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/config.h>
90 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/rtmutex.h>
112 #include <asm/uaccess.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(void *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)
734 /* Guard access to the cache-chain. */
735 static DEFINE_MUTEX(cache_chain_mutex);
736 static struct list_head cache_chain;
739 * vm_enough_memory() looks at this to determine how many slab-allocated pages
740 * are possibly freeable under pressure
742 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
744 atomic_t slab_reclaim_pages;
747 * chicken and egg problem: delay the per-cpu array allocation
748 * until the general caches are up.
758 * used by boot code to determine if it can use slab based allocator
760 int slab_is_available(void)
762 return g_cpucache_up == FULL;
765 static DEFINE_PER_CPU(struct work_struct, reap_work);
767 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
769 return cachep->array[smp_processor_id()];
772 static inline struct kmem_cache *__find_general_cachep(size_t size,
775 struct cache_sizes *csizep = malloc_sizes;
778 /* This happens if someone tries to call
779 * kmem_cache_create(), or __kmalloc(), before
780 * the generic caches are initialized.
782 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
784 while (size > csizep->cs_size)
788 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
789 * has cs_{dma,}cachep==NULL. Thus no special case
790 * for large kmalloc calls required.
792 if (unlikely(gfpflags & GFP_DMA))
793 return csizep->cs_dmacachep;
794 return csizep->cs_cachep;
797 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
799 return __find_general_cachep(size, gfpflags);
802 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
804 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
808 * Calculate the number of objects and left-over bytes for a given buffer size.
810 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
811 size_t align, int flags, size_t *left_over,
816 size_t slab_size = PAGE_SIZE << gfporder;
819 * The slab management structure can be either off the slab or
820 * on it. For the latter case, the memory allocated for a
824 * - One kmem_bufctl_t for each object
825 * - Padding to respect alignment of @align
826 * - @buffer_size bytes for each object
828 * If the slab management structure is off the slab, then the
829 * alignment will already be calculated into the size. Because
830 * the slabs are all pages aligned, the objects will be at the
831 * correct alignment when allocated.
833 if (flags & CFLGS_OFF_SLAB) {
835 nr_objs = slab_size / buffer_size;
837 if (nr_objs > SLAB_LIMIT)
838 nr_objs = SLAB_LIMIT;
841 * Ignore padding for the initial guess. The padding
842 * is at most @align-1 bytes, and @buffer_size is at
843 * least @align. In the worst case, this result will
844 * be one greater than the number of objects that fit
845 * into the memory allocation when taking the padding
848 nr_objs = (slab_size - sizeof(struct slab)) /
849 (buffer_size + sizeof(kmem_bufctl_t));
852 * This calculated number will be either the right
853 * amount, or one greater than what we want.
855 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
859 if (nr_objs > SLAB_LIMIT)
860 nr_objs = SLAB_LIMIT;
862 mgmt_size = slab_mgmt_size(nr_objs, align);
865 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
868 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
870 static void __slab_error(const char *function, struct kmem_cache *cachep,
873 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
874 function, cachep->name, msg);
880 * Special reaping functions for NUMA systems called from cache_reap().
881 * These take care of doing round robin flushing of alien caches (containing
882 * objects freed on different nodes from which they were allocated) and the
883 * flushing of remote pcps by calling drain_node_pages.
885 static DEFINE_PER_CPU(unsigned long, reap_node);
887 static void init_reap_node(int cpu)
891 node = next_node(cpu_to_node(cpu), node_online_map);
892 if (node == MAX_NUMNODES)
893 node = first_node(node_online_map);
895 __get_cpu_var(reap_node) = node;
898 static void next_reap_node(void)
900 int node = __get_cpu_var(reap_node);
903 * Also drain per cpu pages on remote zones
905 if (node != numa_node_id())
906 drain_node_pages(node);
908 node = next_node(node, node_online_map);
909 if (unlikely(node >= MAX_NUMNODES))
910 node = first_node(node_online_map);
911 __get_cpu_var(reap_node) = node;
915 #define init_reap_node(cpu) do { } while (0)
916 #define next_reap_node(void) do { } while (0)
920 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
921 * via the workqueue/eventd.
922 * Add the CPU number into the expiration time to minimize the possibility of
923 * the CPUs getting into lockstep and contending for the global cache chain
926 static void __devinit start_cpu_timer(int cpu)
928 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
931 * When this gets called from do_initcalls via cpucache_init(),
932 * init_workqueues() has already run, so keventd will be setup
935 if (keventd_up() && reap_work->func == NULL) {
937 INIT_WORK(reap_work, cache_reap, NULL);
938 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
942 static struct array_cache *alloc_arraycache(int node, int entries,
945 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
946 struct array_cache *nc = NULL;
948 nc = kmalloc_node(memsize, GFP_KERNEL, node);
952 nc->batchcount = batchcount;
954 spin_lock_init(&nc->lock);
960 * Transfer objects in one arraycache to another.
961 * Locking must be handled by the caller.
963 * Return the number of entries transferred.
965 static int transfer_objects(struct array_cache *to,
966 struct array_cache *from, unsigned int max)
968 /* Figure out how many entries to transfer */
969 int nr = min(min(from->avail, max), to->limit - to->avail);
974 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
984 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
985 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
987 static struct array_cache **alloc_alien_cache(int node, int limit)
989 struct array_cache **ac_ptr;
990 int memsize = sizeof(void *) * MAX_NUMNODES;
995 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
998 if (i == node || !node_online(i)) {
1002 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1004 for (i--; i <= 0; i--)
1014 static void free_alien_cache(struct array_cache **ac_ptr)
1025 static void __drain_alien_cache(struct kmem_cache *cachep,
1026 struct array_cache *ac, int node)
1028 struct kmem_list3 *rl3 = cachep->nodelists[node];
1031 spin_lock(&rl3->list_lock);
1033 * Stuff objects into the remote nodes shared array first.
1034 * That way we could avoid the overhead of putting the objects
1035 * into the free lists and getting them back later.
1038 transfer_objects(rl3->shared, ac, ac->limit);
1040 free_block(cachep, ac->entry, ac->avail, node);
1042 spin_unlock(&rl3->list_lock);
1047 * Called from cache_reap() to regularly drain alien caches round robin.
1049 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1051 int node = __get_cpu_var(reap_node);
1054 struct array_cache *ac = l3->alien[node];
1056 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1057 __drain_alien_cache(cachep, ac, node);
1058 spin_unlock_irq(&ac->lock);
1063 static void drain_alien_cache(struct kmem_cache *cachep,
1064 struct array_cache **alien)
1067 struct array_cache *ac;
1068 unsigned long flags;
1070 for_each_online_node(i) {
1073 spin_lock_irqsave(&ac->lock, flags);
1074 __drain_alien_cache(cachep, ac, i);
1075 spin_unlock_irqrestore(&ac->lock, flags);
1080 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1082 struct slab *slabp = virt_to_slab(objp);
1083 int nodeid = slabp->nodeid;
1084 struct kmem_list3 *l3;
1085 struct array_cache *alien = NULL;
1088 * Make sure we are not freeing a object from another node to the array
1089 * cache on this cpu.
1091 if (likely(slabp->nodeid == numa_node_id()))
1094 l3 = cachep->nodelists[numa_node_id()];
1095 STATS_INC_NODEFREES(cachep);
1096 if (l3->alien && l3->alien[nodeid]) {
1097 alien = l3->alien[nodeid];
1098 spin_lock(&alien->lock);
1099 if (unlikely(alien->avail == alien->limit)) {
1100 STATS_INC_ACOVERFLOW(cachep);
1101 __drain_alien_cache(cachep, alien, nodeid);
1103 alien->entry[alien->avail++] = objp;
1104 spin_unlock(&alien->lock);
1106 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1107 free_block(cachep, &objp, 1, nodeid);
1108 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1115 #define drain_alien_cache(cachep, alien) do { } while (0)
1116 #define reap_alien(cachep, l3) do { } while (0)
1118 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1120 return (struct array_cache **)BAD_ALIEN_MAGIC;
1123 static inline void free_alien_cache(struct array_cache **ac_ptr)
1127 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1134 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1135 unsigned long action, void *hcpu)
1137 long cpu = (long)hcpu;
1138 struct kmem_cache *cachep;
1139 struct kmem_list3 *l3 = NULL;
1140 int node = cpu_to_node(cpu);
1141 int memsize = sizeof(struct kmem_list3);
1144 case CPU_UP_PREPARE:
1145 mutex_lock(&cache_chain_mutex);
1147 * We need to do this right in the beginning since
1148 * alloc_arraycache's are going to use this list.
1149 * kmalloc_node allows us to add the slab to the right
1150 * kmem_list3 and not this cpu's kmem_list3
1153 list_for_each_entry(cachep, &cache_chain, next) {
1155 * Set up the size64 kmemlist for cpu before we can
1156 * begin anything. Make sure some other cpu on this
1157 * node has not already allocated this
1159 if (!cachep->nodelists[node]) {
1160 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1163 kmem_list3_init(l3);
1164 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1165 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1168 * The l3s don't come and go as CPUs come and
1169 * go. cache_chain_mutex is sufficient
1172 cachep->nodelists[node] = l3;
1175 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1176 cachep->nodelists[node]->free_limit =
1177 (1 + nr_cpus_node(node)) *
1178 cachep->batchcount + cachep->num;
1179 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1183 * Now we can go ahead with allocating the shared arrays and
1186 list_for_each_entry(cachep, &cache_chain, next) {
1187 struct array_cache *nc;
1188 struct array_cache *shared;
1189 struct array_cache **alien;
1191 nc = alloc_arraycache(node, cachep->limit,
1192 cachep->batchcount);
1195 shared = alloc_arraycache(node,
1196 cachep->shared * cachep->batchcount,
1201 alien = alloc_alien_cache(node, cachep->limit);
1204 cachep->array[cpu] = nc;
1205 l3 = cachep->nodelists[node];
1208 spin_lock_irq(&l3->list_lock);
1211 * We are serialised from CPU_DEAD or
1212 * CPU_UP_CANCELLED by the cpucontrol lock
1214 l3->shared = shared;
1223 spin_unlock_irq(&l3->list_lock);
1225 free_alien_cache(alien);
1227 mutex_unlock(&cache_chain_mutex);
1230 start_cpu_timer(cpu);
1232 #ifdef CONFIG_HOTPLUG_CPU
1235 * Even if all the cpus of a node are down, we don't free the
1236 * kmem_list3 of any cache. This to avoid a race between
1237 * cpu_down, and a kmalloc allocation from another cpu for
1238 * memory from the node of the cpu going down. The list3
1239 * structure is usually allocated from kmem_cache_create() and
1240 * gets destroyed at kmem_cache_destroy().
