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 unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
438 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
439 return (unsigned long long*) (objp + obj_offset(cachep) -
440 sizeof(unsigned long long));
443 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
445 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
446 if (cachep->flags & SLAB_STORE_USER)
447 return (unsigned long long *)(objp + cachep->size -
448 sizeof(unsigned long long) -
450 return (unsigned long long *) (objp + cachep->size -
451 sizeof(unsigned long long));
454 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
456 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
457 return (void **)(objp + cachep->size - BYTES_PER_WORD);
462 #define obj_offset(x) 0
463 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
469 #ifdef CONFIG_TRACING
470 size_t slab_buffer_size(struct kmem_cache *cachep)
474 EXPORT_SYMBOL(slab_buffer_size);
478 * Do not go above this order unless 0 objects fit into the slab or
479 * overridden on the command line.
481 #define SLAB_MAX_ORDER_HI 1
482 #define SLAB_MAX_ORDER_LO 0
483 static int slab_max_order = SLAB_MAX_ORDER_LO;
484 static bool slab_max_order_set __initdata;
486 static inline struct kmem_cache *page_get_cache(struct page *page)
488 page = compound_head(page);
489 BUG_ON(!PageSlab(page));
490 return page->slab_cache;
493 static inline struct kmem_cache *virt_to_cache(const void *obj)
495 struct page *page = virt_to_head_page(obj);
496 return page->slab_cache;
499 static inline struct slab *virt_to_slab(const void *obj)
501 struct page *page = virt_to_head_page(obj);
503 VM_BUG_ON(!PageSlab(page));
504 return page->slab_page;
507 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
510 return slab->s_mem + cache->size * idx;
514 * We want to avoid an expensive divide : (offset / cache->size)
515 * Using the fact that size is a constant for a particular cache,
516 * we can replace (offset / cache->size) by
517 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
519 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
520 const struct slab *slab, void *obj)
522 u32 offset = (obj - slab->s_mem);
523 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
527 * These are the default caches for kmalloc. Custom caches can have other sizes.
529 struct cache_sizes malloc_sizes[] = {
530 #define CACHE(x) { .cs_size = (x) },
531 #include <linux/kmalloc_sizes.h>
535 EXPORT_SYMBOL(malloc_sizes);
537 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
543 static struct cache_names __initdata cache_names[] = {
544 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
545 #include <linux/kmalloc_sizes.h>
550 static struct arraycache_init initarray_cache __initdata =
551 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
552 static struct arraycache_init initarray_generic =
553 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
555 /* internal cache of cache description objs */
556 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
557 static struct kmem_cache cache_cache = {
558 .nodelists = cache_cache_nodelists,
560 .limit = BOOT_CPUCACHE_ENTRIES,
562 .size = sizeof(struct kmem_cache),
563 .name = "kmem_cache",
566 #define BAD_ALIEN_MAGIC 0x01020304ul
569 * chicken and egg problem: delay the per-cpu array allocation
570 * until the general caches are up.
582 * used by boot code to determine if it can use slab based allocator
584 int slab_is_available(void)
586 return g_cpucache_up >= EARLY;
589 #ifdef CONFIG_LOCKDEP
592 * Slab sometimes uses the kmalloc slabs to store the slab headers
593 * for other slabs "off slab".
594 * The locking for this is tricky in that it nests within the locks
595 * of all other slabs in a few places; to deal with this special
596 * locking we put on-slab caches into a separate lock-class.
598 * We set lock class for alien array caches which are up during init.
599 * The lock annotation will be lost if all cpus of a node goes down and
600 * then comes back up during hotplug
602 static struct lock_class_key on_slab_l3_key;
603 static struct lock_class_key on_slab_alc_key;
605 static struct lock_class_key debugobj_l3_key;
606 static struct lock_class_key debugobj_alc_key;
608 static void slab_set_lock_classes(struct kmem_cache *cachep,
609 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
612 struct array_cache **alc;
613 struct kmem_list3 *l3;
616 l3 = cachep->nodelists[q];
620 lockdep_set_class(&l3->list_lock, l3_key);
623 * FIXME: This check for BAD_ALIEN_MAGIC
624 * should go away when common slab code is taught to
625 * work even without alien caches.
626 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
627 * for alloc_alien_cache,
629 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
633 lockdep_set_class(&alc[r]->lock, alc_key);
637 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
639 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
642 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
646 for_each_online_node(node)
647 slab_set_debugobj_lock_classes_node(cachep, node);
650 static void init_node_lock_keys(int q)
652 struct cache_sizes *s = malloc_sizes;
654 if (g_cpucache_up < LATE)
657 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
658 struct kmem_list3 *l3;
660 l3 = s->cs_cachep->nodelists[q];
661 if (!l3 || OFF_SLAB(s->cs_cachep))
664 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
665 &on_slab_alc_key, q);
669 static inline void init_lock_keys(void)
674 init_node_lock_keys(node);
677 static void init_node_lock_keys(int q)
681 static inline void init_lock_keys(void)
685 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
689 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
695 * Guard access to the cache-chain.
697 static DEFINE_MUTEX(cache_chain_mutex);
698 static struct list_head cache_chain;
700 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
702 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
704 return cachep->array[smp_processor_id()];
707 static inline struct kmem_cache *__find_general_cachep(size_t size,
710 struct cache_sizes *csizep = malloc_sizes;
713 /* This happens if someone tries to call
714 * kmem_cache_create(), or __kmalloc(), before
715 * the generic caches are initialized.
717 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
720 return ZERO_SIZE_PTR;
722 while (size > csizep->cs_size)
726 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
727 * has cs_{dma,}cachep==NULL. Thus no special case
728 * for large kmalloc calls required.
730 #ifdef CONFIG_ZONE_DMA
731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
734 return csizep->cs_cachep;
737 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
739 return __find_general_cachep(size, gfpflags);
742 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
744 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
748 * Calculate the number of objects and left-over bytes for a given buffer size.
750 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
751 size_t align, int flags, size_t *left_over,
756 size_t slab_size = PAGE_SIZE << gfporder;
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
773 if (flags & CFLGS_OFF_SLAB) {
775 nr_objs = slab_size / buffer_size;
777 if (nr_objs > SLAB_LIMIT)
778 nr_objs = SLAB_LIMIT;
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
788 nr_objs = (slab_size - sizeof(struct slab)) /
789 (buffer_size + sizeof(kmem_bufctl_t));
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
795 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
802 mgmt_size = slab_mgmt_size(nr_objs, align);
805 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
808 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
810 static void __slab_error(const char *function, struct kmem_cache *cachep,
813 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
814 function, cachep->name, msg);
819 * By default on NUMA we use alien caches to stage the freeing of
820 * objects allocated from other nodes. This causes massive memory
821 * inefficiencies when using fake NUMA setup to split memory into a
822 * large number of small nodes, so it can be disabled on the command
826 static int use_alien_caches __read_mostly = 1;
827 static int __init noaliencache_setup(char *s)
829 use_alien_caches = 0;
832 __setup("noaliencache", noaliencache_setup);
834 static int __init slab_max_order_setup(char *str)
836 get_option(&str, &slab_max_order);
837 slab_max_order = slab_max_order < 0 ? 0 :
838 min(slab_max_order, MAX_ORDER - 1);
839 slab_max_order_set = true;
843 __setup("slab_max_order=", slab_max_order_setup);
847 * Special reaping functions for NUMA systems called from cache_reap().
848 * These take care of doing round robin flushing of alien caches (containing
849 * objects freed on different nodes from which they were allocated) and the
850 * flushing of remote pcps by calling drain_node_pages.
852 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
854 static void init_reap_node(int cpu)
858 node = next_node(cpu_to_mem(cpu), node_online_map);
859 if (node == MAX_NUMNODES)
860 node = first_node(node_online_map);
862 per_cpu(slab_reap_node, cpu) = node;
865 static void next_reap_node(void)
867 int node = __this_cpu_read(slab_reap_node);
869 node = next_node(node, node_online_map);
870 if (unlikely(node >= MAX_NUMNODES))
871 node = first_node(node_online_map);
872 __this_cpu_write(slab_reap_node, node);
876 #define init_reap_node(cpu) do { } while (0)
877 #define next_reap_node(void) do { } while (0)
881 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
882 * via the workqueue/eventd.
883 * Add the CPU number into the expiration time to minimize the possibility of
884 * the CPUs getting into lockstep and contending for the global cache chain
887 static void __cpuinit start_cpu_timer(int cpu)
889 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
892 * When this gets called from do_initcalls via cpucache_init(),
893 * init_workqueues() has already run, so keventd will be setup
896 if (keventd_up() && reap_work->work.func == NULL) {
898 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
899 schedule_delayed_work_on(cpu, reap_work,
900 __round_jiffies_relative(HZ, cpu));
904 static struct array_cache *alloc_arraycache(int node, int entries,
905 int batchcount, gfp_t gfp)
907 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
908 struct array_cache *nc = NULL;
910 nc = kmalloc_node(memsize, gfp, node);
912 * The array_cache structures contain pointers to free object.
913 * However, when such objects are allocated or transferred to another
914 * cache the pointers are not cleared and they could be counted as
915 * valid references during a kmemleak scan. Therefore, kmemleak must
916 * not scan such objects.
918 kmemleak_no_scan(nc);
922 nc->batchcount = batchcount;
924 spin_lock_init(&nc->lock);
930 * Transfer objects in one arraycache to another.
931 * Locking must be handled by the caller.
933 * Return the number of entries transferred.
935 static int transfer_objects(struct array_cache *to,
936 struct array_cache *from, unsigned int max)
938 /* Figure out how many entries to transfer */
939 int nr = min3(from->avail, max, to->limit - to->avail);
944 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
954 #define drain_alien_cache(cachep, alien) do { } while (0)
955 #define reap_alien(cachep, l3) do { } while (0)
957 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
959 return (struct array_cache **)BAD_ALIEN_MAGIC;
962 static inline void free_alien_cache(struct array_cache **ac_ptr)
966 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
971 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
977 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
978 gfp_t flags, int nodeid)
983 #else /* CONFIG_NUMA */
985 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
986 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
988 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
990 struct array_cache **ac_ptr;
991 int memsize = sizeof(void *) * nr_node_ids;
996 ac_ptr = kzalloc_node(memsize, gfp, node);
999 if (i == node || !node_online(i))
1001 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1003 for (i--; i >= 0; i--)
1013 static void free_alien_cache(struct array_cache **ac_ptr)
1024 static void __drain_alien_cache(struct kmem_cache *cachep,
1025 struct array_cache *ac, int node)
1027 struct kmem_list3 *rl3 = cachep->nodelists[node];
1030 spin_lock(&rl3->list_lock);
1032 * Stuff objects into the remote nodes shared array first.
1033 * That way we could avoid the overhead of putting the objects
1034 * into the free lists and getting them back later.
1037 transfer_objects(rl3->shared, ac, ac->limit);
1039 free_block(cachep, ac->entry, ac->avail, node);
1041 spin_unlock(&rl3->list_lock);
1046 * Called from cache_reap() to regularly drain alien caches round robin.
1048 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1050 int node = __this_cpu_read(slab_reap_node);
1053 struct array_cache *ac = l3->alien[node];
1055 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1056 __drain_alien_cache(cachep, ac, node);
1057 spin_unlock_irq(&ac->lock);
1062 static void drain_alien_cache(struct kmem_cache *cachep,
1063 struct array_cache **alien)
1066 struct array_cache *ac;
1067 unsigned long flags;
1069 for_each_online_node(i) {
1072 spin_lock_irqsave(&ac->lock, flags);
1073 __drain_alien_cache(cachep, ac, i);
1074 spin_unlock_irqrestore(&ac->lock, flags);
1079 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1081 struct slab *slabp = virt_to_slab(objp);
1082 int nodeid = slabp->nodeid;
1083 struct kmem_list3 *l3;
1084 struct array_cache *alien = NULL;
1087 node = numa_mem_id();
1090 * Make sure we are not freeing a object from another node to the array
1091 * cache on this cpu.
