3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
119 #include <asm/cacheflush.h>
120 #include <asm/tlbflush.h>
121 #include <asm/page.h>
124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
125 * 0 for faster, smaller code (especially in the critical paths).
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
133 #ifdef CONFIG_DEBUG_SLAB
136 #define FORCED_DEBUG 1
140 #define FORCED_DEBUG 0
143 /* Shouldn't this be in a header file somewhere? */
144 #define BYTES_PER_WORD sizeof(void *)
145 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147 #ifndef ARCH_KMALLOC_FLAGS
148 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
151 /* Legal flag mask for kmem_cache_create(). */
153 # define CREATE_MASK (SLAB_RED_ZONE | \
154 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
157 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
158 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
159 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
164 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
165 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
171 * Bufctl's are used for linking objs within a slab
174 * This implementation relies on "struct page" for locating the cache &
175 * slab an object belongs to.
176 * This allows the bufctl structure to be small (one int), but limits
177 * the number of objects a slab (not a cache) can contain when off-slab
178 * bufctls are used. The limit is the size of the largest general cache
179 * that does not use off-slab slabs.
180 * For 32bit archs with 4 kB pages, is this 56.
181 * This is not serious, as it is only for large objects, when it is unwise
182 * to have too many per slab.
183 * Note: This limit can be raised by introducing a general cache whose size
184 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
187 typedef unsigned int kmem_bufctl_t;
188 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
189 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
190 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
191 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
196 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
197 * arrange for kmem_freepages to be called via RCU. This is useful if
198 * we need to approach a kernel structure obliquely, from its address
199 * obtained without the usual locking. We can lock the structure to
200 * stabilize it and check it's still at the given address, only if we
201 * can be sure that the memory has not been meanwhile reused for some
202 * other kind of object (which our subsystem's lock might corrupt).
204 * rcu_read_lock before reading the address, then rcu_read_unlock after
205 * taking the spinlock within the structure expected at that address.
208 struct rcu_head head;
209 struct kmem_cache *cachep;
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
228 unsigned short nodeid;
230 struct slab_rcu __slab_cover_slab_rcu;
238 * - LIFO ordering, to hand out cache-warm objects from _alloc
239 * - reduce the number of linked list operations
240 * - reduce spinlock operations
242 * The limit is stored in the per-cpu structure to reduce the data cache
249 unsigned int batchcount;
250 unsigned int touched;
253 * Must have this definition in here for the proper
254 * alignment of array_cache. Also simplifies accessing
260 * bootstrap: The caches do not work without cpuarrays anymore, but the
261 * cpuarrays are allocated from the generic caches...
263 #define BOOT_CPUCACHE_ENTRIES 1
264 struct arraycache_init {
265 struct array_cache cache;
266 void *entries[BOOT_CPUCACHE_ENTRIES];
270 * The slab lists for all objects.
273 struct list_head slabs_partial; /* partial list first, better asm code */
274 struct list_head slabs_full;
275 struct list_head slabs_free;
276 unsigned long free_objects;
277 unsigned int free_limit;
278 unsigned int colour_next; /* Per-node cache coloring */
279 spinlock_t list_lock;
280 struct array_cache *shared; /* shared per node */
281 struct array_cache **alien; /* on other nodes */
282 unsigned long next_reap; /* updated without locking */
283 int free_touched; /* updated without locking */
287 * Need this for bootstrapping a per node allocator.
289 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
290 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
291 #define CACHE_CACHE 0
292 #define SIZE_AC MAX_NUMNODES
293 #define SIZE_L3 (2 * MAX_NUMNODES)
295 static int drain_freelist(struct kmem_cache *cache,
296 struct kmem_list3 *l3, int tofree);
297 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
299 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
300 static void cache_reap(struct work_struct *unused);
303 * This function must be completely optimized away if a constant is passed to
304 * it. Mostly the same as what is in linux/slab.h except it returns an index.
306 static __always_inline int index_of(const size_t size)
308 extern void __bad_size(void);
310 if (__builtin_constant_p(size)) {
318 #include <linux/kmalloc_sizes.h>
326 static int slab_early_init = 1;
328 #define INDEX_AC index_of(sizeof(struct arraycache_init))
329 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
331 static void kmem_list3_init(struct kmem_list3 *parent)
333 INIT_LIST_HEAD(&parent->slabs_full);
334 INIT_LIST_HEAD(&parent->slabs_partial);
335 INIT_LIST_HEAD(&parent->slabs_free);
336 parent->shared = NULL;
337 parent->alien = NULL;
338 parent->colour_next = 0;
339 spin_lock_init(&parent->list_lock);
340 parent->free_objects = 0;
341 parent->free_touched = 0;
344 #define MAKE_LIST(cachep, listp, slab, nodeid) \
346 INIT_LIST_HEAD(listp); \
347 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
350 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
352 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
353 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
354 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
357 #define CFLGS_OFF_SLAB (0x80000000UL)
358 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
360 #define BATCHREFILL_LIMIT 16
362 * Optimization question: fewer reaps means less probability for unnessary
363 * cpucache drain/refill cycles.
365 * OTOH the cpuarrays can contain lots of objects,
366 * which could lock up otherwise freeable slabs.
368 #define REAPTIMEOUT_CPUC (2*HZ)
369 #define REAPTIMEOUT_LIST3 (4*HZ)
372 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
373 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
374 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
375 #define STATS_INC_GROWN(x) ((x)->grown++)
376 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
377 #define STATS_SET_HIGH(x) \
379 if ((x)->num_active > (x)->high_mark) \
380 (x)->high_mark = (x)->num_active; \
382 #define STATS_INC_ERR(x) ((x)->errors++)
383 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
384 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
385 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
386 #define STATS_SET_FREEABLE(x, i) \
388 if ((x)->max_freeable < i) \
389 (x)->max_freeable = i; \
391 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
392 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
393 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
394 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
396 #define STATS_INC_ACTIVE(x) do { } while (0)
397 #define STATS_DEC_ACTIVE(x) do { } while (0)
398 #define STATS_INC_ALLOCED(x) do { } while (0)
399 #define STATS_INC_GROWN(x) do { } while (0)
400 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
401 #define STATS_SET_HIGH(x) do { } while (0)
402 #define STATS_INC_ERR(x) do { } while (0)
403 #define STATS_INC_NODEALLOCS(x) do { } while (0)
404 #define STATS_INC_NODEFREES(x) do { } while (0)
405 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
406 #define STATS_SET_FREEABLE(x, i) do { } while (0)
407 #define STATS_INC_ALLOCHIT(x) do { } while (0)
408 #define STATS_INC_ALLOCMISS(x) do { } while (0)
409 #define STATS_INC_FREEHIT(x) do { } while (0)
410 #define STATS_INC_FREEMISS(x) do { } while (0)
416 * memory layout of objects:
418 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
419 * the end of an object is aligned with the end of the real
420 * allocation. Catches writes behind the end of the allocation.
421 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
423 * cachep->obj_offset: The real object.
424 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
425 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
426 * [BYTES_PER_WORD long]
428 static int obj_offset(struct kmem_cache *cachep)
430 return cachep->obj_offset;
433 static int obj_size(struct kmem_cache *cachep)
435 return cachep->obj_size;
438 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
440 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
441 return (unsigned long long*) (objp + obj_offset(cachep) -
442 sizeof(unsigned long long));
445 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
447 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
448 if (cachep->flags & SLAB_STORE_USER)
449 return (unsigned long long *)(objp + cachep->buffer_size -
450 sizeof(unsigned long long) -
452 return (unsigned long long *) (objp + cachep->buffer_size -
453 sizeof(unsigned long long));
456 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
458 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
459 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
464 #define obj_offset(x) 0
465 #define obj_size(cachep) (cachep->buffer_size)
466 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
467 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
472 #ifdef CONFIG_TRACING
473 size_t slab_buffer_size(struct kmem_cache *cachep)
475 return cachep->buffer_size;
477 EXPORT_SYMBOL(slab_buffer_size);
481 * Do not go above this order unless 0 objects fit into the slab.
483 #define BREAK_GFP_ORDER_HI 1
484 #define BREAK_GFP_ORDER_LO 0
485 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
488 * Functions for storing/retrieving the cachep and or slab from the page
489 * allocator. These are used to find the slab an obj belongs to. With kfree(),
490 * these are used to find the cache which an obj belongs to.
492 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
494 page->lru.next = (struct list_head *)cache;
497 static inline struct kmem_cache *page_get_cache(struct page *page)
499 page = compound_head(page);
500 BUG_ON(!PageSlab(page));
501 return (struct kmem_cache *)page->lru.next;
504 static inline void page_set_slab(struct page *page, struct slab *slab)
506 page->lru.prev = (struct list_head *)slab;
509 static inline struct slab *page_get_slab(struct page *page)
511 BUG_ON(!PageSlab(page));
512 return (struct slab *)page->lru.prev;
515 static inline struct kmem_cache *virt_to_cache(const void *obj)
517 struct page *page = virt_to_head_page(obj);
518 return page_get_cache(page);
521 static inline struct slab *virt_to_slab(const void *obj)
523 struct page *page = virt_to_head_page(obj);
524 return page_get_slab(page);
527 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
530 return slab->s_mem + cache->buffer_size * idx;
534 * We want to avoid an expensive divide : (offset / cache->buffer_size)
535 * Using the fact that buffer_size is a constant for a particular cache,
536 * we can replace (offset / cache->buffer_size) by
537 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
539 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
540 const struct slab *slab, void *obj)
542 u32 offset = (obj - slab->s_mem);
543 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
547 * These are the default caches for kmalloc. Custom caches can have other sizes.
549 struct cache_sizes malloc_sizes[] = {
550 #define CACHE(x) { .cs_size = (x) },
551 #include <linux/kmalloc_sizes.h>
555 EXPORT_SYMBOL(malloc_sizes);
557 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
563 static struct cache_names __initdata cache_names[] = {
564 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
565 #include <linux/kmalloc_sizes.h>
570 static struct arraycache_init initarray_cache __initdata =
571 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
572 static struct arraycache_init initarray_generic =
573 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
575 /* internal cache of cache description objs */
576 static struct kmem_cache cache_cache = {
578 .limit = BOOT_CPUCACHE_ENTRIES,
580 .buffer_size = sizeof(struct kmem_cache),
581 .name = "kmem_cache",
584 #define BAD_ALIEN_MAGIC 0x01020304ul
587 * chicken and egg problem: delay the per-cpu array allocation
588 * until the general caches are up.
599 * used by boot code to determine if it can use slab based allocator
601 int slab_is_available(void)
603 return g_cpucache_up >= EARLY;
606 #ifdef CONFIG_LOCKDEP
609 * Slab sometimes uses the kmalloc slabs to store the slab headers
610 * for other slabs "off slab".
611 * The locking for this is tricky in that it nests within the locks
612 * of all other slabs in a few places; to deal with this special
613 * locking we put on-slab caches into a separate lock-class.
615 * We set lock class for alien array caches which are up during init.
616 * The lock annotation will be lost if all cpus of a node goes down and
617 * then comes back up during hotplug
619 static struct lock_class_key on_slab_l3_key;
620 static struct lock_class_key on_slab_alc_key;
622 static void init_node_lock_keys(int q)
624 struct cache_sizes *s = malloc_sizes;
626 if (g_cpucache_up != FULL)
629 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
630 struct array_cache **alc;
631 struct kmem_list3 *l3;
634 l3 = s->cs_cachep->nodelists[q];
635 if (!l3 || OFF_SLAB(s->cs_cachep))
637 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
640 * FIXME: This check for BAD_ALIEN_MAGIC
641 * should go away when common slab code is taught to
642 * work even without alien caches.
643 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
644 * for alloc_alien_cache,
646 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
650 lockdep_set_class(&alc[r]->lock,
656 static inline void init_lock_keys(void)
661 init_node_lock_keys(node);
664 static void init_node_lock_keys(int q)
668 static inline void init_lock_keys(void)
674 * Guard access to the cache-chain.
676 static DEFINE_MUTEX(cache_chain_mutex);
677 static struct list_head cache_chain;
679 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
681 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
683 return cachep->array[smp_processor_id()];
686 static inline struct kmem_cache *__find_general_cachep(size_t size,
689 struct cache_sizes *csizep = malloc_sizes;
692 /* This happens if someone tries to call
693 * kmem_cache_create(), or __kmalloc(), before
694 * the generic caches are initialized.
