3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
131 #define FORCED_DEBUG 1
135 #define FORCED_DEBUG 0
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
141 #ifndef cache_line_size
142 #define cache_line_size() L1_CACHE_BYTES
145 #ifndef ARCH_KMALLOC_MINALIGN
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
191 * Bufctl's are used for linking objs within a slab
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 typedef unsigned int kmem_bufctl_t;
208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * The slab lists for all objects.
291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
295 unsigned int free_limit;
296 unsigned int colour_next; /* Per-node cache coloring */
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 static int enable_cpucache(struct kmem_cache *cachep);
318 static void cache_reap(struct work_struct *unused);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline int index_of(const size_t size)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size)) {
336 #include "linux/kmalloc_sizes.h"
344 static int slab_early_init = 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3 *parent)
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
356 parent->colour_next = 0;
357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache *array[NR_CPUS];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount;
389 unsigned int buffer_size;
390 u32 reciprocal_buffer_size;
391 /* 3) touched by every alloc & free from the backend */
393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder;
400 /* force GFP flags, e.g. GFP_DMA */
403 size_t colour; /* cache colouring range */
404 unsigned int colour_off; /* colour offset */
405 struct kmem_cache *slabp_cache;
406 unsigned int slab_size;
407 unsigned int dflags; /* dynamic flags */
409 /* constructor func */
410 void (*ctor) (void *, struct kmem_cache *, unsigned long);
412 /* 5) cache creation/removal */
414 struct list_head next;
418 unsigned long num_active;
419 unsigned long num_allocations;
420 unsigned long high_mark;
422 unsigned long reaped;
423 unsigned long errors;
424 unsigned long max_freeable;
425 unsigned long node_allocs;
426 unsigned long node_frees;
427 unsigned long node_overflow;
435 * If debugging is enabled, then the allocator can add additional
436 * fields and/or padding to every object. buffer_size contains the total
437 * object size including these internal fields, the following two
438 * variables contain the offset to the user object and its size.
444 * We put nodelists[] at the end of kmem_cache, because we want to size
445 * this array to nr_node_ids slots instead of MAX_NUMNODES
446 * (see kmem_cache_init())
447 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
448 * is statically defined, so we reserve the max number of nodes.
450 struct kmem_list3 *nodelists[MAX_NUMNODES];
452 * Do not add fields after nodelists[]
456 #define CFLGS_OFF_SLAB (0x80000000UL)
457 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459 #define BATCHREFILL_LIMIT 16
461 * Optimization question: fewer reaps means less probability for unnessary
462 * cpucache drain/refill cycles.
464 * OTOH the cpuarrays can contain lots of objects,
465 * which could lock up otherwise freeable slabs.
467 #define REAPTIMEOUT_CPUC (2*HZ)
468 #define REAPTIMEOUT_LIST3 (4*HZ)
471 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
472 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
473 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
474 #define STATS_INC_GROWN(x) ((x)->grown++)
475 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
476 #define STATS_SET_HIGH(x) \
478 if ((x)->num_active > (x)->high_mark) \
479 (x)->high_mark = (x)->num_active; \
481 #define STATS_INC_ERR(x) ((x)->errors++)
482 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
483 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
484 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
485 #define STATS_SET_FREEABLE(x, i) \
487 if ((x)->max_freeable < i) \
488 (x)->max_freeable = i; \
490 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
491 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
492 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
493 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
495 #define STATS_INC_ACTIVE(x) do { } while (0)
496 #define STATS_DEC_ACTIVE(x) do { } while (0)
497 #define STATS_INC_ALLOCED(x) do { } while (0)
498 #define STATS_INC_GROWN(x) do { } while (0)
499 #define STATS_ADD_REAPED(x,y) do { } while (0)
500 #define STATS_SET_HIGH(x) do { } while (0)
501 #define STATS_INC_ERR(x) do { } while (0)
502 #define STATS_INC_NODEALLOCS(x) do { } while (0)
503 #define STATS_INC_NODEFREES(x) do { } while (0)
504 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
505 #define STATS_SET_FREEABLE(x, i) do { } while (0)
506 #define STATS_INC_ALLOCHIT(x) do { } while (0)
507 #define STATS_INC_ALLOCMISS(x) do { } while (0)
508 #define STATS_INC_FREEHIT(x) do { } while (0)
509 #define STATS_INC_FREEMISS(x) do { } while (0)
515 * memory layout of objects:
517 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
518 * the end of an object is aligned with the end of the real
519 * allocation. Catches writes behind the end of the allocation.
520 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
522 * cachep->obj_offset: The real object.
523 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
524 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
525 * [BYTES_PER_WORD long]
527 static int obj_offset(struct kmem_cache *cachep)
529 return cachep->obj_offset;
532 static int obj_size(struct kmem_cache *cachep)
534 return cachep->obj_size;
537 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
539 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
540 return (unsigned long long*) (objp + obj_offset(cachep) -
541 sizeof(unsigned long long));
544 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
546 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
547 if (cachep->flags & SLAB_STORE_USER)
548 return (unsigned long long *)(objp + cachep->buffer_size -
549 sizeof(unsigned long long) -
551 return (unsigned long long *) (objp + cachep->buffer_size -
552 sizeof(unsigned long long));
555 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
557 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
558 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
563 #define obj_offset(x) 0
564 #define obj_size(cachep) (cachep->buffer_size)
565 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
566 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
572 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
575 #if defined(CONFIG_LARGE_ALLOCS)
576 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
577 #define MAX_GFP_ORDER 13 /* up to 32Mb */
578 #elif defined(CONFIG_MMU)
579 #define MAX_OBJ_ORDER 5 /* 32 pages */
580 #define MAX_GFP_ORDER 5 /* 32 pages */
582 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
583 #define MAX_GFP_ORDER 8 /* up to 1Mb */
587 * Do not go above this order unless 0 objects fit into the slab.
