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;
806 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
808 static void __slab_error(const char *function, struct kmem_cache *cachep,
811 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
812 function, cachep->name, msg);
817 * By default on NUMA we use alien caches to stage the freeing of
818 * objects allocated from other nodes. This causes massive memory
819 * inefficiencies when using fake NUMA setup to split memory into a
820 * large number of small nodes, so it can be disabled on the command
824 static int use_alien_caches __read_mostly = 1;
825 static int __init noaliencache_setup(char *s)
827 use_alien_caches = 0;
830 __setup("noaliencache", noaliencache_setup);
832 static int __init slab_max_order_setup(char *str)
834 get_option(&str, &slab_max_order);
835 slab_max_order = slab_max_order < 0 ? 0 :
836 min(slab_max_order, MAX_ORDER - 1);
837 slab_max_order_set = true;
841 __setup("slab_max_order=", slab_max_order_setup);
845 * Special reaping functions for NUMA systems called from cache_reap().
846 * These take care of doing round robin flushing of alien caches (containing
847 * objects freed on different nodes from which they were allocated) and the
848 * flushing of remote pcps by calling drain_node_pages.
850 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
852 static void init_reap_node(int cpu)
856 node = next_node(cpu_to_mem(cpu), node_online_map);
857 if (node == MAX_NUMNODES)
858 node = first_node(node_online_map);
860 per_cpu(slab_reap_node, cpu) = node;
863 static void next_reap_node(void)
865 int node = __this_cpu_read(slab_reap_node);
867 node = next_node(node, node_online_map);
868 if (unlikely(node >= MAX_NUMNODES))
869 node = first_node(node_online_map);
870 __this_cpu_write(slab_reap_node, node);
874 #define init_reap_node(cpu) do { } while (0)
875 #define next_reap_node(void) do { } while (0)
879 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
880 * via the workqueue/eventd.
881 * Add the CPU number into the expiration time to minimize the possibility of
882 * the CPUs getting into lockstep and contending for the global cache chain
885 static void __cpuinit start_cpu_timer(int cpu)
887 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
890 * When this gets called from do_initcalls via cpucache_init(),
891 * init_workqueues() has already run, so keventd will be setup
894 if (keventd_up() && reap_work->work.func == NULL) {
896 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
897 schedule_delayed_work_on(cpu, reap_work,
898 __round_jiffies_relative(HZ, cpu));
902 static struct array_cache *alloc_arraycache(int node, int entries,
903 int batchcount, gfp_t gfp)
905 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
906 struct array_cache *nc = NULL;
908 nc = kmalloc_node(memsize, gfp, node);
910 * The array_cache structures contain pointers to free object.
911 * However, when such objects are allocated or transferred to another
912 * cache the pointers are not cleared and they could be counted as
913 * valid references during a kmemleak scan. Therefore, kmemleak must
914 * not scan such objects.
916 kmemleak_no_scan(nc);
920 nc->batchcount = batchcount;
922 spin_lock_init(&nc->lock);
927 static inline bool is_slab_pfmemalloc(struct slab *slabp)
929 struct page *page = virt_to_page(slabp->s_mem);
931 return PageSlabPfmemalloc(page);
934 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
935 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
936 struct array_cache *ac)
938 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
942 if (!pfmemalloc_active)
945 spin_lock_irqsave(&l3->list_lock, flags);
946 list_for_each_entry(slabp, &l3->slabs_full, list)
947 if (is_slab_pfmemalloc(slabp))
950 list_for_each_entry(slabp, &l3->slabs_partial, list)
951 if (is_slab_pfmemalloc(slabp))
954 list_for_each_entry(slabp, &l3->slabs_free, list)
955 if (is_slab_pfmemalloc(slabp))
958 pfmemalloc_active = false;
960 spin_unlock_irqrestore(&l3->list_lock, flags);
963 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
964 gfp_t flags, bool force_refill)
967 void *objp = ac->entry[--ac->avail];
969 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
970 if (unlikely(is_obj_pfmemalloc(objp))) {
971 struct kmem_list3 *l3;
973 if (gfp_pfmemalloc_allowed(flags)) {
974 clear_obj_pfmemalloc(&objp);
978 /* The caller cannot use PFMEMALLOC objects, find another one */
979 for (i = 1; i < ac->avail; i++) {
980 /* If a !PFMEMALLOC object is found, swap them */
981 if (!is_obj_pfmemalloc(ac->entry[i])) {
983 ac->entry[i] = ac->entry[ac->avail];
984 ac->entry[ac->avail] = objp;
990 * If there are empty slabs on the slabs_free list and we are
991 * being forced to refill the cache, mark this one !pfmemalloc.
993 l3 = cachep->nodelists[numa_mem_id()];
994 if (!list_empty(&l3->slabs_free) && force_refill) {
995 struct slab *slabp = virt_to_slab(objp);
996 ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem));
997 clear_obj_pfmemalloc(&objp);
998 recheck_pfmemalloc_active(cachep, ac);
1002 /* No !PFMEMALLOC objects available */
1010 static inline void *ac_get_obj(struct kmem_cache *cachep,
1011 struct array_cache *ac, gfp_t flags, bool force_refill)
1015 if (unlikely(sk_memalloc_socks()))
1016 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1018 objp = ac->entry[--ac->avail];
1023 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1026 if (unlikely(pfmemalloc_active)) {
1027 /* Some pfmemalloc slabs exist, check if this is one */
1028 struct page *page = virt_to_page(objp);
1029 if (PageSlabPfmemalloc(page))
1030 set_obj_pfmemalloc(&objp);
1036 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1039 if (unlikely(sk_memalloc_socks()))
1040 objp = __ac_put_obj(cachep, ac, objp);
1042 ac->entry[ac->avail++] = objp;
1046 * Transfer objects in one arraycache to another.
1047 * Locking must be handled by the caller.
1049 * Return the number of entries transferred.
1051 static int transfer_objects(struct array_cache *to,
1052 struct array_cache *from, unsigned int max)
1054 /* Figure out how many entries to transfer */
1055 int nr = min3(from->avail, max, to->limit - to->avail);
1060 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1061 sizeof(void *) *nr);
1070 #define drain_alien_cache(cachep, alien) do { } while (0)
1071 #define reap_alien(cachep, l3) do { } while (0)
1073 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1075 return (struct array_cache **)BAD_ALIEN_MAGIC;
1078 static inline void free_alien_cache(struct array_cache **ac_ptr)
1082 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1087 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1093 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1094 gfp_t flags, int nodeid)
1099 #else /* CONFIG_NUMA */
1101 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1102 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1104 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1106 struct array_cache **ac_ptr;
1107 int memsize = sizeof(void *) * nr_node_ids;
1112 ac_ptr = kzalloc_node(memsize, gfp, node);
1115 if (i == node || !node_online(i))
1117 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1119 for (i--; i >= 0; i--)
1129 static void free_alien_cache(struct array_cache **ac_ptr)
1140 static void __drain_alien_cache(struct kmem_cache *cachep,
1141 struct array_cache *ac, int node)
1143 struct kmem_list3 *rl3 = cachep->nodelists[node];
1146 spin_lock(&rl3->list_lock);
1148 * Stuff objects into the remote nodes shared array first.
1149 * That way we could avoid the overhead of putting the objects
1150 * into the free lists and getting them back later.
1153 transfer_objects(rl3->shared, ac, ac->limit);
1155 free_block(cachep, ac->entry, ac->avail, node);
1157 spin_unlock(&rl3->list_lock);
1162 * Called from cache_reap() to regularly drain alien caches round robin.
1164 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1166 int node = __this_cpu_read(slab_reap_node);
1169 struct array_cache *ac = l3->alien[node];
1171 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1172 __drain_alien_cache(cachep, ac, node);
1173 spin_unlock_irq(&ac->lock);
1178 static void drain_alien_cache(struct kmem_cache *cachep,
1179 struct array_cache **alien)
1182 struct array_cache *ac;
1183 unsigned long flags;
1185 for_each_online_node(i) {
1188 spin_lock_irqsave(&ac->lock, flags);
1189 __drain_alien_cache(cachep, ac, i);
1190 spin_unlock_irqrestore(&ac->lock, flags);
1195 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1197 struct slab *slabp = virt_to_slab(objp);
1198 int nodeid = slabp->nodeid;
1199 struct kmem_list3 *l3;
1200 struct array_cache *alien = NULL;
1203 node = numa_mem_id();
1206 * Make sure we are not freeing a object from another node to the array
1207 * cache on this cpu.
1209 if (likely(slabp->nodeid == node))
1212 l3 = cachep->nodelists[node];
1213 STATS_INC_NODEFREES(cachep);
1214 if (l3->alien && l3->alien[nodeid]) {
1215 alien = l3->alien[nodeid];
1216 spin_lock(&alien->lock);
1217 if (unlikely(alien->avail == alien->limit)) {
1218 STATS_INC_ACOVERFLOW(cachep);
1219 __drain_alien_cache(cachep, alien, nodeid);
1221 ac_put_obj(cachep, alien, objp);
1222 spin_unlock(&alien->lock);
1224 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1225 free_block(cachep, &objp, 1, nodeid);
1226 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1233 * Allocates and initializes nodelists for a node on each slab cache, used for
1234 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1235 * will be allocated off-node since memory is not yet online for the new node.
1236 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1239 * Must hold slab_mutex.
1241 static int init_cache_nodelists_node(int node)
1243 struct kmem_cache *cachep;
1244 struct kmem_list3 *l3;
1245 const int memsize = sizeof(struct kmem_list3);
1247 list_for_each_entry(cachep, &slab_caches, list) {
1249 * Set up the size64 kmemlist for cpu before we can
1250 * begin anything. Make sure some other cpu on this
1251 * node has not already allocated this
1253 if (!cachep->nodelists[node]) {
1254 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1257 kmem_list3_init(l3);
1258 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1259 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1262 * The l3s don't come and go as CPUs come and
1263 * go. slab_mutex is sufficient
1266 cachep->nodelists[node] = l3;
1269 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1270 cachep->nodelists[node]->free_limit =
1271 (1 + nr_cpus_node(node)) *
1272 cachep->batchcount + cachep->num;
1273 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1278 static void __cpuinit cpuup_canceled(long cpu)
1280 struct kmem_cache *cachep;
1281 struct kmem_list3 *l3 = NULL;
1282 int node = cpu_to_mem(cpu);
1283 const struct cpumask *mask = cpumask_of_node(node);
1285 list_for_each_entry(cachep, &slab_caches, list) {
1286 struct array_cache *nc;
1287 struct array_cache *shared;
1288 struct array_cache **alien;
1290 /* cpu is dead; no one can alloc from it. */
1291 nc = cachep->array[cpu];
1292 cachep->array[cpu] = NULL;
1293 l3 = cachep->nodelists[node];
1296 goto free_array_cache;
1298 spin_lock_irq(&l3->list_lock);
1300 /* Free limit for this kmem_list3 */
1301 l3->free_limit -= cachep->batchcount;
1303 free_block(cachep, nc->entry, nc->avail, node);
1305 if (!cpumask_empty(mask)) {
1306 spin_unlock_irq(&l3->list_lock);
1307 goto free_array_cache;
1310 shared = l3->shared;
1312 free_block(cachep, shared->entry,
1313 shared->avail, node);
1320 spin_unlock_irq(&l3->list_lock);
1324 drain_alien_cache(cachep, alien);
1325 free_alien_cache(alien);
1331 * In the previous loop, all the objects were freed to
1332 * the respective cache's slabs, now we can go ahead and
1333 * shrink each nodelist to its limit.
