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_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
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_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_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 */
393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder;
400 /* force GFP flags, e.g. GFP_DMA */
403 size_t colour; /* cache colouring range */
404 unsigned int colour_off; /* colour offset */
405 struct kmem_cache *slabp_cache;
406 unsigned int slab_size;
407 unsigned int dflags; /* dynamic flags */
409 /* constructor func */
410 void (*ctor) (void *, struct kmem_cache *, unsigned long);
412 /* de-constructor func */
413 void (*dtor) (void *, struct kmem_cache *, unsigned long);
415 /* 5) cache creation/removal */
417 struct list_head next;
421 unsigned long num_active;
422 unsigned long num_allocations;
423 unsigned long high_mark;
425 unsigned long reaped;
426 unsigned long errors;
427 unsigned long max_freeable;
428 unsigned long node_allocs;
429 unsigned long node_frees;
430 unsigned long node_overflow;
438 * If debugging is enabled, then the allocator can add additional
439 * fields and/or padding to every object. buffer_size contains the total
440 * object size including these internal fields, the following two
441 * variables contain the offset to the user object and its size.
447 * We put nodelists[] at the end of kmem_cache, because we want to size
448 * this array to nr_node_ids slots instead of MAX_NUMNODES
449 * (see kmem_cache_init())
450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
451 * is statically defined, so we reserve the max number of nodes.
453 struct kmem_list3 *nodelists[MAX_NUMNODES];
455 * Do not add fields after nodelists[]
459 #define CFLGS_OFF_SLAB (0x80000000UL)
460 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
462 #define BATCHREFILL_LIMIT 16
464 * Optimization question: fewer reaps means less probability for unnessary
465 * cpucache drain/refill cycles.
467 * OTOH the cpuarrays can contain lots of objects,
468 * which could lock up otherwise freeable slabs.
470 #define REAPTIMEOUT_CPUC (2*HZ)
471 #define REAPTIMEOUT_LIST3 (4*HZ)
474 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
475 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
476 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
477 #define STATS_INC_GROWN(x) ((x)->grown++)
478 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
479 #define STATS_SET_HIGH(x) \
481 if ((x)->num_active > (x)->high_mark) \
482 (x)->high_mark = (x)->num_active; \
484 #define STATS_INC_ERR(x) ((x)->errors++)
485 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
486 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
487 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
488 #define STATS_SET_FREEABLE(x, i) \
490 if ((x)->max_freeable < i) \
491 (x)->max_freeable = i; \
493 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
494 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
495 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
496 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
498 #define STATS_INC_ACTIVE(x) do { } while (0)
499 #define STATS_DEC_ACTIVE(x) do { } while (0)
500 #define STATS_INC_ALLOCED(x) do { } while (0)
501 #define STATS_INC_GROWN(x) do { } while (0)
502 #define STATS_ADD_REAPED(x,y) do { } while (0)
503 #define STATS_SET_HIGH(x) do { } while (0)
504 #define STATS_INC_ERR(x) do { } while (0)
505 #define STATS_INC_NODEALLOCS(x) do { } while (0)
506 #define STATS_INC_NODEFREES(x) do { } while (0)
507 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
508 #define STATS_SET_FREEABLE(x, i) do { } while (0)
509 #define STATS_INC_ALLOCHIT(x) do { } while (0)
510 #define STATS_INC_ALLOCMISS(x) do { } while (0)
511 #define STATS_INC_FREEHIT(x) do { } while (0)
512 #define STATS_INC_FREEMISS(x) do { } while (0)
518 * memory layout of objects:
520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
521 * the end of an object is aligned with the end of the real
522 * allocation. Catches writes behind the end of the allocation.
523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
525 * cachep->obj_offset: The real object.
526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
528 * [BYTES_PER_WORD long]
530 static int obj_offset(struct kmem_cache *cachep)
532 return cachep->obj_offset;
535 static int obj_size(struct kmem_cache *cachep)
537 return cachep->obj_size;
540 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
542 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
543 return (unsigned long long*) (objp + obj_offset(cachep) -
544 sizeof(unsigned long long));
547 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
549 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
550 if (cachep->flags & SLAB_STORE_USER)
551 return (unsigned long long *)(objp + cachep->buffer_size -
552 sizeof(unsigned long long) -
554 return (unsigned long long *) (objp + cachep->buffer_size -
555 sizeof(unsigned long long));
558 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
560 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
561 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
566 #define obj_offset(x) 0
567 #define obj_size(cachep) (cachep->buffer_size)
568 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
569 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
570 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
575 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
578 #if defined(CONFIG_LARGE_ALLOCS)
579 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
580 #define MAX_GFP_ORDER 13 /* up to 32Mb */
581 #elif defined(CONFIG_MMU)
582 #define MAX_OBJ_ORDER 5 /* 32 pages */
583 #define MAX_GFP_ORDER 5 /* 32 pages */
585 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
586 #define MAX_GFP_ORDER 8 /* up to 1Mb */
590 * Do not go above this order unless 0 objects fit into the slab.
592 #define BREAK_GFP_ORDER_HI 1
593 #define BREAK_GFP_ORDER_LO 0
594 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
597 * Functions for storing/retrieving the cachep and or slab from the page
598 * allocator. These are used to find the slab an obj belongs to. With kfree(),
599 * these are used to find the cache which an obj belongs to.
601 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
603 page->lru.next = (struct list_head *)cache;
606 static inline struct kmem_cache *page_get_cache(struct page *page)
608 page = compound_head(page);
609 BUG_ON(!PageSlab(page));
610 return (struct kmem_cache *)page->lru.next;
613 static inline void page_set_slab(struct page *page, struct slab *slab)
615 page->lru.prev = (struct list_head *)slab;
618 static inline struct slab *page_get_slab(struct page *page)
620 BUG_ON(!PageSlab(page));
621 return (struct slab *)page->lru.prev;
624 static inline struct kmem_cache *virt_to_cache(const void *obj)
626 struct page *page = virt_to_head_page(obj);
627 return page_get_cache(page);
630 static inline struct slab *virt_to_slab(const void *obj)
632 struct page *page = virt_to_head_page(obj);
633 return page_get_slab(page);
636 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
639 return slab->s_mem + cache->buffer_size * idx;
643 * We want to avoid an expensive divide : (offset / cache->buffer_size)
644 * Using the fact that buffer_size is a constant for a particular cache,
645 * we can replace (offset / cache->buffer_size) by
646 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
648 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
649 const struct slab *slab, void *obj)
651 u32 offset = (obj - slab->s_mem);
652 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
656 * These are the default caches for kmalloc. Custom caches can have other sizes.
658 struct cache_sizes malloc_sizes[] = {
659 #define CACHE(x) { .cs_size = (x) },
660 #include <linux/kmalloc_sizes.h>
664 EXPORT_SYMBOL(malloc_sizes);
666 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
672 static struct cache_names __initdata cache_names[] = {
673 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
674 #include <linux/kmalloc_sizes.h>
679 static struct arraycache_init initarray_cache __initdata =
680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
681 static struct arraycache_init initarray_generic =
682 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
684 /* internal cache of cache description objs */
685 static struct kmem_cache cache_cache = {
687 .limit = BOOT_CPUCACHE_ENTRIES,
689 .buffer_size = sizeof(struct kmem_cache),
690 .name = "kmem_cache",
693 #define BAD_ALIEN_MAGIC 0x01020304ul
695 #ifdef CONFIG_LOCKDEP
698 * Slab sometimes uses the kmalloc slabs to store the slab headers
699 * for other slabs "off slab".
700 * The locking for this is tricky in that it nests within the locks
701 * of all other slabs in a few places; to deal with this special
702 * locking we put on-slab caches into a separate lock-class.
704 * We set lock class for alien array caches which are up during init.
705 * The lock annotation will be lost if all cpus of a node goes down and
706 * then comes back up during hotplug
708 static struct lock_class_key on_slab_l3_key;
709 static struct lock_class_key on_slab_alc_key;
711 static inline void init_lock_keys(void)
715 struct cache_sizes *s = malloc_sizes;
717 while (s->cs_size != ULONG_MAX) {
719 struct array_cache **alc;
721 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
722 if (!l3 || OFF_SLAB(s->cs_cachep))
724 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
727 * FIXME: This check for BAD_ALIEN_MAGIC
728 * should go away when common slab code is taught to
729 * work even without alien caches.
730 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
731 * for alloc_alien_cache,
733 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
737 lockdep_set_class(&alc[r]->lock,
745 static inline void init_lock_keys(void)
751 * 1. Guard access to the cache-chain.
752 * 2. Protect sanity of cpu_online_map against cpu hotplug events
754 static DEFINE_MUTEX(cache_chain_mutex);
755 static struct list_head cache_chain;
758 * chicken and egg problem: delay the per-cpu array allocation
759 * until the general caches are up.
769 * used by boot code to determine if it can use slab based allocator
771 int slab_is_available(void)
773 return g_cpucache_up == FULL;
776 static DEFINE_PER_CPU(struct delayed_work, reap_work);
778 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
780 return cachep->array[smp_processor_id()];
783 static inline struct kmem_cache *__find_general_cachep(size_t size,
786 struct cache_sizes *csizep = malloc_sizes;
789 /* This happens if someone tries to call
790 * kmem_cache_create(), or __kmalloc(), before
791 * the generic caches are initialized.
793 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
795 while (size > csizep->cs_size)
799 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
800 * has cs_{dma,}cachep==NULL. Thus no special case
801 * for large kmalloc calls required.
803 #ifdef CONFIG_ZONE_DMA
804 if (unlikely(gfpflags & GFP_DMA))
805 return csizep->cs_dmacachep;
807 return csizep->cs_cachep;
810 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
812 return __find_general_cachep(size, gfpflags);
815 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
817 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
821 * Calculate the number of objects and left-over bytes for a given buffer size.
823 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
824 size_t align, int flags, size_t *left_over,
829 size_t slab_size = PAGE_SIZE << gfporder;
832 * The slab management structure can be either off the slab or
833 * on it. For the latter case, the memory allocated for a
837 * - One kmem_bufctl_t for each object
838 * - Padding to respect alignment of @align
839 * - @buffer_size bytes for each object
841 * If the slab management structure is off the slab, then the
842 * alignment will already be calculated into the size. Because
843 * the slabs are all pages aligned, the objects will be at the
844 * correct alignment when allocated.
846 if (flags & CFLGS_OFF_SLAB) {
848 nr_objs = slab_size / buffer_size;
850 if (nr_objs > SLAB_LIMIT)
851 nr_objs = SLAB_LIMIT;
854 * Ignore padding for the initial guess. The padding
855 * is at most @align-1 bytes, and @buffer_size is at
856 * least @align. In the worst case, this result will
857 * be one greater than the number of objects that fit
858 * into the memory allocation when taking the padding
861 nr_objs = (slab_size - sizeof(struct slab)) /
862 (buffer_size + sizeof(kmem_bufctl_t));
865 * This calculated number will be either the right
866 * amount, or one greater than what we want.
868 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
872 if (nr_objs > SLAB_LIMIT)
873 nr_objs = SLAB_LIMIT;
875 mgmt_size = slab_mgmt_size(nr_objs, align);
878 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
881 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
883 static void __slab_error(const char *function, struct kmem_cache *cachep,
886 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
887 function, cachep->name, msg);
892 * By default on NUMA we use alien caches to stage the freeing of
893 * objects allocated from other nodes. This causes massive memory
894 * inefficiencies when using fake NUMA setup to split memory into a
895 * large number of small nodes, so it can be disabled on the command
899 static int use_alien_caches __read_mostly = 1;
900 static int __init noaliencache_setup(char *s)
902 use_alien_caches = 0;
905 __setup("noaliencache", noaliencache_setup);
909 * Special reaping functions for NUMA systems called from cache_reap().