1243 case CPU_UP_CANCELED:
1244 mutex_lock(&cache_chain_mutex);
1245 list_for_each_entry(cachep, &cache_chain, next) {
1246 struct array_cache *nc;
1247 struct array_cache *shared;
1248 struct array_cache **alien;
1251 mask = node_to_cpumask(node);
1252 /* cpu is dead; no one can alloc from it. */
1253 nc = cachep->array[cpu];
1254 cachep->array[cpu] = NULL;
1255 l3 = cachep->nodelists[node];
1258 goto free_array_cache;
1260 spin_lock_irq(&l3->list_lock);
1262 /* Free limit for this kmem_list3 */
1263 l3->free_limit -= cachep->batchcount;
1265 free_block(cachep, nc->entry, nc->avail, node);
1267 if (!cpus_empty(mask)) {
1268 spin_unlock_irq(&l3->list_lock);
1269 goto free_array_cache;
1272 shared = l3->shared;
1274 free_block(cachep, l3->shared->entry,
1275 l3->shared->avail, node);
1282 spin_unlock_irq(&l3->list_lock);
1286 drain_alien_cache(cachep, alien);
1287 free_alien_cache(alien);
1293 * In the previous loop, all the objects were freed to
1294 * the respective cache's slabs, now we can go ahead and
1295 * shrink each nodelist to its limit.
1297 list_for_each_entry(cachep, &cache_chain, next) {
1298 l3 = cachep->nodelists[node];
1301 drain_freelist(cachep, l3, l3->free_objects);
1303 mutex_unlock(&cache_chain_mutex);
1309 mutex_unlock(&cache_chain_mutex);
1313 static struct notifier_block __cpuinitdata cpucache_notifier = {
1314 &cpuup_callback, NULL, 0
1318 * swap the static kmem_list3 with kmalloced memory
1320 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1323 struct kmem_list3 *ptr;
1325 BUG_ON(cachep->nodelists[nodeid] != list);
1326 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1329 local_irq_disable();
1330 memcpy(ptr, list, sizeof(struct kmem_list3));
1332 * Do not assume that spinlocks can be initialized via memcpy:
1334 spin_lock_init(&ptr->list_lock);
1336 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1337 cachep->nodelists[nodeid] = ptr;
1342 * Initialisation. Called after the page allocator have been initialised and
1343 * before smp_init().
1345 void __init kmem_cache_init(void)
1348 struct cache_sizes *sizes;
1349 struct cache_names *names;
1353 for (i = 0; i < NUM_INIT_LISTS; i++) {
1354 kmem_list3_init(&initkmem_list3[i]);
1355 if (i < MAX_NUMNODES)
1356 cache_cache.nodelists[i] = NULL;
1360 * Fragmentation resistance on low memory - only use bigger
1361 * page orders on machines with more than 32MB of memory.
1363 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1364 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1366 /* Bootstrap is tricky, because several objects are allocated
1367 * from caches that do not exist yet:
1368 * 1) initialize the cache_cache cache: it contains the struct
1369 * kmem_cache structures of all caches, except cache_cache itself:
1370 * cache_cache is statically allocated.
1371 * Initially an __init data area is used for the head array and the
1372 * kmem_list3 structures, it's replaced with a kmalloc allocated
1373 * array at the end of the bootstrap.
1374 * 2) Create the first kmalloc cache.
1375 * The struct kmem_cache for the new cache is allocated normally.
1376 * An __init data area is used for the head array.
1377 * 3) Create the remaining kmalloc caches, with minimally sized
1379 * 4) Replace the __init data head arrays for cache_cache and the first
1380 * kmalloc cache with kmalloc allocated arrays.
1381 * 5) Replace the __init data for kmem_list3 for cache_cache and
1382 * the other cache's with kmalloc allocated memory.
1383 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1386 /* 1) create the cache_cache */
1387 INIT_LIST_HEAD(&cache_chain);
1388 list_add(&cache_cache.next, &cache_chain);
1389 cache_cache.colour_off = cache_line_size();
1390 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1391 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1393 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1396 for (order = 0; order < MAX_ORDER; order++) {
1397 cache_estimate(order, cache_cache.buffer_size,
1398 cache_line_size(), 0, &left_over, &cache_cache.num);
1399 if (cache_cache.num)
1402 BUG_ON(!cache_cache.num);
1403 cache_cache.gfporder = order;
1404 cache_cache.colour = left_over / cache_cache.colour_off;
1405 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1406 sizeof(struct slab), cache_line_size());
1408 /* 2+3) create the kmalloc caches */
1409 sizes = malloc_sizes;
1410 names = cache_names;
1413 * Initialize the caches that provide memory for the array cache and the
1414 * kmem_list3 structures first. Without this, further allocations will
1418 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1419 sizes[INDEX_AC].cs_size,
1420 ARCH_KMALLOC_MINALIGN,
1421 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1424 if (INDEX_AC != INDEX_L3) {
1425 sizes[INDEX_L3].cs_cachep =
1426 kmem_cache_create(names[INDEX_L3].name,
1427 sizes[INDEX_L3].cs_size,
1428 ARCH_KMALLOC_MINALIGN,
1429 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1433 slab_early_init = 0;
1435 while (sizes->cs_size != ULONG_MAX) {
1437 * For performance, all the general caches are L1 aligned.
1438 * This should be particularly beneficial on SMP boxes, as it
1439 * eliminates "false sharing".
1440 * Note for systems short on memory removing the alignment will
1441 * allow tighter packing of the smaller caches.
1443 if (!sizes->cs_cachep) {
1444 sizes->cs_cachep = kmem_cache_create(names->name,
1446 ARCH_KMALLOC_MINALIGN,
1447 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1451 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1453 ARCH_KMALLOC_MINALIGN,
1454 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1460 /* 4) Replace the bootstrap head arrays */
1462 struct array_cache *ptr;
1464 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1466 local_irq_disable();
1467 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1468 memcpy(ptr, cpu_cache_get(&cache_cache),
1469 sizeof(struct arraycache_init));
1471 * Do not assume that spinlocks can be initialized via memcpy:
1473 spin_lock_init(&ptr->lock);
1475 cache_cache.array[smp_processor_id()] = ptr;
1478 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1480 local_irq_disable();
1481 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1482 != &initarray_generic.cache);
1483 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1484 sizeof(struct arraycache_init));
1486 * Do not assume that spinlocks can be initialized via memcpy:
1488 spin_lock_init(&ptr->lock);
1490 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1494 /* 5) Replace the bootstrap kmem_list3's */
1497 /* Replace the static kmem_list3 structures for the boot cpu */
1498 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1501 for_each_online_node(node) {
1502 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1503 &initkmem_list3[SIZE_AC + node], node);
1505 if (INDEX_AC != INDEX_L3) {
1506 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1507 &initkmem_list3[SIZE_L3 + node],
1513 /* 6) resize the head arrays to their final sizes */
1515 struct kmem_cache *cachep;
1516 mutex_lock(&cache_chain_mutex);
1517 list_for_each_entry(cachep, &cache_chain, next)
1518 if (enable_cpucache(cachep))
1520 mutex_unlock(&cache_chain_mutex);
1523 /* Annotate slab for lockdep -- annotate the malloc caches */
1528 g_cpucache_up = FULL;
1531 * Register a cpu startup notifier callback that initializes
1532 * cpu_cache_get for all new cpus
1534 register_cpu_notifier(&cpucache_notifier);
1537 * The reap timers are started later, with a module init call: That part
1538 * of the kernel is not yet operational.
1542 static int __init cpucache_init(void)
1547 * Register the timers that return unneeded pages to the page allocator
1549 for_each_online_cpu(cpu)
1550 start_cpu_timer(cpu);
1553 __initcall(cpucache_init);
1556 * Interface to system's page allocator. No need to hold the cache-lock.
1558 * If we requested dmaable memory, we will get it. Even if we
1559 * did not request dmaable memory, we might get it, but that
1560 * would be relatively rare and ignorable.
1562 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1570 * Nommu uses slab's for process anonymous memory allocations, and thus
1571 * requires __GFP_COMP to properly refcount higher order allocations
1573 flags |= __GFP_COMP;
1575 flags |= cachep->gfpflags;
1577 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1581 nr_pages = (1 << cachep->gfporder);
1582 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1583 atomic_add(nr_pages, &slab_reclaim_pages);
1584 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages);
1585 for (i = 0; i < nr_pages; i++)
1586 __SetPageSlab(page + i);
1587 return page_address(page);
1591 * Interface to system's page release.
