3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
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 <asm/cacheflush.h>
122 #include <asm/tlbflush.h>
123 #include <asm/page.h>
125 #include <trace/events/kmem.h>
128 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
129 * 0 for faster, smaller code (especially in the critical paths).
131 * STATS - 1 to collect stats for /proc/slabinfo.
132 * 0 for faster, smaller code (especially in the critical paths).
134 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
137 #ifdef CONFIG_DEBUG_SLAB
140 #define FORCED_DEBUG 1
144 #define FORCED_DEBUG 0
147 /* Shouldn't this be in a header file somewhere? */
148 #define BYTES_PER_WORD sizeof(void *)
149 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
151 #ifndef ARCH_KMALLOC_FLAGS
152 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
155 /* Legal flag mask for kmem_cache_create(). */
157 # define CREATE_MASK (SLAB_RED_ZONE | \
158 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
161 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
162 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
163 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
165 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
167 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
168 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
169 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
175 * Bufctl's are used for linking objs within a slab
178 * This implementation relies on "struct page" for locating the cache &
179 * slab an object belongs to.
180 * This allows the bufctl structure to be small (one int), but limits
181 * the number of objects a slab (not a cache) can contain when off-slab
182 * bufctls are used. The limit is the size of the largest general cache
183 * that does not use off-slab slabs.
184 * For 32bit archs with 4 kB pages, is this 56.
185 * This is not serious, as it is only for large objects, when it is unwise
186 * to have too many per slab.
187 * Note: This limit can be raised by introducing a general cache whose size
188 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
191 typedef unsigned int kmem_bufctl_t;
192 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
193 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
194 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
195 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
200 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
201 * arrange for kmem_freepages to be called via RCU. This is useful if
202 * we need to approach a kernel structure obliquely, from its address
203 * obtained without the usual locking. We can lock the structure to
204 * stabilize it and check it's still at the given address, only if we
205 * can be sure that the memory has not been meanwhile reused for some
206 * other kind of object (which our subsystem's lock might corrupt).
208 * rcu_read_lock before reading the address, then rcu_read_unlock after
209 * taking the spinlock within the structure expected at that address.
212 struct rcu_head head;
213 struct kmem_cache *cachep;
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
227 struct list_head list;
228 unsigned long colouroff;
229 void *s_mem; /* including colour offset */
230 unsigned int inuse; /* num of objs active in slab */
232 unsigned short nodeid;
234 struct slab_rcu __slab_cover_slab_rcu;
242 * - LIFO ordering, to hand out cache-warm objects from _alloc
243 * - reduce the number of linked list operations
244 * - reduce spinlock operations
246 * The limit is stored in the per-cpu structure to reduce the data cache
253 unsigned int batchcount;
254 unsigned int touched;
257 * Must have this definition in here for the proper
258 * alignment of array_cache. Also simplifies accessing
264 * bootstrap: The caches do not work without cpuarrays anymore, but the
265 * cpuarrays are allocated from the generic caches...
267 #define BOOT_CPUCACHE_ENTRIES 1
268 struct arraycache_init {
269 struct array_cache cache;
270 void *entries[BOOT_CPUCACHE_ENTRIES];
274 * The slab lists for all objects.
277 struct list_head slabs_partial; /* partial list first, better asm code */
278 struct list_head slabs_full;
279 struct list_head slabs_free;
280 unsigned long free_objects;
281 unsigned int free_limit;
282 unsigned int colour_next; /* Per-node cache coloring */
283 spinlock_t list_lock;
284 struct array_cache *shared; /* shared per node */
285 struct array_cache **alien; /* on other nodes */
286 unsigned long next_reap; /* updated without locking */
287 int free_touched; /* updated without locking */
291 * Need this for bootstrapping a per node allocator.
293 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
294 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
295 #define CACHE_CACHE 0
296 #define SIZE_AC MAX_NUMNODES
297 #define SIZE_L3 (2 * MAX_NUMNODES)
299 static int drain_freelist(struct kmem_cache *cache,
300 struct kmem_list3 *l3, int tofree);
301 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
303 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
304 static void cache_reap(struct work_struct *unused);
307 * This function must be completely optimized away if a constant is passed to
308 * it. Mostly the same as what is in linux/slab.h except it returns an index.
310 static __always_inline int index_of(const size_t size)
312 extern void __bad_size(void);
314 if (__builtin_constant_p(size)) {
322 #include <linux/kmalloc_sizes.h>
330 static int slab_early_init = 1;
332 #define INDEX_AC index_of(sizeof(struct arraycache_init))
333 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
335 static void kmem_list3_init(struct kmem_list3 *parent)
337 INIT_LIST_HEAD(&parent->slabs_full);
338 INIT_LIST_HEAD(&parent->slabs_partial);
339 INIT_LIST_HEAD(&parent->slabs_free);
340 parent->shared = NULL;
341 parent->alien = NULL;
342 parent->colour_next = 0;
343 spin_lock_init(&parent->list_lock);
344 parent->free_objects = 0;
345 parent->free_touched = 0;
348 #define MAKE_LIST(cachep, listp, slab, nodeid) \
350 INIT_LIST_HEAD(listp); \
351 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
354 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
356 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
358 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
361 #define CFLGS_OFF_SLAB (0x80000000UL)
362 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
364 #define BATCHREFILL_LIMIT 16
366 * Optimization question: fewer reaps means less probability for unnessary
367 * cpucache drain/refill cycles.
369 * OTOH the cpuarrays can contain lots of objects,
370 * which could lock up otherwise freeable slabs.
372 #define REAPTIMEOUT_CPUC (2*HZ)
373 #define REAPTIMEOUT_LIST3 (4*HZ)
376 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
377 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
378 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
379 #define STATS_INC_GROWN(x) ((x)->grown++)
380 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
381 #define STATS_SET_HIGH(x) \
383 if ((x)->num_active > (x)->high_mark) \
384 (x)->high_mark = (x)->num_active; \
386 #define STATS_INC_ERR(x) ((x)->errors++)
387 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
388 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
389 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
390 #define STATS_SET_FREEABLE(x, i) \
392 if ((x)->max_freeable < i) \
393 (x)->max_freeable = i; \
395 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
396 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
397 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
398 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
400 #define STATS_INC_ACTIVE(x) do { } while (0)
401 #define STATS_DEC_ACTIVE(x) do { } while (0)
402 #define STATS_INC_ALLOCED(x) do { } while (0)
403 #define STATS_INC_GROWN(x) do { } while (0)
404 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
405 #define STATS_SET_HIGH(x) do { } while (0)
406 #define STATS_INC_ERR(x) do { } while (0)
407 #define STATS_INC_NODEALLOCS(x) do { } while (0)
408 #define STATS_INC_NODEFREES(x) do { } while (0)
409 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
410 #define STATS_SET_FREEABLE(x, i) do { } while (0)
411 #define STATS_INC_ALLOCHIT(x) do { } while (0)
412 #define STATS_INC_ALLOCMISS(x) do { } while (0)
413 #define STATS_INC_FREEHIT(x) do { } while (0)
414 #define STATS_INC_FREEMISS(x) do { } while (0)
420 * memory layout of objects:
422 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
423 * the end of an object is aligned with the end of the real
424 * allocation. Catches writes behind the end of the allocation.
425 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
427 * cachep->obj_offset: The real object.
428 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
429 * cachep->size - 1* BYTES_PER_WORD: last caller address
430 * [BYTES_PER_WORD long]
432 static int obj_offset(struct kmem_cache *cachep)
434 return cachep->obj_offset;
437 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
439 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
440 return (unsigned long long*) (objp + obj_offset(cachep) -
441 sizeof(unsigned long long));
444 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
446 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
447 if (cachep->flags & SLAB_STORE_USER)
448 return (unsigned long long *)(objp + cachep->size -
449 sizeof(unsigned long long) -
451 return (unsigned long long *) (objp + cachep->size -
452 sizeof(unsigned long long));
455 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
457 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
458 return (void **)(objp + cachep->size - BYTES_PER_WORD);
463 #define obj_offset(x) 0
464 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
466 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470 #ifdef CONFIG_TRACING
471 size_t slab_buffer_size(struct kmem_cache *cachep)
475 EXPORT_SYMBOL(slab_buffer_size);
479 * Do not go above this order unless 0 objects fit into the slab or
480 * overridden on the command line.
482 #define SLAB_MAX_ORDER_HI 1
483 #define SLAB_MAX_ORDER_LO 0
484 static int slab_max_order = SLAB_MAX_ORDER_LO;
485 static bool slab_max_order_set __initdata;
487 static inline struct kmem_cache *page_get_cache(struct page *page)
489 page = compound_head(page);
490 BUG_ON(!PageSlab(page));
491 return page->slab_cache;
494 static inline struct kmem_cache *virt_to_cache(const void *obj)
496 struct page *page = virt_to_head_page(obj);
497 return page->slab_cache;
500 static inline struct slab *virt_to_slab(const void *obj)
502 struct page *page = virt_to_head_page(obj);
504 VM_BUG_ON(!PageSlab(page));
505 return page->slab_page;
508 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
511 return slab->s_mem + cache->size * idx;
515 * We want to avoid an expensive divide : (offset / cache->size)
516 * Using the fact that size is a constant for a particular cache,
517 * we can replace (offset / cache->size) by
518 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
520 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
521 const struct slab *slab, void *obj)
523 u32 offset = (obj - slab->s_mem);
524 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
528 * These are the default caches for kmalloc. Custom caches can have other sizes.
530 struct cache_sizes malloc_sizes[] = {
531 #define CACHE(x) { .cs_size = (x) },
532 #include <linux/kmalloc_sizes.h>
536 EXPORT_SYMBOL(malloc_sizes);
538 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
544 static struct cache_names __initdata cache_names[] = {
545 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
546 #include <linux/kmalloc_sizes.h>
551 static struct arraycache_init initarray_cache __initdata =
552 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
553 static struct arraycache_init initarray_generic =
554 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
556 /* internal cache of cache description objs */
557 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
558 static struct kmem_cache cache_cache = {
559 .nodelists = cache_cache_nodelists,
561 .limit = BOOT_CPUCACHE_ENTRIES,
563 .size = sizeof(struct kmem_cache),
564 .name = "kmem_cache",
567 #define BAD_ALIEN_MAGIC 0x01020304ul
569 #ifdef CONFIG_LOCKDEP
572 * Slab sometimes uses the kmalloc slabs to store the slab headers
573 * for other slabs "off slab".
574 * The locking for this is tricky in that it nests within the locks
575 * of all other slabs in a few places; to deal with this special
576 * locking we put on-slab caches into a separate lock-class.
578 * We set lock class for alien array caches which are up during init.
579 * The lock annotation will be lost if all cpus of a node goes down and
580 * then comes back up during hotplug
582 static struct lock_class_key on_slab_l3_key;
583 static struct lock_class_key on_slab_alc_key;
585 static struct lock_class_key debugobj_l3_key;
586 static struct lock_class_key debugobj_alc_key;
588 static void slab_set_lock_classes(struct kmem_cache *cachep,
589 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
592 struct array_cache **alc;
593 struct kmem_list3 *l3;
596 l3 = cachep->nodelists[q];
600 lockdep_set_class(&l3->list_lock, l3_key);
603 * FIXME: This check for BAD_ALIEN_MAGIC
604 * should go away when common slab code is taught to
605 * work even without alien caches.
606 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
607 * for alloc_alien_cache,
609 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
613 lockdep_set_class(&alc[r]->lock, alc_key);
617 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
619 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
622 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
626 for_each_online_node(node)
627 slab_set_debugobj_lock_classes_node(cachep, node);
630 static void init_node_lock_keys(int q)
632 struct cache_sizes *s = malloc_sizes;
637 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
638 struct kmem_list3 *l3;
640 l3 = s->cs_cachep->nodelists[q];
641 if (!l3 || OFF_SLAB(s->cs_cachep))
644 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
645 &on_slab_alc_key, q);
649 static inline void init_lock_keys(void)
654 init_node_lock_keys(node);
657 static void init_node_lock_keys(int q)
661 static inline void init_lock_keys(void)
665 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
669 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
675 * Guard access to the cache-chain.
677 static DEFINE_MUTEX(cache_chain_mutex);
678 static struct list_head cache_chain;
680 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
682 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
684 return cachep->array[smp_processor_id()];
687 static inline struct kmem_cache *__find_general_cachep(size_t size,
690 struct cache_sizes *csizep = malloc_sizes;
693 /* This happens if someone tries to call
694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
697 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
700 return ZERO_SIZE_PTR;
702 while (size > csizep->cs_size)
706 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
707 * has cs_{dma,}cachep==NULL. Thus no special case
708 * for large kmalloc calls required.