910 * These take care of doing round robin flushing of alien caches (containing
911 * objects freed on different nodes from which they were allocated) and the
912 * flushing of remote pcps by calling drain_node_pages.
914 static DEFINE_PER_CPU(unsigned long, reap_node);
916 static void init_reap_node(int cpu)
920 node = next_node(cpu_to_node(cpu), node_online_map);
921 if (node == MAX_NUMNODES)
922 node = first_node(node_online_map);
924 per_cpu(reap_node, cpu) = node;
927 static void next_reap_node(void)
929 int node = __get_cpu_var(reap_node);
932 * Also drain per cpu pages on remote zones
934 if (node != numa_node_id())
935 drain_node_pages(node);
937 node = next_node(node, node_online_map);
938 if (unlikely(node >= MAX_NUMNODES))
939 node = first_node(node_online_map);
940 __get_cpu_var(reap_node) = node;
944 #define init_reap_node(cpu) do { } while (0)
945 #define next_reap_node(void) do { } while (0)
949 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
950 * via the workqueue/eventd.
951 * Add the CPU number into the expiration time to minimize the possibility of
952 * the CPUs getting into lockstep and contending for the global cache chain
955 static void __devinit start_cpu_timer(int cpu)
957 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
960 * When this gets called from do_initcalls via cpucache_init(),
961 * init_workqueues() has already run, so keventd will be setup
964 if (keventd_up() && reap_work->work.func == NULL) {
966 INIT_DELAYED_WORK(reap_work, cache_reap);
967 schedule_delayed_work_on(cpu, reap_work,
968 __round_jiffies_relative(HZ, cpu));
972 static struct array_cache *alloc_arraycache(int node, int entries,
975 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
976 struct array_cache *nc = NULL;
978 nc = kmalloc_node(memsize, GFP_KERNEL, node);
982 nc->batchcount = batchcount;
984 spin_lock_init(&nc->lock);
990 * Transfer objects in one arraycache to another.
991 * Locking must be handled by the caller.
993 * Return the number of entries transferred.
995 static int transfer_objects(struct array_cache *to,
996 struct array_cache *from, unsigned int max)
998 /* Figure out how many entries to transfer */
999 int nr = min(min(from->avail, max), to->limit - to->avail);
1004 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1005 sizeof(void *) *nr);
1015 #define drain_alien_cache(cachep, alien) do { } while (0)
1016 #define reap_alien(cachep, l3) do { } while (0)
1018 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1020 return (struct array_cache **)BAD_ALIEN_MAGIC;
1023 static inline void free_alien_cache(struct array_cache **ac_ptr)
1027 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1032 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1038 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1039 gfp_t flags, int nodeid)
1044 #else /* CONFIG_NUMA */
1046 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1047 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1049 static struct array_cache **alloc_alien_cache(int node, int limit)
1051 struct array_cache **ac_ptr;
1052 int memsize = sizeof(void *) * nr_node_ids;
1057 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1060 if (i == node || !node_online(i)) {
1064 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1066 for (i--; i <= 0; i--)
1076 static void free_alien_cache(struct array_cache **ac_ptr)
1087 static void __drain_alien_cache(struct kmem_cache *cachep,
1088 struct array_cache *ac, int node)
1090 struct kmem_list3 *rl3 = cachep->nodelists[node];
1093 spin_lock(&rl3->list_lock);
1095 * Stuff objects into the remote nodes shared array first.
1096 * That way we could avoid the overhead of putting the objects
1097 * into the free lists and getting them back later.
1100 transfer_objects(rl3->shared, ac, ac->limit);
1102 free_block(cachep, ac->entry, ac->avail, node);
1104 spin_unlock(&rl3->list_lock);
1109 * Called from cache_reap() to regularly drain alien caches round robin.
1111 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1113 int node = __get_cpu_var(reap_node);
1116 struct array_cache *ac = l3->alien[node];
1118 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1119 __drain_alien_cache(cachep, ac, node);
1120 spin_unlock_irq(&ac->lock);
1125 static void drain_alien_cache(struct kmem_cache *cachep,
1126 struct array_cache **alien)
1129 struct array_cache *ac;
1130 unsigned long flags;
1132 for_each_online_node(i) {
1135 spin_lock_irqsave(&ac->lock, flags);
1136 __drain_alien_cache(cachep, ac, i);
1137 spin_unlock_irqrestore(&ac->lock, flags);
1142 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1144 struct slab *slabp = virt_to_slab(objp);
1145 int nodeid = slabp->nodeid;
1146 struct kmem_list3 *l3;
1147 struct array_cache *alien = NULL;
1150 node = numa_node_id();
1153 * Make sure we are not freeing a object from another node to the array
1154 * cache on this cpu.
1156 if (likely(slabp->nodeid == node))
1159 l3 = cachep->nodelists[node];
1160 STATS_INC_NODEFREES(cachep);
1161 if (l3->alien && l3->alien[nodeid]) {
1162 alien = l3->alien[nodeid];
1163 spin_lock(&alien->lock);
1164 if (unlikely(alien->avail == alien->limit)) {
1165 STATS_INC_ACOVERFLOW(cachep);
1166 __drain_alien_cache(cachep, alien, nodeid);
1168 alien->entry[alien->avail++] = objp;
1169 spin_unlock(&alien->lock);
1171 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1172 free_block(cachep, &objp, 1, nodeid);
1173 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1179 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1180 unsigned long action, void *hcpu)
1182 long cpu = (long)hcpu;
1183 struct kmem_cache *cachep;
1184 struct kmem_list3 *l3 = NULL;
1185 int node = cpu_to_node(cpu);
1186 int memsize = sizeof(struct kmem_list3);
1189 case CPU_LOCK_ACQUIRE:
1190 mutex_lock(&cache_chain_mutex);
1192 case CPU_UP_PREPARE:
1193 case CPU_UP_PREPARE_FROZEN:
1195 * We need to do this right in the beginning since
1196 * alloc_arraycache's are going to use this list.
1197 * kmalloc_node allows us to add the slab to the right
1198 * kmem_list3 and not this cpu's kmem_list3
1201 list_for_each_entry(cachep, &cache_chain, next) {
1203 * Set up the size64 kmemlist for cpu before we can
1204 * begin anything. Make sure some other cpu on this
1205 * node has not already allocated this
1207 if (!cachep->nodelists[node]) {
1208 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1211 kmem_list3_init(l3);
1212 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1213 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1216 * The l3s don't come and go as CPUs come and
1217 * go. cache_chain_mutex is sufficient
1220 cachep->nodelists[node] = l3;
1223 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1224 cachep->nodelists[node]->free_limit =
1225 (1 + nr_cpus_node(node)) *
1226 cachep->batchcount + cachep->num;
1227 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1231 * Now we can go ahead with allocating the shared arrays and
1234 list_for_each_entry(cachep, &cache_chain, next) {
1235 struct array_cache *nc;
1236 struct array_cache *shared = NULL;
1237 struct array_cache **alien = NULL;
1239 nc = alloc_arraycache(node, cachep->limit,
1240 cachep->batchcount);
1243 if (cachep->shared) {
1244 shared = alloc_arraycache(node,
1245 cachep->shared * cachep->batchcount,
1250 if (use_alien_caches) {
1251 alien = alloc_alien_cache(node, cachep->limit);
1255 cachep->array[cpu] = nc;
1256 l3 = cachep->nodelists[node];
1259 spin_lock_irq(&l3->list_lock);
1262 * We are serialised from CPU_DEAD or
1263 * CPU_UP_CANCELLED by the cpucontrol lock
1265 l3->shared = shared;
1274 spin_unlock_irq(&l3->list_lock);
1276 free_alien_cache(alien);
1280 case CPU_ONLINE_FROZEN:
1281 start_cpu_timer(cpu);
1283 #ifdef CONFIG_HOTPLUG_CPU
1284 case CPU_DOWN_PREPARE:
1285 case CPU_DOWN_PREPARE_FROZEN:
1287 * Shutdown cache reaper. Note that the cache_chain_mutex is
1288 * held so that if cache_reap() is invoked it cannot do
1289 * anything expensive but will only modify reap_work
1290 * and reschedule the timer.
1292 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1293 /* Now the cache_reaper is guaranteed to be not running. */
1294 per_cpu(reap_work, cpu).work.func = NULL;
1296 case CPU_DOWN_FAILED:
1297 case CPU_DOWN_FAILED_FROZEN:
1298 start_cpu_timer(cpu);
1301 case CPU_DEAD_FROZEN:
1303 * Even if all the cpus of a node are down, we don't free the
1304 * kmem_list3 of any cache. This to avoid a race between
1305 * cpu_down, and a kmalloc allocation from another cpu for
1306 * memory from the node of the cpu going down. The list3
1307 * structure is usually allocated from kmem_cache_create() and
1308 * gets destroyed at kmem_cache_destroy().
1312 case CPU_UP_CANCELED:
1313 case CPU_UP_CANCELED_FROZEN:
1314 list_for_each_entry(cachep, &cache_chain, next) {
1315 struct array_cache *nc;
1316 struct array_cache *shared;
1317 struct array_cache **alien;
1320 mask = node_to_cpumask(node);
1321 /* cpu is dead; no one can alloc from it. */
1322 nc = cachep->array[cpu];
1323 cachep->array[cpu] = NULL;
1324 l3 = cachep->nodelists[node];
1327 goto free_array_cache;
1329 spin_lock_irq(&l3->list_lock);
1331 /* Free limit for this kmem_list3 */
1332 l3->free_limit -= cachep->batchcount;
1334 free_block(cachep, nc->entry, nc->avail, node);
1336 if (!cpus_empty(mask)) {
1337 spin_unlock_irq(&l3->list_lock);
1338 goto free_array_cache;
1341 shared = l3->shared;
1343 free_block(cachep, shared->entry,
1344 shared->avail, node);
1351 spin_unlock_irq(&l3->list_lock);
1355 drain_alien_cache(cachep, alien);
1356 free_alien_cache(alien);
1362 * In the previous loop, all the objects were freed to
1363 * the respective cache's slabs, now we can go ahead and
1364 * shrink each nodelist to its limit.
1366 list_for_each_entry(cachep, &cache_chain, next) {
1367 l3 = cachep->nodelists[node];
1370 drain_freelist(cachep, l3, l3->free_objects);
1373 case CPU_LOCK_RELEASE:
1374 mutex_unlock(&cache_chain_mutex);
1382 static struct notifier_block __cpuinitdata cpucache_notifier = {
1383 &cpuup_callback, NULL, 0
1387 * swap the static kmem_list3 with kmalloced memory
1389 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1392 struct kmem_list3 *ptr;
1394 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1397 local_irq_disable();
1398 memcpy(ptr, list, sizeof(struct kmem_list3));
1400 * Do not assume that spinlocks can be initialized via memcpy:
1402 spin_lock_init(&ptr->list_lock);
1404 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1405 cachep->nodelists[nodeid] = ptr;
1410 * Initialisation. Called after the page allocator have been initialised and
1411 * before smp_init().
1413 void __init kmem_cache_init(void)
1416 struct cache_sizes *sizes;
1417 struct cache_names *names;
1422 if (num_possible_nodes() == 1)
1423 use_alien_caches = 0;
1425 for (i = 0; i < NUM_INIT_LISTS; i++) {
1426 kmem_list3_init(&initkmem_list3[i]);
1427 if (i < MAX_NUMNODES)
1428 cache_cache.nodelists[i] = NULL;
1432 * Fragmentation resistance on low memory - only use bigger
1433 * page orders on machines with more than 32MB of memory.
1435 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1436 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1438 /* Bootstrap is tricky, because several objects are allocated
1439 * from caches that do not exist yet:
1440 * 1) initialize the cache_cache cache: it contains the struct
1441 * kmem_cache structures of all caches, except cache_cache itself:
1442 * cache_cache is statically allocated.