1593 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1595 unsigned long i = (1 << cachep->gfporder);
1596 struct page *page = virt_to_page(addr);
1597 const unsigned long nr_freed = i;
1599 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed);
1601 BUG_ON(!PageSlab(page));
1602 __ClearPageSlab(page);
1605 if (current->reclaim_state)
1606 current->reclaim_state->reclaimed_slab += nr_freed;
1607 free_pages((unsigned long)addr, cachep->gfporder);
1608 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1609 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1612 static void kmem_rcu_free(struct rcu_head *head)
1614 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1615 struct kmem_cache *cachep = slab_rcu->cachep;
1617 kmem_freepages(cachep, slab_rcu->addr);
1618 if (OFF_SLAB(cachep))
1619 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1624 #ifdef CONFIG_DEBUG_PAGEALLOC
1625 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1626 unsigned long caller)
1628 int size = obj_size(cachep);
1630 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1632 if (size < 5 * sizeof(unsigned long))
1635 *addr++ = 0x12345678;
1637 *addr++ = smp_processor_id();
1638 size -= 3 * sizeof(unsigned long);
1640 unsigned long *sptr = &caller;
1641 unsigned long svalue;
1643 while (!kstack_end(sptr)) {
1645 if (kernel_text_address(svalue)) {
1647 size -= sizeof(unsigned long);
1648 if (size <= sizeof(unsigned long))
1654 *addr++ = 0x87654321;
1658 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1660 int size = obj_size(cachep);
1661 addr = &((char *)addr)[obj_offset(cachep)];
1663 memset(addr, val, size);
1664 *(unsigned char *)(addr + size - 1) = POISON_END;
1667 static void dump_line(char *data, int offset, int limit)
1670 printk(KERN_ERR "%03x:", offset);
1671 for (i = 0; i < limit; i++)
1672 printk(" %02x", (unsigned char)data[offset + i]);
1679 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1684 if (cachep->flags & SLAB_RED_ZONE) {
1685 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1686 *dbg_redzone1(cachep, objp),
1687 *dbg_redzone2(cachep, objp));
1690 if (cachep->flags & SLAB_STORE_USER) {
1691 printk(KERN_ERR "Last user: [<%p>]",
1692 *dbg_userword(cachep, objp));
1693 print_symbol("(%s)",
1694 (unsigned long)*dbg_userword(cachep, objp));
1697 realobj = (char *)objp + obj_offset(cachep);
1698 size = obj_size(cachep);
1699 for (i = 0; i < size && lines; i += 16, lines--) {
1702 if (i + limit > size)
1704 dump_line(realobj, i, limit);
1708 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1714 realobj = (char *)objp + obj_offset(cachep);
1715 size = obj_size(cachep);
1717 for (i = 0; i < size; i++) {
1718 char exp = POISON_FREE;
1721 if (realobj[i] != exp) {
1727 "Slab corruption: start=%p, len=%d\n",
1729 print_objinfo(cachep, objp, 0);
1731 /* Hexdump the affected line */
1734 if (i + limit > size)
1736 dump_line(realobj, i, limit);
1739 /* Limit to 5 lines */
1745 /* Print some data about the neighboring objects, if they
1748 struct slab *slabp = virt_to_slab(objp);
1751 objnr = obj_to_index(cachep, slabp, objp);
1753 objp = index_to_obj(cachep, slabp, objnr - 1);
1754 realobj = (char *)objp + obj_offset(cachep);
1755 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1757 print_objinfo(cachep, objp, 2);
1759 if (objnr + 1 < cachep->num) {
1760 objp = index_to_obj(cachep, slabp, objnr + 1);
1761 realobj = (char *)objp + obj_offset(cachep);
1762 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1764 print_objinfo(cachep, objp, 2);
1772 * slab_destroy_objs - destroy a slab and its objects
1773 * @cachep: cache pointer being destroyed
1774 * @slabp: slab pointer being destroyed
1776 * Call the registered destructor for each object in a slab that is being
1779 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1782 for (i = 0; i < cachep->num; i++) {
1783 void *objp = index_to_obj(cachep, slabp, i);
1785 if (cachep->flags & SLAB_POISON) {
1786 #ifdef CONFIG_DEBUG_PAGEALLOC
1787 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1789 kernel_map_pages(virt_to_page(objp),
1790 cachep->buffer_size / PAGE_SIZE, 1);
1792 check_poison_obj(cachep, objp);
1794 check_poison_obj(cachep, objp);
1797 if (cachep->flags & SLAB_RED_ZONE) {
1798 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1799 slab_error(cachep, "start of a freed object "
1801 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1802 slab_error(cachep, "end of a freed object "
1805 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1806 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1810 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1814 for (i = 0; i < cachep->num; i++) {
1815 void *objp = index_to_obj(cachep, slabp, i);
1816 (cachep->dtor) (objp, cachep, 0);
1823 * slab_destroy - destroy and release all objects in a slab
1824 * @cachep: cache pointer being destroyed
1825 * @slabp: slab pointer being destroyed
1827 * Destroy all the objs in a slab, and release the mem back to the system.
1828 * Before calling the slab must have been unlinked from the cache. The
1829 * cache-lock is not held/needed.
1831 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1833 void *addr = slabp->s_mem - slabp->colouroff;
1835 slab_destroy_objs(cachep, slabp);
1836 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1837 struct slab_rcu *slab_rcu;
1839 slab_rcu = (struct slab_rcu *)slabp;
1840 slab_rcu->cachep = cachep;
1841 slab_rcu->addr = addr;
1842 call_rcu(&slab_rcu->head, kmem_rcu_free);
1844 kmem_freepages(cachep, addr);
1845 if (OFF_SLAB(cachep))
1846 kmem_cache_free(cachep->slabp_cache, slabp);
1851 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1852 * size of kmem_list3.
1854 static void set_up_list3s(struct kmem_cache *cachep, int index)
1858 for_each_online_node(node) {
1859 cachep->nodelists[node] = &initkmem_list3[index + node];
1860 cachep->nodelists[node]->next_reap = jiffies +
1862 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1866 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1869 struct kmem_list3 *l3;
1871 for_each_online_cpu(i)
1872 kfree(cachep->array[i]);
1874 /* NUMA: free the list3 structures */
1875 for_each_online_node(i) {
1876 l3 = cachep->nodelists[i];
1879 free_alien_cache(l3->alien);
1883 kmem_cache_free(&cache_cache, cachep);
1888 * calculate_slab_order - calculate size (page order) of slabs
1889 * @cachep: pointer to the cache that is being created
1890 * @size: size of objects to be created in this cache.
1891 * @align: required alignment for the objects.
1892 * @flags: slab allocation flags
1894 * Also calculates the number of objects per slab.
1896 * This could be made much more intelligent. For now, try to avoid using
1897 * high order pages for slabs. When the gfp() functions are more friendly
1898 * towards high-order requests, this should be changed.
1900 static size_t calculate_slab_order(struct kmem_cache *cachep,
1901 size_t size, size_t align, unsigned long flags)
1903 unsigned long offslab_limit;
1904 size_t left_over = 0;
1907 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1911 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1915 if (flags & CFLGS_OFF_SLAB) {
1917 * Max number of objs-per-slab for caches which
1918 * use off-slab slabs. Needed to avoid a possible
1919 * looping condition in cache_grow().
1921 offslab_limit = size - sizeof(struct slab);
1922 offslab_limit /= sizeof(kmem_bufctl_t);
1924 if (num > offslab_limit)
1928 /* Found something acceptable - save it away */
1930 cachep->gfporder = gfporder;
1931 left_over = remainder;
1934 * A VFS-reclaimable slab tends to have most allocations
1935 * as GFP_NOFS and we really don't want to have to be allocating
1936 * higher-order pages when we are unable to shrink dcache.
1938 if (flags & SLAB_RECLAIM_ACCOUNT)
1942 * Large number of objects is good, but very large slabs are
1943 * currently bad for the gfp()s.
1945 if (gfporder >= slab_break_gfp_order)
1949 * Acceptable internal fragmentation?
1951 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1957 static int setup_cpu_cache(struct kmem_cache *cachep)
1959 if (g_cpucache_up == FULL)
1960 return enable_cpucache(cachep);
1962 if (g_cpucache_up == NONE) {
1964 * Note: the first kmem_cache_create must create the cache
1965 * that's used by kmalloc(24), otherwise the creation of
1966 * further caches will BUG().
1968 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1971 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1972 * the first cache, then we need to set up all its list3s,
1973 * otherwise the creation of further caches will BUG().
1975 set_up_list3s(cachep, SIZE_AC);
1976 if (INDEX_AC == INDEX_L3)
1977 g_cpucache_up = PARTIAL_L3;
1979 g_cpucache_up = PARTIAL_AC;
1981 cachep->array[smp_processor_id()] =
1982 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1984 if (g_cpucache_up == PARTIAL_AC) {
1985 set_up_list3s(cachep, SIZE_L3);
1986 g_cpucache_up = PARTIAL_L3;
1989 for_each_online_node(node) {
1990 cachep->nodelists[node] =
1991 kmalloc_node(sizeof(struct kmem_list3),
1993 BUG_ON(!cachep->nodelists[node]);
1994 kmem_list3_init(cachep->nodelists[node]);
1998 cachep->nodelists[numa_node_id()]->next_reap =
1999 jiffies + REAPTIMEOUT_LIST3 +
2000 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2002 cpu_cache_get(cachep)->avail = 0;
2003 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2004 cpu_cache_get(cachep)->batchcount = 1;
2005 cpu_cache_get(cachep)->touched = 0;
2006 cachep->batchcount = 1;
2007 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2012 * kmem_cache_create - Create a cache.
2013 * @name: A string which is used in /proc/slabinfo to identify this cache.
2014 * @size: The size of objects to be created in this cache.
2015 * @align: The required alignment for the objects.
2016 * @flags: SLAB flags
2017 * @ctor: A constructor for the objects.
2018 * @dtor: A destructor for the objects.
2020 * Returns a ptr to the cache on success, NULL on failure.
2021 * Cannot be called within a int, but can be interrupted.
2022 * The @ctor is run when new pages are allocated by the cache
2023 * and the @dtor is run before the pages are handed back.
