181 * bufctls are used. The limit is the size of the largest general cache
182 * that does not use off-slab slabs.
183 * For 32bit archs with 4 kB pages, is this 56.
184 * This is not serious, as it is only for large objects, when it is unwise
185 * to have too many per slab.
186 * Note: This limit can be raised by introducing a general cache whose size
187 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
190 typedef unsigned int kmem_bufctl_t;
191 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
192 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
193 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
194 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
199 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
200 * arrange for kmem_freepages to be called via RCU. This is useful if
201 * we need to approach a kernel structure obliquely, from its address
202 * obtained without the usual locking. We can lock the structure to
203 * stabilize it and check it's still at the given address, only if we
204 * can be sure that the memory has not been meanwhile reused for some
205 * other kind of object (which our subsystem's lock might corrupt).
207 * rcu_read_lock before reading the address, then rcu_read_unlock after
208 * taking the spinlock within the structure expected at that address.
211 struct rcu_head head;
212 struct kmem_cache *cachep;
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
226 struct list_head list;
227 unsigned long colouroff;
228 void *s_mem; /* including colour offset */
229 unsigned int inuse; /* num of objs active in slab */
231 unsigned short nodeid;
233 struct slab_rcu __slab_cover_slab_rcu;
241 * - LIFO ordering, to hand out cache-warm objects from _alloc
242 * - reduce the number of linked list operations
243 * - reduce spinlock operations
245 * The limit is stored in the per-cpu structure to reduce the data cache
252 unsigned int batchcount;
253 unsigned int touched;
256 * Must have this definition in here for the proper
257 * alignment of array_cache. Also simplifies accessing
263 * bootstrap: The caches do not work without cpuarrays anymore, but the
264 * cpuarrays are allocated from the generic caches...
266 #define BOOT_CPUCACHE_ENTRIES 1
267 struct arraycache_init {
268 struct array_cache cache;
269 void *entries[BOOT_CPUCACHE_ENTRIES];
273 * The slab lists for all objects.
276 struct list_head slabs_partial; /* partial list first, better asm code */
277 struct list_head slabs_full;
278 struct list_head slabs_free;
279 unsigned long free_objects;
280 unsigned int free_limit;
281 unsigned int colour_next; /* Per-node cache coloring */
282 spinlock_t list_lock;
283 struct array_cache *shared; /* shared per node */
284 struct array_cache **alien; /* on other nodes */
285 unsigned long next_reap; /* updated without locking */
286 int free_touched; /* updated without locking */
290 * Need this for bootstrapping a per node allocator.
292 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
293 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
294 #define CACHE_CACHE 0
295 #define SIZE_AC MAX_NUMNODES
296 #define SIZE_L3 (2 * MAX_NUMNODES)
298 static int drain_freelist(struct kmem_cache *cache,
299 struct kmem_list3 *l3, int tofree);
300 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
302 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
303 static void cache_reap(struct work_struct *unused);
306 * This function must be completely optimized away if a constant is passed to
307 * it. Mostly the same as what is in linux/slab.h except it returns an index.
309 static __always_inline int index_of(const size_t size)
311 extern void __bad_size(void);
313 if (__builtin_constant_p(size)) {
321 #include <linux/kmalloc_sizes.h>
329 static int slab_early_init = 1;
331 #define INDEX_AC index_of(sizeof(struct arraycache_init))
332 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
334 static void kmem_list3_init(struct kmem_list3 *parent)
336 INIT_LIST_HEAD(&parent->slabs_full);
337 INIT_LIST_HEAD(&parent->slabs_partial);
338 INIT_LIST_HEAD(&parent->slabs_free);
339 parent->shared = NULL;
340 parent->alien = NULL;
341 parent->colour_next = 0;
342 spin_lock_init(&parent->list_lock);
343 parent->free_objects = 0;
344 parent->free_touched = 0;
347 #define MAKE_LIST(cachep, listp, slab, nodeid) \
349 INIT_LIST_HEAD(listp); \
350 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
353 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
355 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
356 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
360 #define CFLGS_OFF_SLAB (0x80000000UL)
361 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
363 #define BATCHREFILL_LIMIT 16
365 * Optimization question: fewer reaps means less probability for unnessary
366 * cpucache drain/refill cycles.
368 * OTOH the cpuarrays can contain lots of objects,
369 * which could lock up otherwise freeable slabs.
371 #define REAPTIMEOUT_CPUC (2*HZ)
372 #define REAPTIMEOUT_LIST3 (4*HZ)
375 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
376 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378 #define STATS_INC_GROWN(x) ((x)->grown++)
379 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
380 #define STATS_SET_HIGH(x) \
382 if ((x)->num_active > (x)->high_mark) \
383 (x)->high_mark = (x)->num_active; \
385 #define STATS_INC_ERR(x) ((x)->errors++)
386 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
387 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
388 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
389 #define STATS_SET_FREEABLE(x, i) \
391 if ((x)->max_freeable < i) \
392 (x)->max_freeable = i; \
394 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
395 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
396 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
397 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
399 #define STATS_INC_ACTIVE(x) do { } while (0)
400 #define STATS_DEC_ACTIVE(x) do { } while (0)
401 #define STATS_INC_ALLOCED(x) do { } while (0)
402 #define STATS_INC_GROWN(x) do { } while (0)
403 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
404 #define STATS_SET_HIGH(x) do { } while (0)
405 #define STATS_INC_ERR(x) do { } while (0)
406 #define STATS_INC_NODEALLOCS(x) do { } while (0)
407 #define STATS_INC_NODEFREES(x) do { } while (0)
408 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
409 #define STATS_SET_FREEABLE(x, i) do { } while (0)
410 #define STATS_INC_ALLOCHIT(x) do { } while (0)
411 #define STATS_INC_ALLOCMISS(x) do { } while (0)
412 #define STATS_INC_FREEHIT(x) do { } while (0)
413 #define STATS_INC_FREEMISS(x) do { } while (0)
419 * memory layout of objects:
421 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
422 * the end of an object is aligned with the end of the real
423 * allocation. Catches writes behind the end of the allocation.
424 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
426 * cachep->obj_offset: The real object.
427 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
428 * cachep->size - 1* BYTES_PER_WORD: last caller address
429 * [BYTES_PER_WORD long]
431 static int obj_offset(struct kmem_cache *cachep)
433 return cachep->obj_offset;
436 static int obj_size(struct kmem_cache *cachep)
438 return cachep->object_size;
441 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
443 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
444 return (unsigned long long*) (objp + obj_offset(cachep) -
445 sizeof(unsigned long long));
448 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
450 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
451 if (cachep->flags & SLAB_STORE_USER)
452 return (unsigned long long *)(objp + cachep->size -
453 sizeof(unsigned long long) -
455 return (unsigned long long *) (objp + cachep->size -
456 sizeof(unsigned long long));
459 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
461 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
462 return (void **)(objp + cachep->size - BYTES_PER_WORD);
467 #define obj_offset(x) 0
468 #define obj_size(cachep) (cachep->size)
469 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
470 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
471 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
475 #ifdef CONFIG_TRACING
476 size_t slab_buffer_size(struct kmem_cache *cachep)
480 EXPORT_SYMBOL(slab_buffer_size);
484 * Do not go above this order unless 0 objects fit into the slab or
485 * overridden on the command line.
487 #define SLAB_MAX_ORDER_HI 1
488 #define SLAB_MAX_ORDER_LO 0
489 static int slab_max_order = SLAB_MAX_ORDER_LO;
490 static bool slab_max_order_set __initdata;
492 static inline struct kmem_cache *page_get_cache(struct page *page)
494 page = compound_head(page);
495 BUG_ON(!PageSlab(page));
496 return page->slab_cache;
499 static inline struct kmem_cache *virt_to_cache(const void *obj)
501 struct page *page = virt_to_head_page(obj);
502 return page->slab_cache;
505 static inline struct slab *virt_to_slab(const void *obj)
507 struct page *page = virt_to_head_page(obj);
509 VM_BUG_ON(!PageSlab(page));
510 return page->slab_page;
513 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
516 return slab->s_mem + cache->size * idx;
520 * We want to avoid an expensive divide : (offset / cache->size)
521 * Using the fact that size is a constant for a particular cache,
522 * we can replace (offset / cache->size) by
523 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
525 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
526 const struct slab *slab, void *obj)
528 u32 offset = (obj - slab->s_mem);
529 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
533 * These are the default caches for kmalloc. Custom caches can have other sizes.
535 struct cache_sizes malloc_sizes[] = {
536 #define CACHE(x) { .cs_size = (x) },
537 #include <linux/kmalloc_sizes.h>
541 EXPORT_SYMBOL(malloc_sizes);
543 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
549 static struct cache_names __initdata cache_names[] = {
550 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
551 #include <linux/kmalloc_sizes.h>
556 static struct arraycache_init initarray_cache __initdata =
557 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
558 static struct arraycache_init initarray_generic =
559 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
561 /* internal cache of cache description objs */
562 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
563 static struct kmem_cache cache_cache = {
564 .nodelists = cache_cache_nodelists,
566 .limit = BOOT_CPUCACHE_ENTRIES,
568 .size = sizeof(struct kmem_cache),
569 .name = "kmem_cache",
572 #define BAD_ALIEN_MAGIC 0x01020304ul
575 * chicken and egg problem: delay the per-cpu array allocation
576 * until the general caches are up.
588 * used by boot code to determine if it can use slab based allocator
590 int slab_is_available(void)
592 return g_cpucache_up >= EARLY;
595 #ifdef CONFIG_LOCKDEP
598 * Slab sometimes uses the kmalloc slabs to store the slab headers
599 * for other slabs "off slab".
600 * The locking for this is tricky in that it nests within the locks
601 * of all other slabs in a few places; to deal with this special
602 * locking we put on-slab caches into a separate lock-class.
604 * We set lock class for alien array caches which are up during init.
605 * The lock annotation will be lost if all cpus of a node goes down and
606 * then comes back up during hotplug
608 static struct lock_class_key on_slab_l3_key;
609 static struct lock_class_key on_slab_alc_key;
611 static struct lock_class_key debugobj_l3_key;
612 static struct lock_class_key debugobj_alc_key;
614 static void slab_set_lock_classes(struct kmem_cache *cachep,
615 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
618 struct array_cache **alc;
619 struct kmem_list3 *l3;
622 l3 = cachep->nodelists[q];
626 lockdep_set_class(&l3->list_lock, l3_key);
629 * FIXME: This check for BAD_ALIEN_MAGIC
630 * should go away when common slab code is taught to
631 * work even without alien caches.
632 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
633 * for alloc_alien_cache,
635 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
639 lockdep_set_class(&alc[r]->lock, alc_key);
643 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
645 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
648 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
652 for_each_online_node(node)
653 slab_set_debugobj_lock_classes_node(cachep, node);
656 static void init_node_lock_keys(int q)
658 struct cache_sizes *s = malloc_sizes;
660 if (g_cpucache_up < LATE)
663 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
664 struct kmem_list3 *l3;
666 l3 = s->cs_cachep->nodelists[q];
667 if (!l3 || OFF_SLAB(s->cs_cachep))
670 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
671 &on_slab_alc_key, q);
675 static inline void init_lock_keys(void)
680 init_node_lock_keys(node);
683 static void init_node_lock_keys(int q)
687 static inline void init_lock_keys(void)
691 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
695 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
701 * Guard access to the cache-chain.
703 static DEFINE_MUTEX(cache_chain_mutex);
704 static struct list_head cache_chain;
706 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
708 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
710 return cachep->array[smp_processor_id()];
713 static inline struct kmem_cache *__find_general_cachep(size_t size,
716 struct cache_sizes *csizep = malloc_sizes;
719 /* This happens if someone tries to call
720 * kmem_cache_create(), or __kmalloc(), before
721 * the generic caches are initialized.
723 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
726 return ZERO_SIZE_PTR;
728 while (size > csizep->cs_size)
732 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
733 * has cs_{dma,}cachep==NULL. Thus no special case
734 * for large kmalloc calls required.
736 #ifdef CONFIG_ZONE_DMA
737 if (unlikely(gfpflags & GFP_DMA))
738 return csizep->cs_dmacachep;
740 return csizep->cs_cachep;
743 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
745 return __find_general_cachep(size, gfpflags);
748 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
750 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
754 * Calculate the number of objects and left-over bytes for a given buffer size.