589 #define BREAK_GFP_ORDER_HI 1
590 #define BREAK_GFP_ORDER_LO 0
591 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
594 * Functions for storing/retrieving the cachep and or slab from the page
595 * allocator. These are used to find the slab an obj belongs to. With kfree(),
596 * these are used to find the cache which an obj belongs to.
598 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
600 page->lru.next = (struct list_head *)cache;
603 static inline struct kmem_cache *page_get_cache(struct page *page)
605 page = compound_head(page);
606 BUG_ON(!PageSlab(page));
607 return (struct kmem_cache *)page->lru.next;
610 static inline void page_set_slab(struct page *page, struct slab *slab)
612 page->lru.prev = (struct list_head *)slab;
615 static inline struct slab *page_get_slab(struct page *page)
617 BUG_ON(!PageSlab(page));
618 return (struct slab *)page->lru.prev;
621 static inline struct kmem_cache *virt_to_cache(const void *obj)
623 struct page *page = virt_to_head_page(obj);
624 return page_get_cache(page);
627 static inline struct slab *virt_to_slab(const void *obj)
629 struct page *page = virt_to_head_page(obj);
630 return page_get_slab(page);
633 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
636 return slab->s_mem + cache->buffer_size * idx;
640 * We want to avoid an expensive divide : (offset / cache->buffer_size)
641 * Using the fact that buffer_size is a constant for a particular cache,
642 * we can replace (offset / cache->buffer_size) by
643 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
645 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
646 const struct slab *slab, void *obj)
648 u32 offset = (obj - slab->s_mem);
649 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
653 * These are the default caches for kmalloc. Custom caches can have other sizes.
655 struct cache_sizes malloc_sizes[] = {
656 #define CACHE(x) { .cs_size = (x) },
657 #include <linux/kmalloc_sizes.h>
661 EXPORT_SYMBOL(malloc_sizes);
663 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
669 static struct cache_names __initdata cache_names[] = {
670 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
671 #include <linux/kmalloc_sizes.h>
676 static struct arraycache_init initarray_cache __initdata =
677 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
678 static struct arraycache_init initarray_generic =
679 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
681 /* internal cache of cache description objs */
682 static struct kmem_cache cache_cache = {
684 .limit = BOOT_CPUCACHE_ENTRIES,
686 .buffer_size = sizeof(struct kmem_cache),
687 .name = "kmem_cache",
690 #define BAD_ALIEN_MAGIC 0x01020304ul
692 #ifdef CONFIG_LOCKDEP
695 * Slab sometimes uses the kmalloc slabs to store the slab headers
696 * for other slabs "off slab".
697 * The locking for this is tricky in that it nests within the locks
698 * of all other slabs in a few places; to deal with this special
699 * locking we put on-slab caches into a separate lock-class.
701 * We set lock class for alien array caches which are up during init.
702 * The lock annotation will be lost if all cpus of a node goes down and
703 * then comes back up during hotplug
705 static struct lock_class_key on_slab_l3_key;
706 static struct lock_class_key on_slab_alc_key;
708 static inline void init_lock_keys(void)
712 struct cache_sizes *s = malloc_sizes;
714 while (s->cs_size != ULONG_MAX) {
716 struct array_cache **alc;
718 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
719 if (!l3 || OFF_SLAB(s->cs_cachep))
721 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
724 * FIXME: This check for BAD_ALIEN_MAGIC
725 * should go away when common slab code is taught to
726 * work even without alien caches.
727 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
728 * for alloc_alien_cache,
730 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
734 lockdep_set_class(&alc[r]->lock,
742 static inline void init_lock_keys(void)
748 * 1. Guard access to the cache-chain.
749 * 2. Protect sanity of cpu_online_map against cpu hotplug events
751 static DEFINE_MUTEX(cache_chain_mutex);
752 static struct list_head cache_chain;
755 * chicken and egg problem: delay the per-cpu array allocation
756 * until the general caches are up.
766 * used by boot code to determine if it can use slab based allocator
768 int slab_is_available(void)
770 return g_cpucache_up == FULL;
773 static DEFINE_PER_CPU(struct delayed_work, reap_work);
775 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
777 return cachep->array[smp_processor_id()];
780 static inline struct kmem_cache *__find_general_cachep(size_t size,
783 struct cache_sizes *csizep = malloc_sizes;
786 /* This happens if someone tries to call
787 * kmem_cache_create(), or __kmalloc(), before
788 * the generic caches are initialized.
790 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
792 while (size > csizep->cs_size)
796 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
797 * has cs_{dma,}cachep==NULL. Thus no special case
798 * for large kmalloc calls required.
800 #ifdef CONFIG_ZONE_DMA
801 if (unlikely(gfpflags & GFP_DMA))
802 return csizep->cs_dmacachep;
804 return csizep->cs_cachep;
807 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
809 return __find_general_cachep(size, gfpflags);
812 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
814 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
818 * Calculate the number of objects and left-over bytes for a given buffer size.
820 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
821 size_t align, int flags, size_t *left_over,
826 size_t slab_size = PAGE_SIZE << gfporder;
829 * The slab management structure can be either off the slab or
830 * on it. For the latter case, the memory allocated for a
834 * - One kmem_bufctl_t for each object
835 * - Padding to respect alignment of @align
836 * - @buffer_size bytes for each object
838 * If the slab management structure is off the slab, then the
839 * alignment will already be calculated into the size. Because
840 * the slabs are all pages aligned, the objects will be at the
841 * correct alignment when allocated.
843 if (flags & CFLGS_OFF_SLAB) {
845 nr_objs = slab_size / buffer_size;
847 if (nr_objs > SLAB_LIMIT)
848 nr_objs = SLAB_LIMIT;
851 * Ignore padding for the initial guess. The padding
852 * is at most @align-1 bytes, and @buffer_size is at
853 * least @align. In the worst case, this result will
854 * be one greater than the number of objects that fit
855 * into the memory allocation when taking the padding
858 nr_objs = (slab_size - sizeof(struct slab)) /
859 (buffer_size + sizeof(kmem_bufctl_t));
862 * This calculated number will be either the right
863 * amount, or one greater than what we want.
