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 intializations 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/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
120 * SLAB_RED_ZONE & SLAB_POISON.
121 * 0 for faster, smaller code (especially in the critical paths).
123 * STATS - 1 to collect stats for /proc/slabinfo.
124 * 0 for faster, smaller code (especially in the critical paths).
126 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
129 #ifdef CONFIG_DEBUG_SLAB
132 #define FORCED_DEBUG 1
136 #define FORCED_DEBUG 0
139 /* Shouldn't this be in a header file somewhere? */
140 #define BYTES_PER_WORD sizeof(void *)
142 #ifndef cache_line_size
143 #define cache_line_size() L1_CACHE_BYTES
146 #ifndef ARCH_KMALLOC_MINALIGN
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
153 * Note that this flag disables some debug features.
155 #define ARCH_KMALLOC_MINALIGN 0
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
178 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
183 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
191 * Bufctl's are used for linking objs within a slab
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 typedef unsigned int kmem_bufctl_t;
208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * The slab lists for all objects.
291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
295 unsigned int free_limit;
296 unsigned int colour_next; /* Per-node cache coloring */
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 static int enable_cpucache(struct kmem_cache *cachep);
318 static void cache_reap(struct work_struct *unused);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline int index_of(const size_t size)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size)) {
336 #include "linux/kmalloc_sizes.h"
344 static int slab_early_init = 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3 *parent)
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
356 parent->colour_next = 0;
357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache *array[NR_CPUS];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount;
389 unsigned int buffer_size;
390 u32 reciprocal_buffer_size;
391 /* 3) touched by every alloc & free from the backend */
392 struct kmem_list3 *nodelists[MAX_NUMNODES];
394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
397 /* 4) cache_grow/shrink */
398 /* order of pgs per slab (2^n) */
399 unsigned int gfporder;
401 /* force GFP flags, e.g. GFP_DMA */
404 size_t colour; /* cache colouring range */
405 unsigned int colour_off; /* colour offset */
406 struct kmem_cache *slabp_cache;
407 unsigned int slab_size;
408 unsigned int dflags; /* dynamic flags */
410 /* constructor func */
411 void (*ctor) (void *, struct kmem_cache *, unsigned long);
413 /* de-constructor func */
414 void (*dtor) (void *, struct kmem_cache *, unsigned long);
416 /* 5) cache creation/removal */
418 struct list_head next;
422 unsigned long num_active;
423 unsigned long num_allocations;
424 unsigned long high_mark;
426 unsigned long reaped;
427 unsigned long errors;
428 unsigned long max_freeable;
429 unsigned long node_allocs;
430 unsigned long node_frees;
431 unsigned long node_overflow;
439 * If debugging is enabled, then the allocator can add additional
440 * fields and/or padding to every object. buffer_size contains the total
441 * object size including these internal fields, the following two
442 * variables contain the offset to the user object and its size.
449 #define CFLGS_OFF_SLAB (0x80000000UL)
450 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
452 #define BATCHREFILL_LIMIT 16
454 * Optimization question: fewer reaps means less probability for unnessary
455 * cpucache drain/refill cycles.
457 * OTOH the cpuarrays can contain lots of objects,
458 * which could lock up otherwise freeable slabs.
460 #define REAPTIMEOUT_CPUC (2*HZ)
461 #define REAPTIMEOUT_LIST3 (4*HZ)
464 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
465 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
466 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
467 #define STATS_INC_GROWN(x) ((x)->grown++)
468 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
469 #define STATS_SET_HIGH(x) \
471 if ((x)->num_active > (x)->high_mark) \
472 (x)->high_mark = (x)->num_active; \
474 #define STATS_INC_ERR(x) ((x)->errors++)
475 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
476 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
477 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
478 #define STATS_SET_FREEABLE(x, i) \
480 if ((x)->max_freeable < i) \
481 (x)->max_freeable = i; \
483 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
484 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
485 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
486 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
488 #define STATS_INC_ACTIVE(x) do { } while (0)
489 #define STATS_DEC_ACTIVE(x) do { } while (0)
490 #define STATS_INC_ALLOCED(x) do { } while (0)
491 #define STATS_INC_GROWN(x) do { } while (0)
492 #define STATS_ADD_REAPED(x,y) do { } while (0)
493 #define STATS_SET_HIGH(x) do { } while (0)
494 #define STATS_INC_ERR(x) do { } while (0)
495 #define STATS_INC_NODEALLOCS(x) do { } while (0)
496 #define STATS_INC_NODEFREES(x) do { } while (0)
497 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
498 #define STATS_SET_FREEABLE(x, i) do { } while (0)
499 #define STATS_INC_ALLOCHIT(x) do { } while (0)
500 #define STATS_INC_ALLOCMISS(x) do { } while (0)
501 #define STATS_INC_FREEHIT(x) do { } while (0)
502 #define STATS_INC_FREEMISS(x) do { } while (0)
510 * memory layout of objects:
512 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
513 * the end of an object is aligned with the end of the real
514 * allocation. Catches writes behind the end of the allocation.
515 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
517 * cachep->obj_offset: The real object.
518 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
519 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
520 * [BYTES_PER_WORD long]
522 static int obj_offset(struct kmem_cache *cachep)
524 return cachep->obj_offset;
527 static int obj_size(struct kmem_cache *cachep)
529 return cachep->obj_size;
532 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
534 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
535 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
538 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
541 if (cachep->flags & SLAB_STORE_USER)
542 return (unsigned long *)(objp + cachep->buffer_size -
544 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
547 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
549 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
550 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
555 #define obj_offset(x) 0
556 #define obj_size(cachep) (cachep->buffer_size)
557 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
558 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
559 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
564 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
567 #if defined(CONFIG_LARGE_ALLOCS)
568 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
569 #define MAX_GFP_ORDER 13 /* up to 32Mb */
570 #elif defined(CONFIG_MMU)
571 #define MAX_OBJ_ORDER 5 /* 32 pages */
572 #define MAX_GFP_ORDER 5 /* 32 pages */
574 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
575 #define MAX_GFP_ORDER 8 /* up to 1Mb */
579 * Do not go above this order unless 0 objects fit into the slab.
581 #define BREAK_GFP_ORDER_HI 1
582 #define BREAK_GFP_ORDER_LO 0
583 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
586 * Functions for storing/retrieving the cachep and or slab from the page
587 * allocator. These are used to find the slab an obj belongs to. With kfree(),
588 * these are used to find the cache which an obj belongs to.
590 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
592 page->lru.next = (struct list_head *)cache;
595 static inline struct kmem_cache *page_get_cache(struct page *page)
597 if (unlikely(PageCompound(page)))
598 page = (struct page *)page_private(page);
599 BUG_ON(!PageSlab(page));
600 return (struct kmem_cache *)page->lru.next;
603 static inline void page_set_slab(struct page *page, struct slab *slab)
605 page->lru.prev = (struct list_head *)slab;
608 static inline struct slab *page_get_slab(struct page *page)
610 if (unlikely(PageCompound(page)))
611 page = (struct page *)page_private(page);
612 BUG_ON(!PageSlab(page));
613 return (struct slab *)page->lru.prev;
616 static inline struct kmem_cache *virt_to_cache(const void *obj)
618 struct page *page = virt_to_page(obj);
619 return page_get_cache(page);
622 static inline struct slab *virt_to_slab(const void *obj)
624 struct page *page = virt_to_page(obj);
625 return page_get_slab(page);
628 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
631 return slab->s_mem + cache->buffer_size * idx;
635 * We want to avoid an expensive divide : (offset / cache->buffer_size)
636 * Using the fact that buffer_size is a constant for a particular cache,
637 * we can replace (offset / cache->buffer_size) by
638 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
640 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
641 const struct slab *slab, void *obj)
643 u32 offset = (obj - slab->s_mem);
644 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
648 * These are the default caches for kmalloc. Custom caches can have other sizes.
650 struct cache_sizes malloc_sizes[] = {
651 #define CACHE(x) { .cs_size = (x) },
652 #include <linux/kmalloc_sizes.h>
656 EXPORT_SYMBOL(malloc_sizes);
658 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
664 static struct cache_names __initdata cache_names[] = {
665 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
666 #include <linux/kmalloc_sizes.h>
671 static struct arraycache_init initarray_cache __initdata =
672 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
673 static struct arraycache_init initarray_generic =
674 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
676 /* internal cache of cache description objs */
677 static struct kmem_cache cache_cache = {
679 .limit = BOOT_CPUCACHE_ENTRIES,
681 .buffer_size = sizeof(struct kmem_cache),
682 .name = "kmem_cache",
684 .obj_size = sizeof(struct kmem_cache),
688 #define BAD_ALIEN_MAGIC 0x01020304ul
690 #ifdef CONFIG_LOCKDEP
693 * Slab sometimes uses the kmalloc slabs to store the slab headers
694 * for other slabs "off slab".
695 * The locking for this is tricky in that it nests within the locks
696 * of all other slabs in a few places; to deal with this special
697 * locking we put on-slab caches into a separate lock-class.
699 * We set lock class for alien array caches which are up during init.
700 * The lock annotation will be lost if all cpus of a node goes down and
701 * then comes back up during hotplug
703 static struct lock_class_key on_slab_l3_key;
704 static struct lock_class_key on_slab_alc_key;
706 static inline void init_lock_keys(void)
710 struct cache_sizes *s = malloc_sizes;
712 while (s->cs_size != ULONG_MAX) {
714 struct array_cache **alc;
716 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
717 if (!l3 || OFF_SLAB(s->cs_cachep))
719 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
722 * FIXME: This check for BAD_ALIEN_MAGIC
723 * should go away when common slab code is taught to
724 * work even without alien caches.
725 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
726 * for alloc_alien_cache,
728 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
732 lockdep_set_class(&alc[r]->lock,
740 static inline void init_lock_keys(void)
746 * 1. Guard access to the cache-chain.
747 * 2. Protect sanity of cpu_online_map against cpu hotplug events
749 static DEFINE_MUTEX(cache_chain_mutex);
750 static struct list_head cache_chain;
753 * chicken and egg problem: delay the per-cpu array allocation
754 * until the general caches are up.
764 * used by boot code to determine if it can use slab based allocator
766 int slab_is_available(void)
768 return g_cpucache_up == FULL;
771 static DEFINE_PER_CPU(struct delayed_work, reap_work);
773 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
775 return cachep->array[smp_processor_id()];
778 static inline struct kmem_cache *__find_general_cachep(size_t size,
781 struct cache_sizes *csizep = malloc_sizes;
784 /* This happens if someone tries to call
785 * kmem_cache_create(), or __kmalloc(), before
786 * the generic caches are initialized.
788 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
790 while (size > csizep->cs_size)
794 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
795 * has cs_{dma,}cachep==NULL. Thus no special case
796 * for large kmalloc calls required.
798 if (unlikely(gfpflags & GFP_DMA))
799 return csizep->cs_dmacachep;
800 return csizep->cs_cachep;
803 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
805 return __find_general_cachep(size, gfpflags);
808 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
810 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
814 * Calculate the number of objects and left-over bytes for a given buffer size.
816 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
817 size_t align, int flags, size_t *left_over,
822 size_t slab_size = PAGE_SIZE << gfporder;
825 * The slab management structure can be either off the slab or
826 * on it. For the latter case, the memory allocated for a
830 * - One kmem_bufctl_t for each object
831 * - Padding to respect alignment of @align
832 * - @buffer_size bytes for each object
834 * If the slab management structure is off the slab, then the
835 * alignment will already be calculated into the size. Because
836 * the slabs are all pages aligned, the objects will be at the
837 * correct alignment when allocated.
