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/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
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_MINALIGN
149 * Enforce a minimum alignment for the kmalloc caches.
150 * Usually, the kmalloc caches are cache_line_size() aligned, except when
151 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
152 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
153 * alignment larger than the alignment of a 64-bit integer.
154 * ARCH_KMALLOC_MINALIGN allows that.
155 * Note that increasing this value may disable some debug features.
157 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
160 #ifndef ARCH_SLAB_MINALIGN
162 * Enforce a minimum alignment for all caches.
163 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
164 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
165 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
166 * some debug features.
168 #define ARCH_SLAB_MINALIGN 0
171 #ifndef ARCH_KMALLOC_FLAGS
172 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
175 /* Legal flag mask for kmem_cache_create(). */
177 # define CREATE_MASK (SLAB_RED_ZONE | \
178 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
183 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
185 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
187 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
188 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
189 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
195 * Bufctl's are used for linking objs within a slab
198 * This implementation relies on "struct page" for locating the cache &
199 * slab an object belongs to.
200 * This allows the bufctl structure to be small (one int), but limits
201 * the number of objects a slab (not a cache) can contain when off-slab
202 * bufctls are used. The limit is the size of the largest general cache
203 * that does not use off-slab slabs.
204 * For 32bit archs with 4 kB pages, is this 56.
205 * This is not serious, as it is only for large objects, when it is unwise
206 * to have too many per slab.
207 * Note: This limit can be raised by introducing a general cache whose size
208 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
211 typedef unsigned int kmem_bufctl_t;
212 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
213 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
214 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
215 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
225 struct list_head list;
226 unsigned long colouroff;
227 void *s_mem; /* including colour offset */
228 unsigned int inuse; /* num of objs active in slab */
230 unsigned short nodeid;
236 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
237 * arrange for kmem_freepages to be called via RCU. This is useful if
238 * we need to approach a kernel structure obliquely, from its address
239 * obtained without the usual locking. We can lock the structure to
240 * stabilize it and check it's still at the given address, only if we
241 * can be sure that the memory has not been meanwhile reused for some
242 * other kind of object (which our subsystem's lock might corrupt).
244 * rcu_read_lock before reading the address, then rcu_read_unlock after
245 * taking the spinlock within the structure expected at that address.
247 * We assume struct slab_rcu can overlay struct slab when destroying.
250 struct rcu_head head;
251 struct kmem_cache *cachep;
259 * - LIFO ordering, to hand out cache-warm objects from _alloc
260 * - reduce the number of linked list operations
261 * - reduce spinlock operations
263 * The limit is stored in the per-cpu structure to reduce the data cache
270 unsigned int batchcount;
271 unsigned int touched;
274 * Must have this definition in here for the proper
275 * alignment of array_cache. Also simplifies accessing
281 * bootstrap: The caches do not work without cpuarrays anymore, but the
282 * cpuarrays are allocated from the generic caches...
284 #define BOOT_CPUCACHE_ENTRIES 1
285 struct arraycache_init {
286 struct array_cache cache;
287 void *entries[BOOT_CPUCACHE_ENTRIES];
291 * The slab lists for all objects.
294 struct list_head slabs_partial; /* partial list first, better asm code */
295 struct list_head slabs_full;
296 struct list_head slabs_free;
297 unsigned long free_objects;
298 unsigned int free_limit;
299 unsigned int colour_next; /* Per-node cache coloring */
300 spinlock_t list_lock;
301 struct array_cache *shared; /* shared per node */
302 struct array_cache **alien; /* on other nodes */
303 unsigned long next_reap; /* updated without locking */
304 int free_touched; /* updated without locking */
308 * Need this for bootstrapping a per node allocator.
310 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
311 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
312 #define CACHE_CACHE 0
313 #define SIZE_AC MAX_NUMNODES
314 #define SIZE_L3 (2 * MAX_NUMNODES)
316 static int drain_freelist(struct kmem_cache *cache,
317 struct kmem_list3 *l3, int tofree);
318 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
320 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
321 static void cache_reap(struct work_struct *unused);
324 * This function must be completely optimized away if a constant is passed to
325 * it. Mostly the same as what is in linux/slab.h except it returns an index.
327 static __always_inline int index_of(const size_t size)
329 extern void __bad_size(void);
331 if (__builtin_constant_p(size)) {
339 #include <linux/kmalloc_sizes.h>
347 static int slab_early_init = 1;
349 #define INDEX_AC index_of(sizeof(struct arraycache_init))
350 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
352 static void kmem_list3_init(struct kmem_list3 *parent)
354 INIT_LIST_HEAD(&parent->slabs_full);
355 INIT_LIST_HEAD(&parent->slabs_partial);
356 INIT_LIST_HEAD(&parent->slabs_free);
357 parent->shared = NULL;
358 parent->alien = NULL;
359 parent->colour_next = 0;
360 spin_lock_init(&parent->list_lock);
361 parent->free_objects = 0;
362 parent->free_touched = 0;
365 #define MAKE_LIST(cachep, listp, slab, nodeid) \
367 INIT_LIST_HEAD(listp); \
368 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
371 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
373 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
375 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
378 #define CFLGS_OFF_SLAB (0x80000000UL)
379 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
381 #define BATCHREFILL_LIMIT 16
383 * Optimization question: fewer reaps means less probability for unnessary
384 * cpucache drain/refill cycles.
386 * OTOH the cpuarrays can contain lots of objects,
387 * which could lock up otherwise freeable slabs.
389 #define REAPTIMEOUT_CPUC (2*HZ)
390 #define REAPTIMEOUT_LIST3 (4*HZ)
393 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
394 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
395 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
396 #define STATS_INC_GROWN(x) ((x)->grown++)
397 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
398 #define STATS_SET_HIGH(x) \
400 if ((x)->num_active > (x)->high_mark) \
401 (x)->high_mark = (x)->num_active; \
403 #define STATS_INC_ERR(x) ((x)->errors++)
404 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
405 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
406 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
407 #define STATS_SET_FREEABLE(x, i) \
409 if ((x)->max_freeable < i) \
410 (x)->max_freeable = i; \
412 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
413 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
414 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
415 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
417 #define STATS_INC_ACTIVE(x) do { } while (0)
418 #define STATS_DEC_ACTIVE(x) do { } while (0)
419 #define STATS_INC_ALLOCED(x) do { } while (0)
420 #define STATS_INC_GROWN(x) do { } while (0)
421 #define STATS_ADD_REAPED(x,y) do { } while (0)
422 #define STATS_SET_HIGH(x) do { } while (0)
423 #define STATS_INC_ERR(x) do { } while (0)
424 #define STATS_INC_NODEALLOCS(x) do { } while (0)
425 #define STATS_INC_NODEFREES(x) do { } while (0)
426 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
427 #define STATS_SET_FREEABLE(x, i) do { } while (0)
428 #define STATS_INC_ALLOCHIT(x) do { } while (0)
429 #define STATS_INC_ALLOCMISS(x) do { } while (0)
430 #define STATS_INC_FREEHIT(x) do { } while (0)
431 #define STATS_INC_FREEMISS(x) do { } while (0)
437 * memory layout of objects:
439 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
440 * the end of an object is aligned with the end of the real
441 * allocation. Catches writes behind the end of the allocation.
442 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
444 * cachep->obj_offset: The real object.
445 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
446 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
447 * [BYTES_PER_WORD long]
449 static int obj_offset(struct kmem_cache *cachep)
451 return cachep->obj_offset;
454 static int obj_size(struct kmem_cache *cachep)
456 return cachep->obj_size;
459 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
461 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
462 return (unsigned long long*) (objp + obj_offset(cachep) -
463 sizeof(unsigned long long));
466 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
468 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
469 if (cachep->flags & SLAB_STORE_USER)
470 return (unsigned long long *)(objp + cachep->buffer_size -
471 sizeof(unsigned long long) -
473 return (unsigned long long *) (objp + cachep->buffer_size -
474 sizeof(unsigned long long));
477 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
479 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
480 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
485 #define obj_offset(x) 0
486 #define obj_size(cachep) (cachep->buffer_size)
487 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
488 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
489 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
493 #ifdef CONFIG_TRACING
494 size_t slab_buffer_size(struct kmem_cache *cachep)
496 return cachep->buffer_size;
498 EXPORT_SYMBOL(slab_buffer_size);
502 * Do not go above this order unless 0 objects fit into the slab.
504 #define BREAK_GFP_ORDER_HI 1
505 #define BREAK_GFP_ORDER_LO 0
506 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
509 * Functions for storing/retrieving the cachep and or slab from the page
510 * allocator. These are used to find the slab an obj belongs to. With kfree(),
511 * these are used to find the cache which an obj belongs to.
513 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
515 page->lru.next = (struct list_head *)cache;
518 static inline struct kmem_cache *page_get_cache(struct page *page)
520 page = compound_head(page);
521 BUG_ON(!PageSlab(page));
522 return (struct kmem_cache *)page->lru.next;
525 static inline void page_set_slab(struct page *page, struct slab *slab)
527 page->lru.prev = (struct list_head *)slab;
530 static inline struct slab *page_get_slab(struct page *page)
532 BUG_ON(!PageSlab(page));
533 return (struct slab *)page->lru.prev;
536 static inline struct kmem_cache *virt_to_cache(const void *obj)
538 struct page *page = virt_to_head_page(obj);
539 return page_get_cache(page);
542 static inline struct slab *virt_to_slab(const void *obj)
544 struct page *page = virt_to_head_page(obj);
545 return page_get_slab(page);
548 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
551 return slab->s_mem + cache->buffer_size * idx;
555 * We want to avoid an expensive divide : (offset / cache->buffer_size)
556 * Using the fact that buffer_size is a constant for a particular cache,
557 * we can replace (offset / cache->buffer_size) by
558 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
560 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
561 const struct slab *slab, void *obj)
563 u32 offset = (obj - slab->s_mem);
564 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
568 * These are the default caches for kmalloc. Custom caches can have other sizes.
570 struct cache_sizes malloc_sizes[] = {
571 #define CACHE(x) { .cs_size = (x) },
572 #include <linux/kmalloc_sizes.h>
576 EXPORT_SYMBOL(malloc_sizes);
578 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
584 static struct cache_names __initdata cache_names[] = {
585 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
586 #include <linux/kmalloc_sizes.h>
591 static struct arraycache_init initarray_cache __initdata =
592 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
593 static struct arraycache_init initarray_generic =
594 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
596 /* internal cache of cache description objs */
597 static struct kmem_cache cache_cache = {
599 .limit = BOOT_CPUCACHE_ENTRIES,
601 .buffer_size = sizeof(struct kmem_cache),
602 .name = "kmem_cache",
605 #define BAD_ALIEN_MAGIC 0x01020304ul
608 * chicken and egg problem: delay the per-cpu array allocation
609 * until the general caches are up.
620 * used by boot code to determine if it can use slab based allocator
622 int slab_is_available(void)
624 return g_cpucache_up >= EARLY;
627 #ifdef CONFIG_LOCKDEP
630 * Slab sometimes uses the kmalloc slabs to store the slab headers
631 * for other slabs "off slab".
632 * The locking for this is tricky in that it nests within the locks
633 * of all other slabs in a few places; to deal with this special
634 * locking we put on-slab caches into a separate lock-class.
636 * We set lock class for alien array caches which are up during init.
637 * The lock annotation will be lost if all cpus of a node goes down and
638 * then comes back up during hotplug
640 static struct lock_class_key on_slab_l3_key;
641 static struct lock_class_key on_slab_alc_key;
643 static void init_node_lock_keys(int q)
645 struct cache_sizes *s = malloc_sizes;
647 if (g_cpucache_up != FULL)
650 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
651 struct array_cache **alc;
652 struct kmem_list3 *l3;
655 l3 = s->cs_cachep->nodelists[q];
656 if (!l3 || OFF_SLAB(s->cs_cachep))
658 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
661 * FIXME: This check for BAD_ALIEN_MAGIC
662 * should go away when common slab code is taught to
663 * work even without alien caches.
664 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
665 * for alloc_alien_cache,
667 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
671 lockdep_set_class(&alc[r]->lock,
677 static inline void init_lock_keys(void)
682 init_node_lock_keys(node);
685 static void init_node_lock_keys(int q)
689 static inline void init_lock_keys(void)
695 * Guard access to the cache-chain.
