3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
117 #include <asm/cacheflush.h>
118 #include <asm/tlbflush.h>
119 #include <asm/page.h>
122 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * STATS - 1 to collect stats for /proc/slabinfo.
126 * 0 for faster, smaller code (especially in the critical paths).
128 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 #ifdef CONFIG_DEBUG_SLAB
134 #define FORCED_DEBUG 1
138 #define FORCED_DEBUG 0
141 /* Shouldn't this be in a header file somewhere? */
142 #define BYTES_PER_WORD sizeof(void *)
143 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
183 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
193 * Bufctl's are used for linking objs within a slab
196 * This implementation relies on "struct page" for locating the cache &
197 * slab an object belongs to.
198 * This allows the bufctl structure to be small (one int), but limits
199 * the number of objects a slab (not a cache) can contain when off-slab
200 * bufctls are used. The limit is the size of the largest general cache
201 * that does not use off-slab slabs.
202 * For 32bit archs with 4 kB pages, is this 56.
203 * This is not serious, as it is only for large objects, when it is unwise
204 * to have too many per slab.
205 * Note: This limit can be raised by introducing a general cache whose size
206 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
209 typedef unsigned int kmem_bufctl_t;
210 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
211 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
212 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
213 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
228 unsigned short nodeid;
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
245 * We assume struct slab_rcu can overlay struct slab when destroying.
248 struct rcu_head head;
249 struct kmem_cache *cachep;
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
261 * The limit is stored in the per-cpu structure to reduce the data cache
268 unsigned int batchcount;
269 unsigned int touched;
272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
289 * The slab lists for all objects.
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
309 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
311 #define SIZE_AC MAX_NUMNODES
312 #define SIZE_L3 (2 * MAX_NUMNODES)
314 static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
318 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
319 static void cache_reap(struct work_struct *unused);
322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
325 static __always_inline int index_of(const size_t size)
327 extern void __bad_size(void);
329 if (__builtin_constant_p(size)) {
337 #include <linux/kmalloc_sizes.h>
345 static int slab_early_init = 1;
347 #define INDEX_AC index_of(sizeof(struct arraycache_init))
348 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
350 static void kmem_list3_init(struct kmem_list3 *parent)
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
357 parent->colour_next = 0;
358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
363 #define MAKE_LIST(cachep, listp, slab, nodeid) \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
369 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
383 /* 1) per-cpu data, touched during every alloc/free */
384 struct array_cache *array[NR_CPUS];
385 /* 2) Cache tunables. Protected by cache_chain_mutex */
386 unsigned int batchcount;
390 unsigned int buffer_size;
391 u32 reciprocal_buffer_size;
392 /* 3) touched by every alloc & free from the backend */
394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
397 /* 4) cache_grow/shrink */
398 /* order of pgs per slab (2^n) */
399 unsigned int gfporder;
401 /* force GFP flags, e.g. GFP_DMA */
404 size_t colour; /* cache colouring range */
405 unsigned int colour_off; /* colour offset */
406 struct kmem_cache *slabp_cache;
407 unsigned int slab_size;
408 unsigned int dflags; /* dynamic flags */
410 /* constructor func */
411 void (*ctor)(void *obj);
413 /* 5) cache creation/removal */
415 struct list_head next;
419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
428 unsigned long node_overflow;
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
445 * We put nodelists[] at the end of kmem_cache, because we want to size
446 * this array to nr_node_ids slots instead of MAX_NUMNODES
447 * (see kmem_cache_init())
448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
449 * is statically defined, so we reserve the max number of nodes.
451 struct kmem_list3 *nodelists[MAX_NUMNODES];
453 * Do not add fields after nodelists[]
457 #define CFLGS_OFF_SLAB (0x80000000UL)
458 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
460 #define BATCHREFILL_LIMIT 16
462 * Optimization question: fewer reaps means less probability for unnessary
463 * cpucache drain/refill cycles.
465 * OTOH the cpuarrays can contain lots of objects,
466 * which could lock up otherwise freeable slabs.
468 #define REAPTIMEOUT_CPUC (2*HZ)
469 #define REAPTIMEOUT_LIST3 (4*HZ)
472 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
473 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
474 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
475 #define STATS_INC_GROWN(x) ((x)->grown++)
476 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
477 #define STATS_SET_HIGH(x) \
479 if ((x)->num_active > (x)->high_mark) \
480 (x)->high_mark = (x)->num_active; \
482 #define STATS_INC_ERR(x) ((x)->errors++)
483 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
484 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
485 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
486 #define STATS_SET_FREEABLE(x, i) \
488 if ((x)->max_freeable < i) \
489 (x)->max_freeable = i; \
491 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
492 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
493 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
494 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
496 #define STATS_INC_ACTIVE(x) do { } while (0)
497 #define STATS_DEC_ACTIVE(x) do { } while (0)
498 #define STATS_INC_ALLOCED(x) do { } while (0)
499 #define STATS_INC_GROWN(x) do { } while (0)
500 #define STATS_ADD_REAPED(x,y) do { } while (0)
501 #define STATS_SET_HIGH(x) do { } while (0)
502 #define STATS_INC_ERR(x) do { } while (0)
503 #define STATS_INC_NODEALLOCS(x) do { } while (0)
504 #define STATS_INC_NODEFREES(x) do { } while (0)
505 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
506 #define STATS_SET_FREEABLE(x, i) do { } while (0)
507 #define STATS_INC_ALLOCHIT(x) do { } while (0)
508 #define STATS_INC_ALLOCMISS(x) do { } while (0)
509 #define STATS_INC_FREEHIT(x) do { } while (0)
510 #define STATS_INC_FREEMISS(x) do { } while (0)
516 * memory layout of objects:
518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
519 * the end of an object is aligned with the end of the real
520 * allocation. Catches writes behind the end of the allocation.
521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
523 * cachep->obj_offset: The real object.
524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
526 * [BYTES_PER_WORD long]
528 static int obj_offset(struct kmem_cache *cachep)
530 return cachep->obj_offset;
533 static int obj_size(struct kmem_cache *cachep)
535 return cachep->obj_size;
538 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
541 return (unsigned long long*) (objp + obj_offset(cachep) -
542 sizeof(unsigned long long));
545 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 if (cachep->flags & SLAB_STORE_USER)
549 return (unsigned long long *)(objp + cachep->buffer_size -
550 sizeof(unsigned long long) -
552 return (unsigned long long *) (objp + cachep->buffer_size -
553 sizeof(unsigned long long));
556 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
558 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
564 #define obj_offset(x) 0
565 #define obj_size(cachep) (cachep->buffer_size)
566 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
568 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
572 #ifdef CONFIG_KMEMTRACE
573 size_t slab_buffer_size(struct kmem_cache *cachep)
575 return cachep->buffer_size;
577 EXPORT_SYMBOL(slab_buffer_size);
581 * Do not go above this order unless 0 objects fit into the slab.
583 #define BREAK_GFP_ORDER_HI 1
584 #define BREAK_GFP_ORDER_LO 0
585 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
588 * Functions for storing/retrieving the cachep and or slab from the page
589 * allocator. These are used to find the slab an obj belongs to. With kfree(),
590 * these are used to find the cache which an obj belongs to.
592 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
594 page->lru.next = (struct list_head *)cache;
597 static inline struct kmem_cache *page_get_cache(struct page *page)
599 page = compound_head(page);
600 BUG_ON(!PageSlab(page));
601 return (struct kmem_cache *)page->lru.next;
604 static inline void page_set_slab(struct page *page, struct slab *slab)
606 page->lru.prev = (struct list_head *)slab;
609 static inline struct slab *page_get_slab(struct page *page)
611 BUG_ON(!PageSlab(page));
612 return (struct slab *)page->lru.prev;
615 static inline struct kmem_cache *virt_to_cache(const void *obj)
617 struct page *page = virt_to_head_page(obj);
618 return page_get_cache(page);
621 static inline struct slab *virt_to_slab(const void *obj)
623 struct page *page = virt_to_head_page(obj);
624 return page_get_slab(page);
627 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
630 return slab->s_mem + cache->buffer_size * idx;
634 * We want to avoid an expensive divide : (offset / cache->buffer_size)
635 * Using the fact that buffer_size is a constant for a particular cache,
636 * we can replace (offset / cache->buffer_size) by
637 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
639 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
640 const struct slab *slab, void *obj)
642 u32 offset = (obj - slab->s_mem);
643 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
647 * These are the default caches for kmalloc. Custom caches can have other sizes.
649 struct cache_sizes malloc_sizes[] = {
650 #define CACHE(x) { .cs_size = (x) },
651 #include <linux/kmalloc_sizes.h>
655 EXPORT_SYMBOL(malloc_sizes);
657 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
663 static struct cache_names __initdata cache_names[] = {
664 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
665 #include <linux/kmalloc_sizes.h>
670 static struct arraycache_init initarray_cache __initdata =
671 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
672 static struct arraycache_init initarray_generic =
673 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
675 /* internal cache of cache description objs */
676 static struct kmem_cache cache_cache = {
678 .limit = BOOT_CPUCACHE_ENTRIES,
680 .buffer_size = sizeof(struct kmem_cache),
681 .name = "kmem_cache",
684 #define BAD_ALIEN_MAGIC 0x01020304ul
686 #ifdef CONFIG_LOCKDEP
689 * Slab sometimes uses the kmalloc slabs to store the slab headers
690 * for other slabs "off slab".
691 * The locking for this is tricky in that it nests within the locks
692 * of all other slabs in a few places; to deal with this special
693 * locking we put on-slab caches into a separate lock-class.
695 * We set lock class for alien array caches which are up during init.
696 * The lock annotation will be lost if all cpus of a node goes down and
697 * then comes back up during hotplug
699 static struct lock_class_key on_slab_l3_key;
700 static struct lock_class_key on_slab_alc_key;
702 static inline void init_lock_keys(void)
706 struct cache_sizes *s = malloc_sizes;
708 while (s->cs_size != ULONG_MAX) {
710 struct array_cache **alc;
712 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
713 if (!l3 || OFF_SLAB(s->cs_cachep))
715 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
718 * FIXME: This check for BAD_ALIEN_MAGIC
719 * should go away when common slab code is taught to
720 * work even without alien caches.
721 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
722 * for alloc_alien_cache,
724 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
728 lockdep_set_class(&alc[r]->lock,
736 static inline void init_lock_keys(void)
742 * Guard access to the cache-chain.
744 static DEFINE_MUTEX(cache_chain_mutex);
745 static struct list_head cache_chain;
748 * chicken and egg problem: delay the per-cpu array allocation
749 * until the general caches are up.
759 * used by boot code to determine if it can use slab based allocator
761 int slab_is_available(void)
763 return g_cpucache_up == FULL;
766 static DEFINE_PER_CPU(struct delayed_work, reap_work);
768 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
770 return cachep->array[smp_processor_id()];
773 static inline struct kmem_cache *__find_general_cachep(size_t size,
776 struct cache_sizes *csizep = malloc_sizes;
779 /* This happens if someone tries to call
780 * kmem_cache_create(), or __kmalloc(), before
781 * the generic caches are initialized.
783 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
786 return ZERO_SIZE_PTR;
788 while (size > csizep->cs_size)
792 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
793 * has cs_{dma,}cachep==NULL. Thus no special case
794 * for large kmalloc calls required.
796 #ifdef CONFIG_ZONE_DMA
797 if (unlikely(gfpflags & GFP_DMA))
798 return csizep->cs_dmacachep;
800 return csizep->cs_cachep;
803 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
805 return __find_general_cachep(size, gfpflags);
808 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
810 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
814 * Calculate the number of objects and left-over bytes for a given buffer size.
816 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
817 size_t align, int flags, size_t *left_over,
822 size_t slab_size = PAGE_SIZE << gfporder;
825 * The slab management structure can be either off the slab or
826 * on it. For the latter case, the memory allocated for a
830 * - One kmem_bufctl_t for each object
831 * - Padding to respect alignment of @align
832 * - @buffer_size bytes for each object
834 * If the slab management structure is off the slab, then the
835 * alignment will already be calculated into the size. Because
836 * the slabs are all pages aligned, the objects will be at the
837 * correct alignment when allocated.