839 if (flags & CFLGS_OFF_SLAB) {
841 nr_objs = slab_size / buffer_size;
843 if (nr_objs > SLAB_LIMIT)
844 nr_objs = SLAB_LIMIT;
847 * Ignore padding for the initial guess. The padding
848 * is at most @align-1 bytes, and @buffer_size is at
849 * least @align. In the worst case, this result will
850 * be one greater than the number of objects that fit
851 * into the memory allocation when taking the padding
854 nr_objs = (slab_size - sizeof(struct slab)) /
855 (buffer_size + sizeof(kmem_bufctl_t));
858 * This calculated number will be either the right
859 * amount, or one greater than what we want.
861 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
865 if (nr_objs > SLAB_LIMIT)
866 nr_objs = SLAB_LIMIT;
868 mgmt_size = slab_mgmt_size(nr_objs, align);
871 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
874 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
876 static void __slab_error(const char *function, struct kmem_cache *cachep,
879 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
880 function, cachep->name, msg);
885 * By default on NUMA we use alien caches to stage the freeing of
886 * objects allocated from other nodes. This causes massive memory
887 * inefficiencies when using fake NUMA setup to split memory into a
888 * large number of small nodes, so it can be disabled on the command
892 static int use_alien_caches __read_mostly = 1;
893 static int __init noaliencache_setup(char *s)
895 use_alien_caches = 0;
898 __setup("noaliencache", noaliencache_setup);
902 * Special reaping functions for NUMA systems called from cache_reap().
903 * These take care of doing round robin flushing of alien caches (containing
904 * objects freed on different nodes from which they were allocated) and the
905 * flushing of remote pcps by calling drain_node_pages.
907 static DEFINE_PER_CPU(unsigned long, reap_node);
909 static void init_reap_node(int cpu)
913 node = next_node(cpu_to_node(cpu), node_online_map);
914 if (node == MAX_NUMNODES)
915 node = first_node(node_online_map);
917 per_cpu(reap_node, cpu) = node;
920 static void next_reap_node(void)
922 int node = __get_cpu_var(reap_node);
925 * Also drain per cpu pages on remote zones
927 if (node != numa_node_id())
928 drain_node_pages(node);
930 node = next_node(node, node_online_map);
931 if (unlikely(node >= MAX_NUMNODES))
932 node = first_node(node_online_map);
933 __get_cpu_var(reap_node) = node;
937 #define init_reap_node(cpu) do { } while (0)
938 #define next_reap_node(void) do { } while (0)
942 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
943 * via the workqueue/eventd.
944 * Add the CPU number into the expiration time to minimize the possibility of
945 * the CPUs getting into lockstep and contending for the global cache chain
948 static void __devinit start_cpu_timer(int cpu)
950 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
953 * When this gets called from do_initcalls via cpucache_init(),
954 * init_workqueues() has already run, so keventd will be setup
957 if (keventd_up() && reap_work->work.func == NULL) {
959 INIT_DELAYED_WORK(reap_work, cache_reap);
960 schedule_delayed_work_on(cpu, reap_work,
961 __round_jiffies_relative(HZ, cpu));
965 static struct array_cache *alloc_arraycache(int node, int entries,
968 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
969 struct array_cache *nc = NULL;
971 nc = kmalloc_node(memsize, GFP_KERNEL, node);
975 nc->batchcount = batchcount;
977 spin_lock_init(&nc->lock);
983 * Transfer objects in one arraycache to another.
984 * Locking must be handled by the caller.
986 * Return the number of entries transferred.
988 static int transfer_objects(struct array_cache *to,
989 struct array_cache *from, unsigned int max)
991 /* Figure out how many entries to transfer */
992 int nr = min(min(from->avail, max), to->limit - to->avail);
997 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1008 #define drain_alien_cache(cachep, alien) do { } while (0)
1009 #define reap_alien(cachep, l3) do { } while (0)
1011 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1013 return (struct array_cache **)BAD_ALIEN_MAGIC;
1016 static inline void free_alien_cache(struct array_cache **ac_ptr)
1020 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1025 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1031 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1032 gfp_t flags, int nodeid)
1037 #else /* CONFIG_NUMA */
1039 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1040 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1042 static struct array_cache **alloc_alien_cache(int node, int limit)
1044 struct array_cache **ac_ptr;
1045 int memsize = sizeof(void *) * MAX_NUMNODES;
1050 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1053 if (i == node || !node_online(i)) {
1057 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1059 for (i--; i <= 0; i--)
1069 static void free_alien_cache(struct array_cache **ac_ptr)
1080 static void __drain_alien_cache(struct kmem_cache *cachep,
1081 struct array_cache *ac, int node)
1083 struct kmem_list3 *rl3 = cachep->nodelists[node];
1086 spin_lock(&rl3->list_lock);
1088 * Stuff objects into the remote nodes shared array first.
1089 * That way we could avoid the overhead of putting the objects
1090 * into the free lists and getting them back later.
1093 transfer_objects(rl3->shared, ac, ac->limit);
1095 free_block(cachep, ac->entry, ac->avail, node);
1097 spin_unlock(&rl3->list_lock);
1102 * Called from cache_reap() to regularly drain alien caches round robin.
1104 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1106 int node = __get_cpu_var(reap_node);
1109 struct array_cache *ac = l3->alien[node];
1111 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1112 __drain_alien_cache(cachep, ac, node);
1113 spin_unlock_irq(&ac->lock);
1118 static void drain_alien_cache(struct kmem_cache *cachep,
1119 struct array_cache **alien)
1122 struct array_cache *ac;
1123 unsigned long flags;
1125 for_each_online_node(i) {
1128 spin_lock_irqsave(&ac->lock, flags);
1129 __drain_alien_cache(cachep, ac, i);
1130 spin_unlock_irqrestore(&ac->lock, flags);
1135 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1137 struct slab *slabp = virt_to_slab(objp);
1138 int nodeid = slabp->nodeid;
1139 struct kmem_list3 *l3;
1140 struct array_cache *alien = NULL;
1143 node = numa_node_id();
1146 * Make sure we are not freeing a object from another node to the array
1147 * cache on this cpu.
1149 if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches))
1152 l3 = cachep->nodelists[node];
1153 STATS_INC_NODEFREES(cachep);
1154 if (l3->alien && l3->alien[nodeid]) {
1155 alien = l3->alien[nodeid];
1156 spin_lock(&alien->lock);
1157 if (unlikely(alien->avail == alien->limit)) {
1158 STATS_INC_ACOVERFLOW(cachep);
1159 __drain_alien_cache(cachep, alien, nodeid);
1161 alien->entry[alien->avail++] = objp;
1162 spin_unlock(&alien->lock);
1164 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1165 free_block(cachep, &objp, 1, nodeid);
1166 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1172 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1173 unsigned long action, void *hcpu)
1175 long cpu = (long)hcpu;
1176 struct kmem_cache *cachep;
1177 struct kmem_list3 *l3 = NULL;
1178 int node = cpu_to_node(cpu);
1179 int memsize = sizeof(struct kmem_list3);
1182 case CPU_UP_PREPARE:
1183 mutex_lock(&cache_chain_mutex);
1185 * We need to do this right in the beginning since
1186 * alloc_arraycache's are going to use this list.
1187 * kmalloc_node allows us to add the slab to the right
1188 * kmem_list3 and not this cpu's kmem_list3
1191 list_for_each_entry(cachep, &cache_chain, next) {
1193 * Set up the size64 kmemlist for cpu before we can
1194 * begin anything. Make sure some other cpu on this
1195 * node has not already allocated this
1197 if (!cachep->nodelists[node]) {
1198 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1201 kmem_list3_init(l3);
1202 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1203 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1206 * The l3s don't come and go as CPUs come and
1207 * go. cache_chain_mutex is sufficient
1210 cachep->nodelists[node] = l3;
1213 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1214 cachep->nodelists[node]->free_limit =
1215 (1 + nr_cpus_node(node)) *
1216 cachep->batchcount + cachep->num;
1217 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1221 * Now we can go ahead with allocating the shared arrays and
1224 list_for_each_entry(cachep, &cache_chain, next) {
1225 struct array_cache *nc;
1226 struct array_cache *shared;
1227 struct array_cache **alien = NULL;
1229 nc = alloc_arraycache(node, cachep->limit,
1230 cachep->batchcount);
1233 shared = alloc_arraycache(node,
1234 cachep->shared * cachep->batchcount,
1239 if (use_alien_caches) {
1240 alien = alloc_alien_cache(node, cachep->limit);
1244 cachep->array[cpu] = nc;
1245 l3 = cachep->nodelists[node];
1248 spin_lock_irq(&l3->list_lock);
1251 * We are serialised from CPU_DEAD or
1252 * CPU_UP_CANCELLED by the cpucontrol lock
1254 l3->shared = shared;
1263 spin_unlock_irq(&l3->list_lock);
1265 free_alien_cache(alien);
1269 mutex_unlock(&cache_chain_mutex);
1270 start_cpu_timer(cpu);
1272 #ifdef CONFIG_HOTPLUG_CPU
1273 case CPU_DOWN_PREPARE:
1274 mutex_lock(&cache_chain_mutex);
1276 case CPU_DOWN_FAILED:
1277 mutex_unlock(&cache_chain_mutex);
1281 * Even if all the cpus of a node are down, we don't free the
1282 * kmem_list3 of any cache. This to avoid a race between
1283 * cpu_down, and a kmalloc allocation from another cpu for
1284 * memory from the node of the cpu going down. The list3
1285 * structure is usually allocated from kmem_cache_create() and
1286 * gets destroyed at kmem_cache_destroy().
1290 case CPU_UP_CANCELED:
1291 list_for_each_entry(cachep, &cache_chain, next) {
1292 struct array_cache *nc;
1293 struct array_cache *shared;
1294 struct array_cache **alien;
1297 mask = node_to_cpumask(node);
1298 /* cpu is dead; no one can alloc from it. */
1299 nc = cachep->array[cpu];
1300 cachep->array[cpu] = NULL;
1301 l3 = cachep->nodelists[node];
1304 goto free_array_cache;
1306 spin_lock_irq(&l3->list_lock);
1308 /* Free limit for this kmem_list3 */
1309 l3->free_limit -= cachep->batchcount;
1311 free_block(cachep, nc->entry, nc->avail, node);
1313 if (!cpus_empty(mask)) {
1314 spin_unlock_irq(&l3->list_lock);
1315 goto free_array_cache;
1318 shared = l3->shared;
1320 free_block(cachep, l3->shared->entry,
1321 l3->shared->avail, node);
1328 spin_unlock_irq(&l3->list_lock);
1332 drain_alien_cache(cachep, alien);
1333 free_alien_cache(alien);
1339 * In the previous loop, all the objects were freed to
1340 * the respective cache's slabs, now we can go ahead and
1341 * shrink each nodelist to its limit.
1343 list_for_each_entry(cachep, &cache_chain, next) {
1344 l3 = cachep->nodelists[node];
1347 drain_freelist(cachep, l3, l3->free_objects);
1349 mutex_unlock(&cache_chain_mutex);
1357 static struct notifier_block __cpuinitdata cpucache_notifier = {
1358 &cpuup_callback, NULL, 0
1362 * swap the static kmem_list3 with kmalloced memory
1364 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1367 struct kmem_list3 *ptr;
1369 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1372 local_irq_disable();
1373 memcpy(ptr, list, sizeof(struct kmem_list3));
1375 * Do not assume that spinlocks can be initialized via memcpy:
1377 spin_lock_init(&ptr->list_lock);
1379 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1380 cachep->nodelists[nodeid] = ptr;
1385 * Initialisation. Called after the page allocator have been initialised and
1386 * before smp_init().
1388 void __init kmem_cache_init(void)
1391 struct cache_sizes *sizes;
1392 struct cache_names *names;
1397 for (i = 0; i < NUM_INIT_LISTS; i++) {
1398 kmem_list3_init(&initkmem_list3[i]);
1399 if (i < MAX_NUMNODES)
1400 cache_cache.nodelists[i] = NULL;
1404 * Fragmentation resistance on low memory - only use bigger
1405 * page orders on machines with more than 32MB of memory.