697 static DEFINE_MUTEX(cache_chain_mutex);
698 static struct list_head cache_chain;
700 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
702 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
704 return cachep->array[smp_processor_id()];
707 static inline struct kmem_cache *__find_general_cachep(size_t size,
710 struct cache_sizes *csizep = malloc_sizes;
713 /* This happens if someone tries to call
714 * kmem_cache_create(), or __kmalloc(), before
715 * the generic caches are initialized.
717 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
720 return ZERO_SIZE_PTR;
722 while (size > csizep->cs_size)
726 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
727 * has cs_{dma,}cachep==NULL. Thus no special case
728 * for large kmalloc calls required.
730 #ifdef CONFIG_ZONE_DMA
731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
734 return csizep->cs_cachep;
737 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
739 return __find_general_cachep(size, gfpflags);
742 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
744 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
748 * Calculate the number of objects and left-over bytes for a given buffer size.
750 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
751 size_t align, int flags, size_t *left_over,
756 size_t slab_size = PAGE_SIZE << gfporder;
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
773 if (flags & CFLGS_OFF_SLAB) {
775 nr_objs = slab_size / buffer_size;
777 if (nr_objs > SLAB_LIMIT)
778 nr_objs = SLAB_LIMIT;
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
788 nr_objs = (slab_size - sizeof(struct slab)) /
789 (buffer_size + sizeof(kmem_bufctl_t));
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
795 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
802 mgmt_size = slab_mgmt_size(nr_objs, align);
805 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
808 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
810 static void __slab_error(const char *function, struct kmem_cache *cachep,
813 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
814 function, cachep->name, msg);
819 * By default on NUMA we use alien caches to stage the freeing of
820 * objects allocated from other nodes. This causes massive memory
821 * inefficiencies when using fake NUMA setup to split memory into a
822 * large number of small nodes, so it can be disabled on the command
826 static int use_alien_caches __read_mostly = 1;
827 static int __init noaliencache_setup(char *s)
829 use_alien_caches = 0;
832 __setup("noaliencache", noaliencache_setup);
836 * Special reaping functions for NUMA systems called from cache_reap().
837 * These take care of doing round robin flushing of alien caches (containing
838 * objects freed on different nodes from which they were allocated) and the
839 * flushing of remote pcps by calling drain_node_pages.
841 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
843 static void init_reap_node(int cpu)
847 node = next_node(cpu_to_node(cpu), node_online_map);
848 if (node == MAX_NUMNODES)
849 node = first_node(node_online_map);
851 per_cpu(slab_reap_node, cpu) = node;
854 static void next_reap_node(void)
856 int node = __get_cpu_var(slab_reap_node);
858 node = next_node(node, node_online_map);
859 if (unlikely(node >= MAX_NUMNODES))
860 node = first_node(node_online_map);
861 __get_cpu_var(slab_reap_node) = node;
865 #define init_reap_node(cpu) do { } while (0)
866 #define next_reap_node(void) do { } while (0)
870 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
871 * via the workqueue/eventd.
872 * Add the CPU number into the expiration time to minimize the possibility of
873 * the CPUs getting into lockstep and contending for the global cache chain
876 static void __cpuinit start_cpu_timer(int cpu)
878 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
881 * When this gets called from do_initcalls via cpucache_init(),
882 * init_workqueues() has already run, so keventd will be setup
885 if (keventd_up() && reap_work->work.func == NULL) {
887 INIT_DELAYED_WORK(reap_work, cache_reap);
888 schedule_delayed_work_on(cpu, reap_work,
889 __round_jiffies_relative(HZ, cpu));
893 static struct array_cache *alloc_arraycache(int node, int entries,
894 int batchcount, gfp_t gfp)
896 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
897 struct array_cache *nc = NULL;
899 nc = kmalloc_node(memsize, gfp, node);
901 * The array_cache structures contain pointers to free object.
902 * However, when such objects are allocated or transfered to another
903 * cache the pointers are not cleared and they could be counted as
904 * valid references during a kmemleak scan. Therefore, kmemleak must
905 * not scan such objects.
907 kmemleak_no_scan(nc);
911 nc->batchcount = batchcount;
913 spin_lock_init(&nc->lock);
919 * Transfer objects in one arraycache to another.
920 * Locking must be handled by the caller.
922 * Return the number of entries transferred.
924 static int transfer_objects(struct array_cache *to,
925 struct array_cache *from, unsigned int max)
927 /* Figure out how many entries to transfer */
928 int nr = min(min(from->avail, max), to->limit - to->avail);
933 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
944 #define drain_alien_cache(cachep, alien) do { } while (0)
945 #define reap_alien(cachep, l3) do { } while (0)
947 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
949 return (struct array_cache **)BAD_ALIEN_MAGIC;
952 static inline void free_alien_cache(struct array_cache **ac_ptr)
956 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
961 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
967 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
968 gfp_t flags, int nodeid)
973 #else /* CONFIG_NUMA */
975 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
976 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
978 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
980 struct array_cache **ac_ptr;
981 int memsize = sizeof(void *) * nr_node_ids;
986 ac_ptr = kmalloc_node(memsize, gfp, node);
989 if (i == node || !node_online(i)) {
993 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
995 for (i--; i >= 0; i--)
1005 static void free_alien_cache(struct array_cache **ac_ptr)
1016 static void __drain_alien_cache(struct kmem_cache *cachep,
1017 struct array_cache *ac, int node)
1019 struct kmem_list3 *rl3 = cachep->nodelists[node];
1022 spin_lock(&rl3->list_lock);
1024 * Stuff objects into the remote nodes shared array first.
1025 * That way we could avoid the overhead of putting the objects
1026 * into the free lists and getting them back later.
1029 transfer_objects(rl3->shared, ac, ac->limit);
1031 free_block(cachep, ac->entry, ac->avail, node);
1033 spin_unlock(&rl3->list_lock);
1038 * Called from cache_reap() to regularly drain alien caches round robin.
1040 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1042 int node = __get_cpu_var(slab_reap_node);
1045 struct array_cache *ac = l3->alien[node];
1047 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1048 __drain_alien_cache(cachep, ac, node);
1049 spin_unlock_irq(&ac->lock);
1054 static void drain_alien_cache(struct kmem_cache *cachep,
1055 struct array_cache **alien)
1058 struct array_cache *ac;
1059 unsigned long flags;
1061 for_each_online_node(i) {
1064 spin_lock_irqsave(&ac->lock, flags);
1065 __drain_alien_cache(cachep, ac, i);
1066 spin_unlock_irqrestore(&ac->lock, flags);
1071 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1073 struct slab *slabp = virt_to_slab(objp);
1074 int nodeid = slabp->nodeid;
1075 struct kmem_list3 *l3;
1076 struct array_cache *alien = NULL;
1079 node = numa_node_id();
1082 * Make sure we are not freeing a object from another node to the array
1083 * cache on this cpu.
1085 if (likely(slabp->nodeid == node))
1088 l3 = cachep->nodelists[node];
1089 STATS_INC_NODEFREES(cachep);
1090 if (l3->alien && l3->alien[nodeid]) {
1091 alien = l3->alien[nodeid];
1092 spin_lock(&alien->lock);
1093 if (unlikely(alien->avail == alien->limit)) {
1094 STATS_INC_ACOVERFLOW(cachep);
1095 __drain_alien_cache(cachep, alien, nodeid);
1097 alien->entry[alien->avail++] = objp;
1098 spin_unlock(&alien->lock);
1100 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1101 free_block(cachep, &objp, 1, nodeid);
1102 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1108 static void __cpuinit cpuup_canceled(long cpu)
1110 struct kmem_cache *cachep;
1111 struct kmem_list3 *l3 = NULL;
1112 int node = cpu_to_node(cpu);
1113 const struct cpumask *mask = cpumask_of_node(node);
1115 list_for_each_entry(cachep, &cache_chain, next) {
1116 struct array_cache *nc;
1117 struct array_cache *shared;
1118 struct array_cache **alien;
1120 /* cpu is dead; no one can alloc from it. */
1121 nc = cachep->array[cpu];
1122 cachep->array[cpu] = NULL;
1123 l3 = cachep->nodelists[node];
1126 goto free_array_cache;
1128 spin_lock_irq(&l3->list_lock);
1130 /* Free limit for this kmem_list3 */
1131 l3->free_limit -= cachep->batchcount;
1133 free_block(cachep, nc->entry, nc->avail, node);
1135 if (!cpumask_empty(mask)) {
1136 spin_unlock_irq(&l3->list_lock);
1137 goto free_array_cache;
1140 shared = l3->shared;
1142 free_block(cachep, shared->entry,
1143 shared->avail, node);
1150 spin_unlock_irq(&l3->list_lock);
1154 drain_alien_cache(cachep, alien);
1155 free_alien_cache(alien);
1161 * In the previous loop, all the objects were freed to
1162 * the respective cache's slabs, now we can go ahead and
1163 * shrink each nodelist to its limit.
1165 list_for_each_entry(cachep, &cache_chain, next) {
1166 l3 = cachep->nodelists[node];
1169 drain_freelist(cachep, l3, l3->free_objects);
1173 static int __cpuinit cpuup_prepare(long cpu)
1175 struct kmem_cache *cachep;
1176 struct kmem_list3 *l3 = NULL;
1177 int node = cpu_to_node(cpu);
1178 const int memsize = sizeof(struct kmem_list3);
1181 * We need to do this right in the beginning since
1182 * alloc_arraycache's are going to use this list.
1183 * kmalloc_node allows us to add the slab to the right
1184 * kmem_list3 and not this cpu's kmem_list3
1187 list_for_each_entry(cachep, &cache_chain, next) {
1189 * Set up the size64 kmemlist for cpu before we can
1190 * begin anything. Make sure some other cpu on this
1191 * node has not already allocated this
1193 if (!cachep->nodelists[node]) {
1194 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1197 kmem_list3_init(l3);
1198 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1199 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1202 * The l3s don't come and go as CPUs come and
1203 * go. cache_chain_mutex is sufficient
1206 cachep->nodelists[node] = l3;
1209 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1210 cachep->nodelists[node]->free_limit =
1211 (1 + nr_cpus_node(node)) *
1212 cachep->batchcount + cachep->num;
1213 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1217 * Now we can go ahead with allocating the shared arrays and
1220 list_for_each_entry(cachep, &cache_chain, next) {
1221 struct array_cache *nc;
1222 struct array_cache *shared = NULL;
1223 struct array_cache **alien = NULL;
1225 nc = alloc_arraycache(node, cachep->limit,
1226 cachep->batchcount, GFP_KERNEL);
1229 if (cachep->shared) {
1230 shared = alloc_arraycache(node,
1231 cachep->shared * cachep->batchcount,
1232 0xbaadf00d, GFP_KERNEL);
1238 if (use_alien_caches) {
1239 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1246 cachep->array[cpu] = nc;
1247 l3 = cachep->nodelists[node];
1250 spin_lock_irq(&l3->list_lock);
1253 * We are serialised from CPU_DEAD or
1254 * CPU_UP_CANCELLED by the cpucontrol lock
1256 l3->shared = shared;
1265 spin_unlock_irq(&l3->list_lock);
1267 free_alien_cache(alien);
1269 init_node_lock_keys(node);
1273 cpuup_canceled(cpu);
1277 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1278 unsigned long action, void *hcpu)
1280 long cpu = (long)hcpu;
1284 case CPU_UP_PREPARE:
1285 case CPU_UP_PREPARE_FROZEN:
1286 mutex_lock(&cache_chain_mutex);
1287 err = cpuup_prepare(cpu);
1288 mutex_unlock(&cache_chain_mutex);
1291 case CPU_ONLINE_FROZEN:
1292 start_cpu_timer(cpu);
1294 #ifdef CONFIG_HOTPLUG_CPU
1295 case CPU_DOWN_PREPARE:
1296 case CPU_DOWN_PREPARE_FROZEN:
1298 * Shutdown cache reaper. Note that the cache_chain_mutex is
1299 * held so that if cache_reap() is invoked it cannot do
1300 * anything expensive but will only modify reap_work
1301 * and reschedule the timer.