1335 list_for_each_entry(cachep, &slab_caches, list) {
1336 l3 = cachep->nodelists[node];
1339 drain_freelist(cachep, l3, l3->free_objects);
1343 static int __cpuinit cpuup_prepare(long cpu)
1345 struct kmem_cache *cachep;
1346 struct kmem_list3 *l3 = NULL;
1347 int node = cpu_to_mem(cpu);
1351 * We need to do this right in the beginning since
1352 * alloc_arraycache's are going to use this list.
1353 * kmalloc_node allows us to add the slab to the right
1354 * kmem_list3 and not this cpu's kmem_list3
1356 err = init_cache_nodelists_node(node);
1361 * Now we can go ahead with allocating the shared arrays and
1364 list_for_each_entry(cachep, &slab_caches, list) {
1365 struct array_cache *nc;
1366 struct array_cache *shared = NULL;
1367 struct array_cache **alien = NULL;
1369 nc = alloc_arraycache(node, cachep->limit,
1370 cachep->batchcount, GFP_KERNEL);
1373 if (cachep->shared) {
1374 shared = alloc_arraycache(node,
1375 cachep->shared * cachep->batchcount,
1376 0xbaadf00d, GFP_KERNEL);
1382 if (use_alien_caches) {
1383 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1390 cachep->array[cpu] = nc;
1391 l3 = cachep->nodelists[node];
1394 spin_lock_irq(&l3->list_lock);
1397 * We are serialised from CPU_DEAD or
1398 * CPU_UP_CANCELLED by the cpucontrol lock
1400 l3->shared = shared;
1409 spin_unlock_irq(&l3->list_lock);
1411 free_alien_cache(alien);
1412 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1413 slab_set_debugobj_lock_classes_node(cachep, node);
1415 init_node_lock_keys(node);
1419 cpuup_canceled(cpu);
1423 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1424 unsigned long action, void *hcpu)
1426 long cpu = (long)hcpu;
1430 case CPU_UP_PREPARE:
1431 case CPU_UP_PREPARE_FROZEN:
1432 mutex_lock(&slab_mutex);
1433 err = cpuup_prepare(cpu);
1434 mutex_unlock(&slab_mutex);
1437 case CPU_ONLINE_FROZEN:
1438 start_cpu_timer(cpu);
1440 #ifdef CONFIG_HOTPLUG_CPU
1441 case CPU_DOWN_PREPARE:
1442 case CPU_DOWN_PREPARE_FROZEN:
1444 * Shutdown cache reaper. Note that the slab_mutex is
1445 * held so that if cache_reap() is invoked it cannot do
1446 * anything expensive but will only modify reap_work
1447 * and reschedule the timer.
1449 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1450 /* Now the cache_reaper is guaranteed to be not running. */
1451 per_cpu(slab_reap_work, cpu).work.func = NULL;
1453 case CPU_DOWN_FAILED:
1454 case CPU_DOWN_FAILED_FROZEN:
1455 start_cpu_timer(cpu);
1458 case CPU_DEAD_FROZEN:
1460 * Even if all the cpus of a node are down, we don't free the
1461 * kmem_list3 of any cache. This to avoid a race between
1462 * cpu_down, and a kmalloc allocation from another cpu for
1463 * memory from the node of the cpu going down. The list3
1464 * structure is usually allocated from kmem_cache_create() and
1465 * gets destroyed at kmem_cache_destroy().
1469 case CPU_UP_CANCELED:
1470 case CPU_UP_CANCELED_FROZEN:
1471 mutex_lock(&slab_mutex);
1472 cpuup_canceled(cpu);
1473 mutex_unlock(&slab_mutex);
1476 return notifier_from_errno(err);
1479 static struct notifier_block __cpuinitdata cpucache_notifier = {
1480 &cpuup_callback, NULL, 0
1483 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1485 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1486 * Returns -EBUSY if all objects cannot be drained so that the node is not
1489 * Must hold slab_mutex.
1491 static int __meminit drain_cache_nodelists_node(int node)
1493 struct kmem_cache *cachep;
1496 list_for_each_entry(cachep, &slab_caches, list) {
1497 struct kmem_list3 *l3;
1499 l3 = cachep->nodelists[node];
1503 drain_freelist(cachep, l3, l3->free_objects);
1505 if (!list_empty(&l3->slabs_full) ||
1506 !list_empty(&l3->slabs_partial)) {
1514 static int __meminit slab_memory_callback(struct notifier_block *self,
1515 unsigned long action, void *arg)
1517 struct memory_notify *mnb = arg;
1521 nid = mnb->status_change_nid;
1526 case MEM_GOING_ONLINE:
1527 mutex_lock(&slab_mutex);
1528 ret = init_cache_nodelists_node(nid);
1529 mutex_unlock(&slab_mutex);
1531 case MEM_GOING_OFFLINE:
1532 mutex_lock(&slab_mutex);
1533 ret = drain_cache_nodelists_node(nid);
1534 mutex_unlock(&slab_mutex);
1538 case MEM_CANCEL_ONLINE:
1539 case MEM_CANCEL_OFFLINE:
1543 return notifier_from_errno(ret);
1545 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1548 * swap the static kmem_list3 with kmalloced memory
1550 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1553 struct kmem_list3 *ptr;
1555 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1558 memcpy(ptr, list, sizeof(struct kmem_list3));
1560 * Do not assume that spinlocks can be initialized via memcpy:
1562 spin_lock_init(&ptr->list_lock);
1564 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1565 cachep->nodelists[nodeid] = ptr;
1569 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1570 * size of kmem_list3.
1572 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1576 for_each_online_node(node) {
1577 cachep->nodelists[node] = &initkmem_list3[index + node];
1578 cachep->nodelists[node]->next_reap = jiffies +
1580 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1585 * Initialisation. Called after the page allocator have been initialised and
1586 * before smp_init().
1588 void __init kmem_cache_init(void)
1591 struct cache_sizes *sizes;
1592 struct cache_names *names;
1597 kmem_cache = &kmem_cache_boot;
1599 if (num_possible_nodes() == 1)
1600 use_alien_caches = 0;
1602 for (i = 0; i < NUM_INIT_LISTS; i++) {
1603 kmem_list3_init(&initkmem_list3[i]);
1604 if (i < MAX_NUMNODES)
1605 kmem_cache->nodelists[i] = NULL;
1607 set_up_list3s(kmem_cache, CACHE_CACHE);
1610 * Fragmentation resistance on low memory - only use bigger
1611 * page orders on machines with more than 32MB of memory if
1612 * not overridden on the command line.
1614 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1615 slab_max_order = SLAB_MAX_ORDER_HI;
1617 /* Bootstrap is tricky, because several objects are allocated
1618 * from caches that do not exist yet:
1619 * 1) initialize the kmem_cache cache: it contains the struct
1620 * kmem_cache structures of all caches, except kmem_cache itself:
1621 * kmem_cache is statically allocated.
1622 * Initially an __init data area is used for the head array and the
1623 * kmem_list3 structures, it's replaced with a kmalloc allocated
1624 * array at the end of the bootstrap.
1625 * 2) Create the first kmalloc cache.
1626 * The struct kmem_cache for the new cache is allocated normally.
1627 * An __init data area is used for the head array.
1628 * 3) Create the remaining kmalloc caches, with minimally sized
1630 * 4) Replace the __init data head arrays for kmem_cache and the first
1631 * kmalloc cache with kmalloc allocated arrays.
1632 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1633 * the other cache's with kmalloc allocated memory.
1634 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1637 node = numa_mem_id();
1639 /* 1) create the kmem_cache */
1640 INIT_LIST_HEAD(&slab_caches);
1641 list_add(&kmem_cache->list, &slab_caches);
1642 kmem_cache->colour_off = cache_line_size();
1643 kmem_cache->array[smp_processor_id()] = &initarray_cache.cache;
1644 kmem_cache->nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1647 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1649 kmem_cache->size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1650 nr_node_ids * sizeof(struct kmem_list3 *);
1651 kmem_cache->object_size = kmem_cache->size;
1652 kmem_cache->size = ALIGN(kmem_cache->object_size,
1654 kmem_cache->reciprocal_buffer_size =
1655 reciprocal_value(kmem_cache->size);
1657 for (order = 0; order < MAX_ORDER; order++) {
1658 cache_estimate(order, kmem_cache->size,
1659 cache_line_size(), 0, &left_over, &kmem_cache->num);
1660 if (kmem_cache->num)
1663 BUG_ON(!kmem_cache->num);
1664 kmem_cache->gfporder = order;
1665 kmem_cache->colour = left_over / kmem_cache->colour_off;
1666 kmem_cache->slab_size = ALIGN(kmem_cache->num * sizeof(kmem_bufctl_t) +
1667 sizeof(struct slab), cache_line_size());
1669 /* 2+3) create the kmalloc caches */
1670 sizes = malloc_sizes;
1671 names = cache_names;
1674 * Initialize the caches that provide memory for the array cache and the
1675 * kmem_list3 structures first. Without this, further allocations will
1679 sizes[INDEX_AC].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1680 __kmem_cache_create(sizes[INDEX_AC].cs_cachep, names[INDEX_AC].name,
1681 sizes[INDEX_AC].cs_size,
1682 ARCH_KMALLOC_MINALIGN,
1683 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1686 list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches);
1687 if (INDEX_AC != INDEX_L3) {
1688 sizes[INDEX_L3].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1689 __kmem_cache_create(sizes[INDEX_L3].cs_cachep, names[INDEX_L3].name,
1690 sizes[INDEX_L3].cs_size,
1691 ARCH_KMALLOC_MINALIGN,
1692 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1694 list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches);
1697 slab_early_init = 0;
1699 while (sizes->cs_size != ULONG_MAX) {
1701 * For performance, all the general caches are L1 aligned.
1702 * This should be particularly beneficial on SMP boxes, as it
1703 * eliminates "false sharing".
1704 * Note for systems short on memory removing the alignment will
1705 * allow tighter packing of the smaller caches.