2025 * @name must be valid until the cache is destroyed. This implies that
2026 * the module calling this has to destroy the cache before getting unloaded.
2030 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2031 * to catch references to uninitialised memory.
2033 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2034 * for buffer overruns.
2036 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2037 * cacheline. This can be beneficial if you're counting cycles as closely
2041 kmem_cache_create (const char *name, size_t size, size_t align,
2042 unsigned long flags,
2043 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2044 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2046 size_t left_over, slab_size, ralign;
2047 struct kmem_cache *cachep = NULL, *pc;
2050 * Sanity checks... these are all serious usage bugs.
2052 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2053 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2054 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2060 * Prevent CPUs from coming and going.
2061 * lock_cpu_hotplug() nests outside cache_chain_mutex
2065 mutex_lock(&cache_chain_mutex);
2067 list_for_each_entry(pc, &cache_chain, next) {
2068 mm_segment_t old_fs = get_fs();
2073 * This happens when the module gets unloaded and doesn't
2074 * destroy its slab cache and no-one else reuses the vmalloc
2075 * area of the module. Print a warning.
2078 res = __get_user(tmp, pc->name);
2081 printk("SLAB: cache with size %d has lost its name\n",
2086 if (!strcmp(pc->name, name)) {
2087 printk("kmem_cache_create: duplicate cache %s\n", name);
2094 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2095 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2096 /* No constructor, but inital state check requested */
2097 printk(KERN_ERR "%s: No con, but init state check "
2098 "requested - %s\n", __FUNCTION__, name);
2099 flags &= ~SLAB_DEBUG_INITIAL;
2103 * Enable redzoning and last user accounting, except for caches with
2104 * large objects, if the increased size would increase the object size
2105 * above the next power of two: caches with object sizes just above a
2106 * power of two have a significant amount of internal fragmentation.
2108 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2109 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2110 if (!(flags & SLAB_DESTROY_BY_RCU))
2111 flags |= SLAB_POISON;
2113 if (flags & SLAB_DESTROY_BY_RCU)
2114 BUG_ON(flags & SLAB_POISON);
2116 if (flags & SLAB_DESTROY_BY_RCU)
2120 * Always checks flags, a caller might be expecting debug support which
2123 BUG_ON(flags & ~CREATE_MASK);
2126 * Check that size is in terms of words. This is needed to avoid
2127 * unaligned accesses for some archs when redzoning is used, and makes
2128 * sure any on-slab bufctl's are also correctly aligned.
2130 if (size & (BYTES_PER_WORD - 1)) {
2131 size += (BYTES_PER_WORD - 1);
2132 size &= ~(BYTES_PER_WORD - 1);
2135 /* calculate the final buffer alignment: */
2137 /* 1) arch recommendation: can be overridden for debug */
2138 if (flags & SLAB_HWCACHE_ALIGN) {
2140 * Default alignment: as specified by the arch code. Except if
2141 * an object is really small, then squeeze multiple objects into
2144 ralign = cache_line_size();
2145 while (size <= ralign / 2)
2148 ralign = BYTES_PER_WORD;
2152 * Redzoning and user store require word alignment. Note this will be
2153 * overridden by architecture or caller mandated alignment if either
2154 * is greater than BYTES_PER_WORD.
2156 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2157 ralign = BYTES_PER_WORD;
2159 /* 2) arch mandated alignment: disables debug if necessary */
2160 if (ralign < ARCH_SLAB_MINALIGN) {
2161 ralign = ARCH_SLAB_MINALIGN;
2162 if (ralign > BYTES_PER_WORD)
2163 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2165 /* 3) caller mandated alignment: disables debug if necessary */
2166 if (ralign < align) {
2168 if (ralign > BYTES_PER_WORD)
2169 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2176 /* Get cache's description obj. */
2177 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2182 cachep->obj_size = size;
2185 * Both debugging options require word-alignment which is calculated
2188 if (flags & SLAB_RED_ZONE) {
2189 /* add space for red zone words */
2190 cachep->obj_offset += BYTES_PER_WORD;
2191 size += 2 * BYTES_PER_WORD;
2193 if (flags & SLAB_STORE_USER) {
2194 /* user store requires one word storage behind the end of
2197 size += BYTES_PER_WORD;
2199 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2200 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2201 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2202 cachep->obj_offset += PAGE_SIZE - size;
2209 * Determine if the slab management is 'on' or 'off' slab.
2210 * (bootstrapping cannot cope with offslab caches so don't do
2213 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2215 * Size is large, assume best to place the slab management obj
2216 * off-slab (should allow better packing of objs).
2218 flags |= CFLGS_OFF_SLAB;
2220 size = ALIGN(size, align);
2222 left_over = calculate_slab_order(cachep, size, align, flags);
2225 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2226 kmem_cache_free(&cache_cache, cachep);
2230 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2231 + sizeof(struct slab), align);
2234 * If the slab has been placed off-slab, and we have enough space then
2235 * move it on-slab. This is at the expense of any extra colouring.
2237 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2238 flags &= ~CFLGS_OFF_SLAB;
2239 left_over -= slab_size;
2242 if (flags & CFLGS_OFF_SLAB) {
2243 /* really off slab. No need for manual alignment */
2245 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2248 cachep->colour_off = cache_line_size();
2249 /* Offset must be a multiple of the alignment. */
2250 if (cachep->colour_off < align)
2251 cachep->colour_off = align;
2252 cachep->colour = left_over / cachep->colour_off;
2253 cachep->slab_size = slab_size;
2254 cachep->flags = flags;
2255 cachep->gfpflags = 0;
2256 if (flags & SLAB_CACHE_DMA)
2257 cachep->gfpflags |= GFP_DMA;
2258 cachep->buffer_size = size;
2260 if (flags & CFLGS_OFF_SLAB) {
2261 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2263 * This is a possibility for one of the malloc_sizes caches.
2264 * But since we go off slab only for object size greater than
2265 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2266 * this should not happen at all.
2267 * But leave a BUG_ON for some lucky dude.
2269 BUG_ON(!cachep->slabp_cache);
2271 cachep->ctor = ctor;
2272 cachep->dtor = dtor;
2273 cachep->name = name;
2275 if (setup_cpu_cache(cachep)) {
2276 __kmem_cache_destroy(cachep);
2281 /* cache setup completed, link it into the list */
2282 list_add(&cachep->next, &cache_chain);
2284 if (!cachep && (flags & SLAB_PANIC))
2285 panic("kmem_cache_create(): failed to create slab `%s'\n",
2287 mutex_unlock(&cache_chain_mutex);
2288 unlock_cpu_hotplug();
2291 EXPORT_SYMBOL(kmem_cache_create);
2294 static void check_irq_off(void)
2296 BUG_ON(!irqs_disabled());
2299 static void check_irq_on(void)
2301 BUG_ON(irqs_disabled());
2304 static void check_spinlock_acquired(struct kmem_cache *cachep)
2308 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2312 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2316 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2321 #define check_irq_off() do { } while(0)
2322 #define check_irq_on() do { } while(0)
2323 #define check_spinlock_acquired(x) do { } while(0)
2324 #define check_spinlock_acquired_node(x, y) do { } while(0)
2327 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2328 struct array_cache *ac,
2329 int force, int node);
2331 static void do_drain(void *arg)
2333 struct kmem_cache *cachep = arg;
2334 struct array_cache *ac;
2335 int node = numa_node_id();
2338 ac = cpu_cache_get(cachep);
2339 spin_lock(&cachep->nodelists[node]->list_lock);
2340 free_block(cachep, ac->entry, ac->avail, node);
2341 spin_unlock(&cachep->nodelists[node]->list_lock);
2345 static void drain_cpu_caches(struct kmem_cache *cachep)
2347 struct kmem_list3 *l3;
2350 on_each_cpu(do_drain, cachep, 1, 1);
2352 for_each_online_node(node) {
2353 l3 = cachep->nodelists[node];
2354 if (l3 && l3->alien)
2355 drain_alien_cache(cachep, l3->alien);
2358 for_each_online_node(node) {
2359 l3 = cachep->nodelists[node];
2361 drain_array(cachep, l3, l3->shared, 1, node);
2366 * Remove slabs from the list of free slabs.
2367 * Specify the number of slabs to drain in tofree.
2369 * Returns the actual number of slabs released.
2371 static int drain_freelist(struct kmem_cache *cache,
2372 struct kmem_list3 *l3, int tofree)
2374 struct list_head *p;
2379 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2381 spin_lock_irq(&l3->list_lock);
2382 p = l3->slabs_free.prev;
2383 if (p == &l3->slabs_free) {
2384 spin_unlock_irq(&l3->list_lock);
2388 slabp = list_entry(p, struct slab, list);
2390 BUG_ON(slabp->inuse);
2392 list_del(&slabp->list);
2394 * Safe to drop the lock. The slab is no longer linked
2397 l3->free_objects -= cache->num;
2398 spin_unlock_irq(&l3->list_lock);
2399 slab_destroy(cache, slabp);
2406 static int __cache_shrink(struct kmem_cache *cachep)
2409 struct kmem_list3 *l3;
2411 drain_cpu_caches(cachep);
2414 for_each_online_node(i) {
2415 l3 = cachep->nodelists[i];
2419 drain_freelist(cachep, l3, l3->free_objects);
2421 ret += !list_empty(&l3->slabs_full) ||
2422 !list_empty(&l3->slabs_partial);
2424 return (ret ? 1 : 0);
2428 * kmem_cache_shrink - Shrink a cache.
2429 * @cachep: The cache to shrink.
2431 * Releases as many slabs as possible for a cache.
2432 * To help debugging, a zero exit status indicates all slabs were released.
