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/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
118 #include <asm/cacheflush.h>
119 #include <asm/tlbflush.h>
120 #include <asm/page.h>
123 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
124 * 0 for faster, smaller code (especially in the critical paths).
126 * STATS - 1 to collect stats for /proc/slabinfo.
127 * 0 for faster, smaller code (especially in the critical paths).
129 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 #ifdef CONFIG_DEBUG_SLAB
135 #define FORCED_DEBUG 1
139 #define FORCED_DEBUG 0
142 /* Shouldn't this be in a header file somewhere? */
143 #define BYTES_PER_WORD sizeof(void *)
144 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
146 #ifndef ARCH_KMALLOC_MINALIGN
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
156 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
159 #ifndef ARCH_SLAB_MINALIGN
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
167 #define ARCH_SLAB_MINALIGN 0
170 #ifndef ARCH_KMALLOC_FLAGS
171 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
174 /* Legal flag mask for kmem_cache_create(). */
176 # define CREATE_MASK (SLAB_RED_ZONE | \
177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
182 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
184 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
186 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
188 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
194 * Bufctl's are used for linking objs within a slab
197 * This implementation relies on "struct page" for locating the cache &
198 * slab an object belongs to.
199 * This allows the bufctl structure to be small (one int), but limits
200 * the number of objects a slab (not a cache) can contain when off-slab
201 * bufctls are used. The limit is the size of the largest general cache
202 * that does not use off-slab slabs.
203 * For 32bit archs with 4 kB pages, is this 56.
204 * This is not serious, as it is only for large objects, when it is unwise
205 * to have too many per slab.
206 * Note: This limit can be raised by introducing a general cache whose size
207 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
210 typedef unsigned int kmem_bufctl_t;
211 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
212 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
213 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
214 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
229 unsigned short nodeid;
235 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
236 * arrange for kmem_freepages to be called via RCU. This is useful if
237 * we need to approach a kernel structure obliquely, from its address
238 * obtained without the usual locking. We can lock the structure to
239 * stabilize it and check it's still at the given address, only if we
240 * can be sure that the memory has not been meanwhile reused for some
241 * other kind of object (which our subsystem's lock might corrupt).
243 * rcu_read_lock before reading the address, then rcu_read_unlock after
244 * taking the spinlock within the structure expected at that address.
246 * We assume struct slab_rcu can overlay struct slab when destroying.
249 struct rcu_head head;
250 struct kmem_cache *cachep;
258 * - LIFO ordering, to hand out cache-warm objects from _alloc
259 * - reduce the number of linked list operations
260 * - reduce spinlock operations
262 * The limit is stored in the per-cpu structure to reduce the data cache
269 unsigned int batchcount;
270 unsigned int touched;
273 * Must have this definition in here for the proper
274 * alignment of array_cache. Also simplifies accessing
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
283 #define BOOT_CPUCACHE_ENTRIES 1
284 struct arraycache_init {
285 struct array_cache cache;
286 void *entries[BOOT_CPUCACHE_ENTRIES];
290 * The slab lists for all objects.
293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
297 unsigned int free_limit;
298 unsigned int colour_next; /* Per-node cache coloring */
299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
307 * Need this for bootstrapping a per node allocator.
309 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
310 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
311 #define CACHE_CACHE 0
312 #define SIZE_AC MAX_NUMNODES
313 #define SIZE_L3 (2 * MAX_NUMNODES)
315 static int drain_freelist(struct kmem_cache *cache,
316 struct kmem_list3 *l3, int tofree);
317 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
319 static int enable_cpucache(struct kmem_cache *cachep);
320 static void cache_reap(struct work_struct *unused);
323 * This function must be completely optimized away if a constant is passed to
324 * it. Mostly the same as what is in linux/slab.h except it returns an index.
326 static __always_inline int index_of(const size_t size)
328 extern void __bad_size(void);
330 if (__builtin_constant_p(size)) {
338 #include <linux/kmalloc_sizes.h>
346 static int slab_early_init = 1;
348 #define INDEX_AC index_of(sizeof(struct arraycache_init))
349 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
351 static void kmem_list3_init(struct kmem_list3 *parent)
353 INIT_LIST_HEAD(&parent->slabs_full);
354 INIT_LIST_HEAD(&parent->slabs_partial);
355 INIT_LIST_HEAD(&parent->slabs_free);
356 parent->shared = NULL;
357 parent->alien = NULL;
358 parent->colour_next = 0;
359 spin_lock_init(&parent->list_lock);
360 parent->free_objects = 0;
361 parent->free_touched = 0;
364 #define MAKE_LIST(cachep, listp, slab, nodeid) \
366 INIT_LIST_HEAD(listp); \
367 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
370 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
372 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
384 /* 1) per-cpu data, touched during every alloc/free */
385 struct array_cache *array[NR_CPUS];
386 /* 2) Cache tunables. Protected by cache_chain_mutex */
387 unsigned int batchcount;
391 unsigned int buffer_size;
392 u32 reciprocal_buffer_size;
393 /* 3) touched by every alloc & free from the backend */
395 unsigned int flags; /* constant flags */
396 unsigned int num; /* # of objs per slab */
398 /* 4) cache_grow/shrink */
399 /* order of pgs per slab (2^n) */
400 unsigned int gfporder;
402 /* force GFP flags, e.g. GFP_DMA */
405 size_t colour; /* cache colouring range */
406 unsigned int colour_off; /* colour offset */
407 struct kmem_cache *slabp_cache;
408 unsigned int slab_size;
409 unsigned int dflags; /* dynamic flags */
411 /* constructor func */
412 void (*ctor)(void *obj);
414 /* 5) cache creation/removal */
416 struct list_head next;
420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
429 unsigned long node_overflow;
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
446 * We put nodelists[] at the end of kmem_cache, because we want to size
447 * this array to nr_node_ids slots instead of MAX_NUMNODES
448 * (see kmem_cache_init())
449 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
450 * is statically defined, so we reserve the max number of nodes.
452 struct kmem_list3 *nodelists[MAX_NUMNODES];
454 * Do not add fields after nodelists[]
458 #define CFLGS_OFF_SLAB (0x80000000UL)
459 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
461 #define BATCHREFILL_LIMIT 16
463 * Optimization question: fewer reaps means less probability for unnessary
464 * cpucache drain/refill cycles.
466 * OTOH the cpuarrays can contain lots of objects,
467 * which could lock up otherwise freeable slabs.
469 #define REAPTIMEOUT_CPUC (2*HZ)
470 #define REAPTIMEOUT_LIST3 (4*HZ)
473 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
474 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
475 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
476 #define STATS_INC_GROWN(x) ((x)->grown++)
477 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
478 #define STATS_SET_HIGH(x) \
480 if ((x)->num_active > (x)->high_mark) \
481 (x)->high_mark = (x)->num_active; \
483 #define STATS_INC_ERR(x) ((x)->errors++)
484 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
485 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
486 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
487 #define STATS_SET_FREEABLE(x, i) \
489 if ((x)->max_freeable < i) \
490 (x)->max_freeable = i; \
492 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
493 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
494 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
495 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
497 #define STATS_INC_ACTIVE(x) do { } while (0)
498 #define STATS_DEC_ACTIVE(x) do { } while (0)
499 #define STATS_INC_ALLOCED(x) do { } while (0)
500 #define STATS_INC_GROWN(x) do { } while (0)
501 #define STATS_ADD_REAPED(x,y) do { } while (0)
502 #define STATS_SET_HIGH(x) do { } while (0)
503 #define STATS_INC_ERR(x) do { } while (0)
504 #define STATS_INC_NODEALLOCS(x) do { } while (0)
505 #define STATS_INC_NODEFREES(x) do { } while (0)
506 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
507 #define STATS_SET_FREEABLE(x, i) do { } while (0)
508 #define STATS_INC_ALLOCHIT(x) do { } while (0)
509 #define STATS_INC_ALLOCMISS(x) do { } while (0)
510 #define STATS_INC_FREEHIT(x) do { } while (0)
511 #define STATS_INC_FREEMISS(x) do { } while (0)
517 * memory layout of objects:
519 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
520 * the end of an object is aligned with the end of the real
521 * allocation. Catches writes behind the end of the allocation.
522 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
524 * cachep->obj_offset: The real object.
525 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
526 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
527 * [BYTES_PER_WORD long]
529 static int obj_offset(struct kmem_cache *cachep)
531 return cachep->obj_offset;
534 static int obj_size(struct kmem_cache *cachep)
536 return cachep->obj_size;
539 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
541 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
542 return (unsigned long long*) (objp + obj_offset(cachep) -
543 sizeof(unsigned long long));
546 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
548 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
549 if (cachep->flags & SLAB_STORE_USER)
550 return (unsigned long long *)(objp + cachep->buffer_size -
551 sizeof(unsigned long long) -
553 return (unsigned long long *) (objp + cachep->buffer_size -
554 sizeof(unsigned long long));
557 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
559 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
560 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
565 #define obj_offset(x) 0
566 #define obj_size(cachep) (cachep->buffer_size)
567 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
568 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
569 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
573 #ifdef CONFIG_KMEMTRACE
574 size_t slab_buffer_size(struct kmem_cache *cachep)
576 return cachep->buffer_size;
578 EXPORT_SYMBOL(slab_buffer_size);
582 * Do not go above this order unless 0 objects fit into the slab.
584 #define BREAK_GFP_ORDER_HI 1
585 #define BREAK_GFP_ORDER_LO 0
586 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
589 * Functions for storing/retrieving the cachep and or slab from the page
590 * allocator. These are used to find the slab an obj belongs to. With kfree(),
591 * these are used to find the cache which an obj belongs to.
593 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
595 page->lru.next = (struct list_head *)cache;
598 static inline struct kmem_cache *page_get_cache(struct page *page)
600 page = compound_head(page);
601 BUG_ON(!PageSlab(page));
602 return (struct kmem_cache *)page->lru.next;
605 static inline void page_set_slab(struct page *page, struct slab *slab)
607 page->lru.prev = (struct list_head *)slab;
610 static inline struct slab *page_get_slab(struct page *page)
612 BUG_ON(!PageSlab(page));
613 return (struct slab *)page->lru.prev;
616 static inline struct kmem_cache *virt_to_cache(const void *obj)
618 struct page *page = virt_to_head_page(obj);
619 return page_get_cache(page);
622 static inline struct slab *virt_to_slab(const void *obj)
624 struct page *page = virt_to_head_page(obj);
625 return page_get_slab(page);
628 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
631 return slab->s_mem + cache->buffer_size * idx;
635 * We want to avoid an expensive divide : (offset / cache->buffer_size)
636 * Using the fact that buffer_size is a constant for a particular cache,
637 * we can replace (offset / cache->buffer_size) by
638 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
640 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
641 const struct slab *slab, void *obj)
643 u32 offset = (obj - slab->s_mem);
644 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
648 * These are the default caches for kmalloc. Custom caches can have other sizes.
650 struct cache_sizes malloc_sizes[] = {
651 #define CACHE(x) { .cs_size = (x) },
652 #include <linux/kmalloc_sizes.h>
656 EXPORT_SYMBOL(malloc_sizes);
658 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
664 static struct cache_names __initdata cache_names[] = {
665 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
666 #include <linux/kmalloc_sizes.h>
671 static struct arraycache_init initarray_cache __initdata =
672 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
673 static struct arraycache_init initarray_generic =
674 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
676 /* internal cache of cache description objs */
677 static struct kmem_cache cache_cache = {
679 .limit = BOOT_CPUCACHE_ENTRIES,
681 .buffer_size = sizeof(struct kmem_cache),
682 .name = "kmem_cache",
685 #define BAD_ALIEN_MAGIC 0x01020304ul
687 #ifdef CONFIG_LOCKDEP
690 * Slab sometimes uses the kmalloc slabs to store the slab headers
691 * for other slabs "off slab".
692 * The locking for this is tricky in that it nests within the locks
693 * of all other slabs in a few places; to deal with this special
694 * locking we put on-slab caches into a separate lock-class.
696 * We set lock class for alien array caches which are up during init.