1707 if (!sizes->cs_cachep) {
1708 sizes->cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1709 __kmem_cache_create(sizes->cs_cachep, names->name,
1711 ARCH_KMALLOC_MINALIGN,
1712 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1714 list_add(&sizes->cs_cachep->list, &slab_caches);
1716 #ifdef CONFIG_ZONE_DMA
1717 sizes->cs_dmacachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1718 __kmem_cache_create(sizes->cs_dmacachep,
1721 ARCH_KMALLOC_MINALIGN,
1722 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1725 list_add(&sizes->cs_dmacachep->list, &slab_caches);
1730 /* 4) Replace the bootstrap head arrays */
1732 struct array_cache *ptr;
1734 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1736 BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache);
1737 memcpy(ptr, cpu_cache_get(kmem_cache),
1738 sizeof(struct arraycache_init));
1740 * Do not assume that spinlocks can be initialized via memcpy:
1742 spin_lock_init(&ptr->lock);
1744 kmem_cache->array[smp_processor_id()] = ptr;
1746 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1748 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1749 != &initarray_generic.cache);
1750 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1751 sizeof(struct arraycache_init));
1753 * Do not assume that spinlocks can be initialized via memcpy:
1755 spin_lock_init(&ptr->lock);
1757 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1760 /* 5) Replace the bootstrap kmem_list3's */
1764 for_each_online_node(nid) {
1765 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1767 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1768 &initkmem_list3[SIZE_AC + nid], nid);
1770 if (INDEX_AC != INDEX_L3) {
1771 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1772 &initkmem_list3[SIZE_L3 + nid], nid);
1780 void __init kmem_cache_init_late(void)
1782 struct kmem_cache *cachep;
1786 /* Annotate slab for lockdep -- annotate the malloc caches */
1789 /* 6) resize the head arrays to their final sizes */
1790 mutex_lock(&slab_mutex);
1791 list_for_each_entry(cachep, &slab_caches, list)
1792 if (enable_cpucache(cachep, GFP_NOWAIT))
1794 mutex_unlock(&slab_mutex);
1800 * Register a cpu startup notifier callback that initializes
1801 * cpu_cache_get for all new cpus
1803 register_cpu_notifier(&cpucache_notifier);
1807 * Register a memory hotplug callback that initializes and frees
1810 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1814 * The reap timers are started later, with a module init call: That part
1815 * of the kernel is not yet operational.
1819 static int __init cpucache_init(void)
1824 * Register the timers that return unneeded pages to the page allocator
1826 for_each_online_cpu(cpu)
1827 start_cpu_timer(cpu);
1833 __initcall(cpucache_init);
1835 static noinline void
1836 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1838 struct kmem_list3 *l3;
1840 unsigned long flags;
1844 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1846 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1847 cachep->name, cachep->size, cachep->gfporder);
1849 for_each_online_node(node) {
1850 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1851 unsigned long active_slabs = 0, num_slabs = 0;
1853 l3 = cachep->nodelists[node];
1857 spin_lock_irqsave(&l3->list_lock, flags);
1858 list_for_each_entry(slabp, &l3->slabs_full, list) {
1859 active_objs += cachep->num;
1862 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1863 active_objs += slabp->inuse;
1866 list_for_each_entry(slabp, &l3->slabs_free, list)
1869 free_objects += l3->free_objects;
1870 spin_unlock_irqrestore(&l3->list_lock, flags);
1872 num_slabs += active_slabs;
1873 num_objs = num_slabs * cachep->num;
1875 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1876 node, active_slabs, num_slabs, active_objs, num_objs,
1882 * Interface to system's page allocator. No need to hold the cache-lock.
1884 * If we requested dmaable memory, we will get it. Even if we
1885 * did not request dmaable memory, we might get it, but that
1886 * would be relatively rare and ignorable.
1888 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1896 * Nommu uses slab's for process anonymous memory allocations, and thus
1897 * requires __GFP_COMP to properly refcount higher order allocations
1899 flags |= __GFP_COMP;
1902 flags |= cachep->allocflags;
1903 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1904 flags |= __GFP_RECLAIMABLE;
1906 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1908 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1909 slab_out_of_memory(cachep, flags, nodeid);
1913 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1914 if (unlikely(page->pfmemalloc))
1915 pfmemalloc_active = true;
1917 nr_pages = (1 << cachep->gfporder);
1918 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1919 add_zone_page_state(page_zone(page),
1920 NR_SLAB_RECLAIMABLE, nr_pages);
1922 add_zone_page_state(page_zone(page),
1923 NR_SLAB_UNRECLAIMABLE, nr_pages);
1924 for (i = 0; i < nr_pages; i++) {
1925 __SetPageSlab(page + i);
1927 if (page->pfmemalloc)
1928 SetPageSlabPfmemalloc(page + i);
1931 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1932 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1935 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1937 kmemcheck_mark_unallocated_pages(page, nr_pages);
1940 return page_address(page);
1944 * Interface to system's page release.
1946 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1948 unsigned long i = (1 << cachep->gfporder);
1949 struct page *page = virt_to_page(addr);
1950 const unsigned long nr_freed = i;
1952 kmemcheck_free_shadow(page, cachep->gfporder);
1954 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1955 sub_zone_page_state(page_zone(page),
1956 NR_SLAB_RECLAIMABLE, nr_freed);
1958 sub_zone_page_state(page_zone(page),
1959 NR_SLAB_UNRECLAIMABLE, nr_freed);
1961 BUG_ON(!PageSlab(page));
1962 __ClearPageSlabPfmemalloc(page);
1963 __ClearPageSlab(page);
1966 if (current->reclaim_state)
1967 current->reclaim_state->reclaimed_slab += nr_freed;
1968 free_pages((unsigned long)addr, cachep->gfporder);
1971 static void kmem_rcu_free(struct rcu_head *head)
1973 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1974 struct kmem_cache *cachep = slab_rcu->cachep;
1976 kmem_freepages(cachep, slab_rcu->addr);
1977 if (OFF_SLAB(cachep))
1978 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1983 #ifdef CONFIG_DEBUG_PAGEALLOC
1984 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1985 unsigned long caller)
1987 int size = cachep->object_size;
1989 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1991 if (size < 5 * sizeof(unsigned long))
1994 *addr++ = 0x12345678;
1996 *addr++ = smp_processor_id();
1997 size -= 3 * sizeof(unsigned long);
1999 unsigned long *sptr = &caller;
2000 unsigned long svalue;
2002 while (!kstack_end(sptr)) {
2004 if (kernel_text_address(svalue)) {
2006 size -= sizeof(unsigned long);
2007 if (size <= sizeof(unsigned long))
2013 *addr++ = 0x87654321;
2017 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
2019 int size = cachep->object_size;
2020 addr = &((char *)addr)[obj_offset(cachep)];
2022 memset(addr, val, size);
2023 *(unsigned char *)(addr + size - 1) = POISON_END;
2026 static void dump_line(char *data, int offset, int limit)
2029 unsigned char error = 0;
2032 printk(KERN_ERR "%03x: ", offset);
2033 for (i = 0; i < limit; i++) {
2034 if (data[offset + i] != POISON_FREE) {
2035 error = data[offset + i];
2039 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2040 &data[offset], limit, 1);
2042 if (bad_count == 1) {
2043 error ^= POISON_FREE;
2044 if (!(error & (error - 1))) {
2045 printk(KERN_ERR "Single bit error detected. Probably "
2048 printk(KERN_ERR "Run memtest86+ or a similar memory "
2051 printk(KERN_ERR "Run a memory test tool.\n");
2060 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2065 if (cachep->flags & SLAB_RED_ZONE) {
2066 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2067 *dbg_redzone1(cachep, objp),
2068 *dbg_redzone2(cachep, objp));
2071 if (cachep->flags & SLAB_STORE_USER) {
2072 printk(KERN_ERR "Last user: [<%p>]",
2073 *dbg_userword(cachep, objp));
2074 print_symbol("(%s)",
2075 (unsigned long)*dbg_userword(cachep, objp));
2078 realobj = (char *)objp + obj_offset(cachep);
2079 size = cachep->object_size;
2080 for (i = 0; i < size && lines; i += 16, lines--) {
2083 if (i + limit > size)
2085 dump_line(realobj, i, limit);
2089 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2095 realobj = (char *)objp + obj_offset(cachep);
2096 size = cachep->object_size;
2098 for (i = 0; i < size; i++) {
2099 char exp = POISON_FREE;
2102 if (realobj[i] != exp) {
2108 "Slab corruption (%s): %s start=%p, len=%d\n",
2109 print_tainted(), cachep->name, realobj, size);
2110 print_objinfo(cachep, objp, 0);
2112 /* Hexdump the affected line */
2115 if (i + limit > size)
2117 dump_line(realobj, i, limit);
2120 /* Limit to 5 lines */
2126 /* Print some data about the neighboring objects, if they
2129 struct slab *slabp = virt_to_slab(objp);
2132 objnr = obj_to_index(cachep, slabp, objp);
2134 objp = index_to_obj(cachep, slabp, objnr - 1);
2135 realobj = (char *)objp + obj_offset(cachep);
2136 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2138 print_objinfo(cachep, objp, 2);
2140 if (objnr + 1 < cachep->num) {
2141 objp = index_to_obj(cachep, slabp, objnr + 1);
2142 realobj = (char *)objp + obj_offset(cachep);
2143 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2145 print_objinfo(cachep, objp, 2);
2152 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2155 for (i = 0; i < cachep->num; i++) {
2156 void *objp = index_to_obj(cachep, slabp, i);
2158 if (cachep->flags & SLAB_POISON) {
2159 #ifdef CONFIG_DEBUG_PAGEALLOC
2160 if (cachep->size % PAGE_SIZE == 0 &&
2162 kernel_map_pages(virt_to_page(objp),
2163 cachep->size / PAGE_SIZE, 1);
2165 check_poison_obj(cachep, objp);
2167 check_poison_obj(cachep, objp);
2170 if (cachep->flags & SLAB_RED_ZONE) {
2171 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2172 slab_error(cachep, "start of a freed object "
2174 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2175 slab_error(cachep, "end of a freed object "
2181 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2187 * slab_destroy - destroy and release all objects in a slab
2188 * @cachep: cache pointer being destroyed
2189 * @slabp: slab pointer being destroyed
2191 * Destroy all the objs in a slab, and release the mem back to the system.
2192 * Before calling the slab must have been unlinked from the cache. The
2193 * cache-lock is not held/needed.
2195 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2197 void *addr = slabp->s_mem - slabp->colouroff;
2199 slab_destroy_debugcheck(cachep, slabp);
2200 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2201 struct slab_rcu *slab_rcu;
2203 slab_rcu = (struct slab_rcu *)slabp;
2204 slab_rcu->cachep = cachep;
2205 slab_rcu->addr = addr;
2206 call_rcu(&slab_rcu->head, kmem_rcu_free);
2208 kmem_freepages(cachep, addr);
2209 if (OFF_SLAB(cachep))
2210 kmem_cache_free(cachep->slabp_cache, slabp);
2215 * calculate_slab_order - calculate size (page order) of slabs
2216 * @cachep: pointer to the cache that is being created
2217 * @size: size of objects to be created in this cache.
2218 * @align: required alignment for the objects.
2219 * @flags: slab allocation flags
2221 * Also calculates the number of objects per slab.
