3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <asm/cacheflush.h>
121 #include <asm/tlbflush.h>
122 #include <asm/page.h>
125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
126 * 0 for faster, smaller code (especially in the critical paths).
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
134 #ifdef CONFIG_DEBUG_SLAB
137 #define FORCED_DEBUG 1
141 #define FORCED_DEBUG 0
144 /* Shouldn't this be in a header file somewhere? */
145 #define BYTES_PER_WORD sizeof(void *)
146 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
148 #ifndef ARCH_KMALLOC_FLAGS
149 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
152 /* Legal flag mask for kmem_cache_create(). */
154 # define CREATE_MASK (SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
158 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
159 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
160 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
162 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
164 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
166 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
172 * Bufctl's are used for linking objs within a slab
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
188 typedef unsigned int kmem_bufctl_t;
189 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
191 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
197 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
198 * arrange for kmem_freepages to be called via RCU. This is useful if
199 * we need to approach a kernel structure obliquely, from its address
200 * obtained without the usual locking. We can lock the structure to
201 * stabilize it and check it's still at the given address, only if we
202 * can be sure that the memory has not been meanwhile reused for some
203 * other kind of object (which our subsystem's lock might corrupt).
205 * rcu_read_lock before reading the address, then rcu_read_unlock after
206 * taking the spinlock within the structure expected at that address.
209 struct rcu_head head;
210 struct kmem_cache *cachep;
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
229 unsigned short nodeid;
231 struct slab_rcu __slab_cover_slab_rcu;
239 * - LIFO ordering, to hand out cache-warm objects from _alloc
240 * - reduce the number of linked list operations
241 * - reduce spinlock operations
243 * The limit is stored in the per-cpu structure to reduce the data cache
250 unsigned int batchcount;
251 unsigned int touched;
254 * Must have this definition in here for the proper
255 * alignment of array_cache. Also simplifies accessing
261 * bootstrap: The caches do not work without cpuarrays anymore, but the
262 * cpuarrays are allocated from the generic caches...
264 #define BOOT_CPUCACHE_ENTRIES 1
265 struct arraycache_init {
266 struct array_cache cache;
267 void *entries[BOOT_CPUCACHE_ENTRIES];
271 * The slab lists for all objects.
274 struct list_head slabs_partial; /* partial list first, better asm code */
275 struct list_head slabs_full;
276 struct list_head slabs_free;
277 unsigned long free_objects;
278 unsigned int free_limit;
279 unsigned int colour_next; /* Per-node cache coloring */
280 spinlock_t list_lock;
281 struct array_cache *shared; /* shared per node */
282 struct array_cache **alien; /* on other nodes */
283 unsigned long next_reap; /* updated without locking */
284 int free_touched; /* updated without locking */
288 * Need this for bootstrapping a per node allocator.
290 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
291 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
292 #define CACHE_CACHE 0
293 #define SIZE_AC MAX_NUMNODES
294 #define SIZE_L3 (2 * MAX_NUMNODES)
296 static int drain_freelist(struct kmem_cache *cache,
297 struct kmem_list3 *l3, int tofree);
298 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
300 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
301 static void cache_reap(struct work_struct *unused);
304 * This function must be completely optimized away if a constant is passed to
305 * it. Mostly the same as what is in linux/slab.h except it returns an index.
307 static __always_inline int index_of(const size_t size)
309 extern void __bad_size(void);
311 if (__builtin_constant_p(size)) {
319 #include <linux/kmalloc_sizes.h>
327 static int slab_early_init = 1;
329 #define INDEX_AC index_of(sizeof(struct arraycache_init))
330 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
332 static void kmem_list3_init(struct kmem_list3 *parent)
334 INIT_LIST_HEAD(&parent->slabs_full);
335 INIT_LIST_HEAD(&parent->slabs_partial);
336 INIT_LIST_HEAD(&parent->slabs_free);
337 parent->shared = NULL;
338 parent->alien = NULL;
339 parent->colour_next = 0;
340 spin_lock_init(&parent->list_lock);
341 parent->free_objects = 0;
342 parent->free_touched = 0;
345 #define MAKE_LIST(cachep, listp, slab, nodeid) \
347 INIT_LIST_HEAD(listp); \
348 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
351 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
353 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
354 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
355 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
358 #define CFLGS_OFF_SLAB (0x80000000UL)
359 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
361 #define BATCHREFILL_LIMIT 16
363 * Optimization question: fewer reaps means less probability for unnessary
364 * cpucache drain/refill cycles.
366 * OTOH the cpuarrays can contain lots of objects,
367 * which could lock up otherwise freeable slabs.
369 #define REAPTIMEOUT_CPUC (2*HZ)
370 #define REAPTIMEOUT_LIST3 (4*HZ)
373 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
374 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
375 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
376 #define STATS_INC_GROWN(x) ((x)->grown++)
377 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
378 #define STATS_SET_HIGH(x) \
380 if ((x)->num_active > (x)->high_mark) \
381 (x)->high_mark = (x)->num_active; \
383 #define STATS_INC_ERR(x) ((x)->errors++)
384 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
385 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
386 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
387 #define STATS_SET_FREEABLE(x, i) \
389 if ((x)->max_freeable < i) \
390 (x)->max_freeable = i; \
392 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
393 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
394 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
395 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
397 #define STATS_INC_ACTIVE(x) do { } while (0)
398 #define STATS_DEC_ACTIVE(x) do { } while (0)
399 #define STATS_INC_ALLOCED(x) do { } while (0)
400 #define STATS_INC_GROWN(x) do { } while (0)
401 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
402 #define STATS_SET_HIGH(x) do { } while (0)
403 #define STATS_INC_ERR(x) do { } while (0)
404 #define STATS_INC_NODEALLOCS(x) do { } while (0)
405 #define STATS_INC_NODEFREES(x) do { } while (0)
406 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
407 #define STATS_SET_FREEABLE(x, i) do { } while (0)
408 #define STATS_INC_ALLOCHIT(x) do { } while (0)
409 #define STATS_INC_ALLOCMISS(x) do { } while (0)
410 #define STATS_INC_FREEHIT(x) do { } while (0)
411 #define STATS_INC_FREEMISS(x) do { } while (0)
417 * memory layout of objects:
419 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
420 * the end of an object is aligned with the end of the real
421 * allocation. Catches writes behind the end of the allocation.
422 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
424 * cachep->obj_offset: The real object.
425 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
426 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
427 * [BYTES_PER_WORD long]
429 static int obj_offset(struct kmem_cache *cachep)
431 return cachep->obj_offset;
434 static int obj_size(struct kmem_cache *cachep)
436 return cachep->obj_size;
439 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
441 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
442 return (unsigned long long*) (objp + obj_offset(cachep) -
443 sizeof(unsigned long long));
446 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
448 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449 if (cachep->flags & SLAB_STORE_USER)
450 return (unsigned long long *)(objp + cachep->buffer_size -
451 sizeof(unsigned long long) -
453 return (unsigned long long *) (objp + cachep->buffer_size -
454 sizeof(unsigned long long));
457 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
459 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
460 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
465 #define obj_offset(x) 0
466 #define obj_size(cachep) (cachep->buffer_size)
467 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
469 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
473 #ifdef CONFIG_TRACING
474 size_t slab_buffer_size(struct kmem_cache *cachep)
476 return cachep->buffer_size;
478 EXPORT_SYMBOL(slab_buffer_size);
482 * Do not go above this order unless 0 objects fit into the slab.
484 #define BREAK_GFP_ORDER_HI 1
485 #define BREAK_GFP_ORDER_LO 0
486 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
489 * Functions for storing/retrieving the cachep and or slab from the page
490 * allocator. These are used to find the slab an obj belongs to. With kfree(),
491 * these are used to find the cache which an obj belongs to.
493 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
495 page->lru.next = (struct list_head *)cache;
498 static inline struct kmem_cache *page_get_cache(struct page *page)
500 page = compound_head(page);
501 BUG_ON(!PageSlab(page));
502 return (struct kmem_cache *)page->lru.next;
505 static inline void page_set_slab(struct page *page, struct slab *slab)
507 page->lru.prev = (struct list_head *)slab;
510 static inline struct slab *page_get_slab(struct page *page)
512 BUG_ON(!PageSlab(page));
513 return (struct slab *)page->lru.prev;
516 static inline struct kmem_cache *virt_to_cache(const void *obj)
518 struct page *page = virt_to_head_page(obj);
519 return page_get_cache(page);
522 static inline struct slab *virt_to_slab(const void *obj)
524 struct page *page = virt_to_head_page(obj);
525 return page_get_slab(page);
528 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
531 return slab->s_mem + cache->buffer_size * idx;
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
540 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541 const struct slab *slab, void *obj)
543 u32 offset = (obj - slab->s_mem);
544 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
550 struct cache_sizes malloc_sizes[] = {
551 #define CACHE(x) { .cs_size = (x) },
552 #include <linux/kmalloc_sizes.h>
556 EXPORT_SYMBOL(malloc_sizes);
558 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
564 static struct cache_names __initdata cache_names[] = {
565 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566 #include <linux/kmalloc_sizes.h>
571 static struct arraycache_init initarray_cache __initdata =
572 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
573 static struct arraycache_init initarray_generic =
574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
576 /* internal cache of cache description objs */
577 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
578 static struct kmem_cache cache_cache = {
579 .nodelists = cache_cache_nodelists,
581 .limit = BOOT_CPUCACHE_ENTRIES,
583 .buffer_size = sizeof(struct kmem_cache),
584 .name = "kmem_cache",
587 #define BAD_ALIEN_MAGIC 0x01020304ul
590 * chicken and egg problem: delay the per-cpu array allocation
591 * until the general caches are up.
602 * used by boot code to determine if it can use slab based allocator
604 int slab_is_available(void)
606 return g_cpucache_up >= EARLY;
609 #ifdef CONFIG_LOCKDEP
612 * Slab sometimes uses the kmalloc slabs to store the slab headers
613 * for other slabs "off slab".
614 * The locking for this is tricky in that it nests within the locks
615 * of all other slabs in a few places; to deal with this special
616 * locking we put on-slab caches into a separate lock-class.
618 * We set lock class for alien array caches which are up during init.
619 * The lock annotation will be lost if all cpus of a node goes down and
620 * then comes back up during hotplug
622 static struct lock_class_key on_slab_l3_key;
623 static struct lock_class_key on_slab_alc_key;
625 static struct lock_class_key debugobj_l3_key;
626 static struct lock_class_key debugobj_alc_key;
628 static void slab_set_lock_classes(struct kmem_cache *cachep,
629 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
632 struct array_cache **alc;
633 struct kmem_list3 *l3;
636 l3 = cachep->nodelists[q];
640 lockdep_set_class(&l3->list_lock, l3_key);
643 * FIXME: This check for BAD_ALIEN_MAGIC
644 * should go away when common slab code is taught to
645 * work even without alien caches.
646 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
647 * for alloc_alien_cache,
649 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
653 lockdep_set_class(&alc[r]->lock, alc_key);
657 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
659 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
662 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
666 for_each_online_node(node)
667 slab_set_debugobj_lock_classes_node(cachep, node);
670 static void init_node_lock_keys(int q)
672 struct cache_sizes *s = malloc_sizes;
674 if (g_cpucache_up != FULL)
677 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
678 struct kmem_list3 *l3;
680 l3 = s->cs_cachep->nodelists[q];
681 if (!l3 || OFF_SLAB(s->cs_cachep))
684 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
685 &on_slab_alc_key, q);
689 static inline void init_lock_keys(void)
694 init_node_lock_keys(node);
697 static void init_node_lock_keys(int q)
701 static inline void init_lock_keys(void)
705 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
709 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
715 * Guard access to the cache-chain.
717 static DEFINE_MUTEX(cache_chain_mutex);
718 static struct list_head cache_chain;
720 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
722 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
724 return cachep->array[smp_processor_id()];
727 static inline struct kmem_cache *__find_general_cachep(size_t size,
730 struct cache_sizes *csizep = malloc_sizes;
733 /* This happens if someone tries to call
734 * kmem_cache_create(), or __kmalloc(), before
735 * the generic caches are initialized.
737 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
740 return ZERO_SIZE_PTR;
742 while (size > csizep->cs_size)
746 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
747 * has cs_{dma,}cachep==NULL. Thus no special case
748 * for large kmalloc calls required.