697 * The lock annotation will be lost if all cpus of a node goes down and
698 * then comes back up during hotplug
700 static struct lock_class_key on_slab_l3_key;
701 static struct lock_class_key on_slab_alc_key;
703 static inline void init_lock_keys(void)
707 struct cache_sizes *s = malloc_sizes;
709 while (s->cs_size != ULONG_MAX) {
711 struct array_cache **alc;
713 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
714 if (!l3 || OFF_SLAB(s->cs_cachep))
716 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
719 * FIXME: This check for BAD_ALIEN_MAGIC
720 * should go away when common slab code is taught to
721 * work even without alien caches.
722 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
723 * for alloc_alien_cache,
725 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
729 lockdep_set_class(&alc[r]->lock,
737 static inline void init_lock_keys(void)
743 * Guard access to the cache-chain.
745 static DEFINE_MUTEX(cache_chain_mutex);
746 static struct list_head cache_chain;
749 * chicken and egg problem: delay the per-cpu array allocation
750 * until the general caches are up.
760 * used by boot code to determine if it can use slab based allocator
762 int slab_is_available(void)
764 return g_cpucache_up == FULL;
767 static DEFINE_PER_CPU(struct delayed_work, reap_work);
769 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
771 return cachep->array[smp_processor_id()];
774 static inline struct kmem_cache *__find_general_cachep(size_t size,
777 struct cache_sizes *csizep = malloc_sizes;
780 /* This happens if someone tries to call
781 * kmem_cache_create(), or __kmalloc(), before
782 * the generic caches are initialized.
784 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
787 return ZERO_SIZE_PTR;
789 while (size > csizep->cs_size)
793 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
794 * has cs_{dma,}cachep==NULL. Thus no special case
795 * for large kmalloc calls required.
797 #ifdef CONFIG_ZONE_DMA
798 if (unlikely(gfpflags & GFP_DMA))
799 return csizep->cs_dmacachep;
801 return csizep->cs_cachep;
804 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
806 return __find_general_cachep(size, gfpflags);
809 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
811 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
815 * Calculate the number of objects and left-over bytes for a given buffer size.
817 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
818 size_t align, int flags, size_t *left_over,
823 size_t slab_size = PAGE_SIZE << gfporder;
826 * The slab management structure can be either off the slab or
827 * on it. For the latter case, the memory allocated for a
831 * - One kmem_bufctl_t for each object
832 * - Padding to respect alignment of @align
833 * - @buffer_size bytes for each object
835 * If the slab management structure is off the slab, then the
836 * alignment will already be calculated into the size. Because
837 * the slabs are all pages aligned, the objects will be at the
838 * correct alignment when allocated.
840 if (flags & CFLGS_OFF_SLAB) {
842 nr_objs = slab_size / buffer_size;
844 if (nr_objs > SLAB_LIMIT)
845 nr_objs = SLAB_LIMIT;
848 * Ignore padding for the initial guess. The padding
849 * is at most @align-1 bytes, and @buffer_size is at
850 * least @align. In the worst case, this result will
851 * be one greater than the number of objects that fit
852 * into the memory allocation when taking the padding
855 nr_objs = (slab_size - sizeof(struct slab)) /
856 (buffer_size + sizeof(kmem_bufctl_t));
859 * This calculated number will be either the right
860 * amount, or one greater than what we want.
862 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
866 if (nr_objs > SLAB_LIMIT)
867 nr_objs = SLAB_LIMIT;
869 mgmt_size = slab_mgmt_size(nr_objs, align);
872 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
875 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
877 static void __slab_error(const char *function, struct kmem_cache *cachep,
880 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
881 function, cachep->name, msg);
886 * By default on NUMA we use alien caches to stage the freeing of
887 * objects allocated from other nodes. This causes massive memory
888 * inefficiencies when using fake NUMA setup to split memory into a
889 * large number of small nodes, so it can be disabled on the command
893 static int use_alien_caches __read_mostly = 1;
894 static int numa_platform __read_mostly = 1;
895 static int __init noaliencache_setup(char *s)
897 use_alien_caches = 0;
900 __setup("noaliencache", noaliencache_setup);
904 * Special reaping functions for NUMA systems called from cache_reap().
905 * These take care of doing round robin flushing of alien caches (containing
906 * objects freed on different nodes from which they were allocated) and the
907 * flushing of remote pcps by calling drain_node_pages.
909 static DEFINE_PER_CPU(unsigned long, reap_node);
911 static void init_reap_node(int cpu)
915 node = next_node(cpu_to_node(cpu), node_online_map);
916 if (node == MAX_NUMNODES)
917 node = first_node(node_online_map);
919 per_cpu(reap_node, cpu) = node;
922 static void next_reap_node(void)
924 int node = __get_cpu_var(reap_node);
926 node = next_node(node, node_online_map);
927 if (unlikely(node >= MAX_NUMNODES))
928 node = first_node(node_online_map);
929 __get_cpu_var(reap_node) = node;
933 #define init_reap_node(cpu) do { } while (0)
934 #define next_reap_node(void) do { } while (0)
938 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
939 * via the workqueue/eventd.
940 * Add the CPU number into the expiration time to minimize the possibility of
941 * the CPUs getting into lockstep and contending for the global cache chain
944 static void __cpuinit start_cpu_timer(int cpu)
946 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
949 * When this gets called from do_initcalls via cpucache_init(),
950 * init_workqueues() has already run, so keventd will be setup
953 if (keventd_up() && reap_work->work.func == NULL) {
955 INIT_DELAYED_WORK(reap_work, cache_reap);
956 schedule_delayed_work_on(cpu, reap_work,
957 __round_jiffies_relative(HZ, cpu));
961 static struct array_cache *alloc_arraycache(int node, int entries,
964 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
965 struct array_cache *nc = NULL;
967 nc = kmalloc_node(memsize, GFP_KERNEL, node);
969 * The array_cache structures contain pointers to free object.
970 * However, when such objects are allocated or transfered to another
971 * cache the pointers are not cleared and they could be counted as
972 * valid references during a kmemleak scan. Therefore, kmemleak must
973 * not scan such objects.
975 kmemleak_no_scan(nc);
979 nc->batchcount = batchcount;
981 spin_lock_init(&nc->lock);
987 * Transfer objects in one arraycache to another.
988 * Locking must be handled by the caller.
990 * Return the number of entries transferred.
992 static int transfer_objects(struct array_cache *to,
993 struct array_cache *from, unsigned int max)
995 /* Figure out how many entries to transfer */
996 int nr = min(min(from->avail, max), to->limit - to->avail);
1001 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1002 sizeof(void *) *nr);
1012 #define drain_alien_cache(cachep, alien) do { } while (0)
1013 #define reap_alien(cachep, l3) do { } while (0)
1015 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1017 return (struct array_cache **)BAD_ALIEN_MAGIC;
1020 static inline void free_alien_cache(struct array_cache **ac_ptr)
1024 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1029 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1035 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1036 gfp_t flags, int nodeid)
1041 #else /* CONFIG_NUMA */
1043 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1044 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1046 static struct array_cache **alloc_alien_cache(int node, int limit)
1048 struct array_cache **ac_ptr;
1049 int memsize = sizeof(void *) * nr_node_ids;
1054 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1057 if (i == node || !node_online(i)) {
1061 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1063 for (i--; i >= 0; i--)
1073 static void free_alien_cache(struct array_cache **ac_ptr)
1084 static void __drain_alien_cache(struct kmem_cache *cachep,
1085 struct array_cache *ac, int node)
1087 struct kmem_list3 *rl3 = cachep->nodelists[node];
1090 spin_lock(&rl3->list_lock);
1092 * Stuff objects into the remote nodes shared array first.
1093 * That way we could avoid the overhead of putting the objects
1094 * into the free lists and getting them back later.
1097 transfer_objects(rl3->shared, ac, ac->limit);
1099 free_block(cachep, ac->entry, ac->avail, node);
1101 spin_unlock(&rl3->list_lock);
1106 * Called from cache_reap() to regularly drain alien caches round robin.
1108 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1110 int node = __get_cpu_var(reap_node);
1113 struct array_cache *ac = l3->alien[node];
1115 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1116 __drain_alien_cache(cachep, ac, node);
1117 spin_unlock_irq(&ac->lock);
1122 static void drain_alien_cache(struct kmem_cache *cachep,
1123 struct array_cache **alien)
1126 struct array_cache *ac;
1127 unsigned long flags;
1129 for_each_online_node(i) {
1132 spin_lock_irqsave(&ac->lock, flags);
1133 __drain_alien_cache(cachep, ac, i);
1134 spin_unlock_irqrestore(&ac->lock, flags);
1139 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1141 struct slab *slabp = virt_to_slab(objp);
1142 int nodeid = slabp->nodeid;
1143 struct kmem_list3 *l3;
1144 struct array_cache *alien = NULL;
1147 node = numa_node_id();
1150 * Make sure we are not freeing a object from another node to the array
1151 * cache on this cpu.
1153 if (likely(slabp->nodeid == node))
1156 l3 = cachep->nodelists[node];
1157 STATS_INC_NODEFREES(cachep);
1158 if (l3->alien && l3->alien[nodeid]) {
1159 alien = l3->alien[nodeid];
1160 spin_lock(&alien->lock);
1161 if (unlikely(alien->avail == alien->limit)) {
1162 STATS_INC_ACOVERFLOW(cachep);
1163 __drain_alien_cache(cachep, alien, nodeid);
1165 alien->entry[alien->avail++] = objp;
1166 spin_unlock(&alien->lock);
1168 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1169 free_block(cachep, &objp, 1, nodeid);
1170 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1176 static void __cpuinit cpuup_canceled(long cpu)
1178 struct kmem_cache *cachep;
1179 struct kmem_list3 *l3 = NULL;
1180 int node = cpu_to_node(cpu);
1181 const struct cpumask *mask = cpumask_of_node(node);
1183 list_for_each_entry(cachep, &cache_chain, next) {
1184 struct array_cache *nc;
1185 struct array_cache *shared;
1186 struct array_cache **alien;
1188 /* cpu is dead; no one can alloc from it. */
1189 nc = cachep->array[cpu];
1190 cachep->array[cpu] = NULL;
1191 l3 = cachep->nodelists[node];
1194 goto free_array_cache;
1196 spin_lock_irq(&l3->list_lock);
1198 /* Free limit for this kmem_list3 */
1199 l3->free_limit -= cachep->batchcount;
1201 free_block(cachep, nc->entry, nc->avail, node);
1203 if (!cpus_empty(*mask)) {
1204 spin_unlock_irq(&l3->list_lock);
1205 goto free_array_cache;
1208 shared = l3->shared;
1210 free_block(cachep, shared->entry,
1211 shared->avail, node);
1218 spin_unlock_irq(&l3->list_lock);
1222 drain_alien_cache(cachep, alien);
1223 free_alien_cache(alien);
1229 * In the previous loop, all the objects were freed to
1230 * the respective cache's slabs, now we can go ahead and
1231 * shrink each nodelist to its limit.
1233 list_for_each_entry(cachep, &cache_chain, next) {
1234 l3 = cachep->nodelists[node];
1237 drain_freelist(cachep, l3, l3->free_objects);
1241 static int __cpuinit cpuup_prepare(long cpu)
1243 struct kmem_cache *cachep;
1244 struct kmem_list3 *l3 = NULL;
1245 int node = cpu_to_node(cpu);
1246 const int memsize = sizeof(struct kmem_list3);
1249 * We need to do this right in the beginning since
1250 * alloc_arraycache's are going to use this list.