756 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
757 size_t align, int flags, size_t *left_over,
762 size_t slab_size = PAGE_SIZE << gfporder;
765 * The slab management structure can be either off the slab or
766 * on it. For the latter case, the memory allocated for a
770 * - One kmem_bufctl_t for each object
771 * - Padding to respect alignment of @align
772 * - @buffer_size bytes for each object
774 * If the slab management structure is off the slab, then the
775 * alignment will already be calculated into the size. Because
776 * the slabs are all pages aligned, the objects will be at the
777 * correct alignment when allocated.
779 if (flags & CFLGS_OFF_SLAB) {
781 nr_objs = slab_size / buffer_size;
783 if (nr_objs > SLAB_LIMIT)
784 nr_objs = SLAB_LIMIT;
787 * Ignore padding for the initial guess. The padding
788 * is at most @align-1 bytes, and @buffer_size is at
789 * least @align. In the worst case, this result will
790 * be one greater than the number of objects that fit
791 * into the memory allocation when taking the padding
794 nr_objs = (slab_size - sizeof(struct slab)) /
795 (buffer_size + sizeof(kmem_bufctl_t));
798 * This calculated number will be either the right
799 * amount, or one greater than what we want.
801 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
805 if (nr_objs > SLAB_LIMIT)
806 nr_objs = SLAB_LIMIT;
808 mgmt_size = slab_mgmt_size(nr_objs, align);
811 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
814 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
816 static void __slab_error(const char *function, struct kmem_cache *cachep,
819 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
820 function, cachep->name, msg);
825 * By default on NUMA we use alien caches to stage the freeing of
826 * objects allocated from other nodes. This causes massive memory
827 * inefficiencies when using fake NUMA setup to split memory into a
828 * large number of small nodes, so it can be disabled on the command
832 static int use_alien_caches __read_mostly = 1;
833 static int __init noaliencache_setup(char *s)
835 use_alien_caches = 0;
838 __setup("noaliencache", noaliencache_setup);
840 static int __init slab_max_order_setup(char *str)
842 get_option(&str, &slab_max_order);
843 slab_max_order = slab_max_order < 0 ? 0 :
844 min(slab_max_order, MAX_ORDER - 1);
845 slab_max_order_set = true;
849 __setup("slab_max_order=", slab_max_order_setup);
853 * Special reaping functions for NUMA systems called from cache_reap().
854 * These take care of doing round robin flushing of alien caches (containing
855 * objects freed on different nodes from which they were allocated) and the
856 * flushing of remote pcps by calling drain_node_pages.
858 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
860 static void init_reap_node(int cpu)
864 node = next_node(cpu_to_mem(cpu), node_online_map);
865 if (node == MAX_NUMNODES)
866 node = first_node(node_online_map);
868 per_cpu(slab_reap_node, cpu) = node;
871 static void next_reap_node(void)
873 int node = __this_cpu_read(slab_reap_node);
875 node = next_node(node, node_online_map);
876 if (unlikely(node >= MAX_NUMNODES))
877 node = first_node(node_online_map);
878 __this_cpu_write(slab_reap_node, node);
882 #define init_reap_node(cpu) do { } while (0)
883 #define next_reap_node(void) do { } while (0)
887 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
888 * via the workqueue/eventd.
889 * Add the CPU number into the expiration time to minimize the possibility of
890 * the CPUs getting into lockstep and contending for the global cache chain
893 static void __cpuinit start_cpu_timer(int cpu)
895 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
898 * When this gets called from do_initcalls via cpucache_init(),
899 * init_workqueues() has already run, so keventd will be setup
902 if (keventd_up() && reap_work->work.func == NULL) {
904 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
905 schedule_delayed_work_on(cpu, reap_work,
906 __round_jiffies_relative(HZ, cpu));
910 static struct array_cache *alloc_arraycache(int node, int entries,
911 int batchcount, gfp_t gfp)
913 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
914 struct array_cache *nc = NULL;
916 nc = kmalloc_node(memsize, gfp, node);
918 * The array_cache structures contain pointers to free object.
919 * However, when such objects are allocated or transferred to another
920 * cache the pointers are not cleared and they could be counted as
921 * valid references during a kmemleak scan. Therefore, kmemleak must
922 * not scan such objects.
924 kmemleak_no_scan(nc);
928 nc->batchcount = batchcount;
930 spin_lock_init(&nc->lock);
936 * Transfer objects in one arraycache to another.
937 * Locking must be handled by the caller.
939 * Return the number of entries transferred.
941 static int transfer_objects(struct array_cache *to,
942 struct array_cache *from, unsigned int max)
944 /* Figure out how many entries to transfer */
945 int nr = min3(from->avail, max, to->limit - to->avail);
950 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
960 #define drain_alien_cache(cachep, alien) do { } while (0)
961 #define reap_alien(cachep, l3) do { } while (0)
963 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
965 return (struct array_cache **)BAD_ALIEN_MAGIC;
968 static inline void free_alien_cache(struct array_cache **ac_ptr)
972 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
977 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
983 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
984 gfp_t flags, int nodeid)
989 #else /* CONFIG_NUMA */
991 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
992 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
994 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
996 struct array_cache **ac_ptr;
997 int memsize = sizeof(void *) * nr_node_ids;
1002 ac_ptr = kzalloc_node(memsize, gfp, node);
1005 if (i == node || !node_online(i))
1007 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1009 for (i--; i >= 0; i--)
1019 static void free_alien_cache(struct array_cache **ac_ptr)
1030 static void __drain_alien_cache(struct kmem_cache *cachep,
1031 struct array_cache *ac, int node)
1033 struct kmem_list3 *rl3 = cachep->nodelists[node];
1036 spin_lock(&rl3->list_lock);
1038 * Stuff objects into the remote nodes shared array first.
1039 * That way we could avoid the overhead of putting the objects
1040 * into the free lists and getting them back later.
1043 transfer_objects(rl3->shared, ac, ac->limit);
1045 free_block(cachep, ac->entry, ac->avail, node);
1047 spin_unlock(&rl3->list_lock);
1052 * Called from cache_reap() to regularly drain alien caches round robin.
1054 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1056 int node = __this_cpu_read(slab_reap_node);
1059 struct array_cache *ac = l3->alien[node];
1061 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1062 __drain_alien_cache(cachep, ac, node);
1063 spin_unlock_irq(&ac->lock);
1068 static void drain_alien_cache(struct kmem_cache *cachep,
1069 struct array_cache **alien)
1072 struct array_cache *ac;
1073 unsigned long flags;
1075 for_each_online_node(i) {
1078 spin_lock_irqsave(&ac->lock, flags);
1079 __drain_alien_cache(cachep, ac, i);
1080 spin_unlock_irqrestore(&ac->lock, flags);
1085 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1087 struct slab *slabp = virt_to_slab(objp);
1088 int nodeid = slabp->nodeid;
1089 struct kmem_list3 *l3;
1090 struct array_cache *alien = NULL;
1093 node = numa_mem_id();
1096 * Make sure we are not freeing a object from another node to the array
1097 * cache on this cpu.
1099 if (likely(slabp->nodeid == node))
1102 l3 = cachep->nodelists[node];
1103 STATS_INC_NODEFREES(cachep);
1104 if (l3->alien && l3->alien[nodeid]) {
1105 alien = l3->alien[nodeid];
1106 spin_lock(&alien->lock);
1107 if (unlikely(alien->avail == alien->limit)) {
1108 STATS_INC_ACOVERFLOW(cachep);
1109 __drain_alien_cache(cachep, alien, nodeid);
1111 alien->entry[alien->avail++] = objp;
1112 spin_unlock(&alien->lock);
1114 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1115 free_block(cachep, &objp, 1, nodeid);
1116 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1123 * Allocates and initializes nodelists for a node on each slab cache, used for
1124 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1125 * will be allocated off-node since memory is not yet online for the new node.
1126 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1129 * Must hold cache_chain_mutex.
1131 static int init_cache_nodelists_node(int node)
1133 struct kmem_cache *cachep;
1134 struct kmem_list3 *l3;
1135 const int memsize = sizeof(struct kmem_list3);
1137 list_for_each_entry(cachep, &cache_chain, list) {
1139 * Set up the size64 kmemlist for cpu before we can
1140 * begin anything. Make sure some other cpu on this
1141 * node has not already allocated this
1143 if (!cachep->nodelists[node]) {
1144 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1147 kmem_list3_init(l3);
1148 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1149 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1152 * The l3s don't come and go as CPUs come and
1153 * go. cache_chain_mutex is sufficient
1156 cachep->nodelists[node] = l3;
1159 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1160 cachep->nodelists[node]->free_limit =
1161 (1 + nr_cpus_node(node)) *
1162 cachep->batchcount + cachep->num;
1163 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1168 static void __cpuinit cpuup_canceled(long cpu)
1170 struct kmem_cache *cachep;
1171 struct kmem_list3 *l3 = NULL;
1172 int node = cpu_to_mem(cpu);
1173 const struct cpumask *mask = cpumask_of_node(node);
1175 list_for_each_entry(cachep, &cache_chain, list) {
1176 struct array_cache *nc;
1177 struct array_cache *shared;
1178 struct array_cache **alien;
1180 /* cpu is dead; no one can alloc from it. */
1181 nc = cachep->array[cpu];
1182 cachep->array[cpu] = NULL;
1183 l3 = cachep->nodelists[node];
1186 goto free_array_cache;
1188 spin_lock_irq(&l3->list_lock);
1190 /* Free limit for this kmem_list3 */
1191 l3->free_limit -= cachep->batchcount;
1193 free_block(cachep, nc->entry, nc->avail, node);
1195 if (!cpumask_empty(mask)) {
1196 spin_unlock_irq(&l3->list_lock);
1197 goto free_array_cache;
1200 shared = l3->shared;
1202 free_block(cachep, shared->entry,
1203 shared->avail, node);
1210 spin_unlock_irq(&l3->list_lock);
1214 drain_alien_cache(cachep, alien);
1215 free_alien_cache(alien);
1221 * In the previous loop, all the objects were freed to
1222 * the respective cache's slabs, now we can go ahead and
1223 * shrink each nodelist to its limit.
1225 list_for_each_entry(cachep, &cache_chain, list) {
1226 l3 = cachep->nodelists[node];
1229 drain_freelist(cachep, l3, l3->free_objects);
1233 static int __cpuinit cpuup_prepare(long cpu)
1235 struct kmem_cache *cachep;
1236 struct kmem_list3 *l3 = NULL;
1237 int node = cpu_to_mem(cpu);
1241 * We need to do this right in the beginning since
1242 * alloc_arraycache's are going to use this list.
1243 * kmalloc_node allows us to add the slab to the right
1244 * kmem_list3 and not this cpu's kmem_list3
1246 err = init_cache_nodelists_node(node);
1251 * Now we can go ahead with allocating the shared arrays and
1254 list_for_each_entry(cachep, &cache_chain, list) {
1255 struct array_cache *nc;
1256 struct array_cache *shared = NULL;
1257 struct array_cache **alien = NULL;
1259 nc = alloc_arraycache(node, cachep->limit,
1260 cachep->batchcount, GFP_KERNEL);
1263 if (cachep->shared) {
1264 shared = alloc_arraycache(node,
1265 cachep->shared * cachep->batchcount,
1266 0xbaadf00d, GFP_KERNEL);
1272 if (use_alien_caches) {
1273 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1280 cachep->array[cpu] = nc;
1281 l3 = cachep->nodelists[node];
1284 spin_lock_irq(&l3->list_lock);
1287 * We are serialised from CPU_DEAD or
1288 * CPU_UP_CANCELLED by the cpucontrol lock
1290 l3->shared = shared;
1299 spin_unlock_irq(&l3->list_lock);
1301 free_alien_cache(alien);
1302 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1303 slab_set_debugobj_lock_classes_node(cachep, node);
1305 init_node_lock_keys(node);
1309 cpuup_canceled(cpu);
1313 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1314 unsigned long action, void *hcpu)
1316 long cpu = (long)hcpu;
1320 case CPU_UP_PREPARE:
1321 case CPU_UP_PREPARE_FROZEN:
1322 mutex_lock(&cache_chain_mutex);
1323 err = cpuup_prepare(cpu);
1324 mutex_unlock(&cache_chain_mutex);
1327 case CPU_ONLINE_FROZEN:
1328 start_cpu_timer(cpu);
1330 #ifdef CONFIG_HOTPLUG_CPU
1331 case CPU_DOWN_PREPARE:
1332 case CPU_DOWN_PREPARE_FROZEN:
1334 * Shutdown cache reaper. Note that the cache_chain_mutex is
1335 * held so that if cache_reap() is invoked it cannot do
1336 * anything expensive but will only modify reap_work
1337 * and reschedule the timer.