2434 int kmem_cache_shrink(struct kmem_cache *cachep)
2436 BUG_ON(!cachep || in_interrupt());
2438 return __cache_shrink(cachep);
2440 EXPORT_SYMBOL(kmem_cache_shrink);
2443 * kmem_cache_destroy - delete a cache
2444 * @cachep: the cache to destroy
2446 * Remove a struct kmem_cache object from the slab cache.
2447 * Returns 0 on success.
2449 * It is expected this function will be called by a module when it is
2450 * unloaded. This will remove the cache completely, and avoid a duplicate
2451 * cache being allocated each time a module is loaded and unloaded, if the
2452 * module doesn't have persistent in-kernel storage across loads and unloads.
2454 * The cache must be empty before calling this function.
2456 * The caller must guarantee that noone will allocate memory from the cache
2457 * during the kmem_cache_destroy().
2459 int kmem_cache_destroy(struct kmem_cache *cachep)
2461 BUG_ON(!cachep || in_interrupt());
2463 /* Don't let CPUs to come and go */
2466 /* Find the cache in the chain of caches. */
2467 mutex_lock(&cache_chain_mutex);
2469 * the chain is never empty, cache_cache is never destroyed
2471 list_del(&cachep->next);
2472 mutex_unlock(&cache_chain_mutex);
2474 if (__cache_shrink(cachep)) {
2475 slab_error(cachep, "Can't free all objects");
2476 mutex_lock(&cache_chain_mutex);
2477 list_add(&cachep->next, &cache_chain);
2478 mutex_unlock(&cache_chain_mutex);
2479 unlock_cpu_hotplug();
2483 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2486 __kmem_cache_destroy(cachep);
2487 unlock_cpu_hotplug();
2490 EXPORT_SYMBOL(kmem_cache_destroy);
2493 * Get the memory for a slab management obj.
2494 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2495 * always come from malloc_sizes caches. The slab descriptor cannot
2496 * come from the same cache which is getting created because,
2497 * when we are searching for an appropriate cache for these
2498 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2499 * If we are creating a malloc_sizes cache here it would not be visible to
2500 * kmem_find_general_cachep till the initialization is complete.
2501 * Hence we cannot have slabp_cache same as the original cache.
2503 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2504 int colour_off, gfp_t local_flags,
2509 if (OFF_SLAB(cachep)) {
2510 /* Slab management obj is off-slab. */
2511 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2512 local_flags, nodeid);
2516 slabp = objp + colour_off;
2517 colour_off += cachep->slab_size;
2520 slabp->colouroff = colour_off;
2521 slabp->s_mem = objp + colour_off;
2522 slabp->nodeid = nodeid;
2526 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2528 return (kmem_bufctl_t *) (slabp + 1);
2531 static void cache_init_objs(struct kmem_cache *cachep,
2532 struct slab *slabp, unsigned long ctor_flags)
2536 for (i = 0; i < cachep->num; i++) {
2537 void *objp = index_to_obj(cachep, slabp, i);
2539 /* need to poison the objs? */
2540 if (cachep->flags & SLAB_POISON)
2541 poison_obj(cachep, objp, POISON_FREE);
2542 if (cachep->flags & SLAB_STORE_USER)
2543 *dbg_userword(cachep, objp) = NULL;
2545 if (cachep->flags & SLAB_RED_ZONE) {
2546 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2547 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2550 * Constructors are not allowed to allocate memory from the same
2551 * cache which they are a constructor for. Otherwise, deadlock.
2552 * They must also be threaded.
2554 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2555 cachep->ctor(objp + obj_offset(cachep), cachep,
2558 if (cachep->flags & SLAB_RED_ZONE) {
2559 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2560 slab_error(cachep, "constructor overwrote the"
2561 " end of an object");
2562 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2563 slab_error(cachep, "constructor overwrote the"
2564 " start of an object");
2566 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2567 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2568 kernel_map_pages(virt_to_page(objp),
2569 cachep->buffer_size / PAGE_SIZE, 0);
2572 cachep->ctor(objp, cachep, ctor_flags);
2574 slab_bufctl(slabp)[i] = i + 1;
2576 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2580 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2582 if (flags & SLAB_DMA)
2583 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2585 BUG_ON(cachep->gfpflags & GFP_DMA);
2588 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2591 void *objp = index_to_obj(cachep, slabp, slabp->free);
2595 next = slab_bufctl(slabp)[slabp->free];
2597 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2598 WARN_ON(slabp->nodeid != nodeid);
2605 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2606 void *objp, int nodeid)
2608 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2611 /* Verify that the slab belongs to the intended node */
2612 WARN_ON(slabp->nodeid != nodeid);
2614 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2615 printk(KERN_ERR "slab: double free detected in cache "
2616 "'%s', objp %p\n", cachep->name, objp);
2620 slab_bufctl(slabp)[objnr] = slabp->free;
2621 slabp->free = objnr;
2626 * Map pages beginning at addr to the given cache and slab. This is required
2627 * for the slab allocator to be able to lookup the cache and slab of a
2628 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2630 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2636 page = virt_to_page(addr);
2639 if (likely(!PageCompound(page)))
2640 nr_pages <<= cache->gfporder;
2643 page_set_cache(page, cache);
2644 page_set_slab(page, slab);
2646 } while (--nr_pages);
2650 * Grow (by 1) the number of slabs within a cache. This is called by
2651 * kmem_cache_alloc() when there are no active objs left in a cache.
2653 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2659 unsigned long ctor_flags;
2660 struct kmem_list3 *l3;
2663 * Be lazy and only check for valid flags here, keeping it out of the
2664 * critical path in kmem_cache_alloc().
2666 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2667 if (flags & SLAB_NO_GROW)
2670 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2671 local_flags = (flags & SLAB_LEVEL_MASK);
2672 if (!(local_flags & __GFP_WAIT))
2674 * Not allowed to sleep. Need to tell a constructor about
2675 * this - it might need to know...
2677 ctor_flags |= SLAB_CTOR_ATOMIC;
2679 /* Take the l3 list lock to change the colour_next on this node */
2681 l3 = cachep->nodelists[nodeid];
2682 spin_lock(&l3->list_lock);
2684 /* Get colour for the slab, and cal the next value. */
2685 offset = l3->colour_next;
2687 if (l3->colour_next >= cachep->colour)
2688 l3->colour_next = 0;
2689 spin_unlock(&l3->list_lock);
2691 offset *= cachep->colour_off;
2693 if (local_flags & __GFP_WAIT)
2697 * The test for missing atomic flag is performed here, rather than
2698 * the more obvious place, simply to reduce the critical path length
2699 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2700 * will eventually be caught here (where it matters).
2702 kmem_flagcheck(cachep, flags);
2705 * Get mem for the objs. Attempt to allocate a physical page from
2708 objp = kmem_getpages(cachep, flags, nodeid);
2712 /* Get slab management. */
2713 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2717 slabp->nodeid = nodeid;
2718 slab_map_pages(cachep, slabp, objp);
2720 cache_init_objs(cachep, slabp, ctor_flags);
2722 if (local_flags & __GFP_WAIT)
2723 local_irq_disable();
2725 spin_lock(&l3->list_lock);
2727 /* Make slab active. */
2728 list_add_tail(&slabp->list, &(l3->slabs_free));
2729 STATS_INC_GROWN(cachep);
2730 l3->free_objects += cachep->num;
2731 spin_unlock(&l3->list_lock);
2734 kmem_freepages(cachep, objp);
2736 if (local_flags & __GFP_WAIT)
2737 local_irq_disable();
2744 * Perform extra freeing checks:
2745 * - detect bad pointers.
2746 * - POISON/RED_ZONE checking
2747 * - destructor calls, for caches with POISON+dtor
2749 static void kfree_debugcheck(const void *objp)
2753 if (!virt_addr_valid(objp)) {
2754 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2755 (unsigned long)objp);
2758 page = virt_to_page(objp);
2759 if (!PageSlab(page)) {
2760 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2761 (unsigned long)objp);
2766 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2768 unsigned long redzone1, redzone2;
2770 redzone1 = *dbg_redzone1(cache, obj);
2771 redzone2 = *dbg_redzone2(cache, obj);
2776 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2779 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2780 slab_error(cache, "double free detected");
2782 slab_error(cache, "memory outside object was overwritten");
2784 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2785 obj, redzone1, redzone2);
2788 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2795 objp -= obj_offset(cachep);
2796 kfree_debugcheck(objp);
2797 page = virt_to_page(objp);
2799 slabp = page_get_slab(page);
2801 if (cachep->flags & SLAB_RED_ZONE) {
2802 verify_redzone_free(cachep, objp);
2803 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2804 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2806 if (cachep->flags & SLAB_STORE_USER)
2807 *dbg_userword(cachep, objp) = caller;
2809 objnr = obj_to_index(cachep, slabp, objp);
2811 BUG_ON(objnr >= cachep->num);
2812 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2814 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2816 * Need to call the slab's constructor so the caller can
2817 * perform a verify of its state (debugging). Called without
2818 * the cache-lock held.