1407 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1408 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1410 /* Bootstrap is tricky, because several objects are allocated
1411 * from caches that do not exist yet:
1412 * 1) initialize the cache_cache cache: it contains the struct
1413 * kmem_cache structures of all caches, except cache_cache itself:
1414 * cache_cache is statically allocated.
1415 * Initially an __init data area is used for the head array and the
1416 * kmem_list3 structures, it's replaced with a kmalloc allocated
1417 * array at the end of the bootstrap.
1418 * 2) Create the first kmalloc cache.
1419 * The struct kmem_cache for the new cache is allocated normally.
1420 * An __init data area is used for the head array.
1421 * 3) Create the remaining kmalloc caches, with minimally sized
1423 * 4) Replace the __init data head arrays for cache_cache and the first
1424 * kmalloc cache with kmalloc allocated arrays.
1425 * 5) Replace the __init data for kmem_list3 for cache_cache and
1426 * the other cache's with kmalloc allocated memory.
1427 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1430 node = numa_node_id();
1432 /* 1) create the cache_cache */
1433 INIT_LIST_HEAD(&cache_chain);
1434 list_add(&cache_cache.next, &cache_chain);
1435 cache_cache.colour_off = cache_line_size();
1436 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1437 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1439 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1441 cache_cache.reciprocal_buffer_size =
1442 reciprocal_value(cache_cache.buffer_size);
1444 for (order = 0; order < MAX_ORDER; order++) {
1445 cache_estimate(order, cache_cache.buffer_size,
1446 cache_line_size(), 0, &left_over, &cache_cache.num);
1447 if (cache_cache.num)
1450 BUG_ON(!cache_cache.num);
1451 cache_cache.gfporder = order;
1452 cache_cache.colour = left_over / cache_cache.colour_off;
1453 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1454 sizeof(struct slab), cache_line_size());
1456 /* 2+3) create the kmalloc caches */
1457 sizes = malloc_sizes;
1458 names = cache_names;
1461 * Initialize the caches that provide memory for the array cache and the
1462 * kmem_list3 structures first. Without this, further allocations will
1466 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1467 sizes[INDEX_AC].cs_size,
1468 ARCH_KMALLOC_MINALIGN,
1469 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1472 if (INDEX_AC != INDEX_L3) {
1473 sizes[INDEX_L3].cs_cachep =
1474 kmem_cache_create(names[INDEX_L3].name,
1475 sizes[INDEX_L3].cs_size,
1476 ARCH_KMALLOC_MINALIGN,
1477 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1481 slab_early_init = 0;
1483 while (sizes->cs_size != ULONG_MAX) {
1485 * For performance, all the general caches are L1 aligned.
1486 * This should be particularly beneficial on SMP boxes, as it
1487 * eliminates "false sharing".
1488 * Note for systems short on memory removing the alignment will
1489 * allow tighter packing of the smaller caches.
1491 if (!sizes->cs_cachep) {
1492 sizes->cs_cachep = kmem_cache_create(names->name,
1494 ARCH_KMALLOC_MINALIGN,
1495 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1499 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1501 ARCH_KMALLOC_MINALIGN,
1502 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1508 /* 4) Replace the bootstrap head arrays */
1510 struct array_cache *ptr;
1512 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1514 local_irq_disable();
1515 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1516 memcpy(ptr, cpu_cache_get(&cache_cache),
1517 sizeof(struct arraycache_init));
1519 * Do not assume that spinlocks can be initialized via memcpy:
1521 spin_lock_init(&ptr->lock);
1523 cache_cache.array[smp_processor_id()] = ptr;
1526 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1528 local_irq_disable();
1529 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1530 != &initarray_generic.cache);
1531 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1532 sizeof(struct arraycache_init));
1534 * Do not assume that spinlocks can be initialized via memcpy:
1536 spin_lock_init(&ptr->lock);
1538 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1542 /* 5) Replace the bootstrap kmem_list3's */
1546 /* Replace the static kmem_list3 structures for the boot cpu */
1547 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1549 for_each_online_node(nid) {
1550 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1551 &initkmem_list3[SIZE_AC + nid], nid);
1553 if (INDEX_AC != INDEX_L3) {
1554 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1555 &initkmem_list3[SIZE_L3 + nid], nid);
1560 /* 6) resize the head arrays to their final sizes */
1562 struct kmem_cache *cachep;
1563 mutex_lock(&cache_chain_mutex);
1564 list_for_each_entry(cachep, &cache_chain, next)
1565 if (enable_cpucache(cachep))
1567 mutex_unlock(&cache_chain_mutex);
1570 /* Annotate slab for lockdep -- annotate the malloc caches */
1575 g_cpucache_up = FULL;
1578 * Register a cpu startup notifier callback that initializes
1579 * cpu_cache_get for all new cpus
1581 register_cpu_notifier(&cpucache_notifier);
1584 * The reap timers are started later, with a module init call: That part
1585 * of the kernel is not yet operational.
1589 static int __init cpucache_init(void)
1594 * Register the timers that return unneeded pages to the page allocator
1596 for_each_online_cpu(cpu)
1597 start_cpu_timer(cpu);
1600 __initcall(cpucache_init);
1603 * Interface to system's page allocator. No need to hold the cache-lock.
1605 * If we requested dmaable memory, we will get it. Even if we
1606 * did not request dmaable memory, we might get it, but that
1607 * would be relatively rare and ignorable.
1609 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1617 * Nommu uses slab's for process anonymous memory allocations, and thus
1618 * requires __GFP_COMP to properly refcount higher order allocations
1620 flags |= __GFP_COMP;
1623 flags |= cachep->gfpflags;
1625 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1629 nr_pages = (1 << cachep->gfporder);
1630 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1631 add_zone_page_state(page_zone(page),
1632 NR_SLAB_RECLAIMABLE, nr_pages);
1634 add_zone_page_state(page_zone(page),
1635 NR_SLAB_UNRECLAIMABLE, nr_pages);
1636 for (i = 0; i < nr_pages; i++)
1637 __SetPageSlab(page + i);
1638 return page_address(page);
1642 * Interface to system's page release.
1644 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1646 unsigned long i = (1 << cachep->gfporder);
1647 struct page *page = virt_to_page(addr);
1648 const unsigned long nr_freed = i;
1650 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1651 sub_zone_page_state(page_zone(page),
1652 NR_SLAB_RECLAIMABLE, nr_freed);
1654 sub_zone_page_state(page_zone(page),
1655 NR_SLAB_UNRECLAIMABLE, nr_freed);
1657 BUG_ON(!PageSlab(page));
1658 __ClearPageSlab(page);
1661 if (current->reclaim_state)
1662 current->reclaim_state->reclaimed_slab += nr_freed;
1663 free_pages((unsigned long)addr, cachep->gfporder);
1666 static void kmem_rcu_free(struct rcu_head *head)
1668 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1669 struct kmem_cache *cachep = slab_rcu->cachep;
1671 kmem_freepages(cachep, slab_rcu->addr);
1672 if (OFF_SLAB(cachep))
1673 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1678 #ifdef CONFIG_DEBUG_PAGEALLOC
1679 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1680 unsigned long caller)
1682 int size = obj_size(cachep);
1684 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1686 if (size < 5 * sizeof(unsigned long))
1689 *addr++ = 0x12345678;
1691 *addr++ = smp_processor_id();
1692 size -= 3 * sizeof(unsigned long);
1694 unsigned long *sptr = &caller;
1695 unsigned long svalue;
1697 while (!kstack_end(sptr)) {
1699 if (kernel_text_address(svalue)) {
1701 size -= sizeof(unsigned long);
1702 if (size <= sizeof(unsigned long))
1708 *addr++ = 0x87654321;
1712 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1714 int size = obj_size(cachep);
1715 addr = &((char *)addr)[obj_offset(cachep)];
1717 memset(addr, val, size);
1718 *(unsigned char *)(addr + size - 1) = POISON_END;
1721 static void dump_line(char *data, int offset, int limit)
1724 unsigned char error = 0;
1727 printk(KERN_ERR "%03x:", offset);
1728 for (i = 0; i < limit; i++) {
1729 if (data[offset + i] != POISON_FREE) {
1730 error = data[offset + i];
1733 printk(" %02x", (unsigned char)data[offset + i]);
1737 if (bad_count == 1) {
1738 error ^= POISON_FREE;
1739 if (!(error & (error - 1))) {
1740 printk(KERN_ERR "Single bit error detected. Probably "
1743 printk(KERN_ERR "Run memtest86+ or a similar memory "
1746 printk(KERN_ERR "Run a memory test tool.\n");
1755 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1760 if (cachep->flags & SLAB_RED_ZONE) {
1761 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1762 *dbg_redzone1(cachep, objp),
1763 *dbg_redzone2(cachep, objp));
1766 if (cachep->flags & SLAB_STORE_USER) {
1767 printk(KERN_ERR "Last user: [<%p>]",
1768 *dbg_userword(cachep, objp));
1769 print_symbol("(%s)",
1770 (unsigned long)*dbg_userword(cachep, objp));
1773 realobj = (char *)objp + obj_offset(cachep);
1774 size = obj_size(cachep);
1775 for (i = 0; i < size && lines; i += 16, lines--) {
1778 if (i + limit > size)
1780 dump_line(realobj, i, limit);
1784 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1790 realobj = (char *)objp + obj_offset(cachep);
1791 size = obj_size(cachep);
1793 for (i = 0; i < size; i++) {
1794 char exp = POISON_FREE;
1797 if (realobj[i] != exp) {
1803 "Slab corruption: (%s) start=%p, len=%d\n",
1804 print_tainted(), realobj, size);
1805 print_objinfo(cachep, objp, 0);
1808 /* Hexdump the affected line */
1811 if (i + limit > size)
1813 dump_line(realobj, i, limit);
1816 /* Limit to 5 lines */
1822 /* Print some data about the neighboring objects, if they
1825 struct slab *slabp = virt_to_slab(objp);
1828 objnr = obj_to_index(cachep, slabp, objp);
1830 objp = index_to_obj(cachep, slabp, objnr - 1);
1831 realobj = (char *)objp + obj_offset(cachep);
1832 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1834 print_objinfo(cachep, objp, 2);
1836 if (objnr + 1 < cachep->num) {
1837 objp = index_to_obj(cachep, slabp, objnr + 1);
1838 realobj = (char *)objp + obj_offset(cachep);
1839 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1841 print_objinfo(cachep, objp, 2);
1849 * slab_destroy_objs - destroy a slab and its objects
1850 * @cachep: cache pointer being destroyed
1851 * @slabp: slab pointer being destroyed
1853 * Call the registered destructor for each object in a slab that is being
1856 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1859 for (i = 0; i < cachep->num; i++) {
1860 void *objp = index_to_obj(cachep, slabp, i);
1862 if (cachep->flags & SLAB_POISON) {
1863 #ifdef CONFIG_DEBUG_PAGEALLOC
1864 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1866 kernel_map_pages(virt_to_page(objp),
1867 cachep->buffer_size / PAGE_SIZE, 1);
1869 check_poison_obj(cachep, objp);
1871 check_poison_obj(cachep, objp);
1874 if (cachep->flags & SLAB_RED_ZONE) {
1875 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1876 slab_error(cachep, "start of a freed object "
1878 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1879 slab_error(cachep, "end of a freed object "
1882 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1883 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1887 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1891 for (i = 0; i < cachep->num; i++) {
1892 void *objp = index_to_obj(cachep, slabp, i);
1893 (cachep->dtor) (objp, cachep, 0);
1900 * slab_destroy - destroy and release all objects in a slab
1901 * @cachep: cache pointer being destroyed
1902 * @slabp: slab pointer being destroyed
1904 * Destroy all the objs in a slab, and release the mem back to the system.
1905 * Before calling the slab must have been unlinked from the cache. The
1906 * cache-lock is not held/needed.