696 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
699 return ZERO_SIZE_PTR;
701 while (size > csizep->cs_size)
705 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
706 * has cs_{dma,}cachep==NULL. Thus no special case
707 * for large kmalloc calls required.
709 #ifdef CONFIG_ZONE_DMA
710 if (unlikely(gfpflags & GFP_DMA))
711 return csizep->cs_dmacachep;
713 return csizep->cs_cachep;
716 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
718 return __find_general_cachep(size, gfpflags);
721 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
723 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
727 * Calculate the number of objects and left-over bytes for a given buffer size.
729 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
730 size_t align, int flags, size_t *left_over,
735 size_t slab_size = PAGE_SIZE << gfporder;
738 * The slab management structure can be either off the slab or
739 * on it. For the latter case, the memory allocated for a
743 * - One kmem_bufctl_t for each object
744 * - Padding to respect alignment of @align
745 * - @buffer_size bytes for each object
747 * If the slab management structure is off the slab, then the
748 * alignment will already be calculated into the size. Because
749 * the slabs are all pages aligned, the objects will be at the
750 * correct alignment when allocated.
752 if (flags & CFLGS_OFF_SLAB) {
754 nr_objs = slab_size / buffer_size;
756 if (nr_objs > SLAB_LIMIT)
757 nr_objs = SLAB_LIMIT;
760 * Ignore padding for the initial guess. The padding
761 * is at most @align-1 bytes, and @buffer_size is at
762 * least @align. In the worst case, this result will
763 * be one greater than the number of objects that fit
764 * into the memory allocation when taking the padding
767 nr_objs = (slab_size - sizeof(struct slab)) /
768 (buffer_size + sizeof(kmem_bufctl_t));
771 * This calculated number will be either the right
772 * amount, or one greater than what we want.
774 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
778 if (nr_objs > SLAB_LIMIT)
779 nr_objs = SLAB_LIMIT;
781 mgmt_size = slab_mgmt_size(nr_objs, align);
784 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
787 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
789 static void __slab_error(const char *function, struct kmem_cache *cachep,
792 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
793 function, cachep->name, msg);
798 * By default on NUMA we use alien caches to stage the freeing of
799 * objects allocated from other nodes. This causes massive memory
800 * inefficiencies when using fake NUMA setup to split memory into a
801 * large number of small nodes, so it can be disabled on the command
805 static int use_alien_caches __read_mostly = 1;
806 static int __init noaliencache_setup(char *s)
808 use_alien_caches = 0;
811 __setup("noaliencache", noaliencache_setup);
815 * Special reaping functions for NUMA systems called from cache_reap().
816 * These take care of doing round robin flushing of alien caches (containing
817 * objects freed on different nodes from which they were allocated) and the
818 * flushing of remote pcps by calling drain_node_pages.
820 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
822 static void init_reap_node(int cpu)
826 node = next_node(cpu_to_mem(cpu), node_online_map);
827 if (node == MAX_NUMNODES)
828 node = first_node(node_online_map);
830 per_cpu(slab_reap_node, cpu) = node;
833 static void next_reap_node(void)
835 int node = __this_cpu_read(slab_reap_node);
837 node = next_node(node, node_online_map);
838 if (unlikely(node >= MAX_NUMNODES))
839 node = first_node(node_online_map);
840 __this_cpu_write(slab_reap_node, node);
844 #define init_reap_node(cpu) do { } while (0)
845 #define next_reap_node(void) do { } while (0)
849 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
850 * via the workqueue/eventd.
851 * Add the CPU number into the expiration time to minimize the possibility of
852 * the CPUs getting into lockstep and contending for the global cache chain
855 static void __cpuinit start_cpu_timer(int cpu)
857 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
860 * When this gets called from do_initcalls via cpucache_init(),
861 * init_workqueues() has already run, so keventd will be setup
864 if (keventd_up() && reap_work->work.func == NULL) {
866 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
867 schedule_delayed_work_on(cpu, reap_work,
868 __round_jiffies_relative(HZ, cpu));
872 static struct array_cache *alloc_arraycache(int node, int entries,
873 int batchcount, gfp_t gfp)
875 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
876 struct array_cache *nc = NULL;
878 nc = kmalloc_node(memsize, gfp, node);
880 * The array_cache structures contain pointers to free object.
881 * However, when such objects are allocated or transfered to another
882 * cache the pointers are not cleared and they could be counted as
883 * valid references during a kmemleak scan. Therefore, kmemleak must
884 * not scan such objects.
886 kmemleak_no_scan(nc);
890 nc->batchcount = batchcount;
892 spin_lock_init(&nc->lock);
898 * Transfer objects in one arraycache to another.
899 * Locking must be handled by the caller.
901 * Return the number of entries transferred.
903 static int transfer_objects(struct array_cache *to,
904 struct array_cache *from, unsigned int max)
906 /* Figure out how many entries to transfer */
907 int nr = min3(from->avail, max, to->limit - to->avail);
912 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
922 #define drain_alien_cache(cachep, alien) do { } while (0)
923 #define reap_alien(cachep, l3) do { } while (0)
925 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
927 return (struct array_cache **)BAD_ALIEN_MAGIC;
930 static inline void free_alien_cache(struct array_cache **ac_ptr)
934 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
939 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
945 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
946 gfp_t flags, int nodeid)
951 #else /* CONFIG_NUMA */
953 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
954 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
956 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
958 struct array_cache **ac_ptr;
959 int memsize = sizeof(void *) * nr_node_ids;
964 ac_ptr = kzalloc_node(memsize, gfp, node);
967 if (i == node || !node_online(i))
969 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
971 for (i--; i >= 0; i--)
981 static void free_alien_cache(struct array_cache **ac_ptr)
992 static void __drain_alien_cache(struct kmem_cache *cachep,
993 struct array_cache *ac, int node)
995 struct kmem_list3 *rl3 = cachep->nodelists[node];
998 spin_lock(&rl3->list_lock);
1000 * Stuff objects into the remote nodes shared array first.
1001 * That way we could avoid the overhead of putting the objects
1002 * into the free lists and getting them back later.
1005 transfer_objects(rl3->shared, ac, ac->limit);
1007 free_block(cachep, ac->entry, ac->avail, node);
1009 spin_unlock(&rl3->list_lock);
1014 * Called from cache_reap() to regularly drain alien caches round robin.
1016 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1018 int node = __this_cpu_read(slab_reap_node);
1021 struct array_cache *ac = l3->alien[node];
1023 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1024 __drain_alien_cache(cachep, ac, node);
1025 spin_unlock_irq(&ac->lock);
1030 static void drain_alien_cache(struct kmem_cache *cachep,
1031 struct array_cache **alien)
1034 struct array_cache *ac;
1035 unsigned long flags;
1037 for_each_online_node(i) {
1040 spin_lock_irqsave(&ac->lock, flags);
1041 __drain_alien_cache(cachep, ac, i);
1042 spin_unlock_irqrestore(&ac->lock, flags);
1047 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1049 struct slab *slabp = virt_to_slab(objp);
1050 int nodeid = slabp->nodeid;
1051 struct kmem_list3 *l3;
1052 struct array_cache *alien = NULL;
1055 node = numa_mem_id();
1058 * Make sure we are not freeing a object from another node to the array
1059 * cache on this cpu.
1061 if (likely(slabp->nodeid == node))
1064 l3 = cachep->nodelists[node];
1065 STATS_INC_NODEFREES(cachep);
1066 if (l3->alien && l3->alien[nodeid]) {
1067 alien = l3->alien[nodeid];
1068 spin_lock(&alien->lock);
1069 if (unlikely(alien->avail == alien->limit)) {
1070 STATS_INC_ACOVERFLOW(cachep);
1071 __drain_alien_cache(cachep, alien, nodeid);
1073 alien->entry[alien->avail++] = objp;
1074 spin_unlock(&alien->lock);
1076 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1077 free_block(cachep, &objp, 1, nodeid);
1078 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1085 * Allocates and initializes nodelists for a node on each slab cache, used for
1086 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1087 * will be allocated off-node since memory is not yet online for the new node.
1088 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1091 * Must hold cache_chain_mutex.
1093 static int init_cache_nodelists_node(int node)
1095 struct kmem_cache *cachep;
1096 struct kmem_list3 *l3;
1097 const int memsize = sizeof(struct kmem_list3);
1099 list_for_each_entry(cachep, &cache_chain, next) {
1101 * Set up the size64 kmemlist for cpu before we can
1102 * begin anything. Make sure some other cpu on this
1103 * node has not already allocated this
1105 if (!cachep->nodelists[node]) {
1106 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1109 kmem_list3_init(l3);
1110 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1111 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1114 * The l3s don't come and go as CPUs come and
1115 * go. cache_chain_mutex is sufficient
1118 cachep->nodelists[node] = l3;
1121 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1122 cachep->nodelists[node]->free_limit =
1123 (1 + nr_cpus_node(node)) *
1124 cachep->batchcount + cachep->num;
1125 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1130 static void __cpuinit cpuup_canceled(long cpu)
1132 struct kmem_cache *cachep;
1133 struct kmem_list3 *l3 = NULL;
1134 int node = cpu_to_mem(cpu);
1135 const struct cpumask *mask = cpumask_of_node(node);
1137 list_for_each_entry(cachep, &cache_chain, next) {
1138 struct array_cache *nc;
1139 struct array_cache *shared;
1140 struct array_cache **alien;
1142 /* cpu is dead; no one can alloc from it. */
1143 nc = cachep->array[cpu];
1144 cachep->array[cpu] = NULL;
1145 l3 = cachep->nodelists[node];
1148 goto free_array_cache;
1150 spin_lock_irq(&l3->list_lock);
1152 /* Free limit for this kmem_list3 */
1153 l3->free_limit -= cachep->batchcount;
1155 free_block(cachep, nc->entry, nc->avail, node);
1157 if (!cpumask_empty(mask)) {
1158 spin_unlock_irq(&l3->list_lock);
1159 goto free_array_cache;
1162 shared = l3->shared;
1164 free_block(cachep, shared->entry,
1165 shared->avail, node);
1172 spin_unlock_irq(&l3->list_lock);
1176 drain_alien_cache(cachep, alien);
1177 free_alien_cache(alien);
1183 * In the previous loop, all the objects were freed to
1184 * the respective cache's slabs, now we can go ahead and
1185 * shrink each nodelist to its limit.
1187 list_for_each_entry(cachep, &cache_chain, next) {
1188 l3 = cachep->nodelists[node];
1191 drain_freelist(cachep, l3, l3->free_objects);
1195 static int __cpuinit cpuup_prepare(long cpu)
1197 struct kmem_cache *cachep;
1198 struct kmem_list3 *l3 = NULL;
1199 int node = cpu_to_mem(cpu);
1203 * We need to do this right in the beginning since
1204 * alloc_arraycache's are going to use this list.
1205 * kmalloc_node allows us to add the slab to the right
1206 * kmem_list3 and not this cpu's kmem_list3
1208 err = init_cache_nodelists_node(node);
1213 * Now we can go ahead with allocating the shared arrays and
1216 list_for_each_entry(cachep, &cache_chain, next) {
1217 struct array_cache *nc;
1218 struct array_cache *shared = NULL;
1219 struct array_cache **alien = NULL;
1221 nc = alloc_arraycache(node, cachep->limit,
1222 cachep->batchcount, GFP_KERNEL);
1225 if (cachep->shared) {
1226 shared = alloc_arraycache(node,
1227 cachep->shared * cachep->batchcount,
1228 0xbaadf00d, GFP_KERNEL);
1234 if (use_alien_caches) {
1235 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1242 cachep->array[cpu] = nc;
1243 l3 = cachep->nodelists[node];
1246 spin_lock_irq(&l3->list_lock);
1249 * We are serialised from CPU_DEAD or
1250 * CPU_UP_CANCELLED by the cpucontrol lock
1252 l3->shared = shared;
1261 spin_unlock_irq(&l3->list_lock);
1263 free_alien_cache(alien);
1265 init_node_lock_keys(node);
1269 cpuup_canceled(cpu);
1273 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1274 unsigned long action, void *hcpu)
1276 long cpu = (long)hcpu;
1280 case CPU_UP_PREPARE:
1281 case CPU_UP_PREPARE_FROZEN:
1282 mutex_lock(&cache_chain_mutex);
1283 err = cpuup_prepare(cpu);
1284 mutex_unlock(&cache_chain_mutex);
1287 case CPU_ONLINE_FROZEN:
1288 start_cpu_timer(cpu);
1290 #ifdef CONFIG_HOTPLUG_CPU
1291 case CPU_DOWN_PREPARE:
1292 case CPU_DOWN_PREPARE_FROZEN:
1294 * Shutdown cache reaper. Note that the cache_chain_mutex is
1295 * held so that if cache_reap() is invoked it cannot do
1296 * anything expensive but will only modify reap_work
1297 * and reschedule the timer.