1443 * Initially an __init data area is used for the head array and the
1444 * kmem_list3 structures, it's replaced with a kmalloc allocated
1445 * array at the end of the bootstrap.
1446 * 2) Create the first kmalloc cache.
1447 * The struct kmem_cache for the new cache is allocated normally.
1448 * An __init data area is used for the head array.
1449 * 3) Create the remaining kmalloc caches, with minimally sized
1451 * 4) Replace the __init data head arrays for cache_cache and the first
1452 * kmalloc cache with kmalloc allocated arrays.
1453 * 5) Replace the __init data for kmem_list3 for cache_cache and
1454 * the other cache's with kmalloc allocated memory.
1455 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1458 node = numa_node_id();
1460 /* 1) create the cache_cache */
1461 INIT_LIST_HEAD(&cache_chain);
1462 list_add(&cache_cache.next, &cache_chain);
1463 cache_cache.colour_off = cache_line_size();
1464 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1465 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1468 * struct kmem_cache size depends on nr_node_ids, which
1469 * can be less than MAX_NUMNODES.
1471 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1472 nr_node_ids * sizeof(struct kmem_list3 *);
1474 cache_cache.obj_size = cache_cache.buffer_size;
1476 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1478 cache_cache.reciprocal_buffer_size =
1479 reciprocal_value(cache_cache.buffer_size);
1481 for (order = 0; order < MAX_ORDER; order++) {
1482 cache_estimate(order, cache_cache.buffer_size,
1483 cache_line_size(), 0, &left_over, &cache_cache.num);
1484 if (cache_cache.num)
1487 BUG_ON(!cache_cache.num);
1488 cache_cache.gfporder = order;
1489 cache_cache.colour = left_over / cache_cache.colour_off;
1490 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1491 sizeof(struct slab), cache_line_size());
1493 /* 2+3) create the kmalloc caches */
1494 sizes = malloc_sizes;
1495 names = cache_names;
1498 * Initialize the caches that provide memory for the array cache and the
1499 * kmem_list3 structures first. Without this, further allocations will
1503 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1504 sizes[INDEX_AC].cs_size,
1505 ARCH_KMALLOC_MINALIGN,
1506 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1509 if (INDEX_AC != INDEX_L3) {
1510 sizes[INDEX_L3].cs_cachep =
1511 kmem_cache_create(names[INDEX_L3].name,
1512 sizes[INDEX_L3].cs_size,
1513 ARCH_KMALLOC_MINALIGN,
1514 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1518 slab_early_init = 0;
1520 while (sizes->cs_size != ULONG_MAX) {
1522 * For performance, all the general caches are L1 aligned.
1523 * This should be particularly beneficial on SMP boxes, as it
1524 * eliminates "false sharing".
1525 * Note for systems short on memory removing the alignment will
1526 * allow tighter packing of the smaller caches.
1528 if (!sizes->cs_cachep) {
1529 sizes->cs_cachep = kmem_cache_create(names->name,
1531 ARCH_KMALLOC_MINALIGN,
1532 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1535 #ifdef CONFIG_ZONE_DMA
1536 sizes->cs_dmacachep = kmem_cache_create(
1539 ARCH_KMALLOC_MINALIGN,
1540 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1547 /* 4) Replace the bootstrap head arrays */
1549 struct array_cache *ptr;
1551 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1553 local_irq_disable();
1554 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1555 memcpy(ptr, cpu_cache_get(&cache_cache),
1556 sizeof(struct arraycache_init));
1558 * Do not assume that spinlocks can be initialized via memcpy:
1560 spin_lock_init(&ptr->lock);
1562 cache_cache.array[smp_processor_id()] = ptr;
1565 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1567 local_irq_disable();
1568 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1569 != &initarray_generic.cache);
1570 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1571 sizeof(struct arraycache_init));
1573 * Do not assume that spinlocks can be initialized via memcpy:
1575 spin_lock_init(&ptr->lock);
1577 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1581 /* 5) Replace the bootstrap kmem_list3's */
1585 /* Replace the static kmem_list3 structures for the boot cpu */
1586 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1588 for_each_online_node(nid) {
1589 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1590 &initkmem_list3[SIZE_AC + nid], nid);
1592 if (INDEX_AC != INDEX_L3) {
1593 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1594 &initkmem_list3[SIZE_L3 + nid], nid);
1599 /* 6) resize the head arrays to their final sizes */
1601 struct kmem_cache *cachep;
1602 mutex_lock(&cache_chain_mutex);
1603 list_for_each_entry(cachep, &cache_chain, next)
1604 if (enable_cpucache(cachep))
1606 mutex_unlock(&cache_chain_mutex);
1609 /* Annotate slab for lockdep -- annotate the malloc caches */
1614 g_cpucache_up = FULL;
1617 * Register a cpu startup notifier callback that initializes
1618 * cpu_cache_get for all new cpus
1620 register_cpu_notifier(&cpucache_notifier);
1623 * The reap timers are started later, with a module init call: That part
1624 * of the kernel is not yet operational.
1628 static int __init cpucache_init(void)
1633 * Register the timers that return unneeded pages to the page allocator
1635 for_each_online_cpu(cpu)
1636 start_cpu_timer(cpu);
1639 __initcall(cpucache_init);
1642 * Interface to system's page allocator. No need to hold the cache-lock.
1644 * If we requested dmaable memory, we will get it. Even if we
1645 * did not request dmaable memory, we might get it, but that
1646 * would be relatively rare and ignorable.
1648 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1656 * Nommu uses slab's for process anonymous memory allocations, and thus
1657 * requires __GFP_COMP to properly refcount higher order allocations
1659 flags |= __GFP_COMP;
1662 flags |= cachep->gfpflags;
1664 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1668 nr_pages = (1 << cachep->gfporder);
1669 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1670 add_zone_page_state(page_zone(page),
1671 NR_SLAB_RECLAIMABLE, nr_pages);
1673 add_zone_page_state(page_zone(page),
1674 NR_SLAB_UNRECLAIMABLE, nr_pages);
1675 for (i = 0; i < nr_pages; i++)
1676 __SetPageSlab(page + i);
1677 return page_address(page);
1681 * Interface to system's page release.
1683 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1685 unsigned long i = (1 << cachep->gfporder);
1686 struct page *page = virt_to_page(addr);
1687 const unsigned long nr_freed = i;
1689 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1690 sub_zone_page_state(page_zone(page),
1691 NR_SLAB_RECLAIMABLE, nr_freed);
1693 sub_zone_page_state(page_zone(page),
1694 NR_SLAB_UNRECLAIMABLE, nr_freed);
1696 BUG_ON(!PageSlab(page));
1697 __ClearPageSlab(page);
1700 if (current->reclaim_state)
1701 current->reclaim_state->reclaimed_slab += nr_freed;
1702 free_pages((unsigned long)addr, cachep->gfporder);
1705 static void kmem_rcu_free(struct rcu_head *head)
1707 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1708 struct kmem_cache *cachep = slab_rcu->cachep;
1710 kmem_freepages(cachep, slab_rcu->addr);
1711 if (OFF_SLAB(cachep))
1712 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1717 #ifdef CONFIG_DEBUG_PAGEALLOC
1718 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1719 unsigned long caller)
1721 int size = obj_size(cachep);
1723 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1725 if (size < 5 * sizeof(unsigned long))
1728 *addr++ = 0x12345678;
1730 *addr++ = smp_processor_id();
1731 size -= 3 * sizeof(unsigned long);
1733 unsigned long *sptr = &caller;
1734 unsigned long svalue;
1736 while (!kstack_end(sptr)) {
1738 if (kernel_text_address(svalue)) {
1740 size -= sizeof(unsigned long);
1741 if (size <= sizeof(unsigned long))
1747 *addr++ = 0x87654321;
1751 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1753 int size = obj_size(cachep);
1754 addr = &((char *)addr)[obj_offset(cachep)];
1756 memset(addr, val, size);
1757 *(unsigned char *)(addr + size - 1) = POISON_END;
1760 static void dump_line(char *data, int offset, int limit)
1763 unsigned char error = 0;
1766 printk(KERN_ERR "%03x:", offset);
1767 for (i = 0; i < limit; i++) {
1768 if (data[offset + i] != POISON_FREE) {
1769 error = data[offset + i];
1772 printk(" %02x", (unsigned char)data[offset + i]);
1776 if (bad_count == 1) {
1777 error ^= POISON_FREE;
1778 if (!(error & (error - 1))) {
1779 printk(KERN_ERR "Single bit error detected. Probably "
1782 printk(KERN_ERR "Run memtest86+ or a similar memory "
1785 printk(KERN_ERR "Run a memory test tool.\n");
1794 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1799 if (cachep->flags & SLAB_RED_ZONE) {
1800 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1801 *dbg_redzone1(cachep, objp),
1802 *dbg_redzone2(cachep, objp));
1805 if (cachep->flags & SLAB_STORE_USER) {
1806 printk(KERN_ERR "Last user: [<%p>]",
1807 *dbg_userword(cachep, objp));
1808 print_symbol("(%s)",
1809 (unsigned long)*dbg_userword(cachep, objp));
1812 realobj = (char *)objp + obj_offset(cachep);
1813 size = obj_size(cachep);
1814 for (i = 0; i < size && lines; i += 16, lines--) {
1817 if (i + limit > size)
1819 dump_line(realobj, i, limit);
1823 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1829 realobj = (char *)objp + obj_offset(cachep);
1830 size = obj_size(cachep);
1832 for (i = 0; i < size; i++) {
1833 char exp = POISON_FREE;
1836 if (realobj[i] != exp) {
1842 "Slab corruption: %s start=%p, len=%d\n",
1843 cachep->name, realobj, size);
1844 print_objinfo(cachep, objp, 0);
1846 /* Hexdump the affected line */
1849 if (i + limit > size)
1851 dump_line(realobj, i, limit);
1854 /* Limit to 5 lines */
1860 /* Print some data about the neighboring objects, if they
1863 struct slab *slabp = virt_to_slab(objp);
1866 objnr = obj_to_index(cachep, slabp, objp);
1868 objp = index_to_obj(cachep, slabp, objnr - 1);
1869 realobj = (char *)objp + obj_offset(cachep);
1870 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1872 print_objinfo(cachep, objp, 2);
1874 if (objnr + 1 < cachep->num) {
1875 objp = index_to_obj(cachep, slabp, objnr + 1);
1876 realobj = (char *)objp + obj_offset(cachep);
1877 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1879 print_objinfo(cachep, objp, 2);
1887 * slab_destroy_objs - destroy a slab and its objects
1888 * @cachep: cache pointer being destroyed
1889 * @slabp: slab pointer being destroyed
1891 * Call the registered destructor for each object in a slab that is being
1894 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1897 for (i = 0; i < cachep->num; i++) {
1898 void *objp = index_to_obj(cachep, slabp, i);
1900 if (cachep->flags & SLAB_POISON) {
1901 #ifdef CONFIG_DEBUG_PAGEALLOC
1902 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1904 kernel_map_pages(virt_to_page(objp),
1905 cachep->buffer_size / PAGE_SIZE, 1);
1907 check_poison_obj(cachep, objp);
1909 check_poison_obj(cachep, objp);
1912 if (cachep->flags & SLAB_RED_ZONE) {
1913 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1914 slab_error(cachep, "start of a freed object "
1916 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1917 slab_error(cachep, "end of a freed object "
1920 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1921 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1925 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1929 for (i = 0; i < cachep->num; i++) {
1930 void *objp = index_to_obj(cachep, slabp, i);
1931 (cachep->dtor) (objp, cachep, 0);
1938 * slab_destroy - destroy and release all objects in a slab
1939 * @cachep: cache pointer being destroyed
1940 * @slabp: slab pointer being destroyed
1942 * Destroy all the objs in a slab, and release the mem back to the system.