750 #ifdef CONFIG_ZONE_DMA
751 if (unlikely(gfpflags & GFP_DMA))
752 return csizep->cs_dmacachep;
754 return csizep->cs_cachep;
757 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
759 return __find_general_cachep(size, gfpflags);
762 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
764 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
768 * Calculate the number of objects and left-over bytes for a given buffer size.
770 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
771 size_t align, int flags, size_t *left_over,
776 size_t slab_size = PAGE_SIZE << gfporder;
779 * The slab management structure can be either off the slab or
780 * on it. For the latter case, the memory allocated for a
784 * - One kmem_bufctl_t for each object
785 * - Padding to respect alignment of @align
786 * - @buffer_size bytes for each object
788 * If the slab management structure is off the slab, then the
789 * alignment will already be calculated into the size. Because
790 * the slabs are all pages aligned, the objects will be at the
791 * correct alignment when allocated.
793 if (flags & CFLGS_OFF_SLAB) {
795 nr_objs = slab_size / buffer_size;
797 if (nr_objs > SLAB_LIMIT)
798 nr_objs = SLAB_LIMIT;
801 * Ignore padding for the initial guess. The padding
802 * is at most @align-1 bytes, and @buffer_size is at
803 * least @align. In the worst case, this result will
804 * be one greater than the number of objects that fit
805 * into the memory allocation when taking the padding
808 nr_objs = (slab_size - sizeof(struct slab)) /
809 (buffer_size + sizeof(kmem_bufctl_t));
812 * This calculated number will be either the right
813 * amount, or one greater than what we want.
815 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
819 if (nr_objs > SLAB_LIMIT)
820 nr_objs = SLAB_LIMIT;
822 mgmt_size = slab_mgmt_size(nr_objs, align);
825 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
828 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
830 static void __slab_error(const char *function, struct kmem_cache *cachep,
833 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
834 function, cachep->name, msg);
839 * By default on NUMA we use alien caches to stage the freeing of
840 * objects allocated from other nodes. This causes massive memory
841 * inefficiencies when using fake NUMA setup to split memory into a
842 * large number of small nodes, so it can be disabled on the command
846 static int use_alien_caches __read_mostly = 1;
847 static int __init noaliencache_setup(char *s)
849 use_alien_caches = 0;
852 __setup("noaliencache", noaliencache_setup);
856 * Special reaping functions for NUMA systems called from cache_reap().
857 * These take care of doing round robin flushing of alien caches (containing
858 * objects freed on different nodes from which they were allocated) and the
859 * flushing of remote pcps by calling drain_node_pages.
861 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
863 static void init_reap_node(int cpu)
867 node = next_node(cpu_to_mem(cpu), node_online_map);
868 if (node == MAX_NUMNODES)
869 node = first_node(node_online_map);
871 per_cpu(slab_reap_node, cpu) = node;
874 static void next_reap_node(void)
876 int node = __this_cpu_read(slab_reap_node);
878 node = next_node(node, node_online_map);
879 if (unlikely(node >= MAX_NUMNODES))
880 node = first_node(node_online_map);
881 __this_cpu_write(slab_reap_node, node);
885 #define init_reap_node(cpu) do { } while (0)
886 #define next_reap_node(void) do { } while (0)
890 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
891 * via the workqueue/eventd.
892 * Add the CPU number into the expiration time to minimize the possibility of
893 * the CPUs getting into lockstep and contending for the global cache chain
896 static void __cpuinit start_cpu_timer(int cpu)
898 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
901 * When this gets called from do_initcalls via cpucache_init(),
902 * init_workqueues() has already run, so keventd will be setup
905 if (keventd_up() && reap_work->work.func == NULL) {
907 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
908 schedule_delayed_work_on(cpu, reap_work,
909 __round_jiffies_relative(HZ, cpu));
913 static struct array_cache *alloc_arraycache(int node, int entries,
914 int batchcount, gfp_t gfp)
916 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
917 struct array_cache *nc = NULL;
919 nc = kmalloc_node(memsize, gfp, node);
921 * The array_cache structures contain pointers to free object.
922 * However, when such objects are allocated or transferred to another
923 * cache the pointers are not cleared and they could be counted as
924 * valid references during a kmemleak scan. Therefore, kmemleak must
925 * not scan such objects.
927 kmemleak_no_scan(nc);
931 nc->batchcount = batchcount;
933 spin_lock_init(&nc->lock);
939 * Transfer objects in one arraycache to another.
940 * Locking must be handled by the caller.
942 * Return the number of entries transferred.
944 static int transfer_objects(struct array_cache *to,
945 struct array_cache *from, unsigned int max)
947 /* Figure out how many entries to transfer */
948 int nr = min3(from->avail, max, to->limit - to->avail);
953 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
963 #define drain_alien_cache(cachep, alien) do { } while (0)
964 #define reap_alien(cachep, l3) do { } while (0)
966 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
968 return (struct array_cache **)BAD_ALIEN_MAGIC;
971 static inline void free_alien_cache(struct array_cache **ac_ptr)
975 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
980 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
986 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
987 gfp_t flags, int nodeid)
992 #else /* CONFIG_NUMA */
994 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
995 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
997 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
999 struct array_cache **ac_ptr;
1000 int memsize = sizeof(void *) * nr_node_ids;
1005 ac_ptr = kzalloc_node(memsize, gfp, node);
1008 if (i == node || !node_online(i))
1010 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1012 for (i--; i >= 0; i--)
1022 static void free_alien_cache(struct array_cache **ac_ptr)
1033 static void __drain_alien_cache(struct kmem_cache *cachep,
1034 struct array_cache *ac, int node)
1036 struct kmem_list3 *rl3 = cachep->nodelists[node];
1039 spin_lock(&rl3->list_lock);
1041 * Stuff objects into the remote nodes shared array first.
1042 * That way we could avoid the overhead of putting the objects
1043 * into the free lists and getting them back later.
1046 transfer_objects(rl3->shared, ac, ac->limit);
1048 free_block(cachep, ac->entry, ac->avail, node);
1050 spin_unlock(&rl3->list_lock);
1055 * Called from cache_reap() to regularly drain alien caches round robin.
1057 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1059 int node = __this_cpu_read(slab_reap_node);
1062 struct array_cache *ac = l3->alien[node];
1064 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1065 __drain_alien_cache(cachep, ac, node);
1066 spin_unlock_irq(&ac->lock);
1071 static void drain_alien_cache(struct kmem_cache *cachep,
1072 struct array_cache **alien)
1075 struct array_cache *ac;
1076 unsigned long flags;
1078 for_each_online_node(i) {
1081 spin_lock_irqsave(&ac->lock, flags);
1082 __drain_alien_cache(cachep, ac, i);
1083 spin_unlock_irqrestore(&ac->lock, flags);
1088 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1090 struct slab *slabp = virt_to_slab(objp);
1091 int nodeid = slabp->nodeid;
1092 struct kmem_list3 *l3;
1093 struct array_cache *alien = NULL;
1096 node = numa_mem_id();
1099 * Make sure we are not freeing a object from another node to the array
1100 * cache on this cpu.
1102 if (likely(slabp->nodeid == node))
1105 l3 = cachep->nodelists[node];
1106 STATS_INC_NODEFREES(cachep);
1107 if (l3->alien && l3->alien[nodeid]) {
1108 alien = l3->alien[nodeid];
1109 spin_lock(&alien->lock);
1110 if (unlikely(alien->avail == alien->limit)) {
1111 STATS_INC_ACOVERFLOW(cachep);
1112 __drain_alien_cache(cachep, alien, nodeid);
1114 alien->entry[alien->avail++] = objp;
1115 spin_unlock(&alien->lock);
1117 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1118 free_block(cachep, &objp, 1, nodeid);
1119 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1126 * Allocates and initializes nodelists for a node on each slab cache, used for
1127 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1128 * will be allocated off-node since memory is not yet online for the new node.
1129 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1132 * Must hold cache_chain_mutex.
1134 static int init_cache_nodelists_node(int node)
1136 struct kmem_cache *cachep;
1137 struct kmem_list3 *l3;
1138 const int memsize = sizeof(struct kmem_list3);
1140 list_for_each_entry(cachep, &cache_chain, next) {
1142 * Set up the size64 kmemlist for cpu before we can
1143 * begin anything. Make sure some other cpu on this
1144 * node has not already allocated this
1146 if (!cachep->nodelists[node]) {
1147 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1150 kmem_list3_init(l3);
1151 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1152 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1155 * The l3s don't come and go as CPUs come and
1156 * go. cache_chain_mutex is sufficient
1159 cachep->nodelists[node] = l3;
1162 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1163 cachep->nodelists[node]->free_limit =
1164 (1 + nr_cpus_node(node)) *
1165 cachep->batchcount + cachep->num;
1166 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1171 static void __cpuinit cpuup_canceled(long cpu)
1173 struct kmem_cache *cachep;
1174 struct kmem_list3 *l3 = NULL;
1175 int node = cpu_to_mem(cpu);
1176 const struct cpumask *mask = cpumask_of_node(node);
1178 list_for_each_entry(cachep, &cache_chain, next) {
1179 struct array_cache *nc;
1180 struct array_cache *shared;
1181 struct array_cache **alien;
1183 /* cpu is dead; no one can alloc from it. */
1184 nc = cachep->array[cpu];
1185 cachep->array[cpu] = NULL;
1186 l3 = cachep->nodelists[node];
1189 goto free_array_cache;
1191 spin_lock_irq(&l3->list_lock);
1193 /* Free limit for this kmem_list3 */
1194 l3->free_limit -= cachep->batchcount;
1196 free_block(cachep, nc->entry, nc->avail, node);
1198 if (!cpumask_empty(mask)) {
1199 spin_unlock_irq(&l3->list_lock);
1200 goto free_array_cache;
1203 shared = l3->shared;
1205 free_block(cachep, shared->entry,
1206 shared->avail, node);
1213 spin_unlock_irq(&l3->list_lock);
1217 drain_alien_cache(cachep, alien);
1218 free_alien_cache(alien);
1224 * In the previous loop, all the objects were freed to
1225 * the respective cache's slabs, now we can go ahead and
1226 * shrink each nodelist to its limit.
1228 list_for_each_entry(cachep, &cache_chain, next) {
1229 l3 = cachep->nodelists[node];
1232 drain_freelist(cachep, l3, l3->free_objects);
1236 static int __cpuinit cpuup_prepare(long cpu)
1238 struct kmem_cache *cachep;
1239 struct kmem_list3 *l3 = NULL;
1240 int node = cpu_to_mem(cpu);
1244 * We need to do this right in the beginning since
1245 * alloc_arraycache's are going to use this list.
1246 * kmalloc_node allows us to add the slab to the right
1247 * kmem_list3 and not this cpu's kmem_list3
1249 err = init_cache_nodelists_node(node);
1254 * Now we can go ahead with allocating the shared arrays and
1257 list_for_each_entry(cachep, &cache_chain, next) {
1258 struct array_cache *nc;
1259 struct array_cache *shared = NULL;
1260 struct array_cache **alien = NULL;
1262 nc = alloc_arraycache(node, cachep->limit,
1263 cachep->batchcount, GFP_KERNEL);
1266 if (cachep->shared) {
1267 shared = alloc_arraycache(node,
1268 cachep->shared * cachep->batchcount,
1269 0xbaadf00d, GFP_KERNEL);
1275 if (use_alien_caches) {
1276 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1283 cachep->array[cpu] = nc;
1284 l3 = cachep->nodelists[node];
1287 spin_lock_irq(&l3->list_lock);
1290 * We are serialised from CPU_DEAD or
1291 * CPU_UP_CANCELLED by the cpucontrol lock
1293 l3->shared = shared;
1302 spin_unlock_irq(&l3->list_lock);
1304 free_alien_cache(alien);
1305 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1306 slab_set_debugobj_lock_classes_node(cachep, node);
1308 init_node_lock_keys(node);
1312 cpuup_canceled(cpu);
1316 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1317 unsigned long action, void *hcpu)
1319 long cpu = (long)hcpu;
1323 case CPU_UP_PREPARE:
1324 case CPU_UP_PREPARE_FROZEN:
1325 mutex_lock(&cache_chain_mutex);
1326 err = cpuup_prepare(cpu);
1327 mutex_unlock(&cache_chain_mutex);
1330 case CPU_ONLINE_FROZEN:
1331 start_cpu_timer(cpu);
1333 #ifdef CONFIG_HOTPLUG_CPU
1334 case CPU_DOWN_PREPARE:
1335 case CPU_DOWN_PREPARE_FROZEN:
1337 * Shutdown cache reaper. Note that the cache_chain_mutex is
1338 * held so that if cache_reap() is invoked it cannot do
1339 * anything expensive but will only modify reap_work
1340 * and reschedule the timer.