2820 cachep->ctor(objp + obj_offset(cachep),
2821 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2823 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2824 /* we want to cache poison the object,
2825 * call the destruction callback
2827 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2829 #ifdef CONFIG_DEBUG_SLAB_LEAK
2830 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2832 if (cachep->flags & SLAB_POISON) {
2833 #ifdef CONFIG_DEBUG_PAGEALLOC
2834 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2835 store_stackinfo(cachep, objp, (unsigned long)caller);
2836 kernel_map_pages(virt_to_page(objp),
2837 cachep->buffer_size / PAGE_SIZE, 0);
2839 poison_obj(cachep, objp, POISON_FREE);
2842 poison_obj(cachep, objp, POISON_FREE);
2848 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2853 /* Check slab's freelist to see if this obj is there. */
2854 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2856 if (entries > cachep->num || i >= cachep->num)
2859 if (entries != cachep->num - slabp->inuse) {
2861 printk(KERN_ERR "slab: Internal list corruption detected in "
2862 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2863 cachep->name, cachep->num, slabp, slabp->inuse);
2865 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2868 printk("\n%03x:", i);
2869 printk(" %02x", ((unsigned char *)slabp)[i]);
2876 #define kfree_debugcheck(x) do { } while(0)
2877 #define cache_free_debugcheck(x,objp,z) (objp)
2878 #define check_slabp(x,y) do { } while(0)
2881 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2884 struct kmem_list3 *l3;
2885 struct array_cache *ac;
2888 ac = cpu_cache_get(cachep);
2890 batchcount = ac->batchcount;
2891 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2893 * If there was little recent activity on this cache, then
2894 * perform only a partial refill. Otherwise we could generate
2897 batchcount = BATCHREFILL_LIMIT;
2899 l3 = cachep->nodelists[numa_node_id()];
2901 BUG_ON(ac->avail > 0 || !l3);
2902 spin_lock(&l3->list_lock);
2904 /* See if we can refill from the shared array */
2905 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2908 while (batchcount > 0) {
2909 struct list_head *entry;
2911 /* Get slab alloc is to come from. */
2912 entry = l3->slabs_partial.next;
2913 if (entry == &l3->slabs_partial) {
2914 l3->free_touched = 1;
2915 entry = l3->slabs_free.next;
2916 if (entry == &l3->slabs_free)
2920 slabp = list_entry(entry, struct slab, list);
2921 check_slabp(cachep, slabp);
2922 check_spinlock_acquired(cachep);
2923 while (slabp->inuse < cachep->num && batchcount--) {
2924 STATS_INC_ALLOCED(cachep);
2925 STATS_INC_ACTIVE(cachep);
2926 STATS_SET_HIGH(cachep);
2928 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2931 check_slabp(cachep, slabp);
2933 /* move slabp to correct slabp list: */
2934 list_del(&slabp->list);
2935 if (slabp->free == BUFCTL_END)
2936 list_add(&slabp->list, &l3->slabs_full);
2938 list_add(&slabp->list, &l3->slabs_partial);
2942 l3->free_objects -= ac->avail;
2944 spin_unlock(&l3->list_lock);
2946 if (unlikely(!ac->avail)) {
2948 x = cache_grow(cachep, flags, numa_node_id());
2950 /* cache_grow can reenable interrupts, then ac could change. */
2951 ac = cpu_cache_get(cachep);
2952 if (!x && ac->avail == 0) /* no objects in sight? abort */
2955 if (!ac->avail) /* objects refilled by interrupt? */
2959 return ac->entry[--ac->avail];
2962 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2965 might_sleep_if(flags & __GFP_WAIT);
2967 kmem_flagcheck(cachep, flags);
2972 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2973 gfp_t flags, void *objp, void *caller)
2977 if (cachep->flags & SLAB_POISON) {
2978 #ifdef CONFIG_DEBUG_PAGEALLOC
2979 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2980 kernel_map_pages(virt_to_page(objp),
2981 cachep->buffer_size / PAGE_SIZE, 1);
2983 check_poison_obj(cachep, objp);
2985 check_poison_obj(cachep, objp);
2987 poison_obj(cachep, objp, POISON_INUSE);
2989 if (cachep->flags & SLAB_STORE_USER)
2990 *dbg_userword(cachep, objp) = caller;
2992 if (cachep->flags & SLAB_RED_ZONE) {
2993 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2994 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2995 slab_error(cachep, "double free, or memory outside"
2996 " object was overwritten");
2998 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2999 objp, *dbg_redzone1(cachep, objp),
3000 *dbg_redzone2(cachep, objp));
3002 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3003 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3005 #ifdef CONFIG_DEBUG_SLAB_LEAK
3010 slabp = page_get_slab(virt_to_page(objp));
3011 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3012 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3015 objp += obj_offset(cachep);
3016 if (cachep->ctor && cachep->flags & SLAB_POISON) {
3017 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
3019 if (!(flags & __GFP_WAIT))
3020 ctor_flags |= SLAB_CTOR_ATOMIC;
3022 cachep->ctor(objp, cachep, ctor_flags);
3027 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3030 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3033 struct array_cache *ac;
3036 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3037 objp = alternate_node_alloc(cachep, flags);
3044 ac = cpu_cache_get(cachep);
3045 if (likely(ac->avail)) {
3046 STATS_INC_ALLOCHIT(cachep);
3048 objp = ac->entry[--ac->avail];
3050 STATS_INC_ALLOCMISS(cachep);
3051 objp = cache_alloc_refill(cachep, flags);
3056 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3057 gfp_t flags, void *caller)
3059 unsigned long save_flags;
3062 cache_alloc_debugcheck_before(cachep, flags);
3064 local_irq_save(save_flags);
3065 objp = ____cache_alloc(cachep, flags);
3066 local_irq_restore(save_flags);
3067 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3075 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3077 * If we are in_interrupt, then process context, including cpusets and
3078 * mempolicy, may not apply and should not be used for allocation policy.
3080 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3082 int nid_alloc, nid_here;
3086 nid_alloc = nid_here = numa_node_id();
3087 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3088 nid_alloc = cpuset_mem_spread_node();
3089 else if (current->mempolicy)
3090 nid_alloc = slab_node(current->mempolicy);
3091 if (nid_alloc != nid_here)
3092 return __cache_alloc_node(cachep, flags, nid_alloc);
3097 * A interface to enable slab creation on nodeid
3099 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3102 struct list_head *entry;
3104 struct kmem_list3 *l3;
3108 l3 = cachep->nodelists[nodeid];
3113 spin_lock(&l3->list_lock);
3114 entry = l3->slabs_partial.next;
3115 if (entry == &l3->slabs_partial) {
3116 l3->free_touched = 1;
3117 entry = l3->slabs_free.next;
3118 if (entry == &l3->slabs_free)
3122 slabp = list_entry(entry, struct slab, list);
3123 check_spinlock_acquired_node(cachep, nodeid);
3124 check_slabp(cachep, slabp);
3126 STATS_INC_NODEALLOCS(cachep);
3127 STATS_INC_ACTIVE(cachep);
3128 STATS_SET_HIGH(cachep);
3130 BUG_ON(slabp->inuse == cachep->num);
3132 obj = slab_get_obj(cachep, slabp, nodeid);
3133 check_slabp(cachep, slabp);
3135 /* move slabp to correct slabp list: */
3136 list_del(&slabp->list);
3138 if (slabp->free == BUFCTL_END)
3139 list_add(&slabp->list, &l3->slabs_full);
3141 list_add(&slabp->list, &l3->slabs_partial);
3143 spin_unlock(&l3->list_lock);
3147 spin_unlock(&l3->list_lock);
3148 x = cache_grow(cachep, flags, nodeid);
3160 * Caller needs to acquire correct kmem_list's list_lock
3162 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3166 struct kmem_list3 *l3;
3168 for (i = 0; i < nr_objects; i++) {
3169 void *objp = objpp[i];
3172 slabp = virt_to_slab(objp);
3173 l3 = cachep->nodelists[node];
3174 list_del(&slabp->list);
3175 check_spinlock_acquired_node(cachep, node);
3176 check_slabp(cachep, slabp);
3177 slab_put_obj(cachep, slabp, objp, node);
3178 STATS_DEC_ACTIVE(cachep);
3180 check_slabp(cachep, slabp);
3182 /* fixup slab chains */
3183 if (slabp->inuse == 0) {
3184 if (l3->free_objects > l3->free_limit) {
3185 l3->free_objects -= cachep->num;
3186 /* No need to drop any previously held
3187 * lock here, even if we have a off-slab slab
3188 * descriptor it is guaranteed to come from
3189 * a different cache, refer to comments before
3192 slab_destroy(cachep, slabp);
3194 list_add(&slabp->list, &l3->slabs_free);
3197 /* Unconditionally move a slab to the end of the
3198 * partial list on free - maximum time for the
3199 * other objects to be freed, too.
3201 list_add_tail(&slabp->list, &l3->slabs_partial);
3206 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3209 struct kmem_list3 *l3;
3210 int node = numa_node_id();
3212 batchcount = ac->batchcount;
3214 BUG_ON(!batchcount || batchcount > ac->avail);
3217 l3 = cachep->nodelists[node];
3218 spin_lock(&l3->list_lock);
3220 struct array_cache *shared_array = l3->shared;
3221 int max = shared_array->limit - shared_array->avail;
3223 if (batchcount > max)
3225 memcpy(&(shared_array->entry[shared_array->avail]),
3226 ac->entry, sizeof(void *) * batchcount);
3227 shared_array->avail += batchcount;
3232 free_block(cachep, ac->entry, batchcount, node);
3237 struct list_head *p;
3239 p = l3->slabs_free.next;
3240 while (p != &(l3->slabs_free)) {
3243 slabp = list_entry(p, struct slab, list);
3244 BUG_ON(slabp->inuse);
3249 STATS_SET_FREEABLE(cachep, i);
3252 spin_unlock(&l3->list_lock);
3253 ac->avail -= batchcount;
3254 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3258 * Release an obj back to its cache. If the obj has a constructed state, it must
3259 * be in this state _before_ it is released. Called with disabled ints.
3261 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3263 struct array_cache *ac = cpu_cache_get(cachep);
3266 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3268 if (cache_free_alien(cachep, objp))
3271 if (likely(ac->avail < ac->limit)) {
3272 STATS_INC_FREEHIT(cachep);
3273 ac->entry[ac->avail++] = objp;
3276 STATS_INC_FREEMISS(cachep);
3277 cache_flusharray(cachep, ac);
3278 ac->entry[ac->avail++] = objp;
3283 * kmem_cache_alloc - Allocate an object
3284 * @cachep: The cache to allocate from.