710 #ifdef CONFIG_ZONE_DMA
711 if (unlikely(gfpflags & GFP_DMA))
712 return csizep->cs_dmacachep;
714 return csizep->cs_cachep;
717 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
719 return __find_general_cachep(size, gfpflags);
722 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
724 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
728 * Calculate the number of objects and left-over bytes for a given buffer size.
730 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
731 size_t align, int flags, size_t *left_over,
736 size_t slab_size = PAGE_SIZE << gfporder;
739 * The slab management structure can be either off the slab or
740 * on it. For the latter case, the memory allocated for a
744 * - One kmem_bufctl_t for each object
745 * - Padding to respect alignment of @align
746 * - @buffer_size bytes for each object
748 * If the slab management structure is off the slab, then the
749 * alignment will already be calculated into the size. Because
750 * the slabs are all pages aligned, the objects will be at the
751 * correct alignment when allocated.
753 if (flags & CFLGS_OFF_SLAB) {
755 nr_objs = slab_size / buffer_size;
757 if (nr_objs > SLAB_LIMIT)
758 nr_objs = SLAB_LIMIT;
761 * Ignore padding for the initial guess. The padding
762 * is at most @align-1 bytes, and @buffer_size is at
763 * least @align. In the worst case, this result will
764 * be one greater than the number of objects that fit
765 * into the memory allocation when taking the padding
768 nr_objs = (slab_size - sizeof(struct slab)) /
769 (buffer_size + sizeof(kmem_bufctl_t));
772 * This calculated number will be either the right
773 * amount, or one greater than what we want.
775 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
782 mgmt_size = slab_mgmt_size(nr_objs, align);
785 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
788 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
790 static void __slab_error(const char *function, struct kmem_cache *cachep,
793 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
794 function, cachep->name, msg);
799 * By default on NUMA we use alien caches to stage the freeing of
800 * objects allocated from other nodes. This causes massive memory
801 * inefficiencies when using fake NUMA setup to split memory into a
802 * large number of small nodes, so it can be disabled on the command
806 static int use_alien_caches __read_mostly = 1;
807 static int __init noaliencache_setup(char *s)
809 use_alien_caches = 0;
812 __setup("noaliencache", noaliencache_setup);
814 static int __init slab_max_order_setup(char *str)
816 get_option(&str, &slab_max_order);
817 slab_max_order = slab_max_order < 0 ? 0 :
818 min(slab_max_order, MAX_ORDER - 1);
819 slab_max_order_set = true;
823 __setup("slab_max_order=", slab_max_order_setup);
827 * Special reaping functions for NUMA systems called from cache_reap().
828 * These take care of doing round robin flushing of alien caches (containing
829 * objects freed on different nodes from which they were allocated) and the
830 * flushing of remote pcps by calling drain_node_pages.
832 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
834 static void init_reap_node(int cpu)
838 node = next_node(cpu_to_mem(cpu), node_online_map);
839 if (node == MAX_NUMNODES)
840 node = first_node(node_online_map);
842 per_cpu(slab_reap_node, cpu) = node;
845 static void next_reap_node(void)
847 int node = __this_cpu_read(slab_reap_node);
849 node = next_node(node, node_online_map);
850 if (unlikely(node >= MAX_NUMNODES))
851 node = first_node(node_online_map);
852 __this_cpu_write(slab_reap_node, node);
856 #define init_reap_node(cpu) do { } while (0)
857 #define next_reap_node(void) do { } while (0)
861 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
862 * via the workqueue/eventd.
863 * Add the CPU number into the expiration time to minimize the possibility of
864 * the CPUs getting into lockstep and contending for the global cache chain
867 static void __cpuinit start_cpu_timer(int cpu)
869 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
872 * When this gets called from do_initcalls via cpucache_init(),
873 * init_workqueues() has already run, so keventd will be setup
876 if (keventd_up() && reap_work->work.func == NULL) {
878 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
879 schedule_delayed_work_on(cpu, reap_work,
880 __round_jiffies_relative(HZ, cpu));
884 static struct array_cache *alloc_arraycache(int node, int entries,
885 int batchcount, gfp_t gfp)
887 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
888 struct array_cache *nc = NULL;
890 nc = kmalloc_node(memsize, gfp, node);
892 * The array_cache structures contain pointers to free object.
893 * However, when such objects are allocated or transferred to another
894 * cache the pointers are not cleared and they could be counted as
895 * valid references during a kmemleak scan. Therefore, kmemleak must
896 * not scan such objects.
898 kmemleak_no_scan(nc);
902 nc->batchcount = batchcount;
904 spin_lock_init(&nc->lock);
910 * Transfer objects in one arraycache to another.
911 * Locking must be handled by the caller.
913 * Return the number of entries transferred.
915 static int transfer_objects(struct array_cache *to,
916 struct array_cache *from, unsigned int max)
918 /* Figure out how many entries to transfer */
919 int nr = min3(from->avail, max, to->limit - to->avail);
924 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
934 #define drain_alien_cache(cachep, alien) do { } while (0)
935 #define reap_alien(cachep, l3) do { } while (0)
937 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
939 return (struct array_cache **)BAD_ALIEN_MAGIC;
942 static inline void free_alien_cache(struct array_cache **ac_ptr)
946 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
951 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
957 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
958 gfp_t flags, int nodeid)
963 #else /* CONFIG_NUMA */
965 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
966 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
968 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
970 struct array_cache **ac_ptr;
971 int memsize = sizeof(void *) * nr_node_ids;
976 ac_ptr = kzalloc_node(memsize, gfp, node);
979 if (i == node || !node_online(i))
981 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
983 for (i--; i >= 0; i--)
993 static void free_alien_cache(struct array_cache **ac_ptr)
1004 static void __drain_alien_cache(struct kmem_cache *cachep,
1005 struct array_cache *ac, int node)
1007 struct kmem_list3 *rl3 = cachep->nodelists[node];
1010 spin_lock(&rl3->list_lock);
1012 * Stuff objects into the remote nodes shared array first.
1013 * That way we could avoid the overhead of putting the objects
1014 * into the free lists and getting them back later.
1017 transfer_objects(rl3->shared, ac, ac->limit);
1019 free_block(cachep, ac->entry, ac->avail, node);
1021 spin_unlock(&rl3->list_lock);
1026 * Called from cache_reap() to regularly drain alien caches round robin.
1028 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1030 int node = __this_cpu_read(slab_reap_node);
1033 struct array_cache *ac = l3->alien[node];
1035 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1036 __drain_alien_cache(cachep, ac, node);
1037 spin_unlock_irq(&ac->lock);
1042 static void drain_alien_cache(struct kmem_cache *cachep,
1043 struct array_cache **alien)
1046 struct array_cache *ac;
1047 unsigned long flags;
1049 for_each_online_node(i) {
1052 spin_lock_irqsave(&ac->lock, flags);
1053 __drain_alien_cache(cachep, ac, i);
1054 spin_unlock_irqrestore(&ac->lock, flags);
1059 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1061 struct slab *slabp = virt_to_slab(objp);
1062 int nodeid = slabp->nodeid;
1063 struct kmem_list3 *l3;
1064 struct array_cache *alien = NULL;
1067 node = numa_mem_id();
1070 * Make sure we are not freeing a object from another node to the array
1071 * cache on this cpu.
1073 if (likely(slabp->nodeid == node))
1076 l3 = cachep->nodelists[node];
1077 STATS_INC_NODEFREES(cachep);
1078 if (l3->alien && l3->alien[nodeid]) {
1079 alien = l3->alien[nodeid];
1080 spin_lock(&alien->lock);
1081 if (unlikely(alien->avail == alien->limit)) {
1082 STATS_INC_ACOVERFLOW(cachep);
1083 __drain_alien_cache(cachep, alien, nodeid);
1085 alien->entry[alien->avail++] = objp;
1086 spin_unlock(&alien->lock);
1088 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1089 free_block(cachep, &objp, 1, nodeid);
1090 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1097 * Allocates and initializes nodelists for a node on each slab cache, used for
1098 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1099 * will be allocated off-node since memory is not yet online for the new node.
1100 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1103 * Must hold cache_chain_mutex.
1105 static int init_cache_nodelists_node(int node)
1107 struct kmem_cache *cachep;
1108 struct kmem_list3 *l3;
1109 const int memsize = sizeof(struct kmem_list3);
1111 list_for_each_entry(cachep, &cache_chain, list) {
1113 * Set up the size64 kmemlist for cpu before we can
1114 * begin anything. Make sure some other cpu on this
1115 * node has not already allocated this
1117 if (!cachep->nodelists[node]) {
1118 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1121 kmem_list3_init(l3);
1122 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1123 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1126 * The l3s don't come and go as CPUs come and
1127 * go. cache_chain_mutex is sufficient
1130 cachep->nodelists[node] = l3;
1133 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1134 cachep->nodelists[node]->free_limit =
1135 (1 + nr_cpus_node(node)) *
1136 cachep->batchcount + cachep->num;
1137 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1142 static void __cpuinit cpuup_canceled(long cpu)
1144 struct kmem_cache *cachep;
1145 struct kmem_list3 *l3 = NULL;
1146 int node = cpu_to_mem(cpu);
1147 const struct cpumask *mask = cpumask_of_node(node);
1149 list_for_each_entry(cachep, &cache_chain, list) {
1150 struct array_cache *nc;
1151 struct array_cache *shared;
1152 struct array_cache **alien;
1154 /* cpu is dead; no one can alloc from it. */
1155 nc = cachep->array[cpu];
1156 cachep->array[cpu] = NULL;
1157 l3 = cachep->nodelists[node];
1160 goto free_array_cache;
1162 spin_lock_irq(&l3->list_lock);
1164 /* Free limit for this kmem_list3 */
1165 l3->free_limit -= cachep->batchcount;
1167 free_block(cachep, nc->entry, nc->avail, node);
1169 if (!cpumask_empty(mask)) {
1170 spin_unlock_irq(&l3->list_lock);
1171 goto free_array_cache;
1174 shared = l3->shared;
1176 free_block(cachep, shared->entry,
1177 shared->avail, node);
1184 spin_unlock_irq(&l3->list_lock);
1188 drain_alien_cache(cachep, alien);
1189 free_alien_cache(alien);
1195 * In the previous loop, all the objects were freed to
1196 * the respective cache's slabs, now we can go ahead and
1197 * shrink each nodelist to its limit.
1199 list_for_each_entry(cachep, &cache_chain, list) {
1200 l3 = cachep->nodelists[node];
1203 drain_freelist(cachep, l3, l3->free_objects);
1207 static int __cpuinit cpuup_prepare(long cpu)
1209 struct kmem_cache *cachep;
1210 struct kmem_list3 *l3 = NULL;
1211 int node = cpu_to_mem(cpu);
1215 * We need to do this right in the beginning since
1216 * alloc_arraycache's are going to use this list.
1217 * kmalloc_node allows us to add the slab to the right
1218 * kmem_list3 and not this cpu's kmem_list3
1220 err = init_cache_nodelists_node(node);
1225 * Now we can go ahead with allocating the shared arrays and
1228 list_for_each_entry(cachep, &cache_chain, list) {
1229 struct array_cache *nc;
1230 struct array_cache *shared = NULL;
1231 struct array_cache **alien = NULL;
1233 nc = alloc_arraycache(node, cachep->limit,
1234 cachep->batchcount, GFP_KERNEL);
1237 if (cachep->shared) {
1238 shared = alloc_arraycache(node,
1239 cachep->shared * cachep->batchcount,
1240 0xbaadf00d, GFP_KERNEL);
1246 if (use_alien_caches) {
1247 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1254 cachep->array[cpu] = nc;
1255 l3 = cachep->nodelists[node];
1258 spin_lock_irq(&l3->list_lock);
1261 * We are serialised from CPU_DEAD or
1262 * CPU_UP_CANCELLED by the cpucontrol lock
1264 l3->shared = shared;
1273 spin_unlock_irq(&l3->list_lock);
1275 free_alien_cache(alien);
1276 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1277 slab_set_debugobj_lock_classes_node(cachep, node);
1279 init_node_lock_keys(node);
1283 cpuup_canceled(cpu);
1287 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1288 unsigned long action, void *hcpu)
1290 long cpu = (long)hcpu;
1294 case CPU_UP_PREPARE:
1295 case CPU_UP_PREPARE_FROZEN:
1296 mutex_lock(&cache_chain_mutex);
1297 err = cpuup_prepare(cpu);
1298 mutex_unlock(&cache_chain_mutex);
1301 case CPU_ONLINE_FROZEN:
1302 start_cpu_timer(cpu);
1304 #ifdef CONFIG_HOTPLUG_CPU
1305 case CPU_DOWN_PREPARE:
1306 case CPU_DOWN_PREPARE_FROZEN:
1308 * Shutdown cache reaper. Note that the cache_chain_mutex is
1309 * held so that if cache_reap() is invoked it cannot do
1310 * anything expensive but will only modify reap_work
1311 * and reschedule the timer.
