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;})
502 * Do not go above this order unless 0 objects fit into the slab or
503 * overridden on the command line.
505 #define SLAB_MAX_ORDER_HI 1
506 #define SLAB_MAX_ORDER_LO 0
507 static int slab_max_order = SLAB_MAX_ORDER_LO;
508 static bool slab_max_order_set __initdata;
510 static inline struct kmem_cache *virt_to_cache(const void *obj)
512 struct page *page = virt_to_head_page(obj);
513 return page->slab_cache;
516 static inline struct slab *virt_to_slab(const void *obj)
518 struct page *page = virt_to_head_page(obj);
520 VM_BUG_ON(!PageSlab(page));
521 return page->slab_page;
524 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
527 return slab->s_mem + cache->size * idx;
531 * We want to avoid an expensive divide : (offset / cache->size)
532 * Using the fact that size is a constant for a particular cache,
533 * we can replace (offset / cache->size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
537 const struct slab *slab, void *obj)
539 u32 offset = (obj - slab->s_mem);
540 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 struct cache_sizes malloc_sizes[] = {
547 #define CACHE(x) { .cs_size = (x) },
548 #include <linux/kmalloc_sizes.h>
552 EXPORT_SYMBOL(malloc_sizes);
554 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
560 static struct cache_names __initdata cache_names[] = {
561 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562 #include <linux/kmalloc_sizes.h>
567 static struct arraycache_init initarray_cache __initdata =
568 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
569 static struct arraycache_init initarray_generic =
570 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
572 /* internal cache of cache description objs */
573 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
574 static struct kmem_cache cache_cache = {
575 .nodelists = cache_cache_nodelists,
577 .limit = BOOT_CPUCACHE_ENTRIES,
579 .size = sizeof(struct kmem_cache),
580 .name = "kmem_cache",
583 #define BAD_ALIEN_MAGIC 0x01020304ul
585 #ifdef CONFIG_LOCKDEP
588 * Slab sometimes uses the kmalloc slabs to store the slab headers
589 * for other slabs "off slab".
590 * The locking for this is tricky in that it nests within the locks
591 * of all other slabs in a few places; to deal with this special
592 * locking we put on-slab caches into a separate lock-class.
594 * We set lock class for alien array caches which are up during init.
595 * The lock annotation will be lost if all cpus of a node goes down and
596 * then comes back up during hotplug
598 static struct lock_class_key on_slab_l3_key;
599 static struct lock_class_key on_slab_alc_key;
601 static struct lock_class_key debugobj_l3_key;
602 static struct lock_class_key debugobj_alc_key;
604 static void slab_set_lock_classes(struct kmem_cache *cachep,
605 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
608 struct array_cache **alc;
609 struct kmem_list3 *l3;
612 l3 = cachep->nodelists[q];
616 lockdep_set_class(&l3->list_lock, l3_key);
619 * FIXME: This check for BAD_ALIEN_MAGIC
620 * should go away when common slab code is taught to
621 * work even without alien caches.
622 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
623 * for alloc_alien_cache,
625 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
629 lockdep_set_class(&alc[r]->lock, alc_key);
633 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
635 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
638 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
642 for_each_online_node(node)
643 slab_set_debugobj_lock_classes_node(cachep, node);
646 static void init_node_lock_keys(int q)
648 struct cache_sizes *s = malloc_sizes;
653 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
654 struct kmem_list3 *l3;
656 l3 = s->cs_cachep->nodelists[q];
657 if (!l3 || OFF_SLAB(s->cs_cachep))
660 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
661 &on_slab_alc_key, q);
665 static inline void init_lock_keys(void)
670 init_node_lock_keys(node);
673 static void init_node_lock_keys(int q)
677 static inline void init_lock_keys(void)
681 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
685 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
690 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
692 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
694 return cachep->array[smp_processor_id()];
697 static inline struct kmem_cache *__find_general_cachep(size_t size,
700 struct cache_sizes *csizep = malloc_sizes;
703 /* This happens if someone tries to call
704 * kmem_cache_create(), or __kmalloc(), before
705 * the generic caches are initialized.
707 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
710 return ZERO_SIZE_PTR;
712 while (size > csizep->cs_size)
716 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
717 * has cs_{dma,}cachep==NULL. Thus no special case
718 * for large kmalloc calls required.
720 #ifdef CONFIG_ZONE_DMA
721 if (unlikely(gfpflags & GFP_DMA))
722 return csizep->cs_dmacachep;
724 return csizep->cs_cachep;
727 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
729 return __find_general_cachep(size, gfpflags);
732 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
734 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
738 * Calculate the number of objects and left-over bytes for a given buffer size.
740 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
741 size_t align, int flags, size_t *left_over,
746 size_t slab_size = PAGE_SIZE << gfporder;
749 * The slab management structure can be either off the slab or
750 * on it. For the latter case, the memory allocated for a
754 * - One kmem_bufctl_t for each object
755 * - Padding to respect alignment of @align
756 * - @buffer_size bytes for each object
758 * If the slab management structure is off the slab, then the
759 * alignment will already be calculated into the size. Because
760 * the slabs are all pages aligned, the objects will be at the
761 * correct alignment when allocated.
763 if (flags & CFLGS_OFF_SLAB) {
765 nr_objs = slab_size / buffer_size;
767 if (nr_objs > SLAB_LIMIT)
768 nr_objs = SLAB_LIMIT;
771 * Ignore padding for the initial guess. The padding
772 * is at most @align-1 bytes, and @buffer_size is at
773 * least @align. In the worst case, this result will
774 * be one greater than the number of objects that fit
775 * into the memory allocation when taking the padding
778 nr_objs = (slab_size - sizeof(struct slab)) /
779 (buffer_size + sizeof(kmem_bufctl_t));
782 * This calculated number will be either the right
783 * amount, or one greater than what we want.
785 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
789 if (nr_objs > SLAB_LIMIT)
790 nr_objs = SLAB_LIMIT;
792 mgmt_size = slab_mgmt_size(nr_objs, align);
795 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
798 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
800 static void __slab_error(const char *function, struct kmem_cache *cachep,
803 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
804 function, cachep->name, msg);
806 add_taint(TAINT_BAD_PAGE);
810 * By default on NUMA we use alien caches to stage the freeing of
811 * objects allocated from other nodes. This causes massive memory
812 * inefficiencies when using fake NUMA setup to split memory into a
813 * large number of small nodes, so it can be disabled on the command
817 static int use_alien_caches __read_mostly = 1;
818 static int __init noaliencache_setup(char *s)
820 use_alien_caches = 0;
823 __setup("noaliencache", noaliencache_setup);
825 static int __init slab_max_order_setup(char *str)
827 get_option(&str, &slab_max_order);
828 slab_max_order = slab_max_order < 0 ? 0 :
829 min(slab_max_order, MAX_ORDER - 1);
830 slab_max_order_set = true;
834 __setup("slab_max_order=", slab_max_order_setup);
838 * Special reaping functions for NUMA systems called from cache_reap().
839 * These take care of doing round robin flushing of alien caches (containing
840 * objects freed on different nodes from which they were allocated) and the
841 * flushing of remote pcps by calling drain_node_pages.
843 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
845 static void init_reap_node(int cpu)
849 node = next_node(cpu_to_mem(cpu), node_online_map);
850 if (node == MAX_NUMNODES)
851 node = first_node(node_online_map);
853 per_cpu(slab_reap_node, cpu) = node;
856 static void next_reap_node(void)
858 int node = __this_cpu_read(slab_reap_node);
860 node = next_node(node, node_online_map);
861 if (unlikely(node >= MAX_NUMNODES))
862 node = first_node(node_online_map);
863 __this_cpu_write(slab_reap_node, node);
867 #define init_reap_node(cpu) do { } while (0)
868 #define next_reap_node(void) do { } while (0)
872 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
873 * via the workqueue/eventd.
874 * Add the CPU number into the expiration time to minimize the possibility of
875 * the CPUs getting into lockstep and contending for the global cache chain
878 static void __cpuinit start_cpu_timer(int cpu)
880 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
883 * When this gets called from do_initcalls via cpucache_init(),
884 * init_workqueues() has already run, so keventd will be setup
887 if (keventd_up() && reap_work->work.func == NULL) {
889 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
890 schedule_delayed_work_on(cpu, reap_work,
891 __round_jiffies_relative(HZ, cpu));
895 static struct array_cache *alloc_arraycache(int node, int entries,
896 int batchcount, gfp_t gfp)
898 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
899 struct array_cache *nc = NULL;
901 nc = kmalloc_node(memsize, gfp, node);
903 * The array_cache structures contain pointers to free object.
904 * However, when such objects are allocated or transferred to another
905 * cache the pointers are not cleared and they could be counted as
906 * valid references during a kmemleak scan. Therefore, kmemleak must
907 * not scan such objects.
909 kmemleak_no_scan(nc);
913 nc->batchcount = batchcount;
915 spin_lock_init(&nc->lock);
920 static inline bool is_slab_pfmemalloc(struct slab *slabp)
922 struct page *page = virt_to_page(slabp->s_mem);
924 return PageSlabPfmemalloc(page);
927 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
928 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
929 struct array_cache *ac)
931 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
935 if (!pfmemalloc_active)
938 spin_lock_irqsave(&l3->list_lock, flags);
939 list_for_each_entry(slabp, &l3->slabs_full, list)
940 if (is_slab_pfmemalloc(slabp))
943 list_for_each_entry(slabp, &l3->slabs_partial, list)
944 if (is_slab_pfmemalloc(slabp))
947 list_for_each_entry(slabp, &l3->slabs_free, list)
948 if (is_slab_pfmemalloc(slabp))
951 pfmemalloc_active = false;
953 spin_unlock_irqrestore(&l3->list_lock, flags);
956 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
957 gfp_t flags, bool force_refill)
960 void *objp = ac->entry[--ac->avail];
962 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
963 if (unlikely(is_obj_pfmemalloc(objp))) {
964 struct kmem_list3 *l3;
966 if (gfp_pfmemalloc_allowed(flags)) {
967 clear_obj_pfmemalloc(&objp);
971 /* The caller cannot use PFMEMALLOC objects, find another one */
972 for (i = 1; i < ac->avail; i++) {
973 /* If a !PFMEMALLOC object is found, swap them */
974 if (!is_obj_pfmemalloc(ac->entry[i])) {
976 ac->entry[i] = ac->entry[ac->avail];
977 ac->entry[ac->avail] = objp;
983 * If there are empty slabs on the slabs_free list and we are
984 * being forced to refill the cache, mark this one !pfmemalloc.
986 l3 = cachep->nodelists[numa_mem_id()];
987 if (!list_empty(&l3->slabs_free) && force_refill) {
988 struct slab *slabp = virt_to_slab(objp);
989 ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem));
990 clear_obj_pfmemalloc(&objp);
991 recheck_pfmemalloc_active(cachep, ac);
995 /* No !PFMEMALLOC objects available */
1003 static inline void *ac_get_obj(struct kmem_cache *cachep,
1004 struct array_cache *ac, gfp_t flags, bool force_refill)
1008 if (unlikely(sk_memalloc_socks()))
1009 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1011 objp = ac->entry[--ac->avail];
1016 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1019 if (unlikely(pfmemalloc_active)) {
1020 /* Some pfmemalloc slabs exist, check if this is one */
1021 struct page *page = virt_to_page(objp);
1022 if (PageSlabPfmemalloc(page))
1023 set_obj_pfmemalloc(&objp);
1029 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1032 if (unlikely(sk_memalloc_socks()))
1033 objp = __ac_put_obj(cachep, ac, objp);
1035 ac->entry[ac->avail++] = objp;
1039 * Transfer objects in one arraycache to another.
1040 * Locking must be handled by the caller.
1042 * Return the number of entries transferred.