3285 * @flags: See kmalloc().
3287 * Allocate an object from this cache. The flags are only relevant
3288 * if the cache has no available objects.
3290 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3292 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3294 EXPORT_SYMBOL(kmem_cache_alloc);
3297 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3298 * @cache: The cache to allocate from.
3299 * @flags: See kmalloc().
3301 * Allocate an object from this cache and set the allocated memory to zero.
3302 * The flags are only relevant if the cache has no available objects.
3304 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3306 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3308 memset(ret, 0, obj_size(cache));
3311 EXPORT_SYMBOL(kmem_cache_zalloc);
3314 * kmem_ptr_validate - check if an untrusted pointer might
3316 * @cachep: the cache we're checking against
3317 * @ptr: pointer to validate
3319 * This verifies that the untrusted pointer looks sane:
3320 * it is _not_ a guarantee that the pointer is actually
3321 * part of the slab cache in question, but it at least
3322 * validates that the pointer can be dereferenced and
3323 * looks half-way sane.
3325 * Currently only used for dentry validation.
3327 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3329 unsigned long addr = (unsigned long)ptr;
3330 unsigned long min_addr = PAGE_OFFSET;
3331 unsigned long align_mask = BYTES_PER_WORD - 1;
3332 unsigned long size = cachep->buffer_size;
3335 if (unlikely(addr < min_addr))
3337 if (unlikely(addr > (unsigned long)high_memory - size))
3339 if (unlikely(addr & align_mask))
3341 if (unlikely(!kern_addr_valid(addr)))
3343 if (unlikely(!kern_addr_valid(addr + size - 1)))
3345 page = virt_to_page(ptr);
3346 if (unlikely(!PageSlab(page)))
3348 if (unlikely(page_get_cache(page) != cachep))
3357 * kmem_cache_alloc_node - Allocate an object on the specified node
3358 * @cachep: The cache to allocate from.
3359 * @flags: See kmalloc().
3360 * @nodeid: node number of the target node.
3362 * Identical to kmem_cache_alloc, except that this function is slow
3363 * and can sleep. And it will allocate memory on the given node, which
3364 * can improve the performance for cpu bound structures.
3365 * New and improved: it will now make sure that the object gets
3366 * put on the correct node list so that there is no false sharing.
3368 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3370 unsigned long save_flags;
3373 cache_alloc_debugcheck_before(cachep, flags);
3374 local_irq_save(save_flags);
3376 if (nodeid == -1 || nodeid == numa_node_id() ||
3377 !cachep->nodelists[nodeid])
3378 ptr = ____cache_alloc(cachep, flags);
3380 ptr = __cache_alloc_node(cachep, flags, nodeid);
3381 local_irq_restore(save_flags);
3383 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3384 __builtin_return_address(0));
3388 EXPORT_SYMBOL(kmem_cache_alloc_node);
3390 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3392 struct kmem_cache *cachep;
3394 cachep = kmem_find_general_cachep(size, flags);
3395 if (unlikely(cachep == NULL))
3397 return kmem_cache_alloc_node(cachep, flags, node);
3399 EXPORT_SYMBOL(__kmalloc_node);
3403 * __do_kmalloc - allocate memory
3404 * @size: how many bytes of memory are required.
3405 * @flags: the type of memory to allocate (see kmalloc).
3406 * @caller: function caller for debug tracking of the caller
3408 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3411 struct kmem_cache *cachep;
3413 /* If you want to save a few bytes .text space: replace
3415 * Then kmalloc uses the uninlined functions instead of the inline
3418 cachep = __find_general_cachep(size, flags);
3419 if (unlikely(cachep == NULL))
3421 return __cache_alloc(cachep, flags, caller);
3425 void *__kmalloc(size_t size, gfp_t flags)
3427 #ifndef CONFIG_DEBUG_SLAB
3428 return __do_kmalloc(size, flags, NULL);
3430 return __do_kmalloc(size, flags, __builtin_return_address(0));
3433 EXPORT_SYMBOL(__kmalloc);
3435 #ifdef CONFIG_DEBUG_SLAB
3436 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3438 return __do_kmalloc(size, flags, caller);
3440 EXPORT_SYMBOL(__kmalloc_track_caller);
3444 * kmem_cache_free - Deallocate an object
3445 * @cachep: The cache the allocation was from.
3446 * @objp: The previously allocated object.
3448 * Free an object which was previously allocated from this
3451 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3453 unsigned long flags;
3455 BUG_ON(virt_to_cache(objp) != cachep);
3457 local_irq_save(flags);
3458 __cache_free(cachep, objp);
3459 local_irq_restore(flags);
3461 EXPORT_SYMBOL(kmem_cache_free);
3464 * kfree - free previously allocated memory
3465 * @objp: pointer returned by kmalloc.
3467 * If @objp is NULL, no operation is performed.
3469 * Don't free memory not originally allocated by kmalloc()
3470 * or you will run into trouble.
3472 void kfree(const void *objp)
3474 struct kmem_cache *c;
3475 unsigned long flags;
3477 if (unlikely(!objp))
3479 local_irq_save(flags);
3480 kfree_debugcheck(objp);
3481 c = virt_to_cache(objp);
3482 debug_check_no_locks_freed(objp, obj_size(c));
3483 __cache_free(c, (void *)objp);
3484 local_irq_restore(flags);
3486 EXPORT_SYMBOL(kfree);
3488 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3490 return obj_size(cachep);
3492 EXPORT_SYMBOL(kmem_cache_size);
3494 const char *kmem_cache_name(struct kmem_cache *cachep)
3496 return cachep->name;
3498 EXPORT_SYMBOL_GPL(kmem_cache_name);
3501 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3503 static int alloc_kmemlist(struct kmem_cache *cachep)
3506 struct kmem_list3 *l3;
3507 struct array_cache *new_shared;
3508 struct array_cache **new_alien;
3510 for_each_online_node(node) {
3512 new_alien = alloc_alien_cache(node, cachep->limit);
3516 new_shared = alloc_arraycache(node,
3517 cachep->shared*cachep->batchcount,
3520 free_alien_cache(new_alien);
3524 l3 = cachep->nodelists[node];
3526 struct array_cache *shared = l3->shared;
3528 spin_lock_irq(&l3->list_lock);
3531 free_block(cachep, shared->entry,
3532 shared->avail, node);
3534 l3->shared = new_shared;
3536 l3->alien = new_alien;
3539 l3->free_limit = (1 + nr_cpus_node(node)) *
3540 cachep->batchcount + cachep->num;
3541 spin_unlock_irq(&l3->list_lock);
3543 free_alien_cache(new_alien);
3546 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3548 free_alien_cache(new_alien);
3553 kmem_list3_init(l3);
3554 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3555 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3556 l3->shared = new_shared;
3557 l3->alien = new_alien;
3558 l3->free_limit = (1 + nr_cpus_node(node)) *
3559 cachep->batchcount + cachep->num;
3560 cachep->nodelists[node] = l3;
3565 if (!cachep->next.next) {
3566 /* Cache is not active yet. Roll back what we did */
3569 if (cachep->nodelists[node]) {
3570 l3 = cachep->nodelists[node];
3573 free_alien_cache(l3->alien);
3575 cachep->nodelists[node] = NULL;
3583 struct ccupdate_struct {
3584 struct kmem_cache *cachep;
3585 struct array_cache *new[NR_CPUS];
3588 static void do_ccupdate_local(void *info)
3590 struct ccupdate_struct *new = info;
3591 struct array_cache *old;
3594 old = cpu_cache_get(new->cachep);
3596 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3597 new->new[smp_processor_id()] = old;
3600 /* Always called with the cache_chain_mutex held */
3601 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3602 int batchcount, int shared)
3604 struct ccupdate_struct *new;
3607 new = kzalloc(sizeof(*new), GFP_KERNEL);
3611 for_each_online_cpu(i) {
3612 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3615 for (i--; i >= 0; i--)
3621 new->cachep = cachep;
3623 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3626 cachep->batchcount = batchcount;
3627 cachep->limit = limit;
3628 cachep->shared = shared;
3630 for_each_online_cpu(i) {
3631 struct array_cache *ccold = new->new[i];
3634 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3635 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3636 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3640 return alloc_kmemlist(cachep);
3643 /* Called with cache_chain_mutex held always */
3644 static int enable_cpucache(struct kmem_cache *cachep)
3650 * The head array serves three purposes:
3651 * - create a LIFO ordering, i.e. return objects that are cache-warm
3652 * - reduce the number of spinlock operations.
3653 * - reduce the number of linked list operations on the slab and
3654 * bufctl chains: array operations are cheaper.
3655 * The numbers are guessed, we should auto-tune as described by
3658 if (cachep->buffer_size > 131072)
3660 else if (cachep->buffer_size > PAGE_SIZE)
3662 else if (cachep->buffer_size > 1024)
3664 else if (cachep->buffer_size > 256)
3670 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3671 * allocation behaviour: Most allocs on one cpu, most free operations
3672 * on another cpu. For these cases, an efficient object passing between
3673 * cpus is necessary. This is provided by a shared array. The array
3674 * replaces Bonwick's magazine layer.
3675 * On uniprocessor, it's functionally equivalent (but less efficient)
3676 * to a larger limit. Thus disabled by default.
3680 if (cachep->buffer_size <= PAGE_SIZE)
3686 * With debugging enabled, large batchcount lead to excessively long
3687 * periods with disabled local interrupts. Limit the batchcount
3692 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3694 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3695 cachep->name, -err);
3700 * Drain an array if it contains any elements taking the l3 lock only if
3701 * necessary. Note that the l3 listlock also protects the array_cache
3702 * if drain_array() is used on the shared array.