1313 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1314 /* Now the cache_reaper is guaranteed to be not running. */
1315 per_cpu(slab_reap_work, cpu).work.func = NULL;
1317 case CPU_DOWN_FAILED:
1318 case CPU_DOWN_FAILED_FROZEN:
1319 start_cpu_timer(cpu);
1322 case CPU_DEAD_FROZEN:
1324 * Even if all the cpus of a node are down, we don't free the
1325 * kmem_list3 of any cache. This to avoid a race between
1326 * cpu_down, and a kmalloc allocation from another cpu for
1327 * memory from the node of the cpu going down. The list3
1328 * structure is usually allocated from kmem_cache_create() and
1329 * gets destroyed at kmem_cache_destroy().
1333 case CPU_UP_CANCELED:
1334 case CPU_UP_CANCELED_FROZEN:
1335 mutex_lock(&cache_chain_mutex);
1336 cpuup_canceled(cpu);
1337 mutex_unlock(&cache_chain_mutex);
1340 return notifier_from_errno(err);
1343 static struct notifier_block __cpuinitdata cpucache_notifier = {
1344 &cpuup_callback, NULL, 0
1347 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1349 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1350 * Returns -EBUSY if all objects cannot be drained so that the node is not
1353 * Must hold cache_chain_mutex.
1355 static int __meminit drain_cache_nodelists_node(int node)
1357 struct kmem_cache *cachep;
1360 list_for_each_entry(cachep, &cache_chain, list) {
1361 struct kmem_list3 *l3;
1363 l3 = cachep->nodelists[node];
1367 drain_freelist(cachep, l3, l3->free_objects);
1369 if (!list_empty(&l3->slabs_full) ||
1370 !list_empty(&l3->slabs_partial)) {
1378 static int __meminit slab_memory_callback(struct notifier_block *self,
1379 unsigned long action, void *arg)
1381 struct memory_notify *mnb = arg;
1385 nid = mnb->status_change_nid;
1390 case MEM_GOING_ONLINE:
1391 mutex_lock(&cache_chain_mutex);
1392 ret = init_cache_nodelists_node(nid);
1393 mutex_unlock(&cache_chain_mutex);
1395 case MEM_GOING_OFFLINE:
1396 mutex_lock(&cache_chain_mutex);
1397 ret = drain_cache_nodelists_node(nid);
1398 mutex_unlock(&cache_chain_mutex);
1402 case MEM_CANCEL_ONLINE:
1403 case MEM_CANCEL_OFFLINE:
1407 return notifier_from_errno(ret);
1409 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1412 * swap the static kmem_list3 with kmalloced memory
1414 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1417 struct kmem_list3 *ptr;
1419 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1422 memcpy(ptr, list, sizeof(struct kmem_list3));
1424 * Do not assume that spinlocks can be initialized via memcpy:
1426 spin_lock_init(&ptr->list_lock);
1428 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1429 cachep->nodelists[nodeid] = ptr;
1433 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1434 * size of kmem_list3.
1436 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1440 for_each_online_node(node) {
1441 cachep->nodelists[node] = &initkmem_list3[index + node];
1442 cachep->nodelists[node]->next_reap = jiffies +
1444 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1449 * Initialisation. Called after the page allocator have been initialised and
1450 * before smp_init().
1452 void __init kmem_cache_init(void)
1455 struct cache_sizes *sizes;
1456 struct cache_names *names;
1461 if (num_possible_nodes() == 1)
1462 use_alien_caches = 0;
1464 for (i = 0; i < NUM_INIT_LISTS; i++) {
1465 kmem_list3_init(&initkmem_list3[i]);
1466 if (i < MAX_NUMNODES)
1467 cache_cache.nodelists[i] = NULL;
1469 set_up_list3s(&cache_cache, CACHE_CACHE);
1472 * Fragmentation resistance on low memory - only use bigger
1473 * page orders on machines with more than 32MB of memory if
1474 * not overridden on the command line.
1476 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1477 slab_max_order = SLAB_MAX_ORDER_HI;
1479 /* Bootstrap is tricky, because several objects are allocated
1480 * from caches that do not exist yet:
1481 * 1) initialize the cache_cache cache: it contains the struct
1482 * kmem_cache structures of all caches, except cache_cache itself:
1483 * cache_cache is statically allocated.
1484 * Initially an __init data area is used for the head array and the
1485 * kmem_list3 structures, it's replaced with a kmalloc allocated
1486 * array at the end of the bootstrap.
1487 * 2) Create the first kmalloc cache.
1488 * The struct kmem_cache for the new cache is allocated normally.
1489 * An __init data area is used for the head array.
1490 * 3) Create the remaining kmalloc caches, with minimally sized
1492 * 4) Replace the __init data head arrays for cache_cache and the first
1493 * kmalloc cache with kmalloc allocated arrays.
1494 * 5) Replace the __init data for kmem_list3 for cache_cache and
1495 * the other cache's with kmalloc allocated memory.
1496 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1499 node = numa_mem_id();
1501 /* 1) create the cache_cache */
1502 INIT_LIST_HEAD(&cache_chain);
1503 list_add(&cache_cache.list, &cache_chain);
1504 cache_cache.colour_off = cache_line_size();
1505 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1506 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1509 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1511 cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1512 nr_node_ids * sizeof(struct kmem_list3 *);
1513 cache_cache.object_size = cache_cache.size;
1514 cache_cache.size = ALIGN(cache_cache.size,
1516 cache_cache.reciprocal_buffer_size =
1517 reciprocal_value(cache_cache.size);
1519 for (order = 0; order < MAX_ORDER; order++) {
1520 cache_estimate(order, cache_cache.size,
1521 cache_line_size(), 0, &left_over, &cache_cache.num);
1522 if (cache_cache.num)
1525 BUG_ON(!cache_cache.num);
1526 cache_cache.gfporder = order;
1527 cache_cache.colour = left_over / cache_cache.colour_off;
1528 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1529 sizeof(struct slab), cache_line_size());
1531 /* 2+3) create the kmalloc caches */
1532 sizes = malloc_sizes;
1533 names = cache_names;
1536 * Initialize the caches that provide memory for the array cache and the
1537 * kmem_list3 structures first. Without this, further allocations will
1541 sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name,
1542 sizes[INDEX_AC].cs_size,
1543 ARCH_KMALLOC_MINALIGN,
1544 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1547 if (INDEX_AC != INDEX_L3) {
1548 sizes[INDEX_L3].cs_cachep =
1549 __kmem_cache_create(names[INDEX_L3].name,
1550 sizes[INDEX_L3].cs_size,
1551 ARCH_KMALLOC_MINALIGN,
1552 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1556 slab_early_init = 0;
1558 while (sizes->cs_size != ULONG_MAX) {
1560 * For performance, all the general caches are L1 aligned.
1561 * This should be particularly beneficial on SMP boxes, as it
1562 * eliminates "false sharing".
1563 * Note for systems short on memory removing the alignment will
1564 * allow tighter packing of the smaller caches.
1566 if (!sizes->cs_cachep) {
1567 sizes->cs_cachep = __kmem_cache_create(names->name,
1569 ARCH_KMALLOC_MINALIGN,
1570 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1573 #ifdef CONFIG_ZONE_DMA
1574 sizes->cs_dmacachep = __kmem_cache_create(
1577 ARCH_KMALLOC_MINALIGN,
1578 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1585 /* 4) Replace the bootstrap head arrays */
1587 struct array_cache *ptr;
1589 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1591 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1592 memcpy(ptr, cpu_cache_get(&cache_cache),
1593 sizeof(struct arraycache_init));
1595 * Do not assume that spinlocks can be initialized via memcpy:
1597 spin_lock_init(&ptr->lock);
1599 cache_cache.array[smp_processor_id()] = ptr;
1601 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1603 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1604 != &initarray_generic.cache);
1605 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1606 sizeof(struct arraycache_init));
1608 * Do not assume that spinlocks can be initialized via memcpy:
1610 spin_lock_init(&ptr->lock);
1612 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1615 /* 5) Replace the bootstrap kmem_list3's */
1619 for_each_online_node(nid) {
1620 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1622 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1623 &initkmem_list3[SIZE_AC + nid], nid);
1625 if (INDEX_AC != INDEX_L3) {
1626 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1627 &initkmem_list3[SIZE_L3 + nid], nid);
1635 void __init kmem_cache_init_late(void)
1637 struct kmem_cache *cachep;
1641 /* Annotate slab for lockdep -- annotate the malloc caches */
1644 /* 6) resize the head arrays to their final sizes */
1645 mutex_lock(&cache_chain_mutex);
1646 list_for_each_entry(cachep, &cache_chain, list)
1647 if (enable_cpucache(cachep, GFP_NOWAIT))
1649 mutex_unlock(&cache_chain_mutex);
1655 * Register a cpu startup notifier callback that initializes
1656 * cpu_cache_get for all new cpus
1658 register_cpu_notifier(&cpucache_notifier);
1662 * Register a memory hotplug callback that initializes and frees
1665 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1669 * The reap timers are started later, with a module init call: That part
1670 * of the kernel is not yet operational.
1674 static int __init cpucache_init(void)
1679 * Register the timers that return unneeded pages to the page allocator
1681 for_each_online_cpu(cpu)
1682 start_cpu_timer(cpu);
1688 __initcall(cpucache_init);
1690 static noinline void
1691 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1693 struct kmem_list3 *l3;
1695 unsigned long flags;
1699 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1701 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1702 cachep->name, cachep->size, cachep->gfporder);
1704 for_each_online_node(node) {
1705 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1706 unsigned long active_slabs = 0, num_slabs = 0;
1708 l3 = cachep->nodelists[node];
1712 spin_lock_irqsave(&l3->list_lock, flags);
1713 list_for_each_entry(slabp, &l3->slabs_full, list) {
1714 active_objs += cachep->num;
1717 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1718 active_objs += slabp->inuse;
1721 list_for_each_entry(slabp, &l3->slabs_free, list)
1724 free_objects += l3->free_objects;
1725 spin_unlock_irqrestore(&l3->list_lock, flags);
1727 num_slabs += active_slabs;
1728 num_objs = num_slabs * cachep->num;
1730 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1731 node, active_slabs, num_slabs, active_objs, num_objs,
1737 * Interface to system's page allocator. No need to hold the cache-lock.
1739 * If we requested dmaable memory, we will get it. Even if we
1740 * did not request dmaable memory, we might get it, but that
1741 * would be relatively rare and ignorable.
1743 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1751 * Nommu uses slab's for process anonymous memory allocations, and thus
1752 * requires __GFP_COMP to properly refcount higher order allocations
1754 flags |= __GFP_COMP;
1757 flags |= cachep->allocflags;
1758 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1759 flags |= __GFP_RECLAIMABLE;
1761 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1763 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1764 slab_out_of_memory(cachep, flags, nodeid);
1768 nr_pages = (1 << cachep->gfporder);
1769 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1770 add_zone_page_state(page_zone(page),
1771 NR_SLAB_RECLAIMABLE, nr_pages);
1773 add_zone_page_state(page_zone(page),
1774 NR_SLAB_UNRECLAIMABLE, nr_pages);
1775 for (i = 0; i < nr_pages; i++)
1776 __SetPageSlab(page + i);
1778 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1779 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1782 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1784 kmemcheck_mark_unallocated_pages(page, nr_pages);
1787 return page_address(page);
1791 * Interface to system's page release.