1044 static int transfer_objects(struct array_cache *to,
1045 struct array_cache *from, unsigned int max)
1047 /* Figure out how many entries to transfer */
1048 int nr = min3(from->avail, max, to->limit - to->avail);
1053 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1054 sizeof(void *) *nr);
1063 #define drain_alien_cache(cachep, alien) do { } while (0)
1064 #define reap_alien(cachep, l3) do { } while (0)
1066 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1068 return (struct array_cache **)BAD_ALIEN_MAGIC;
1071 static inline void free_alien_cache(struct array_cache **ac_ptr)
1075 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1080 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1086 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1087 gfp_t flags, int nodeid)
1092 #else /* CONFIG_NUMA */
1094 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1095 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1097 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1099 struct array_cache **ac_ptr;
1100 int memsize = sizeof(void *) * nr_node_ids;
1105 ac_ptr = kzalloc_node(memsize, gfp, node);
1108 if (i == node || !node_online(i))
1110 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1112 for (i--; i >= 0; i--)
1122 static void free_alien_cache(struct array_cache **ac_ptr)
1133 static void __drain_alien_cache(struct kmem_cache *cachep,
1134 struct array_cache *ac, int node)
1136 struct kmem_list3 *rl3 = cachep->nodelists[node];
1139 spin_lock(&rl3->list_lock);
1141 * Stuff objects into the remote nodes shared array first.
1142 * That way we could avoid the overhead of putting the objects
1143 * into the free lists and getting them back later.
1146 transfer_objects(rl3->shared, ac, ac->limit);
1148 free_block(cachep, ac->entry, ac->avail, node);
1150 spin_unlock(&rl3->list_lock);
1155 * Called from cache_reap() to regularly drain alien caches round robin.
1157 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1159 int node = __this_cpu_read(slab_reap_node);
1162 struct array_cache *ac = l3->alien[node];
1164 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1165 __drain_alien_cache(cachep, ac, node);
1166 spin_unlock_irq(&ac->lock);
1171 static void drain_alien_cache(struct kmem_cache *cachep,
1172 struct array_cache **alien)
1175 struct array_cache *ac;
1176 unsigned long flags;
1178 for_each_online_node(i) {
1181 spin_lock_irqsave(&ac->lock, flags);
1182 __drain_alien_cache(cachep, ac, i);
1183 spin_unlock_irqrestore(&ac->lock, flags);
1188 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1190 struct slab *slabp = virt_to_slab(objp);
1191 int nodeid = slabp->nodeid;
1192 struct kmem_list3 *l3;
1193 struct array_cache *alien = NULL;
1196 node = numa_mem_id();
1199 * Make sure we are not freeing a object from another node to the array
1200 * cache on this cpu.
1202 if (likely(slabp->nodeid == node))
1205 l3 = cachep->nodelists[node];
1206 STATS_INC_NODEFREES(cachep);
1207 if (l3->alien && l3->alien[nodeid]) {
1208 alien = l3->alien[nodeid];
1209 spin_lock(&alien->lock);
1210 if (unlikely(alien->avail == alien->limit)) {
1211 STATS_INC_ACOVERFLOW(cachep);
1212 __drain_alien_cache(cachep, alien, nodeid);
1214 ac_put_obj(cachep, alien, objp);
1215 spin_unlock(&alien->lock);
1217 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1218 free_block(cachep, &objp, 1, nodeid);
1219 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1226 * Allocates and initializes nodelists for a node on each slab cache, used for
1227 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1228 * will be allocated off-node since memory is not yet online for the new node.
1229 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1232 * Must hold slab_mutex.
1234 static int init_cache_nodelists_node(int node)
1236 struct kmem_cache *cachep;
1237 struct kmem_list3 *l3;
1238 const int memsize = sizeof(struct kmem_list3);
1240 list_for_each_entry(cachep, &slab_caches, list) {
1242 * Set up the size64 kmemlist for cpu before we can
1243 * begin anything. Make sure some other cpu on this
1244 * node has not already allocated this
1246 if (!cachep->nodelists[node]) {
1247 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1250 kmem_list3_init(l3);
1251 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1252 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1255 * The l3s don't come and go as CPUs come and
1256 * go. slab_mutex is sufficient
1259 cachep->nodelists[node] = l3;
1262 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1263 cachep->nodelists[node]->free_limit =
1264 (1 + nr_cpus_node(node)) *
1265 cachep->batchcount + cachep->num;
1266 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1271 static void __cpuinit cpuup_canceled(long cpu)
1273 struct kmem_cache *cachep;
1274 struct kmem_list3 *l3 = NULL;
1275 int node = cpu_to_mem(cpu);
1276 const struct cpumask *mask = cpumask_of_node(node);
1278 list_for_each_entry(cachep, &slab_caches, list) {
1279 struct array_cache *nc;
1280 struct array_cache *shared;
1281 struct array_cache **alien;
1283 /* cpu is dead; no one can alloc from it. */
1284 nc = cachep->array[cpu];
1285 cachep->array[cpu] = NULL;
1286 l3 = cachep->nodelists[node];
1289 goto free_array_cache;
1291 spin_lock_irq(&l3->list_lock);
1293 /* Free limit for this kmem_list3 */
1294 l3->free_limit -= cachep->batchcount;
1296 free_block(cachep, nc->entry, nc->avail, node);
1298 if (!cpumask_empty(mask)) {
1299 spin_unlock_irq(&l3->list_lock);
1300 goto free_array_cache;
1303 shared = l3->shared;
1305 free_block(cachep, shared->entry,
1306 shared->avail, node);
1313 spin_unlock_irq(&l3->list_lock);
1317 drain_alien_cache(cachep, alien);
1318 free_alien_cache(alien);
1324 * In the previous loop, all the objects were freed to
1325 * the respective cache's slabs, now we can go ahead and
1326 * shrink each nodelist to its limit.
1328 list_for_each_entry(cachep, &slab_caches, list) {
1329 l3 = cachep->nodelists[node];
1332 drain_freelist(cachep, l3, l3->free_objects);
1336 static int __cpuinit cpuup_prepare(long cpu)
1338 struct kmem_cache *cachep;
1339 struct kmem_list3 *l3 = NULL;
1340 int node = cpu_to_mem(cpu);
1344 * We need to do this right in the beginning since
1345 * alloc_arraycache's are going to use this list.
1346 * kmalloc_node allows us to add the slab to the right
1347 * kmem_list3 and not this cpu's kmem_list3
1349 err = init_cache_nodelists_node(node);
1354 * Now we can go ahead with allocating the shared arrays and
1357 list_for_each_entry(cachep, &slab_caches, list) {
1358 struct array_cache *nc;
1359 struct array_cache *shared = NULL;
1360 struct array_cache **alien = NULL;
1362 nc = alloc_arraycache(node, cachep->limit,
1363 cachep->batchcount, GFP_KERNEL);
1366 if (cachep->shared) {
1367 shared = alloc_arraycache(node,
1368 cachep->shared * cachep->batchcount,
1369 0xbaadf00d, GFP_KERNEL);
1375 if (use_alien_caches) {
1376 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1383 cachep->array[cpu] = nc;
1384 l3 = cachep->nodelists[node];
1387 spin_lock_irq(&l3->list_lock);
1390 * We are serialised from CPU_DEAD or
1391 * CPU_UP_CANCELLED by the cpucontrol lock
1393 l3->shared = shared;
1402 spin_unlock_irq(&l3->list_lock);
1404 free_alien_cache(alien);
1405 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1406 slab_set_debugobj_lock_classes_node(cachep, node);
1408 init_node_lock_keys(node);
1412 cpuup_canceled(cpu);
1416 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1417 unsigned long action, void *hcpu)
1419 long cpu = (long)hcpu;
1423 case CPU_UP_PREPARE:
1424 case CPU_UP_PREPARE_FROZEN:
1425 mutex_lock(&slab_mutex);
1426 err = cpuup_prepare(cpu);
1427 mutex_unlock(&slab_mutex);
1430 case CPU_ONLINE_FROZEN:
1431 start_cpu_timer(cpu);
1433 #ifdef CONFIG_HOTPLUG_CPU
1434 case CPU_DOWN_PREPARE:
1435 case CPU_DOWN_PREPARE_FROZEN:
1437 * Shutdown cache reaper. Note that the slab_mutex is
1438 * held so that if cache_reap() is invoked it cannot do
1439 * anything expensive but will only modify reap_work
1440 * and reschedule the timer.
1442 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1443 /* Now the cache_reaper is guaranteed to be not running. */
1444 per_cpu(slab_reap_work, cpu).work.func = NULL;
1446 case CPU_DOWN_FAILED:
1447 case CPU_DOWN_FAILED_FROZEN:
1448 start_cpu_timer(cpu);
1451 case CPU_DEAD_FROZEN:
1453 * Even if all the cpus of a node are down, we don't free the
1454 * kmem_list3 of any cache. This to avoid a race between
1455 * cpu_down, and a kmalloc allocation from another cpu for
1456 * memory from the node of the cpu going down. The list3
1457 * structure is usually allocated from kmem_cache_create() and
1458 * gets destroyed at kmem_cache_destroy().
1462 case CPU_UP_CANCELED:
1463 case CPU_UP_CANCELED_FROZEN:
1464 mutex_lock(&slab_mutex);
1465 cpuup_canceled(cpu);
1466 mutex_unlock(&slab_mutex);
1469 return notifier_from_errno(err);
1472 static struct notifier_block __cpuinitdata cpucache_notifier = {
1473 &cpuup_callback, NULL, 0
1476 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1478 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1479 * Returns -EBUSY if all objects cannot be drained so that the node is not
1482 * Must hold slab_mutex.
1484 static int __meminit drain_cache_nodelists_node(int node)
1486 struct kmem_cache *cachep;
1489 list_for_each_entry(cachep, &slab_caches, list) {
1490 struct kmem_list3 *l3;
1492 l3 = cachep->nodelists[node];
1496 drain_freelist(cachep, l3, l3->free_objects);
1498 if (!list_empty(&l3->slabs_full) ||
1499 !list_empty(&l3->slabs_partial)) {
1507 static int __meminit slab_memory_callback(struct notifier_block *self,
1508 unsigned long action, void *arg)
1510 struct memory_notify *mnb = arg;
1514 nid = mnb->status_change_nid;
1519 case MEM_GOING_ONLINE:
1520 mutex_lock(&slab_mutex);
1521 ret = init_cache_nodelists_node(nid);
1522 mutex_unlock(&slab_mutex);
1524 case MEM_GOING_OFFLINE:
1525 mutex_lock(&slab_mutex);
1526 ret = drain_cache_nodelists_node(nid);
1527 mutex_unlock(&slab_mutex);
1531 case MEM_CANCEL_ONLINE:
1532 case MEM_CANCEL_OFFLINE:
1536 return notifier_from_errno(ret);
1538 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1541 * swap the static kmem_list3 with kmalloced memory
1543 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1546 struct kmem_list3 *ptr;
1548 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1551 memcpy(ptr, list, sizeof(struct kmem_list3));
1553 * Do not assume that spinlocks can be initialized via memcpy:
1555 spin_lock_init(&ptr->list_lock);
1557 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1558 cachep->nodelists[nodeid] = ptr;
1562 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1563 * size of kmem_list3.
1565 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1569 for_each_online_node(node) {
1570 cachep->nodelists[node] = &initkmem_list3[index + node];
1571 cachep->nodelists[node]->next_reap = jiffies +
1573 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1578 * Initialisation. Called after the page allocator have been initialised and
1579 * before smp_init().