865 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
869 if (nr_objs > SLAB_LIMIT)
870 nr_objs = SLAB_LIMIT;
872 mgmt_size = slab_mgmt_size(nr_objs, align);
875 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
878 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
880 static void __slab_error(const char *function, struct kmem_cache *cachep,
883 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
884 function, cachep->name, msg);
889 * By default on NUMA we use alien caches to stage the freeing of
890 * objects allocated from other nodes. This causes massive memory
891 * inefficiencies when using fake NUMA setup to split memory into a
892 * large number of small nodes, so it can be disabled on the command
896 static int use_alien_caches __read_mostly = 1;
897 static int __init noaliencache_setup(char *s)
899 use_alien_caches = 0;
902 __setup("noaliencache", noaliencache_setup);
906 * Special reaping functions for NUMA systems called from cache_reap().
907 * These take care of doing round robin flushing of alien caches (containing
908 * objects freed on different nodes from which they were allocated) and the
909 * flushing of remote pcps by calling drain_node_pages.
911 static DEFINE_PER_CPU(unsigned long, reap_node);
913 static void init_reap_node(int cpu)
917 node = next_node(cpu_to_node(cpu), node_online_map);
918 if (node == MAX_NUMNODES)
919 node = first_node(node_online_map);
921 per_cpu(reap_node, cpu) = node;
924 static void next_reap_node(void)
926 int node = __get_cpu_var(reap_node);
928 node = next_node(node, node_online_map);
929 if (unlikely(node >= MAX_NUMNODES))
930 node = first_node(node_online_map);
931 __get_cpu_var(reap_node) = node;
935 #define init_reap_node(cpu) do { } while (0)
936 #define next_reap_node(void) do { } while (0)
940 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
941 * via the workqueue/eventd.
942 * Add the CPU number into the expiration time to minimize the possibility of
943 * the CPUs getting into lockstep and contending for the global cache chain
946 static void __devinit start_cpu_timer(int cpu)
948 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
951 * When this gets called from do_initcalls via cpucache_init(),
952 * init_workqueues() has already run, so keventd will be setup
955 if (keventd_up() && reap_work->work.func == NULL) {
957 INIT_DELAYED_WORK(reap_work, cache_reap);
958 schedule_delayed_work_on(cpu, reap_work,
959 __round_jiffies_relative(HZ, cpu));
963 static struct array_cache *alloc_arraycache(int node, int entries,
966 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
967 struct array_cache *nc = NULL;
969 nc = kmalloc_node(memsize, GFP_KERNEL, node);
973 nc->batchcount = batchcount;
975 spin_lock_init(&nc->lock);
981 * Transfer objects in one arraycache to another.
982 * Locking must be handled by the caller.
984 * Return the number of entries transferred.
986 static int transfer_objects(struct array_cache *to,
987 struct array_cache *from, unsigned int max)
989 /* Figure out how many entries to transfer */
990 int nr = min(min(from->avail, max), to->limit - to->avail);
995 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1006 #define drain_alien_cache(cachep, alien) do { } while (0)
1007 #define reap_alien(cachep, l3) do { } while (0)
1009 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1011 return (struct array_cache **)BAD_ALIEN_MAGIC;
1014 static inline void free_alien_cache(struct array_cache **ac_ptr)
1018 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1023 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1029 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1030 gfp_t flags, int nodeid)
1035 #else /* CONFIG_NUMA */
1037 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1038 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1040 static struct array_cache **alloc_alien_cache(int node, int limit)
1042 struct array_cache **ac_ptr;
1043 int memsize = sizeof(void *) * nr_node_ids;
1048 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1051 if (i == node || !node_online(i)) {
1055 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1057 for (i--; i <= 0; i--)
1067 static void free_alien_cache(struct array_cache **ac_ptr)
1078 static void __drain_alien_cache(struct kmem_cache *cachep,
1079 struct array_cache *ac, int node)
1081 struct kmem_list3 *rl3 = cachep->nodelists[node];
1084 spin_lock(&rl3->list_lock);
1086 * Stuff objects into the remote nodes shared array first.
1087 * That way we could avoid the overhead of putting the objects
1088 * into the free lists and getting them back later.
1091 transfer_objects(rl3->shared, ac, ac->limit);
1093 free_block(cachep, ac->entry, ac->avail, node);
1095 spin_unlock(&rl3->list_lock);
1100 * Called from cache_reap() to regularly drain alien caches round robin.
1102 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1104 int node = __get_cpu_var(reap_node);
1107 struct array_cache *ac = l3->alien[node];
1109 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1110 __drain_alien_cache(cachep, ac, node);
1111 spin_unlock_irq(&ac->lock);
1116 static void drain_alien_cache(struct kmem_cache *cachep,
1117 struct array_cache **alien)
1120 struct array_cache *ac;
1121 unsigned long flags;
1123 for_each_online_node(i) {
1126 spin_lock_irqsave(&ac->lock, flags);
1127 __drain_alien_cache(cachep, ac, i);
1128 spin_unlock_irqrestore(&ac->lock, flags);
1133 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1135 struct slab *slabp = virt_to_slab(objp);
1136 int nodeid = slabp->nodeid;
1137 struct kmem_list3 *l3;
1138 struct array_cache *alien = NULL;
1141 node = numa_node_id();
1144 * Make sure we are not freeing a object from another node to the array
1145 * cache on this cpu.