1908 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1910 void *addr = slabp->s_mem - slabp->colouroff;
1912 slab_destroy_objs(cachep, slabp);
1913 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1914 struct slab_rcu *slab_rcu;
1916 slab_rcu = (struct slab_rcu *)slabp;
1917 slab_rcu->cachep = cachep;
1918 slab_rcu->addr = addr;
1919 call_rcu(&slab_rcu->head, kmem_rcu_free);
1921 kmem_freepages(cachep, addr);
1922 if (OFF_SLAB(cachep))
1923 kmem_cache_free(cachep->slabp_cache, slabp);
1928 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1929 * size of kmem_list3.
1931 static void set_up_list3s(struct kmem_cache *cachep, int index)
1935 for_each_online_node(node) {
1936 cachep->nodelists[node] = &initkmem_list3[index + node];
1937 cachep->nodelists[node]->next_reap = jiffies +
1939 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1943 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1946 struct kmem_list3 *l3;
1948 for_each_online_cpu(i)
1949 kfree(cachep->array[i]);
1951 /* NUMA: free the list3 structures */
1952 for_each_online_node(i) {
1953 l3 = cachep->nodelists[i];
1956 free_alien_cache(l3->alien);
1960 kmem_cache_free(&cache_cache, cachep);
1965 * calculate_slab_order - calculate size (page order) of slabs
1966 * @cachep: pointer to the cache that is being created
1967 * @size: size of objects to be created in this cache.
1968 * @align: required alignment for the objects.
1969 * @flags: slab allocation flags
1971 * Also calculates the number of objects per slab.
1973 * This could be made much more intelligent. For now, try to avoid using
1974 * high order pages for slabs. When the gfp() functions are more friendly
1975 * towards high-order requests, this should be changed.
1977 static size_t calculate_slab_order(struct kmem_cache *cachep,
1978 size_t size, size_t align, unsigned long flags)
1980 unsigned long offslab_limit;
1981 size_t left_over = 0;
1984 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1988 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1992 if (flags & CFLGS_OFF_SLAB) {
1994 * Max number of objs-per-slab for caches which
1995 * use off-slab slabs. Needed to avoid a possible
1996 * looping condition in cache_grow().
1998 offslab_limit = size - sizeof(struct slab);
1999 offslab_limit /= sizeof(kmem_bufctl_t);
2001 if (num > offslab_limit)
2005 /* Found something acceptable - save it away */
2007 cachep->gfporder = gfporder;
2008 left_over = remainder;
2011 * A VFS-reclaimable slab tends to have most allocations
2012 * as GFP_NOFS and we really don't want to have to be allocating
2013 * higher-order pages when we are unable to shrink dcache.
2015 if (flags & SLAB_RECLAIM_ACCOUNT)
2019 * Large number of objects is good, but very large slabs are
2020 * currently bad for the gfp()s.
2022 if (gfporder >= slab_break_gfp_order)
2026 * Acceptable internal fragmentation?
2028 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2034 static int setup_cpu_cache(struct kmem_cache *cachep)
2036 if (g_cpucache_up == FULL)
2037 return enable_cpucache(cachep);
2039 if (g_cpucache_up == NONE) {
2041 * Note: the first kmem_cache_create must create the cache
2042 * that's used by kmalloc(24), otherwise the creation of
2043 * further caches will BUG().
2045 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2048 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2049 * the first cache, then we need to set up all its list3s,
2050 * otherwise the creation of further caches will BUG().
2052 set_up_list3s(cachep, SIZE_AC);
2053 if (INDEX_AC == INDEX_L3)
2054 g_cpucache_up = PARTIAL_L3;
2056 g_cpucache_up = PARTIAL_AC;
2058 cachep->array[smp_processor_id()] =
2059 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2061 if (g_cpucache_up == PARTIAL_AC) {
2062 set_up_list3s(cachep, SIZE_L3);
2063 g_cpucache_up = PARTIAL_L3;
2066 for_each_online_node(node) {
2067 cachep->nodelists[node] =
2068 kmalloc_node(sizeof(struct kmem_list3),
2070 BUG_ON(!cachep->nodelists[node]);
2071 kmem_list3_init(cachep->nodelists[node]);
2075 cachep->nodelists[numa_node_id()]->next_reap =
2076 jiffies + REAPTIMEOUT_LIST3 +
2077 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2079 cpu_cache_get(cachep)->avail = 0;
2080 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2081 cpu_cache_get(cachep)->batchcount = 1;
2082 cpu_cache_get(cachep)->touched = 0;
2083 cachep->batchcount = 1;
2084 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2089 * kmem_cache_create - Create a cache.
2090 * @name: A string which is used in /proc/slabinfo to identify this cache.
2091 * @size: The size of objects to be created in this cache.
2092 * @align: The required alignment for the objects.
2093 * @flags: SLAB flags
2094 * @ctor: A constructor for the objects.
2095 * @dtor: A destructor for the objects.
2097 * Returns a ptr to the cache on success, NULL on failure.
2098 * Cannot be called within a int, but can be interrupted.
2099 * The @ctor is run when new pages are allocated by the cache
2100 * and the @dtor is run before the pages are handed back.
2102 * @name must be valid until the cache is destroyed. This implies that
2103 * the module calling this has to destroy the cache before getting unloaded.
2107 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2108 * to catch references to uninitialised memory.
2110 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2111 * for buffer overruns.
2113 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2114 * cacheline. This can be beneficial if you're counting cycles as closely
2118 kmem_cache_create (const char *name, size_t size, size_t align,
2119 unsigned long flags,
2120 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2121 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2123 size_t left_over, slab_size, ralign;
2124 struct kmem_cache *cachep = NULL, *pc;
2127 * Sanity checks... these are all serious usage bugs.
2129 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2130 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2131 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2137 * We use cache_chain_mutex to ensure a consistent view of
2138 * cpu_online_map as well. Please see cpuup_callback
2140 mutex_lock(&cache_chain_mutex);
2142 list_for_each_entry(pc, &cache_chain, next) {
2147 * This happens when the module gets unloaded and doesn't
2148 * destroy its slab cache and no-one else reuses the vmalloc
2149 * area of the module. Print a warning.
2151 res = probe_kernel_address(pc->name, tmp);
2153 printk("SLAB: cache with size %d has lost its name\n",
2158 if (!strcmp(pc->name, name)) {
2159 printk("kmem_cache_create: duplicate cache %s\n", name);
2166 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2167 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2168 /* No constructor, but inital state check requested */
2169 printk(KERN_ERR "%s: No con, but init state check "
2170 "requested - %s\n", __FUNCTION__, name);
2171 flags &= ~SLAB_DEBUG_INITIAL;
2175 * Enable redzoning and last user accounting, except for caches with
2176 * large objects, if the increased size would increase the object size
2177 * above the next power of two: caches with object sizes just above a
2178 * power of two have a significant amount of internal fragmentation.
2180 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2181 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2182 if (!(flags & SLAB_DESTROY_BY_RCU))
2183 flags |= SLAB_POISON;
2185 if (flags & SLAB_DESTROY_BY_RCU)
2186 BUG_ON(flags & SLAB_POISON);
2188 if (flags & SLAB_DESTROY_BY_RCU)
2192 * Always checks flags, a caller might be expecting debug support which
2195 BUG_ON(flags & ~CREATE_MASK);
2198 * Check that size is in terms of words. This is needed to avoid
2199 * unaligned accesses for some archs when redzoning is used, and makes
2200 * sure any on-slab bufctl's are also correctly aligned.
2202 if (size & (BYTES_PER_WORD - 1)) {
2203 size += (BYTES_PER_WORD - 1);
2204 size &= ~(BYTES_PER_WORD - 1);
2207 /* calculate the final buffer alignment: */
2209 /* 1) arch recommendation: can be overridden for debug */
2210 if (flags & SLAB_HWCACHE_ALIGN) {
2212 * Default alignment: as specified by the arch code. Except if
2213 * an object is really small, then squeeze multiple objects into
2216 ralign = cache_line_size();
2217 while (size <= ralign / 2)
2220 ralign = BYTES_PER_WORD;
2224 * Redzoning and user store require word alignment. Note this will be
2225 * overridden by architecture or caller mandated alignment if either
2226 * is greater than BYTES_PER_WORD.
2228 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2229 ralign = BYTES_PER_WORD;
2231 /* 2) arch mandated alignment */
2232 if (ralign < ARCH_SLAB_MINALIGN) {
2233 ralign = ARCH_SLAB_MINALIGN;
2235 /* 3) caller mandated alignment */
2236 if (ralign < align) {
2239 /* disable debug if necessary */
2240 if (ralign > BYTES_PER_WORD)
2241 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2247 /* Get cache's description obj. */
2248 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2253 cachep->obj_size = size;
2256 * Both debugging options require word-alignment which is calculated
2259 if (flags & SLAB_RED_ZONE) {
2260 /* add space for red zone words */
2261 cachep->obj_offset += BYTES_PER_WORD;
2262 size += 2 * BYTES_PER_WORD;
2264 if (flags & SLAB_STORE_USER) {
2265 /* user store requires one word storage behind the end of
2268 size += BYTES_PER_WORD;
2270 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2271 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2272 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2273 cachep->obj_offset += PAGE_SIZE - size;
2280 * Determine if the slab management is 'on' or 'off' slab.
2281 * (bootstrapping cannot cope with offslab caches so don't do
2284 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2286 * Size is large, assume best to place the slab management obj
2287 * off-slab (should allow better packing of objs).
2289 flags |= CFLGS_OFF_SLAB;
2291 size = ALIGN(size, align);
2293 left_over = calculate_slab_order(cachep, size, align, flags);
2296 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2297 kmem_cache_free(&cache_cache, cachep);
2301 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2302 + sizeof(struct slab), align);
2305 * If the slab has been placed off-slab, and we have enough space then
2306 * move it on-slab. This is at the expense of any extra colouring.
2308 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2309 flags &= ~CFLGS_OFF_SLAB;
2310 left_over -= slab_size;
2313 if (flags & CFLGS_OFF_SLAB) {
2314 /* really off slab. No need for manual alignment */
2316 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2319 cachep->colour_off = cache_line_size();
2320 /* Offset must be a multiple of the alignment. */
2321 if (cachep->colour_off < align)
2322 cachep->colour_off = align;
2323 cachep->colour = left_over / cachep->colour_off;
2324 cachep->slab_size = slab_size;
2325 cachep->flags = flags;
2326 cachep->gfpflags = 0;
2327 if (flags & SLAB_CACHE_DMA)
2328 cachep->gfpflags |= GFP_DMA;
2329 cachep->buffer_size = size;
2330 cachep->reciprocal_buffer_size = reciprocal_value(size);
2332 if (flags & CFLGS_OFF_SLAB) {
2333 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2335 * This is a possibility for one of the malloc_sizes caches.
2336 * But since we go off slab only for object size greater than
2337 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2338 * this should not happen at all.
2339 * But leave a BUG_ON for some lucky dude.