1581 void __init kmem_cache_init(void)
1584 struct cache_sizes *sizes;
1585 struct cache_names *names;
1590 if (num_possible_nodes() == 1)
1591 use_alien_caches = 0;
1593 for (i = 0; i < NUM_INIT_LISTS; i++) {
1594 kmem_list3_init(&initkmem_list3[i]);
1595 if (i < MAX_NUMNODES)
1596 cache_cache.nodelists[i] = NULL;
1598 set_up_list3s(&cache_cache, CACHE_CACHE);
1601 * Fragmentation resistance on low memory - only use bigger
1602 * page orders on machines with more than 32MB of memory if
1603 * not overridden on the command line.
1605 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1606 slab_max_order = SLAB_MAX_ORDER_HI;
1608 /* Bootstrap is tricky, because several objects are allocated
1609 * from caches that do not exist yet:
1610 * 1) initialize the cache_cache cache: it contains the struct
1611 * kmem_cache structures of all caches, except cache_cache itself:
1612 * cache_cache is statically allocated.
1613 * Initially an __init data area is used for the head array and the
1614 * kmem_list3 structures, it's replaced with a kmalloc allocated
1615 * array at the end of the bootstrap.
1616 * 2) Create the first kmalloc cache.
1617 * The struct kmem_cache for the new cache is allocated normally.
1618 * An __init data area is used for the head array.
1619 * 3) Create the remaining kmalloc caches, with minimally sized
1621 * 4) Replace the __init data head arrays for cache_cache and the first
1622 * kmalloc cache with kmalloc allocated arrays.
1623 * 5) Replace the __init data for kmem_list3 for cache_cache and
1624 * the other cache's with kmalloc allocated memory.
1625 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1628 node = numa_mem_id();
1630 /* 1) create the cache_cache */
1631 INIT_LIST_HEAD(&slab_caches);
1632 list_add(&cache_cache.list, &slab_caches);
1633 cache_cache.colour_off = cache_line_size();
1634 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1635 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1638 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1640 cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1641 nr_node_ids * sizeof(struct kmem_list3 *);
1642 cache_cache.object_size = cache_cache.size;
1643 cache_cache.size = ALIGN(cache_cache.size,
1645 cache_cache.reciprocal_buffer_size =
1646 reciprocal_value(cache_cache.size);
1648 for (order = 0; order < MAX_ORDER; order++) {
1649 cache_estimate(order, cache_cache.size,
1650 cache_line_size(), 0, &left_over, &cache_cache.num);
1651 if (cache_cache.num)
1654 BUG_ON(!cache_cache.num);
1655 cache_cache.gfporder = order;
1656 cache_cache.colour = left_over / cache_cache.colour_off;
1657 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1658 sizeof(struct slab), cache_line_size());
1660 /* 2+3) create the kmalloc caches */
1661 sizes = malloc_sizes;
1662 names = cache_names;
1665 * Initialize the caches that provide memory for the array cache and the
1666 * kmem_list3 structures first. Without this, further allocations will
1670 sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name,
1671 sizes[INDEX_AC].cs_size,
1672 ARCH_KMALLOC_MINALIGN,
1673 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1676 if (INDEX_AC != INDEX_L3) {
1677 sizes[INDEX_L3].cs_cachep =
1678 __kmem_cache_create(names[INDEX_L3].name,
1679 sizes[INDEX_L3].cs_size,
1680 ARCH_KMALLOC_MINALIGN,
1681 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1685 slab_early_init = 0;
1687 while (sizes->cs_size != ULONG_MAX) {
1689 * For performance, all the general caches are L1 aligned.
1690 * This should be particularly beneficial on SMP boxes, as it
1691 * eliminates "false sharing".
1692 * Note for systems short on memory removing the alignment will
1693 * allow tighter packing of the smaller caches.
1695 if (!sizes->cs_cachep) {
1696 sizes->cs_cachep = __kmem_cache_create(names->name,
1698 ARCH_KMALLOC_MINALIGN,
1699 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1702 #ifdef CONFIG_ZONE_DMA
1703 sizes->cs_dmacachep = __kmem_cache_create(
1706 ARCH_KMALLOC_MINALIGN,
1707 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1714 /* 4) Replace the bootstrap head arrays */
1716 struct array_cache *ptr;
1718 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1720 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1721 memcpy(ptr, cpu_cache_get(&cache_cache),
1722 sizeof(struct arraycache_init));
1724 * Do not assume that spinlocks can be initialized via memcpy:
1726 spin_lock_init(&ptr->lock);
1728 cache_cache.array[smp_processor_id()] = ptr;
1730 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1732 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1733 != &initarray_generic.cache);
1734 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1735 sizeof(struct arraycache_init));
1737 * Do not assume that spinlocks can be initialized via memcpy:
1739 spin_lock_init(&ptr->lock);
1741 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1744 /* 5) Replace the bootstrap kmem_list3's */
1748 for_each_online_node(nid) {
1749 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1751 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1752 &initkmem_list3[SIZE_AC + nid], nid);
1754 if (INDEX_AC != INDEX_L3) {
1755 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1756 &initkmem_list3[SIZE_L3 + nid], nid);
1764 void __init kmem_cache_init_late(void)
1766 struct kmem_cache *cachep;
1770 /* 6) resize the head arrays to their final sizes */
1771 mutex_lock(&slab_mutex);
1772 list_for_each_entry(cachep, &slab_caches, list)
1773 if (enable_cpucache(cachep, GFP_NOWAIT))
1775 mutex_unlock(&slab_mutex);
1777 /* Annotate slab for lockdep -- annotate the malloc caches */
1784 * Register a cpu startup notifier callback that initializes
1785 * cpu_cache_get for all new cpus
1787 register_cpu_notifier(&cpucache_notifier);
1791 * Register a memory hotplug callback that initializes and frees
1794 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1798 * The reap timers are started later, with a module init call: That part
1799 * of the kernel is not yet operational.
1803 static int __init cpucache_init(void)
1808 * Register the timers that return unneeded pages to the page allocator
1810 for_each_online_cpu(cpu)
1811 start_cpu_timer(cpu);
1817 __initcall(cpucache_init);
1819 static noinline void
1820 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1822 struct kmem_list3 *l3;
1824 unsigned long flags;
1828 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1830 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1831 cachep->name, cachep->size, cachep->gfporder);
1833 for_each_online_node(node) {
1834 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1835 unsigned long active_slabs = 0, num_slabs = 0;
1837 l3 = cachep->nodelists[node];
1841 spin_lock_irqsave(&l3->list_lock, flags);
1842 list_for_each_entry(slabp, &l3->slabs_full, list) {
1843 active_objs += cachep->num;
1846 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1847 active_objs += slabp->inuse;
1850 list_for_each_entry(slabp, &l3->slabs_free, list)
1853 free_objects += l3->free_objects;
1854 spin_unlock_irqrestore(&l3->list_lock, flags);
1856 num_slabs += active_slabs;
1857 num_objs = num_slabs * cachep->num;
1859 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1860 node, active_slabs, num_slabs, active_objs, num_objs,
1866 * Interface to system's page allocator. No need to hold the cache-lock.
1868 * If we requested dmaable memory, we will get it. Even if we
1869 * did not request dmaable memory, we might get it, but that
1870 * would be relatively rare and ignorable.
1872 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1880 * Nommu uses slab's for process anonymous memory allocations, and thus
1881 * requires __GFP_COMP to properly refcount higher order allocations
1883 flags |= __GFP_COMP;
1886 flags |= cachep->allocflags;
1887 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1888 flags |= __GFP_RECLAIMABLE;
1890 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1892 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1893 slab_out_of_memory(cachep, flags, nodeid);
1897 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1898 if (unlikely(page->pfmemalloc))
1899 pfmemalloc_active = true;
1901 nr_pages = (1 << cachep->gfporder);
1902 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1903 add_zone_page_state(page_zone(page),
1904 NR_SLAB_RECLAIMABLE, nr_pages);
1906 add_zone_page_state(page_zone(page),
1907 NR_SLAB_UNRECLAIMABLE, nr_pages);
1908 for (i = 0; i < nr_pages; i++) {
1909 __SetPageSlab(page + i);
1911 if (page->pfmemalloc)
1912 SetPageSlabPfmemalloc(page + i);
1915 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1916 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1919 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1921 kmemcheck_mark_unallocated_pages(page, nr_pages);
1924 return page_address(page);
1928 * Interface to system's page release.
1930 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1932 unsigned long i = (1 << cachep->gfporder);
1933 struct page *page = virt_to_page(addr);
1934 const unsigned long nr_freed = i;
1936 kmemcheck_free_shadow(page, cachep->gfporder);
1938 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1939 sub_zone_page_state(page_zone(page),
1940 NR_SLAB_RECLAIMABLE, nr_freed);
1942 sub_zone_page_state(page_zone(page),
1943 NR_SLAB_UNRECLAIMABLE, nr_freed);
1945 BUG_ON(!PageSlab(page));
1946 __ClearPageSlabPfmemalloc(page);
1947 __ClearPageSlab(page);
1950 if (current->reclaim_state)
1951 current->reclaim_state->reclaimed_slab += nr_freed;
1952 free_pages((unsigned long)addr, cachep->gfporder);
1955 static void kmem_rcu_free(struct rcu_head *head)
1957 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1958 struct kmem_cache *cachep = slab_rcu->cachep;
1960 kmem_freepages(cachep, slab_rcu->addr);
1961 if (OFF_SLAB(cachep))
1962 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1967 #ifdef CONFIG_DEBUG_PAGEALLOC
1968 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1969 unsigned long caller)
1971 int size = cachep->object_size;
1973 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1975 if (size < 5 * sizeof(unsigned long))
1978 *addr++ = 0x12345678;
1980 *addr++ = smp_processor_id();
1981 size -= 3 * sizeof(unsigned long);
1983 unsigned long *sptr = &caller;
1984 unsigned long svalue;
1986 while (!kstack_end(sptr)) {
1988 if (kernel_text_address(svalue)) {
1990 size -= sizeof(unsigned long);
1991 if (size <= sizeof(unsigned long))
1997 *addr++ = 0x87654321;
2001 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
2003 int size = cachep->object_size;
2004 addr = &((char *)addr)[obj_offset(cachep)];
2006 memset(addr, val, size);
2007 *(unsigned char *)(addr + size - 1) = POISON_END;
2010 static void dump_line(char *data, int offset, int limit)
2013 unsigned char error = 0;
2016 printk(KERN_ERR "%03x: ", offset);
2017 for (i = 0; i < limit; i++) {
2018 if (data[offset + i] != POISON_FREE) {
2019 error = data[offset + i];
2023 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2024 &data[offset], limit, 1);
2026 if (bad_count == 1) {
2027 error ^= POISON_FREE;
2028 if (!(error & (error - 1))) {
2029 printk(KERN_ERR "Single bit error detected. Probably "
2032 printk(KERN_ERR "Run memtest86+ or a similar memory "
2035 printk(KERN_ERR "Run a memory test tool.\n");
2044 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2049 if (cachep->flags & SLAB_RED_ZONE) {
2050 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2051 *dbg_redzone1(cachep, objp),
2052 *dbg_redzone2(cachep, objp));
2055 if (cachep->flags & SLAB_STORE_USER) {
2056 printk(KERN_ERR "Last user: [<%p>]",
2057 *dbg_userword(cachep, objp));
2058 print_symbol("(%s)",
2059 (unsigned long)*dbg_userword(cachep, objp));
2062 realobj = (char *)objp + obj_offset(cachep);
2063 size = cachep->object_size;
2064 for (i = 0; i < size && lines; i += 16, lines--) {
2067 if (i + limit > size)
2069 dump_line(realobj, i, limit);
2073 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2079 realobj = (char *)objp + obj_offset(cachep);
2080 size = cachep->object_size;
2082 for (i = 0; i < size; i++) {
2083 char exp = POISON_FREE;
2086 if (realobj[i] != exp) {
2092 "Slab corruption (%s): %s start=%p, len=%d\n",
2093 print_tainted(), cachep->name, realobj, size);
2094 print_objinfo(cachep, objp, 0);
2096 /* Hexdump the affected line */
2099 if (i + limit > size)
2101 dump_line(realobj, i, limit);
2104 /* Limit to 5 lines */
2110 /* Print some data about the neighboring objects, if they
2113 struct slab *slabp = virt_to_slab(objp);
2116 objnr = obj_to_index(cachep, slabp, objp);
2118 objp = index_to_obj(cachep, slabp, objnr - 1);
2119 realobj = (char *)objp + obj_offset(cachep);
2120 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2122 print_objinfo(cachep, objp, 2);
2124 if (objnr + 1 < cachep->num) {
2125 objp = index_to_obj(cachep, slabp, objnr + 1);
2126 realobj = (char *)objp + obj_offset(cachep);
2127 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2129 print_objinfo(cachep, objp, 2);
2136 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2139 for (i = 0; i < cachep->num; i++) {
2140 void *objp = index_to_obj(cachep, slabp, i);
2142 if (cachep->flags & SLAB_POISON) {
2143 #ifdef CONFIG_DEBUG_PAGEALLOC
2144 if (cachep->size % PAGE_SIZE == 0 &&
2146 kernel_map_pages(virt_to_page(objp),
2147 cachep->size / PAGE_SIZE, 1);
2149 check_poison_obj(cachep, objp);
2151 check_poison_obj(cachep, objp);
2154 if (cachep->flags & SLAB_RED_ZONE) {
2155 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2156 slab_error(cachep, "start of a freed object "
2158 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2159 slab_error(cachep, "end of a freed object "
2165 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2171 * slab_destroy - destroy and release all objects in a slab
2172 * @cachep: cache pointer being destroyed
2173 * @slabp: slab pointer being destroyed
2175 * Destroy all the objs in a slab, and release the mem back to the system.