839 if (flags & CFLGS_OFF_SLAB) {
841 nr_objs = slab_size / buffer_size;
843 if (nr_objs > SLAB_LIMIT)
844 nr_objs = SLAB_LIMIT;
847 * Ignore padding for the initial guess. The padding
848 * is at most @align-1 bytes, and @buffer_size is at
849 * least @align. In the worst case, this result will
850 * be one greater than the number of objects that fit
851 * into the memory allocation when taking the padding
854 nr_objs = (slab_size - sizeof(struct slab)) /
855 (buffer_size + sizeof(kmem_bufctl_t));
858 * This calculated number will be either the right
859 * amount, or one greater than what we want.
861 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
865 if (nr_objs > SLAB_LIMIT)
866 nr_objs = SLAB_LIMIT;
868 mgmt_size = slab_mgmt_size(nr_objs, align);
871 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
874 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
876 static void __slab_error(const char *function, struct kmem_cache *cachep,
879 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
880 function, cachep->name, msg);
885 * By default on NUMA we use alien caches to stage the freeing of
886 * objects allocated from other nodes. This causes massive memory
887 * inefficiencies when using fake NUMA setup to split memory into a
888 * large number of small nodes, so it can be disabled on the command
892 static int use_alien_caches __read_mostly = 1;
893 static int numa_platform __read_mostly = 1;
894 static int __init noaliencache_setup(char *s)
896 use_alien_caches = 0;
899 __setup("noaliencache", noaliencache_setup);
903 * Special reaping functions for NUMA systems called from cache_reap().
904 * These take care of doing round robin flushing of alien caches (containing
905 * objects freed on different nodes from which they were allocated) and the
906 * flushing of remote pcps by calling drain_node_pages.
908 static DEFINE_PER_CPU(unsigned long, reap_node);
910 static void init_reap_node(int cpu)
914 node = next_node(cpu_to_node(cpu), node_online_map);
915 if (node == MAX_NUMNODES)
916 node = first_node(node_online_map);
918 per_cpu(reap_node, cpu) = node;
921 static void next_reap_node(void)
923 int node = __get_cpu_var(reap_node);
925 node = next_node(node, node_online_map);
926 if (unlikely(node >= MAX_NUMNODES))
927 node = first_node(node_online_map);
928 __get_cpu_var(reap_node) = node;
932 #define init_reap_node(cpu) do { } while (0)
933 #define next_reap_node(void) do { } while (0)
937 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
938 * via the workqueue/eventd.
939 * Add the CPU number into the expiration time to minimize the possibility of
940 * the CPUs getting into lockstep and contending for the global cache chain
943 static void __cpuinit start_cpu_timer(int cpu)
945 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
948 * When this gets called from do_initcalls via cpucache_init(),
949 * init_workqueues() has already run, so keventd will be setup
952 if (keventd_up() && reap_work->work.func == NULL) {
954 INIT_DELAYED_WORK(reap_work, cache_reap);
955 schedule_delayed_work_on(cpu, reap_work,
956 __round_jiffies_relative(HZ, cpu));
960 static struct array_cache *alloc_arraycache(int node, int entries,
961 int batchcount, gfp_t gfp)
963 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
964 struct array_cache *nc = NULL;
966 nc = kmalloc_node(memsize, gfp, node);
970 nc->batchcount = batchcount;
972 spin_lock_init(&nc->lock);
978 * Transfer objects in one arraycache to another.
979 * Locking must be handled by the caller.
981 * Return the number of entries transferred.
983 static int transfer_objects(struct array_cache *to,
984 struct array_cache *from, unsigned int max)
986 /* Figure out how many entries to transfer */
987 int nr = min(min(from->avail, max), to->limit - to->avail);
992 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1003 #define drain_alien_cache(cachep, alien) do { } while (0)
1004 #define reap_alien(cachep, l3) do { } while (0)
1006 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1008 return (struct array_cache **)BAD_ALIEN_MAGIC;
1011 static inline void free_alien_cache(struct array_cache **ac_ptr)
1015 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1020 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1026 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1027 gfp_t flags, int nodeid)
1032 #else /* CONFIG_NUMA */
1034 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1035 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1037 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1039 struct array_cache **ac_ptr;
1040 int memsize = sizeof(void *) * nr_node_ids;
1045 ac_ptr = kmalloc_node(memsize, gfp, node);
1048 if (i == node || !node_online(i)) {
1052 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1054 for (i--; i >= 0; i--)
1064 static void free_alien_cache(struct array_cache **ac_ptr)
1075 static void __drain_alien_cache(struct kmem_cache *cachep,
1076 struct array_cache *ac, int node)
1078 struct kmem_list3 *rl3 = cachep->nodelists[node];
1081 spin_lock(&rl3->list_lock);
1083 * Stuff objects into the remote nodes shared array first.
1084 * That way we could avoid the overhead of putting the objects
1085 * into the free lists and getting them back later.
1088 transfer_objects(rl3->shared, ac, ac->limit);
1090 free_block(cachep, ac->entry, ac->avail, node);
1092 spin_unlock(&rl3->list_lock);
1097 * Called from cache_reap() to regularly drain alien caches round robin.
1099 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1101 int node = __get_cpu_var(reap_node);
1104 struct array_cache *ac = l3->alien[node];
1106 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1107 __drain_alien_cache(cachep, ac, node);
1108 spin_unlock_irq(&ac->lock);
1113 static void drain_alien_cache(struct kmem_cache *cachep,
1114 struct array_cache **alien)
1117 struct array_cache *ac;
1118 unsigned long flags;
1120 for_each_online_node(i) {
1123 spin_lock_irqsave(&ac->lock, flags);
1124 __drain_alien_cache(cachep, ac, i);
1125 spin_unlock_irqrestore(&ac->lock, flags);
1130 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1132 struct slab *slabp = virt_to_slab(objp);
1133 int nodeid = slabp->nodeid;
1134 struct kmem_list3 *l3;
1135 struct array_cache *alien = NULL;
1138 node = numa_node_id();
1141 * Make sure we are not freeing a object from another node to the array
1142 * cache on this cpu.
1144 if (likely(slabp->nodeid == node))
1147 l3 = cachep->nodelists[node];
1148 STATS_INC_NODEFREES(cachep);
1149 if (l3->alien && l3->alien[nodeid]) {
1150 alien = l3->alien[nodeid];
1151 spin_lock(&alien->lock);
1152 if (unlikely(alien->avail == alien->limit)) {
1153 STATS_INC_ACOVERFLOW(cachep);
1154 __drain_alien_cache(cachep, alien, nodeid);
1156 alien->entry[alien->avail++] = objp;
1157 spin_unlock(&alien->lock);
1159 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1160 free_block(cachep, &objp, 1, nodeid);
1161 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1167 static void __cpuinit cpuup_canceled(long cpu)
1169 struct kmem_cache *cachep;
1170 struct kmem_list3 *l3 = NULL;
1171 int node = cpu_to_node(cpu);
1172 const struct cpumask *mask = cpumask_of_node(node);
1174 list_for_each_entry(cachep, &cache_chain, next) {
1175 struct array_cache *nc;
1176 struct array_cache *shared;
1177 struct array_cache **alien;
1179 /* cpu is dead; no one can alloc from it. */
1180 nc = cachep->array[cpu];
1181 cachep->array[cpu] = NULL;
1182 l3 = cachep->nodelists[node];
1185 goto free_array_cache;
1187 spin_lock_irq(&l3->list_lock);
1189 /* Free limit for this kmem_list3 */
1190 l3->free_limit -= cachep->batchcount;
1192 free_block(cachep, nc->entry, nc->avail, node);
1194 if (!cpus_empty(*mask)) {
1195 spin_unlock_irq(&l3->list_lock);
1196 goto free_array_cache;
1199 shared = l3->shared;
1201 free_block(cachep, shared->entry,
1202 shared->avail, node);
1209 spin_unlock_irq(&l3->list_lock);
1213 drain_alien_cache(cachep, alien);
1214 free_alien_cache(alien);
1220 * In the previous loop, all the objects were freed to
1221 * the respective cache's slabs, now we can go ahead and
1222 * shrink each nodelist to its limit.
1224 list_for_each_entry(cachep, &cache_chain, next) {
1225 l3 = cachep->nodelists[node];
1228 drain_freelist(cachep, l3, l3->free_objects);
1232 static int __cpuinit cpuup_prepare(long cpu)
1234 struct kmem_cache *cachep;
1235 struct kmem_list3 *l3 = NULL;
1236 int node = cpu_to_node(cpu);
1237 const int memsize = sizeof(struct kmem_list3);
1240 * We need to do this right in the beginning since
1241 * alloc_arraycache's are going to use this list.
1242 * kmalloc_node allows us to add the slab to the right
1243 * kmem_list3 and not this cpu's kmem_list3
1246 list_for_each_entry(cachep, &cache_chain, next) {
1248 * Set up the size64 kmemlist for cpu before we can
1249 * begin anything. Make sure some other cpu on this
1250 * node has not already allocated this
1252 if (!cachep->nodelists[node]) {
1253 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1256 kmem_list3_init(l3);
1257 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1258 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1261 * The l3s don't come and go as CPUs come and
1262 * go. cache_chain_mutex is sufficient
1265 cachep->nodelists[node] = l3;
1268 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1269 cachep->nodelists[node]->free_limit =
1270 (1 + nr_cpus_node(node)) *
1271 cachep->batchcount + cachep->num;
1272 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1276 * Now we can go ahead with allocating the shared arrays and
1279 list_for_each_entry(cachep, &cache_chain, next) {
1280 struct array_cache *nc;
1281 struct array_cache *shared = NULL;
1282 struct array_cache **alien = NULL;
1284 nc = alloc_arraycache(node, cachep->limit,
1285 cachep->batchcount, GFP_KERNEL);
1288 if (cachep->shared) {
1289 shared = alloc_arraycache(node,
1290 cachep->shared * cachep->batchcount,
1291 0xbaadf00d, GFP_KERNEL);
1297 if (use_alien_caches) {
1298 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1305 cachep->array[cpu] = nc;
1306 l3 = cachep->nodelists[node];
1309 spin_lock_irq(&l3->list_lock);
1312 * We are serialised from CPU_DEAD or
1313 * CPU_UP_CANCELLED by the cpucontrol lock
1315 l3->shared = shared;
1324 spin_unlock_irq(&l3->list_lock);
1326 free_alien_cache(alien);
1330 cpuup_canceled(cpu);
1334 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1335 unsigned long action, void *hcpu)
1337 long cpu = (long)hcpu;
1341 case CPU_UP_PREPARE:
1342 case CPU_UP_PREPARE_FROZEN:
1343 mutex_lock(&cache_chain_mutex);
1344 err = cpuup_prepare(cpu);
1345 mutex_unlock(&cache_chain_mutex);
1348 case CPU_ONLINE_FROZEN:
1349 start_cpu_timer(cpu);
1351 #ifdef CONFIG_HOTPLUG_CPU
1352 case CPU_DOWN_PREPARE:
1353 case CPU_DOWN_PREPARE_FROZEN:
1355 * Shutdown cache reaper. Note that the cache_chain_mutex is
1356 * held so that if cache_reap() is invoked it cannot do
1357 * anything expensive but will only modify reap_work
1358 * and reschedule the timer.
1360 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1361 /* Now the cache_reaper is guaranteed to be not running. */
1362 per_cpu(reap_work, cpu).work.func = NULL;
1364 case CPU_DOWN_FAILED:
1365 case CPU_DOWN_FAILED_FROZEN:
1366 start_cpu_timer(cpu);
1369 case CPU_DEAD_FROZEN:
1371 * Even if all the cpus of a node are down, we don't free the
1372 * kmem_list3 of any cache. This to avoid a race between
1373 * cpu_down, and a kmalloc allocation from another cpu for
1374 * memory from the node of the cpu going down. The list3
1375 * structure is usually allocated from kmem_cache_create() and
1376 * gets destroyed at kmem_cache_destroy().