1093 if (likely(slabp->nodeid == node))
1096 l3 = cachep->nodelists[node];
1097 STATS_INC_NODEFREES(cachep);
1098 if (l3->alien && l3->alien[nodeid]) {
1099 alien = l3->alien[nodeid];
1100 spin_lock(&alien->lock);
1101 if (unlikely(alien->avail == alien->limit)) {
1102 STATS_INC_ACOVERFLOW(cachep);
1103 __drain_alien_cache(cachep, alien, nodeid);
1105 alien->entry[alien->avail++] = objp;
1106 spin_unlock(&alien->lock);
1108 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1109 free_block(cachep, &objp, 1, nodeid);
1110 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1117 * Allocates and initializes nodelists for a node on each slab cache, used for
1118 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1119 * will be allocated off-node since memory is not yet online for the new node.
1120 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1123 * Must hold cache_chain_mutex.
1125 static int init_cache_nodelists_node(int node)
1127 struct kmem_cache *cachep;
1128 struct kmem_list3 *l3;
1129 const int memsize = sizeof(struct kmem_list3);
1131 list_for_each_entry(cachep, &cache_chain, list) {
1133 * Set up the size64 kmemlist for cpu before we can
1134 * begin anything. Make sure some other cpu on this
1135 * node has not already allocated this
1137 if (!cachep->nodelists[node]) {
1138 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1141 kmem_list3_init(l3);
1142 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1143 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1146 * The l3s don't come and go as CPUs come and
1147 * go. cache_chain_mutex is sufficient
1150 cachep->nodelists[node] = l3;
1153 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1154 cachep->nodelists[node]->free_limit =
1155 (1 + nr_cpus_node(node)) *
1156 cachep->batchcount + cachep->num;
1157 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1162 static void __cpuinit cpuup_canceled(long cpu)
1164 struct kmem_cache *cachep;
1165 struct kmem_list3 *l3 = NULL;
1166 int node = cpu_to_mem(cpu);
1167 const struct cpumask *mask = cpumask_of_node(node);
1169 list_for_each_entry(cachep, &cache_chain, list) {
1170 struct array_cache *nc;
1171 struct array_cache *shared;
1172 struct array_cache **alien;
1174 /* cpu is dead; no one can alloc from it. */
1175 nc = cachep->array[cpu];
1176 cachep->array[cpu] = NULL;
1177 l3 = cachep->nodelists[node];
1180 goto free_array_cache;
1182 spin_lock_irq(&l3->list_lock);
1184 /* Free limit for this kmem_list3 */
1185 l3->free_limit -= cachep->batchcount;
1187 free_block(cachep, nc->entry, nc->avail, node);
1189 if (!cpumask_empty(mask)) {
1190 spin_unlock_irq(&l3->list_lock);
1191 goto free_array_cache;
1194 shared = l3->shared;
1196 free_block(cachep, shared->entry,
1197 shared->avail, node);
1204 spin_unlock_irq(&l3->list_lock);
1208 drain_alien_cache(cachep, alien);
1209 free_alien_cache(alien);
1215 * In the previous loop, all the objects were freed to
1216 * the respective cache's slabs, now we can go ahead and
1217 * shrink each nodelist to its limit.
1219 list_for_each_entry(cachep, &cache_chain, list) {
1220 l3 = cachep->nodelists[node];
1223 drain_freelist(cachep, l3, l3->free_objects);
1227 static int __cpuinit cpuup_prepare(long cpu)
1229 struct kmem_cache *cachep;
1230 struct kmem_list3 *l3 = NULL;
1231 int node = cpu_to_mem(cpu);
1235 * We need to do this right in the beginning since
1236 * alloc_arraycache's are going to use this list.
1237 * kmalloc_node allows us to add the slab to the right
1238 * kmem_list3 and not this cpu's kmem_list3
1240 err = init_cache_nodelists_node(node);
1245 * Now we can go ahead with allocating the shared arrays and
1248 list_for_each_entry(cachep, &cache_chain, list) {
1249 struct array_cache *nc;
1250 struct array_cache *shared = NULL;
1251 struct array_cache **alien = NULL;
1253 nc = alloc_arraycache(node, cachep->limit,
1254 cachep->batchcount, GFP_KERNEL);
1257 if (cachep->shared) {
1258 shared = alloc_arraycache(node,
1259 cachep->shared * cachep->batchcount,
1260 0xbaadf00d, GFP_KERNEL);
1266 if (use_alien_caches) {
1267 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1274 cachep->array[cpu] = nc;
1275 l3 = cachep->nodelists[node];
1278 spin_lock_irq(&l3->list_lock);
1281 * We are serialised from CPU_DEAD or
1282 * CPU_UP_CANCELLED by the cpucontrol lock
1284 l3->shared = shared;
1293 spin_unlock_irq(&l3->list_lock);
1295 free_alien_cache(alien);
1296 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1297 slab_set_debugobj_lock_classes_node(cachep, node);
1299 init_node_lock_keys(node);
1303 cpuup_canceled(cpu);
1307 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1308 unsigned long action, void *hcpu)
1310 long cpu = (long)hcpu;
1314 case CPU_UP_PREPARE:
1315 case CPU_UP_PREPARE_FROZEN:
1316 mutex_lock(&cache_chain_mutex);
1317 err = cpuup_prepare(cpu);
1318 mutex_unlock(&cache_chain_mutex);
1321 case CPU_ONLINE_FROZEN:
1322 start_cpu_timer(cpu);
1324 #ifdef CONFIG_HOTPLUG_CPU
1325 case CPU_DOWN_PREPARE:
1326 case CPU_DOWN_PREPARE_FROZEN:
1328 * Shutdown cache reaper. Note that the cache_chain_mutex is
1329 * held so that if cache_reap() is invoked it cannot do
1330 * anything expensive but will only modify reap_work
1331 * and reschedule the timer.
1333 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1334 /* Now the cache_reaper is guaranteed to be not running. */
1335 per_cpu(slab_reap_work, cpu).work.func = NULL;
1337 case CPU_DOWN_FAILED:
1338 case CPU_DOWN_FAILED_FROZEN:
1339 start_cpu_timer(cpu);
1342 case CPU_DEAD_FROZEN:
1344 * Even if all the cpus of a node are down, we don't free the
1345 * kmem_list3 of any cache. This to avoid a race between
1346 * cpu_down, and a kmalloc allocation from another cpu for
1347 * memory from the node of the cpu going down. The list3
1348 * structure is usually allocated from kmem_cache_create() and
1349 * gets destroyed at kmem_cache_destroy().
1353 case CPU_UP_CANCELED:
1354 case CPU_UP_CANCELED_FROZEN:
1355 mutex_lock(&cache_chain_mutex);
1356 cpuup_canceled(cpu);
1357 mutex_unlock(&cache_chain_mutex);
1360 return notifier_from_errno(err);
1363 static struct notifier_block __cpuinitdata cpucache_notifier = {
1364 &cpuup_callback, NULL, 0
1367 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1369 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1370 * Returns -EBUSY if all objects cannot be drained so that the node is not
1373 * Must hold cache_chain_mutex.
1375 static int __meminit drain_cache_nodelists_node(int node)
1377 struct kmem_cache *cachep;
1380 list_for_each_entry(cachep, &cache_chain, list) {
1381 struct kmem_list3 *l3;
1383 l3 = cachep->nodelists[node];
1387 drain_freelist(cachep, l3, l3->free_objects);
1389 if (!list_empty(&l3->slabs_full) ||
1390 !list_empty(&l3->slabs_partial)) {
1398 static int __meminit slab_memory_callback(struct notifier_block *self,
1399 unsigned long action, void *arg)
1401 struct memory_notify *mnb = arg;
1405 nid = mnb->status_change_nid;
1410 case MEM_GOING_ONLINE:
1411 mutex_lock(&cache_chain_mutex);
1412 ret = init_cache_nodelists_node(nid);
1413 mutex_unlock(&cache_chain_mutex);
1415 case MEM_GOING_OFFLINE:
1416 mutex_lock(&cache_chain_mutex);
1417 ret = drain_cache_nodelists_node(nid);
1418 mutex_unlock(&cache_chain_mutex);
1422 case MEM_CANCEL_ONLINE:
1423 case MEM_CANCEL_OFFLINE:
1427 return notifier_from_errno(ret);
1429 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1432 * swap the static kmem_list3 with kmalloced memory
1434 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1437 struct kmem_list3 *ptr;
1439 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1442 memcpy(ptr, list, sizeof(struct kmem_list3));
1444 * Do not assume that spinlocks can be initialized via memcpy:
1446 spin_lock_init(&ptr->list_lock);
1448 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1449 cachep->nodelists[nodeid] = ptr;
1453 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1454 * size of kmem_list3.
1456 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1460 for_each_online_node(node) {
1461 cachep->nodelists[node] = &initkmem_list3[index + node];
1462 cachep->nodelists[node]->next_reap = jiffies +
1464 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1469 * Initialisation. Called after the page allocator have been initialised and
1470 * before smp_init().
1472 void __init kmem_cache_init(void)
1475 struct cache_sizes *sizes;
1476 struct cache_names *names;
1481 if (num_possible_nodes() == 1)
1482 use_alien_caches = 0;
1484 for (i = 0; i < NUM_INIT_LISTS; i++) {
1485 kmem_list3_init(&initkmem_list3[i]);
1486 if (i < MAX_NUMNODES)
1487 cache_cache.nodelists[i] = NULL;
1489 set_up_list3s(&cache_cache, CACHE_CACHE);
1492 * Fragmentation resistance on low memory - only use bigger
1493 * page orders on machines with more than 32MB of memory if
1494 * not overridden on the command line.
1496 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1497 slab_max_order = SLAB_MAX_ORDER_HI;
1499 /* Bootstrap is tricky, because several objects are allocated
1500 * from caches that do not exist yet:
1501 * 1) initialize the cache_cache cache: it contains the struct
1502 * kmem_cache structures of all caches, except cache_cache itself:
1503 * cache_cache is statically allocated.
1504 * Initially an __init data area is used for the head array and the
1505 * kmem_list3 structures, it's replaced with a kmalloc allocated
1506 * array at the end of the bootstrap.
1507 * 2) Create the first kmalloc cache.
1508 * The struct kmem_cache for the new cache is allocated normally.
1509 * An __init data area is used for the head array.
1510 * 3) Create the remaining kmalloc caches, with minimally sized
1512 * 4) Replace the __init data head arrays for cache_cache and the first
1513 * kmalloc cache with kmalloc allocated arrays.
1514 * 5) Replace the __init data for kmem_list3 for cache_cache and
1515 * the other cache's with kmalloc allocated memory.
