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>
113 #include <asm/cacheflush.h>
114 #include <asm/tlbflush.h>
115 #include <asm/page.h>
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * 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 BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
154 #define ARCH_KMALLOC_MINALIGN 0
157 #ifndef ARCH_SLAB_MINALIGN
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
165 #define ARCH_SLAB_MINALIGN 0
168 #ifndef ARCH_KMALLOC_FLAGS
169 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 /* Legal flag mask for kmem_cache_create(). */
174 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
181 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
190 * Bufctl's are used for linking objs within a slab
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 typedef unsigned int kmem_bufctl_t;
207 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
209 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
225 unsigned short nodeid;
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
242 * We assume struct slab_rcu can overlay struct slab when destroying.
245 struct rcu_head head;
246 struct kmem_cache *cachep;
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
258 * The limit is stored in the per-cpu structure to reduce the data cache
265 unsigned int batchcount;
266 unsigned int touched;
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
272 * [0] is for gcc 2.95. It should really be [].
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
280 #define BOOT_CPUCACHE_ENTRIES 1
281 struct arraycache_init {
282 struct array_cache cache;
283 void *entries[BOOT_CPUCACHE_ENTRIES];
287 * The slab lists for all objects.
290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
294 unsigned int free_limit;
295 unsigned int colour_next; /* Per-node cache coloring */
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
304 * Need this for bootstrapping a per node allocator.
306 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308 #define CACHE_CACHE 0
310 #define SIZE_L3 (1 + MAX_NUMNODES)
312 static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 static int enable_cpucache(struct kmem_cache *cachep);
317 static void cache_reap(struct work_struct *unused);
320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
323 static __always_inline int index_of(const size_t size)
325 extern void __bad_size(void);
327 if (__builtin_constant_p(size)) {
335 #include "linux/kmalloc_sizes.h"
343 static int slab_early_init = 1;
345 #define INDEX_AC index_of(sizeof(struct arraycache_init))
346 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
348 static void kmem_list3_init(struct kmem_list3 *parent)
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
355 parent->colour_next = 0;
356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
361 #define MAKE_LIST(cachep, listp, slab, nodeid) \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
367 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 /* 1) per-cpu data, touched during every alloc/free */
382 struct array_cache *array[NR_CPUS];
383 /* 2) Cache tunables. Protected by cache_chain_mutex */
384 unsigned int batchcount;
388 unsigned int buffer_size;
389 /* 3) touched by every alloc & free from the backend */
390 struct kmem_list3 *nodelists[MAX_NUMNODES];
392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
395 /* 4) cache_grow/shrink */
396 /* order of pgs per slab (2^n) */
397 unsigned int gfporder;
399 /* force GFP flags, e.g. GFP_DMA */
402 size_t colour; /* cache colouring range */
403 unsigned int colour_off; /* colour offset */
404 struct kmem_cache *slabp_cache;
405 unsigned int slab_size;
406 unsigned int dflags; /* dynamic flags */
408 /* constructor func */
409 void (*ctor) (void *, struct kmem_cache *, unsigned long);
411 /* de-constructor func */
412 void (*dtor) (void *, struct kmem_cache *, unsigned long);
414 /* 5) cache creation/removal */
416 struct list_head next;
420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
429 unsigned long node_overflow;
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
447 #define CFLGS_OFF_SLAB (0x80000000UL)
448 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
450 #define BATCHREFILL_LIMIT 16
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
455 * OTOH the cpuarrays can contain lots of objects,
456 * which could lock up otherwise freeable slabs.
458 #define REAPTIMEOUT_CPUC (2*HZ)
459 #define REAPTIMEOUT_LIST3 (4*HZ)
462 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
463 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465 #define STATS_INC_GROWN(x) ((x)->grown++)
466 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
467 #define STATS_SET_HIGH(x) \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
472 #define STATS_INC_ERR(x) ((x)->errors++)
473 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
474 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
475 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
476 #define STATS_SET_FREEABLE(x, i) \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
481 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
486 #define STATS_INC_ACTIVE(x) do { } while (0)
487 #define STATS_DEC_ACTIVE(x) do { } while (0)
488 #define STATS_INC_ALLOCED(x) do { } while (0)
489 #define STATS_INC_GROWN(x) do { } while (0)
490 #define STATS_ADD_REAPED(x,y) do { } while (0)
491 #define STATS_SET_HIGH(x) do { } while (0)
492 #define STATS_INC_ERR(x) do { } while (0)
493 #define STATS_INC_NODEALLOCS(x) do { } while (0)
494 #define STATS_INC_NODEFREES(x) do { } while (0)
495 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
496 #define STATS_SET_FREEABLE(x, i) do { } while (0)
497 #define STATS_INC_ALLOCHIT(x) do { } while (0)
498 #define STATS_INC_ALLOCMISS(x) do { } while (0)
499 #define STATS_INC_FREEHIT(x) do { } while (0)
500 #define STATS_INC_FREEMISS(x) do { } while (0)
506 * memory layout of objects:
508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
518 static int obj_offset(struct kmem_cache *cachep)
520 return cachep->obj_offset;
523 static int obj_size(struct kmem_cache *cachep)
525 return cachep->obj_size;
528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
538 return (unsigned long *)(objp + cachep->buffer_size -
540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
543 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
551 #define obj_offset(x) 0
552 #define obj_size(cachep) (cachep->buffer_size)
553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 #if defined(CONFIG_LARGE_ALLOCS)
564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
565 #define MAX_GFP_ORDER 13 /* up to 32Mb */
566 #elif defined(CONFIG_MMU)
567 #define MAX_OBJ_ORDER 5 /* 32 pages */
568 #define MAX_GFP_ORDER 5 /* 32 pages */
570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
571 #define MAX_GFP_ORDER 8 /* up to 1Mb */
575 * Do not go above this order unless 0 objects fit into the slab.
577 #define BREAK_GFP_ORDER_HI 1
578 #define BREAK_GFP_ORDER_LO 0
579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
588 page->lru.next = (struct list_head *)cache;
591 static inline struct kmem_cache *page_get_cache(struct page *page)
593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
595 BUG_ON(!PageSlab(page));
596 return (struct kmem_cache *)page->lru.next;
599 static inline void page_set_slab(struct page *page, struct slab *slab)
601 page->lru.prev = (struct list_head *)slab;
604 static inline struct slab *page_get_slab(struct page *page)
606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
608 BUG_ON(!PageSlab(page));
609 return (struct slab *)page->lru.prev;
612 static inline struct kmem_cache *virt_to_cache(const void *obj)
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
618 static inline struct slab *virt_to_slab(const void *obj)
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
624 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
627 return slab->s_mem + cache->buffer_size * idx;
630 static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 struct cache_sizes malloc_sizes[] = {
640 #define CACHE(x) { .cs_size = (x) },
641 #include <linux/kmalloc_sizes.h>
645 EXPORT_SYMBOL(malloc_sizes);
647 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
653 static struct cache_names __initdata cache_names[] = {
654 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655 #include <linux/kmalloc_sizes.h>
660 static struct arraycache_init initarray_cache __initdata =
661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
662 static struct arraycache_init initarray_generic =
663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
665 /* internal cache of cache description objs */
666 static struct kmem_cache cache_cache = {
668 .limit = BOOT_CPUCACHE_ENTRIES,
670 .buffer_size = sizeof(struct kmem_cache),
671 .name = "kmem_cache",
673 .obj_size = sizeof(struct kmem_cache),
677 #define BAD_ALIEN_MAGIC 0x01020304ul
679 #ifdef CONFIG_LOCKDEP
682 * Slab sometimes uses the kmalloc slabs to store the slab headers
683 * for other slabs "off slab".
684 * The locking for this is tricky in that it nests within the locks
685 * of all other slabs in a few places; to deal with this special
686 * locking we put on-slab caches into a separate lock-class.
688 * We set lock class for alien array caches which are up during init.
689 * The lock annotation will be lost if all cpus of a node goes down and
690 * then comes back up during hotplug
692 static struct lock_class_key on_slab_l3_key;
693 static struct lock_class_key on_slab_alc_key;
695 static inline void init_lock_keys(void)
699 struct cache_sizes *s = malloc_sizes;
701 while (s->cs_size != ULONG_MAX) {
703 struct array_cache **alc;
705 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
706 if (!l3 || OFF_SLAB(s->cs_cachep))
708 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
711 * FIXME: This check for BAD_ALIEN_MAGIC
712 * should go away when common slab code is taught to
713 * work even without alien caches.
714 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
715 * for alloc_alien_cache,
717 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
721 lockdep_set_class(&alc[r]->lock,
729 static inline void init_lock_keys(void)
735 * 1. Guard access to the cache-chain.
736 * 2. Protect sanity of cpu_online_map against cpu hotplug events
738 static DEFINE_MUTEX(cache_chain_mutex);
739 static struct list_head cache_chain;
742 * chicken and egg problem: delay the per-cpu array allocation
743 * until the general caches are up.
753 * used by boot code to determine if it can use slab based allocator
755 int slab_is_available(void)
757 return g_cpucache_up == FULL;
760 static DEFINE_PER_CPU(struct delayed_work, reap_work);
762 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
764 return cachep->array[smp_processor_id()];
767 static inline struct kmem_cache *__find_general_cachep(size_t size,
770 struct cache_sizes *csizep = malloc_sizes;
773 /* This happens if someone tries to call
774 * kmem_cache_create(), or __kmalloc(), before
775 * the generic caches are initialized.
777 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
779 while (size > csizep->cs_size)
783 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
784 * has cs_{dma,}cachep==NULL. Thus no special case
785 * for large kmalloc calls required.
787 if (unlikely(gfpflags & GFP_DMA))
788 return csizep->cs_dmacachep;
789 return csizep->cs_cachep;
792 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
794 return __find_general_cachep(size, gfpflags);
797 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
799 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
803 * Calculate the number of objects and left-over bytes for a given buffer size.
805 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
806 size_t align, int flags, size_t *left_over,
811 size_t slab_size = PAGE_SIZE << gfporder;
814 * The slab management structure can be either off the slab or
815 * on it. For the latter case, the memory allocated for a
819 * - One kmem_bufctl_t for each object
820 * - Padding to respect alignment of @align
821 * - @buffer_size bytes for each object
823 * If the slab management structure is off the slab, then the
824 * alignment will already be calculated into the size. Because
825 * the slabs are all pages aligned, the objects will be at the
826 * correct alignment when allocated.
828 if (flags & CFLGS_OFF_SLAB) {
830 nr_objs = slab_size / buffer_size;
832 if (nr_objs > SLAB_LIMIT)
833 nr_objs = SLAB_LIMIT;
836 * Ignore padding for the initial guess. The padding
837 * is at most @align-1 bytes, and @buffer_size is at
838 * least @align. In the worst case, this result will
839 * be one greater than the number of objects that fit
840 * into the memory allocation when taking the padding
843 nr_objs = (slab_size - sizeof(struct slab)) /
844 (buffer_size + sizeof(kmem_bufctl_t));
847 * This calculated number will be either the right
848 * amount, or one greater than what we want.
850 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
854 if (nr_objs > SLAB_LIMIT)
855 nr_objs = SLAB_LIMIT;
857 mgmt_size = slab_mgmt_size(nr_objs, align);
860 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
863 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
865 static void __slab_error(const char *function, struct kmem_cache *cachep,
868 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
869 function, cachep->name, msg);
874 * By default on NUMA we use alien caches to stage the freeing of
875 * objects allocated from other nodes. This causes massive memory
876 * inefficiencies when using fake NUMA setup to split memory into a
877 * large number of small nodes, so it can be disabled on the command
881 static int use_alien_caches __read_mostly = 1;
882 static int __init noaliencache_setup(char *s)
884 use_alien_caches = 0;
887 __setup("noaliencache", noaliencache_setup);
891 * Special reaping functions for NUMA systems called from cache_reap().