1339 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1340 /* Now the cache_reaper is guaranteed to be not running. */
1341 per_cpu(slab_reap_work, cpu).work.func = NULL;
1343 case CPU_DOWN_FAILED:
1344 case CPU_DOWN_FAILED_FROZEN:
1345 start_cpu_timer(cpu);
1348 case CPU_DEAD_FROZEN:
1350 * Even if all the cpus of a node are down, we don't free the
1351 * kmem_list3 of any cache. This to avoid a race between
1352 * cpu_down, and a kmalloc allocation from another cpu for
1353 * memory from the node of the cpu going down. The list3
1354 * structure is usually allocated from kmem_cache_create() and
1355 * gets destroyed at kmem_cache_destroy().
1359 case CPU_UP_CANCELED:
1360 case CPU_UP_CANCELED_FROZEN:
1361 mutex_lock(&cache_chain_mutex);
1362 cpuup_canceled(cpu);
1363 mutex_unlock(&cache_chain_mutex);
1366 return notifier_from_errno(err);
1369 static struct notifier_block __cpuinitdata cpucache_notifier = {
1370 &cpuup_callback, NULL, 0
1373 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1375 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1376 * Returns -EBUSY if all objects cannot be drained so that the node is not
1379 * Must hold cache_chain_mutex.
1381 static int __meminit drain_cache_nodelists_node(int node)
1383 struct kmem_cache *cachep;
1386 list_for_each_entry(cachep, &cache_chain, list) {
1387 struct kmem_list3 *l3;
1389 l3 = cachep->nodelists[node];
1393 drain_freelist(cachep, l3, l3->free_objects);
1395 if (!list_empty(&l3->slabs_full) ||
1396 !list_empty(&l3->slabs_partial)) {
1404 static int __meminit slab_memory_callback(struct notifier_block *self,
1405 unsigned long action, void *arg)
1407 struct memory_notify *mnb = arg;
1411 nid = mnb->status_change_nid;
1416 case MEM_GOING_ONLINE:
1417 mutex_lock(&cache_chain_mutex);
1418 ret = init_cache_nodelists_node(nid);
1419 mutex_unlock(&cache_chain_mutex);
1421 case MEM_GOING_OFFLINE:
1422 mutex_lock(&cache_chain_mutex);
1423 ret = drain_cache_nodelists_node(nid);
1424 mutex_unlock(&cache_chain_mutex);
1428 case MEM_CANCEL_ONLINE:
1429 case MEM_CANCEL_OFFLINE:
1433 return notifier_from_errno(ret);
1435 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1438 * swap the static kmem_list3 with kmalloced memory
1440 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1443 struct kmem_list3 *ptr;
1445 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1448 memcpy(ptr, list, sizeof(struct kmem_list3));
1450 * Do not assume that spinlocks can be initialized via memcpy:
1452 spin_lock_init(&ptr->list_lock);
1454 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1455 cachep->nodelists[nodeid] = ptr;
1459 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1460 * size of kmem_list3.
1462 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1466 for_each_online_node(node) {
1467 cachep->nodelists[node] = &initkmem_list3[index + node];
1468 cachep->nodelists[node]->next_reap = jiffies +
1470 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1475 * Initialisation. Called after the page allocator have been initialised and
1476 * before smp_init().
1478 void __init kmem_cache_init(void)
1481 struct cache_sizes *sizes;
1482 struct cache_names *names;
1487 if (num_possible_nodes() == 1)
1488 use_alien_caches = 0;
1490 for (i = 0; i < NUM_INIT_LISTS; i++) {
1491 kmem_list3_init(&initkmem_list3[i]);
1492 if (i < MAX_NUMNODES)
1493 cache_cache.nodelists[i] = NULL;
1495 set_up_list3s(&cache_cache, CACHE_CACHE);
1498 * Fragmentation resistance on low memory - only use bigger
1499 * page orders on machines with more than 32MB of memory if
1500 * not overridden on the command line.
1502 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1503 slab_max_order = SLAB_MAX_ORDER_HI;
1505 /* Bootstrap is tricky, because several objects are allocated
1506 * from caches that do not exist yet:
1507 * 1) initialize the cache_cache cache: it contains the struct
1508 * kmem_cache structures of all caches, except cache_cache itself:
1509 * cache_cache is statically allocated.
1510 * Initially an __init data area is used for the head array and the
1511 * kmem_list3 structures, it's replaced with a kmalloc allocated
1512 * array at the end of the bootstrap.
1513 * 2) Create the first kmalloc cache.
1514 * The struct kmem_cache for the new cache is allocated normally.
1515 * An __init data area is used for the head array.
1516 * 3) Create the remaining kmalloc caches, with minimally sized
1518 * 4) Replace the __init data head arrays for cache_cache and the first
1519 * kmalloc cache with kmalloc allocated arrays.
1520 * 5) Replace the __init data for kmem_list3 for cache_cache and
1521 * the other cache's with kmalloc allocated memory.
1522 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1525 node = numa_mem_id();
1527 /* 1) create the cache_cache */
1528 INIT_LIST_HEAD(&cache_chain);
1529 list_add(&cache_cache.list, &cache_chain);
1530 cache_cache.colour_off = cache_line_size();
1531 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1532 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1535 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1537 cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1538 nr_node_ids * sizeof(struct kmem_list3 *);
1539 cache_cache.object_size = cache_cache.size;
1540 cache_cache.size = ALIGN(cache_cache.size,
1542 cache_cache.reciprocal_buffer_size =
1543 reciprocal_value(cache_cache.size);
1545 for (order = 0; order < MAX_ORDER; order++) {
1546 cache_estimate(order, cache_cache.size,
1547 cache_line_size(), 0, &left_over, &cache_cache.num);
1548 if (cache_cache.num)
1551 BUG_ON(!cache_cache.num);
1552 cache_cache.gfporder = order;
1553 cache_cache.colour = left_over / cache_cache.colour_off;
1554 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1555 sizeof(struct slab), cache_line_size());
1557 /* 2+3) create the kmalloc caches */
1558 sizes = malloc_sizes;
1559 names = cache_names;
1562 * Initialize the caches that provide memory for the array cache and the
1563 * kmem_list3 structures first. Without this, further allocations will
1567 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1568 sizes[INDEX_AC].cs_size,
1569 ARCH_KMALLOC_MINALIGN,
1570 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1573 if (INDEX_AC != INDEX_L3) {
1574 sizes[INDEX_L3].cs_cachep =
1575 kmem_cache_create(names[INDEX_L3].name,
1576 sizes[INDEX_L3].cs_size,
1577 ARCH_KMALLOC_MINALIGN,
1578 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1582 slab_early_init = 0;
1584 while (sizes->cs_size != ULONG_MAX) {
1586 * For performance, all the general caches are L1 aligned.
1587 * This should be particularly beneficial on SMP boxes, as it
1588 * eliminates "false sharing".
1589 * Note for systems short on memory removing the alignment will
1590 * allow tighter packing of the smaller caches.
1592 if (!sizes->cs_cachep) {
1593 sizes->cs_cachep = kmem_cache_create(names->name,
1595 ARCH_KMALLOC_MINALIGN,
1596 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1599 #ifdef CONFIG_ZONE_DMA
1600 sizes->cs_dmacachep = kmem_cache_create(
1603 ARCH_KMALLOC_MINALIGN,
1604 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1611 /* 4) Replace the bootstrap head arrays */
1613 struct array_cache *ptr;
1615 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1617 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1618 memcpy(ptr, cpu_cache_get(&cache_cache),
1619 sizeof(struct arraycache_init));
1621 * Do not assume that spinlocks can be initialized via memcpy:
1623 spin_lock_init(&ptr->lock);
1625 cache_cache.array[smp_processor_id()] = ptr;
1627 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1629 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1630 != &initarray_generic.cache);
1631 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1632 sizeof(struct arraycache_init));
1634 * Do not assume that spinlocks can be initialized via memcpy:
1636 spin_lock_init(&ptr->lock);
1638 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1641 /* 5) Replace the bootstrap kmem_list3's */
1645 for_each_online_node(nid) {
1646 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1648 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1649 &initkmem_list3[SIZE_AC + nid], nid);
1651 if (INDEX_AC != INDEX_L3) {
1652 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1653 &initkmem_list3[SIZE_L3 + nid], nid);
1658 g_cpucache_up = EARLY;
1661 void __init kmem_cache_init_late(void)
1663 struct kmem_cache *cachep;
1665 g_cpucache_up = LATE;
1667 /* Annotate slab for lockdep -- annotate the malloc caches */
1670 /* 6) resize the head arrays to their final sizes */
1671 mutex_lock(&cache_chain_mutex);
1672 list_for_each_entry(cachep, &cache_chain, list)
1673 if (enable_cpucache(cachep, GFP_NOWAIT))
1675 mutex_unlock(&cache_chain_mutex);
1678 g_cpucache_up = FULL;
1681 * Register a cpu startup notifier callback that initializes
1682 * cpu_cache_get for all new cpus
1684 register_cpu_notifier(&cpucache_notifier);
1688 * Register a memory hotplug callback that initializes and frees
1691 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1695 * The reap timers are started later, with a module init call: That part
1696 * of the kernel is not yet operational.
1700 static int __init cpucache_init(void)
1705 * Register the timers that return unneeded pages to the page allocator
1707 for_each_online_cpu(cpu)
1708 start_cpu_timer(cpu);
1711 __initcall(cpucache_init);
1713 static noinline void
1714 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1716 struct kmem_list3 *l3;
1718 unsigned long flags;
1722 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1724 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1725 cachep->name, cachep->size, cachep->gfporder);
1727 for_each_online_node(node) {
1728 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1729 unsigned long active_slabs = 0, num_slabs = 0;
1731 l3 = cachep->nodelists[node];
1735 spin_lock_irqsave(&l3->list_lock, flags);
1736 list_for_each_entry(slabp, &l3->slabs_full, list) {
1737 active_objs += cachep->num;
1740 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1741 active_objs += slabp->inuse;
1744 list_for_each_entry(slabp, &l3->slabs_free, list)
1747 free_objects += l3->free_objects;
1748 spin_unlock_irqrestore(&l3->list_lock, flags);
1750 num_slabs += active_slabs;
1751 num_objs = num_slabs * cachep->num;
1753 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1754 node, active_slabs, num_slabs, active_objs, num_objs,
1760 * Interface to system's page allocator. No need to hold the cache-lock.
1762 * If we requested dmaable memory, we will get it. Even if we
1763 * did not request dmaable memory, we might get it, but that
1764 * would be relatively rare and ignorable.
1766 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1774 * Nommu uses slab's for process anonymous memory allocations, and thus
1775 * requires __GFP_COMP to properly refcount higher order allocations
1777 flags |= __GFP_COMP;
1780 flags |= cachep->gfpflags;
1781 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1782 flags |= __GFP_RECLAIMABLE;
1784 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1786 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1787 slab_out_of_memory(cachep, flags, nodeid);
1791 nr_pages = (1 << cachep->gfporder);
1792 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1793 add_zone_page_state(page_zone(page),
1794 NR_SLAB_RECLAIMABLE, nr_pages);
1796 add_zone_page_state(page_zone(page),
1797 NR_SLAB_UNRECLAIMABLE, nr_pages);
1798 for (i = 0; i < nr_pages; i++)
1799 __SetPageSlab(page + i);
1801 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1802 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1805 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1807 kmemcheck_mark_unallocated_pages(page, nr_pages);
1810 return page_address(page);
1814 * Interface to system's page release.