1299 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1300 /* Now the cache_reaper is guaranteed to be not running. */
1301 per_cpu(slab_reap_work, cpu).work.func = NULL;
1303 case CPU_DOWN_FAILED:
1304 case CPU_DOWN_FAILED_FROZEN:
1305 start_cpu_timer(cpu);
1308 case CPU_DEAD_FROZEN:
1310 * Even if all the cpus of a node are down, we don't free the
1311 * kmem_list3 of any cache. This to avoid a race between
1312 * cpu_down, and a kmalloc allocation from another cpu for
1313 * memory from the node of the cpu going down. The list3
1314 * structure is usually allocated from kmem_cache_create() and
1315 * gets destroyed at kmem_cache_destroy().
1319 case CPU_UP_CANCELED:
1320 case CPU_UP_CANCELED_FROZEN:
1321 mutex_lock(&cache_chain_mutex);
1322 cpuup_canceled(cpu);
1323 mutex_unlock(&cache_chain_mutex);
1326 return notifier_from_errno(err);
1329 static struct notifier_block __cpuinitdata cpucache_notifier = {
1330 &cpuup_callback, NULL, 0
1333 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1335 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1336 * Returns -EBUSY if all objects cannot be drained so that the node is not
1339 * Must hold cache_chain_mutex.
1341 static int __meminit drain_cache_nodelists_node(int node)
1343 struct kmem_cache *cachep;
1346 list_for_each_entry(cachep, &cache_chain, next) {
1347 struct kmem_list3 *l3;
1349 l3 = cachep->nodelists[node];
1353 drain_freelist(cachep, l3, l3->free_objects);
1355 if (!list_empty(&l3->slabs_full) ||
1356 !list_empty(&l3->slabs_partial)) {
1364 static int __meminit slab_memory_callback(struct notifier_block *self,
1365 unsigned long action, void *arg)
1367 struct memory_notify *mnb = arg;
1371 nid = mnb->status_change_nid;
1376 case MEM_GOING_ONLINE:
1377 mutex_lock(&cache_chain_mutex);
1378 ret = init_cache_nodelists_node(nid);
1379 mutex_unlock(&cache_chain_mutex);
1381 case MEM_GOING_OFFLINE:
1382 mutex_lock(&cache_chain_mutex);
1383 ret = drain_cache_nodelists_node(nid);
1384 mutex_unlock(&cache_chain_mutex);
1388 case MEM_CANCEL_ONLINE:
1389 case MEM_CANCEL_OFFLINE:
1393 return ret ? notifier_from_errno(ret) : NOTIFY_OK;
1395 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1398 * swap the static kmem_list3 with kmalloced memory
1400 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1403 struct kmem_list3 *ptr;
1405 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1408 memcpy(ptr, list, sizeof(struct kmem_list3));
1410 * Do not assume that spinlocks can be initialized via memcpy:
1412 spin_lock_init(&ptr->list_lock);
1414 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1415 cachep->nodelists[nodeid] = ptr;
1419 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1420 * size of kmem_list3.
1422 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1426 for_each_online_node(node) {
1427 cachep->nodelists[node] = &initkmem_list3[index + node];
1428 cachep->nodelists[node]->next_reap = jiffies +
1430 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1435 * Initialisation. Called after the page allocator have been initialised and
1436 * before smp_init().
1438 void __init kmem_cache_init(void)
1441 struct cache_sizes *sizes;
1442 struct cache_names *names;
1447 if (num_possible_nodes() == 1)
1448 use_alien_caches = 0;
1450 for (i = 0; i < NUM_INIT_LISTS; i++) {
1451 kmem_list3_init(&initkmem_list3[i]);
1452 if (i < MAX_NUMNODES)
1453 cache_cache.nodelists[i] = NULL;
1455 set_up_list3s(&cache_cache, CACHE_CACHE);
1458 * Fragmentation resistance on low memory - only use bigger
1459 * page orders on machines with more than 32MB of memory.
1461 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1462 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1464 /* Bootstrap is tricky, because several objects are allocated
1465 * from caches that do not exist yet:
1466 * 1) initialize the cache_cache cache: it contains the struct
1467 * kmem_cache structures of all caches, except cache_cache itself:
1468 * cache_cache is statically allocated.
1469 * Initially an __init data area is used for the head array and the
1470 * kmem_list3 structures, it's replaced with a kmalloc allocated
1471 * array at the end of the bootstrap.
1472 * 2) Create the first kmalloc cache.
1473 * The struct kmem_cache for the new cache is allocated normally.
1474 * An __init data area is used for the head array.
1475 * 3) Create the remaining kmalloc caches, with minimally sized
1477 * 4) Replace the __init data head arrays for cache_cache and the first
1478 * kmalloc cache with kmalloc allocated arrays.
1479 * 5) Replace the __init data for kmem_list3 for cache_cache and
1480 * the other cache's with kmalloc allocated memory.
1481 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1484 node = numa_mem_id();
1486 /* 1) create the cache_cache */
1487 INIT_LIST_HEAD(&cache_chain);
1488 list_add(&cache_cache.next, &cache_chain);
1489 cache_cache.colour_off = cache_line_size();
1490 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1491 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1494 * struct kmem_cache size depends on nr_node_ids, which
1495 * can be less than MAX_NUMNODES.
1497 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1498 nr_node_ids * sizeof(struct kmem_list3 *);
1500 cache_cache.obj_size = cache_cache.buffer_size;
1502 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1504 cache_cache.reciprocal_buffer_size =
1505 reciprocal_value(cache_cache.buffer_size);
1507 for (order = 0; order < MAX_ORDER; order++) {
1508 cache_estimate(order, cache_cache.buffer_size,
1509 cache_line_size(), 0, &left_over, &cache_cache.num);
1510 if (cache_cache.num)
1513 BUG_ON(!cache_cache.num);
1514 cache_cache.gfporder = order;
1515 cache_cache.colour = left_over / cache_cache.colour_off;
1516 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1517 sizeof(struct slab), cache_line_size());
1519 /* 2+3) create the kmalloc caches */
1520 sizes = malloc_sizes;
1521 names = cache_names;
1524 * Initialize the caches that provide memory for the array cache and the
1525 * kmem_list3 structures first. Without this, further allocations will
1529 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1530 sizes[INDEX_AC].cs_size,
1531 ARCH_KMALLOC_MINALIGN,
1532 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1535 if (INDEX_AC != INDEX_L3) {
1536 sizes[INDEX_L3].cs_cachep =
1537 kmem_cache_create(names[INDEX_L3].name,
1538 sizes[INDEX_L3].cs_size,
1539 ARCH_KMALLOC_MINALIGN,
1540 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1544 slab_early_init = 0;
1546 while (sizes->cs_size != ULONG_MAX) {
1548 * For performance, all the general caches are L1 aligned.
1549 * This should be particularly beneficial on SMP boxes, as it
1550 * eliminates "false sharing".
1551 * Note for systems short on memory removing the alignment will
1552 * allow tighter packing of the smaller caches.
1554 if (!sizes->cs_cachep) {
1555 sizes->cs_cachep = kmem_cache_create(names->name,
1557 ARCH_KMALLOC_MINALIGN,
1558 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1561 #ifdef CONFIG_ZONE_DMA
1562 sizes->cs_dmacachep = kmem_cache_create(
1565 ARCH_KMALLOC_MINALIGN,
1566 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1573 /* 4) Replace the bootstrap head arrays */
1575 struct array_cache *ptr;
1577 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1579 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1580 memcpy(ptr, cpu_cache_get(&cache_cache),
1581 sizeof(struct arraycache_init));
1583 * Do not assume that spinlocks can be initialized via memcpy:
1585 spin_lock_init(&ptr->lock);
1587 cache_cache.array[smp_processor_id()] = ptr;
1589 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1591 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1592 != &initarray_generic.cache);
1593 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1594 sizeof(struct arraycache_init));
1596 * Do not assume that spinlocks can be initialized via memcpy:
1598 spin_lock_init(&ptr->lock);
1600 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1603 /* 5) Replace the bootstrap kmem_list3's */
1607 for_each_online_node(nid) {
1608 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1610 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1611 &initkmem_list3[SIZE_AC + nid], nid);
1613 if (INDEX_AC != INDEX_L3) {
1614 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1615 &initkmem_list3[SIZE_L3 + nid], nid);
1620 g_cpucache_up = EARLY;
1623 void __init kmem_cache_init_late(void)
1625 struct kmem_cache *cachep;
1627 /* 6) resize the head arrays to their final sizes */
1628 mutex_lock(&cache_chain_mutex);
1629 list_for_each_entry(cachep, &cache_chain, next)
1630 if (enable_cpucache(cachep, GFP_NOWAIT))
1632 mutex_unlock(&cache_chain_mutex);
1635 g_cpucache_up = FULL;
1637 /* Annotate slab for lockdep -- annotate the malloc caches */
1641 * Register a cpu startup notifier callback that initializes
1642 * cpu_cache_get for all new cpus
1644 register_cpu_notifier(&cpucache_notifier);
1648 * Register a memory hotplug callback that initializes and frees
1651 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1655 * The reap timers are started later, with a module init call: That part
1656 * of the kernel is not yet operational.
1660 static int __init cpucache_init(void)
1665 * Register the timers that return unneeded pages to the page allocator
1667 for_each_online_cpu(cpu)
1668 start_cpu_timer(cpu);
1671 __initcall(cpucache_init);
1674 * Interface to system's page allocator. No need to hold the cache-lock.
1676 * If we requested dmaable memory, we will get it. Even if we
1677 * did not request dmaable memory, we might get it, but that
1678 * would be relatively rare and ignorable.
1680 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1688 * Nommu uses slab's for process anonymous memory allocations, and thus
1689 * requires __GFP_COMP to properly refcount higher order allocations
1691 flags |= __GFP_COMP;
1694 flags |= cachep->gfpflags;
1695 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1696 flags |= __GFP_RECLAIMABLE;
1698 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1702 nr_pages = (1 << cachep->gfporder);
1703 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1704 add_zone_page_state(page_zone(page),
1705 NR_SLAB_RECLAIMABLE, nr_pages);
1707 add_zone_page_state(page_zone(page),
1708 NR_SLAB_UNRECLAIMABLE, nr_pages);
1709 for (i = 0; i < nr_pages; i++)
1710 __SetPageSlab(page + i);
1712 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1713 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1716 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1718 kmemcheck_mark_unallocated_pages(page, nr_pages);
1721 return page_address(page);
1725 * Interface to system's page release.