1147 if (likely(slabp->nodeid == node))
1150 l3 = cachep->nodelists[node];
1151 STATS_INC_NODEFREES(cachep);
1152 if (l3->alien && l3->alien[nodeid]) {
1153 alien = l3->alien[nodeid];
1154 spin_lock(&alien->lock);
1155 if (unlikely(alien->avail == alien->limit)) {
1156 STATS_INC_ACOVERFLOW(cachep);
1157 __drain_alien_cache(cachep, alien, nodeid);
1159 alien->entry[alien->avail++] = objp;
1160 spin_unlock(&alien->lock);
1162 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1163 free_block(cachep, &objp, 1, nodeid);
1164 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1170 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1171 unsigned long action, void *hcpu)
1173 long cpu = (long)hcpu;
1174 struct kmem_cache *cachep;
1175 struct kmem_list3 *l3 = NULL;
1176 int node = cpu_to_node(cpu);
1177 int memsize = sizeof(struct kmem_list3);
1180 case CPU_LOCK_ACQUIRE:
1181 mutex_lock(&cache_chain_mutex);
1183 case CPU_UP_PREPARE:
1184 case CPU_UP_PREPARE_FROZEN:
1186 * We need to do this right in the beginning since
1187 * alloc_arraycache's are going to use this list.
1188 * kmalloc_node allows us to add the slab to the right
1189 * kmem_list3 and not this cpu's kmem_list3
1192 list_for_each_entry(cachep, &cache_chain, next) {
1194 * Set up the size64 kmemlist for cpu before we can
1195 * begin anything. Make sure some other cpu on this
1196 * node has not already allocated this
1198 if (!cachep->nodelists[node]) {
1199 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1202 kmem_list3_init(l3);
1203 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1204 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1207 * The l3s don't come and go as CPUs come and
1208 * go. cache_chain_mutex is sufficient
1211 cachep->nodelists[node] = l3;
1214 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1215 cachep->nodelists[node]->free_limit =
1216 (1 + nr_cpus_node(node)) *
1217 cachep->batchcount + cachep->num;
1218 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1222 * Now we can go ahead with allocating the shared arrays and
1225 list_for_each_entry(cachep, &cache_chain, next) {
1226 struct array_cache *nc;
1227 struct array_cache *shared = NULL;
1228 struct array_cache **alien = NULL;
1230 nc = alloc_arraycache(node, cachep->limit,
1231 cachep->batchcount);
1234 if (cachep->shared) {
1235 shared = alloc_arraycache(node,
1236 cachep->shared * cachep->batchcount,
1241 if (use_alien_caches) {
1242 alien = alloc_alien_cache(node, cachep->limit);
1246 cachep->array[cpu] = nc;
1247 l3 = cachep->nodelists[node];
1250 spin_lock_irq(&l3->list_lock);
1253 * We are serialised from CPU_DEAD or
1254 * CPU_UP_CANCELLED by the cpucontrol lock
1256 l3->shared = shared;
1265 spin_unlock_irq(&l3->list_lock);
1267 free_alien_cache(alien);
1271 case CPU_ONLINE_FROZEN:
1272 start_cpu_timer(cpu);
1274 #ifdef CONFIG_HOTPLUG_CPU
1275 case CPU_DOWN_PREPARE:
1276 case CPU_DOWN_PREPARE_FROZEN:
1278 * Shutdown cache reaper. Note that the cache_chain_mutex is
1279 * held so that if cache_reap() is invoked it cannot do
1280 * anything expensive but will only modify reap_work
1281 * and reschedule the timer.
1283 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1284 /* Now the cache_reaper is guaranteed to be not running. */
1285 per_cpu(reap_work, cpu).work.func = NULL;
1287 case CPU_DOWN_FAILED:
1288 case CPU_DOWN_FAILED_FROZEN:
1289 start_cpu_timer(cpu);
1292 case CPU_DEAD_FROZEN:
1294 * Even if all the cpus of a node are down, we don't free the
1295 * kmem_list3 of any cache. This to avoid a race between
1296 * cpu_down, and a kmalloc allocation from another cpu for
1297 * memory from the node of the cpu going down. The list3
1298 * structure is usually allocated from kmem_cache_create() and
1299 * gets destroyed at kmem_cache_destroy().
1303 case CPU_UP_CANCELED:
1304 case CPU_UP_CANCELED_FROZEN:
1305 list_for_each_entry(cachep, &cache_chain, next) {
1306 struct array_cache *nc;
1307 struct array_cache *shared;
1308 struct array_cache **alien;
1311 mask = node_to_cpumask(node);
1312 /* cpu is dead; no one can alloc from it. */
1313 nc = cachep->array[cpu];
1314 cachep->array[cpu] = NULL;
1315 l3 = cachep->nodelists[node];
1318 goto free_array_cache;
1320 spin_lock_irq(&l3->list_lock);
1322 /* Free limit for this kmem_list3 */
1323 l3->free_limit -= cachep->batchcount;
1325 free_block(cachep, nc->entry, nc->avail, node);
1327 if (!cpus_empty(mask)) {
1328 spin_unlock_irq(&l3->list_lock);
1329 goto free_array_cache;
1332 shared = l3->shared;
1334 free_block(cachep, shared->entry,
1335 shared->avail, node);
1342 spin_unlock_irq(&l3->list_lock);
1346 drain_alien_cache(cachep, alien);
1347 free_alien_cache(alien);
1353 * In the previous loop, all the objects were freed to
1354 * the respective cache's slabs, now we can go ahead and
1355 * shrink each nodelist to its limit.
1357 list_for_each_entry(cachep, &cache_chain, next) {
1358 l3 = cachep->nodelists[node];
1361 drain_freelist(cachep, l3, l3->free_objects);
1364 case CPU_LOCK_RELEASE:
1365 mutex_unlock(&cache_chain_mutex);
1373 static struct notifier_block __cpuinitdata cpucache_notifier = {
1374 &cpuup_callback, NULL, 0
1378 * swap the static kmem_list3 with kmalloced memory
1380 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1383 struct kmem_list3 *ptr;
1385 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1388 local_irq_disable();
1389 memcpy(ptr, list, sizeof(struct kmem_list3));
1391 * Do not assume that spinlocks can be initialized via memcpy:
1393 spin_lock_init(&ptr->list_lock);
1395 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1396 cachep->nodelists[nodeid] = ptr;
1401 * Initialisation. Called after the page allocator have been initialised and
1402 * before smp_init().