1516 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1519 node = numa_mem_id();
1521 /* 1) create the cache_cache */
1522 INIT_LIST_HEAD(&cache_chain);
1523 list_add(&cache_cache.list, &cache_chain);
1524 cache_cache.colour_off = cache_line_size();
1525 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1526 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1529 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1531 cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1532 nr_node_ids * sizeof(struct kmem_list3 *);
1533 cache_cache.object_size = cache_cache.size;
1534 cache_cache.size = ALIGN(cache_cache.size,
1536 cache_cache.reciprocal_buffer_size =
1537 reciprocal_value(cache_cache.size);
1539 for (order = 0; order < MAX_ORDER; order++) {
1540 cache_estimate(order, cache_cache.size,
1541 cache_line_size(), 0, &left_over, &cache_cache.num);
1542 if (cache_cache.num)
1545 BUG_ON(!cache_cache.num);
1546 cache_cache.gfporder = order;
1547 cache_cache.colour = left_over / cache_cache.colour_off;
1548 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1549 sizeof(struct slab), cache_line_size());
1551 /* 2+3) create the kmalloc caches */
1552 sizes = malloc_sizes;
1553 names = cache_names;
1556 * Initialize the caches that provide memory for the array cache and the
1557 * kmem_list3 structures first. Without this, further allocations will
1561 sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name,
1562 sizes[INDEX_AC].cs_size,
1563 ARCH_KMALLOC_MINALIGN,
1564 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1567 if (INDEX_AC != INDEX_L3) {
1568 sizes[INDEX_L3].cs_cachep =
1569 __kmem_cache_create(names[INDEX_L3].name,
1570 sizes[INDEX_L3].cs_size,
1571 ARCH_KMALLOC_MINALIGN,
1572 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1576 slab_early_init = 0;
1578 while (sizes->cs_size != ULONG_MAX) {
1580 * For performance, all the general caches are L1 aligned.
1581 * This should be particularly beneficial on SMP boxes, as it
1582 * eliminates "false sharing".
1583 * Note for systems short on memory removing the alignment will
1584 * allow tighter packing of the smaller caches.
1586 if (!sizes->cs_cachep) {
1587 sizes->cs_cachep = __kmem_cache_create(names->name,
1589 ARCH_KMALLOC_MINALIGN,
1590 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1593 #ifdef CONFIG_ZONE_DMA
1594 sizes->cs_dmacachep = __kmem_cache_create(
1597 ARCH_KMALLOC_MINALIGN,
1598 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1605 /* 4) Replace the bootstrap head arrays */
1607 struct array_cache *ptr;
1609 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1611 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1612 memcpy(ptr, cpu_cache_get(&cache_cache),
1613 sizeof(struct arraycache_init));
1615 * Do not assume that spinlocks can be initialized via memcpy:
1617 spin_lock_init(&ptr->lock);
1619 cache_cache.array[smp_processor_id()] = ptr;
1621 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1623 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1624 != &initarray_generic.cache);
1625 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1626 sizeof(struct arraycache_init));
1628 * Do not assume that spinlocks can be initialized via memcpy:
1630 spin_lock_init(&ptr->lock);
1632 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1635 /* 5) Replace the bootstrap kmem_list3's */
1639 for_each_online_node(nid) {
1640 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1642 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1643 &initkmem_list3[SIZE_AC + nid], nid);
1645 if (INDEX_AC != INDEX_L3) {
1646 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1647 &initkmem_list3[SIZE_L3 + nid], nid);
1652 g_cpucache_up = EARLY;
1655 void __init kmem_cache_init_late(void)
1657 struct kmem_cache *cachep;
1659 g_cpucache_up = LATE;
1661 /* Annotate slab for lockdep -- annotate the malloc caches */
1664 /* 6) resize the head arrays to their final sizes */
1665 mutex_lock(&cache_chain_mutex);
1666 list_for_each_entry(cachep, &cache_chain, list)
1667 if (enable_cpucache(cachep, GFP_NOWAIT))
1669 mutex_unlock(&cache_chain_mutex);
1672 * Register a cpu startup notifier callback that initializes
1673 * cpu_cache_get for all new cpus
1675 register_cpu_notifier(&cpucache_notifier);
1679 * Register a memory hotplug callback that initializes and frees
1682 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1686 * The reap timers are started later, with a module init call: That part
1687 * of the kernel is not yet operational.
1691 static int __init cpucache_init(void)
1696 * Register the timers that return unneeded pages to the page allocator
1698 for_each_online_cpu(cpu)
1699 start_cpu_timer(cpu);
1702 g_cpucache_up = FULL;
1705 __initcall(cpucache_init);
1707 static noinline void
1708 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1710 struct kmem_list3 *l3;
1712 unsigned long flags;
1716 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1718 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1719 cachep->name, cachep->size, cachep->gfporder);
1721 for_each_online_node(node) {
1722 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1723 unsigned long active_slabs = 0, num_slabs = 0;
1725 l3 = cachep->nodelists[node];
1729 spin_lock_irqsave(&l3->list_lock, flags);
1730 list_for_each_entry(slabp, &l3->slabs_full, list) {
1731 active_objs += cachep->num;
1734 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1735 active_objs += slabp->inuse;
1738 list_for_each_entry(slabp, &l3->slabs_free, list)
1741 free_objects += l3->free_objects;
1742 spin_unlock_irqrestore(&l3->list_lock, flags);
1744 num_slabs += active_slabs;
1745 num_objs = num_slabs * cachep->num;
1747 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1748 node, active_slabs, num_slabs, active_objs, num_objs,
1754 * Interface to system's page allocator. No need to hold the cache-lock.
1756 * If we requested dmaable memory, we will get it. Even if we
1757 * did not request dmaable memory, we might get it, but that
1758 * would be relatively rare and ignorable.
1760 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1768 * Nommu uses slab's for process anonymous memory allocations, and thus
1769 * requires __GFP_COMP to properly refcount higher order allocations
1771 flags |= __GFP_COMP;
1774 flags |= cachep->allocflags;
1775 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1776 flags |= __GFP_RECLAIMABLE;
1778 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1780 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1781 slab_out_of_memory(cachep, flags, nodeid);
1785 nr_pages = (1 << cachep->gfporder);
1786 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1787 add_zone_page_state(page_zone(page),
1788 NR_SLAB_RECLAIMABLE, nr_pages);
1790 add_zone_page_state(page_zone(page),
1791 NR_SLAB_UNRECLAIMABLE, nr_pages);
1792 for (i = 0; i < nr_pages; i++)
1793 __SetPageSlab(page + i);
1795 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1796 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1799 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1801 kmemcheck_mark_unallocated_pages(page, nr_pages);
1804 return page_address(page);
1808 * Interface to system's page release.
1810 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1812 unsigned long i = (1 << cachep->gfporder);
1813 struct page *page = virt_to_page(addr);
1814 const unsigned long nr_freed = i;
1816 kmemcheck_free_shadow(page, cachep->gfporder);
1818 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1819 sub_zone_page_state(page_zone(page),
1820 NR_SLAB_RECLAIMABLE, nr_freed);
1822 sub_zone_page_state(page_zone(page),
1823 NR_SLAB_UNRECLAIMABLE, nr_freed);
1825 BUG_ON(!PageSlab(page));
1826 __ClearPageSlab(page);
1829 if (current->reclaim_state)
1830 current->reclaim_state->reclaimed_slab += nr_freed;
1831 free_pages((unsigned long)addr, cachep->gfporder);
1834 static void kmem_rcu_free(struct rcu_head *head)
1836 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1837 struct kmem_cache *cachep = slab_rcu->cachep;
1839 kmem_freepages(cachep, slab_rcu->addr);
1840 if (OFF_SLAB(cachep))
1841 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1846 #ifdef CONFIG_DEBUG_PAGEALLOC
1847 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1848 unsigned long caller)
1850 int size = cachep->object_size;
1852 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1854 if (size < 5 * sizeof(unsigned long))
1857 *addr++ = 0x12345678;
1859 *addr++ = smp_processor_id();
1860 size -= 3 * sizeof(unsigned long);
1862 unsigned long *sptr = &caller;
1863 unsigned long svalue;
1865 while (!kstack_end(sptr)) {
1867 if (kernel_text_address(svalue)) {
1869 size -= sizeof(unsigned long);
1870 if (size <= sizeof(unsigned long))
1876 *addr++ = 0x87654321;
1880 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1882 int size = cachep->object_size;
1883 addr = &((char *)addr)[obj_offset(cachep)];
1885 memset(addr, val, size);
1886 *(unsigned char *)(addr + size - 1) = POISON_END;
1889 static void dump_line(char *data, int offset, int limit)
1892 unsigned char error = 0;
1895 printk(KERN_ERR "%03x: ", offset);
1896 for (i = 0; i < limit; i++) {
1897 if (data[offset + i] != POISON_FREE) {
1898 error = data[offset + i];
1902 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1903 &data[offset], limit, 1);
1905 if (bad_count == 1) {
1906 error ^= POISON_FREE;
1907 if (!(error & (error - 1))) {
1908 printk(KERN_ERR "Single bit error detected. Probably "
1911 printk(KERN_ERR "Run memtest86+ or a similar memory "
1914 printk(KERN_ERR "Run a memory test tool.\n");
1923 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1928 if (cachep->flags & SLAB_RED_ZONE) {
1929 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1930 *dbg_redzone1(cachep, objp),
1931 *dbg_redzone2(cachep, objp));
1934 if (cachep->flags & SLAB_STORE_USER) {
1935 printk(KERN_ERR "Last user: [<%p>]",
1936 *dbg_userword(cachep, objp));
1937 print_symbol("(%s)",
1938 (unsigned long)*dbg_userword(cachep, objp));
1941 realobj = (char *)objp + obj_offset(cachep);
1942 size = cachep->object_size;
1943 for (i = 0; i < size && lines; i += 16, lines--) {
1946 if (i + limit > size)
1948 dump_line(realobj, i, limit);
1952 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1958 realobj = (char *)objp + obj_offset(cachep);
1959 size = cachep->object_size;
1961 for (i = 0; i < size; i++) {
1962 char exp = POISON_FREE;
1965 if (realobj[i] != exp) {
1971 "Slab corruption (%s): %s start=%p, len=%d\n",
1972 print_tainted(), cachep->name, realobj, size);
1973 print_objinfo(cachep, objp, 0);
1975 /* Hexdump the affected line */
1978 if (i + limit > size)
1980 dump_line(realobj, i, limit);
1983 /* Limit to 5 lines */
1989 /* Print some data about the neighboring objects, if they
1992 struct slab *slabp = virt_to_slab(objp);
1995 objnr = obj_to_index(cachep, slabp, objp);
1997 objp = index_to_obj(cachep, slabp, objnr - 1);
1998 realobj = (char *)objp + obj_offset(cachep);
1999 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2001 print_objinfo(cachep, objp, 2);
2003 if (objnr + 1 < cachep->num) {
2004 objp = index_to_obj(cachep, slabp, objnr + 1);
2005 realobj = (char *)objp + obj_offset(cachep);
2006 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2008 print_objinfo(cachep, objp, 2);
2015 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2018 for (i = 0; i < cachep->num; i++) {
2019 void *objp = index_to_obj(cachep, slabp, i);
2021 if (cachep->flags & SLAB_POISON) {
2022 #ifdef CONFIG_DEBUG_PAGEALLOC
2023 if (cachep->size % PAGE_SIZE == 0 &&
2025 kernel_map_pages(virt_to_page(objp),
2026 cachep->size / PAGE_SIZE, 1);
2028 check_poison_obj(cachep, objp);
2030 check_poison_obj(cachep, objp);
2033 if (cachep->flags & SLAB_RED_ZONE) {
2034 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2035 slab_error(cachep, "start of a freed object "
2037 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2038 slab_error(cachep, "end of a freed object "
2044 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2050 * slab_destroy - destroy and release all objects in a slab
2051 * @cachep: cache pointer being destroyed
2052 * @slabp: slab pointer being destroyed
2054 * Destroy all the objs in a slab, and release the mem back to the system.
2055 * Before calling the slab must have been unlinked from the cache. The
2056 * cache-lock is not held/needed.