1303 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
1304 /* Now the cache_reaper is guaranteed to be not running. */
1305 per_cpu(slab_reap_work, cpu).work.func = NULL;
1307 case CPU_DOWN_FAILED:
1308 case CPU_DOWN_FAILED_FROZEN:
1309 start_cpu_timer(cpu);
1312 case CPU_DEAD_FROZEN:
1314 * Even if all the cpus of a node are down, we don't free the
1315 * kmem_list3 of any cache. This to avoid a race between
1316 * cpu_down, and a kmalloc allocation from another cpu for
1317 * memory from the node of the cpu going down. The list3
1318 * structure is usually allocated from kmem_cache_create() and
1319 * gets destroyed at kmem_cache_destroy().
1323 case CPU_UP_CANCELED:
1324 case CPU_UP_CANCELED_FROZEN:
1325 mutex_lock(&cache_chain_mutex);
1326 cpuup_canceled(cpu);
1327 mutex_unlock(&cache_chain_mutex);
1330 return err ? NOTIFY_BAD : NOTIFY_OK;
1333 static struct notifier_block __cpuinitdata cpucache_notifier = {
1334 &cpuup_callback, NULL, 0
1338 * swap the static kmem_list3 with kmalloced memory
1340 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1343 struct kmem_list3 *ptr;
1345 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1348 memcpy(ptr, list, sizeof(struct kmem_list3));
1350 * Do not assume that spinlocks can be initialized via memcpy:
1352 spin_lock_init(&ptr->list_lock);
1354 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1355 cachep->nodelists[nodeid] = ptr;
1359 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1360 * size of kmem_list3.
1362 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1366 for_each_online_node(node) {
1367 cachep->nodelists[node] = &initkmem_list3[index + node];
1368 cachep->nodelists[node]->next_reap = jiffies +
1370 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1375 * Initialisation. Called after the page allocator have been initialised and
1376 * before smp_init().
1378 void __init kmem_cache_init(void)
1381 struct cache_sizes *sizes;
1382 struct cache_names *names;
1387 if (num_possible_nodes() == 1)
1388 use_alien_caches = 0;
1390 for (i = 0; i < NUM_INIT_LISTS; i++) {
1391 kmem_list3_init(&initkmem_list3[i]);
1392 if (i < MAX_NUMNODES)
1393 cache_cache.nodelists[i] = NULL;
1395 set_up_list3s(&cache_cache, CACHE_CACHE);
1398 * Fragmentation resistance on low memory - only use bigger
1399 * page orders on machines with more than 32MB of memory.
1401 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1402 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1404 /* Bootstrap is tricky, because several objects are allocated
1405 * from caches that do not exist yet:
1406 * 1) initialize the cache_cache cache: it contains the struct
1407 * kmem_cache structures of all caches, except cache_cache itself:
1408 * cache_cache is statically allocated.
1409 * Initially an __init data area is used for the head array and the
1410 * kmem_list3 structures, it's replaced with a kmalloc allocated
1411 * array at the end of the bootstrap.
1412 * 2) Create the first kmalloc cache.
1413 * The struct kmem_cache for the new cache is allocated normally.
1414 * An __init data area is used for the head array.
1415 * 3) Create the remaining kmalloc caches, with minimally sized
1417 * 4) Replace the __init data head arrays for cache_cache and the first
1418 * kmalloc cache with kmalloc allocated arrays.
1419 * 5) Replace the __init data for kmem_list3 for cache_cache and
1420 * the other cache's with kmalloc allocated memory.
1421 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1424 node = numa_node_id();
1426 /* 1) create the cache_cache */
1427 INIT_LIST_HEAD(&cache_chain);
1428 list_add(&cache_cache.next, &cache_chain);
1429 cache_cache.colour_off = cache_line_size();
1430 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1431 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1434 * struct kmem_cache size depends on nr_node_ids, which
1435 * can be less than MAX_NUMNODES.
1437 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1438 nr_node_ids * sizeof(struct kmem_list3 *);
1440 cache_cache.obj_size = cache_cache.buffer_size;
1442 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1444 cache_cache.reciprocal_buffer_size =
1445 reciprocal_value(cache_cache.buffer_size);
1447 for (order = 0; order < MAX_ORDER; order++) {
1448 cache_estimate(order, cache_cache.buffer_size,
1449 cache_line_size(), 0, &left_over, &cache_cache.num);
1450 if (cache_cache.num)
1453 BUG_ON(!cache_cache.num);
1454 cache_cache.gfporder = order;
1455 cache_cache.colour = left_over / cache_cache.colour_off;
1456 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1457 sizeof(struct slab), cache_line_size());
1459 /* 2+3) create the kmalloc caches */
1460 sizes = malloc_sizes;
1461 names = cache_names;
1464 * Initialize the caches that provide memory for the array cache and the
1465 * kmem_list3 structures first. Without this, further allocations will
1469 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1470 sizes[INDEX_AC].cs_size,
1471 ARCH_KMALLOC_MINALIGN,
1472 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1475 if (INDEX_AC != INDEX_L3) {
1476 sizes[INDEX_L3].cs_cachep =
1477 kmem_cache_create(names[INDEX_L3].name,
1478 sizes[INDEX_L3].cs_size,
1479 ARCH_KMALLOC_MINALIGN,
1480 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1484 slab_early_init = 0;
1486 while (sizes->cs_size != ULONG_MAX) {
1488 * For performance, all the general caches are L1 aligned.
1489 * This should be particularly beneficial on SMP boxes, as it
1490 * eliminates "false sharing".
1491 * Note for systems short on memory removing the alignment will
1492 * allow tighter packing of the smaller caches.
1494 if (!sizes->cs_cachep) {
1495 sizes->cs_cachep = kmem_cache_create(names->name,
1497 ARCH_KMALLOC_MINALIGN,
1498 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1501 #ifdef CONFIG_ZONE_DMA
1502 sizes->cs_dmacachep = kmem_cache_create(
1505 ARCH_KMALLOC_MINALIGN,
1506 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1513 /* 4) Replace the bootstrap head arrays */
1515 struct array_cache *ptr;
1517 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1519 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1520 memcpy(ptr, cpu_cache_get(&cache_cache),
1521 sizeof(struct arraycache_init));
1523 * Do not assume that spinlocks can be initialized via memcpy:
1525 spin_lock_init(&ptr->lock);
1527 cache_cache.array[smp_processor_id()] = ptr;
1529 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1531 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1532 != &initarray_generic.cache);
1533 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1534 sizeof(struct arraycache_init));
1536 * Do not assume that spinlocks can be initialized via memcpy:
1538 spin_lock_init(&ptr->lock);
1540 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1543 /* 5) Replace the bootstrap kmem_list3's */
1547 for_each_online_node(nid) {
1548 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1550 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1551 &initkmem_list3[SIZE_AC + nid], nid);
1553 if (INDEX_AC != INDEX_L3) {
1554 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1555 &initkmem_list3[SIZE_L3 + nid], nid);
1560 g_cpucache_up = EARLY;
1563 void __init kmem_cache_init_late(void)
1565 struct kmem_cache *cachep;
1567 /* 6) resize the head arrays to their final sizes */
1568 mutex_lock(&cache_chain_mutex);
1569 list_for_each_entry(cachep, &cache_chain, next)
1570 if (enable_cpucache(cachep, GFP_NOWAIT))
1572 mutex_unlock(&cache_chain_mutex);
1575 g_cpucache_up = FULL;
1577 /* Annotate slab for lockdep -- annotate the malloc caches */
1581 * Register a cpu startup notifier callback that initializes
1582 * cpu_cache_get for all new cpus
1584 register_cpu_notifier(&cpucache_notifier);
1587 * The reap timers are started later, with a module init call: That part
1588 * of the kernel is not yet operational.
1592 static int __init cpucache_init(void)
1597 * Register the timers that return unneeded pages to the page allocator
1599 for_each_online_cpu(cpu)
1600 start_cpu_timer(cpu);
1603 __initcall(cpucache_init);
1606 * Interface to system's page allocator. No need to hold the cache-lock.
1608 * If we requested dmaable memory, we will get it. Even if we
1609 * did not request dmaable memory, we might get it, but that
1610 * would be relatively rare and ignorable.
1612 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1620 * Nommu uses slab's for process anonymous memory allocations, and thus
1621 * requires __GFP_COMP to properly refcount higher order allocations
1623 flags |= __GFP_COMP;
1626 flags |= cachep->gfpflags;
1627 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1628 flags |= __GFP_RECLAIMABLE;
1630 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1634 nr_pages = (1 << cachep->gfporder);
1635 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1636 add_zone_page_state(page_zone(page),
1637 NR_SLAB_RECLAIMABLE, nr_pages);
1639 add_zone_page_state(page_zone(page),
1640 NR_SLAB_UNRECLAIMABLE, nr_pages);
1641 for (i = 0; i < nr_pages; i++)
1642 __SetPageSlab(page + i);
1644 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1645 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1648 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1650 kmemcheck_mark_unallocated_pages(page, nr_pages);
1653 return page_address(page);
1657 * Interface to system's page release.
1659 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1661 unsigned long i = (1 << cachep->gfporder);
1662 struct page *page = virt_to_page(addr);
1663 const unsigned long nr_freed = i;
1665 kmemcheck_free_shadow(page, cachep->gfporder);
1667 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1668 sub_zone_page_state(page_zone(page),
1669 NR_SLAB_RECLAIMABLE, nr_freed);
1671 sub_zone_page_state(page_zone(page),
1672 NR_SLAB_UNRECLAIMABLE, nr_freed);
1674 BUG_ON(!PageSlab(page));
1675 __ClearPageSlab(page);
1678 if (current->reclaim_state)
1679 current->reclaim_state->reclaimed_slab += nr_freed;
1680 free_pages((unsigned long)addr, cachep->gfporder);
1683 static void kmem_rcu_free(struct rcu_head *head)
1685 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1686 struct kmem_cache *cachep = slab_rcu->cachep;
1688 kmem_freepages(cachep, slab_rcu->addr);
1689 if (OFF_SLAB(cachep))
1690 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1695 #ifdef CONFIG_DEBUG_PAGEALLOC
1696 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1697 unsigned long caller)
1699 int size = obj_size(cachep);
1701 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1703 if (size < 5 * sizeof(unsigned long))
1706 *addr++ = 0x12345678;
1708 *addr++ = smp_processor_id();
1709 size -= 3 * sizeof(unsigned long);
1711 unsigned long *sptr = &caller;
1712 unsigned long svalue;
1714 while (!kstack_end(sptr)) {
1716 if (kernel_text_address(svalue)) {
1718 size -= sizeof(unsigned long);
1719 if (size <= sizeof(unsigned long))
1725 *addr++ = 0x87654321;
1729 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1731 int size = obj_size(cachep);
1732 addr = &((char *)addr)[obj_offset(cachep)];
1734 memset(addr, val, size);
1735 *(unsigned char *)(addr + size - 1) = POISON_END;
1738 static void dump_line(char *data, int offset, int limit)
1741 unsigned char error = 0;
1744 printk(KERN_ERR "%03x:", offset);
1745 for (i = 0; i < limit; i++) {
1746 if (data[offset + i] != POISON_FREE) {
1747 error = data[offset + i];
1750 printk(" %02x", (unsigned char)data[offset + i]);
1754 if (bad_count == 1) {
1755 error ^= POISON_FREE;
1756 if (!(error & (error - 1))) {
1757 printk(KERN_ERR "Single bit error detected. Probably "
1760 printk(KERN_ERR "Run memtest86+ or a similar memory "
1763 printk(KERN_ERR "Run a memory test tool.\n");
1772 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1777 if (cachep->flags & SLAB_RED_ZONE) {
1778 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1779 *dbg_redzone1(cachep, objp),
1780 *dbg_redzone2(cachep, objp));
1783 if (cachep->flags & SLAB_STORE_USER) {
1784 printk(KERN_ERR "Last user: [<%p>]",
1785 *dbg_userword(cachep, objp));
1786 print_symbol("(%s)",
1787 (unsigned long)*dbg_userword(cachep, objp));
1790 realobj = (char *)objp + obj_offset(cachep);
1791 size = obj_size(cachep);
1792 for (i = 0; i < size && lines; i += 16, lines--) {
1795 if (i + limit > size)
1797 dump_line(realobj, i, limit);
1801 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1807 realobj = (char *)objp + obj_offset(cachep);
1808 size = obj_size(cachep);
1810 for (i = 0; i < size; i++) {
1811 char exp = POISON_FREE;
1814 if (realobj[i] != exp) {
1820 "Slab corruption: %s start=%p, len=%d\n",
1821 cachep->name, realobj, size);
1822 print_objinfo(cachep, objp, 0);
1824 /* Hexdump the affected line */
1827 if (i + limit > size)
1829 dump_line(realobj, i, limit);
1832 /* Limit to 5 lines */
1838 /* Print some data about the neighboring objects, if they
1841 struct slab *slabp = virt_to_slab(objp);
1844 objnr = obj_to_index(cachep, slabp, objp);
1846 objp = index_to_obj(cachep, slabp, objnr - 1);
1847 realobj = (char *)objp + obj_offset(cachep);
1848 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1850 print_objinfo(cachep, objp, 2);
1852 if (objnr + 1 < cachep->num) {
1853 objp = index_to_obj(cachep, slabp, objnr + 1);
1854 realobj = (char *)objp + obj_offset(cachep);
1855 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1857 print_objinfo(cachep, objp, 2);
1864 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1867 for (i = 0; i < cachep->num; i++) {
1868 void *objp = index_to_obj(cachep, slabp, i);
1870 if (cachep->flags & SLAB_POISON) {
1871 #ifdef CONFIG_DEBUG_PAGEALLOC
1872 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1874 kernel_map_pages(virt_to_page(objp),
1875 cachep->buffer_size / PAGE_SIZE, 1);
1877 check_poison_obj(cachep, objp);
1879 check_poison_obj(cachep, objp);
1882 if (cachep->flags & SLAB_RED_ZONE) {
1883 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1884 slab_error(cachep, "start of a freed object "
1886 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1887 slab_error(cachep, "end of a freed object "
1893 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1899 * slab_destroy - destroy and release all objects in a slab
1900 * @cachep: cache pointer being destroyed
1901 * @slabp: slab pointer being destroyed
1903 * Destroy all the objs in a slab, and release the mem back to the system.