1380 case CPU_UP_CANCELED:
1381 case CPU_UP_CANCELED_FROZEN:
1382 mutex_lock(&cache_chain_mutex);
1383 cpuup_canceled(cpu);
1384 mutex_unlock(&cache_chain_mutex);
1387 return err ? NOTIFY_BAD : NOTIFY_OK;
1390 static struct notifier_block __cpuinitdata cpucache_notifier = {
1391 &cpuup_callback, NULL, 0
1395 * swap the static kmem_list3 with kmalloced memory
1397 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1400 struct kmem_list3 *ptr;
1402 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1405 memcpy(ptr, list, sizeof(struct kmem_list3));
1407 * Do not assume that spinlocks can be initialized via memcpy:
1409 spin_lock_init(&ptr->list_lock);
1411 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1412 cachep->nodelists[nodeid] = ptr;
1416 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1417 * size of kmem_list3.
1419 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1423 for_each_online_node(node) {
1424 cachep->nodelists[node] = &initkmem_list3[index + node];
1425 cachep->nodelists[node]->next_reap = jiffies +
1427 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1432 * Initialisation. Called after the page allocator have been initialised and
1433 * before smp_init().
1435 void __init kmem_cache_init(void)
1438 struct cache_sizes *sizes;
1439 struct cache_names *names;
1444 if (num_possible_nodes() == 1) {
1445 use_alien_caches = 0;
1449 for (i = 0; i < NUM_INIT_LISTS; i++) {
1450 kmem_list3_init(&initkmem_list3[i]);
1451 if (i < MAX_NUMNODES)
1452 cache_cache.nodelists[i] = NULL;
1454 set_up_list3s(&cache_cache, CACHE_CACHE);
1457 * Fragmentation resistance on low memory - only use bigger
1458 * page orders on machines with more than 32MB of memory.
1460 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1461 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1463 /* Bootstrap is tricky, because several objects are allocated
1464 * from caches that do not exist yet:
1465 * 1) initialize the cache_cache cache: it contains the struct
1466 * kmem_cache structures of all caches, except cache_cache itself:
1467 * cache_cache is statically allocated.
1468 * Initially an __init data area is used for the head array and the
1469 * kmem_list3 structures, it's replaced with a kmalloc allocated
1470 * array at the end of the bootstrap.
1471 * 2) Create the first kmalloc cache.
1472 * The struct kmem_cache for the new cache is allocated normally.
1473 * An __init data area is used for the head array.
1474 * 3) Create the remaining kmalloc caches, with minimally sized
1476 * 4) Replace the __init data head arrays for cache_cache and the first
1477 * kmalloc cache with kmalloc allocated arrays.
1478 * 5) Replace the __init data for kmem_list3 for cache_cache and
1479 * the other cache's with kmalloc allocated memory.
1480 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1483 node = numa_node_id();
1485 /* 1) create the cache_cache */
1486 INIT_LIST_HEAD(&cache_chain);
1487 list_add(&cache_cache.next, &cache_chain);
1488 cache_cache.colour_off = cache_line_size();
1489 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1490 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1493 * struct kmem_cache size depends on nr_node_ids, which
1494 * can be less than MAX_NUMNODES.
1496 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1497 nr_node_ids * sizeof(struct kmem_list3 *);
1499 cache_cache.obj_size = cache_cache.buffer_size;
1501 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1503 cache_cache.reciprocal_buffer_size =
1504 reciprocal_value(cache_cache.buffer_size);
1506 for (order = 0; order < MAX_ORDER; order++) {
1507 cache_estimate(order, cache_cache.buffer_size,
1508 cache_line_size(), 0, &left_over, &cache_cache.num);
1509 if (cache_cache.num)
1512 BUG_ON(!cache_cache.num);
1513 cache_cache.gfporder = order;
1514 cache_cache.colour = left_over / cache_cache.colour_off;
1515 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1516 sizeof(struct slab), cache_line_size());
1518 /* 2+3) create the kmalloc caches */
1519 sizes = malloc_sizes;
1520 names = cache_names;
1523 * Initialize the caches that provide memory for the array cache and the
1524 * kmem_list3 structures first. Without this, further allocations will
1528 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1529 sizes[INDEX_AC].cs_size,
1530 ARCH_KMALLOC_MINALIGN,
1531 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1534 if (INDEX_AC != INDEX_L3) {
1535 sizes[INDEX_L3].cs_cachep =
1536 kmem_cache_create(names[INDEX_L3].name,
1537 sizes[INDEX_L3].cs_size,
1538 ARCH_KMALLOC_MINALIGN,
1539 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1543 slab_early_init = 0;
1545 while (sizes->cs_size != ULONG_MAX) {
1547 * For performance, all the general caches are L1 aligned.
1548 * This should be particularly beneficial on SMP boxes, as it
1549 * eliminates "false sharing".
1550 * Note for systems short on memory removing the alignment will
1551 * allow tighter packing of the smaller caches.
1553 if (!sizes->cs_cachep) {
1554 sizes->cs_cachep = kmem_cache_create(names->name,
1556 ARCH_KMALLOC_MINALIGN,
1557 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1560 #ifdef CONFIG_ZONE_DMA
1561 sizes->cs_dmacachep = kmem_cache_create(
1564 ARCH_KMALLOC_MINALIGN,
1565 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1572 /* 4) Replace the bootstrap head arrays */
1574 struct array_cache *ptr;
1576 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1578 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1579 memcpy(ptr, cpu_cache_get(&cache_cache),
1580 sizeof(struct arraycache_init));
1582 * Do not assume that spinlocks can be initialized via memcpy:
1584 spin_lock_init(&ptr->lock);
1586 cache_cache.array[smp_processor_id()] = ptr;
1588 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1590 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1591 != &initarray_generic.cache);
1592 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1593 sizeof(struct arraycache_init));
1595 * Do not assume that spinlocks can be initialized via memcpy:
1597 spin_lock_init(&ptr->lock);
1599 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1602 /* 5) Replace the bootstrap kmem_list3's */
1606 for_each_online_node(nid) {
1607 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1609 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1610 &initkmem_list3[SIZE_AC + nid], nid);
1612 if (INDEX_AC != INDEX_L3) {
1613 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1614 &initkmem_list3[SIZE_L3 + nid], nid);
1619 /* 6) resize the head arrays to their final sizes */
1621 struct kmem_cache *cachep;
1622 mutex_lock(&cache_chain_mutex);
1623 list_for_each_entry(cachep, &cache_chain, next)
1624 if (enable_cpucache(cachep, GFP_NOWAIT))
1626 mutex_unlock(&cache_chain_mutex);
1629 /* Annotate slab for lockdep -- annotate the malloc caches */
1634 g_cpucache_up = FULL;
1637 * Register a cpu startup notifier callback that initializes
1638 * cpu_cache_get for all new cpus
1640 register_cpu_notifier(&cpucache_notifier);
1643 * The reap timers are started later, with a module init call: That part
1644 * of the kernel is not yet operational.
1648 static int __init cpucache_init(void)
1653 * Register the timers that return unneeded pages to the page allocator
1655 for_each_online_cpu(cpu)
1656 start_cpu_timer(cpu);
1659 __initcall(cpucache_init);
1662 * Interface to system's page allocator. No need to hold the cache-lock.
1664 * If we requested dmaable memory, we will get it. Even if we
1665 * did not request dmaable memory, we might get it, but that
1666 * would be relatively rare and ignorable.
1668 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1676 * Nommu uses slab's for process anonymous memory allocations, and thus
1677 * requires __GFP_COMP to properly refcount higher order allocations
1679 flags |= __GFP_COMP;
1682 flags |= cachep->gfpflags;
1683 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1684 flags |= __GFP_RECLAIMABLE;
1686 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1690 nr_pages = (1 << cachep->gfporder);
1691 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1692 add_zone_page_state(page_zone(page),
1693 NR_SLAB_RECLAIMABLE, nr_pages);
1695 add_zone_page_state(page_zone(page),
1696 NR_SLAB_UNRECLAIMABLE, nr_pages);
1697 for (i = 0; i < nr_pages; i++)
1698 __SetPageSlab(page + i);
1699 return page_address(page);
1703 * Interface to system's page release.
1705 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1707 unsigned long i = (1 << cachep->gfporder);
1708 struct page *page = virt_to_page(addr);
1709 const unsigned long nr_freed = i;
1711 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1712 sub_zone_page_state(page_zone(page),
1713 NR_SLAB_RECLAIMABLE, nr_freed);
1715 sub_zone_page_state(page_zone(page),
1716 NR_SLAB_UNRECLAIMABLE, nr_freed);
1718 BUG_ON(!PageSlab(page));
1719 __ClearPageSlab(page);
1722 if (current->reclaim_state)
1723 current->reclaim_state->reclaimed_slab += nr_freed;
1724 free_pages((unsigned long)addr, cachep->gfporder);
1727 static void kmem_rcu_free(struct rcu_head *head)
1729 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1730 struct kmem_cache *cachep = slab_rcu->cachep;
1732 kmem_freepages(cachep, slab_rcu->addr);
1733 if (OFF_SLAB(cachep))
1734 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1739 #ifdef CONFIG_DEBUG_PAGEALLOC
1740 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1741 unsigned long caller)
1743 int size = obj_size(cachep);
1745 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1747 if (size < 5 * sizeof(unsigned long))
1750 *addr++ = 0x12345678;
1752 *addr++ = smp_processor_id();
1753 size -= 3 * sizeof(unsigned long);
1755 unsigned long *sptr = &caller;
1756 unsigned long svalue;
1758 while (!kstack_end(sptr)) {
1760 if (kernel_text_address(svalue)) {
1762 size -= sizeof(unsigned long);
1763 if (size <= sizeof(unsigned long))
1769 *addr++ = 0x87654321;
1773 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1775 int size = obj_size(cachep);
1776 addr = &((char *)addr)[obj_offset(cachep)];
1778 memset(addr, val, size);
1779 *(unsigned char *)(addr + size - 1) = POISON_END;
1782 static void dump_line(char *data, int offset, int limit)
1785 unsigned char error = 0;
1788 printk(KERN_ERR "%03x:", offset);
1789 for (i = 0; i < limit; i++) {
1790 if (data[offset + i] != POISON_FREE) {
1791 error = data[offset + i];
1794 printk(" %02x", (unsigned char)data[offset + i]);
1798 if (bad_count == 1) {
1799 error ^= POISON_FREE;
1800 if (!(error & (error - 1))) {
1801 printk(KERN_ERR "Single bit error detected. Probably "
1804 printk(KERN_ERR "Run memtest86+ or a similar memory "
1807 printk(KERN_ERR "Run a memory test tool.\n");
1816 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1821 if (cachep->flags & SLAB_RED_ZONE) {
1822 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1823 *dbg_redzone1(cachep, objp),
1824 *dbg_redzone2(cachep, objp));
1827 if (cachep->flags & SLAB_STORE_USER) {
1828 printk(KERN_ERR "Last user: [<%p>]",
1829 *dbg_userword(cachep, objp));
1830 print_symbol("(%s)",
1831 (unsigned long)*dbg_userword(cachep, objp));
1834 realobj = (char *)objp + obj_offset(cachep);
1835 size = obj_size(cachep);
1836 for (i = 0; i < size && lines; i += 16, lines--) {
1839 if (i + limit > size)
1841 dump_line(realobj, i, limit);
1845 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1851 realobj = (char *)objp + obj_offset(cachep);
1852 size = obj_size(cachep);
1854 for (i = 0; i < size; i++) {
1855 char exp = POISON_FREE;
1858 if (realobj[i] != exp) {
1864 "Slab corruption: %s start=%p, len=%d\n",
1865 cachep->name, realobj, size);
1866 print_objinfo(cachep, objp, 0);
1868 /* Hexdump the affected line */
1871 if (i + limit > size)
1873 dump_line(realobj, i, limit);
1876 /* Limit to 5 lines */
1882 /* Print some data about the neighboring objects, if they
1885 struct slab *slabp = virt_to_slab(objp);
1888 objnr = obj_to_index(cachep, slabp, objp);
1890 objp = index_to_obj(cachep, slabp, objnr - 1);
1891 realobj = (char *)objp + obj_offset(cachep);
1892 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1894 print_objinfo(cachep, objp, 2);
1896 if (objnr + 1 < cachep->num) {
1897 objp = index_to_obj(cachep, slabp, objnr + 1);
1898 realobj = (char *)objp + obj_offset(cachep);
1899 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1901 print_objinfo(cachep, objp, 2);
1908 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1911 for (i = 0; i < cachep->num; i++) {
1912 void *objp = index_to_obj(cachep, slabp, i);
1914 if (cachep->flags & SLAB_POISON) {
1915 #ifdef CONFIG_DEBUG_PAGEALLOC
1916 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1918 kernel_map_pages(virt_to_page(objp),
1919 cachep->buffer_size / PAGE_SIZE, 1);
1921 check_poison_obj(cachep, objp);
1923 check_poison_obj(cachep, objp);
1926 if (cachep->flags & SLAB_RED_ZONE) {
1927 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1928 slab_error(cachep, "start of a freed object "
1930 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1931 slab_error(cachep, "end of a freed object "
1937 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1943 * slab_destroy - destroy and release all objects in a slab
1944 * @cachep: cache pointer being destroyed
1945 * @slabp: slab pointer being destroyed
1947 * Destroy all the objs in a slab, and release the mem back to the system.