1251 * kmalloc_node allows us to add the slab to the right
1252 * kmem_list3 and not this cpu's kmem_list3
1255 list_for_each_entry(cachep, &cache_chain, next) {
1257 * Set up the size64 kmemlist for cpu before we can
1258 * begin anything. Make sure some other cpu on this
1259 * node has not already allocated this
1261 if (!cachep->nodelists[node]) {
1262 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1265 kmem_list3_init(l3);
1266 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1267 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1270 * The l3s don't come and go as CPUs come and
1271 * go. cache_chain_mutex is sufficient
1274 cachep->nodelists[node] = l3;
1277 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1278 cachep->nodelists[node]->free_limit =
1279 (1 + nr_cpus_node(node)) *
1280 cachep->batchcount + cachep->num;
1281 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1285 * Now we can go ahead with allocating the shared arrays and
1288 list_for_each_entry(cachep, &cache_chain, next) {
1289 struct array_cache *nc;
1290 struct array_cache *shared = NULL;
1291 struct array_cache **alien = NULL;
1293 nc = alloc_arraycache(node, cachep->limit,
1294 cachep->batchcount);
1297 if (cachep->shared) {
1298 shared = alloc_arraycache(node,
1299 cachep->shared * cachep->batchcount,
1306 if (use_alien_caches) {
1307 alien = alloc_alien_cache(node, cachep->limit);
1314 cachep->array[cpu] = nc;
1315 l3 = cachep->nodelists[node];
1318 spin_lock_irq(&l3->list_lock);
1321 * We are serialised from CPU_DEAD or
1322 * CPU_UP_CANCELLED by the cpucontrol lock
1324 l3->shared = shared;
1333 spin_unlock_irq(&l3->list_lock);
1335 free_alien_cache(alien);
1339 cpuup_canceled(cpu);
1343 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1344 unsigned long action, void *hcpu)
1346 long cpu = (long)hcpu;
1350 case CPU_UP_PREPARE:
1351 case CPU_UP_PREPARE_FROZEN:
1352 mutex_lock(&cache_chain_mutex);
1353 err = cpuup_prepare(cpu);
1354 mutex_unlock(&cache_chain_mutex);
1357 case CPU_ONLINE_FROZEN:
1358 start_cpu_timer(cpu);
1360 #ifdef CONFIG_HOTPLUG_CPU
1361 case CPU_DOWN_PREPARE:
1362 case CPU_DOWN_PREPARE_FROZEN:
1364 * Shutdown cache reaper. Note that the cache_chain_mutex is
1365 * held so that if cache_reap() is invoked it cannot do
1366 * anything expensive but will only modify reap_work
1367 * and reschedule the timer.
1369 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1370 /* Now the cache_reaper is guaranteed to be not running. */
1371 per_cpu(reap_work, cpu).work.func = NULL;
1373 case CPU_DOWN_FAILED:
1374 case CPU_DOWN_FAILED_FROZEN:
1375 start_cpu_timer(cpu);
1378 case CPU_DEAD_FROZEN:
1380 * Even if all the cpus of a node are down, we don't free the
1381 * kmem_list3 of any cache. This to avoid a race between
1382 * cpu_down, and a kmalloc allocation from another cpu for
1383 * memory from the node of the cpu going down. The list3
1384 * structure is usually allocated from kmem_cache_create() and
1385 * gets destroyed at kmem_cache_destroy().
1389 case CPU_UP_CANCELED:
1390 case CPU_UP_CANCELED_FROZEN:
1391 mutex_lock(&cache_chain_mutex);
1392 cpuup_canceled(cpu);
1393 mutex_unlock(&cache_chain_mutex);
1396 return err ? NOTIFY_BAD : NOTIFY_OK;
1399 static struct notifier_block __cpuinitdata cpucache_notifier = {
1400 &cpuup_callback, NULL, 0
1404 * swap the static kmem_list3 with kmalloced memory
1406 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1409 struct kmem_list3 *ptr;
1411 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1414 local_irq_disable();
1415 memcpy(ptr, list, sizeof(struct kmem_list3));
1417 * Do not assume that spinlocks can be initialized via memcpy:
1419 spin_lock_init(&ptr->list_lock);
1421 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1422 cachep->nodelists[nodeid] = ptr;
1427 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1428 * size of kmem_list3.
1430 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1434 for_each_online_node(node) {
1435 cachep->nodelists[node] = &initkmem_list3[index + node];
1436 cachep->nodelists[node]->next_reap = jiffies +
1438 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1443 * Initialisation. Called after the page allocator have been initialised and
1444 * before smp_init().
1446 void __init kmem_cache_init(void)
1449 struct cache_sizes *sizes;
1450 struct cache_names *names;
1455 if (num_possible_nodes() == 1) {
1456 use_alien_caches = 0;
1460 for (i = 0; i < NUM_INIT_LISTS; i++) {
1461 kmem_list3_init(&initkmem_list3[i]);
1462 if (i < MAX_NUMNODES)
1463 cache_cache.nodelists[i] = NULL;
1465 set_up_list3s(&cache_cache, CACHE_CACHE);
1468 * Fragmentation resistance on low memory - only use bigger
1469 * page orders on machines with more than 32MB of memory.
1471 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1472 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1474 /* Bootstrap is tricky, because several objects are allocated
1475 * from caches that do not exist yet:
1476 * 1) initialize the cache_cache cache: it contains the struct
1477 * kmem_cache structures of all caches, except cache_cache itself:
1478 * cache_cache is statically allocated.
1479 * Initially an __init data area is used for the head array and the
1480 * kmem_list3 structures, it's replaced with a kmalloc allocated
1481 * array at the end of the bootstrap.
1482 * 2) Create the first kmalloc cache.
1483 * The struct kmem_cache for the new cache is allocated normally.
1484 * An __init data area is used for the head array.
1485 * 3) Create the remaining kmalloc caches, with minimally sized
1487 * 4) Replace the __init data head arrays for cache_cache and the first
1488 * kmalloc cache with kmalloc allocated arrays.
1489 * 5) Replace the __init data for kmem_list3 for cache_cache and
1490 * the other cache's with kmalloc allocated memory.
1491 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1494 node = numa_node_id();
1496 /* 1) create the cache_cache */
1497 INIT_LIST_HEAD(&cache_chain);
1498 list_add(&cache_cache.next, &cache_chain);
1499 cache_cache.colour_off = cache_line_size();
1500 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1501 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1504 * struct kmem_cache size depends on nr_node_ids, which
1505 * can be less than MAX_NUMNODES.
1507 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1508 nr_node_ids * sizeof(struct kmem_list3 *);
1510 cache_cache.obj_size = cache_cache.buffer_size;
1512 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1514 cache_cache.reciprocal_buffer_size =
1515 reciprocal_value(cache_cache.buffer_size);
1517 for (order = 0; order < MAX_ORDER; order++) {
1518 cache_estimate(order, cache_cache.buffer_size,
1519 cache_line_size(), 0, &left_over, &cache_cache.num);
1520 if (cache_cache.num)
1523 BUG_ON(!cache_cache.num);
1524 cache_cache.gfporder = order;
1525 cache_cache.colour = left_over / cache_cache.colour_off;
1526 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1527 sizeof(struct slab), cache_line_size());
1529 /* 2+3) create the kmalloc caches */
1530 sizes = malloc_sizes;
1531 names = cache_names;
1534 * Initialize the caches that provide memory for the array cache and the
1535 * kmem_list3 structures first. Without this, further allocations will
1539 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1540 sizes[INDEX_AC].cs_size,
1541 ARCH_KMALLOC_MINALIGN,
1542 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1545 if (INDEX_AC != INDEX_L3) {
1546 sizes[INDEX_L3].cs_cachep =
1547 kmem_cache_create(names[INDEX_L3].name,
1548 sizes[INDEX_L3].cs_size,
1549 ARCH_KMALLOC_MINALIGN,
1550 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1554 slab_early_init = 0;
1556 while (sizes->cs_size != ULONG_MAX) {
1558 * For performance, all the general caches are L1 aligned.
1559 * This should be particularly beneficial on SMP boxes, as it
1560 * eliminates "false sharing".
1561 * Note for systems short on memory removing the alignment will
1562 * allow tighter packing of the smaller caches.
1564 if (!sizes->cs_cachep) {
1565 sizes->cs_cachep = kmem_cache_create(names->name,
1567 ARCH_KMALLOC_MINALIGN,
1568 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1571 #ifdef CONFIG_ZONE_DMA
1572 sizes->cs_dmacachep = kmem_cache_create(
1575 ARCH_KMALLOC_MINALIGN,
1576 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1583 /* 4) Replace the bootstrap head arrays */
1585 struct array_cache *ptr;
1587 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1589 local_irq_disable();
1590 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1591 memcpy(ptr, cpu_cache_get(&cache_cache),
1592 sizeof(struct arraycache_init));
1594 * Do not assume that spinlocks can be initialized via memcpy:
1596 spin_lock_init(&ptr->lock);
1598 cache_cache.array[smp_processor_id()] = ptr;
1601 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1603 local_irq_disable();
1604 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1605 != &initarray_generic.cache);
1606 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1607 sizeof(struct arraycache_init));
1609 * Do not assume that spinlocks can be initialized via memcpy:
1611 spin_lock_init(&ptr->lock);
1613 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1617 /* 5) Replace the bootstrap kmem_list3's */
1621 for_each_online_node(nid) {
1622 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1624 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1625 &initkmem_list3[SIZE_AC + nid], nid);
1627 if (INDEX_AC != INDEX_L3) {
1628 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1629 &initkmem_list3[SIZE_L3 + nid], nid);
1634 /* 6) resize the head arrays to their final sizes */
1636 struct kmem_cache *cachep;
1637 mutex_lock(&cache_chain_mutex);
1638 list_for_each_entry(cachep, &cache_chain, next)
1639 if (enable_cpucache(cachep))
1641 mutex_unlock(&cache_chain_mutex);
1644 /* Annotate slab for lockdep -- annotate the malloc caches */
1649 g_cpucache_up = FULL;
1652 * Register a cpu startup notifier callback that initializes
1653 * cpu_cache_get for all new cpus
1655 register_cpu_notifier(&cpucache_notifier);
1658 * The reap timers are started later, with a module init call: That part
1659 * of the kernel is not yet operational.
1663 static int __init cpucache_init(void)
1668 * Register the timers that return unneeded pages to the page allocator
1670 for_each_online_cpu(cpu)
1671 start_cpu_timer(cpu);
1674 __initcall(cpucache_init);
1677 * Interface to system's page allocator. No need to hold the cache-lock.
1679 * If we requested dmaable memory, we will get it. Even if we
1680 * did not request dmaable memory, we might get it, but that
1681 * would be relatively rare and ignorable.
1683 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1691 * Nommu uses slab's for process anonymous memory allocations, and thus
1692 * requires __GFP_COMP to properly refcount higher order allocations
1694 flags |= __GFP_COMP;
1697 flags |= cachep->gfpflags;
1698 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1699 flags |= __GFP_RECLAIMABLE;
1701 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1705 nr_pages = (1 << cachep->gfporder);
1706 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1707 add_zone_page_state(page_zone(page),
1708 NR_SLAB_RECLAIMABLE, nr_pages);
1710 add_zone_page_state(page_zone(page),
1711 NR_SLAB_UNRECLAIMABLE, nr_pages);
1712 for (i = 0; i < nr_pages; i++)
1713 __SetPageSlab(page + i);
1714 return page_address(page);
1718 * Interface to system's page release.