1727 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1729 unsigned long i = (1 << cachep->gfporder);
1730 struct page *page = virt_to_page(addr);
1731 const unsigned long nr_freed = i;
1733 kmemcheck_free_shadow(page, cachep->gfporder);
1735 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1736 sub_zone_page_state(page_zone(page),
1737 NR_SLAB_RECLAIMABLE, nr_freed);
1739 sub_zone_page_state(page_zone(page),
1740 NR_SLAB_UNRECLAIMABLE, nr_freed);
1742 BUG_ON(!PageSlab(page));
1743 __ClearPageSlab(page);
1746 if (current->reclaim_state)
1747 current->reclaim_state->reclaimed_slab += nr_freed;
1748 free_pages((unsigned long)addr, cachep->gfporder);
1751 static void kmem_rcu_free(struct rcu_head *head)
1753 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1754 struct kmem_cache *cachep = slab_rcu->cachep;
1756 kmem_freepages(cachep, slab_rcu->addr);
1757 if (OFF_SLAB(cachep))
1758 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1763 #ifdef CONFIG_DEBUG_PAGEALLOC
1764 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1765 unsigned long caller)
1767 int size = obj_size(cachep);
1769 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1771 if (size < 5 * sizeof(unsigned long))
1774 *addr++ = 0x12345678;
1776 *addr++ = smp_processor_id();
1777 size -= 3 * sizeof(unsigned long);
1779 unsigned long *sptr = &caller;
1780 unsigned long svalue;
1782 while (!kstack_end(sptr)) {
1784 if (kernel_text_address(svalue)) {
1786 size -= sizeof(unsigned long);
1787 if (size <= sizeof(unsigned long))
1793 *addr++ = 0x87654321;
1797 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1799 int size = obj_size(cachep);
1800 addr = &((char *)addr)[obj_offset(cachep)];
1802 memset(addr, val, size);
1803 *(unsigned char *)(addr + size - 1) = POISON_END;
1806 static void dump_line(char *data, int offset, int limit)
1809 unsigned char error = 0;
1812 printk(KERN_ERR "%03x:", offset);
1813 for (i = 0; i < limit; i++) {
1814 if (data[offset + i] != POISON_FREE) {
1815 error = data[offset + i];
1818 printk(" %02x", (unsigned char)data[offset + i]);
1822 if (bad_count == 1) {
1823 error ^= POISON_FREE;
1824 if (!(error & (error - 1))) {
1825 printk(KERN_ERR "Single bit error detected. Probably "
1828 printk(KERN_ERR "Run memtest86+ or a similar memory "
1831 printk(KERN_ERR "Run a memory test tool.\n");
1840 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1845 if (cachep->flags & SLAB_RED_ZONE) {
1846 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1847 *dbg_redzone1(cachep, objp),
1848 *dbg_redzone2(cachep, objp));
1851 if (cachep->flags & SLAB_STORE_USER) {
1852 printk(KERN_ERR "Last user: [<%p>]",
1853 *dbg_userword(cachep, objp));
1854 print_symbol("(%s)",
1855 (unsigned long)*dbg_userword(cachep, objp));
1858 realobj = (char *)objp + obj_offset(cachep);
1859 size = obj_size(cachep);
1860 for (i = 0; i < size && lines; i += 16, lines--) {
1863 if (i + limit > size)
1865 dump_line(realobj, i, limit);
1869 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1875 realobj = (char *)objp + obj_offset(cachep);
1876 size = obj_size(cachep);
1878 for (i = 0; i < size; i++) {
1879 char exp = POISON_FREE;
1882 if (realobj[i] != exp) {
1888 "Slab corruption: %s start=%p, len=%d\n",
1889 cachep->name, realobj, size);
1890 print_objinfo(cachep, objp, 0);
1892 /* Hexdump the affected line */
1895 if (i + limit > size)
1897 dump_line(realobj, i, limit);
1900 /* Limit to 5 lines */
1906 /* Print some data about the neighboring objects, if they
1909 struct slab *slabp = virt_to_slab(objp);
1912 objnr = obj_to_index(cachep, slabp, objp);
1914 objp = index_to_obj(cachep, slabp, objnr - 1);
1915 realobj = (char *)objp + obj_offset(cachep);
1916 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1918 print_objinfo(cachep, objp, 2);
1920 if (objnr + 1 < cachep->num) {
1921 objp = index_to_obj(cachep, slabp, objnr + 1);
1922 realobj = (char *)objp + obj_offset(cachep);
1923 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1925 print_objinfo(cachep, objp, 2);
1932 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1935 for (i = 0; i < cachep->num; i++) {
1936 void *objp = index_to_obj(cachep, slabp, i);
1938 if (cachep->flags & SLAB_POISON) {
1939 #ifdef CONFIG_DEBUG_PAGEALLOC
1940 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1942 kernel_map_pages(virt_to_page(objp),
1943 cachep->buffer_size / PAGE_SIZE, 1);
1945 check_poison_obj(cachep, objp);
1947 check_poison_obj(cachep, objp);
1950 if (cachep->flags & SLAB_RED_ZONE) {
1951 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1952 slab_error(cachep, "start of a freed object "
1954 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1955 slab_error(cachep, "end of a freed object "
1961 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1967 * slab_destroy - destroy and release all objects in a slab
1968 * @cachep: cache pointer being destroyed
1969 * @slabp: slab pointer being destroyed
1971 * Destroy all the objs in a slab, and release the mem back to the system.
1972 * Before calling the slab must have been unlinked from the cache. The
1973 * cache-lock is not held/needed.
1975 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1977 void *addr = slabp->s_mem - slabp->colouroff;
1979 slab_destroy_debugcheck(cachep, slabp);
1980 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1981 struct slab_rcu *slab_rcu;
1983 slab_rcu = (struct slab_rcu *)slabp;
1984 slab_rcu->cachep = cachep;
1985 slab_rcu->addr = addr;
1986 call_rcu(&slab_rcu->head, kmem_rcu_free);
1988 kmem_freepages(cachep, addr);
1989 if (OFF_SLAB(cachep))
1990 kmem_cache_free(cachep->slabp_cache, slabp);
1994 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1997 struct kmem_list3 *l3;
1999 for_each_online_cpu(i)
2000 kfree(cachep->array[i]);
2002 /* NUMA: free the list3 structures */
2003 for_each_online_node(i) {
2004 l3 = cachep->nodelists[i];
2007 free_alien_cache(l3->alien);
2011 kmem_cache_free(&cache_cache, cachep);
2016 * calculate_slab_order - calculate size (page order) of slabs
2017 * @cachep: pointer to the cache that is being created
2018 * @size: size of objects to be created in this cache.
2019 * @align: required alignment for the objects.
2020 * @flags: slab allocation flags
2022 * Also calculates the number of objects per slab.
2024 * This could be made much more intelligent. For now, try to avoid using
2025 * high order pages for slabs. When the gfp() functions are more friendly
2026 * towards high-order requests, this should be changed.
2028 static size_t calculate_slab_order(struct kmem_cache *cachep,
2029 size_t size, size_t align, unsigned long flags)
2031 unsigned long offslab_limit;
2032 size_t left_over = 0;
2035 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2039 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2043 if (flags & CFLGS_OFF_SLAB) {
2045 * Max number of objs-per-slab for caches which
2046 * use off-slab slabs. Needed to avoid a possible
2047 * looping condition in cache_grow().
2049 offslab_limit = size - sizeof(struct slab);
2050 offslab_limit /= sizeof(kmem_bufctl_t);
2052 if (num > offslab_limit)
2056 /* Found something acceptable - save it away */
2058 cachep->gfporder = gfporder;
2059 left_over = remainder;
2062 * A VFS-reclaimable slab tends to have most allocations
2063 * as GFP_NOFS and we really don't want to have to be allocating
2064 * higher-order pages when we are unable to shrink dcache.
2066 if (flags & SLAB_RECLAIM_ACCOUNT)
2070 * Large number of objects is good, but very large slabs are
2071 * currently bad for the gfp()s.
2073 if (gfporder >= slab_break_gfp_order)
2077 * Acceptable internal fragmentation?
2079 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2085 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2087 if (g_cpucache_up == FULL)
2088 return enable_cpucache(cachep, gfp);
2090 if (g_cpucache_up == NONE) {
2092 * Note: the first kmem_cache_create must create the cache
2093 * that's used by kmalloc(24), otherwise the creation of
2094 * further caches will BUG().
2096 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2099 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2100 * the first cache, then we need to set up all its list3s,
2101 * otherwise the creation of further caches will BUG().
2103 set_up_list3s(cachep, SIZE_AC);
2104 if (INDEX_AC == INDEX_L3)
2105 g_cpucache_up = PARTIAL_L3;
2107 g_cpucache_up = PARTIAL_AC;
2109 cachep->array[smp_processor_id()] =
2110 kmalloc(sizeof(struct arraycache_init), gfp);
2112 if (g_cpucache_up == PARTIAL_AC) {
2113 set_up_list3s(cachep, SIZE_L3);
2114 g_cpucache_up = PARTIAL_L3;
2117 for_each_online_node(node) {
2118 cachep->nodelists[node] =
2119 kmalloc_node(sizeof(struct kmem_list3),
2121 BUG_ON(!cachep->nodelists[node]);
2122 kmem_list3_init(cachep->nodelists[node]);
2126 cachep->nodelists[numa_mem_id()]->next_reap =
2127 jiffies + REAPTIMEOUT_LIST3 +
2128 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2130 cpu_cache_get(cachep)->avail = 0;
2131 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2132 cpu_cache_get(cachep)->batchcount = 1;
2133 cpu_cache_get(cachep)->touched = 0;
2134 cachep->batchcount = 1;
2135 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2140 * kmem_cache_create - Create a cache.
2141 * @name: A string which is used in /proc/slabinfo to identify this cache.
2142 * @size: The size of objects to be created in this cache.
2143 * @align: The required alignment for the objects.
2144 * @flags: SLAB flags
2145 * @ctor: A constructor for the objects.
2147 * Returns a ptr to the cache on success, NULL on failure.
2148 * Cannot be called within a int, but can be interrupted.
2149 * The @ctor is run when new pages are allocated by the cache.
2151 * @name must be valid until the cache is destroyed. This implies that
2152 * the module calling this has to destroy the cache before getting unloaded.
2153 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2154 * therefore applications must manage it themselves.
2158 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2159 * to catch references to uninitialised memory.
2161 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2162 * for buffer overruns.
2164 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2165 * cacheline. This can be beneficial if you're counting cycles as closely
2169 kmem_cache_create (const char *name, size_t size, size_t align,
2170 unsigned long flags, void (*ctor)(void *))
2172 size_t left_over, slab_size, ralign;
2173 struct kmem_cache *cachep = NULL, *pc;
2177 * Sanity checks... these are all serious usage bugs.
2179 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2180 size > KMALLOC_MAX_SIZE) {
2181 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2187 * We use cache_chain_mutex to ensure a consistent view of
2188 * cpu_online_mask as well. Please see cpuup_callback
2190 if (slab_is_available()) {
2192 mutex_lock(&cache_chain_mutex);
2195 list_for_each_entry(pc, &cache_chain, next) {
2200 * This happens when the module gets unloaded and doesn't
2201 * destroy its slab cache and no-one else reuses the vmalloc
2202 * area of the module. Print a warning.
2204 res = probe_kernel_address(pc->name, tmp);
2207 "SLAB: cache with size %d has lost its name\n",
2212 if (!strcmp(pc->name, name)) {
2214 "kmem_cache_create: duplicate cache %s\n", name);
2221 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2224 * Enable redzoning and last user accounting, except for caches with
2225 * large objects, if the increased size would increase the object size
2226 * above the next power of two: caches with object sizes just above a
2227 * power of two have a significant amount of internal fragmentation.
2229 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2230 2 * sizeof(unsigned long long)))
2231 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2232 if (!(flags & SLAB_DESTROY_BY_RCU))
2233 flags |= SLAB_POISON;
2235 if (flags & SLAB_DESTROY_BY_RCU)
2236 BUG_ON(flags & SLAB_POISON);
2239 * Always checks flags, a caller might be expecting debug support which
2242 BUG_ON(flags & ~CREATE_MASK);
2245 * Check that size is in terms of words. This is needed to avoid
2246 * unaligned accesses for some archs when redzoning is used, and makes
2247 * sure any on-slab bufctl's are also correctly aligned.
2249 if (size & (BYTES_PER_WORD - 1)) {
2250 size += (BYTES_PER_WORD - 1);
2251 size &= ~(BYTES_PER_WORD - 1);
2254 /* calculate the final buffer alignment: */
2256 /* 1) arch recommendation: can be overridden for debug */
2257 if (flags & SLAB_HWCACHE_ALIGN) {
2259 * Default alignment: as specified by the arch code. Except if
2260 * an object is really small, then squeeze multiple objects into
2263 ralign = cache_line_size();
2264 while (size <= ralign / 2)
2267 ralign = BYTES_PER_WORD;
2271 * Redzoning and user store require word alignment or possibly larger.
2272 * Note this will be overridden by architecture or caller mandated
2273 * alignment if either is greater than BYTES_PER_WORD.
