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 'slab_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>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
121 #include <net/sock.h>
123 #include <asm/cacheflush.h>
124 #include <asm/tlbflush.h>
125 #include <asm/page.h>
127 #include <trace/events/kmem.h>
129 #include "internal.h"
132 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
133 * 0 for faster, smaller code (especially in the critical paths).
135 * STATS - 1 to collect stats for /proc/slabinfo.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 #ifdef CONFIG_DEBUG_SLAB
144 #define FORCED_DEBUG 1
148 #define FORCED_DEBUG 0
151 /* Shouldn't this be in a header file somewhere? */
152 #define BYTES_PER_WORD sizeof(void *)
153 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155 #ifndef ARCH_KMALLOC_FLAGS
156 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 * true if a page was allocated from pfmemalloc reserves for network-based
163 static bool pfmemalloc_active __read_mostly;
165 /* Legal flag mask for kmem_cache_create(). */
167 # define CREATE_MASK (SLAB_RED_ZONE | \
168 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
171 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
172 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
173 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
175 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
179 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
185 * Bufctl's are used for linking objs within a slab
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 typedef unsigned int kmem_bufctl_t;
202 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
204 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
205 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
210 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
211 * arrange for kmem_freepages to be called via RCU. This is useful if
212 * we need to approach a kernel structure obliquely, from its address
213 * obtained without the usual locking. We can lock the structure to
214 * stabilize it and check it's still at the given address, only if we
215 * can be sure that the memory has not been meanwhile reused for some
216 * other kind of object (which our subsystem's lock might corrupt).
218 * rcu_read_lock before reading the address, then rcu_read_unlock after
219 * taking the spinlock within the structure expected at that address.
222 struct rcu_head head;
223 struct kmem_cache *cachep;
230 * Manages the objs in a slab. Placed either at the beginning of mem allocated
231 * for a slab, or allocated from an general cache.
232 * Slabs are chained into three list: fully used, partial, fully free slabs.
237 struct list_head list;
238 unsigned long colouroff;
239 void *s_mem; /* including colour offset */
240 unsigned int inuse; /* num of objs active in slab */
242 unsigned short nodeid;
244 struct slab_rcu __slab_cover_slab_rcu;
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
256 * The limit is stored in the per-cpu structure to reduce the data cache
263 unsigned int batchcount;
264 unsigned int touched;
267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
271 * Entries should not be directly dereferenced as
272 * entries belonging to slabs marked pfmemalloc will
273 * have the lower bits set SLAB_OBJ_PFMEMALLOC
277 #define SLAB_OBJ_PFMEMALLOC 1
278 static inline bool is_obj_pfmemalloc(void *objp)
280 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
283 static inline void set_obj_pfmemalloc(void **objp)
285 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
289 static inline void clear_obj_pfmemalloc(void **objp)
291 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
295 * bootstrap: The caches do not work without cpuarrays anymore, but the
296 * cpuarrays are allocated from the generic caches...
298 #define BOOT_CPUCACHE_ENTRIES 1
299 struct arraycache_init {
300 struct array_cache cache;
301 void *entries[BOOT_CPUCACHE_ENTRIES];
305 * The slab lists for all objects.
308 struct list_head slabs_partial; /* partial list first, better asm code */
309 struct list_head slabs_full;
310 struct list_head slabs_free;
311 unsigned long free_objects;
312 unsigned int free_limit;
313 unsigned int colour_next; /* Per-node cache coloring */
314 spinlock_t list_lock;
315 struct array_cache *shared; /* shared per node */
316 struct array_cache **alien; /* on other nodes */
317 unsigned long next_reap; /* updated without locking */
318 int free_touched; /* updated without locking */
322 * Need this for bootstrapping a per node allocator.
324 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
325 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
326 #define CACHE_CACHE 0
327 #define SIZE_AC MAX_NUMNODES
328 #define SIZE_L3 (2 * MAX_NUMNODES)
330 static int drain_freelist(struct kmem_cache *cache,
331 struct kmem_list3 *l3, int tofree);
332 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
334 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
335 static void cache_reap(struct work_struct *unused);
338 * This function must be completely optimized away if a constant is passed to
339 * it. Mostly the same as what is in linux/slab.h except it returns an index.
341 static __always_inline int index_of(const size_t size)
343 extern void __bad_size(void);
345 if (__builtin_constant_p(size)) {
353 #include <linux/kmalloc_sizes.h>
361 static int slab_early_init = 1;
363 #define INDEX_AC index_of(sizeof(struct arraycache_init))
364 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
366 static void kmem_list3_init(struct kmem_list3 *parent)
368 INIT_LIST_HEAD(&parent->slabs_full);
369 INIT_LIST_HEAD(&parent->slabs_partial);
370 INIT_LIST_HEAD(&parent->slabs_free);
371 parent->shared = NULL;
372 parent->alien = NULL;
373 parent->colour_next = 0;
374 spin_lock_init(&parent->list_lock);
375 parent->free_objects = 0;
376 parent->free_touched = 0;
379 #define MAKE_LIST(cachep, listp, slab, nodeid) \
381 INIT_LIST_HEAD(listp); \
382 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
385 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
387 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
388 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
389 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
392 #define CFLGS_OFF_SLAB (0x80000000UL)
393 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
395 #define BATCHREFILL_LIMIT 16
397 * Optimization question: fewer reaps means less probability for unnessary
398 * cpucache drain/refill cycles.
400 * OTOH the cpuarrays can contain lots of objects,
401 * which could lock up otherwise freeable slabs.
403 #define REAPTIMEOUT_CPUC (2*HZ)
404 #define REAPTIMEOUT_LIST3 (4*HZ)
407 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
408 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
409 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
410 #define STATS_INC_GROWN(x) ((x)->grown++)
411 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
412 #define STATS_SET_HIGH(x) \
414 if ((x)->num_active > (x)->high_mark) \
415 (x)->high_mark = (x)->num_active; \
417 #define STATS_INC_ERR(x) ((x)->errors++)
418 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
419 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
420 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
421 #define STATS_SET_FREEABLE(x, i) \
423 if ((x)->max_freeable < i) \
424 (x)->max_freeable = i; \
426 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
427 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
428 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
429 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
431 #define STATS_INC_ACTIVE(x) do { } while (0)
432 #define STATS_DEC_ACTIVE(x) do { } while (0)
433 #define STATS_INC_ALLOCED(x) do { } while (0)
434 #define STATS_INC_GROWN(x) do { } while (0)
435 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
436 #define STATS_SET_HIGH(x) do { } while (0)
437 #define STATS_INC_ERR(x) do { } while (0)
438 #define STATS_INC_NODEALLOCS(x) do { } while (0)
439 #define STATS_INC_NODEFREES(x) do { } while (0)
440 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
441 #define STATS_SET_FREEABLE(x, i) do { } while (0)
442 #define STATS_INC_ALLOCHIT(x) do { } while (0)
443 #define STATS_INC_ALLOCMISS(x) do { } while (0)
444 #define STATS_INC_FREEHIT(x) do { } while (0)
445 #define STATS_INC_FREEMISS(x) do { } while (0)
451 * memory layout of objects:
453 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
454 * the end of an object is aligned with the end of the real
455 * allocation. Catches writes behind the end of the allocation.
456 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
458 * cachep->obj_offset: The real object.
459 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
460 * cachep->size - 1* BYTES_PER_WORD: last caller address
461 * [BYTES_PER_WORD long]
463 static int obj_offset(struct kmem_cache *cachep)
465 return cachep->obj_offset;
468 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
470 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
471 return (unsigned long long*) (objp + obj_offset(cachep) -
472 sizeof(unsigned long long));
475 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
477 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
478 if (cachep->flags & SLAB_STORE_USER)
479 return (unsigned long long *)(objp + cachep->size -
480 sizeof(unsigned long long) -
482 return (unsigned long long *) (objp + cachep->size -
483 sizeof(unsigned long long));
486 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
488 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
489 return (void **)(objp + cachep->size - BYTES_PER_WORD);
494 #define obj_offset(x) 0
495 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
496 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
497 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
501 #ifdef CONFIG_TRACING
502 size_t slab_buffer_size(struct kmem_cache *cachep)
506 EXPORT_SYMBOL(slab_buffer_size);
510 * Do not go above this order unless 0 objects fit into the slab or
511 * overridden on the command line.
513 #define SLAB_MAX_ORDER_HI 1
514 #define SLAB_MAX_ORDER_LO 0
515 static int slab_max_order = SLAB_MAX_ORDER_LO;
516 static bool slab_max_order_set __initdata;
518 static inline struct kmem_cache *virt_to_cache(const void *obj)
520 struct page *page = virt_to_head_page(obj);
521 return page->slab_cache;
524 static inline struct slab *virt_to_slab(const void *obj)
526 struct page *page = virt_to_head_page(obj);
528 VM_BUG_ON(!PageSlab(page));
529 return page->slab_page;
532 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
535 return slab->s_mem + cache->size * idx;
539 * We want to avoid an expensive divide : (offset / cache->size)
540 * Using the fact that size is a constant for a particular cache,
541 * we can replace (offset / cache->size) by
542 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
544 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
545 const struct slab *slab, void *obj)
547 u32 offset = (obj - slab->s_mem);
548 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
552 * These are the default caches for kmalloc. Custom caches can have other sizes.
554 struct cache_sizes malloc_sizes[] = {
555 #define CACHE(x) { .cs_size = (x) },
556 #include <linux/kmalloc_sizes.h>
560 EXPORT_SYMBOL(malloc_sizes);
562 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
568 static struct cache_names __initdata cache_names[] = {
569 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
570 #include <linux/kmalloc_sizes.h>
575 static struct arraycache_init initarray_cache __initdata =
576 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
577 static struct arraycache_init initarray_generic =
578 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
580 /* internal cache of cache description objs */
581 static struct kmem_list3 *kmem_cache_nodelists[MAX_NUMNODES];
582 static struct kmem_cache kmem_cache_boot = {
583 .nodelists = kmem_cache_nodelists,
585 .limit = BOOT_CPUCACHE_ENTRIES,
587 .size = sizeof(struct kmem_cache),
588 .name = "kmem_cache",
591 #define BAD_ALIEN_MAGIC 0x01020304ul
593 #ifdef CONFIG_LOCKDEP
596 * Slab sometimes uses the kmalloc slabs to store the slab headers
597 * for other slabs "off slab".
598 * The locking for this is tricky in that it nests within the locks
599 * of all other slabs in a few places; to deal with this special
600 * locking we put on-slab caches into a separate lock-class.
602 * We set lock class for alien array caches which are up during init.
603 * The lock annotation will be lost if all cpus of a node goes down and
604 * then comes back up during hotplug
606 static struct lock_class_key on_slab_l3_key;
607 static struct lock_class_key on_slab_alc_key;
609 static struct lock_class_key debugobj_l3_key;
610 static struct lock_class_key debugobj_alc_key;
612 static void slab_set_lock_classes(struct kmem_cache *cachep,
613 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
616 struct array_cache **alc;
617 struct kmem_list3 *l3;
620 l3 = cachep->nodelists[q];
624 lockdep_set_class(&l3->list_lock, l3_key);
627 * FIXME: This check for BAD_ALIEN_MAGIC
628 * should go away when common slab code is taught to
629 * work even without alien caches.
630 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
631 * for alloc_alien_cache,
633 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
637 lockdep_set_class(&alc[r]->lock, alc_key);
641 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
643 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
646 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
650 for_each_online_node(node)
651 slab_set_debugobj_lock_classes_node(cachep, node);
654 static void init_node_lock_keys(int q)
656 struct cache_sizes *s = malloc_sizes;
661 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
662 struct kmem_list3 *l3;
664 l3 = s->cs_cachep->nodelists[q];
665 if (!l3 || OFF_SLAB(s->cs_cachep))
668 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
669 &on_slab_alc_key, q);
673 static inline void init_lock_keys(void)
678 init_node_lock_keys(node);
681 static void init_node_lock_keys(int q)
685 static inline void init_lock_keys(void)
689 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
693 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
698 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
700 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
702 return cachep->array[smp_processor_id()];
705 static inline struct kmem_cache *__find_general_cachep(size_t size,
708 struct cache_sizes *csizep = malloc_sizes;
711 /* This happens if someone tries to call
712 * kmem_cache_create(), or __kmalloc(), before
713 * the generic caches are initialized.