1720 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1722 unsigned long i = (1 << cachep->gfporder);
1723 struct page *page = virt_to_page(addr);
1724 const unsigned long nr_freed = i;
1726 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1727 sub_zone_page_state(page_zone(page),
1728 NR_SLAB_RECLAIMABLE, nr_freed);
1730 sub_zone_page_state(page_zone(page),
1731 NR_SLAB_UNRECLAIMABLE, nr_freed);
1733 BUG_ON(!PageSlab(page));
1734 __ClearPageSlab(page);
1737 if (current->reclaim_state)
1738 current->reclaim_state->reclaimed_slab += nr_freed;
1739 free_pages((unsigned long)addr, cachep->gfporder);
1742 static void kmem_rcu_free(struct rcu_head *head)
1744 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1745 struct kmem_cache *cachep = slab_rcu->cachep;
1747 kmem_freepages(cachep, slab_rcu->addr);
1748 if (OFF_SLAB(cachep))
1749 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1754 #ifdef CONFIG_DEBUG_PAGEALLOC
1755 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1756 unsigned long caller)
1758 int size = obj_size(cachep);
1760 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1762 if (size < 5 * sizeof(unsigned long))
1765 *addr++ = 0x12345678;
1767 *addr++ = smp_processor_id();
1768 size -= 3 * sizeof(unsigned long);
1770 unsigned long *sptr = &caller;
1771 unsigned long svalue;
1773 while (!kstack_end(sptr)) {
1775 if (kernel_text_address(svalue)) {
1777 size -= sizeof(unsigned long);
1778 if (size <= sizeof(unsigned long))
1784 *addr++ = 0x87654321;
1788 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1790 int size = obj_size(cachep);
1791 addr = &((char *)addr)[obj_offset(cachep)];
1793 memset(addr, val, size);
1794 *(unsigned char *)(addr + size - 1) = POISON_END;
1797 static void dump_line(char *data, int offset, int limit)
1800 unsigned char error = 0;
1803 printk(KERN_ERR "%03x:", offset);
1804 for (i = 0; i < limit; i++) {
1805 if (data[offset + i] != POISON_FREE) {
1806 error = data[offset + i];
1809 printk(" %02x", (unsigned char)data[offset + i]);
1813 if (bad_count == 1) {
1814 error ^= POISON_FREE;
1815 if (!(error & (error - 1))) {
1816 printk(KERN_ERR "Single bit error detected. Probably "
1819 printk(KERN_ERR "Run memtest86+ or a similar memory "
1822 printk(KERN_ERR "Run a memory test tool.\n");
1831 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1836 if (cachep->flags & SLAB_RED_ZONE) {
1837 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1838 *dbg_redzone1(cachep, objp),
1839 *dbg_redzone2(cachep, objp));
1842 if (cachep->flags & SLAB_STORE_USER) {
1843 printk(KERN_ERR "Last user: [<%p>]",
1844 *dbg_userword(cachep, objp));
1845 print_symbol("(%s)",
1846 (unsigned long)*dbg_userword(cachep, objp));
1849 realobj = (char *)objp + obj_offset(cachep);
1850 size = obj_size(cachep);
1851 for (i = 0; i < size && lines; i += 16, lines--) {
1854 if (i + limit > size)
1856 dump_line(realobj, i, limit);
1860 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1866 realobj = (char *)objp + obj_offset(cachep);
1867 size = obj_size(cachep);
1869 for (i = 0; i < size; i++) {
1870 char exp = POISON_FREE;
1873 if (realobj[i] != exp) {
1879 "Slab corruption: %s start=%p, len=%d\n",
1880 cachep->name, realobj, size);
1881 print_objinfo(cachep, objp, 0);
1883 /* Hexdump the affected line */
1886 if (i + limit > size)
1888 dump_line(realobj, i, limit);
1891 /* Limit to 5 lines */
1897 /* Print some data about the neighboring objects, if they
1900 struct slab *slabp = virt_to_slab(objp);
1903 objnr = obj_to_index(cachep, slabp, objp);
1905 objp = index_to_obj(cachep, slabp, objnr - 1);
1906 realobj = (char *)objp + obj_offset(cachep);
1907 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1909 print_objinfo(cachep, objp, 2);
1911 if (objnr + 1 < cachep->num) {
1912 objp = index_to_obj(cachep, slabp, objnr + 1);
1913 realobj = (char *)objp + obj_offset(cachep);
1914 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1916 print_objinfo(cachep, objp, 2);
1923 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1926 for (i = 0; i < cachep->num; i++) {
1927 void *objp = index_to_obj(cachep, slabp, i);
1929 if (cachep->flags & SLAB_POISON) {
1930 #ifdef CONFIG_DEBUG_PAGEALLOC
1931 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1933 kernel_map_pages(virt_to_page(objp),
1934 cachep->buffer_size / PAGE_SIZE, 1);
1936 check_poison_obj(cachep, objp);
1938 check_poison_obj(cachep, objp);
1941 if (cachep->flags & SLAB_RED_ZONE) {
1942 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1943 slab_error(cachep, "start of a freed object "
1945 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1946 slab_error(cachep, "end of a freed object "
1952 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1958 * slab_destroy - destroy and release all objects in a slab
1959 * @cachep: cache pointer being destroyed
1960 * @slabp: slab pointer being destroyed
1962 * Destroy all the objs in a slab, and release the mem back to the system.
1963 * Before calling the slab must have been unlinked from the cache. The
1964 * cache-lock is not held/needed.
1966 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1968 void *addr = slabp->s_mem - slabp->colouroff;
1970 slab_destroy_debugcheck(cachep, slabp);
1971 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1972 struct slab_rcu *slab_rcu;
1974 slab_rcu = (struct slab_rcu *)slabp;
1975 slab_rcu->cachep = cachep;
1976 slab_rcu->addr = addr;
1977 call_rcu(&slab_rcu->head, kmem_rcu_free);
1979 kmem_freepages(cachep, addr);
1980 if (OFF_SLAB(cachep))
1981 kmem_cache_free(cachep->slabp_cache, slabp);
1985 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1988 struct kmem_list3 *l3;
1990 for_each_online_cpu(i)
1991 kfree(cachep->array[i]);
1993 /* NUMA: free the list3 structures */
1994 for_each_online_node(i) {
1995 l3 = cachep->nodelists[i];
1998 free_alien_cache(l3->alien);
2002 kmem_cache_free(&cache_cache, cachep);
2007 * calculate_slab_order - calculate size (page order) of slabs
2008 * @cachep: pointer to the cache that is being created
2009 * @size: size of objects to be created in this cache.
2010 * @align: required alignment for the objects.
2011 * @flags: slab allocation flags
2013 * Also calculates the number of objects per slab.
2015 * This could be made much more intelligent. For now, try to avoid using
2016 * high order pages for slabs. When the gfp() functions are more friendly
2017 * towards high-order requests, this should be changed.
2019 static size_t calculate_slab_order(struct kmem_cache *cachep,
2020 size_t size, size_t align, unsigned long flags)
2022 unsigned long offslab_limit;
2023 size_t left_over = 0;
2026 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2030 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2034 if (flags & CFLGS_OFF_SLAB) {
2036 * Max number of objs-per-slab for caches which
2037 * use off-slab slabs. Needed to avoid a possible
2038 * looping condition in cache_grow().
2040 offslab_limit = size - sizeof(struct slab);
2041 offslab_limit /= sizeof(kmem_bufctl_t);
2043 if (num > offslab_limit)
2047 /* Found something acceptable - save it away */
2049 cachep->gfporder = gfporder;
2050 left_over = remainder;
2053 * A VFS-reclaimable slab tends to have most allocations
2054 * as GFP_NOFS and we really don't want to have to be allocating
2055 * higher-order pages when we are unable to shrink dcache.
2057 if (flags & SLAB_RECLAIM_ACCOUNT)
2061 * Large number of objects is good, but very large slabs are
2062 * currently bad for the gfp()s.
2064 if (gfporder >= slab_break_gfp_order)
2068 * Acceptable internal fragmentation?
2070 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2076 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
2078 if (g_cpucache_up == FULL)
2079 return enable_cpucache(cachep);
2081 if (g_cpucache_up == NONE) {
2083 * Note: the first kmem_cache_create must create the cache
2084 * that's used by kmalloc(24), otherwise the creation of
2085 * further caches will BUG().
2087 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2090 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2091 * the first cache, then we need to set up all its list3s,
2092 * otherwise the creation of further caches will BUG().
2094 set_up_list3s(cachep, SIZE_AC);
2095 if (INDEX_AC == INDEX_L3)
2096 g_cpucache_up = PARTIAL_L3;
2098 g_cpucache_up = PARTIAL_AC;
2100 cachep->array[smp_processor_id()] =
2101 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2103 if (g_cpucache_up == PARTIAL_AC) {
2104 set_up_list3s(cachep, SIZE_L3);
2105 g_cpucache_up = PARTIAL_L3;
2108 for_each_online_node(node) {
2109 cachep->nodelists[node] =
2110 kmalloc_node(sizeof(struct kmem_list3),
2112 BUG_ON(!cachep->nodelists[node]);
2113 kmem_list3_init(cachep->nodelists[node]);
2117 cachep->nodelists[numa_node_id()]->next_reap =
2118 jiffies + REAPTIMEOUT_LIST3 +
2119 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2121 cpu_cache_get(cachep)->avail = 0;
2122 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2123 cpu_cache_get(cachep)->batchcount = 1;
2124 cpu_cache_get(cachep)->touched = 0;
2125 cachep->batchcount = 1;
2126 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2131 * kmem_cache_create - Create a cache.
2132 * @name: A string which is used in /proc/slabinfo to identify this cache.
2133 * @size: The size of objects to be created in this cache.
2134 * @align: The required alignment for the objects.
2135 * @flags: SLAB flags
2136 * @ctor: A constructor for the objects.
2138 * Returns a ptr to the cache on success, NULL on failure.
2139 * Cannot be called within a int, but can be interrupted.
2140 * The @ctor is run when new pages are allocated by the cache.
2142 * @name must be valid until the cache is destroyed. This implies that
2143 * the module calling this has to destroy the cache before getting unloaded.
2144 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2145 * therefore applications must manage it themselves.
2149 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2150 * to catch references to uninitialised memory.
2152 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2153 * for buffer overruns.
2155 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2156 * cacheline. This can be beneficial if you're counting cycles as closely
2160 kmem_cache_create (const char *name, size_t size, size_t align,
2161 unsigned long flags, void (*ctor)(void *))
2163 size_t left_over, slab_size, ralign;
2164 struct kmem_cache *cachep = NULL, *pc;
2167 * Sanity checks... these are all serious usage bugs.
2169 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2170 size > KMALLOC_MAX_SIZE) {
2171 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2177 * We use cache_chain_mutex to ensure a consistent view of
2178 * cpu_online_mask as well. Please see cpuup_callback
2181 mutex_lock(&cache_chain_mutex);
2183 list_for_each_entry(pc, &cache_chain, next) {
2188 * This happens when the module gets unloaded and doesn't
2189 * destroy its slab cache and no-one else reuses the vmalloc
2190 * area of the module. Print a warning.
2192 res = probe_kernel_address(pc->name, tmp);
2195 "SLAB: cache with size %d has lost its name\n",
2200 if (!strcmp(pc->name, name)) {
2202 "kmem_cache_create: duplicate cache %s\n", name);
2209 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2212 * Enable redzoning and last user accounting, except for caches with
2213 * large objects, if the increased size would increase the object size
2214 * above the next power of two: caches with object sizes just above a
2215 * power of two have a significant amount of internal fragmentation.
2217 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2218 2 * sizeof(unsigned long long)))
2219 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2220 if (!(flags & SLAB_DESTROY_BY_RCU))
2221 flags |= SLAB_POISON;
2223 if (flags & SLAB_DESTROY_BY_RCU)
2224 BUG_ON(flags & SLAB_POISON);
2227 * Always checks flags, a caller might be expecting debug support which
2230 BUG_ON(flags & ~CREATE_MASK);
2233 * Check that size is in terms of words. This is needed to avoid
2234 * unaligned accesses for some archs when redzoning is used, and makes
2235 * sure any on-slab bufctl's are also correctly aligned.
2237 if (size & (BYTES_PER_WORD - 1)) {
2238 size += (BYTES_PER_WORD - 1);
2239 size &= ~(BYTES_PER_WORD - 1);
2242 /* calculate the final buffer alignment: */
2244 /* 1) arch recommendation: can be overridden for debug */
2245 if (flags & SLAB_HWCACHE_ALIGN) {
2247 * Default alignment: as specified by the arch code. Except if
2248 * an object is really small, then squeeze multiple objects into
2251 ralign = cache_line_size();
2252 while (size <= ralign / 2)
2255 ralign = BYTES_PER_WORD;
2259 * Redzoning and user store require word alignment or possibly larger.
2260 * Note this will be overridden by architecture or caller mandated
2261 * alignment if either is greater than BYTES_PER_WORD.
2263 if (flags & SLAB_STORE_USER)
2264 ralign = BYTES_PER_WORD;
2266 if (flags & SLAB_RED_ZONE) {
2267 ralign = REDZONE_ALIGN;
2268 /* If redzoning, ensure that the second redzone is suitably
2269 * aligned, by adjusting the object size accordingly. */
2270 size += REDZONE_ALIGN - 1;
2271 size &= ~(REDZONE_ALIGN - 1);
2274 /* 2) arch mandated alignment */
2275 if (ralign < ARCH_SLAB_MINALIGN) {
2276 ralign = ARCH_SLAB_MINALIGN;
2278 /* 3) caller mandated alignment */
2279 if (ralign < align) {
2282 /* disable debug if necessary */
2283 if (ralign > __alignof__(unsigned long long))
2284 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2290 /* Get cache's description obj. */
2291 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2296 cachep->obj_size = size;
2299 * Both debugging options require word-alignment which is calculated
2302 if (flags & SLAB_RED_ZONE) {
2303 /* add space for red zone words */
2304 cachep->obj_offset += sizeof(unsigned long long);
2305 size += 2 * sizeof(unsigned long long);
2307 if (flags & SLAB_STORE_USER) {
2308 /* user store requires one word storage behind the end of
2309 * the real object. But if the second red zone needs to be
2310 * aligned to 64 bits, we must allow that much space.