2223 * This could be made much more intelligent. For now, try to avoid using
2224 * high order pages for slabs. When the gfp() functions are more friendly
2225 * towards high-order requests, this should be changed.
2227 static size_t calculate_slab_order(struct kmem_cache *cachep,
2228 size_t size, size_t align, unsigned long flags)
2230 unsigned long offslab_limit;
2231 size_t left_over = 0;
2234 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2238 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2242 if (flags & CFLGS_OFF_SLAB) {
2244 * Max number of objs-per-slab for caches which
2245 * use off-slab slabs. Needed to avoid a possible
2246 * looping condition in cache_grow().
2248 offslab_limit = size - sizeof(struct slab);
2249 offslab_limit /= sizeof(kmem_bufctl_t);
2251 if (num > offslab_limit)
2255 /* Found something acceptable - save it away */
2257 cachep->gfporder = gfporder;
2258 left_over = remainder;
2261 * A VFS-reclaimable slab tends to have most allocations
2262 * as GFP_NOFS and we really don't want to have to be allocating
2263 * higher-order pages when we are unable to shrink dcache.
2265 if (flags & SLAB_RECLAIM_ACCOUNT)
2269 * Large number of objects is good, but very large slabs are
2270 * currently bad for the gfp()s.
2272 if (gfporder >= slab_max_order)
2276 * Acceptable internal fragmentation?
2278 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2284 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2286 if (slab_state >= FULL)
2287 return enable_cpucache(cachep, gfp);
2289 if (slab_state == DOWN) {
2291 * Note: the first kmem_cache_create must create the cache
2292 * that's used by kmalloc(24), otherwise the creation of
2293 * further caches will BUG().
2295 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2298 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2299 * the first cache, then we need to set up all its list3s,
2300 * otherwise the creation of further caches will BUG().
2302 set_up_list3s(cachep, SIZE_AC);
2303 if (INDEX_AC == INDEX_L3)
2304 slab_state = PARTIAL_L3;
2306 slab_state = PARTIAL_ARRAYCACHE;
2308 cachep->array[smp_processor_id()] =
2309 kmalloc(sizeof(struct arraycache_init), gfp);
2311 if (slab_state == PARTIAL_ARRAYCACHE) {
2312 set_up_list3s(cachep, SIZE_L3);
2313 slab_state = PARTIAL_L3;
2316 for_each_online_node(node) {
2317 cachep->nodelists[node] =
2318 kmalloc_node(sizeof(struct kmem_list3),
2320 BUG_ON(!cachep->nodelists[node]);
2321 kmem_list3_init(cachep->nodelists[node]);
2325 cachep->nodelists[numa_mem_id()]->next_reap =
2326 jiffies + REAPTIMEOUT_LIST3 +
2327 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2329 cpu_cache_get(cachep)->avail = 0;
2330 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2331 cpu_cache_get(cachep)->batchcount = 1;
2332 cpu_cache_get(cachep)->touched = 0;
2333 cachep->batchcount = 1;
2334 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2339 * __kmem_cache_create - Create a cache.
2340 * @name: A string which is used in /proc/slabinfo to identify this cache.
2341 * @size: The size of objects to be created in this cache.
2342 * @align: The required alignment for the objects.
2343 * @flags: SLAB flags
2344 * @ctor: A constructor for the objects.
2346 * Returns a ptr to the cache on success, NULL on failure.
2347 * Cannot be called within a int, but can be interrupted.
2348 * The @ctor is run when new pages are allocated by the cache.
2352 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2353 * to catch references to uninitialised memory.
2355 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2356 * for buffer overruns.
2358 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2359 * cacheline. This can be beneficial if you're counting cycles as closely
2363 __kmem_cache_create (struct kmem_cache *cachep, const char *name, size_t size, size_t align,
2364 unsigned long flags, void (*ctor)(void *))
2366 size_t left_over, slab_size, ralign;
2373 * Enable redzoning and last user accounting, except for caches with
2374 * large objects, if the increased size would increase the object size
2375 * above the next power of two: caches with object sizes just above a
2376 * power of two have a significant amount of internal fragmentation.
2378 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2379 2 * sizeof(unsigned long long)))
2380 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2381 if (!(flags & SLAB_DESTROY_BY_RCU))
2382 flags |= SLAB_POISON;
2384 if (flags & SLAB_DESTROY_BY_RCU)
2385 BUG_ON(flags & SLAB_POISON);
2388 * Always checks flags, a caller might be expecting debug support which
2391 BUG_ON(flags & ~CREATE_MASK);
2394 * Check that size is in terms of words. This is needed to avoid
2395 * unaligned accesses for some archs when redzoning is used, and makes
2396 * sure any on-slab bufctl's are also correctly aligned.
2398 if (size & (BYTES_PER_WORD - 1)) {
2399 size += (BYTES_PER_WORD - 1);
2400 size &= ~(BYTES_PER_WORD - 1);
2403 /* calculate the final buffer alignment: */
2405 /* 1) arch recommendation: can be overridden for debug */
2406 if (flags & SLAB_HWCACHE_ALIGN) {
2408 * Default alignment: as specified by the arch code. Except if
2409 * an object is really small, then squeeze multiple objects into
2412 ralign = cache_line_size();
2413 while (size <= ralign / 2)
2416 ralign = BYTES_PER_WORD;
2420 * Redzoning and user store require word alignment or possibly larger.
2421 * Note this will be overridden by architecture or caller mandated
2422 * alignment if either is greater than BYTES_PER_WORD.
2424 if (flags & SLAB_STORE_USER)
2425 ralign = BYTES_PER_WORD;
2427 if (flags & SLAB_RED_ZONE) {
2428 ralign = REDZONE_ALIGN;
2429 /* If redzoning, ensure that the second redzone is suitably
2430 * aligned, by adjusting the object size accordingly. */
2431 size += REDZONE_ALIGN - 1;
2432 size &= ~(REDZONE_ALIGN - 1);
2435 /* 2) arch mandated alignment */
2436 if (ralign < ARCH_SLAB_MINALIGN) {
2437 ralign = ARCH_SLAB_MINALIGN;
2439 /* 3) caller mandated alignment */
2440 if (ralign < align) {
2443 /* disable debug if necessary */
2444 if (ralign > __alignof__(unsigned long long))
2445 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2451 if (slab_is_available())
2456 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2457 cachep->object_size = size;
2458 cachep->align = align;
2462 * Both debugging options require word-alignment which is calculated
2465 if (flags & SLAB_RED_ZONE) {
2466 /* add space for red zone words */
2467 cachep->obj_offset += sizeof(unsigned long long);
2468 size += 2 * sizeof(unsigned long long);
2470 if (flags & SLAB_STORE_USER) {
2471 /* user store requires one word storage behind the end of
2472 * the real object. But if the second red zone needs to be
2473 * aligned to 64 bits, we must allow that much space.
2475 if (flags & SLAB_RED_ZONE)
2476 size += REDZONE_ALIGN;
2478 size += BYTES_PER_WORD;
2480 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2481 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2482 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2483 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2490 * Determine if the slab management is 'on' or 'off' slab.
2491 * (bootstrapping cannot cope with offslab caches so don't do
2492 * it too early on. Always use on-slab management when
2493 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2495 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2496 !(flags & SLAB_NOLEAKTRACE))
2498 * Size is large, assume best to place the slab management obj
2499 * off-slab (should allow better packing of objs).
2501 flags |= CFLGS_OFF_SLAB;
2503 size = ALIGN(size, align);
2505 left_over = calculate_slab_order(cachep, size, align, flags);
2509 "kmem_cache_create: couldn't create cache %s.\n", name);
2512 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2513 + sizeof(struct slab), align);
2516 * If the slab has been placed off-slab, and we have enough space then
2517 * move it on-slab. This is at the expense of any extra colouring.
2519 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2520 flags &= ~CFLGS_OFF_SLAB;
2521 left_over -= slab_size;
2524 if (flags & CFLGS_OFF_SLAB) {
2525 /* really off slab. No need for manual alignment */
2527 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2529 #ifdef CONFIG_PAGE_POISONING
2530 /* If we're going to use the generic kernel_map_pages()
2531 * poisoning, then it's going to smash the contents of
2532 * the redzone and userword anyhow, so switch them off.
2534 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2535 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2539 cachep->colour_off = cache_line_size();
2540 /* Offset must be a multiple of the alignment. */
2541 if (cachep->colour_off < align)
2542 cachep->colour_off = align;
2543 cachep->colour = left_over / cachep->colour_off;
2544 cachep->slab_size = slab_size;
2545 cachep->flags = flags;
2546 cachep->allocflags = 0;
2547 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2548 cachep->allocflags |= GFP_DMA;
2549 cachep->size = size;
2550 cachep->reciprocal_buffer_size = reciprocal_value(size);
2552 if (flags & CFLGS_OFF_SLAB) {
2553 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2555 * This is a possibility for one of the malloc_sizes caches.
2556 * But since we go off slab only for object size greater than
2557 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2558 * this should not happen at all.
2559 * But leave a BUG_ON for some lucky dude.
2561 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2563 cachep->ctor = ctor;
2564 cachep->name = name;
2565 cachep->refcount = 1;
2567 err = setup_cpu_cache(cachep, gfp);
2569 __kmem_cache_shutdown(cachep);
2573 if (flags & SLAB_DEBUG_OBJECTS) {
2575 * Would deadlock through slab_destroy()->call_rcu()->
2576 * debug_object_activate()->kmem_cache_alloc().
2578 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2580 slab_set_debugobj_lock_classes(cachep);
2587 static void check_irq_off(void)
2589 BUG_ON(!irqs_disabled());
2592 static void check_irq_on(void)
2594 BUG_ON(irqs_disabled());
2597 static void check_spinlock_acquired(struct kmem_cache *cachep)
2601 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2605 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2609 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2614 #define check_irq_off() do { } while(0)
2615 #define check_irq_on() do { } while(0)
2616 #define check_spinlock_acquired(x) do { } while(0)
2617 #define check_spinlock_acquired_node(x, y) do { } while(0)
2620 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2621 struct array_cache *ac,
2622 int force, int node);
2624 static void do_drain(void *arg)
2626 struct kmem_cache *cachep = arg;
2627 struct array_cache *ac;
2628 int node = numa_mem_id();
2631 ac = cpu_cache_get(cachep);
2632 spin_lock(&cachep->nodelists[node]->list_lock);
2633 free_block(cachep, ac->entry, ac->avail, node);
2634 spin_unlock(&cachep->nodelists[node]->list_lock);
2638 static void drain_cpu_caches(struct kmem_cache *cachep)
2640 struct kmem_list3 *l3;
2643 on_each_cpu(do_drain, cachep, 1);
2645 for_each_online_node(node) {
2646 l3 = cachep->nodelists[node];
2647 if (l3 && l3->alien)
2648 drain_alien_cache(cachep, l3->alien);
2651 for_each_online_node(node) {
2652 l3 = cachep->nodelists[node];
2654 drain_array(cachep, l3, l3->shared, 1, node);
2659 * Remove slabs from the list of free slabs.