1943 * Before calling the slab must have been unlinked from the cache. The
1944 * cache-lock is not held/needed.
1946 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1948 void *addr = slabp->s_mem - slabp->colouroff;
1950 slab_destroy_objs(cachep, slabp);
1951 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1952 struct slab_rcu *slab_rcu;
1954 slab_rcu = (struct slab_rcu *)slabp;
1955 slab_rcu->cachep = cachep;
1956 slab_rcu->addr = addr;
1957 call_rcu(&slab_rcu->head, kmem_rcu_free);
1959 kmem_freepages(cachep, addr);
1960 if (OFF_SLAB(cachep))
1961 kmem_cache_free(cachep->slabp_cache, slabp);
1966 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1967 * size of kmem_list3.
1969 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1973 for_each_online_node(node) {
1974 cachep->nodelists[node] = &initkmem_list3[index + node];
1975 cachep->nodelists[node]->next_reap = jiffies +
1977 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1981 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1984 struct kmem_list3 *l3;
1986 for_each_online_cpu(i)
1987 kfree(cachep->array[i]);
1989 /* NUMA: free the list3 structures */
1990 for_each_online_node(i) {
1991 l3 = cachep->nodelists[i];
1994 free_alien_cache(l3->alien);
1998 kmem_cache_free(&cache_cache, cachep);
2003 * calculate_slab_order - calculate size (page order) of slabs
2004 * @cachep: pointer to the cache that is being created
2005 * @size: size of objects to be created in this cache.
2006 * @align: required alignment for the objects.
2007 * @flags: slab allocation flags
2009 * Also calculates the number of objects per slab.
2011 * This could be made much more intelligent. For now, try to avoid using
2012 * high order pages for slabs. When the gfp() functions are more friendly
2013 * towards high-order requests, this should be changed.
2015 static size_t calculate_slab_order(struct kmem_cache *cachep,
2016 size_t size, size_t align, unsigned long flags)
2018 unsigned long offslab_limit;
2019 size_t left_over = 0;
2022 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
2026 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2030 if (flags & CFLGS_OFF_SLAB) {
2032 * Max number of objs-per-slab for caches which
2033 * use off-slab slabs. Needed to avoid a possible
2034 * looping condition in cache_grow().
2036 offslab_limit = size - sizeof(struct slab);
2037 offslab_limit /= sizeof(kmem_bufctl_t);
2039 if (num > offslab_limit)
2043 /* Found something acceptable - save it away */
2045 cachep->gfporder = gfporder;
2046 left_over = remainder;
2049 * A VFS-reclaimable slab tends to have most allocations
2050 * as GFP_NOFS and we really don't want to have to be allocating
2051 * higher-order pages when we are unable to shrink dcache.
2053 if (flags & SLAB_RECLAIM_ACCOUNT)
2057 * Large number of objects is good, but very large slabs are
2058 * currently bad for the gfp()s.
2060 if (gfporder >= slab_break_gfp_order)
2064 * Acceptable internal fragmentation?
2066 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2072 static int setup_cpu_cache(struct kmem_cache *cachep)
2074 if (g_cpucache_up == FULL)
2075 return enable_cpucache(cachep);
2077 if (g_cpucache_up == NONE) {
2079 * Note: the first kmem_cache_create must create the cache
2080 * that's used by kmalloc(24), otherwise the creation of
2081 * further caches will BUG().
2083 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2086 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2087 * the first cache, then we need to set up all its list3s,
2088 * otherwise the creation of further caches will BUG().
2090 set_up_list3s(cachep, SIZE_AC);
2091 if (INDEX_AC == INDEX_L3)
2092 g_cpucache_up = PARTIAL_L3;
2094 g_cpucache_up = PARTIAL_AC;
2096 cachep->array[smp_processor_id()] =
2097 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2099 if (g_cpucache_up == PARTIAL_AC) {
2100 set_up_list3s(cachep, SIZE_L3);
2101 g_cpucache_up = PARTIAL_L3;
2104 for_each_online_node(node) {
2105 cachep->nodelists[node] =
2106 kmalloc_node(sizeof(struct kmem_list3),
2108 BUG_ON(!cachep->nodelists[node]);
2109 kmem_list3_init(cachep->nodelists[node]);
2113 cachep->nodelists[numa_node_id()]->next_reap =
2114 jiffies + REAPTIMEOUT_LIST3 +
2115 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2117 cpu_cache_get(cachep)->avail = 0;
2118 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2119 cpu_cache_get(cachep)->batchcount = 1;
2120 cpu_cache_get(cachep)->touched = 0;
2121 cachep->batchcount = 1;
2122 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2127 * kmem_cache_create - Create a cache.
2128 * @name: A string which is used in /proc/slabinfo to identify this cache.
2129 * @size: The size of objects to be created in this cache.
2130 * @align: The required alignment for the objects.
2131 * @flags: SLAB flags
2132 * @ctor: A constructor for the objects.
2133 * @dtor: A destructor for the objects.
2135 * Returns a ptr to the cache on success, NULL on failure.
2136 * Cannot be called within a int, but can be interrupted.
2137 * The @ctor is run when new pages are allocated by the cache
2138 * and the @dtor is run before the pages are handed back.
2140 * @name must be valid until the cache is destroyed. This implies that
2141 * the module calling this has to destroy the cache before getting unloaded.
2145 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2146 * to catch references to uninitialised memory.
2148 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2149 * for buffer overruns.
2151 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2152 * cacheline. This can be beneficial if you're counting cycles as closely
2156 kmem_cache_create (const char *name, size_t size, size_t align,
2157 unsigned long flags,
2158 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2159 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2161 size_t left_over, slab_size, ralign;
2162 struct kmem_cache *cachep = NULL, *pc;
2165 * Sanity checks... these are all serious usage bugs.
2167 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2168 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2169 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2175 * We use cache_chain_mutex to ensure a consistent view of
2176 * cpu_online_map as well. Please see cpuup_callback
2178 mutex_lock(&cache_chain_mutex);
2180 list_for_each_entry(pc, &cache_chain, next) {
2185 * This happens when the module gets unloaded and doesn't
2186 * destroy its slab cache and no-one else reuses the vmalloc
2187 * area of the module. Print a warning.
2189 res = probe_kernel_address(pc->name, tmp);
2192 "SLAB: cache with size %d has lost its name\n",
2197 if (!strcmp(pc->name, name)) {
2199 "kmem_cache_create: duplicate cache %s\n", name);
2206 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2209 * Enable redzoning and last user accounting, except for caches with
2210 * large objects, if the increased size would increase the object size
2211 * above the next power of two: caches with object sizes just above a
2212 * power of two have a significant amount of internal fragmentation.
2214 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2215 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2216 if (!(flags & SLAB_DESTROY_BY_RCU))
2217 flags |= SLAB_POISON;
2219 if (flags & SLAB_DESTROY_BY_RCU)
2220 BUG_ON(flags & SLAB_POISON);
2222 if (flags & SLAB_DESTROY_BY_RCU)
2226 * Always checks flags, a caller might be expecting debug support which
2229 BUG_ON(flags & ~CREATE_MASK);
2232 * Check that size is in terms of words. This is needed to avoid
2233 * unaligned accesses for some archs when redzoning is used, and makes
2234 * sure any on-slab bufctl's are also correctly aligned.
2236 if (size & (BYTES_PER_WORD - 1)) {
2237 size += (BYTES_PER_WORD - 1);
2238 size &= ~(BYTES_PER_WORD - 1);
2241 /* calculate the final buffer alignment: */
2243 /* 1) arch recommendation: can be overridden for debug */
2244 if (flags & SLAB_HWCACHE_ALIGN) {
2246 * Default alignment: as specified by the arch code. Except if
2247 * an object is really small, then squeeze multiple objects into
2250 ralign = cache_line_size();
2251 while (size <= ralign / 2)
2254 ralign = BYTES_PER_WORD;
2258 * Redzoning and user store require word alignment. Note this will be
2259 * overridden by architecture or caller mandated alignment if either
2260 * is greater than BYTES_PER_WORD.
2262 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2263 ralign = __alignof__(unsigned long long);
2265 /* 2) arch mandated alignment */
2266 if (ralign < ARCH_SLAB_MINALIGN) {
2267 ralign = ARCH_SLAB_MINALIGN;
2269 /* 3) caller mandated alignment */
2270 if (ralign < align) {
2273 /* disable debug if necessary */
2274 if (ralign > __alignof__(unsigned long long))
2275 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2281 /* Get cache's description obj. */
2282 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2287 cachep->obj_size = size;
2290 * Both debugging options require word-alignment which is calculated
2293 if (flags & SLAB_RED_ZONE) {
2294 /* add space for red zone words */
2295 cachep->obj_offset += sizeof(unsigned long long);
2296 size += 2 * sizeof(unsigned long long);
2298 if (flags & SLAB_STORE_USER) {
2299 /* user store requires one word storage behind the end of
2302 size += BYTES_PER_WORD;
2304 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2305 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2306 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2307 cachep->obj_offset += PAGE_SIZE - size;
2314 * Determine if the slab management is 'on' or 'off' slab.
2315 * (bootstrapping cannot cope with offslab caches so don't do
2318 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2320 * Size is large, assume best to place the slab management obj
2321 * off-slab (should allow better packing of objs).
2323 flags |= CFLGS_OFF_SLAB;
2325 size = ALIGN(size, align);
2327 left_over = calculate_slab_order(cachep, size, align, flags);
2331 "kmem_cache_create: couldn't create cache %s.\n", name);
2332 kmem_cache_free(&cache_cache, cachep);
2336 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2337 + sizeof(struct slab), align);
2340 * If the slab has been placed off-slab, and we have enough space then
2341 * move it on-slab. This is at the expense of any extra colouring.
2343 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2344 flags &= ~CFLGS_OFF_SLAB;
2345 left_over -= slab_size;
2348 if (flags & CFLGS_OFF_SLAB) {
2349 /* really off slab. No need for manual alignment */
2351 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2354 cachep->colour_off = cache_line_size();
2355 /* Offset must be a multiple of the alignment. */
2356 if (cachep->colour_off < align)
2357 cachep->colour_off = align;
2358 cachep->colour = left_over / cachep->colour_off;
2359 cachep->slab_size = slab_size;
2360 cachep->flags = flags;
2361 cachep->gfpflags = 0;
2362 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2363 cachep->gfpflags |= GFP_DMA;
2364 cachep->buffer_size = size;
2365 cachep->reciprocal_buffer_size = reciprocal_value(size);
2367 if (flags & CFLGS_OFF_SLAB) {
2368 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2370 * This is a possibility for one of the malloc_sizes caches.
2371 * But since we go off slab only for object size greater than
2372 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2373 * this should not happen at all.
2374 * But leave a BUG_ON for some lucky dude.