1342 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1343 /* Now the cache_reaper is guaranteed to be not running. */
1344 per_cpu(slab_reap_work, cpu).work.func = NULL;
1346 case CPU_DOWN_FAILED:
1347 case CPU_DOWN_FAILED_FROZEN:
1348 start_cpu_timer(cpu);
1351 case CPU_DEAD_FROZEN:
1353 * Even if all the cpus of a node are down, we don't free the
1354 * kmem_list3 of any cache. This to avoid a race between
1355 * cpu_down, and a kmalloc allocation from another cpu for
1356 * memory from the node of the cpu going down. The list3
1357 * structure is usually allocated from kmem_cache_create() and
1358 * gets destroyed at kmem_cache_destroy().
1362 case CPU_UP_CANCELED:
1363 case CPU_UP_CANCELED_FROZEN:
1364 mutex_lock(&cache_chain_mutex);
1365 cpuup_canceled(cpu);
1366 mutex_unlock(&cache_chain_mutex);
1369 return notifier_from_errno(err);
1372 static struct notifier_block __cpuinitdata cpucache_notifier = {
1373 &cpuup_callback, NULL, 0
1376 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1378 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1379 * Returns -EBUSY if all objects cannot be drained so that the node is not
1382 * Must hold cache_chain_mutex.
1384 static int __meminit drain_cache_nodelists_node(int node)
1386 struct kmem_cache *cachep;
1389 list_for_each_entry(cachep, &cache_chain, next) {
1390 struct kmem_list3 *l3;
1392 l3 = cachep->nodelists[node];
1396 drain_freelist(cachep, l3, l3->free_objects);
1398 if (!list_empty(&l3->slabs_full) ||
1399 !list_empty(&l3->slabs_partial)) {
1407 static int __meminit slab_memory_callback(struct notifier_block *self,
1408 unsigned long action, void *arg)
1410 struct memory_notify *mnb = arg;
1414 nid = mnb->status_change_nid;
1419 case MEM_GOING_ONLINE:
1420 mutex_lock(&cache_chain_mutex);
1421 ret = init_cache_nodelists_node(nid);
1422 mutex_unlock(&cache_chain_mutex);
1424 case MEM_GOING_OFFLINE:
1425 mutex_lock(&cache_chain_mutex);
1426 ret = drain_cache_nodelists_node(nid);
1427 mutex_unlock(&cache_chain_mutex);
1431 case MEM_CANCEL_ONLINE:
1432 case MEM_CANCEL_OFFLINE:
1436 return notifier_from_errno(ret);
1438 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1441 * swap the static kmem_list3 with kmalloced memory
1443 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1446 struct kmem_list3 *ptr;
1448 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1451 memcpy(ptr, list, sizeof(struct kmem_list3));
1453 * Do not assume that spinlocks can be initialized via memcpy:
1455 spin_lock_init(&ptr->list_lock);
1457 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1458 cachep->nodelists[nodeid] = ptr;
1462 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1463 * size of kmem_list3.
1465 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1469 for_each_online_node(node) {
1470 cachep->nodelists[node] = &initkmem_list3[index + node];
1471 cachep->nodelists[node]->next_reap = jiffies +
1473 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1478 * Initialisation. Called after the page allocator have been initialised and
1479 * before smp_init().
1481 void __init kmem_cache_init(void)
1484 struct cache_sizes *sizes;
1485 struct cache_names *names;
1490 if (num_possible_nodes() == 1)
1491 use_alien_caches = 0;
1493 for (i = 0; i < NUM_INIT_LISTS; i++) {
1494 kmem_list3_init(&initkmem_list3[i]);
1495 if (i < MAX_NUMNODES)
1496 cache_cache.nodelists[i] = NULL;
1498 set_up_list3s(&cache_cache, CACHE_CACHE);
1501 * Fragmentation resistance on low memory - only use bigger
1502 * page orders on machines with more than 32MB of memory.
1504 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1505 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1507 /* Bootstrap is tricky, because several objects are allocated
1508 * from caches that do not exist yet:
1509 * 1) initialize the cache_cache cache: it contains the struct
1510 * kmem_cache structures of all caches, except cache_cache itself:
1511 * cache_cache is statically allocated.
1512 * Initially an __init data area is used for the head array and the
1513 * kmem_list3 structures, it's replaced with a kmalloc allocated
1514 * array at the end of the bootstrap.
1515 * 2) Create the first kmalloc cache.
1516 * The struct kmem_cache for the new cache is allocated normally.
1517 * An __init data area is used for the head array.
1518 * 3) Create the remaining kmalloc caches, with minimally sized
1520 * 4) Replace the __init data head arrays for cache_cache and the first
1521 * kmalloc cache with kmalloc allocated arrays.
1522 * 5) Replace the __init data for kmem_list3 for cache_cache and
1523 * the other cache's with kmalloc allocated memory.
1524 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1527 node = numa_mem_id();
1529 /* 1) create the cache_cache */
1530 INIT_LIST_HEAD(&cache_chain);
1531 list_add(&cache_cache.next, &cache_chain);
1532 cache_cache.colour_off = cache_line_size();
1533 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1534 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1537 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1539 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1540 nr_node_ids * sizeof(struct kmem_list3 *);
1542 cache_cache.obj_size = cache_cache.buffer_size;
1544 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1546 cache_cache.reciprocal_buffer_size =
1547 reciprocal_value(cache_cache.buffer_size);
1549 for (order = 0; order < MAX_ORDER; order++) {
1550 cache_estimate(order, cache_cache.buffer_size,
1551 cache_line_size(), 0, &left_over, &cache_cache.num);
1552 if (cache_cache.num)
1555 BUG_ON(!cache_cache.num);
1556 cache_cache.gfporder = order;
1557 cache_cache.colour = left_over / cache_cache.colour_off;
1558 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1559 sizeof(struct slab), cache_line_size());
1561 /* 2+3) create the kmalloc caches */
1562 sizes = malloc_sizes;
1563 names = cache_names;
1566 * Initialize the caches that provide memory for the array cache and the
1567 * kmem_list3 structures first. Without this, further allocations will
1571 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1572 sizes[INDEX_AC].cs_size,
1573 ARCH_KMALLOC_MINALIGN,
1574 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1577 if (INDEX_AC != INDEX_L3) {
1578 sizes[INDEX_L3].cs_cachep =
1579 kmem_cache_create(names[INDEX_L3].name,
1580 sizes[INDEX_L3].cs_size,
1581 ARCH_KMALLOC_MINALIGN,
1582 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1586 slab_early_init = 0;
1588 while (sizes->cs_size != ULONG_MAX) {
1590 * For performance, all the general caches are L1 aligned.
1591 * This should be particularly beneficial on SMP boxes, as it
1592 * eliminates "false sharing".
1593 * Note for systems short on memory removing the alignment will
1594 * allow tighter packing of the smaller caches.
1596 if (!sizes->cs_cachep) {
1597 sizes->cs_cachep = kmem_cache_create(names->name,
1599 ARCH_KMALLOC_MINALIGN,
1600 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1603 #ifdef CONFIG_ZONE_DMA
1604 sizes->cs_dmacachep = kmem_cache_create(
1607 ARCH_KMALLOC_MINALIGN,
1608 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1615 /* 4) Replace the bootstrap head arrays */
1617 struct array_cache *ptr;
1619 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1621 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1622 memcpy(ptr, cpu_cache_get(&cache_cache),
1623 sizeof(struct arraycache_init));
1625 * Do not assume that spinlocks can be initialized via memcpy:
1627 spin_lock_init(&ptr->lock);
1629 cache_cache.array[smp_processor_id()] = ptr;
1631 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1633 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1634 != &initarray_generic.cache);
1635 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1636 sizeof(struct arraycache_init));
1638 * Do not assume that spinlocks can be initialized via memcpy:
1640 spin_lock_init(&ptr->lock);
1642 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1645 /* 5) Replace the bootstrap kmem_list3's */
1649 for_each_online_node(nid) {
1650 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1652 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1653 &initkmem_list3[SIZE_AC + nid], nid);
1655 if (INDEX_AC != INDEX_L3) {
1656 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1657 &initkmem_list3[SIZE_L3 + nid], nid);
1662 g_cpucache_up = EARLY;
1665 void __init kmem_cache_init_late(void)
1667 struct kmem_cache *cachep;
1669 /* Annotate slab for lockdep -- annotate the malloc caches */
1672 /* 6) resize the head arrays to their final sizes */
1673 mutex_lock(&cache_chain_mutex);
1674 list_for_each_entry(cachep, &cache_chain, next)
1675 if (enable_cpucache(cachep, GFP_NOWAIT))
1677 mutex_unlock(&cache_chain_mutex);
1680 g_cpucache_up = FULL;
1683 * Register a cpu startup notifier callback that initializes
1684 * cpu_cache_get for all new cpus
1686 register_cpu_notifier(&cpucache_notifier);
1690 * Register a memory hotplug callback that initializes and frees
1693 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1697 * The reap timers are started later, with a module init call: That part
1698 * of the kernel is not yet operational.
1702 static int __init cpucache_init(void)
1707 * Register the timers that return unneeded pages to the page allocator
1709 for_each_online_cpu(cpu)
1710 start_cpu_timer(cpu);
1713 __initcall(cpucache_init);
1716 * Interface to system's page allocator. No need to hold the cache-lock.
1718 * If we requested dmaable memory, we will get it. Even if we
1719 * did not request dmaable memory, we might get it, but that
1720 * would be relatively rare and ignorable.
1722 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1730 * Nommu uses slab's for process anonymous memory allocations, and thus
1731 * requires __GFP_COMP to properly refcount higher order allocations
1733 flags |= __GFP_COMP;
1736 flags |= cachep->gfpflags;
1737 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1738 flags |= __GFP_RECLAIMABLE;
1740 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1744 nr_pages = (1 << cachep->gfporder);
1745 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1746 add_zone_page_state(page_zone(page),
1747 NR_SLAB_RECLAIMABLE, nr_pages);
1749 add_zone_page_state(page_zone(page),
1750 NR_SLAB_UNRECLAIMABLE, nr_pages);
1751 for (i = 0; i < nr_pages; i++)
1752 __SetPageSlab(page + i);
1754 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1755 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1758 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1760 kmemcheck_mark_unallocated_pages(page, nr_pages);
1763 return page_address(page);
1767 * Interface to system's page release.