1904 * Before calling the slab must have been unlinked from the cache. The
1905 * cache-lock is not held/needed.
1907 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1909 void *addr = slabp->s_mem - slabp->colouroff;
1911 slab_destroy_debugcheck(cachep, slabp);
1912 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1913 struct slab_rcu *slab_rcu;
1915 slab_rcu = (struct slab_rcu *)slabp;
1916 slab_rcu->cachep = cachep;
1917 slab_rcu->addr = addr;
1918 call_rcu(&slab_rcu->head, kmem_rcu_free);
1920 kmem_freepages(cachep, addr);
1921 if (OFF_SLAB(cachep))
1922 kmem_cache_free(cachep->slabp_cache, slabp);
1926 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1929 struct kmem_list3 *l3;
1931 for_each_online_cpu(i)
1932 kfree(cachep->array[i]);
1934 /* NUMA: free the list3 structures */
1935 for_each_online_node(i) {
1936 l3 = cachep->nodelists[i];
1939 free_alien_cache(l3->alien);
1943 kmem_cache_free(&cache_cache, cachep);
1948 * calculate_slab_order - calculate size (page order) of slabs
1949 * @cachep: pointer to the cache that is being created
1950 * @size: size of objects to be created in this cache.
1951 * @align: required alignment for the objects.
1952 * @flags: slab allocation flags
1954 * Also calculates the number of objects per slab.
1956 * This could be made much more intelligent. For now, try to avoid using
1957 * high order pages for slabs. When the gfp() functions are more friendly
1958 * towards high-order requests, this should be changed.
1960 static size_t calculate_slab_order(struct kmem_cache *cachep,
1961 size_t size, size_t align, unsigned long flags)
1963 unsigned long offslab_limit;
1964 size_t left_over = 0;
1967 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1971 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1975 if (flags & CFLGS_OFF_SLAB) {
1977 * Max number of objs-per-slab for caches which
1978 * use off-slab slabs. Needed to avoid a possible
1979 * looping condition in cache_grow().
1981 offslab_limit = size - sizeof(struct slab);
1982 offslab_limit /= sizeof(kmem_bufctl_t);
1984 if (num > offslab_limit)
1988 /* Found something acceptable - save it away */
1990 cachep->gfporder = gfporder;
1991 left_over = remainder;
1994 * A VFS-reclaimable slab tends to have most allocations
1995 * as GFP_NOFS and we really don't want to have to be allocating
1996 * higher-order pages when we are unable to shrink dcache.
1998 if (flags & SLAB_RECLAIM_ACCOUNT)
2002 * Large number of objects is good, but very large slabs are
2003 * currently bad for the gfp()s.
2005 if (gfporder >= slab_break_gfp_order)
2009 * Acceptable internal fragmentation?
2011 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2017 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2019 if (g_cpucache_up == FULL)
2020 return enable_cpucache(cachep, gfp);
2022 if (g_cpucache_up == NONE) {
2024 * Note: the first kmem_cache_create must create the cache
2025 * that's used by kmalloc(24), otherwise the creation of
2026 * further caches will BUG().
2028 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2031 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2032 * the first cache, then we need to set up all its list3s,
2033 * otherwise the creation of further caches will BUG().
2035 set_up_list3s(cachep, SIZE_AC);
2036 if (INDEX_AC == INDEX_L3)
2037 g_cpucache_up = PARTIAL_L3;
2039 g_cpucache_up = PARTIAL_AC;
2041 cachep->array[smp_processor_id()] =
2042 kmalloc(sizeof(struct arraycache_init), gfp);
2044 if (g_cpucache_up == PARTIAL_AC) {
2045 set_up_list3s(cachep, SIZE_L3);
2046 g_cpucache_up = PARTIAL_L3;
2049 for_each_online_node(node) {
2050 cachep->nodelists[node] =
2051 kmalloc_node(sizeof(struct kmem_list3),
2053 BUG_ON(!cachep->nodelists[node]);
2054 kmem_list3_init(cachep->nodelists[node]);
2058 cachep->nodelists[numa_node_id()]->next_reap =
2059 jiffies + REAPTIMEOUT_LIST3 +
2060 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2062 cpu_cache_get(cachep)->avail = 0;
2063 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2064 cpu_cache_get(cachep)->batchcount = 1;
2065 cpu_cache_get(cachep)->touched = 0;
2066 cachep->batchcount = 1;
2067 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2072 * kmem_cache_create - Create a cache.
2073 * @name: A string which is used in /proc/slabinfo to identify this cache.
2074 * @size: The size of objects to be created in this cache.
2075 * @align: The required alignment for the objects.
2076 * @flags: SLAB flags
2077 * @ctor: A constructor for the objects.
2079 * Returns a ptr to the cache on success, NULL on failure.
2080 * Cannot be called within a int, but can be interrupted.
2081 * The @ctor is run when new pages are allocated by the cache.
2083 * @name must be valid until the cache is destroyed. This implies that
2084 * the module calling this has to destroy the cache before getting unloaded.
2085 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2086 * therefore applications must manage it themselves.
2090 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2091 * to catch references to uninitialised memory.
2093 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2094 * for buffer overruns.
2096 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2097 * cacheline. This can be beneficial if you're counting cycles as closely
2101 kmem_cache_create (const char *name, size_t size, size_t align,
2102 unsigned long flags, void (*ctor)(void *))
2104 size_t left_over, slab_size, ralign;
2105 struct kmem_cache *cachep = NULL, *pc;
2109 * Sanity checks... these are all serious usage bugs.
2111 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2112 size > KMALLOC_MAX_SIZE) {
2113 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2119 * We use cache_chain_mutex to ensure a consistent view of
2120 * cpu_online_mask as well. Please see cpuup_callback
2122 if (slab_is_available()) {
2124 mutex_lock(&cache_chain_mutex);
2127 list_for_each_entry(pc, &cache_chain, next) {
2132 * This happens when the module gets unloaded and doesn't
2133 * destroy its slab cache and no-one else reuses the vmalloc
2134 * area of the module. Print a warning.
2136 res = probe_kernel_address(pc->name, tmp);
2139 "SLAB: cache with size %d has lost its name\n",
2144 if (!strcmp(pc->name, name)) {
2146 "kmem_cache_create: duplicate cache %s\n", name);
2153 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2156 * Enable redzoning and last user accounting, except for caches with
2157 * large objects, if the increased size would increase the object size
2158 * above the next power of two: caches with object sizes just above a
2159 * power of two have a significant amount of internal fragmentation.
2161 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2162 2 * sizeof(unsigned long long)))
2163 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2164 if (!(flags & SLAB_DESTROY_BY_RCU))
2165 flags |= SLAB_POISON;
2167 if (flags & SLAB_DESTROY_BY_RCU)
2168 BUG_ON(flags & SLAB_POISON);
2171 * Always checks flags, a caller might be expecting debug support which
2174 BUG_ON(flags & ~CREATE_MASK);
2177 * Check that size is in terms of words. This is needed to avoid
2178 * unaligned accesses for some archs when redzoning is used, and makes
2179 * sure any on-slab bufctl's are also correctly aligned.
2181 if (size & (BYTES_PER_WORD - 1)) {
2182 size += (BYTES_PER_WORD - 1);
2183 size &= ~(BYTES_PER_WORD - 1);
2186 /* calculate the final buffer alignment: */
2188 /* 1) arch recommendation: can be overridden for debug */
2189 if (flags & SLAB_HWCACHE_ALIGN) {
2191 * Default alignment: as specified by the arch code. Except if
2192 * an object is really small, then squeeze multiple objects into
2195 ralign = cache_line_size();
2196 while (size <= ralign / 2)
2199 ralign = BYTES_PER_WORD;
2203 * Redzoning and user store require word alignment or possibly larger.
2204 * Note this will be overridden by architecture or caller mandated
2205 * alignment if either is greater than BYTES_PER_WORD.
2207 if (flags & SLAB_STORE_USER)
2208 ralign = BYTES_PER_WORD;
2210 if (flags & SLAB_RED_ZONE) {
2211 ralign = REDZONE_ALIGN;
2212 /* If redzoning, ensure that the second redzone is suitably
2213 * aligned, by adjusting the object size accordingly. */
2214 size += REDZONE_ALIGN - 1;
2215 size &= ~(REDZONE_ALIGN - 1);
2218 /* 2) arch mandated alignment */
2219 if (ralign < ARCH_SLAB_MINALIGN) {
2220 ralign = ARCH_SLAB_MINALIGN;
2222 /* 3) caller mandated alignment */
2223 if (ralign < align) {
2226 /* disable debug if necessary */
2227 if (ralign > __alignof__(unsigned long long))
2228 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2234 if (slab_is_available())
2239 /* Get cache's description obj. */
2240 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2245 cachep->obj_size = size;
2248 * Both debugging options require word-alignment which is calculated
2251 if (flags & SLAB_RED_ZONE) {
2252 /* add space for red zone words */
2253 cachep->obj_offset += sizeof(unsigned long long);
2254 size += 2 * sizeof(unsigned long long);
2256 if (flags & SLAB_STORE_USER) {
2257 /* user store requires one word storage behind the end of
2258 * the real object. But if the second red zone needs to be
2259 * aligned to 64 bits, we must allow that much space.
2261 if (flags & SLAB_RED_ZONE)
2262 size += REDZONE_ALIGN;
2264 size += BYTES_PER_WORD;
2266 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2267 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2268 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2269 cachep->obj_offset += PAGE_SIZE - size;
2276 * Determine if the slab management is 'on' or 'off' slab.
2277 * (bootstrapping cannot cope with offslab caches so don't do
2278 * it too early on. Always use on-slab management when
2279 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2281 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2282 !(flags & SLAB_NOLEAKTRACE))
2284 * Size is large, assume best to place the slab management obj
2285 * off-slab (should allow better packing of objs).