1948 * Before calling the slab must have been unlinked from the cache. The
1949 * cache-lock is not held/needed.
1951 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1953 void *addr = slabp->s_mem - slabp->colouroff;
1955 slab_destroy_debugcheck(cachep, slabp);
1956 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1957 struct slab_rcu *slab_rcu;
1959 slab_rcu = (struct slab_rcu *)slabp;
1960 slab_rcu->cachep = cachep;
1961 slab_rcu->addr = addr;
1962 call_rcu(&slab_rcu->head, kmem_rcu_free);
1964 kmem_freepages(cachep, addr);
1965 if (OFF_SLAB(cachep))
1966 kmem_cache_free(cachep->slabp_cache, slabp);
1970 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1973 struct kmem_list3 *l3;
1975 for_each_online_cpu(i)
1976 kfree(cachep->array[i]);
1978 /* NUMA: free the list3 structures */
1979 for_each_online_node(i) {
1980 l3 = cachep->nodelists[i];
1983 free_alien_cache(l3->alien);
1987 kmem_cache_free(&cache_cache, cachep);
1992 * calculate_slab_order - calculate size (page order) of slabs
1993 * @cachep: pointer to the cache that is being created
1994 * @size: size of objects to be created in this cache.
1995 * @align: required alignment for the objects.
1996 * @flags: slab allocation flags
1998 * Also calculates the number of objects per slab.
2000 * This could be made much more intelligent. For now, try to avoid using
2001 * high order pages for slabs. When the gfp() functions are more friendly
2002 * towards high-order requests, this should be changed.
2004 static size_t calculate_slab_order(struct kmem_cache *cachep,
2005 size_t size, size_t align, unsigned long flags)
2007 unsigned long offslab_limit;
2008 size_t left_over = 0;
2011 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2015 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2019 if (flags & CFLGS_OFF_SLAB) {
2021 * Max number of objs-per-slab for caches which
2022 * use off-slab slabs. Needed to avoid a possible
2023 * looping condition in cache_grow().
2025 offslab_limit = size - sizeof(struct slab);
2026 offslab_limit /= sizeof(kmem_bufctl_t);
2028 if (num > offslab_limit)
2032 /* Found something acceptable - save it away */
2034 cachep->gfporder = gfporder;
2035 left_over = remainder;
2038 * A VFS-reclaimable slab tends to have most allocations
2039 * as GFP_NOFS and we really don't want to have to be allocating
2040 * higher-order pages when we are unable to shrink dcache.
2042 if (flags & SLAB_RECLAIM_ACCOUNT)
2046 * Large number of objects is good, but very large slabs are
2047 * currently bad for the gfp()s.
2049 if (gfporder >= slab_break_gfp_order)
2053 * Acceptable internal fragmentation?
2055 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2061 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2063 if (g_cpucache_up == FULL)
2064 return enable_cpucache(cachep, gfp);
2066 if (g_cpucache_up == NONE) {
2068 * Note: the first kmem_cache_create must create the cache
2069 * that's used by kmalloc(24), otherwise the creation of
2070 * further caches will BUG().
2072 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2075 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2076 * the first cache, then we need to set up all its list3s,
2077 * otherwise the creation of further caches will BUG().
2079 set_up_list3s(cachep, SIZE_AC);
2080 if (INDEX_AC == INDEX_L3)
2081 g_cpucache_up = PARTIAL_L3;
2083 g_cpucache_up = PARTIAL_AC;
2085 cachep->array[smp_processor_id()] =
2086 kmalloc(sizeof(struct arraycache_init), gfp);
2088 if (g_cpucache_up == PARTIAL_AC) {
2089 set_up_list3s(cachep, SIZE_L3);
2090 g_cpucache_up = PARTIAL_L3;
2093 for_each_online_node(node) {
2094 cachep->nodelists[node] =
2095 kmalloc_node(sizeof(struct kmem_list3),
2097 BUG_ON(!cachep->nodelists[node]);
2098 kmem_list3_init(cachep->nodelists[node]);
2102 cachep->nodelists[numa_node_id()]->next_reap =
2103 jiffies + REAPTIMEOUT_LIST3 +
2104 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2106 cpu_cache_get(cachep)->avail = 0;
2107 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2108 cpu_cache_get(cachep)->batchcount = 1;
2109 cpu_cache_get(cachep)->touched = 0;
2110 cachep->batchcount = 1;
2111 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2116 * kmem_cache_create - Create a cache.
2117 * @name: A string which is used in /proc/slabinfo to identify this cache.
2118 * @size: The size of objects to be created in this cache.
2119 * @align: The required alignment for the objects.
2120 * @flags: SLAB flags
2121 * @ctor: A constructor for the objects.
2123 * Returns a ptr to the cache on success, NULL on failure.
2124 * Cannot be called within a int, but can be interrupted.
2125 * The @ctor is run when new pages are allocated by the cache.
2127 * @name must be valid until the cache is destroyed. This implies that
2128 * the module calling this has to destroy the cache before getting unloaded.
2129 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2130 * therefore applications must manage it themselves.
2134 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2135 * to catch references to uninitialised memory.
2137 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2138 * for buffer overruns.
2140 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2141 * cacheline. This can be beneficial if you're counting cycles as closely
2145 kmem_cache_create (const char *name, size_t size, size_t align,
2146 unsigned long flags, void (*ctor)(void *))
2148 size_t left_over, slab_size, ralign;
2149 struct kmem_cache *cachep = NULL, *pc;
2153 * Sanity checks... these are all serious usage bugs.
2155 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2156 size > KMALLOC_MAX_SIZE) {
2157 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2163 * We use cache_chain_mutex to ensure a consistent view of
2164 * cpu_online_mask as well. Please see cpuup_callback
2166 if (slab_is_available()) {
2168 mutex_lock(&cache_chain_mutex);
2171 list_for_each_entry(pc, &cache_chain, next) {
2176 * This happens when the module gets unloaded and doesn't
2177 * destroy its slab cache and no-one else reuses the vmalloc
2178 * area of the module. Print a warning.
2180 res = probe_kernel_address(pc->name, tmp);
2183 "SLAB: cache with size %d has lost its name\n",
2188 if (!strcmp(pc->name, name)) {
2190 "kmem_cache_create: duplicate cache %s\n", name);
2197 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2200 * Enable redzoning and last user accounting, except for caches with
2201 * large objects, if the increased size would increase the object size
2202 * above the next power of two: caches with object sizes just above a
2203 * power of two have a significant amount of internal fragmentation.
2205 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2206 2 * sizeof(unsigned long long)))
2207 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2208 if (!(flags & SLAB_DESTROY_BY_RCU))
2209 flags |= SLAB_POISON;
2211 if (flags & SLAB_DESTROY_BY_RCU)
2212 BUG_ON(flags & SLAB_POISON);
2215 * Always checks flags, a caller might be expecting debug support which
2218 BUG_ON(flags & ~CREATE_MASK);
2221 * Check that size is in terms of words. This is needed to avoid
2222 * unaligned accesses for some archs when redzoning is used, and makes
2223 * sure any on-slab bufctl's are also correctly aligned.
2225 if (size & (BYTES_PER_WORD - 1)) {
2226 size += (BYTES_PER_WORD - 1);
2227 size &= ~(BYTES_PER_WORD - 1);
2230 /* calculate the final buffer alignment: */
2232 /* 1) arch recommendation: can be overridden for debug */
2233 if (flags & SLAB_HWCACHE_ALIGN) {
2235 * Default alignment: as specified by the arch code. Except if
2236 * an object is really small, then squeeze multiple objects into
2239 ralign = cache_line_size();
2240 while (size <= ralign / 2)
2243 ralign = BYTES_PER_WORD;
2247 * Redzoning and user store require word alignment or possibly larger.
2248 * Note this will be overridden by architecture or caller mandated
2249 * alignment if either is greater than BYTES_PER_WORD.
2251 if (flags & SLAB_STORE_USER)
2252 ralign = BYTES_PER_WORD;
2254 if (flags & SLAB_RED_ZONE) {
2255 ralign = REDZONE_ALIGN;
2256 /* If redzoning, ensure that the second redzone is suitably
2257 * aligned, by adjusting the object size accordingly. */
2258 size += REDZONE_ALIGN - 1;
2259 size &= ~(REDZONE_ALIGN - 1);
2262 /* 2) arch mandated alignment */
2263 if (ralign < ARCH_SLAB_MINALIGN) {
2264 ralign = ARCH_SLAB_MINALIGN;
2266 /* 3) caller mandated alignment */
2267 if (ralign < align) {
2270 /* disable debug if necessary */
2271 if (ralign > __alignof__(unsigned long long))
2272 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2278 if (slab_is_available())
2283 /* Get cache's description obj. */
2284 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2289 cachep->obj_size = size;
2292 * Both debugging options require word-alignment which is calculated
2295 if (flags & SLAB_RED_ZONE) {
2296 /* add space for red zone words */
2297 cachep->obj_offset += sizeof(unsigned long long);
2298 size += 2 * sizeof(unsigned long long);
2300 if (flags & SLAB_STORE_USER) {
2301 /* user store requires one word storage behind the end of
2302 * the real object. But if the second red zone needs to be
2303 * aligned to 64 bits, we must allow that much space.
2305 if (flags & SLAB_RED_ZONE)
2306 size += REDZONE_ALIGN;
2308 size += BYTES_PER_WORD;
2310 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2311 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2312 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2313 cachep->obj_offset += PAGE_SIZE - size;
2320 * Determine if the slab management is 'on' or 'off' slab.
2321 * (bootstrapping cannot cope with offslab caches so don't do
2324 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2326 * Size is large, assume best to place the slab management obj
2327 * off-slab (should allow better packing of objs).
2329 flags |= CFLGS_OFF_SLAB;
2331 size = ALIGN(size, align);
2333 left_over = calculate_slab_order(cachep, size, align, flags);
2337 "kmem_cache_create: couldn't create cache %s.\n", name);
2338 kmem_cache_free(&cache_cache, cachep);
2342 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2343 + sizeof(struct slab), align);
2346 * If the slab has been placed off-slab, and we have enough space then
2347 * move it on-slab. This is at the expense of any extra colouring.