1793 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1795 unsigned long i = (1 << cachep->gfporder);
1796 struct page *page = virt_to_page(addr);
1797 const unsigned long nr_freed = i;
1799 kmemcheck_free_shadow(page, cachep->gfporder);
1801 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1802 sub_zone_page_state(page_zone(page),
1803 NR_SLAB_RECLAIMABLE, nr_freed);
1805 sub_zone_page_state(page_zone(page),
1806 NR_SLAB_UNRECLAIMABLE, nr_freed);
1808 BUG_ON(!PageSlab(page));
1809 __ClearPageSlab(page);
1812 if (current->reclaim_state)
1813 current->reclaim_state->reclaimed_slab += nr_freed;
1814 free_pages((unsigned long)addr, cachep->gfporder);
1817 static void kmem_rcu_free(struct rcu_head *head)
1819 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1820 struct kmem_cache *cachep = slab_rcu->cachep;
1822 kmem_freepages(cachep, slab_rcu->addr);
1823 if (OFF_SLAB(cachep))
1824 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1829 #ifdef CONFIG_DEBUG_PAGEALLOC
1830 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1831 unsigned long caller)
1833 int size = cachep->object_size;
1835 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1837 if (size < 5 * sizeof(unsigned long))
1840 *addr++ = 0x12345678;
1842 *addr++ = smp_processor_id();
1843 size -= 3 * sizeof(unsigned long);
1845 unsigned long *sptr = &caller;
1846 unsigned long svalue;
1848 while (!kstack_end(sptr)) {
1850 if (kernel_text_address(svalue)) {
1852 size -= sizeof(unsigned long);
1853 if (size <= sizeof(unsigned long))
1859 *addr++ = 0x87654321;
1863 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1865 int size = cachep->object_size;
1866 addr = &((char *)addr)[obj_offset(cachep)];
1868 memset(addr, val, size);
1869 *(unsigned char *)(addr + size - 1) = POISON_END;
1872 static void dump_line(char *data, int offset, int limit)
1875 unsigned char error = 0;
1878 printk(KERN_ERR "%03x: ", offset);
1879 for (i = 0; i < limit; i++) {
1880 if (data[offset + i] != POISON_FREE) {
1881 error = data[offset + i];
1885 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1886 &data[offset], limit, 1);
1888 if (bad_count == 1) {
1889 error ^= POISON_FREE;
1890 if (!(error & (error - 1))) {
1891 printk(KERN_ERR "Single bit error detected. Probably "
1894 printk(KERN_ERR "Run memtest86+ or a similar memory "
1897 printk(KERN_ERR "Run a memory test tool.\n");
1906 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1911 if (cachep->flags & SLAB_RED_ZONE) {
1912 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1913 *dbg_redzone1(cachep, objp),
1914 *dbg_redzone2(cachep, objp));
1917 if (cachep->flags & SLAB_STORE_USER) {
1918 printk(KERN_ERR "Last user: [<%p>]",
1919 *dbg_userword(cachep, objp));
1920 print_symbol("(%s)",
1921 (unsigned long)*dbg_userword(cachep, objp));
1924 realobj = (char *)objp + obj_offset(cachep);
1925 size = cachep->object_size;
1926 for (i = 0; i < size && lines; i += 16, lines--) {
1929 if (i + limit > size)
1931 dump_line(realobj, i, limit);
1935 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1941 realobj = (char *)objp + obj_offset(cachep);
1942 size = cachep->object_size;
1944 for (i = 0; i < size; i++) {
1945 char exp = POISON_FREE;
1948 if (realobj[i] != exp) {
1954 "Slab corruption (%s): %s start=%p, len=%d\n",
1955 print_tainted(), cachep->name, realobj, size);
1956 print_objinfo(cachep, objp, 0);
1958 /* Hexdump the affected line */
1961 if (i + limit > size)
1963 dump_line(realobj, i, limit);
1966 /* Limit to 5 lines */
1972 /* Print some data about the neighboring objects, if they
1975 struct slab *slabp = virt_to_slab(objp);
1978 objnr = obj_to_index(cachep, slabp, objp);
1980 objp = index_to_obj(cachep, slabp, objnr - 1);
1981 realobj = (char *)objp + obj_offset(cachep);
1982 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1984 print_objinfo(cachep, objp, 2);
1986 if (objnr + 1 < cachep->num) {
1987 objp = index_to_obj(cachep, slabp, objnr + 1);
1988 realobj = (char *)objp + obj_offset(cachep);
1989 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1991 print_objinfo(cachep, objp, 2);
1998 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2001 for (i = 0; i < cachep->num; i++) {
2002 void *objp = index_to_obj(cachep, slabp, i);
2004 if (cachep->flags & SLAB_POISON) {
2005 #ifdef CONFIG_DEBUG_PAGEALLOC
2006 if (cachep->size % PAGE_SIZE == 0 &&
2008 kernel_map_pages(virt_to_page(objp),
2009 cachep->size / PAGE_SIZE, 1);
2011 check_poison_obj(cachep, objp);
2013 check_poison_obj(cachep, objp);
2016 if (cachep->flags & SLAB_RED_ZONE) {
2017 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2018 slab_error(cachep, "start of a freed object "
2020 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2021 slab_error(cachep, "end of a freed object "
2027 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2033 * slab_destroy - destroy and release all objects in a slab
2034 * @cachep: cache pointer being destroyed
2035 * @slabp: slab pointer being destroyed
2037 * Destroy all the objs in a slab, and release the mem back to the system.
2038 * Before calling the slab must have been unlinked from the cache. The
2039 * cache-lock is not held/needed.
2041 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2043 void *addr = slabp->s_mem - slabp->colouroff;
2045 slab_destroy_debugcheck(cachep, slabp);
2046 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2047 struct slab_rcu *slab_rcu;
2049 slab_rcu = (struct slab_rcu *)slabp;
2050 slab_rcu->cachep = cachep;
2051 slab_rcu->addr = addr;
2052 call_rcu(&slab_rcu->head, kmem_rcu_free);
2054 kmem_freepages(cachep, addr);
2055 if (OFF_SLAB(cachep))
2056 kmem_cache_free(cachep->slabp_cache, slabp);
2060 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2063 struct kmem_list3 *l3;
2065 for_each_online_cpu(i)
2066 kfree(cachep->array[i]);
2068 /* NUMA: free the list3 structures */
2069 for_each_online_node(i) {
2070 l3 = cachep->nodelists[i];
2073 free_alien_cache(l3->alien);
2077 kmem_cache_free(&cache_cache, cachep);
2082 * calculate_slab_order - calculate size (page order) of slabs
2083 * @cachep: pointer to the cache that is being created
2084 * @size: size of objects to be created in this cache.
2085 * @align: required alignment for the objects.
2086 * @flags: slab allocation flags
2088 * Also calculates the number of objects per slab.
2090 * This could be made much more intelligent. For now, try to avoid using
2091 * high order pages for slabs. When the gfp() functions are more friendly
2092 * towards high-order requests, this should be changed.
2094 static size_t calculate_slab_order(struct kmem_cache *cachep,
2095 size_t size, size_t align, unsigned long flags)
2097 unsigned long offslab_limit;
2098 size_t left_over = 0;
2101 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2105 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2109 if (flags & CFLGS_OFF_SLAB) {
2111 * Max number of objs-per-slab for caches which
2112 * use off-slab slabs. Needed to avoid a possible
2113 * looping condition in cache_grow().
2115 offslab_limit = size - sizeof(struct slab);
2116 offslab_limit /= sizeof(kmem_bufctl_t);
2118 if (num > offslab_limit)
2122 /* Found something acceptable - save it away */
2124 cachep->gfporder = gfporder;
2125 left_over = remainder;
2128 * A VFS-reclaimable slab tends to have most allocations
2129 * as GFP_NOFS and we really don't want to have to be allocating
2130 * higher-order pages when we are unable to shrink dcache.
2132 if (flags & SLAB_RECLAIM_ACCOUNT)
2136 * Large number of objects is good, but very large slabs are
2137 * currently bad for the gfp()s.
2139 if (gfporder >= slab_max_order)
2143 * Acceptable internal fragmentation?
2145 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2151 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2153 if (slab_state >= FULL)
2154 return enable_cpucache(cachep, gfp);
2156 if (slab_state == DOWN) {
2158 * Note: the first kmem_cache_create must create the cache
2159 * that's used by kmalloc(24), otherwise the creation of
2160 * further caches will BUG().
2162 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2165 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2166 * the first cache, then we need to set up all its list3s,
2167 * otherwise the creation of further caches will BUG().
2169 set_up_list3s(cachep, SIZE_AC);
2170 if (INDEX_AC == INDEX_L3)
2171 slab_state = PARTIAL_L3;
2173 slab_state = PARTIAL_ARRAYCACHE;
2175 cachep->array[smp_processor_id()] =
2176 kmalloc(sizeof(struct arraycache_init), gfp);
2178 if (slab_state == PARTIAL_ARRAYCACHE) {
2179 set_up_list3s(cachep, SIZE_L3);
2180 slab_state = PARTIAL_L3;
2183 for_each_online_node(node) {
2184 cachep->nodelists[node] =
2185 kmalloc_node(sizeof(struct kmem_list3),
2187 BUG_ON(!cachep->nodelists[node]);
2188 kmem_list3_init(cachep->nodelists[node]);
2192 cachep->nodelists[numa_mem_id()]->next_reap =
2193 jiffies + REAPTIMEOUT_LIST3 +
2194 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2196 cpu_cache_get(cachep)->avail = 0;
2197 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2198 cpu_cache_get(cachep)->batchcount = 1;
2199 cpu_cache_get(cachep)->touched = 0;
2200 cachep->batchcount = 1;
2201 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2206 * __kmem_cache_create - Create a cache.
2207 * @name: A string which is used in /proc/slabinfo to identify this cache.
2208 * @size: The size of objects to be created in this cache.
2209 * @align: The required alignment for the objects.
2210 * @flags: SLAB flags
2211 * @ctor: A constructor for the objects.
2213 * Returns a ptr to the cache on success, NULL on failure.
2214 * Cannot be called within a int, but can be interrupted.
2215 * The @ctor is run when new pages are allocated by the cache.
2217 * @name must be valid until the cache is destroyed. This implies that
2218 * the module calling this has to destroy the cache before getting unloaded.
2222 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2223 * to catch references to uninitialised memory.
2225 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2226 * for buffer overruns.
2228 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2229 * cacheline. This can be beneficial if you're counting cycles as closely
2233 __kmem_cache_create (const char *name, size_t size, size_t align,
2234 unsigned long flags, void (*ctor)(void *))
2236 size_t left_over, slab_size, ralign;
2237 struct kmem_cache *cachep = NULL, *pc;
2241 * Sanity checks... these are all serious usage bugs.
2243 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2244 size > KMALLOC_MAX_SIZE) {
2245 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2251 * We use cache_chain_mutex to ensure a consistent view of
2252 * cpu_online_mask as well. Please see cpuup_callback
2254 if (slab_is_available()) {
2256 mutex_lock(&cache_chain_mutex);
2259 list_for_each_entry(pc, &cache_chain, list) {
2264 * This happens when the module gets unloaded and doesn't
2265 * destroy its slab cache and no-one else reuses the vmalloc
2266 * area of the module. Print a warning.
2268 res = probe_kernel_address(pc->name, tmp);
2271 "SLAB: cache with size %d has lost its name\n",
2276 if (!strcmp(pc->name, name)) {
2278 "kmem_cache_create: duplicate cache %s\n", name);
2285 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2288 * Enable redzoning and last user accounting, except for caches with
2289 * large objects, if the increased size would increase the object size
2290 * above the next power of two: caches with object sizes just above a
2291 * power of two have a significant amount of internal fragmentation.
2293 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2294 2 * sizeof(unsigned long long)))
2295 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2296 if (!(flags & SLAB_DESTROY_BY_RCU))
2297 flags |= SLAB_POISON;
2299 if (flags & SLAB_DESTROY_BY_RCU)
2300 BUG_ON(flags & SLAB_POISON);
2303 * Always checks flags, a caller might be expecting debug support which
2306 BUG_ON(flags & ~CREATE_MASK);
2309 * Check that size is in terms of words. This is needed to avoid
2310 * unaligned accesses for some archs when redzoning is used, and makes
2311 * sure any on-slab bufctl's are also correctly aligned.
2313 if (size & (BYTES_PER_WORD - 1)) {
2314 size += (BYTES_PER_WORD - 1);
2315 size &= ~(BYTES_PER_WORD - 1);
2318 /* calculate the final buffer alignment: */
2320 /* 1) arch recommendation: can be overridden for debug */
2321 if (flags & SLAB_HWCACHE_ALIGN) {
2323 * Default alignment: as specified by the arch code. Except if
2324 * an object is really small, then squeeze multiple objects into
2327 ralign = cache_line_size();
2328 while (size <= ralign / 2)
2331 ralign = BYTES_PER_WORD;
2335 * Redzoning and user store require word alignment or possibly larger.
2336 * Note this will be overridden by architecture or caller mandated
2337 * alignment if either is greater than BYTES_PER_WORD.