2176 * Before calling the slab must have been unlinked from the cache. The
2177 * cache-lock is not held/needed.
2179 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2181 void *addr = slabp->s_mem - slabp->colouroff;
2183 slab_destroy_debugcheck(cachep, slabp);
2184 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2185 struct slab_rcu *slab_rcu;
2187 slab_rcu = (struct slab_rcu *)slabp;
2188 slab_rcu->cachep = cachep;
2189 slab_rcu->addr = addr;
2190 call_rcu(&slab_rcu->head, kmem_rcu_free);
2192 kmem_freepages(cachep, addr);
2193 if (OFF_SLAB(cachep))
2194 kmem_cache_free(cachep->slabp_cache, slabp);
2198 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2201 struct kmem_list3 *l3;
2203 for_each_online_cpu(i)
2204 kfree(cachep->array[i]);
2206 /* NUMA: free the list3 structures */
2207 for_each_online_node(i) {
2208 l3 = cachep->nodelists[i];
2211 free_alien_cache(l3->alien);
2215 kmem_cache_free(&cache_cache, cachep);
2220 * calculate_slab_order - calculate size (page order) of slabs
2221 * @cachep: pointer to the cache that is being created
2222 * @size: size of objects to be created in this cache.
2223 * @align: required alignment for the objects.
2224 * @flags: slab allocation flags
2226 * Also calculates the number of objects per slab.
2228 * This could be made much more intelligent. For now, try to avoid using
2229 * high order pages for slabs. When the gfp() functions are more friendly
2230 * towards high-order requests, this should be changed.
2232 static size_t calculate_slab_order(struct kmem_cache *cachep,
2233 size_t size, size_t align, unsigned long flags)
2235 unsigned long offslab_limit;
2236 size_t left_over = 0;
2239 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2243 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2247 if (flags & CFLGS_OFF_SLAB) {
2249 * Max number of objs-per-slab for caches which
2250 * use off-slab slabs. Needed to avoid a possible
2251 * looping condition in cache_grow().
2253 offslab_limit = size - sizeof(struct slab);
2254 offslab_limit /= sizeof(kmem_bufctl_t);
2256 if (num > offslab_limit)
2260 /* Found something acceptable - save it away */
2262 cachep->gfporder = gfporder;
2263 left_over = remainder;
2266 * A VFS-reclaimable slab tends to have most allocations
2267 * as GFP_NOFS and we really don't want to have to be allocating
2268 * higher-order pages when we are unable to shrink dcache.
2270 if (flags & SLAB_RECLAIM_ACCOUNT)
2274 * Large number of objects is good, but very large slabs are
2275 * currently bad for the gfp()s.
2277 if (gfporder >= slab_max_order)
2281 * Acceptable internal fragmentation?
2283 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2289 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2291 if (slab_state >= FULL)
2292 return enable_cpucache(cachep, gfp);
2294 if (slab_state == DOWN) {
2296 * Note: the first kmem_cache_create must create the cache
2297 * that's used by kmalloc(24), otherwise the creation of
2298 * further caches will BUG().
2300 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2303 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2304 * the first cache, then we need to set up all its list3s,
2305 * otherwise the creation of further caches will BUG().
2307 set_up_list3s(cachep, SIZE_AC);
2308 if (INDEX_AC == INDEX_L3)
2309 slab_state = PARTIAL_L3;
2311 slab_state = PARTIAL_ARRAYCACHE;
2313 cachep->array[smp_processor_id()] =
2314 kmalloc(sizeof(struct arraycache_init), gfp);
2316 if (slab_state == PARTIAL_ARRAYCACHE) {
2317 set_up_list3s(cachep, SIZE_L3);
2318 slab_state = PARTIAL_L3;
2321 for_each_online_node(node) {
2322 cachep->nodelists[node] =
2323 kmalloc_node(sizeof(struct kmem_list3),
2325 BUG_ON(!cachep->nodelists[node]);
2326 kmem_list3_init(cachep->nodelists[node]);
2330 cachep->nodelists[numa_mem_id()]->next_reap =
2331 jiffies + REAPTIMEOUT_LIST3 +
2332 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2334 cpu_cache_get(cachep)->avail = 0;
2335 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2336 cpu_cache_get(cachep)->batchcount = 1;
2337 cpu_cache_get(cachep)->touched = 0;
2338 cachep->batchcount = 1;
2339 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2344 * __kmem_cache_create - Create a cache.
2345 * @name: A string which is used in /proc/slabinfo to identify this cache.
2346 * @size: The size of objects to be created in this cache.
2347 * @align: The required alignment for the objects.
2348 * @flags: SLAB flags
2349 * @ctor: A constructor for the objects.
2351 * Returns a ptr to the cache on success, NULL on failure.
2352 * Cannot be called within a int, but can be interrupted.
2353 * The @ctor is run when new pages are allocated by the cache.
2355 * @name must be valid until the cache is destroyed. This implies that
2356 * the module calling this has to destroy the cache before getting unloaded.
2360 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2361 * to catch references to uninitialised memory.
2363 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2364 * for buffer overruns.
2366 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2367 * cacheline. This can be beneficial if you're counting cycles as closely
2371 __kmem_cache_create (const char *name, size_t size, size_t align,
2372 unsigned long flags, void (*ctor)(void *))
2374 size_t left_over, slab_size, ralign;
2375 struct kmem_cache *cachep = NULL;
2381 * Enable redzoning and last user accounting, except for caches with
2382 * large objects, if the increased size would increase the object size
2383 * above the next power of two: caches with object sizes just above a
2384 * power of two have a significant amount of internal fragmentation.
2386 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2387 2 * sizeof(unsigned long long)))
2388 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2389 if (!(flags & SLAB_DESTROY_BY_RCU))
2390 flags |= SLAB_POISON;
2392 if (flags & SLAB_DESTROY_BY_RCU)
2393 BUG_ON(flags & SLAB_POISON);
2396 * Always checks flags, a caller might be expecting debug support which
2399 BUG_ON(flags & ~CREATE_MASK);
2402 * Check that size is in terms of words. This is needed to avoid
2403 * unaligned accesses for some archs when redzoning is used, and makes
2404 * sure any on-slab bufctl's are also correctly aligned.
2406 if (size & (BYTES_PER_WORD - 1)) {
2407 size += (BYTES_PER_WORD - 1);
2408 size &= ~(BYTES_PER_WORD - 1);
2411 /* calculate the final buffer alignment: */
2413 /* 1) arch recommendation: can be overridden for debug */
2414 if (flags & SLAB_HWCACHE_ALIGN) {
2416 * Default alignment: as specified by the arch code. Except if
2417 * an object is really small, then squeeze multiple objects into
2420 ralign = cache_line_size();
2421 while (size <= ralign / 2)
2424 ralign = BYTES_PER_WORD;
2428 * Redzoning and user store require word alignment or possibly larger.
2429 * Note this will be overridden by architecture or caller mandated
2430 * alignment if either is greater than BYTES_PER_WORD.
2432 if (flags & SLAB_STORE_USER)
2433 ralign = BYTES_PER_WORD;
2435 if (flags & SLAB_RED_ZONE) {
2436 ralign = REDZONE_ALIGN;
2437 /* If redzoning, ensure that the second redzone is suitably
2438 * aligned, by adjusting the object size accordingly. */
2439 size += REDZONE_ALIGN - 1;
2440 size &= ~(REDZONE_ALIGN - 1);
2443 /* 2) arch mandated alignment */
2444 if (ralign < ARCH_SLAB_MINALIGN) {
2445 ralign = ARCH_SLAB_MINALIGN;
2447 /* 3) caller mandated alignment */
2448 if (ralign < align) {
2451 /* disable debug if necessary */
2452 if (ralign > __alignof__(unsigned long long))
2453 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2459 if (slab_is_available())
2464 /* Get cache's description obj. */
2465 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2469 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2470 cachep->object_size = size;
2471 cachep->align = align;
2475 * Both debugging options require word-alignment which is calculated
2478 if (flags & SLAB_RED_ZONE) {
2479 /* add space for red zone words */
2480 cachep->obj_offset += sizeof(unsigned long long);
2481 size += 2 * sizeof(unsigned long long);
2483 if (flags & SLAB_STORE_USER) {
2484 /* user store requires one word storage behind the end of
2485 * the real object. But if the second red zone needs to be
2486 * aligned to 64 bits, we must allow that much space.
2488 if (flags & SLAB_RED_ZONE)
2489 size += REDZONE_ALIGN;
2491 size += BYTES_PER_WORD;
2493 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2494 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2495 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2496 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2503 * Determine if the slab management is 'on' or 'off' slab.
2504 * (bootstrapping cannot cope with offslab caches so don't do
2505 * it too early on. Always use on-slab management when
2506 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2508 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2509 !(flags & SLAB_NOLEAKTRACE))
2511 * Size is large, assume best to place the slab management obj
2512 * off-slab (should allow better packing of objs).
2514 flags |= CFLGS_OFF_SLAB;
2516 size = ALIGN(size, align);
2518 left_over = calculate_slab_order(cachep, size, align, flags);
2522 "kmem_cache_create: couldn't create cache %s.\n", name);
2523 kmem_cache_free(&cache_cache, cachep);
2526 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2527 + sizeof(struct slab), align);
2530 * If the slab has been placed off-slab, and we have enough space then
2531 * move it on-slab. This is at the expense of any extra colouring.
2533 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2534 flags &= ~CFLGS_OFF_SLAB;
2535 left_over -= slab_size;
2538 if (flags & CFLGS_OFF_SLAB) {
2539 /* really off slab. No need for manual alignment */
2541 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2543 #ifdef CONFIG_PAGE_POISONING
2544 /* If we're going to use the generic kernel_map_pages()
2545 * poisoning, then it's going to smash the contents of
2546 * the redzone and userword anyhow, so switch them off.