2341 BUG_ON(!cachep->slabp_cache);
2343 cachep->ctor = ctor;
2344 cachep->dtor = dtor;
2345 cachep->name = name;
2347 if (setup_cpu_cache(cachep)) {
2348 __kmem_cache_destroy(cachep);
2353 /* cache setup completed, link it into the list */
2354 list_add(&cachep->next, &cache_chain);
2356 if (!cachep && (flags & SLAB_PANIC))
2357 panic("kmem_cache_create(): failed to create slab `%s'\n",
2359 mutex_unlock(&cache_chain_mutex);
2362 EXPORT_SYMBOL(kmem_cache_create);
2365 static void check_irq_off(void)
2367 BUG_ON(!irqs_disabled());
2370 static void check_irq_on(void)
2372 BUG_ON(irqs_disabled());
2375 static void check_spinlock_acquired(struct kmem_cache *cachep)
2379 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2383 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2387 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2392 #define check_irq_off() do { } while(0)
2393 #define check_irq_on() do { } while(0)
2394 #define check_spinlock_acquired(x) do { } while(0)
2395 #define check_spinlock_acquired_node(x, y) do { } while(0)
2398 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2399 struct array_cache *ac,
2400 int force, int node);
2402 static void do_drain(void *arg)
2404 struct kmem_cache *cachep = arg;
2405 struct array_cache *ac;
2406 int node = numa_node_id();
2409 ac = cpu_cache_get(cachep);
2410 spin_lock(&cachep->nodelists[node]->list_lock);
2411 free_block(cachep, ac->entry, ac->avail, node);
2412 spin_unlock(&cachep->nodelists[node]->list_lock);
2416 static void drain_cpu_caches(struct kmem_cache *cachep)
2418 struct kmem_list3 *l3;
2421 on_each_cpu(do_drain, cachep, 1, 1);
2423 for_each_online_node(node) {
2424 l3 = cachep->nodelists[node];
2425 if (l3 && l3->alien)
2426 drain_alien_cache(cachep, l3->alien);
2429 for_each_online_node(node) {
2430 l3 = cachep->nodelists[node];
2432 drain_array(cachep, l3, l3->shared, 1, node);
2437 * Remove slabs from the list of free slabs.
2438 * Specify the number of slabs to drain in tofree.
2440 * Returns the actual number of slabs released.
2442 static int drain_freelist(struct kmem_cache *cache,
2443 struct kmem_list3 *l3, int tofree)
2445 struct list_head *p;
2450 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2452 spin_lock_irq(&l3->list_lock);
2453 p = l3->slabs_free.prev;
2454 if (p == &l3->slabs_free) {
2455 spin_unlock_irq(&l3->list_lock);
2459 slabp = list_entry(p, struct slab, list);
2461 BUG_ON(slabp->inuse);
2463 list_del(&slabp->list);
2465 * Safe to drop the lock. The slab is no longer linked
2468 l3->free_objects -= cache->num;
2469 spin_unlock_irq(&l3->list_lock);
2470 slab_destroy(cache, slabp);
2477 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2478 static int __cache_shrink(struct kmem_cache *cachep)
2481 struct kmem_list3 *l3;
2483 drain_cpu_caches(cachep);
2486 for_each_online_node(i) {
2487 l3 = cachep->nodelists[i];
2491 drain_freelist(cachep, l3, l3->free_objects);
2493 ret += !list_empty(&l3->slabs_full) ||
2494 !list_empty(&l3->slabs_partial);
2496 return (ret ? 1 : 0);
2500 * kmem_cache_shrink - Shrink a cache.
2501 * @cachep: The cache to shrink.
2503 * Releases as many slabs as possible for a cache.
2504 * To help debugging, a zero exit status indicates all slabs were released.
2506 int kmem_cache_shrink(struct kmem_cache *cachep)
2509 BUG_ON(!cachep || in_interrupt());
2511 mutex_lock(&cache_chain_mutex);
2512 ret = __cache_shrink(cachep);
2513 mutex_unlock(&cache_chain_mutex);
2516 EXPORT_SYMBOL(kmem_cache_shrink);
2519 * kmem_cache_destroy - delete a cache
2520 * @cachep: the cache to destroy
2522 * Remove a struct kmem_cache object from the slab cache.
2524 * It is expected this function will be called by a module when it is
2525 * unloaded. This will remove the cache completely, and avoid a duplicate
2526 * cache being allocated each time a module is loaded and unloaded, if the
2527 * module doesn't have persistent in-kernel storage across loads and unloads.
2529 * The cache must be empty before calling this function.
2531 * The caller must guarantee that noone will allocate memory from the cache
2532 * during the kmem_cache_destroy().
2534 void kmem_cache_destroy(struct kmem_cache *cachep)
2536 BUG_ON(!cachep || in_interrupt());
2538 /* Find the cache in the chain of caches. */
2539 mutex_lock(&cache_chain_mutex);
2541 * the chain is never empty, cache_cache is never destroyed
2543 list_del(&cachep->next);
2544 if (__cache_shrink(cachep)) {
2545 slab_error(cachep, "Can't free all objects");
2546 list_add(&cachep->next, &cache_chain);
2547 mutex_unlock(&cache_chain_mutex);
2551 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2554 __kmem_cache_destroy(cachep);
2555 mutex_unlock(&cache_chain_mutex);
2557 EXPORT_SYMBOL(kmem_cache_destroy);
2560 * Get the memory for a slab management obj.
2561 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2562 * always come from malloc_sizes caches. The slab descriptor cannot
2563 * come from the same cache which is getting created because,
2564 * when we are searching for an appropriate cache for these
2565 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2566 * If we are creating a malloc_sizes cache here it would not be visible to
2567 * kmem_find_general_cachep till the initialization is complete.
2568 * Hence we cannot have slabp_cache same as the original cache.
2570 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2571 int colour_off, gfp_t local_flags,
2576 if (OFF_SLAB(cachep)) {
2577 /* Slab management obj is off-slab. */
2578 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2579 local_flags & ~GFP_THISNODE, nodeid);
2583 slabp = objp + colour_off;
2584 colour_off += cachep->slab_size;
2587 slabp->colouroff = colour_off;
2588 slabp->s_mem = objp + colour_off;
2589 slabp->nodeid = nodeid;
2593 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2595 return (kmem_bufctl_t *) (slabp + 1);
2598 static void cache_init_objs(struct kmem_cache *cachep,
2599 struct slab *slabp, unsigned long ctor_flags)
2603 for (i = 0; i < cachep->num; i++) {
2604 void *objp = index_to_obj(cachep, slabp, i);
2606 /* need to poison the objs? */
2607 if (cachep->flags & SLAB_POISON)
2608 poison_obj(cachep, objp, POISON_FREE);
2609 if (cachep->flags & SLAB_STORE_USER)
2610 *dbg_userword(cachep, objp) = NULL;
2612 if (cachep->flags & SLAB_RED_ZONE) {
2613 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2614 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2617 * Constructors are not allowed to allocate memory from the same
2618 * cache which they are a constructor for. Otherwise, deadlock.
2619 * They must also be threaded.
2621 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2622 cachep->ctor(objp + obj_offset(cachep), cachep,
2625 if (cachep->flags & SLAB_RED_ZONE) {
2626 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2627 slab_error(cachep, "constructor overwrote the"
2628 " end of an object");
2629 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2630 slab_error(cachep, "constructor overwrote the"
2631 " start of an object");
2633 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2634 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2635 kernel_map_pages(virt_to_page(objp),
2636 cachep->buffer_size / PAGE_SIZE, 0);
2639 cachep->ctor(objp, cachep, ctor_flags);
2641 slab_bufctl(slabp)[i] = i + 1;
2643 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2647 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2649 if (flags & GFP_DMA)
2650 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2652 BUG_ON(cachep->gfpflags & GFP_DMA);
2655 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2658 void *objp = index_to_obj(cachep, slabp, slabp->free);
2662 next = slab_bufctl(slabp)[slabp->free];
2664 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2665 WARN_ON(slabp->nodeid != nodeid);
2672 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2673 void *objp, int nodeid)
2675 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2678 /* Verify that the slab belongs to the intended node */
2679 WARN_ON(slabp->nodeid != nodeid);
2681 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2682 printk(KERN_ERR "slab: double free detected in cache "
2683 "'%s', objp %p\n", cachep->name, objp);
2687 slab_bufctl(slabp)[objnr] = slabp->free;
2688 slabp->free = objnr;
2693 * Map pages beginning at addr to the given cache and slab. This is required
2694 * for the slab allocator to be able to lookup the cache and slab of a
2695 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2697 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2703 page = virt_to_page(addr);
2706 if (likely(!PageCompound(page)))
2707 nr_pages <<= cache->gfporder;
2710 page_set_cache(page, cache);
2711 page_set_slab(page, slab);
2713 } while (--nr_pages);
2717 * Grow (by 1) the number of slabs within a cache. This is called by
2718 * kmem_cache_alloc() when there are no active objs left in a cache.
2720 static int cache_grow(struct kmem_cache *cachep,
2721 gfp_t flags, int nodeid, void *objp)
2726 unsigned long ctor_flags;
2727 struct kmem_list3 *l3;
2730 * Be lazy and only check for valid flags here, keeping it out of the
2731 * critical path in kmem_cache_alloc().
2733 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW));
2734 if (flags & __GFP_NO_GROW)
2737 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2738 local_flags = (flags & GFP_LEVEL_MASK);
2739 if (!(local_flags & __GFP_WAIT))
2741 * Not allowed to sleep. Need to tell a constructor about
2742 * this - it might need to know...
2744 ctor_flags |= SLAB_CTOR_ATOMIC;
2746 /* Take the l3 list lock to change the colour_next on this node */
2748 l3 = cachep->nodelists[nodeid];
2749 spin_lock(&l3->list_lock);
2751 /* Get colour for the slab, and cal the next value. */
2752 offset = l3->colour_next;
2754 if (l3->colour_next >= cachep->colour)
2755 l3->colour_next = 0;
2756 spin_unlock(&l3->list_lock);
2758 offset *= cachep->colour_off;
2760 if (local_flags & __GFP_WAIT)
2764 * The test for missing atomic flag is performed here, rather than
2765 * the more obvious place, simply to reduce the critical path length
2766 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2767 * will eventually be caught here (where it matters).
2769 kmem_flagcheck(cachep, flags);
2772 * Get mem for the objs. Attempt to allocate a physical page from
2776 objp = kmem_getpages(cachep, flags, nodeid);
2780 /* Get slab management. */
2781 slabp = alloc_slabmgmt(cachep, objp, offset,
2782 local_flags & ~GFP_THISNODE, nodeid);
2786 slabp->nodeid = nodeid;
2787 slab_map_pages(cachep, slabp, objp);
2789 cache_init_objs(cachep, slabp, ctor_flags);
2791 if (local_flags & __GFP_WAIT)
2792 local_irq_disable();
2794 spin_lock(&l3->list_lock);
2796 /* Make slab active. */
2797 list_add_tail(&slabp->list, &(l3->slabs_free));
2798 STATS_INC_GROWN(cachep);
2799 l3->free_objects += cachep->num;
2800 spin_unlock(&l3->list_lock);
2803 kmem_freepages(cachep, objp);
2805 if (local_flags & __GFP_WAIT)
2806 local_irq_disable();
2813 * Perform extra freeing checks:
2814 * - detect bad pointers.
2815 * - POISON/RED_ZONE checking
2816 * - destructor calls, for caches with POISON+dtor
2818 static void kfree_debugcheck(const void *objp)
2822 if (!virt_addr_valid(objp)) {
2823 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2824 (unsigned long)objp);
2827 page = virt_to_page(objp);
2828 if (!PageSlab(page)) {
2829 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2830 (unsigned long)objp);
2835 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2837 unsigned long redzone1, redzone2;
2839 redzone1 = *dbg_redzone1(cache, obj);
2840 redzone2 = *dbg_redzone2(cache, obj);
2845 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2848 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2849 slab_error(cache, "double free detected");
2851 slab_error(cache, "memory outside object was overwritten");
2853 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2854 obj, redzone1, redzone2);
2857 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2864 objp -= obj_offset(cachep);
2865 kfree_debugcheck(objp);
2866 page = virt_to_page(objp);
2868 slabp = page_get_slab(page);
2870 if (cachep->flags & SLAB_RED_ZONE) {
2871 verify_redzone_free(cachep, objp);
2872 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2873 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2875 if (cachep->flags & SLAB_STORE_USER)
2876 *dbg_userword(cachep, objp) = caller;
2878 objnr = obj_to_index(cachep, slabp, objp);
2880 BUG_ON(objnr >= cachep->num);
2881 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2883 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2885 * Need to call the slab's constructor so the caller can
2886 * perform a verify of its state (debugging). Called without
2887 * the cache-lock held.