2312 if (flags & SLAB_RED_ZONE)
2313 size += REDZONE_ALIGN;
2315 size += BYTES_PER_WORD;
2317 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2318 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2319 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2320 cachep->obj_offset += PAGE_SIZE - size;
2327 * Determine if the slab management is 'on' or 'off' slab.
2328 * (bootstrapping cannot cope with offslab caches so don't do
2331 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2333 * Size is large, assume best to place the slab management obj
2334 * off-slab (should allow better packing of objs).
2336 flags |= CFLGS_OFF_SLAB;
2338 size = ALIGN(size, align);
2340 left_over = calculate_slab_order(cachep, size, align, flags);
2344 "kmem_cache_create: couldn't create cache %s.\n", name);
2345 kmem_cache_free(&cache_cache, cachep);
2349 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2350 + sizeof(struct slab), align);
2353 * If the slab has been placed off-slab, and we have enough space then
2354 * move it on-slab. This is at the expense of any extra colouring.
2356 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2357 flags &= ~CFLGS_OFF_SLAB;
2358 left_over -= slab_size;
2361 if (flags & CFLGS_OFF_SLAB) {
2362 /* really off slab. No need for manual alignment */
2364 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2367 cachep->colour_off = cache_line_size();
2368 /* Offset must be a multiple of the alignment. */
2369 if (cachep->colour_off < align)
2370 cachep->colour_off = align;
2371 cachep->colour = left_over / cachep->colour_off;
2372 cachep->slab_size = slab_size;
2373 cachep->flags = flags;
2374 cachep->gfpflags = 0;
2375 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2376 cachep->gfpflags |= GFP_DMA;
2377 cachep->buffer_size = size;
2378 cachep->reciprocal_buffer_size = reciprocal_value(size);
2380 if (flags & CFLGS_OFF_SLAB) {
2381 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2383 * This is a possibility for one of the malloc_sizes caches.
2384 * But since we go off slab only for object size greater than
2385 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2386 * this should not happen at all.
2387 * But leave a BUG_ON for some lucky dude.
2389 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2391 cachep->ctor = ctor;
2392 cachep->name = name;
2394 if (setup_cpu_cache(cachep)) {
2395 __kmem_cache_destroy(cachep);
2400 /* cache setup completed, link it into the list */
2401 list_add(&cachep->next, &cache_chain);
2403 if (!cachep && (flags & SLAB_PANIC))
2404 panic("kmem_cache_create(): failed to create slab `%s'\n",
2406 mutex_unlock(&cache_chain_mutex);
2410 EXPORT_SYMBOL(kmem_cache_create);
2413 static void check_irq_off(void)
2415 BUG_ON(!irqs_disabled());
2418 static void check_irq_on(void)
2420 BUG_ON(irqs_disabled());
2423 static void check_spinlock_acquired(struct kmem_cache *cachep)
2427 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2431 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2435 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2440 #define check_irq_off() do { } while(0)
2441 #define check_irq_on() do { } while(0)
2442 #define check_spinlock_acquired(x) do { } while(0)
2443 #define check_spinlock_acquired_node(x, y) do { } while(0)
2446 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2447 struct array_cache *ac,
2448 int force, int node);
2450 static void do_drain(void *arg)
2452 struct kmem_cache *cachep = arg;
2453 struct array_cache *ac;
2454 int node = numa_node_id();
2457 ac = cpu_cache_get(cachep);
2458 spin_lock(&cachep->nodelists[node]->list_lock);
2459 free_block(cachep, ac->entry, ac->avail, node);
2460 spin_unlock(&cachep->nodelists[node]->list_lock);
2464 static void drain_cpu_caches(struct kmem_cache *cachep)
2466 struct kmem_list3 *l3;
2469 on_each_cpu(do_drain, cachep, 1);
2471 for_each_online_node(node) {
2472 l3 = cachep->nodelists[node];
2473 if (l3 && l3->alien)
2474 drain_alien_cache(cachep, l3->alien);
2477 for_each_online_node(node) {
2478 l3 = cachep->nodelists[node];
2480 drain_array(cachep, l3, l3->shared, 1, node);
2485 * Remove slabs from the list of free slabs.
2486 * Specify the number of slabs to drain in tofree.
2488 * Returns the actual number of slabs released.
2490 static int drain_freelist(struct kmem_cache *cache,
2491 struct kmem_list3 *l3, int tofree)
2493 struct list_head *p;
2498 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2500 spin_lock_irq(&l3->list_lock);
2501 p = l3->slabs_free.prev;
2502 if (p == &l3->slabs_free) {
2503 spin_unlock_irq(&l3->list_lock);
2507 slabp = list_entry(p, struct slab, list);
2509 BUG_ON(slabp->inuse);
2511 list_del(&slabp->list);
2513 * Safe to drop the lock. The slab is no longer linked
2516 l3->free_objects -= cache->num;
2517 spin_unlock_irq(&l3->list_lock);
2518 slab_destroy(cache, slabp);
2525 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2526 static int __cache_shrink(struct kmem_cache *cachep)
2529 struct kmem_list3 *l3;
2531 drain_cpu_caches(cachep);
2534 for_each_online_node(i) {
2535 l3 = cachep->nodelists[i];
2539 drain_freelist(cachep, l3, l3->free_objects);
2541 ret += !list_empty(&l3->slabs_full) ||
2542 !list_empty(&l3->slabs_partial);
2544 return (ret ? 1 : 0);
2548 * kmem_cache_shrink - Shrink a cache.
2549 * @cachep: The cache to shrink.
2551 * Releases as many slabs as possible for a cache.
2552 * To help debugging, a zero exit status indicates all slabs were released.
2554 int kmem_cache_shrink(struct kmem_cache *cachep)
2557 BUG_ON(!cachep || in_interrupt());
2560 mutex_lock(&cache_chain_mutex);
2561 ret = __cache_shrink(cachep);
2562 mutex_unlock(&cache_chain_mutex);
2566 EXPORT_SYMBOL(kmem_cache_shrink);
2569 * kmem_cache_destroy - delete a cache
2570 * @cachep: the cache to destroy
2572 * Remove a &struct kmem_cache object from the slab cache.
2574 * It is expected this function will be called by a module when it is
2575 * unloaded. This will remove the cache completely, and avoid a duplicate
2576 * cache being allocated each time a module is loaded and unloaded, if the
2577 * module doesn't have persistent in-kernel storage across loads and unloads.
2579 * The cache must be empty before calling this function.
2581 * The caller must guarantee that noone will allocate memory from the cache
2582 * during the kmem_cache_destroy().
2584 void kmem_cache_destroy(struct kmem_cache *cachep)
2586 BUG_ON(!cachep || in_interrupt());
2588 /* Find the cache in the chain of caches. */
2590 mutex_lock(&cache_chain_mutex);
2592 * the chain is never empty, cache_cache is never destroyed
2594 list_del(&cachep->next);
2595 if (__cache_shrink(cachep)) {
2596 slab_error(cachep, "Can't free all objects");
2597 list_add(&cachep->next, &cache_chain);
2598 mutex_unlock(&cache_chain_mutex);
2603 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2606 __kmem_cache_destroy(cachep);
2607 mutex_unlock(&cache_chain_mutex);
2610 EXPORT_SYMBOL(kmem_cache_destroy);
2613 * Get the memory for a slab management obj.
2614 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2615 * always come from malloc_sizes caches. The slab descriptor cannot
2616 * come from the same cache which is getting created because,
2617 * when we are searching for an appropriate cache for these
2618 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2619 * If we are creating a malloc_sizes cache here it would not be visible to
2620 * kmem_find_general_cachep till the initialization is complete.
2621 * Hence we cannot have slabp_cache same as the original cache.
2623 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2624 int colour_off, gfp_t local_flags,
2629 if (OFF_SLAB(cachep)) {
2630 /* Slab management obj is off-slab. */
2631 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2632 local_flags, nodeid);
2634 * If the first object in the slab is leaked (it's allocated
2635 * but no one has a reference to it), we want to make sure
2636 * kmemleak does not treat the ->s_mem pointer as a reference
2637 * to the object. Otherwise we will not report the leak.
2639 kmemleak_scan_area(slabp, offsetof(struct slab, list),
2640 sizeof(struct list_head), local_flags);
2644 slabp = objp + colour_off;
2645 colour_off += cachep->slab_size;
2648 slabp->colouroff = colour_off;
2649 slabp->s_mem = objp + colour_off;
2650 slabp->nodeid = nodeid;
2655 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2657 return (kmem_bufctl_t *) (slabp + 1);
2660 static void cache_init_objs(struct kmem_cache *cachep,
2665 for (i = 0; i < cachep->num; i++) {
2666 void *objp = index_to_obj(cachep, slabp, i);
2668 /* need to poison the objs? */
2669 if (cachep->flags & SLAB_POISON)
2670 poison_obj(cachep, objp, POISON_FREE);
2671 if (cachep->flags & SLAB_STORE_USER)
2672 *dbg_userword(cachep, objp) = NULL;
2674 if (cachep->flags & SLAB_RED_ZONE) {
2675 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2676 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2679 * Constructors are not allowed to allocate memory from the same
2680 * cache which they are a constructor for. Otherwise, deadlock.
2681 * They must also be threaded.
2683 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2684 cachep->ctor(objp + obj_offset(cachep));
2686 if (cachep->flags & SLAB_RED_ZONE) {
2687 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2688 slab_error(cachep, "constructor overwrote the"
2689 " end of an object");
2690 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2691 slab_error(cachep, "constructor overwrote the"
2692 " start of an object");
2694 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2695 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2696 kernel_map_pages(virt_to_page(objp),
2697 cachep->buffer_size / PAGE_SIZE, 0);
2702 slab_bufctl(slabp)[i] = i + 1;
2704 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2707 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2709 if (CONFIG_ZONE_DMA_FLAG) {
2710 if (flags & GFP_DMA)
2711 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2713 BUG_ON(cachep->gfpflags & GFP_DMA);
2717 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2720 void *objp = index_to_obj(cachep, slabp, slabp->free);
2724 next = slab_bufctl(slabp)[slabp->free];
2726 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2727 WARN_ON(slabp->nodeid != nodeid);
2734 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2735 void *objp, int nodeid)
2737 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2740 /* Verify that the slab belongs to the intended node */
2741 WARN_ON(slabp->nodeid != nodeid);
2743 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2744 printk(KERN_ERR "slab: double free detected in cache "
2745 "'%s', objp %p\n", cachep->name, objp);
2749 slab_bufctl(slabp)[objnr] = slabp->free;
2750 slabp->free = objnr;
2755 * Map pages beginning at addr to the given cache and slab. This is required
2756 * for the slab allocator to be able to lookup the cache and slab of a
2757 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2759 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2765 page = virt_to_page(addr);
2768 if (likely(!PageCompound(page)))
2769 nr_pages <<= cache->gfporder;
2772 page_set_cache(page, cache);
2773 page_set_slab(page, slab);
2775 } while (--nr_pages);
2779 * Grow (by 1) the number of slabs within a cache. This is called by
2780 * kmem_cache_alloc() when there are no active objs left in a cache.
2782 static int cache_grow(struct kmem_cache *cachep,
2783 gfp_t flags, int nodeid, void *objp)
2788 struct kmem_list3 *l3;
2791 * Be lazy and only check for valid flags here, keeping it out of the
2792 * critical path in kmem_cache_alloc().
2794 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2795 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2797 /* Take the l3 list lock to change the colour_next on this node */
2799 l3 = cachep->nodelists[nodeid];
2800 spin_lock(&l3->list_lock);
2802 /* Get colour for the slab, and cal the next value. */
2803 offset = l3->colour_next;
2805 if (l3->colour_next >= cachep->colour)
2806 l3->colour_next = 0;
2807 spin_unlock(&l3->list_lock);
2809 offset *= cachep->colour_off;
2811 if (local_flags & __GFP_WAIT)
2815 * The test for missing atomic flag is performed here, rather than
2816 * the more obvious place, simply to reduce the critical path length
2817 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2818 * will eventually be caught here (where it matters).