892 * These take care of doing round robin flushing of alien caches (containing
893 * objects freed on different nodes from which they were allocated) and the
894 * flushing of remote pcps by calling drain_node_pages.
896 static DEFINE_PER_CPU(unsigned long, reap_node);
898 static void init_reap_node(int cpu)
902 node = next_node(cpu_to_node(cpu), node_online_map);
903 if (node == MAX_NUMNODES)
904 node = first_node(node_online_map);
906 per_cpu(reap_node, cpu) = node;
909 static void next_reap_node(void)
911 int node = __get_cpu_var(reap_node);
914 * Also drain per cpu pages on remote zones
916 if (node != numa_node_id())
917 drain_node_pages(node);
919 node = next_node(node, node_online_map);
920 if (unlikely(node >= MAX_NUMNODES))
921 node = first_node(node_online_map);
922 __get_cpu_var(reap_node) = node;
926 #define init_reap_node(cpu) do { } while (0)
927 #define next_reap_node(void) do { } while (0)
931 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
932 * via the workqueue/eventd.
933 * Add the CPU number into the expiration time to minimize the possibility of
934 * the CPUs getting into lockstep and contending for the global cache chain
937 static void __devinit start_cpu_timer(int cpu)
939 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
942 * When this gets called from do_initcalls via cpucache_init(),
943 * init_workqueues() has already run, so keventd will be setup
946 if (keventd_up() && reap_work->work.func == NULL) {
948 INIT_DELAYED_WORK(reap_work, cache_reap);
949 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
953 static struct array_cache *alloc_arraycache(int node, int entries,
956 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
957 struct array_cache *nc = NULL;
959 nc = kmalloc_node(memsize, GFP_KERNEL, node);
963 nc->batchcount = batchcount;
965 spin_lock_init(&nc->lock);
971 * Transfer objects in one arraycache to another.
972 * Locking must be handled by the caller.
974 * Return the number of entries transferred.
976 static int transfer_objects(struct array_cache *to,
977 struct array_cache *from, unsigned int max)
979 /* Figure out how many entries to transfer */
980 int nr = min(min(from->avail, max), to->limit - to->avail);
985 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
996 #define drain_alien_cache(cachep, alien) do { } while (0)
997 #define reap_alien(cachep, l3) do { } while (0)
999 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1001 return (struct array_cache **)BAD_ALIEN_MAGIC;
1004 static inline void free_alien_cache(struct array_cache **ac_ptr)
1008 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1013 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1019 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1020 gfp_t flags, int nodeid)
1025 #else /* CONFIG_NUMA */
1027 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1028 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1030 static struct array_cache **alloc_alien_cache(int node, int limit)
1032 struct array_cache **ac_ptr;
1033 int memsize = sizeof(void *) * MAX_NUMNODES;
1038 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1041 if (i == node || !node_online(i)) {
1045 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1047 for (i--; i <= 0; i--)
1057 static void free_alien_cache(struct array_cache **ac_ptr)
1068 static void __drain_alien_cache(struct kmem_cache *cachep,
1069 struct array_cache *ac, int node)
1071 struct kmem_list3 *rl3 = cachep->nodelists[node];
1074 spin_lock(&rl3->list_lock);
1076 * Stuff objects into the remote nodes shared array first.
1077 * That way we could avoid the overhead of putting the objects
1078 * into the free lists and getting them back later.
1081 transfer_objects(rl3->shared, ac, ac->limit);
1083 free_block(cachep, ac->entry, ac->avail, node);
1085 spin_unlock(&rl3->list_lock);
1090 * Called from cache_reap() to regularly drain alien caches round robin.
1092 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1094 int node = __get_cpu_var(reap_node);
1097 struct array_cache *ac = l3->alien[node];
1099 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1100 __drain_alien_cache(cachep, ac, node);
1101 spin_unlock_irq(&ac->lock);
1106 static void drain_alien_cache(struct kmem_cache *cachep,
1107 struct array_cache **alien)
1110 struct array_cache *ac;
1111 unsigned long flags;
1113 for_each_online_node(i) {
1116 spin_lock_irqsave(&ac->lock, flags);
1117 __drain_alien_cache(cachep, ac, i);
1118 spin_unlock_irqrestore(&ac->lock, flags);
1123 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1125 struct slab *slabp = virt_to_slab(objp);
1126 int nodeid = slabp->nodeid;
1127 struct kmem_list3 *l3;
1128 struct array_cache *alien = NULL;
1131 node = numa_node_id();
1134 * Make sure we are not freeing a object from another node to the array
1135 * cache on this cpu.
1137 if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches))
1140 l3 = cachep->nodelists[node];
1141 STATS_INC_NODEFREES(cachep);
1142 if (l3->alien && l3->alien[nodeid]) {
1143 alien = l3->alien[nodeid];
1144 spin_lock(&alien->lock);
1145 if (unlikely(alien->avail == alien->limit)) {
1146 STATS_INC_ACOVERFLOW(cachep);
1147 __drain_alien_cache(cachep, alien, nodeid);
1149 alien->entry[alien->avail++] = objp;
1150 spin_unlock(&alien->lock);
1152 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1153 free_block(cachep, &objp, 1, nodeid);
1154 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1160 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1161 unsigned long action, void *hcpu)
1163 long cpu = (long)hcpu;
1164 struct kmem_cache *cachep;
1165 struct kmem_list3 *l3 = NULL;
1166 int node = cpu_to_node(cpu);
1167 int memsize = sizeof(struct kmem_list3);
1170 case CPU_UP_PREPARE:
1171 mutex_lock(&cache_chain_mutex);
1173 * We need to do this right in the beginning since
1174 * alloc_arraycache's are going to use this list.
1175 * kmalloc_node allows us to add the slab to the right
1176 * kmem_list3 and not this cpu's kmem_list3
1179 list_for_each_entry(cachep, &cache_chain, next) {
1181 * Set up the size64 kmemlist for cpu before we can
1182 * begin anything. Make sure some other cpu on this
1183 * node has not already allocated this
1185 if (!cachep->nodelists[node]) {
1186 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1189 kmem_list3_init(l3);
1190 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1191 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1194 * The l3s don't come and go as CPUs come and
1195 * go. cache_chain_mutex is sufficient
1198 cachep->nodelists[node] = l3;
1201 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1202 cachep->nodelists[node]->free_limit =
1203 (1 + nr_cpus_node(node)) *
1204 cachep->batchcount + cachep->num;
1205 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1209 * Now we can go ahead with allocating the shared arrays and
1212 list_for_each_entry(cachep, &cache_chain, next) {
1213 struct array_cache *nc;
1214 struct array_cache *shared;
1215 struct array_cache **alien = NULL;
1217 nc = alloc_arraycache(node, cachep->limit,
1218 cachep->batchcount);
1221 shared = alloc_arraycache(node,
1222 cachep->shared * cachep->batchcount,
1227 if (use_alien_caches) {
1228 alien = alloc_alien_cache(node, cachep->limit);
1232 cachep->array[cpu] = nc;
1233 l3 = cachep->nodelists[node];
1236 spin_lock_irq(&l3->list_lock);
1239 * We are serialised from CPU_DEAD or
1240 * CPU_UP_CANCELLED by the cpucontrol lock
1242 l3->shared = shared;
1251 spin_unlock_irq(&l3->list_lock);
1253 free_alien_cache(alien);
1257 mutex_unlock(&cache_chain_mutex);
1258 start_cpu_timer(cpu);
1260 #ifdef CONFIG_HOTPLUG_CPU
1261 case CPU_DOWN_PREPARE:
1262 mutex_lock(&cache_chain_mutex);
1264 case CPU_DOWN_FAILED:
1265 mutex_unlock(&cache_chain_mutex);
1269 * Even if all the cpus of a node are down, we don't free the
1270 * kmem_list3 of any cache. This to avoid a race between
1271 * cpu_down, and a kmalloc allocation from another cpu for
1272 * memory from the node of the cpu going down. The list3
1273 * structure is usually allocated from kmem_cache_create() and
1274 * gets destroyed at kmem_cache_destroy().
1278 case CPU_UP_CANCELED:
1279 list_for_each_entry(cachep, &cache_chain, next) {
1280 struct array_cache *nc;
1281 struct array_cache *shared;
1282 struct array_cache **alien;
1285 mask = node_to_cpumask(node);
1286 /* cpu is dead; no one can alloc from it. */
1287 nc = cachep->array[cpu];
1288 cachep->array[cpu] = NULL;
1289 l3 = cachep->nodelists[node];
1292 goto free_array_cache;
1294 spin_lock_irq(&l3->list_lock);
1296 /* Free limit for this kmem_list3 */
1297 l3->free_limit -= cachep->batchcount;
1299 free_block(cachep, nc->entry, nc->avail, node);
1301 if (!cpus_empty(mask)) {
1302 spin_unlock_irq(&l3->list_lock);
1303 goto free_array_cache;
1306 shared = l3->shared;
1308 free_block(cachep, l3->shared->entry,
1309 l3->shared->avail, node);
1316 spin_unlock_irq(&l3->list_lock);
1320 drain_alien_cache(cachep, alien);
1321 free_alien_cache(alien);
1327 * In the previous loop, all the objects were freed to
1328 * the respective cache's slabs, now we can go ahead and
1329 * shrink each nodelist to its limit.
1331 list_for_each_entry(cachep, &cache_chain, next) {
1332 l3 = cachep->nodelists[node];
1335 drain_freelist(cachep, l3, l3->free_objects);
1337 mutex_unlock(&cache_chain_mutex);
1345 static struct notifier_block __cpuinitdata cpucache_notifier = {
1346 &cpuup_callback, NULL, 0
1350 * swap the static kmem_list3 with kmalloced memory
1352 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1355 struct kmem_list3 *ptr;
1357 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1360 local_irq_disable();
1361 memcpy(ptr, list, sizeof(struct kmem_list3));
1363 * Do not assume that spinlocks can be initialized via memcpy:
1365 spin_lock_init(&ptr->list_lock);
1367 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1368 cachep->nodelists[nodeid] = ptr;
1373 * Initialisation. Called after the page allocator have been initialised and
1374 * before smp_init().
1376 void __init kmem_cache_init(void)
1379 struct cache_sizes *sizes;
1380 struct cache_names *names;
1385 for (i = 0; i < NUM_INIT_LISTS; i++) {
1386 kmem_list3_init(&initkmem_list3[i]);
1387 if (i < MAX_NUMNODES)
1388 cache_cache.nodelists[i] = NULL;
1392 * Fragmentation resistance on low memory - only use bigger
1393 * page orders on machines with more than 32MB of memory.
1395 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1396 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1398 /* Bootstrap is tricky, because several objects are allocated
1399 * from caches that do not exist yet:
1400 * 1) initialize the cache_cache cache: it contains the struct
1401 * kmem_cache structures of all caches, except cache_cache itself:
1402 * cache_cache is statically allocated.
1403 * Initially an __init data area is used for the head array and the
1404 * kmem_list3 structures, it's replaced with a kmalloc allocated
1405 * array at the end of the bootstrap.
1406 * 2) Create the first kmalloc cache.
1407 * The struct kmem_cache for the new cache is allocated normally.
1408 * An __init data area is used for the head array.
1409 * 3) Create the remaining kmalloc caches, with minimally sized
1411 * 4) Replace the __init data head arrays for cache_cache and the first
1412 * kmalloc cache with kmalloc allocated arrays.