2660 * Specify the number of slabs to drain in tofree.
2662 * Returns the actual number of slabs released.
2664 static int drain_freelist(struct kmem_cache *cache,
2665 struct kmem_list3 *l3, int tofree)
2667 struct list_head *p;
2672 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2674 spin_lock_irq(&l3->list_lock);
2675 p = l3->slabs_free.prev;
2676 if (p == &l3->slabs_free) {
2677 spin_unlock_irq(&l3->list_lock);
2681 slabp = list_entry(p, struct slab, list);
2683 BUG_ON(slabp->inuse);
2685 list_del(&slabp->list);
2687 * Safe to drop the lock. The slab is no longer linked
2690 l3->free_objects -= cache->num;
2691 spin_unlock_irq(&l3->list_lock);
2692 slab_destroy(cache, slabp);
2699 /* Called with slab_mutex held to protect against cpu hotplug */
2700 static int __cache_shrink(struct kmem_cache *cachep)
2703 struct kmem_list3 *l3;
2705 drain_cpu_caches(cachep);
2708 for_each_online_node(i) {
2709 l3 = cachep->nodelists[i];
2713 drain_freelist(cachep, l3, l3->free_objects);
2715 ret += !list_empty(&l3->slabs_full) ||
2716 !list_empty(&l3->slabs_partial);
2718 return (ret ? 1 : 0);
2722 * kmem_cache_shrink - Shrink a cache.
2723 * @cachep: The cache to shrink.
2725 * Releases as many slabs as possible for a cache.
2726 * To help debugging, a zero exit status indicates all slabs were released.
2728 int kmem_cache_shrink(struct kmem_cache *cachep)
2731 BUG_ON(!cachep || in_interrupt());
2734 mutex_lock(&slab_mutex);
2735 ret = __cache_shrink(cachep);
2736 mutex_unlock(&slab_mutex);
2740 EXPORT_SYMBOL(kmem_cache_shrink);
2742 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2745 struct kmem_list3 *l3;
2746 int rc = __cache_shrink(cachep);
2751 for_each_online_cpu(i)
2752 kfree(cachep->array[i]);
2754 /* NUMA: free the list3 structures */
2755 for_each_online_node(i) {
2756 l3 = cachep->nodelists[i];
2759 free_alien_cache(l3->alien);
2767 * Get the memory for a slab management obj.
2768 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2769 * always come from malloc_sizes caches. The slab descriptor cannot
2770 * come from the same cache which is getting created because,
2771 * when we are searching for an appropriate cache for these
2772 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2773 * If we are creating a malloc_sizes cache here it would not be visible to
2774 * kmem_find_general_cachep till the initialization is complete.
2775 * Hence we cannot have slabp_cache same as the original cache.
2777 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2778 int colour_off, gfp_t local_flags,
2783 if (OFF_SLAB(cachep)) {
2784 /* Slab management obj is off-slab. */
2785 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2786 local_flags, nodeid);
2788 * If the first object in the slab is leaked (it's allocated
2789 * but no one has a reference to it), we want to make sure
2790 * kmemleak does not treat the ->s_mem pointer as a reference
2791 * to the object. Otherwise we will not report the leak.
2793 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2798 slabp = objp + colour_off;
2799 colour_off += cachep->slab_size;
2802 slabp->colouroff = colour_off;
2803 slabp->s_mem = objp + colour_off;
2804 slabp->nodeid = nodeid;
2809 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2811 return (kmem_bufctl_t *) (slabp + 1);
2814 static void cache_init_objs(struct kmem_cache *cachep,
2819 for (i = 0; i < cachep->num; i++) {
2820 void *objp = index_to_obj(cachep, slabp, i);
2822 /* need to poison the objs? */
2823 if (cachep->flags & SLAB_POISON)
2824 poison_obj(cachep, objp, POISON_FREE);
2825 if (cachep->flags & SLAB_STORE_USER)
2826 *dbg_userword(cachep, objp) = NULL;
2828 if (cachep->flags & SLAB_RED_ZONE) {
2829 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2830 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2833 * Constructors are not allowed to allocate memory from the same
2834 * cache which they are a constructor for. Otherwise, deadlock.
2835 * They must also be threaded.
2837 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2838 cachep->ctor(objp + obj_offset(cachep));
2840 if (cachep->flags & SLAB_RED_ZONE) {
2841 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2842 slab_error(cachep, "constructor overwrote the"
2843 " end of an object");
2844 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2845 slab_error(cachep, "constructor overwrote the"
2846 " start of an object");
2848 if ((cachep->size % PAGE_SIZE) == 0 &&
2849 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2850 kernel_map_pages(virt_to_page(objp),
2851 cachep->size / PAGE_SIZE, 0);
2856 slab_bufctl(slabp)[i] = i + 1;
2858 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2861 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2863 if (CONFIG_ZONE_DMA_FLAG) {
2864 if (flags & GFP_DMA)
2865 BUG_ON(!(cachep->allocflags & GFP_DMA));
2867 BUG_ON(cachep->allocflags & GFP_DMA);
2871 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2874 void *objp = index_to_obj(cachep, slabp, slabp->free);
2878 next = slab_bufctl(slabp)[slabp->free];
2880 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2881 WARN_ON(slabp->nodeid != nodeid);
2888 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2889 void *objp, int nodeid)
2891 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2894 /* Verify that the slab belongs to the intended node */
2895 WARN_ON(slabp->nodeid != nodeid);
2897 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2898 printk(KERN_ERR "slab: double free detected in cache "
2899 "'%s', objp %p\n", cachep->name, objp);
2903 slab_bufctl(slabp)[objnr] = slabp->free;
2904 slabp->free = objnr;
2909 * Map pages beginning at addr to the given cache and slab. This is required
2910 * for the slab allocator to be able to lookup the cache and slab of a
2911 * virtual address for kfree, ksize, and slab debugging.
2913 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2919 page = virt_to_page(addr);
2922 if (likely(!PageCompound(page)))
2923 nr_pages <<= cache->gfporder;
2926 page->slab_cache = cache;
2927 page->slab_page = slab;
2929 } while (--nr_pages);
2933 * Grow (by 1) the number of slabs within a cache. This is called by
2934 * kmem_cache_alloc() when there are no active objs left in a cache.
2936 static int cache_grow(struct kmem_cache *cachep,
2937 gfp_t flags, int nodeid, void *objp)
2942 struct kmem_list3 *l3;
2945 * Be lazy and only check for valid flags here, keeping it out of the
2946 * critical path in kmem_cache_alloc().
2948 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2949 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2951 /* Take the l3 list lock to change the colour_next on this node */
2953 l3 = cachep->nodelists[nodeid];
2954 spin_lock(&l3->list_lock);
2956 /* Get colour for the slab, and cal the next value. */
2957 offset = l3->colour_next;
2959 if (l3->colour_next >= cachep->colour)
2960 l3->colour_next = 0;
2961 spin_unlock(&l3->list_lock);
2963 offset *= cachep->colour_off;
2965 if (local_flags & __GFP_WAIT)
2969 * The test for missing atomic flag is performed here, rather than
2970 * the more obvious place, simply to reduce the critical path length
2971 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2972 * will eventually be caught here (where it matters).
2974 kmem_flagcheck(cachep, flags);
2977 * Get mem for the objs. Attempt to allocate a physical page from
2981 objp = kmem_getpages(cachep, local_flags, nodeid);
2985 /* Get slab management. */
2986 slabp = alloc_slabmgmt(cachep, objp, offset,
2987 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2991 slab_map_pages(cachep, slabp, objp);
2993 cache_init_objs(cachep, slabp);
2995 if (local_flags & __GFP_WAIT)
2996 local_irq_disable();
2998 spin_lock(&l3->list_lock);
3000 /* Make slab active. */
3001 list_add_tail(&slabp->list, &(l3->slabs_free));
3002 STATS_INC_GROWN(cachep);
3003 l3->free_objects += cachep->num;
3004 spin_unlock(&l3->list_lock);
3007 kmem_freepages(cachep, objp);
3009 if (local_flags & __GFP_WAIT)
3010 local_irq_disable();
3017 * Perform extra freeing checks:
3018 * - detect bad pointers.