2820 kmem_flagcheck(cachep, flags);
2823 * Get mem for the objs. Attempt to allocate a physical page from
2827 objp = kmem_getpages(cachep, local_flags, nodeid);
2831 /* Get slab management. */
2832 slabp = alloc_slabmgmt(cachep, objp, offset,
2833 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2837 slab_map_pages(cachep, slabp, objp);
2839 cache_init_objs(cachep, slabp);
2841 if (local_flags & __GFP_WAIT)
2842 local_irq_disable();
2844 spin_lock(&l3->list_lock);
2846 /* Make slab active. */
2847 list_add_tail(&slabp->list, &(l3->slabs_free));
2848 STATS_INC_GROWN(cachep);
2849 l3->free_objects += cachep->num;
2850 spin_unlock(&l3->list_lock);
2853 kmem_freepages(cachep, objp);
2855 if (local_flags & __GFP_WAIT)
2856 local_irq_disable();
2863 * Perform extra freeing checks:
2864 * - detect bad pointers.
2865 * - POISON/RED_ZONE checking
2867 static void kfree_debugcheck(const void *objp)
2869 if (!virt_addr_valid(objp)) {
2870 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2871 (unsigned long)objp);
2876 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2878 unsigned long long redzone1, redzone2;
2880 redzone1 = *dbg_redzone1(cache, obj);
2881 redzone2 = *dbg_redzone2(cache, obj);
2886 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2889 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2890 slab_error(cache, "double free detected");
2892 slab_error(cache, "memory outside object was overwritten");
2894 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2895 obj, redzone1, redzone2);
2898 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2905 BUG_ON(virt_to_cache(objp) != cachep);
2907 objp -= obj_offset(cachep);
2908 kfree_debugcheck(objp);
2909 page = virt_to_head_page(objp);
2911 slabp = page_get_slab(page);
2913 if (cachep->flags & SLAB_RED_ZONE) {
2914 verify_redzone_free(cachep, objp);
2915 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2916 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2918 if (cachep->flags & SLAB_STORE_USER)
2919 *dbg_userword(cachep, objp) = caller;
2921 objnr = obj_to_index(cachep, slabp, objp);
2923 BUG_ON(objnr >= cachep->num);
2924 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2926 #ifdef CONFIG_DEBUG_SLAB_LEAK
2927 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2929 if (cachep->flags & SLAB_POISON) {
2930 #ifdef CONFIG_DEBUG_PAGEALLOC
2931 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2932 store_stackinfo(cachep, objp, (unsigned long)caller);
2933 kernel_map_pages(virt_to_page(objp),
2934 cachep->buffer_size / PAGE_SIZE, 0);
2936 poison_obj(cachep, objp, POISON_FREE);
2939 poison_obj(cachep, objp, POISON_FREE);
2945 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2950 /* Check slab's freelist to see if this obj is there. */
2951 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2953 if (entries > cachep->num || i >= cachep->num)
2956 if (entries != cachep->num - slabp->inuse) {
2958 printk(KERN_ERR "slab: Internal list corruption detected in "
2959 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2960 cachep->name, cachep->num, slabp, slabp->inuse);
2962 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2965 printk("\n%03x:", i);
2966 printk(" %02x", ((unsigned char *)slabp)[i]);
2973 #define kfree_debugcheck(x) do { } while(0)
2974 #define cache_free_debugcheck(x,objp,z) (objp)
2975 #define check_slabp(x,y) do { } while(0)
2978 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2981 struct kmem_list3 *l3;
2982 struct array_cache *ac;
2987 node = numa_node_id();
2988 ac = cpu_cache_get(cachep);
2989 batchcount = ac->batchcount;
2990 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2992 * If there was little recent activity on this cache, then
2993 * perform only a partial refill. Otherwise we could generate
2996 batchcount = BATCHREFILL_LIMIT;
2998 l3 = cachep->nodelists[node];
3000 BUG_ON(ac->avail > 0 || !l3);
3001 spin_lock(&l3->list_lock);
3003 /* See if we can refill from the shared array */
3004 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
3007 while (batchcount > 0) {
3008 struct list_head *entry;
3010 /* Get slab alloc is to come from. */
3011 entry = l3->slabs_partial.next;
3012 if (entry == &l3->slabs_partial) {
3013 l3->free_touched = 1;
3014 entry = l3->slabs_free.next;
3015 if (entry == &l3->slabs_free)
3019 slabp = list_entry(entry, struct slab, list);
3020 check_slabp(cachep, slabp);
3021 check_spinlock_acquired(cachep);
3024 * The slab was either on partial or free list so
3025 * there must be at least one object available for
3028 BUG_ON(slabp->inuse >= cachep->num);
3030 while (slabp->inuse < cachep->num && batchcount--) {
3031 STATS_INC_ALLOCED(cachep);
3032 STATS_INC_ACTIVE(cachep);
3033 STATS_SET_HIGH(cachep);
3035 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3038 check_slabp(cachep, slabp);
3040 /* move slabp to correct slabp list: */
3041 list_del(&slabp->list);
3042 if (slabp->free == BUFCTL_END)
3043 list_add(&slabp->list, &l3->slabs_full);
3045 list_add(&slabp->list, &l3->slabs_partial);
3049 l3->free_objects -= ac->avail;
3051 spin_unlock(&l3->list_lock);
3053 if (unlikely(!ac->avail)) {
3055 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3057 /* cache_grow can reenable interrupts, then ac could change. */
3058 ac = cpu_cache_get(cachep);
3059 if (!x && ac->avail == 0) /* no objects in sight? abort */
3062 if (!ac->avail) /* objects refilled by interrupt? */
3066 return ac->entry[--ac->avail];
3069 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3072 might_sleep_if(flags & __GFP_WAIT);
3074 kmem_flagcheck(cachep, flags);
3079 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3080 gfp_t flags, void *objp, void *caller)
3084 if (cachep->flags & SLAB_POISON) {
3085 #ifdef CONFIG_DEBUG_PAGEALLOC
3086 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3087 kernel_map_pages(virt_to_page(objp),
3088 cachep->buffer_size / PAGE_SIZE, 1);
3090 check_poison_obj(cachep, objp);
3092 check_poison_obj(cachep, objp);
3094 poison_obj(cachep, objp, POISON_INUSE);
3096 if (cachep->flags & SLAB_STORE_USER)
3097 *dbg_userword(cachep, objp) = caller;
3099 if (cachep->flags & SLAB_RED_ZONE) {
3100 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3101 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3102 slab_error(cachep, "double free, or memory outside"
3103 " object was overwritten");
3105 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3106 objp, *dbg_redzone1(cachep, objp),
3107 *dbg_redzone2(cachep, objp));
3109 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3110 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3112 #ifdef CONFIG_DEBUG_SLAB_LEAK
3117 slabp = page_get_slab(virt_to_head_page(objp));
3118 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3119 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3122 objp += obj_offset(cachep);
3123 if (cachep->ctor && cachep->flags & SLAB_POISON)
3125 #if ARCH_SLAB_MINALIGN
3126 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3127 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3128 objp, ARCH_SLAB_MINALIGN);
3134 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3137 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3139 if (cachep == &cache_cache)
3142 return should_failslab(obj_size(cachep), flags);
3145 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3148 struct array_cache *ac;
3152 ac = cpu_cache_get(cachep);
3153 if (likely(ac->avail)) {
3154 STATS_INC_ALLOCHIT(cachep);
3156 objp = ac->entry[--ac->avail];
3158 STATS_INC_ALLOCMISS(cachep);
3159 objp = cache_alloc_refill(cachep, flags);
3162 * To avoid a false negative, if an object that is in one of the
3163 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3164 * treat the array pointers as a reference to the object.
3166 kmemleak_erase(&ac->entry[ac->avail]);
3172 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3174 * If we are in_interrupt, then process context, including cpusets and
3175 * mempolicy, may not apply and should not be used for allocation policy.
3177 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3179 int nid_alloc, nid_here;
3181 if (in_interrupt() || (flags & __GFP_THISNODE))
3183 nid_alloc = nid_here = numa_node_id();
3184 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3185 nid_alloc = cpuset_mem_spread_node();
3186 else if (current->mempolicy)
3187 nid_alloc = slab_node(current->mempolicy);
3188 if (nid_alloc != nid_here)
3189 return ____cache_alloc_node(cachep, flags, nid_alloc);
3194 * Fallback function if there was no memory available and no objects on a
3195 * certain node and fall back is permitted. First we scan all the
3196 * available nodelists for available objects. If that fails then we
3197 * perform an allocation without specifying a node. This allows the page
3198 * allocator to do its reclaim / fallback magic. We then insert the
3199 * slab into the proper nodelist and then allocate from it.
3201 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3203 struct zonelist *zonelist;
3207 enum zone_type high_zoneidx = gfp_zone(flags);
3211 if (flags & __GFP_THISNODE)
3214 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3215 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3219 * Look through allowed nodes for objects available
3220 * from existing per node queues.
3222 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3223 nid = zone_to_nid(zone);
3225 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3226 cache->nodelists[nid] &&
3227 cache->nodelists[nid]->free_objects) {
3228 obj = ____cache_alloc_node(cache,
3229 flags | GFP_THISNODE, nid);
3237 * This allocation will be performed within the constraints
3238 * of the current cpuset / memory policy requirements.
3239 * We may trigger various forms of reclaim on the allowed
3240 * set and go into memory reserves if necessary.
3242 if (local_flags & __GFP_WAIT)
3244 kmem_flagcheck(cache, flags);
3245 obj = kmem_getpages(cache, local_flags, -1);
3246 if (local_flags & __GFP_WAIT)
3247 local_irq_disable();
3250 * Insert into the appropriate per node queues
3252 nid = page_to_nid(virt_to_page(obj));
3253 if (cache_grow(cache, flags, nid, obj)) {
3254 obj = ____cache_alloc_node(cache,
3255 flags | GFP_THISNODE, nid);
3258 * Another processor may allocate the
3259 * objects in the slab since we are
3260 * not holding any locks.
3264 /* cache_grow already freed obj */
3273 * A interface to enable slab creation on nodeid
3275 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3278 struct list_head *entry;
3280 struct kmem_list3 *l3;
3284 l3 = cachep->nodelists[nodeid];
3289 spin_lock(&l3->list_lock);
3290 entry = l3->slabs_partial.next;
3291 if (entry == &l3->slabs_partial) {
3292 l3->free_touched = 1;
3293 entry = l3->slabs_free.next;
3294 if (entry == &l3->slabs_free)
3298 slabp = list_entry(entry, struct slab, list);
3299 check_spinlock_acquired_node(cachep, nodeid);
3300 check_slabp(cachep, slabp);
3302 STATS_INC_NODEALLOCS(cachep);
3303 STATS_INC_ACTIVE(cachep);
3304 STATS_SET_HIGH(cachep);
3306 BUG_ON(slabp->inuse == cachep->num);
3308 obj = slab_get_obj(cachep, slabp, nodeid);
3309 check_slabp(cachep, slabp);
3311 /* move slabp to correct slabp list: */
3312 list_del(&slabp->list);
3314 if (slabp->free == BUFCTL_END)
3315 list_add(&slabp->list, &l3->slabs_full);
3317 list_add(&slabp->list, &l3->slabs_partial);
3319 spin_unlock(&l3->list_lock);
3323 spin_unlock(&l3->list_lock);
3324 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3328 return fallback_alloc(cachep, flags);
3335 * kmem_cache_alloc_node - Allocate an object on the specified node
3336 * @cachep: The cache to allocate from.
3337 * @flags: See kmalloc().
3338 * @nodeid: node number of the target node.
3339 * @caller: return address of caller, used for debug information
3341 * Identical to kmem_cache_alloc but it will allocate memory on the given
3342 * node, which can improve the performance for cpu bound structures.
3344 * Fallback to other node is possible if __GFP_THISNODE is not set.
3346 static __always_inline void *
3347 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3350 unsigned long save_flags;
3353 lockdep_trace_alloc(flags);
3355 if (slab_should_failslab(cachep, flags))
3358 cache_alloc_debugcheck_before(cachep, flags);
3359 local_irq_save(save_flags);
3361 if (unlikely(nodeid == -1))
3362 nodeid = numa_node_id();
3364 if (unlikely(!cachep->nodelists[nodeid])) {
3365 /* Node not bootstrapped yet */
3366 ptr = fallback_alloc(cachep, flags);
3370 if (nodeid == numa_node_id()) {
3372 * Use the locally cached objects if possible.
3373 * However ____cache_alloc does not allow fallback
3374 * to other nodes. It may fail while we still have
3375 * objects on other nodes available.