1413 * 5) Replace the __init data for kmem_list3 for cache_cache and
1414 * the other cache's with kmalloc allocated memory.
1415 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1418 node = numa_node_id();
1420 /* 1) create the cache_cache */
1421 INIT_LIST_HEAD(&cache_chain);
1422 list_add(&cache_cache.next, &cache_chain);
1423 cache_cache.colour_off = cache_line_size();
1424 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1425 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1427 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1430 for (order = 0; order < MAX_ORDER; order++) {
1431 cache_estimate(order, cache_cache.buffer_size,
1432 cache_line_size(), 0, &left_over, &cache_cache.num);
1433 if (cache_cache.num)
1436 BUG_ON(!cache_cache.num);
1437 cache_cache.gfporder = order;
1438 cache_cache.colour = left_over / cache_cache.colour_off;
1439 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1440 sizeof(struct slab), cache_line_size());
1442 /* 2+3) create the kmalloc caches */
1443 sizes = malloc_sizes;
1444 names = cache_names;
1447 * Initialize the caches that provide memory for the array cache and the
1448 * kmem_list3 structures first. Without this, further allocations will
1452 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1453 sizes[INDEX_AC].cs_size,
1454 ARCH_KMALLOC_MINALIGN,
1455 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1458 if (INDEX_AC != INDEX_L3) {
1459 sizes[INDEX_L3].cs_cachep =
1460 kmem_cache_create(names[INDEX_L3].name,
1461 sizes[INDEX_L3].cs_size,
1462 ARCH_KMALLOC_MINALIGN,
1463 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1467 slab_early_init = 0;
1469 while (sizes->cs_size != ULONG_MAX) {
1471 * For performance, all the general caches are L1 aligned.
1472 * This should be particularly beneficial on SMP boxes, as it
1473 * eliminates "false sharing".
1474 * Note for systems short on memory removing the alignment will
1475 * allow tighter packing of the smaller caches.
1477 if (!sizes->cs_cachep) {
1478 sizes->cs_cachep = kmem_cache_create(names->name,
1480 ARCH_KMALLOC_MINALIGN,
1481 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1485 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1487 ARCH_KMALLOC_MINALIGN,
1488 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1494 /* 4) Replace the bootstrap head arrays */
1496 struct array_cache *ptr;
1498 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1500 local_irq_disable();
1501 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1502 memcpy(ptr, cpu_cache_get(&cache_cache),
1503 sizeof(struct arraycache_init));
1505 * Do not assume that spinlocks can be initialized via memcpy:
1507 spin_lock_init(&ptr->lock);
1509 cache_cache.array[smp_processor_id()] = ptr;
1512 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1514 local_irq_disable();
1515 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1516 != &initarray_generic.cache);
1517 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1518 sizeof(struct arraycache_init));
1520 * Do not assume that spinlocks can be initialized via memcpy:
1522 spin_lock_init(&ptr->lock);
1524 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1528 /* 5) Replace the bootstrap kmem_list3's */
1532 /* Replace the static kmem_list3 structures for the boot cpu */
1533 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1535 for_each_online_node(nid) {
1536 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1537 &initkmem_list3[SIZE_AC + nid], nid);
1539 if (INDEX_AC != INDEX_L3) {
1540 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1541 &initkmem_list3[SIZE_L3 + nid], nid);
1546 /* 6) resize the head arrays to their final sizes */
1548 struct kmem_cache *cachep;
1549 mutex_lock(&cache_chain_mutex);
1550 list_for_each_entry(cachep, &cache_chain, next)
1551 if (enable_cpucache(cachep))
1553 mutex_unlock(&cache_chain_mutex);
1556 /* Annotate slab for lockdep -- annotate the malloc caches */
1561 g_cpucache_up = FULL;
1564 * Register a cpu startup notifier callback that initializes
1565 * cpu_cache_get for all new cpus
1567 register_cpu_notifier(&cpucache_notifier);
1570 * The reap timers are started later, with a module init call: That part
1571 * of the kernel is not yet operational.
1575 static int __init cpucache_init(void)
1580 * Register the timers that return unneeded pages to the page allocator
1582 for_each_online_cpu(cpu)
1583 start_cpu_timer(cpu);
1586 __initcall(cpucache_init);
1589 * Interface to system's page allocator. No need to hold the cache-lock.
1591 * If we requested dmaable memory, we will get it. Even if we
1592 * did not request dmaable memory, we might get it, but that
1593 * would be relatively rare and ignorable.
1595 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1603 * Nommu uses slab's for process anonymous memory allocations, and thus
1604 * requires __GFP_COMP to properly refcount higher order allocations
1606 flags |= __GFP_COMP;
1609 flags |= cachep->gfpflags;
1611 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1615 nr_pages = (1 << cachep->gfporder);
1616 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1617 add_zone_page_state(page_zone(page),
1618 NR_SLAB_RECLAIMABLE, nr_pages);
1620 add_zone_page_state(page_zone(page),
1621 NR_SLAB_UNRECLAIMABLE, nr_pages);
1622 for (i = 0; i < nr_pages; i++)
1623 __SetPageSlab(page + i);
1624 return page_address(page);
1628 * Interface to system's page release.
1630 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1632 unsigned long i = (1 << cachep->gfporder);
1633 struct page *page = virt_to_page(addr);
1634 const unsigned long nr_freed = i;
1636 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1637 sub_zone_page_state(page_zone(page),
1638 NR_SLAB_RECLAIMABLE, nr_freed);
1640 sub_zone_page_state(page_zone(page),
1641 NR_SLAB_UNRECLAIMABLE, nr_freed);
1643 BUG_ON(!PageSlab(page));
1644 __ClearPageSlab(page);
1647 if (current->reclaim_state)
1648 current->reclaim_state->reclaimed_slab += nr_freed;
1649 free_pages((unsigned long)addr, cachep->gfporder);
1652 static void kmem_rcu_free(struct rcu_head *head)
1654 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1655 struct kmem_cache *cachep = slab_rcu->cachep;
1657 kmem_freepages(cachep, slab_rcu->addr);
1658 if (OFF_SLAB(cachep))
1659 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1664 #ifdef CONFIG_DEBUG_PAGEALLOC
1665 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1666 unsigned long caller)
1668 int size = obj_size(cachep);
1670 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1672 if (size < 5 * sizeof(unsigned long))
1675 *addr++ = 0x12345678;
1677 *addr++ = smp_processor_id();
1678 size -= 3 * sizeof(unsigned long);
1680 unsigned long *sptr = &caller;
1681 unsigned long svalue;
1683 while (!kstack_end(sptr)) {
1685 if (kernel_text_address(svalue)) {
1687 size -= sizeof(unsigned long);
1688 if (size <= sizeof(unsigned long))
1694 *addr++ = 0x87654321;
1698 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1700 int size = obj_size(cachep);
1701 addr = &((char *)addr)[obj_offset(cachep)];
1703 memset(addr, val, size);
1704 *(unsigned char *)(addr + size - 1) = POISON_END;
1707 static void dump_line(char *data, int offset, int limit)
1710 unsigned char error = 0;
1713 printk(KERN_ERR "%03x:", offset);
1714 for (i = 0; i < limit; i++) {
1715 if (data[offset + i] != POISON_FREE) {
1716 error = data[offset + i];
1719 printk(" %02x", (unsigned char)data[offset + i]);
1723 if (bad_count == 1) {
1724 error ^= POISON_FREE;
1725 if (!(error & (error - 1))) {
1726 printk(KERN_ERR "Single bit error detected. Probably "
1729 printk(KERN_ERR "Run memtest86+ or a similar memory "
1732 printk(KERN_ERR "Run a memory test tool.\n");
1741 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1746 if (cachep->flags & SLAB_RED_ZONE) {
1747 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1748 *dbg_redzone1(cachep, objp),
1749 *dbg_redzone2(cachep, objp));
1752 if (cachep->flags & SLAB_STORE_USER) {
1753 printk(KERN_ERR "Last user: [<%p>]",
1754 *dbg_userword(cachep, objp));
1755 print_symbol("(%s)",
1756 (unsigned long)*dbg_userword(cachep, objp));
1759 realobj = (char *)objp + obj_offset(cachep);
1760 size = obj_size(cachep);
1761 for (i = 0; i < size && lines; i += 16, lines--) {
1764 if (i + limit > size)
1766 dump_line(realobj, i, limit);
1770 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1776 realobj = (char *)objp + obj_offset(cachep);
1777 size = obj_size(cachep);
1779 for (i = 0; i < size; i++) {
1780 char exp = POISON_FREE;
1783 if (realobj[i] != exp) {
1789 "Slab corruption: start=%p, len=%d\n",
1791 print_objinfo(cachep, objp, 0);
1793 /* Hexdump the affected line */
1796 if (i + limit > size)
1798 dump_line(realobj, i, limit);
1801 /* Limit to 5 lines */
1807 /* Print some data about the neighboring objects, if they
1810 struct slab *slabp = virt_to_slab(objp);
1813 objnr = obj_to_index(cachep, slabp, objp);
1815 objp = index_to_obj(cachep, slabp, objnr - 1);
1816 realobj = (char *)objp + obj_offset(cachep);
1817 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1819 print_objinfo(cachep, objp, 2);
1821 if (objnr + 1 < cachep->num) {
1822 objp = index_to_obj(cachep, slabp, objnr + 1);
1823 realobj = (char *)objp + obj_offset(cachep);
1824 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1826 print_objinfo(cachep, objp, 2);
1834 * slab_destroy_objs - destroy a slab and its objects
1835 * @cachep: cache pointer being destroyed
1836 * @slabp: slab pointer being destroyed
1838 * Call the registered destructor for each object in a slab that is being
1841 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1844 for (i = 0; i < cachep->num; i++) {
1845 void *objp = index_to_obj(cachep, slabp, i);
1847 if (cachep->flags & SLAB_POISON) {
1848 #ifdef CONFIG_DEBUG_PAGEALLOC
1849 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1851 kernel_map_pages(virt_to_page(objp),
1852 cachep->buffer_size / PAGE_SIZE, 1);
1854 check_poison_obj(cachep, objp);
1856 check_poison_obj(cachep, objp);
1859 if (cachep->flags & SLAB_RED_ZONE) {
1860 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1861 slab_error(cachep, "start of a freed object "
1863 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1864 slab_error(cachep, "end of a freed object "
1867 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1868 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1872 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1876 for (i = 0; i < cachep->num; i++) {
1877 void *objp = index_to_obj(cachep, slabp, i);
1878 (cachep->dtor) (objp, cachep, 0);
1885 * slab_destroy - destroy and release all objects in a slab
1886 * @cachep: cache pointer being destroyed
1887 * @slabp: slab pointer being destroyed
1889 * Destroy all the objs in a slab, and release the mem back to the system.
1890 * Before calling the slab must have been unlinked from the cache. The
1891 * cache-lock is not held/needed.
1893 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1895 void *addr = slabp->s_mem - slabp->colouroff;
1897 slab_destroy_objs(cachep, slabp);
1898 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1899 struct slab_rcu *slab_rcu;
1901 slab_rcu = (struct slab_rcu *)slabp;
1902 slab_rcu->cachep = cachep;
1903 slab_rcu->addr = addr;
1904 call_rcu(&slab_rcu->head, kmem_rcu_free);
1906 kmem_freepages(cachep, addr);
1907 if (OFF_SLAB(cachep))
1908 kmem_cache_free(cachep->slabp_cache, slabp);
1913 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1914 * size of kmem_list3.