2275 if (flags & SLAB_STORE_USER)
2276 ralign = BYTES_PER_WORD;
2278 if (flags & SLAB_RED_ZONE) {
2279 ralign = REDZONE_ALIGN;
2280 /* If redzoning, ensure that the second redzone is suitably
2281 * aligned, by adjusting the object size accordingly. */
2282 size += REDZONE_ALIGN - 1;
2283 size &= ~(REDZONE_ALIGN - 1);
2286 /* 2) arch mandated alignment */
2287 if (ralign < ARCH_SLAB_MINALIGN) {
2288 ralign = ARCH_SLAB_MINALIGN;
2290 /* 3) caller mandated alignment */
2291 if (ralign < align) {
2294 /* disable debug if not aligning with REDZONE_ALIGN */
2295 if (ralign & (__alignof__(unsigned long long) - 1))
2296 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2302 if (slab_is_available())
2307 /* Get cache's description obj. */
2308 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2313 cachep->obj_size = size;
2316 * Both debugging options require word-alignment which is calculated
2319 if (flags & SLAB_RED_ZONE) {
2320 /* add space for red zone words */
2321 cachep->obj_offset += align;
2322 size += align + sizeof(unsigned long long);
2324 if (flags & SLAB_STORE_USER) {
2325 /* user store requires one word storage behind the end of
2326 * the real object. But if the second red zone needs to be
2327 * aligned to 64 bits, we must allow that much space.
2329 if (flags & SLAB_RED_ZONE)
2330 size += REDZONE_ALIGN;
2332 size += BYTES_PER_WORD;
2334 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2335 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2336 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2337 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2344 * Determine if the slab management is 'on' or 'off' slab.
2345 * (bootstrapping cannot cope with offslab caches so don't do
2346 * it too early on. Always use on-slab management when
2347 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2349 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2350 !(flags & SLAB_NOLEAKTRACE))
2352 * Size is large, assume best to place the slab management obj
2353 * off-slab (should allow better packing of objs).
2355 flags |= CFLGS_OFF_SLAB;
2357 size = ALIGN(size, align);
2359 left_over = calculate_slab_order(cachep, size, align, flags);
2363 "kmem_cache_create: couldn't create cache %s.\n", name);
2364 kmem_cache_free(&cache_cache, cachep);
2368 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2369 + sizeof(struct slab), align);
2372 * If the slab has been placed off-slab, and we have enough space then
2373 * move it on-slab. This is at the expense of any extra colouring.
2375 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2376 flags &= ~CFLGS_OFF_SLAB;
2377 left_over -= slab_size;
2380 if (flags & CFLGS_OFF_SLAB) {
2381 /* really off slab. No need for manual alignment */
2383 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2385 #ifdef CONFIG_PAGE_POISONING
2386 /* If we're going to use the generic kernel_map_pages()
2387 * poisoning, then it's going to smash the contents of
2388 * the redzone and userword anyhow, so switch them off.
2390 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2391 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2395 cachep->colour_off = cache_line_size();
2396 /* Offset must be a multiple of the alignment. */
2397 if (cachep->colour_off < align)
2398 cachep->colour_off = align;
2399 cachep->colour = left_over / cachep->colour_off;
2400 cachep->slab_size = slab_size;
2401 cachep->flags = flags;
2402 cachep->gfpflags = 0;
2403 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2404 cachep->gfpflags |= GFP_DMA;
2405 cachep->buffer_size = size;
2406 cachep->reciprocal_buffer_size = reciprocal_value(size);
2408 if (flags & CFLGS_OFF_SLAB) {
2409 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2411 * This is a possibility for one of the malloc_sizes caches.
2412 * But since we go off slab only for object size greater than
2413 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2414 * this should not happen at all.
2415 * But leave a BUG_ON for some lucky dude.
2417 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2419 cachep->ctor = ctor;
2420 cachep->name = name;
2422 if (setup_cpu_cache(cachep, gfp)) {
2423 __kmem_cache_destroy(cachep);
2428 /* cache setup completed, link it into the list */
2429 list_add(&cachep->next, &cache_chain);
2431 if (!cachep && (flags & SLAB_PANIC))
2432 panic("kmem_cache_create(): failed to create slab `%s'\n",
2434 if (slab_is_available()) {
2435 mutex_unlock(&cache_chain_mutex);
2440 EXPORT_SYMBOL(kmem_cache_create);
2443 static void check_irq_off(void)
2445 BUG_ON(!irqs_disabled());
2448 static void check_irq_on(void)
2450 BUG_ON(irqs_disabled());
2453 static void check_spinlock_acquired(struct kmem_cache *cachep)
2457 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2461 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2465 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2470 #define check_irq_off() do { } while(0)
2471 #define check_irq_on() do { } while(0)
2472 #define check_spinlock_acquired(x) do { } while(0)
2473 #define check_spinlock_acquired_node(x, y) do { } while(0)
2476 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2477 struct array_cache *ac,
2478 int force, int node);
2480 static void do_drain(void *arg)
2482 struct kmem_cache *cachep = arg;
2483 struct array_cache *ac;
2484 int node = numa_mem_id();
2487 ac = cpu_cache_get(cachep);
2488 spin_lock(&cachep->nodelists[node]->list_lock);
2489 free_block(cachep, ac->entry, ac->avail, node);
2490 spin_unlock(&cachep->nodelists[node]->list_lock);
2494 static void drain_cpu_caches(struct kmem_cache *cachep)
2496 struct kmem_list3 *l3;
2499 on_each_cpu(do_drain, cachep, 1);
2501 for_each_online_node(node) {
2502 l3 = cachep->nodelists[node];
2503 if (l3 && l3->alien)
2504 drain_alien_cache(cachep, l3->alien);
2507 for_each_online_node(node) {
2508 l3 = cachep->nodelists[node];
2510 drain_array(cachep, l3, l3->shared, 1, node);
2515 * Remove slabs from the list of free slabs.
2516 * Specify the number of slabs to drain in tofree.
2518 * Returns the actual number of slabs released.
2520 static int drain_freelist(struct kmem_cache *cache,
2521 struct kmem_list3 *l3, int tofree)
2523 struct list_head *p;
2528 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2530 spin_lock_irq(&l3->list_lock);
2531 p = l3->slabs_free.prev;
2532 if (p == &l3->slabs_free) {
2533 spin_unlock_irq(&l3->list_lock);
2537 slabp = list_entry(p, struct slab, list);
2539 BUG_ON(slabp->inuse);
2541 list_del(&slabp->list);
2543 * Safe to drop the lock. The slab is no longer linked
2546 l3->free_objects -= cache->num;
2547 spin_unlock_irq(&l3->list_lock);
2548 slab_destroy(cache, slabp);
2555 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2556 static int __cache_shrink(struct kmem_cache *cachep)
2559 struct kmem_list3 *l3;
2561 drain_cpu_caches(cachep);
2564 for_each_online_node(i) {
2565 l3 = cachep->nodelists[i];
2569 drain_freelist(cachep, l3, l3->free_objects);
2571 ret += !list_empty(&l3->slabs_full) ||
2572 !list_empty(&l3->slabs_partial);
2574 return (ret ? 1 : 0);
2578 * kmem_cache_shrink - Shrink a cache.
2579 * @cachep: The cache to shrink.
2581 * Releases as many slabs as possible for a cache.
2582 * To help debugging, a zero exit status indicates all slabs were released.
2584 int kmem_cache_shrink(struct kmem_cache *cachep)
2587 BUG_ON(!cachep || in_interrupt());
2590 mutex_lock(&cache_chain_mutex);
2591 ret = __cache_shrink(cachep);
2592 mutex_unlock(&cache_chain_mutex);
2596 EXPORT_SYMBOL(kmem_cache_shrink);
2599 * kmem_cache_destroy - delete a cache
2600 * @cachep: the cache to destroy
2602 * Remove a &struct kmem_cache object from the slab cache.
2604 * It is expected this function will be called by a module when it is
2605 * unloaded. This will remove the cache completely, and avoid a duplicate
2606 * cache being allocated each time a module is loaded and unloaded, if the
2607 * module doesn't have persistent in-kernel storage across loads and unloads.
2609 * The cache must be empty before calling this function.
2611 * The caller must guarantee that noone will allocate memory from the cache
2612 * during the kmem_cache_destroy().
2614 void kmem_cache_destroy(struct kmem_cache *cachep)
2616 BUG_ON(!cachep || in_interrupt());
2618 /* Find the cache in the chain of caches. */
2620 mutex_lock(&cache_chain_mutex);
2622 * the chain is never empty, cache_cache is never destroyed
2624 list_del(&cachep->next);
2625 if (__cache_shrink(cachep)) {
2626 slab_error(cachep, "Can't free all objects");
2627 list_add(&cachep->next, &cache_chain);
2628 mutex_unlock(&cache_chain_mutex);
2633 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2636 __kmem_cache_destroy(cachep);
2637 mutex_unlock(&cache_chain_mutex);
2640 EXPORT_SYMBOL(kmem_cache_destroy);
2643 * Get the memory for a slab management obj.
2644 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2645 * always come from malloc_sizes caches. The slab descriptor cannot
2646 * come from the same cache which is getting created because,
2647 * when we are searching for an appropriate cache for these
2648 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2649 * If we are creating a malloc_sizes cache here it would not be visible to
2650 * kmem_find_general_cachep till the initialization is complete.
2651 * Hence we cannot have slabp_cache same as the original cache.
2653 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2654 int colour_off, gfp_t local_flags,
2659 if (OFF_SLAB(cachep)) {
2660 /* Slab management obj is off-slab. */
2661 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2662 local_flags, nodeid);
2664 * If the first object in the slab is leaked (it's allocated
2665 * but no one has a reference to it), we want to make sure
2666 * kmemleak does not treat the ->s_mem pointer as a reference
2667 * to the object. Otherwise we will not report the leak.
2669 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2674 slabp = objp + colour_off;
2675 colour_off += cachep->slab_size;
2678 slabp->colouroff = colour_off;
2679 slabp->s_mem = objp + colour_off;
2680 slabp->nodeid = nodeid;
2685 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2687 return (kmem_bufctl_t *) (slabp + 1);
2690 static void cache_init_objs(struct kmem_cache *cachep,
2695 for (i = 0; i < cachep->num; i++) {
2696 void *objp = index_to_obj(cachep, slabp, i);
2698 /* need to poison the objs? */
2699 if (cachep->flags & SLAB_POISON)
2700 poison_obj(cachep, objp, POISON_FREE);
2701 if (cachep->flags & SLAB_STORE_USER)
2702 *dbg_userword(cachep, objp) = NULL;
2704 if (cachep->flags & SLAB_RED_ZONE) {
2705 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2706 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2709 * Constructors are not allowed to allocate memory from the same
2710 * cache which they are a constructor for. Otherwise, deadlock.
2711 * They must also be threaded.
2713 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2714 cachep->ctor(objp + obj_offset(cachep));
2716 if (cachep->flags & SLAB_RED_ZONE) {
2717 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2718 slab_error(cachep, "constructor overwrote the"
2719 " end of an object");
2720 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2721 slab_error(cachep, "constructor overwrote the"
2722 " start of an object");
2724 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2725 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2726 kernel_map_pages(virt_to_page(objp),
2727 cachep->buffer_size / PAGE_SIZE, 0);
2732 slab_bufctl(slabp)[i] = i + 1;
2734 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2737 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2739 if (CONFIG_ZONE_DMA_FLAG) {
2740 if (flags & GFP_DMA)
2741 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2743 BUG_ON(cachep->gfpflags & GFP_DMA);
2747 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2750 void *objp = index_to_obj(cachep, slabp, slabp->free);
2754 next = slab_bufctl(slabp)[slabp->free];
2756 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2757 WARN_ON(slabp->nodeid != nodeid);
2764 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2765 void *objp, int nodeid)
2767 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2770 /* Verify that the slab belongs to the intended node */
2771 WARN_ON(slabp->nodeid != nodeid);
2773 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2774 printk(KERN_ERR "slab: double free detected in cache "
2775 "'%s', objp %p\n", cachep->name, objp);
2779 slab_bufctl(slabp)[objnr] = slabp->free;
2780 slabp->free = objnr;
2785 * Map pages beginning at addr to the given cache and slab. This is required
2786 * for the slab allocator to be able to lookup the cache and slab of a
2787 * virtual address for kfree, ksize, and slab debugging.
2789 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2795 page = virt_to_page(addr);
2798 if (likely(!PageCompound(page)))
2799 nr_pages <<= cache->gfporder;
2802 page_set_cache(page, cache);
2803 page_set_slab(page, slab);
2805 } while (--nr_pages);
2809 * Grow (by 1) the number of slabs within a cache. This is called by
2810 * kmem_cache_alloc() when there are no active objs left in a cache.
2812 static int cache_grow(struct kmem_cache *cachep,
2813 gfp_t flags, int nodeid, void *objp)
2818 struct kmem_list3 *l3;
2821 * Be lazy and only check for valid flags here, keeping it out of the
2822 * critical path in kmem_cache_alloc().