2889 cachep->ctor(objp + obj_offset(cachep),
2890 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2892 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2893 /* we want to cache poison the object,
2894 * call the destruction callback
2896 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2898 #ifdef CONFIG_DEBUG_SLAB_LEAK
2899 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2901 if (cachep->flags & SLAB_POISON) {
2902 #ifdef CONFIG_DEBUG_PAGEALLOC
2903 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2904 store_stackinfo(cachep, objp, (unsigned long)caller);
2905 kernel_map_pages(virt_to_page(objp),
2906 cachep->buffer_size / PAGE_SIZE, 0);
2908 poison_obj(cachep, objp, POISON_FREE);
2911 poison_obj(cachep, objp, POISON_FREE);
2917 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2922 /* Check slab's freelist to see if this obj is there. */
2923 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2925 if (entries > cachep->num || i >= cachep->num)
2928 if (entries != cachep->num - slabp->inuse) {
2930 printk(KERN_ERR "slab: Internal list corruption detected in "
2931 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2932 cachep->name, cachep->num, slabp, slabp->inuse);
2934 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2937 printk("\n%03x:", i);
2938 printk(" %02x", ((unsigned char *)slabp)[i]);
2945 #define kfree_debugcheck(x) do { } while(0)
2946 #define cache_free_debugcheck(x,objp,z) (objp)
2947 #define check_slabp(x,y) do { } while(0)
2950 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2953 struct kmem_list3 *l3;
2954 struct array_cache *ac;
2957 node = numa_node_id();
2960 ac = cpu_cache_get(cachep);
2962 batchcount = ac->batchcount;
2963 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2965 * If there was little recent activity on this cache, then
2966 * perform only a partial refill. Otherwise we could generate
2969 batchcount = BATCHREFILL_LIMIT;
2971 l3 = cachep->nodelists[node];
2973 BUG_ON(ac->avail > 0 || !l3);
2974 spin_lock(&l3->list_lock);
2976 /* See if we can refill from the shared array */
2977 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2980 while (batchcount > 0) {
2981 struct list_head *entry;
2983 /* Get slab alloc is to come from. */
2984 entry = l3->slabs_partial.next;
2985 if (entry == &l3->slabs_partial) {
2986 l3->free_touched = 1;
2987 entry = l3->slabs_free.next;
2988 if (entry == &l3->slabs_free)
2992 slabp = list_entry(entry, struct slab, list);
2993 check_slabp(cachep, slabp);
2994 check_spinlock_acquired(cachep);
2995 while (slabp->inuse < cachep->num && batchcount--) {
2996 STATS_INC_ALLOCED(cachep);
2997 STATS_INC_ACTIVE(cachep);
2998 STATS_SET_HIGH(cachep);
3000 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3003 check_slabp(cachep, slabp);
3005 /* move slabp to correct slabp list: */
3006 list_del(&slabp->list);
3007 if (slabp->free == BUFCTL_END)
3008 list_add(&slabp->list, &l3->slabs_full);
3010 list_add(&slabp->list, &l3->slabs_partial);
3014 l3->free_objects -= ac->avail;
3016 spin_unlock(&l3->list_lock);
3018 if (unlikely(!ac->avail)) {
3020 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3022 /* cache_grow can reenable interrupts, then ac could change. */
3023 ac = cpu_cache_get(cachep);
3024 if (!x && ac->avail == 0) /* no objects in sight? abort */
3027 if (!ac->avail) /* objects refilled by interrupt? */
3031 return ac->entry[--ac->avail];
3034 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3037 might_sleep_if(flags & __GFP_WAIT);
3039 kmem_flagcheck(cachep, flags);
3044 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3045 gfp_t flags, void *objp, void *caller)
3049 if (cachep->flags & SLAB_POISON) {
3050 #ifdef CONFIG_DEBUG_PAGEALLOC
3051 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3052 kernel_map_pages(virt_to_page(objp),
3053 cachep->buffer_size / PAGE_SIZE, 1);
3055 check_poison_obj(cachep, objp);
3057 check_poison_obj(cachep, objp);
3059 poison_obj(cachep, objp, POISON_INUSE);
3061 if (cachep->flags & SLAB_STORE_USER)
3062 *dbg_userword(cachep, objp) = caller;
3064 if (cachep->flags & SLAB_RED_ZONE) {
3065 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3066 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3067 slab_error(cachep, "double free, or memory outside"
3068 " object was overwritten");
3070 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3071 objp, *dbg_redzone1(cachep, objp),
3072 *dbg_redzone2(cachep, objp));
3074 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3075 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3077 #ifdef CONFIG_DEBUG_SLAB_LEAK
3082 slabp = page_get_slab(virt_to_page(objp));
3083 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3084 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3087 objp += obj_offset(cachep);
3088 if (cachep->ctor && cachep->flags & SLAB_POISON) {
3089 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
3091 if (!(flags & __GFP_WAIT))
3092 ctor_flags |= SLAB_CTOR_ATOMIC;
3094 cachep->ctor(objp, cachep, ctor_flags);
3096 #if ARCH_SLAB_MINALIGN
3097 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3098 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3099 objp, ARCH_SLAB_MINALIGN);
3105 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3108 #ifdef CONFIG_FAILSLAB
3110 static struct failslab_attr {
3112 struct fault_attr attr;
3114 u32 ignore_gfp_wait;
3115 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3116 struct dentry *ignore_gfp_wait_file;
3120 .attr = FAULT_ATTR_INITIALIZER,
3121 .ignore_gfp_wait = 1,
3124 static int __init setup_failslab(char *str)
3126 return setup_fault_attr(&failslab.attr, str);
3128 __setup("failslab=", setup_failslab);
3130 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3132 if (cachep == &cache_cache)
3134 if (flags & __GFP_NOFAIL)
3136 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3139 return should_fail(&failslab.attr, obj_size(cachep));
3142 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3144 static int __init failslab_debugfs(void)
3146 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3150 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3153 dir = failslab.attr.dentries.dir;
3155 failslab.ignore_gfp_wait_file =
3156 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3157 &failslab.ignore_gfp_wait);
3159 if (!failslab.ignore_gfp_wait_file) {
3161 debugfs_remove(failslab.ignore_gfp_wait_file);
3162 cleanup_fault_attr_dentries(&failslab.attr);
3168 late_initcall(failslab_debugfs);
3170 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3172 #else /* CONFIG_FAILSLAB */
3174 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3179 #endif /* CONFIG_FAILSLAB */
3181 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3184 struct array_cache *ac;
3188 if (should_failslab(cachep, flags))
3191 ac = cpu_cache_get(cachep);
3192 if (likely(ac->avail)) {
3193 STATS_INC_ALLOCHIT(cachep);
3195 objp = ac->entry[--ac->avail];
3197 STATS_INC_ALLOCMISS(cachep);
3198 objp = cache_alloc_refill(cachep, flags);
3203 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3204 gfp_t flags, void *caller)
3206 unsigned long save_flags;
3209 cache_alloc_debugcheck_before(cachep, flags);
3211 local_irq_save(save_flags);
3213 if (unlikely(NUMA_BUILD &&
3214 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
3215 objp = alternate_node_alloc(cachep, flags);
3218 objp = ____cache_alloc(cachep, flags);
3220 * We may just have run out of memory on the local node.
3221 * ____cache_alloc_node() knows how to locate memory on other nodes
3223 if (NUMA_BUILD && !objp)
3224 objp = ____cache_alloc_node(cachep, flags, numa_node_id());
3226 vx_slab_alloc(cachep, flags);
3227 local_irq_restore(save_flags);
3228 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3236 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3238 * If we are in_interrupt, then process context, including cpusets and
3239 * mempolicy, may not apply and should not be used for allocation policy.
3241 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3243 int nid_alloc, nid_here;
3245 if (in_interrupt() || (flags & __GFP_THISNODE))
3247 nid_alloc = nid_here = numa_node_id();
3248 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3249 nid_alloc = cpuset_mem_spread_node();
3250 else if (current->mempolicy)
3251 nid_alloc = slab_node(current->mempolicy);
3252 if (nid_alloc != nid_here)
3253 return ____cache_alloc_node(cachep, flags, nid_alloc);
3258 * Fallback function if there was no memory available and no objects on a
3259 * certain node and fall back is permitted. First we scan all the
3260 * available nodelists for available objects. If that fails then we
3261 * perform an allocation without specifying a node. This allows the page
3262 * allocator to do its reclaim / fallback magic. We then insert the
3263 * slab into the proper nodelist and then allocate from it.
3265 void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3267 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3268 ->node_zonelists[gfp_zone(flags)];
3272 gfp_t local_flags = (flags & GFP_LEVEL_MASK);
3276 * Look through allowed nodes for objects available
3277 * from existing per node queues.
3279 for (z = zonelist->zones; *z && !obj; z++) {
3280 nid = zone_to_nid(*z);
3282 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3283 cache->nodelists[nid] &&
3284 cache->nodelists[nid]->free_objects)
3285 obj = ____cache_alloc_node(cache,
3286 flags | GFP_THISNODE, nid);
3289 if (!obj && !(flags & __GFP_NO_GROW)) {
3291 * This allocation will be performed within the constraints
3292 * of the current cpuset / memory policy requirements.
3293 * We may trigger various forms of reclaim on the allowed
3294 * set and go into memory reserves if necessary.
3296 if (local_flags & __GFP_WAIT)
3298 kmem_flagcheck(cache, flags);
3299 obj = kmem_getpages(cache, flags, -1);
3300 if (local_flags & __GFP_WAIT)
3301 local_irq_disable();
3304 * Insert into the appropriate per node queues
3306 nid = page_to_nid(virt_to_page(obj));
3307 if (cache_grow(cache, flags, nid, obj)) {
3308 obj = ____cache_alloc_node(cache,
3309 flags | GFP_THISNODE, nid);
3312 * Another processor may allocate the
3313 * objects in the slab since we are
3314 * not holding any locks.
3318 /* cache_grow already freed obj */
3327 * A interface to enable slab creation on nodeid
3329 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3332 struct list_head *entry;
3334 struct kmem_list3 *l3;
3338 l3 = cachep->nodelists[nodeid];
3343 spin_lock(&l3->list_lock);
3344 entry = l3->slabs_partial.next;
3345 if (entry == &l3->slabs_partial) {
3346 l3->free_touched = 1;
3347 entry = l3->slabs_free.next;
3348 if (entry == &l3->slabs_free)
3352 slabp = list_entry(entry, struct slab, list);
3353 check_spinlock_acquired_node(cachep, nodeid);
3354 check_slabp(cachep, slabp);
3356 STATS_INC_NODEALLOCS(cachep);
3357 STATS_INC_ACTIVE(cachep);
3358 STATS_SET_HIGH(cachep);
3360 BUG_ON(slabp->inuse == cachep->num);
3362 obj = slab_get_obj(cachep, slabp, nodeid);
3363 check_slabp(cachep, slabp);
3364 vx_slab_alloc(cachep, flags);
3366 /* move slabp to correct slabp list: */
3367 list_del(&slabp->list);
3369 if (slabp->free == BUFCTL_END)
3370 list_add(&slabp->list, &l3->slabs_full);
3372 list_add(&slabp->list, &l3->slabs_partial);
3374 spin_unlock(&l3->list_lock);
3378 spin_unlock(&l3->list_lock);
3379 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3383 if (!(flags & __GFP_THISNODE))
3384 /* Unable to grow the cache. Fall back to other nodes. */
3385 return fallback_alloc(cachep, flags);
3395 * Caller needs to acquire correct kmem_list's list_lock
3397 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3401 struct kmem_list3 *l3;
3403 for (i = 0; i < nr_objects; i++) {
3404 void *objp = objpp[i];
3407 slabp = virt_to_slab(objp);
3408 l3 = cachep->nodelists[node];
3409 list_del(&slabp->list);
3410 check_spinlock_acquired_node(cachep, node);
3411 check_slabp(cachep, slabp);
3412 slab_put_obj(cachep, slabp, objp, node);
3413 STATS_DEC_ACTIVE(cachep);
3415 check_slabp(cachep, slabp);
3417 /* fixup slab chains */
3418 if (slabp->inuse == 0) {
3419 if (l3->free_objects > l3->free_limit) {
3420 l3->free_objects -= cachep->num;
3421 /* No need to drop any previously held
3422 * lock here, even if we have a off-slab slab
3423 * descriptor it is guaranteed to come from
3424 * a different cache, refer to comments before
3427 slab_destroy(cachep, slabp);
3429 list_add(&slabp->list, &l3->slabs_free);
3432 /* Unconditionally move a slab to the end of the
3433 * partial list on free - maximum time for the
3434 * other objects to be freed, too.