2339 if (flags & SLAB_STORE_USER)
2340 ralign = BYTES_PER_WORD;
2342 if (flags & SLAB_RED_ZONE) {
2343 ralign = REDZONE_ALIGN;
2344 /* If redzoning, ensure that the second redzone is suitably
2345 * aligned, by adjusting the object size accordingly. */
2346 size += REDZONE_ALIGN - 1;
2347 size &= ~(REDZONE_ALIGN - 1);
2350 /* 2) arch mandated alignment */
2351 if (ralign < ARCH_SLAB_MINALIGN) {
2352 ralign = ARCH_SLAB_MINALIGN;
2354 /* 3) caller mandated alignment */
2355 if (ralign < align) {
2358 /* disable debug if necessary */
2359 if (ralign > __alignof__(unsigned long long))
2360 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2366 if (slab_is_available())
2371 /* Get cache's description obj. */
2372 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2376 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2377 cachep->object_size = size;
2378 cachep->align = align;
2382 * Both debugging options require word-alignment which is calculated
2385 if (flags & SLAB_RED_ZONE) {
2386 /* add space for red zone words */
2387 cachep->obj_offset += sizeof(unsigned long long);
2388 size += 2 * sizeof(unsigned long long);
2390 if (flags & SLAB_STORE_USER) {
2391 /* user store requires one word storage behind the end of
2392 * the real object. But if the second red zone needs to be
2393 * aligned to 64 bits, we must allow that much space.
2395 if (flags & SLAB_RED_ZONE)
2396 size += REDZONE_ALIGN;
2398 size += BYTES_PER_WORD;
2400 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2401 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2402 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2403 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2410 * Determine if the slab management is 'on' or 'off' slab.
2411 * (bootstrapping cannot cope with offslab caches so don't do
2412 * it too early on. Always use on-slab management when
2413 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2415 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2416 !(flags & SLAB_NOLEAKTRACE))
2418 * Size is large, assume best to place the slab management obj
2419 * off-slab (should allow better packing of objs).
2421 flags |= CFLGS_OFF_SLAB;
2423 size = ALIGN(size, align);
2425 left_over = calculate_slab_order(cachep, size, align, flags);
2429 "kmem_cache_create: couldn't create cache %s.\n", name);
2430 kmem_cache_free(&cache_cache, cachep);
2433 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2434 + sizeof(struct slab), align);
2437 * If the slab has been placed off-slab, and we have enough space then
2438 * move it on-slab. This is at the expense of any extra colouring.
2440 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2441 flags &= ~CFLGS_OFF_SLAB;
2442 left_over -= slab_size;
2445 if (flags & CFLGS_OFF_SLAB) {
2446 /* really off slab. No need for manual alignment */
2448 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2450 #ifdef CONFIG_PAGE_POISONING
2451 /* If we're going to use the generic kernel_map_pages()
2452 * poisoning, then it's going to smash the contents of
2453 * the redzone and userword anyhow, so switch them off.
2455 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2456 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2460 cachep->colour_off = cache_line_size();
2461 /* Offset must be a multiple of the alignment. */
2462 if (cachep->colour_off < align)
2463 cachep->colour_off = align;
2464 cachep->colour = left_over / cachep->colour_off;
2465 cachep->slab_size = slab_size;
2466 cachep->flags = flags;
2467 cachep->allocflags = 0;
2468 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2469 cachep->allocflags |= GFP_DMA;
2470 cachep->size = size;
2471 cachep->reciprocal_buffer_size = reciprocal_value(size);
2473 if (flags & CFLGS_OFF_SLAB) {
2474 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2476 * This is a possibility for one of the malloc_sizes caches.
2477 * But since we go off slab only for object size greater than
2478 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2479 * this should not happen at all.
2480 * But leave a BUG_ON for some lucky dude.
2482 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2484 cachep->ctor = ctor;
2485 cachep->name = name;
2487 if (setup_cpu_cache(cachep, gfp)) {
2488 __kmem_cache_destroy(cachep);
2492 if (flags & SLAB_DEBUG_OBJECTS) {
2494 * Would deadlock through slab_destroy()->call_rcu()->
2495 * debug_object_activate()->kmem_cache_alloc().
2497 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2499 slab_set_debugobj_lock_classes(cachep);
2502 /* cache setup completed, link it into the list */
2503 list_add(&cachep->list, &cache_chain);
2505 if (slab_is_available()) {
2506 mutex_unlock(&cache_chain_mutex);
2513 static void check_irq_off(void)
2515 BUG_ON(!irqs_disabled());
2518 static void check_irq_on(void)
2520 BUG_ON(irqs_disabled());
2523 static void check_spinlock_acquired(struct kmem_cache *cachep)
2527 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2531 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2535 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2540 #define check_irq_off() do { } while(0)
2541 #define check_irq_on() do { } while(0)
2542 #define check_spinlock_acquired(x) do { } while(0)
2543 #define check_spinlock_acquired_node(x, y) do { } while(0)
2546 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2547 struct array_cache *ac,
2548 int force, int node);
2550 static void do_drain(void *arg)
2552 struct kmem_cache *cachep = arg;
2553 struct array_cache *ac;
2554 int node = numa_mem_id();
2557 ac = cpu_cache_get(cachep);
2558 spin_lock(&cachep->nodelists[node]->list_lock);
2559 free_block(cachep, ac->entry, ac->avail, node);
2560 spin_unlock(&cachep->nodelists[node]->list_lock);
2564 static void drain_cpu_caches(struct kmem_cache *cachep)
2566 struct kmem_list3 *l3;
2569 on_each_cpu(do_drain, cachep, 1);
2571 for_each_online_node(node) {
2572 l3 = cachep->nodelists[node];
2573 if (l3 && l3->alien)
2574 drain_alien_cache(cachep, l3->alien);
2577 for_each_online_node(node) {
2578 l3 = cachep->nodelists[node];
2580 drain_array(cachep, l3, l3->shared, 1, node);
2585 * Remove slabs from the list of free slabs.
2586 * Specify the number of slabs to drain in tofree.
2588 * Returns the actual number of slabs released.
2590 static int drain_freelist(struct kmem_cache *cache,
2591 struct kmem_list3 *l3, int tofree)
2593 struct list_head *p;
2598 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2600 spin_lock_irq(&l3->list_lock);
2601 p = l3->slabs_free.prev;
2602 if (p == &l3->slabs_free) {
2603 spin_unlock_irq(&l3->list_lock);
2607 slabp = list_entry(p, struct slab, list);
2609 BUG_ON(slabp->inuse);
2611 list_del(&slabp->list);
2613 * Safe to drop the lock. The slab is no longer linked
2616 l3->free_objects -= cache->num;
2617 spin_unlock_irq(&l3->list_lock);
2618 slab_destroy(cache, slabp);
2625 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2626 static int __cache_shrink(struct kmem_cache *cachep)
2629 struct kmem_list3 *l3;
2631 drain_cpu_caches(cachep);
2634 for_each_online_node(i) {
2635 l3 = cachep->nodelists[i];
2639 drain_freelist(cachep, l3, l3->free_objects);
2641 ret += !list_empty(&l3->slabs_full) ||
2642 !list_empty(&l3->slabs_partial);
2644 return (ret ? 1 : 0);
2648 * kmem_cache_shrink - Shrink a cache.
2649 * @cachep: The cache to shrink.
2651 * Releases as many slabs as possible for a cache.
2652 * To help debugging, a zero exit status indicates all slabs were released.
2654 int kmem_cache_shrink(struct kmem_cache *cachep)
2657 BUG_ON(!cachep || in_interrupt());
2660 mutex_lock(&cache_chain_mutex);
2661 ret = __cache_shrink(cachep);
2662 mutex_unlock(&cache_chain_mutex);
2666 EXPORT_SYMBOL(kmem_cache_shrink);
2669 * kmem_cache_destroy - delete a cache
2670 * @cachep: the cache to destroy
2672 * Remove a &struct kmem_cache object from the slab cache.
2674 * It is expected this function will be called by a module when it is
2675 * unloaded. This will remove the cache completely, and avoid a duplicate
2676 * cache being allocated each time a module is loaded and unloaded, if the
2677 * module doesn't have persistent in-kernel storage across loads and unloads.
2679 * The cache must be empty before calling this function.
2681 * The caller must guarantee that no one will allocate memory from the cache
2682 * during the kmem_cache_destroy().
2684 void kmem_cache_destroy(struct kmem_cache *cachep)
2686 BUG_ON(!cachep || in_interrupt());
2688 /* Find the cache in the chain of caches. */
2690 mutex_lock(&cache_chain_mutex);
2692 * the chain is never empty, cache_cache is never destroyed
2694 list_del(&cachep->list);
2695 if (__cache_shrink(cachep)) {
2696 slab_error(cachep, "Can't free all objects");
2697 list_add(&cachep->list, &cache_chain);
2698 mutex_unlock(&cache_chain_mutex);
2703 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2706 __kmem_cache_destroy(cachep);
2707 mutex_unlock(&cache_chain_mutex);
2710 EXPORT_SYMBOL(kmem_cache_destroy);
2713 * Get the memory for a slab management obj.
2714 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2715 * always come from malloc_sizes caches. The slab descriptor cannot
2716 * come from the same cache which is getting created because,
2717 * when we are searching for an appropriate cache for these
2718 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2719 * If we are creating a malloc_sizes cache here it would not be visible to
2720 * kmem_find_general_cachep till the initialization is complete.
2721 * Hence we cannot have slabp_cache same as the original cache.
2723 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2724 int colour_off, gfp_t local_flags,
2729 if (OFF_SLAB(cachep)) {
2730 /* Slab management obj is off-slab. */
2731 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2732 local_flags, nodeid);
2734 * If the first object in the slab is leaked (it's allocated
2735 * but no one has a reference to it), we want to make sure
2736 * kmemleak does not treat the ->s_mem pointer as a reference
2737 * to the object. Otherwise we will not report the leak.
2739 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2744 slabp = objp + colour_off;
2745 colour_off += cachep->slab_size;
2748 slabp->colouroff = colour_off;
2749 slabp->s_mem = objp + colour_off;
2750 slabp->nodeid = nodeid;
2755 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2757 return (kmem_bufctl_t *) (slabp + 1);
2760 static void cache_init_objs(struct kmem_cache *cachep,
2765 for (i = 0; i < cachep->num; i++) {
2766 void *objp = index_to_obj(cachep, slabp, i);
2768 /* need to poison the objs? */
2769 if (cachep->flags & SLAB_POISON)
2770 poison_obj(cachep, objp, POISON_FREE);
2771 if (cachep->flags & SLAB_STORE_USER)
2772 *dbg_userword(cachep, objp) = NULL;
2774 if (cachep->flags & SLAB_RED_ZONE) {
2775 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2776 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2779 * Constructors are not allowed to allocate memory from the same
2780 * cache which they are a constructor for. Otherwise, deadlock.
2781 * They must also be threaded.
2783 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2784 cachep->ctor(objp + obj_offset(cachep));
2786 if (cachep->flags & SLAB_RED_ZONE) {
2787 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2788 slab_error(cachep, "constructor overwrote the"
2789 " end of an object");
2790 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2791 slab_error(cachep, "constructor overwrote the"
2792 " start of an object");
2794 if ((cachep->size % PAGE_SIZE) == 0 &&
2795 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2796 kernel_map_pages(virt_to_page(objp),
2797 cachep->size / PAGE_SIZE, 0);
2802 slab_bufctl(slabp)[i] = i + 1;
2804 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2807 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2809 if (CONFIG_ZONE_DMA_FLAG) {
2810 if (flags & GFP_DMA)
2811 BUG_ON(!(cachep->allocflags & GFP_DMA));
2813 BUG_ON(cachep->allocflags & GFP_DMA);
2817 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2820 void *objp = index_to_obj(cachep, slabp, slabp->free);
2824 next = slab_bufctl(slabp)[slabp->free];
2826 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2827 WARN_ON(slabp->nodeid != nodeid);
2834 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2835 void *objp, int nodeid)
2837 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2840 /* Verify that the slab belongs to the intended node */
2841 WARN_ON(slabp->nodeid != nodeid);
2843 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2844 printk(KERN_ERR "slab: double free detected in cache "
2845 "'%s', objp %p\n", cachep->name, objp);
2849 slab_bufctl(slabp)[objnr] = slabp->free;
2850 slabp->free = objnr;
2855 * Map pages beginning at addr to the given cache and slab. This is required
2856 * for the slab allocator to be able to lookup the cache and slab of a
2857 * virtual address for kfree, ksize, and slab debugging.