2058 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2060 void *addr = slabp->s_mem - slabp->colouroff;
2062 slab_destroy_debugcheck(cachep, slabp);
2063 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2064 struct slab_rcu *slab_rcu;
2066 slab_rcu = (struct slab_rcu *)slabp;
2067 slab_rcu->cachep = cachep;
2068 slab_rcu->addr = addr;
2069 call_rcu(&slab_rcu->head, kmem_rcu_free);
2071 kmem_freepages(cachep, addr);
2072 if (OFF_SLAB(cachep))
2073 kmem_cache_free(cachep->slabp_cache, slabp);
2077 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2080 struct kmem_list3 *l3;
2082 for_each_online_cpu(i)
2083 kfree(cachep->array[i]);
2085 /* NUMA: free the list3 structures */
2086 for_each_online_node(i) {
2087 l3 = cachep->nodelists[i];
2090 free_alien_cache(l3->alien);
2094 kmem_cache_free(&cache_cache, cachep);
2099 * calculate_slab_order - calculate size (page order) of slabs
2100 * @cachep: pointer to the cache that is being created
2101 * @size: size of objects to be created in this cache.
2102 * @align: required alignment for the objects.
2103 * @flags: slab allocation flags
2105 * Also calculates the number of objects per slab.
2107 * This could be made much more intelligent. For now, try to avoid using
2108 * high order pages for slabs. When the gfp() functions are more friendly
2109 * towards high-order requests, this should be changed.
2111 static size_t calculate_slab_order(struct kmem_cache *cachep,
2112 size_t size, size_t align, unsigned long flags)
2114 unsigned long offslab_limit;
2115 size_t left_over = 0;
2118 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2122 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2126 if (flags & CFLGS_OFF_SLAB) {
2128 * Max number of objs-per-slab for caches which
2129 * use off-slab slabs. Needed to avoid a possible
2130 * looping condition in cache_grow().
2132 offslab_limit = size - sizeof(struct slab);
2133 offslab_limit /= sizeof(kmem_bufctl_t);
2135 if (num > offslab_limit)
2139 /* Found something acceptable - save it away */
2141 cachep->gfporder = gfporder;
2142 left_over = remainder;
2145 * A VFS-reclaimable slab tends to have most allocations
2146 * as GFP_NOFS and we really don't want to have to be allocating
2147 * higher-order pages when we are unable to shrink dcache.
2149 if (flags & SLAB_RECLAIM_ACCOUNT)
2153 * Large number of objects is good, but very large slabs are
2154 * currently bad for the gfp()s.
2156 if (gfporder >= slab_max_order)
2160 * Acceptable internal fragmentation?
2162 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2168 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2170 if (g_cpucache_up >= LATE)
2171 return enable_cpucache(cachep, gfp);
2173 if (g_cpucache_up == NONE) {
2175 * Note: the first kmem_cache_create must create the cache
2176 * that's used by kmalloc(24), otherwise the creation of
2177 * further caches will BUG().
2179 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2182 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2183 * the first cache, then we need to set up all its list3s,
2184 * otherwise the creation of further caches will BUG().
2186 set_up_list3s(cachep, SIZE_AC);
2187 if (INDEX_AC == INDEX_L3)
2188 g_cpucache_up = PARTIAL_L3;
2190 g_cpucache_up = PARTIAL_AC;
2192 cachep->array[smp_processor_id()] =
2193 kmalloc(sizeof(struct arraycache_init), gfp);
2195 if (g_cpucache_up == PARTIAL_AC) {
2196 set_up_list3s(cachep, SIZE_L3);
2197 g_cpucache_up = PARTIAL_L3;
2200 for_each_online_node(node) {
2201 cachep->nodelists[node] =
2202 kmalloc_node(sizeof(struct kmem_list3),
2204 BUG_ON(!cachep->nodelists[node]);
2205 kmem_list3_init(cachep->nodelists[node]);
2209 cachep->nodelists[numa_mem_id()]->next_reap =
2210 jiffies + REAPTIMEOUT_LIST3 +
2211 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2213 cpu_cache_get(cachep)->avail = 0;
2214 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2215 cpu_cache_get(cachep)->batchcount = 1;
2216 cpu_cache_get(cachep)->touched = 0;
2217 cachep->batchcount = 1;
2218 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2223 * __kmem_cache_create - Create a cache.
2224 * @name: A string which is used in /proc/slabinfo to identify this cache.
2225 * @size: The size of objects to be created in this cache.
2226 * @align: The required alignment for the objects.
2227 * @flags: SLAB flags
2228 * @ctor: A constructor for the objects.
2230 * Returns a ptr to the cache on success, NULL on failure.
2231 * Cannot be called within a int, but can be interrupted.
2232 * The @ctor is run when new pages are allocated by the cache.
2234 * @name must be valid until the cache is destroyed. This implies that
2235 * the module calling this has to destroy the cache before getting unloaded.
2239 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2240 * to catch references to uninitialised memory.
2242 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2243 * for buffer overruns.
2245 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2246 * cacheline. This can be beneficial if you're counting cycles as closely
2250 __kmem_cache_create (const char *name, size_t size, size_t align,
2251 unsigned long flags, void (*ctor)(void *))
2253 size_t left_over, slab_size, ralign;
2254 struct kmem_cache *cachep = NULL, *pc;
2258 * Sanity checks... these are all serious usage bugs.
2260 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2261 size > KMALLOC_MAX_SIZE) {
2262 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2268 * We use cache_chain_mutex to ensure a consistent view of
2269 * cpu_online_mask as well. Please see cpuup_callback
2271 if (slab_is_available()) {
2273 mutex_lock(&cache_chain_mutex);
2276 list_for_each_entry(pc, &cache_chain, list) {
2281 * This happens when the module gets unloaded and doesn't
2282 * destroy its slab cache and no-one else reuses the vmalloc
2283 * area of the module. Print a warning.
2285 res = probe_kernel_address(pc->name, tmp);
2288 "SLAB: cache with size %d has lost its name\n",
2293 if (!strcmp(pc->name, name)) {
2295 "kmem_cache_create: duplicate cache %s\n", name);
2302 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2305 * Enable redzoning and last user accounting, except for caches with
2306 * large objects, if the increased size would increase the object size
2307 * above the next power of two: caches with object sizes just above a
2308 * power of two have a significant amount of internal fragmentation.
2310 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2311 2 * sizeof(unsigned long long)))
2312 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2313 if (!(flags & SLAB_DESTROY_BY_RCU))
2314 flags |= SLAB_POISON;
2316 if (flags & SLAB_DESTROY_BY_RCU)
2317 BUG_ON(flags & SLAB_POISON);
2320 * Always checks flags, a caller might be expecting debug support which
2323 BUG_ON(flags & ~CREATE_MASK);
2326 * Check that size is in terms of words. This is needed to avoid
2327 * unaligned accesses for some archs when redzoning is used, and makes
2328 * sure any on-slab bufctl's are also correctly aligned.
2330 if (size & (BYTES_PER_WORD - 1)) {
2331 size += (BYTES_PER_WORD - 1);
2332 size &= ~(BYTES_PER_WORD - 1);
2335 /* calculate the final buffer alignment: */
2337 /* 1) arch recommendation: can be overridden for debug */
2338 if (flags & SLAB_HWCACHE_ALIGN) {
2340 * Default alignment: as specified by the arch code. Except if
2341 * an object is really small, then squeeze multiple objects into
2344 ralign = cache_line_size();
2345 while (size <= ralign / 2)
2348 ralign = BYTES_PER_WORD;
2352 * Redzoning and user store require word alignment or possibly larger.
2353 * Note this will be overridden by architecture or caller mandated
2354 * alignment if either is greater than BYTES_PER_WORD.
2356 if (flags & SLAB_STORE_USER)
2357 ralign = BYTES_PER_WORD;
2359 if (flags & SLAB_RED_ZONE) {
2360 ralign = REDZONE_ALIGN;
2361 /* If redzoning, ensure that the second redzone is suitably
2362 * aligned, by adjusting the object size accordingly. */
2363 size += REDZONE_ALIGN - 1;
2364 size &= ~(REDZONE_ALIGN - 1);
2367 /* 2) arch mandated alignment */
2368 if (ralign < ARCH_SLAB_MINALIGN) {
2369 ralign = ARCH_SLAB_MINALIGN;
2371 /* 3) caller mandated alignment */
2372 if (ralign < align) {
2375 /* disable debug if necessary */
2376 if (ralign > __alignof__(unsigned long long))
2377 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2383 if (slab_is_available())
2388 /* Get cache's description obj. */
2389 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2393 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2394 cachep->object_size = size;
2395 cachep->align = align;
2399 * Both debugging options require word-alignment which is calculated
2402 if (flags & SLAB_RED_ZONE) {
2403 /* add space for red zone words */
2404 cachep->obj_offset += sizeof(unsigned long long);
2405 size += 2 * sizeof(unsigned long long);
2407 if (flags & SLAB_STORE_USER) {
2408 /* user store requires one word storage behind the end of
2409 * the real object. But if the second red zone needs to be
2410 * aligned to 64 bits, we must allow that much space.
2412 if (flags & SLAB_RED_ZONE)
2413 size += REDZONE_ALIGN;
2415 size += BYTES_PER_WORD;
2417 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2418 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2419 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2420 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2427 * Determine if the slab management is 'on' or 'off' slab.
2428 * (bootstrapping cannot cope with offslab caches so don't do
2429 * it too early on. Always use on-slab management when
2430 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2432 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2433 !(flags & SLAB_NOLEAKTRACE))
2435 * Size is large, assume best to place the slab management obj
2436 * off-slab (should allow better packing of objs).
2438 flags |= CFLGS_OFF_SLAB;
2440 size = ALIGN(size, align);
2442 left_over = calculate_slab_order(cachep, size, align, flags);
2446 "kmem_cache_create: couldn't create cache %s.\n", name);
2447 kmem_cache_free(&cache_cache, cachep);
2450 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2451 + sizeof(struct slab), align);
2454 * If the slab has been placed off-slab, and we have enough space then
2455 * move it on-slab. This is at the expense of any extra colouring.
2457 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2458 flags &= ~CFLGS_OFF_SLAB;
2459 left_over -= slab_size;
2462 if (flags & CFLGS_OFF_SLAB) {
2463 /* really off slab. No need for manual alignment */
2465 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2467 #ifdef CONFIG_PAGE_POISONING
2468 /* If we're going to use the generic kernel_map_pages()
2469 * poisoning, then it's going to smash the contents of
2470 * the redzone and userword anyhow, so switch them off.
2472 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2473 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2477 cachep->colour_off = cache_line_size();
2478 /* Offset must be a multiple of the alignment. */
2479 if (cachep->colour_off < align)
2480 cachep->colour_off = align;
2481 cachep->colour = left_over / cachep->colour_off;
2482 cachep->slab_size = slab_size;
2483 cachep->flags = flags;
2484 cachep->allocflags = 0;
2485 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2486 cachep->allocflags |= GFP_DMA;
2487 cachep->size = size;
2488 cachep->reciprocal_buffer_size = reciprocal_value(size);
2490 if (flags & CFLGS_OFF_SLAB) {
2491 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2493 * This is a possibility for one of the malloc_sizes caches.
2494 * But since we go off slab only for object size greater than
2495 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2496 * this should not happen at all.
2497 * But leave a BUG_ON for some lucky dude.