1916 static void set_up_list3s(struct kmem_cache *cachep, int index)
1920 for_each_online_node(node) {
1921 cachep->nodelists[node] = &initkmem_list3[index + node];
1922 cachep->nodelists[node]->next_reap = jiffies +
1924 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1928 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1931 struct kmem_list3 *l3;
1933 for_each_online_cpu(i)
1934 kfree(cachep->array[i]);
1936 /* NUMA: free the list3 structures */
1937 for_each_online_node(i) {
1938 l3 = cachep->nodelists[i];
1941 free_alien_cache(l3->alien);
1945 kmem_cache_free(&cache_cache, cachep);
1950 * calculate_slab_order - calculate size (page order) of slabs
1951 * @cachep: pointer to the cache that is being created
1952 * @size: size of objects to be created in this cache.
1953 * @align: required alignment for the objects.
1954 * @flags: slab allocation flags
1956 * Also calculates the number of objects per slab.
1958 * This could be made much more intelligent. For now, try to avoid using
1959 * high order pages for slabs. When the gfp() functions are more friendly
1960 * towards high-order requests, this should be changed.
1962 static size_t calculate_slab_order(struct kmem_cache *cachep,
1963 size_t size, size_t align, unsigned long flags)
1965 unsigned long offslab_limit;
1966 size_t left_over = 0;
1969 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1973 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1977 if (flags & CFLGS_OFF_SLAB) {
1979 * Max number of objs-per-slab for caches which
1980 * use off-slab slabs. Needed to avoid a possible
1981 * looping condition in cache_grow().
1983 offslab_limit = size - sizeof(struct slab);
1984 offslab_limit /= sizeof(kmem_bufctl_t);
1986 if (num > offslab_limit)
1990 /* Found something acceptable - save it away */
1992 cachep->gfporder = gfporder;
1993 left_over = remainder;
1996 * A VFS-reclaimable slab tends to have most allocations
1997 * as GFP_NOFS and we really don't want to have to be allocating
1998 * higher-order pages when we are unable to shrink dcache.
2000 if (flags & SLAB_RECLAIM_ACCOUNT)
2004 * Large number of objects is good, but very large slabs are
2005 * currently bad for the gfp()s.
2007 if (gfporder >= slab_break_gfp_order)
2011 * Acceptable internal fragmentation?
2013 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2019 static int setup_cpu_cache(struct kmem_cache *cachep)
2021 if (g_cpucache_up == FULL)
2022 return enable_cpucache(cachep);
2024 if (g_cpucache_up == NONE) {
2026 * Note: the first kmem_cache_create must create the cache
2027 * that's used by kmalloc(24), otherwise the creation of
2028 * further caches will BUG().
2030 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2033 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2034 * the first cache, then we need to set up all its list3s,
2035 * otherwise the creation of further caches will BUG().
2037 set_up_list3s(cachep, SIZE_AC);
2038 if (INDEX_AC == INDEX_L3)
2039 g_cpucache_up = PARTIAL_L3;
2041 g_cpucache_up = PARTIAL_AC;
2043 cachep->array[smp_processor_id()] =
2044 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2046 if (g_cpucache_up == PARTIAL_AC) {
2047 set_up_list3s(cachep, SIZE_L3);
2048 g_cpucache_up = PARTIAL_L3;
2051 for_each_online_node(node) {
2052 cachep->nodelists[node] =
2053 kmalloc_node(sizeof(struct kmem_list3),
2055 BUG_ON(!cachep->nodelists[node]);
2056 kmem_list3_init(cachep->nodelists[node]);
2060 cachep->nodelists[numa_node_id()]->next_reap =
2061 jiffies + REAPTIMEOUT_LIST3 +
2062 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2064 cpu_cache_get(cachep)->avail = 0;
2065 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2066 cpu_cache_get(cachep)->batchcount = 1;
2067 cpu_cache_get(cachep)->touched = 0;
2068 cachep->batchcount = 1;
2069 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2074 * kmem_cache_create - Create a cache.
2075 * @name: A string which is used in /proc/slabinfo to identify this cache.
2076 * @size: The size of objects to be created in this cache.
2077 * @align: The required alignment for the objects.
2078 * @flags: SLAB flags
2079 * @ctor: A constructor for the objects.
2080 * @dtor: A destructor for the objects.
2082 * Returns a ptr to the cache on success, NULL on failure.
2083 * Cannot be called within a int, but can be interrupted.
2084 * The @ctor is run when new pages are allocated by the cache
2085 * and the @dtor is run before the pages are handed back.
2087 * @name must be valid until the cache is destroyed. This implies that
2088 * the module calling this has to destroy the cache before getting unloaded.
2092 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2093 * to catch references to uninitialised memory.
2095 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2096 * for buffer overruns.
2098 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2099 * cacheline. This can be beneficial if you're counting cycles as closely
2103 kmem_cache_create (const char *name, size_t size, size_t align,
2104 unsigned long flags,
2105 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2106 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2108 size_t left_over, slab_size, ralign;
2109 struct kmem_cache *cachep = NULL, *pc;
2112 * Sanity checks... these are all serious usage bugs.
2114 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2115 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2116 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2122 * We use cache_chain_mutex to ensure a consistent view of
2123 * cpu_online_map as well. Please see cpuup_callback
2125 mutex_lock(&cache_chain_mutex);
2127 list_for_each_entry(pc, &cache_chain, next) {
2132 * This happens when the module gets unloaded and doesn't
2133 * destroy its slab cache and no-one else reuses the vmalloc
2134 * area of the module. Print a warning.
2136 res = probe_kernel_address(pc->name, tmp);
2138 printk("SLAB: cache with size %d has lost its name\n",
2143 if (!strcmp(pc->name, name)) {
2144 printk("kmem_cache_create: duplicate cache %s\n", name);
2151 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2152 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2153 /* No constructor, but inital state check requested */
2154 printk(KERN_ERR "%s: No con, but init state check "
2155 "requested - %s\n", __FUNCTION__, name);
2156 flags &= ~SLAB_DEBUG_INITIAL;
2160 * Enable redzoning and last user accounting, except for caches with
2161 * large objects, if the increased size would increase the object size
2162 * above the next power of two: caches with object sizes just above a
2163 * power of two have a significant amount of internal fragmentation.
2165 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2166 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2167 if (!(flags & SLAB_DESTROY_BY_RCU))
2168 flags |= SLAB_POISON;
2170 if (flags & SLAB_DESTROY_BY_RCU)
2171 BUG_ON(flags & SLAB_POISON);
2173 if (flags & SLAB_DESTROY_BY_RCU)
2177 * Always checks flags, a caller might be expecting debug support which
2180 BUG_ON(flags & ~CREATE_MASK);
2183 * Check that size is in terms of words. This is needed to avoid
2184 * unaligned accesses for some archs when redzoning is used, and makes
2185 * sure any on-slab bufctl's are also correctly aligned.
2187 if (size & (BYTES_PER_WORD - 1)) {
2188 size += (BYTES_PER_WORD - 1);
2189 size &= ~(BYTES_PER_WORD - 1);
2192 /* calculate the final buffer alignment: */
2194 /* 1) arch recommendation: can be overridden for debug */
2195 if (flags & SLAB_HWCACHE_ALIGN) {
2197 * Default alignment: as specified by the arch code. Except if
2198 * an object is really small, then squeeze multiple objects into
2201 ralign = cache_line_size();
2202 while (size <= ralign / 2)
2205 ralign = BYTES_PER_WORD;
2209 * Redzoning and user store require word alignment. Note this will be
2210 * overridden by architecture or caller mandated alignment if either
2211 * is greater than BYTES_PER_WORD.
2213 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2214 ralign = BYTES_PER_WORD;
2216 /* 2) arch mandated alignment */
2217 if (ralign < ARCH_SLAB_MINALIGN) {
2218 ralign = ARCH_SLAB_MINALIGN;
2220 /* 3) caller mandated alignment */
2221 if (ralign < align) {
2224 /* disable debug if necessary */
2225 if (ralign > BYTES_PER_WORD)
2226 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2232 /* Get cache's description obj. */
2233 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2238 cachep->obj_size = size;
2241 * Both debugging options require word-alignment which is calculated
2244 if (flags & SLAB_RED_ZONE) {
2245 /* add space for red zone words */
2246 cachep->obj_offset += BYTES_PER_WORD;
2247 size += 2 * BYTES_PER_WORD;
2249 if (flags & SLAB_STORE_USER) {
2250 /* user store requires one word storage behind the end of
2253 size += BYTES_PER_WORD;
2255 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2256 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2257 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2258 cachep->obj_offset += PAGE_SIZE - size;
2265 * Determine if the slab management is 'on' or 'off' slab.
2266 * (bootstrapping cannot cope with offslab caches so don't do
2269 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2271 * Size is large, assume best to place the slab management obj
2272 * off-slab (should allow better packing of objs).
2274 flags |= CFLGS_OFF_SLAB;
2276 size = ALIGN(size, align);
2278 left_over = calculate_slab_order(cachep, size, align, flags);
2281 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2282 kmem_cache_free(&cache_cache, cachep);
2286 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2287 + sizeof(struct slab), align);
2290 * If the slab has been placed off-slab, and we have enough space then
2291 * move it on-slab. This is at the expense of any extra colouring.
2293 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2294 flags &= ~CFLGS_OFF_SLAB;
2295 left_over -= slab_size;
2298 if (flags & CFLGS_OFF_SLAB) {
2299 /* really off slab. No need for manual alignment */
2301 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2304 cachep->colour_off = cache_line_size();
2305 /* Offset must be a multiple of the alignment. */
2306 if (cachep->colour_off < align)
2307 cachep->colour_off = align;
2308 cachep->colour = left_over / cachep->colour_off;
2309 cachep->slab_size = slab_size;
2310 cachep->flags = flags;
2311 cachep->gfpflags = 0;
2312 if (flags & SLAB_CACHE_DMA)
2313 cachep->gfpflags |= GFP_DMA;
2314 cachep->buffer_size = size;
2316 if (flags & CFLGS_OFF_SLAB) {
2317 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2319 * This is a possibility for one of the malloc_sizes caches.
2320 * But since we go off slab only for object size greater than
2321 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2322 * this should not happen at all.
2323 * But leave a BUG_ON for some lucky dude.
2325 BUG_ON(!cachep->slabp_cache);
2327 cachep->ctor = ctor;
2328 cachep->dtor = dtor;
2329 cachep->name = name;
2331 if (setup_cpu_cache(cachep)) {
2332 __kmem_cache_destroy(cachep);
2337 /* cache setup completed, link it into the list */
2338 list_add(&cachep->next, &cache_chain);
2340 if (!cachep && (flags & SLAB_PANIC))
2341 panic("kmem_cache_create(): failed to create slab `%s'\n",
2343 mutex_unlock(&cache_chain_mutex);
2346 EXPORT_SYMBOL(kmem_cache_create);
2349 static void check_irq_off(void)
2351 BUG_ON(!irqs_disabled());
2354 static void check_irq_on(void)
2356 BUG_ON(irqs_disabled());
2359 static void check_spinlock_acquired(struct kmem_cache *cachep)
2363 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2367 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2371 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2376 #define check_irq_off() do { } while(0)
2377 #define check_irq_on() do { } while(0)
2378 #define check_spinlock_acquired(x) do { } while(0)
2379 #define check_spinlock_acquired_node(x, y) do { } while(0)
2382 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2383 struct array_cache *ac,
2384 int force, int node);
2386 static void do_drain(void *arg)
2388 struct kmem_cache *cachep = arg;
2389 struct array_cache *ac;
2390 int node = numa_node_id();
2393 ac = cpu_cache_get(cachep);
2394 spin_lock(&cachep->nodelists[node]->list_lock);
2395 free_block(cachep, ac->entry, ac->avail, node);
2396 spin_unlock(&cachep->nodelists[node]->list_lock);
2400 static void drain_cpu_caches(struct kmem_cache *cachep)
2402 struct kmem_list3 *l3;
2405 on_each_cpu(do_drain, cachep, 1, 1);
2407 for_each_online_node(node) {
2408 l3 = cachep->nodelists[node];
2409 if (l3 && l3->alien)
2410 drain_alien_cache(cachep, l3->alien);
2413 for_each_online_node(node) {
2414 l3 = cachep->nodelists[node];
2416 drain_array(cachep, l3, l3->shared, 1, node);
2421 * Remove slabs from the list of free slabs.