1404 void __init kmem_cache_init(void)
1407 struct cache_sizes *sizes;
1408 struct cache_names *names;
1413 if (num_possible_nodes() == 1)
1414 use_alien_caches = 0;
1416 for (i = 0; i < NUM_INIT_LISTS; i++) {
1417 kmem_list3_init(&initkmem_list3[i]);
1418 if (i < MAX_NUMNODES)
1419 cache_cache.nodelists[i] = NULL;
1423 * Fragmentation resistance on low memory - only use bigger
1424 * page orders on machines with more than 32MB of memory.
1426 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1427 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1429 /* Bootstrap is tricky, because several objects are allocated
1430 * from caches that do not exist yet:
1431 * 1) initialize the cache_cache cache: it contains the struct
1432 * kmem_cache structures of all caches, except cache_cache itself:
1433 * cache_cache is statically allocated.
1434 * Initially an __init data area is used for the head array and the
1435 * kmem_list3 structures, it's replaced with a kmalloc allocated
1436 * array at the end of the bootstrap.
1437 * 2) Create the first kmalloc cache.
1438 * The struct kmem_cache for the new cache is allocated normally.
1439 * An __init data area is used for the head array.
1440 * 3) Create the remaining kmalloc caches, with minimally sized
1442 * 4) Replace the __init data head arrays for cache_cache and the first
1443 * kmalloc cache with kmalloc allocated arrays.
1444 * 5) Replace the __init data for kmem_list3 for cache_cache and
1445 * the other cache's with kmalloc allocated memory.
1446 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1449 node = numa_node_id();
1451 /* 1) create the cache_cache */
1452 INIT_LIST_HEAD(&cache_chain);
1453 list_add(&cache_cache.next, &cache_chain);
1454 cache_cache.colour_off = cache_line_size();
1455 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1456 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1459 * struct kmem_cache size depends on nr_node_ids, which
1460 * can be less than MAX_NUMNODES.
1462 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1463 nr_node_ids * sizeof(struct kmem_list3 *);
1465 cache_cache.obj_size = cache_cache.buffer_size;
1467 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1469 cache_cache.reciprocal_buffer_size =
1470 reciprocal_value(cache_cache.buffer_size);
1472 for (order = 0; order < MAX_ORDER; order++) {
1473 cache_estimate(order, cache_cache.buffer_size,
1474 cache_line_size(), 0, &left_over, &cache_cache.num);
1475 if (cache_cache.num)
1478 BUG_ON(!cache_cache.num);
1479 cache_cache.gfporder = order;
1480 cache_cache.colour = left_over / cache_cache.colour_off;
1481 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1482 sizeof(struct slab), cache_line_size());
1484 /* 2+3) create the kmalloc caches */
1485 sizes = malloc_sizes;
1486 names = cache_names;
1489 * Initialize the caches that provide memory for the array cache and the
1490 * kmem_list3 structures first. Without this, further allocations will
1494 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1495 sizes[INDEX_AC].cs_size,
1496 ARCH_KMALLOC_MINALIGN,
1497 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1500 if (INDEX_AC != INDEX_L3) {
1501 sizes[INDEX_L3].cs_cachep =
1502 kmem_cache_create(names[INDEX_L3].name,
1503 sizes[INDEX_L3].cs_size,
1504 ARCH_KMALLOC_MINALIGN,
1505 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1509 slab_early_init = 0;
1511 while (sizes->cs_size != ULONG_MAX) {
1513 * For performance, all the general caches are L1 aligned.
1514 * This should be particularly beneficial on SMP boxes, as it
1515 * eliminates "false sharing".
1516 * Note for systems short on memory removing the alignment will
1517 * allow tighter packing of the smaller caches.
1519 if (!sizes->cs_cachep) {
1520 sizes->cs_cachep = kmem_cache_create(names->name,
1522 ARCH_KMALLOC_MINALIGN,
1523 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1526 #ifdef CONFIG_ZONE_DMA
1527 sizes->cs_dmacachep = kmem_cache_create(
1530 ARCH_KMALLOC_MINALIGN,
1531 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1538 /* 4) Replace the bootstrap head arrays */
1540 struct array_cache *ptr;
1542 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1544 local_irq_disable();
1545 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1546 memcpy(ptr, cpu_cache_get(&cache_cache),
1547 sizeof(struct arraycache_init));
1549 * Do not assume that spinlocks can be initialized via memcpy:
1551 spin_lock_init(&ptr->lock);
1553 cache_cache.array[smp_processor_id()] = ptr;
1556 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1558 local_irq_disable();
1559 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1560 != &initarray_generic.cache);
1561 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1562 sizeof(struct arraycache_init));
1564 * Do not assume that spinlocks can be initialized via memcpy:
1566 spin_lock_init(&ptr->lock);
1568 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1572 /* 5) Replace the bootstrap kmem_list3's */
1576 /* Replace the static kmem_list3 structures for the boot cpu */
1577 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1579 for_each_online_node(nid) {
1580 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1581 &initkmem_list3[SIZE_AC + nid], nid);
1583 if (INDEX_AC != INDEX_L3) {
1584 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1585 &initkmem_list3[SIZE_L3 + nid], nid);
1590 /* 6) resize the head arrays to their final sizes */
1592 struct kmem_cache *cachep;
1593 mutex_lock(&cache_chain_mutex);
1594 list_for_each_entry(cachep, &cache_chain, next)
1595 if (enable_cpucache(cachep))
1597 mutex_unlock(&cache_chain_mutex);
1600 /* Annotate slab for lockdep -- annotate the malloc caches */
1605 g_cpucache_up = FULL;
1608 * Register a cpu startup notifier callback that initializes
1609 * cpu_cache_get for all new cpus
1611 register_cpu_notifier(&cpucache_notifier);
1614 * The reap timers are started later, with a module init call: That part
1615 * of the kernel is not yet operational.