2499 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2501 cachep->ctor = ctor;
2502 cachep->name = name;
2504 if (setup_cpu_cache(cachep, gfp)) {
2505 __kmem_cache_destroy(cachep);
2509 if (flags & SLAB_DEBUG_OBJECTS) {
2511 * Would deadlock through slab_destroy()->call_rcu()->
2512 * debug_object_activate()->kmem_cache_alloc().
2514 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2516 slab_set_debugobj_lock_classes(cachep);
2519 /* cache setup completed, link it into the list */
2520 list_add(&cachep->list, &cache_chain);
2522 if (slab_is_available()) {
2523 mutex_unlock(&cache_chain_mutex);
2530 static void check_irq_off(void)
2532 BUG_ON(!irqs_disabled());
2535 static void check_irq_on(void)
2537 BUG_ON(irqs_disabled());
2540 static void check_spinlock_acquired(struct kmem_cache *cachep)
2544 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2548 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2552 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2557 #define check_irq_off() do { } while(0)
2558 #define check_irq_on() do { } while(0)
2559 #define check_spinlock_acquired(x) do { } while(0)
2560 #define check_spinlock_acquired_node(x, y) do { } while(0)
2563 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2564 struct array_cache *ac,
2565 int force, int node);
2567 static void do_drain(void *arg)
2569 struct kmem_cache *cachep = arg;
2570 struct array_cache *ac;
2571 int node = numa_mem_id();
2574 ac = cpu_cache_get(cachep);
2575 spin_lock(&cachep->nodelists[node]->list_lock);
2576 free_block(cachep, ac->entry, ac->avail, node);
2577 spin_unlock(&cachep->nodelists[node]->list_lock);
2581 static void drain_cpu_caches(struct kmem_cache *cachep)
2583 struct kmem_list3 *l3;
2586 on_each_cpu(do_drain, cachep, 1);
2588 for_each_online_node(node) {
2589 l3 = cachep->nodelists[node];
2590 if (l3 && l3->alien)
2591 drain_alien_cache(cachep, l3->alien);
2594 for_each_online_node(node) {
2595 l3 = cachep->nodelists[node];
2597 drain_array(cachep, l3, l3->shared, 1, node);
2602 * Remove slabs from the list of free slabs.
2603 * Specify the number of slabs to drain in tofree.
2605 * Returns the actual number of slabs released.
2607 static int drain_freelist(struct kmem_cache *cache,
2608 struct kmem_list3 *l3, int tofree)
2610 struct list_head *p;
2615 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2617 spin_lock_irq(&l3->list_lock);
2618 p = l3->slabs_free.prev;
2619 if (p == &l3->slabs_free) {
2620 spin_unlock_irq(&l3->list_lock);
2624 slabp = list_entry(p, struct slab, list);
2626 BUG_ON(slabp->inuse);
2628 list_del(&slabp->list);
2630 * Safe to drop the lock. The slab is no longer linked
2633 l3->free_objects -= cache->num;
2634 spin_unlock_irq(&l3->list_lock);
2635 slab_destroy(cache, slabp);
2642 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2643 static int __cache_shrink(struct kmem_cache *cachep)
2646 struct kmem_list3 *l3;
2648 drain_cpu_caches(cachep);
2651 for_each_online_node(i) {
2652 l3 = cachep->nodelists[i];
2656 drain_freelist(cachep, l3, l3->free_objects);
2658 ret += !list_empty(&l3->slabs_full) ||
2659 !list_empty(&l3->slabs_partial);
2661 return (ret ? 1 : 0);
2665 * kmem_cache_shrink - Shrink a cache.
2666 * @cachep: The cache to shrink.
2668 * Releases as many slabs as possible for a cache.
2669 * To help debugging, a zero exit status indicates all slabs were released.
2671 int kmem_cache_shrink(struct kmem_cache *cachep)
2674 BUG_ON(!cachep || in_interrupt());
2677 mutex_lock(&cache_chain_mutex);
2678 ret = __cache_shrink(cachep);
2679 mutex_unlock(&cache_chain_mutex);
2683 EXPORT_SYMBOL(kmem_cache_shrink);
2686 * kmem_cache_destroy - delete a cache
2687 * @cachep: the cache to destroy
2689 * Remove a &struct kmem_cache object from the slab cache.
2691 * It is expected this function will be called by a module when it is
2692 * unloaded. This will remove the cache completely, and avoid a duplicate
2693 * cache being allocated each time a module is loaded and unloaded, if the
2694 * module doesn't have persistent in-kernel storage across loads and unloads.
2696 * The cache must be empty before calling this function.
2698 * The caller must guarantee that no one will allocate memory from the cache
2699 * during the kmem_cache_destroy().
2701 void kmem_cache_destroy(struct kmem_cache *cachep)
2703 BUG_ON(!cachep || in_interrupt());
2705 /* Find the cache in the chain of caches. */
2707 mutex_lock(&cache_chain_mutex);
2709 * the chain is never empty, cache_cache is never destroyed
2711 list_del(&cachep->list);
2712 if (__cache_shrink(cachep)) {
2713 slab_error(cachep, "Can't free all objects");
2714 list_add(&cachep->list, &cache_chain);
2715 mutex_unlock(&cache_chain_mutex);
2720 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2723 __kmem_cache_destroy(cachep);
2724 mutex_unlock(&cache_chain_mutex);
2727 EXPORT_SYMBOL(kmem_cache_destroy);
2730 * Get the memory for a slab management obj.
2731 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2732 * always come from malloc_sizes caches. The slab descriptor cannot
2733 * come from the same cache which is getting created because,
2734 * when we are searching for an appropriate cache for these
2735 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2736 * If we are creating a malloc_sizes cache here it would not be visible to
2737 * kmem_find_general_cachep till the initialization is complete.
2738 * Hence we cannot have slabp_cache same as the original cache.
2740 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2741 int colour_off, gfp_t local_flags,
2746 if (OFF_SLAB(cachep)) {
2747 /* Slab management obj is off-slab. */
2748 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2749 local_flags, nodeid);
2751 * If the first object in the slab is leaked (it's allocated
2752 * but no one has a reference to it), we want to make sure
2753 * kmemleak does not treat the ->s_mem pointer as a reference
2754 * to the object. Otherwise we will not report the leak.
2756 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2761 slabp = objp + colour_off;
2762 colour_off += cachep->slab_size;
2765 slabp->colouroff = colour_off;
2766 slabp->s_mem = objp + colour_off;
2767 slabp->nodeid = nodeid;
2772 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2774 return (kmem_bufctl_t *) (slabp + 1);
2777 static void cache_init_objs(struct kmem_cache *cachep,
2782 for (i = 0; i < cachep->num; i++) {
2783 void *objp = index_to_obj(cachep, slabp, i);
2785 /* need to poison the objs? */
2786 if (cachep->flags & SLAB_POISON)
2787 poison_obj(cachep, objp, POISON_FREE);
2788 if (cachep->flags & SLAB_STORE_USER)
2789 *dbg_userword(cachep, objp) = NULL;
2791 if (cachep->flags & SLAB_RED_ZONE) {
2792 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2793 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2796 * Constructors are not allowed to allocate memory from the same
2797 * cache which they are a constructor for. Otherwise, deadlock.
2798 * They must also be threaded.
2800 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2801 cachep->ctor(objp + obj_offset(cachep));
2803 if (cachep->flags & SLAB_RED_ZONE) {
2804 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2805 slab_error(cachep, "constructor overwrote the"
2806 " end of an object");
2807 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2808 slab_error(cachep, "constructor overwrote the"
2809 " start of an object");
2811 if ((cachep->size % PAGE_SIZE) == 0 &&
2812 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2813 kernel_map_pages(virt_to_page(objp),
2814 cachep->size / PAGE_SIZE, 0);
2819 slab_bufctl(slabp)[i] = i + 1;
2821 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2824 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2826 if (CONFIG_ZONE_DMA_FLAG) {
2827 if (flags & GFP_DMA)
2828 BUG_ON(!(cachep->allocflags & GFP_DMA));
2830 BUG_ON(cachep->allocflags & GFP_DMA);
2834 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2837 void *objp = index_to_obj(cachep, slabp, slabp->free);
2841 next = slab_bufctl(slabp)[slabp->free];
2843 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2844 WARN_ON(slabp->nodeid != nodeid);
2851 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2852 void *objp, int nodeid)
2854 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2857 /* Verify that the slab belongs to the intended node */
2858 WARN_ON(slabp->nodeid != nodeid);
2860 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2861 printk(KERN_ERR "slab: double free detected in cache "
2862 "'%s', objp %p\n", cachep->name, objp);
2866 slab_bufctl(slabp)[objnr] = slabp->free;
2867 slabp->free = objnr;
2872 * Map pages beginning at addr to the given cache and slab. This is required
2873 * for the slab allocator to be able to lookup the cache and slab of a
2874 * virtual address for kfree, ksize, and slab debugging.
2876 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2882 page = virt_to_page(addr);
2885 if (likely(!PageCompound(page)))
2886 nr_pages <<= cache->gfporder;
2889 page->slab_cache = cache;
2890 page->slab_page = slab;
2892 } while (--nr_pages);
2896 * Grow (by 1) the number of slabs within a cache. This is called by
2897 * kmem_cache_alloc() when there are no active objs left in a cache.
2899 static int cache_grow(struct kmem_cache *cachep,
2900 gfp_t flags, int nodeid, void *objp)
2905 struct kmem_list3 *l3;
2908 * Be lazy and only check for valid flags here, keeping it out of the
2909 * critical path in kmem_cache_alloc().
2911 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2912 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2914 /* Take the l3 list lock to change the colour_next on this node */
2916 l3 = cachep->nodelists[nodeid];
2917 spin_lock(&l3->list_lock);
2919 /* Get colour for the slab, and cal the next value. */
2920 offset = l3->colour_next;
2922 if (l3->colour_next >= cachep->colour)
2923 l3->colour_next = 0;
2924 spin_unlock(&l3->list_lock);
2926 offset *= cachep->colour_off;
2928 if (local_flags & __GFP_WAIT)
2932 * The test for missing atomic flag is performed here, rather than
2933 * the more obvious place, simply to reduce the critical path length
2934 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2935 * will eventually be caught here (where it matters).
2937 kmem_flagcheck(cachep, flags);
2940 * Get mem for the objs. Attempt to allocate a physical page from
2944 objp = kmem_getpages(cachep, local_flags, nodeid);
2948 /* Get slab management. */
2949 slabp = alloc_slabmgmt(cachep, objp, offset,
2950 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2954 slab_map_pages(cachep, slabp, objp);
2956 cache_init_objs(cachep, slabp);
2958 if (local_flags & __GFP_WAIT)
2959 local_irq_disable();
2961 spin_lock(&l3->list_lock);
2963 /* Make slab active. */
2964 list_add_tail(&slabp->list, &(l3->slabs_free));
2965 STATS_INC_GROWN(cachep);
2966 l3->free_objects += cachep->num;
2967 spin_unlock(&l3->list_lock);
2970 kmem_freepages(cachep, objp);
2972 if (local_flags & __GFP_WAIT)
2973 local_irq_disable();
2980 * Perform extra freeing checks:
2981 * - detect bad pointers.