2548 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2549 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2553 cachep->colour_off = cache_line_size();
2554 /* Offset must be a multiple of the alignment. */
2555 if (cachep->colour_off < align)
2556 cachep->colour_off = align;
2557 cachep->colour = left_over / cachep->colour_off;
2558 cachep->slab_size = slab_size;
2559 cachep->flags = flags;
2560 cachep->allocflags = 0;
2561 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2562 cachep->allocflags |= GFP_DMA;
2563 cachep->size = size;
2564 cachep->reciprocal_buffer_size = reciprocal_value(size);
2566 if (flags & CFLGS_OFF_SLAB) {
2567 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2569 * This is a possibility for one of the malloc_sizes caches.
2570 * But since we go off slab only for object size greater than
2571 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2572 * this should not happen at all.
2573 * But leave a BUG_ON for some lucky dude.
2575 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2577 cachep->ctor = ctor;
2578 cachep->name = name;
2580 if (setup_cpu_cache(cachep, gfp)) {
2581 __kmem_cache_destroy(cachep);
2585 if (flags & SLAB_DEBUG_OBJECTS) {
2587 * Would deadlock through slab_destroy()->call_rcu()->
2588 * debug_object_activate()->kmem_cache_alloc().
2590 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2592 slab_set_debugobj_lock_classes(cachep);
2595 /* cache setup completed, link it into the list */
2596 list_add(&cachep->list, &slab_caches);
2601 static void check_irq_off(void)
2603 BUG_ON(!irqs_disabled());
2606 static void check_irq_on(void)
2608 BUG_ON(irqs_disabled());
2611 static void check_spinlock_acquired(struct kmem_cache *cachep)
2615 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2619 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2623 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2628 #define check_irq_off() do { } while(0)
2629 #define check_irq_on() do { } while(0)
2630 #define check_spinlock_acquired(x) do { } while(0)
2631 #define check_spinlock_acquired_node(x, y) do { } while(0)
2634 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2635 struct array_cache *ac,
2636 int force, int node);
2638 static void do_drain(void *arg)
2640 struct kmem_cache *cachep = arg;
2641 struct array_cache *ac;
2642 int node = numa_mem_id();
2645 ac = cpu_cache_get(cachep);
2646 spin_lock(&cachep->nodelists[node]->list_lock);
2647 free_block(cachep, ac->entry, ac->avail, node);
2648 spin_unlock(&cachep->nodelists[node]->list_lock);
2652 static void drain_cpu_caches(struct kmem_cache *cachep)
2654 struct kmem_list3 *l3;
2657 on_each_cpu(do_drain, cachep, 1);
2659 for_each_online_node(node) {
2660 l3 = cachep->nodelists[node];
2661 if (l3 && l3->alien)
2662 drain_alien_cache(cachep, l3->alien);
2665 for_each_online_node(node) {
2666 l3 = cachep->nodelists[node];
2668 drain_array(cachep, l3, l3->shared, 1, node);
2673 * Remove slabs from the list of free slabs.
2674 * Specify the number of slabs to drain in tofree.
2676 * Returns the actual number of slabs released.
2678 static int drain_freelist(struct kmem_cache *cache,
2679 struct kmem_list3 *l3, int tofree)
2681 struct list_head *p;
2686 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2688 spin_lock_irq(&l3->list_lock);
2689 p = l3->slabs_free.prev;
2690 if (p == &l3->slabs_free) {
2691 spin_unlock_irq(&l3->list_lock);
2695 slabp = list_entry(p, struct slab, list);
2697 BUG_ON(slabp->inuse);
2699 list_del(&slabp->list);
2701 * Safe to drop the lock. The slab is no longer linked
2704 l3->free_objects -= cache->num;
2705 spin_unlock_irq(&l3->list_lock);
2706 slab_destroy(cache, slabp);
2713 /* Called with slab_mutex held to protect against cpu hotplug */
2714 static int __cache_shrink(struct kmem_cache *cachep)
2717 struct kmem_list3 *l3;
2719 drain_cpu_caches(cachep);
2722 for_each_online_node(i) {
2723 l3 = cachep->nodelists[i];
2727 drain_freelist(cachep, l3, l3->free_objects);
2729 ret += !list_empty(&l3->slabs_full) ||
2730 !list_empty(&l3->slabs_partial);
2732 return (ret ? 1 : 0);
2736 * kmem_cache_shrink - Shrink a cache.
2737 * @cachep: The cache to shrink.
2739 * Releases as many slabs as possible for a cache.
2740 * To help debugging, a zero exit status indicates all slabs were released.
2742 int kmem_cache_shrink(struct kmem_cache *cachep)
2745 BUG_ON(!cachep || in_interrupt());
2748 mutex_lock(&slab_mutex);
2749 ret = __cache_shrink(cachep);
2750 mutex_unlock(&slab_mutex);
2754 EXPORT_SYMBOL(kmem_cache_shrink);
2757 * kmem_cache_destroy - delete a cache
2758 * @cachep: the cache to destroy
2760 * Remove a &struct kmem_cache object from the slab cache.
2762 * It is expected this function will be called by a module when it is
2763 * unloaded. This will remove the cache completely, and avoid a duplicate
2764 * cache being allocated each time a module is loaded and unloaded, if the
2765 * module doesn't have persistent in-kernel storage across loads and unloads.
2767 * The cache must be empty before calling this function.
2769 * The caller must guarantee that no one will allocate memory from the cache
2770 * during the kmem_cache_destroy().
2772 void kmem_cache_destroy(struct kmem_cache *cachep)
2774 BUG_ON(!cachep || in_interrupt());
2776 /* Find the cache in the chain of caches. */
2778 mutex_lock(&slab_mutex);
2780 * the chain is never empty, cache_cache is never destroyed
2782 list_del(&cachep->list);
2783 if (__cache_shrink(cachep)) {
2784 slab_error(cachep, "Can't free all objects");
2785 list_add(&cachep->list, &slab_caches);
2786 mutex_unlock(&slab_mutex);
2791 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2794 __kmem_cache_destroy(cachep);
2795 mutex_unlock(&slab_mutex);
2798 EXPORT_SYMBOL(kmem_cache_destroy);
2801 * Get the memory for a slab management obj.
2802 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2803 * always come from malloc_sizes caches. The slab descriptor cannot
2804 * come from the same cache which is getting created because,
2805 * when we are searching for an appropriate cache for these
2806 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2807 * If we are creating a malloc_sizes cache here it would not be visible to
2808 * kmem_find_general_cachep till the initialization is complete.
2809 * Hence we cannot have slabp_cache same as the original cache.
2811 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2812 int colour_off, gfp_t local_flags,
2817 if (OFF_SLAB(cachep)) {
2818 /* Slab management obj is off-slab. */
2819 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2820 local_flags, nodeid);
2822 * If the first object in the slab is leaked (it's allocated
2823 * but no one has a reference to it), we want to make sure
2824 * kmemleak does not treat the ->s_mem pointer as a reference
2825 * to the object. Otherwise we will not report the leak.
2827 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2832 slabp = objp + colour_off;
2833 colour_off += cachep->slab_size;
2836 slabp->colouroff = colour_off;
2837 slabp->s_mem = objp + colour_off;
2838 slabp->nodeid = nodeid;
2843 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2845 return (kmem_bufctl_t *) (slabp + 1);
2848 static void cache_init_objs(struct kmem_cache *cachep,
2853 for (i = 0; i < cachep->num; i++) {
2854 void *objp = index_to_obj(cachep, slabp, i);
2856 /* need to poison the objs? */
2857 if (cachep->flags & SLAB_POISON)
2858 poison_obj(cachep, objp, POISON_FREE);
2859 if (cachep->flags & SLAB_STORE_USER)
2860 *dbg_userword(cachep, objp) = NULL;
2862 if (cachep->flags & SLAB_RED_ZONE) {
2863 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2864 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2867 * Constructors are not allowed to allocate memory from the same
2868 * cache which they are a constructor for. Otherwise, deadlock.
2869 * They must also be threaded.
2871 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2872 cachep->ctor(objp + obj_offset(cachep));
2874 if (cachep->flags & SLAB_RED_ZONE) {
2875 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2876 slab_error(cachep, "constructor overwrote the"
2877 " end of an object");
2878 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2879 slab_error(cachep, "constructor overwrote the"
2880 " start of an object");
2882 if ((cachep->size % PAGE_SIZE) == 0 &&
2883 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2884 kernel_map_pages(virt_to_page(objp),
2885 cachep->size / PAGE_SIZE, 0);
2890 slab_bufctl(slabp)[i] = i + 1;
2892 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2895 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2897 if (CONFIG_ZONE_DMA_FLAG) {
2898 if (flags & GFP_DMA)
2899 BUG_ON(!(cachep->allocflags & GFP_DMA));
2901 BUG_ON(cachep->allocflags & GFP_DMA);
2905 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2908 void *objp = index_to_obj(cachep, slabp, slabp->free);
2912 next = slab_bufctl(slabp)[slabp->free];
2914 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2915 WARN_ON(slabp->nodeid != nodeid);
2922 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2923 void *objp, int nodeid)
2925 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2928 /* Verify that the slab belongs to the intended node */
2929 WARN_ON(slabp->nodeid != nodeid);
2931 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2932 printk(KERN_ERR "slab: double free detected in cache "
2933 "'%s', objp %p\n", cachep->name, objp);
2937 slab_bufctl(slabp)[objnr] = slabp->free;
2938 slabp->free = objnr;
2943 * Map pages beginning at addr to the given cache and slab. This is required
2944 * for the slab allocator to be able to lookup the cache and slab of a
2945 * virtual address for kfree, ksize, and slab debugging.
2947 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2953 page = virt_to_page(addr);
2956 if (likely(!PageCompound(page)))
2957 nr_pages <<= cache->gfporder;
2960 page->slab_cache = cache;
2961 page->slab_page = slab;
2963 } while (--nr_pages);
2967 * Grow (by 1) the number of slabs within a cache. This is called by
2968 * kmem_cache_alloc() when there are no active objs left in a cache.
2970 static int cache_grow(struct kmem_cache *cachep,
2971 gfp_t flags, int nodeid, void *objp)
2976 struct kmem_list3 *l3;
2979 * Be lazy and only check for valid flags here, keeping it out of the
2980 * critical path in kmem_cache_alloc().
2982 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2983 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2985 /* Take the l3 list lock to change the colour_next on this node */
2987 l3 = cachep->nodelists[nodeid];
2988 spin_lock(&l3->list_lock);
2990 /* Get colour for the slab, and cal the next value. */
2991 offset = l3->colour_next;
2993 if (l3->colour_next >= cachep->colour)
2994 l3->colour_next = 0;
2995 spin_unlock(&l3->list_lock);
2997 offset *= cachep->colour_off;
2999 if (local_flags & __GFP_WAIT)
3003 * The test for missing atomic flag is performed here, rather than
3004 * the more obvious place, simply to reduce the critical path length
3005 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
3006 * will eventually be caught here (where it matters).
3008 kmem_flagcheck(cachep, flags);
3011 * Get mem for the objs. Attempt to allocate a physical page from
3015 objp = kmem_getpages(cachep, local_flags, nodeid);
3019 /* Get slab management. */
3020 slabp = alloc_slabmgmt(cachep, objp, offset,
3021 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
3025 slab_map_pages(cachep, slabp, objp);
3027 cache_init_objs(cachep, slabp);
3029 if (local_flags & __GFP_WAIT)
3030 local_irq_disable();
3032 spin_lock(&l3->list_lock);
3034 /* Make slab active. */
3035 list_add_tail(&slabp->list, &(l3->slabs_free));
3036 STATS_INC_GROWN(cachep);
3037 l3->free_objects += cachep->num;
3038 spin_unlock(&l3->list_lock);
3041 kmem_freepages(cachep, objp);
3043 if (local_flags & __GFP_WAIT)
3044 local_irq_disable();
3051 * Perform extra freeing checks:
3052 * - detect bad pointers.