1769 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1771 unsigned long i = (1 << cachep->gfporder);
1772 struct page *page = virt_to_page(addr);
1773 const unsigned long nr_freed = i;
1775 kmemcheck_free_shadow(page, cachep->gfporder);
1777 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1778 sub_zone_page_state(page_zone(page),
1779 NR_SLAB_RECLAIMABLE, nr_freed);
1781 sub_zone_page_state(page_zone(page),
1782 NR_SLAB_UNRECLAIMABLE, nr_freed);
1784 BUG_ON(!PageSlab(page));
1785 __ClearPageSlab(page);
1788 if (current->reclaim_state)
1789 current->reclaim_state->reclaimed_slab += nr_freed;
1790 free_pages((unsigned long)addr, cachep->gfporder);
1793 static void kmem_rcu_free(struct rcu_head *head)
1795 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1796 struct kmem_cache *cachep = slab_rcu->cachep;
1798 kmem_freepages(cachep, slab_rcu->addr);
1799 if (OFF_SLAB(cachep))
1800 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1805 #ifdef CONFIG_DEBUG_PAGEALLOC
1806 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1807 unsigned long caller)
1809 int size = obj_size(cachep);
1811 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1813 if (size < 5 * sizeof(unsigned long))
1816 *addr++ = 0x12345678;
1818 *addr++ = smp_processor_id();
1819 size -= 3 * sizeof(unsigned long);
1821 unsigned long *sptr = &caller;
1822 unsigned long svalue;
1824 while (!kstack_end(sptr)) {
1826 if (kernel_text_address(svalue)) {
1828 size -= sizeof(unsigned long);
1829 if (size <= sizeof(unsigned long))
1835 *addr++ = 0x87654321;
1839 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1841 int size = obj_size(cachep);
1842 addr = &((char *)addr)[obj_offset(cachep)];
1844 memset(addr, val, size);
1845 *(unsigned char *)(addr + size - 1) = POISON_END;
1848 static void dump_line(char *data, int offset, int limit)
1851 unsigned char error = 0;
1854 printk(KERN_ERR "%03x: ", offset);
1855 for (i = 0; i < limit; i++) {
1856 if (data[offset + i] != POISON_FREE) {
1857 error = data[offset + i];
1861 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1862 &data[offset], limit, 1);
1864 if (bad_count == 1) {
1865 error ^= POISON_FREE;
1866 if (!(error & (error - 1))) {
1867 printk(KERN_ERR "Single bit error detected. Probably "
1870 printk(KERN_ERR "Run memtest86+ or a similar memory "
1873 printk(KERN_ERR "Run a memory test tool.\n");
1882 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1887 if (cachep->flags & SLAB_RED_ZONE) {
1888 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1889 *dbg_redzone1(cachep, objp),
1890 *dbg_redzone2(cachep, objp));
1893 if (cachep->flags & SLAB_STORE_USER) {
1894 printk(KERN_ERR "Last user: [<%p>]",
1895 *dbg_userword(cachep, objp));
1896 print_symbol("(%s)",
1897 (unsigned long)*dbg_userword(cachep, objp));
1900 realobj = (char *)objp + obj_offset(cachep);
1901 size = obj_size(cachep);
1902 for (i = 0; i < size && lines; i += 16, lines--) {
1905 if (i + limit > size)
1907 dump_line(realobj, i, limit);
1911 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1917 realobj = (char *)objp + obj_offset(cachep);
1918 size = obj_size(cachep);
1920 for (i = 0; i < size; i++) {
1921 char exp = POISON_FREE;
1924 if (realobj[i] != exp) {
1930 "Slab corruption: %s start=%p, len=%d\n",
1931 cachep->name, realobj, size);
1932 print_objinfo(cachep, objp, 0);
1934 /* Hexdump the affected line */
1937 if (i + limit > size)
1939 dump_line(realobj, i, limit);
1942 /* Limit to 5 lines */
1948 /* Print some data about the neighboring objects, if they
1951 struct slab *slabp = virt_to_slab(objp);
1954 objnr = obj_to_index(cachep, slabp, objp);
1956 objp = index_to_obj(cachep, slabp, objnr - 1);
1957 realobj = (char *)objp + obj_offset(cachep);
1958 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1960 print_objinfo(cachep, objp, 2);
1962 if (objnr + 1 < cachep->num) {
1963 objp = index_to_obj(cachep, slabp, objnr + 1);
1964 realobj = (char *)objp + obj_offset(cachep);
1965 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1967 print_objinfo(cachep, objp, 2);
1974 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1977 for (i = 0; i < cachep->num; i++) {
1978 void *objp = index_to_obj(cachep, slabp, i);
1980 if (cachep->flags & SLAB_POISON) {
1981 #ifdef CONFIG_DEBUG_PAGEALLOC
1982 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1984 kernel_map_pages(virt_to_page(objp),
1985 cachep->buffer_size / PAGE_SIZE, 1);
1987 check_poison_obj(cachep, objp);
1989 check_poison_obj(cachep, objp);
1992 if (cachep->flags & SLAB_RED_ZONE) {
1993 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1994 slab_error(cachep, "start of a freed object "
1996 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1997 slab_error(cachep, "end of a freed object "
2003 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2009 * slab_destroy - destroy and release all objects in a slab
2010 * @cachep: cache pointer being destroyed
2011 * @slabp: slab pointer being destroyed
2013 * Destroy all the objs in a slab, and release the mem back to the system.
2014 * Before calling the slab must have been unlinked from the cache. The
2015 * cache-lock is not held/needed.
2017 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2019 void *addr = slabp->s_mem - slabp->colouroff;
2021 slab_destroy_debugcheck(cachep, slabp);
2022 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2023 struct slab_rcu *slab_rcu;
2025 slab_rcu = (struct slab_rcu *)slabp;
2026 slab_rcu->cachep = cachep;
2027 slab_rcu->addr = addr;
2028 call_rcu(&slab_rcu->head, kmem_rcu_free);
2030 kmem_freepages(cachep, addr);
2031 if (OFF_SLAB(cachep))
2032 kmem_cache_free(cachep->slabp_cache, slabp);
2036 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2039 struct kmem_list3 *l3;
2041 for_each_online_cpu(i)
2042 kfree(cachep->array[i]);
2044 /* NUMA: free the list3 structures */
2045 for_each_online_node(i) {
2046 l3 = cachep->nodelists[i];
2049 free_alien_cache(l3->alien);
2053 kmem_cache_free(&cache_cache, cachep);
2058 * calculate_slab_order - calculate size (page order) of slabs
2059 * @cachep: pointer to the cache that is being created
2060 * @size: size of objects to be created in this cache.
2061 * @align: required alignment for the objects.
2062 * @flags: slab allocation flags
2064 * Also calculates the number of objects per slab.
2066 * This could be made much more intelligent. For now, try to avoid using
2067 * high order pages for slabs. When the gfp() functions are more friendly
2068 * towards high-order requests, this should be changed.
2070 static size_t calculate_slab_order(struct kmem_cache *cachep,
2071 size_t size, size_t align, unsigned long flags)
2073 unsigned long offslab_limit;
2074 size_t left_over = 0;
2077 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2081 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2085 if (flags & CFLGS_OFF_SLAB) {
2087 * Max number of objs-per-slab for caches which
2088 * use off-slab slabs. Needed to avoid a possible
2089 * looping condition in cache_grow().
2091 offslab_limit = size - sizeof(struct slab);
2092 offslab_limit /= sizeof(kmem_bufctl_t);
2094 if (num > offslab_limit)
2098 /* Found something acceptable - save it away */
2100 cachep->gfporder = gfporder;
2101 left_over = remainder;
2104 * A VFS-reclaimable slab tends to have most allocations
2105 * as GFP_NOFS and we really don't want to have to be allocating
2106 * higher-order pages when we are unable to shrink dcache.
2108 if (flags & SLAB_RECLAIM_ACCOUNT)
2112 * Large number of objects is good, but very large slabs are
2113 * currently bad for the gfp()s.
2115 if (gfporder >= slab_break_gfp_order)
2119 * Acceptable internal fragmentation?
2121 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2127 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2129 if (g_cpucache_up == FULL)
2130 return enable_cpucache(cachep, gfp);
2132 if (g_cpucache_up == NONE) {
2134 * Note: the first kmem_cache_create must create the cache
2135 * that's used by kmalloc(24), otherwise the creation of
2136 * further caches will BUG().
2138 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2141 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2142 * the first cache, then we need to set up all its list3s,
2143 * otherwise the creation of further caches will BUG().
2145 set_up_list3s(cachep, SIZE_AC);
2146 if (INDEX_AC == INDEX_L3)
2147 g_cpucache_up = PARTIAL_L3;
2149 g_cpucache_up = PARTIAL_AC;
2151 cachep->array[smp_processor_id()] =
2152 kmalloc(sizeof(struct arraycache_init), gfp);
2154 if (g_cpucache_up == PARTIAL_AC) {
2155 set_up_list3s(cachep, SIZE_L3);
2156 g_cpucache_up = PARTIAL_L3;
2159 for_each_online_node(node) {
2160 cachep->nodelists[node] =
2161 kmalloc_node(sizeof(struct kmem_list3),
2163 BUG_ON(!cachep->nodelists[node]);
2164 kmem_list3_init(cachep->nodelists[node]);
2168 cachep->nodelists[numa_mem_id()]->next_reap =
2169 jiffies + REAPTIMEOUT_LIST3 +
2170 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2172 cpu_cache_get(cachep)->avail = 0;
2173 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2174 cpu_cache_get(cachep)->batchcount = 1;
2175 cpu_cache_get(cachep)->touched = 0;
2176 cachep->batchcount = 1;
2177 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2182 * kmem_cache_create - Create a cache.
2183 * @name: A string which is used in /proc/slabinfo to identify this cache.
2184 * @size: The size of objects to be created in this cache.
2185 * @align: The required alignment for the objects.
2186 * @flags: SLAB flags
2187 * @ctor: A constructor for the objects.
2189 * Returns a ptr to the cache on success, NULL on failure.
2190 * Cannot be called within a int, but can be interrupted.
2191 * The @ctor is run when new pages are allocated by the cache.
2193 * @name must be valid until the cache is destroyed. This implies that
2194 * the module calling this has to destroy the cache before getting unloaded.
2198 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2199 * to catch references to uninitialised memory.
2201 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2202 * for buffer overruns.
2204 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2205 * cacheline. This can be beneficial if you're counting cycles as closely
2209 kmem_cache_create (const char *name, size_t size, size_t align,
2210 unsigned long flags, void (*ctor)(void *))
2212 size_t left_over, slab_size, ralign;
2213 struct kmem_cache *cachep = NULL, *pc;
2217 * Sanity checks... these are all serious usage bugs.
2219 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2220 size > KMALLOC_MAX_SIZE) {
2221 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2227 * We use cache_chain_mutex to ensure a consistent view of
2228 * cpu_online_mask as well. Please see cpuup_callback
2230 if (slab_is_available()) {
2232 mutex_lock(&cache_chain_mutex);
2235 list_for_each_entry(pc, &cache_chain, next) {
2240 * This happens when the module gets unloaded and doesn't
2241 * destroy its slab cache and no-one else reuses the vmalloc
2242 * area of the module. Print a warning.
2244 res = probe_kernel_address(pc->name, tmp);
2247 "SLAB: cache with size %d has lost its name\n",
2252 if (!strcmp(pc->name, name)) {
2254 "kmem_cache_create: duplicate cache %s\n", name);
2261 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2264 * Enable redzoning and last user accounting, except for caches with
2265 * large objects, if the increased size would increase the object size
2266 * above the next power of two: caches with object sizes just above a
2267 * power of two have a significant amount of internal fragmentation.
2269 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2270 2 * sizeof(unsigned long long)))
2271 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2272 if (!(flags & SLAB_DESTROY_BY_RCU))
2273 flags |= SLAB_POISON;
2275 if (flags & SLAB_DESTROY_BY_RCU)
2276 BUG_ON(flags & SLAB_POISON);
2279 * Always checks flags, a caller might be expecting debug support which
2282 BUG_ON(flags & ~CREATE_MASK);
2285 * Check that size is in terms of words. This is needed to avoid
2286 * unaligned accesses for some archs when redzoning is used, and makes
2287 * sure any on-slab bufctl's are also correctly aligned.
2289 if (size & (BYTES_PER_WORD - 1)) {
2290 size += (BYTES_PER_WORD - 1);
2291 size &= ~(BYTES_PER_WORD - 1);
2294 /* calculate the final buffer alignment: */
2296 /* 1) arch recommendation: can be overridden for debug */
2297 if (flags & SLAB_HWCACHE_ALIGN) {
2299 * Default alignment: as specified by the arch code. Except if
2300 * an object is really small, then squeeze multiple objects into
2303 ralign = cache_line_size();
2304 while (size <= ralign / 2)
2307 ralign = BYTES_PER_WORD;
2311 * Redzoning and user store require word alignment or possibly larger.
2312 * Note this will be overridden by architecture or caller mandated
2313 * alignment if either is greater than BYTES_PER_WORD.
2315 if (flags & SLAB_STORE_USER)
2316 ralign = BYTES_PER_WORD;
2318 if (flags & SLAB_RED_ZONE) {
2319 ralign = REDZONE_ALIGN;
2320 /* If redzoning, ensure that the second redzone is suitably
2321 * aligned, by adjusting the object size accordingly. */
2322 size += REDZONE_ALIGN - 1;
2323 size &= ~(REDZONE_ALIGN - 1);
2326 /* 2) arch mandated alignment */
2327 if (ralign < ARCH_SLAB_MINALIGN) {
2328 ralign = ARCH_SLAB_MINALIGN;
2330 /* 3) caller mandated alignment */
2331 if (ralign < align) {
2334 /* disable debug if necessary */
2335 if (ralign > __alignof__(unsigned long long))
2336 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2342 if (slab_is_available())
2347 /* Get cache's description obj. */
2348 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2352 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2354 cachep->obj_size = size;
2357 * Both debugging options require word-alignment which is calculated
2360 if (flags & SLAB_RED_ZONE) {
2361 /* add space for red zone words */
2362 cachep->obj_offset += sizeof(unsigned long long);
2363 size += 2 * sizeof(unsigned long long);
2365 if (flags & SLAB_STORE_USER) {
2366 /* user store requires one word storage behind the end of
2367 * the real object. But if the second red zone needs to be
2368 * aligned to 64 bits, we must allow that much space.