1816 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1818 unsigned long i = (1 << cachep->gfporder);
1819 struct page *page = virt_to_page(addr);
1820 const unsigned long nr_freed = i;
1822 kmemcheck_free_shadow(page, cachep->gfporder);
1824 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1825 sub_zone_page_state(page_zone(page),
1826 NR_SLAB_RECLAIMABLE, nr_freed);
1828 sub_zone_page_state(page_zone(page),
1829 NR_SLAB_UNRECLAIMABLE, nr_freed);
1831 BUG_ON(!PageSlab(page));
1832 __ClearPageSlab(page);
1835 if (current->reclaim_state)
1836 current->reclaim_state->reclaimed_slab += nr_freed;
1837 free_pages((unsigned long)addr, cachep->gfporder);
1840 static void kmem_rcu_free(struct rcu_head *head)
1842 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1843 struct kmem_cache *cachep = slab_rcu->cachep;
1845 kmem_freepages(cachep, slab_rcu->addr);
1846 if (OFF_SLAB(cachep))
1847 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1852 #ifdef CONFIG_DEBUG_PAGEALLOC
1853 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1854 unsigned long caller)
1856 int size = obj_size(cachep);
1858 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1860 if (size < 5 * sizeof(unsigned long))
1863 *addr++ = 0x12345678;
1865 *addr++ = smp_processor_id();
1866 size -= 3 * sizeof(unsigned long);
1868 unsigned long *sptr = &caller;
1869 unsigned long svalue;
1871 while (!kstack_end(sptr)) {
1873 if (kernel_text_address(svalue)) {
1875 size -= sizeof(unsigned long);
1876 if (size <= sizeof(unsigned long))
1882 *addr++ = 0x87654321;
1886 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1888 int size = obj_size(cachep);
1889 addr = &((char *)addr)[obj_offset(cachep)];
1891 memset(addr, val, size);
1892 *(unsigned char *)(addr + size - 1) = POISON_END;
1895 static void dump_line(char *data, int offset, int limit)
1898 unsigned char error = 0;
1901 printk(KERN_ERR "%03x: ", offset);
1902 for (i = 0; i < limit; i++) {
1903 if (data[offset + i] != POISON_FREE) {
1904 error = data[offset + i];
1908 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1909 &data[offset], limit, 1);
1911 if (bad_count == 1) {
1912 error ^= POISON_FREE;
1913 if (!(error & (error - 1))) {
1914 printk(KERN_ERR "Single bit error detected. Probably "
1917 printk(KERN_ERR "Run memtest86+ or a similar memory "
1920 printk(KERN_ERR "Run a memory test tool.\n");
1929 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1934 if (cachep->flags & SLAB_RED_ZONE) {
1935 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1936 *dbg_redzone1(cachep, objp),
1937 *dbg_redzone2(cachep, objp));
1940 if (cachep->flags & SLAB_STORE_USER) {
1941 printk(KERN_ERR "Last user: [<%p>]",
1942 *dbg_userword(cachep, objp));
1943 print_symbol("(%s)",
1944 (unsigned long)*dbg_userword(cachep, objp));
1947 realobj = (char *)objp + obj_offset(cachep);
1948 size = obj_size(cachep);
1949 for (i = 0; i < size && lines; i += 16, lines--) {
1952 if (i + limit > size)
1954 dump_line(realobj, i, limit);
1958 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1964 realobj = (char *)objp + obj_offset(cachep);
1965 size = obj_size(cachep);
1967 for (i = 0; i < size; i++) {
1968 char exp = POISON_FREE;
1971 if (realobj[i] != exp) {
1977 "Slab corruption (%s): %s start=%p, len=%d\n",
1978 print_tainted(), cachep->name, realobj, size);
1979 print_objinfo(cachep, objp, 0);
1981 /* Hexdump the affected line */
1984 if (i + limit > size)
1986 dump_line(realobj, i, limit);
1989 /* Limit to 5 lines */
1995 /* Print some data about the neighboring objects, if they
1998 struct slab *slabp = virt_to_slab(objp);
2001 objnr = obj_to_index(cachep, slabp, objp);
2003 objp = index_to_obj(cachep, slabp, objnr - 1);
2004 realobj = (char *)objp + obj_offset(cachep);
2005 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2007 print_objinfo(cachep, objp, 2);
2009 if (objnr + 1 < cachep->num) {
2010 objp = index_to_obj(cachep, slabp, objnr + 1);
2011 realobj = (char *)objp + obj_offset(cachep);
2012 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2014 print_objinfo(cachep, objp, 2);
2021 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2024 for (i = 0; i < cachep->num; i++) {
2025 void *objp = index_to_obj(cachep, slabp, i);
2027 if (cachep->flags & SLAB_POISON) {
2028 #ifdef CONFIG_DEBUG_PAGEALLOC
2029 if (cachep->size % PAGE_SIZE == 0 &&
2031 kernel_map_pages(virt_to_page(objp),
2032 cachep->size / PAGE_SIZE, 1);
2034 check_poison_obj(cachep, objp);
2036 check_poison_obj(cachep, objp);
2039 if (cachep->flags & SLAB_RED_ZONE) {
2040 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2041 slab_error(cachep, "start of a freed object "
2043 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2044 slab_error(cachep, "end of a freed object "
2050 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2056 * slab_destroy - destroy and release all objects in a slab
2057 * @cachep: cache pointer being destroyed
2058 * @slabp: slab pointer being destroyed
2060 * Destroy all the objs in a slab, and release the mem back to the system.
2061 * Before calling the slab must have been unlinked from the cache. The
2062 * cache-lock is not held/needed.
2064 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2066 void *addr = slabp->s_mem - slabp->colouroff;
2068 slab_destroy_debugcheck(cachep, slabp);
2069 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2070 struct slab_rcu *slab_rcu;
2072 slab_rcu = (struct slab_rcu *)slabp;
2073 slab_rcu->cachep = cachep;
2074 slab_rcu->addr = addr;
2075 call_rcu(&slab_rcu->head, kmem_rcu_free);
2077 kmem_freepages(cachep, addr);
2078 if (OFF_SLAB(cachep))
2079 kmem_cache_free(cachep->slabp_cache, slabp);
2083 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2086 struct kmem_list3 *l3;
2088 for_each_online_cpu(i)
2089 kfree(cachep->array[i]);
2091 /* NUMA: free the list3 structures */
2092 for_each_online_node(i) {
2093 l3 = cachep->nodelists[i];
2096 free_alien_cache(l3->alien);
2100 kmem_cache_free(&cache_cache, cachep);
2105 * calculate_slab_order - calculate size (page order) of slabs
2106 * @cachep: pointer to the cache that is being created
2107 * @size: size of objects to be created in this cache.
2108 * @align: required alignment for the objects.
2109 * @flags: slab allocation flags
2111 * Also calculates the number of objects per slab.
2113 * This could be made much more intelligent. For now, try to avoid using
2114 * high order pages for slabs. When the gfp() functions are more friendly
2115 * towards high-order requests, this should be changed.
2117 static size_t calculate_slab_order(struct kmem_cache *cachep,
2118 size_t size, size_t align, unsigned long flags)
2120 unsigned long offslab_limit;
2121 size_t left_over = 0;
2124 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2128 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2132 if (flags & CFLGS_OFF_SLAB) {
2134 * Max number of objs-per-slab for caches which
2135 * use off-slab slabs. Needed to avoid a possible
2136 * looping condition in cache_grow().
2138 offslab_limit = size - sizeof(struct slab);
2139 offslab_limit /= sizeof(kmem_bufctl_t);
2141 if (num > offslab_limit)
2145 /* Found something acceptable - save it away */
2147 cachep->gfporder = gfporder;
2148 left_over = remainder;
2151 * A VFS-reclaimable slab tends to have most allocations
2152 * as GFP_NOFS and we really don't want to have to be allocating
2153 * higher-order pages when we are unable to shrink dcache.
2155 if (flags & SLAB_RECLAIM_ACCOUNT)
2159 * Large number of objects is good, but very large slabs are
2160 * currently bad for the gfp()s.
2162 if (gfporder >= slab_max_order)
2166 * Acceptable internal fragmentation?
2168 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2174 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2176 if (g_cpucache_up == FULL)
2177 return enable_cpucache(cachep, gfp);
2179 if (g_cpucache_up == NONE) {
2181 * Note: the first kmem_cache_create must create the cache
2182 * that's used by kmalloc(24), otherwise the creation of
2183 * further caches will BUG().
2185 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2188 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2189 * the first cache, then we need to set up all its list3s,
2190 * otherwise the creation of further caches will BUG().
2192 set_up_list3s(cachep, SIZE_AC);
2193 if (INDEX_AC == INDEX_L3)
2194 g_cpucache_up = PARTIAL_L3;
2196 g_cpucache_up = PARTIAL_AC;
2198 cachep->array[smp_processor_id()] =
2199 kmalloc(sizeof(struct arraycache_init), gfp);
2201 if (g_cpucache_up == PARTIAL_AC) {
2202 set_up_list3s(cachep, SIZE_L3);
2203 g_cpucache_up = PARTIAL_L3;
2206 for_each_online_node(node) {
2207 cachep->nodelists[node] =
2208 kmalloc_node(sizeof(struct kmem_list3),
2210 BUG_ON(!cachep->nodelists[node]);
2211 kmem_list3_init(cachep->nodelists[node]);
2215 cachep->nodelists[numa_mem_id()]->next_reap =
2216 jiffies + REAPTIMEOUT_LIST3 +
2217 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2219 cpu_cache_get(cachep)->avail = 0;
2220 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2221 cpu_cache_get(cachep)->batchcount = 1;
2222 cpu_cache_get(cachep)->touched = 0;
2223 cachep->batchcount = 1;
2224 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2229 * kmem_cache_create - Create a cache.
2230 * @name: A string which is used in /proc/slabinfo to identify this cache.
2231 * @size: The size of objects to be created in this cache.
2232 * @align: The required alignment for the objects.
2233 * @flags: SLAB flags
2234 * @ctor: A constructor for the objects.
2236 * Returns a ptr to the cache on success, NULL on failure.
2237 * Cannot be called within a int, but can be interrupted.
2238 * The @ctor is run when new pages are allocated by the cache.
2240 * @name must be valid until the cache is destroyed. This implies that
2241 * the module calling this has to destroy the cache before getting unloaded.
2245 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2246 * to catch references to uninitialised memory.
2248 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2249 * for buffer overruns.
2251 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2252 * cacheline. This can be beneficial if you're counting cycles as closely
2256 kmem_cache_create (const char *name, size_t size, size_t align,
2257 unsigned long flags, void (*ctor)(void *))
2259 size_t left_over, slab_size, ralign;
2260 struct kmem_cache *cachep = NULL, *pc;
2264 * Sanity checks... these are all serious usage bugs.
2266 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2267 size > KMALLOC_MAX_SIZE) {
2268 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2274 * We use cache_chain_mutex to ensure a consistent view of
2275 * cpu_online_mask as well. Please see cpuup_callback
2277 if (slab_is_available()) {
2279 mutex_lock(&cache_chain_mutex);
2282 list_for_each_entry(pc, &cache_chain, list) {
2287 * This happens when the module gets unloaded and doesn't
2288 * destroy its slab cache and no-one else reuses the vmalloc
2289 * area of the module. Print a warning.
2291 res = probe_kernel_address(pc->name, tmp);
2294 "SLAB: cache with size %d has lost its name\n",
2299 if (!strcmp(pc->name, name)) {
2301 "kmem_cache_create: duplicate cache %s\n", name);
2308 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2311 * Enable redzoning and last user accounting, except for caches with
2312 * large objects, if the increased size would increase the object size
2313 * above the next power of two: caches with object sizes just above a
2314 * power of two have a significant amount of internal fragmentation.
2316 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2317 2 * sizeof(unsigned long long)))
2318 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2319 if (!(flags & SLAB_DESTROY_BY_RCU))
2320 flags |= SLAB_POISON;
2322 if (flags & SLAB_DESTROY_BY_RCU)
2323 BUG_ON(flags & SLAB_POISON);
2326 * Always checks flags, a caller might be expecting debug support which
2329 BUG_ON(flags & ~CREATE_MASK);
2332 * Check that size is in terms of words. This is needed to avoid
2333 * unaligned accesses for some archs when redzoning is used, and makes
2334 * sure any on-slab bufctl's are also correctly aligned.
2336 if (size & (BYTES_PER_WORD - 1)) {
2337 size += (BYTES_PER_WORD - 1);
2338 size &= ~(BYTES_PER_WORD - 1);
2341 /* calculate the final buffer alignment: */
2343 /* 1) arch recommendation: can be overridden for debug */
2344 if (flags & SLAB_HWCACHE_ALIGN) {
2346 * Default alignment: as specified by the arch code. Except if
2347 * an object is really small, then squeeze multiple objects into
2350 ralign = cache_line_size();
2351 while (size <= ralign / 2)
2354 ralign = BYTES_PER_WORD;
2358 * Redzoning and user store require word alignment or possibly larger.
2359 * Note this will be overridden by architecture or caller mandated
2360 * alignment if either is greater than BYTES_PER_WORD.