3436 list_add_tail(&slabp->list, &l3->slabs_partial);
3441 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3444 struct kmem_list3 *l3;
3445 int node = numa_node_id();
3447 batchcount = ac->batchcount;
3449 BUG_ON(!batchcount || batchcount > ac->avail);
3452 l3 = cachep->nodelists[node];
3453 spin_lock(&l3->list_lock);
3455 struct array_cache *shared_array = l3->shared;
3456 int max = shared_array->limit - shared_array->avail;
3458 if (batchcount > max)
3460 memcpy(&(shared_array->entry[shared_array->avail]),
3461 ac->entry, sizeof(void *) * batchcount);
3462 shared_array->avail += batchcount;
3467 free_block(cachep, ac->entry, batchcount, node);
3472 struct list_head *p;
3474 p = l3->slabs_free.next;
3475 while (p != &(l3->slabs_free)) {
3478 slabp = list_entry(p, struct slab, list);
3479 BUG_ON(slabp->inuse);
3484 STATS_SET_FREEABLE(cachep, i);
3487 spin_unlock(&l3->list_lock);
3488 ac->avail -= batchcount;
3489 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3493 * Release an obj back to its cache. If the obj has a constructed state, it must
3494 * be in this state _before_ it is released. Called with disabled ints.
3496 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3498 struct array_cache *ac = cpu_cache_get(cachep);
3501 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3502 vx_slab_free(cachep);
3504 if (cache_free_alien(cachep, objp))
3507 if (likely(ac->avail < ac->limit)) {
3508 STATS_INC_FREEHIT(cachep);
3509 ac->entry[ac->avail++] = objp;
3512 STATS_INC_FREEMISS(cachep);
3513 cache_flusharray(cachep, ac);
3514 ac->entry[ac->avail++] = objp;
3519 * kmem_cache_alloc - Allocate an object
3520 * @cachep: The cache to allocate from.
3521 * @flags: See kmalloc().
3523 * Allocate an object from this cache. The flags are only relevant
3524 * if the cache has no available objects.
3526 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3528 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3530 EXPORT_SYMBOL(kmem_cache_alloc);
3533 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3534 * @cache: The cache to allocate from.
3535 * @flags: See kmalloc().
3537 * Allocate an object from this cache and set the allocated memory to zero.
3538 * The flags are only relevant if the cache has no available objects.
3540 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3542 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3544 memset(ret, 0, obj_size(cache));
3547 EXPORT_SYMBOL(kmem_cache_zalloc);
3550 * kmem_ptr_validate - check if an untrusted pointer might
3552 * @cachep: the cache we're checking against
3553 * @ptr: pointer to validate
3555 * This verifies that the untrusted pointer looks sane:
3556 * it is _not_ a guarantee that the pointer is actually
3557 * part of the slab cache in question, but it at least
3558 * validates that the pointer can be dereferenced and
3559 * looks half-way sane.
3561 * Currently only used for dentry validation.
3563 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3565 unsigned long addr = (unsigned long)ptr;
3566 unsigned long min_addr = PAGE_OFFSET;
3567 unsigned long align_mask = BYTES_PER_WORD - 1;
3568 unsigned long size = cachep->buffer_size;
3571 if (unlikely(addr < min_addr))
3573 if (unlikely(addr > (unsigned long)high_memory - size))
3575 if (unlikely(addr & align_mask))
3577 if (unlikely(!kern_addr_valid(addr)))
3579 if (unlikely(!kern_addr_valid(addr + size - 1)))
3581 page = virt_to_page(ptr);
3582 if (unlikely(!PageSlab(page)))
3584 if (unlikely(page_get_cache(page) != cachep))
3593 * kmem_cache_alloc_node - Allocate an object on the specified node
3594 * @cachep: The cache to allocate from.
3595 * @flags: See kmalloc().
3596 * @nodeid: node number of the target node.
3597 * @caller: return address of caller, used for debug information
3599 * Identical to kmem_cache_alloc but it will allocate memory on the given
3600 * node, which can improve the performance for cpu bound structures.
3602 * Fallback to other node is possible if __GFP_THISNODE is not set.
3604 static __always_inline void *
3605 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3606 int nodeid, void *caller)
3608 unsigned long save_flags;
3611 cache_alloc_debugcheck_before(cachep, flags);
3612 local_irq_save(save_flags);
3614 if (unlikely(nodeid == -1))
3615 nodeid = numa_node_id();
3617 if (likely(cachep->nodelists[nodeid])) {
3618 if (nodeid == numa_node_id()) {
3620 * Use the locally cached objects if possible.
3621 * However ____cache_alloc does not allow fallback
3622 * to other nodes. It may fail while we still have
3623 * objects on other nodes available.
3625 ptr = ____cache_alloc(cachep, flags);
3628 /* ___cache_alloc_node can fall back to other nodes */
3629 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3632 /* Node not bootstrapped yet */
3633 if (!(flags & __GFP_THISNODE))
3634 ptr = fallback_alloc(cachep, flags);
3637 local_irq_restore(save_flags);
3638 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3643 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3645 return __cache_alloc_node(cachep, flags, nodeid,
3646 __builtin_return_address(0));
3648 EXPORT_SYMBOL(kmem_cache_alloc_node);
3650 static __always_inline void *
3651 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3653 struct kmem_cache *cachep;
3655 cachep = kmem_find_general_cachep(size, flags);
3656 if (unlikely(cachep == NULL))
3658 return kmem_cache_alloc_node(cachep, flags, node);
3661 #ifdef CONFIG_DEBUG_SLAB
3662 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3664 return __do_kmalloc_node(size, flags, node,
3665 __builtin_return_address(0));
3667 EXPORT_SYMBOL(__kmalloc_node);
3669 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3670 int node, void *caller)
3672 return __do_kmalloc_node(size, flags, node, caller);
3674 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3676 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3678 return __do_kmalloc_node(size, flags, node, NULL);
3680 EXPORT_SYMBOL(__kmalloc_node);
3681 #endif /* CONFIG_DEBUG_SLAB */
3682 #endif /* CONFIG_NUMA */
3685 * __do_kmalloc - allocate memory
3686 * @size: how many bytes of memory are required.
3687 * @flags: the type of memory to allocate (see kmalloc).
3688 * @caller: function caller for debug tracking of the caller
3690 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3693 struct kmem_cache *cachep;
3695 /* If you want to save a few bytes .text space: replace
3697 * Then kmalloc uses the uninlined functions instead of the inline
3700 cachep = __find_general_cachep(size, flags);
3701 if (unlikely(cachep == NULL))
3703 return __cache_alloc(cachep, flags, caller);
3707 #ifdef CONFIG_DEBUG_SLAB
3708 void *__kmalloc(size_t size, gfp_t flags)
3710 return __do_kmalloc(size, flags, __builtin_return_address(0));
3712 EXPORT_SYMBOL(__kmalloc);
3714 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3716 return __do_kmalloc(size, flags, caller);
3718 EXPORT_SYMBOL(__kmalloc_track_caller);
3721 void *__kmalloc(size_t size, gfp_t flags)
3723 return __do_kmalloc(size, flags, NULL);
3725 EXPORT_SYMBOL(__kmalloc);
3729 * kmem_cache_free - Deallocate an object
3730 * @cachep: The cache the allocation was from.
3731 * @objp: The previously allocated object.
3733 * Free an object which was previously allocated from this
3736 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3738 unsigned long flags;
3740 BUG_ON(virt_to_cache(objp) != cachep);
3742 local_irq_save(flags);
3743 __cache_free(cachep, objp);
3744 local_irq_restore(flags);
3746 EXPORT_SYMBOL(kmem_cache_free);
3749 * kfree - free previously allocated memory
3750 * @objp: pointer returned by kmalloc.
3752 * If @objp is NULL, no operation is performed.
3754 * Don't free memory not originally allocated by kmalloc()
3755 * or you will run into trouble.
3757 void kfree(const void *objp)
3759 struct kmem_cache *c;
3760 unsigned long flags;
3762 if (unlikely(!objp))
3764 local_irq_save(flags);
3765 kfree_debugcheck(objp);
3766 c = virt_to_cache(objp);
3767 debug_check_no_locks_freed(objp, obj_size(c));
3768 __cache_free(c, (void *)objp);
3769 local_irq_restore(flags);
3771 EXPORT_SYMBOL(kfree);
3773 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3775 return obj_size(cachep);
3777 EXPORT_SYMBOL(kmem_cache_size);
3779 const char *kmem_cache_name(struct kmem_cache *cachep)
3781 return cachep->name;
3783 EXPORT_SYMBOL_GPL(kmem_cache_name);
3786 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3788 static int alloc_kmemlist(struct kmem_cache *cachep)
3791 struct kmem_list3 *l3;
3792 struct array_cache *new_shared;
3793 struct array_cache **new_alien = NULL;
3795 for_each_online_node(node) {
3797 if (use_alien_caches) {
3798 new_alien = alloc_alien_cache(node, cachep->limit);
3803 new_shared = alloc_arraycache(node,
3804 cachep->shared*cachep->batchcount,
3807 free_alien_cache(new_alien);
3811 l3 = cachep->nodelists[node];
3813 struct array_cache *shared = l3->shared;
3815 spin_lock_irq(&l3->list_lock);
3818 free_block(cachep, shared->entry,
3819 shared->avail, node);
3821 l3->shared = new_shared;
3823 l3->alien = new_alien;
3826 l3->free_limit = (1 + nr_cpus_node(node)) *
3827 cachep->batchcount + cachep->num;
3828 spin_unlock_irq(&l3->list_lock);
3830 free_alien_cache(new_alien);
3833 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3835 free_alien_cache(new_alien);
3840 kmem_list3_init(l3);
3841 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3842 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3843 l3->shared = new_shared;
3844 l3->alien = new_alien;
3845 l3->free_limit = (1 + nr_cpus_node(node)) *
3846 cachep->batchcount + cachep->num;
3847 cachep->nodelists[node] = l3;
3852 if (!cachep->next.next) {
3853 /* Cache is not active yet. Roll back what we did */
3856 if (cachep->nodelists[node]) {
3857 l3 = cachep->nodelists[node];
3860 free_alien_cache(l3->alien);
3862 cachep->nodelists[node] = NULL;
3870 struct ccupdate_struct {
3871 struct kmem_cache *cachep;
3872 struct array_cache *new[NR_CPUS];
3875 static void do_ccupdate_local(void *info)
3877 struct ccupdate_struct *new = info;
3878 struct array_cache *old;
3881 old = cpu_cache_get(new->cachep);
3883 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3884 new->new[smp_processor_id()] = old;
3887 /* Always called with the cache_chain_mutex held */
3888 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3889 int batchcount, int shared)
3891 struct ccupdate_struct *new;
3894 new = kzalloc(sizeof(*new), GFP_KERNEL);
3898 for_each_online_cpu(i) {
3899 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3902 for (i--; i >= 0; i--)
3908 new->cachep = cachep;
3910 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3913 cachep->batchcount = batchcount;
3914 cachep->limit = limit;
3915 cachep->shared = shared;
3917 for_each_online_cpu(i) {
3918 struct array_cache *ccold = new->new[i];
3921 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3922 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3923 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3927 return alloc_kmemlist(cachep);
3930 /* Called with cache_chain_mutex held always */
3931 static int enable_cpucache(struct kmem_cache *cachep)
3937 * The head array serves three purposes:
3938 * - create a LIFO ordering, i.e. return objects that are cache-warm
3939 * - reduce the number of spinlock operations.