3019 * - POISON/RED_ZONE checking
3021 static void kfree_debugcheck(const void *objp)
3023 if (!virt_addr_valid(objp)) {
3024 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3025 (unsigned long)objp);
3030 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3032 unsigned long long redzone1, redzone2;
3034 redzone1 = *dbg_redzone1(cache, obj);
3035 redzone2 = *dbg_redzone2(cache, obj);
3040 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3043 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3044 slab_error(cache, "double free detected");
3046 slab_error(cache, "memory outside object was overwritten");
3048 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3049 obj, redzone1, redzone2);
3052 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3059 BUG_ON(virt_to_cache(objp) != cachep);
3061 objp -= obj_offset(cachep);
3062 kfree_debugcheck(objp);
3063 page = virt_to_head_page(objp);
3065 slabp = page->slab_page;
3067 if (cachep->flags & SLAB_RED_ZONE) {
3068 verify_redzone_free(cachep, objp);
3069 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3070 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3072 if (cachep->flags & SLAB_STORE_USER)
3073 *dbg_userword(cachep, objp) = caller;
3075 objnr = obj_to_index(cachep, slabp, objp);
3077 BUG_ON(objnr >= cachep->num);
3078 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3080 #ifdef CONFIG_DEBUG_SLAB_LEAK
3081 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3083 if (cachep->flags & SLAB_POISON) {
3084 #ifdef CONFIG_DEBUG_PAGEALLOC
3085 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3086 store_stackinfo(cachep, objp, (unsigned long)caller);
3087 kernel_map_pages(virt_to_page(objp),
3088 cachep->size / PAGE_SIZE, 0);
3090 poison_obj(cachep, objp, POISON_FREE);
3093 poison_obj(cachep, objp, POISON_FREE);
3099 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3104 /* Check slab's freelist to see if this obj is there. */
3105 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3107 if (entries > cachep->num || i >= cachep->num)
3110 if (entries != cachep->num - slabp->inuse) {
3112 printk(KERN_ERR "slab: Internal list corruption detected in "
3113 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3114 cachep->name, cachep->num, slabp, slabp->inuse,
3116 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3117 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3123 #define kfree_debugcheck(x) do { } while(0)
3124 #define cache_free_debugcheck(x,objp,z) (objp)
3125 #define check_slabp(x,y) do { } while(0)
3128 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3132 struct kmem_list3 *l3;
3133 struct array_cache *ac;
3137 node = numa_mem_id();
3138 if (unlikely(force_refill))
3141 ac = cpu_cache_get(cachep);
3142 batchcount = ac->batchcount;
3143 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3145 * If there was little recent activity on this cache, then
3146 * perform only a partial refill. Otherwise we could generate
3149 batchcount = BATCHREFILL_LIMIT;
3151 l3 = cachep->nodelists[node];
3153 BUG_ON(ac->avail > 0 || !l3);
3154 spin_lock(&l3->list_lock);
3156 /* See if we can refill from the shared array */
3157 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3158 l3->shared->touched = 1;
3162 while (batchcount > 0) {
3163 struct list_head *entry;
3165 /* Get slab alloc is to come from. */
3166 entry = l3->slabs_partial.next;
3167 if (entry == &l3->slabs_partial) {
3168 l3->free_touched = 1;
3169 entry = l3->slabs_free.next;
3170 if (entry == &l3->slabs_free)
3174 slabp = list_entry(entry, struct slab, list);
3175 check_slabp(cachep, slabp);
3176 check_spinlock_acquired(cachep);
3179 * The slab was either on partial or free list so
3180 * there must be at least one object available for
3183 BUG_ON(slabp->inuse >= cachep->num);
3185 while (slabp->inuse < cachep->num && batchcount--) {
3186 STATS_INC_ALLOCED(cachep);
3187 STATS_INC_ACTIVE(cachep);
3188 STATS_SET_HIGH(cachep);
3190 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3193 check_slabp(cachep, slabp);
3195 /* move slabp to correct slabp list: */
3196 list_del(&slabp->list);
3197 if (slabp->free == BUFCTL_END)
3198 list_add(&slabp->list, &l3->slabs_full);
3200 list_add(&slabp->list, &l3->slabs_partial);
3204 l3->free_objects -= ac->avail;
3206 spin_unlock(&l3->list_lock);
3208 if (unlikely(!ac->avail)) {
3211 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3213 /* cache_grow can reenable interrupts, then ac could change. */
3214 ac = cpu_cache_get(cachep);
3216 /* no objects in sight? abort */
3217 if (!x && (ac->avail == 0 || force_refill))
3220 if (!ac->avail) /* objects refilled by interrupt? */
3225 return ac_get_obj(cachep, ac, flags, force_refill);
3228 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3231 might_sleep_if(flags & __GFP_WAIT);
3233 kmem_flagcheck(cachep, flags);
3238 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3239 gfp_t flags, void *objp, void *caller)
3243 if (cachep->flags & SLAB_POISON) {
3244 #ifdef CONFIG_DEBUG_PAGEALLOC
3245 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3246 kernel_map_pages(virt_to_page(objp),
3247 cachep->size / PAGE_SIZE, 1);
3249 check_poison_obj(cachep, objp);
3251 check_poison_obj(cachep, objp);
3253 poison_obj(cachep, objp, POISON_INUSE);
3255 if (cachep->flags & SLAB_STORE_USER)
3256 *dbg_userword(cachep, objp) = caller;
3258 if (cachep->flags & SLAB_RED_ZONE) {
3259 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3260 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3261 slab_error(cachep, "double free, or memory outside"
3262 " object was overwritten");
3264 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3265 objp, *dbg_redzone1(cachep, objp),
3266 *dbg_redzone2(cachep, objp));
3268 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3269 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3271 #ifdef CONFIG_DEBUG_SLAB_LEAK
3276 slabp = virt_to_head_page(objp)->slab_page;
3277 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3278 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3281 objp += obj_offset(cachep);
3282 if (cachep->ctor && cachep->flags & SLAB_POISON)
3284 if (ARCH_SLAB_MINALIGN &&
3285 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3286 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3287 objp, (int)ARCH_SLAB_MINALIGN);
3292 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3295 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3297 if (cachep == kmem_cache)
3300 return should_failslab(cachep->object_size, flags, cachep->flags);
3303 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3306 struct array_cache *ac;
3307 bool force_refill = false;
3311 ac = cpu_cache_get(cachep);
3312 if (likely(ac->avail)) {
3314 objp = ac_get_obj(cachep, ac, flags, false);
3317 * Allow for the possibility all avail objects are not allowed
3318 * by the current flags
3321 STATS_INC_ALLOCHIT(cachep);
3324 force_refill = true;
3327 STATS_INC_ALLOCMISS(cachep);
3328 objp = cache_alloc_refill(cachep, flags, force_refill);
3330 * the 'ac' may be updated by cache_alloc_refill(),
3331 * and kmemleak_erase() requires its correct value.
3333 ac = cpu_cache_get(cachep);
3337 * To avoid a false negative, if an object that is in one of the
3338 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3339 * treat the array pointers as a reference to the object.
3342 kmemleak_erase(&ac->entry[ac->avail]);
3348 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3350 * If we are in_interrupt, then process context, including cpusets and
3351 * mempolicy, may not apply and should not be used for allocation policy.
3353 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3355 int nid_alloc, nid_here;
3357 if (in_interrupt() || (flags & __GFP_THISNODE))
3359 nid_alloc = nid_here = numa_mem_id();
3360 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3361 nid_alloc = cpuset_slab_spread_node();
3362 else if (current->mempolicy)
3363 nid_alloc = slab_node();
3364 if (nid_alloc != nid_here)
3365 return ____cache_alloc_node(cachep, flags, nid_alloc);
3370 * Fallback function if there was no memory available and no objects on a
3371 * certain node and fall back is permitted. First we scan all the
3372 * available nodelists for available objects. If that fails then we
3373 * perform an allocation without specifying a node. This allows the page
3374 * allocator to do its reclaim / fallback magic. We then insert the
3375 * slab into the proper nodelist and then allocate from it.
3377 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3379 struct zonelist *zonelist;
3383 enum zone_type high_zoneidx = gfp_zone(flags);
3386 unsigned int cpuset_mems_cookie;
3388 if (flags & __GFP_THISNODE)
3391 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3394 cpuset_mems_cookie = get_mems_allowed();
3395 zonelist = node_zonelist(slab_node(), flags);
3399 * Look through allowed nodes for objects available
3400 * from existing per node queues.
3402 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3403 nid = zone_to_nid(zone);
3405 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3406 cache->nodelists[nid] &&
3407 cache->nodelists[nid]->free_objects) {
3408 obj = ____cache_alloc_node(cache,
3409 flags | GFP_THISNODE, nid);
3417 * This allocation will be performed within the constraints
3418 * of the current cpuset / memory policy requirements.
3419 * We may trigger various forms of reclaim on the allowed
3420 * set and go into memory reserves if necessary.
3422 if (local_flags & __GFP_WAIT)
3424 kmem_flagcheck(cache, flags);
3425 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3426 if (local_flags & __GFP_WAIT)
3427 local_irq_disable();
3430 * Insert into the appropriate per node queues
3432 nid = page_to_nid(virt_to_page(obj));
3433 if (cache_grow(cache, flags, nid, obj)) {
3434 obj = ____cache_alloc_node(cache,
3435 flags | GFP_THISNODE, nid);
3438 * Another processor may allocate the
3439 * objects in the slab since we are
3440 * not holding any locks.
3444 /* cache_grow already freed obj */
3450 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3456 * A interface to enable slab creation on nodeid
3458 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3461 struct list_head *entry;
3463 struct kmem_list3 *l3;
3467 l3 = cachep->nodelists[nodeid];
3472 spin_lock(&l3->list_lock);
3473 entry = l3->slabs_partial.next;
3474 if (entry == &l3->slabs_partial) {
3475 l3->free_touched = 1;
3476 entry = l3->slabs_free.next;
3477 if (entry == &l3->slabs_free)
3481 slabp = list_entry(entry, struct slab, list);
3482 check_spinlock_acquired_node(cachep, nodeid);
3483 check_slabp(cachep, slabp);
3485 STATS_INC_NODEALLOCS(cachep);
3486 STATS_INC_ACTIVE(cachep);
3487 STATS_SET_HIGH(cachep);
3489 BUG_ON(slabp->inuse == cachep->num);
3491 obj = slab_get_obj(cachep, slabp, nodeid);
3492 check_slabp(cachep, slabp);
3494 /* move slabp to correct slabp list: */
3495 list_del(&slabp->list);
3497 if (slabp->free == BUFCTL_END)
3498 list_add(&slabp->list, &l3->slabs_full);
3500 list_add(&slabp->list, &l3->slabs_partial);
3502 spin_unlock(&l3->list_lock);
3506 spin_unlock(&l3->list_lock);
3507 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3511 return fallback_alloc(cachep, flags);
3518 * kmem_cache_alloc_node - Allocate an object on the specified node
3519 * @cachep: The cache to allocate from.
3520 * @flags: See kmalloc().
3521 * @nodeid: node number of the target node.
3522 * @caller: return address of caller, used for debug information
3524 * Identical to kmem_cache_alloc but it will allocate memory on the given
3525 * node, which can improve the performance for cpu bound structures.
3527 * Fallback to other node is possible if __GFP_THISNODE is not set.
3529 static __always_inline void *
3530 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3533 unsigned long save_flags;
3535 int slab_node = numa_mem_id();
3537 flags &= gfp_allowed_mask;
3539 lockdep_trace_alloc(flags);
3541 if (slab_should_failslab(cachep, flags))
3544 cache_alloc_debugcheck_before(cachep, flags);
3545 local_irq_save(save_flags);
3547 if (nodeid == NUMA_NO_NODE)
3550 if (unlikely(!cachep->nodelists[nodeid])) {
3551 /* Node not bootstrapped yet */
3552 ptr = fallback_alloc(cachep, flags);
3556 if (nodeid == slab_node) {
3558 * Use the locally cached objects if possible.
3559 * However ____cache_alloc does not allow fallback
3560 * to other nodes. It may fail while we still have
3561 * objects on other nodes available.
3563 ptr = ____cache_alloc(cachep, flags);
3567 /* ___cache_alloc_node can fall back to other nodes */
3568 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3570 local_irq_restore(save_flags);
3571 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3572 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3576 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3578 if (unlikely((flags & __GFP_ZERO) && ptr))
3579 memset(ptr, 0, cachep->object_size);
3584 static __always_inline void *
3585 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3589 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3590 objp = alternate_node_alloc(cache, flags);
3594 objp = ____cache_alloc(cache, flags);
3597 * We may just have run out of memory on the local node.