715 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
718 return ZERO_SIZE_PTR;
720 while (size > csizep->cs_size)
724 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
725 * has cs_{dma,}cachep==NULL. Thus no special case
726 * for large kmalloc calls required.
728 #ifdef CONFIG_ZONE_DMA
729 if (unlikely(gfpflags & GFP_DMA))
730 return csizep->cs_dmacachep;
732 return csizep->cs_cachep;
735 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
737 return __find_general_cachep(size, gfpflags);
740 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
742 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
746 * Calculate the number of objects and left-over bytes for a given buffer size.
748 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
749 size_t align, int flags, size_t *left_over,
754 size_t slab_size = PAGE_SIZE << gfporder;
757 * The slab management structure can be either off the slab or
758 * on it. For the latter case, the memory allocated for a
762 * - One kmem_bufctl_t for each object
763 * - Padding to respect alignment of @align
764 * - @buffer_size bytes for each object
766 * If the slab management structure is off the slab, then the
767 * alignment will already be calculated into the size. Because
768 * the slabs are all pages aligned, the objects will be at the
769 * correct alignment when allocated.
771 if (flags & CFLGS_OFF_SLAB) {
773 nr_objs = slab_size / buffer_size;
775 if (nr_objs > SLAB_LIMIT)
776 nr_objs = SLAB_LIMIT;
779 * Ignore padding for the initial guess. The padding
780 * is at most @align-1 bytes, and @buffer_size is at
781 * least @align. In the worst case, this result will
782 * be one greater than the number of objects that fit
783 * into the memory allocation when taking the padding
786 nr_objs = (slab_size - sizeof(struct slab)) /
787 (buffer_size + sizeof(kmem_bufctl_t));
790 * This calculated number will be either the right
791 * amount, or one greater than what we want.
793 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
797 if (nr_objs > SLAB_LIMIT)
798 nr_objs = SLAB_LIMIT;
800 mgmt_size = slab_mgmt_size(nr_objs, align);
803 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
807 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
809 static void __slab_error(const char *function, struct kmem_cache *cachep,
812 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
813 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);
929 static inline bool is_slab_pfmemalloc(struct slab *slabp)
931 struct page *page = virt_to_page(slabp->s_mem);
933 return PageSlabPfmemalloc(page);
936 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
937 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
938 struct array_cache *ac)
940 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
944 if (!pfmemalloc_active)
947 spin_lock_irqsave(&l3->list_lock, flags);
948 list_for_each_entry(slabp, &l3->slabs_full, list)
949 if (is_slab_pfmemalloc(slabp))
952 list_for_each_entry(slabp, &l3->slabs_partial, list)
953 if (is_slab_pfmemalloc(slabp))
956 list_for_each_entry(slabp, &l3->slabs_free, list)
957 if (is_slab_pfmemalloc(slabp))
960 pfmemalloc_active = false;
962 spin_unlock_irqrestore(&l3->list_lock, flags);
965 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
966 gfp_t flags, bool force_refill)
969 void *objp = ac->entry[--ac->avail];
971 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
972 if (unlikely(is_obj_pfmemalloc(objp))) {
973 struct kmem_list3 *l3;
975 if (gfp_pfmemalloc_allowed(flags)) {
976 clear_obj_pfmemalloc(&objp);
980 /* The caller cannot use PFMEMALLOC objects, find another one */
981 for (i = 1; i < ac->avail; i++) {
982 /* If a !PFMEMALLOC object is found, swap them */
983 if (!is_obj_pfmemalloc(ac->entry[i])) {
985 ac->entry[i] = ac->entry[ac->avail];
986 ac->entry[ac->avail] = objp;
992 * If there are empty slabs on the slabs_free list and we are
993 * being forced to refill the cache, mark this one !pfmemalloc.
995 l3 = cachep->nodelists[numa_mem_id()];
996 if (!list_empty(&l3->slabs_free) && force_refill) {
997 struct slab *slabp = virt_to_slab(objp);
998 ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem));
999 clear_obj_pfmemalloc(&objp);
1000 recheck_pfmemalloc_active(cachep, ac);
1004 /* No !PFMEMALLOC objects available */
1012 static inline void *ac_get_obj(struct kmem_cache *cachep,
1013 struct array_cache *ac, gfp_t flags, bool force_refill)
1017 if (unlikely(sk_memalloc_socks()))
1018 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1020 objp = ac->entry[--ac->avail];
1025 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1028 if (unlikely(pfmemalloc_active)) {
1029 /* Some pfmemalloc slabs exist, check if this is one */
1030 struct page *page = virt_to_page(objp);
1031 if (PageSlabPfmemalloc(page))
1032 set_obj_pfmemalloc(&objp);
1038 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1041 if (unlikely(sk_memalloc_socks()))
1042 objp = __ac_put_obj(cachep, ac, objp);
1044 ac->entry[ac->avail++] = objp;
1048 * Transfer objects in one arraycache to another.
1049 * Locking must be handled by the caller.
1051 * Return the number of entries transferred.
1053 static int transfer_objects(struct array_cache *to,
1054 struct array_cache *from, unsigned int max)
1056 /* Figure out how many entries to transfer */
1057 int nr = min3(from->avail, max, to->limit - to->avail);
1062 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1063 sizeof(void *) *nr);
1072 #define drain_alien_cache(cachep, alien) do { } while (0)
1073 #define reap_alien(cachep, l3) do { } while (0)
1075 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1077 return (struct array_cache **)BAD_ALIEN_MAGIC;
1080 static inline void free_alien_cache(struct array_cache **ac_ptr)
1084 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1089 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1095 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1096 gfp_t flags, int nodeid)
1101 #else /* CONFIG_NUMA */
1103 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1104 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1106 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1108 struct array_cache **ac_ptr;
1109 int memsize = sizeof(void *) * nr_node_ids;
1114 ac_ptr = kzalloc_node(memsize, gfp, node);
1117 if (i == node || !node_online(i))
1119 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1121 for (i--; i >= 0; i--)
1131 static void free_alien_cache(struct array_cache **ac_ptr)
1142 static void __drain_alien_cache(struct kmem_cache *cachep,
1143 struct array_cache *ac, int node)
1145 struct kmem_list3 *rl3 = cachep->nodelists[node];
1148 spin_lock(&rl3->list_lock);
1150 * Stuff objects into the remote nodes shared array first.
1151 * That way we could avoid the overhead of putting the objects
1152 * into the free lists and getting them back later.
1155 transfer_objects(rl3->shared, ac, ac->limit);
1157 free_block(cachep, ac->entry, ac->avail, node);
1159 spin_unlock(&rl3->list_lock);
1164 * Called from cache_reap() to regularly drain alien caches round robin.
1166 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1168 int node = __this_cpu_read(slab_reap_node);
1171 struct array_cache *ac = l3->alien[node];
1173 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1174 __drain_alien_cache(cachep, ac, node);
1175 spin_unlock_irq(&ac->lock);
1180 static void drain_alien_cache(struct kmem_cache *cachep,
1181 struct array_cache **alien)
1184 struct array_cache *ac;
1185 unsigned long flags;
1187 for_each_online_node(i) {
1190 spin_lock_irqsave(&ac->lock, flags);
1191 __drain_alien_cache(cachep, ac, i);
1192 spin_unlock_irqrestore(&ac->lock, flags);
1197 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1199 struct slab *slabp = virt_to_slab(objp);
1200 int nodeid = slabp->nodeid;
1201 struct kmem_list3 *l3;
1202 struct array_cache *alien = NULL;
1205 node = numa_mem_id();
1208 * Make sure we are not freeing a object from another node to the array
1209 * cache on this cpu.
1211 if (likely(slabp->nodeid == node))
1214 l3 = cachep->nodelists[node];
1215 STATS_INC_NODEFREES(cachep);
1216 if (l3->alien && l3->alien[nodeid]) {
1217 alien = l3->alien[nodeid];
1218 spin_lock(&alien->lock);
1219 if (unlikely(alien->avail == alien->limit)) {
1220 STATS_INC_ACOVERFLOW(cachep);
1221 __drain_alien_cache(cachep, alien, nodeid);
1223 ac_put_obj(cachep, alien, objp);
1224 spin_unlock(&alien->lock);
1226 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1227 free_block(cachep, &objp, 1, nodeid);
1228 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1235 * Allocates and initializes nodelists for a node on each slab cache, used for
1236 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1237 * will be allocated off-node since memory is not yet online for the new node.
1238 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1241 * Must hold slab_mutex.
1243 static int init_cache_nodelists_node(int node)
1245 struct kmem_cache *cachep;
1246 struct kmem_list3 *l3;
1247 const int memsize = sizeof(struct kmem_list3);
1249 list_for_each_entry(cachep, &slab_caches, list) {
1251 * Set up the size64 kmemlist for cpu before we can
1252 * begin anything. Make sure some other cpu on this
1253 * node has not already allocated this
1255 if (!cachep->nodelists[node]) {
1256 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1259 kmem_list3_init(l3);
1260 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1261 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1264 * The l3s don't come and go as CPUs come and
1265 * go. slab_mutex is sufficient
1268 cachep->nodelists[node] = l3;
1271 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1272 cachep->nodelists[node]->free_limit =
1273 (1 + nr_cpus_node(node)) *
1274 cachep->batchcount + cachep->num;
1275 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1280 static void __cpuinit cpuup_canceled(long cpu)
1282 struct kmem_cache *cachep;
1283 struct kmem_list3 *l3 = NULL;
1284 int node = cpu_to_mem(cpu);
1285 const struct cpumask *mask = cpumask_of_node(node);
1287 list_for_each_entry(cachep, &slab_caches, list) {
1288 struct array_cache *nc;
1289 struct array_cache *shared;
1290 struct array_cache **alien;
1292 /* cpu is dead; no one can alloc from it. */
1293 nc = cachep->array[cpu];
1294 cachep->array[cpu] = NULL;
1295 l3 = cachep->nodelists[node];
1298 goto free_array_cache;
1300 spin_lock_irq(&l3->list_lock);
1302 /* Free limit for this kmem_list3 */
1303 l3->free_limit -= cachep->batchcount;
1305 free_block(cachep, nc->entry, nc->avail, node);
1307 if (!cpumask_empty(mask)) {
1308 spin_unlock_irq(&l3->list_lock);
1309 goto free_array_cache;
1312 shared = l3->shared;
1314 free_block(cachep, shared->entry,
1315 shared->avail, node);
1322 spin_unlock_irq(&l3->list_lock);
1326 drain_alien_cache(cachep, alien);
1327 free_alien_cache(alien);
1333 * In the previous loop, all the objects were freed to
1334 * the respective cache's slabs, now we can go ahead and
1335 * shrink each nodelist to its limit.
1337 list_for_each_entry(cachep, &slab_caches, list) {
1338 l3 = cachep->nodelists[node];
1341 drain_freelist(cachep, l3, l3->free_objects);
1345 static int __cpuinit cpuup_prepare(long cpu)
1347 struct kmem_cache *cachep;
1348 struct kmem_list3 *l3 = NULL;
1349 int node = cpu_to_mem(cpu);
1353 * We need to do this right in the beginning since
1354 * alloc_arraycache's are going to use this list.
1355 * kmalloc_node allows us to add the slab to the right
1356 * kmem_list3 and not this cpu's kmem_list3
1358 err = init_cache_nodelists_node(node);
1363 * Now we can go ahead with allocating the shared arrays and
1366 list_for_each_entry(cachep, &slab_caches, list) {
1367 struct array_cache *nc;
1368 struct array_cache *shared = NULL;
1369 struct array_cache **alien = NULL;
1371 nc = alloc_arraycache(node, cachep->limit,
1372 cachep->batchcount, GFP_KERNEL);
1375 if (cachep->shared) {
1376 shared = alloc_arraycache(node,
1377 cachep->shared * cachep->batchcount,
1378 0xbaadf00d, GFP_KERNEL);
1384 if (use_alien_caches) {
1385 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1392 cachep->array[cpu] = nc;
1393 l3 = cachep->nodelists[node];
1396 spin_lock_irq(&l3->list_lock);
1399 * We are serialised from CPU_DEAD or
1400 * CPU_UP_CANCELLED by the cpucontrol lock
1402 l3->shared = shared;
1411 spin_unlock_irq(&l3->list_lock);
1413 free_alien_cache(alien);
1414 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1415 slab_set_debugobj_lock_classes_node(cachep, node);
1417 init_node_lock_keys(node);
1421 cpuup_canceled(cpu);
1425 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1426 unsigned long action, void *hcpu)
1428 long cpu = (long)hcpu;
1432 case CPU_UP_PREPARE:
1433 case CPU_UP_PREPARE_FROZEN:
1434 mutex_lock(&slab_mutex);
1435 err = cpuup_prepare(cpu);
1436 mutex_unlock(&slab_mutex);
1439 case CPU_ONLINE_FROZEN:
1440 start_cpu_timer(cpu);
1442 #ifdef CONFIG_HOTPLUG_CPU
1443 case CPU_DOWN_PREPARE:
1444 case CPU_DOWN_PREPARE_FROZEN:
1446 * Shutdown cache reaper. Note that the slab_mutex is
1447 * held so that if cache_reap() is invoked it cannot do
1448 * anything expensive but will only modify reap_work
1449 * and reschedule the timer.