2422 * Specify the number of slabs to drain in tofree.
2424 * Returns the actual number of slabs released.
2426 static int drain_freelist(struct kmem_cache *cache,
2427 struct kmem_list3 *l3, int tofree)
2429 struct list_head *p;
2434 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2436 spin_lock_irq(&l3->list_lock);
2437 p = l3->slabs_free.prev;
2438 if (p == &l3->slabs_free) {
2439 spin_unlock_irq(&l3->list_lock);
2443 slabp = list_entry(p, struct slab, list);
2445 BUG_ON(slabp->inuse);
2447 list_del(&slabp->list);
2449 * Safe to drop the lock. The slab is no longer linked
2452 l3->free_objects -= cache->num;
2453 spin_unlock_irq(&l3->list_lock);
2454 slab_destroy(cache, slabp);
2461 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2462 static int __cache_shrink(struct kmem_cache *cachep)
2465 struct kmem_list3 *l3;
2467 drain_cpu_caches(cachep);
2470 for_each_online_node(i) {
2471 l3 = cachep->nodelists[i];
2475 drain_freelist(cachep, l3, l3->free_objects);
2477 ret += !list_empty(&l3->slabs_full) ||
2478 !list_empty(&l3->slabs_partial);
2480 return (ret ? 1 : 0);
2484 * kmem_cache_shrink - Shrink a cache.
2485 * @cachep: The cache to shrink.
2487 * Releases as many slabs as possible for a cache.
2488 * To help debugging, a zero exit status indicates all slabs were released.
2490 int kmem_cache_shrink(struct kmem_cache *cachep)
2493 BUG_ON(!cachep || in_interrupt());
2495 mutex_lock(&cache_chain_mutex);
2496 ret = __cache_shrink(cachep);
2497 mutex_unlock(&cache_chain_mutex);
2500 EXPORT_SYMBOL(kmem_cache_shrink);
2503 * kmem_cache_destroy - delete a cache
2504 * @cachep: the cache to destroy
2506 * Remove a struct kmem_cache object from the slab cache.
2508 * It is expected this function will be called by a module when it is
2509 * unloaded. This will remove the cache completely, and avoid a duplicate
2510 * cache being allocated each time a module is loaded and unloaded, if the
2511 * module doesn't have persistent in-kernel storage across loads and unloads.
2513 * The cache must be empty before calling this function.
2515 * The caller must guarantee that noone will allocate memory from the cache
2516 * during the kmem_cache_destroy().
2518 void kmem_cache_destroy(struct kmem_cache *cachep)
2520 BUG_ON(!cachep || in_interrupt());
2522 /* Find the cache in the chain of caches. */
2523 mutex_lock(&cache_chain_mutex);
2525 * the chain is never empty, cache_cache is never destroyed
2527 list_del(&cachep->next);
2528 if (__cache_shrink(cachep)) {
2529 slab_error(cachep, "Can't free all objects");
2530 list_add(&cachep->next, &cache_chain);
2531 mutex_unlock(&cache_chain_mutex);
2535 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2538 __kmem_cache_destroy(cachep);
2539 mutex_unlock(&cache_chain_mutex);
2541 EXPORT_SYMBOL(kmem_cache_destroy);
2544 * Get the memory for a slab management obj.
2545 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2546 * always come from malloc_sizes caches. The slab descriptor cannot
2547 * come from the same cache which is getting created because,
2548 * when we are searching for an appropriate cache for these
2549 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2550 * If we are creating a malloc_sizes cache here it would not be visible to
2551 * kmem_find_general_cachep till the initialization is complete.
2552 * Hence we cannot have slabp_cache same as the original cache.
2554 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2555 int colour_off, gfp_t local_flags,
2560 if (OFF_SLAB(cachep)) {
2561 /* Slab management obj is off-slab. */
2562 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2563 local_flags & ~GFP_THISNODE, nodeid);
2567 slabp = objp + colour_off;
2568 colour_off += cachep->slab_size;
2571 slabp->colouroff = colour_off;
2572 slabp->s_mem = objp + colour_off;
2573 slabp->nodeid = nodeid;
2577 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2579 return (kmem_bufctl_t *) (slabp + 1);
2582 static void cache_init_objs(struct kmem_cache *cachep,
2583 struct slab *slabp, unsigned long ctor_flags)
2587 for (i = 0; i < cachep->num; i++) {
2588 void *objp = index_to_obj(cachep, slabp, i);
2590 /* need to poison the objs? */
2591 if (cachep->flags & SLAB_POISON)
2592 poison_obj(cachep, objp, POISON_FREE);
2593 if (cachep->flags & SLAB_STORE_USER)
2594 *dbg_userword(cachep, objp) = NULL;
2596 if (cachep->flags & SLAB_RED_ZONE) {
2597 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2598 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2601 * Constructors are not allowed to allocate memory from the same
2602 * cache which they are a constructor for. Otherwise, deadlock.
2603 * They must also be threaded.
2605 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2606 cachep->ctor(objp + obj_offset(cachep), cachep,
2609 if (cachep->flags & SLAB_RED_ZONE) {
2610 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2611 slab_error(cachep, "constructor overwrote the"
2612 " end of an object");
2613 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2614 slab_error(cachep, "constructor overwrote the"
2615 " start of an object");
2617 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2618 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2619 kernel_map_pages(virt_to_page(objp),
2620 cachep->buffer_size / PAGE_SIZE, 0);
2623 cachep->ctor(objp, cachep, ctor_flags);
2625 slab_bufctl(slabp)[i] = i + 1;
2627 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2631 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2633 if (flags & GFP_DMA)
2634 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2636 BUG_ON(cachep->gfpflags & GFP_DMA);
2639 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2642 void *objp = index_to_obj(cachep, slabp, slabp->free);
2646 next = slab_bufctl(slabp)[slabp->free];
2648 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2649 WARN_ON(slabp->nodeid != nodeid);
2656 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2657 void *objp, int nodeid)
2659 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2662 /* Verify that the slab belongs to the intended node */
2663 WARN_ON(slabp->nodeid != nodeid);
2665 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2666 printk(KERN_ERR "slab: double free detected in cache "
2667 "'%s', objp %p\n", cachep->name, objp);
2671 slab_bufctl(slabp)[objnr] = slabp->free;
2672 slabp->free = objnr;
2677 * Map pages beginning at addr to the given cache and slab. This is required
2678 * for the slab allocator to be able to lookup the cache and slab of a
2679 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2681 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2687 page = virt_to_page(addr);
2690 if (likely(!PageCompound(page)))
2691 nr_pages <<= cache->gfporder;
2694 page_set_cache(page, cache);
2695 page_set_slab(page, slab);
2697 } while (--nr_pages);
2701 * Grow (by 1) the number of slabs within a cache. This is called by
2702 * kmem_cache_alloc() when there are no active objs left in a cache.
2704 static int cache_grow(struct kmem_cache *cachep,
2705 gfp_t flags, int nodeid, void *objp)
2710 unsigned long ctor_flags;
2711 struct kmem_list3 *l3;
2714 * Be lazy and only check for valid flags here, keeping it out of the
2715 * critical path in kmem_cache_alloc().
2717 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW));
2718 if (flags & __GFP_NO_GROW)
2721 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2722 local_flags = (flags & GFP_LEVEL_MASK);
2723 if (!(local_flags & __GFP_WAIT))
2725 * Not allowed to sleep. Need to tell a constructor about
2726 * this - it might need to know...
2728 ctor_flags |= SLAB_CTOR_ATOMIC;
2730 /* Take the l3 list lock to change the colour_next on this node */
2732 l3 = cachep->nodelists[nodeid];
2733 spin_lock(&l3->list_lock);
2735 /* Get colour for the slab, and cal the next value. */
2736 offset = l3->colour_next;
2738 if (l3->colour_next >= cachep->colour)
2739 l3->colour_next = 0;
2740 spin_unlock(&l3->list_lock);
2742 offset *= cachep->colour_off;
2744 if (local_flags & __GFP_WAIT)
2748 * The test for missing atomic flag is performed here, rather than
2749 * the more obvious place, simply to reduce the critical path length
2750 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2751 * will eventually be caught here (where it matters).
2753 kmem_flagcheck(cachep, flags);
2756 * Get mem for the objs. Attempt to allocate a physical page from
2760 objp = kmem_getpages(cachep, flags, nodeid);
2764 /* Get slab management. */
2765 slabp = alloc_slabmgmt(cachep, objp, offset,
2766 local_flags & ~GFP_THISNODE, nodeid);
2770 slabp->nodeid = nodeid;
2771 slab_map_pages(cachep, slabp, objp);
2773 cache_init_objs(cachep, slabp, ctor_flags);
2775 if (local_flags & __GFP_WAIT)
2776 local_irq_disable();
2778 spin_lock(&l3->list_lock);
2780 /* Make slab active. */
2781 list_add_tail(&slabp->list, &(l3->slabs_free));
2782 STATS_INC_GROWN(cachep);
2783 l3->free_objects += cachep->num;
2784 spin_unlock(&l3->list_lock);
2787 kmem_freepages(cachep, objp);
2789 if (local_flags & __GFP_WAIT)
2790 local_irq_disable();
2797 * Perform extra freeing checks:
2798 * - detect bad pointers.
2799 * - POISON/RED_ZONE checking
2800 * - destructor calls, for caches with POISON+dtor
2802 static void kfree_debugcheck(const void *objp)
2806 if (!virt_addr_valid(objp)) {
2807 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2808 (unsigned long)objp);
2811 page = virt_to_page(objp);
2812 if (!PageSlab(page)) {
2813 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2814 (unsigned long)objp);
2819 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2821 unsigned long redzone1, redzone2;
2823 redzone1 = *dbg_redzone1(cache, obj);
2824 redzone2 = *dbg_redzone2(cache, obj);
2829 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2832 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2833 slab_error(cache, "double free detected");
2835 slab_error(cache, "memory outside object was overwritten");
2837 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2838 obj, redzone1, redzone2);
2841 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2848 objp -= obj_offset(cachep);
2849 kfree_debugcheck(objp);
2850 page = virt_to_page(objp);
2852 slabp = page_get_slab(page);
2854 if (cachep->flags & SLAB_RED_ZONE) {
2855 verify_redzone_free(cachep, objp);
2856 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2857 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2859 if (cachep->flags & SLAB_STORE_USER)
2860 *dbg_userword(cachep, objp) = caller;
2862 objnr = obj_to_index(cachep, slabp, objp);
2864 BUG_ON(objnr >= cachep->num);
2865 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2867 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2869 * Need to call the slab's constructor so the caller can
2870 * perform a verify of its state (debugging). Called without
2871 * the cache-lock held.