2287 flags |= CFLGS_OFF_SLAB;
2289 size = ALIGN(size, align);
2291 left_over = calculate_slab_order(cachep, size, align, flags);
2295 "kmem_cache_create: couldn't create cache %s.\n", name);
2296 kmem_cache_free(&cache_cache, cachep);
2300 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2301 + sizeof(struct slab), align);
2304 * If the slab has been placed off-slab, and we have enough space then
2305 * move it on-slab. This is at the expense of any extra colouring.
2307 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2308 flags &= ~CFLGS_OFF_SLAB;
2309 left_over -= slab_size;
2312 if (flags & CFLGS_OFF_SLAB) {
2313 /* really off slab. No need for manual alignment */
2315 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2317 #ifdef CONFIG_PAGE_POISONING
2318 /* If we're going to use the generic kernel_map_pages()
2319 * poisoning, then it's going to smash the contents of
2320 * the redzone and userword anyhow, so switch them off.
2322 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2323 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2327 cachep->colour_off = cache_line_size();
2328 /* Offset must be a multiple of the alignment. */
2329 if (cachep->colour_off < align)
2330 cachep->colour_off = align;
2331 cachep->colour = left_over / cachep->colour_off;
2332 cachep->slab_size = slab_size;
2333 cachep->flags = flags;
2334 cachep->gfpflags = 0;
2335 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2336 cachep->gfpflags |= GFP_DMA;
2337 cachep->buffer_size = size;
2338 cachep->reciprocal_buffer_size = reciprocal_value(size);
2340 if (flags & CFLGS_OFF_SLAB) {
2341 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2343 * This is a possibility for one of the malloc_sizes caches.
2344 * But since we go off slab only for object size greater than
2345 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2346 * this should not happen at all.
2347 * But leave a BUG_ON for some lucky dude.
2349 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2351 cachep->ctor = ctor;
2352 cachep->name = name;
2354 if (setup_cpu_cache(cachep, gfp)) {
2355 __kmem_cache_destroy(cachep);
2360 /* cache setup completed, link it into the list */
2361 list_add(&cachep->next, &cache_chain);
2363 if (!cachep && (flags & SLAB_PANIC))
2364 panic("kmem_cache_create(): failed to create slab `%s'\n",
2366 if (slab_is_available()) {
2367 mutex_unlock(&cache_chain_mutex);
2372 EXPORT_SYMBOL(kmem_cache_create);
2375 static void check_irq_off(void)
2377 BUG_ON(!irqs_disabled());
2380 static void check_irq_on(void)
2382 BUG_ON(irqs_disabled());
2385 static void check_spinlock_acquired(struct kmem_cache *cachep)
2389 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2393 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2397 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2402 #define check_irq_off() do { } while(0)
2403 #define check_irq_on() do { } while(0)
2404 #define check_spinlock_acquired(x) do { } while(0)
2405 #define check_spinlock_acquired_node(x, y) do { } while(0)
2408 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2409 struct array_cache *ac,
2410 int force, int node);
2412 static void do_drain(void *arg)
2414 struct kmem_cache *cachep = arg;
2415 struct array_cache *ac;
2416 int node = numa_node_id();
2419 ac = cpu_cache_get(cachep);
2420 spin_lock(&cachep->nodelists[node]->list_lock);
2421 free_block(cachep, ac->entry, ac->avail, node);
2422 spin_unlock(&cachep->nodelists[node]->list_lock);
2426 static void drain_cpu_caches(struct kmem_cache *cachep)
2428 struct kmem_list3 *l3;
2431 on_each_cpu(do_drain, cachep, 1);
2433 for_each_online_node(node) {
2434 l3 = cachep->nodelists[node];
2435 if (l3 && l3->alien)
2436 drain_alien_cache(cachep, l3->alien);
2439 for_each_online_node(node) {
2440 l3 = cachep->nodelists[node];
2442 drain_array(cachep, l3, l3->shared, 1, node);
2447 * Remove slabs from the list of free slabs.
2448 * Specify the number of slabs to drain in tofree.
2450 * Returns the actual number of slabs released.
2452 static int drain_freelist(struct kmem_cache *cache,
2453 struct kmem_list3 *l3, int tofree)
2455 struct list_head *p;
2460 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2462 spin_lock_irq(&l3->list_lock);
2463 p = l3->slabs_free.prev;
2464 if (p == &l3->slabs_free) {
2465 spin_unlock_irq(&l3->list_lock);
2469 slabp = list_entry(p, struct slab, list);
2471 BUG_ON(slabp->inuse);
2473 list_del(&slabp->list);
2475 * Safe to drop the lock. The slab is no longer linked
2478 l3->free_objects -= cache->num;
2479 spin_unlock_irq(&l3->list_lock);
2480 slab_destroy(cache, slabp);
2487 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2488 static int __cache_shrink(struct kmem_cache *cachep)
2491 struct kmem_list3 *l3;
2493 drain_cpu_caches(cachep);
2496 for_each_online_node(i) {
2497 l3 = cachep->nodelists[i];
2501 drain_freelist(cachep, l3, l3->free_objects);
2503 ret += !list_empty(&l3->slabs_full) ||
2504 !list_empty(&l3->slabs_partial);
2506 return (ret ? 1 : 0);
2510 * kmem_cache_shrink - Shrink a cache.
2511 * @cachep: The cache to shrink.
2513 * Releases as many slabs as possible for a cache.
2514 * To help debugging, a zero exit status indicates all slabs were released.
2516 int kmem_cache_shrink(struct kmem_cache *cachep)
2519 BUG_ON(!cachep || in_interrupt());
2522 mutex_lock(&cache_chain_mutex);
2523 ret = __cache_shrink(cachep);
2524 mutex_unlock(&cache_chain_mutex);
2528 EXPORT_SYMBOL(kmem_cache_shrink);
2531 * kmem_cache_destroy - delete a cache
2532 * @cachep: the cache to destroy
2534 * Remove a &struct kmem_cache object from the slab cache.
2536 * It is expected this function will be called by a module when it is
2537 * unloaded. This will remove the cache completely, and avoid a duplicate
2538 * cache being allocated each time a module is loaded and unloaded, if the
2539 * module doesn't have persistent in-kernel storage across loads and unloads.
2541 * The cache must be empty before calling this function.
2543 * The caller must guarantee that noone will allocate memory from the cache
2544 * during the kmem_cache_destroy().
2546 void kmem_cache_destroy(struct kmem_cache *cachep)
2548 BUG_ON(!cachep || in_interrupt());
2550 /* Find the cache in the chain of caches. */
2552 mutex_lock(&cache_chain_mutex);
2554 * the chain is never empty, cache_cache is never destroyed
2556 list_del(&cachep->next);
2557 if (__cache_shrink(cachep)) {
2558 slab_error(cachep, "Can't free all objects");
2559 list_add(&cachep->next, &cache_chain);
2560 mutex_unlock(&cache_chain_mutex);
2565 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2568 __kmem_cache_destroy(cachep);
2569 mutex_unlock(&cache_chain_mutex);
2572 EXPORT_SYMBOL(kmem_cache_destroy);
2575 * Get the memory for a slab management obj.
2576 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2577 * always come from malloc_sizes caches. The slab descriptor cannot
2578 * come from the same cache which is getting created because,
2579 * when we are searching for an appropriate cache for these
2580 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2581 * If we are creating a malloc_sizes cache here it would not be visible to
2582 * kmem_find_general_cachep till the initialization is complete.
2583 * Hence we cannot have slabp_cache same as the original cache.
2585 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2586 int colour_off, gfp_t local_flags,
2591 if (OFF_SLAB(cachep)) {
2592 /* Slab management obj is off-slab. */
2593 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2594 local_flags, nodeid);
2596 * If the first object in the slab is leaked (it's allocated
2597 * but no one has a reference to it), we want to make sure
2598 * kmemleak does not treat the ->s_mem pointer as a reference
2599 * to the object. Otherwise we will not report the leak.
2601 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2606 slabp = objp + colour_off;
2607 colour_off += cachep->slab_size;
2610 slabp->colouroff = colour_off;
2611 slabp->s_mem = objp + colour_off;
2612 slabp->nodeid = nodeid;
2617 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2619 return (kmem_bufctl_t *) (slabp + 1);
2622 static void cache_init_objs(struct kmem_cache *cachep,
2627 for (i = 0; i < cachep->num; i++) {
2628 void *objp = index_to_obj(cachep, slabp, i);
2630 /* need to poison the objs? */
2631 if (cachep->flags & SLAB_POISON)
2632 poison_obj(cachep, objp, POISON_FREE);
2633 if (cachep->flags & SLAB_STORE_USER)
2634 *dbg_userword(cachep, objp) = NULL;
2636 if (cachep->flags & SLAB_RED_ZONE) {
2637 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2638 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2641 * Constructors are not allowed to allocate memory from the same
2642 * cache which they are a constructor for. Otherwise, deadlock.
2643 * They must also be threaded.
2645 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2646 cachep->ctor(objp + obj_offset(cachep));
2648 if (cachep->flags & SLAB_RED_ZONE) {
2649 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2650 slab_error(cachep, "constructor overwrote the"
2651 " end of an object");
2652 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2653 slab_error(cachep, "constructor overwrote the"
2654 " start of an object");
2656 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2657 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2658 kernel_map_pages(virt_to_page(objp),
2659 cachep->buffer_size / PAGE_SIZE, 0);
2664 slab_bufctl(slabp)[i] = i + 1;
2666 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2669 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2671 if (CONFIG_ZONE_DMA_FLAG) {
2672 if (flags & GFP_DMA)
2673 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2675 BUG_ON(cachep->gfpflags & GFP_DMA);
2679 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2682 void *objp = index_to_obj(cachep, slabp, slabp->free);
2686 next = slab_bufctl(slabp)[slabp->free];
2688 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2689 WARN_ON(slabp->nodeid != nodeid);
2696 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2697 void *objp, int nodeid)
2699 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2702 /* Verify that the slab belongs to the intended node */
2703 WARN_ON(slabp->nodeid != nodeid);
2705 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2706 printk(KERN_ERR "slab: double free detected in cache "
2707 "'%s', objp %p\n", cachep->name, objp);
2711 slab_bufctl(slabp)[objnr] = slabp->free;
2712 slabp->free = objnr;
2717 * Map pages beginning at addr to the given cache and slab. This is required
2718 * for the slab allocator to be able to lookup the cache and slab of a
2719 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2721 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2727 page = virt_to_page(addr);
2730 if (likely(!PageCompound(page)))
2731 nr_pages <<= cache->gfporder;
2734 page_set_cache(page, cache);
2735 page_set_slab(page, slab);
2737 } while (--nr_pages);
2741 * Grow (by 1) the number of slabs within a cache. This is called by
2742 * kmem_cache_alloc() when there are no active objs left in a cache.
2744 static int cache_grow(struct kmem_cache *cachep,
2745 gfp_t flags, int nodeid, void *objp)
2750 struct kmem_list3 *l3;
2753 * Be lazy and only check for valid flags here, keeping it out of the
2754 * critical path in kmem_cache_alloc().
2756 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2757 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2759 /* Take the l3 list lock to change the colour_next on this node */
2761 l3 = cachep->nodelists[nodeid];
2762 spin_lock(&l3->list_lock);
2764 /* Get colour for the slab, and cal the next value. */
2765 offset = l3->colour_next;
2767 if (l3->colour_next >= cachep->colour)
2768 l3->colour_next = 0;
2769 spin_unlock(&l3->list_lock);
2771 offset *= cachep->colour_off;
2773 if (local_flags & __GFP_WAIT)
2777 * The test for missing atomic flag is performed here, rather than
2778 * the more obvious place, simply to reduce the critical path length
2779 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2780 * will eventually be caught here (where it matters).
2782 kmem_flagcheck(cachep, flags);
2785 * Get mem for the objs. Attempt to allocate a physical page from
2789 objp = kmem_getpages(cachep, local_flags, nodeid);
2793 /* Get slab management. */
2794 slabp = alloc_slabmgmt(cachep, objp, offset,
2795 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2799 slab_map_pages(cachep, slabp, objp);
2801 cache_init_objs(cachep, slabp);
2803 if (local_flags & __GFP_WAIT)
2804 local_irq_disable();
2806 spin_lock(&l3->list_lock);
2808 /* Make slab active. */
2809 list_add_tail(&slabp->list, &(l3->slabs_free));
2810 STATS_INC_GROWN(cachep);
2811 l3->free_objects += cachep->num;
2812 spin_unlock(&l3->list_lock);
2815 kmem_freepages(cachep, objp);
2817 if (local_flags & __GFP_WAIT)
2818 local_irq_disable();
2825 * Perform extra freeing checks:
2826 * - detect bad pointers.