1451 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1452 /* Now the cache_reaper is guaranteed to be not running. */
1453 per_cpu(slab_reap_work, cpu).work.func = NULL;
1455 case CPU_DOWN_FAILED:
1456 case CPU_DOWN_FAILED_FROZEN:
1457 start_cpu_timer(cpu);
1460 case CPU_DEAD_FROZEN:
1462 * Even if all the cpus of a node are down, we don't free the
1463 * kmem_list3 of any cache. This to avoid a race between
1464 * cpu_down, and a kmalloc allocation from another cpu for
1465 * memory from the node of the cpu going down. The list3
1466 * structure is usually allocated from kmem_cache_create() and
1467 * gets destroyed at kmem_cache_destroy().
1471 case CPU_UP_CANCELED:
1472 case CPU_UP_CANCELED_FROZEN:
1473 mutex_lock(&slab_mutex);
1474 cpuup_canceled(cpu);
1475 mutex_unlock(&slab_mutex);
1478 return notifier_from_errno(err);
1481 static struct notifier_block __cpuinitdata cpucache_notifier = {
1482 &cpuup_callback, NULL, 0
1485 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1487 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1488 * Returns -EBUSY if all objects cannot be drained so that the node is not
1491 * Must hold slab_mutex.
1493 static int __meminit drain_cache_nodelists_node(int node)
1495 struct kmem_cache *cachep;
1498 list_for_each_entry(cachep, &slab_caches, list) {
1499 struct kmem_list3 *l3;
1501 l3 = cachep->nodelists[node];
1505 drain_freelist(cachep, l3, l3->free_objects);
1507 if (!list_empty(&l3->slabs_full) ||
1508 !list_empty(&l3->slabs_partial)) {
1516 static int __meminit slab_memory_callback(struct notifier_block *self,
1517 unsigned long action, void *arg)
1519 struct memory_notify *mnb = arg;
1523 nid = mnb->status_change_nid;
1528 case MEM_GOING_ONLINE:
1529 mutex_lock(&slab_mutex);
1530 ret = init_cache_nodelists_node(nid);
1531 mutex_unlock(&slab_mutex);
1533 case MEM_GOING_OFFLINE:
1534 mutex_lock(&slab_mutex);
1535 ret = drain_cache_nodelists_node(nid);
1536 mutex_unlock(&slab_mutex);
1540 case MEM_CANCEL_ONLINE:
1541 case MEM_CANCEL_OFFLINE:
1545 return notifier_from_errno(ret);
1547 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1550 * swap the static kmem_list3 with kmalloced memory
1552 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1555 struct kmem_list3 *ptr;
1557 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1560 memcpy(ptr, list, sizeof(struct kmem_list3));
1562 * Do not assume that spinlocks can be initialized via memcpy:
1564 spin_lock_init(&ptr->list_lock);
1566 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1567 cachep->nodelists[nodeid] = ptr;
1571 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1572 * size of kmem_list3.
1574 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1578 for_each_online_node(node) {
1579 cachep->nodelists[node] = &initkmem_list3[index + node];
1580 cachep->nodelists[node]->next_reap = jiffies +
1582 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1587 * Initialisation. Called after the page allocator have been initialised and
1588 * before smp_init().
1590 void __init kmem_cache_init(void)
1593 struct cache_sizes *sizes;
1594 struct cache_names *names;
1599 kmem_cache = &kmem_cache_boot;
1601 if (num_possible_nodes() == 1)
1602 use_alien_caches = 0;
1604 for (i = 0; i < NUM_INIT_LISTS; i++) {
1605 kmem_list3_init(&initkmem_list3[i]);
1606 if (i < MAX_NUMNODES)
1607 kmem_cache->nodelists[i] = NULL;
1609 set_up_list3s(kmem_cache, CACHE_CACHE);
1612 * Fragmentation resistance on low memory - only use bigger
1613 * page orders on machines with more than 32MB of memory if
1614 * not overridden on the command line.
1616 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1617 slab_max_order = SLAB_MAX_ORDER_HI;
1619 /* Bootstrap is tricky, because several objects are allocated
1620 * from caches that do not exist yet:
1621 * 1) initialize the kmem_cache cache: it contains the struct
1622 * kmem_cache structures of all caches, except kmem_cache itself:
1623 * kmem_cache is statically allocated.
1624 * Initially an __init data area is used for the head array and the
1625 * kmem_list3 structures, it's replaced with a kmalloc allocated
1626 * array at the end of the bootstrap.
1627 * 2) Create the first kmalloc cache.
1628 * The struct kmem_cache for the new cache is allocated normally.
1629 * An __init data area is used for the head array.
1630 * 3) Create the remaining kmalloc caches, with minimally sized
1632 * 4) Replace the __init data head arrays for kmem_cache and the first
1633 * kmalloc cache with kmalloc allocated arrays.
1634 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1635 * the other cache's with kmalloc allocated memory.
1636 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1639 node = numa_mem_id();
1641 /* 1) create the kmem_cache */
1642 INIT_LIST_HEAD(&slab_caches);
1643 list_add(&kmem_cache->list, &slab_caches);
1644 kmem_cache->colour_off = cache_line_size();
1645 kmem_cache->array[smp_processor_id()] = &initarray_cache.cache;
1646 kmem_cache->nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1649 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1651 kmem_cache->size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1652 nr_node_ids * sizeof(struct kmem_list3 *);
1653 kmem_cache->object_size = kmem_cache->size;
1654 kmem_cache->size = ALIGN(kmem_cache->object_size,
1656 kmem_cache->reciprocal_buffer_size =
1657 reciprocal_value(kmem_cache->size);
1659 for (order = 0; order < MAX_ORDER; order++) {
1660 cache_estimate(order, kmem_cache->size,
1661 cache_line_size(), 0, &left_over, &kmem_cache->num);
1662 if (kmem_cache->num)
1665 BUG_ON(!kmem_cache->num);
1666 kmem_cache->gfporder = order;
1667 kmem_cache->colour = left_over / kmem_cache->colour_off;
1668 kmem_cache->slab_size = ALIGN(kmem_cache->num * sizeof(kmem_bufctl_t) +
1669 sizeof(struct slab), cache_line_size());
1671 /* 2+3) create the kmalloc caches */
1672 sizes = malloc_sizes;
1673 names = cache_names;
1676 * Initialize the caches that provide memory for the array cache and the
1677 * kmem_list3 structures first. Without this, further allocations will
1681 sizes[INDEX_AC].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1682 sizes[INDEX_AC].cs_cachep->name = names[INDEX_AC].name;
1683 sizes[INDEX_AC].cs_cachep->size = sizes[INDEX_AC].cs_size;
1684 sizes[INDEX_AC].cs_cachep->object_size = sizes[INDEX_AC].cs_size;
1685 sizes[INDEX_AC].cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1686 __kmem_cache_create(sizes[INDEX_AC].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1687 list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches);
1689 if (INDEX_AC != INDEX_L3) {
1690 sizes[INDEX_L3].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1691 sizes[INDEX_L3].cs_cachep->name = names[INDEX_L3].name;
1692 sizes[INDEX_L3].cs_cachep->size = sizes[INDEX_L3].cs_size;
1693 sizes[INDEX_L3].cs_cachep->object_size = sizes[INDEX_L3].cs_size;
1694 sizes[INDEX_L3].cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1695 __kmem_cache_create(sizes[INDEX_L3].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1696 list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches);
1699 slab_early_init = 0;
1701 while (sizes->cs_size != ULONG_MAX) {
1703 * For performance, all the general caches are L1 aligned.
1704 * This should be particularly beneficial on SMP boxes, as it
1705 * eliminates "false sharing".
1706 * Note for systems short on memory removing the alignment will
1707 * allow tighter packing of the smaller caches.
1709 if (!sizes->cs_cachep) {
1710 sizes->cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1711 sizes->cs_cachep->name = names->name;
1712 sizes->cs_cachep->size = sizes->cs_size;
1713 sizes->cs_cachep->object_size = sizes->cs_size;
1714 sizes->cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1715 __kmem_cache_create(sizes->cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1716 list_add(&sizes->cs_cachep->list, &slab_caches);
1718 #ifdef CONFIG_ZONE_DMA
1719 sizes->cs_dmacachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1720 sizes->cs_dmacachep->name = names->name_dma;
1721 sizes->cs_dmacachep->size = sizes->cs_size;
1722 sizes->cs_dmacachep->object_size = sizes->cs_size;
1723 sizes->cs_dmacachep->align = ARCH_KMALLOC_MINALIGN;
1724 __kmem_cache_create(sizes->cs_dmacachep,
1725 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| SLAB_PANIC);
1726 list_add(&sizes->cs_dmacachep->list, &slab_caches);
1731 /* 4) Replace the bootstrap head arrays */
1733 struct array_cache *ptr;
1735 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1737 BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache);
1738 memcpy(ptr, cpu_cache_get(kmem_cache),
1739 sizeof(struct arraycache_init));
1741 * Do not assume that spinlocks can be initialized via memcpy:
1743 spin_lock_init(&ptr->lock);
1745 kmem_cache->array[smp_processor_id()] = ptr;
1747 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1749 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1750 != &initarray_generic.cache);
1751 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1752 sizeof(struct arraycache_init));
1754 * Do not assume that spinlocks can be initialized via memcpy:
1756 spin_lock_init(&ptr->lock);
1758 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1761 /* 5) Replace the bootstrap kmem_list3's */
1765 for_each_online_node(nid) {
1766 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1768 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1769 &initkmem_list3[SIZE_AC + nid], nid);
1771 if (INDEX_AC != INDEX_L3) {
1772 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1773 &initkmem_list3[SIZE_L3 + nid], nid);
1781 void __init kmem_cache_init_late(void)
1783 struct kmem_cache *cachep;
1787 /* Annotate slab for lockdep -- annotate the malloc caches */
1790 /* 6) resize the head arrays to their final sizes */
1791 mutex_lock(&slab_mutex);
1792 list_for_each_entry(cachep, &slab_caches, list)
1793 if (enable_cpucache(cachep, GFP_NOWAIT))
1795 mutex_unlock(&slab_mutex);
1801 * Register a cpu startup notifier callback that initializes
1802 * cpu_cache_get for all new cpus
1804 register_cpu_notifier(&cpucache_notifier);
1808 * Register a memory hotplug callback that initializes and frees
1811 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1815 * The reap timers are started later, with a module init call: That part
1816 * of the kernel is not yet operational.
1820 static int __init cpucache_init(void)
1825 * Register the timers that return unneeded pages to the page allocator
1827 for_each_online_cpu(cpu)
1828 start_cpu_timer(cpu);
1834 __initcall(cpucache_init);
1836 static noinline void
1837 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1839 struct kmem_list3 *l3;
1841 unsigned long flags;
1845 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1847 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1848 cachep->name, cachep->size, cachep->gfporder);
1850 for_each_online_node(node) {
1851 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1852 unsigned long active_slabs = 0, num_slabs = 0;
1854 l3 = cachep->nodelists[node];
1858 spin_lock_irqsave(&l3->list_lock, flags);
1859 list_for_each_entry(slabp, &l3->slabs_full, list) {
1860 active_objs += cachep->num;
1863 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1864 active_objs += slabp->inuse;
1867 list_for_each_entry(slabp, &l3->slabs_free, list)
1870 free_objects += l3->free_objects;
1871 spin_unlock_irqrestore(&l3->list_lock, flags);
1873 num_slabs += active_slabs;
1874 num_objs = num_slabs * cachep->num;
1876 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1877 node, active_slabs, num_slabs, active_objs, num_objs,
1883 * Interface to system's page allocator. No need to hold the cache-lock.