2370 if (flags & SLAB_RED_ZONE)
2371 size += REDZONE_ALIGN;
2373 size += BYTES_PER_WORD;
2375 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2376 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2377 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2378 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2385 * Determine if the slab management is 'on' or 'off' slab.
2386 * (bootstrapping cannot cope with offslab caches so don't do
2387 * it too early on. Always use on-slab management when
2388 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2390 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2391 !(flags & SLAB_NOLEAKTRACE))
2393 * Size is large, assume best to place the slab management obj
2394 * off-slab (should allow better packing of objs).
2396 flags |= CFLGS_OFF_SLAB;
2398 size = ALIGN(size, align);
2400 left_over = calculate_slab_order(cachep, size, align, flags);
2404 "kmem_cache_create: couldn't create cache %s.\n", name);
2405 kmem_cache_free(&cache_cache, cachep);
2409 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2410 + sizeof(struct slab), align);
2413 * If the slab has been placed off-slab, and we have enough space then
2414 * move it on-slab. This is at the expense of any extra colouring.
2416 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2417 flags &= ~CFLGS_OFF_SLAB;
2418 left_over -= slab_size;
2421 if (flags & CFLGS_OFF_SLAB) {
2422 /* really off slab. No need for manual alignment */
2424 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2426 #ifdef CONFIG_PAGE_POISONING
2427 /* If we're going to use the generic kernel_map_pages()
2428 * poisoning, then it's going to smash the contents of
2429 * the redzone and userword anyhow, so switch them off.
2431 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2432 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2436 cachep->colour_off = cache_line_size();
2437 /* Offset must be a multiple of the alignment. */
2438 if (cachep->colour_off < align)
2439 cachep->colour_off = align;
2440 cachep->colour = left_over / cachep->colour_off;
2441 cachep->slab_size = slab_size;
2442 cachep->flags = flags;
2443 cachep->gfpflags = 0;
2444 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2445 cachep->gfpflags |= GFP_DMA;
2446 cachep->buffer_size = size;
2447 cachep->reciprocal_buffer_size = reciprocal_value(size);
2449 if (flags & CFLGS_OFF_SLAB) {
2450 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2452 * This is a possibility for one of the malloc_sizes caches.
2453 * But since we go off slab only for object size greater than
2454 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2455 * this should not happen at all.
2456 * But leave a BUG_ON for some lucky dude.
2458 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2460 cachep->ctor = ctor;
2461 cachep->name = name;
2463 if (setup_cpu_cache(cachep, gfp)) {
2464 __kmem_cache_destroy(cachep);
2469 if (flags & SLAB_DEBUG_OBJECTS) {
2471 * Would deadlock through slab_destroy()->call_rcu()->
2472 * debug_object_activate()->kmem_cache_alloc().
2474 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2476 slab_set_debugobj_lock_classes(cachep);
2479 /* cache setup completed, link it into the list */
2480 list_add(&cachep->next, &cache_chain);
2482 if (!cachep && (flags & SLAB_PANIC))
2483 panic("kmem_cache_create(): failed to create slab `%s'\n",
2485 if (slab_is_available()) {
2486 mutex_unlock(&cache_chain_mutex);
2491 EXPORT_SYMBOL(kmem_cache_create);
2494 static void check_irq_off(void)
2496 BUG_ON(!irqs_disabled());
2499 static void check_irq_on(void)
2501 BUG_ON(irqs_disabled());
2504 static void check_spinlock_acquired(struct kmem_cache *cachep)
2508 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2512 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2516 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2521 #define check_irq_off() do { } while(0)
2522 #define check_irq_on() do { } while(0)
2523 #define check_spinlock_acquired(x) do { } while(0)
2524 #define check_spinlock_acquired_node(x, y) do { } while(0)
2527 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2528 struct array_cache *ac,
2529 int force, int node);
2531 static void do_drain(void *arg)
2533 struct kmem_cache *cachep = arg;
2534 struct array_cache *ac;
2535 int node = numa_mem_id();
2538 ac = cpu_cache_get(cachep);
2539 spin_lock(&cachep->nodelists[node]->list_lock);
2540 free_block(cachep, ac->entry, ac->avail, node);
2541 spin_unlock(&cachep->nodelists[node]->list_lock);
2545 static void drain_cpu_caches(struct kmem_cache *cachep)
2547 struct kmem_list3 *l3;
2550 on_each_cpu(do_drain, cachep, 1);
2552 for_each_online_node(node) {
2553 l3 = cachep->nodelists[node];
2554 if (l3 && l3->alien)
2555 drain_alien_cache(cachep, l3->alien);
2558 for_each_online_node(node) {
2559 l3 = cachep->nodelists[node];
2561 drain_array(cachep, l3, l3->shared, 1, node);
2566 * Remove slabs from the list of free slabs.
2567 * Specify the number of slabs to drain in tofree.
2569 * Returns the actual number of slabs released.
2571 static int drain_freelist(struct kmem_cache *cache,
2572 struct kmem_list3 *l3, int tofree)
2574 struct list_head *p;
2579 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2581 spin_lock_irq(&l3->list_lock);
2582 p = l3->slabs_free.prev;
2583 if (p == &l3->slabs_free) {
2584 spin_unlock_irq(&l3->list_lock);
2588 slabp = list_entry(p, struct slab, list);
2590 BUG_ON(slabp->inuse);
2592 list_del(&slabp->list);
2594 * Safe to drop the lock. The slab is no longer linked
2597 l3->free_objects -= cache->num;
2598 spin_unlock_irq(&l3->list_lock);
2599 slab_destroy(cache, slabp);
2606 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2607 static int __cache_shrink(struct kmem_cache *cachep)
2610 struct kmem_list3 *l3;
2612 drain_cpu_caches(cachep);
2615 for_each_online_node(i) {
2616 l3 = cachep->nodelists[i];
2620 drain_freelist(cachep, l3, l3->free_objects);
2622 ret += !list_empty(&l3->slabs_full) ||
2623 !list_empty(&l3->slabs_partial);
2625 return (ret ? 1 : 0);
2629 * kmem_cache_shrink - Shrink a cache.
2630 * @cachep: The cache to shrink.
2632 * Releases as many slabs as possible for a cache.
2633 * To help debugging, a zero exit status indicates all slabs were released.
2635 int kmem_cache_shrink(struct kmem_cache *cachep)
2638 BUG_ON(!cachep || in_interrupt());
2641 mutex_lock(&cache_chain_mutex);
2642 ret = __cache_shrink(cachep);
2643 mutex_unlock(&cache_chain_mutex);
2647 EXPORT_SYMBOL(kmem_cache_shrink);
2650 * kmem_cache_destroy - delete a cache
2651 * @cachep: the cache to destroy
2653 * Remove a &struct kmem_cache object from the slab cache.
2655 * It is expected this function will be called by a module when it is
2656 * unloaded. This will remove the cache completely, and avoid a duplicate
2657 * cache being allocated each time a module is loaded and unloaded, if the
2658 * module doesn't have persistent in-kernel storage across loads and unloads.
2660 * The cache must be empty before calling this function.
2662 * The caller must guarantee that no one will allocate memory from the cache
2663 * during the kmem_cache_destroy().
2665 void kmem_cache_destroy(struct kmem_cache *cachep)
2667 BUG_ON(!cachep || in_interrupt());
2669 /* Find the cache in the chain of caches. */
2671 mutex_lock(&cache_chain_mutex);
2673 * the chain is never empty, cache_cache is never destroyed
2675 list_del(&cachep->next);
2676 if (__cache_shrink(cachep)) {
2677 slab_error(cachep, "Can't free all objects");
2678 list_add(&cachep->next, &cache_chain);
2679 mutex_unlock(&cache_chain_mutex);
2684 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2687 __kmem_cache_destroy(cachep);
2688 mutex_unlock(&cache_chain_mutex);
2691 EXPORT_SYMBOL(kmem_cache_destroy);
2694 * Get the memory for a slab management obj.
2695 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2696 * always come from malloc_sizes caches. The slab descriptor cannot
2697 * come from the same cache which is getting created because,
2698 * when we are searching for an appropriate cache for these
2699 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2700 * If we are creating a malloc_sizes cache here it would not be visible to
2701 * kmem_find_general_cachep till the initialization is complete.
2702 * Hence we cannot have slabp_cache same as the original cache.
2704 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2705 int colour_off, gfp_t local_flags,
2710 if (OFF_SLAB(cachep)) {
2711 /* Slab management obj is off-slab. */
2712 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2713 local_flags, nodeid);
2715 * If the first object in the slab is leaked (it's allocated
2716 * but no one has a reference to it), we want to make sure
2717 * kmemleak does not treat the ->s_mem pointer as a reference
2718 * to the object. Otherwise we will not report the leak.
2720 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2725 slabp = objp + colour_off;
2726 colour_off += cachep->slab_size;
2729 slabp->colouroff = colour_off;
2730 slabp->s_mem = objp + colour_off;
2731 slabp->nodeid = nodeid;
2736 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2738 return (kmem_bufctl_t *) (slabp + 1);
2741 static void cache_init_objs(struct kmem_cache *cachep,
2746 for (i = 0; i < cachep->num; i++) {
2747 void *objp = index_to_obj(cachep, slabp, i);
2749 /* need to poison the objs? */
2750 if (cachep->flags & SLAB_POISON)
2751 poison_obj(cachep, objp, POISON_FREE);
2752 if (cachep->flags & SLAB_STORE_USER)
2753 *dbg_userword(cachep, objp) = NULL;
2755 if (cachep->flags & SLAB_RED_ZONE) {
2756 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2757 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2760 * Constructors are not allowed to allocate memory from the same
2761 * cache which they are a constructor for. Otherwise, deadlock.
2762 * They must also be threaded.
2764 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2765 cachep->ctor(objp + obj_offset(cachep));
2767 if (cachep->flags & SLAB_RED_ZONE) {
2768 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2769 slab_error(cachep, "constructor overwrote the"
2770 " end of an object");
2771 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2772 slab_error(cachep, "constructor overwrote the"
2773 " start of an object");
2775 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2776 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2777 kernel_map_pages(virt_to_page(objp),
2778 cachep->buffer_size / PAGE_SIZE, 0);
2783 slab_bufctl(slabp)[i] = i + 1;
2785 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2788 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2790 if (CONFIG_ZONE_DMA_FLAG) {
2791 if (flags & GFP_DMA)
2792 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2794 BUG_ON(cachep->gfpflags & GFP_DMA);
2798 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2801 void *objp = index_to_obj(cachep, slabp, slabp->free);
2805 next = slab_bufctl(slabp)[slabp->free];
2807 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2808 WARN_ON(slabp->nodeid != nodeid);
2815 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2816 void *objp, int nodeid)
2818 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2821 /* Verify that the slab belongs to the intended node */
2822 WARN_ON(slabp->nodeid != nodeid);
2824 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2825 printk(KERN_ERR "slab: double free detected in cache "
2826 "'%s', objp %p\n", cachep->name, objp);
2830 slab_bufctl(slabp)[objnr] = slabp->free;
2831 slabp->free = objnr;
2836 * Map pages beginning at addr to the given cache and slab. This is required
2837 * for the slab allocator to be able to lookup the cache and slab of a
2838 * virtual address for kfree, ksize, and slab debugging.
2840 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2846 page = virt_to_page(addr);
2849 if (likely(!PageCompound(page)))
2850 nr_pages <<= cache->gfporder;
2853 page_set_cache(page, cache);
2854 page_set_slab(page, slab);
2856 } while (--nr_pages);
2860 * Grow (by 1) the number of slabs within a cache. This is called by
2861 * kmem_cache_alloc() when there are no active objs left in a cache.
2863 static int cache_grow(struct kmem_cache *cachep,
2864 gfp_t flags, int nodeid, void *objp)
2869 struct kmem_list3 *l3;
2872 * Be lazy and only check for valid flags here, keeping it out of the
2873 * critical path in kmem_cache_alloc().