2349 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2350 flags &= ~CFLGS_OFF_SLAB;
2351 left_over -= slab_size;
2354 if (flags & CFLGS_OFF_SLAB) {
2355 /* really off slab. No need for manual alignment */
2357 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2360 cachep->colour_off = cache_line_size();
2361 /* Offset must be a multiple of the alignment. */
2362 if (cachep->colour_off < align)
2363 cachep->colour_off = align;
2364 cachep->colour = left_over / cachep->colour_off;
2365 cachep->slab_size = slab_size;
2366 cachep->flags = flags;
2367 cachep->gfpflags = 0;
2368 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2369 cachep->gfpflags |= GFP_DMA;
2370 cachep->buffer_size = size;
2371 cachep->reciprocal_buffer_size = reciprocal_value(size);
2373 if (flags & CFLGS_OFF_SLAB) {
2374 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2376 * This is a possibility for one of the malloc_sizes caches.
2377 * But since we go off slab only for object size greater than
2378 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2379 * this should not happen at all.
2380 * But leave a BUG_ON for some lucky dude.
2382 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2384 cachep->ctor = ctor;
2385 cachep->name = name;
2387 if (setup_cpu_cache(cachep, gfp)) {
2388 __kmem_cache_destroy(cachep);
2393 /* cache setup completed, link it into the list */
2394 list_add(&cachep->next, &cache_chain);
2396 if (!cachep && (flags & SLAB_PANIC))
2397 panic("kmem_cache_create(): failed to create slab `%s'\n",
2399 if (slab_is_available()) {
2400 mutex_unlock(&cache_chain_mutex);
2405 EXPORT_SYMBOL(kmem_cache_create);
2408 static void check_irq_off(void)
2410 BUG_ON(!irqs_disabled());
2413 static void check_irq_on(void)
2415 BUG_ON(irqs_disabled());
2418 static void check_spinlock_acquired(struct kmem_cache *cachep)
2422 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2426 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2430 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2435 #define check_irq_off() do { } while(0)
2436 #define check_irq_on() do { } while(0)
2437 #define check_spinlock_acquired(x) do { } while(0)
2438 #define check_spinlock_acquired_node(x, y) do { } while(0)
2441 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2442 struct array_cache *ac,
2443 int force, int node);
2445 static void do_drain(void *arg)
2447 struct kmem_cache *cachep = arg;
2448 struct array_cache *ac;
2449 int node = numa_node_id();
2452 ac = cpu_cache_get(cachep);
2453 spin_lock(&cachep->nodelists[node]->list_lock);
2454 free_block(cachep, ac->entry, ac->avail, node);
2455 spin_unlock(&cachep->nodelists[node]->list_lock);
2459 static void drain_cpu_caches(struct kmem_cache *cachep)
2461 struct kmem_list3 *l3;
2464 on_each_cpu(do_drain, cachep, 1);
2466 for_each_online_node(node) {
2467 l3 = cachep->nodelists[node];
2468 if (l3 && l3->alien)
2469 drain_alien_cache(cachep, l3->alien);
2472 for_each_online_node(node) {
2473 l3 = cachep->nodelists[node];
2475 drain_array(cachep, l3, l3->shared, 1, node);
2480 * Remove slabs from the list of free slabs.
2481 * Specify the number of slabs to drain in tofree.
2483 * Returns the actual number of slabs released.
2485 static int drain_freelist(struct kmem_cache *cache,
2486 struct kmem_list3 *l3, int tofree)
2488 struct list_head *p;
2493 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2495 spin_lock_irq(&l3->list_lock);
2496 p = l3->slabs_free.prev;
2497 if (p == &l3->slabs_free) {
2498 spin_unlock_irq(&l3->list_lock);
2502 slabp = list_entry(p, struct slab, list);
2504 BUG_ON(slabp->inuse);
2506 list_del(&slabp->list);
2508 * Safe to drop the lock. The slab is no longer linked
2511 l3->free_objects -= cache->num;
2512 spin_unlock_irq(&l3->list_lock);
2513 slab_destroy(cache, slabp);
2520 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2521 static int __cache_shrink(struct kmem_cache *cachep)
2524 struct kmem_list3 *l3;
2526 drain_cpu_caches(cachep);
2529 for_each_online_node(i) {
2530 l3 = cachep->nodelists[i];
2534 drain_freelist(cachep, l3, l3->free_objects);
2536 ret += !list_empty(&l3->slabs_full) ||
2537 !list_empty(&l3->slabs_partial);
2539 return (ret ? 1 : 0);
2543 * kmem_cache_shrink - Shrink a cache.
2544 * @cachep: The cache to shrink.
2546 * Releases as many slabs as possible for a cache.
2547 * To help debugging, a zero exit status indicates all slabs were released.
2549 int kmem_cache_shrink(struct kmem_cache *cachep)
2552 BUG_ON(!cachep || in_interrupt());
2555 mutex_lock(&cache_chain_mutex);
2556 ret = __cache_shrink(cachep);
2557 mutex_unlock(&cache_chain_mutex);
2561 EXPORT_SYMBOL(kmem_cache_shrink);
2564 * kmem_cache_destroy - delete a cache
2565 * @cachep: the cache to destroy
2567 * Remove a &struct kmem_cache object from the slab cache.
2569 * It is expected this function will be called by a module when it is
2570 * unloaded. This will remove the cache completely, and avoid a duplicate
2571 * cache being allocated each time a module is loaded and unloaded, if the
2572 * module doesn't have persistent in-kernel storage across loads and unloads.
2574 * The cache must be empty before calling this function.
2576 * The caller must guarantee that noone will allocate memory from the cache
2577 * during the kmem_cache_destroy().
2579 void kmem_cache_destroy(struct kmem_cache *cachep)
2581 BUG_ON(!cachep || in_interrupt());
2583 /* Find the cache in the chain of caches. */
2585 mutex_lock(&cache_chain_mutex);
2587 * the chain is never empty, cache_cache is never destroyed
2589 list_del(&cachep->next);
2590 if (__cache_shrink(cachep)) {
2591 slab_error(cachep, "Can't free all objects");
2592 list_add(&cachep->next, &cache_chain);
2593 mutex_unlock(&cache_chain_mutex);
2598 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2601 __kmem_cache_destroy(cachep);
2602 mutex_unlock(&cache_chain_mutex);
2605 EXPORT_SYMBOL(kmem_cache_destroy);
2608 * Get the memory for a slab management obj.
2609 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2610 * always come from malloc_sizes caches. The slab descriptor cannot
2611 * come from the same cache which is getting created because,
2612 * when we are searching for an appropriate cache for these
2613 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2614 * If we are creating a malloc_sizes cache here it would not be visible to
2615 * kmem_find_general_cachep till the initialization is complete.
2616 * Hence we cannot have slabp_cache same as the original cache.
2618 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2619 int colour_off, gfp_t local_flags,
2624 if (OFF_SLAB(cachep)) {
2625 /* Slab management obj is off-slab. */
2626 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2627 local_flags, nodeid);
2631 slabp = objp + colour_off;
2632 colour_off += cachep->slab_size;
2635 slabp->colouroff = colour_off;
2636 slabp->s_mem = objp + colour_off;
2637 slabp->nodeid = nodeid;
2642 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2644 return (kmem_bufctl_t *) (slabp + 1);
2647 static void cache_init_objs(struct kmem_cache *cachep,
2652 for (i = 0; i < cachep->num; i++) {
2653 void *objp = index_to_obj(cachep, slabp, i);
2655 /* need to poison the objs? */
2656 if (cachep->flags & SLAB_POISON)
2657 poison_obj(cachep, objp, POISON_FREE);
2658 if (cachep->flags & SLAB_STORE_USER)
2659 *dbg_userword(cachep, objp) = NULL;
2661 if (cachep->flags & SLAB_RED_ZONE) {
2662 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2663 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2666 * Constructors are not allowed to allocate memory from the same
2667 * cache which they are a constructor for. Otherwise, deadlock.
2668 * They must also be threaded.
2670 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2671 cachep->ctor(objp + obj_offset(cachep));
2673 if (cachep->flags & SLAB_RED_ZONE) {
2674 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2675 slab_error(cachep, "constructor overwrote the"
2676 " end of an object");
2677 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2678 slab_error(cachep, "constructor overwrote the"
2679 " start of an object");
2681 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2682 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2683 kernel_map_pages(virt_to_page(objp),
2684 cachep->buffer_size / PAGE_SIZE, 0);
2689 slab_bufctl(slabp)[i] = i + 1;
2691 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2694 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2696 if (CONFIG_ZONE_DMA_FLAG) {
2697 if (flags & GFP_DMA)
2698 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2700 BUG_ON(cachep->gfpflags & GFP_DMA);
2704 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2707 void *objp = index_to_obj(cachep, slabp, slabp->free);
2711 next = slab_bufctl(slabp)[slabp->free];
2713 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2714 WARN_ON(slabp->nodeid != nodeid);
2721 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2722 void *objp, int nodeid)
2724 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2727 /* Verify that the slab belongs to the intended node */
2728 WARN_ON(slabp->nodeid != nodeid);
2730 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2731 printk(KERN_ERR "slab: double free detected in cache "
2732 "'%s', objp %p\n", cachep->name, objp);
2736 slab_bufctl(slabp)[objnr] = slabp->free;
2737 slabp->free = objnr;
2742 * Map pages beginning at addr to the given cache and slab. This is required
2743 * for the slab allocator to be able to lookup the cache and slab of a
2744 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2746 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2752 page = virt_to_page(addr);
2755 if (likely(!PageCompound(page)))
2756 nr_pages <<= cache->gfporder;
2759 page_set_cache(page, cache);
2760 page_set_slab(page, slab);
2762 } while (--nr_pages);
2766 * Grow (by 1) the number of slabs within a cache. This is called by
2767 * kmem_cache_alloc() when there are no active objs left in a cache.
2769 static int cache_grow(struct kmem_cache *cachep,
2770 gfp_t flags, int nodeid, void *objp)
2775 struct kmem_list3 *l3;
2778 * Be lazy and only check for valid flags here, keeping it out of the
2779 * critical path in kmem_cache_alloc().
2781 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2782 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2784 /* Take the l3 list lock to change the colour_next on this node */
2786 l3 = cachep->nodelists[nodeid];
2787 spin_lock(&l3->list_lock);
2789 /* Get colour for the slab, and cal the next value. */
2790 offset = l3->colour_next;
2792 if (l3->colour_next >= cachep->colour)
2793 l3->colour_next = 0;
2794 spin_unlock(&l3->list_lock);
2796 offset *= cachep->colour_off;
2798 if (local_flags & __GFP_WAIT)
2802 * The test for missing atomic flag is performed here, rather than
2803 * the more obvious place, simply to reduce the critical path length
2804 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2805 * will eventually be caught here (where it matters).
2807 kmem_flagcheck(cachep, flags);
2810 * Get mem for the objs. Attempt to allocate a physical page from
2814 objp = kmem_getpages(cachep, local_flags, nodeid);
2818 /* Get slab management. */
2819 slabp = alloc_slabmgmt(cachep, objp, offset,
2820 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2824 slab_map_pages(cachep, slabp, objp);
2826 cache_init_objs(cachep, slabp);
2828 if (local_flags & __GFP_WAIT)
2829 local_irq_disable();
2831 spin_lock(&l3->list_lock);
2833 /* Make slab active. */
2834 list_add_tail(&slabp->list, &(l3->slabs_free));
2835 STATS_INC_GROWN(cachep);
2836 l3->free_objects += cachep->num;
2837 spin_unlock(&l3->list_lock);
2840 kmem_freepages(cachep, objp);
2842 if (local_flags & __GFP_WAIT)
2843 local_irq_disable();
2850 * Perform extra freeing checks:
2851 * - detect bad pointers.