1885 * If we requested dmaable memory, we will get it. Even if we
1886 * did not request dmaable memory, we might get it, but that
1887 * would be relatively rare and ignorable.
1889 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1897 * Nommu uses slab's for process anonymous memory allocations, and thus
1898 * requires __GFP_COMP to properly refcount higher order allocations
1900 flags |= __GFP_COMP;
1903 flags |= cachep->allocflags;
1904 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1905 flags |= __GFP_RECLAIMABLE;
1907 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1909 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1910 slab_out_of_memory(cachep, flags, nodeid);
1914 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1915 if (unlikely(page->pfmemalloc))
1916 pfmemalloc_active = true;
1918 nr_pages = (1 << cachep->gfporder);
1919 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1920 add_zone_page_state(page_zone(page),
1921 NR_SLAB_RECLAIMABLE, nr_pages);
1923 add_zone_page_state(page_zone(page),
1924 NR_SLAB_UNRECLAIMABLE, nr_pages);
1925 for (i = 0; i < nr_pages; i++) {
1926 __SetPageSlab(page + i);
1928 if (page->pfmemalloc)
1929 SetPageSlabPfmemalloc(page + i);
1932 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1933 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1936 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1938 kmemcheck_mark_unallocated_pages(page, nr_pages);
1941 return page_address(page);
1945 * Interface to system's page release.
1947 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1949 unsigned long i = (1 << cachep->gfporder);
1950 struct page *page = virt_to_page(addr);
1951 const unsigned long nr_freed = i;
1953 kmemcheck_free_shadow(page, cachep->gfporder);
1955 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1956 sub_zone_page_state(page_zone(page),
1957 NR_SLAB_RECLAIMABLE, nr_freed);
1959 sub_zone_page_state(page_zone(page),
1960 NR_SLAB_UNRECLAIMABLE, nr_freed);
1962 BUG_ON(!PageSlab(page));
1963 __ClearPageSlabPfmemalloc(page);
1964 __ClearPageSlab(page);
1967 if (current->reclaim_state)
1968 current->reclaim_state->reclaimed_slab += nr_freed;
1969 free_pages((unsigned long)addr, cachep->gfporder);
1972 static void kmem_rcu_free(struct rcu_head *head)
1974 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1975 struct kmem_cache *cachep = slab_rcu->cachep;
1977 kmem_freepages(cachep, slab_rcu->addr);
1978 if (OFF_SLAB(cachep))
1979 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1984 #ifdef CONFIG_DEBUG_PAGEALLOC
1985 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1986 unsigned long caller)
1988 int size = cachep->object_size;
1990 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1992 if (size < 5 * sizeof(unsigned long))
1995 *addr++ = 0x12345678;
1997 *addr++ = smp_processor_id();
1998 size -= 3 * sizeof(unsigned long);
2000 unsigned long *sptr = &caller;
2001 unsigned long svalue;
2003 while (!kstack_end(sptr)) {
2005 if (kernel_text_address(svalue)) {
2007 size -= sizeof(unsigned long);
2008 if (size <= sizeof(unsigned long))
2014 *addr++ = 0x87654321;
2018 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
2020 int size = cachep->object_size;
2021 addr = &((char *)addr)[obj_offset(cachep)];
2023 memset(addr, val, size);
2024 *(unsigned char *)(addr + size - 1) = POISON_END;
2027 static void dump_line(char *data, int offset, int limit)
2030 unsigned char error = 0;
2033 printk(KERN_ERR "%03x: ", offset);
2034 for (i = 0; i < limit; i++) {
2035 if (data[offset + i] != POISON_FREE) {
2036 error = data[offset + i];
2040 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2041 &data[offset], limit, 1);
2043 if (bad_count == 1) {
2044 error ^= POISON_FREE;
2045 if (!(error & (error - 1))) {
2046 printk(KERN_ERR "Single bit error detected. Probably "
2049 printk(KERN_ERR "Run memtest86+ or a similar memory "
2052 printk(KERN_ERR "Run a memory test tool.\n");
2061 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2066 if (cachep->flags & SLAB_RED_ZONE) {
2067 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2068 *dbg_redzone1(cachep, objp),
2069 *dbg_redzone2(cachep, objp));
2072 if (cachep->flags & SLAB_STORE_USER) {
2073 printk(KERN_ERR "Last user: [<%p>]",
2074 *dbg_userword(cachep, objp));
2075 print_symbol("(%s)",
2076 (unsigned long)*dbg_userword(cachep, objp));
2079 realobj = (char *)objp + obj_offset(cachep);
2080 size = cachep->object_size;
2081 for (i = 0; i < size && lines; i += 16, lines--) {
2084 if (i + limit > size)
2086 dump_line(realobj, i, limit);
2090 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2096 realobj = (char *)objp + obj_offset(cachep);
2097 size = cachep->object_size;
2099 for (i = 0; i < size; i++) {
2100 char exp = POISON_FREE;
2103 if (realobj[i] != exp) {
2109 "Slab corruption (%s): %s start=%p, len=%d\n",
2110 print_tainted(), cachep->name, realobj, size);
2111 print_objinfo(cachep, objp, 0);
2113 /* Hexdump the affected line */
2116 if (i + limit > size)
2118 dump_line(realobj, i, limit);
2121 /* Limit to 5 lines */
2127 /* Print some data about the neighboring objects, if they
2130 struct slab *slabp = virt_to_slab(objp);
2133 objnr = obj_to_index(cachep, slabp, objp);
2135 objp = index_to_obj(cachep, slabp, objnr - 1);
2136 realobj = (char *)objp + obj_offset(cachep);
2137 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2139 print_objinfo(cachep, objp, 2);
2141 if (objnr + 1 < cachep->num) {
2142 objp = index_to_obj(cachep, slabp, objnr + 1);
2143 realobj = (char *)objp + obj_offset(cachep);
2144 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2146 print_objinfo(cachep, objp, 2);
2153 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2156 for (i = 0; i < cachep->num; i++) {
2157 void *objp = index_to_obj(cachep, slabp, i);
2159 if (cachep->flags & SLAB_POISON) {
2160 #ifdef CONFIG_DEBUG_PAGEALLOC
2161 if (cachep->size % PAGE_SIZE == 0 &&
2163 kernel_map_pages(virt_to_page(objp),
2164 cachep->size / PAGE_SIZE, 1);
2166 check_poison_obj(cachep, objp);
2168 check_poison_obj(cachep, objp);
2171 if (cachep->flags & SLAB_RED_ZONE) {
2172 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2173 slab_error(cachep, "start of a freed object "
2175 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2176 slab_error(cachep, "end of a freed object "
2182 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2188 * slab_destroy - destroy and release all objects in a slab
2189 * @cachep: cache pointer being destroyed
2190 * @slabp: slab pointer being destroyed
2192 * Destroy all the objs in a slab, and release the mem back to the system.
2193 * Before calling the slab must have been unlinked from the cache. The
2194 * cache-lock is not held/needed.
2196 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2198 void *addr = slabp->s_mem - slabp->colouroff;
2200 slab_destroy_debugcheck(cachep, slabp);
2201 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2202 struct slab_rcu *slab_rcu;
2204 slab_rcu = (struct slab_rcu *)slabp;
2205 slab_rcu->cachep = cachep;
2206 slab_rcu->addr = addr;
2207 call_rcu(&slab_rcu->head, kmem_rcu_free);
2209 kmem_freepages(cachep, addr);
2210 if (OFF_SLAB(cachep))
2211 kmem_cache_free(cachep->slabp_cache, slabp);
2216 * calculate_slab_order - calculate size (page order) of slabs
2217 * @cachep: pointer to the cache that is being created
2218 * @size: size of objects to be created in this cache.
2219 * @align: required alignment for the objects.
2220 * @flags: slab allocation flags
2222 * Also calculates the number of objects per slab.
2224 * This could be made much more intelligent. For now, try to avoid using
2225 * high order pages for slabs. When the gfp() functions are more friendly
2226 * towards high-order requests, this should be changed.
2228 static size_t calculate_slab_order(struct kmem_cache *cachep,
2229 size_t size, size_t align, unsigned long flags)
2231 unsigned long offslab_limit;
2232 size_t left_over = 0;
2235 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2239 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2243 if (flags & CFLGS_OFF_SLAB) {
2245 * Max number of objs-per-slab for caches which
2246 * use off-slab slabs. Needed to avoid a possible
2247 * looping condition in cache_grow().
2249 offslab_limit = size - sizeof(struct slab);
2250 offslab_limit /= sizeof(kmem_bufctl_t);
2252 if (num > offslab_limit)
2256 /* Found something acceptable - save it away */
2258 cachep->gfporder = gfporder;
2259 left_over = remainder;
2262 * A VFS-reclaimable slab tends to have most allocations
2263 * as GFP_NOFS and we really don't want to have to be allocating
2264 * higher-order pages when we are unable to shrink dcache.
2266 if (flags & SLAB_RECLAIM_ACCOUNT)
2270 * Large number of objects is good, but very large slabs are
2271 * currently bad for the gfp()s.
2273 if (gfporder >= slab_max_order)
2277 * Acceptable internal fragmentation?
2279 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2285 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2287 if (slab_state >= FULL)
2288 return enable_cpucache(cachep, gfp);
2290 if (slab_state == DOWN) {
2292 * Note: the first kmem_cache_create must create the cache
2293 * that's used by kmalloc(24), otherwise the creation of
2294 * further caches will BUG().
2296 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2299 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2300 * the first cache, then we need to set up all its list3s,
2301 * otherwise the creation of further caches will BUG().
2303 set_up_list3s(cachep, SIZE_AC);
2304 if (INDEX_AC == INDEX_L3)
2305 slab_state = PARTIAL_L3;
2307 slab_state = PARTIAL_ARRAYCACHE;
2309 cachep->array[smp_processor_id()] =
2310 kmalloc(sizeof(struct arraycache_init), gfp);
2312 if (slab_state == PARTIAL_ARRAYCACHE) {
2313 set_up_list3s(cachep, SIZE_L3);
2314 slab_state = PARTIAL_L3;
2317 for_each_online_node(node) {
2318 cachep->nodelists[node] =
2319 kmalloc_node(sizeof(struct kmem_list3),
2321 BUG_ON(!cachep->nodelists[node]);
2322 kmem_list3_init(cachep->nodelists[node]);
2326 cachep->nodelists[numa_mem_id()]->next_reap =
2327 jiffies + REAPTIMEOUT_LIST3 +
2328 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2330 cpu_cache_get(cachep)->avail = 0;
2331 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2332 cpu_cache_get(cachep)->batchcount = 1;
2333 cpu_cache_get(cachep)->touched = 0;
2334 cachep->batchcount = 1;
2335 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2340 * __kmem_cache_create - Create a cache.
2341 * @name: A string which is used in /proc/slabinfo to identify this cache.
2342 * @size: The size of objects to be created in this cache.
2343 * @align: The required alignment for the objects.
2344 * @flags: SLAB flags
2345 * @ctor: A constructor for the objects.
2347 * Returns a ptr to the cache on success, NULL on failure.
2348 * Cannot be called within a int, but can be interrupted.
2349 * The @ctor is run when new pages are allocated by the cache.
2353 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2354 * to catch references to uninitialised memory.
2356 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2357 * for buffer overruns.
2359 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2360 * cacheline. This can be beneficial if you're counting cycles as closely
2364 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2366 size_t left_over, slab_size, ralign;
2369 size_t size = cachep->size;
2374 * Enable redzoning and last user accounting, except for caches with
2375 * large objects, if the increased size would increase the object size
2376 * above the next power of two: caches with object sizes just above a
2377 * power of two have a significant amount of internal fragmentation.
2379 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2380 2 * sizeof(unsigned long long)))
2381 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2382 if (!(flags & SLAB_DESTROY_BY_RCU))
2383 flags |= SLAB_POISON;
2385 if (flags & SLAB_DESTROY_BY_RCU)
2386 BUG_ON(flags & SLAB_POISON);
2389 * Always checks flags, a caller might be expecting debug support which
2392 BUG_ON(flags & ~CREATE_MASK);
2395 * Check that size is in terms of words. This is needed to avoid
2396 * unaligned accesses for some archs when redzoning is used, and makes
2397 * sure any on-slab bufctl's are also correctly aligned.