2859 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2865 page = virt_to_page(addr);
2868 if (likely(!PageCompound(page)))
2869 nr_pages <<= cache->gfporder;
2872 page->slab_cache = cache;
2873 page->slab_page = slab;
2875 } while (--nr_pages);
2879 * Grow (by 1) the number of slabs within a cache. This is called by
2880 * kmem_cache_alloc() when there are no active objs left in a cache.
2882 static int cache_grow(struct kmem_cache *cachep,
2883 gfp_t flags, int nodeid, void *objp)
2888 struct kmem_list3 *l3;
2891 * Be lazy and only check for valid flags here, keeping it out of the
2892 * critical path in kmem_cache_alloc().
2894 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2895 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2897 /* Take the l3 list lock to change the colour_next on this node */
2899 l3 = cachep->nodelists[nodeid];
2900 spin_lock(&l3->list_lock);
2902 /* Get colour for the slab, and cal the next value. */
2903 offset = l3->colour_next;
2905 if (l3->colour_next >= cachep->colour)
2906 l3->colour_next = 0;
2907 spin_unlock(&l3->list_lock);
2909 offset *= cachep->colour_off;
2911 if (local_flags & __GFP_WAIT)
2915 * The test for missing atomic flag is performed here, rather than
2916 * the more obvious place, simply to reduce the critical path length
2917 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2918 * will eventually be caught here (where it matters).
2920 kmem_flagcheck(cachep, flags);
2923 * Get mem for the objs. Attempt to allocate a physical page from
2927 objp = kmem_getpages(cachep, local_flags, nodeid);
2931 /* Get slab management. */
2932 slabp = alloc_slabmgmt(cachep, objp, offset,
2933 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2937 slab_map_pages(cachep, slabp, objp);
2939 cache_init_objs(cachep, slabp);
2941 if (local_flags & __GFP_WAIT)
2942 local_irq_disable();
2944 spin_lock(&l3->list_lock);
2946 /* Make slab active. */
2947 list_add_tail(&slabp->list, &(l3->slabs_free));
2948 STATS_INC_GROWN(cachep);
2949 l3->free_objects += cachep->num;
2950 spin_unlock(&l3->list_lock);
2953 kmem_freepages(cachep, objp);
2955 if (local_flags & __GFP_WAIT)
2956 local_irq_disable();
2963 * Perform extra freeing checks:
2964 * - detect bad pointers.
2965 * - POISON/RED_ZONE checking
2967 static void kfree_debugcheck(const void *objp)
2969 if (!virt_addr_valid(objp)) {
2970 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2971 (unsigned long)objp);
2976 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2978 unsigned long long redzone1, redzone2;
2980 redzone1 = *dbg_redzone1(cache, obj);
2981 redzone2 = *dbg_redzone2(cache, obj);
2986 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2989 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2990 slab_error(cache, "double free detected");
2992 slab_error(cache, "memory outside object was overwritten");
2994 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2995 obj, redzone1, redzone2);
2998 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3005 BUG_ON(virt_to_cache(objp) != cachep);
3007 objp -= obj_offset(cachep);
3008 kfree_debugcheck(objp);
3009 page = virt_to_head_page(objp);
3011 slabp = page->slab_page;
3013 if (cachep->flags & SLAB_RED_ZONE) {
3014 verify_redzone_free(cachep, objp);
3015 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3016 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3018 if (cachep->flags & SLAB_STORE_USER)
3019 *dbg_userword(cachep, objp) = caller;
3021 objnr = obj_to_index(cachep, slabp, objp);
3023 BUG_ON(objnr >= cachep->num);
3024 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3026 #ifdef CONFIG_DEBUG_SLAB_LEAK
3027 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3029 if (cachep->flags & SLAB_POISON) {
3030 #ifdef CONFIG_DEBUG_PAGEALLOC
3031 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3032 store_stackinfo(cachep, objp, (unsigned long)caller);
3033 kernel_map_pages(virt_to_page(objp),
3034 cachep->size / PAGE_SIZE, 0);
3036 poison_obj(cachep, objp, POISON_FREE);
3039 poison_obj(cachep, objp, POISON_FREE);
3045 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3050 /* Check slab's freelist to see if this obj is there. */
3051 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3053 if (entries > cachep->num || i >= cachep->num)
3056 if (entries != cachep->num - slabp->inuse) {
3058 printk(KERN_ERR "slab: Internal list corruption detected in "
3059 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3060 cachep->name, cachep->num, slabp, slabp->inuse,
3062 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3063 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3069 #define kfree_debugcheck(x) do { } while(0)
3070 #define cache_free_debugcheck(x,objp,z) (objp)
3071 #define check_slabp(x,y) do { } while(0)
3074 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3077 struct kmem_list3 *l3;
3078 struct array_cache *ac;
3083 node = numa_mem_id();
3084 ac = cpu_cache_get(cachep);
3085 batchcount = ac->batchcount;
3086 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3088 * If there was little recent activity on this cache, then
3089 * perform only a partial refill. Otherwise we could generate
3092 batchcount = BATCHREFILL_LIMIT;
3094 l3 = cachep->nodelists[node];
3096 BUG_ON(ac->avail > 0 || !l3);
3097 spin_lock(&l3->list_lock);
3099 /* See if we can refill from the shared array */
3100 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3101 l3->shared->touched = 1;
3105 while (batchcount > 0) {
3106 struct list_head *entry;
3108 /* Get slab alloc is to come from. */
3109 entry = l3->slabs_partial.next;
3110 if (entry == &l3->slabs_partial) {
3111 l3->free_touched = 1;
3112 entry = l3->slabs_free.next;
3113 if (entry == &l3->slabs_free)
3117 slabp = list_entry(entry, struct slab, list);
3118 check_slabp(cachep, slabp);
3119 check_spinlock_acquired(cachep);
3122 * The slab was either on partial or free list so
3123 * there must be at least one object available for
3126 BUG_ON(slabp->inuse >= cachep->num);
3128 while (slabp->inuse < cachep->num && batchcount--) {
3129 STATS_INC_ALLOCED(cachep);
3130 STATS_INC_ACTIVE(cachep);
3131 STATS_SET_HIGH(cachep);
3133 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3136 check_slabp(cachep, slabp);
3138 /* move slabp to correct slabp list: */
3139 list_del(&slabp->list);
3140 if (slabp->free == BUFCTL_END)
3141 list_add(&slabp->list, &l3->slabs_full);
3143 list_add(&slabp->list, &l3->slabs_partial);
3147 l3->free_objects -= ac->avail;
3149 spin_unlock(&l3->list_lock);
3151 if (unlikely(!ac->avail)) {
3153 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3155 /* cache_grow can reenable interrupts, then ac could change. */
3156 ac = cpu_cache_get(cachep);
3157 if (!x && ac->avail == 0) /* no objects in sight? abort */
3160 if (!ac->avail) /* objects refilled by interrupt? */
3164 return ac->entry[--ac->avail];
3167 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3170 might_sleep_if(flags & __GFP_WAIT);
3172 kmem_flagcheck(cachep, flags);
3177 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3178 gfp_t flags, void *objp, void *caller)
3182 if (cachep->flags & SLAB_POISON) {
3183 #ifdef CONFIG_DEBUG_PAGEALLOC
3184 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3185 kernel_map_pages(virt_to_page(objp),
3186 cachep->size / PAGE_SIZE, 1);
3188 check_poison_obj(cachep, objp);
3190 check_poison_obj(cachep, objp);
3192 poison_obj(cachep, objp, POISON_INUSE);
3194 if (cachep->flags & SLAB_STORE_USER)
3195 *dbg_userword(cachep, objp) = caller;
3197 if (cachep->flags & SLAB_RED_ZONE) {
3198 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3199 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3200 slab_error(cachep, "double free, or memory outside"
3201 " object was overwritten");
3203 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3204 objp, *dbg_redzone1(cachep, objp),
3205 *dbg_redzone2(cachep, objp));
3207 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3208 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3210 #ifdef CONFIG_DEBUG_SLAB_LEAK
3215 slabp = virt_to_head_page(objp)->slab_page;
3216 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3217 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3220 objp += obj_offset(cachep);
3221 if (cachep->ctor && cachep->flags & SLAB_POISON)
3223 if (ARCH_SLAB_MINALIGN &&
3224 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3225 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3226 objp, (int)ARCH_SLAB_MINALIGN);
3231 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3234 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3236 if (cachep == &cache_cache)
3239 return should_failslab(cachep->object_size, flags, cachep->flags);
3242 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3245 struct array_cache *ac;
3249 ac = cpu_cache_get(cachep);
3250 if (likely(ac->avail)) {
3251 STATS_INC_ALLOCHIT(cachep);
3253 objp = ac->entry[--ac->avail];
3255 STATS_INC_ALLOCMISS(cachep);
3256 objp = cache_alloc_refill(cachep, flags);
3258 * the 'ac' may be updated by cache_alloc_refill(),
3259 * and kmemleak_erase() requires its correct value.
3261 ac = cpu_cache_get(cachep);
3264 * To avoid a false negative, if an object that is in one of the
3265 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3266 * treat the array pointers as a reference to the object.
3269 kmemleak_erase(&ac->entry[ac->avail]);
3275 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3277 * If we are in_interrupt, then process context, including cpusets and
3278 * mempolicy, may not apply and should not be used for allocation policy.
3280 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3282 int nid_alloc, nid_here;
3284 if (in_interrupt() || (flags & __GFP_THISNODE))
3286 nid_alloc = nid_here = numa_mem_id();
3287 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3288 nid_alloc = cpuset_slab_spread_node();
3289 else if (current->mempolicy)
3290 nid_alloc = slab_node();
3291 if (nid_alloc != nid_here)
3292 return ____cache_alloc_node(cachep, flags, nid_alloc);
3297 * Fallback function if there was no memory available and no objects on a
3298 * certain node and fall back is permitted. First we scan all the
3299 * available nodelists for available objects. If that fails then we
3300 * perform an allocation without specifying a node. This allows the page
3301 * allocator to do its reclaim / fallback magic. We then insert the
3302 * slab into the proper nodelist and then allocate from it.
3304 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3306 struct zonelist *zonelist;
3310 enum zone_type high_zoneidx = gfp_zone(flags);
3313 unsigned int cpuset_mems_cookie;
3315 if (flags & __GFP_THISNODE)
3318 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3321 cpuset_mems_cookie = get_mems_allowed();
3322 zonelist = node_zonelist(slab_node(), flags);
3326 * Look through allowed nodes for objects available
3327 * from existing per node queues.
3329 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3330 nid = zone_to_nid(zone);
3332 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3333 cache->nodelists[nid] &&
3334 cache->nodelists[nid]->free_objects) {
3335 obj = ____cache_alloc_node(cache,
3336 flags | GFP_THISNODE, nid);
3344 * This allocation will be performed within the constraints
3345 * of the current cpuset / memory policy requirements.
3346 * We may trigger various forms of reclaim on the allowed
3347 * set and go into memory reserves if necessary.
3349 if (local_flags & __GFP_WAIT)
3351 kmem_flagcheck(cache, flags);
3352 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3353 if (local_flags & __GFP_WAIT)
3354 local_irq_disable();
3357 * Insert into the appropriate per node queues
3359 nid = page_to_nid(virt_to_page(obj));
3360 if (cache_grow(cache, flags, nid, obj)) {
3361 obj = ____cache_alloc_node(cache,
3362 flags | GFP_THISNODE, nid);
3365 * Another processor may allocate the
3366 * objects in the slab since we are
3367 * not holding any locks.
3371 /* cache_grow already freed obj */
3377 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3383 * A interface to enable slab creation on nodeid
3385 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3388 struct list_head *entry;
3390 struct kmem_list3 *l3;
3394 l3 = cachep->nodelists[nodeid];
3399 spin_lock(&l3->list_lock);
3400 entry = l3->slabs_partial.next;
3401 if (entry == &l3->slabs_partial) {
3402 l3->free_touched = 1;
3403 entry = l3->slabs_free.next;
3404 if (entry == &l3->slabs_free)
3408 slabp = list_entry(entry, struct slab, list);
3409 check_spinlock_acquired_node(cachep, nodeid);
3410 check_slabp(cachep, slabp);
3412 STATS_INC_NODEALLOCS(cachep);
3413 STATS_INC_ACTIVE(cachep);
3414 STATS_SET_HIGH(cachep);
3416 BUG_ON(slabp->inuse == cachep->num);
3418 obj = slab_get_obj(cachep, slabp, nodeid);
3419 check_slabp(cachep, slabp);
3421 /* move slabp to correct slabp list: */
3422 list_del(&slabp->list);
3424 if (slabp->free == BUFCTL_END)
3425 list_add(&slabp->list, &l3->slabs_full);
3427 list_add(&slabp->list, &l3->slabs_partial);
3429 spin_unlock(&l3->list_lock);
3433 spin_unlock(&l3->list_lock);
3434 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3438 return fallback_alloc(cachep, flags);
3445 * kmem_cache_alloc_node - Allocate an object on the specified node
3446 * @cachep: The cache to allocate from.