3053 * - POISON/RED_ZONE checking
3055 static void kfree_debugcheck(const void *objp)
3057 if (!virt_addr_valid(objp)) {
3058 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3059 (unsigned long)objp);
3064 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3066 unsigned long long redzone1, redzone2;
3068 redzone1 = *dbg_redzone1(cache, obj);
3069 redzone2 = *dbg_redzone2(cache, obj);
3074 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3077 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3078 slab_error(cache, "double free detected");
3080 slab_error(cache, "memory outside object was overwritten");
3082 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3083 obj, redzone1, redzone2);
3086 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3087 unsigned long caller)
3093 BUG_ON(virt_to_cache(objp) != cachep);
3095 objp -= obj_offset(cachep);
3096 kfree_debugcheck(objp);
3097 page = virt_to_head_page(objp);
3099 slabp = page->slab_page;
3101 if (cachep->flags & SLAB_RED_ZONE) {
3102 verify_redzone_free(cachep, objp);
3103 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3104 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3106 if (cachep->flags & SLAB_STORE_USER)
3107 *dbg_userword(cachep, objp) = (void *)caller;
3109 objnr = obj_to_index(cachep, slabp, objp);
3111 BUG_ON(objnr >= cachep->num);
3112 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3114 #ifdef CONFIG_DEBUG_SLAB_LEAK
3115 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3117 if (cachep->flags & SLAB_POISON) {
3118 #ifdef CONFIG_DEBUG_PAGEALLOC
3119 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3120 store_stackinfo(cachep, objp, caller);
3121 kernel_map_pages(virt_to_page(objp),
3122 cachep->size / PAGE_SIZE, 0);
3124 poison_obj(cachep, objp, POISON_FREE);
3127 poison_obj(cachep, objp, POISON_FREE);
3133 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3138 /* Check slab's freelist to see if this obj is there. */
3139 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3141 if (entries > cachep->num || i >= cachep->num)
3144 if (entries != cachep->num - slabp->inuse) {
3146 printk(KERN_ERR "slab: Internal list corruption detected in "
3147 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3148 cachep->name, cachep->num, slabp, slabp->inuse,
3150 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3151 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3157 #define kfree_debugcheck(x) do { } while(0)
3158 #define cache_free_debugcheck(x,objp,z) (objp)
3159 #define check_slabp(x,y) do { } while(0)
3162 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3166 struct kmem_list3 *l3;
3167 struct array_cache *ac;
3171 node = numa_mem_id();
3172 if (unlikely(force_refill))
3175 ac = cpu_cache_get(cachep);
3176 batchcount = ac->batchcount;
3177 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3179 * If there was little recent activity on this cache, then
3180 * perform only a partial refill. Otherwise we could generate
3183 batchcount = BATCHREFILL_LIMIT;
3185 l3 = cachep->nodelists[node];
3187 BUG_ON(ac->avail > 0 || !l3);
3188 spin_lock(&l3->list_lock);
3190 /* See if we can refill from the shared array */
3191 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3192 l3->shared->touched = 1;
3196 while (batchcount > 0) {
3197 struct list_head *entry;
3199 /* Get slab alloc is to come from. */
3200 entry = l3->slabs_partial.next;
3201 if (entry == &l3->slabs_partial) {
3202 l3->free_touched = 1;
3203 entry = l3->slabs_free.next;
3204 if (entry == &l3->slabs_free)
3208 slabp = list_entry(entry, struct slab, list);
3209 check_slabp(cachep, slabp);
3210 check_spinlock_acquired(cachep);
3213 * The slab was either on partial or free list so
3214 * there must be at least one object available for
3217 BUG_ON(slabp->inuse >= cachep->num);
3219 while (slabp->inuse < cachep->num && batchcount--) {
3220 STATS_INC_ALLOCED(cachep);
3221 STATS_INC_ACTIVE(cachep);
3222 STATS_SET_HIGH(cachep);
3224 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3227 check_slabp(cachep, slabp);
3229 /* move slabp to correct slabp list: */
3230 list_del(&slabp->list);
3231 if (slabp->free == BUFCTL_END)
3232 list_add(&slabp->list, &l3->slabs_full);
3234 list_add(&slabp->list, &l3->slabs_partial);
3238 l3->free_objects -= ac->avail;
3240 spin_unlock(&l3->list_lock);
3242 if (unlikely(!ac->avail)) {
3245 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3247 /* cache_grow can reenable interrupts, then ac could change. */
3248 ac = cpu_cache_get(cachep);
3250 /* no objects in sight? abort */
3251 if (!x && (ac->avail == 0 || force_refill))
3254 if (!ac->avail) /* objects refilled by interrupt? */
3259 return ac_get_obj(cachep, ac, flags, force_refill);
3262 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3265 might_sleep_if(flags & __GFP_WAIT);
3267 kmem_flagcheck(cachep, flags);
3272 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3273 gfp_t flags, void *objp, unsigned long caller)
3277 if (cachep->flags & SLAB_POISON) {
3278 #ifdef CONFIG_DEBUG_PAGEALLOC
3279 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3280 kernel_map_pages(virt_to_page(objp),
3281 cachep->size / PAGE_SIZE, 1);
3283 check_poison_obj(cachep, objp);
3285 check_poison_obj(cachep, objp);
3287 poison_obj(cachep, objp, POISON_INUSE);
3289 if (cachep->flags & SLAB_STORE_USER)
3290 *dbg_userword(cachep, objp) = (void *)caller;
3292 if (cachep->flags & SLAB_RED_ZONE) {
3293 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3294 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3295 slab_error(cachep, "double free, or memory outside"
3296 " object was overwritten");
3298 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3299 objp, *dbg_redzone1(cachep, objp),
3300 *dbg_redzone2(cachep, objp));
3302 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3303 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3305 #ifdef CONFIG_DEBUG_SLAB_LEAK
3310 slabp = virt_to_head_page(objp)->slab_page;
3311 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3312 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3315 objp += obj_offset(cachep);
3316 if (cachep->ctor && cachep->flags & SLAB_POISON)
3318 if (ARCH_SLAB_MINALIGN &&
3319 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3320 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3321 objp, (int)ARCH_SLAB_MINALIGN);
3326 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3329 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3331 if (cachep == &cache_cache)
3334 return should_failslab(cachep->object_size, flags, cachep->flags);
3337 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3340 struct array_cache *ac;
3341 bool force_refill = false;
3345 ac = cpu_cache_get(cachep);
3346 if (likely(ac->avail)) {
3348 objp = ac_get_obj(cachep, ac, flags, false);
3351 * Allow for the possibility all avail objects are not allowed
3352 * by the current flags
3355 STATS_INC_ALLOCHIT(cachep);
3358 force_refill = true;
3361 STATS_INC_ALLOCMISS(cachep);
3362 objp = cache_alloc_refill(cachep, flags, force_refill);
3364 * the 'ac' may be updated by cache_alloc_refill(),
3365 * and kmemleak_erase() requires its correct value.
3367 ac = cpu_cache_get(cachep);
3371 * To avoid a false negative, if an object that is in one of the
3372 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3373 * treat the array pointers as a reference to the object.
3376 kmemleak_erase(&ac->entry[ac->avail]);
3382 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3384 * If we are in_interrupt, then process context, including cpusets and
3385 * mempolicy, may not apply and should not be used for allocation policy.
3387 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3389 int nid_alloc, nid_here;
3391 if (in_interrupt() || (flags & __GFP_THISNODE))
3393 nid_alloc = nid_here = numa_mem_id();
3394 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3395 nid_alloc = cpuset_slab_spread_node();
3396 else if (current->mempolicy)
3397 nid_alloc = slab_node();
3398 if (nid_alloc != nid_here)
3399 return ____cache_alloc_node(cachep, flags, nid_alloc);
3404 * Fallback function if there was no memory available and no objects on a
3405 * certain node and fall back is permitted. First we scan all the
3406 * available nodelists for available objects. If that fails then we
3407 * perform an allocation without specifying a node. This allows the page
3408 * allocator to do its reclaim / fallback magic. We then insert the
3409 * slab into the proper nodelist and then allocate from it.
3411 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3413 struct zonelist *zonelist;
3417 enum zone_type high_zoneidx = gfp_zone(flags);
3420 unsigned int cpuset_mems_cookie;
3422 if (flags & __GFP_THISNODE)
3425 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3428 cpuset_mems_cookie = get_mems_allowed();
3429 zonelist = node_zonelist(slab_node(), flags);
3433 * Look through allowed nodes for objects available
3434 * from existing per node queues.
3436 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3437 nid = zone_to_nid(zone);
3439 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3440 cache->nodelists[nid] &&
3441 cache->nodelists[nid]->free_objects) {
3442 obj = ____cache_alloc_node(cache,
3443 flags | GFP_THISNODE, nid);
3451 * This allocation will be performed within the constraints
3452 * of the current cpuset / memory policy requirements.
3453 * We may trigger various forms of reclaim on the allowed
3454 * set and go into memory reserves if necessary.
3456 if (local_flags & __GFP_WAIT)
3458 kmem_flagcheck(cache, flags);
3459 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3460 if (local_flags & __GFP_WAIT)
3461 local_irq_disable();
3464 * Insert into the appropriate per node queues
3466 nid = page_to_nid(virt_to_page(obj));
3467 if (cache_grow(cache, flags, nid, obj)) {
3468 obj = ____cache_alloc_node(cache,
3469 flags | GFP_THISNODE, nid);
3472 * Another processor may allocate the
3473 * objects in the slab since we are
3474 * not holding any locks.
3478 /* cache_grow already freed obj */
3484 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3490 * A interface to enable slab creation on nodeid
3492 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3495 struct list_head *entry;
3497 struct kmem_list3 *l3;
3501 l3 = cachep->nodelists[nodeid];
3506 spin_lock(&l3->list_lock);
3507 entry = l3->slabs_partial.next;
3508 if (entry == &l3->slabs_partial) {
3509 l3->free_touched = 1;
3510 entry = l3->slabs_free.next;
3511 if (entry == &l3->slabs_free)
3515 slabp = list_entry(entry, struct slab, list);
3516 check_spinlock_acquired_node(cachep, nodeid);
3517 check_slabp(cachep, slabp);
3519 STATS_INC_NODEALLOCS(cachep);
3520 STATS_INC_ACTIVE(cachep);
3521 STATS_SET_HIGH(cachep);
3523 BUG_ON(slabp->inuse == cachep->num);
3525 obj = slab_get_obj(cachep, slabp, nodeid);
3526 check_slabp(cachep, slabp);
3528 /* move slabp to correct slabp list: */
3529 list_del(&slabp->list);
3531 if (slabp->free == BUFCTL_END)
3532 list_add(&slabp->list, &l3->slabs_full);
3534 list_add(&slabp->list, &l3->slabs_partial);
3536 spin_unlock(&l3->list_lock);
3540 spin_unlock(&l3->list_lock);
3541 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3545 return fallback_alloc(cachep, flags);
3552 * kmem_cache_alloc_node - Allocate an object on the specified node
3553 * @cachep: The cache to allocate from.
3554 * @flags: See kmalloc().
3555 * @nodeid: node number of the target node.
3556 * @caller: return address of caller, used for debug information
3558 * Identical to kmem_cache_alloc but it will allocate memory on the given
3559 * node, which can improve the performance for cpu bound structures.
3561 * Fallback to other node is possible if __GFP_THISNODE is not set.
3563 static __always_inline void *
3564 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3565 unsigned long caller)
3567 unsigned long save_flags;
3569 int slab_node = numa_mem_id();
3571 flags &= gfp_allowed_mask;
3573 lockdep_trace_alloc(flags);
3575 if (slab_should_failslab(cachep, flags))
3578 cache_alloc_debugcheck_before(cachep, flags);
3579 local_irq_save(save_flags);
3581 if (nodeid == NUMA_NO_NODE)
3584 if (unlikely(!cachep->nodelists[nodeid])) {
3585 /* Node not bootstrapped yet */
3586 ptr = fallback_alloc(cachep, flags);
3590 if (nodeid == slab_node) {
3592 * Use the locally cached objects if possible.