2399 if (size & (BYTES_PER_WORD - 1)) {
2400 size += (BYTES_PER_WORD - 1);
2401 size &= ~(BYTES_PER_WORD - 1);
2404 /* calculate the final buffer alignment: */
2406 /* 1) arch recommendation: can be overridden for debug */
2407 if (flags & SLAB_HWCACHE_ALIGN) {
2409 * Default alignment: as specified by the arch code. Except if
2410 * an object is really small, then squeeze multiple objects into
2413 ralign = cache_line_size();
2414 while (size <= ralign / 2)
2417 ralign = BYTES_PER_WORD;
2421 * Redzoning and user store require word alignment or possibly larger.
2422 * Note this will be overridden by architecture or caller mandated
2423 * alignment if either is greater than BYTES_PER_WORD.
2425 if (flags & SLAB_STORE_USER)
2426 ralign = BYTES_PER_WORD;
2428 if (flags & SLAB_RED_ZONE) {
2429 ralign = REDZONE_ALIGN;
2430 /* If redzoning, ensure that the second redzone is suitably
2431 * aligned, by adjusting the object size accordingly. */
2432 size += REDZONE_ALIGN - 1;
2433 size &= ~(REDZONE_ALIGN - 1);
2436 /* 2) arch mandated alignment */
2437 if (ralign < ARCH_SLAB_MINALIGN) {
2438 ralign = ARCH_SLAB_MINALIGN;
2440 /* 3) caller mandated alignment */
2441 if (ralign < cachep->align) {
2442 ralign = cachep->align;
2444 /* disable debug if necessary */
2445 if (ralign > __alignof__(unsigned long long))
2446 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2450 cachep->align = ralign;
2452 if (slab_is_available())
2457 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2461 * Both debugging options require word-alignment which is calculated
2464 if (flags & SLAB_RED_ZONE) {
2465 /* add space for red zone words */
2466 cachep->obj_offset += sizeof(unsigned long long);
2467 size += 2 * sizeof(unsigned long long);
2469 if (flags & SLAB_STORE_USER) {
2470 /* user store requires one word storage behind the end of
2471 * the real object. But if the second red zone needs to be
2472 * aligned to 64 bits, we must allow that much space.
2474 if (flags & SLAB_RED_ZONE)
2475 size += REDZONE_ALIGN;
2477 size += BYTES_PER_WORD;
2479 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2480 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2481 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2482 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2489 * Determine if the slab management is 'on' or 'off' slab.
2490 * (bootstrapping cannot cope with offslab caches so don't do
2491 * it too early on. Always use on-slab management when
2492 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2494 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2495 !(flags & SLAB_NOLEAKTRACE))
2497 * Size is large, assume best to place the slab management obj
2498 * off-slab (should allow better packing of objs).
2500 flags |= CFLGS_OFF_SLAB;
2502 size = ALIGN(size, cachep->align);
2504 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2509 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2510 + sizeof(struct slab), cachep->align);
2513 * If the slab has been placed off-slab, and we have enough space then
2514 * move it on-slab. This is at the expense of any extra colouring.
2516 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2517 flags &= ~CFLGS_OFF_SLAB;
2518 left_over -= slab_size;
2521 if (flags & CFLGS_OFF_SLAB) {
2522 /* really off slab. No need for manual alignment */
2524 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2526 #ifdef CONFIG_PAGE_POISONING
2527 /* If we're going to use the generic kernel_map_pages()
2528 * poisoning, then it's going to smash the contents of
2529 * the redzone and userword anyhow, so switch them off.
2531 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2532 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2536 cachep->colour_off = cache_line_size();
2537 /* Offset must be a multiple of the alignment. */
2538 if (cachep->colour_off < cachep->align)
2539 cachep->colour_off = cachep->align;
2540 cachep->colour = left_over / cachep->colour_off;
2541 cachep->slab_size = slab_size;
2542 cachep->flags = flags;
2543 cachep->allocflags = 0;
2544 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2545 cachep->allocflags |= GFP_DMA;
2546 cachep->size = size;
2547 cachep->reciprocal_buffer_size = reciprocal_value(size);
2549 if (flags & CFLGS_OFF_SLAB) {
2550 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2552 * This is a possibility for one of the malloc_sizes caches.
2553 * But since we go off slab only for object size greater than
2554 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2555 * this should not happen at all.
2556 * But leave a BUG_ON for some lucky dude.
2558 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2561 err = setup_cpu_cache(cachep, gfp);
2563 __kmem_cache_shutdown(cachep);
2567 if (flags & SLAB_DEBUG_OBJECTS) {
2569 * Would deadlock through slab_destroy()->call_rcu()->
2570 * debug_object_activate()->kmem_cache_alloc().
2572 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2574 slab_set_debugobj_lock_classes(cachep);
2581 static void check_irq_off(void)
2583 BUG_ON(!irqs_disabled());
2586 static void check_irq_on(void)
2588 BUG_ON(irqs_disabled());
2591 static void check_spinlock_acquired(struct kmem_cache *cachep)
2595 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2599 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2603 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2608 #define check_irq_off() do { } while(0)
2609 #define check_irq_on() do { } while(0)
2610 #define check_spinlock_acquired(x) do { } while(0)
2611 #define check_spinlock_acquired_node(x, y) do { } while(0)
2614 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2615 struct array_cache *ac,
2616 int force, int node);
2618 static void do_drain(void *arg)
2620 struct kmem_cache *cachep = arg;
2621 struct array_cache *ac;
2622 int node = numa_mem_id();
2625 ac = cpu_cache_get(cachep);
2626 spin_lock(&cachep->nodelists[node]->list_lock);
2627 free_block(cachep, ac->entry, ac->avail, node);
2628 spin_unlock(&cachep->nodelists[node]->list_lock);
2632 static void drain_cpu_caches(struct kmem_cache *cachep)
2634 struct kmem_list3 *l3;
2637 on_each_cpu(do_drain, cachep, 1);
2639 for_each_online_node(node) {
2640 l3 = cachep->nodelists[node];
2641 if (l3 && l3->alien)
2642 drain_alien_cache(cachep, l3->alien);
2645 for_each_online_node(node) {
2646 l3 = cachep->nodelists[node];
2648 drain_array(cachep, l3, l3->shared, 1, node);
2653 * Remove slabs from the list of free slabs.
2654 * Specify the number of slabs to drain in tofree.
2656 * Returns the actual number of slabs released.
2658 static int drain_freelist(struct kmem_cache *cache,
2659 struct kmem_list3 *l3, int tofree)
2661 struct list_head *p;
2666 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2668 spin_lock_irq(&l3->list_lock);
2669 p = l3->slabs_free.prev;
2670 if (p == &l3->slabs_free) {
2671 spin_unlock_irq(&l3->list_lock);
2675 slabp = list_entry(p, struct slab, list);
2677 BUG_ON(slabp->inuse);
2679 list_del(&slabp->list);
2681 * Safe to drop the lock. The slab is no longer linked
2684 l3->free_objects -= cache->num;
2685 spin_unlock_irq(&l3->list_lock);
2686 slab_destroy(cache, slabp);
2693 /* Called with slab_mutex held to protect against cpu hotplug */
2694 static int __cache_shrink(struct kmem_cache *cachep)
2697 struct kmem_list3 *l3;
2699 drain_cpu_caches(cachep);
2702 for_each_online_node(i) {
2703 l3 = cachep->nodelists[i];
2707 drain_freelist(cachep, l3, l3->free_objects);
2709 ret += !list_empty(&l3->slabs_full) ||
2710 !list_empty(&l3->slabs_partial);
2712 return (ret ? 1 : 0);
2716 * kmem_cache_shrink - Shrink a cache.
2717 * @cachep: The cache to shrink.
2719 * Releases as many slabs as possible for a cache.
2720 * To help debugging, a zero exit status indicates all slabs were released.
2722 int kmem_cache_shrink(struct kmem_cache *cachep)
2725 BUG_ON(!cachep || in_interrupt());
2728 mutex_lock(&slab_mutex);
2729 ret = __cache_shrink(cachep);
2730 mutex_unlock(&slab_mutex);
2734 EXPORT_SYMBOL(kmem_cache_shrink);
2736 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2739 struct kmem_list3 *l3;
2740 int rc = __cache_shrink(cachep);
2745 for_each_online_cpu(i)
2746 kfree(cachep->array[i]);
2748 /* NUMA: free the list3 structures */
2749 for_each_online_node(i) {
2750 l3 = cachep->nodelists[i];
2753 free_alien_cache(l3->alien);
2761 * Get the memory for a slab management obj.
2762 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2763 * always come from malloc_sizes caches. The slab descriptor cannot
2764 * come from the same cache which is getting created because,
2765 * when we are searching for an appropriate cache for these
2766 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2767 * If we are creating a malloc_sizes cache here it would not be visible to
2768 * kmem_find_general_cachep till the initialization is complete.
2769 * Hence we cannot have slabp_cache same as the original cache.
2771 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2772 int colour_off, gfp_t local_flags,
2777 if (OFF_SLAB(cachep)) {
2778 /* Slab management obj is off-slab. */
2779 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2780 local_flags, nodeid);
2782 * If the first object in the slab is leaked (it's allocated
2783 * but no one has a reference to it), we want to make sure
2784 * kmemleak does not treat the ->s_mem pointer as a reference
2785 * to the object. Otherwise we will not report the leak.
2787 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2792 slabp = objp + colour_off;
2793 colour_off += cachep->slab_size;
2796 slabp->colouroff = colour_off;
2797 slabp->s_mem = objp + colour_off;
2798 slabp->nodeid = nodeid;
2803 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2805 return (kmem_bufctl_t *) (slabp + 1);
2808 static void cache_init_objs(struct kmem_cache *cachep,
2813 for (i = 0; i < cachep->num; i++) {
2814 void *objp = index_to_obj(cachep, slabp, i);
2816 /* need to poison the objs? */
2817 if (cachep->flags & SLAB_POISON)
2818 poison_obj(cachep, objp, POISON_FREE);
2819 if (cachep->flags & SLAB_STORE_USER)
2820 *dbg_userword(cachep, objp) = NULL;
2822 if (cachep->flags & SLAB_RED_ZONE) {
2823 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2824 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2827 * Constructors are not allowed to allocate memory from the same
2828 * cache which they are a constructor for. Otherwise, deadlock.
2829 * They must also be threaded.
2831 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2832 cachep->ctor(objp + obj_offset(cachep));
2834 if (cachep->flags & SLAB_RED_ZONE) {
2835 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2836 slab_error(cachep, "constructor overwrote the"
2837 " end of an object");
2838 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2839 slab_error(cachep, "constructor overwrote the"
2840 " start of an object");
2842 if ((cachep->size % PAGE_SIZE) == 0 &&
2843 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2844 kernel_map_pages(virt_to_page(objp),
2845 cachep->size / PAGE_SIZE, 0);
2850 slab_bufctl(slabp)[i] = i + 1;
2852 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2855 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2857 if (CONFIG_ZONE_DMA_FLAG) {
2858 if (flags & GFP_DMA)
2859 BUG_ON(!(cachep->allocflags & GFP_DMA));
2861 BUG_ON(cachep->allocflags & GFP_DMA);
2865 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2868 void *objp = index_to_obj(cachep, slabp, slabp->free);
2872 next = slab_bufctl(slabp)[slabp->free];
2874 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2875 WARN_ON(slabp->nodeid != nodeid);
2882 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2883 void *objp, int nodeid)
2885 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2888 /* Verify that the slab belongs to the intended node */
2889 WARN_ON(slabp->nodeid != nodeid);
2891 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2892 printk(KERN_ERR "slab: double free detected in cache "
2893 "'%s', objp %p\n", cachep->name, objp);
2897 slab_bufctl(slabp)[objnr] = slabp->free;
2898 slabp->free = objnr;
2903 * Map pages beginning at addr to the given cache and slab. This is required
2904 * for the slab allocator to be able to lookup the cache and slab of a
2905 * virtual address for kfree, ksize, and slab debugging.
2907 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2913 page = virt_to_page(addr);
2916 if (likely(!PageCompound(page)))
2917 nr_pages <<= cache->gfporder;
2920 page->slab_cache = cache;
2921 page->slab_page = slab;
2923 } while (--nr_pages);
2927 * Grow (by 1) the number of slabs within a cache. This is called by
2928 * kmem_cache_alloc() when there are no active objs left in a cache.
2930 static int cache_grow(struct kmem_cache *cachep,
2931 gfp_t flags, int nodeid, void *objp)
2936 struct kmem_list3 *l3;
2939 * Be lazy and only check for valid flags here, keeping it out of the
2940 * critical path in kmem_cache_alloc().