1619 static int __init cpucache_init(void)
1624 * Register the timers that return unneeded pages to the page allocator
1626 for_each_online_cpu(cpu)
1627 start_cpu_timer(cpu);
1630 __initcall(cpucache_init);
1633 * Interface to system's page allocator. No need to hold the cache-lock.
1635 * If we requested dmaable memory, we will get it. Even if we
1636 * did not request dmaable memory, we might get it, but that
1637 * would be relatively rare and ignorable.
1639 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1647 * Nommu uses slab's for process anonymous memory allocations, and thus
1648 * requires __GFP_COMP to properly refcount higher order allocations
1650 flags |= __GFP_COMP;
1653 flags |= cachep->gfpflags;
1655 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1659 nr_pages = (1 << cachep->gfporder);
1660 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1661 add_zone_page_state(page_zone(page),
1662 NR_SLAB_RECLAIMABLE, nr_pages);
1664 add_zone_page_state(page_zone(page),
1665 NR_SLAB_UNRECLAIMABLE, nr_pages);
1666 for (i = 0; i < nr_pages; i++)
1667 __SetPageSlab(page + i);
1668 return page_address(page);
1672 * Interface to system's page release.
1674 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1676 unsigned long i = (1 << cachep->gfporder);
1677 struct page *page = virt_to_page(addr);
1678 const unsigned long nr_freed = i;
1680 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1681 sub_zone_page_state(page_zone(page),
1682 NR_SLAB_RECLAIMABLE, nr_freed);
1684 sub_zone_page_state(page_zone(page),
1685 NR_SLAB_UNRECLAIMABLE, nr_freed);
1687 BUG_ON(!PageSlab(page));
1688 __ClearPageSlab(page);
1691 if (current->reclaim_state)
1692 current->reclaim_state->reclaimed_slab += nr_freed;
1693 free_pages((unsigned long)addr, cachep->gfporder);
1696 static void kmem_rcu_free(struct rcu_head *head)
1698 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1699 struct kmem_cache *cachep = slab_rcu->cachep;
1701 kmem_freepages(cachep, slab_rcu->addr);
1702 if (OFF_SLAB(cachep))
1703 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1708 #ifdef CONFIG_DEBUG_PAGEALLOC
1709 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1710 unsigned long caller)
1712 int size = obj_size(cachep);
1714 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1716 if (size < 5 * sizeof(unsigned long))
1719 *addr++ = 0x12345678;
1721 *addr++ = smp_processor_id();
1722 size -= 3 * sizeof(unsigned long);
1724 unsigned long *sptr = &caller;
1725 unsigned long svalue;
1727 while (!kstack_end(sptr)) {
1729 if (kernel_text_address(svalue)) {
1731 size -= sizeof(unsigned long);
1732 if (size <= sizeof(unsigned long))
1738 *addr++ = 0x87654321;
1742 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1744 int size = obj_size(cachep);
1745 addr = &((char *)addr)[obj_offset(cachep)];
1747 memset(addr, val, size);
1748 *(unsigned char *)(addr + size - 1) = POISON_END;
1751 static void dump_line(char *data, int offset, int limit)
1754 unsigned char error = 0;
1757 printk(KERN_ERR "%03x:", offset);
1758 for (i = 0; i < limit; i++) {
1759 if (data[offset + i] != POISON_FREE) {
1760 error = data[offset + i];
1763 printk(" %02x", (unsigned char)data[offset + i]);
1767 if (bad_count == 1) {
1768 error ^= POISON_FREE;
1769 if (!(error & (error - 1))) {
1770 printk(KERN_ERR "Single bit error detected. Probably "
1773 printk(KERN_ERR "Run memtest86+ or a similar memory "
1776 printk(KERN_ERR "Run a memory test tool.\n");
1785 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1790 if (cachep->flags & SLAB_RED_ZONE) {
1791 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1792 *dbg_redzone1(cachep, objp),
1793 *dbg_redzone2(cachep, objp));
1796 if (cachep->flags & SLAB_STORE_USER) {
1797 printk(KERN_ERR "Last user: [<%p>]",
1798 *dbg_userword(cachep, objp));
1799 print_symbol("(%s)",
1800 (unsigned long)*dbg_userword(cachep, objp));
1803 realobj = (char *)objp + obj_offset(cachep);
1804 size = obj_size(cachep);
1805 for (i = 0; i < size && lines; i += 16, lines--) {
1808 if (i + limit > size)
1810 dump_line(realobj, i, limit);
1814 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1820 realobj = (char *)objp + obj_offset(cachep);
1821 size = obj_size(cachep);
1823 for (i = 0; i < size; i++) {
1824 char exp = POISON_FREE;
1827 if (realobj[i] != exp) {
1833 "Slab corruption: %s start=%p, len=%d\n",
1834 cachep->name, realobj, size);
1835 print_objinfo(cachep, objp, 0);
1837 /* Hexdump the affected line */
1840 if (i + limit > size)
1842 dump_line(realobj, i, limit);
1845 /* Limit to 5 lines */
1851 /* Print some data about the neighboring objects, if they
1854 struct slab *slabp = virt_to_slab(objp);
1857 objnr = obj_to_index(cachep, slabp, objp);
1859 objp = index_to_obj(cachep, slabp, objnr - 1);
1860 realobj = (char *)objp + obj_offset(cachep);
1861 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1863 print_objinfo(cachep, objp, 2);
1865 if (objnr + 1 < cachep->num) {
1866 objp = index_to_obj(cachep, slabp, objnr + 1);
1867 realobj = (char *)objp + obj_offset(cachep);
1868 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1870 print_objinfo(cachep, objp, 2);
1878 * slab_destroy_objs - destroy a slab and its objects
1879 * @cachep: cache pointer being destroyed
1880 * @slabp: slab pointer being destroyed
1882 * Call the registered destructor for each object in a slab that is being
1885 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1888 for (i = 0; i < cachep->num; i++) {
1889 void *objp = index_to_obj(cachep, slabp, i);
1891 if (cachep->flags & SLAB_POISON) {
1892 #ifdef CONFIG_DEBUG_PAGEALLOC
1893 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1895 kernel_map_pages(virt_to_page(objp),
1896 cachep->buffer_size / PAGE_SIZE, 1);
1898 check_poison_obj(cachep, objp);
1900 check_poison_obj(cachep, objp);
1903 if (cachep->flags & SLAB_RED_ZONE) {
1904 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1905 slab_error(cachep, "start of a freed object "
1907 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1908 slab_error(cachep, "end of a freed object "
1914 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1920 * slab_destroy - destroy and release all objects in a slab
1921 * @cachep: cache pointer being destroyed
1922 * @slabp: slab pointer being destroyed
1924 * Destroy all the objs in a slab, and release the mem back to the system.