3593 * However ____cache_alloc does not allow fallback
3594 * to other nodes. It may fail while we still have
3595 * objects on other nodes available.
3597 ptr = ____cache_alloc(cachep, flags);
3601 /* ___cache_alloc_node can fall back to other nodes */
3602 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3604 local_irq_restore(save_flags);
3605 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3606 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3610 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3612 if (unlikely((flags & __GFP_ZERO) && ptr))
3613 memset(ptr, 0, cachep->object_size);
3618 static __always_inline void *
3619 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3623 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3624 objp = alternate_node_alloc(cache, flags);
3628 objp = ____cache_alloc(cache, flags);
3631 * We may just have run out of memory on the local node.
3632 * ____cache_alloc_node() knows how to locate memory on other nodes
3635 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3642 static __always_inline void *
3643 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3645 return ____cache_alloc(cachep, flags);
3648 #endif /* CONFIG_NUMA */
3650 static __always_inline void *
3651 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3653 unsigned long save_flags;
3656 flags &= gfp_allowed_mask;
3658 lockdep_trace_alloc(flags);
3660 if (slab_should_failslab(cachep, flags))
3663 cache_alloc_debugcheck_before(cachep, flags);
3664 local_irq_save(save_flags);
3665 objp = __do_cache_alloc(cachep, flags);
3666 local_irq_restore(save_flags);
3667 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3668 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3673 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3675 if (unlikely((flags & __GFP_ZERO) && objp))
3676 memset(objp, 0, cachep->object_size);
3682 * Caller needs to acquire correct kmem_list's list_lock
3684 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3688 struct kmem_list3 *l3;
3690 for (i = 0; i < nr_objects; i++) {
3694 clear_obj_pfmemalloc(&objpp[i]);
3697 slabp = virt_to_slab(objp);
3698 l3 = cachep->nodelists[node];
3699 list_del(&slabp->list);
3700 check_spinlock_acquired_node(cachep, node);
3701 check_slabp(cachep, slabp);
3702 slab_put_obj(cachep, slabp, objp, node);
3703 STATS_DEC_ACTIVE(cachep);
3705 check_slabp(cachep, slabp);
3707 /* fixup slab chains */
3708 if (slabp->inuse == 0) {
3709 if (l3->free_objects > l3->free_limit) {
3710 l3->free_objects -= cachep->num;
3711 /* No need to drop any previously held
3712 * lock here, even if we have a off-slab slab
3713 * descriptor it is guaranteed to come from
3714 * a different cache, refer to comments before
3717 slab_destroy(cachep, slabp);
3719 list_add(&slabp->list, &l3->slabs_free);
3722 /* Unconditionally move a slab to the end of the
3723 * partial list on free - maximum time for the
3724 * other objects to be freed, too.
3726 list_add_tail(&slabp->list, &l3->slabs_partial);
3731 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3734 struct kmem_list3 *l3;
3735 int node = numa_mem_id();
3737 batchcount = ac->batchcount;
3739 BUG_ON(!batchcount || batchcount > ac->avail);
3742 l3 = cachep->nodelists[node];
3743 spin_lock(&l3->list_lock);
3745 struct array_cache *shared_array = l3->shared;
3746 int max = shared_array->limit - shared_array->avail;
3748 if (batchcount > max)
3750 memcpy(&(shared_array->entry[shared_array->avail]),
3751 ac->entry, sizeof(void *) * batchcount);
3752 shared_array->avail += batchcount;
3757 free_block(cachep, ac->entry, batchcount, node);
3762 struct list_head *p;
3764 p = l3->slabs_free.next;
3765 while (p != &(l3->slabs_free)) {
3768 slabp = list_entry(p, struct slab, list);
3769 BUG_ON(slabp->inuse);
3774 STATS_SET_FREEABLE(cachep, i);
3777 spin_unlock(&l3->list_lock);
3778 ac->avail -= batchcount;
3779 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3783 * Release an obj back to its cache. If the obj has a constructed state, it must
3784 * be in this state _before_ it is released. Called with disabled ints.
3786 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3787 unsigned long caller)
3789 struct array_cache *ac = cpu_cache_get(cachep);
3792 kmemleak_free_recursive(objp, cachep->flags);
3793 objp = cache_free_debugcheck(cachep, objp, caller);
3795 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3798 * Skip calling cache_free_alien() when the platform is not numa.
3799 * This will avoid cache misses that happen while accessing slabp (which
3800 * is per page memory reference) to get nodeid. Instead use a global
3801 * variable to skip the call, which is mostly likely to be present in
3804 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3807 if (likely(ac->avail < ac->limit)) {
3808 STATS_INC_FREEHIT(cachep);
3810 STATS_INC_FREEMISS(cachep);
3811 cache_flusharray(cachep, ac);
3814 ac_put_obj(cachep, ac, objp);
3818 * kmem_cache_alloc - Allocate an object
3819 * @cachep: The cache to allocate from.
3820 * @flags: See kmalloc().
3822 * Allocate an object from this cache. The flags are only relevant
3823 * if the cache has no available objects.
3825 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3827 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3829 trace_kmem_cache_alloc(_RET_IP_, ret,
3830 cachep->object_size, cachep->size, flags);
3834 EXPORT_SYMBOL(kmem_cache_alloc);
3836 #ifdef CONFIG_TRACING
3838 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3842 ret = slab_alloc(cachep, flags, _RET_IP_);
3844 trace_kmalloc(_RET_IP_, ret,
3845 size, cachep->size, flags);
3848 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3852 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3854 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3856 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3857 cachep->object_size, cachep->size,
3862 EXPORT_SYMBOL(kmem_cache_alloc_node);
3864 #ifdef CONFIG_TRACING
3865 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3872 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP);
3874 trace_kmalloc_node(_RET_IP_, ret,
3879 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3882 static __always_inline void *
3883 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3885 struct kmem_cache *cachep;
3887 cachep = kmem_find_general_cachep(size, flags);
3888 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3890 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3893 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3894 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3896 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3898 EXPORT_SYMBOL(__kmalloc_node);
3900 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3901 int node, unsigned long caller)
3903 return __do_kmalloc_node(size, flags, node, caller);
3905 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3907 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3909 return __do_kmalloc_node(size, flags, node, 0);
3911 EXPORT_SYMBOL(__kmalloc_node);
3912 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3913 #endif /* CONFIG_NUMA */
3916 * __do_kmalloc - allocate memory
3917 * @size: how many bytes of memory are required.
3918 * @flags: the type of memory to allocate (see kmalloc).
3919 * @caller: function caller for debug tracking of the caller
3921 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3922 unsigned long caller)
3924 struct kmem_cache *cachep;
3927 /* If you want to save a few bytes .text space: replace
3929 * Then kmalloc uses the uninlined functions instead of the inline
3932 cachep = __find_general_cachep(size, flags);
3933 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3935 ret = slab_alloc(cachep, flags, caller);
3937 trace_kmalloc(caller, ret,
3938 size, cachep->size, flags);
3944 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3945 void *__kmalloc(size_t size, gfp_t flags)
3947 return __do_kmalloc(size, flags, _RET_IP_);
3949 EXPORT_SYMBOL(__kmalloc);
3951 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3953 return __do_kmalloc(size, flags, caller);
3955 EXPORT_SYMBOL(__kmalloc_track_caller);
3958 void *__kmalloc(size_t size, gfp_t flags)
3960 return __do_kmalloc(size, flags, 0);
3962 EXPORT_SYMBOL(__kmalloc);
3966 * kmem_cache_free - Deallocate an object
3967 * @cachep: The cache the allocation was from.
3968 * @objp: The previously allocated object.
3970 * Free an object which was previously allocated from this
3973 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3975 unsigned long flags;
3977 local_irq_save(flags);
3978 debug_check_no_locks_freed(objp, cachep->object_size);
3979 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3980 debug_check_no_obj_freed(objp, cachep->object_size);
3981 __cache_free(cachep, objp, _RET_IP_);
3982 local_irq_restore(flags);
3984 trace_kmem_cache_free(_RET_IP_, objp);
3986 EXPORT_SYMBOL(kmem_cache_free);
3989 * kfree - free previously allocated memory
3990 * @objp: pointer returned by kmalloc.
3992 * If @objp is NULL, no operation is performed.
3994 * Don't free memory not originally allocated by kmalloc()
3995 * or you will run into trouble.
3997 void kfree(const void *objp)
3999 struct kmem_cache *c;
4000 unsigned long flags;
4002 trace_kfree(_RET_IP_, objp);
4004 if (unlikely(ZERO_OR_NULL_PTR(objp)))
4006 local_irq_save(flags);
4007 kfree_debugcheck(objp);
4008 c = virt_to_cache(objp);
4009 debug_check_no_locks_freed(objp, c->object_size);
4011 debug_check_no_obj_freed(objp, c->object_size);
4012 __cache_free(c, (void *)objp, _RET_IP_);
4013 local_irq_restore(flags);
4015 EXPORT_SYMBOL(kfree);
4017 unsigned int kmem_cache_size(struct kmem_cache *cachep)
4019 return cachep->object_size;
4021 EXPORT_SYMBOL(kmem_cache_size);
4024 * This initializes kmem_list3 or resizes various caches for all nodes.
4026 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
4029 struct kmem_list3 *l3;
4030 struct array_cache *new_shared;
4031 struct array_cache **new_alien = NULL;
4033 for_each_online_node(node) {
4035 if (use_alien_caches) {
4036 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
4042 if (cachep->shared) {
4043 new_shared = alloc_arraycache(node,
4044 cachep->shared*cachep->batchcount,
4047 free_alien_cache(new_alien);
4052 l3 = cachep->nodelists[node];
4054 struct array_cache *shared = l3->shared;
4056 spin_lock_irq(&l3->list_lock);
4059 free_block(cachep, shared->entry,
4060 shared->avail, node);
4062 l3->shared = new_shared;
4064 l3->alien = new_alien;
4067 l3->free_limit = (1 + nr_cpus_node(node)) *
4068 cachep->batchcount + cachep->num;
4069 spin_unlock_irq(&l3->list_lock);
4071 free_alien_cache(new_alien);
4074 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4076 free_alien_cache(new_alien);
4081 kmem_list3_init(l3);
4082 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4083 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4084 l3->shared = new_shared;
4085 l3->alien = new_alien;
4086 l3->free_limit = (1 + nr_cpus_node(node)) *
4087 cachep->batchcount + cachep->num;
4088 cachep->nodelists[node] = l3;
4093 if (!cachep->list.next) {
4094 /* Cache is not active yet. Roll back what we did */
4097 if (cachep->nodelists[node]) {
4098 l3 = cachep->nodelists[node];
4101 free_alien_cache(l3->alien);
4103 cachep->nodelists[node] = NULL;
4111 struct ccupdate_struct {
4112 struct kmem_cache *cachep;
4113 struct array_cache *new[0];
4116 static void do_ccupdate_local(void *info)
4118 struct ccupdate_struct *new = info;
4119 struct array_cache *old;
4122 old = cpu_cache_get(new->cachep);
4124 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4125 new->new[smp_processor_id()] = old;
4128 /* Always called with the slab_mutex held */
4129 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4130 int batchcount, int shared, gfp_t gfp)
4132 struct ccupdate_struct *new;
4135 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4140 for_each_online_cpu(i) {
4141 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4144 for (i--; i >= 0; i--)
4150 new->cachep = cachep;
4152 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4155 cachep->batchcount = batchcount;
4156 cachep->limit = limit;
4157 cachep->shared = shared;
4159 for_each_online_cpu(i) {
4160 struct array_cache *ccold = new->new[i];
4163 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4164 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4165 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4169 return alloc_kmemlist(cachep, gfp);
4172 /* Called with slab_mutex held always */
4173 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4179 * The head array serves three purposes:
4180 * - create a LIFO ordering, i.e. return objects that are cache-warm
4181 * - reduce the number of spinlock operations.