2942 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2943 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2945 /* Take the l3 list lock to change the colour_next on this node */
2947 l3 = cachep->nodelists[nodeid];
2948 spin_lock(&l3->list_lock);
2950 /* Get colour for the slab, and cal the next value. */
2951 offset = l3->colour_next;
2953 if (l3->colour_next >= cachep->colour)
2954 l3->colour_next = 0;
2955 spin_unlock(&l3->list_lock);
2957 offset *= cachep->colour_off;
2959 if (local_flags & __GFP_WAIT)
2963 * The test for missing atomic flag is performed here, rather than
2964 * the more obvious place, simply to reduce the critical path length
2965 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2966 * will eventually be caught here (where it matters).
2968 kmem_flagcheck(cachep, flags);
2971 * Get mem for the objs. Attempt to allocate a physical page from
2975 objp = kmem_getpages(cachep, local_flags, nodeid);
2979 /* Get slab management. */
2980 slabp = alloc_slabmgmt(cachep, objp, offset,
2981 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2985 slab_map_pages(cachep, slabp, objp);
2987 cache_init_objs(cachep, slabp);
2989 if (local_flags & __GFP_WAIT)
2990 local_irq_disable();
2992 spin_lock(&l3->list_lock);
2994 /* Make slab active. */
2995 list_add_tail(&slabp->list, &(l3->slabs_free));
2996 STATS_INC_GROWN(cachep);
2997 l3->free_objects += cachep->num;
2998 spin_unlock(&l3->list_lock);
3001 kmem_freepages(cachep, objp);
3003 if (local_flags & __GFP_WAIT)
3004 local_irq_disable();
3011 * Perform extra freeing checks:
3012 * - detect bad pointers.
3013 * - POISON/RED_ZONE checking
3015 static void kfree_debugcheck(const void *objp)
3017 if (!virt_addr_valid(objp)) {
3018 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3019 (unsigned long)objp);
3024 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3026 unsigned long long redzone1, redzone2;
3028 redzone1 = *dbg_redzone1(cache, obj);
3029 redzone2 = *dbg_redzone2(cache, obj);
3034 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3037 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3038 slab_error(cache, "double free detected");
3040 slab_error(cache, "memory outside object was overwritten");
3042 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3043 obj, redzone1, redzone2);
3046 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3053 BUG_ON(virt_to_cache(objp) != cachep);
3055 objp -= obj_offset(cachep);
3056 kfree_debugcheck(objp);
3057 page = virt_to_head_page(objp);
3059 slabp = page->slab_page;
3061 if (cachep->flags & SLAB_RED_ZONE) {
3062 verify_redzone_free(cachep, objp);
3063 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3064 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3066 if (cachep->flags & SLAB_STORE_USER)
3067 *dbg_userword(cachep, objp) = caller;
3069 objnr = obj_to_index(cachep, slabp, objp);
3071 BUG_ON(objnr >= cachep->num);
3072 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3074 #ifdef CONFIG_DEBUG_SLAB_LEAK
3075 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3077 if (cachep->flags & SLAB_POISON) {
3078 #ifdef CONFIG_DEBUG_PAGEALLOC
3079 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3080 store_stackinfo(cachep, objp, (unsigned long)caller);
3081 kernel_map_pages(virt_to_page(objp),
3082 cachep->size / PAGE_SIZE, 0);
3084 poison_obj(cachep, objp, POISON_FREE);
3087 poison_obj(cachep, objp, POISON_FREE);
3093 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3098 /* Check slab's freelist to see if this obj is there. */
3099 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3101 if (entries > cachep->num || i >= cachep->num)
3104 if (entries != cachep->num - slabp->inuse) {
3106 printk(KERN_ERR "slab: Internal list corruption detected in "
3107 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3108 cachep->name, cachep->num, slabp, slabp->inuse,
3110 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3111 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3117 #define kfree_debugcheck(x) do { } while(0)
3118 #define cache_free_debugcheck(x,objp,z) (objp)
3119 #define check_slabp(x,y) do { } while(0)
3122 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3126 struct kmem_list3 *l3;
3127 struct array_cache *ac;
3131 node = numa_mem_id();
3132 if (unlikely(force_refill))
3135 ac = cpu_cache_get(cachep);
3136 batchcount = ac->batchcount;
3137 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3139 * If there was little recent activity on this cache, then
3140 * perform only a partial refill. Otherwise we could generate
3143 batchcount = BATCHREFILL_LIMIT;
3145 l3 = cachep->nodelists[node];
3147 BUG_ON(ac->avail > 0 || !l3);
3148 spin_lock(&l3->list_lock);
3150 /* See if we can refill from the shared array */
3151 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3152 l3->shared->touched = 1;
3156 while (batchcount > 0) {
3157 struct list_head *entry;
3159 /* Get slab alloc is to come from. */
3160 entry = l3->slabs_partial.next;
3161 if (entry == &l3->slabs_partial) {
3162 l3->free_touched = 1;
3163 entry = l3->slabs_free.next;
3164 if (entry == &l3->slabs_free)
3168 slabp = list_entry(entry, struct slab, list);
3169 check_slabp(cachep, slabp);
3170 check_spinlock_acquired(cachep);
3173 * The slab was either on partial or free list so
3174 * there must be at least one object available for
3177 BUG_ON(slabp->inuse >= cachep->num);
3179 while (slabp->inuse < cachep->num && batchcount--) {
3180 STATS_INC_ALLOCED(cachep);
3181 STATS_INC_ACTIVE(cachep);
3182 STATS_SET_HIGH(cachep);
3184 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3187 check_slabp(cachep, slabp);
3189 /* move slabp to correct slabp list: */
3190 list_del(&slabp->list);
3191 if (slabp->free == BUFCTL_END)
3192 list_add(&slabp->list, &l3->slabs_full);
3194 list_add(&slabp->list, &l3->slabs_partial);
3198 l3->free_objects -= ac->avail;
3200 spin_unlock(&l3->list_lock);
3202 if (unlikely(!ac->avail)) {
3205 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3207 /* cache_grow can reenable interrupts, then ac could change. */
3208 ac = cpu_cache_get(cachep);
3210 /* no objects in sight? abort */
3211 if (!x && (ac->avail == 0 || force_refill))
3214 if (!ac->avail) /* objects refilled by interrupt? */
3219 return ac_get_obj(cachep, ac, flags, force_refill);
3222 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3225 might_sleep_if(flags & __GFP_WAIT);
3227 kmem_flagcheck(cachep, flags);
3232 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3233 gfp_t flags, void *objp, void *caller)
3237 if (cachep->flags & SLAB_POISON) {
3238 #ifdef CONFIG_DEBUG_PAGEALLOC
3239 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3240 kernel_map_pages(virt_to_page(objp),
3241 cachep->size / PAGE_SIZE, 1);
3243 check_poison_obj(cachep, objp);
3245 check_poison_obj(cachep, objp);
3247 poison_obj(cachep, objp, POISON_INUSE);
3249 if (cachep->flags & SLAB_STORE_USER)
3250 *dbg_userword(cachep, objp) = caller;
3252 if (cachep->flags & SLAB_RED_ZONE) {
3253 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3254 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3255 slab_error(cachep, "double free, or memory outside"
3256 " object was overwritten");
3258 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3259 objp, *dbg_redzone1(cachep, objp),
3260 *dbg_redzone2(cachep, objp));
3262 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3263 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3265 #ifdef CONFIG_DEBUG_SLAB_LEAK
3270 slabp = virt_to_head_page(objp)->slab_page;
3271 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3272 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3275 objp += obj_offset(cachep);
3276 if (cachep->ctor && cachep->flags & SLAB_POISON)
3278 if (ARCH_SLAB_MINALIGN &&
3279 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3280 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3281 objp, (int)ARCH_SLAB_MINALIGN);
3286 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3289 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3291 if (cachep == kmem_cache)
3294 return should_failslab(cachep->object_size, flags, cachep->flags);
3297 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3300 struct array_cache *ac;
3301 bool force_refill = false;
3305 ac = cpu_cache_get(cachep);
3306 if (likely(ac->avail)) {
3308 objp = ac_get_obj(cachep, ac, flags, false);
3311 * Allow for the possibility all avail objects are not allowed
3312 * by the current flags
3315 STATS_INC_ALLOCHIT(cachep);
3318 force_refill = true;
3321 STATS_INC_ALLOCMISS(cachep);
3322 objp = cache_alloc_refill(cachep, flags, force_refill);
3324 * the 'ac' may be updated by cache_alloc_refill(),
3325 * and kmemleak_erase() requires its correct value.
3327 ac = cpu_cache_get(cachep);
3331 * To avoid a false negative, if an object that is in one of the
3332 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3333 * treat the array pointers as a reference to the object.
3336 kmemleak_erase(&ac->entry[ac->avail]);
3342 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3344 * If we are in_interrupt, then process context, including cpusets and
3345 * mempolicy, may not apply and should not be used for allocation policy.
3347 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3349 int nid_alloc, nid_here;
3351 if (in_interrupt() || (flags & __GFP_THISNODE))
3353 nid_alloc = nid_here = numa_mem_id();
3354 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3355 nid_alloc = cpuset_slab_spread_node();
3356 else if (current->mempolicy)
3357 nid_alloc = slab_node();
3358 if (nid_alloc != nid_here)
3359 return ____cache_alloc_node(cachep, flags, nid_alloc);
3364 * Fallback function if there was no memory available and no objects on a
3365 * certain node and fall back is permitted. First we scan all the
3366 * available nodelists for available objects. If that fails then we
3367 * perform an allocation without specifying a node. This allows the page
3368 * allocator to do its reclaim / fallback magic. We then insert the
3369 * slab into the proper nodelist and then allocate from it.
3371 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3373 struct zonelist *zonelist;
3377 enum zone_type high_zoneidx = gfp_zone(flags);
3380 unsigned int cpuset_mems_cookie;
3382 if (flags & __GFP_THISNODE)
3385 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3388 cpuset_mems_cookie = get_mems_allowed();
3389 zonelist = node_zonelist(slab_node(), flags);
3393 * Look through allowed nodes for objects available
3394 * from existing per node queues.
3396 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3397 nid = zone_to_nid(zone);
3399 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3400 cache->nodelists[nid] &&
3401 cache->nodelists[nid]->free_objects) {
3402 obj = ____cache_alloc_node(cache,
3403 flags | GFP_THISNODE, nid);
3411 * This allocation will be performed within the constraints
3412 * of the current cpuset / memory policy requirements.
3413 * We may trigger various forms of reclaim on the allowed
3414 * set and go into memory reserves if necessary.
3416 if (local_flags & __GFP_WAIT)
3418 kmem_flagcheck(cache, flags);
3419 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3420 if (local_flags & __GFP_WAIT)
3421 local_irq_disable();
3424 * Insert into the appropriate per node queues
3426 nid = page_to_nid(virt_to_page(obj));
3427 if (cache_grow(cache, flags, nid, obj)) {
3428 obj = ____cache_alloc_node(cache,
3429 flags | GFP_THISNODE, nid);
3432 * Another processor may allocate the
3433 * objects in the slab since we are
3434 * not holding any locks.
3438 /* cache_grow already freed obj */
3444 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3450 * A interface to enable slab creation on nodeid
3452 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3455 struct list_head *entry;
3457 struct kmem_list3 *l3;
3461 l3 = cachep->nodelists[nodeid];
3466 spin_lock(&l3->list_lock);
3467 entry = l3->slabs_partial.next;
3468 if (entry == &l3->slabs_partial) {
3469 l3->free_touched = 1;
3470 entry = l3->slabs_free.next;
3471 if (entry == &l3->slabs_free)
3475 slabp = list_entry(entry, struct slab, list);
3476 check_spinlock_acquired_node(cachep, nodeid);
3477 check_slabp(cachep, slabp);
3479 STATS_INC_NODEALLOCS(cachep);
3480 STATS_INC_ACTIVE(cachep);
3481 STATS_SET_HIGH(cachep);
3483 BUG_ON(slabp->inuse == cachep->num);
3485 obj = slab_get_obj(cachep, slabp, nodeid);
3486 check_slabp(cachep, slabp);
3488 /* move slabp to correct slabp list: */
3489 list_del(&slabp->list);
3491 if (slabp->free == BUFCTL_END)
3492 list_add(&slabp->list, &l3->slabs_full);
3494 list_add(&slabp->list, &l3->slabs_partial);
3496 spin_unlock(&l3->list_lock);
3500 spin_unlock(&l3->list_lock);
3501 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3505 return fallback_alloc(cachep, flags);
3512 * kmem_cache_alloc_node - Allocate an object on the specified node
3513 * @cachep: The cache to allocate from.