2875 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2876 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2878 /* Take the l3 list lock to change the colour_next on this node */
2880 l3 = cachep->nodelists[nodeid];
2881 spin_lock(&l3->list_lock);
2883 /* Get colour for the slab, and cal the next value. */
2884 offset = l3->colour_next;
2886 if (l3->colour_next >= cachep->colour)
2887 l3->colour_next = 0;
2888 spin_unlock(&l3->list_lock);
2890 offset *= cachep->colour_off;
2892 if (local_flags & __GFP_WAIT)
2896 * The test for missing atomic flag is performed here, rather than
2897 * the more obvious place, simply to reduce the critical path length
2898 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2899 * will eventually be caught here (where it matters).
2901 kmem_flagcheck(cachep, flags);
2904 * Get mem for the objs. Attempt to allocate a physical page from
2908 objp = kmem_getpages(cachep, local_flags, nodeid);
2912 /* Get slab management. */
2913 slabp = alloc_slabmgmt(cachep, objp, offset,
2914 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2918 slab_map_pages(cachep, slabp, objp);
2920 cache_init_objs(cachep, slabp);
2922 if (local_flags & __GFP_WAIT)
2923 local_irq_disable();
2925 spin_lock(&l3->list_lock);
2927 /* Make slab active. */
2928 list_add_tail(&slabp->list, &(l3->slabs_free));
2929 STATS_INC_GROWN(cachep);
2930 l3->free_objects += cachep->num;
2931 spin_unlock(&l3->list_lock);
2934 kmem_freepages(cachep, objp);
2936 if (local_flags & __GFP_WAIT)
2937 local_irq_disable();
2944 * Perform extra freeing checks:
2945 * - detect bad pointers.
2946 * - POISON/RED_ZONE checking
2948 static void kfree_debugcheck(const void *objp)
2950 if (!virt_addr_valid(objp)) {
2951 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2952 (unsigned long)objp);
2957 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2959 unsigned long long redzone1, redzone2;
2961 redzone1 = *dbg_redzone1(cache, obj);
2962 redzone2 = *dbg_redzone2(cache, obj);
2967 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2970 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2971 slab_error(cache, "double free detected");
2973 slab_error(cache, "memory outside object was overwritten");
2975 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2976 obj, redzone1, redzone2);
2979 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2986 BUG_ON(virt_to_cache(objp) != cachep);
2988 objp -= obj_offset(cachep);
2989 kfree_debugcheck(objp);
2990 page = virt_to_head_page(objp);
2992 slabp = page_get_slab(page);
2994 if (cachep->flags & SLAB_RED_ZONE) {
2995 verify_redzone_free(cachep, objp);
2996 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2997 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2999 if (cachep->flags & SLAB_STORE_USER)
3000 *dbg_userword(cachep, objp) = caller;
3002 objnr = obj_to_index(cachep, slabp, objp);
3004 BUG_ON(objnr >= cachep->num);
3005 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3007 #ifdef CONFIG_DEBUG_SLAB_LEAK
3008 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3010 if (cachep->flags & SLAB_POISON) {
3011 #ifdef CONFIG_DEBUG_PAGEALLOC
3012 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3013 store_stackinfo(cachep, objp, (unsigned long)caller);
3014 kernel_map_pages(virt_to_page(objp),
3015 cachep->buffer_size / PAGE_SIZE, 0);
3017 poison_obj(cachep, objp, POISON_FREE);
3020 poison_obj(cachep, objp, POISON_FREE);
3026 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3031 /* Check slab's freelist to see if this obj is there. */
3032 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3034 if (entries > cachep->num || i >= cachep->num)
3037 if (entries != cachep->num - slabp->inuse) {
3039 printk(KERN_ERR "slab: Internal list corruption detected in "
3040 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3041 cachep->name, cachep->num, slabp, slabp->inuse);
3042 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3043 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3049 #define kfree_debugcheck(x) do { } while(0)
3050 #define cache_free_debugcheck(x,objp,z) (objp)
3051 #define check_slabp(x,y) do { } while(0)
3054 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3057 struct kmem_list3 *l3;
3058 struct array_cache *ac;
3063 node = numa_mem_id();
3064 ac = cpu_cache_get(cachep);
3065 batchcount = ac->batchcount;
3066 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3068 * If there was little recent activity on this cache, then
3069 * perform only a partial refill. Otherwise we could generate
3072 batchcount = BATCHREFILL_LIMIT;
3074 l3 = cachep->nodelists[node];
3076 BUG_ON(ac->avail > 0 || !l3);
3077 spin_lock(&l3->list_lock);
3079 /* See if we can refill from the shared array */
3080 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3081 l3->shared->touched = 1;
3085 while (batchcount > 0) {
3086 struct list_head *entry;
3088 /* Get slab alloc is to come from. */
3089 entry = l3->slabs_partial.next;
3090 if (entry == &l3->slabs_partial) {
3091 l3->free_touched = 1;
3092 entry = l3->slabs_free.next;
3093 if (entry == &l3->slabs_free)
3097 slabp = list_entry(entry, struct slab, list);
3098 check_slabp(cachep, slabp);
3099 check_spinlock_acquired(cachep);
3102 * The slab was either on partial or free list so
3103 * there must be at least one object available for
3106 BUG_ON(slabp->inuse >= cachep->num);
3108 while (slabp->inuse < cachep->num && batchcount--) {
3109 STATS_INC_ALLOCED(cachep);
3110 STATS_INC_ACTIVE(cachep);
3111 STATS_SET_HIGH(cachep);
3113 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3116 check_slabp(cachep, slabp);
3118 /* move slabp to correct slabp list: */
3119 list_del(&slabp->list);
3120 if (slabp->free == BUFCTL_END)
3121 list_add(&slabp->list, &l3->slabs_full);
3123 list_add(&slabp->list, &l3->slabs_partial);
3127 l3->free_objects -= ac->avail;
3129 spin_unlock(&l3->list_lock);
3131 if (unlikely(!ac->avail)) {
3133 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3135 /* cache_grow can reenable interrupts, then ac could change. */
3136 ac = cpu_cache_get(cachep);
3137 if (!x && ac->avail == 0) /* no objects in sight? abort */
3140 if (!ac->avail) /* objects refilled by interrupt? */
3144 return ac->entry[--ac->avail];
3147 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3150 might_sleep_if(flags & __GFP_WAIT);
3152 kmem_flagcheck(cachep, flags);
3157 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3158 gfp_t flags, void *objp, void *caller)
3162 if (cachep->flags & SLAB_POISON) {
3163 #ifdef CONFIG_DEBUG_PAGEALLOC
3164 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3165 kernel_map_pages(virt_to_page(objp),
3166 cachep->buffer_size / PAGE_SIZE, 1);
3168 check_poison_obj(cachep, objp);
3170 check_poison_obj(cachep, objp);
3172 poison_obj(cachep, objp, POISON_INUSE);
3174 if (cachep->flags & SLAB_STORE_USER)
3175 *dbg_userword(cachep, objp) = caller;
3177 if (cachep->flags & SLAB_RED_ZONE) {
3178 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3179 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3180 slab_error(cachep, "double free, or memory outside"
3181 " object was overwritten");
3183 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3184 objp, *dbg_redzone1(cachep, objp),
3185 *dbg_redzone2(cachep, objp));
3187 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3188 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3190 #ifdef CONFIG_DEBUG_SLAB_LEAK
3195 slabp = page_get_slab(virt_to_head_page(objp));
3196 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3197 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3200 objp += obj_offset(cachep);
3201 if (cachep->ctor && cachep->flags & SLAB_POISON)
3203 if (ARCH_SLAB_MINALIGN &&
3204 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3205 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3206 objp, (int)ARCH_SLAB_MINALIGN);
3211 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3214 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3216 if (cachep == &cache_cache)
3219 return should_failslab(obj_size(cachep), flags, cachep->flags);
3222 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3225 struct array_cache *ac;
3229 ac = cpu_cache_get(cachep);
3230 if (likely(ac->avail)) {
3231 STATS_INC_ALLOCHIT(cachep);
3233 objp = ac->entry[--ac->avail];
3235 STATS_INC_ALLOCMISS(cachep);
3236 objp = cache_alloc_refill(cachep, flags);
3238 * the 'ac' may be updated by cache_alloc_refill(),
3239 * and kmemleak_erase() requires its correct value.
3241 ac = cpu_cache_get(cachep);
3244 * To avoid a false negative, if an object that is in one of the
3245 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3246 * treat the array pointers as a reference to the object.
3249 kmemleak_erase(&ac->entry[ac->avail]);
3255 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3257 * If we are in_interrupt, then process context, including cpusets and
3258 * mempolicy, may not apply and should not be used for allocation policy.
3260 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3262 int nid_alloc, nid_here;
3264 if (in_interrupt() || (flags & __GFP_THISNODE))
3266 nid_alloc = nid_here = numa_mem_id();
3268 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3269 nid_alloc = cpuset_slab_spread_node();
3270 else if (current->mempolicy)
3271 nid_alloc = slab_node(current->mempolicy);
3273 if (nid_alloc != nid_here)
3274 return ____cache_alloc_node(cachep, flags, nid_alloc);
3279 * Fallback function if there was no memory available and no objects on a
3280 * certain node and fall back is permitted. First we scan all the
3281 * available nodelists for available objects. If that fails then we
3282 * perform an allocation without specifying a node. This allows the page
3283 * allocator to do its reclaim / fallback magic. We then insert the
3284 * slab into the proper nodelist and then allocate from it.
3286 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3288 struct zonelist *zonelist;
3292 enum zone_type high_zoneidx = gfp_zone(flags);
3296 if (flags & __GFP_THISNODE)
3300 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3301 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3305 * Look through allowed nodes for objects available
3306 * from existing per node queues.
3308 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3309 nid = zone_to_nid(zone);
3311 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3312 cache->nodelists[nid] &&
3313 cache->nodelists[nid]->free_objects) {
3314 obj = ____cache_alloc_node(cache,
3315 flags | GFP_THISNODE, nid);
3323 * This allocation will be performed within the constraints
3324 * of the current cpuset / memory policy requirements.
3325 * We may trigger various forms of reclaim on the allowed
3326 * set and go into memory reserves if necessary.
3328 if (local_flags & __GFP_WAIT)
3330 kmem_flagcheck(cache, flags);
3331 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3332 if (local_flags & __GFP_WAIT)
3333 local_irq_disable();
3336 * Insert into the appropriate per node queues
3338 nid = page_to_nid(virt_to_page(obj));
3339 if (cache_grow(cache, flags, nid, obj)) {
3340 obj = ____cache_alloc_node(cache,
3341 flags | GFP_THISNODE, nid);
3344 * Another processor may allocate the
3345 * objects in the slab since we are
3346 * not holding any locks.
3350 /* cache_grow already freed obj */
3360 * A interface to enable slab creation on nodeid
3362 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3365 struct list_head *entry;
3367 struct kmem_list3 *l3;
3371 l3 = cachep->nodelists[nodeid];
3376 spin_lock(&l3->list_lock);
3377 entry = l3->slabs_partial.next;
3378 if (entry == &l3->slabs_partial) {
3379 l3->free_touched = 1;
3380 entry = l3->slabs_free.next;
3381 if (entry == &l3->slabs_free)
3385 slabp = list_entry(entry, struct slab, list);
3386 check_spinlock_acquired_node(cachep, nodeid);
3387 check_slabp(cachep, slabp);
3389 STATS_INC_NODEALLOCS(cachep);
3390 STATS_INC_ACTIVE(cachep);
3391 STATS_SET_HIGH(cachep);
3393 BUG_ON(slabp->inuse == cachep->num);
3395 obj = slab_get_obj(cachep, slabp, nodeid);
3396 check_slabp(cachep, slabp);
3398 /* move slabp to correct slabp list: */
3399 list_del(&slabp->list);
3401 if (slabp->free == BUFCTL_END)
3402 list_add(&slabp->list, &l3->slabs_full);
3404 list_add(&slabp->list, &l3->slabs_partial);
3406 spin_unlock(&l3->list_lock);
3410 spin_unlock(&l3->list_lock);
3411 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3415 return fallback_alloc(cachep, flags);
3422 * kmem_cache_alloc_node - Allocate an object on the specified node
3423 * @cachep: The cache to allocate from.