3447 * @flags: See kmalloc().
3448 * @nodeid: node number of the target node.
3449 * @caller: return address of caller, used for debug information
3451 * Identical to kmem_cache_alloc but it will allocate memory on the given
3452 * node, which can improve the performance for cpu bound structures.
3454 * Fallback to other node is possible if __GFP_THISNODE is not set.
3456 static __always_inline void *
3457 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3460 unsigned long save_flags;
3462 int slab_node = numa_mem_id();
3464 flags &= gfp_allowed_mask;
3466 lockdep_trace_alloc(flags);
3468 if (slab_should_failslab(cachep, flags))
3471 cache_alloc_debugcheck_before(cachep, flags);
3472 local_irq_save(save_flags);
3474 if (nodeid == NUMA_NO_NODE)
3477 if (unlikely(!cachep->nodelists[nodeid])) {
3478 /* Node not bootstrapped yet */
3479 ptr = fallback_alloc(cachep, flags);
3483 if (nodeid == slab_node) {
3485 * Use the locally cached objects if possible.
3486 * However ____cache_alloc does not allow fallback
3487 * to other nodes. It may fail while we still have
3488 * objects on other nodes available.
3490 ptr = ____cache_alloc(cachep, flags);
3494 /* ___cache_alloc_node can fall back to other nodes */
3495 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3497 local_irq_restore(save_flags);
3498 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3499 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3503 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3505 if (unlikely((flags & __GFP_ZERO) && ptr))
3506 memset(ptr, 0, cachep->object_size);
3511 static __always_inline void *
3512 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3516 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3517 objp = alternate_node_alloc(cache, flags);
3521 objp = ____cache_alloc(cache, flags);
3524 * We may just have run out of memory on the local node.
3525 * ____cache_alloc_node() knows how to locate memory on other nodes
3528 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3535 static __always_inline void *
3536 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3538 return ____cache_alloc(cachep, flags);
3541 #endif /* CONFIG_NUMA */
3543 static __always_inline void *
3544 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3546 unsigned long save_flags;
3549 flags &= gfp_allowed_mask;
3551 lockdep_trace_alloc(flags);
3553 if (slab_should_failslab(cachep, flags))
3556 cache_alloc_debugcheck_before(cachep, flags);
3557 local_irq_save(save_flags);
3558 objp = __do_cache_alloc(cachep, flags);
3559 local_irq_restore(save_flags);
3560 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3561 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3566 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3568 if (unlikely((flags & __GFP_ZERO) && objp))
3569 memset(objp, 0, cachep->object_size);
3575 * Caller needs to acquire correct kmem_list's list_lock
3577 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3581 struct kmem_list3 *l3;
3583 for (i = 0; i < nr_objects; i++) {
3584 void *objp = objpp[i];
3587 slabp = virt_to_slab(objp);
3588 l3 = cachep->nodelists[node];
3589 list_del(&slabp->list);
3590 check_spinlock_acquired_node(cachep, node);
3591 check_slabp(cachep, slabp);
3592 slab_put_obj(cachep, slabp, objp, node);
3593 STATS_DEC_ACTIVE(cachep);
3595 check_slabp(cachep, slabp);
3597 /* fixup slab chains */
3598 if (slabp->inuse == 0) {
3599 if (l3->free_objects > l3->free_limit) {
3600 l3->free_objects -= cachep->num;
3601 /* No need to drop any previously held
3602 * lock here, even if we have a off-slab slab
3603 * descriptor it is guaranteed to come from
3604 * a different cache, refer to comments before
3607 slab_destroy(cachep, slabp);
3609 list_add(&slabp->list, &l3->slabs_free);
3612 /* Unconditionally move a slab to the end of the
3613 * partial list on free - maximum time for the
3614 * other objects to be freed, too.
3616 list_add_tail(&slabp->list, &l3->slabs_partial);
3621 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3624 struct kmem_list3 *l3;
3625 int node = numa_mem_id();
3627 batchcount = ac->batchcount;
3629 BUG_ON(!batchcount || batchcount > ac->avail);
3632 l3 = cachep->nodelists[node];
3633 spin_lock(&l3->list_lock);
3635 struct array_cache *shared_array = l3->shared;
3636 int max = shared_array->limit - shared_array->avail;
3638 if (batchcount > max)
3640 memcpy(&(shared_array->entry[shared_array->avail]),
3641 ac->entry, sizeof(void *) * batchcount);
3642 shared_array->avail += batchcount;
3647 free_block(cachep, ac->entry, batchcount, node);
3652 struct list_head *p;
3654 p = l3->slabs_free.next;
3655 while (p != &(l3->slabs_free)) {
3658 slabp = list_entry(p, struct slab, list);
3659 BUG_ON(slabp->inuse);
3664 STATS_SET_FREEABLE(cachep, i);
3667 spin_unlock(&l3->list_lock);
3668 ac->avail -= batchcount;
3669 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3673 * Release an obj back to its cache. If the obj has a constructed state, it must
3674 * be in this state _before_ it is released. Called with disabled ints.
3676 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3679 struct array_cache *ac = cpu_cache_get(cachep);
3682 kmemleak_free_recursive(objp, cachep->flags);
3683 objp = cache_free_debugcheck(cachep, objp, caller);
3685 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3688 * Skip calling cache_free_alien() when the platform is not numa.
3689 * This will avoid cache misses that happen while accessing slabp (which
3690 * is per page memory reference) to get nodeid. Instead use a global
3691 * variable to skip the call, which is mostly likely to be present in
3694 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3697 if (likely(ac->avail < ac->limit)) {
3698 STATS_INC_FREEHIT(cachep);
3700 STATS_INC_FREEMISS(cachep);
3701 cache_flusharray(cachep, ac);
3704 ac->entry[ac->avail++] = objp;
3708 * kmem_cache_alloc - Allocate an object
3709 * @cachep: The cache to allocate from.
3710 * @flags: See kmalloc().
3712 * Allocate an object from this cache. The flags are only relevant
3713 * if the cache has no available objects.
3715 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3717 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3719 trace_kmem_cache_alloc(_RET_IP_, ret,
3720 cachep->object_size, cachep->size, flags);
3724 EXPORT_SYMBOL(kmem_cache_alloc);
3726 #ifdef CONFIG_TRACING
3728 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3732 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3734 trace_kmalloc(_RET_IP_, ret,
3735 size, slab_buffer_size(cachep), flags);
3738 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3742 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3744 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3745 __builtin_return_address(0));
3747 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3748 cachep->object_size, cachep->size,
3753 EXPORT_SYMBOL(kmem_cache_alloc_node);
3755 #ifdef CONFIG_TRACING
3756 void *kmem_cache_alloc_node_trace(size_t size,
3757 struct kmem_cache *cachep,
3763 ret = __cache_alloc_node(cachep, flags, nodeid,
3764 __builtin_return_address(0));
3765 trace_kmalloc_node(_RET_IP_, ret,
3766 size, slab_buffer_size(cachep),
3770 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3773 static __always_inline void *
3774 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3776 struct kmem_cache *cachep;
3778 cachep = kmem_find_general_cachep(size, flags);
3779 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3781 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3784 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3785 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3787 return __do_kmalloc_node(size, flags, node,
3788 __builtin_return_address(0));
3790 EXPORT_SYMBOL(__kmalloc_node);
3792 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3793 int node, unsigned long caller)
3795 return __do_kmalloc_node(size, flags, node, (void *)caller);
3797 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3799 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3801 return __do_kmalloc_node(size, flags, node, NULL);
3803 EXPORT_SYMBOL(__kmalloc_node);
3804 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3805 #endif /* CONFIG_NUMA */
3808 * __do_kmalloc - allocate memory
3809 * @size: how many bytes of memory are required.
3810 * @flags: the type of memory to allocate (see kmalloc).
3811 * @caller: function caller for debug tracking of the caller
3813 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3816 struct kmem_cache *cachep;
3819 /* If you want to save a few bytes .text space: replace
3821 * Then kmalloc uses the uninlined functions instead of the inline
3824 cachep = __find_general_cachep(size, flags);
3825 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3827 ret = __cache_alloc(cachep, flags, caller);
3829 trace_kmalloc((unsigned long) caller, ret,
3830 size, cachep->size, flags);
3836 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3837 void *__kmalloc(size_t size, gfp_t flags)
3839 return __do_kmalloc(size, flags, __builtin_return_address(0));
3841 EXPORT_SYMBOL(__kmalloc);
3843 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3845 return __do_kmalloc(size, flags, (void *)caller);
3847 EXPORT_SYMBOL(__kmalloc_track_caller);
3850 void *__kmalloc(size_t size, gfp_t flags)
3852 return __do_kmalloc(size, flags, NULL);
3854 EXPORT_SYMBOL(__kmalloc);
3858 * kmem_cache_free - Deallocate an object
3859 * @cachep: The cache the allocation was from.
3860 * @objp: The previously allocated object.
3862 * Free an object which was previously allocated from this
3865 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3867 unsigned long flags;
3869 local_irq_save(flags);
3870 debug_check_no_locks_freed(objp, cachep->object_size);
3871 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3872 debug_check_no_obj_freed(objp, cachep->object_size);
3873 __cache_free(cachep, objp, __builtin_return_address(0));
3874 local_irq_restore(flags);
3876 trace_kmem_cache_free(_RET_IP_, objp);
3878 EXPORT_SYMBOL(kmem_cache_free);
3881 * kfree - free previously allocated memory
3882 * @objp: pointer returned by kmalloc.
3884 * If @objp is NULL, no operation is performed.
3886 * Don't free memory not originally allocated by kmalloc()
3887 * or you will run into trouble.
3889 void kfree(const void *objp)
3891 struct kmem_cache *c;
3892 unsigned long flags;
3894 trace_kfree(_RET_IP_, objp);
3896 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3898 local_irq_save(flags);
3899 kfree_debugcheck(objp);
3900 c = virt_to_cache(objp);
3901 debug_check_no_locks_freed(objp, c->object_size);
3903 debug_check_no_obj_freed(objp, c->object_size);
3904 __cache_free(c, (void *)objp, __builtin_return_address(0));
3905 local_irq_restore(flags);
3907 EXPORT_SYMBOL(kfree);
3909 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3911 return cachep->object_size;
3913 EXPORT_SYMBOL(kmem_cache_size);
3916 * This initializes kmem_list3 or resizes various caches for all nodes.
3918 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3921 struct kmem_list3 *l3;
3922 struct array_cache *new_shared;
3923 struct array_cache **new_alien = NULL;
3925 for_each_online_node(node) {
3927 if (use_alien_caches) {
3928 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3934 if (cachep->shared) {
3935 new_shared = alloc_arraycache(node,
3936 cachep->shared*cachep->batchcount,
3939 free_alien_cache(new_alien);
3944 l3 = cachep->nodelists[node];
3946 struct array_cache *shared = l3->shared;
3948 spin_lock_irq(&l3->list_lock);
3951 free_block(cachep, shared->entry,
3952 shared->avail, node);
3954 l3->shared = new_shared;
3956 l3->alien = new_alien;
3959 l3->free_limit = (1 + nr_cpus_node(node)) *
3960 cachep->batchcount + cachep->num;
3961 spin_unlock_irq(&l3->list_lock);
3963 free_alien_cache(new_alien);
3966 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3968 free_alien_cache(new_alien);
3973 kmem_list3_init(l3);
3974 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3975 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3976 l3->shared = new_shared;
3977 l3->alien = new_alien;
3978 l3->free_limit = (1 + nr_cpus_node(node)) *
3979 cachep->batchcount + cachep->num;
3980 cachep->nodelists[node] = l3;
3985 if (!cachep->list.next) {
3986 /* Cache is not active yet. Roll back what we did */
3989 if (cachep->nodelists[node]) {
3990 l3 = cachep->nodelists[node];
3993 free_alien_cache(l3->alien);
3995 cachep->nodelists[node] = NULL;
4003 struct ccupdate_struct {
4004 struct kmem_cache *cachep;
4005 struct array_cache *new[0];
4008 static void do_ccupdate_local(void *info)
4010 struct ccupdate_struct *new = info;
4011 struct array_cache *old;
4014 old = cpu_cache_get(new->cachep);
4016 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4017 new->new[smp_processor_id()] = old;
4020 /* Always called with the cache_chain_mutex held */
4021 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4022 int batchcount, int shared, gfp_t gfp)
4024 struct ccupdate_struct *new;
4027 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4032 for_each_online_cpu(i) {
4033 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4036 for (i--; i >= 0; i--)
4042 new->cachep = cachep;
4044 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4047 cachep->batchcount = batchcount;
4048 cachep->limit = limit;
4049 cachep->shared = shared;
4051 for_each_online_cpu(i) {
4052 struct array_cache *ccold = new->new[i];
4055 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4056 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4057 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4061 return alloc_kmemlist(cachep, gfp);
4064 /* Called with cache_chain_mutex held always */
4065 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4071 * The head array serves three purposes:
4072 * - create a LIFO ordering, i.e. return objects that are cache-warm
4073 * - reduce the number of spinlock operations.