3514 * @flags: See kmalloc().
3515 * @nodeid: node number of the target node.
3516 * @caller: return address of caller, used for debug information
3518 * Identical to kmem_cache_alloc but it will allocate memory on the given
3519 * node, which can improve the performance for cpu bound structures.
3521 * Fallback to other node is possible if __GFP_THISNODE is not set.
3523 static __always_inline void *
3524 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3527 unsigned long save_flags;
3529 int slab_node = numa_mem_id();
3531 flags &= gfp_allowed_mask;
3533 lockdep_trace_alloc(flags);
3535 if (slab_should_failslab(cachep, flags))
3538 cache_alloc_debugcheck_before(cachep, flags);
3539 local_irq_save(save_flags);
3541 if (nodeid == NUMA_NO_NODE)
3544 if (unlikely(!cachep->nodelists[nodeid])) {
3545 /* Node not bootstrapped yet */
3546 ptr = fallback_alloc(cachep, flags);
3550 if (nodeid == slab_node) {
3552 * Use the locally cached objects if possible.
3553 * However ____cache_alloc does not allow fallback
3554 * to other nodes. It may fail while we still have
3555 * objects on other nodes available.
3557 ptr = ____cache_alloc(cachep, flags);
3561 /* ___cache_alloc_node can fall back to other nodes */
3562 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3564 local_irq_restore(save_flags);
3565 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3566 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3570 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3572 if (unlikely((flags & __GFP_ZERO) && ptr))
3573 memset(ptr, 0, cachep->object_size);
3578 static __always_inline void *
3579 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3583 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3584 objp = alternate_node_alloc(cache, flags);
3588 objp = ____cache_alloc(cache, flags);
3591 * We may just have run out of memory on the local node.
3592 * ____cache_alloc_node() knows how to locate memory on other nodes
3595 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3602 static __always_inline void *
3603 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3605 return ____cache_alloc(cachep, flags);
3608 #endif /* CONFIG_NUMA */
3610 static __always_inline void *
3611 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3613 unsigned long save_flags;
3616 flags &= gfp_allowed_mask;
3618 lockdep_trace_alloc(flags);
3620 if (slab_should_failslab(cachep, flags))
3623 cache_alloc_debugcheck_before(cachep, flags);
3624 local_irq_save(save_flags);
3625 objp = __do_cache_alloc(cachep, flags);
3626 local_irq_restore(save_flags);
3627 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3628 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3633 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3635 if (unlikely((flags & __GFP_ZERO) && objp))
3636 memset(objp, 0, cachep->object_size);
3642 * Caller needs to acquire correct kmem_list's list_lock
3644 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3648 struct kmem_list3 *l3;
3650 for (i = 0; i < nr_objects; i++) {
3654 clear_obj_pfmemalloc(&objpp[i]);
3657 slabp = virt_to_slab(objp);
3658 l3 = cachep->nodelists[node];
3659 list_del(&slabp->list);
3660 check_spinlock_acquired_node(cachep, node);
3661 check_slabp(cachep, slabp);
3662 slab_put_obj(cachep, slabp, objp, node);
3663 STATS_DEC_ACTIVE(cachep);
3665 check_slabp(cachep, slabp);
3667 /* fixup slab chains */
3668 if (slabp->inuse == 0) {
3669 if (l3->free_objects > l3->free_limit) {
3670 l3->free_objects -= cachep->num;
3671 /* No need to drop any previously held
3672 * lock here, even if we have a off-slab slab
3673 * descriptor it is guaranteed to come from
3674 * a different cache, refer to comments before
3677 slab_destroy(cachep, slabp);
3679 list_add(&slabp->list, &l3->slabs_free);
3682 /* Unconditionally move a slab to the end of the
3683 * partial list on free - maximum time for the
3684 * other objects to be freed, too.
3686 list_add_tail(&slabp->list, &l3->slabs_partial);
3691 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3694 struct kmem_list3 *l3;
3695 int node = numa_mem_id();
3697 batchcount = ac->batchcount;
3699 BUG_ON(!batchcount || batchcount > ac->avail);
3702 l3 = cachep->nodelists[node];
3703 spin_lock(&l3->list_lock);
3705 struct array_cache *shared_array = l3->shared;
3706 int max = shared_array->limit - shared_array->avail;
3708 if (batchcount > max)
3710 memcpy(&(shared_array->entry[shared_array->avail]),
3711 ac->entry, sizeof(void *) * batchcount);
3712 shared_array->avail += batchcount;
3717 free_block(cachep, ac->entry, batchcount, node);
3722 struct list_head *p;
3724 p = l3->slabs_free.next;
3725 while (p != &(l3->slabs_free)) {
3728 slabp = list_entry(p, struct slab, list);
3729 BUG_ON(slabp->inuse);
3734 STATS_SET_FREEABLE(cachep, i);
3737 spin_unlock(&l3->list_lock);
3738 ac->avail -= batchcount;
3739 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3743 * Release an obj back to its cache. If the obj has a constructed state, it must
3744 * be in this state _before_ it is released. Called with disabled ints.
3746 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3749 struct array_cache *ac = cpu_cache_get(cachep);
3752 kmemleak_free_recursive(objp, cachep->flags);
3753 objp = cache_free_debugcheck(cachep, objp, caller);
3755 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3758 * Skip calling cache_free_alien() when the platform is not numa.
3759 * This will avoid cache misses that happen while accessing slabp (which
3760 * is per page memory reference) to get nodeid. Instead use a global
3761 * variable to skip the call, which is mostly likely to be present in
3764 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3767 if (likely(ac->avail < ac->limit)) {
3768 STATS_INC_FREEHIT(cachep);
3770 STATS_INC_FREEMISS(cachep);
3771 cache_flusharray(cachep, ac);
3774 ac_put_obj(cachep, ac, objp);
3778 * kmem_cache_alloc - Allocate an object
3779 * @cachep: The cache to allocate from.
3780 * @flags: See kmalloc().
3782 * Allocate an object from this cache. The flags are only relevant
3783 * if the cache has no available objects.
3785 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3787 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3789 trace_kmem_cache_alloc(_RET_IP_, ret,
3790 cachep->object_size, cachep->size, flags);
3794 EXPORT_SYMBOL(kmem_cache_alloc);
3796 #ifdef CONFIG_TRACING
3798 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3802 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3804 trace_kmalloc(_RET_IP_, ret,
3805 size, slab_buffer_size(cachep), flags);
3808 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3812 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3814 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3815 __builtin_return_address(0));
3817 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3818 cachep->object_size, cachep->size,
3823 EXPORT_SYMBOL(kmem_cache_alloc_node);
3825 #ifdef CONFIG_TRACING
3826 void *kmem_cache_alloc_node_trace(size_t size,
3827 struct kmem_cache *cachep,
3833 ret = __cache_alloc_node(cachep, flags, nodeid,
3834 __builtin_return_address(0));
3835 trace_kmalloc_node(_RET_IP_, ret,
3836 size, slab_buffer_size(cachep),
3840 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3843 static __always_inline void *
3844 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3846 struct kmem_cache *cachep;
3848 cachep = kmem_find_general_cachep(size, flags);
3849 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3851 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3854 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3855 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3857 return __do_kmalloc_node(size, flags, node,
3858 __builtin_return_address(0));
3860 EXPORT_SYMBOL(__kmalloc_node);
3862 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3863 int node, unsigned long caller)
3865 return __do_kmalloc_node(size, flags, node, (void *)caller);
3867 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3869 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3871 return __do_kmalloc_node(size, flags, node, NULL);
3873 EXPORT_SYMBOL(__kmalloc_node);
3874 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3875 #endif /* CONFIG_NUMA */
3878 * __do_kmalloc - allocate memory
3879 * @size: how many bytes of memory are required.
3880 * @flags: the type of memory to allocate (see kmalloc).
3881 * @caller: function caller for debug tracking of the caller
3883 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3886 struct kmem_cache *cachep;
3889 /* If you want to save a few bytes .text space: replace
3891 * Then kmalloc uses the uninlined functions instead of the inline
3894 cachep = __find_general_cachep(size, flags);
3895 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3897 ret = __cache_alloc(cachep, flags, caller);
3899 trace_kmalloc((unsigned long) caller, ret,
3900 size, cachep->size, flags);
3906 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3907 void *__kmalloc(size_t size, gfp_t flags)
3909 return __do_kmalloc(size, flags, __builtin_return_address(0));
3911 EXPORT_SYMBOL(__kmalloc);
3913 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3915 return __do_kmalloc(size, flags, (void *)caller);
3917 EXPORT_SYMBOL(__kmalloc_track_caller);
3920 void *__kmalloc(size_t size, gfp_t flags)
3922 return __do_kmalloc(size, flags, NULL);
3924 EXPORT_SYMBOL(__kmalloc);
3928 * kmem_cache_free - Deallocate an object
3929 * @cachep: The cache the allocation was from.
3930 * @objp: The previously allocated object.
3932 * Free an object which was previously allocated from this
3935 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3937 unsigned long flags;
3939 local_irq_save(flags);
3940 debug_check_no_locks_freed(objp, cachep->object_size);
3941 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3942 debug_check_no_obj_freed(objp, cachep->object_size);
3943 __cache_free(cachep, objp, __builtin_return_address(0));
3944 local_irq_restore(flags);
3946 trace_kmem_cache_free(_RET_IP_, objp);
3948 EXPORT_SYMBOL(kmem_cache_free);
3951 * kfree - free previously allocated memory
3952 * @objp: pointer returned by kmalloc.
3954 * If @objp is NULL, no operation is performed.
3956 * Don't free memory not originally allocated by kmalloc()
3957 * or you will run into trouble.
3959 void kfree(const void *objp)
3961 struct kmem_cache *c;
3962 unsigned long flags;
3964 trace_kfree(_RET_IP_, objp);
3966 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3968 local_irq_save(flags);
3969 kfree_debugcheck(objp);
3970 c = virt_to_cache(objp);
3971 debug_check_no_locks_freed(objp, c->object_size);
3973 debug_check_no_obj_freed(objp, c->object_size);
3974 __cache_free(c, (void *)objp, __builtin_return_address(0));
3975 local_irq_restore(flags);
3977 EXPORT_SYMBOL(kfree);
3979 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3981 return cachep->object_size;
3983 EXPORT_SYMBOL(kmem_cache_size);
3986 * This initializes kmem_list3 or resizes various caches for all nodes.
3988 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3991 struct kmem_list3 *l3;
3992 struct array_cache *new_shared;
3993 struct array_cache **new_alien = NULL;
3995 for_each_online_node(node) {
3997 if (use_alien_caches) {
3998 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
4004 if (cachep->shared) {
4005 new_shared = alloc_arraycache(node,
4006 cachep->shared*cachep->batchcount,
4009 free_alien_cache(new_alien);
4014 l3 = cachep->nodelists[node];
4016 struct array_cache *shared = l3->shared;
4018 spin_lock_irq(&l3->list_lock);
4021 free_block(cachep, shared->entry,
4022 shared->avail, node);
4024 l3->shared = new_shared;
4026 l3->alien = new_alien;
4029 l3->free_limit = (1 + nr_cpus_node(node)) *
4030 cachep->batchcount + cachep->num;
4031 spin_unlock_irq(&l3->list_lock);
4033 free_alien_cache(new_alien);
4036 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4038 free_alien_cache(new_alien);
4043 kmem_list3_init(l3);
4044 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4045 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4046 l3->shared = new_shared;
4047 l3->alien = new_alien;
4048 l3->free_limit = (1 + nr_cpus_node(node)) *
4049 cachep->batchcount + cachep->num;
4050 cachep->nodelists[node] = l3;
4055 if (!cachep->list.next) {
4056 /* Cache is not active yet. Roll back what we did */
4059 if (cachep->nodelists[node]) {
4060 l3 = cachep->nodelists[node];
4063 free_alien_cache(l3->alien);
4065 cachep->nodelists[node] = NULL;
4073 struct ccupdate_struct {
4074 struct kmem_cache *cachep;
4075 struct array_cache *new[0];
4078 static void do_ccupdate_local(void *info)
4080 struct ccupdate_struct *new = info;
4081 struct array_cache *old;
4084 old = cpu_cache_get(new->cachep);
4086 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4087 new->new[smp_processor_id()] = old;
4090 /* Always called with the slab_mutex held */
4091 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4092 int batchcount, int shared, gfp_t gfp)
4094 struct ccupdate_struct *new;
4097 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4102 for_each_online_cpu(i) {
4103 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4106 for (i--; i >= 0; i--)
4112 new->cachep = cachep;
4114 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4117 cachep->batchcount = batchcount;
4118 cachep->limit = limit;
4119 cachep->shared = shared;
4121 for_each_online_cpu(i) {
4122 struct array_cache *ccold = new->new[i];
4125 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4126 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4127 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4131 return alloc_kmemlist(cachep, gfp);
4134 /* Called with slab_mutex held always */
4135 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4141 * The head array serves three purposes:
4142 * - create a LIFO ordering, i.e. return objects that are cache-warm
4143 * - reduce the number of spinlock operations.