2376 BUG_ON(!cachep->slabp_cache);
2378 cachep->ctor = ctor;
2379 cachep->dtor = dtor;
2380 cachep->name = name;
2382 if (setup_cpu_cache(cachep)) {
2383 __kmem_cache_destroy(cachep);
2388 /* cache setup completed, link it into the list */
2389 list_add(&cachep->next, &cache_chain);
2391 if (!cachep && (flags & SLAB_PANIC))
2392 panic("kmem_cache_create(): failed to create slab `%s'\n",
2394 mutex_unlock(&cache_chain_mutex);
2397 EXPORT_SYMBOL(kmem_cache_create);
2400 static void check_irq_off(void)
2402 BUG_ON(!irqs_disabled());
2405 static void check_irq_on(void)
2407 BUG_ON(irqs_disabled());
2410 static void check_spinlock_acquired(struct kmem_cache *cachep)
2414 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2418 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2422 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2427 #define check_irq_off() do { } while(0)
2428 #define check_irq_on() do { } while(0)
2429 #define check_spinlock_acquired(x) do { } while(0)
2430 #define check_spinlock_acquired_node(x, y) do { } while(0)
2433 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2434 struct array_cache *ac,
2435 int force, int node);
2437 static void do_drain(void *arg)
2439 struct kmem_cache *cachep = arg;
2440 struct array_cache *ac;
2441 int node = numa_node_id();
2444 ac = cpu_cache_get(cachep);
2445 spin_lock(&cachep->nodelists[node]->list_lock);
2446 free_block(cachep, ac->entry, ac->avail, node);
2447 spin_unlock(&cachep->nodelists[node]->list_lock);
2451 static void drain_cpu_caches(struct kmem_cache *cachep)
2453 struct kmem_list3 *l3;
2456 on_each_cpu(do_drain, cachep, 1, 1);
2458 for_each_online_node(node) {
2459 l3 = cachep->nodelists[node];
2460 if (l3 && l3->alien)
2461 drain_alien_cache(cachep, l3->alien);
2464 for_each_online_node(node) {
2465 l3 = cachep->nodelists[node];
2467 drain_array(cachep, l3, l3->shared, 1, node);
2472 * Remove slabs from the list of free slabs.
2473 * Specify the number of slabs to drain in tofree.
2475 * Returns the actual number of slabs released.
2477 static int drain_freelist(struct kmem_cache *cache,
2478 struct kmem_list3 *l3, int tofree)
2480 struct list_head *p;
2485 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2487 spin_lock_irq(&l3->list_lock);
2488 p = l3->slabs_free.prev;
2489 if (p == &l3->slabs_free) {
2490 spin_unlock_irq(&l3->list_lock);
2494 slabp = list_entry(p, struct slab, list);
2496 BUG_ON(slabp->inuse);
2498 list_del(&slabp->list);
2500 * Safe to drop the lock. The slab is no longer linked
2503 l3->free_objects -= cache->num;
2504 spin_unlock_irq(&l3->list_lock);
2505 slab_destroy(cache, slabp);
2512 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2513 static int __cache_shrink(struct kmem_cache *cachep)
2516 struct kmem_list3 *l3;
2518 drain_cpu_caches(cachep);
2521 for_each_online_node(i) {
2522 l3 = cachep->nodelists[i];
2526 drain_freelist(cachep, l3, l3->free_objects);
2528 ret += !list_empty(&l3->slabs_full) ||
2529 !list_empty(&l3->slabs_partial);
2531 return (ret ? 1 : 0);
2535 * kmem_cache_shrink - Shrink a cache.
2536 * @cachep: The cache to shrink.
2538 * Releases as many slabs as possible for a cache.
2539 * To help debugging, a zero exit status indicates all slabs were released.
2541 int kmem_cache_shrink(struct kmem_cache *cachep)
2544 BUG_ON(!cachep || in_interrupt());
2546 mutex_lock(&cache_chain_mutex);
2547 ret = __cache_shrink(cachep);
2548 mutex_unlock(&cache_chain_mutex);
2551 EXPORT_SYMBOL(kmem_cache_shrink);
2554 * kmem_cache_destroy - delete a cache
2555 * @cachep: the cache to destroy
2557 * Remove a &struct kmem_cache object from the slab cache.
2559 * It is expected this function will be called by a module when it is
2560 * unloaded. This will remove the cache completely, and avoid a duplicate
2561 * cache being allocated each time a module is loaded and unloaded, if the
2562 * module doesn't have persistent in-kernel storage across loads and unloads.
2564 * The cache must be empty before calling this function.
2566 * The caller must guarantee that noone will allocate memory from the cache
2567 * during the kmem_cache_destroy().
2569 void kmem_cache_destroy(struct kmem_cache *cachep)
2571 BUG_ON(!cachep || in_interrupt());
2573 /* Find the cache in the chain of caches. */
2574 mutex_lock(&cache_chain_mutex);
2576 * the chain is never empty, cache_cache is never destroyed
2578 list_del(&cachep->next);
2579 if (__cache_shrink(cachep)) {
2580 slab_error(cachep, "Can't free all objects");
2581 list_add(&cachep->next, &cache_chain);
2582 mutex_unlock(&cache_chain_mutex);
2586 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2589 __kmem_cache_destroy(cachep);
2590 mutex_unlock(&cache_chain_mutex);
2592 EXPORT_SYMBOL(kmem_cache_destroy);
2595 * Get the memory for a slab management obj.
2596 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2597 * always come from malloc_sizes caches. The slab descriptor cannot
2598 * come from the same cache which is getting created because,
2599 * when we are searching for an appropriate cache for these
2600 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2601 * If we are creating a malloc_sizes cache here it would not be visible to
2602 * kmem_find_general_cachep till the initialization is complete.
2603 * Hence we cannot have slabp_cache same as the original cache.
2605 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2606 int colour_off, gfp_t local_flags,
2611 if (OFF_SLAB(cachep)) {
2612 /* Slab management obj is off-slab. */
2613 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2614 local_flags & ~GFP_THISNODE, nodeid);
2618 slabp = objp + colour_off;
2619 colour_off += cachep->slab_size;
2622 slabp->colouroff = colour_off;
2623 slabp->s_mem = objp + colour_off;
2624 slabp->nodeid = nodeid;
2628 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2630 return (kmem_bufctl_t *) (slabp + 1);
2633 static void cache_init_objs(struct kmem_cache *cachep,
2634 struct slab *slabp, unsigned long ctor_flags)
2638 for (i = 0; i < cachep->num; i++) {
2639 void *objp = index_to_obj(cachep, slabp, i);
2641 /* need to poison the objs? */
2642 if (cachep->flags & SLAB_POISON)
2643 poison_obj(cachep, objp, POISON_FREE);
2644 if (cachep->flags & SLAB_STORE_USER)
2645 *dbg_userword(cachep, objp) = NULL;
2647 if (cachep->flags & SLAB_RED_ZONE) {
2648 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2649 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2652 * Constructors are not allowed to allocate memory from the same
2653 * cache which they are a constructor for. Otherwise, deadlock.
2654 * They must also be threaded.
2656 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2657 cachep->ctor(objp + obj_offset(cachep), cachep,
2660 if (cachep->flags & SLAB_RED_ZONE) {
2661 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2662 slab_error(cachep, "constructor overwrote the"
2663 " end of an object");
2664 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2665 slab_error(cachep, "constructor overwrote the"
2666 " start of an object");
2668 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2669 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2670 kernel_map_pages(virt_to_page(objp),
2671 cachep->buffer_size / PAGE_SIZE, 0);
2674 cachep->ctor(objp, cachep, ctor_flags);
2676 slab_bufctl(slabp)[i] = i + 1;
2678 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2682 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2684 if (CONFIG_ZONE_DMA_FLAG) {
2685 if (flags & GFP_DMA)
2686 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2688 BUG_ON(cachep->gfpflags & GFP_DMA);
2692 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2695 void *objp = index_to_obj(cachep, slabp, slabp->free);
2699 next = slab_bufctl(slabp)[slabp->free];
2701 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2702 WARN_ON(slabp->nodeid != nodeid);
2709 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2710 void *objp, int nodeid)
2712 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2715 /* Verify that the slab belongs to the intended node */
2716 WARN_ON(slabp->nodeid != nodeid);
2718 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2719 printk(KERN_ERR "slab: double free detected in cache "
2720 "'%s', objp %p\n", cachep->name, objp);
2724 slab_bufctl(slabp)[objnr] = slabp->free;
2725 slabp->free = objnr;
2730 * Map pages beginning at addr to the given cache and slab. This is required
2731 * for the slab allocator to be able to lookup the cache and slab of a
2732 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2734 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2740 page = virt_to_page(addr);
2743 if (likely(!PageCompound(page)))
2744 nr_pages <<= cache->gfporder;
2747 page_set_cache(page, cache);
2748 page_set_slab(page, slab);
2750 } while (--nr_pages);
2754 * Grow (by 1) the number of slabs within a cache. This is called by
2755 * kmem_cache_alloc() when there are no active objs left in a cache.
2757 static int cache_grow(struct kmem_cache *cachep,
2758 gfp_t flags, int nodeid, void *objp)
2763 unsigned long ctor_flags;
2764 struct kmem_list3 *l3;
2767 * Be lazy and only check for valid flags here, keeping it out of the
2768 * critical path in kmem_cache_alloc().
2770 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
2772 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2773 local_flags = (flags & GFP_LEVEL_MASK);
2774 /* Take the l3 list lock to change the colour_next on this node */
2776 l3 = cachep->nodelists[nodeid];
2777 spin_lock(&l3->list_lock);
2779 /* Get colour for the slab, and cal the next value. */
2780 offset = l3->colour_next;
2782 if (l3->colour_next >= cachep->colour)
2783 l3->colour_next = 0;
2784 spin_unlock(&l3->list_lock);
2786 offset *= cachep->colour_off;
2788 if (local_flags & __GFP_WAIT)
2792 * The test for missing atomic flag is performed here, rather than
2793 * the more obvious place, simply to reduce the critical path length
2794 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2795 * will eventually be caught here (where it matters).
2797 kmem_flagcheck(cachep, flags);
2800 * Get mem for the objs. Attempt to allocate a physical page from
2804 objp = kmem_getpages(cachep, flags, nodeid);
2808 /* Get slab management. */
2809 slabp = alloc_slabmgmt(cachep, objp, offset,
2810 local_flags & ~GFP_THISNODE, nodeid);
2814 slabp->nodeid = nodeid;
2815 slab_map_pages(cachep, slabp, objp);
2817 cache_init_objs(cachep, slabp, ctor_flags);
2819 if (local_flags & __GFP_WAIT)
2820 local_irq_disable();
2822 spin_lock(&l3->list_lock);
2824 /* Make slab active. */
2825 list_add_tail(&slabp->list, &(l3->slabs_free));
2826 STATS_INC_GROWN(cachep);
2827 l3->free_objects += cachep->num;
2828 spin_unlock(&l3->list_lock);
2831 kmem_freepages(cachep, objp);
2833 if (local_flags & __GFP_WAIT)
2834 local_irq_disable();
2841 * Perform extra freeing checks:
2842 * - detect bad pointers.