2362 if (flags & SLAB_STORE_USER)
2363 ralign = BYTES_PER_WORD;
2365 if (flags & SLAB_RED_ZONE) {
2366 ralign = REDZONE_ALIGN;
2367 /* If redzoning, ensure that the second redzone is suitably
2368 * aligned, by adjusting the object size accordingly. */
2369 size += REDZONE_ALIGN - 1;
2370 size &= ~(REDZONE_ALIGN - 1);
2373 /* 2) arch mandated alignment */
2374 if (ralign < ARCH_SLAB_MINALIGN) {
2375 ralign = ARCH_SLAB_MINALIGN;
2377 /* 3) caller mandated alignment */
2378 if (ralign < align) {
2381 /* disable debug if necessary */
2382 if (ralign > __alignof__(unsigned long long))
2383 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2389 if (slab_is_available())
2394 /* Get cache's description obj. */
2395 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2399 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2400 cachep->object_size = size;
2401 cachep->align = align;
2405 * Both debugging options require word-alignment which is calculated
2408 if (flags & SLAB_RED_ZONE) {
2409 /* add space for red zone words */
2410 cachep->obj_offset += sizeof(unsigned long long);
2411 size += 2 * sizeof(unsigned long long);
2413 if (flags & SLAB_STORE_USER) {
2414 /* user store requires one word storage behind the end of
2415 * the real object. But if the second red zone needs to be
2416 * aligned to 64 bits, we must allow that much space.
2418 if (flags & SLAB_RED_ZONE)
2419 size += REDZONE_ALIGN;
2421 size += BYTES_PER_WORD;
2423 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2424 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2425 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2426 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2433 * Determine if the slab management is 'on' or 'off' slab.
2434 * (bootstrapping cannot cope with offslab caches so don't do
2435 * it too early on. Always use on-slab management when
2436 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2438 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2439 !(flags & SLAB_NOLEAKTRACE))
2441 * Size is large, assume best to place the slab management obj
2442 * off-slab (should allow better packing of objs).
2444 flags |= CFLGS_OFF_SLAB;
2446 size = ALIGN(size, align);
2448 left_over = calculate_slab_order(cachep, size, align, flags);
2452 "kmem_cache_create: couldn't create cache %s.\n", name);
2453 kmem_cache_free(&cache_cache, cachep);
2457 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2458 + sizeof(struct slab), align);
2461 * If the slab has been placed off-slab, and we have enough space then
2462 * move it on-slab. This is at the expense of any extra colouring.
2464 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2465 flags &= ~CFLGS_OFF_SLAB;
2466 left_over -= slab_size;
2469 if (flags & CFLGS_OFF_SLAB) {
2470 /* really off slab. No need for manual alignment */
2472 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2474 #ifdef CONFIG_PAGE_POISONING
2475 /* If we're going to use the generic kernel_map_pages()
2476 * poisoning, then it's going to smash the contents of
2477 * the redzone and userword anyhow, so switch them off.
2479 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2480 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2484 cachep->colour_off = cache_line_size();
2485 /* Offset must be a multiple of the alignment. */
2486 if (cachep->colour_off < align)
2487 cachep->colour_off = align;
2488 cachep->colour = left_over / cachep->colour_off;
2489 cachep->slab_size = slab_size;
2490 cachep->flags = flags;
2491 cachep->gfpflags = 0;
2492 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2493 cachep->gfpflags |= GFP_DMA;
2494 cachep->size = size;
2495 cachep->reciprocal_buffer_size = reciprocal_value(size);
2497 if (flags & CFLGS_OFF_SLAB) {
2498 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2500 * This is a possibility for one of the malloc_sizes caches.
2501 * But since we go off slab only for object size greater than
2502 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2503 * this should not happen at all.
2504 * But leave a BUG_ON for some lucky dude.
2506 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2508 cachep->ctor = ctor;
2509 cachep->name = name;
2511 if (setup_cpu_cache(cachep, gfp)) {
2512 __kmem_cache_destroy(cachep);
2517 if (flags & SLAB_DEBUG_OBJECTS) {
2519 * Would deadlock through slab_destroy()->call_rcu()->
2520 * debug_object_activate()->kmem_cache_alloc().
2522 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2524 slab_set_debugobj_lock_classes(cachep);
2527 /* cache setup completed, link it into the list */
2528 list_add(&cachep->list, &cache_chain);
2530 if (!cachep && (flags & SLAB_PANIC))
2531 panic("kmem_cache_create(): failed to create slab `%s'\n",
2533 if (slab_is_available()) {
2534 mutex_unlock(&cache_chain_mutex);
2539 EXPORT_SYMBOL(kmem_cache_create);
2542 static void check_irq_off(void)
2544 BUG_ON(!irqs_disabled());
2547 static void check_irq_on(void)
2549 BUG_ON(irqs_disabled());
2552 static void check_spinlock_acquired(struct kmem_cache *cachep)
2556 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2560 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2564 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2569 #define check_irq_off() do { } while(0)
2570 #define check_irq_on() do { } while(0)
2571 #define check_spinlock_acquired(x) do { } while(0)
2572 #define check_spinlock_acquired_node(x, y) do { } while(0)
2575 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2576 struct array_cache *ac,
2577 int force, int node);
2579 static void do_drain(void *arg)
2581 struct kmem_cache *cachep = arg;
2582 struct array_cache *ac;
2583 int node = numa_mem_id();
2586 ac = cpu_cache_get(cachep);
2587 spin_lock(&cachep->nodelists[node]->list_lock);
2588 free_block(cachep, ac->entry, ac->avail, node);
2589 spin_unlock(&cachep->nodelists[node]->list_lock);
2593 static void drain_cpu_caches(struct kmem_cache *cachep)
2595 struct kmem_list3 *l3;
2598 on_each_cpu(do_drain, cachep, 1);
2600 for_each_online_node(node) {
2601 l3 = cachep->nodelists[node];
2602 if (l3 && l3->alien)
2603 drain_alien_cache(cachep, l3->alien);
2606 for_each_online_node(node) {
2607 l3 = cachep->nodelists[node];
2609 drain_array(cachep, l3, l3->shared, 1, node);
2614 * Remove slabs from the list of free slabs.
2615 * Specify the number of slabs to drain in tofree.
2617 * Returns the actual number of slabs released.
2619 static int drain_freelist(struct kmem_cache *cache,
2620 struct kmem_list3 *l3, int tofree)
2622 struct list_head *p;
2627 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2629 spin_lock_irq(&l3->list_lock);
2630 p = l3->slabs_free.prev;
2631 if (p == &l3->slabs_free) {
2632 spin_unlock_irq(&l3->list_lock);
2636 slabp = list_entry(p, struct slab, list);
2638 BUG_ON(slabp->inuse);
2640 list_del(&slabp->list);
2642 * Safe to drop the lock. The slab is no longer linked
2645 l3->free_objects -= cache->num;
2646 spin_unlock_irq(&l3->list_lock);
2647 slab_destroy(cache, slabp);
2654 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2655 static int __cache_shrink(struct kmem_cache *cachep)
2658 struct kmem_list3 *l3;
2660 drain_cpu_caches(cachep);
2663 for_each_online_node(i) {
2664 l3 = cachep->nodelists[i];
2668 drain_freelist(cachep, l3, l3->free_objects);
2670 ret += !list_empty(&l3->slabs_full) ||
2671 !list_empty(&l3->slabs_partial);
2673 return (ret ? 1 : 0);
2677 * kmem_cache_shrink - Shrink a cache.
2678 * @cachep: The cache to shrink.
2680 * Releases as many slabs as possible for a cache.
2681 * To help debugging, a zero exit status indicates all slabs were released.
2683 int kmem_cache_shrink(struct kmem_cache *cachep)
2686 BUG_ON(!cachep || in_interrupt());
2689 mutex_lock(&cache_chain_mutex);
2690 ret = __cache_shrink(cachep);
2691 mutex_unlock(&cache_chain_mutex);
2695 EXPORT_SYMBOL(kmem_cache_shrink);
2698 * kmem_cache_destroy - delete a cache
2699 * @cachep: the cache to destroy
2701 * Remove a &struct kmem_cache object from the slab cache.
2703 * It is expected this function will be called by a module when it is
2704 * unloaded. This will remove the cache completely, and avoid a duplicate
2705 * cache being allocated each time a module is loaded and unloaded, if the
2706 * module doesn't have persistent in-kernel storage across loads and unloads.
2708 * The cache must be empty before calling this function.
2710 * The caller must guarantee that no one will allocate memory from the cache
2711 * during the kmem_cache_destroy().
2713 void kmem_cache_destroy(struct kmem_cache *cachep)
2715 BUG_ON(!cachep || in_interrupt());
2717 /* Find the cache in the chain of caches. */
2719 mutex_lock(&cache_chain_mutex);
2721 * the chain is never empty, cache_cache is never destroyed
2723 list_del(&cachep->list);
2724 if (__cache_shrink(cachep)) {
2725 slab_error(cachep, "Can't free all objects");
2726 list_add(&cachep->list, &cache_chain);
2727 mutex_unlock(&cache_chain_mutex);
2732 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2735 __kmem_cache_destroy(cachep);
2736 mutex_unlock(&cache_chain_mutex);
2739 EXPORT_SYMBOL(kmem_cache_destroy);
2742 * Get the memory for a slab management obj.
2743 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2744 * always come from malloc_sizes caches. The slab descriptor cannot
2745 * come from the same cache which is getting created because,
2746 * when we are searching for an appropriate cache for these
2747 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2748 * If we are creating a malloc_sizes cache here it would not be visible to
2749 * kmem_find_general_cachep till the initialization is complete.
2750 * Hence we cannot have slabp_cache same as the original cache.
2752 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2753 int colour_off, gfp_t local_flags,
2758 if (OFF_SLAB(cachep)) {
2759 /* Slab management obj is off-slab. */
2760 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2761 local_flags, nodeid);
2763 * If the first object in the slab is leaked (it's allocated
2764 * but no one has a reference to it), we want to make sure
2765 * kmemleak does not treat the ->s_mem pointer as a reference
2766 * to the object. Otherwise we will not report the leak.
2768 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2773 slabp = objp + colour_off;
2774 colour_off += cachep->slab_size;
2777 slabp->colouroff = colour_off;
2778 slabp->s_mem = objp + colour_off;
2779 slabp->nodeid = nodeid;
2784 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2786 return (kmem_bufctl_t *) (slabp + 1);
2789 static void cache_init_objs(struct kmem_cache *cachep,
2794 for (i = 0; i < cachep->num; i++) {
2795 void *objp = index_to_obj(cachep, slabp, i);
2797 /* need to poison the objs? */
2798 if (cachep->flags & SLAB_POISON)
2799 poison_obj(cachep, objp, POISON_FREE);
2800 if (cachep->flags & SLAB_STORE_USER)
2801 *dbg_userword(cachep, objp) = NULL;
2803 if (cachep->flags & SLAB_RED_ZONE) {
2804 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2805 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2808 * Constructors are not allowed to allocate memory from the same
2809 * cache which they are a constructor for. Otherwise, deadlock.
2810 * They must also be threaded.
2812 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2813 cachep->ctor(objp + obj_offset(cachep));
2815 if (cachep->flags & SLAB_RED_ZONE) {
2816 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2817 slab_error(cachep, "constructor overwrote the"
2818 " end of an object");
2819 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2820 slab_error(cachep, "constructor overwrote the"
2821 " start of an object");
2823 if ((cachep->size % PAGE_SIZE) == 0 &&
2824 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2825 kernel_map_pages(virt_to_page(objp),
2826 cachep->size / PAGE_SIZE, 0);
2831 slab_bufctl(slabp)[i] = i + 1;
2833 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2836 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2838 if (CONFIG_ZONE_DMA_FLAG) {
2839 if (flags & GFP_DMA)
2840 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2842 BUG_ON(cachep->gfpflags & GFP_DMA);
2846 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2849 void *objp = index_to_obj(cachep, slabp, slabp->free);
2853 next = slab_bufctl(slabp)[slabp->free];
2855 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2856 WARN_ON(slabp->nodeid != nodeid);
2863 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2864 void *objp, int nodeid)
2866 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2869 /* Verify that the slab belongs to the intended node */
2870 WARN_ON(slabp->nodeid != nodeid);
2872 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2873 printk(KERN_ERR "slab: double free detected in cache "
2874 "'%s', objp %p\n", cachep->name, objp);
2878 slab_bufctl(slabp)[objnr] = slabp->free;
2879 slabp->free = objnr;
2884 * Map pages beginning at addr to the given cache and slab. This is required
2885 * for the slab allocator to be able to lookup the cache and slab of a
2886 * virtual address for kfree, ksize, and slab debugging.