1925 * Before calling the slab must have been unlinked from the cache. The
1926 * cache-lock is not held/needed.
1928 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1930 void *addr = slabp->s_mem - slabp->colouroff;
1932 slab_destroy_objs(cachep, slabp);
1933 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1934 struct slab_rcu *slab_rcu;
1936 slab_rcu = (struct slab_rcu *)slabp;
1937 slab_rcu->cachep = cachep;
1938 slab_rcu->addr = addr;
1939 call_rcu(&slab_rcu->head, kmem_rcu_free);
1941 kmem_freepages(cachep, addr);
1942 if (OFF_SLAB(cachep))
1943 kmem_cache_free(cachep->slabp_cache, slabp);
1948 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1949 * size of kmem_list3.
1951 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1955 for_each_online_node(node) {
1956 cachep->nodelists[node] = &initkmem_list3[index + node];
1957 cachep->nodelists[node]->next_reap = jiffies +
1959 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1963 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1966 struct kmem_list3 *l3;
1968 for_each_online_cpu(i)
1969 kfree(cachep->array[i]);
1971 /* NUMA: free the list3 structures */
1972 for_each_online_node(i) {
1973 l3 = cachep->nodelists[i];
1976 free_alien_cache(l3->alien);
1980 kmem_cache_free(&cache_cache, cachep);
1985 * calculate_slab_order - calculate size (page order) of slabs
1986 * @cachep: pointer to the cache that is being created
1987 * @size: size of objects to be created in this cache.
1988 * @align: required alignment for the objects.
1989 * @flags: slab allocation flags
1991 * Also calculates the number of objects per slab.
1993 * This could be made much more intelligent. For now, try to avoid using
1994 * high order pages for slabs. When the gfp() functions are more friendly
1995 * towards high-order requests, this should be changed.
1997 static size_t calculate_slab_order(struct kmem_cache *cachep,
1998 size_t size, size_t align, unsigned long flags)
2000 unsigned long offslab_limit;
2001 size_t left_over = 0;
2004 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
2008 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2012 if (flags & CFLGS_OFF_SLAB) {
2014 * Max number of objs-per-slab for caches which
2015 * use off-slab slabs. Needed to avoid a possible
2016 * looping condition in cache_grow().
2018 offslab_limit = size - sizeof(struct slab);
2019 offslab_limit /= sizeof(kmem_bufctl_t);
2021 if (num > offslab_limit)
2025 /* Found something acceptable - save it away */
2027 cachep->gfporder = gfporder;
2028 left_over = remainder;
2031 * A VFS-reclaimable slab tends to have most allocations
2032 * as GFP_NOFS and we really don't want to have to be allocating
2033 * higher-order pages when we are unable to shrink dcache.
2035 if (flags & SLAB_RECLAIM_ACCOUNT)
2039 * Large number of objects is good, but very large slabs are
2040 * currently bad for the gfp()s.
2042 if (gfporder >= slab_break_gfp_order)
2046 * Acceptable internal fragmentation?
2048 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2054 static int setup_cpu_cache(struct kmem_cache *cachep)
2056 if (g_cpucache_up == FULL)
2057 return enable_cpucache(cachep);
2059 if (g_cpucache_up == NONE) {
2061 * Note: the first kmem_cache_create must create the cache
2062 * that's used by kmalloc(24), otherwise the creation of
2063 * further caches will BUG().
2065 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2068 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2069 * the first cache, then we need to set up all its list3s,
2070 * otherwise the creation of further caches will BUG().
2072 set_up_list3s(cachep, SIZE_AC);
2073 if (INDEX_AC == INDEX_L3)
2074 g_cpucache_up = PARTIAL_L3;
2076 g_cpucache_up = PARTIAL_AC;
2078 cachep->array[smp_processor_id()] =
2079 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2081 if (g_cpucache_up == PARTIAL_AC) {
2082 set_up_list3s(cachep, SIZE_L3);
2083 g_cpucache_up = PARTIAL_L3;
2086 for_each_online_node(node) {
2087 cachep->nodelists[node] =
2088 kmalloc_node(sizeof(struct kmem_list3),
2090 BUG_ON(!cachep->nodelists[node]);
2091 kmem_list3_init(cachep->nodelists[node]);
2095 cachep->nodelists[numa_node_id()]->next_reap =
2096 jiffies + REAPTIMEOUT_LIST3 +
2097 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2099 cpu_cache_get(cachep)->avail = 0;
2100 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2101 cpu_cache_get(cachep)->batchcount = 1;
2102 cpu_cache_get(cachep)->touched = 0;
2103 cachep->batchcount = 1;
2104 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2109 * kmem_cache_create - Create a cache.