3704 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3705 struct array_cache *ac, int force, int node)
3709 if (!ac || !ac->avail)
3711 if (ac->touched && !force) {
3714 spin_lock_irq(&l3->list_lock);
3716 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3717 if (tofree > ac->avail)
3718 tofree = (ac->avail + 1) / 2;
3719 free_block(cachep, ac->entry, tofree, node);
3720 ac->avail -= tofree;
3721 memmove(ac->entry, &(ac->entry[tofree]),
3722 sizeof(void *) * ac->avail);
3724 spin_unlock_irq(&l3->list_lock);
3729 * cache_reap - Reclaim memory from caches.
3730 * @unused: unused parameter
3732 * Called from workqueue/eventd every few seconds.
3734 * - clear the per-cpu caches for this CPU.
3735 * - return freeable pages to the main free memory pool.
3737 * If we cannot acquire the cache chain mutex then just give up - we'll try
3738 * again on the next iteration.
3740 static void cache_reap(void *unused)
3742 struct kmem_cache *searchp;
3743 struct kmem_list3 *l3;
3744 int node = numa_node_id();
3746 if (!mutex_trylock(&cache_chain_mutex)) {
3747 /* Give up. Setup the next iteration. */
3748 schedule_delayed_work(&__get_cpu_var(reap_work),
3753 list_for_each_entry(searchp, &cache_chain, next) {
3757 * We only take the l3 lock if absolutely necessary and we
3758 * have established with reasonable certainty that
3759 * we can do some work if the lock was obtained.
3761 l3 = searchp->nodelists[node];
3763 reap_alien(searchp, l3);
3765 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3768 * These are racy checks but it does not matter
3769 * if we skip one check or scan twice.
3771 if (time_after(l3->next_reap, jiffies))
3774 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3776 drain_array(searchp, l3, l3->shared, 0, node);
3778 if (l3->free_touched)
3779 l3->free_touched = 0;
3783 freed = drain_freelist(searchp, l3, (l3->free_limit +
3784 5 * searchp->num - 1) / (5 * searchp->num));
3785 STATS_ADD_REAPED(searchp, freed);
3791 mutex_unlock(&cache_chain_mutex);
3793 refresh_cpu_vm_stats(smp_processor_id());
3794 /* Set up the next iteration */
3795 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3798 #ifdef CONFIG_PROC_FS
3800 static void print_slabinfo_header(struct seq_file *m)
3803 * Output format version, so at least we can change it
3804 * without _too_ many complaints.
3807 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3809 seq_puts(m, "slabinfo - version: 2.1\n");
3811 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3812 "<objperslab> <pagesperslab>");
3813 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3814 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3816 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3817 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3818 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3823 static void *s_start(struct seq_file *m, loff_t *pos)
3826 struct list_head *p;
3828 mutex_lock(&cache_chain_mutex);
3830 print_slabinfo_header(m);
3831 p = cache_chain.next;
3834 if (p == &cache_chain)
3837 return list_entry(p, struct kmem_cache, next);
3840 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3842 struct kmem_cache *cachep = p;
3844 return cachep->next.next == &cache_chain ?
3845 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3848 static void s_stop(struct seq_file *m, void *p)
3850 mutex_unlock(&cache_chain_mutex);
3853 static int s_show(struct seq_file *m, void *p)
3855 struct kmem_cache *cachep = p;
3857 unsigned long active_objs;
3858 unsigned long num_objs;
3859 unsigned long active_slabs = 0;
3860 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3864 struct kmem_list3 *l3;
3868 for_each_online_node(node) {
3869 l3 = cachep->nodelists[node];
3874 spin_lock_irq(&l3->list_lock);
3876 list_for_each_entry(slabp, &l3->slabs_full, list) {
3877 if (slabp->inuse != cachep->num && !error)
3878 error = "slabs_full accounting error";
3879 active_objs += cachep->num;
3882 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3883 if (slabp->inuse == cachep->num && !error)
3884 error = "slabs_partial inuse accounting error";
3885 if (!slabp->inuse && !error)
3886 error = "slabs_partial/inuse accounting error";
3887 active_objs += slabp->inuse;
3890 list_for_each_entry(slabp, &l3->slabs_free, list) {
3891 if (slabp->inuse && !error)
3892 error = "slabs_free/inuse accounting error";
3895 free_objects += l3->free_objects;
3897 shared_avail += l3->shared->avail;
3899 spin_unlock_irq(&l3->list_lock);
3901 num_slabs += active_slabs;
3902 num_objs = num_slabs * cachep->num;
3903 if (num_objs - active_objs != free_objects && !error)
3904 error = "free_objects accounting error";
3906 name = cachep->name;
3908 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3910 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3911 name, active_objs, num_objs, cachep->buffer_size,
3912 cachep->num, (1 << cachep->gfporder));
3913 seq_printf(m, " : tunables %4u %4u %4u",
3914 cachep->limit, cachep->batchcount, cachep->shared);
3915 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3916 active_slabs, num_slabs, shared_avail);
3919 unsigned long high = cachep->high_mark;
3920 unsigned long allocs = cachep->num_allocations;
3921 unsigned long grown = cachep->grown;
3922 unsigned long reaped = cachep->reaped;
3923 unsigned long errors = cachep->errors;
3924 unsigned long max_freeable = cachep->max_freeable;
3925 unsigned long node_allocs = cachep->node_allocs;
3926 unsigned long node_frees = cachep->node_frees;
3927 unsigned long overflows = cachep->node_overflow;
3929 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3930 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3931 reaped, errors, max_freeable, node_allocs,
3932 node_frees, overflows);
3936 unsigned long allochit = atomic_read(&cachep->allochit);
3937 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3938 unsigned long freehit = atomic_read(&cachep->freehit);
3939 unsigned long freemiss = atomic_read(&cachep->freemiss);
3941 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3942 allochit, allocmiss, freehit, freemiss);
3950 * slabinfo_op - iterator that generates /proc/slabinfo
3959 * num-pages-per-slab
3960 * + further values on SMP and with statistics enabled
3963 struct seq_operations slabinfo_op = {
3970 #define MAX_SLABINFO_WRITE 128
3972 * slabinfo_write - Tuning for the slab allocator
3974 * @buffer: user buffer
3975 * @count: data length
3978 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3979 size_t count, loff_t *ppos)
3981 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3982 int limit, batchcount, shared, res;
3983 struct kmem_cache *cachep;
3985 if (count > MAX_SLABINFO_WRITE)
3987 if (copy_from_user(&kbuf, buffer, count))
3989 kbuf[MAX_SLABINFO_WRITE] = '\0';
3991 tmp = strchr(kbuf, ' ');
3996 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3999 /* Find the cache in the chain of caches. */
4000 mutex_lock(&cache_chain_mutex);
4002 list_for_each_entry(cachep, &cache_chain, next) {
4003 if (!strcmp(cachep->name, kbuf)) {
4004 if (limit < 1 || batchcount < 1 ||
4005 batchcount > limit || shared < 0) {
4008 res = do_tune_cpucache(cachep, limit,
4009 batchcount, shared);
4014 mutex_unlock(&cache_chain_mutex);
4020 #ifdef CONFIG_DEBUG_SLAB_LEAK
4022 static void *leaks_start(struct seq_file *m, loff_t *pos)
4025 struct list_head *p;
4027 mutex_lock(&cache_chain_mutex);
4028 p = cache_chain.next;
4031 if (p == &cache_chain)
4034 return list_entry(p, struct kmem_cache, next);
4037 static inline int add_caller(unsigned long *n, unsigned long v)
4047 unsigned long *q = p + 2 * i;
4061 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4067 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4073 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4074 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4076 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4081 static void show_symbol(struct seq_file *m, unsigned long address)
4083 #ifdef CONFIG_KALLSYMS
4086 unsigned long offset, size;
4087 char namebuf[KSYM_NAME_LEN+1];
4089 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4092 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4094 seq_printf(m, " [%s]", modname);
4098 seq_printf(m, "%p", (void *)address);
4101 static int leaks_show(struct seq_file *m, void *p)
4103 struct kmem_cache *cachep = p;
4105 struct kmem_list3 *l3;
4107 unsigned long *n = m->private;
4111 if (!(cachep->flags & SLAB_STORE_USER))
4113 if (!(cachep->flags & SLAB_RED_ZONE))
4116 /* OK, we can do it */
4120 for_each_online_node(node) {
4121 l3 = cachep->nodelists[node];
4126 spin_lock_irq(&l3->list_lock);
4128 list_for_each_entry(slabp, &l3->slabs_full, list)
4129 handle_slab(n, cachep, slabp);
4130 list_for_each_entry(slabp, &l3->slabs_partial, list)
4131 handle_slab(n, cachep, slabp);
4132 spin_unlock_irq(&l3->list_lock);
4134 name = cachep->name;
4136 /* Increase the buffer size */
4137 mutex_unlock(&cache_chain_mutex);
4138 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4140 /* Too bad, we are really out */
4142 mutex_lock(&cache_chain_mutex);
4145 *(unsigned long *)m->private = n[0] * 2;
4147 mutex_lock(&cache_chain_mutex);
4148 /* Now make sure this entry will be retried */
4152 for (i = 0; i < n[1]; i++) {
4153 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4154 show_symbol(m, n[2*i+2]);
4161 struct seq_operations slabstats_op = {
4162 .start = leaks_start,
4171 * ksize - get the actual amount of memory allocated for a given object
4172 * @objp: Pointer to the object
4174 * kmalloc may internally round up allocations and return more memory
4175 * than requested. ksize() can be used to determine the actual amount of
4176 * memory allocated. The caller may use this additional memory, even though
4177 * a smaller amount of memory was initially specified with the kmalloc call.
4178 * The caller must guarantee that objp points to a valid object previously
4179 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4180 * must not be freed during the duration of the call.
4182 unsigned int ksize(const void *objp)
4184 if (unlikely(objp == NULL))
4187 return obj_size(virt_to_cache(objp));