2827 * - POISON/RED_ZONE checking
2829 static void kfree_debugcheck(const void *objp)
2831 if (!virt_addr_valid(objp)) {
2832 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2833 (unsigned long)objp);
2838 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2840 unsigned long long redzone1, redzone2;
2842 redzone1 = *dbg_redzone1(cache, obj);
2843 redzone2 = *dbg_redzone2(cache, obj);
2848 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2851 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2852 slab_error(cache, "double free detected");
2854 slab_error(cache, "memory outside object was overwritten");
2856 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2857 obj, redzone1, redzone2);
2860 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2867 BUG_ON(virt_to_cache(objp) != cachep);
2869 objp -= obj_offset(cachep);
2870 kfree_debugcheck(objp);
2871 page = virt_to_head_page(objp);
2873 slabp = page_get_slab(page);
2875 if (cachep->flags & SLAB_RED_ZONE) {
2876 verify_redzone_free(cachep, objp);
2877 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2878 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2880 if (cachep->flags & SLAB_STORE_USER)
2881 *dbg_userword(cachep, objp) = caller;
2883 objnr = obj_to_index(cachep, slabp, objp);
2885 BUG_ON(objnr >= cachep->num);
2886 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2888 #ifdef CONFIG_DEBUG_SLAB_LEAK
2889 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2891 if (cachep->flags & SLAB_POISON) {
2892 #ifdef CONFIG_DEBUG_PAGEALLOC
2893 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2894 store_stackinfo(cachep, objp, (unsigned long)caller);
2895 kernel_map_pages(virt_to_page(objp),
2896 cachep->buffer_size / PAGE_SIZE, 0);
2898 poison_obj(cachep, objp, POISON_FREE);
2901 poison_obj(cachep, objp, POISON_FREE);
2907 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2912 /* Check slab's freelist to see if this obj is there. */
2913 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2915 if (entries > cachep->num || i >= cachep->num)
2918 if (entries != cachep->num - slabp->inuse) {
2920 printk(KERN_ERR "slab: Internal list corruption detected in "
2921 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2922 cachep->name, cachep->num, slabp, slabp->inuse);
2924 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2927 printk("\n%03x:", i);
2928 printk(" %02x", ((unsigned char *)slabp)[i]);
2935 #define kfree_debugcheck(x) do { } while(0)
2936 #define cache_free_debugcheck(x,objp,z) (objp)
2937 #define check_slabp(x,y) do { } while(0)
2940 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2943 struct kmem_list3 *l3;
2944 struct array_cache *ac;
2949 node = numa_node_id();
2950 ac = cpu_cache_get(cachep);
2951 batchcount = ac->batchcount;
2952 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2954 * If there was little recent activity on this cache, then
2955 * perform only a partial refill. Otherwise we could generate
2958 batchcount = BATCHREFILL_LIMIT;
2960 l3 = cachep->nodelists[node];
2962 BUG_ON(ac->avail > 0 || !l3);
2963 spin_lock(&l3->list_lock);
2965 /* See if we can refill from the shared array */
2966 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2969 while (batchcount > 0) {
2970 struct list_head *entry;
2972 /* Get slab alloc is to come from. */
2973 entry = l3->slabs_partial.next;
2974 if (entry == &l3->slabs_partial) {
2975 l3->free_touched = 1;
2976 entry = l3->slabs_free.next;
2977 if (entry == &l3->slabs_free)
2981 slabp = list_entry(entry, struct slab, list);
2982 check_slabp(cachep, slabp);
2983 check_spinlock_acquired(cachep);
2986 * The slab was either on partial or free list so
2987 * there must be at least one object available for
2990 BUG_ON(slabp->inuse >= cachep->num);
2992 while (slabp->inuse < cachep->num && batchcount--) {
2993 STATS_INC_ALLOCED(cachep);
2994 STATS_INC_ACTIVE(cachep);
2995 STATS_SET_HIGH(cachep);
2997 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3000 check_slabp(cachep, slabp);
3002 /* move slabp to correct slabp list: */
3003 list_del(&slabp->list);
3004 if (slabp->free == BUFCTL_END)
3005 list_add(&slabp->list, &l3->slabs_full);
3007 list_add(&slabp->list, &l3->slabs_partial);
3011 l3->free_objects -= ac->avail;
3013 spin_unlock(&l3->list_lock);
3015 if (unlikely(!ac->avail)) {
3017 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3019 /* cache_grow can reenable interrupts, then ac could change. */
3020 ac = cpu_cache_get(cachep);
3021 if (!x && ac->avail == 0) /* no objects in sight? abort */
3024 if (!ac->avail) /* objects refilled by interrupt? */
3028 return ac->entry[--ac->avail];
3031 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3034 might_sleep_if(flags & __GFP_WAIT);
3036 kmem_flagcheck(cachep, flags);
3041 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3042 gfp_t flags, void *objp, void *caller)
3046 if (cachep->flags & SLAB_POISON) {
3047 #ifdef CONFIG_DEBUG_PAGEALLOC
3048 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3049 kernel_map_pages(virt_to_page(objp),
3050 cachep->buffer_size / PAGE_SIZE, 1);
3052 check_poison_obj(cachep, objp);
3054 check_poison_obj(cachep, objp);
3056 poison_obj(cachep, objp, POISON_INUSE);
3058 if (cachep->flags & SLAB_STORE_USER)
3059 *dbg_userword(cachep, objp) = caller;
3061 if (cachep->flags & SLAB_RED_ZONE) {
3062 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3063 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3064 slab_error(cachep, "double free, or memory outside"
3065 " object was overwritten");
3067 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3068 objp, *dbg_redzone1(cachep, objp),
3069 *dbg_redzone2(cachep, objp));
3071 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3072 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3074 #ifdef CONFIG_DEBUG_SLAB_LEAK
3079 slabp = page_get_slab(virt_to_head_page(objp));
3080 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3081 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3084 objp += obj_offset(cachep);
3085 if (cachep->ctor && cachep->flags & SLAB_POISON)
3087 #if ARCH_SLAB_MINALIGN
3088 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3089 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3090 objp, ARCH_SLAB_MINALIGN);
3096 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3099 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3101 if (cachep == &cache_cache)
3104 return should_failslab(obj_size(cachep), flags);
3107 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3110 struct array_cache *ac;
3114 ac = cpu_cache_get(cachep);
3115 if (likely(ac->avail)) {
3116 STATS_INC_ALLOCHIT(cachep);
3118 objp = ac->entry[--ac->avail];
3120 STATS_INC_ALLOCMISS(cachep);
3121 objp = cache_alloc_refill(cachep, flags);
3123 * the 'ac' may be updated by cache_alloc_refill(),
3124 * and kmemleak_erase() requires its correct value.
3126 ac = cpu_cache_get(cachep);
3129 * To avoid a false negative, if an object that is in one of the
3130 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3131 * treat the array pointers as a reference to the object.
3134 kmemleak_erase(&ac->entry[ac->avail]);
3140 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3142 * If we are in_interrupt, then process context, including cpusets and
3143 * mempolicy, may not apply and should not be used for allocation policy.
3145 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3147 int nid_alloc, nid_here;
3149 if (in_interrupt() || (flags & __GFP_THISNODE))
3151 nid_alloc = nid_here = numa_node_id();
3152 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3153 nid_alloc = cpuset_mem_spread_node();
3154 else if (current->mempolicy)
3155 nid_alloc = slab_node(current->mempolicy);
3156 if (nid_alloc != nid_here)
3157 return ____cache_alloc_node(cachep, flags, nid_alloc);
3162 * Fallback function if there was no memory available and no objects on a
3163 * certain node and fall back is permitted. First we scan all the
3164 * available nodelists for available objects. If that fails then we
3165 * perform an allocation without specifying a node. This allows the page
3166 * allocator to do its reclaim / fallback magic. We then insert the
3167 * slab into the proper nodelist and then allocate from it.
3169 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3171 struct zonelist *zonelist;
3175 enum zone_type high_zoneidx = gfp_zone(flags);
3179 if (flags & __GFP_THISNODE)
3182 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3183 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3187 * Look through allowed nodes for objects available
3188 * from existing per node queues.
3190 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3191 nid = zone_to_nid(zone);
3193 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3194 cache->nodelists[nid] &&
3195 cache->nodelists[nid]->free_objects) {
3196 obj = ____cache_alloc_node(cache,
3197 flags | GFP_THISNODE, nid);
3205 * This allocation will be performed within the constraints
3206 * of the current cpuset / memory policy requirements.
3207 * We may trigger various forms of reclaim on the allowed
3208 * set and go into memory reserves if necessary.
3210 if (local_flags & __GFP_WAIT)
3212 kmem_flagcheck(cache, flags);
3213 obj = kmem_getpages(cache, local_flags, numa_node_id());
3214 if (local_flags & __GFP_WAIT)
3215 local_irq_disable();
3218 * Insert into the appropriate per node queues
3220 nid = page_to_nid(virt_to_page(obj));
3221 if (cache_grow(cache, flags, nid, obj)) {
3222 obj = ____cache_alloc_node(cache,
3223 flags | GFP_THISNODE, nid);
3226 * Another processor may allocate the
3227 * objects in the slab since we are
3228 * not holding any locks.
3232 /* cache_grow already freed obj */
3241 * A interface to enable slab creation on nodeid
3243 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3246 struct list_head *entry;
3248 struct kmem_list3 *l3;
3252 l3 = cachep->nodelists[nodeid];
3257 spin_lock(&l3->list_lock);
3258 entry = l3->slabs_partial.next;
3259 if (entry == &l3->slabs_partial) {
3260 l3->free_touched = 1;
3261 entry = l3->slabs_free.next;
3262 if (entry == &l3->slabs_free)
3266 slabp = list_entry(entry, struct slab, list);
3267 check_spinlock_acquired_node(cachep, nodeid);
3268 check_slabp(cachep, slabp);
3270 STATS_INC_NODEALLOCS(cachep);
3271 STATS_INC_ACTIVE(cachep);
3272 STATS_SET_HIGH(cachep);
3274 BUG_ON(slabp->inuse == cachep->num);
3276 obj = slab_get_obj(cachep, slabp, nodeid);
3277 check_slabp(cachep, slabp);
3279 /* move slabp to correct slabp list: */
3280 list_del(&slabp->list);
3282 if (slabp->free == BUFCTL_END)
3283 list_add(&slabp->list, &l3->slabs_full);
3285 list_add(&slabp->list, &l3->slabs_partial);
3287 spin_unlock(&l3->list_lock);
3291 spin_unlock(&l3->list_lock);
3292 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3296 return fallback_alloc(cachep, flags);
3303 * kmem_cache_alloc_node - Allocate an object on the specified node
3304 * @cachep: The cache to allocate from.
3305 * @flags: See kmalloc().
3306 * @nodeid: node number of the target node.
3307 * @caller: return address of caller, used for debug information
3309 * Identical to kmem_cache_alloc but it will allocate memory on the given
3310 * node, which can improve the performance for cpu bound structures.
3312 * Fallback to other node is possible if __GFP_THISNODE is not set.
3314 static __always_inline void *
3315 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3318 unsigned long save_flags;
3321 flags &= gfp_allowed_mask;
3323 lockdep_trace_alloc(flags);
3325 if (slab_should_failslab(cachep, flags))
3328 cache_alloc_debugcheck_before(cachep, flags);
3329 local_irq_save(save_flags);
3332 nodeid = numa_node_id();
3334 if (unlikely(!cachep->nodelists[nodeid])) {
3335 /* Node not bootstrapped yet */
3336 ptr = fallback_alloc(cachep, flags);
3340 if (nodeid == numa_node_id()) {
3342 * Use the locally cached objects if possible.
3343 * However ____cache_alloc does not allow fallback
3344 * to other nodes. It may fail while we still have
3345 * objects on other nodes available.