2824 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2825 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2827 /* Take the l3 list lock to change the colour_next on this node */
2829 l3 = cachep->nodelists[nodeid];
2830 spin_lock(&l3->list_lock);
2832 /* Get colour for the slab, and cal the next value. */
2833 offset = l3->colour_next;
2835 if (l3->colour_next >= cachep->colour)
2836 l3->colour_next = 0;
2837 spin_unlock(&l3->list_lock);
2839 offset *= cachep->colour_off;
2841 if (local_flags & __GFP_WAIT)
2845 * The test for missing atomic flag is performed here, rather than
2846 * the more obvious place, simply to reduce the critical path length
2847 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2848 * will eventually be caught here (where it matters).
2850 kmem_flagcheck(cachep, flags);
2853 * Get mem for the objs. Attempt to allocate a physical page from
2857 objp = kmem_getpages(cachep, local_flags, nodeid);
2861 /* Get slab management. */
2862 slabp = alloc_slabmgmt(cachep, objp, offset,
2863 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2867 slab_map_pages(cachep, slabp, objp);
2869 cache_init_objs(cachep, slabp);
2871 if (local_flags & __GFP_WAIT)
2872 local_irq_disable();
2874 spin_lock(&l3->list_lock);
2876 /* Make slab active. */
2877 list_add_tail(&slabp->list, &(l3->slabs_free));
2878 STATS_INC_GROWN(cachep);
2879 l3->free_objects += cachep->num;
2880 spin_unlock(&l3->list_lock);
2883 kmem_freepages(cachep, objp);
2885 if (local_flags & __GFP_WAIT)
2886 local_irq_disable();
2893 * Perform extra freeing checks:
2894 * - detect bad pointers.
2895 * - POISON/RED_ZONE checking
2897 static void kfree_debugcheck(const void *objp)
2899 if (!virt_addr_valid(objp)) {
2900 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2901 (unsigned long)objp);
2906 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2908 unsigned long long redzone1, redzone2;
2910 redzone1 = *dbg_redzone1(cache, obj);
2911 redzone2 = *dbg_redzone2(cache, obj);
2916 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2919 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2920 slab_error(cache, "double free detected");
2922 slab_error(cache, "memory outside object was overwritten");
2924 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2925 obj, redzone1, redzone2);
2928 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2935 BUG_ON(virt_to_cache(objp) != cachep);
2937 objp -= obj_offset(cachep);
2938 kfree_debugcheck(objp);
2939 page = virt_to_head_page(objp);
2941 slabp = page_get_slab(page);
2943 if (cachep->flags & SLAB_RED_ZONE) {
2944 verify_redzone_free(cachep, objp);
2945 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2946 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2948 if (cachep->flags & SLAB_STORE_USER)
2949 *dbg_userword(cachep, objp) = caller;
2951 objnr = obj_to_index(cachep, slabp, objp);
2953 BUG_ON(objnr >= cachep->num);
2954 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2956 #ifdef CONFIG_DEBUG_SLAB_LEAK
2957 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2959 if (cachep->flags & SLAB_POISON) {
2960 #ifdef CONFIG_DEBUG_PAGEALLOC
2961 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2962 store_stackinfo(cachep, objp, (unsigned long)caller);
2963 kernel_map_pages(virt_to_page(objp),
2964 cachep->buffer_size / PAGE_SIZE, 0);
2966 poison_obj(cachep, objp, POISON_FREE);
2969 poison_obj(cachep, objp, POISON_FREE);
2975 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2980 /* Check slab's freelist to see if this obj is there. */
2981 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2983 if (entries > cachep->num || i >= cachep->num)
2986 if (entries != cachep->num - slabp->inuse) {
2988 printk(KERN_ERR "slab: Internal list corruption detected in "
2989 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2990 cachep->name, cachep->num, slabp, slabp->inuse);
2992 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2995 printk("\n%03x:", i);
2996 printk(" %02x", ((unsigned char *)slabp)[i]);
3003 #define kfree_debugcheck(x) do { } while(0)
3004 #define cache_free_debugcheck(x,objp,z) (objp)
3005 #define check_slabp(x,y) do { } while(0)
3008 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3011 struct kmem_list3 *l3;
3012 struct array_cache *ac;
3017 node = numa_mem_id();
3018 ac = cpu_cache_get(cachep);
3019 batchcount = ac->batchcount;
3020 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3022 * If there was little recent activity on this cache, then
3023 * perform only a partial refill. Otherwise we could generate
3026 batchcount = BATCHREFILL_LIMIT;
3028 l3 = cachep->nodelists[node];
3030 BUG_ON(ac->avail > 0 || !l3);
3031 spin_lock(&l3->list_lock);
3033 /* See if we can refill from the shared array */
3034 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3035 l3->shared->touched = 1;
3039 while (batchcount > 0) {
3040 struct list_head *entry;
3042 /* Get slab alloc is to come from. */
3043 entry = l3->slabs_partial.next;
3044 if (entry == &l3->slabs_partial) {
3045 l3->free_touched = 1;
3046 entry = l3->slabs_free.next;
3047 if (entry == &l3->slabs_free)
3051 slabp = list_entry(entry, struct slab, list);
3052 check_slabp(cachep, slabp);
3053 check_spinlock_acquired(cachep);
3056 * The slab was either on partial or free list so
3057 * there must be at least one object available for
3060 BUG_ON(slabp->inuse >= cachep->num);
3062 while (slabp->inuse < cachep->num && batchcount--) {
3063 STATS_INC_ALLOCED(cachep);
3064 STATS_INC_ACTIVE(cachep);
3065 STATS_SET_HIGH(cachep);
3067 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3070 check_slabp(cachep, slabp);
3072 /* move slabp to correct slabp list: */
3073 list_del(&slabp->list);
3074 if (slabp->free == BUFCTL_END)
3075 list_add(&slabp->list, &l3->slabs_full);
3077 list_add(&slabp->list, &l3->slabs_partial);
3081 l3->free_objects -= ac->avail;
3083 spin_unlock(&l3->list_lock);
3085 if (unlikely(!ac->avail)) {
3087 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3089 /* cache_grow can reenable interrupts, then ac could change. */
3090 ac = cpu_cache_get(cachep);
3091 if (!x && ac->avail == 0) /* no objects in sight? abort */
3094 if (!ac->avail) /* objects refilled by interrupt? */
3098 return ac->entry[--ac->avail];
3101 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3104 might_sleep_if(flags & __GFP_WAIT);
3106 kmem_flagcheck(cachep, flags);
3111 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3112 gfp_t flags, void *objp, void *caller)
3116 if (cachep->flags & SLAB_POISON) {
3117 #ifdef CONFIG_DEBUG_PAGEALLOC
3118 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3119 kernel_map_pages(virt_to_page(objp),
3120 cachep->buffer_size / PAGE_SIZE, 1);
3122 check_poison_obj(cachep, objp);
3124 check_poison_obj(cachep, objp);
3126 poison_obj(cachep, objp, POISON_INUSE);
3128 if (cachep->flags & SLAB_STORE_USER)
3129 *dbg_userword(cachep, objp) = caller;
3131 if (cachep->flags & SLAB_RED_ZONE) {
3132 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3133 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3134 slab_error(cachep, "double free, or memory outside"
3135 " object was overwritten");
3137 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3138 objp, *dbg_redzone1(cachep, objp),
3139 *dbg_redzone2(cachep, objp));
3141 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3142 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3144 #ifdef CONFIG_DEBUG_SLAB_LEAK
3149 slabp = page_get_slab(virt_to_head_page(objp));
3150 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3151 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3154 objp += obj_offset(cachep);
3155 if (cachep->ctor && cachep->flags & SLAB_POISON)
3157 #if ARCH_SLAB_MINALIGN
3158 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3159 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3160 objp, ARCH_SLAB_MINALIGN);
3166 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3169 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3171 if (cachep == &cache_cache)
3174 return should_failslab(obj_size(cachep), flags, cachep->flags);
3177 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3180 struct array_cache *ac;
3184 ac = cpu_cache_get(cachep);
3185 if (likely(ac->avail)) {
3186 STATS_INC_ALLOCHIT(cachep);
3188 objp = ac->entry[--ac->avail];
3190 STATS_INC_ALLOCMISS(cachep);
3191 objp = cache_alloc_refill(cachep, flags);
3193 * the 'ac' may be updated by cache_alloc_refill(),
3194 * and kmemleak_erase() requires its correct value.
3196 ac = cpu_cache_get(cachep);
3199 * To avoid a false negative, if an object that is in one of the
3200 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3201 * treat the array pointers as a reference to the object.
3204 kmemleak_erase(&ac->entry[ac->avail]);
3210 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3212 * If we are in_interrupt, then process context, including cpusets and
3213 * mempolicy, may not apply and should not be used for allocation policy.
3215 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3217 int nid_alloc, nid_here;
3219 if (in_interrupt() || (flags & __GFP_THISNODE))
3221 nid_alloc = nid_here = numa_mem_id();
3223 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3224 nid_alloc = cpuset_slab_spread_node();
3225 else if (current->mempolicy)
3226 nid_alloc = slab_node(current->mempolicy);
3228 if (nid_alloc != nid_here)
3229 return ____cache_alloc_node(cachep, flags, nid_alloc);
3234 * Fallback function if there was no memory available and no objects on a
3235 * certain node and fall back is permitted. First we scan all the
3236 * available nodelists for available objects. If that fails then we
3237 * perform an allocation without specifying a node. This allows the page
3238 * allocator to do its reclaim / fallback magic. We then insert the
3239 * slab into the proper nodelist and then allocate from it.
3241 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3243 struct zonelist *zonelist;
3247 enum zone_type high_zoneidx = gfp_zone(flags);
3251 if (flags & __GFP_THISNODE)
3255 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3256 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3260 * Look through allowed nodes for objects available
3261 * from existing per node queues.
3263 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3264 nid = zone_to_nid(zone);
3266 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3267 cache->nodelists[nid] &&
3268 cache->nodelists[nid]->free_objects) {
3269 obj = ____cache_alloc_node(cache,
3270 flags | GFP_THISNODE, nid);
3278 * This allocation will be performed within the constraints
3279 * of the current cpuset / memory policy requirements.
3280 * We may trigger various forms of reclaim on the allowed
3281 * set and go into memory reserves if necessary.
3283 if (local_flags & __GFP_WAIT)
3285 kmem_flagcheck(cache, flags);
3286 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3287 if (local_flags & __GFP_WAIT)
3288 local_irq_disable();
3291 * Insert into the appropriate per node queues
3293 nid = page_to_nid(virt_to_page(obj));
3294 if (cache_grow(cache, flags, nid, obj)) {
3295 obj = ____cache_alloc_node(cache,
3296 flags | GFP_THISNODE, nid);
3299 * Another processor may allocate the
3300 * objects in the slab since we are
3301 * not holding any locks.
3305 /* cache_grow already freed obj */
3315 * A interface to enable slab creation on nodeid
3317 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3320 struct list_head *entry;
3322 struct kmem_list3 *l3;
3326 l3 = cachep->nodelists[nodeid];
3331 spin_lock(&l3->list_lock);
3332 entry = l3->slabs_partial.next;
3333 if (entry == &l3->slabs_partial) {
3334 l3->free_touched = 1;
3335 entry = l3->slabs_free.next;
3336 if (entry == &l3->slabs_free)
3340 slabp = list_entry(entry, struct slab, list);
3341 check_spinlock_acquired_node(cachep, nodeid);
3342 check_slabp(cachep, slabp);
3344 STATS_INC_NODEALLOCS(cachep);
3345 STATS_INC_ACTIVE(cachep);
3346 STATS_SET_HIGH(cachep);
3348 BUG_ON(slabp->inuse == cachep->num);
3350 obj = slab_get_obj(cachep, slabp, nodeid);
3351 check_slabp(cachep, slabp);
3353 /* move slabp to correct slabp list: */
3354 list_del(&slabp->list);
3356 if (slabp->free == BUFCTL_END)
3357 list_add(&slabp->list, &l3->slabs_full);
3359 list_add(&slabp->list, &l3->slabs_partial);
3361 spin_unlock(&l3->list_lock);
3365 spin_unlock(&l3->list_lock);
3366 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3370 return fallback_alloc(cachep, flags);
3377 * kmem_cache_alloc_node - Allocate an object on the specified node
3378 * @cachep: The cache to allocate from.