2873 cachep->ctor(objp + obj_offset(cachep),
2874 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2876 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2877 /* we want to cache poison the object,
2878 * call the destruction callback
2880 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2882 #ifdef CONFIG_DEBUG_SLAB_LEAK
2883 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2885 if (cachep->flags & SLAB_POISON) {
2886 #ifdef CONFIG_DEBUG_PAGEALLOC
2887 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2888 store_stackinfo(cachep, objp, (unsigned long)caller);
2889 kernel_map_pages(virt_to_page(objp),
2890 cachep->buffer_size / PAGE_SIZE, 0);
2892 poison_obj(cachep, objp, POISON_FREE);
2895 poison_obj(cachep, objp, POISON_FREE);
2901 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2906 /* Check slab's freelist to see if this obj is there. */
2907 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2909 if (entries > cachep->num || i >= cachep->num)
2912 if (entries != cachep->num - slabp->inuse) {
2914 printk(KERN_ERR "slab: Internal list corruption detected in "
2915 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2916 cachep->name, cachep->num, slabp, slabp->inuse);
2918 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2921 printk("\n%03x:", i);
2922 printk(" %02x", ((unsigned char *)slabp)[i]);
2929 #define kfree_debugcheck(x) do { } while(0)
2930 #define cache_free_debugcheck(x,objp,z) (objp)
2931 #define check_slabp(x,y) do { } while(0)
2934 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2937 struct kmem_list3 *l3;
2938 struct array_cache *ac;
2941 node = numa_node_id();
2944 ac = cpu_cache_get(cachep);
2946 batchcount = ac->batchcount;
2947 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2949 * If there was little recent activity on this cache, then
2950 * perform only a partial refill. Otherwise we could generate
2953 batchcount = BATCHREFILL_LIMIT;
2955 l3 = cachep->nodelists[node];
2957 BUG_ON(ac->avail > 0 || !l3);
2958 spin_lock(&l3->list_lock);
2960 /* See if we can refill from the shared array */
2961 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2964 while (batchcount > 0) {
2965 struct list_head *entry;
2967 /* Get slab alloc is to come from. */
2968 entry = l3->slabs_partial.next;
2969 if (entry == &l3->slabs_partial) {
2970 l3->free_touched = 1;
2971 entry = l3->slabs_free.next;
2972 if (entry == &l3->slabs_free)
2976 slabp = list_entry(entry, struct slab, list);
2977 check_slabp(cachep, slabp);
2978 check_spinlock_acquired(cachep);
2979 while (slabp->inuse < cachep->num && batchcount--) {
2980 STATS_INC_ALLOCED(cachep);
2981 STATS_INC_ACTIVE(cachep);
2982 STATS_SET_HIGH(cachep);
2984 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2987 check_slabp(cachep, slabp);
2989 /* move slabp to correct slabp list: */
2990 list_del(&slabp->list);
2991 if (slabp->free == BUFCTL_END)
2992 list_add(&slabp->list, &l3->slabs_full);
2994 list_add(&slabp->list, &l3->slabs_partial);
2998 l3->free_objects -= ac->avail;
3000 spin_unlock(&l3->list_lock);
3002 if (unlikely(!ac->avail)) {
3004 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3006 /* cache_grow can reenable interrupts, then ac could change. */
3007 ac = cpu_cache_get(cachep);
3008 if (!x && ac->avail == 0) /* no objects in sight? abort */
3011 if (!ac->avail) /* objects refilled by interrupt? */
3015 return ac->entry[--ac->avail];
3018 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3021 might_sleep_if(flags & __GFP_WAIT);
3023 kmem_flagcheck(cachep, flags);
3028 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3029 gfp_t flags, void *objp, void *caller)
3033 if (cachep->flags & SLAB_POISON) {
3034 #ifdef CONFIG_DEBUG_PAGEALLOC
3035 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3036 kernel_map_pages(virt_to_page(objp),
3037 cachep->buffer_size / PAGE_SIZE, 1);
3039 check_poison_obj(cachep, objp);
3041 check_poison_obj(cachep, objp);
3043 poison_obj(cachep, objp, POISON_INUSE);
3045 if (cachep->flags & SLAB_STORE_USER)
3046 *dbg_userword(cachep, objp) = caller;
3048 if (cachep->flags & SLAB_RED_ZONE) {
3049 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3050 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3051 slab_error(cachep, "double free, or memory outside"
3052 " object was overwritten");
3054 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3055 objp, *dbg_redzone1(cachep, objp),
3056 *dbg_redzone2(cachep, objp));
3058 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3059 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3061 #ifdef CONFIG_DEBUG_SLAB_LEAK
3066 slabp = page_get_slab(virt_to_page(objp));
3067 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3068 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3071 objp += obj_offset(cachep);
3072 if (cachep->ctor && cachep->flags & SLAB_POISON) {
3073 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
3075 if (!(flags & __GFP_WAIT))
3076 ctor_flags |= SLAB_CTOR_ATOMIC;
3078 cachep->ctor(objp, cachep, ctor_flags);
3080 #if ARCH_SLAB_MINALIGN
3081 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3082 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3083 objp, ARCH_SLAB_MINALIGN);
3089 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3092 #ifdef CONFIG_FAILSLAB
3094 static struct failslab_attr {
3096 struct fault_attr attr;
3098 u32 ignore_gfp_wait;
3099 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3100 struct dentry *ignore_gfp_wait_file;
3104 .attr = FAULT_ATTR_INITIALIZER,
3105 .ignore_gfp_wait = 1,
3108 static int __init setup_failslab(char *str)
3110 return setup_fault_attr(&failslab.attr, str);
3112 __setup("failslab=", setup_failslab);
3114 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3116 if (cachep == &cache_cache)
3118 if (flags & __GFP_NOFAIL)
3120 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3123 return should_fail(&failslab.attr, obj_size(cachep));
3126 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3128 static int __init failslab_debugfs(void)
3130 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3134 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3137 dir = failslab.attr.dentries.dir;
3139 failslab.ignore_gfp_wait_file =
3140 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3141 &failslab.ignore_gfp_wait);
3143 if (!failslab.ignore_gfp_wait_file) {
3145 debugfs_remove(failslab.ignore_gfp_wait_file);
3146 cleanup_fault_attr_dentries(&failslab.attr);
3152 late_initcall(failslab_debugfs);
3154 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3156 #else /* CONFIG_FAILSLAB */
3158 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3163 #endif /* CONFIG_FAILSLAB */
3165 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3168 struct array_cache *ac;
3172 if (should_failslab(cachep, flags))
3175 ac = cpu_cache_get(cachep);
3176 if (likely(ac->avail)) {
3177 STATS_INC_ALLOCHIT(cachep);
3179 objp = ac->entry[--ac->avail];
3181 STATS_INC_ALLOCMISS(cachep);
3182 objp = cache_alloc_refill(cachep, flags);
3187 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3188 gfp_t flags, void *caller)
3190 unsigned long save_flags;
3193 cache_alloc_debugcheck_before(cachep, flags);
3195 local_irq_save(save_flags);
3197 if (unlikely(NUMA_BUILD &&
3198 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
3199 objp = alternate_node_alloc(cachep, flags);
3202 objp = ____cache_alloc(cachep, flags);
3204 * We may just have run out of memory on the local node.
3205 * ____cache_alloc_node() knows how to locate memory on other nodes
3207 if (NUMA_BUILD && !objp)
3208 objp = ____cache_alloc_node(cachep, flags, numa_node_id());
3209 local_irq_restore(save_flags);
3210 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3218 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3220 * If we are in_interrupt, then process context, including cpusets and
3221 * mempolicy, may not apply and should not be used for allocation policy.
3223 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3225 int nid_alloc, nid_here;
3227 if (in_interrupt() || (flags & __GFP_THISNODE))
3229 nid_alloc = nid_here = numa_node_id();
3230 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3231 nid_alloc = cpuset_mem_spread_node();
3232 else if (current->mempolicy)
3233 nid_alloc = slab_node(current->mempolicy);
3234 if (nid_alloc != nid_here)
3235 return ____cache_alloc_node(cachep, flags, nid_alloc);
3240 * Fallback function if there was no memory available and no objects on a
3241 * certain node and fall back is permitted. First we scan all the
3242 * available nodelists for available objects. If that fails then we
3243 * perform an allocation without specifying a node. This allows the page
3244 * allocator to do its reclaim / fallback magic. We then insert the
3245 * slab into the proper nodelist and then allocate from it.
3247 void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3249 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3250 ->node_zonelists[gfp_zone(flags)];
3257 * Look through allowed nodes for objects available
3258 * from existing per node queues.
3260 for (z = zonelist->zones; *z && !obj; z++) {
3261 nid = zone_to_nid(*z);
3263 if (cpuset_zone_allowed(*z, flags | __GFP_HARDWALL) &&
3264 cache->nodelists[nid] &&
3265 cache->nodelists[nid]->free_objects)
3266 obj = ____cache_alloc_node(cache,
3267 flags | GFP_THISNODE, nid);
3272 * This allocation will be performed within the constraints
3273 * of the current cpuset / memory policy requirements.
3274 * We may trigger various forms of reclaim on the allowed
3275 * set and go into memory reserves if necessary.
3277 obj = kmem_getpages(cache, flags, -1);
3280 * Insert into the appropriate per node queues
3282 nid = page_to_nid(virt_to_page(obj));
3283 if (cache_grow(cache, flags, nid, obj)) {
3284 obj = ____cache_alloc_node(cache,
3285 flags | GFP_THISNODE, nid);
3288 * Another processor may allocate the
3289 * objects in the slab since we are
3290 * not holding any locks.
3294 kmem_freepages(cache, obj);
3303 * A interface to enable slab creation on nodeid
3305 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3308 struct list_head *entry;
3310 struct kmem_list3 *l3;
3314 l3 = cachep->nodelists[nodeid];
3319 spin_lock(&l3->list_lock);
3320 entry = l3->slabs_partial.next;
3321 if (entry == &l3->slabs_partial) {
3322 l3->free_touched = 1;
3323 entry = l3->slabs_free.next;
3324 if (entry == &l3->slabs_free)
3328 slabp = list_entry(entry, struct slab, list);
3329 check_spinlock_acquired_node(cachep, nodeid);
3330 check_slabp(cachep, slabp);
3332 STATS_INC_NODEALLOCS(cachep);
3333 STATS_INC_ACTIVE(cachep);
3334 STATS_SET_HIGH(cachep);
3336 BUG_ON(slabp->inuse == cachep->num);
3338 obj = slab_get_obj(cachep, slabp, nodeid);
3339 check_slabp(cachep, slabp);
3341 /* move slabp to correct slabp list: */
3342 list_del(&slabp->list);
3344 if (slabp->free == BUFCTL_END)
3345 list_add(&slabp->list, &l3->slabs_full);
3347 list_add(&slabp->list, &l3->slabs_partial);
3349 spin_unlock(&l3->list_lock);
3353 spin_unlock(&l3->list_lock);
3354 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3358 if (!(flags & __GFP_THISNODE))
3359 /* Unable to grow the cache. Fall back to other nodes. */
3360 return fallback_alloc(cachep, flags);
3370 * Caller needs to acquire correct kmem_list's list_lock
3372 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3376 struct kmem_list3 *l3;
3378 for (i = 0; i < nr_objects; i++) {
3379 void *objp = objpp[i];
3382 slabp = virt_to_slab(objp);
3383 l3 = cachep->nodelists[node];
3384 list_del(&slabp->list);
3385 check_spinlock_acquired_node(cachep, node);
3386 check_slabp(cachep, slabp);
3387 slab_put_obj(cachep, slabp, objp, node);
3388 STATS_DEC_ACTIVE(cachep);
3390 check_slabp(cachep, slabp);
3392 /* fixup slab chains */
3393 if (slabp->inuse == 0) {
3394 if (l3->free_objects > l3->free_limit) {
3395 l3->free_objects -= cachep->num;
3396 /* No need to drop any previously held
3397 * lock here, even if we have a off-slab slab
3398 * descriptor it is guaranteed to come from
3399 * a different cache, refer to comments before
3402 slab_destroy(cachep, slabp);
3404 list_add(&slabp->list, &l3->slabs_free);
3407 /* Unconditionally move a slab to the end of the
3408 * partial list on free - maximum time for the
3409 * other objects to be freed, too.