2843 * - POISON/RED_ZONE checking
2844 * - destructor calls, for caches with POISON+dtor
2846 static void kfree_debugcheck(const void *objp)
2848 if (!virt_addr_valid(objp)) {
2849 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2850 (unsigned long)objp);
2855 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2857 unsigned long long redzone1, redzone2;
2859 redzone1 = *dbg_redzone1(cache, obj);
2860 redzone2 = *dbg_redzone2(cache, obj);
2865 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2868 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2869 slab_error(cache, "double free detected");
2871 slab_error(cache, "memory outside object was overwritten");
2873 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2874 obj, redzone1, redzone2);
2877 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2884 objp -= obj_offset(cachep);
2885 kfree_debugcheck(objp);
2886 page = virt_to_head_page(objp);
2888 slabp = page_get_slab(page);
2890 if (cachep->flags & SLAB_RED_ZONE) {
2891 verify_redzone_free(cachep, objp);
2892 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2893 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2895 if (cachep->flags & SLAB_STORE_USER)
2896 *dbg_userword(cachep, objp) = caller;
2898 objnr = obj_to_index(cachep, slabp, objp);
2900 BUG_ON(objnr >= cachep->num);
2901 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2903 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2904 /* we want to cache poison the object,
2905 * call the destruction callback
2907 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2909 #ifdef CONFIG_DEBUG_SLAB_LEAK
2910 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2912 if (cachep->flags & SLAB_POISON) {
2913 #ifdef CONFIG_DEBUG_PAGEALLOC
2914 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2915 store_stackinfo(cachep, objp, (unsigned long)caller);
2916 kernel_map_pages(virt_to_page(objp),
2917 cachep->buffer_size / PAGE_SIZE, 0);
2919 poison_obj(cachep, objp, POISON_FREE);
2922 poison_obj(cachep, objp, POISON_FREE);
2928 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2933 /* Check slab's freelist to see if this obj is there. */
2934 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2936 if (entries > cachep->num || i >= cachep->num)
2939 if (entries != cachep->num - slabp->inuse) {
2941 printk(KERN_ERR "slab: Internal list corruption detected in "
2942 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2943 cachep->name, cachep->num, slabp, slabp->inuse);
2945 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2948 printk("\n%03x:", i);
2949 printk(" %02x", ((unsigned char *)slabp)[i]);
2956 #define kfree_debugcheck(x) do { } while(0)
2957 #define cache_free_debugcheck(x,objp,z) (objp)
2958 #define check_slabp(x,y) do { } while(0)
2961 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2964 struct kmem_list3 *l3;
2965 struct array_cache *ac;
2968 node = numa_node_id();
2971 ac = cpu_cache_get(cachep);
2973 batchcount = ac->batchcount;
2974 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2976 * If there was little recent activity on this cache, then
2977 * perform only a partial refill. Otherwise we could generate
2980 batchcount = BATCHREFILL_LIMIT;
2982 l3 = cachep->nodelists[node];
2984 BUG_ON(ac->avail > 0 || !l3);
2985 spin_lock(&l3->list_lock);
2987 /* See if we can refill from the shared array */
2988 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2991 while (batchcount > 0) {
2992 struct list_head *entry;
2994 /* Get slab alloc is to come from. */
2995 entry = l3->slabs_partial.next;
2996 if (entry == &l3->slabs_partial) {
2997 l3->free_touched = 1;
2998 entry = l3->slabs_free.next;
2999 if (entry == &l3->slabs_free)
3003 slabp = list_entry(entry, struct slab, list);
3004 check_slabp(cachep, slabp);
3005 check_spinlock_acquired(cachep);
3008 * The slab was either on partial or free list so
3009 * there must be at least one object available for
3012 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3014 while (slabp->inuse < cachep->num && batchcount--) {
3015 STATS_INC_ALLOCED(cachep);
3016 STATS_INC_ACTIVE(cachep);
3017 STATS_SET_HIGH(cachep);
3019 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3022 check_slabp(cachep, slabp);
3024 /* move slabp to correct slabp list: */
3025 list_del(&slabp->list);
3026 if (slabp->free == BUFCTL_END)
3027 list_add(&slabp->list, &l3->slabs_full);
3029 list_add(&slabp->list, &l3->slabs_partial);
3033 l3->free_objects -= ac->avail;
3035 spin_unlock(&l3->list_lock);
3037 if (unlikely(!ac->avail)) {
3039 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3041 /* cache_grow can reenable interrupts, then ac could change. */
3042 ac = cpu_cache_get(cachep);
3043 if (!x && ac->avail == 0) /* no objects in sight? abort */
3046 if (!ac->avail) /* objects refilled by interrupt? */
3050 return ac->entry[--ac->avail];
3053 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3056 might_sleep_if(flags & __GFP_WAIT);
3058 kmem_flagcheck(cachep, flags);
3063 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3064 gfp_t flags, void *objp, void *caller)
3068 if (cachep->flags & SLAB_POISON) {
3069 #ifdef CONFIG_DEBUG_PAGEALLOC
3070 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3071 kernel_map_pages(virt_to_page(objp),
3072 cachep->buffer_size / PAGE_SIZE, 1);
3074 check_poison_obj(cachep, objp);
3076 check_poison_obj(cachep, objp);
3078 poison_obj(cachep, objp, POISON_INUSE);
3080 if (cachep->flags & SLAB_STORE_USER)
3081 *dbg_userword(cachep, objp) = caller;
3083 if (cachep->flags & SLAB_RED_ZONE) {
3084 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3085 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3086 slab_error(cachep, "double free, or memory outside"
3087 " object was overwritten");
3089 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3090 objp, *dbg_redzone1(cachep, objp),
3091 *dbg_redzone2(cachep, objp));
3093 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3094 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3096 #ifdef CONFIG_DEBUG_SLAB_LEAK
3101 slabp = page_get_slab(virt_to_head_page(objp));
3102 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3103 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3106 objp += obj_offset(cachep);
3107 if (cachep->ctor && cachep->flags & SLAB_POISON)
3108 cachep->ctor(objp, cachep, SLAB_CTOR_CONSTRUCTOR);
3109 #if ARCH_SLAB_MINALIGN
3110 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3111 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3112 objp, ARCH_SLAB_MINALIGN);
3118 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3121 #ifdef CONFIG_FAILSLAB
3123 static struct failslab_attr {
3125 struct fault_attr attr;
3127 u32 ignore_gfp_wait;
3128 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3129 struct dentry *ignore_gfp_wait_file;
3133 .attr = FAULT_ATTR_INITIALIZER,
3134 .ignore_gfp_wait = 1,
3137 static int __init setup_failslab(char *str)
3139 return setup_fault_attr(&failslab.attr, str);
3141 __setup("failslab=", setup_failslab);
3143 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3145 if (cachep == &cache_cache)
3147 if (flags & __GFP_NOFAIL)
3149 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3152 return should_fail(&failslab.attr, obj_size(cachep));
3155 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3157 static int __init failslab_debugfs(void)
3159 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3163 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3166 dir = failslab.attr.dentries.dir;
3168 failslab.ignore_gfp_wait_file =
3169 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3170 &failslab.ignore_gfp_wait);
3172 if (!failslab.ignore_gfp_wait_file) {
3174 debugfs_remove(failslab.ignore_gfp_wait_file);
3175 cleanup_fault_attr_dentries(&failslab.attr);
3181 late_initcall(failslab_debugfs);
3183 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3185 #else /* CONFIG_FAILSLAB */
3187 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3192 #endif /* CONFIG_FAILSLAB */
3194 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3197 struct array_cache *ac;
3201 ac = cpu_cache_get(cachep);
3202 if (likely(ac->avail)) {
3203 STATS_INC_ALLOCHIT(cachep);
3205 objp = ac->entry[--ac->avail];
3207 STATS_INC_ALLOCMISS(cachep);
3208 objp = cache_alloc_refill(cachep, flags);
3215 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3217 * If we are in_interrupt, then process context, including cpusets and
3218 * mempolicy, may not apply and should not be used for allocation policy.
3220 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3222 int nid_alloc, nid_here;
3224 if (in_interrupt() || (flags & __GFP_THISNODE))
3226 nid_alloc = nid_here = numa_node_id();
3227 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3228 nid_alloc = cpuset_mem_spread_node();
3229 else if (current->mempolicy)
3230 nid_alloc = slab_node(current->mempolicy);
3231 if (nid_alloc != nid_here)
3232 return ____cache_alloc_node(cachep, flags, nid_alloc);
3237 * Fallback function if there was no memory available and no objects on a
3238 * certain node and fall back is permitted. First we scan all the
3239 * available nodelists for available objects. If that fails then we
3240 * perform an allocation without specifying a node. This allows the page
3241 * allocator to do its reclaim / fallback magic. We then insert the
3242 * slab into the proper nodelist and then allocate from it.
3244 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3246 struct zonelist *zonelist;
3252 if (flags & __GFP_THISNODE)
3255 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3256 ->node_zonelists[gfp_zone(flags)];
3257 local_flags = (flags & GFP_LEVEL_MASK);
3261 * Look through allowed nodes for objects available
3262 * from existing per node queues.
3264 for (z = zonelist->zones; *z && !obj; z++) {
3265 nid = zone_to_nid(*z);
3267 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3268 cache->nodelists[nid] &&
3269 cache->nodelists[nid]->free_objects)
3270 obj = ____cache_alloc_node(cache,
3271 flags | GFP_THISNODE, nid);
3276 * This allocation will be performed within the constraints
3277 * of the current cpuset / memory policy requirements.
3278 * We may trigger various forms of reclaim on the allowed
3279 * set and go into memory reserves if necessary.
3281 if (local_flags & __GFP_WAIT)
3283 kmem_flagcheck(cache, flags);
3284 obj = kmem_getpages(cache, flags, -1);
3285 if (local_flags & __GFP_WAIT)
3286 local_irq_disable();
3289 * Insert into the appropriate per node queues
3291 nid = page_to_nid(virt_to_page(obj));
3292 if (cache_grow(cache, flags, nid, obj)) {
3293 obj = ____cache_alloc_node(cache,
3294 flags | GFP_THISNODE, nid);
3297 * Another processor may allocate the
3298 * objects in the slab since we are
3299 * not holding any locks.
3303 /* cache_grow already freed obj */
3312 * A interface to enable slab creation on nodeid
3314 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3317 struct list_head *entry;
3319 struct kmem_list3 *l3;
3323 l3 = cachep->nodelists[nodeid];
3328 spin_lock(&l3->list_lock);
3329 entry = l3->slabs_partial.next;
3330 if (entry == &l3->slabs_partial) {
3331 l3->free_touched = 1;
3332 entry = l3->slabs_free.next;
3333 if (entry == &l3->slabs_free)
3337 slabp = list_entry(entry, struct slab, list);
3338 check_spinlock_acquired_node(cachep, nodeid);
3339 check_slabp(cachep, slabp);
3341 STATS_INC_NODEALLOCS(cachep);
3342 STATS_INC_ACTIVE(cachep);
3343 STATS_SET_HIGH(cachep);
3345 BUG_ON(slabp->inuse == cachep->num);
3347 obj = slab_get_obj(cachep, slabp, nodeid);
3348 check_slabp(cachep, slabp);
3350 /* move slabp to correct slabp list: */
3351 list_del(&slabp->list);
3353 if (slabp->free == BUFCTL_END)
3354 list_add(&slabp->list, &l3->slabs_full);
3356 list_add(&slabp->list, &l3->slabs_partial);
3358 spin_unlock(&l3->list_lock);
3362 spin_unlock(&l3->list_lock);
3363 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3367 return fallback_alloc(cachep, flags);
3374 * kmem_cache_alloc_node - Allocate an object on the specified node
3375 * @cachep: The cache to allocate from.
3376 * @flags: See kmalloc().
3377 * @nodeid: node number of the target node.
3378 * @caller: return address of caller, used for debug information
3380 * Identical to kmem_cache_alloc but it will allocate memory on the given
3381 * node, which can improve the performance for cpu bound structures.
3383 * Fallback to other node is possible if __GFP_THISNODE is not set.
3385 static __always_inline void *
3386 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3389 unsigned long save_flags;
3392 if (should_failslab(cachep, flags))
3395 cache_alloc_debugcheck_before(cachep, flags);
3396 local_irq_save(save_flags);
3398 if (unlikely(nodeid == -1))
3399 nodeid = numa_node_id();
3401 if (unlikely(!cachep->nodelists[nodeid])) {
3402 /* Node not bootstrapped yet */
3403 ptr = fallback_alloc(cachep, flags);
3407 if (nodeid == numa_node_id()) {
3409 * Use the locally cached objects if possible.
3410 * However ____cache_alloc does not allow fallback
3411 * to other nodes. It may fail while we still have
3412 * objects on other nodes available.
3414 ptr = ____cache_alloc(cachep, flags);
3418 /* ___cache_alloc_node can fall back to other nodes */
3419 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3421 local_irq_restore(save_flags);
3422 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3427 static __always_inline void *
3428 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3432 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3433 objp = alternate_node_alloc(cache, flags);
3437 objp = ____cache_alloc(cache, flags);
3440 * We may just have run out of memory on the local node.