3424 * @flags: See kmalloc().
3425 * @nodeid: node number of the target node.
3426 * @caller: return address of caller, used for debug information
3428 * Identical to kmem_cache_alloc but it will allocate memory on the given
3429 * node, which can improve the performance for cpu bound structures.
3431 * Fallback to other node is possible if __GFP_THISNODE is not set.
3433 static __always_inline void *
3434 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3437 unsigned long save_flags;
3439 int slab_node = numa_mem_id();
3441 flags &= gfp_allowed_mask;
3443 lockdep_trace_alloc(flags);
3445 if (slab_should_failslab(cachep, flags))
3448 cache_alloc_debugcheck_before(cachep, flags);
3449 local_irq_save(save_flags);
3451 if (nodeid == NUMA_NO_NODE)
3454 if (unlikely(!cachep->nodelists[nodeid])) {
3455 /* Node not bootstrapped yet */
3456 ptr = fallback_alloc(cachep, flags);
3460 if (nodeid == slab_node) {
3462 * Use the locally cached objects if possible.
3463 * However ____cache_alloc does not allow fallback
3464 * to other nodes. It may fail while we still have
3465 * objects on other nodes available.
3467 ptr = ____cache_alloc(cachep, flags);
3471 /* ___cache_alloc_node can fall back to other nodes */
3472 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3474 local_irq_restore(save_flags);
3475 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3476 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3480 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3482 if (unlikely((flags & __GFP_ZERO) && ptr))
3483 memset(ptr, 0, obj_size(cachep));
3488 static __always_inline void *
3489 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3493 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3494 objp = alternate_node_alloc(cache, flags);
3498 objp = ____cache_alloc(cache, flags);
3501 * We may just have run out of memory on the local node.
3502 * ____cache_alloc_node() knows how to locate memory on other nodes
3505 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3512 static __always_inline void *
3513 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3515 return ____cache_alloc(cachep, flags);
3518 #endif /* CONFIG_NUMA */
3520 static __always_inline void *
3521 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3523 unsigned long save_flags;
3526 flags &= gfp_allowed_mask;
3528 lockdep_trace_alloc(flags);
3530 if (slab_should_failslab(cachep, flags))
3533 cache_alloc_debugcheck_before(cachep, flags);
3534 local_irq_save(save_flags);
3535 objp = __do_cache_alloc(cachep, flags);
3536 local_irq_restore(save_flags);
3537 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3538 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3543 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3545 if (unlikely((flags & __GFP_ZERO) && objp))
3546 memset(objp, 0, obj_size(cachep));
3552 * Caller needs to acquire correct kmem_list's list_lock
3554 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3558 struct kmem_list3 *l3;
3560 for (i = 0; i < nr_objects; i++) {
3561 void *objp = objpp[i];
3564 slabp = virt_to_slab(objp);
3565 l3 = cachep->nodelists[node];
3566 list_del(&slabp->list);
3567 check_spinlock_acquired_node(cachep, node);
3568 check_slabp(cachep, slabp);
3569 slab_put_obj(cachep, slabp, objp, node);
3570 STATS_DEC_ACTIVE(cachep);
3572 check_slabp(cachep, slabp);
3574 /* fixup slab chains */
3575 if (slabp->inuse == 0) {
3576 if (l3->free_objects > l3->free_limit) {
3577 l3->free_objects -= cachep->num;
3578 /* No need to drop any previously held
3579 * lock here, even if we have a off-slab slab
3580 * descriptor it is guaranteed to come from
3581 * a different cache, refer to comments before
3584 slab_destroy(cachep, slabp);
3586 list_add(&slabp->list, &l3->slabs_free);
3589 /* Unconditionally move a slab to the end of the
3590 * partial list on free - maximum time for the
3591 * other objects to be freed, too.
3593 list_add_tail(&slabp->list, &l3->slabs_partial);
3598 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3601 struct kmem_list3 *l3;
3602 int node = numa_mem_id();
3604 batchcount = ac->batchcount;
3606 BUG_ON(!batchcount || batchcount > ac->avail);
3609 l3 = cachep->nodelists[node];
3610 spin_lock(&l3->list_lock);
3612 struct array_cache *shared_array = l3->shared;
3613 int max = shared_array->limit - shared_array->avail;
3615 if (batchcount > max)
3617 memcpy(&(shared_array->entry[shared_array->avail]),
3618 ac->entry, sizeof(void *) * batchcount);
3619 shared_array->avail += batchcount;
3624 free_block(cachep, ac->entry, batchcount, node);
3629 struct list_head *p;
3631 p = l3->slabs_free.next;
3632 while (p != &(l3->slabs_free)) {
3635 slabp = list_entry(p, struct slab, list);
3636 BUG_ON(slabp->inuse);
3641 STATS_SET_FREEABLE(cachep, i);
3644 spin_unlock(&l3->list_lock);
3645 ac->avail -= batchcount;
3646 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3650 * Release an obj back to its cache. If the obj has a constructed state, it must
3651 * be in this state _before_ it is released. Called with disabled ints.
3653 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3656 struct array_cache *ac = cpu_cache_get(cachep);
3659 kmemleak_free_recursive(objp, cachep->flags);
3660 objp = cache_free_debugcheck(cachep, objp, caller);
3662 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3665 * Skip calling cache_free_alien() when the platform is not numa.
3666 * This will avoid cache misses that happen while accessing slabp (which
3667 * is per page memory reference) to get nodeid. Instead use a global
3668 * variable to skip the call, which is mostly likely to be present in
3671 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3674 if (likely(ac->avail < ac->limit)) {
3675 STATS_INC_FREEHIT(cachep);
3676 ac->entry[ac->avail++] = objp;
3679 STATS_INC_FREEMISS(cachep);
3680 cache_flusharray(cachep, ac);
3681 ac->entry[ac->avail++] = objp;
3686 * kmem_cache_alloc - Allocate an object
3687 * @cachep: The cache to allocate from.
3688 * @flags: See kmalloc().
3690 * Allocate an object from this cache. The flags are only relevant
3691 * if the cache has no available objects.
3693 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3695 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3697 trace_kmem_cache_alloc(_RET_IP_, ret,
3698 obj_size(cachep), cachep->buffer_size, flags);
3702 EXPORT_SYMBOL(kmem_cache_alloc);
3704 #ifdef CONFIG_TRACING
3706 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3710 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3712 trace_kmalloc(_RET_IP_, ret,
3713 size, slab_buffer_size(cachep), flags);
3716 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3720 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3722 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3723 __builtin_return_address(0));
3725 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3726 obj_size(cachep), cachep->buffer_size,
3731 EXPORT_SYMBOL(kmem_cache_alloc_node);
3733 #ifdef CONFIG_TRACING
3734 void *kmem_cache_alloc_node_trace(size_t size,
3735 struct kmem_cache *cachep,
3741 ret = __cache_alloc_node(cachep, flags, nodeid,
3742 __builtin_return_address(0));
3743 trace_kmalloc_node(_RET_IP_, ret,
3744 size, slab_buffer_size(cachep),
3748 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3751 static __always_inline void *
3752 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3754 struct kmem_cache *cachep;
3756 cachep = kmem_find_general_cachep(size, flags);
3757 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3759 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3762 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3763 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3765 return __do_kmalloc_node(size, flags, node,
3766 __builtin_return_address(0));
3768 EXPORT_SYMBOL(__kmalloc_node);
3770 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3771 int node, unsigned long caller)
3773 return __do_kmalloc_node(size, flags, node, (void *)caller);
3775 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3777 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3779 return __do_kmalloc_node(size, flags, node, NULL);
3781 EXPORT_SYMBOL(__kmalloc_node);
3782 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3783 #endif /* CONFIG_NUMA */
3786 * __do_kmalloc - allocate memory
3787 * @size: how many bytes of memory are required.
3788 * @flags: the type of memory to allocate (see kmalloc).
3789 * @caller: function caller for debug tracking of the caller
3791 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3794 struct kmem_cache *cachep;
3797 /* If you want to save a few bytes .text space: replace
3799 * Then kmalloc uses the uninlined functions instead of the inline
3802 cachep = __find_general_cachep(size, flags);
3803 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3805 ret = __cache_alloc(cachep, flags, caller);
3807 trace_kmalloc((unsigned long) caller, ret,
3808 size, cachep->buffer_size, flags);
3814 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3815 void *__kmalloc(size_t size, gfp_t flags)
3817 return __do_kmalloc(size, flags, __builtin_return_address(0));
3819 EXPORT_SYMBOL(__kmalloc);
3821 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3823 return __do_kmalloc(size, flags, (void *)caller);
3825 EXPORT_SYMBOL(__kmalloc_track_caller);
3828 void *__kmalloc(size_t size, gfp_t flags)
3830 return __do_kmalloc(size, flags, NULL);
3832 EXPORT_SYMBOL(__kmalloc);
3836 * kmem_cache_free - Deallocate an object
3837 * @cachep: The cache the allocation was from.
3838 * @objp: The previously allocated object.
3840 * Free an object which was previously allocated from this
3843 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3845 unsigned long flags;
3847 local_irq_save(flags);
3848 debug_check_no_locks_freed(objp, obj_size(cachep));
3849 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3850 debug_check_no_obj_freed(objp, obj_size(cachep));
3851 __cache_free(cachep, objp, __builtin_return_address(0));
3852 local_irq_restore(flags);
3854 trace_kmem_cache_free(_RET_IP_, objp);
3856 EXPORT_SYMBOL(kmem_cache_free);
3859 * kfree - free previously allocated memory
3860 * @objp: pointer returned by kmalloc.
3862 * If @objp is NULL, no operation is performed.
3864 * Don't free memory not originally allocated by kmalloc()
3865 * or you will run into trouble.
3867 void kfree(const void *objp)
3869 struct kmem_cache *c;
3870 unsigned long flags;
3872 trace_kfree(_RET_IP_, objp);
3874 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3876 local_irq_save(flags);
3877 kfree_debugcheck(objp);
3878 c = virt_to_cache(objp);
3879 debug_check_no_locks_freed(objp, obj_size(c));
3880 debug_check_no_obj_freed(objp, obj_size(c));
3881 __cache_free(c, (void *)objp, __builtin_return_address(0));
3882 local_irq_restore(flags);
3884 EXPORT_SYMBOL(kfree);
3886 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3888 return obj_size(cachep);
3890 EXPORT_SYMBOL(kmem_cache_size);
3893 * This initializes kmem_list3 or resizes various caches for all nodes.
3895 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3898 struct kmem_list3 *l3;
3899 struct array_cache *new_shared;
3900 struct array_cache **new_alien = NULL;
3902 for_each_online_node(node) {
3904 if (use_alien_caches) {
3905 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3911 if (cachep->shared) {
3912 new_shared = alloc_arraycache(node,
3913 cachep->shared*cachep->batchcount,
3916 free_alien_cache(new_alien);
3921 l3 = cachep->nodelists[node];
3923 struct array_cache *shared = l3->shared;
3925 spin_lock_irq(&l3->list_lock);
3928 free_block(cachep, shared->entry,
3929 shared->avail, node);
3931 l3->shared = new_shared;
3933 l3->alien = new_alien;
3936 l3->free_limit = (1 + nr_cpus_node(node)) *
3937 cachep->batchcount + cachep->num;
3938 spin_unlock_irq(&l3->list_lock);
3940 free_alien_cache(new_alien);
3943 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3945 free_alien_cache(new_alien);
3950 kmem_list3_init(l3);
3951 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3952 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3953 l3->shared = new_shared;
3954 l3->alien = new_alien;
3955 l3->free_limit = (1 + nr_cpus_node(node)) *
3956 cachep->batchcount + cachep->num;
3957 cachep->nodelists[node] = l3;
3962 if (!cachep->next.next) {
3963 /* Cache is not active yet. Roll back what we did */
3966 if (cachep->nodelists[node]) {
3967 l3 = cachep->nodelists[node];
3970 free_alien_cache(l3->alien);
3972 cachep->nodelists[node] = NULL;
3980 struct ccupdate_struct {
3981 struct kmem_cache *cachep;
3982 struct array_cache *new[0];
3985 static void do_ccupdate_local(void *info)
3987 struct ccupdate_struct *new = info;
3988 struct array_cache *old;
3991 old = cpu_cache_get(new->cachep);
3993 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3994 new->new[smp_processor_id()] = old;
3997 /* Always called with the cache_chain_mutex held */
3998 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3999 int batchcount, int shared, gfp_t gfp)
4001 struct ccupdate_struct *new;
4004 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4009 for_each_online_cpu(i) {
4010 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4013 for (i--; i >= 0; i--)
4019 new->cachep = cachep;
4021 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4024 cachep->batchcount = batchcount;
4025 cachep->limit = limit;
4026 cachep->shared = shared;
4028 for_each_online_cpu(i) {
4029 struct array_cache *ccold = new->new[i];
4032 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4033 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4034 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4038 return alloc_kmemlist(cachep, gfp);
4041 /* Called with cache_chain_mutex held always */
4042 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4048 * The head array serves three purposes:
4049 * - create a LIFO ordering, i.e. return objects that are cache-warm
4050 * - reduce the number of spinlock operations.