3598 * ____cache_alloc_node() knows how to locate memory on other nodes
3601 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3608 static __always_inline void *
3609 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3611 return ____cache_alloc(cachep, flags);
3614 #endif /* CONFIG_NUMA */
3616 static __always_inline void *
3617 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3619 unsigned long save_flags;
3622 flags &= gfp_allowed_mask;
3624 lockdep_trace_alloc(flags);
3626 if (slab_should_failslab(cachep, flags))
3629 cache_alloc_debugcheck_before(cachep, flags);
3630 local_irq_save(save_flags);
3631 objp = __do_cache_alloc(cachep, flags);
3632 local_irq_restore(save_flags);
3633 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3634 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3639 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3641 if (unlikely((flags & __GFP_ZERO) && objp))
3642 memset(objp, 0, cachep->object_size);
3648 * Caller needs to acquire correct kmem_list's list_lock
3650 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3654 struct kmem_list3 *l3;
3656 for (i = 0; i < nr_objects; i++) {
3660 clear_obj_pfmemalloc(&objpp[i]);
3663 slabp = virt_to_slab(objp);
3664 l3 = cachep->nodelists[node];
3665 list_del(&slabp->list);
3666 check_spinlock_acquired_node(cachep, node);
3667 check_slabp(cachep, slabp);
3668 slab_put_obj(cachep, slabp, objp, node);
3669 STATS_DEC_ACTIVE(cachep);
3671 check_slabp(cachep, slabp);
3673 /* fixup slab chains */
3674 if (slabp->inuse == 0) {
3675 if (l3->free_objects > l3->free_limit) {
3676 l3->free_objects -= cachep->num;
3677 /* No need to drop any previously held
3678 * lock here, even if we have a off-slab slab
3679 * descriptor it is guaranteed to come from
3680 * a different cache, refer to comments before
3683 slab_destroy(cachep, slabp);
3685 list_add(&slabp->list, &l3->slabs_free);
3688 /* Unconditionally move a slab to the end of the
3689 * partial list on free - maximum time for the
3690 * other objects to be freed, too.
3692 list_add_tail(&slabp->list, &l3->slabs_partial);
3697 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3700 struct kmem_list3 *l3;
3701 int node = numa_mem_id();
3703 batchcount = ac->batchcount;
3705 BUG_ON(!batchcount || batchcount > ac->avail);
3708 l3 = cachep->nodelists[node];
3709 spin_lock(&l3->list_lock);
3711 struct array_cache *shared_array = l3->shared;
3712 int max = shared_array->limit - shared_array->avail;
3714 if (batchcount > max)
3716 memcpy(&(shared_array->entry[shared_array->avail]),
3717 ac->entry, sizeof(void *) * batchcount);
3718 shared_array->avail += batchcount;
3723 free_block(cachep, ac->entry, batchcount, node);
3728 struct list_head *p;
3730 p = l3->slabs_free.next;
3731 while (p != &(l3->slabs_free)) {
3734 slabp = list_entry(p, struct slab, list);
3735 BUG_ON(slabp->inuse);
3740 STATS_SET_FREEABLE(cachep, i);
3743 spin_unlock(&l3->list_lock);
3744 ac->avail -= batchcount;
3745 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3749 * Release an obj back to its cache. If the obj has a constructed state, it must
3750 * be in this state _before_ it is released. Called with disabled ints.
3752 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3755 struct array_cache *ac = cpu_cache_get(cachep);
3758 kmemleak_free_recursive(objp, cachep->flags);
3759 objp = cache_free_debugcheck(cachep, objp, caller);
3761 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3764 * Skip calling cache_free_alien() when the platform is not numa.
3765 * This will avoid cache misses that happen while accessing slabp (which
3766 * is per page memory reference) to get nodeid. Instead use a global
3767 * variable to skip the call, which is mostly likely to be present in
3770 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3773 if (likely(ac->avail < ac->limit)) {
3774 STATS_INC_FREEHIT(cachep);
3776 STATS_INC_FREEMISS(cachep);
3777 cache_flusharray(cachep, ac);
3780 ac_put_obj(cachep, ac, objp);
3784 * kmem_cache_alloc - Allocate an object
3785 * @cachep: The cache to allocate from.
3786 * @flags: See kmalloc().
3788 * Allocate an object from this cache. The flags are only relevant
3789 * if the cache has no available objects.
3791 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3793 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3795 trace_kmem_cache_alloc(_RET_IP_, ret,
3796 cachep->object_size, cachep->size, flags);
3800 EXPORT_SYMBOL(kmem_cache_alloc);
3802 #ifdef CONFIG_TRACING
3804 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3808 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3810 trace_kmalloc(_RET_IP_, ret,
3811 size, slab_buffer_size(cachep), flags);
3814 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3818 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3820 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3821 __builtin_return_address(0));
3823 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3824 cachep->object_size, cachep->size,
3829 EXPORT_SYMBOL(kmem_cache_alloc_node);
3831 #ifdef CONFIG_TRACING
3832 void *kmem_cache_alloc_node_trace(size_t size,
3833 struct kmem_cache *cachep,
3839 ret = __cache_alloc_node(cachep, flags, nodeid,
3840 __builtin_return_address(0));
3841 trace_kmalloc_node(_RET_IP_, ret,
3842 size, slab_buffer_size(cachep),
3846 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3849 static __always_inline void *
3850 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3852 struct kmem_cache *cachep;
3854 cachep = kmem_find_general_cachep(size, flags);
3855 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3857 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3860 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3861 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3863 return __do_kmalloc_node(size, flags, node,
3864 __builtin_return_address(0));
3866 EXPORT_SYMBOL(__kmalloc_node);
3868 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3869 int node, unsigned long caller)
3871 return __do_kmalloc_node(size, flags, node, (void *)caller);
3873 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3875 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3877 return __do_kmalloc_node(size, flags, node, NULL);
3879 EXPORT_SYMBOL(__kmalloc_node);
3880 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3881 #endif /* CONFIG_NUMA */
3884 * __do_kmalloc - allocate memory
3885 * @size: how many bytes of memory are required.
3886 * @flags: the type of memory to allocate (see kmalloc).
3887 * @caller: function caller for debug tracking of the caller
3889 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3892 struct kmem_cache *cachep;
3895 /* If you want to save a few bytes .text space: replace
3897 * Then kmalloc uses the uninlined functions instead of the inline
3900 cachep = __find_general_cachep(size, flags);
3901 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3903 ret = __cache_alloc(cachep, flags, caller);
3905 trace_kmalloc((unsigned long) caller, ret,
3906 size, cachep->size, flags);
3912 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3913 void *__kmalloc(size_t size, gfp_t flags)
3915 return __do_kmalloc(size, flags, __builtin_return_address(0));
3917 EXPORT_SYMBOL(__kmalloc);
3919 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3921 return __do_kmalloc(size, flags, (void *)caller);
3923 EXPORT_SYMBOL(__kmalloc_track_caller);
3926 void *__kmalloc(size_t size, gfp_t flags)
3928 return __do_kmalloc(size, flags, NULL);
3930 EXPORT_SYMBOL(__kmalloc);
3934 * kmem_cache_free - Deallocate an object
3935 * @cachep: The cache the allocation was from.
3936 * @objp: The previously allocated object.
3938 * Free an object which was previously allocated from this
3941 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3943 unsigned long flags;
3945 local_irq_save(flags);
3946 debug_check_no_locks_freed(objp, cachep->object_size);
3947 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3948 debug_check_no_obj_freed(objp, cachep->object_size);
3949 __cache_free(cachep, objp, __builtin_return_address(0));
3950 local_irq_restore(flags);
3952 trace_kmem_cache_free(_RET_IP_, objp);
3954 EXPORT_SYMBOL(kmem_cache_free);
3957 * kfree - free previously allocated memory
3958 * @objp: pointer returned by kmalloc.
3960 * If @objp is NULL, no operation is performed.
3962 * Don't free memory not originally allocated by kmalloc()
3963 * or you will run into trouble.
3965 void kfree(const void *objp)
3967 struct kmem_cache *c;
3968 unsigned long flags;
3970 trace_kfree(_RET_IP_, objp);
3972 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3974 local_irq_save(flags);
3975 kfree_debugcheck(objp);
3976 c = virt_to_cache(objp);
3977 debug_check_no_locks_freed(objp, c->object_size);
3979 debug_check_no_obj_freed(objp, c->object_size);
3980 __cache_free(c, (void *)objp, __builtin_return_address(0));
3981 local_irq_restore(flags);
3983 EXPORT_SYMBOL(kfree);
3985 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3987 return cachep->object_size;
3989 EXPORT_SYMBOL(kmem_cache_size);
3992 * This initializes kmem_list3 or resizes various caches for all nodes.
3994 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3997 struct kmem_list3 *l3;
3998 struct array_cache *new_shared;
3999 struct array_cache **new_alien = NULL;
4001 for_each_online_node(node) {
4003 if (use_alien_caches) {
4004 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
4010 if (cachep->shared) {
4011 new_shared = alloc_arraycache(node,
4012 cachep->shared*cachep->batchcount,
4015 free_alien_cache(new_alien);
4020 l3 = cachep->nodelists[node];
4022 struct array_cache *shared = l3->shared;
4024 spin_lock_irq(&l3->list_lock);
4027 free_block(cachep, shared->entry,
4028 shared->avail, node);
4030 l3->shared = new_shared;
4032 l3->alien = new_alien;
4035 l3->free_limit = (1 + nr_cpus_node(node)) *
4036 cachep->batchcount + cachep->num;
4037 spin_unlock_irq(&l3->list_lock);
4039 free_alien_cache(new_alien);
4042 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4044 free_alien_cache(new_alien);
4049 kmem_list3_init(l3);
4050 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4051 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4052 l3->shared = new_shared;
4053 l3->alien = new_alien;
4054 l3->free_limit = (1 + nr_cpus_node(node)) *
4055 cachep->batchcount + cachep->num;
4056 cachep->nodelists[node] = l3;
4061 if (!cachep->list.next) {
4062 /* Cache is not active yet. Roll back what we did */
4065 if (cachep->nodelists[node]) {
4066 l3 = cachep->nodelists[node];
4069 free_alien_cache(l3->alien);
4071 cachep->nodelists[node] = NULL;
4079 struct ccupdate_struct {
4080 struct kmem_cache *cachep;
4081 struct array_cache *new[0];
4084 static void do_ccupdate_local(void *info)
4086 struct ccupdate_struct *new = info;
4087 struct array_cache *old;
4090 old = cpu_cache_get(new->cachep);
4092 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4093 new->new[smp_processor_id()] = old;
4096 /* Always called with the slab_mutex held */
4097 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4098 int batchcount, int shared, gfp_t gfp)
4100 struct ccupdate_struct *new;
4103 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4108 for_each_online_cpu(i) {
4109 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4112 for (i--; i >= 0; i--)
4118 new->cachep = cachep;
4120 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4123 cachep->batchcount = batchcount;
4124 cachep->limit = limit;
4125 cachep->shared = shared;
4127 for_each_online_cpu(i) {
4128 struct array_cache *ccold = new->new[i];
4131 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4132 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4133 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4137 return alloc_kmemlist(cachep, gfp);
4140 /* Called with slab_mutex held always */
4141 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4147 * The head array serves three purposes:
4148 * - create a LIFO ordering, i.e. return objects that are cache-warm
4149 * - reduce the number of spinlock operations.
4150 * - reduce the number of linked list operations on the slab and
4151 * bufctl chains: array operations are cheaper.
4152 * The numbers are guessed, we should auto-tune as described by
4155 if (cachep->size > 131072)
4157 else if (cachep->size > PAGE_SIZE)
4159 else if (cachep->size > 1024)
4161 else if (cachep->size > 256)
4167 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4168 * allocation behaviour: Most allocs on one cpu, most free operations
4169 * on another cpu. For these cases, an efficient object passing between
4170 * cpus is necessary. This is provided by a shared array. The array
4171 * replaces Bonwick's magazine layer.