3379 * @flags: See kmalloc().
3380 * @nodeid: node number of the target node.
3381 * @caller: return address of caller, used for debug information
3383 * Identical to kmem_cache_alloc but it will allocate memory on the given
3384 * node, which can improve the performance for cpu bound structures.
3386 * Fallback to other node is possible if __GFP_THISNODE is not set.
3388 static __always_inline void *
3389 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3392 unsigned long save_flags;
3394 int slab_node = numa_mem_id();
3396 flags &= gfp_allowed_mask;
3398 lockdep_trace_alloc(flags);
3400 if (slab_should_failslab(cachep, flags))
3403 cache_alloc_debugcheck_before(cachep, flags);
3404 local_irq_save(save_flags);
3409 if (unlikely(!cachep->nodelists[nodeid])) {
3410 /* Node not bootstrapped yet */
3411 ptr = fallback_alloc(cachep, flags);
3415 if (nodeid == slab_node) {
3417 * Use the locally cached objects if possible.
3418 * However ____cache_alloc does not allow fallback
3419 * to other nodes. It may fail while we still have
3420 * objects on other nodes available.
3422 ptr = ____cache_alloc(cachep, flags);
3426 /* ___cache_alloc_node can fall back to other nodes */
3427 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3429 local_irq_restore(save_flags);
3430 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3431 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3435 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3437 if (unlikely((flags & __GFP_ZERO) && ptr))
3438 memset(ptr, 0, obj_size(cachep));
3443 static __always_inline void *
3444 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3448 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3449 objp = alternate_node_alloc(cache, flags);
3453 objp = ____cache_alloc(cache, flags);
3456 * We may just have run out of memory on the local node.
3457 * ____cache_alloc_node() knows how to locate memory on other nodes
3460 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3467 static __always_inline void *
3468 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3470 return ____cache_alloc(cachep, flags);
3473 #endif /* CONFIG_NUMA */
3475 static __always_inline void *
3476 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3478 unsigned long save_flags;
3481 flags &= gfp_allowed_mask;
3483 lockdep_trace_alloc(flags);
3485 if (slab_should_failslab(cachep, flags))
3488 cache_alloc_debugcheck_before(cachep, flags);
3489 local_irq_save(save_flags);
3490 objp = __do_cache_alloc(cachep, flags);
3491 local_irq_restore(save_flags);
3492 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3493 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3498 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3500 if (unlikely((flags & __GFP_ZERO) && objp))
3501 memset(objp, 0, obj_size(cachep));
3507 * Caller needs to acquire correct kmem_list's list_lock
3509 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3513 struct kmem_list3 *l3;
3515 for (i = 0; i < nr_objects; i++) {
3516 void *objp = objpp[i];
3519 slabp = virt_to_slab(objp);
3520 l3 = cachep->nodelists[node];
3521 list_del(&slabp->list);
3522 check_spinlock_acquired_node(cachep, node);
3523 check_slabp(cachep, slabp);
3524 slab_put_obj(cachep, slabp, objp, node);
3525 STATS_DEC_ACTIVE(cachep);
3527 check_slabp(cachep, slabp);
3529 /* fixup slab chains */
3530 if (slabp->inuse == 0) {
3531 if (l3->free_objects > l3->free_limit) {
3532 l3->free_objects -= cachep->num;
3533 /* No need to drop any previously held
3534 * lock here, even if we have a off-slab slab
3535 * descriptor it is guaranteed to come from
3536 * a different cache, refer to comments before
3539 slab_destroy(cachep, slabp);
3541 list_add(&slabp->list, &l3->slabs_free);
3544 /* Unconditionally move a slab to the end of the
3545 * partial list on free - maximum time for the
3546 * other objects to be freed, too.
3548 list_add_tail(&slabp->list, &l3->slabs_partial);
3553 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3556 struct kmem_list3 *l3;
3557 int node = numa_mem_id();
3559 batchcount = ac->batchcount;
3561 BUG_ON(!batchcount || batchcount > ac->avail);
3564 l3 = cachep->nodelists[node];
3565 spin_lock(&l3->list_lock);
3567 struct array_cache *shared_array = l3->shared;
3568 int max = shared_array->limit - shared_array->avail;
3570 if (batchcount > max)
3572 memcpy(&(shared_array->entry[shared_array->avail]),
3573 ac->entry, sizeof(void *) * batchcount);
3574 shared_array->avail += batchcount;
3579 free_block(cachep, ac->entry, batchcount, node);
3584 struct list_head *p;
3586 p = l3->slabs_free.next;
3587 while (p != &(l3->slabs_free)) {
3590 slabp = list_entry(p, struct slab, list);
3591 BUG_ON(slabp->inuse);
3596 STATS_SET_FREEABLE(cachep, i);
3599 spin_unlock(&l3->list_lock);
3600 ac->avail -= batchcount;
3601 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3605 * Release an obj back to its cache. If the obj has a constructed state, it must
3606 * be in this state _before_ it is released. Called with disabled ints.
3608 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3610 struct array_cache *ac = cpu_cache_get(cachep);
3613 kmemleak_free_recursive(objp, cachep->flags);
3614 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3616 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3619 * Skip calling cache_free_alien() when the platform is not numa.
3620 * This will avoid cache misses that happen while accessing slabp (which
3621 * is per page memory reference) to get nodeid. Instead use a global
3622 * variable to skip the call, which is mostly likely to be present in
3625 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3628 if (likely(ac->avail < ac->limit)) {
3629 STATS_INC_FREEHIT(cachep);
3630 ac->entry[ac->avail++] = objp;
3633 STATS_INC_FREEMISS(cachep);
3634 cache_flusharray(cachep, ac);
3635 ac->entry[ac->avail++] = objp;
3640 * kmem_cache_alloc - Allocate an object
3641 * @cachep: The cache to allocate from.
3642 * @flags: See kmalloc().
3644 * Allocate an object from this cache. The flags are only relevant
3645 * if the cache has no available objects.
3647 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3649 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3651 trace_kmem_cache_alloc(_RET_IP_, ret,
3652 obj_size(cachep), cachep->buffer_size, flags);
3656 EXPORT_SYMBOL(kmem_cache_alloc);
3658 #ifdef CONFIG_TRACING
3660 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3664 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3666 trace_kmalloc(_RET_IP_, ret,
3667 size, slab_buffer_size(cachep), flags);
3670 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3674 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3676 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3677 __builtin_return_address(0));
3679 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3680 obj_size(cachep), cachep->buffer_size,
3685 EXPORT_SYMBOL(kmem_cache_alloc_node);
3687 #ifdef CONFIG_TRACING
3688 void *kmem_cache_alloc_node_trace(size_t size,
3689 struct kmem_cache *cachep,
3695 ret = __cache_alloc_node(cachep, flags, nodeid,
3696 __builtin_return_address(0));
3697 trace_kmalloc_node(_RET_IP_, ret,
3698 size, slab_buffer_size(cachep),
3702 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3705 static __always_inline void *
3706 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3708 struct kmem_cache *cachep;
3710 cachep = kmem_find_general_cachep(size, flags);
3711 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3713 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3716 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3717 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3719 return __do_kmalloc_node(size, flags, node,
3720 __builtin_return_address(0));
3722 EXPORT_SYMBOL(__kmalloc_node);
3724 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3725 int node, unsigned long caller)
3727 return __do_kmalloc_node(size, flags, node, (void *)caller);
3729 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3731 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3733 return __do_kmalloc_node(size, flags, node, NULL);
3735 EXPORT_SYMBOL(__kmalloc_node);
3736 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3737 #endif /* CONFIG_NUMA */
3740 * __do_kmalloc - allocate memory
3741 * @size: how many bytes of memory are required.
3742 * @flags: the type of memory to allocate (see kmalloc).
3743 * @caller: function caller for debug tracking of the caller
3745 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3748 struct kmem_cache *cachep;
3751 /* If you want to save a few bytes .text space: replace
3753 * Then kmalloc uses the uninlined functions instead of the inline
3756 cachep = __find_general_cachep(size, flags);
3757 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3759 ret = __cache_alloc(cachep, flags, caller);
3761 trace_kmalloc((unsigned long) caller, ret,
3762 size, cachep->buffer_size, flags);
3768 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3769 void *__kmalloc(size_t size, gfp_t flags)
3771 return __do_kmalloc(size, flags, __builtin_return_address(0));
3773 EXPORT_SYMBOL(__kmalloc);
3775 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3777 return __do_kmalloc(size, flags, (void *)caller);
3779 EXPORT_SYMBOL(__kmalloc_track_caller);
3782 void *__kmalloc(size_t size, gfp_t flags)
3784 return __do_kmalloc(size, flags, NULL);
3786 EXPORT_SYMBOL(__kmalloc);
3790 * kmem_cache_free - Deallocate an object
3791 * @cachep: The cache the allocation was from.
3792 * @objp: The previously allocated object.
3794 * Free an object which was previously allocated from this
3797 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3799 unsigned long flags;
3801 local_irq_save(flags);
3802 debug_check_no_locks_freed(objp, obj_size(cachep));
3803 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3804 debug_check_no_obj_freed(objp, obj_size(cachep));
3805 __cache_free(cachep, objp);
3806 local_irq_restore(flags);
3808 trace_kmem_cache_free(_RET_IP_, objp);
3810 EXPORT_SYMBOL(kmem_cache_free);
3813 * kfree - free previously allocated memory
3814 * @objp: pointer returned by kmalloc.
3816 * If @objp is NULL, no operation is performed.
3818 * Don't free memory not originally allocated by kmalloc()
3819 * or you will run into trouble.
3821 void kfree(const void *objp)
3823 struct kmem_cache *c;
3824 unsigned long flags;
3826 trace_kfree(_RET_IP_, objp);
3828 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3830 local_irq_save(flags);
3831 kfree_debugcheck(objp);
3832 c = virt_to_cache(objp);
3833 debug_check_no_locks_freed(objp, obj_size(c));
3834 debug_check_no_obj_freed(objp, obj_size(c));
3835 __cache_free(c, (void *)objp);
3836 local_irq_restore(flags);
3838 EXPORT_SYMBOL(kfree);
3840 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3842 return obj_size(cachep);
3844 EXPORT_SYMBOL(kmem_cache_size);
3846 const char *kmem_cache_name(struct kmem_cache *cachep)
3848 return cachep->name;
3850 EXPORT_SYMBOL_GPL(kmem_cache_name);
3853 * This initializes kmem_list3 or resizes various caches for all nodes.
3855 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3858 struct kmem_list3 *l3;
3859 struct array_cache *new_shared;
3860 struct array_cache **new_alien = NULL;
3862 for_each_online_node(node) {
3864 if (use_alien_caches) {
3865 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3871 if (cachep->shared) {
3872 new_shared = alloc_arraycache(node,
3873 cachep->shared*cachep->batchcount,
3876 free_alien_cache(new_alien);
3881 l3 = cachep->nodelists[node];
3883 struct array_cache *shared = l3->shared;
3885 spin_lock_irq(&l3->list_lock);
3888 free_block(cachep, shared->entry,
3889 shared->avail, node);
3891 l3->shared = new_shared;
3893 l3->alien = new_alien;
3896 l3->free_limit = (1 + nr_cpus_node(node)) *
3897 cachep->batchcount + cachep->num;
3898 spin_unlock_irq(&l3->list_lock);
3900 free_alien_cache(new_alien);
3903 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3905 free_alien_cache(new_alien);
3910 kmem_list3_init(l3);
3911 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3912 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3913 l3->shared = new_shared;
3914 l3->alien = new_alien;
3915 l3->free_limit = (1 + nr_cpus_node(node)) *
3916 cachep->batchcount + cachep->num;
3917 cachep->nodelists[node] = l3;
3922 if (!cachep->next.next) {
3923 /* Cache is not active yet. Roll back what we did */
3926 if (cachep->nodelists[node]) {
3927 l3 = cachep->nodelists[node];
3930 free_alien_cache(l3->alien);
3932 cachep->nodelists[node] = NULL;
3940 struct ccupdate_struct {
3941 struct kmem_cache *cachep;
3942 struct array_cache *new[NR_CPUS];
3945 static void do_ccupdate_local(void *info)
3947 struct ccupdate_struct *new = info;
3948 struct array_cache *old;
3951 old = cpu_cache_get(new->cachep);
3953 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3954 new->new[smp_processor_id()] = old;
3957 /* Always called with the cache_chain_mutex held */
3958 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3959 int batchcount, int shared, gfp_t gfp)
3961 struct ccupdate_struct *new;
3964 new = kzalloc(sizeof(*new), gfp);
3968 for_each_online_cpu(i) {
3969 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3972 for (i--; i >= 0; i--)
3978 new->cachep = cachep;
3980 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3983 cachep->batchcount = batchcount;
3984 cachep->limit = limit;
3985 cachep->shared = shared;
3987 for_each_online_cpu(i) {
3988 struct array_cache *ccold = new->new[i];
3991 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3992 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3993 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3997 return alloc_kmemlist(cachep, gfp);
4000 /* Called with cache_chain_mutex held always */
4001 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4007 * The head array serves three purposes:
4008 * - create a LIFO ordering, i.e. return objects that are cache-warm
4009 * - reduce the number of spinlock operations.