3377 ptr = ____cache_alloc(cachep, flags);
3381 /* ___cache_alloc_node can fall back to other nodes */
3382 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3384 local_irq_restore(save_flags);
3385 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3386 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3389 if (unlikely((flags & __GFP_ZERO) && ptr))
3390 memset(ptr, 0, obj_size(cachep));
3395 static __always_inline void *
3396 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3400 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3401 objp = alternate_node_alloc(cache, flags);
3405 objp = ____cache_alloc(cache, flags);
3408 * We may just have run out of memory on the local node.
3409 * ____cache_alloc_node() knows how to locate memory on other nodes
3412 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3419 static __always_inline void *
3420 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3422 return ____cache_alloc(cachep, flags);
3425 #endif /* CONFIG_NUMA */
3427 static __always_inline void *
3428 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3430 unsigned long save_flags;
3433 lockdep_trace_alloc(flags);
3435 if (slab_should_failslab(cachep, flags))
3438 cache_alloc_debugcheck_before(cachep, flags);
3439 local_irq_save(save_flags);
3440 objp = __do_cache_alloc(cachep, flags);
3441 local_irq_restore(save_flags);
3442 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3443 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3447 if (unlikely((flags & __GFP_ZERO) && objp))
3448 memset(objp, 0, obj_size(cachep));
3454 * Caller needs to acquire correct kmem_list's list_lock
3456 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3460 struct kmem_list3 *l3;
3462 for (i = 0; i < nr_objects; i++) {
3463 void *objp = objpp[i];
3466 slabp = virt_to_slab(objp);
3467 l3 = cachep->nodelists[node];
3468 list_del(&slabp->list);
3469 check_spinlock_acquired_node(cachep, node);
3470 check_slabp(cachep, slabp);
3471 slab_put_obj(cachep, slabp, objp, node);
3472 STATS_DEC_ACTIVE(cachep);
3474 check_slabp(cachep, slabp);
3476 /* fixup slab chains */
3477 if (slabp->inuse == 0) {
3478 if (l3->free_objects > l3->free_limit) {
3479 l3->free_objects -= cachep->num;
3480 /* No need to drop any previously held
3481 * lock here, even if we have a off-slab slab
3482 * descriptor it is guaranteed to come from
3483 * a different cache, refer to comments before
3486 slab_destroy(cachep, slabp);
3488 list_add(&slabp->list, &l3->slabs_free);
3491 /* Unconditionally move a slab to the end of the
3492 * partial list on free - maximum time for the
3493 * other objects to be freed, too.
3495 list_add_tail(&slabp->list, &l3->slabs_partial);
3500 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3503 struct kmem_list3 *l3;
3504 int node = numa_node_id();
3506 batchcount = ac->batchcount;
3508 BUG_ON(!batchcount || batchcount > ac->avail);
3511 l3 = cachep->nodelists[node];
3512 spin_lock(&l3->list_lock);
3514 struct array_cache *shared_array = l3->shared;
3515 int max = shared_array->limit - shared_array->avail;
3517 if (batchcount > max)
3519 memcpy(&(shared_array->entry[shared_array->avail]),
3520 ac->entry, sizeof(void *) * batchcount);
3521 shared_array->avail += batchcount;
3526 free_block(cachep, ac->entry, batchcount, node);
3531 struct list_head *p;
3533 p = l3->slabs_free.next;
3534 while (p != &(l3->slabs_free)) {
3537 slabp = list_entry(p, struct slab, list);
3538 BUG_ON(slabp->inuse);
3543 STATS_SET_FREEABLE(cachep, i);
3546 spin_unlock(&l3->list_lock);
3547 ac->avail -= batchcount;
3548 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3552 * Release an obj back to its cache. If the obj has a constructed state, it must
3553 * be in this state _before_ it is released. Called with disabled ints.
3555 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3557 struct array_cache *ac = cpu_cache_get(cachep);
3560 kmemleak_free_recursive(objp, cachep->flags);
3561 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3564 * Skip calling cache_free_alien() when the platform is not numa.
3565 * This will avoid cache misses that happen while accessing slabp (which
3566 * is per page memory reference) to get nodeid. Instead use a global
3567 * variable to skip the call, which is mostly likely to be present in
3570 if (numa_platform && cache_free_alien(cachep, objp))
3573 if (likely(ac->avail < ac->limit)) {
3574 STATS_INC_FREEHIT(cachep);
3575 ac->entry[ac->avail++] = objp;
3578 STATS_INC_FREEMISS(cachep);
3579 cache_flusharray(cachep, ac);
3580 ac->entry[ac->avail++] = objp;
3585 * kmem_cache_alloc - Allocate an object
3586 * @cachep: The cache to allocate from.
3587 * @flags: See kmalloc().
3589 * Allocate an object from this cache. The flags are only relevant
3590 * if the cache has no available objects.
3592 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3594 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3596 trace_kmem_cache_alloc(_RET_IP_, ret,
3597 obj_size(cachep), cachep->buffer_size, flags);
3601 EXPORT_SYMBOL(kmem_cache_alloc);
3603 #ifdef CONFIG_KMEMTRACE
3604 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3606 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3608 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3612 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3613 * @cachep: the cache we're checking against
3614 * @ptr: pointer to validate
3616 * This verifies that the untrusted pointer looks sane;
3617 * it is _not_ a guarantee that the pointer is actually
3618 * part of the slab cache in question, but it at least
3619 * validates that the pointer can be dereferenced and
3620 * looks half-way sane.
3622 * Currently only used for dentry validation.
3624 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3626 unsigned long addr = (unsigned long)ptr;
3627 unsigned long min_addr = PAGE_OFFSET;
3628 unsigned long align_mask = BYTES_PER_WORD - 1;
3629 unsigned long size = cachep->buffer_size;
3632 if (unlikely(addr < min_addr))
3634 if (unlikely(addr > (unsigned long)high_memory - size))
3636 if (unlikely(addr & align_mask))
3638 if (unlikely(!kern_addr_valid(addr)))
3640 if (unlikely(!kern_addr_valid(addr + size - 1)))
3642 page = virt_to_page(ptr);
3643 if (unlikely(!PageSlab(page)))
3645 if (unlikely(page_get_cache(page) != cachep))
3653 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3655 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3656 __builtin_return_address(0));
3658 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3659 obj_size(cachep), cachep->buffer_size,
3664 EXPORT_SYMBOL(kmem_cache_alloc_node);
3666 #ifdef CONFIG_KMEMTRACE
3667 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3671 return __cache_alloc_node(cachep, flags, nodeid,
3672 __builtin_return_address(0));
3674 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3677 static __always_inline void *
3678 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3680 struct kmem_cache *cachep;
3683 cachep = kmem_find_general_cachep(size, flags);
3684 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3686 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3688 trace_kmalloc_node((unsigned long) caller, ret,
3689 size, cachep->buffer_size, flags, node);
3694 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3695 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3697 return __do_kmalloc_node(size, flags, node,
3698 __builtin_return_address(0));
3700 EXPORT_SYMBOL(__kmalloc_node);
3702 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3703 int node, unsigned long caller)
3705 return __do_kmalloc_node(size, flags, node, (void *)caller);
3707 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3709 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3711 return __do_kmalloc_node(size, flags, node, NULL);
3713 EXPORT_SYMBOL(__kmalloc_node);
3714 #endif /* CONFIG_DEBUG_SLAB */
3715 #endif /* CONFIG_NUMA */
3718 * __do_kmalloc - allocate memory
3719 * @size: how many bytes of memory are required.
3720 * @flags: the type of memory to allocate (see kmalloc).
3721 * @caller: function caller for debug tracking of the caller
3723 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3726 struct kmem_cache *cachep;
3729 /* If you want to save a few bytes .text space: replace
3731 * Then kmalloc uses the uninlined functions instead of the inline
3734 cachep = __find_general_cachep(size, flags);
3735 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3737 ret = __cache_alloc(cachep, flags, caller);
3739 trace_kmalloc((unsigned long) caller, ret,
3740 size, cachep->buffer_size, flags);
3746 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3747 void *__kmalloc(size_t size, gfp_t flags)
3749 return __do_kmalloc(size, flags, __builtin_return_address(0));
3751 EXPORT_SYMBOL(__kmalloc);
3753 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3755 return __do_kmalloc(size, flags, (void *)caller);
3757 EXPORT_SYMBOL(__kmalloc_track_caller);
3760 void *__kmalloc(size_t size, gfp_t flags)
3762 return __do_kmalloc(size, flags, NULL);
3764 EXPORT_SYMBOL(__kmalloc);
3768 * kmem_cache_free - Deallocate an object
3769 * @cachep: The cache the allocation was from.
3770 * @objp: The previously allocated object.
3772 * Free an object which was previously allocated from this
3775 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3777 unsigned long flags;
3779 local_irq_save(flags);
3780 debug_check_no_locks_freed(objp, obj_size(cachep));
3781 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3782 debug_check_no_obj_freed(objp, obj_size(cachep));
3783 __cache_free(cachep, objp);
3784 local_irq_restore(flags);
3786 trace_kmem_cache_free(_RET_IP_, objp);
3788 EXPORT_SYMBOL(kmem_cache_free);
3791 * kfree - free previously allocated memory
3792 * @objp: pointer returned by kmalloc.
3794 * If @objp is NULL, no operation is performed.
3796 * Don't free memory not originally allocated by kmalloc()
3797 * or you will run into trouble.
3799 void kfree(const void *objp)
3801 struct kmem_cache *c;
3802 unsigned long flags;
3804 trace_kfree(_RET_IP_, objp);
3806 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3808 local_irq_save(flags);
3809 kfree_debugcheck(objp);
3810 c = virt_to_cache(objp);
3811 debug_check_no_locks_freed(objp, obj_size(c));
3812 debug_check_no_obj_freed(objp, obj_size(c));
3813 __cache_free(c, (void *)objp);
3814 local_irq_restore(flags);
3816 EXPORT_SYMBOL(kfree);
3818 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3820 return obj_size(cachep);
3822 EXPORT_SYMBOL(kmem_cache_size);
3824 const char *kmem_cache_name(struct kmem_cache *cachep)
3826 return cachep->name;
3828 EXPORT_SYMBOL_GPL(kmem_cache_name);
3831 * This initializes kmem_list3 or resizes various caches for all nodes.
3833 static int alloc_kmemlist(struct kmem_cache *cachep)
3836 struct kmem_list3 *l3;
3837 struct array_cache *new_shared;
3838 struct array_cache **new_alien = NULL;
3840 for_each_online_node(node) {
3842 if (use_alien_caches) {
3843 new_alien = alloc_alien_cache(node, cachep->limit);
3849 if (cachep->shared) {
3850 new_shared = alloc_arraycache(node,
3851 cachep->shared*cachep->batchcount,
3854 free_alien_cache(new_alien);
3859 l3 = cachep->nodelists[node];
3861 struct array_cache *shared = l3->shared;
3863 spin_lock_irq(&l3->list_lock);
3866 free_block(cachep, shared->entry,
3867 shared->avail, node);
3869 l3->shared = new_shared;
3871 l3->alien = new_alien;
3874 l3->free_limit = (1 + nr_cpus_node(node)) *
3875 cachep->batchcount + cachep->num;
3876 spin_unlock_irq(&l3->list_lock);
3878 free_alien_cache(new_alien);
3881 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3883 free_alien_cache(new_alien);
3888 kmem_list3_init(l3);
3889 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3890 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3891 l3->shared = new_shared;
3892 l3->alien = new_alien;
3893 l3->free_limit = (1 + nr_cpus_node(node)) *
3894 cachep->batchcount + cachep->num;
3895 cachep->nodelists[node] = l3;
3900 if (!cachep->next.next) {
3901 /* Cache is not active yet. Roll back what we did */
3904 if (cachep->nodelists[node]) {
3905 l3 = cachep->nodelists[node];
3908 free_alien_cache(l3->alien);
3910 cachep->nodelists[node] = NULL;
3918 struct ccupdate_struct {
3919 struct kmem_cache *cachep;
3920 struct array_cache *new[NR_CPUS];
3923 static void do_ccupdate_local(void *info)
3925 struct ccupdate_struct *new = info;
3926 struct array_cache *old;
3929 old = cpu_cache_get(new->cachep);
3931 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3932 new->new[smp_processor_id()] = old;
3935 /* Always called with the cache_chain_mutex held */
3936 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3937 int batchcount, int shared)
3939 struct ccupdate_struct *new;
3942 new = kzalloc(sizeof(*new), GFP_KERNEL);
3946 for_each_online_cpu(i) {
3947 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3950 for (i--; i >= 0; i--)
3956 new->cachep = cachep;
3958 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3961 cachep->batchcount = batchcount;
3962 cachep->limit = limit;
3963 cachep->shared = shared;
3965 for_each_online_cpu(i) {
3966 struct array_cache *ccold = new->new[i];
3969 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3970 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3971 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3975 return alloc_kmemlist(cachep);
3978 /* Called with cache_chain_mutex held always */
3979 static int enable_cpucache(struct kmem_cache *cachep)
3985 * The head array serves three purposes:
3986 * - create a LIFO ordering, i.e. return objects that are cache-warm
3987 * - reduce the number of spinlock operations.