3347 ptr = ____cache_alloc(cachep, flags);
3351 /* ___cache_alloc_node can fall back to other nodes */
3352 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3354 local_irq_restore(save_flags);
3355 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3356 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3360 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3362 if (unlikely((flags & __GFP_ZERO) && ptr))
3363 memset(ptr, 0, obj_size(cachep));
3368 static __always_inline void *
3369 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3373 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3374 objp = alternate_node_alloc(cache, flags);
3378 objp = ____cache_alloc(cache, flags);
3381 * We may just have run out of memory on the local node.
3382 * ____cache_alloc_node() knows how to locate memory on other nodes
3385 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3392 static __always_inline void *
3393 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3395 return ____cache_alloc(cachep, flags);
3398 #endif /* CONFIG_NUMA */
3400 static __always_inline void *
3401 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3403 unsigned long save_flags;
3406 flags &= gfp_allowed_mask;
3408 lockdep_trace_alloc(flags);
3410 if (slab_should_failslab(cachep, flags))
3413 cache_alloc_debugcheck_before(cachep, flags);
3414 local_irq_save(save_flags);
3415 objp = __do_cache_alloc(cachep, flags);
3416 local_irq_restore(save_flags);
3417 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3418 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3423 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3425 if (unlikely((flags & __GFP_ZERO) && objp))
3426 memset(objp, 0, obj_size(cachep));
3432 * Caller needs to acquire correct kmem_list's list_lock
3434 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3438 struct kmem_list3 *l3;
3440 for (i = 0; i < nr_objects; i++) {
3441 void *objp = objpp[i];
3444 slabp = virt_to_slab(objp);
3445 l3 = cachep->nodelists[node];
3446 list_del(&slabp->list);
3447 check_spinlock_acquired_node(cachep, node);
3448 check_slabp(cachep, slabp);
3449 slab_put_obj(cachep, slabp, objp, node);
3450 STATS_DEC_ACTIVE(cachep);
3452 check_slabp(cachep, slabp);
3454 /* fixup slab chains */
3455 if (slabp->inuse == 0) {
3456 if (l3->free_objects > l3->free_limit) {
3457 l3->free_objects -= cachep->num;
3458 /* No need to drop any previously held
3459 * lock here, even if we have a off-slab slab
3460 * descriptor it is guaranteed to come from
3461 * a different cache, refer to comments before
3464 slab_destroy(cachep, slabp);
3466 list_add(&slabp->list, &l3->slabs_free);
3469 /* Unconditionally move a slab to the end of the
3470 * partial list on free - maximum time for the
3471 * other objects to be freed, too.
3473 list_add_tail(&slabp->list, &l3->slabs_partial);
3478 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3481 struct kmem_list3 *l3;
3482 int node = numa_node_id();
3484 batchcount = ac->batchcount;
3486 BUG_ON(!batchcount || batchcount > ac->avail);
3489 l3 = cachep->nodelists[node];
3490 spin_lock(&l3->list_lock);
3492 struct array_cache *shared_array = l3->shared;
3493 int max = shared_array->limit - shared_array->avail;
3495 if (batchcount > max)
3497 memcpy(&(shared_array->entry[shared_array->avail]),
3498 ac->entry, sizeof(void *) * batchcount);
3499 shared_array->avail += batchcount;
3504 free_block(cachep, ac->entry, batchcount, node);
3509 struct list_head *p;
3511 p = l3->slabs_free.next;
3512 while (p != &(l3->slabs_free)) {
3515 slabp = list_entry(p, struct slab, list);
3516 BUG_ON(slabp->inuse);
3521 STATS_SET_FREEABLE(cachep, i);
3524 spin_unlock(&l3->list_lock);
3525 ac->avail -= batchcount;
3526 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3530 * Release an obj back to its cache. If the obj has a constructed state, it must
3531 * be in this state _before_ it is released. Called with disabled ints.
3533 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3535 struct array_cache *ac = cpu_cache_get(cachep);
3538 kmemleak_free_recursive(objp, cachep->flags);
3539 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3541 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3544 * Skip calling cache_free_alien() when the platform is not numa.
3545 * This will avoid cache misses that happen while accessing slabp (which
3546 * is per page memory reference) to get nodeid. Instead use a global
3547 * variable to skip the call, which is mostly likely to be present in
3550 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3553 if (likely(ac->avail < ac->limit)) {
3554 STATS_INC_FREEHIT(cachep);
3555 ac->entry[ac->avail++] = objp;
3558 STATS_INC_FREEMISS(cachep);
3559 cache_flusharray(cachep, ac);
3560 ac->entry[ac->avail++] = objp;
3565 * kmem_cache_alloc - Allocate an object
3566 * @cachep: The cache to allocate from.
3567 * @flags: See kmalloc().
3569 * Allocate an object from this cache. The flags are only relevant
3570 * if the cache has no available objects.
3572 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3574 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3576 trace_kmem_cache_alloc(_RET_IP_, ret,
3577 obj_size(cachep), cachep->buffer_size, flags);
3581 EXPORT_SYMBOL(kmem_cache_alloc);
3583 #ifdef CONFIG_TRACING
3584 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3586 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3588 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3592 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3593 * @cachep: the cache we're checking against
3594 * @ptr: pointer to validate
3596 * This verifies that the untrusted pointer looks sane;
3597 * it is _not_ a guarantee that the pointer is actually
3598 * part of the slab cache in question, but it at least
3599 * validates that the pointer can be dereferenced and
3600 * looks half-way sane.
3602 * Currently only used for dentry validation.
3604 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3606 unsigned long addr = (unsigned long)ptr;
3607 unsigned long min_addr = PAGE_OFFSET;
3608 unsigned long align_mask = BYTES_PER_WORD - 1;
3609 unsigned long size = cachep->buffer_size;
3612 if (unlikely(addr < min_addr))
3614 if (unlikely(addr > (unsigned long)high_memory - size))
3616 if (unlikely(addr & align_mask))
3618 if (unlikely(!kern_addr_valid(addr)))
3620 if (unlikely(!kern_addr_valid(addr + size - 1)))
3622 page = virt_to_page(ptr);
3623 if (unlikely(!PageSlab(page)))
3625 if (unlikely(page_get_cache(page) != cachep))
3633 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3635 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3636 __builtin_return_address(0));
3638 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3639 obj_size(cachep), cachep->buffer_size,
3644 EXPORT_SYMBOL(kmem_cache_alloc_node);
3646 #ifdef CONFIG_TRACING
3647 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3651 return __cache_alloc_node(cachep, flags, nodeid,
3652 __builtin_return_address(0));
3654 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3657 static __always_inline void *
3658 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3660 struct kmem_cache *cachep;
3663 cachep = kmem_find_general_cachep(size, flags);
3664 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3666 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3668 trace_kmalloc_node((unsigned long) caller, ret,
3669 size, cachep->buffer_size, flags, node);
3674 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3675 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3677 return __do_kmalloc_node(size, flags, node,
3678 __builtin_return_address(0));
3680 EXPORT_SYMBOL(__kmalloc_node);
3682 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3683 int node, unsigned long caller)
3685 return __do_kmalloc_node(size, flags, node, (void *)caller);
3687 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3689 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3691 return __do_kmalloc_node(size, flags, node, NULL);
3693 EXPORT_SYMBOL(__kmalloc_node);
3694 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3695 #endif /* CONFIG_NUMA */
3698 * __do_kmalloc - allocate memory
3699 * @size: how many bytes of memory are required.
3700 * @flags: the type of memory to allocate (see kmalloc).
3701 * @caller: function caller for debug tracking of the caller
3703 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3706 struct kmem_cache *cachep;
3709 /* If you want to save a few bytes .text space: replace
3711 * Then kmalloc uses the uninlined functions instead of the inline
3714 cachep = __find_general_cachep(size, flags);
3715 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3717 ret = __cache_alloc(cachep, flags, caller);
3719 trace_kmalloc((unsigned long) caller, ret,
3720 size, cachep->buffer_size, flags);
3726 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3727 void *__kmalloc(size_t size, gfp_t flags)
3729 return __do_kmalloc(size, flags, __builtin_return_address(0));
3731 EXPORT_SYMBOL(__kmalloc);
3733 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3735 return __do_kmalloc(size, flags, (void *)caller);
3737 EXPORT_SYMBOL(__kmalloc_track_caller);
3740 void *__kmalloc(size_t size, gfp_t flags)
3742 return __do_kmalloc(size, flags, NULL);
3744 EXPORT_SYMBOL(__kmalloc);
3748 * kmem_cache_free - Deallocate an object
3749 * @cachep: The cache the allocation was from.
3750 * @objp: The previously allocated object.
3752 * Free an object which was previously allocated from this
3755 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3757 unsigned long flags;
3759 local_irq_save(flags);
3760 debug_check_no_locks_freed(objp, obj_size(cachep));
3761 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3762 debug_check_no_obj_freed(objp, obj_size(cachep));
3763 __cache_free(cachep, objp);
3764 local_irq_restore(flags);
3766 trace_kmem_cache_free(_RET_IP_, objp);
3768 EXPORT_SYMBOL(kmem_cache_free);
3771 * kfree - free previously allocated memory
3772 * @objp: pointer returned by kmalloc.
3774 * If @objp is NULL, no operation is performed.
3776 * Don't free memory not originally allocated by kmalloc()
3777 * or you will run into trouble.
3779 void kfree(const void *objp)
3781 struct kmem_cache *c;
3782 unsigned long flags;
3784 trace_kfree(_RET_IP_, objp);
3786 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3788 local_irq_save(flags);
3789 kfree_debugcheck(objp);
3790 c = virt_to_cache(objp);
3791 debug_check_no_locks_freed(objp, obj_size(c));
3792 debug_check_no_obj_freed(objp, obj_size(c));
3793 __cache_free(c, (void *)objp);
3794 local_irq_restore(flags);
3796 EXPORT_SYMBOL(kfree);
3798 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3800 return obj_size(cachep);
3802 EXPORT_SYMBOL(kmem_cache_size);
3804 const char *kmem_cache_name(struct kmem_cache *cachep)
3806 return cachep->name;
3808 EXPORT_SYMBOL_GPL(kmem_cache_name);
3811 * This initializes kmem_list3 or resizes various caches for all nodes.
3813 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3816 struct kmem_list3 *l3;
3817 struct array_cache *new_shared;
3818 struct array_cache **new_alien = NULL;
3820 for_each_online_node(node) {
3822 if (use_alien_caches) {
3823 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3829 if (cachep->shared) {
3830 new_shared = alloc_arraycache(node,
3831 cachep->shared*cachep->batchcount,
3834 free_alien_cache(new_alien);
3839 l3 = cachep->nodelists[node];
3841 struct array_cache *shared = l3->shared;
3843 spin_lock_irq(&l3->list_lock);
3846 free_block(cachep, shared->entry,
3847 shared->avail, node);
3849 l3->shared = new_shared;
3851 l3->alien = new_alien;
3854 l3->free_limit = (1 + nr_cpus_node(node)) *
3855 cachep->batchcount + cachep->num;
3856 spin_unlock_irq(&l3->list_lock);
3858 free_alien_cache(new_alien);
3861 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3863 free_alien_cache(new_alien);
3868 kmem_list3_init(l3);
3869 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3870 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3871 l3->shared = new_shared;
3872 l3->alien = new_alien;
3873 l3->free_limit = (1 + nr_cpus_node(node)) *
3874 cachep->batchcount + cachep->num;
3875 cachep->nodelists[node] = l3;
3880 if (!cachep->next.next) {
3881 /* Cache is not active yet. Roll back what we did */
3884 if (cachep->nodelists[node]) {
3885 l3 = cachep->nodelists[node];
3888 free_alien_cache(l3->alien);
3890 cachep->nodelists[node] = NULL;
3898 struct ccupdate_struct {
3899 struct kmem_cache *cachep;
3900 struct array_cache *new[NR_CPUS];
3903 static void do_ccupdate_local(void *info)
3905 struct ccupdate_struct *new = info;
3906 struct array_cache *old;
3909 old = cpu_cache_get(new->cachep);
3911 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3912 new->new[smp_processor_id()] = old;
3915 /* Always called with the cache_chain_mutex held */
3916 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3917 int batchcount, int shared, gfp_t gfp)
3919 struct ccupdate_struct *new;
3922 new = kzalloc(sizeof(*new), gfp);
3926 for_each_online_cpu(i) {
3927 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3930 for (i--; i >= 0; i--)
3936 new->cachep = cachep;
3938 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3941 cachep->batchcount = batchcount;
3942 cachep->limit = limit;
3943 cachep->shared = shared;
3945 for_each_online_cpu(i) {
3946 struct array_cache *ccold = new->new[i];
3949 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3950 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3951 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3955 return alloc_kmemlist(cachep, gfp);
3958 /* Called with cache_chain_mutex held always */
3959 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3965 * The head array serves three purposes:
3966 * - create a LIFO ordering, i.e. return objects that are cache-warm
3967 * - reduce the number of spinlock operations.