4010 * - reduce the number of linked list operations on the slab and
4011 * bufctl chains: array operations are cheaper.
4012 * The numbers are guessed, we should auto-tune as described by
4015 if (cachep->buffer_size > 131072)
4017 else if (cachep->buffer_size > PAGE_SIZE)
4019 else if (cachep->buffer_size > 1024)
4021 else if (cachep->buffer_size > 256)
4027 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4028 * allocation behaviour: Most allocs on one cpu, most free operations
4029 * on another cpu. For these cases, an efficient object passing between
4030 * cpus is necessary. This is provided by a shared array. The array
4031 * replaces Bonwick's magazine layer.
4032 * On uniprocessor, it's functionally equivalent (but less efficient)
4033 * to a larger limit. Thus disabled by default.
4036 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4041 * With debugging enabled, large batchcount lead to excessively long
4042 * periods with disabled local interrupts. Limit the batchcount
4047 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4049 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4050 cachep->name, -err);
4055 * Drain an array if it contains any elements taking the l3 lock only if
4056 * necessary. Note that the l3 listlock also protects the array_cache
4057 * if drain_array() is used on the shared array.
4059 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4060 struct array_cache *ac, int force, int node)
4064 if (!ac || !ac->avail)
4066 if (ac->touched && !force) {
4069 spin_lock_irq(&l3->list_lock);
4071 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4072 if (tofree > ac->avail)
4073 tofree = (ac->avail + 1) / 2;
4074 free_block(cachep, ac->entry, tofree, node);
4075 ac->avail -= tofree;
4076 memmove(ac->entry, &(ac->entry[tofree]),
4077 sizeof(void *) * ac->avail);
4079 spin_unlock_irq(&l3->list_lock);
4084 * cache_reap - Reclaim memory from caches.
4085 * @w: work descriptor
4087 * Called from workqueue/eventd every few seconds.
4089 * - clear the per-cpu caches for this CPU.
4090 * - return freeable pages to the main free memory pool.
4092 * If we cannot acquire the cache chain mutex then just give up - we'll try
4093 * again on the next iteration.
4095 static void cache_reap(struct work_struct *w)
4097 struct kmem_cache *searchp;
4098 struct kmem_list3 *l3;
4099 int node = numa_mem_id();
4100 struct delayed_work *work = to_delayed_work(w);
4102 if (!mutex_trylock(&cache_chain_mutex))
4103 /* Give up. Setup the next iteration. */
4106 list_for_each_entry(searchp, &cache_chain, next) {
4110 * We only take the l3 lock if absolutely necessary and we
4111 * have established with reasonable certainty that
4112 * we can do some work if the lock was obtained.
4114 l3 = searchp->nodelists[node];
4116 reap_alien(searchp, l3);
4118 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4121 * These are racy checks but it does not matter
4122 * if we skip one check or scan twice.
4124 if (time_after(l3->next_reap, jiffies))
4127 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4129 drain_array(searchp, l3, l3->shared, 0, node);
4131 if (l3->free_touched)
4132 l3->free_touched = 0;
4136 freed = drain_freelist(searchp, l3, (l3->free_limit +
4137 5 * searchp->num - 1) / (5 * searchp->num));
4138 STATS_ADD_REAPED(searchp, freed);
4144 mutex_unlock(&cache_chain_mutex);
4147 /* Set up the next iteration */
4148 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4151 #ifdef CONFIG_SLABINFO
4153 static void print_slabinfo_header(struct seq_file *m)
4156 * Output format version, so at least we can change it
4157 * without _too_ many complaints.
4160 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4162 seq_puts(m, "slabinfo - version: 2.1\n");
4164 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4165 "<objperslab> <pagesperslab>");
4166 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4167 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4169 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4170 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4171 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4176 static void *s_start(struct seq_file *m, loff_t *pos)
4180 mutex_lock(&cache_chain_mutex);
4182 print_slabinfo_header(m);
4184 return seq_list_start(&cache_chain, *pos);
4187 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4189 return seq_list_next(p, &cache_chain, pos);
4192 static void s_stop(struct seq_file *m, void *p)
4194 mutex_unlock(&cache_chain_mutex);
4197 static int s_show(struct seq_file *m, void *p)
4199 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4201 unsigned long active_objs;
4202 unsigned long num_objs;
4203 unsigned long active_slabs = 0;
4204 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4208 struct kmem_list3 *l3;
4212 for_each_online_node(node) {
4213 l3 = cachep->nodelists[node];
4218 spin_lock_irq(&l3->list_lock);
4220 list_for_each_entry(slabp, &l3->slabs_full, list) {
4221 if (slabp->inuse != cachep->num && !error)
4222 error = "slabs_full accounting error";
4223 active_objs += cachep->num;
4226 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4227 if (slabp->inuse == cachep->num && !error)
4228 error = "slabs_partial inuse accounting error";
4229 if (!slabp->inuse && !error)
4230 error = "slabs_partial/inuse accounting error";
4231 active_objs += slabp->inuse;
4234 list_for_each_entry(slabp, &l3->slabs_free, list) {
4235 if (slabp->inuse && !error)
4236 error = "slabs_free/inuse accounting error";
4239 free_objects += l3->free_objects;
4241 shared_avail += l3->shared->avail;
4243 spin_unlock_irq(&l3->list_lock);
4245 num_slabs += active_slabs;
4246 num_objs = num_slabs * cachep->num;
4247 if (num_objs - active_objs != free_objects && !error)
4248 error = "free_objects accounting error";
4250 name = cachep->name;
4252 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4254 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4255 name, active_objs, num_objs, cachep->buffer_size,
4256 cachep->num, (1 << cachep->gfporder));
4257 seq_printf(m, " : tunables %4u %4u %4u",
4258 cachep->limit, cachep->batchcount, cachep->shared);
4259 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4260 active_slabs, num_slabs, shared_avail);
4263 unsigned long high = cachep->high_mark;
4264 unsigned long allocs = cachep->num_allocations;
4265 unsigned long grown = cachep->grown;
4266 unsigned long reaped = cachep->reaped;
4267 unsigned long errors = cachep->errors;
4268 unsigned long max_freeable = cachep->max_freeable;
4269 unsigned long node_allocs = cachep->node_allocs;
4270 unsigned long node_frees = cachep->node_frees;
4271 unsigned long overflows = cachep->node_overflow;
4273 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4274 "%4lu %4lu %4lu %4lu %4lu",
4275 allocs, high, grown,
4276 reaped, errors, max_freeable, node_allocs,
4277 node_frees, overflows);
4281 unsigned long allochit = atomic_read(&cachep->allochit);
4282 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4283 unsigned long freehit = atomic_read(&cachep->freehit);
4284 unsigned long freemiss = atomic_read(&cachep->freemiss);
4286 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4287 allochit, allocmiss, freehit, freemiss);
4295 * slabinfo_op - iterator that generates /proc/slabinfo
4304 * num-pages-per-slab
4305 * + further values on SMP and with statistics enabled
4308 static const struct seq_operations slabinfo_op = {
4315 #define MAX_SLABINFO_WRITE 128
4317 * slabinfo_write - Tuning for the slab allocator
4319 * @buffer: user buffer
4320 * @count: data length
4323 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4324 size_t count, loff_t *ppos)
4326 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4327 int limit, batchcount, shared, res;
4328 struct kmem_cache *cachep;
4330 if (count > MAX_SLABINFO_WRITE)
4332 if (copy_from_user(&kbuf, buffer, count))
4334 kbuf[MAX_SLABINFO_WRITE] = '\0';
4336 tmp = strchr(kbuf, ' ');
4341 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4344 /* Find the cache in the chain of caches. */
4345 mutex_lock(&cache_chain_mutex);
4347 list_for_each_entry(cachep, &cache_chain, next) {
4348 if (!strcmp(cachep->name, kbuf)) {
4349 if (limit < 1 || batchcount < 1 ||
4350 batchcount > limit || shared < 0) {
4353 res = do_tune_cpucache(cachep, limit,
4360 mutex_unlock(&cache_chain_mutex);
4366 static int slabinfo_open(struct inode *inode, struct file *file)
4368 return seq_open(file, &slabinfo_op);
4371 static const struct file_operations proc_slabinfo_operations = {
4372 .open = slabinfo_open,
4374 .write = slabinfo_write,
4375 .llseek = seq_lseek,
4376 .release = seq_release,
4379 #ifdef CONFIG_DEBUG_SLAB_LEAK
4381 static void *leaks_start(struct seq_file *m, loff_t *pos)
4383 mutex_lock(&cache_chain_mutex);
4384 return seq_list_start(&cache_chain, *pos);
4387 static inline int add_caller(unsigned long *n, unsigned long v)
4397 unsigned long *q = p + 2 * i;
4411 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4417 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4423 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4424 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4426 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4431 static void show_symbol(struct seq_file *m, unsigned long address)
4433 #ifdef CONFIG_KALLSYMS
4434 unsigned long offset, size;
4435 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4437 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4438 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4440 seq_printf(m, " [%s]", modname);
4444 seq_printf(m, "%p", (void *)address);
4447 static int leaks_show(struct seq_file *m, void *p)
4449 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4451 struct kmem_list3 *l3;
4453 unsigned long *n = m->private;
4457 if (!(cachep->flags & SLAB_STORE_USER))
4459 if (!(cachep->flags & SLAB_RED_ZONE))
4462 /* OK, we can do it */
4466 for_each_online_node(node) {
4467 l3 = cachep->nodelists[node];
4472 spin_lock_irq(&l3->list_lock);
4474 list_for_each_entry(slabp, &l3->slabs_full, list)
4475 handle_slab(n, cachep, slabp);
4476 list_for_each_entry(slabp, &l3->slabs_partial, list)
4477 handle_slab(n, cachep, slabp);
4478 spin_unlock_irq(&l3->list_lock);
4480 name = cachep->name;
4482 /* Increase the buffer size */
4483 mutex_unlock(&cache_chain_mutex);
4484 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4486 /* Too bad, we are really out */
4488 mutex_lock(&cache_chain_mutex);
4491 *(unsigned long *)m->private = n[0] * 2;
4493 mutex_lock(&cache_chain_mutex);
4494 /* Now make sure this entry will be retried */
4498 for (i = 0; i < n[1]; i++) {
4499 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4500 show_symbol(m, n[2*i+2]);
4507 static const struct seq_operations slabstats_op = {
4508 .start = leaks_start,
4514 static int slabstats_open(struct inode *inode, struct file *file)
4516 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4519 ret = seq_open(file, &slabstats_op);
4521 struct seq_file *m = file->private_data;
4522 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4531 static const struct file_operations proc_slabstats_operations = {
4532 .open = slabstats_open,
4534 .llseek = seq_lseek,
4535 .release = seq_release_private,
4539 static int __init slab_proc_init(void)
4541 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4542 #ifdef CONFIG_DEBUG_SLAB_LEAK
4543 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4547 module_init(slab_proc_init);
4551 * ksize - get the actual amount of memory allocated for a given object
4552 * @objp: Pointer to the object
4554 * kmalloc may internally round up allocations and return more memory
4555 * than requested. ksize() can be used to determine the actual amount of
4556 * memory allocated. The caller may use this additional memory, even though
4557 * a smaller amount of memory was initially specified with the kmalloc call.
4558 * The caller must guarantee that objp points to a valid object previously
4559 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4560 * must not be freed during the duration of the call.
4562 size_t ksize(const void *objp)
4565 if (unlikely(objp == ZERO_SIZE_PTR))
4568 return obj_size(virt_to_cache(objp));
4570 EXPORT_SYMBOL(ksize);