3988 * - reduce the number of linked list operations on the slab and
3989 * bufctl chains: array operations are cheaper.
3990 * The numbers are guessed, we should auto-tune as described by
3993 if (cachep->buffer_size > 131072)
3995 else if (cachep->buffer_size > PAGE_SIZE)
3997 else if (cachep->buffer_size > 1024)
3999 else if (cachep->buffer_size > 256)
4005 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4006 * allocation behaviour: Most allocs on one cpu, most free operations
4007 * on another cpu. For these cases, an efficient object passing between
4008 * cpus is necessary. This is provided by a shared array. The array
4009 * replaces Bonwick's magazine layer.
4010 * On uniprocessor, it's functionally equivalent (but less efficient)
4011 * to a larger limit. Thus disabled by default.
4014 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4019 * With debugging enabled, large batchcount lead to excessively long
4020 * periods with disabled local interrupts. Limit the batchcount
4025 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4027 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4028 cachep->name, -err);
4033 * Drain an array if it contains any elements taking the l3 lock only if
4034 * necessary. Note that the l3 listlock also protects the array_cache
4035 * if drain_array() is used on the shared array.
4037 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4038 struct array_cache *ac, int force, int node)
4042 if (!ac || !ac->avail)
4044 if (ac->touched && !force) {
4047 spin_lock_irq(&l3->list_lock);
4049 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4050 if (tofree > ac->avail)
4051 tofree = (ac->avail + 1) / 2;
4052 free_block(cachep, ac->entry, tofree, node);
4053 ac->avail -= tofree;
4054 memmove(ac->entry, &(ac->entry[tofree]),
4055 sizeof(void *) * ac->avail);
4057 spin_unlock_irq(&l3->list_lock);
4062 * cache_reap - Reclaim memory from caches.
4063 * @w: work descriptor
4065 * Called from workqueue/eventd every few seconds.
4067 * - clear the per-cpu caches for this CPU.
4068 * - return freeable pages to the main free memory pool.
4070 * If we cannot acquire the cache chain mutex then just give up - we'll try
4071 * again on the next iteration.
4073 static void cache_reap(struct work_struct *w)
4075 struct kmem_cache *searchp;
4076 struct kmem_list3 *l3;
4077 int node = numa_node_id();
4078 struct delayed_work *work = to_delayed_work(w);
4080 if (!mutex_trylock(&cache_chain_mutex))
4081 /* Give up. Setup the next iteration. */
4084 list_for_each_entry(searchp, &cache_chain, next) {
4088 * We only take the l3 lock if absolutely necessary and we
4089 * have established with reasonable certainty that
4090 * we can do some work if the lock was obtained.
4092 l3 = searchp->nodelists[node];
4094 reap_alien(searchp, l3);
4096 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4099 * These are racy checks but it does not matter
4100 * if we skip one check or scan twice.
4102 if (time_after(l3->next_reap, jiffies))
4105 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4107 drain_array(searchp, l3, l3->shared, 0, node);
4109 if (l3->free_touched)
4110 l3->free_touched = 0;
4114 freed = drain_freelist(searchp, l3, (l3->free_limit +
4115 5 * searchp->num - 1) / (5 * searchp->num));
4116 STATS_ADD_REAPED(searchp, freed);
4122 mutex_unlock(&cache_chain_mutex);
4125 /* Set up the next iteration */
4126 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4129 #ifdef CONFIG_SLABINFO
4131 static void print_slabinfo_header(struct seq_file *m)
4134 * Output format version, so at least we can change it
4135 * without _too_ many complaints.
4138 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4140 seq_puts(m, "slabinfo - version: 2.1\n");
4142 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4143 "<objperslab> <pagesperslab>");
4144 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4145 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4147 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4148 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4149 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4154 static void *s_start(struct seq_file *m, loff_t *pos)
4158 mutex_lock(&cache_chain_mutex);
4160 print_slabinfo_header(m);
4162 return seq_list_start(&cache_chain, *pos);
4165 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4167 return seq_list_next(p, &cache_chain, pos);
4170 static void s_stop(struct seq_file *m, void *p)
4172 mutex_unlock(&cache_chain_mutex);
4175 static int s_show(struct seq_file *m, void *p)
4177 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4179 unsigned long active_objs;
4180 unsigned long num_objs;
4181 unsigned long active_slabs = 0;
4182 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4186 struct kmem_list3 *l3;
4190 for_each_online_node(node) {
4191 l3 = cachep->nodelists[node];
4196 spin_lock_irq(&l3->list_lock);
4198 list_for_each_entry(slabp, &l3->slabs_full, list) {
4199 if (slabp->inuse != cachep->num && !error)
4200 error = "slabs_full accounting error";
4201 active_objs += cachep->num;
4204 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4205 if (slabp->inuse == cachep->num && !error)
4206 error = "slabs_partial inuse accounting error";
4207 if (!slabp->inuse && !error)
4208 error = "slabs_partial/inuse accounting error";
4209 active_objs += slabp->inuse;
4212 list_for_each_entry(slabp, &l3->slabs_free, list) {
4213 if (slabp->inuse && !error)
4214 error = "slabs_free/inuse accounting error";
4217 free_objects += l3->free_objects;
4219 shared_avail += l3->shared->avail;
4221 spin_unlock_irq(&l3->list_lock);
4223 num_slabs += active_slabs;
4224 num_objs = num_slabs * cachep->num;
4225 if (num_objs - active_objs != free_objects && !error)
4226 error = "free_objects accounting error";
4228 name = cachep->name;
4230 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4232 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4233 name, active_objs, num_objs, cachep->buffer_size,
4234 cachep->num, (1 << cachep->gfporder));
4235 seq_printf(m, " : tunables %4u %4u %4u",
4236 cachep->limit, cachep->batchcount, cachep->shared);
4237 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4238 active_slabs, num_slabs, shared_avail);
4241 unsigned long high = cachep->high_mark;
4242 unsigned long allocs = cachep->num_allocations;
4243 unsigned long grown = cachep->grown;
4244 unsigned long reaped = cachep->reaped;
4245 unsigned long errors = cachep->errors;
4246 unsigned long max_freeable = cachep->max_freeable;
4247 unsigned long node_allocs = cachep->node_allocs;
4248 unsigned long node_frees = cachep->node_frees;
4249 unsigned long overflows = cachep->node_overflow;
4251 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4252 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4253 reaped, errors, max_freeable, node_allocs,
4254 node_frees, overflows);
4258 unsigned long allochit = atomic_read(&cachep->allochit);
4259 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4260 unsigned long freehit = atomic_read(&cachep->freehit);
4261 unsigned long freemiss = atomic_read(&cachep->freemiss);
4263 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4264 allochit, allocmiss, freehit, freemiss);
4272 * slabinfo_op - iterator that generates /proc/slabinfo
4281 * num-pages-per-slab
4282 * + further values on SMP and with statistics enabled
4285 static const struct seq_operations slabinfo_op = {
4292 #define MAX_SLABINFO_WRITE 128
4294 * slabinfo_write - Tuning for the slab allocator
4296 * @buffer: user buffer
4297 * @count: data length
4300 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4301 size_t count, loff_t *ppos)
4303 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4304 int limit, batchcount, shared, res;
4305 struct kmem_cache *cachep;
4307 if (count > MAX_SLABINFO_WRITE)
4309 if (copy_from_user(&kbuf, buffer, count))
4311 kbuf[MAX_SLABINFO_WRITE] = '\0';
4313 tmp = strchr(kbuf, ' ');
4318 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4321 /* Find the cache in the chain of caches. */
4322 mutex_lock(&cache_chain_mutex);
4324 list_for_each_entry(cachep, &cache_chain, next) {
4325 if (!strcmp(cachep->name, kbuf)) {
4326 if (limit < 1 || batchcount < 1 ||
4327 batchcount > limit || shared < 0) {
4330 res = do_tune_cpucache(cachep, limit,
4331 batchcount, shared);
4336 mutex_unlock(&cache_chain_mutex);
4342 static int slabinfo_open(struct inode *inode, struct file *file)
4344 return seq_open(file, &slabinfo_op);
4347 static const struct file_operations proc_slabinfo_operations = {
4348 .open = slabinfo_open,
4350 .write = slabinfo_write,
4351 .llseek = seq_lseek,
4352 .release = seq_release,
4355 #ifdef CONFIG_DEBUG_SLAB_LEAK
4357 static void *leaks_start(struct seq_file *m, loff_t *pos)
4359 mutex_lock(&cache_chain_mutex);
4360 return seq_list_start(&cache_chain, *pos);
4363 static inline int add_caller(unsigned long *n, unsigned long v)
4373 unsigned long *q = p + 2 * i;
4387 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4393 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4399 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4400 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4402 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4407 static void show_symbol(struct seq_file *m, unsigned long address)
4409 #ifdef CONFIG_KALLSYMS
4410 unsigned long offset, size;
4411 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4413 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4414 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4416 seq_printf(m, " [%s]", modname);
4420 seq_printf(m, "%p", (void *)address);
4423 static int leaks_show(struct seq_file *m, void *p)
4425 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4427 struct kmem_list3 *l3;
4429 unsigned long *n = m->private;
4433 if (!(cachep->flags & SLAB_STORE_USER))
4435 if (!(cachep->flags & SLAB_RED_ZONE))
4438 /* OK, we can do it */
4442 for_each_online_node(node) {
4443 l3 = cachep->nodelists[node];
4448 spin_lock_irq(&l3->list_lock);
4450 list_for_each_entry(slabp, &l3->slabs_full, list)
4451 handle_slab(n, cachep, slabp);
4452 list_for_each_entry(slabp, &l3->slabs_partial, list)
4453 handle_slab(n, cachep, slabp);
4454 spin_unlock_irq(&l3->list_lock);
4456 name = cachep->name;
4458 /* Increase the buffer size */
4459 mutex_unlock(&cache_chain_mutex);
4460 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4462 /* Too bad, we are really out */
4464 mutex_lock(&cache_chain_mutex);
4467 *(unsigned long *)m->private = n[0] * 2;
4469 mutex_lock(&cache_chain_mutex);
4470 /* Now make sure this entry will be retried */
4474 for (i = 0; i < n[1]; i++) {
4475 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4476 show_symbol(m, n[2*i+2]);
4483 static const struct seq_operations slabstats_op = {
4484 .start = leaks_start,
4490 static int slabstats_open(struct inode *inode, struct file *file)
4492 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4495 ret = seq_open(file, &slabstats_op);
4497 struct seq_file *m = file->private_data;
4498 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4507 static const struct file_operations proc_slabstats_operations = {
4508 .open = slabstats_open,
4510 .llseek = seq_lseek,
4511 .release = seq_release_private,
4515 static int __init slab_proc_init(void)
4517 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4518 #ifdef CONFIG_DEBUG_SLAB_LEAK
4519 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4523 module_init(slab_proc_init);
4527 * ksize - get the actual amount of memory allocated for a given object
4528 * @objp: Pointer to the object
4530 * kmalloc may internally round up allocations and return more memory
4531 * than requested. ksize() can be used to determine the actual amount of
4532 * memory allocated. The caller may use this additional memory, even though
4533 * a smaller amount of memory was initially specified with the kmalloc call.
4534 * The caller must guarantee that objp points to a valid object previously
4535 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4536 * must not be freed during the duration of the call.
4538 size_t ksize(const void *objp)
4541 if (unlikely(objp == ZERO_SIZE_PTR))
4544 return obj_size(virt_to_cache(objp));
4546 EXPORT_SYMBOL(ksize);