3940 * - reduce the number of linked list operations on the slab and
3941 * bufctl chains: array operations are cheaper.
3942 * The numbers are guessed, we should auto-tune as described by
3945 if (cachep->buffer_size > 131072)
3947 else if (cachep->buffer_size > PAGE_SIZE)
3949 else if (cachep->buffer_size > 1024)
3951 else if (cachep->buffer_size > 256)
3957 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3958 * allocation behaviour: Most allocs on one cpu, most free operations
3959 * on another cpu. For these cases, an efficient object passing between
3960 * cpus is necessary. This is provided by a shared array. The array
3961 * replaces Bonwick's magazine layer.
3962 * On uniprocessor, it's functionally equivalent (but less efficient)
3963 * to a larger limit. Thus disabled by default.
3967 if (cachep->buffer_size <= PAGE_SIZE)
3973 * With debugging enabled, large batchcount lead to excessively long
3974 * periods with disabled local interrupts. Limit the batchcount
3979 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3981 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3982 cachep->name, -err);
3987 * Drain an array if it contains any elements taking the l3 lock only if
3988 * necessary. Note that the l3 listlock also protects the array_cache
3989 * if drain_array() is used on the shared array.
3991 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3992 struct array_cache *ac, int force, int node)
3996 if (!ac || !ac->avail)
3998 if (ac->touched && !force) {
4001 spin_lock_irq(&l3->list_lock);
4003 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4004 if (tofree > ac->avail)
4005 tofree = (ac->avail + 1) / 2;
4006 free_block(cachep, ac->entry, tofree, node);
4007 ac->avail -= tofree;
4008 memmove(ac->entry, &(ac->entry[tofree]),
4009 sizeof(void *) * ac->avail);
4011 spin_unlock_irq(&l3->list_lock);
4016 * cache_reap - Reclaim memory from caches.
4017 * @unused: unused parameter
4019 * Called from workqueue/eventd every few seconds.
4021 * - clear the per-cpu caches for this CPU.
4022 * - return freeable pages to the main free memory pool.
4024 * If we cannot acquire the cache chain mutex then just give up - we'll try
4025 * again on the next iteration.
4027 static void cache_reap(struct work_struct *unused)
4029 struct kmem_cache *searchp;
4030 struct kmem_list3 *l3;
4031 int node = numa_node_id();
4033 if (!mutex_trylock(&cache_chain_mutex)) {
4034 /* Give up. Setup the next iteration. */
4035 schedule_delayed_work(&__get_cpu_var(reap_work),
4036 round_jiffies_relative(REAPTIMEOUT_CPUC));
4040 list_for_each_entry(searchp, &cache_chain, next) {
4044 * We only take the l3 lock if absolutely necessary and we
4045 * have established with reasonable certainty that
4046 * we can do some work if the lock was obtained.
4048 l3 = searchp->nodelists[node];
4050 reap_alien(searchp, l3);
4052 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4055 * These are racy checks but it does not matter
4056 * if we skip one check or scan twice.
4058 if (time_after(l3->next_reap, jiffies))
4061 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4063 drain_array(searchp, l3, l3->shared, 0, node);
4065 if (l3->free_touched)
4066 l3->free_touched = 0;
4070 freed = drain_freelist(searchp, l3, (l3->free_limit +
4071 5 * searchp->num - 1) / (5 * searchp->num));
4072 STATS_ADD_REAPED(searchp, freed);
4078 mutex_unlock(&cache_chain_mutex);
4080 refresh_cpu_vm_stats(smp_processor_id());
4081 /* Set up the next iteration */
4082 schedule_delayed_work(&__get_cpu_var(reap_work),
4083 round_jiffies_relative(REAPTIMEOUT_CPUC));
4086 #ifdef CONFIG_PROC_FS
4088 static void print_slabinfo_header(struct seq_file *m)
4091 * Output format version, so at least we can change it
4092 * without _too_ many complaints.
4095 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4097 seq_puts(m, "slabinfo - version: 2.1\n");
4099 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4100 "<objperslab> <pagesperslab>");
4101 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4102 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4104 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4105 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4106 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4111 static void *s_start(struct seq_file *m, loff_t *pos)
4114 struct list_head *p;
4116 mutex_lock(&cache_chain_mutex);
4118 print_slabinfo_header(m);
4119 p = cache_chain.next;
4122 if (p == &cache_chain)
4125 return list_entry(p, struct kmem_cache, next);
4128 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4130 struct kmem_cache *cachep = p;
4132 return cachep->next.next == &cache_chain ?
4133 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
4136 static void s_stop(struct seq_file *m, void *p)
4138 mutex_unlock(&cache_chain_mutex);
4141 static int s_show(struct seq_file *m, void *p)
4143 struct kmem_cache *cachep = p;
4145 unsigned long active_objs;
4146 unsigned long num_objs;
4147 unsigned long active_slabs = 0;
4148 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4152 struct kmem_list3 *l3;
4156 for_each_online_node(node) {
4157 l3 = cachep->nodelists[node];
4162 spin_lock_irq(&l3->list_lock);
4164 list_for_each_entry(slabp, &l3->slabs_full, list) {
4165 if (slabp->inuse != cachep->num && !error)
4166 error = "slabs_full accounting error";
4167 active_objs += cachep->num;
4170 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4171 if (slabp->inuse == cachep->num && !error)
4172 error = "slabs_partial inuse accounting error";
4173 if (!slabp->inuse && !error)
4174 error = "slabs_partial/inuse accounting error";
4175 active_objs += slabp->inuse;
4178 list_for_each_entry(slabp, &l3->slabs_free, list) {
4179 if (slabp->inuse && !error)
4180 error = "slabs_free/inuse accounting error";
4183 free_objects += l3->free_objects;
4185 shared_avail += l3->shared->avail;
4187 spin_unlock_irq(&l3->list_lock);
4189 num_slabs += active_slabs;
4190 num_objs = num_slabs * cachep->num;
4191 if (num_objs - active_objs != free_objects && !error)
4192 error = "free_objects accounting error";
4194 name = cachep->name;
4196 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4198 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4199 name, active_objs, num_objs, cachep->buffer_size,
4200 cachep->num, (1 << cachep->gfporder));
4201 seq_printf(m, " : tunables %4u %4u %4u",
4202 cachep->limit, cachep->batchcount, cachep->shared);
4203 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4204 active_slabs, num_slabs, shared_avail);
4207 unsigned long high = cachep->high_mark;
4208 unsigned long allocs = cachep->num_allocations;
4209 unsigned long grown = cachep->grown;
4210 unsigned long reaped = cachep->reaped;
4211 unsigned long errors = cachep->errors;
4212 unsigned long max_freeable = cachep->max_freeable;
4213 unsigned long node_allocs = cachep->node_allocs;
4214 unsigned long node_frees = cachep->node_frees;
4215 unsigned long overflows = cachep->node_overflow;
4217 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4218 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4219 reaped, errors, max_freeable, node_allocs,
4220 node_frees, overflows);
4224 unsigned long allochit = atomic_read(&cachep->allochit);
4225 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4226 unsigned long freehit = atomic_read(&cachep->freehit);
4227 unsigned long freemiss = atomic_read(&cachep->freemiss);
4229 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4230 allochit, allocmiss, freehit, freemiss);
4238 * slabinfo_op - iterator that generates /proc/slabinfo
4247 * num-pages-per-slab
4248 * + further values on SMP and with statistics enabled
4251 const struct seq_operations slabinfo_op = {
4258 #define MAX_SLABINFO_WRITE 128
4260 * slabinfo_write - Tuning for the slab allocator
4262 * @buffer: user buffer
4263 * @count: data length
4266 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4267 size_t count, loff_t *ppos)
4269 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4270 int limit, batchcount, shared, res;
4271 struct kmem_cache *cachep;
4273 if (count > MAX_SLABINFO_WRITE)
4275 if (copy_from_user(&kbuf, buffer, count))
4277 kbuf[MAX_SLABINFO_WRITE] = '\0';
4279 tmp = strchr(kbuf, ' ');
4284 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4287 /* Find the cache in the chain of caches. */
4288 mutex_lock(&cache_chain_mutex);
4290 list_for_each_entry(cachep, &cache_chain, next) {
4291 if (!strcmp(cachep->name, kbuf)) {
4292 if (limit < 1 || batchcount < 1 ||
4293 batchcount > limit || shared < 0) {
4296 res = do_tune_cpucache(cachep, limit,
4297 batchcount, shared);
4302 mutex_unlock(&cache_chain_mutex);
4308 #ifdef CONFIG_DEBUG_SLAB_LEAK
4310 static void *leaks_start(struct seq_file *m, loff_t *pos)
4313 struct list_head *p;
4315 mutex_lock(&cache_chain_mutex);
4316 p = cache_chain.next;
4319 if (p == &cache_chain)
4322 return list_entry(p, struct kmem_cache, next);
4325 static inline int add_caller(unsigned long *n, unsigned long v)
4335 unsigned long *q = p + 2 * i;
4349 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4355 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4361 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4362 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4364 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4369 static void show_symbol(struct seq_file *m, unsigned long address)
4371 #ifdef CONFIG_KALLSYMS
4374 unsigned long offset, size;
4375 char namebuf[KSYM_NAME_LEN+1];
4377 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4380 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4382 seq_printf(m, " [%s]", modname);
4386 seq_printf(m, "%p", (void *)address);
4389 static int leaks_show(struct seq_file *m, void *p)
4391 struct kmem_cache *cachep = p;
4393 struct kmem_list3 *l3;
4395 unsigned long *n = m->private;
4399 if (!(cachep->flags & SLAB_STORE_USER))
4401 if (!(cachep->flags & SLAB_RED_ZONE))
4404 /* OK, we can do it */
4408 for_each_online_node(node) {
4409 l3 = cachep->nodelists[node];
4414 spin_lock_irq(&l3->list_lock);
4416 list_for_each_entry(slabp, &l3->slabs_full, list)
4417 handle_slab(n, cachep, slabp);
4418 list_for_each_entry(slabp, &l3->slabs_partial, list)
4419 handle_slab(n, cachep, slabp);
4420 spin_unlock_irq(&l3->list_lock);
4422 name = cachep->name;
4424 /* Increase the buffer size */
4425 mutex_unlock(&cache_chain_mutex);
4426 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4428 /* Too bad, we are really out */
4430 mutex_lock(&cache_chain_mutex);
4433 *(unsigned long *)m->private = n[0] * 2;
4435 mutex_lock(&cache_chain_mutex);
4436 /* Now make sure this entry will be retried */
4440 for (i = 0; i < n[1]; i++) {
4441 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4442 show_symbol(m, n[2*i+2]);
4449 const struct seq_operations slabstats_op = {
4450 .start = leaks_start,
4459 * ksize - get the actual amount of memory allocated for a given object
4460 * @objp: Pointer to the object
4462 * kmalloc may internally round up allocations and return more memory
4463 * than requested. ksize() can be used to determine the actual amount of
4464 * memory allocated. The caller may use this additional memory, even though
4465 * a smaller amount of memory was initially specified with the kmalloc call.
4466 * The caller must guarantee that objp points to a valid object previously
4467 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4468 * must not be freed during the duration of the call.
4470 unsigned int ksize(const void *objp)
4472 if (unlikely(objp == NULL))
4475 return obj_size(virt_to_cache(objp));