4051 * - reduce the number of linked list operations on the slab and
4052 * bufctl chains: array operations are cheaper.
4053 * The numbers are guessed, we should auto-tune as described by
4056 if (cachep->buffer_size > 131072)
4058 else if (cachep->buffer_size > PAGE_SIZE)
4060 else if (cachep->buffer_size > 1024)
4062 else if (cachep->buffer_size > 256)
4068 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4069 * allocation behaviour: Most allocs on one cpu, most free operations
4070 * on another cpu. For these cases, an efficient object passing between
4071 * cpus is necessary. This is provided by a shared array. The array
4072 * replaces Bonwick's magazine layer.
4073 * On uniprocessor, it's functionally equivalent (but less efficient)
4074 * to a larger limit. Thus disabled by default.
4077 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4082 * With debugging enabled, large batchcount lead to excessively long
4083 * periods with disabled local interrupts. Limit the batchcount
4088 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4090 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4091 cachep->name, -err);
4096 * Drain an array if it contains any elements taking the l3 lock only if
4097 * necessary. Note that the l3 listlock also protects the array_cache
4098 * if drain_array() is used on the shared array.
4100 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4101 struct array_cache *ac, int force, int node)
4105 if (!ac || !ac->avail)
4107 if (ac->touched && !force) {
4110 spin_lock_irq(&l3->list_lock);
4112 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4113 if (tofree > ac->avail)
4114 tofree = (ac->avail + 1) / 2;
4115 free_block(cachep, ac->entry, tofree, node);
4116 ac->avail -= tofree;
4117 memmove(ac->entry, &(ac->entry[tofree]),
4118 sizeof(void *) * ac->avail);
4120 spin_unlock_irq(&l3->list_lock);
4125 * cache_reap - Reclaim memory from caches.
4126 * @w: work descriptor
4128 * Called from workqueue/eventd every few seconds.
4130 * - clear the per-cpu caches for this CPU.
4131 * - return freeable pages to the main free memory pool.
4133 * If we cannot acquire the cache chain mutex then just give up - we'll try
4134 * again on the next iteration.
4136 static void cache_reap(struct work_struct *w)
4138 struct kmem_cache *searchp;
4139 struct kmem_list3 *l3;
4140 int node = numa_mem_id();
4141 struct delayed_work *work = to_delayed_work(w);
4143 if (!mutex_trylock(&cache_chain_mutex))
4144 /* Give up. Setup the next iteration. */
4147 list_for_each_entry(searchp, &cache_chain, next) {
4151 * We only take the l3 lock if absolutely necessary and we
4152 * have established with reasonable certainty that
4153 * we can do some work if the lock was obtained.
4155 l3 = searchp->nodelists[node];
4157 reap_alien(searchp, l3);
4159 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4162 * These are racy checks but it does not matter
4163 * if we skip one check or scan twice.
4165 if (time_after(l3->next_reap, jiffies))
4168 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4170 drain_array(searchp, l3, l3->shared, 0, node);
4172 if (l3->free_touched)
4173 l3->free_touched = 0;
4177 freed = drain_freelist(searchp, l3, (l3->free_limit +
4178 5 * searchp->num - 1) / (5 * searchp->num));
4179 STATS_ADD_REAPED(searchp, freed);
4185 mutex_unlock(&cache_chain_mutex);
4188 /* Set up the next iteration */
4189 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4192 #ifdef CONFIG_SLABINFO
4194 static void print_slabinfo_header(struct seq_file *m)
4197 * Output format version, so at least we can change it
4198 * without _too_ many complaints.
4201 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4203 seq_puts(m, "slabinfo - version: 2.1\n");
4205 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4206 "<objperslab> <pagesperslab>");
4207 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4208 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4210 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4211 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4212 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4217 static void *s_start(struct seq_file *m, loff_t *pos)
4221 mutex_lock(&cache_chain_mutex);
4223 print_slabinfo_header(m);
4225 return seq_list_start(&cache_chain, *pos);
4228 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4230 return seq_list_next(p, &cache_chain, pos);
4233 static void s_stop(struct seq_file *m, void *p)
4235 mutex_unlock(&cache_chain_mutex);
4238 static int s_show(struct seq_file *m, void *p)
4240 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4242 unsigned long active_objs;
4243 unsigned long num_objs;
4244 unsigned long active_slabs = 0;
4245 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4249 struct kmem_list3 *l3;
4253 for_each_online_node(node) {
4254 l3 = cachep->nodelists[node];
4259 spin_lock_irq(&l3->list_lock);
4261 list_for_each_entry(slabp, &l3->slabs_full, list) {
4262 if (slabp->inuse != cachep->num && !error)
4263 error = "slabs_full accounting error";
4264 active_objs += cachep->num;
4267 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4268 if (slabp->inuse == cachep->num && !error)
4269 error = "slabs_partial inuse accounting error";
4270 if (!slabp->inuse && !error)
4271 error = "slabs_partial/inuse accounting error";
4272 active_objs += slabp->inuse;
4275 list_for_each_entry(slabp, &l3->slabs_free, list) {
4276 if (slabp->inuse && !error)
4277 error = "slabs_free/inuse accounting error";
4280 free_objects += l3->free_objects;
4282 shared_avail += l3->shared->avail;
4284 spin_unlock_irq(&l3->list_lock);
4286 num_slabs += active_slabs;
4287 num_objs = num_slabs * cachep->num;
4288 if (num_objs - active_objs != free_objects && !error)
4289 error = "free_objects accounting error";
4291 name = cachep->name;
4293 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4295 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4296 name, active_objs, num_objs, cachep->buffer_size,
4297 cachep->num, (1 << cachep->gfporder));
4298 seq_printf(m, " : tunables %4u %4u %4u",
4299 cachep->limit, cachep->batchcount, cachep->shared);
4300 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4301 active_slabs, num_slabs, shared_avail);
4304 unsigned long high = cachep->high_mark;
4305 unsigned long allocs = cachep->num_allocations;
4306 unsigned long grown = cachep->grown;
4307 unsigned long reaped = cachep->reaped;
4308 unsigned long errors = cachep->errors;
4309 unsigned long max_freeable = cachep->max_freeable;
4310 unsigned long node_allocs = cachep->node_allocs;
4311 unsigned long node_frees = cachep->node_frees;
4312 unsigned long overflows = cachep->node_overflow;
4314 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4315 "%4lu %4lu %4lu %4lu %4lu",
4316 allocs, high, grown,
4317 reaped, errors, max_freeable, node_allocs,
4318 node_frees, overflows);
4322 unsigned long allochit = atomic_read(&cachep->allochit);
4323 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4324 unsigned long freehit = atomic_read(&cachep->freehit);
4325 unsigned long freemiss = atomic_read(&cachep->freemiss);
4327 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4328 allochit, allocmiss, freehit, freemiss);
4336 * slabinfo_op - iterator that generates /proc/slabinfo
4345 * num-pages-per-slab
4346 * + further values on SMP and with statistics enabled
4349 static const struct seq_operations slabinfo_op = {
4356 #define MAX_SLABINFO_WRITE 128
4358 * slabinfo_write - Tuning for the slab allocator
4360 * @buffer: user buffer
4361 * @count: data length
4364 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4365 size_t count, loff_t *ppos)
4367 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4368 int limit, batchcount, shared, res;
4369 struct kmem_cache *cachep;
4371 if (count > MAX_SLABINFO_WRITE)
4373 if (copy_from_user(&kbuf, buffer, count))
4375 kbuf[MAX_SLABINFO_WRITE] = '\0';
4377 tmp = strchr(kbuf, ' ');
4382 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4385 /* Find the cache in the chain of caches. */
4386 mutex_lock(&cache_chain_mutex);
4388 list_for_each_entry(cachep, &cache_chain, next) {
4389 if (!strcmp(cachep->name, kbuf)) {
4390 if (limit < 1 || batchcount < 1 ||
4391 batchcount > limit || shared < 0) {
4394 res = do_tune_cpucache(cachep, limit,
4401 mutex_unlock(&cache_chain_mutex);
4407 static int slabinfo_open(struct inode *inode, struct file *file)
4409 return seq_open(file, &slabinfo_op);
4412 static const struct file_operations proc_slabinfo_operations = {
4413 .open = slabinfo_open,
4415 .write = slabinfo_write,
4416 .llseek = seq_lseek,
4417 .release = seq_release,
4420 #ifdef CONFIG_DEBUG_SLAB_LEAK
4422 static void *leaks_start(struct seq_file *m, loff_t *pos)
4424 mutex_lock(&cache_chain_mutex);
4425 return seq_list_start(&cache_chain, *pos);
4428 static inline int add_caller(unsigned long *n, unsigned long v)
4438 unsigned long *q = p + 2 * i;
4452 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4458 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4464 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4465 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4467 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4472 static void show_symbol(struct seq_file *m, unsigned long address)
4474 #ifdef CONFIG_KALLSYMS
4475 unsigned long offset, size;
4476 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4478 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4479 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4481 seq_printf(m, " [%s]", modname);
4485 seq_printf(m, "%p", (void *)address);
4488 static int leaks_show(struct seq_file *m, void *p)
4490 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4492 struct kmem_list3 *l3;
4494 unsigned long *n = m->private;
4498 if (!(cachep->flags & SLAB_STORE_USER))
4500 if (!(cachep->flags & SLAB_RED_ZONE))
4503 /* OK, we can do it */
4507 for_each_online_node(node) {
4508 l3 = cachep->nodelists[node];
4513 spin_lock_irq(&l3->list_lock);
4515 list_for_each_entry(slabp, &l3->slabs_full, list)
4516 handle_slab(n, cachep, slabp);
4517 list_for_each_entry(slabp, &l3->slabs_partial, list)
4518 handle_slab(n, cachep, slabp);
4519 spin_unlock_irq(&l3->list_lock);
4521 name = cachep->name;
4523 /* Increase the buffer size */
4524 mutex_unlock(&cache_chain_mutex);
4525 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4527 /* Too bad, we are really out */
4529 mutex_lock(&cache_chain_mutex);
4532 *(unsigned long *)m->private = n[0] * 2;
4534 mutex_lock(&cache_chain_mutex);
4535 /* Now make sure this entry will be retried */
4539 for (i = 0; i < n[1]; i++) {
4540 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4541 show_symbol(m, n[2*i+2]);
4548 static const struct seq_operations slabstats_op = {
4549 .start = leaks_start,
4555 static int slabstats_open(struct inode *inode, struct file *file)
4557 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4560 ret = seq_open(file, &slabstats_op);
4562 struct seq_file *m = file->private_data;
4563 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4572 static const struct file_operations proc_slabstats_operations = {
4573 .open = slabstats_open,
4575 .llseek = seq_lseek,
4576 .release = seq_release_private,
4580 static int __init slab_proc_init(void)
4582 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4583 #ifdef CONFIG_DEBUG_SLAB_LEAK
4584 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4588 module_init(slab_proc_init);
4592 * ksize - get the actual amount of memory allocated for a given object
4593 * @objp: Pointer to the object
4595 * kmalloc may internally round up allocations and return more memory
4596 * than requested. ksize() can be used to determine the actual amount of
4597 * memory allocated. The caller may use this additional memory, even though
4598 * a smaller amount of memory was initially specified with the kmalloc call.
4599 * The caller must guarantee that objp points to a valid object previously
4600 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4601 * must not be freed during the duration of the call.
4603 size_t ksize(const void *objp)
4606 if (unlikely(objp == ZERO_SIZE_PTR))
4609 return obj_size(virt_to_cache(objp));
4611 EXPORT_SYMBOL(ksize);