2852 * - POISON/RED_ZONE checking
2854 static void kfree_debugcheck(const void *objp)
2856 if (!virt_addr_valid(objp)) {
2857 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2858 (unsigned long)objp);
2863 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2865 unsigned long long redzone1, redzone2;
2867 redzone1 = *dbg_redzone1(cache, obj);
2868 redzone2 = *dbg_redzone2(cache, obj);
2873 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2876 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2877 slab_error(cache, "double free detected");
2879 slab_error(cache, "memory outside object was overwritten");
2881 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2882 obj, redzone1, redzone2);
2885 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2892 BUG_ON(virt_to_cache(objp) != cachep);
2894 objp -= obj_offset(cachep);
2895 kfree_debugcheck(objp);
2896 page = virt_to_head_page(objp);
2898 slabp = page_get_slab(page);
2900 if (cachep->flags & SLAB_RED_ZONE) {
2901 verify_redzone_free(cachep, objp);
2902 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2903 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2905 if (cachep->flags & SLAB_STORE_USER)
2906 *dbg_userword(cachep, objp) = caller;
2908 objnr = obj_to_index(cachep, slabp, objp);
2910 BUG_ON(objnr >= cachep->num);
2911 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2913 #ifdef CONFIG_DEBUG_SLAB_LEAK
2914 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2916 if (cachep->flags & SLAB_POISON) {
2917 #ifdef CONFIG_DEBUG_PAGEALLOC
2918 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2919 store_stackinfo(cachep, objp, (unsigned long)caller);
2920 kernel_map_pages(virt_to_page(objp),
2921 cachep->buffer_size / PAGE_SIZE, 0);
2923 poison_obj(cachep, objp, POISON_FREE);
2926 poison_obj(cachep, objp, POISON_FREE);
2932 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2937 /* Check slab's freelist to see if this obj is there. */
2938 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2940 if (entries > cachep->num || i >= cachep->num)
2943 if (entries != cachep->num - slabp->inuse) {
2945 printk(KERN_ERR "slab: Internal list corruption detected in "
2946 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2947 cachep->name, cachep->num, slabp, slabp->inuse);
2949 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2952 printk("\n%03x:", i);
2953 printk(" %02x", ((unsigned char *)slabp)[i]);
2960 #define kfree_debugcheck(x) do { } while(0)
2961 #define cache_free_debugcheck(x,objp,z) (objp)
2962 #define check_slabp(x,y) do { } while(0)
2965 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2968 struct kmem_list3 *l3;
2969 struct array_cache *ac;
2974 node = numa_node_id();
2975 ac = cpu_cache_get(cachep);
2976 batchcount = ac->batchcount;
2977 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2979 * If there was little recent activity on this cache, then
2980 * perform only a partial refill. Otherwise we could generate
2983 batchcount = BATCHREFILL_LIMIT;
2985 l3 = cachep->nodelists[node];
2987 BUG_ON(ac->avail > 0 || !l3);
2988 spin_lock(&l3->list_lock);
2990 /* See if we can refill from the shared array */
2991 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2994 while (batchcount > 0) {
2995 struct list_head *entry;
2997 /* Get slab alloc is to come from. */
2998 entry = l3->slabs_partial.next;
2999 if (entry == &l3->slabs_partial) {
3000 l3->free_touched = 1;
3001 entry = l3->slabs_free.next;
3002 if (entry == &l3->slabs_free)
3006 slabp = list_entry(entry, struct slab, list);
3007 check_slabp(cachep, slabp);
3008 check_spinlock_acquired(cachep);
3011 * The slab was either on partial or free list so
3012 * there must be at least one object available for
3015 BUG_ON(slabp->inuse >= cachep->num);
3017 while (slabp->inuse < cachep->num && batchcount--) {
3018 STATS_INC_ALLOCED(cachep);
3019 STATS_INC_ACTIVE(cachep);
3020 STATS_SET_HIGH(cachep);
3022 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3025 check_slabp(cachep, slabp);
3027 /* move slabp to correct slabp list: */
3028 list_del(&slabp->list);
3029 if (slabp->free == BUFCTL_END)
3030 list_add(&slabp->list, &l3->slabs_full);
3032 list_add(&slabp->list, &l3->slabs_partial);
3036 l3->free_objects -= ac->avail;
3038 spin_unlock(&l3->list_lock);
3040 if (unlikely(!ac->avail)) {
3042 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3044 /* cache_grow can reenable interrupts, then ac could change. */
3045 ac = cpu_cache_get(cachep);
3046 if (!x && ac->avail == 0) /* no objects in sight? abort */
3049 if (!ac->avail) /* objects refilled by interrupt? */
3053 return ac->entry[--ac->avail];
3056 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3059 might_sleep_if(flags & __GFP_WAIT);
3061 kmem_flagcheck(cachep, flags);
3066 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3067 gfp_t flags, void *objp, void *caller)
3071 if (cachep->flags & SLAB_POISON) {
3072 #ifdef CONFIG_DEBUG_PAGEALLOC
3073 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3074 kernel_map_pages(virt_to_page(objp),
3075 cachep->buffer_size / PAGE_SIZE, 1);
3077 check_poison_obj(cachep, objp);
3079 check_poison_obj(cachep, objp);
3081 poison_obj(cachep, objp, POISON_INUSE);
3083 if (cachep->flags & SLAB_STORE_USER)
3084 *dbg_userword(cachep, objp) = caller;
3086 if (cachep->flags & SLAB_RED_ZONE) {
3087 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3088 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3089 slab_error(cachep, "double free, or memory outside"
3090 " object was overwritten");
3092 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3093 objp, *dbg_redzone1(cachep, objp),
3094 *dbg_redzone2(cachep, objp));
3096 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3097 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3099 #ifdef CONFIG_DEBUG_SLAB_LEAK
3104 slabp = page_get_slab(virt_to_head_page(objp));
3105 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3106 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3109 objp += obj_offset(cachep);
3110 if (cachep->ctor && cachep->flags & SLAB_POISON)
3112 #if ARCH_SLAB_MINALIGN
3113 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3114 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3115 objp, ARCH_SLAB_MINALIGN);
3121 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3124 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3126 if (cachep == &cache_cache)
3129 return should_failslab(obj_size(cachep), flags);
3132 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3135 struct array_cache *ac;
3139 ac = cpu_cache_get(cachep);
3140 if (likely(ac->avail)) {
3141 STATS_INC_ALLOCHIT(cachep);
3143 objp = ac->entry[--ac->avail];
3145 STATS_INC_ALLOCMISS(cachep);
3146 objp = cache_alloc_refill(cachep, flags);
3153 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3155 * If we are in_interrupt, then process context, including cpusets and
3156 * mempolicy, may not apply and should not be used for allocation policy.
3158 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3160 int nid_alloc, nid_here;
3162 if (in_interrupt() || (flags & __GFP_THISNODE))
3164 nid_alloc = nid_here = numa_node_id();
3165 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3166 nid_alloc = cpuset_mem_spread_node();
3167 else if (current->mempolicy)
3168 nid_alloc = slab_node(current->mempolicy);
3169 if (nid_alloc != nid_here)
3170 return ____cache_alloc_node(cachep, flags, nid_alloc);
3175 * Fallback function if there was no memory available and no objects on a
3176 * certain node and fall back is permitted. First we scan all the
3177 * available nodelists for available objects. If that fails then we
3178 * perform an allocation without specifying a node. This allows the page
3179 * allocator to do its reclaim / fallback magic. We then insert the
3180 * slab into the proper nodelist and then allocate from it.
3182 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3184 struct zonelist *zonelist;
3188 enum zone_type high_zoneidx = gfp_zone(flags);
3192 if (flags & __GFP_THISNODE)
3195 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3196 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3200 * Look through allowed nodes for objects available
3201 * from existing per node queues.
3203 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3204 nid = zone_to_nid(zone);
3206 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3207 cache->nodelists[nid] &&
3208 cache->nodelists[nid]->free_objects) {
3209 obj = ____cache_alloc_node(cache,
3210 flags | GFP_THISNODE, nid);
3218 * This allocation will be performed within the constraints
3219 * of the current cpuset / memory policy requirements.
3220 * We may trigger various forms of reclaim on the allowed
3221 * set and go into memory reserves if necessary.
3223 if (local_flags & __GFP_WAIT)
3225 kmem_flagcheck(cache, flags);
3226 obj = kmem_getpages(cache, local_flags, -1);
3227 if (local_flags & __GFP_WAIT)
3228 local_irq_disable();
3231 * Insert into the appropriate per node queues
3233 nid = page_to_nid(virt_to_page(obj));
3234 if (cache_grow(cache, flags, nid, obj)) {
3235 obj = ____cache_alloc_node(cache,
3236 flags | GFP_THISNODE, nid);
3239 * Another processor may allocate the
3240 * objects in the slab since we are
3241 * not holding any locks.
3245 /* cache_grow already freed obj */
3254 * A interface to enable slab creation on nodeid
3256 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3259 struct list_head *entry;
3261 struct kmem_list3 *l3;
3265 l3 = cachep->nodelists[nodeid];
3270 spin_lock(&l3->list_lock);
3271 entry = l3->slabs_partial.next;
3272 if (entry == &l3->slabs_partial) {
3273 l3->free_touched = 1;
3274 entry = l3->slabs_free.next;
3275 if (entry == &l3->slabs_free)
3279 slabp = list_entry(entry, struct slab, list);
3280 check_spinlock_acquired_node(cachep, nodeid);
3281 check_slabp(cachep, slabp);
3283 STATS_INC_NODEALLOCS(cachep);
3284 STATS_INC_ACTIVE(cachep);
3285 STATS_SET_HIGH(cachep);
3287 BUG_ON(slabp->inuse == cachep->num);
3289 obj = slab_get_obj(cachep, slabp, nodeid);
3290 check_slabp(cachep, slabp);
3292 /* move slabp to correct slabp list: */
3293 list_del(&slabp->list);
3295 if (slabp->free == BUFCTL_END)
3296 list_add(&slabp->list, &l3->slabs_full);
3298 list_add(&slabp->list, &l3->slabs_partial);
3300 spin_unlock(&l3->list_lock);
3304 spin_unlock(&l3->list_lock);
3305 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3309 return fallback_alloc(cachep, flags);
3316 * kmem_cache_alloc_node - Allocate an object on the specified node
3317 * @cachep: The cache to allocate from.
3318 * @flags: See kmalloc().
3319 * @nodeid: node number of the target node.
3320 * @caller: return address of caller, used for debug information
3322 * Identical to kmem_cache_alloc but it will allocate memory on the given
3323 * node, which can improve the performance for cpu bound structures.
3325 * Fallback to other node is possible if __GFP_THISNODE is not set.
3327 static __always_inline void *
3328 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3331 unsigned long save_flags;
3334 lockdep_trace_alloc(flags);
3336 if (slab_should_failslab(cachep, flags))
3339 cache_alloc_debugcheck_before(cachep, flags);
3340 local_irq_save(save_flags);
3342 if (unlikely(nodeid == -1))
3343 nodeid = numa_node_id();
3345 if (unlikely(!cachep->nodelists[nodeid])) {
3346 /* Node not bootstrapped yet */
3347 ptr = fallback_alloc(cachep, flags);
3351 if (nodeid == numa_node_id()) {
3353 * Use the locally cached objects if possible.
3354 * However ____cache_alloc does not allow fallback
3355 * to other nodes. It may fail while we still have
3356 * objects on other nodes available.
3358 ptr = ____cache_alloc(cachep, flags);
3362 /* ___cache_alloc_node can fall back to other nodes */
3363 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3365 local_irq_restore(save_flags);
3366 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3368 if (unlikely((flags & __GFP_ZERO) && ptr))
3369 memset(ptr, 0, obj_size(cachep));
3374 static __always_inline void *
3375 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3379 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3380 objp = alternate_node_alloc(cache, flags);
3384 objp = ____cache_alloc(cache, flags);
3387 * We may just have run out of memory on the local node.