4182 * - reduce the number of linked list operations on the slab and
4183 * bufctl chains: array operations are cheaper.
4184 * The numbers are guessed, we should auto-tune as described by
4187 if (cachep->size > 131072)
4189 else if (cachep->size > PAGE_SIZE)
4191 else if (cachep->size > 1024)
4193 else if (cachep->size > 256)
4199 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4200 * allocation behaviour: Most allocs on one cpu, most free operations
4201 * on another cpu. For these cases, an efficient object passing between
4202 * cpus is necessary. This is provided by a shared array. The array
4203 * replaces Bonwick's magazine layer.
4204 * On uniprocessor, it's functionally equivalent (but less efficient)
4205 * to a larger limit. Thus disabled by default.
4208 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4213 * With debugging enabled, large batchcount lead to excessively long
4214 * periods with disabled local interrupts. Limit the batchcount
4219 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4221 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4222 cachep->name, -err);
4227 * Drain an array if it contains any elements taking the l3 lock only if
4228 * necessary. Note that the l3 listlock also protects the array_cache
4229 * if drain_array() is used on the shared array.
4231 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4232 struct array_cache *ac, int force, int node)
4236 if (!ac || !ac->avail)
4238 if (ac->touched && !force) {
4241 spin_lock_irq(&l3->list_lock);
4243 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4244 if (tofree > ac->avail)
4245 tofree = (ac->avail + 1) / 2;
4246 free_block(cachep, ac->entry, tofree, node);
4247 ac->avail -= tofree;
4248 memmove(ac->entry, &(ac->entry[tofree]),
4249 sizeof(void *) * ac->avail);
4251 spin_unlock_irq(&l3->list_lock);
4256 * cache_reap - Reclaim memory from caches.
4257 * @w: work descriptor
4259 * Called from workqueue/eventd every few seconds.
4261 * - clear the per-cpu caches for this CPU.
4262 * - return freeable pages to the main free memory pool.
4264 * If we cannot acquire the cache chain mutex then just give up - we'll try
4265 * again on the next iteration.
4267 static void cache_reap(struct work_struct *w)
4269 struct kmem_cache *searchp;
4270 struct kmem_list3 *l3;
4271 int node = numa_mem_id();
4272 struct delayed_work *work = to_delayed_work(w);
4274 if (!mutex_trylock(&slab_mutex))
4275 /* Give up. Setup the next iteration. */
4278 list_for_each_entry(searchp, &slab_caches, list) {
4282 * We only take the l3 lock if absolutely necessary and we
4283 * have established with reasonable certainty that
4284 * we can do some work if the lock was obtained.
4286 l3 = searchp->nodelists[node];
4288 reap_alien(searchp, l3);
4290 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4293 * These are racy checks but it does not matter
4294 * if we skip one check or scan twice.
4296 if (time_after(l3->next_reap, jiffies))
4299 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4301 drain_array(searchp, l3, l3->shared, 0, node);
4303 if (l3->free_touched)
4304 l3->free_touched = 0;
4308 freed = drain_freelist(searchp, l3, (l3->free_limit +
4309 5 * searchp->num - 1) / (5 * searchp->num));
4310 STATS_ADD_REAPED(searchp, freed);
4316 mutex_unlock(&slab_mutex);
4319 /* Set up the next iteration */
4320 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4323 #ifdef CONFIG_SLABINFO
4325 static void print_slabinfo_header(struct seq_file *m)
4328 * Output format version, so at least we can change it
4329 * without _too_ many complaints.
4332 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4334 seq_puts(m, "slabinfo - version: 2.1\n");
4336 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4337 "<objperslab> <pagesperslab>");
4338 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4339 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4341 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4342 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4343 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4348 static void *s_start(struct seq_file *m, loff_t *pos)
4352 mutex_lock(&slab_mutex);
4354 print_slabinfo_header(m);
4356 return seq_list_start(&slab_caches, *pos);
4359 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4361 return seq_list_next(p, &slab_caches, pos);
4364 static void s_stop(struct seq_file *m, void *p)
4366 mutex_unlock(&slab_mutex);
4369 static int s_show(struct seq_file *m, void *p)
4371 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4373 unsigned long active_objs;
4374 unsigned long num_objs;
4375 unsigned long active_slabs = 0;
4376 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4380 struct kmem_list3 *l3;
4384 for_each_online_node(node) {
4385 l3 = cachep->nodelists[node];
4390 spin_lock_irq(&l3->list_lock);
4392 list_for_each_entry(slabp, &l3->slabs_full, list) {
4393 if (slabp->inuse != cachep->num && !error)
4394 error = "slabs_full accounting error";
4395 active_objs += cachep->num;
4398 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4399 if (slabp->inuse == cachep->num && !error)
4400 error = "slabs_partial inuse accounting error";
4401 if (!slabp->inuse && !error)
4402 error = "slabs_partial/inuse accounting error";
4403 active_objs += slabp->inuse;
4406 list_for_each_entry(slabp, &l3->slabs_free, list) {
4407 if (slabp->inuse && !error)
4408 error = "slabs_free/inuse accounting error";
4411 free_objects += l3->free_objects;
4413 shared_avail += l3->shared->avail;
4415 spin_unlock_irq(&l3->list_lock);
4417 num_slabs += active_slabs;
4418 num_objs = num_slabs * cachep->num;
4419 if (num_objs - active_objs != free_objects && !error)
4420 error = "free_objects accounting error";
4422 name = cachep->name;
4424 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4426 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4427 name, active_objs, num_objs, cachep->size,
4428 cachep->num, (1 << cachep->gfporder));
4429 seq_printf(m, " : tunables %4u %4u %4u",
4430 cachep->limit, cachep->batchcount, cachep->shared);
4431 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4432 active_slabs, num_slabs, shared_avail);
4435 unsigned long high = cachep->high_mark;
4436 unsigned long allocs = cachep->num_allocations;
4437 unsigned long grown = cachep->grown;
4438 unsigned long reaped = cachep->reaped;
4439 unsigned long errors = cachep->errors;
4440 unsigned long max_freeable = cachep->max_freeable;
4441 unsigned long node_allocs = cachep->node_allocs;
4442 unsigned long node_frees = cachep->node_frees;
4443 unsigned long overflows = cachep->node_overflow;
4445 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4446 "%4lu %4lu %4lu %4lu %4lu",
4447 allocs, high, grown,
4448 reaped, errors, max_freeable, node_allocs,
4449 node_frees, overflows);
4453 unsigned long allochit = atomic_read(&cachep->allochit);
4454 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4455 unsigned long freehit = atomic_read(&cachep->freehit);
4456 unsigned long freemiss = atomic_read(&cachep->freemiss);
4458 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4459 allochit, allocmiss, freehit, freemiss);
4467 * slabinfo_op - iterator that generates /proc/slabinfo
4476 * num-pages-per-slab
4477 * + further values on SMP and with statistics enabled
4480 static const struct seq_operations slabinfo_op = {
4487 #define MAX_SLABINFO_WRITE 128
4489 * slabinfo_write - Tuning for the slab allocator
4491 * @buffer: user buffer
4492 * @count: data length
4495 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4496 size_t count, loff_t *ppos)
4498 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4499 int limit, batchcount, shared, res;
4500 struct kmem_cache *cachep;
4502 if (count > MAX_SLABINFO_WRITE)
4504 if (copy_from_user(&kbuf, buffer, count))
4506 kbuf[MAX_SLABINFO_WRITE] = '\0';
4508 tmp = strchr(kbuf, ' ');
4513 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4516 /* Find the cache in the chain of caches. */
4517 mutex_lock(&slab_mutex);
4519 list_for_each_entry(cachep, &slab_caches, list) {
4520 if (!strcmp(cachep->name, kbuf)) {
4521 if (limit < 1 || batchcount < 1 ||
4522 batchcount > limit || shared < 0) {
4525 res = do_tune_cpucache(cachep, limit,
4532 mutex_unlock(&slab_mutex);
4538 static int slabinfo_open(struct inode *inode, struct file *file)
4540 return seq_open(file, &slabinfo_op);
4543 static const struct file_operations proc_slabinfo_operations = {
4544 .open = slabinfo_open,
4546 .write = slabinfo_write,
4547 .llseek = seq_lseek,
4548 .release = seq_release,
4551 #ifdef CONFIG_DEBUG_SLAB_LEAK
4553 static void *leaks_start(struct seq_file *m, loff_t *pos)
4555 mutex_lock(&slab_mutex);
4556 return seq_list_start(&slab_caches, *pos);
4559 static inline int add_caller(unsigned long *n, unsigned long v)
4569 unsigned long *q = p + 2 * i;
4583 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4589 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4595 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4596 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4598 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4603 static void show_symbol(struct seq_file *m, unsigned long address)
4605 #ifdef CONFIG_KALLSYMS
4606 unsigned long offset, size;
4607 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4609 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4610 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4612 seq_printf(m, " [%s]", modname);
4616 seq_printf(m, "%p", (void *)address);
4619 static int leaks_show(struct seq_file *m, void *p)
4621 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4623 struct kmem_list3 *l3;
4625 unsigned long *n = m->private;
4629 if (!(cachep->flags & SLAB_STORE_USER))
4631 if (!(cachep->flags & SLAB_RED_ZONE))
4634 /* OK, we can do it */
4638 for_each_online_node(node) {
4639 l3 = cachep->nodelists[node];
4644 spin_lock_irq(&l3->list_lock);
4646 list_for_each_entry(slabp, &l3->slabs_full, list)
4647 handle_slab(n, cachep, slabp);
4648 list_for_each_entry(slabp, &l3->slabs_partial, list)
4649 handle_slab(n, cachep, slabp);
4650 spin_unlock_irq(&l3->list_lock);
4652 name = cachep->name;
4654 /* Increase the buffer size */
4655 mutex_unlock(&slab_mutex);
4656 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4658 /* Too bad, we are really out */
4660 mutex_lock(&slab_mutex);
4663 *(unsigned long *)m->private = n[0] * 2;
4665 mutex_lock(&slab_mutex);
4666 /* Now make sure this entry will be retried */
4670 for (i = 0; i < n[1]; i++) {
4671 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4672 show_symbol(m, n[2*i+2]);
4679 static const struct seq_operations slabstats_op = {
4680 .start = leaks_start,
4686 static int slabstats_open(struct inode *inode, struct file *file)
4688 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4691 ret = seq_open(file, &slabstats_op);
4693 struct seq_file *m = file->private_data;
4694 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4703 static const struct file_operations proc_slabstats_operations = {
4704 .open = slabstats_open,
4706 .llseek = seq_lseek,
4707 .release = seq_release_private,
4711 static int __init slab_proc_init(void)
4713 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4714 #ifdef CONFIG_DEBUG_SLAB_LEAK
4715 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4719 module_init(slab_proc_init);
4723 * ksize - get the actual amount of memory allocated for a given object
4724 * @objp: Pointer to the object
4726 * kmalloc may internally round up allocations and return more memory
4727 * than requested. ksize() can be used to determine the actual amount of
4728 * memory allocated. The caller may use this additional memory, even though
4729 * a smaller amount of memory was initially specified with the kmalloc call.
4730 * The caller must guarantee that objp points to a valid object previously
4731 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4732 * must not be freed during the duration of the call.
4734 size_t ksize(const void *objp)
4737 if (unlikely(objp == ZERO_SIZE_PTR))
4740 return virt_to_cache(objp)->object_size;
4742 EXPORT_SYMBOL(ksize);