3968 * - reduce the number of linked list operations on the slab and
3969 * bufctl chains: array operations are cheaper.
3970 * The numbers are guessed, we should auto-tune as described by
3973 if (cachep->buffer_size > 131072)
3975 else if (cachep->buffer_size > PAGE_SIZE)
3977 else if (cachep->buffer_size > 1024)
3979 else if (cachep->buffer_size > 256)
3985 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3986 * allocation behaviour: Most allocs on one cpu, most free operations
3987 * on another cpu. For these cases, an efficient object passing between
3988 * cpus is necessary. This is provided by a shared array. The array
3989 * replaces Bonwick's magazine layer.
3990 * On uniprocessor, it's functionally equivalent (but less efficient)
3991 * to a larger limit. Thus disabled by default.
3994 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3999 * With debugging enabled, large batchcount lead to excessively long
4000 * periods with disabled local interrupts. Limit the batchcount
4005 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4007 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4008 cachep->name, -err);
4013 * Drain an array if it contains any elements taking the l3 lock only if
4014 * necessary. Note that the l3 listlock also protects the array_cache
4015 * if drain_array() is used on the shared array.
4017 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4018 struct array_cache *ac, int force, int node)
4022 if (!ac || !ac->avail)
4024 if (ac->touched && !force) {
4027 spin_lock_irq(&l3->list_lock);
4029 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4030 if (tofree > ac->avail)
4031 tofree = (ac->avail + 1) / 2;
4032 free_block(cachep, ac->entry, tofree, node);
4033 ac->avail -= tofree;
4034 memmove(ac->entry, &(ac->entry[tofree]),
4035 sizeof(void *) * ac->avail);
4037 spin_unlock_irq(&l3->list_lock);
4042 * cache_reap - Reclaim memory from caches.
4043 * @w: work descriptor
4045 * Called from workqueue/eventd every few seconds.
4047 * - clear the per-cpu caches for this CPU.
4048 * - return freeable pages to the main free memory pool.
4050 * If we cannot acquire the cache chain mutex then just give up - we'll try
4051 * again on the next iteration.
4053 static void cache_reap(struct work_struct *w)
4055 struct kmem_cache *searchp;
4056 struct kmem_list3 *l3;
4057 int node = numa_node_id();
4058 struct delayed_work *work = to_delayed_work(w);
4060 if (!mutex_trylock(&cache_chain_mutex))
4061 /* Give up. Setup the next iteration. */
4064 list_for_each_entry(searchp, &cache_chain, next) {
4068 * We only take the l3 lock if absolutely necessary and we
4069 * have established with reasonable certainty that
4070 * we can do some work if the lock was obtained.
4072 l3 = searchp->nodelists[node];
4074 reap_alien(searchp, l3);
4076 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4079 * These are racy checks but it does not matter
4080 * if we skip one check or scan twice.
4082 if (time_after(l3->next_reap, jiffies))
4085 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4087 drain_array(searchp, l3, l3->shared, 0, node);
4089 if (l3->free_touched)
4090 l3->free_touched = 0;
4094 freed = drain_freelist(searchp, l3, (l3->free_limit +
4095 5 * searchp->num - 1) / (5 * searchp->num));
4096 STATS_ADD_REAPED(searchp, freed);
4102 mutex_unlock(&cache_chain_mutex);
4105 /* Set up the next iteration */
4106 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4109 #ifdef CONFIG_SLABINFO
4111 static void print_slabinfo_header(struct seq_file *m)
4114 * Output format version, so at least we can change it
4115 * without _too_ many complaints.
4118 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4120 seq_puts(m, "slabinfo - version: 2.1\n");
4122 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4123 "<objperslab> <pagesperslab>");
4124 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4125 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4127 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4128 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4129 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4134 static void *s_start(struct seq_file *m, loff_t *pos)
4138 mutex_lock(&cache_chain_mutex);
4140 print_slabinfo_header(m);
4142 return seq_list_start(&cache_chain, *pos);
4145 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4147 return seq_list_next(p, &cache_chain, pos);
4150 static void s_stop(struct seq_file *m, void *p)
4152 mutex_unlock(&cache_chain_mutex);
4155 static int s_show(struct seq_file *m, void *p)
4157 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4159 unsigned long active_objs;
4160 unsigned long num_objs;
4161 unsigned long active_slabs = 0;
4162 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4166 struct kmem_list3 *l3;
4170 for_each_online_node(node) {
4171 l3 = cachep->nodelists[node];
4176 spin_lock_irq(&l3->list_lock);
4178 list_for_each_entry(slabp, &l3->slabs_full, list) {
4179 if (slabp->inuse != cachep->num && !error)
4180 error = "slabs_full accounting error";
4181 active_objs += cachep->num;
4184 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4185 if (slabp->inuse == cachep->num && !error)
4186 error = "slabs_partial inuse accounting error";
4187 if (!slabp->inuse && !error)
4188 error = "slabs_partial/inuse accounting error";
4189 active_objs += slabp->inuse;
4192 list_for_each_entry(slabp, &l3->slabs_free, list) {
4193 if (slabp->inuse && !error)
4194 error = "slabs_free/inuse accounting error";
4197 free_objects += l3->free_objects;
4199 shared_avail += l3->shared->avail;
4201 spin_unlock_irq(&l3->list_lock);
4203 num_slabs += active_slabs;
4204 num_objs = num_slabs * cachep->num;
4205 if (num_objs - active_objs != free_objects && !error)
4206 error = "free_objects accounting error";
4208 name = cachep->name;
4210 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4212 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4213 name, active_objs, num_objs, cachep->buffer_size,
4214 cachep->num, (1 << cachep->gfporder));
4215 seq_printf(m, " : tunables %4u %4u %4u",
4216 cachep->limit, cachep->batchcount, cachep->shared);
4217 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4218 active_slabs, num_slabs, shared_avail);
4221 unsigned long high = cachep->high_mark;
4222 unsigned long allocs = cachep->num_allocations;
4223 unsigned long grown = cachep->grown;
4224 unsigned long reaped = cachep->reaped;
4225 unsigned long errors = cachep->errors;
4226 unsigned long max_freeable = cachep->max_freeable;
4227 unsigned long node_allocs = cachep->node_allocs;
4228 unsigned long node_frees = cachep->node_frees;
4229 unsigned long overflows = cachep->node_overflow;
4231 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4232 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4233 reaped, errors, max_freeable, node_allocs,
4234 node_frees, overflows);
4238 unsigned long allochit = atomic_read(&cachep->allochit);
4239 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4240 unsigned long freehit = atomic_read(&cachep->freehit);
4241 unsigned long freemiss = atomic_read(&cachep->freemiss);
4243 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4244 allochit, allocmiss, freehit, freemiss);
4252 * slabinfo_op - iterator that generates /proc/slabinfo
4261 * num-pages-per-slab
4262 * + further values on SMP and with statistics enabled
4265 static const struct seq_operations slabinfo_op = {
4272 #define MAX_SLABINFO_WRITE 128
4274 * slabinfo_write - Tuning for the slab allocator
4276 * @buffer: user buffer
4277 * @count: data length
4280 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4281 size_t count, loff_t *ppos)
4283 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4284 int limit, batchcount, shared, res;
4285 struct kmem_cache *cachep;
4287 if (count > MAX_SLABINFO_WRITE)
4289 if (copy_from_user(&kbuf, buffer, count))
4291 kbuf[MAX_SLABINFO_WRITE] = '\0';
4293 tmp = strchr(kbuf, ' ');
4298 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4301 /* Find the cache in the chain of caches. */
4302 mutex_lock(&cache_chain_mutex);
4304 list_for_each_entry(cachep, &cache_chain, next) {
4305 if (!strcmp(cachep->name, kbuf)) {
4306 if (limit < 1 || batchcount < 1 ||
4307 batchcount > limit || shared < 0) {
4310 res = do_tune_cpucache(cachep, limit,
4317 mutex_unlock(&cache_chain_mutex);
4323 static int slabinfo_open(struct inode *inode, struct file *file)
4325 return seq_open(file, &slabinfo_op);
4328 static const struct file_operations proc_slabinfo_operations = {
4329 .open = slabinfo_open,
4331 .write = slabinfo_write,
4332 .llseek = seq_lseek,
4333 .release = seq_release,
4336 #ifdef CONFIG_DEBUG_SLAB_LEAK
4338 static void *leaks_start(struct seq_file *m, loff_t *pos)
4340 mutex_lock(&cache_chain_mutex);
4341 return seq_list_start(&cache_chain, *pos);
4344 static inline int add_caller(unsigned long *n, unsigned long v)
4354 unsigned long *q = p + 2 * i;
4368 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4374 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4380 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4381 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4383 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4388 static void show_symbol(struct seq_file *m, unsigned long address)
4390 #ifdef CONFIG_KALLSYMS
4391 unsigned long offset, size;
4392 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4394 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4395 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4397 seq_printf(m, " [%s]", modname);
4401 seq_printf(m, "%p", (void *)address);
4404 static int leaks_show(struct seq_file *m, void *p)
4406 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4408 struct kmem_list3 *l3;
4410 unsigned long *n = m->private;
4414 if (!(cachep->flags & SLAB_STORE_USER))
4416 if (!(cachep->flags & SLAB_RED_ZONE))
4419 /* OK, we can do it */
4423 for_each_online_node(node) {
4424 l3 = cachep->nodelists[node];
4429 spin_lock_irq(&l3->list_lock);
4431 list_for_each_entry(slabp, &l3->slabs_full, list)
4432 handle_slab(n, cachep, slabp);
4433 list_for_each_entry(slabp, &l3->slabs_partial, list)
4434 handle_slab(n, cachep, slabp);
4435 spin_unlock_irq(&l3->list_lock);
4437 name = cachep->name;
4439 /* Increase the buffer size */
4440 mutex_unlock(&cache_chain_mutex);
4441 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4443 /* Too bad, we are really out */
4445 mutex_lock(&cache_chain_mutex);
4448 *(unsigned long *)m->private = n[0] * 2;
4450 mutex_lock(&cache_chain_mutex);
4451 /* Now make sure this entry will be retried */
4455 for (i = 0; i < n[1]; i++) {
4456 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4457 show_symbol(m, n[2*i+2]);
4464 static const struct seq_operations slabstats_op = {
4465 .start = leaks_start,
4471 static int slabstats_open(struct inode *inode, struct file *file)
4473 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4476 ret = seq_open(file, &slabstats_op);
4478 struct seq_file *m = file->private_data;
4479 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4488 static const struct file_operations proc_slabstats_operations = {
4489 .open = slabstats_open,
4491 .llseek = seq_lseek,
4492 .release = seq_release_private,
4496 static int __init slab_proc_init(void)
4498 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4499 #ifdef CONFIG_DEBUG_SLAB_LEAK
4500 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4504 module_init(slab_proc_init);
4508 * ksize - get the actual amount of memory allocated for a given object
4509 * @objp: Pointer to the object
4511 * kmalloc may internally round up allocations and return more memory
4512 * than requested. ksize() can be used to determine the actual amount of
4513 * memory allocated. The caller may use this additional memory, even though
4514 * a smaller amount of memory was initially specified with the kmalloc call.
4515 * The caller must guarantee that objp points to a valid object previously
4516 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4517 * must not be freed during the duration of the call.
4519 size_t ksize(const void *objp)
4522 if (unlikely(objp == ZERO_SIZE_PTR))
4525 return obj_size(virt_to_cache(objp));
4527 EXPORT_SYMBOL(ksize);