2982 * - POISON/RED_ZONE checking
2984 static void kfree_debugcheck(const void *objp)
2986 if (!virt_addr_valid(objp)) {
2987 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2988 (unsigned long)objp);
2993 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2995 unsigned long long redzone1, redzone2;
2997 redzone1 = *dbg_redzone1(cache, obj);
2998 redzone2 = *dbg_redzone2(cache, obj);
3003 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3006 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3007 slab_error(cache, "double free detected");
3009 slab_error(cache, "memory outside object was overwritten");
3011 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3012 obj, redzone1, redzone2);
3015 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3022 BUG_ON(virt_to_cache(objp) != cachep);
3024 objp -= obj_offset(cachep);
3025 kfree_debugcheck(objp);
3026 page = virt_to_head_page(objp);
3028 slabp = page->slab_page;
3030 if (cachep->flags & SLAB_RED_ZONE) {
3031 verify_redzone_free(cachep, objp);
3032 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3033 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3035 if (cachep->flags & SLAB_STORE_USER)
3036 *dbg_userword(cachep, objp) = caller;
3038 objnr = obj_to_index(cachep, slabp, objp);
3040 BUG_ON(objnr >= cachep->num);
3041 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3043 #ifdef CONFIG_DEBUG_SLAB_LEAK
3044 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3046 if (cachep->flags & SLAB_POISON) {
3047 #ifdef CONFIG_DEBUG_PAGEALLOC
3048 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3049 store_stackinfo(cachep, objp, (unsigned long)caller);
3050 kernel_map_pages(virt_to_page(objp),
3051 cachep->size / PAGE_SIZE, 0);
3053 poison_obj(cachep, objp, POISON_FREE);
3056 poison_obj(cachep, objp, POISON_FREE);
3062 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3067 /* Check slab's freelist to see if this obj is there. */
3068 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3070 if (entries > cachep->num || i >= cachep->num)
3073 if (entries != cachep->num - slabp->inuse) {
3075 printk(KERN_ERR "slab: Internal list corruption detected in "
3076 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3077 cachep->name, cachep->num, slabp, slabp->inuse,
3079 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3080 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3086 #define kfree_debugcheck(x) do { } while(0)
3087 #define cache_free_debugcheck(x,objp,z) (objp)
3088 #define check_slabp(x,y) do { } while(0)
3091 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3094 struct kmem_list3 *l3;
3095 struct array_cache *ac;
3100 node = numa_mem_id();
3101 ac = cpu_cache_get(cachep);
3102 batchcount = ac->batchcount;
3103 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3105 * If there was little recent activity on this cache, then
3106 * perform only a partial refill. Otherwise we could generate
3109 batchcount = BATCHREFILL_LIMIT;
3111 l3 = cachep->nodelists[node];
3113 BUG_ON(ac->avail > 0 || !l3);
3114 spin_lock(&l3->list_lock);
3116 /* See if we can refill from the shared array */
3117 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3118 l3->shared->touched = 1;
3122 while (batchcount > 0) {
3123 struct list_head *entry;
3125 /* Get slab alloc is to come from. */
3126 entry = l3->slabs_partial.next;
3127 if (entry == &l3->slabs_partial) {
3128 l3->free_touched = 1;
3129 entry = l3->slabs_free.next;
3130 if (entry == &l3->slabs_free)
3134 slabp = list_entry(entry, struct slab, list);
3135 check_slabp(cachep, slabp);
3136 check_spinlock_acquired(cachep);
3139 * The slab was either on partial or free list so
3140 * there must be at least one object available for
3143 BUG_ON(slabp->inuse >= cachep->num);
3145 while (slabp->inuse < cachep->num && batchcount--) {
3146 STATS_INC_ALLOCED(cachep);
3147 STATS_INC_ACTIVE(cachep);
3148 STATS_SET_HIGH(cachep);
3150 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3153 check_slabp(cachep, slabp);
3155 /* move slabp to correct slabp list: */
3156 list_del(&slabp->list);
3157 if (slabp->free == BUFCTL_END)
3158 list_add(&slabp->list, &l3->slabs_full);
3160 list_add(&slabp->list, &l3->slabs_partial);
3164 l3->free_objects -= ac->avail;
3166 spin_unlock(&l3->list_lock);
3168 if (unlikely(!ac->avail)) {
3170 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3172 /* cache_grow can reenable interrupts, then ac could change. */
3173 ac = cpu_cache_get(cachep);
3174 if (!x && ac->avail == 0) /* no objects in sight? abort */
3177 if (!ac->avail) /* objects refilled by interrupt? */
3181 return ac->entry[--ac->avail];
3184 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3187 might_sleep_if(flags & __GFP_WAIT);
3189 kmem_flagcheck(cachep, flags);
3194 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3195 gfp_t flags, void *objp, void *caller)
3199 if (cachep->flags & SLAB_POISON) {
3200 #ifdef CONFIG_DEBUG_PAGEALLOC
3201 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3202 kernel_map_pages(virt_to_page(objp),
3203 cachep->size / PAGE_SIZE, 1);
3205 check_poison_obj(cachep, objp);
3207 check_poison_obj(cachep, objp);
3209 poison_obj(cachep, objp, POISON_INUSE);
3211 if (cachep->flags & SLAB_STORE_USER)
3212 *dbg_userword(cachep, objp) = caller;
3214 if (cachep->flags & SLAB_RED_ZONE) {
3215 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3216 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3217 slab_error(cachep, "double free, or memory outside"
3218 " object was overwritten");
3220 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3221 objp, *dbg_redzone1(cachep, objp),
3222 *dbg_redzone2(cachep, objp));
3224 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3225 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3227 #ifdef CONFIG_DEBUG_SLAB_LEAK
3232 slabp = virt_to_head_page(objp)->slab_page;
3233 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3234 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3237 objp += obj_offset(cachep);
3238 if (cachep->ctor && cachep->flags & SLAB_POISON)
3240 if (ARCH_SLAB_MINALIGN &&
3241 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3242 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3243 objp, (int)ARCH_SLAB_MINALIGN);
3248 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3251 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3253 if (cachep == &cache_cache)
3256 return should_failslab(cachep->object_size, flags, cachep->flags);
3259 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3262 struct array_cache *ac;
3266 ac = cpu_cache_get(cachep);
3267 if (likely(ac->avail)) {
3268 STATS_INC_ALLOCHIT(cachep);
3270 objp = ac->entry[--ac->avail];
3272 STATS_INC_ALLOCMISS(cachep);
3273 objp = cache_alloc_refill(cachep, flags);
3275 * the 'ac' may be updated by cache_alloc_refill(),
3276 * and kmemleak_erase() requires its correct value.
3278 ac = cpu_cache_get(cachep);
3281 * To avoid a false negative, if an object that is in one of the
3282 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3283 * treat the array pointers as a reference to the object.
3286 kmemleak_erase(&ac->entry[ac->avail]);
3292 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3294 * If we are in_interrupt, then process context, including cpusets and
3295 * mempolicy, may not apply and should not be used for allocation policy.
3297 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3299 int nid_alloc, nid_here;
3301 if (in_interrupt() || (flags & __GFP_THISNODE))
3303 nid_alloc = nid_here = numa_mem_id();
3304 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3305 nid_alloc = cpuset_slab_spread_node();
3306 else if (current->mempolicy)
3307 nid_alloc = slab_node();
3308 if (nid_alloc != nid_here)
3309 return ____cache_alloc_node(cachep, flags, nid_alloc);
3314 * Fallback function if there was no memory available and no objects on a
3315 * certain node and fall back is permitted. First we scan all the
3316 * available nodelists for available objects. If that fails then we
3317 * perform an allocation without specifying a node. This allows the page
3318 * allocator to do its reclaim / fallback magic. We then insert the
3319 * slab into the proper nodelist and then allocate from it.
3321 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3323 struct zonelist *zonelist;
3327 enum zone_type high_zoneidx = gfp_zone(flags);
3330 unsigned int cpuset_mems_cookie;
3332 if (flags & __GFP_THISNODE)
3335 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3338 cpuset_mems_cookie = get_mems_allowed();
3339 zonelist = node_zonelist(slab_node(), flags);
3343 * Look through allowed nodes for objects available
3344 * from existing per node queues.
3346 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3347 nid = zone_to_nid(zone);
3349 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3350 cache->nodelists[nid] &&
3351 cache->nodelists[nid]->free_objects) {
3352 obj = ____cache_alloc_node(cache,
3353 flags | GFP_THISNODE, nid);
3361 * This allocation will be performed within the constraints
3362 * of the current cpuset / memory policy requirements.
3363 * We may trigger various forms of reclaim on the allowed
3364 * set and go into memory reserves if necessary.
3366 if (local_flags & __GFP_WAIT)
3368 kmem_flagcheck(cache, flags);
3369 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3370 if (local_flags & __GFP_WAIT)
3371 local_irq_disable();
3374 * Insert into the appropriate per node queues
3376 nid = page_to_nid(virt_to_page(obj));
3377 if (cache_grow(cache, flags, nid, obj)) {
3378 obj = ____cache_alloc_node(cache,
3379 flags | GFP_THISNODE, nid);
3382 * Another processor may allocate the
3383 * objects in the slab since we are
3384 * not holding any locks.
3388 /* cache_grow already freed obj */
3394 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3400 * A interface to enable slab creation on nodeid
3402 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3405 struct list_head *entry;
3407 struct kmem_list3 *l3;
3411 l3 = cachep->nodelists[nodeid];
3416 spin_lock(&l3->list_lock);
3417 entry = l3->slabs_partial.next;
3418 if (entry == &l3->slabs_partial) {
3419 l3->free_touched = 1;
3420 entry = l3->slabs_free.next;
3421 if (entry == &l3->slabs_free)
3425 slabp = list_entry(entry, struct slab, list);
3426 check_spinlock_acquired_node(cachep, nodeid);
3427 check_slabp(cachep, slabp);
3429 STATS_INC_NODEALLOCS(cachep);
3430 STATS_INC_ACTIVE(cachep);
3431 STATS_SET_HIGH(cachep);
3433 BUG_ON(slabp->inuse == cachep->num);
3435 obj = slab_get_obj(cachep, slabp, nodeid);
3436 check_slabp(cachep, slabp);
3438 /* move slabp to correct slabp list: */
3439 list_del(&slabp->list);
3441 if (slabp->free == BUFCTL_END)
3442 list_add(&slabp->list, &l3->slabs_full);
3444 list_add(&slabp->list, &l3->slabs_partial);
3446 spin_unlock(&l3->list_lock);
3450 spin_unlock(&l3->list_lock);
3451 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3455 return fallback_alloc(cachep, flags);
3462 * kmem_cache_alloc_node - Allocate an object on the specified node
3463 * @cachep: The cache to allocate from.
3464 * @flags: See kmalloc().
3465 * @nodeid: node number of the target node.
3466 * @caller: return address of caller, used for debug information
3468 * Identical to kmem_cache_alloc but it will allocate memory on the given
3469 * node, which can improve the performance for cpu bound structures.
3471 * Fallback to other node is possible if __GFP_THISNODE is not set.
3473 static __always_inline void *
3474 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3477 unsigned long save_flags;
3479 int slab_node = numa_mem_id();
3481 flags &= gfp_allowed_mask;
3483 lockdep_trace_alloc(flags);
3485 if (slab_should_failslab(cachep, flags))
3488 cache_alloc_debugcheck_before(cachep, flags);
3489 local_irq_save(save_flags);
3491 if (nodeid == NUMA_NO_NODE)
3494 if (unlikely(!cachep->nodelists[nodeid])) {
3495 /* Node not bootstrapped yet */
3496 ptr = fallback_alloc(cachep, flags);
3500 if (nodeid == slab_node) {
3502 * Use the locally cached objects if possible.
3503 * However ____cache_alloc does not allow fallback
3504 * to other nodes. It may fail while we still have
3505 * objects on other nodes available.