3411 list_add_tail(&slabp->list, &l3->slabs_partial);
3416 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3419 struct kmem_list3 *l3;
3420 int node = numa_node_id();
3422 batchcount = ac->batchcount;
3424 BUG_ON(!batchcount || batchcount > ac->avail);
3427 l3 = cachep->nodelists[node];
3428 spin_lock(&l3->list_lock);
3430 struct array_cache *shared_array = l3->shared;
3431 int max = shared_array->limit - shared_array->avail;
3433 if (batchcount > max)
3435 memcpy(&(shared_array->entry[shared_array->avail]),
3436 ac->entry, sizeof(void *) * batchcount);
3437 shared_array->avail += batchcount;
3442 free_block(cachep, ac->entry, batchcount, node);
3447 struct list_head *p;
3449 p = l3->slabs_free.next;
3450 while (p != &(l3->slabs_free)) {
3453 slabp = list_entry(p, struct slab, list);
3454 BUG_ON(slabp->inuse);
3459 STATS_SET_FREEABLE(cachep, i);
3462 spin_unlock(&l3->list_lock);
3463 ac->avail -= batchcount;
3464 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3468 * Release an obj back to its cache. If the obj has a constructed state, it must
3469 * be in this state _before_ it is released. Called with disabled ints.
3471 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3473 struct array_cache *ac = cpu_cache_get(cachep);
3476 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3478 if (cache_free_alien(cachep, objp))
3481 if (likely(ac->avail < ac->limit)) {
3482 STATS_INC_FREEHIT(cachep);
3483 ac->entry[ac->avail++] = objp;
3486 STATS_INC_FREEMISS(cachep);
3487 cache_flusharray(cachep, ac);
3488 ac->entry[ac->avail++] = objp;
3493 * kmem_cache_alloc - Allocate an object
3494 * @cachep: The cache to allocate from.
3495 * @flags: See kmalloc().
3497 * Allocate an object from this cache. The flags are only relevant
3498 * if the cache has no available objects.
3500 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3502 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3504 EXPORT_SYMBOL(kmem_cache_alloc);
3507 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3508 * @cache: The cache to allocate from.
3509 * @flags: See kmalloc().
3511 * Allocate an object from this cache and set the allocated memory to zero.
3512 * The flags are only relevant if the cache has no available objects.
3514 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3516 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3518 memset(ret, 0, obj_size(cache));
3521 EXPORT_SYMBOL(kmem_cache_zalloc);
3524 * kmem_ptr_validate - check if an untrusted pointer might
3526 * @cachep: the cache we're checking against
3527 * @ptr: pointer to validate
3529 * This verifies that the untrusted pointer looks sane:
3530 * it is _not_ a guarantee that the pointer is actually
3531 * part of the slab cache in question, but it at least
3532 * validates that the pointer can be dereferenced and
3533 * looks half-way sane.
3535 * Currently only used for dentry validation.
3537 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3539 unsigned long addr = (unsigned long)ptr;
3540 unsigned long min_addr = PAGE_OFFSET;
3541 unsigned long align_mask = BYTES_PER_WORD - 1;
3542 unsigned long size = cachep->buffer_size;
3545 if (unlikely(addr < min_addr))
3547 if (unlikely(addr > (unsigned long)high_memory - size))
3549 if (unlikely(addr & align_mask))
3551 if (unlikely(!kern_addr_valid(addr)))
3553 if (unlikely(!kern_addr_valid(addr + size - 1)))
3555 page = virt_to_page(ptr);
3556 if (unlikely(!PageSlab(page)))
3558 if (unlikely(page_get_cache(page) != cachep))
3567 * kmem_cache_alloc_node - Allocate an object on the specified node
3568 * @cachep: The cache to allocate from.
3569 * @flags: See kmalloc().
3570 * @nodeid: node number of the target node.
3572 * Identical to kmem_cache_alloc but it will allocate memory on the given
3573 * node, which can improve the performance for cpu bound structures.
3575 * Fallback to other node is possible if __GFP_THISNODE is not set.
3577 static __always_inline void *
3578 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3579 int nodeid, void *caller)
3581 unsigned long save_flags;
3584 cache_alloc_debugcheck_before(cachep, flags);
3585 local_irq_save(save_flags);
3587 if (unlikely(nodeid == -1))
3588 nodeid = numa_node_id();
3590 if (likely(cachep->nodelists[nodeid])) {
3591 if (nodeid == numa_node_id()) {
3593 * Use the locally cached objects if possible.
3594 * However ____cache_alloc does not allow fallback
3595 * to other nodes. It may fail while we still have
3596 * objects on other nodes available.
3598 ptr = ____cache_alloc(cachep, flags);
3601 /* ___cache_alloc_node can fall back to other nodes */
3602 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3605 /* Node not bootstrapped yet */
3606 if (!(flags & __GFP_THISNODE))
3607 ptr = fallback_alloc(cachep, flags);
3610 local_irq_restore(save_flags);
3611 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3616 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3618 return __cache_alloc_node(cachep, flags, nodeid,
3619 __builtin_return_address(0));
3621 EXPORT_SYMBOL(kmem_cache_alloc_node);
3623 static __always_inline void *
3624 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3626 struct kmem_cache *cachep;
3628 cachep = kmem_find_general_cachep(size, flags);
3629 if (unlikely(cachep == NULL))
3631 return kmem_cache_alloc_node(cachep, flags, node);
3634 #ifdef CONFIG_DEBUG_SLAB
3635 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3637 return __do_kmalloc_node(size, flags, node,
3638 __builtin_return_address(0));
3640 EXPORT_SYMBOL(__kmalloc_node);
3642 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3643 int node, void *caller)
3645 return __do_kmalloc_node(size, flags, node, caller);
3647 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3649 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3651 return __do_kmalloc_node(size, flags, node, NULL);
3653 EXPORT_SYMBOL(__kmalloc_node);
3654 #endif /* CONFIG_DEBUG_SLAB */
3655 #endif /* CONFIG_NUMA */
3658 * __do_kmalloc - allocate memory
3659 * @size: how many bytes of memory are required.
3660 * @flags: the type of memory to allocate (see kmalloc).
3661 * @caller: function caller for debug tracking of the caller
3663 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3666 struct kmem_cache *cachep;
3668 /* If you want to save a few bytes .text space: replace
3670 * Then kmalloc uses the uninlined functions instead of the inline
3673 cachep = __find_general_cachep(size, flags);
3674 if (unlikely(cachep == NULL))
3676 return __cache_alloc(cachep, flags, caller);
3680 #ifdef CONFIG_DEBUG_SLAB
3681 void *__kmalloc(size_t size, gfp_t flags)
3683 return __do_kmalloc(size, flags, __builtin_return_address(0));
3685 EXPORT_SYMBOL(__kmalloc);
3687 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3689 return __do_kmalloc(size, flags, caller);
3691 EXPORT_SYMBOL(__kmalloc_track_caller);
3694 void *__kmalloc(size_t size, gfp_t flags)
3696 return __do_kmalloc(size, flags, NULL);
3698 EXPORT_SYMBOL(__kmalloc);
3702 * kmem_cache_free - Deallocate an object
3703 * @cachep: The cache the allocation was from.
3704 * @objp: The previously allocated object.
3706 * Free an object which was previously allocated from this
3709 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3711 unsigned long flags;
3713 BUG_ON(virt_to_cache(objp) != cachep);
3715 local_irq_save(flags);
3716 __cache_free(cachep, objp);
3717 local_irq_restore(flags);
3719 EXPORT_SYMBOL(kmem_cache_free);
3722 * kfree - free previously allocated memory
3723 * @objp: pointer returned by kmalloc.
3725 * If @objp is NULL, no operation is performed.
3727 * Don't free memory not originally allocated by kmalloc()
3728 * or you will run into trouble.
3730 void kfree(const void *objp)
3732 struct kmem_cache *c;
3733 unsigned long flags;
3735 if (unlikely(!objp))
3737 local_irq_save(flags);
3738 kfree_debugcheck(objp);
3739 c = virt_to_cache(objp);
3740 debug_check_no_locks_freed(objp, obj_size(c));
3741 __cache_free(c, (void *)objp);
3742 local_irq_restore(flags);
3744 EXPORT_SYMBOL(kfree);
3746 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3748 return obj_size(cachep);
3750 EXPORT_SYMBOL(kmem_cache_size);
3752 const char *kmem_cache_name(struct kmem_cache *cachep)
3754 return cachep->name;
3756 EXPORT_SYMBOL_GPL(kmem_cache_name);
3759 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3761 static int alloc_kmemlist(struct kmem_cache *cachep)
3764 struct kmem_list3 *l3;
3765 struct array_cache *new_shared;
3766 struct array_cache **new_alien = NULL;
3768 for_each_online_node(node) {
3770 if (use_alien_caches) {
3771 new_alien = alloc_alien_cache(node, cachep->limit);
3776 new_shared = alloc_arraycache(node,
3777 cachep->shared*cachep->batchcount,
3780 free_alien_cache(new_alien);
3784 l3 = cachep->nodelists[node];
3786 struct array_cache *shared = l3->shared;
3788 spin_lock_irq(&l3->list_lock);
3791 free_block(cachep, shared->entry,
3792 shared->avail, node);
3794 l3->shared = new_shared;
3796 l3->alien = new_alien;
3799 l3->free_limit = (1 + nr_cpus_node(node)) *
3800 cachep->batchcount + cachep->num;
3801 spin_unlock_irq(&l3->list_lock);
3803 free_alien_cache(new_alien);
3806 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3808 free_alien_cache(new_alien);
3813 kmem_list3_init(l3);
3814 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3815 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3816 l3->shared = new_shared;
3817 l3->alien = new_alien;
3818 l3->free_limit = (1 + nr_cpus_node(node)) *
3819 cachep->batchcount + cachep->num;
3820 cachep->nodelists[node] = l3;
3825 if (!cachep->next.next) {
3826 /* Cache is not active yet. Roll back what we did */
3829 if (cachep->nodelists[node]) {
3830 l3 = cachep->nodelists[node];
3833 free_alien_cache(l3->alien);
3835 cachep->nodelists[node] = NULL;
3843 struct ccupdate_struct {
3844 struct kmem_cache *cachep;
3845 struct array_cache *new[NR_CPUS];
3848 static void do_ccupdate_local(void *info)
3850 struct ccupdate_struct *new = info;
3851 struct array_cache *old;
3854 old = cpu_cache_get(new->cachep);
3856 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3857 new->new[smp_processor_id()] = old;
3860 /* Always called with the cache_chain_mutex held */
3861 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3862 int batchcount, int shared)
3864 struct ccupdate_struct *new;
3867 new = kzalloc(sizeof(*new), GFP_KERNEL);
3871 for_each_online_cpu(i) {
3872 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3875 for (i--; i >= 0; i--)
3881 new->cachep = cachep;
3883 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3886 cachep->batchcount = batchcount;
3887 cachep->limit = limit;
3888 cachep->shared = shared;
3890 for_each_online_cpu(i) {
3891 struct array_cache *ccold = new->new[i];
3894 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3895 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3896 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3900 return alloc_kmemlist(cachep);
3903 /* Called with cache_chain_mutex held always */
3904 static int enable_cpucache(struct kmem_cache *cachep)
3910 * The head array serves three purposes:
3911 * - create a LIFO ordering, i.e. return objects that are cache-warm
3912 * - reduce the number of spinlock operations.