2888 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2894 page = virt_to_page(addr);
2897 if (likely(!PageCompound(page)))
2898 nr_pages <<= cache->gfporder;
2901 page->slab_cache = cache;
2902 page->slab_page = slab;
2904 } while (--nr_pages);
2908 * Grow (by 1) the number of slabs within a cache. This is called by
2909 * kmem_cache_alloc() when there are no active objs left in a cache.
2911 static int cache_grow(struct kmem_cache *cachep,
2912 gfp_t flags, int nodeid, void *objp)
2917 struct kmem_list3 *l3;
2920 * Be lazy and only check for valid flags here, keeping it out of the
2921 * critical path in kmem_cache_alloc().
2923 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2924 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2926 /* Take the l3 list lock to change the colour_next on this node */
2928 l3 = cachep->nodelists[nodeid];
2929 spin_lock(&l3->list_lock);
2931 /* Get colour for the slab, and cal the next value. */
2932 offset = l3->colour_next;
2934 if (l3->colour_next >= cachep->colour)
2935 l3->colour_next = 0;
2936 spin_unlock(&l3->list_lock);
2938 offset *= cachep->colour_off;
2940 if (local_flags & __GFP_WAIT)
2944 * The test for missing atomic flag is performed here, rather than
2945 * the more obvious place, simply to reduce the critical path length
2946 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2947 * will eventually be caught here (where it matters).
2949 kmem_flagcheck(cachep, flags);
2952 * Get mem for the objs. Attempt to allocate a physical page from
2956 objp = kmem_getpages(cachep, local_flags, nodeid);
2960 /* Get slab management. */
2961 slabp = alloc_slabmgmt(cachep, objp, offset,
2962 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2966 slab_map_pages(cachep, slabp, objp);
2968 cache_init_objs(cachep, slabp);
2970 if (local_flags & __GFP_WAIT)
2971 local_irq_disable();
2973 spin_lock(&l3->list_lock);
2975 /* Make slab active. */
2976 list_add_tail(&slabp->list, &(l3->slabs_free));
2977 STATS_INC_GROWN(cachep);
2978 l3->free_objects += cachep->num;
2979 spin_unlock(&l3->list_lock);
2982 kmem_freepages(cachep, objp);
2984 if (local_flags & __GFP_WAIT)
2985 local_irq_disable();
2992 * Perform extra freeing checks:
2993 * - detect bad pointers.
2994 * - POISON/RED_ZONE checking
2996 static void kfree_debugcheck(const void *objp)
2998 if (!virt_addr_valid(objp)) {
2999 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3000 (unsigned long)objp);
3005 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3007 unsigned long long redzone1, redzone2;
3009 redzone1 = *dbg_redzone1(cache, obj);
3010 redzone2 = *dbg_redzone2(cache, obj);
3015 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3018 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3019 slab_error(cache, "double free detected");
3021 slab_error(cache, "memory outside object was overwritten");
3023 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3024 obj, redzone1, redzone2);
3027 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3034 BUG_ON(virt_to_cache(objp) != cachep);
3036 objp -= obj_offset(cachep);
3037 kfree_debugcheck(objp);
3038 page = virt_to_head_page(objp);
3040 slabp = page->slab_page;
3042 if (cachep->flags & SLAB_RED_ZONE) {
3043 verify_redzone_free(cachep, objp);
3044 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3045 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3047 if (cachep->flags & SLAB_STORE_USER)
3048 *dbg_userword(cachep, objp) = caller;
3050 objnr = obj_to_index(cachep, slabp, objp);
3052 BUG_ON(objnr >= cachep->num);
3053 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3055 #ifdef CONFIG_DEBUG_SLAB_LEAK
3056 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3058 if (cachep->flags & SLAB_POISON) {
3059 #ifdef CONFIG_DEBUG_PAGEALLOC
3060 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3061 store_stackinfo(cachep, objp, (unsigned long)caller);
3062 kernel_map_pages(virt_to_page(objp),
3063 cachep->size / PAGE_SIZE, 0);
3065 poison_obj(cachep, objp, POISON_FREE);
3068 poison_obj(cachep, objp, POISON_FREE);
3074 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3079 /* Check slab's freelist to see if this obj is there. */
3080 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3082 if (entries > cachep->num || i >= cachep->num)
3085 if (entries != cachep->num - slabp->inuse) {
3087 printk(KERN_ERR "slab: Internal list corruption detected in "
3088 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3089 cachep->name, cachep->num, slabp, slabp->inuse,
3091 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3092 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3098 #define kfree_debugcheck(x) do { } while(0)
3099 #define cache_free_debugcheck(x,objp,z) (objp)
3100 #define check_slabp(x,y) do { } while(0)
3103 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3106 struct kmem_list3 *l3;
3107 struct array_cache *ac;
3112 node = numa_mem_id();
3113 ac = cpu_cache_get(cachep);
3114 batchcount = ac->batchcount;
3115 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3117 * If there was little recent activity on this cache, then
3118 * perform only a partial refill. Otherwise we could generate
3121 batchcount = BATCHREFILL_LIMIT;
3123 l3 = cachep->nodelists[node];
3125 BUG_ON(ac->avail > 0 || !l3);
3126 spin_lock(&l3->list_lock);
3128 /* See if we can refill from the shared array */
3129 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3130 l3->shared->touched = 1;
3134 while (batchcount > 0) {
3135 struct list_head *entry;
3137 /* Get slab alloc is to come from. */
3138 entry = l3->slabs_partial.next;
3139 if (entry == &l3->slabs_partial) {
3140 l3->free_touched = 1;
3141 entry = l3->slabs_free.next;
3142 if (entry == &l3->slabs_free)
3146 slabp = list_entry(entry, struct slab, list);
3147 check_slabp(cachep, slabp);
3148 check_spinlock_acquired(cachep);
3151 * The slab was either on partial or free list so
3152 * there must be at least one object available for
3155 BUG_ON(slabp->inuse >= cachep->num);
3157 while (slabp->inuse < cachep->num && batchcount--) {
3158 STATS_INC_ALLOCED(cachep);
3159 STATS_INC_ACTIVE(cachep);
3160 STATS_SET_HIGH(cachep);
3162 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3165 check_slabp(cachep, slabp);
3167 /* move slabp to correct slabp list: */
3168 list_del(&slabp->list);
3169 if (slabp->free == BUFCTL_END)
3170 list_add(&slabp->list, &l3->slabs_full);
3172 list_add(&slabp->list, &l3->slabs_partial);
3176 l3->free_objects -= ac->avail;
3178 spin_unlock(&l3->list_lock);
3180 if (unlikely(!ac->avail)) {
3182 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3184 /* cache_grow can reenable interrupts, then ac could change. */
3185 ac = cpu_cache_get(cachep);
3186 if (!x && ac->avail == 0) /* no objects in sight? abort */
3189 if (!ac->avail) /* objects refilled by interrupt? */
3193 return ac->entry[--ac->avail];
3196 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3199 might_sleep_if(flags & __GFP_WAIT);
3201 kmem_flagcheck(cachep, flags);
3206 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3207 gfp_t flags, void *objp, void *caller)
3211 if (cachep->flags & SLAB_POISON) {
3212 #ifdef CONFIG_DEBUG_PAGEALLOC
3213 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3214 kernel_map_pages(virt_to_page(objp),
3215 cachep->size / PAGE_SIZE, 1);
3217 check_poison_obj(cachep, objp);
3219 check_poison_obj(cachep, objp);
3221 poison_obj(cachep, objp, POISON_INUSE);
3223 if (cachep->flags & SLAB_STORE_USER)
3224 *dbg_userword(cachep, objp) = caller;
3226 if (cachep->flags & SLAB_RED_ZONE) {
3227 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3228 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3229 slab_error(cachep, "double free, or memory outside"
3230 " object was overwritten");
3232 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3233 objp, *dbg_redzone1(cachep, objp),
3234 *dbg_redzone2(cachep, objp));
3236 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3237 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3239 #ifdef CONFIG_DEBUG_SLAB_LEAK
3244 slabp = virt_to_head_page(objp)->slab_page;
3245 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3246 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3249 objp += obj_offset(cachep);
3250 if (cachep->ctor && cachep->flags & SLAB_POISON)
3252 if (ARCH_SLAB_MINALIGN &&
3253 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3254 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3255 objp, (int)ARCH_SLAB_MINALIGN);
3260 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3263 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3265 if (cachep == &cache_cache)
3268 return should_failslab(obj_size(cachep), flags, cachep->flags);
3271 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3274 struct array_cache *ac;
3278 ac = cpu_cache_get(cachep);
3279 if (likely(ac->avail)) {
3280 STATS_INC_ALLOCHIT(cachep);
3282 objp = ac->entry[--ac->avail];
3284 STATS_INC_ALLOCMISS(cachep);
3285 objp = cache_alloc_refill(cachep, flags);
3287 * the 'ac' may be updated by cache_alloc_refill(),
3288 * and kmemleak_erase() requires its correct value.
3290 ac = cpu_cache_get(cachep);
3293 * To avoid a false negative, if an object that is in one of the
3294 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3295 * treat the array pointers as a reference to the object.
3298 kmemleak_erase(&ac->entry[ac->avail]);
3304 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3306 * If we are in_interrupt, then process context, including cpusets and
3307 * mempolicy, may not apply and should not be used for allocation policy.
3309 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3311 int nid_alloc, nid_here;
3313 if (in_interrupt() || (flags & __GFP_THISNODE))
3315 nid_alloc = nid_here = numa_mem_id();
3316 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3317 nid_alloc = cpuset_slab_spread_node();
3318 else if (current->mempolicy)
3319 nid_alloc = slab_node(current->mempolicy);
3320 if (nid_alloc != nid_here)
3321 return ____cache_alloc_node(cachep, flags, nid_alloc);
3326 * Fallback function if there was no memory available and no objects on a
3327 * certain node and fall back is permitted. First we scan all the
3328 * available nodelists for available objects. If that fails then we
3329 * perform an allocation without specifying a node. This allows the page
3330 * allocator to do its reclaim / fallback magic. We then insert the
3331 * slab into the proper nodelist and then allocate from it.
3333 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3335 struct zonelist *zonelist;
3339 enum zone_type high_zoneidx = gfp_zone(flags);
3342 unsigned int cpuset_mems_cookie;
3344 if (flags & __GFP_THISNODE)
3347 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3350 cpuset_mems_cookie = get_mems_allowed();
3351 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3355 * Look through allowed nodes for objects available
3356 * from existing per node queues.
3358 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3359 nid = zone_to_nid(zone);
3361 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3362 cache->nodelists[nid] &&
3363 cache->nodelists[nid]->free_objects) {
3364 obj = ____cache_alloc_node(cache,
3365 flags | GFP_THISNODE, nid);
3373 * This allocation will be performed within the constraints
3374 * of the current cpuset / memory policy requirements.
3375 * We may trigger various forms of reclaim on the allowed
3376 * set and go into memory reserves if necessary.
3378 if (local_flags & __GFP_WAIT)
3380 kmem_flagcheck(cache, flags);
3381 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3382 if (local_flags & __GFP_WAIT)
3383 local_irq_disable();
3386 * Insert into the appropriate per node queues
3388 nid = page_to_nid(virt_to_page(obj));
3389 if (cache_grow(cache, flags, nid, obj)) {
3390 obj = ____cache_alloc_node(cache,
3391 flags | GFP_THISNODE, nid);
3394 * Another processor may allocate the
3395 * objects in the slab since we are
3396 * not holding any locks.
3400 /* cache_grow already freed obj */
3406 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3412 * A interface to enable slab creation on nodeid
3414 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3417 struct list_head *entry;
3419 struct kmem_list3 *l3;
3423 l3 = cachep->nodelists[nodeid];
3428 spin_lock(&l3->list_lock);
3429 entry = l3->slabs_partial.next;
3430 if (entry == &l3->slabs_partial) {
3431 l3->free_touched = 1;
3432 entry = l3->slabs_free.next;
3433 if (entry == &l3->slabs_free)
3437 slabp = list_entry(entry, struct slab, list);
3438 check_spinlock_acquired_node(cachep, nodeid);
3439 check_slabp(cachep, slabp);
3441 STATS_INC_NODEALLOCS(cachep);
3442 STATS_INC_ACTIVE(cachep);
3443 STATS_SET_HIGH(cachep);
3445 BUG_ON(slabp->inuse == cachep->num);
3447 obj = slab_get_obj(cachep, slabp, nodeid);
3448 check_slabp(cachep, slabp);
3450 /* move slabp to correct slabp list: */
3451 list_del(&slabp->list);
3453 if (slabp->free == BUFCTL_END)
3454 list_add(&slabp->list, &l3->slabs_full);
3456 list_add(&slabp->list, &l3->slabs_partial);
3458 spin_unlock(&l3->list_lock);
3462 spin_unlock(&l3->list_lock);
3463 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3467 return fallback_alloc(cachep, flags);
3474 * kmem_cache_alloc_node - Allocate an object on the specified node
3475 * @cachep: The cache to allocate from.