3388 * ____cache_alloc_node() knows how to locate memory on other nodes
3391 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3398 static __always_inline void *
3399 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3401 return ____cache_alloc(cachep, flags);
3404 #endif /* CONFIG_NUMA */
3406 static __always_inline void *
3407 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3409 unsigned long save_flags;
3412 lockdep_trace_alloc(flags);
3414 if (slab_should_failslab(cachep, flags))
3417 cache_alloc_debugcheck_before(cachep, flags);
3418 local_irq_save(save_flags);
3419 objp = __do_cache_alloc(cachep, flags);
3420 local_irq_restore(save_flags);
3421 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3424 if (unlikely((flags & __GFP_ZERO) && objp))
3425 memset(objp, 0, obj_size(cachep));
3431 * Caller needs to acquire correct kmem_list's list_lock
3433 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3437 struct kmem_list3 *l3;
3439 for (i = 0; i < nr_objects; i++) {
3440 void *objp = objpp[i];
3443 slabp = virt_to_slab(objp);
3444 l3 = cachep->nodelists[node];
3445 list_del(&slabp->list);
3446 check_spinlock_acquired_node(cachep, node);
3447 check_slabp(cachep, slabp);
3448 slab_put_obj(cachep, slabp, objp, node);
3449 STATS_DEC_ACTIVE(cachep);
3451 check_slabp(cachep, slabp);
3453 /* fixup slab chains */
3454 if (slabp->inuse == 0) {
3455 if (l3->free_objects > l3->free_limit) {
3456 l3->free_objects -= cachep->num;
3457 /* No need to drop any previously held
3458 * lock here, even if we have a off-slab slab
3459 * descriptor it is guaranteed to come from
3460 * a different cache, refer to comments before
3463 slab_destroy(cachep, slabp);
3465 list_add(&slabp->list, &l3->slabs_free);
3468 /* Unconditionally move a slab to the end of the
3469 * partial list on free - maximum time for the
3470 * other objects to be freed, too.
3472 list_add_tail(&slabp->list, &l3->slabs_partial);
3477 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3480 struct kmem_list3 *l3;
3481 int node = numa_node_id();
3483 batchcount = ac->batchcount;
3485 BUG_ON(!batchcount || batchcount > ac->avail);
3488 l3 = cachep->nodelists[node];
3489 spin_lock(&l3->list_lock);
3491 struct array_cache *shared_array = l3->shared;
3492 int max = shared_array->limit - shared_array->avail;
3494 if (batchcount > max)
3496 memcpy(&(shared_array->entry[shared_array->avail]),
3497 ac->entry, sizeof(void *) * batchcount);
3498 shared_array->avail += batchcount;
3503 free_block(cachep, ac->entry, batchcount, node);
3508 struct list_head *p;
3510 p = l3->slabs_free.next;
3511 while (p != &(l3->slabs_free)) {
3514 slabp = list_entry(p, struct slab, list);
3515 BUG_ON(slabp->inuse);
3520 STATS_SET_FREEABLE(cachep, i);
3523 spin_unlock(&l3->list_lock);
3524 ac->avail -= batchcount;
3525 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3529 * Release an obj back to its cache. If the obj has a constructed state, it must
3530 * be in this state _before_ it is released. Called with disabled ints.
3532 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3534 struct array_cache *ac = cpu_cache_get(cachep);
3537 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3540 * Skip calling cache_free_alien() when the platform is not numa.
3541 * This will avoid cache misses that happen while accessing slabp (which
3542 * is per page memory reference) to get nodeid. Instead use a global
3543 * variable to skip the call, which is mostly likely to be present in
3546 if (numa_platform && cache_free_alien(cachep, objp))
3549 if (likely(ac->avail < ac->limit)) {
3550 STATS_INC_FREEHIT(cachep);
3551 ac->entry[ac->avail++] = objp;
3554 STATS_INC_FREEMISS(cachep);
3555 cache_flusharray(cachep, ac);
3556 ac->entry[ac->avail++] = objp;
3561 * kmem_cache_alloc - Allocate an object
3562 * @cachep: The cache to allocate from.
3563 * @flags: See kmalloc().
3565 * Allocate an object from this cache. The flags are only relevant
3566 * if the cache has no available objects.
3568 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3570 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3572 trace_kmem_cache_alloc(_RET_IP_, ret,
3573 obj_size(cachep), cachep->buffer_size, flags);
3577 EXPORT_SYMBOL(kmem_cache_alloc);
3579 #ifdef CONFIG_KMEMTRACE
3580 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3582 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3584 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3588 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3589 * @cachep: the cache we're checking against
3590 * @ptr: pointer to validate
3592 * This verifies that the untrusted pointer looks sane;
3593 * it is _not_ a guarantee that the pointer is actually
3594 * part of the slab cache in question, but it at least
3595 * validates that the pointer can be dereferenced and
3596 * looks half-way sane.
3598 * Currently only used for dentry validation.
3600 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3602 unsigned long addr = (unsigned long)ptr;
3603 unsigned long min_addr = PAGE_OFFSET;
3604 unsigned long align_mask = BYTES_PER_WORD - 1;
3605 unsigned long size = cachep->buffer_size;
3608 if (unlikely(addr < min_addr))
3610 if (unlikely(addr > (unsigned long)high_memory - size))
3612 if (unlikely(addr & align_mask))
3614 if (unlikely(!kern_addr_valid(addr)))
3616 if (unlikely(!kern_addr_valid(addr + size - 1)))
3618 page = virt_to_page(ptr);
3619 if (unlikely(!PageSlab(page)))
3621 if (unlikely(page_get_cache(page) != cachep))
3629 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3631 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3632 __builtin_return_address(0));
3634 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3635 obj_size(cachep), cachep->buffer_size,
3640 EXPORT_SYMBOL(kmem_cache_alloc_node);
3642 #ifdef CONFIG_KMEMTRACE
3643 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3647 return __cache_alloc_node(cachep, flags, nodeid,
3648 __builtin_return_address(0));
3650 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3653 static __always_inline void *
3654 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3656 struct kmem_cache *cachep;
3659 cachep = kmem_find_general_cachep(size, flags);
3660 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3662 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3664 trace_kmalloc_node((unsigned long) caller, ret,
3665 size, cachep->buffer_size, flags, node);
3670 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3671 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3673 return __do_kmalloc_node(size, flags, node,
3674 __builtin_return_address(0));
3676 EXPORT_SYMBOL(__kmalloc_node);
3678 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3679 int node, unsigned long caller)
3681 return __do_kmalloc_node(size, flags, node, (void *)caller);
3683 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3685 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3687 return __do_kmalloc_node(size, flags, node, NULL);
3689 EXPORT_SYMBOL(__kmalloc_node);
3690 #endif /* CONFIG_DEBUG_SLAB */
3691 #endif /* CONFIG_NUMA */
3694 * __do_kmalloc - allocate memory
3695 * @size: how many bytes of memory are required.
3696 * @flags: the type of memory to allocate (see kmalloc).
3697 * @caller: function caller for debug tracking of the caller
3699 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3702 struct kmem_cache *cachep;
3705 /* If you want to save a few bytes .text space: replace
3707 * Then kmalloc uses the uninlined functions instead of the inline
3710 cachep = __find_general_cachep(size, flags);
3711 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3713 ret = __cache_alloc(cachep, flags, caller);
3715 trace_kmalloc((unsigned long) caller, ret,
3716 size, cachep->buffer_size, flags);
3722 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3723 void *__kmalloc(size_t size, gfp_t flags)
3725 return __do_kmalloc(size, flags, __builtin_return_address(0));
3727 EXPORT_SYMBOL(__kmalloc);
3729 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3731 return __do_kmalloc(size, flags, (void *)caller);
3733 EXPORT_SYMBOL(__kmalloc_track_caller);
3736 void *__kmalloc(size_t size, gfp_t flags)
3738 return __do_kmalloc(size, flags, NULL);
3740 EXPORT_SYMBOL(__kmalloc);
3744 * kmem_cache_free - Deallocate an object
3745 * @cachep: The cache the allocation was from.
3746 * @objp: The previously allocated object.
3748 * Free an object which was previously allocated from this
3751 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3753 unsigned long flags;
3755 local_irq_save(flags);
3756 debug_check_no_locks_freed(objp, obj_size(cachep));
3757 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3758 debug_check_no_obj_freed(objp, obj_size(cachep));
3759 __cache_free(cachep, objp);
3760 local_irq_restore(flags);
3762 trace_kmem_cache_free(_RET_IP_, objp);
3764 EXPORT_SYMBOL(kmem_cache_free);
3767 * kfree - free previously allocated memory
3768 * @objp: pointer returned by kmalloc.
3770 * If @objp is NULL, no operation is performed.
3772 * Don't free memory not originally allocated by kmalloc()
3773 * or you will run into trouble.
3775 void kfree(const void *objp)
3777 struct kmem_cache *c;
3778 unsigned long flags;
3780 trace_kfree(_RET_IP_, objp);
3782 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3784 local_irq_save(flags);
3785 kfree_debugcheck(objp);
3786 c = virt_to_cache(objp);
3787 debug_check_no_locks_freed(objp, obj_size(c));
3788 debug_check_no_obj_freed(objp, obj_size(c));
3789 __cache_free(c, (void *)objp);
3790 local_irq_restore(flags);
3792 EXPORT_SYMBOL(kfree);
3794 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3796 return obj_size(cachep);
3798 EXPORT_SYMBOL(kmem_cache_size);
3800 const char *kmem_cache_name(struct kmem_cache *cachep)
3802 return cachep->name;
3804 EXPORT_SYMBOL_GPL(kmem_cache_name);
3807 * This initializes kmem_list3 or resizes various caches for all nodes.
3809 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3812 struct kmem_list3 *l3;
3813 struct array_cache *new_shared;
3814 struct array_cache **new_alien = NULL;
3816 for_each_online_node(node) {
3818 if (use_alien_caches) {
3819 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3825 if (cachep->shared) {
3826 new_shared = alloc_arraycache(node,
3827 cachep->shared*cachep->batchcount,
3830 free_alien_cache(new_alien);
3835 l3 = cachep->nodelists[node];
3837 struct array_cache *shared = l3->shared;
3839 spin_lock_irq(&l3->list_lock);
3842 free_block(cachep, shared->entry,
3843 shared->avail, node);
3845 l3->shared = new_shared;
3847 l3->alien = new_alien;
3850 l3->free_limit = (1 + nr_cpus_node(node)) *
3851 cachep->batchcount + cachep->num;
3852 spin_unlock_irq(&l3->list_lock);
3854 free_alien_cache(new_alien);
3857 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3859 free_alien_cache(new_alien);
3864 kmem_list3_init(l3);
3865 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3866 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3867 l3->shared = new_shared;
3868 l3->alien = new_alien;
3869 l3->free_limit = (1 + nr_cpus_node(node)) *
3870 cachep->batchcount + cachep->num;
3871 cachep->nodelists[node] = l3;
3876 if (!cachep->next.next) {
3877 /* Cache is not active yet. Roll back what we did */
3880 if (cachep->nodelists[node]) {
3881 l3 = cachep->nodelists[node];
3884 free_alien_cache(l3->alien);
3886 cachep->nodelists[node] = NULL;
3894 struct ccupdate_struct {
3895 struct kmem_cache *cachep;
3896 struct array_cache *new[NR_CPUS];
3899 static void do_ccupdate_local(void *info)
3901 struct ccupdate_struct *new = info;
3902 struct array_cache *old;
3905 old = cpu_cache_get(new->cachep);
3907 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3908 new->new[smp_processor_id()] = old;
3911 /* Always called with the cache_chain_mutex held */
3912 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3913 int batchcount, int shared, gfp_t gfp)
3915 struct ccupdate_struct *new;
3918 new = kzalloc(sizeof(*new), gfp);
3922 for_each_online_cpu(i) {
3923 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3926 for (i--; i >= 0; i--)
3932 new->cachep = cachep;
3934 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3937 cachep->batchcount = batchcount;
3938 cachep->limit = limit;
3939 cachep->shared = shared;
3941 for_each_online_cpu(i) {
3942 struct array_cache *ccold = new->new[i];
3945 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3946 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3947 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3951 return alloc_kmemlist(cachep, gfp);
3954 /* Called with cache_chain_mutex held always */
3955 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3961 * The head array serves three purposes:
3962 * - create a LIFO ordering, i.e. return objects that are cache-warm
3963 * - reduce the number of spinlock operations.