3913 * - reduce the number of linked list operations on the slab and
3914 * bufctl chains: array operations are cheaper.
3915 * The numbers are guessed, we should auto-tune as described by
3918 if (cachep->buffer_size > 131072)
3920 else if (cachep->buffer_size > PAGE_SIZE)
3922 else if (cachep->buffer_size > 1024)
3924 else if (cachep->buffer_size > 256)
3930 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3931 * allocation behaviour: Most allocs on one cpu, most free operations
3932 * on another cpu. For these cases, an efficient object passing between
3933 * cpus is necessary. This is provided by a shared array. The array
3934 * replaces Bonwick's magazine layer.
3935 * On uniprocessor, it's functionally equivalent (but less efficient)
3936 * to a larger limit. Thus disabled by default.
3940 if (cachep->buffer_size <= PAGE_SIZE)
3946 * With debugging enabled, large batchcount lead to excessively long
3947 * periods with disabled local interrupts. Limit the batchcount
3952 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3954 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3955 cachep->name, -err);
3960 * Drain an array if it contains any elements taking the l3 lock only if
3961 * necessary. Note that the l3 listlock also protects the array_cache
3962 * if drain_array() is used on the shared array.
3964 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3965 struct array_cache *ac, int force, int node)
3969 if (!ac || !ac->avail)
3971 if (ac->touched && !force) {
3974 spin_lock_irq(&l3->list_lock);
3976 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3977 if (tofree > ac->avail)
3978 tofree = (ac->avail + 1) / 2;
3979 free_block(cachep, ac->entry, tofree, node);
3980 ac->avail -= tofree;
3981 memmove(ac->entry, &(ac->entry[tofree]),
3982 sizeof(void *) * ac->avail);
3984 spin_unlock_irq(&l3->list_lock);
3989 * cache_reap - Reclaim memory from caches.
3990 * @unused: unused parameter
3992 * Called from workqueue/eventd every few seconds.
3994 * - clear the per-cpu caches for this CPU.
3995 * - return freeable pages to the main free memory pool.
3997 * If we cannot acquire the cache chain mutex then just give up - we'll try
3998 * again on the next iteration.
4000 static void cache_reap(struct work_struct *unused)
4002 struct kmem_cache *searchp;
4003 struct kmem_list3 *l3;
4004 int node = numa_node_id();
4006 if (!mutex_trylock(&cache_chain_mutex)) {
4007 /* Give up. Setup the next iteration. */
4008 schedule_delayed_work(&__get_cpu_var(reap_work),
4013 list_for_each_entry(searchp, &cache_chain, next) {
4017 * We only take the l3 lock if absolutely necessary and we
4018 * have established with reasonable certainty that
4019 * we can do some work if the lock was obtained.
4021 l3 = searchp->nodelists[node];
4023 reap_alien(searchp, l3);
4025 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4028 * These are racy checks but it does not matter
4029 * if we skip one check or scan twice.
4031 if (time_after(l3->next_reap, jiffies))
4034 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4036 drain_array(searchp, l3, l3->shared, 0, node);
4038 if (l3->free_touched)
4039 l3->free_touched = 0;
4043 freed = drain_freelist(searchp, l3, (l3->free_limit +
4044 5 * searchp->num - 1) / (5 * searchp->num));
4045 STATS_ADD_REAPED(searchp, freed);
4051 mutex_unlock(&cache_chain_mutex);
4053 refresh_cpu_vm_stats(smp_processor_id());
4054 /* Set up the next iteration */
4055 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
4058 #ifdef CONFIG_PROC_FS
4060 static void print_slabinfo_header(struct seq_file *m)
4063 * Output format version, so at least we can change it
4064 * without _too_ many complaints.
4067 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4069 seq_puts(m, "slabinfo - version: 2.1\n");
4071 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4072 "<objperslab> <pagesperslab>");
4073 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4074 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4076 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4077 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4078 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4083 static void *s_start(struct seq_file *m, loff_t *pos)
4086 struct list_head *p;
4088 mutex_lock(&cache_chain_mutex);
4090 print_slabinfo_header(m);
4091 p = cache_chain.next;
4094 if (p == &cache_chain)
4097 return list_entry(p, struct kmem_cache, next);
4100 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4102 struct kmem_cache *cachep = p;
4104 return cachep->next.next == &cache_chain ?
4105 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
4108 static void s_stop(struct seq_file *m, void *p)
4110 mutex_unlock(&cache_chain_mutex);
4113 static int s_show(struct seq_file *m, void *p)
4115 struct kmem_cache *cachep = p;
4117 unsigned long active_objs;
4118 unsigned long num_objs;
4119 unsigned long active_slabs = 0;
4120 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4124 struct kmem_list3 *l3;
4128 for_each_online_node(node) {
4129 l3 = cachep->nodelists[node];
4134 spin_lock_irq(&l3->list_lock);
4136 list_for_each_entry(slabp, &l3->slabs_full, list) {
4137 if (slabp->inuse != cachep->num && !error)
4138 error = "slabs_full accounting error";
4139 active_objs += cachep->num;
4142 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4143 if (slabp->inuse == cachep->num && !error)
4144 error = "slabs_partial inuse accounting error";
4145 if (!slabp->inuse && !error)
4146 error = "slabs_partial/inuse accounting error";
4147 active_objs += slabp->inuse;
4150 list_for_each_entry(slabp, &l3->slabs_free, list) {
4151 if (slabp->inuse && !error)
4152 error = "slabs_free/inuse accounting error";
4155 free_objects += l3->free_objects;
4157 shared_avail += l3->shared->avail;
4159 spin_unlock_irq(&l3->list_lock);
4161 num_slabs += active_slabs;
4162 num_objs = num_slabs * cachep->num;
4163 if (num_objs - active_objs != free_objects && !error)
4164 error = "free_objects accounting error";
4166 name = cachep->name;
4168 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4170 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4171 name, active_objs, num_objs, cachep->buffer_size,
4172 cachep->num, (1 << cachep->gfporder));
4173 seq_printf(m, " : tunables %4u %4u %4u",
4174 cachep->limit, cachep->batchcount, cachep->shared);
4175 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4176 active_slabs, num_slabs, shared_avail);
4179 unsigned long high = cachep->high_mark;
4180 unsigned long allocs = cachep->num_allocations;
4181 unsigned long grown = cachep->grown;
4182 unsigned long reaped = cachep->reaped;
4183 unsigned long errors = cachep->errors;
4184 unsigned long max_freeable = cachep->max_freeable;
4185 unsigned long node_allocs = cachep->node_allocs;
4186 unsigned long node_frees = cachep->node_frees;
4187 unsigned long overflows = cachep->node_overflow;
4189 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4190 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4191 reaped, errors, max_freeable, node_allocs,
4192 node_frees, overflows);
4196 unsigned long allochit = atomic_read(&cachep->allochit);
4197 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4198 unsigned long freehit = atomic_read(&cachep->freehit);
4199 unsigned long freemiss = atomic_read(&cachep->freemiss);
4201 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4202 allochit, allocmiss, freehit, freemiss);
4210 * slabinfo_op - iterator that generates /proc/slabinfo
4219 * num-pages-per-slab
4220 * + further values on SMP and with statistics enabled
4223 const struct seq_operations slabinfo_op = {
4230 #define MAX_SLABINFO_WRITE 128
4232 * slabinfo_write - Tuning for the slab allocator
4234 * @buffer: user buffer
4235 * @count: data length
4238 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4239 size_t count, loff_t *ppos)
4241 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4242 int limit, batchcount, shared, res;
4243 struct kmem_cache *cachep;
4245 if (count > MAX_SLABINFO_WRITE)
4247 if (copy_from_user(&kbuf, buffer, count))
4249 kbuf[MAX_SLABINFO_WRITE] = '\0';
4251 tmp = strchr(kbuf, ' ');
4256 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4259 /* Find the cache in the chain of caches. */
4260 mutex_lock(&cache_chain_mutex);
4262 list_for_each_entry(cachep, &cache_chain, next) {
4263 if (!strcmp(cachep->name, kbuf)) {
4264 if (limit < 1 || batchcount < 1 ||
4265 batchcount > limit || shared < 0) {
4268 res = do_tune_cpucache(cachep, limit,
4269 batchcount, shared);
4274 mutex_unlock(&cache_chain_mutex);
4280 #ifdef CONFIG_DEBUG_SLAB_LEAK
4282 static void *leaks_start(struct seq_file *m, loff_t *pos)
4285 struct list_head *p;
4287 mutex_lock(&cache_chain_mutex);
4288 p = cache_chain.next;
4291 if (p == &cache_chain)
4294 return list_entry(p, struct kmem_cache, next);
4297 static inline int add_caller(unsigned long *n, unsigned long v)
4307 unsigned long *q = p + 2 * i;
4321 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4327 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4333 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4334 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4336 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4341 static void show_symbol(struct seq_file *m, unsigned long address)
4343 #ifdef CONFIG_KALLSYMS
4346 unsigned long offset, size;
4347 char namebuf[KSYM_NAME_LEN+1];
4349 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4352 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4354 seq_printf(m, " [%s]", modname);
4358 seq_printf(m, "%p", (void *)address);
4361 static int leaks_show(struct seq_file *m, void *p)
4363 struct kmem_cache *cachep = p;
4365 struct kmem_list3 *l3;
4367 unsigned long *n = m->private;
4371 if (!(cachep->flags & SLAB_STORE_USER))
4373 if (!(cachep->flags & SLAB_RED_ZONE))
4376 /* OK, we can do it */
4380 for_each_online_node(node) {
4381 l3 = cachep->nodelists[node];
4386 spin_lock_irq(&l3->list_lock);
4388 list_for_each_entry(slabp, &l3->slabs_full, list)
4389 handle_slab(n, cachep, slabp);
4390 list_for_each_entry(slabp, &l3->slabs_partial, list)
4391 handle_slab(n, cachep, slabp);
4392 spin_unlock_irq(&l3->list_lock);
4394 name = cachep->name;
4396 /* Increase the buffer size */
4397 mutex_unlock(&cache_chain_mutex);
4398 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4400 /* Too bad, we are really out */
4402 mutex_lock(&cache_chain_mutex);
4405 *(unsigned long *)m->private = n[0] * 2;
4407 mutex_lock(&cache_chain_mutex);
4408 /* Now make sure this entry will be retried */
4412 for (i = 0; i < n[1]; i++) {
4413 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4414 show_symbol(m, n[2*i+2]);
4421 const struct seq_operations slabstats_op = {
4422 .start = leaks_start,
4431 * ksize - get the actual amount of memory allocated for a given object
4432 * @objp: Pointer to the object
4434 * kmalloc may internally round up allocations and return more memory
4435 * than requested. ksize() can be used to determine the actual amount of
4436 * memory allocated. The caller may use this additional memory, even though
4437 * a smaller amount of memory was initially specified with the kmalloc call.
4438 * The caller must guarantee that objp points to a valid object previously
4439 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4440 * must not be freed during the duration of the call.
4442 unsigned int ksize(const void *objp)
4444 if (unlikely(objp == NULL))
4447 return obj_size(virt_to_cache(objp));