3441 * ____cache_alloc_node() knows how to locate memory on other nodes
3444 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3451 static __always_inline void *
3452 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3454 return ____cache_alloc(cachep, flags);
3457 #endif /* CONFIG_NUMA */
3459 static __always_inline void *
3460 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3462 unsigned long save_flags;
3465 if (should_failslab(cachep, flags))
3468 cache_alloc_debugcheck_before(cachep, flags);
3469 local_irq_save(save_flags);
3470 objp = __do_cache_alloc(cachep, flags);
3471 local_irq_restore(save_flags);
3472 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3479 * Caller needs to acquire correct kmem_list's list_lock
3481 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3485 struct kmem_list3 *l3;
3487 for (i = 0; i < nr_objects; i++) {
3488 void *objp = objpp[i];
3491 slabp = virt_to_slab(objp);
3492 l3 = cachep->nodelists[node];
3493 list_del(&slabp->list);
3494 check_spinlock_acquired_node(cachep, node);
3495 check_slabp(cachep, slabp);
3496 slab_put_obj(cachep, slabp, objp, node);
3497 STATS_DEC_ACTIVE(cachep);
3499 check_slabp(cachep, slabp);
3501 /* fixup slab chains */
3502 if (slabp->inuse == 0) {
3503 if (l3->free_objects > l3->free_limit) {
3504 l3->free_objects -= cachep->num;
3505 /* No need to drop any previously held
3506 * lock here, even if we have a off-slab slab
3507 * descriptor it is guaranteed to come from
3508 * a different cache, refer to comments before
3511 slab_destroy(cachep, slabp);
3513 list_add(&slabp->list, &l3->slabs_free);
3516 /* Unconditionally move a slab to the end of the
3517 * partial list on free - maximum time for the
3518 * other objects to be freed, too.
3520 list_add_tail(&slabp->list, &l3->slabs_partial);
3525 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3528 struct kmem_list3 *l3;
3529 int node = numa_node_id();
3531 batchcount = ac->batchcount;
3533 BUG_ON(!batchcount || batchcount > ac->avail);
3536 l3 = cachep->nodelists[node];
3537 spin_lock(&l3->list_lock);
3539 struct array_cache *shared_array = l3->shared;
3540 int max = shared_array->limit - shared_array->avail;
3542 if (batchcount > max)
3544 memcpy(&(shared_array->entry[shared_array->avail]),
3545 ac->entry, sizeof(void *) * batchcount);
3546 shared_array->avail += batchcount;
3551 free_block(cachep, ac->entry, batchcount, node);
3556 struct list_head *p;
3558 p = l3->slabs_free.next;
3559 while (p != &(l3->slabs_free)) {
3562 slabp = list_entry(p, struct slab, list);
3563 BUG_ON(slabp->inuse);
3568 STATS_SET_FREEABLE(cachep, i);
3571 spin_unlock(&l3->list_lock);
3572 ac->avail -= batchcount;
3573 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3577 * Release an obj back to its cache. If the obj has a constructed state, it must
3578 * be in this state _before_ it is released. Called with disabled ints.
3580 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3582 struct array_cache *ac = cpu_cache_get(cachep);
3585 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3587 if (use_alien_caches && cache_free_alien(cachep, objp))
3590 if (likely(ac->avail < ac->limit)) {
3591 STATS_INC_FREEHIT(cachep);
3592 ac->entry[ac->avail++] = objp;
3595 STATS_INC_FREEMISS(cachep);
3596 cache_flusharray(cachep, ac);
3597 ac->entry[ac->avail++] = objp;
3602 * kmem_cache_alloc - Allocate an object
3603 * @cachep: The cache to allocate from.
3604 * @flags: See kmalloc().
3606 * Allocate an object from this cache. The flags are only relevant
3607 * if the cache has no available objects.
3609 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3611 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3613 EXPORT_SYMBOL(kmem_cache_alloc);
3616 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3617 * @cache: The cache to allocate from.
3618 * @flags: See kmalloc().
3620 * Allocate an object from this cache and set the allocated memory to zero.
3621 * The flags are only relevant if the cache has no available objects.
3623 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3625 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3627 memset(ret, 0, obj_size(cache));
3630 EXPORT_SYMBOL(kmem_cache_zalloc);
3633 * kmem_ptr_validate - check if an untrusted pointer might
3635 * @cachep: the cache we're checking against
3636 * @ptr: pointer to validate
3638 * This verifies that the untrusted pointer looks sane:
3639 * it is _not_ a guarantee that the pointer is actually
3640 * part of the slab cache in question, but it at least
3641 * validates that the pointer can be dereferenced and
3642 * looks half-way sane.
3644 * Currently only used for dentry validation.
3646 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3648 unsigned long addr = (unsigned long)ptr;
3649 unsigned long min_addr = PAGE_OFFSET;
3650 unsigned long align_mask = BYTES_PER_WORD - 1;
3651 unsigned long size = cachep->buffer_size;
3654 if (unlikely(addr < min_addr))
3656 if (unlikely(addr > (unsigned long)high_memory - size))
3658 if (unlikely(addr & align_mask))
3660 if (unlikely(!kern_addr_valid(addr)))
3662 if (unlikely(!kern_addr_valid(addr + size - 1)))
3664 page = virt_to_page(ptr);
3665 if (unlikely(!PageSlab(page)))
3667 if (unlikely(page_get_cache(page) != cachep))
3675 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3677 return __cache_alloc_node(cachep, flags, nodeid,
3678 __builtin_return_address(0));
3680 EXPORT_SYMBOL(kmem_cache_alloc_node);
3682 static __always_inline void *
3683 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3685 struct kmem_cache *cachep;
3687 cachep = kmem_find_general_cachep(size, flags);
3688 if (unlikely(cachep == NULL))
3690 return kmem_cache_alloc_node(cachep, flags, node);
3693 #ifdef CONFIG_DEBUG_SLAB
3694 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3696 return __do_kmalloc_node(size, flags, node,
3697 __builtin_return_address(0));
3699 EXPORT_SYMBOL(__kmalloc_node);
3701 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3702 int node, void *caller)
3704 return __do_kmalloc_node(size, flags, node, caller);
3706 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3708 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3710 return __do_kmalloc_node(size, flags, node, NULL);
3712 EXPORT_SYMBOL(__kmalloc_node);
3713 #endif /* CONFIG_DEBUG_SLAB */
3714 #endif /* CONFIG_NUMA */
3717 * __do_kmalloc - allocate memory
3718 * @size: how many bytes of memory are required.
3719 * @flags: the type of memory to allocate (see kmalloc).
3720 * @caller: function caller for debug tracking of the caller
3722 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3725 struct kmem_cache *cachep;
3727 /* If you want to save a few bytes .text space: replace
3729 * Then kmalloc uses the uninlined functions instead of the inline
3732 cachep = __find_general_cachep(size, flags);
3733 if (unlikely(cachep == NULL))
3735 return __cache_alloc(cachep, flags, caller);
3739 #ifdef CONFIG_DEBUG_SLAB
3740 void *__kmalloc(size_t size, gfp_t flags)
3742 return __do_kmalloc(size, flags, __builtin_return_address(0));
3744 EXPORT_SYMBOL(__kmalloc);
3746 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3748 return __do_kmalloc(size, flags, caller);
3750 EXPORT_SYMBOL(__kmalloc_track_caller);
3753 void *__kmalloc(size_t size, gfp_t flags)
3755 return __do_kmalloc(size, flags, NULL);
3757 EXPORT_SYMBOL(__kmalloc);
3761 * krealloc - reallocate memory. The contents will remain unchanged.
3762 * @p: object to reallocate memory for.
3763 * @new_size: how many bytes of memory are required.
3764 * @flags: the type of memory to allocate.
3766 * The contents of the object pointed to are preserved up to the
3767 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3768 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3769 * %NULL pointer, the object pointed to is freed.
3771 void *krealloc(const void *p, size_t new_size, gfp_t flags)
3773 struct kmem_cache *cache, *new_cache;
3777 return kmalloc_track_caller(new_size, flags);
3779 if (unlikely(!new_size)) {
3784 cache = virt_to_cache(p);
3785 new_cache = __find_general_cachep(new_size, flags);
3788 * If new size fits in the current cache, bail out.
3790 if (likely(cache == new_cache))
3794 * We are on the slow-path here so do not use __cache_alloc
3795 * because it bloats kernel text.
3797 ret = kmalloc_track_caller(new_size, flags);
3799 memcpy(ret, p, min(new_size, ksize(p)));
3804 EXPORT_SYMBOL(krealloc);
3807 * kmem_cache_free - Deallocate an object
3808 * @cachep: The cache the allocation was from.
3809 * @objp: The previously allocated object.
3811 * Free an object which was previously allocated from this
3814 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3816 unsigned long flags;
3818 BUG_ON(virt_to_cache(objp) != cachep);
3820 local_irq_save(flags);
3821 debug_check_no_locks_freed(objp, obj_size(cachep));
3822 __cache_free(cachep, objp);
3823 local_irq_restore(flags);
3825 EXPORT_SYMBOL(kmem_cache_free);
3828 * kfree - free previously allocated memory
3829 * @objp: pointer returned by kmalloc.
3831 * If @objp is NULL, no operation is performed.
3833 * Don't free memory not originally allocated by kmalloc()
3834 * or you will run into trouble.
3836 void kfree(const void *objp)
3838 struct kmem_cache *c;
3839 unsigned long flags;
3841 if (unlikely(!objp))
3843 local_irq_save(flags);
3844 kfree_debugcheck(objp);
3845 c = virt_to_cache(objp);
3846 debug_check_no_locks_freed(objp, obj_size(c));
3847 __cache_free(c, (void *)objp);
3848 local_irq_restore(flags);
3850 EXPORT_SYMBOL(kfree);
3852 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3854 return obj_size(cachep);
3856 EXPORT_SYMBOL(kmem_cache_size);
3858 const char *kmem_cache_name(struct kmem_cache *cachep)
3860 return cachep->name;
3862 EXPORT_SYMBOL_GPL(kmem_cache_name);
3865 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3867 static int alloc_kmemlist(struct kmem_cache *cachep)
3870 struct kmem_list3 *l3;
3871 struct array_cache *new_shared;
3872 struct array_cache **new_alien = NULL;
3874 for_each_online_node(node) {
3876 if (use_alien_caches) {
3877 new_alien = alloc_alien_cache(node, cachep->limit);
3883 if (cachep->shared) {
3884 new_shared = alloc_arraycache(node,
3885 cachep->shared*cachep->batchcount,
3888 free_alien_cache(new_alien);
3893 l3 = cachep->nodelists[node];
3895 struct array_cache *shared = l3->shared;
3897 spin_lock_irq(&l3->list_lock);
3900 free_block(cachep, shared->entry,
3901 shared->avail, node);
3903 l3->shared = new_shared;
3905 l3->alien = new_alien;
3908 l3->free_limit = (1 + nr_cpus_node(node)) *
3909 cachep->batchcount + cachep->num;
3910 spin_unlock_irq(&l3->list_lock);
3912 free_alien_cache(new_alien);
3915 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3917 free_alien_cache(new_alien);
3922 kmem_list3_init(l3);
3923 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3924 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3925 l3->shared = new_shared;
3926 l3->alien = new_alien;
3927 l3->free_limit = (1 + nr_cpus_node(node)) *
3928 cachep->batchcount + cachep->num;
3929 cachep->nodelists[node] = l3;
3934 if (!cachep->next.next) {
3935 /* Cache is not active yet. Roll back what we did */
3938 if (cachep->nodelists[node]) {
3939 l3 = cachep->nodelists[node];
3942 free_alien_cache(l3->alien);
3944 cachep->nodelists[node] = NULL;
3952 struct ccupdate_struct {
3953 struct kmem_cache *cachep;
3954 struct array_cache *new[NR_CPUS];
3957 static void do_ccupdate_local(void *info)
3959 struct ccupdate_struct *new = info;
3960 struct array_cache *old;
3963 old = cpu_cache_get(new->cachep);
3965 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3966 new->new[smp_processor_id()] = old;
3969 /* Always called with the cache_chain_mutex held */
3970 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3971 int batchcount, int shared)
3973 struct ccupdate_struct *new;
3976 new = kzalloc(sizeof(*new), GFP_KERNEL);
3980 for_each_online_cpu(i) {
3981 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3984 for (i--; i >= 0; i--)
3990 new->cachep = cachep;
3992 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3995 cachep->batchcount = batchcount;
3996 cachep->limit = limit;
3997 cachep->shared = shared;
3999 for_each_online_cpu(i) {
4000 struct array_cache *ccold = new->new[i];
4003 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
4004 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
4005 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
4009 return alloc_kmemlist(cachep);
4012 /* Called with cache_chain_mutex held always */
4013 static int enable_cpucache(struct kmem_cache *cachep)
4019 * The head array serves three purposes:
4020 * - create a LIFO ordering, i.e. return objects that are cache-warm
4021 * - reduce the number of spinlock operations.