4074 * - reduce the number of linked list operations on the slab and
4075 * bufctl chains: array operations are cheaper.
4076 * The numbers are guessed, we should auto-tune as described by
4079 if (cachep->size > 131072)
4081 else if (cachep->size > PAGE_SIZE)
4083 else if (cachep->size > 1024)
4085 else if (cachep->size > 256)
4091 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4092 * allocation behaviour: Most allocs on one cpu, most free operations
4093 * on another cpu. For these cases, an efficient object passing between
4094 * cpus is necessary. This is provided by a shared array. The array
4095 * replaces Bonwick's magazine layer.
4096 * On uniprocessor, it's functionally equivalent (but less efficient)
4097 * to a larger limit. Thus disabled by default.
4100 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4105 * With debugging enabled, large batchcount lead to excessively long
4106 * periods with disabled local interrupts. Limit the batchcount
4111 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4113 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4114 cachep->name, -err);
4119 * Drain an array if it contains any elements taking the l3 lock only if
4120 * necessary. Note that the l3 listlock also protects the array_cache
4121 * if drain_array() is used on the shared array.
4123 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4124 struct array_cache *ac, int force, int node)
4128 if (!ac || !ac->avail)
4130 if (ac->touched && !force) {
4133 spin_lock_irq(&l3->list_lock);
4135 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4136 if (tofree > ac->avail)
4137 tofree = (ac->avail + 1) / 2;
4138 free_block(cachep, ac->entry, tofree, node);
4139 ac->avail -= tofree;
4140 memmove(ac->entry, &(ac->entry[tofree]),
4141 sizeof(void *) * ac->avail);
4143 spin_unlock_irq(&l3->list_lock);
4148 * cache_reap - Reclaim memory from caches.
4149 * @w: work descriptor
4151 * Called from workqueue/eventd every few seconds.
4153 * - clear the per-cpu caches for this CPU.
4154 * - return freeable pages to the main free memory pool.
4156 * If we cannot acquire the cache chain mutex then just give up - we'll try
4157 * again on the next iteration.
4159 static void cache_reap(struct work_struct *w)
4161 struct kmem_cache *searchp;
4162 struct kmem_list3 *l3;
4163 int node = numa_mem_id();
4164 struct delayed_work *work = to_delayed_work(w);
4166 if (!mutex_trylock(&cache_chain_mutex))
4167 /* Give up. Setup the next iteration. */
4170 list_for_each_entry(searchp, &cache_chain, list) {
4174 * We only take the l3 lock if absolutely necessary and we
4175 * have established with reasonable certainty that
4176 * we can do some work if the lock was obtained.
4178 l3 = searchp->nodelists[node];
4180 reap_alien(searchp, l3);
4182 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4185 * These are racy checks but it does not matter
4186 * if we skip one check or scan twice.
4188 if (time_after(l3->next_reap, jiffies))
4191 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4193 drain_array(searchp, l3, l3->shared, 0, node);
4195 if (l3->free_touched)
4196 l3->free_touched = 0;
4200 freed = drain_freelist(searchp, l3, (l3->free_limit +
4201 5 * searchp->num - 1) / (5 * searchp->num));
4202 STATS_ADD_REAPED(searchp, freed);
4208 mutex_unlock(&cache_chain_mutex);
4211 /* Set up the next iteration */
4212 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4215 #ifdef CONFIG_SLABINFO
4217 static void print_slabinfo_header(struct seq_file *m)
4220 * Output format version, so at least we can change it
4221 * without _too_ many complaints.
4224 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4226 seq_puts(m, "slabinfo - version: 2.1\n");
4228 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4229 "<objperslab> <pagesperslab>");
4230 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4231 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4233 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4234 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4235 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4240 static void *s_start(struct seq_file *m, loff_t *pos)
4244 mutex_lock(&cache_chain_mutex);
4246 print_slabinfo_header(m);
4248 return seq_list_start(&cache_chain, *pos);
4251 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4253 return seq_list_next(p, &cache_chain, pos);
4256 static void s_stop(struct seq_file *m, void *p)
4258 mutex_unlock(&cache_chain_mutex);
4261 static int s_show(struct seq_file *m, void *p)
4263 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4265 unsigned long active_objs;
4266 unsigned long num_objs;
4267 unsigned long active_slabs = 0;
4268 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4272 struct kmem_list3 *l3;
4276 for_each_online_node(node) {
4277 l3 = cachep->nodelists[node];
4282 spin_lock_irq(&l3->list_lock);
4284 list_for_each_entry(slabp, &l3->slabs_full, list) {
4285 if (slabp->inuse != cachep->num && !error)
4286 error = "slabs_full accounting error";
4287 active_objs += cachep->num;
4290 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4291 if (slabp->inuse == cachep->num && !error)
4292 error = "slabs_partial inuse accounting error";
4293 if (!slabp->inuse && !error)
4294 error = "slabs_partial/inuse accounting error";
4295 active_objs += slabp->inuse;
4298 list_for_each_entry(slabp, &l3->slabs_free, list) {
4299 if (slabp->inuse && !error)
4300 error = "slabs_free/inuse accounting error";
4303 free_objects += l3->free_objects;
4305 shared_avail += l3->shared->avail;
4307 spin_unlock_irq(&l3->list_lock);
4309 num_slabs += active_slabs;
4310 num_objs = num_slabs * cachep->num;
4311 if (num_objs - active_objs != free_objects && !error)
4312 error = "free_objects accounting error";
4314 name = cachep->name;
4316 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4318 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4319 name, active_objs, num_objs, cachep->size,
4320 cachep->num, (1 << cachep->gfporder));
4321 seq_printf(m, " : tunables %4u %4u %4u",
4322 cachep->limit, cachep->batchcount, cachep->shared);
4323 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4324 active_slabs, num_slabs, shared_avail);
4327 unsigned long high = cachep->high_mark;
4328 unsigned long allocs = cachep->num_allocations;
4329 unsigned long grown = cachep->grown;
4330 unsigned long reaped = cachep->reaped;
4331 unsigned long errors = cachep->errors;
4332 unsigned long max_freeable = cachep->max_freeable;
4333 unsigned long node_allocs = cachep->node_allocs;
4334 unsigned long node_frees = cachep->node_frees;
4335 unsigned long overflows = cachep->node_overflow;
4337 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4338 "%4lu %4lu %4lu %4lu %4lu",
4339 allocs, high, grown,
4340 reaped, errors, max_freeable, node_allocs,
4341 node_frees, overflows);
4345 unsigned long allochit = atomic_read(&cachep->allochit);
4346 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4347 unsigned long freehit = atomic_read(&cachep->freehit);
4348 unsigned long freemiss = atomic_read(&cachep->freemiss);
4350 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4351 allochit, allocmiss, freehit, freemiss);
4359 * slabinfo_op - iterator that generates /proc/slabinfo
4368 * num-pages-per-slab
4369 * + further values on SMP and with statistics enabled
4372 static const struct seq_operations slabinfo_op = {
4379 #define MAX_SLABINFO_WRITE 128
4381 * slabinfo_write - Tuning for the slab allocator
4383 * @buffer: user buffer
4384 * @count: data length
4387 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4388 size_t count, loff_t *ppos)
4390 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4391 int limit, batchcount, shared, res;
4392 struct kmem_cache *cachep;
4394 if (count > MAX_SLABINFO_WRITE)
4396 if (copy_from_user(&kbuf, buffer, count))
4398 kbuf[MAX_SLABINFO_WRITE] = '\0';
4400 tmp = strchr(kbuf, ' ');
4405 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4408 /* Find the cache in the chain of caches. */
4409 mutex_lock(&cache_chain_mutex);
4411 list_for_each_entry(cachep, &cache_chain, list) {
4412 if (!strcmp(cachep->name, kbuf)) {
4413 if (limit < 1 || batchcount < 1 ||
4414 batchcount > limit || shared < 0) {
4417 res = do_tune_cpucache(cachep, limit,
4424 mutex_unlock(&cache_chain_mutex);
4430 static int slabinfo_open(struct inode *inode, struct file *file)
4432 return seq_open(file, &slabinfo_op);
4435 static const struct file_operations proc_slabinfo_operations = {
4436 .open = slabinfo_open,
4438 .write = slabinfo_write,
4439 .llseek = seq_lseek,
4440 .release = seq_release,
4443 #ifdef CONFIG_DEBUG_SLAB_LEAK
4445 static void *leaks_start(struct seq_file *m, loff_t *pos)
4447 mutex_lock(&cache_chain_mutex);
4448 return seq_list_start(&cache_chain, *pos);
4451 static inline int add_caller(unsigned long *n, unsigned long v)
4461 unsigned long *q = p + 2 * i;
4475 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4481 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4487 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4488 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4490 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4495 static void show_symbol(struct seq_file *m, unsigned long address)
4497 #ifdef CONFIG_KALLSYMS
4498 unsigned long offset, size;
4499 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4501 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4502 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4504 seq_printf(m, " [%s]", modname);
4508 seq_printf(m, "%p", (void *)address);
4511 static int leaks_show(struct seq_file *m, void *p)
4513 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4515 struct kmem_list3 *l3;
4517 unsigned long *n = m->private;
4521 if (!(cachep->flags & SLAB_STORE_USER))
4523 if (!(cachep->flags & SLAB_RED_ZONE))
4526 /* OK, we can do it */
4530 for_each_online_node(node) {
4531 l3 = cachep->nodelists[node];
4536 spin_lock_irq(&l3->list_lock);
4538 list_for_each_entry(slabp, &l3->slabs_full, list)
4539 handle_slab(n, cachep, slabp);
4540 list_for_each_entry(slabp, &l3->slabs_partial, list)
4541 handle_slab(n, cachep, slabp);
4542 spin_unlock_irq(&l3->list_lock);
4544 name = cachep->name;
4546 /* Increase the buffer size */
4547 mutex_unlock(&cache_chain_mutex);
4548 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4550 /* Too bad, we are really out */
4552 mutex_lock(&cache_chain_mutex);
4555 *(unsigned long *)m->private = n[0] * 2;
4557 mutex_lock(&cache_chain_mutex);
4558 /* Now make sure this entry will be retried */
4562 for (i = 0; i < n[1]; i++) {
4563 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4564 show_symbol(m, n[2*i+2]);
4571 static const struct seq_operations slabstats_op = {
4572 .start = leaks_start,
4578 static int slabstats_open(struct inode *inode, struct file *file)
4580 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4583 ret = seq_open(file, &slabstats_op);
4585 struct seq_file *m = file->private_data;
4586 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4595 static const struct file_operations proc_slabstats_operations = {
4596 .open = slabstats_open,
4598 .llseek = seq_lseek,
4599 .release = seq_release_private,
4603 static int __init slab_proc_init(void)
4605 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4606 #ifdef CONFIG_DEBUG_SLAB_LEAK
4607 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4611 module_init(slab_proc_init);
4615 * ksize - get the actual amount of memory allocated for a given object
4616 * @objp: Pointer to the object
4618 * kmalloc may internally round up allocations and return more memory
4619 * than requested. ksize() can be used to determine the actual amount of
4620 * memory allocated. The caller may use this additional memory, even though
4621 * a smaller amount of memory was initially specified with the kmalloc call.
4622 * The caller must guarantee that objp points to a valid object previously
4623 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4624 * must not be freed during the duration of the call.
4626 size_t ksize(const void *objp)
4629 if (unlikely(objp == ZERO_SIZE_PTR))
4632 return virt_to_cache(objp)->object_size;
4634 EXPORT_SYMBOL(ksize);