3476 * @flags: See kmalloc().
3477 * @nodeid: node number of the target node.
3478 * @caller: return address of caller, used for debug information
3480 * Identical to kmem_cache_alloc but it will allocate memory on the given
3481 * node, which can improve the performance for cpu bound structures.
3483 * Fallback to other node is possible if __GFP_THISNODE is not set.
3485 static __always_inline void *
3486 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3489 unsigned long save_flags;
3491 int slab_node = numa_mem_id();
3493 flags &= gfp_allowed_mask;
3495 lockdep_trace_alloc(flags);
3497 if (slab_should_failslab(cachep, flags))
3500 cache_alloc_debugcheck_before(cachep, flags);
3501 local_irq_save(save_flags);
3503 if (nodeid == NUMA_NO_NODE)
3506 if (unlikely(!cachep->nodelists[nodeid])) {
3507 /* Node not bootstrapped yet */
3508 ptr = fallback_alloc(cachep, flags);
3512 if (nodeid == slab_node) {
3514 * Use the locally cached objects if possible.
3515 * However ____cache_alloc does not allow fallback
3516 * to other nodes. It may fail while we still have
3517 * objects on other nodes available.
3519 ptr = ____cache_alloc(cachep, flags);
3523 /* ___cache_alloc_node can fall back to other nodes */
3524 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3526 local_irq_restore(save_flags);
3527 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3528 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3532 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3534 if (unlikely((flags & __GFP_ZERO) && ptr))
3535 memset(ptr, 0, obj_size(cachep));
3540 static __always_inline void *
3541 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3545 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3546 objp = alternate_node_alloc(cache, flags);
3550 objp = ____cache_alloc(cache, flags);
3553 * We may just have run out of memory on the local node.
3554 * ____cache_alloc_node() knows how to locate memory on other nodes
3557 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3564 static __always_inline void *
3565 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3567 return ____cache_alloc(cachep, flags);
3570 #endif /* CONFIG_NUMA */
3572 static __always_inline void *
3573 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3575 unsigned long save_flags;
3578 flags &= gfp_allowed_mask;
3580 lockdep_trace_alloc(flags);
3582 if (slab_should_failslab(cachep, flags))
3585 cache_alloc_debugcheck_before(cachep, flags);
3586 local_irq_save(save_flags);
3587 objp = __do_cache_alloc(cachep, flags);
3588 local_irq_restore(save_flags);
3589 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3590 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3595 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3597 if (unlikely((flags & __GFP_ZERO) && objp))
3598 memset(objp, 0, obj_size(cachep));
3604 * Caller needs to acquire correct kmem_list's list_lock
3606 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3610 struct kmem_list3 *l3;
3612 for (i = 0; i < nr_objects; i++) {
3613 void *objp = objpp[i];
3616 slabp = virt_to_slab(objp);
3617 l3 = cachep->nodelists[node];
3618 list_del(&slabp->list);
3619 check_spinlock_acquired_node(cachep, node);
3620 check_slabp(cachep, slabp);
3621 slab_put_obj(cachep, slabp, objp, node);
3622 STATS_DEC_ACTIVE(cachep);
3624 check_slabp(cachep, slabp);
3626 /* fixup slab chains */
3627 if (slabp->inuse == 0) {
3628 if (l3->free_objects > l3->free_limit) {
3629 l3->free_objects -= cachep->num;
3630 /* No need to drop any previously held
3631 * lock here, even if we have a off-slab slab
3632 * descriptor it is guaranteed to come from
3633 * a different cache, refer to comments before
3636 slab_destroy(cachep, slabp);
3638 list_add(&slabp->list, &l3->slabs_free);
3641 /* Unconditionally move a slab to the end of the
3642 * partial list on free - maximum time for the
3643 * other objects to be freed, too.
3645 list_add_tail(&slabp->list, &l3->slabs_partial);
3650 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3653 struct kmem_list3 *l3;
3654 int node = numa_mem_id();
3656 batchcount = ac->batchcount;
3658 BUG_ON(!batchcount || batchcount > ac->avail);
3661 l3 = cachep->nodelists[node];
3662 spin_lock(&l3->list_lock);
3664 struct array_cache *shared_array = l3->shared;
3665 int max = shared_array->limit - shared_array->avail;
3667 if (batchcount > max)
3669 memcpy(&(shared_array->entry[shared_array->avail]),
3670 ac->entry, sizeof(void *) * batchcount);
3671 shared_array->avail += batchcount;
3676 free_block(cachep, ac->entry, batchcount, node);
3681 struct list_head *p;
3683 p = l3->slabs_free.next;
3684 while (p != &(l3->slabs_free)) {
3687 slabp = list_entry(p, struct slab, list);
3688 BUG_ON(slabp->inuse);
3693 STATS_SET_FREEABLE(cachep, i);
3696 spin_unlock(&l3->list_lock);
3697 ac->avail -= batchcount;
3698 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3702 * Release an obj back to its cache. If the obj has a constructed state, it must
3703 * be in this state _before_ it is released. Called with disabled ints.
3705 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3708 struct array_cache *ac = cpu_cache_get(cachep);
3711 kmemleak_free_recursive(objp, cachep->flags);
3712 objp = cache_free_debugcheck(cachep, objp, caller);
3714 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3717 * Skip calling cache_free_alien() when the platform is not numa.
3718 * This will avoid cache misses that happen while accessing slabp (which
3719 * is per page memory reference) to get nodeid. Instead use a global
3720 * variable to skip the call, which is mostly likely to be present in
3723 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3726 if (likely(ac->avail < ac->limit)) {
3727 STATS_INC_FREEHIT(cachep);
3729 STATS_INC_FREEMISS(cachep);
3730 cache_flusharray(cachep, ac);
3733 ac->entry[ac->avail++] = objp;
3737 * kmem_cache_alloc - Allocate an object
3738 * @cachep: The cache to allocate from.
3739 * @flags: See kmalloc().
3741 * Allocate an object from this cache. The flags are only relevant
3742 * if the cache has no available objects.
3744 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3746 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3748 trace_kmem_cache_alloc(_RET_IP_, ret,
3749 obj_size(cachep), cachep->size, flags);
3753 EXPORT_SYMBOL(kmem_cache_alloc);
3755 #ifdef CONFIG_TRACING
3757 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3761 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3763 trace_kmalloc(_RET_IP_, ret,
3764 size, slab_buffer_size(cachep), flags);
3767 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3771 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3773 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3774 __builtin_return_address(0));
3776 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3777 obj_size(cachep), cachep->size,
3782 EXPORT_SYMBOL(kmem_cache_alloc_node);
3784 #ifdef CONFIG_TRACING
3785 void *kmem_cache_alloc_node_trace(size_t size,
3786 struct kmem_cache *cachep,
3792 ret = __cache_alloc_node(cachep, flags, nodeid,
3793 __builtin_return_address(0));
3794 trace_kmalloc_node(_RET_IP_, ret,
3795 size, slab_buffer_size(cachep),
3799 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3802 static __always_inline void *
3803 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3805 struct kmem_cache *cachep;
3807 cachep = kmem_find_general_cachep(size, flags);
3808 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3810 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3813 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3814 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3816 return __do_kmalloc_node(size, flags, node,
3817 __builtin_return_address(0));
3819 EXPORT_SYMBOL(__kmalloc_node);
3821 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3822 int node, unsigned long caller)
3824 return __do_kmalloc_node(size, flags, node, (void *)caller);
3826 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3828 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3830 return __do_kmalloc_node(size, flags, node, NULL);
3832 EXPORT_SYMBOL(__kmalloc_node);
3833 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3834 #endif /* CONFIG_NUMA */
3837 * __do_kmalloc - allocate memory
3838 * @size: how many bytes of memory are required.
3839 * @flags: the type of memory to allocate (see kmalloc).
3840 * @caller: function caller for debug tracking of the caller
3842 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3845 struct kmem_cache *cachep;
3848 /* If you want to save a few bytes .text space: replace
3850 * Then kmalloc uses the uninlined functions instead of the inline
3853 cachep = __find_general_cachep(size, flags);
3854 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3856 ret = __cache_alloc(cachep, flags, caller);
3858 trace_kmalloc((unsigned long) caller, ret,
3859 size, cachep->size, flags);
3865 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3866 void *__kmalloc(size_t size, gfp_t flags)
3868 return __do_kmalloc(size, flags, __builtin_return_address(0));
3870 EXPORT_SYMBOL(__kmalloc);
3872 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3874 return __do_kmalloc(size, flags, (void *)caller);
3876 EXPORT_SYMBOL(__kmalloc_track_caller);
3879 void *__kmalloc(size_t size, gfp_t flags)
3881 return __do_kmalloc(size, flags, NULL);
3883 EXPORT_SYMBOL(__kmalloc);
3887 * kmem_cache_free - Deallocate an object
3888 * @cachep: The cache the allocation was from.
3889 * @objp: The previously allocated object.
3891 * Free an object which was previously allocated from this
3894 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3896 unsigned long flags;
3898 local_irq_save(flags);
3899 debug_check_no_locks_freed(objp, obj_size(cachep));
3900 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3901 debug_check_no_obj_freed(objp, obj_size(cachep));
3902 __cache_free(cachep, objp, __builtin_return_address(0));
3903 local_irq_restore(flags);
3905 trace_kmem_cache_free(_RET_IP_, objp);
3907 EXPORT_SYMBOL(kmem_cache_free);
3910 * kfree - free previously allocated memory
3911 * @objp: pointer returned by kmalloc.
3913 * If @objp is NULL, no operation is performed.
3915 * Don't free memory not originally allocated by kmalloc()
3916 * or you will run into trouble.
3918 void kfree(const void *objp)
3920 struct kmem_cache *c;
3921 unsigned long flags;
3923 trace_kfree(_RET_IP_, objp);
3925 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3927 local_irq_save(flags);
3928 kfree_debugcheck(objp);
3929 c = virt_to_cache(objp);
3930 debug_check_no_locks_freed(objp, obj_size(c));
3931 debug_check_no_obj_freed(objp, obj_size(c));
3932 __cache_free(c, (void *)objp, __builtin_return_address(0));
3933 local_irq_restore(flags);
3935 EXPORT_SYMBOL(kfree);
3937 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3939 return obj_size(cachep);
3941 EXPORT_SYMBOL(kmem_cache_size);
3944 * This initializes kmem_list3 or resizes various caches for all nodes.
3946 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3949 struct kmem_list3 *l3;
3950 struct array_cache *new_shared;
3951 struct array_cache **new_alien = NULL;
3953 for_each_online_node(node) {
3955 if (use_alien_caches) {
3956 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3962 if (cachep->shared) {
3963 new_shared = alloc_arraycache(node,
3964 cachep->shared*cachep->batchcount,
3967 free_alien_cache(new_alien);
3972 l3 = cachep->nodelists[node];
3974 struct array_cache *shared = l3->shared;
3976 spin_lock_irq(&l3->list_lock);
3979 free_block(cachep, shared->entry,
3980 shared->avail, node);
3982 l3->shared = new_shared;
3984 l3->alien = new_alien;
3987 l3->free_limit = (1 + nr_cpus_node(node)) *
3988 cachep->batchcount + cachep->num;
3989 spin_unlock_irq(&l3->list_lock);
3991 free_alien_cache(new_alien);
3994 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3996 free_alien_cache(new_alien);
4001 kmem_list3_init(l3);
4002 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4003 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4004 l3->shared = new_shared;
4005 l3->alien = new_alien;
4006 l3->free_limit = (1 + nr_cpus_node(node)) *
4007 cachep->batchcount + cachep->num;
4008 cachep->nodelists[node] = l3;
4013 if (!cachep->list.next) {
4014 /* Cache is not active yet. Roll back what we did */
4017 if (cachep->nodelists[node]) {
4018 l3 = cachep->nodelists[node];
4021 free_alien_cache(l3->alien);
4023 cachep->nodelists[node] = NULL;
4031 struct ccupdate_struct {
4032 struct kmem_cache *cachep;
4033 struct array_cache *new[0];
4036 static void do_ccupdate_local(void *info)
4038 struct ccupdate_struct *new = info;
4039 struct array_cache *old;
4042 old = cpu_cache_get(new->cachep);
4044 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4045 new->new[smp_processor_id()] = old;
4048 /* Always called with the cache_chain_mutex held */
4049 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4050 int batchcount, int shared, gfp_t gfp)
4052 struct ccupdate_struct *new;
4055 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4060 for_each_online_cpu(i) {
4061 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4064 for (i--; i >= 0; i--)
4070 new->cachep = cachep;
4072 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4075 cachep->batchcount = batchcount;
4076 cachep->limit = limit;
4077 cachep->shared = shared;
4079 for_each_online_cpu(i) {
4080 struct array_cache *ccold = new->new[i];
4083 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4084 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4085 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4089 return alloc_kmemlist(cachep, gfp);
4092 /* Called with cache_chain_mutex held always */
4093 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4099 * The head array serves three purposes:
4100 * - create a LIFO ordering, i.e. return objects that are cache-warm
4101 * - reduce the number of spinlock operations.