3507 ptr = ____cache_alloc(cachep, flags);
3511 /* ___cache_alloc_node can fall back to other nodes */
3512 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3514 local_irq_restore(save_flags);
3515 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3516 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3520 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3522 if (unlikely((flags & __GFP_ZERO) && ptr))
3523 memset(ptr, 0, cachep->object_size);
3528 static __always_inline void *
3529 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3533 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3534 objp = alternate_node_alloc(cache, flags);
3538 objp = ____cache_alloc(cache, flags);
3541 * We may just have run out of memory on the local node.
3542 * ____cache_alloc_node() knows how to locate memory on other nodes
3545 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3552 static __always_inline void *
3553 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3555 return ____cache_alloc(cachep, flags);
3558 #endif /* CONFIG_NUMA */
3560 static __always_inline void *
3561 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3563 unsigned long save_flags;
3566 flags &= gfp_allowed_mask;
3568 lockdep_trace_alloc(flags);
3570 if (slab_should_failslab(cachep, flags))
3573 cache_alloc_debugcheck_before(cachep, flags);
3574 local_irq_save(save_flags);
3575 objp = __do_cache_alloc(cachep, flags);
3576 local_irq_restore(save_flags);
3577 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3578 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3583 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3585 if (unlikely((flags & __GFP_ZERO) && objp))
3586 memset(objp, 0, cachep->object_size);
3592 * Caller needs to acquire correct kmem_list's list_lock
3594 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3598 struct kmem_list3 *l3;
3600 for (i = 0; i < nr_objects; i++) {
3601 void *objp = objpp[i];
3604 slabp = virt_to_slab(objp);
3605 l3 = cachep->nodelists[node];
3606 list_del(&slabp->list);
3607 check_spinlock_acquired_node(cachep, node);
3608 check_slabp(cachep, slabp);
3609 slab_put_obj(cachep, slabp, objp, node);
3610 STATS_DEC_ACTIVE(cachep);
3612 check_slabp(cachep, slabp);
3614 /* fixup slab chains */
3615 if (slabp->inuse == 0) {
3616 if (l3->free_objects > l3->free_limit) {
3617 l3->free_objects -= cachep->num;
3618 /* No need to drop any previously held
3619 * lock here, even if we have a off-slab slab
3620 * descriptor it is guaranteed to come from
3621 * a different cache, refer to comments before
3624 slab_destroy(cachep, slabp);
3626 list_add(&slabp->list, &l3->slabs_free);
3629 /* Unconditionally move a slab to the end of the
3630 * partial list on free - maximum time for the
3631 * other objects to be freed, too.
3633 list_add_tail(&slabp->list, &l3->slabs_partial);
3638 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3641 struct kmem_list3 *l3;
3642 int node = numa_mem_id();
3644 batchcount = ac->batchcount;
3646 BUG_ON(!batchcount || batchcount > ac->avail);
3649 l3 = cachep->nodelists[node];
3650 spin_lock(&l3->list_lock);
3652 struct array_cache *shared_array = l3->shared;
3653 int max = shared_array->limit - shared_array->avail;
3655 if (batchcount > max)
3657 memcpy(&(shared_array->entry[shared_array->avail]),
3658 ac->entry, sizeof(void *) * batchcount);
3659 shared_array->avail += batchcount;
3664 free_block(cachep, ac->entry, batchcount, node);
3669 struct list_head *p;
3671 p = l3->slabs_free.next;
3672 while (p != &(l3->slabs_free)) {
3675 slabp = list_entry(p, struct slab, list);
3676 BUG_ON(slabp->inuse);
3681 STATS_SET_FREEABLE(cachep, i);
3684 spin_unlock(&l3->list_lock);
3685 ac->avail -= batchcount;
3686 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3690 * Release an obj back to its cache. If the obj has a constructed state, it must
3691 * be in this state _before_ it is released. Called with disabled ints.
3693 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3696 struct array_cache *ac = cpu_cache_get(cachep);
3699 kmemleak_free_recursive(objp, cachep->flags);
3700 objp = cache_free_debugcheck(cachep, objp, caller);
3702 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3705 * Skip calling cache_free_alien() when the platform is not numa.
3706 * This will avoid cache misses that happen while accessing slabp (which
3707 * is per page memory reference) to get nodeid. Instead use a global
3708 * variable to skip the call, which is mostly likely to be present in
3711 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3714 if (likely(ac->avail < ac->limit)) {
3715 STATS_INC_FREEHIT(cachep);
3717 STATS_INC_FREEMISS(cachep);
3718 cache_flusharray(cachep, ac);
3721 ac->entry[ac->avail++] = objp;
3725 * kmem_cache_alloc - Allocate an object
3726 * @cachep: The cache to allocate from.
3727 * @flags: See kmalloc().
3729 * Allocate an object from this cache. The flags are only relevant
3730 * if the cache has no available objects.
3732 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3734 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3736 trace_kmem_cache_alloc(_RET_IP_, ret,
3737 cachep->object_size, cachep->size, flags);
3741 EXPORT_SYMBOL(kmem_cache_alloc);
3743 #ifdef CONFIG_TRACING
3745 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3749 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3751 trace_kmalloc(_RET_IP_, ret,
3752 size, slab_buffer_size(cachep), flags);
3755 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3759 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3761 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3762 __builtin_return_address(0));
3764 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3765 cachep->object_size, cachep->size,
3770 EXPORT_SYMBOL(kmem_cache_alloc_node);
3772 #ifdef CONFIG_TRACING
3773 void *kmem_cache_alloc_node_trace(size_t size,
3774 struct kmem_cache *cachep,
3780 ret = __cache_alloc_node(cachep, flags, nodeid,
3781 __builtin_return_address(0));
3782 trace_kmalloc_node(_RET_IP_, ret,
3783 size, slab_buffer_size(cachep),
3787 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3790 static __always_inline void *
3791 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3793 struct kmem_cache *cachep;
3795 cachep = kmem_find_general_cachep(size, flags);
3796 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3798 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3801 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3802 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3804 return __do_kmalloc_node(size, flags, node,
3805 __builtin_return_address(0));
3807 EXPORT_SYMBOL(__kmalloc_node);
3809 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3810 int node, unsigned long caller)
3812 return __do_kmalloc_node(size, flags, node, (void *)caller);
3814 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3816 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3818 return __do_kmalloc_node(size, flags, node, NULL);
3820 EXPORT_SYMBOL(__kmalloc_node);
3821 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3822 #endif /* CONFIG_NUMA */
3825 * __do_kmalloc - allocate memory
3826 * @size: how many bytes of memory are required.
3827 * @flags: the type of memory to allocate (see kmalloc).
3828 * @caller: function caller for debug tracking of the caller
3830 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3833 struct kmem_cache *cachep;
3836 /* If you want to save a few bytes .text space: replace
3838 * Then kmalloc uses the uninlined functions instead of the inline
3841 cachep = __find_general_cachep(size, flags);
3842 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3844 ret = __cache_alloc(cachep, flags, caller);
3846 trace_kmalloc((unsigned long) caller, ret,
3847 size, cachep->size, flags);
3853 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3854 void *__kmalloc(size_t size, gfp_t flags)
3856 return __do_kmalloc(size, flags, __builtin_return_address(0));
3858 EXPORT_SYMBOL(__kmalloc);
3860 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3862 return __do_kmalloc(size, flags, (void *)caller);
3864 EXPORT_SYMBOL(__kmalloc_track_caller);
3867 void *__kmalloc(size_t size, gfp_t flags)
3869 return __do_kmalloc(size, flags, NULL);
3871 EXPORT_SYMBOL(__kmalloc);
3875 * kmem_cache_free - Deallocate an object
3876 * @cachep: The cache the allocation was from.
3877 * @objp: The previously allocated object.
3879 * Free an object which was previously allocated from this
3882 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3884 unsigned long flags;
3886 local_irq_save(flags);
3887 debug_check_no_locks_freed(objp, cachep->object_size);
3888 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3889 debug_check_no_obj_freed(objp, cachep->object_size);
3890 __cache_free(cachep, objp, __builtin_return_address(0));
3891 local_irq_restore(flags);
3893 trace_kmem_cache_free(_RET_IP_, objp);
3895 EXPORT_SYMBOL(kmem_cache_free);
3898 * kfree - free previously allocated memory
3899 * @objp: pointer returned by kmalloc.
3901 * If @objp is NULL, no operation is performed.
3903 * Don't free memory not originally allocated by kmalloc()
3904 * or you will run into trouble.
3906 void kfree(const void *objp)
3908 struct kmem_cache *c;
3909 unsigned long flags;
3911 trace_kfree(_RET_IP_, objp);
3913 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3915 local_irq_save(flags);
3916 kfree_debugcheck(objp);
3917 c = virt_to_cache(objp);
3918 debug_check_no_locks_freed(objp, c->object_size);
3920 debug_check_no_obj_freed(objp, c->object_size);
3921 __cache_free(c, (void *)objp, __builtin_return_address(0));
3922 local_irq_restore(flags);
3924 EXPORT_SYMBOL(kfree);
3926 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3928 return cachep->object_size;
3930 EXPORT_SYMBOL(kmem_cache_size);
3933 * This initializes kmem_list3 or resizes various caches for all nodes.
3935 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3938 struct kmem_list3 *l3;
3939 struct array_cache *new_shared;
3940 struct array_cache **new_alien = NULL;
3942 for_each_online_node(node) {
3944 if (use_alien_caches) {
3945 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3951 if (cachep->shared) {
3952 new_shared = alloc_arraycache(node,
3953 cachep->shared*cachep->batchcount,
3956 free_alien_cache(new_alien);
3961 l3 = cachep->nodelists[node];
3963 struct array_cache *shared = l3->shared;
3965 spin_lock_irq(&l3->list_lock);
3968 free_block(cachep, shared->entry,
3969 shared->avail, node);
3971 l3->shared = new_shared;
3973 l3->alien = new_alien;
3976 l3->free_limit = (1 + nr_cpus_node(node)) *
3977 cachep->batchcount + cachep->num;
3978 spin_unlock_irq(&l3->list_lock);
3980 free_alien_cache(new_alien);
3983 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3985 free_alien_cache(new_alien);
3990 kmem_list3_init(l3);
3991 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3992 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3993 l3->shared = new_shared;
3994 l3->alien = new_alien;
3995 l3->free_limit = (1 + nr_cpus_node(node)) *
3996 cachep->batchcount + cachep->num;
3997 cachep->nodelists[node] = l3;
4002 if (!cachep->list.next) {
4003 /* Cache is not active yet. Roll back what we did */
4006 if (cachep->nodelists[node]) {
4007 l3 = cachep->nodelists[node];
4010 free_alien_cache(l3->alien);
4012 cachep->nodelists[node] = NULL;
4020 struct ccupdate_struct {
4021 struct kmem_cache *cachep;
4022 struct array_cache *new[0];
4025 static void do_ccupdate_local(void *info)
4027 struct ccupdate_struct *new = info;
4028 struct array_cache *old;
4031 old = cpu_cache_get(new->cachep);
4033 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4034 new->new[smp_processor_id()] = old;
4037 /* Always called with the cache_chain_mutex held */
4038 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4039 int batchcount, int shared, gfp_t gfp)
4041 struct ccupdate_struct *new;
4044 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4049 for_each_online_cpu(i) {
4050 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4053 for (i--; i >= 0; i--)
4059 new->cachep = cachep;
4061 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4064 cachep->batchcount = batchcount;
4065 cachep->limit = limit;
4066 cachep->shared = shared;
4068 for_each_online_cpu(i) {
4069 struct array_cache *ccold = new->new[i];
4072 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4073 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4074 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4078 return alloc_kmemlist(cachep, gfp);
4081 /* Called with cache_chain_mutex held always */
4082 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4088 * The head array serves three purposes:
4089 * - create a LIFO ordering, i.e. return objects that are cache-warm
4090 * - reduce the number of spinlock operations.