3964 * - reduce the number of linked list operations on the slab and
3965 * bufctl chains: array operations are cheaper.
3966 * The numbers are guessed, we should auto-tune as described by
3969 if (cachep->buffer_size > 131072)
3971 else if (cachep->buffer_size > PAGE_SIZE)
3973 else if (cachep->buffer_size > 1024)
3975 else if (cachep->buffer_size > 256)
3981 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3982 * allocation behaviour: Most allocs on one cpu, most free operations
3983 * on another cpu. For these cases, an efficient object passing between
3984 * cpus is necessary. This is provided by a shared array. The array
3985 * replaces Bonwick's magazine layer.
3986 * On uniprocessor, it's functionally equivalent (but less efficient)
3987 * to a larger limit. Thus disabled by default.
3990 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3995 * With debugging enabled, large batchcount lead to excessively long
3996 * periods with disabled local interrupts. Limit the batchcount
4001 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4003 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4004 cachep->name, -err);
4009 * Drain an array if it contains any elements taking the l3 lock only if
4010 * necessary. Note that the l3 listlock also protects the array_cache
4011 * if drain_array() is used on the shared array.
4013 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4014 struct array_cache *ac, int force, int node)
4018 if (!ac || !ac->avail)
4020 if (ac->touched && !force) {
4023 spin_lock_irq(&l3->list_lock);
4025 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4026 if (tofree > ac->avail)
4027 tofree = (ac->avail + 1) / 2;
4028 free_block(cachep, ac->entry, tofree, node);
4029 ac->avail -= tofree;
4030 memmove(ac->entry, &(ac->entry[tofree]),
4031 sizeof(void *) * ac->avail);
4033 spin_unlock_irq(&l3->list_lock);
4038 * cache_reap - Reclaim memory from caches.
4039 * @w: work descriptor
4041 * Called from workqueue/eventd every few seconds.
4043 * - clear the per-cpu caches for this CPU.
4044 * - return freeable pages to the main free memory pool.
4046 * If we cannot acquire the cache chain mutex then just give up - we'll try
4047 * again on the next iteration.
4049 static void cache_reap(struct work_struct *w)
4051 struct kmem_cache *searchp;
4052 struct kmem_list3 *l3;
4053 int node = numa_node_id();
4054 struct delayed_work *work = to_delayed_work(w);
4056 if (!mutex_trylock(&cache_chain_mutex))
4057 /* Give up. Setup the next iteration. */
4060 list_for_each_entry(searchp, &cache_chain, next) {
4064 * We only take the l3 lock if absolutely necessary and we
4065 * have established with reasonable certainty that
4066 * we can do some work if the lock was obtained.
4068 l3 = searchp->nodelists[node];
4070 reap_alien(searchp, l3);
4072 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4075 * These are racy checks but it does not matter
4076 * if we skip one check or scan twice.
4078 if (time_after(l3->next_reap, jiffies))
4081 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4083 drain_array(searchp, l3, l3->shared, 0, node);
4085 if (l3->free_touched)
4086 l3->free_touched = 0;
4090 freed = drain_freelist(searchp, l3, (l3->free_limit +
4091 5 * searchp->num - 1) / (5 * searchp->num));
4092 STATS_ADD_REAPED(searchp, freed);
4098 mutex_unlock(&cache_chain_mutex);
4101 /* Set up the next iteration */
4102 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4105 #ifdef CONFIG_SLABINFO
4107 static void print_slabinfo_header(struct seq_file *m)
4110 * Output format version, so at least we can change it
4111 * without _too_ many complaints.
4114 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4116 seq_puts(m, "slabinfo - version: 2.1\n");
4118 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4119 "<objperslab> <pagesperslab>");
4120 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4121 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4123 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4124 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4125 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4130 static void *s_start(struct seq_file *m, loff_t *pos)
4134 mutex_lock(&cache_chain_mutex);
4136 print_slabinfo_header(m);
4138 return seq_list_start(&cache_chain, *pos);
4141 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4143 return seq_list_next(p, &cache_chain, pos);
4146 static void s_stop(struct seq_file *m, void *p)
4148 mutex_unlock(&cache_chain_mutex);
4151 static int s_show(struct seq_file *m, void *p)
4153 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4155 unsigned long active_objs;
4156 unsigned long num_objs;
4157 unsigned long active_slabs = 0;
4158 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4162 struct kmem_list3 *l3;
4166 for_each_online_node(node) {
4167 l3 = cachep->nodelists[node];
4172 spin_lock_irq(&l3->list_lock);
4174 list_for_each_entry(slabp, &l3->slabs_full, list) {
4175 if (slabp->inuse != cachep->num && !error)
4176 error = "slabs_full accounting error";
4177 active_objs += cachep->num;
4180 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4181 if (slabp->inuse == cachep->num && !error)
4182 error = "slabs_partial inuse accounting error";
4183 if (!slabp->inuse && !error)
4184 error = "slabs_partial/inuse accounting error";
4185 active_objs += slabp->inuse;
4188 list_for_each_entry(slabp, &l3->slabs_free, list) {
4189 if (slabp->inuse && !error)
4190 error = "slabs_free/inuse accounting error";
4193 free_objects += l3->free_objects;
4195 shared_avail += l3->shared->avail;
4197 spin_unlock_irq(&l3->list_lock);
4199 num_slabs += active_slabs;
4200 num_objs = num_slabs * cachep->num;
4201 if (num_objs - active_objs != free_objects && !error)
4202 error = "free_objects accounting error";
4204 name = cachep->name;
4206 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4208 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4209 name, active_objs, num_objs, cachep->buffer_size,
4210 cachep->num, (1 << cachep->gfporder));
4211 seq_printf(m, " : tunables %4u %4u %4u",
4212 cachep->limit, cachep->batchcount, cachep->shared);
4213 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4214 active_slabs, num_slabs, shared_avail);
4217 unsigned long high = cachep->high_mark;
4218 unsigned long allocs = cachep->num_allocations;
4219 unsigned long grown = cachep->grown;
4220 unsigned long reaped = cachep->reaped;
4221 unsigned long errors = cachep->errors;
4222 unsigned long max_freeable = cachep->max_freeable;
4223 unsigned long node_allocs = cachep->node_allocs;
4224 unsigned long node_frees = cachep->node_frees;
4225 unsigned long overflows = cachep->node_overflow;
4227 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4228 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4229 reaped, errors, max_freeable, node_allocs,
4230 node_frees, overflows);
4234 unsigned long allochit = atomic_read(&cachep->allochit);
4235 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4236 unsigned long freehit = atomic_read(&cachep->freehit);
4237 unsigned long freemiss = atomic_read(&cachep->freemiss);
4239 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4240 allochit, allocmiss, freehit, freemiss);
4248 * slabinfo_op - iterator that generates /proc/slabinfo
4257 * num-pages-per-slab
4258 * + further values on SMP and with statistics enabled
4261 static const struct seq_operations slabinfo_op = {
4268 #define MAX_SLABINFO_WRITE 128
4270 * slabinfo_write - Tuning for the slab allocator
4272 * @buffer: user buffer
4273 * @count: data length
4276 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4277 size_t count, loff_t *ppos)
4279 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4280 int limit, batchcount, shared, res;
4281 struct kmem_cache *cachep;
4283 if (count > MAX_SLABINFO_WRITE)
4285 if (copy_from_user(&kbuf, buffer, count))
4287 kbuf[MAX_SLABINFO_WRITE] = '\0';
4289 tmp = strchr(kbuf, ' ');
4294 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4297 /* Find the cache in the chain of caches. */
4298 mutex_lock(&cache_chain_mutex);
4300 list_for_each_entry(cachep, &cache_chain, next) {
4301 if (!strcmp(cachep->name, kbuf)) {
4302 if (limit < 1 || batchcount < 1 ||
4303 batchcount > limit || shared < 0) {
4306 res = do_tune_cpucache(cachep, limit,
4313 mutex_unlock(&cache_chain_mutex);
4319 static int slabinfo_open(struct inode *inode, struct file *file)
4321 return seq_open(file, &slabinfo_op);
4324 static const struct file_operations proc_slabinfo_operations = {
4325 .open = slabinfo_open,
4327 .write = slabinfo_write,
4328 .llseek = seq_lseek,
4329 .release = seq_release,
4332 #ifdef CONFIG_DEBUG_SLAB_LEAK
4334 static void *leaks_start(struct seq_file *m, loff_t *pos)
4336 mutex_lock(&cache_chain_mutex);
4337 return seq_list_start(&cache_chain, *pos);
4340 static inline int add_caller(unsigned long *n, unsigned long v)
4350 unsigned long *q = p + 2 * i;
4364 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4370 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4376 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4377 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4379 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4384 static void show_symbol(struct seq_file *m, unsigned long address)
4386 #ifdef CONFIG_KALLSYMS
4387 unsigned long offset, size;
4388 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4390 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4391 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4393 seq_printf(m, " [%s]", modname);
4397 seq_printf(m, "%p", (void *)address);
4400 static int leaks_show(struct seq_file *m, void *p)
4402 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4404 struct kmem_list3 *l3;
4406 unsigned long *n = m->private;
4410 if (!(cachep->flags & SLAB_STORE_USER))
4412 if (!(cachep->flags & SLAB_RED_ZONE))
4415 /* OK, we can do it */
4419 for_each_online_node(node) {
4420 l3 = cachep->nodelists[node];
4425 spin_lock_irq(&l3->list_lock);
4427 list_for_each_entry(slabp, &l3->slabs_full, list)
4428 handle_slab(n, cachep, slabp);
4429 list_for_each_entry(slabp, &l3->slabs_partial, list)
4430 handle_slab(n, cachep, slabp);
4431 spin_unlock_irq(&l3->list_lock);
4433 name = cachep->name;
4435 /* Increase the buffer size */
4436 mutex_unlock(&cache_chain_mutex);
4437 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4439 /* Too bad, we are really out */
4441 mutex_lock(&cache_chain_mutex);
4444 *(unsigned long *)m->private = n[0] * 2;
4446 mutex_lock(&cache_chain_mutex);
4447 /* Now make sure this entry will be retried */
4451 for (i = 0; i < n[1]; i++) {
4452 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4453 show_symbol(m, n[2*i+2]);
4460 static const struct seq_operations slabstats_op = {
4461 .start = leaks_start,
4467 static int slabstats_open(struct inode *inode, struct file *file)
4469 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4472 ret = seq_open(file, &slabstats_op);
4474 struct seq_file *m = file->private_data;
4475 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4484 static const struct file_operations proc_slabstats_operations = {
4485 .open = slabstats_open,
4487 .llseek = seq_lseek,
4488 .release = seq_release_private,
4492 static int __init slab_proc_init(void)
4494 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4495 #ifdef CONFIG_DEBUG_SLAB_LEAK
4496 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4500 module_init(slab_proc_init);
4504 * ksize - get the actual amount of memory allocated for a given object
4505 * @objp: Pointer to the object
4507 * kmalloc may internally round up allocations and return more memory
4508 * than requested. ksize() can be used to determine the actual amount of
4509 * memory allocated. The caller may use this additional memory, even though
4510 * a smaller amount of memory was initially specified with the kmalloc call.
4511 * The caller must guarantee that objp points to a valid object previously
4512 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4513 * must not be freed during the duration of the call.
4515 size_t ksize(const void *objp)
4518 if (unlikely(objp == ZERO_SIZE_PTR))
4521 return obj_size(virt_to_cache(objp));
4523 EXPORT_SYMBOL(ksize);