4022 * - reduce the number of linked list operations on the slab and
4023 * bufctl chains: array operations are cheaper.
4024 * The numbers are guessed, we should auto-tune as described by
4027 if (cachep->buffer_size > 131072)
4029 else if (cachep->buffer_size > PAGE_SIZE)
4031 else if (cachep->buffer_size > 1024)
4033 else if (cachep->buffer_size > 256)
4039 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4040 * allocation behaviour: Most allocs on one cpu, most free operations
4041 * on another cpu. For these cases, an efficient object passing between
4042 * cpus is necessary. This is provided by a shared array. The array
4043 * replaces Bonwick's magazine layer.
4044 * On uniprocessor, it's functionally equivalent (but less efficient)
4045 * to a larger limit. Thus disabled by default.
4048 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4053 * With debugging enabled, large batchcount lead to excessively long
4054 * periods with disabled local interrupts. Limit the batchcount
4059 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4061 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4062 cachep->name, -err);
4067 * Drain an array if it contains any elements taking the l3 lock only if
4068 * necessary. Note that the l3 listlock also protects the array_cache
4069 * if drain_array() is used on the shared array.
4071 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4072 struct array_cache *ac, int force, int node)
4076 if (!ac || !ac->avail)
4078 if (ac->touched && !force) {
4081 spin_lock_irq(&l3->list_lock);
4083 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4084 if (tofree > ac->avail)
4085 tofree = (ac->avail + 1) / 2;
4086 free_block(cachep, ac->entry, tofree, node);
4087 ac->avail -= tofree;
4088 memmove(ac->entry, &(ac->entry[tofree]),
4089 sizeof(void *) * ac->avail);
4091 spin_unlock_irq(&l3->list_lock);
4096 * cache_reap - Reclaim memory from caches.
4097 * @w: work descriptor
4099 * Called from workqueue/eventd every few seconds.
4101 * - clear the per-cpu caches for this CPU.
4102 * - return freeable pages to the main free memory pool.
4104 * If we cannot acquire the cache chain mutex then just give up - we'll try
4105 * again on the next iteration.
4107 static void cache_reap(struct work_struct *w)
4109 struct kmem_cache *searchp;
4110 struct kmem_list3 *l3;
4111 int node = numa_node_id();
4112 struct delayed_work *work =
4113 container_of(w, struct delayed_work, work);
4115 if (!mutex_trylock(&cache_chain_mutex))
4116 /* Give up. Setup the next iteration. */
4119 list_for_each_entry(searchp, &cache_chain, next) {
4123 * We only take the l3 lock if absolutely necessary and we
4124 * have established with reasonable certainty that
4125 * we can do some work if the lock was obtained.
4127 l3 = searchp->nodelists[node];
4129 reap_alien(searchp, l3);
4131 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4134 * These are racy checks but it does not matter
4135 * if we skip one check or scan twice.
4137 if (time_after(l3->next_reap, jiffies))
4140 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4142 drain_array(searchp, l3, l3->shared, 0, node);
4144 if (l3->free_touched)
4145 l3->free_touched = 0;
4149 freed = drain_freelist(searchp, l3, (l3->free_limit +
4150 5 * searchp->num - 1) / (5 * searchp->num));
4151 STATS_ADD_REAPED(searchp, freed);
4157 mutex_unlock(&cache_chain_mutex);
4160 /* Set up the next iteration */
4161 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4164 #ifdef CONFIG_PROC_FS
4166 static void print_slabinfo_header(struct seq_file *m)
4169 * Output format version, so at least we can change it
4170 * without _too_ many complaints.
4173 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4175 seq_puts(m, "slabinfo - version: 2.1\n");
4177 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4178 "<objperslab> <pagesperslab>");
4179 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4180 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4182 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4183 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4184 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4189 static void *s_start(struct seq_file *m, loff_t *pos)
4192 struct list_head *p;
4194 mutex_lock(&cache_chain_mutex);
4196 print_slabinfo_header(m);
4197 p = cache_chain.next;
4200 if (p == &cache_chain)
4203 return list_entry(p, struct kmem_cache, next);
4206 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4208 struct kmem_cache *cachep = p;
4210 return cachep->next.next == &cache_chain ?
4211 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
4214 static void s_stop(struct seq_file *m, void *p)
4216 mutex_unlock(&cache_chain_mutex);
4219 static int s_show(struct seq_file *m, void *p)
4221 struct kmem_cache *cachep = p;
4223 unsigned long active_objs;
4224 unsigned long num_objs;
4225 unsigned long active_slabs = 0;
4226 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4230 struct kmem_list3 *l3;
4234 for_each_online_node(node) {
4235 l3 = cachep->nodelists[node];
4240 spin_lock_irq(&l3->list_lock);
4242 list_for_each_entry(slabp, &l3->slabs_full, list) {
4243 if (slabp->inuse != cachep->num && !error)
4244 error = "slabs_full accounting error";
4245 active_objs += cachep->num;
4248 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4249 if (slabp->inuse == cachep->num && !error)
4250 error = "slabs_partial inuse accounting error";
4251 if (!slabp->inuse && !error)
4252 error = "slabs_partial/inuse accounting error";
4253 active_objs += slabp->inuse;
4256 list_for_each_entry(slabp, &l3->slabs_free, list) {
4257 if (slabp->inuse && !error)
4258 error = "slabs_free/inuse accounting error";
4261 free_objects += l3->free_objects;
4263 shared_avail += l3->shared->avail;
4265 spin_unlock_irq(&l3->list_lock);
4267 num_slabs += active_slabs;
4268 num_objs = num_slabs * cachep->num;
4269 if (num_objs - active_objs != free_objects && !error)
4270 error = "free_objects accounting error";
4272 name = cachep->name;
4274 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4276 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4277 name, active_objs, num_objs, cachep->buffer_size,
4278 cachep->num, (1 << cachep->gfporder));
4279 seq_printf(m, " : tunables %4u %4u %4u",
4280 cachep->limit, cachep->batchcount, cachep->shared);
4281 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4282 active_slabs, num_slabs, shared_avail);
4285 unsigned long high = cachep->high_mark;
4286 unsigned long allocs = cachep->num_allocations;
4287 unsigned long grown = cachep->grown;
4288 unsigned long reaped = cachep->reaped;
4289 unsigned long errors = cachep->errors;
4290 unsigned long max_freeable = cachep->max_freeable;
4291 unsigned long node_allocs = cachep->node_allocs;
4292 unsigned long node_frees = cachep->node_frees;
4293 unsigned long overflows = cachep->node_overflow;
4295 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4296 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4297 reaped, errors, max_freeable, node_allocs,
4298 node_frees, overflows);
4302 unsigned long allochit = atomic_read(&cachep->allochit);
4303 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4304 unsigned long freehit = atomic_read(&cachep->freehit);
4305 unsigned long freemiss = atomic_read(&cachep->freemiss);
4307 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4308 allochit, allocmiss, freehit, freemiss);
4316 * slabinfo_op - iterator that generates /proc/slabinfo
4325 * num-pages-per-slab
4326 * + further values on SMP and with statistics enabled
4329 const struct seq_operations slabinfo_op = {
4336 #define MAX_SLABINFO_WRITE 128
4338 * slabinfo_write - Tuning for the slab allocator
4340 * @buffer: user buffer
4341 * @count: data length
4344 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4345 size_t count, loff_t *ppos)
4347 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4348 int limit, batchcount, shared, res;
4349 struct kmem_cache *cachep;
4351 if (count > MAX_SLABINFO_WRITE)
4353 if (copy_from_user(&kbuf, buffer, count))
4355 kbuf[MAX_SLABINFO_WRITE] = '\0';
4357 tmp = strchr(kbuf, ' ');
4362 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4365 /* Find the cache in the chain of caches. */
4366 mutex_lock(&cache_chain_mutex);
4368 list_for_each_entry(cachep, &cache_chain, next) {
4369 if (!strcmp(cachep->name, kbuf)) {
4370 if (limit < 1 || batchcount < 1 ||
4371 batchcount > limit || shared < 0) {
4374 res = do_tune_cpucache(cachep, limit,
4375 batchcount, shared);
4380 mutex_unlock(&cache_chain_mutex);
4386 #ifdef CONFIG_DEBUG_SLAB_LEAK
4388 static void *leaks_start(struct seq_file *m, loff_t *pos)
4391 struct list_head *p;
4393 mutex_lock(&cache_chain_mutex);
4394 p = cache_chain.next;
4397 if (p == &cache_chain)
4400 return list_entry(p, struct kmem_cache, next);
4403 static inline int add_caller(unsigned long *n, unsigned long v)
4413 unsigned long *q = p + 2 * i;
4427 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4433 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4439 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4440 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4442 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4447 static void show_symbol(struct seq_file *m, unsigned long address)
4449 #ifdef CONFIG_KALLSYMS
4450 unsigned long offset, size;
4451 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1];
4453 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4454 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4456 seq_printf(m, " [%s]", modname);
4460 seq_printf(m, "%p", (void *)address);
4463 static int leaks_show(struct seq_file *m, void *p)
4465 struct kmem_cache *cachep = p;
4467 struct kmem_list3 *l3;
4469 unsigned long *n = m->private;
4473 if (!(cachep->flags & SLAB_STORE_USER))
4475 if (!(cachep->flags & SLAB_RED_ZONE))
4478 /* OK, we can do it */
4482 for_each_online_node(node) {
4483 l3 = cachep->nodelists[node];
4488 spin_lock_irq(&l3->list_lock);
4490 list_for_each_entry(slabp, &l3->slabs_full, list)
4491 handle_slab(n, cachep, slabp);
4492 list_for_each_entry(slabp, &l3->slabs_partial, list)
4493 handle_slab(n, cachep, slabp);
4494 spin_unlock_irq(&l3->list_lock);
4496 name = cachep->name;
4498 /* Increase the buffer size */
4499 mutex_unlock(&cache_chain_mutex);
4500 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4502 /* Too bad, we are really out */
4504 mutex_lock(&cache_chain_mutex);
4507 *(unsigned long *)m->private = n[0] * 2;
4509 mutex_lock(&cache_chain_mutex);
4510 /* Now make sure this entry will be retried */
4514 for (i = 0; i < n[1]; i++) {
4515 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4516 show_symbol(m, n[2*i+2]);
4523 const struct seq_operations slabstats_op = {
4524 .start = leaks_start,
4533 * ksize - get the actual amount of memory allocated for a given object
4534 * @objp: Pointer to the object
4536 * kmalloc may internally round up allocations and return more memory
4537 * than requested. ksize() can be used to determine the actual amount of
4538 * memory allocated. The caller may use this additional memory, even though
4539 * a smaller amount of memory was initially specified with the kmalloc call.
4540 * The caller must guarantee that objp points to a valid object previously
4541 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4542 * must not be freed during the duration of the call.
4544 size_t ksize(const void *objp)
4546 if (unlikely(objp == NULL))
4549 return obj_size(virt_to_cache(objp));