4091 * - reduce the number of linked list operations on the slab and
4092 * bufctl chains: array operations are cheaper.
4093 * The numbers are guessed, we should auto-tune as described by
4096 if (cachep->size > 131072)
4098 else if (cachep->size > PAGE_SIZE)
4100 else if (cachep->size > 1024)
4102 else if (cachep->size > 256)
4108 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4109 * allocation behaviour: Most allocs on one cpu, most free operations
4110 * on another cpu. For these cases, an efficient object passing between
4111 * cpus is necessary. This is provided by a shared array. The array
4112 * replaces Bonwick's magazine layer.
4113 * On uniprocessor, it's functionally equivalent (but less efficient)
4114 * to a larger limit. Thus disabled by default.
4117 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4122 * With debugging enabled, large batchcount lead to excessively long
4123 * periods with disabled local interrupts. Limit the batchcount
4128 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4130 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4131 cachep->name, -err);
4136 * Drain an array if it contains any elements taking the l3 lock only if
4137 * necessary. Note that the l3 listlock also protects the array_cache
4138 * if drain_array() is used on the shared array.
4140 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4141 struct array_cache *ac, int force, int node)
4145 if (!ac || !ac->avail)
4147 if (ac->touched && !force) {
4150 spin_lock_irq(&l3->list_lock);
4152 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4153 if (tofree > ac->avail)
4154 tofree = (ac->avail + 1) / 2;
4155 free_block(cachep, ac->entry, tofree, node);
4156 ac->avail -= tofree;
4157 memmove(ac->entry, &(ac->entry[tofree]),
4158 sizeof(void *) * ac->avail);
4160 spin_unlock_irq(&l3->list_lock);
4165 * cache_reap - Reclaim memory from caches.
4166 * @w: work descriptor
4168 * Called from workqueue/eventd every few seconds.
4170 * - clear the per-cpu caches for this CPU.
4171 * - return freeable pages to the main free memory pool.
4173 * If we cannot acquire the cache chain mutex then just give up - we'll try
4174 * again on the next iteration.
4176 static void cache_reap(struct work_struct *w)
4178 struct kmem_cache *searchp;
4179 struct kmem_list3 *l3;
4180 int node = numa_mem_id();
4181 struct delayed_work *work = to_delayed_work(w);
4183 if (!mutex_trylock(&cache_chain_mutex))
4184 /* Give up. Setup the next iteration. */
4187 list_for_each_entry(searchp, &cache_chain, list) {
4191 * We only take the l3 lock if absolutely necessary and we
4192 * have established with reasonable certainty that
4193 * we can do some work if the lock was obtained.
4195 l3 = searchp->nodelists[node];
4197 reap_alien(searchp, l3);
4199 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4202 * These are racy checks but it does not matter
4203 * if we skip one check or scan twice.
4205 if (time_after(l3->next_reap, jiffies))
4208 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4210 drain_array(searchp, l3, l3->shared, 0, node);
4212 if (l3->free_touched)
4213 l3->free_touched = 0;
4217 freed = drain_freelist(searchp, l3, (l3->free_limit +
4218 5 * searchp->num - 1) / (5 * searchp->num));
4219 STATS_ADD_REAPED(searchp, freed);
4225 mutex_unlock(&cache_chain_mutex);
4228 /* Set up the next iteration */
4229 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4232 #ifdef CONFIG_SLABINFO
4234 static void print_slabinfo_header(struct seq_file *m)
4237 * Output format version, so at least we can change it
4238 * without _too_ many complaints.
4241 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4243 seq_puts(m, "slabinfo - version: 2.1\n");
4245 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4246 "<objperslab> <pagesperslab>");
4247 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4248 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4250 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4251 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4252 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4257 static void *s_start(struct seq_file *m, loff_t *pos)
4261 mutex_lock(&cache_chain_mutex);
4263 print_slabinfo_header(m);
4265 return seq_list_start(&cache_chain, *pos);
4268 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4270 return seq_list_next(p, &cache_chain, pos);
4273 static void s_stop(struct seq_file *m, void *p)
4275 mutex_unlock(&cache_chain_mutex);
4278 static int s_show(struct seq_file *m, void *p)
4280 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4282 unsigned long active_objs;
4283 unsigned long num_objs;
4284 unsigned long active_slabs = 0;
4285 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4289 struct kmem_list3 *l3;
4293 for_each_online_node(node) {
4294 l3 = cachep->nodelists[node];
4299 spin_lock_irq(&l3->list_lock);
4301 list_for_each_entry(slabp, &l3->slabs_full, list) {
4302 if (slabp->inuse != cachep->num && !error)
4303 error = "slabs_full accounting error";
4304 active_objs += cachep->num;
4307 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4308 if (slabp->inuse == cachep->num && !error)
4309 error = "slabs_partial inuse accounting error";
4310 if (!slabp->inuse && !error)
4311 error = "slabs_partial/inuse accounting error";
4312 active_objs += slabp->inuse;
4315 list_for_each_entry(slabp, &l3->slabs_free, list) {
4316 if (slabp->inuse && !error)
4317 error = "slabs_free/inuse accounting error";
4320 free_objects += l3->free_objects;
4322 shared_avail += l3->shared->avail;
4324 spin_unlock_irq(&l3->list_lock);
4326 num_slabs += active_slabs;
4327 num_objs = num_slabs * cachep->num;
4328 if (num_objs - active_objs != free_objects && !error)
4329 error = "free_objects accounting error";
4331 name = cachep->name;
4333 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4335 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4336 name, active_objs, num_objs, cachep->size,
4337 cachep->num, (1 << cachep->gfporder));
4338 seq_printf(m, " : tunables %4u %4u %4u",
4339 cachep->limit, cachep->batchcount, cachep->shared);
4340 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4341 active_slabs, num_slabs, shared_avail);
4344 unsigned long high = cachep->high_mark;
4345 unsigned long allocs = cachep->num_allocations;
4346 unsigned long grown = cachep->grown;
4347 unsigned long reaped = cachep->reaped;
4348 unsigned long errors = cachep->errors;
4349 unsigned long max_freeable = cachep->max_freeable;
4350 unsigned long node_allocs = cachep->node_allocs;
4351 unsigned long node_frees = cachep->node_frees;
4352 unsigned long overflows = cachep->node_overflow;
4354 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4355 "%4lu %4lu %4lu %4lu %4lu",
4356 allocs, high, grown,
4357 reaped, errors, max_freeable, node_allocs,
4358 node_frees, overflows);
4362 unsigned long allochit = atomic_read(&cachep->allochit);
4363 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4364 unsigned long freehit = atomic_read(&cachep->freehit);
4365 unsigned long freemiss = atomic_read(&cachep->freemiss);
4367 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4368 allochit, allocmiss, freehit, freemiss);
4376 * slabinfo_op - iterator that generates /proc/slabinfo
4385 * num-pages-per-slab
4386 * + further values on SMP and with statistics enabled
4389 static const struct seq_operations slabinfo_op = {
4396 #define MAX_SLABINFO_WRITE 128
4398 * slabinfo_write - Tuning for the slab allocator
4400 * @buffer: user buffer
4401 * @count: data length
4404 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4405 size_t count, loff_t *ppos)
4407 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4408 int limit, batchcount, shared, res;
4409 struct kmem_cache *cachep;
4411 if (count > MAX_SLABINFO_WRITE)
4413 if (copy_from_user(&kbuf, buffer, count))
4415 kbuf[MAX_SLABINFO_WRITE] = '\0';
4417 tmp = strchr(kbuf, ' ');
4422 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4425 /* Find the cache in the chain of caches. */
4426 mutex_lock(&cache_chain_mutex);
4428 list_for_each_entry(cachep, &cache_chain, list) {
4429 if (!strcmp(cachep->name, kbuf)) {
4430 if (limit < 1 || batchcount < 1 ||
4431 batchcount > limit || shared < 0) {
4434 res = do_tune_cpucache(cachep, limit,
4441 mutex_unlock(&cache_chain_mutex);
4447 static int slabinfo_open(struct inode *inode, struct file *file)
4449 return seq_open(file, &slabinfo_op);
4452 static const struct file_operations proc_slabinfo_operations = {
4453 .open = slabinfo_open,
4455 .write = slabinfo_write,
4456 .llseek = seq_lseek,
4457 .release = seq_release,
4460 #ifdef CONFIG_DEBUG_SLAB_LEAK
4462 static void *leaks_start(struct seq_file *m, loff_t *pos)
4464 mutex_lock(&cache_chain_mutex);
4465 return seq_list_start(&cache_chain, *pos);
4468 static inline int add_caller(unsigned long *n, unsigned long v)
4478 unsigned long *q = p + 2 * i;
4492 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4498 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4504 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4505 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4507 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4512 static void show_symbol(struct seq_file *m, unsigned long address)
4514 #ifdef CONFIG_KALLSYMS
4515 unsigned long offset, size;
4516 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4518 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4519 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4521 seq_printf(m, " [%s]", modname);
4525 seq_printf(m, "%p", (void *)address);
4528 static int leaks_show(struct seq_file *m, void *p)
4530 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4532 struct kmem_list3 *l3;
4534 unsigned long *n = m->private;
4538 if (!(cachep->flags & SLAB_STORE_USER))
4540 if (!(cachep->flags & SLAB_RED_ZONE))
4543 /* OK, we can do it */
4547 for_each_online_node(node) {
4548 l3 = cachep->nodelists[node];
4553 spin_lock_irq(&l3->list_lock);
4555 list_for_each_entry(slabp, &l3->slabs_full, list)
4556 handle_slab(n, cachep, slabp);
4557 list_for_each_entry(slabp, &l3->slabs_partial, list)
4558 handle_slab(n, cachep, slabp);
4559 spin_unlock_irq(&l3->list_lock);
4561 name = cachep->name;
4563 /* Increase the buffer size */
4564 mutex_unlock(&cache_chain_mutex);
4565 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4567 /* Too bad, we are really out */
4569 mutex_lock(&cache_chain_mutex);
4572 *(unsigned long *)m->private = n[0] * 2;
4574 mutex_lock(&cache_chain_mutex);
4575 /* Now make sure this entry will be retried */
4579 for (i = 0; i < n[1]; i++) {
4580 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4581 show_symbol(m, n[2*i+2]);
4588 static const struct seq_operations slabstats_op = {
4589 .start = leaks_start,
4595 static int slabstats_open(struct inode *inode, struct file *file)
4597 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4600 ret = seq_open(file, &slabstats_op);
4602 struct seq_file *m = file->private_data;
4603 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4612 static const struct file_operations proc_slabstats_operations = {
4613 .open = slabstats_open,
4615 .llseek = seq_lseek,
4616 .release = seq_release_private,
4620 static int __init slab_proc_init(void)
4622 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4623 #ifdef CONFIG_DEBUG_SLAB_LEAK
4624 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4628 module_init(slab_proc_init);
4632 * ksize - get the actual amount of memory allocated for a given object
4633 * @objp: Pointer to the object
4635 * kmalloc may internally round up allocations and return more memory
4636 * than requested. ksize() can be used to determine the actual amount of
4637 * memory allocated. The caller may use this additional memory, even though
4638 * a smaller amount of memory was initially specified with the kmalloc call.
4639 * The caller must guarantee that objp points to a valid object previously
4640 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4641 * must not be freed during the duration of the call.
4643 size_t ksize(const void *objp)
4646 if (unlikely(objp == ZERO_SIZE_PTR))
4649 return virt_to_cache(objp)->object_size;
4651 EXPORT_SYMBOL(ksize);