4144 * - reduce the number of linked list operations on the slab and
4145 * bufctl chains: array operations are cheaper.
4146 * The numbers are guessed, we should auto-tune as described by
4149 if (cachep->size > 131072)
4151 else if (cachep->size > PAGE_SIZE)
4153 else if (cachep->size > 1024)
4155 else if (cachep->size > 256)
4161 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4162 * allocation behaviour: Most allocs on one cpu, most free operations
4163 * on another cpu. For these cases, an efficient object passing between
4164 * cpus is necessary. This is provided by a shared array. The array
4165 * replaces Bonwick's magazine layer.
4166 * On uniprocessor, it's functionally equivalent (but less efficient)
4167 * to a larger limit. Thus disabled by default.
4170 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4175 * With debugging enabled, large batchcount lead to excessively long
4176 * periods with disabled local interrupts. Limit the batchcount
4181 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4183 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4184 cachep->name, -err);
4189 * Drain an array if it contains any elements taking the l3 lock only if
4190 * necessary. Note that the l3 listlock also protects the array_cache
4191 * if drain_array() is used on the shared array.
4193 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4194 struct array_cache *ac, int force, int node)
4198 if (!ac || !ac->avail)
4200 if (ac->touched && !force) {
4203 spin_lock_irq(&l3->list_lock);
4205 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4206 if (tofree > ac->avail)
4207 tofree = (ac->avail + 1) / 2;
4208 free_block(cachep, ac->entry, tofree, node);
4209 ac->avail -= tofree;
4210 memmove(ac->entry, &(ac->entry[tofree]),
4211 sizeof(void *) * ac->avail);
4213 spin_unlock_irq(&l3->list_lock);
4218 * cache_reap - Reclaim memory from caches.
4219 * @w: work descriptor
4221 * Called from workqueue/eventd every few seconds.
4223 * - clear the per-cpu caches for this CPU.
4224 * - return freeable pages to the main free memory pool.
4226 * If we cannot acquire the cache chain mutex then just give up - we'll try
4227 * again on the next iteration.
4229 static void cache_reap(struct work_struct *w)
4231 struct kmem_cache *searchp;
4232 struct kmem_list3 *l3;
4233 int node = numa_mem_id();
4234 struct delayed_work *work = to_delayed_work(w);
4236 if (!mutex_trylock(&slab_mutex))
4237 /* Give up. Setup the next iteration. */
4240 list_for_each_entry(searchp, &slab_caches, list) {
4244 * We only take the l3 lock if absolutely necessary and we
4245 * have established with reasonable certainty that
4246 * we can do some work if the lock was obtained.
4248 l3 = searchp->nodelists[node];
4250 reap_alien(searchp, l3);
4252 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4255 * These are racy checks but it does not matter
4256 * if we skip one check or scan twice.
4258 if (time_after(l3->next_reap, jiffies))
4261 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4263 drain_array(searchp, l3, l3->shared, 0, node);
4265 if (l3->free_touched)
4266 l3->free_touched = 0;
4270 freed = drain_freelist(searchp, l3, (l3->free_limit +
4271 5 * searchp->num - 1) / (5 * searchp->num));
4272 STATS_ADD_REAPED(searchp, freed);
4278 mutex_unlock(&slab_mutex);
4281 /* Set up the next iteration */
4282 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4285 #ifdef CONFIG_SLABINFO
4287 static void print_slabinfo_header(struct seq_file *m)
4290 * Output format version, so at least we can change it
4291 * without _too_ many complaints.
4294 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4296 seq_puts(m, "slabinfo - version: 2.1\n");
4298 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4299 "<objperslab> <pagesperslab>");
4300 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4301 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4303 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4304 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4305 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4310 static void *s_start(struct seq_file *m, loff_t *pos)
4314 mutex_lock(&slab_mutex);
4316 print_slabinfo_header(m);
4318 return seq_list_start(&slab_caches, *pos);
4321 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4323 return seq_list_next(p, &slab_caches, pos);
4326 static void s_stop(struct seq_file *m, void *p)
4328 mutex_unlock(&slab_mutex);
4331 static int s_show(struct seq_file *m, void *p)
4333 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4335 unsigned long active_objs;
4336 unsigned long num_objs;
4337 unsigned long active_slabs = 0;
4338 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4342 struct kmem_list3 *l3;
4346 for_each_online_node(node) {
4347 l3 = cachep->nodelists[node];
4352 spin_lock_irq(&l3->list_lock);
4354 list_for_each_entry(slabp, &l3->slabs_full, list) {
4355 if (slabp->inuse != cachep->num && !error)
4356 error = "slabs_full accounting error";
4357 active_objs += cachep->num;
4360 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4361 if (slabp->inuse == cachep->num && !error)
4362 error = "slabs_partial inuse accounting error";
4363 if (!slabp->inuse && !error)
4364 error = "slabs_partial/inuse accounting error";
4365 active_objs += slabp->inuse;
4368 list_for_each_entry(slabp, &l3->slabs_free, list) {
4369 if (slabp->inuse && !error)
4370 error = "slabs_free/inuse accounting error";
4373 free_objects += l3->free_objects;
4375 shared_avail += l3->shared->avail;
4377 spin_unlock_irq(&l3->list_lock);
4379 num_slabs += active_slabs;
4380 num_objs = num_slabs * cachep->num;
4381 if (num_objs - active_objs != free_objects && !error)
4382 error = "free_objects accounting error";
4384 name = cachep->name;
4386 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4388 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4389 name, active_objs, num_objs, cachep->size,
4390 cachep->num, (1 << cachep->gfporder));
4391 seq_printf(m, " : tunables %4u %4u %4u",
4392 cachep->limit, cachep->batchcount, cachep->shared);
4393 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4394 active_slabs, num_slabs, shared_avail);
4397 unsigned long high = cachep->high_mark;
4398 unsigned long allocs = cachep->num_allocations;
4399 unsigned long grown = cachep->grown;
4400 unsigned long reaped = cachep->reaped;
4401 unsigned long errors = cachep->errors;
4402 unsigned long max_freeable = cachep->max_freeable;
4403 unsigned long node_allocs = cachep->node_allocs;
4404 unsigned long node_frees = cachep->node_frees;
4405 unsigned long overflows = cachep->node_overflow;
4407 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4408 "%4lu %4lu %4lu %4lu %4lu",
4409 allocs, high, grown,
4410 reaped, errors, max_freeable, node_allocs,
4411 node_frees, overflows);
4415 unsigned long allochit = atomic_read(&cachep->allochit);
4416 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4417 unsigned long freehit = atomic_read(&cachep->freehit);
4418 unsigned long freemiss = atomic_read(&cachep->freemiss);
4420 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4421 allochit, allocmiss, freehit, freemiss);
4429 * slabinfo_op - iterator that generates /proc/slabinfo
4438 * num-pages-per-slab
4439 * + further values on SMP and with statistics enabled
4442 static const struct seq_operations slabinfo_op = {
4449 #define MAX_SLABINFO_WRITE 128
4451 * slabinfo_write - Tuning for the slab allocator
4453 * @buffer: user buffer
4454 * @count: data length
4457 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4458 size_t count, loff_t *ppos)
4460 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4461 int limit, batchcount, shared, res;
4462 struct kmem_cache *cachep;
4464 if (count > MAX_SLABINFO_WRITE)
4466 if (copy_from_user(&kbuf, buffer, count))
4468 kbuf[MAX_SLABINFO_WRITE] = '\0';
4470 tmp = strchr(kbuf, ' ');
4475 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4478 /* Find the cache in the chain of caches. */
4479 mutex_lock(&slab_mutex);
4481 list_for_each_entry(cachep, &slab_caches, list) {
4482 if (!strcmp(cachep->name, kbuf)) {
4483 if (limit < 1 || batchcount < 1 ||
4484 batchcount > limit || shared < 0) {
4487 res = do_tune_cpucache(cachep, limit,
4494 mutex_unlock(&slab_mutex);
4500 static int slabinfo_open(struct inode *inode, struct file *file)
4502 return seq_open(file, &slabinfo_op);
4505 static const struct file_operations proc_slabinfo_operations = {
4506 .open = slabinfo_open,
4508 .write = slabinfo_write,
4509 .llseek = seq_lseek,
4510 .release = seq_release,
4513 #ifdef CONFIG_DEBUG_SLAB_LEAK
4515 static void *leaks_start(struct seq_file *m, loff_t *pos)
4517 mutex_lock(&slab_mutex);
4518 return seq_list_start(&slab_caches, *pos);
4521 static inline int add_caller(unsigned long *n, unsigned long v)
4531 unsigned long *q = p + 2 * i;
4545 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4551 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4557 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4558 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4560 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4565 static void show_symbol(struct seq_file *m, unsigned long address)
4567 #ifdef CONFIG_KALLSYMS
4568 unsigned long offset, size;
4569 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4571 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4572 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4574 seq_printf(m, " [%s]", modname);
4578 seq_printf(m, "%p", (void *)address);
4581 static int leaks_show(struct seq_file *m, void *p)
4583 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4585 struct kmem_list3 *l3;
4587 unsigned long *n = m->private;
4591 if (!(cachep->flags & SLAB_STORE_USER))
4593 if (!(cachep->flags & SLAB_RED_ZONE))
4596 /* OK, we can do it */
4600 for_each_online_node(node) {
4601 l3 = cachep->nodelists[node];
4606 spin_lock_irq(&l3->list_lock);
4608 list_for_each_entry(slabp, &l3->slabs_full, list)
4609 handle_slab(n, cachep, slabp);
4610 list_for_each_entry(slabp, &l3->slabs_partial, list)
4611 handle_slab(n, cachep, slabp);
4612 spin_unlock_irq(&l3->list_lock);
4614 name = cachep->name;
4616 /* Increase the buffer size */
4617 mutex_unlock(&slab_mutex);
4618 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4620 /* Too bad, we are really out */
4622 mutex_lock(&slab_mutex);
4625 *(unsigned long *)m->private = n[0] * 2;
4627 mutex_lock(&slab_mutex);
4628 /* Now make sure this entry will be retried */
4632 for (i = 0; i < n[1]; i++) {
4633 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4634 show_symbol(m, n[2*i+2]);
4641 static const struct seq_operations slabstats_op = {
4642 .start = leaks_start,
4648 static int slabstats_open(struct inode *inode, struct file *file)
4650 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4653 ret = seq_open(file, &slabstats_op);
4655 struct seq_file *m = file->private_data;
4656 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4665 static const struct file_operations proc_slabstats_operations = {
4666 .open = slabstats_open,
4668 .llseek = seq_lseek,
4669 .release = seq_release_private,
4673 static int __init slab_proc_init(void)
4675 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4676 #ifdef CONFIG_DEBUG_SLAB_LEAK
4677 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4681 module_init(slab_proc_init);
4685 * ksize - get the actual amount of memory allocated for a given object
4686 * @objp: Pointer to the object
4688 * kmalloc may internally round up allocations and return more memory
4689 * than requested. ksize() can be used to determine the actual amount of
4690 * memory allocated. The caller may use this additional memory, even though
4691 * a smaller amount of memory was initially specified with the kmalloc call.
4692 * The caller must guarantee that objp points to a valid object previously
4693 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4694 * must not be freed during the duration of the call.
4696 size_t ksize(const void *objp)
4699 if (unlikely(objp == ZERO_SIZE_PTR))
4702 return virt_to_cache(objp)->object_size;
4704 EXPORT_SYMBOL(ksize);