4172 * On uniprocessor, it's functionally equivalent (but less efficient)
4173 * to a larger limit. Thus disabled by default.
4176 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4181 * With debugging enabled, large batchcount lead to excessively long
4182 * periods with disabled local interrupts. Limit the batchcount
4187 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4189 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4190 cachep->name, -err);
4195 * Drain an array if it contains any elements taking the l3 lock only if
4196 * necessary. Note that the l3 listlock also protects the array_cache
4197 * if drain_array() is used on the shared array.
4199 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4200 struct array_cache *ac, int force, int node)
4204 if (!ac || !ac->avail)
4206 if (ac->touched && !force) {
4209 spin_lock_irq(&l3->list_lock);
4211 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4212 if (tofree > ac->avail)
4213 tofree = (ac->avail + 1) / 2;
4214 free_block(cachep, ac->entry, tofree, node);
4215 ac->avail -= tofree;
4216 memmove(ac->entry, &(ac->entry[tofree]),
4217 sizeof(void *) * ac->avail);
4219 spin_unlock_irq(&l3->list_lock);
4224 * cache_reap - Reclaim memory from caches.
4225 * @w: work descriptor
4227 * Called from workqueue/eventd every few seconds.
4229 * - clear the per-cpu caches for this CPU.
4230 * - return freeable pages to the main free memory pool.
4232 * If we cannot acquire the cache chain mutex then just give up - we'll try
4233 * again on the next iteration.
4235 static void cache_reap(struct work_struct *w)
4237 struct kmem_cache *searchp;
4238 struct kmem_list3 *l3;
4239 int node = numa_mem_id();
4240 struct delayed_work *work = to_delayed_work(w);
4242 if (!mutex_trylock(&slab_mutex))
4243 /* Give up. Setup the next iteration. */
4246 list_for_each_entry(searchp, &slab_caches, list) {
4250 * We only take the l3 lock if absolutely necessary and we
4251 * have established with reasonable certainty that
4252 * we can do some work if the lock was obtained.
4254 l3 = searchp->nodelists[node];
4256 reap_alien(searchp, l3);
4258 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4261 * These are racy checks but it does not matter
4262 * if we skip one check or scan twice.
4264 if (time_after(l3->next_reap, jiffies))
4267 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4269 drain_array(searchp, l3, l3->shared, 0, node);
4271 if (l3->free_touched)
4272 l3->free_touched = 0;
4276 freed = drain_freelist(searchp, l3, (l3->free_limit +
4277 5 * searchp->num - 1) / (5 * searchp->num));
4278 STATS_ADD_REAPED(searchp, freed);
4284 mutex_unlock(&slab_mutex);
4287 /* Set up the next iteration */
4288 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4291 #ifdef CONFIG_SLABINFO
4293 static void print_slabinfo_header(struct seq_file *m)
4296 * Output format version, so at least we can change it
4297 * without _too_ many complaints.
4300 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4302 seq_puts(m, "slabinfo - version: 2.1\n");
4304 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4305 "<objperslab> <pagesperslab>");
4306 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4307 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4309 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4310 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4311 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4316 static void *s_start(struct seq_file *m, loff_t *pos)
4320 mutex_lock(&slab_mutex);
4322 print_slabinfo_header(m);
4324 return seq_list_start(&slab_caches, *pos);
4327 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4329 return seq_list_next(p, &slab_caches, pos);
4332 static void s_stop(struct seq_file *m, void *p)
4334 mutex_unlock(&slab_mutex);
4337 static int s_show(struct seq_file *m, void *p)
4339 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4341 unsigned long active_objs;
4342 unsigned long num_objs;
4343 unsigned long active_slabs = 0;
4344 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4348 struct kmem_list3 *l3;
4352 for_each_online_node(node) {
4353 l3 = cachep->nodelists[node];
4358 spin_lock_irq(&l3->list_lock);
4360 list_for_each_entry(slabp, &l3->slabs_full, list) {
4361 if (slabp->inuse != cachep->num && !error)
4362 error = "slabs_full accounting error";
4363 active_objs += cachep->num;
4366 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4367 if (slabp->inuse == cachep->num && !error)
4368 error = "slabs_partial inuse accounting error";
4369 if (!slabp->inuse && !error)
4370 error = "slabs_partial/inuse accounting error";
4371 active_objs += slabp->inuse;
4374 list_for_each_entry(slabp, &l3->slabs_free, list) {
4375 if (slabp->inuse && !error)
4376 error = "slabs_free/inuse accounting error";
4379 free_objects += l3->free_objects;
4381 shared_avail += l3->shared->avail;
4383 spin_unlock_irq(&l3->list_lock);
4385 num_slabs += active_slabs;
4386 num_objs = num_slabs * cachep->num;
4387 if (num_objs - active_objs != free_objects && !error)
4388 error = "free_objects accounting error";
4390 name = cachep->name;
4392 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4394 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4395 name, active_objs, num_objs, cachep->size,
4396 cachep->num, (1 << cachep->gfporder));
4397 seq_printf(m, " : tunables %4u %4u %4u",
4398 cachep->limit, cachep->batchcount, cachep->shared);
4399 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4400 active_slabs, num_slabs, shared_avail);
4403 unsigned long high = cachep->high_mark;
4404 unsigned long allocs = cachep->num_allocations;
4405 unsigned long grown = cachep->grown;
4406 unsigned long reaped = cachep->reaped;
4407 unsigned long errors = cachep->errors;
4408 unsigned long max_freeable = cachep->max_freeable;
4409 unsigned long node_allocs = cachep->node_allocs;
4410 unsigned long node_frees = cachep->node_frees;
4411 unsigned long overflows = cachep->node_overflow;
4413 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4414 "%4lu %4lu %4lu %4lu %4lu",
4415 allocs, high, grown,
4416 reaped, errors, max_freeable, node_allocs,
4417 node_frees, overflows);
4421 unsigned long allochit = atomic_read(&cachep->allochit);
4422 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4423 unsigned long freehit = atomic_read(&cachep->freehit);
4424 unsigned long freemiss = atomic_read(&cachep->freemiss);
4426 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4427 allochit, allocmiss, freehit, freemiss);
4435 * slabinfo_op - iterator that generates /proc/slabinfo
4444 * num-pages-per-slab
4445 * + further values on SMP and with statistics enabled
4448 static const struct seq_operations slabinfo_op = {
4455 #define MAX_SLABINFO_WRITE 128
4457 * slabinfo_write - Tuning for the slab allocator
4459 * @buffer: user buffer
4460 * @count: data length
4463 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4464 size_t count, loff_t *ppos)
4466 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4467 int limit, batchcount, shared, res;
4468 struct kmem_cache *cachep;
4470 if (count > MAX_SLABINFO_WRITE)
4472 if (copy_from_user(&kbuf, buffer, count))
4474 kbuf[MAX_SLABINFO_WRITE] = '\0';
4476 tmp = strchr(kbuf, ' ');
4481 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4484 /* Find the cache in the chain of caches. */
4485 mutex_lock(&slab_mutex);
4487 list_for_each_entry(cachep, &slab_caches, list) {
4488 if (!strcmp(cachep->name, kbuf)) {
4489 if (limit < 1 || batchcount < 1 ||
4490 batchcount > limit || shared < 0) {
4493 res = do_tune_cpucache(cachep, limit,
4500 mutex_unlock(&slab_mutex);
4506 static int slabinfo_open(struct inode *inode, struct file *file)
4508 return seq_open(file, &slabinfo_op);
4511 static const struct file_operations proc_slabinfo_operations = {
4512 .open = slabinfo_open,
4514 .write = slabinfo_write,
4515 .llseek = seq_lseek,
4516 .release = seq_release,
4519 #ifdef CONFIG_DEBUG_SLAB_LEAK
4521 static void *leaks_start(struct seq_file *m, loff_t *pos)
4523 mutex_lock(&slab_mutex);
4524 return seq_list_start(&slab_caches, *pos);
4527 static inline int add_caller(unsigned long *n, unsigned long v)
4537 unsigned long *q = p + 2 * i;
4551 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4557 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4563 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4564 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4566 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4571 static void show_symbol(struct seq_file *m, unsigned long address)
4573 #ifdef CONFIG_KALLSYMS
4574 unsigned long offset, size;
4575 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4577 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4578 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4580 seq_printf(m, " [%s]", modname);
4584 seq_printf(m, "%p", (void *)address);
4587 static int leaks_show(struct seq_file *m, void *p)
4589 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4591 struct kmem_list3 *l3;
4593 unsigned long *n = m->private;
4597 if (!(cachep->flags & SLAB_STORE_USER))
4599 if (!(cachep->flags & SLAB_RED_ZONE))
4602 /* OK, we can do it */
4606 for_each_online_node(node) {
4607 l3 = cachep->nodelists[node];
4612 spin_lock_irq(&l3->list_lock);
4614 list_for_each_entry(slabp, &l3->slabs_full, list)
4615 handle_slab(n, cachep, slabp);
4616 list_for_each_entry(slabp, &l3->slabs_partial, list)
4617 handle_slab(n, cachep, slabp);
4618 spin_unlock_irq(&l3->list_lock);
4620 name = cachep->name;
4622 /* Increase the buffer size */
4623 mutex_unlock(&slab_mutex);
4624 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4626 /* Too bad, we are really out */
4628 mutex_lock(&slab_mutex);
4631 *(unsigned long *)m->private = n[0] * 2;
4633 mutex_lock(&slab_mutex);
4634 /* Now make sure this entry will be retried */
4638 for (i = 0; i < n[1]; i++) {
4639 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4640 show_symbol(m, n[2*i+2]);
4647 static const struct seq_operations slabstats_op = {
4648 .start = leaks_start,
4654 static int slabstats_open(struct inode *inode, struct file *file)
4656 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4659 ret = seq_open(file, &slabstats_op);
4661 struct seq_file *m = file->private_data;
4662 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4671 static const struct file_operations proc_slabstats_operations = {
4672 .open = slabstats_open,
4674 .llseek = seq_lseek,
4675 .release = seq_release_private,
4679 static int __init slab_proc_init(void)
4681 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4682 #ifdef CONFIG_DEBUG_SLAB_LEAK
4683 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4687 module_init(slab_proc_init);
4691 * ksize - get the actual amount of memory allocated for a given object
4692 * @objp: Pointer to the object
4694 * kmalloc may internally round up allocations and return more memory
4695 * than requested. ksize() can be used to determine the actual amount of
4696 * memory allocated. The caller may use this additional memory, even though
4697 * a smaller amount of memory was initially specified with the kmalloc call.
4698 * The caller must guarantee that objp points to a valid object previously
4699 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4700 * must not be freed during the duration of the call.
4702 size_t ksize(const void *objp)
4705 if (unlikely(objp == ZERO_SIZE_PTR))
4708 return virt_to_cache(objp)->object_size;
4710 EXPORT_SYMBOL(ksize);