4102 * - reduce the number of linked list operations on the slab and
4103 * bufctl chains: array operations are cheaper.
4104 * The numbers are guessed, we should auto-tune as described by
4107 if (cachep->size > 131072)
4109 else if (cachep->size > PAGE_SIZE)
4111 else if (cachep->size > 1024)
4113 else if (cachep->size > 256)
4119 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4120 * allocation behaviour: Most allocs on one cpu, most free operations
4121 * on another cpu. For these cases, an efficient object passing between
4122 * cpus is necessary. This is provided by a shared array. The array
4123 * replaces Bonwick's magazine layer.
4124 * On uniprocessor, it's functionally equivalent (but less efficient)
4125 * to a larger limit. Thus disabled by default.
4128 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4133 * With debugging enabled, large batchcount lead to excessively long
4134 * periods with disabled local interrupts. Limit the batchcount
4139 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4141 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4142 cachep->name, -err);
4147 * Drain an array if it contains any elements taking the l3 lock only if
4148 * necessary. Note that the l3 listlock also protects the array_cache
4149 * if drain_array() is used on the shared array.
4151 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4152 struct array_cache *ac, int force, int node)
4156 if (!ac || !ac->avail)
4158 if (ac->touched && !force) {
4161 spin_lock_irq(&l3->list_lock);
4163 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4164 if (tofree > ac->avail)
4165 tofree = (ac->avail + 1) / 2;
4166 free_block(cachep, ac->entry, tofree, node);
4167 ac->avail -= tofree;
4168 memmove(ac->entry, &(ac->entry[tofree]),
4169 sizeof(void *) * ac->avail);
4171 spin_unlock_irq(&l3->list_lock);
4176 * cache_reap - Reclaim memory from caches.
4177 * @w: work descriptor
4179 * Called from workqueue/eventd every few seconds.
4181 * - clear the per-cpu caches for this CPU.
4182 * - return freeable pages to the main free memory pool.
4184 * If we cannot acquire the cache chain mutex then just give up - we'll try
4185 * again on the next iteration.
4187 static void cache_reap(struct work_struct *w)
4189 struct kmem_cache *searchp;
4190 struct kmem_list3 *l3;
4191 int node = numa_mem_id();
4192 struct delayed_work *work = to_delayed_work(w);
4194 if (!mutex_trylock(&cache_chain_mutex))
4195 /* Give up. Setup the next iteration. */
4198 list_for_each_entry(searchp, &cache_chain, list) {
4202 * We only take the l3 lock if absolutely necessary and we
4203 * have established with reasonable certainty that
4204 * we can do some work if the lock was obtained.
4206 l3 = searchp->nodelists[node];
4208 reap_alien(searchp, l3);
4210 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4213 * These are racy checks but it does not matter
4214 * if we skip one check or scan twice.
4216 if (time_after(l3->next_reap, jiffies))
4219 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4221 drain_array(searchp, l3, l3->shared, 0, node);
4223 if (l3->free_touched)
4224 l3->free_touched = 0;
4228 freed = drain_freelist(searchp, l3, (l3->free_limit +
4229 5 * searchp->num - 1) / (5 * searchp->num));
4230 STATS_ADD_REAPED(searchp, freed);
4236 mutex_unlock(&cache_chain_mutex);
4239 /* Set up the next iteration */
4240 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4243 #ifdef CONFIG_SLABINFO
4245 static void print_slabinfo_header(struct seq_file *m)
4248 * Output format version, so at least we can change it
4249 * without _too_ many complaints.
4252 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4254 seq_puts(m, "slabinfo - version: 2.1\n");
4256 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4257 "<objperslab> <pagesperslab>");
4258 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4259 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4261 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4262 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4263 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4268 static void *s_start(struct seq_file *m, loff_t *pos)
4272 mutex_lock(&cache_chain_mutex);
4274 print_slabinfo_header(m);
4276 return seq_list_start(&cache_chain, *pos);
4279 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4281 return seq_list_next(p, &cache_chain, pos);
4284 static void s_stop(struct seq_file *m, void *p)
4286 mutex_unlock(&cache_chain_mutex);
4289 static int s_show(struct seq_file *m, void *p)
4291 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4293 unsigned long active_objs;
4294 unsigned long num_objs;
4295 unsigned long active_slabs = 0;
4296 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4300 struct kmem_list3 *l3;
4304 for_each_online_node(node) {
4305 l3 = cachep->nodelists[node];
4310 spin_lock_irq(&l3->list_lock);
4312 list_for_each_entry(slabp, &l3->slabs_full, list) {
4313 if (slabp->inuse != cachep->num && !error)
4314 error = "slabs_full accounting error";
4315 active_objs += cachep->num;
4318 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4319 if (slabp->inuse == cachep->num && !error)
4320 error = "slabs_partial inuse accounting error";
4321 if (!slabp->inuse && !error)
4322 error = "slabs_partial/inuse accounting error";
4323 active_objs += slabp->inuse;
4326 list_for_each_entry(slabp, &l3->slabs_free, list) {
4327 if (slabp->inuse && !error)
4328 error = "slabs_free/inuse accounting error";
4331 free_objects += l3->free_objects;
4333 shared_avail += l3->shared->avail;
4335 spin_unlock_irq(&l3->list_lock);
4337 num_slabs += active_slabs;
4338 num_objs = num_slabs * cachep->num;
4339 if (num_objs - active_objs != free_objects && !error)
4340 error = "free_objects accounting error";
4342 name = cachep->name;
4344 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4346 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4347 name, active_objs, num_objs, cachep->size,
4348 cachep->num, (1 << cachep->gfporder));
4349 seq_printf(m, " : tunables %4u %4u %4u",
4350 cachep->limit, cachep->batchcount, cachep->shared);
4351 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4352 active_slabs, num_slabs, shared_avail);
4355 unsigned long high = cachep->high_mark;
4356 unsigned long allocs = cachep->num_allocations;
4357 unsigned long grown = cachep->grown;
4358 unsigned long reaped = cachep->reaped;
4359 unsigned long errors = cachep->errors;
4360 unsigned long max_freeable = cachep->max_freeable;
4361 unsigned long node_allocs = cachep->node_allocs;
4362 unsigned long node_frees = cachep->node_frees;
4363 unsigned long overflows = cachep->node_overflow;
4365 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4366 "%4lu %4lu %4lu %4lu %4lu",
4367 allocs, high, grown,
4368 reaped, errors, max_freeable, node_allocs,
4369 node_frees, overflows);
4373 unsigned long allochit = atomic_read(&cachep->allochit);
4374 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4375 unsigned long freehit = atomic_read(&cachep->freehit);
4376 unsigned long freemiss = atomic_read(&cachep->freemiss);
4378 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4379 allochit, allocmiss, freehit, freemiss);
4387 * slabinfo_op - iterator that generates /proc/slabinfo
4396 * num-pages-per-slab
4397 * + further values on SMP and with statistics enabled
4400 static const struct seq_operations slabinfo_op = {
4407 #define MAX_SLABINFO_WRITE 128
4409 * slabinfo_write - Tuning for the slab allocator
4411 * @buffer: user buffer
4412 * @count: data length
4415 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4416 size_t count, loff_t *ppos)
4418 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4419 int limit, batchcount, shared, res;
4420 struct kmem_cache *cachep;
4422 if (count > MAX_SLABINFO_WRITE)
4424 if (copy_from_user(&kbuf, buffer, count))
4426 kbuf[MAX_SLABINFO_WRITE] = '\0';
4428 tmp = strchr(kbuf, ' ');
4433 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4436 /* Find the cache in the chain of caches. */
4437 mutex_lock(&cache_chain_mutex);
4439 list_for_each_entry(cachep, &cache_chain, list) {
4440 if (!strcmp(cachep->name, kbuf)) {
4441 if (limit < 1 || batchcount < 1 ||
4442 batchcount > limit || shared < 0) {
4445 res = do_tune_cpucache(cachep, limit,
4452 mutex_unlock(&cache_chain_mutex);
4458 static int slabinfo_open(struct inode *inode, struct file *file)
4460 return seq_open(file, &slabinfo_op);
4463 static const struct file_operations proc_slabinfo_operations = {
4464 .open = slabinfo_open,
4466 .write = slabinfo_write,
4467 .llseek = seq_lseek,
4468 .release = seq_release,
4471 #ifdef CONFIG_DEBUG_SLAB_LEAK
4473 static void *leaks_start(struct seq_file *m, loff_t *pos)
4475 mutex_lock(&cache_chain_mutex);
4476 return seq_list_start(&cache_chain, *pos);
4479 static inline int add_caller(unsigned long *n, unsigned long v)
4489 unsigned long *q = p + 2 * i;
4503 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4509 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4515 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4516 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4518 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4523 static void show_symbol(struct seq_file *m, unsigned long address)
4525 #ifdef CONFIG_KALLSYMS
4526 unsigned long offset, size;
4527 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4529 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4530 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4532 seq_printf(m, " [%s]", modname);
4536 seq_printf(m, "%p", (void *)address);
4539 static int leaks_show(struct seq_file *m, void *p)
4541 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4543 struct kmem_list3 *l3;
4545 unsigned long *n = m->private;
4549 if (!(cachep->flags & SLAB_STORE_USER))
4551 if (!(cachep->flags & SLAB_RED_ZONE))
4554 /* OK, we can do it */
4558 for_each_online_node(node) {
4559 l3 = cachep->nodelists[node];
4564 spin_lock_irq(&l3->list_lock);
4566 list_for_each_entry(slabp, &l3->slabs_full, list)
4567 handle_slab(n, cachep, slabp);
4568 list_for_each_entry(slabp, &l3->slabs_partial, list)
4569 handle_slab(n, cachep, slabp);
4570 spin_unlock_irq(&l3->list_lock);
4572 name = cachep->name;
4574 /* Increase the buffer size */
4575 mutex_unlock(&cache_chain_mutex);
4576 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4578 /* Too bad, we are really out */
4580 mutex_lock(&cache_chain_mutex);
4583 *(unsigned long *)m->private = n[0] * 2;
4585 mutex_lock(&cache_chain_mutex);
4586 /* Now make sure this entry will be retried */
4590 for (i = 0; i < n[1]; i++) {
4591 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4592 show_symbol(m, n[2*i+2]);
4599 static const struct seq_operations slabstats_op = {
4600 .start = leaks_start,
4606 static int slabstats_open(struct inode *inode, struct file *file)
4608 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4611 ret = seq_open(file, &slabstats_op);
4613 struct seq_file *m = file->private_data;
4614 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4623 static const struct file_operations proc_slabstats_operations = {
4624 .open = slabstats_open,
4626 .llseek = seq_lseek,
4627 .release = seq_release_private,
4631 static int __init slab_proc_init(void)
4633 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4634 #ifdef CONFIG_DEBUG_SLAB_LEAK
4635 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4639 module_init(slab_proc_init);
4643 * ksize - get the actual amount of memory allocated for a given object
4644 * @objp: Pointer to the object
4646 * kmalloc may internally round up allocations and return more memory
4647 * than requested. ksize() can be used to determine the actual amount of
4648 * memory allocated. The caller may use this additional memory, even though
4649 * a smaller amount of memory was initially specified with the kmalloc call.
4650 * The caller must guarantee that objp points to a valid object previously
4651 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4652 * must not be freed during the duration of the call.
4654 size_t ksize(const void *objp)
4657 if (unlikely(objp == ZERO_SIZE_PTR))
4660 return obj_size(virt_to_cache(objp));
4662 EXPORT_SYMBOL(ksize);