2110 * @name: A string which is used in /proc/slabinfo to identify this cache.
2111 * @size: The size of objects to be created in this cache.
2112 * @align: The required alignment for the objects.
2113 * @flags: SLAB flags
2114 * @ctor: A constructor for the objects.
2115 * @dtor: A destructor for the objects (not implemented anymore).
2117 * Returns a ptr to the cache on success, NULL on failure.
2118 * Cannot be called within a int, but can be interrupted.
2119 * The @ctor is run when new pages are allocated by the cache
2120 * and the @dtor is run before the pages are handed back.
2122 * @name must be valid until the cache is destroyed. This implies that
2123 * the module calling this has to destroy the cache before getting unloaded.
2127 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2128 * to catch references to uninitialised memory.
2130 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2131 * for buffer overruns.
2133 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2134 * cacheline. This can be beneficial if you're counting cycles as closely
2138 kmem_cache_create (const char *name, size_t size, size_t align,
2139 unsigned long flags,
2140 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2141 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2143 size_t left_over, slab_size, ralign;
2144 struct kmem_cache *cachep = NULL, *pc;
2147 * Sanity checks... these are all serious usage bugs.
2149 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2150 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || dtor) {
2151 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2157 * We use cache_chain_mutex to ensure a consistent view of
2158 * cpu_online_map as well. Please see cpuup_callback
2160 mutex_lock(&cache_chain_mutex);
2162 list_for_each_entry(pc, &cache_chain, next) {
2167 * This happens when the module gets unloaded and doesn't
2168 * destroy its slab cache and no-one else reuses the vmalloc
2169 * area of the module. Print a warning.
2171 res = probe_kernel_address(pc->name, tmp);
2174 "SLAB: cache with size %d has lost its name\n",
2179 if (!strcmp(pc->name, name)) {
2181 "kmem_cache_create: duplicate cache %s\n", name);
2188 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2191 * Enable redzoning and last user accounting, except for caches with
2192 * large objects, if the increased size would increase the object size
2193 * above the next power of two: caches with object sizes just above a
2194 * power of two have a significant amount of internal fragmentation.
2196 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2197 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2198 if (!(flags & SLAB_DESTROY_BY_RCU))
2199 flags |= SLAB_POISON;
2201 if (flags & SLAB_DESTROY_BY_RCU)
2202 BUG_ON(flags & SLAB_POISON);
2205 * Always checks flags, a caller might be expecting debug support which
2208 BUG_ON(flags & ~CREATE_MASK);
2211 * Check that size is in terms of words. This is needed to avoid
2212 * unaligned accesses for some archs when redzoning is used, and makes
2213 * sure any on-slab bufctl's are also correctly aligned.
2215 if (size & (BYTES_PER_WORD - 1)) {
2216 size += (BYTES_PER_WORD - 1);
2217 size &= ~(BYTES_PER_WORD - 1);
2220 /* calculate the final buffer alignment: */
2222 /* 1) arch recommendation: can be overridden for debug */
2223 if (flags & SLAB_HWCACHE_ALIGN) {
2225 * Default alignment: as specified by the arch code. Except if
2226 * an object is really small, then squeeze multiple objects into
2229 ralign = cache_line_size();
2230 while (size <= ralign / 2)
2233 ralign = BYTES_PER_WORD;
2237 * Redzoning and user store require word alignment. Note this will be
2238 * overridden by architecture or caller mandated alignment if either
2239 * is greater than BYTES_PER_WORD.
2241 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2242 ralign = __alignof__(unsigned long long);
2244 /* 2) arch mandated alignment */
2245 if (ralign < ARCH_SLAB_MINALIGN) {
2246 ralign = ARCH_SLAB_MINALIGN;
2248 /* 3) caller mandated alignment */
2249 if (ralign < align) {
2252 /* disable debug if necessary */
2253 if (ralign > __alignof__(unsigned long long))
2254 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2260 /* Get cache's description obj. */
2261 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2266 cachep->obj_size = size;
2269 * Both debugging options require word-alignment which is calculated
2272 if (flags & SLAB_RED_ZONE) {
2273 /* add space for red zone words */
2274 cachep->obj_offset += sizeof(unsigned long long);
2275 size += 2 * sizeof(unsigned long long);
2277 if (flags & SLAB_STORE_USER) {
2278 /* user store requires one word storage behind the end of
2281 size += BYTES_PER_WORD;
2283 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2284 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2285 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2286 cachep->obj_offset += PAGE_SIZE - size;
2293 * Determine if the slab management is 'on' or 'off' slab.
2294 * (bootstrapping cannot cope with offslab caches so don't do
2297 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2299 * Size is large, assume best to place the slab management obj
2300 * off-slab (should allow better packing of objs).
2302 flags |= CFLGS_OFF_SLAB;
2304 size = ALIGN(size, align);
2306 left_over = calculate_slab_order(cachep, size, align, flags);
2310 "kmem_cache_create: couldn't create cache %s.\n", name);
2311 kmem_cache_free(&cache_cache, cachep);
2315 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2316 + sizeof(struct slab), align);
2319 * If the slab has been placed off-slab, and we have enough space then
2320 * move it on-slab. This is at the expense of any extra colouring.
2322 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2323 flags &= ~CFLGS_OFF_SLAB;
2324 left_over -= slab_size;
2327 if (flags & CFLGS_OFF_SLAB) {
2328 /* really off slab. No need for manual alignment */
2330 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2333 cachep->colour_off = cache_line_size();
2334 /* Offset must be a multiple of the alignment. */
2335 if (cachep->colour_off < align)
2336 cachep->colour_off = align;
2337 cachep->colour = left_over / cachep->colour_off;
2338 cachep->slab_size = slab_size;
2339 cachep->flags = flags;
2340 cachep->gfpflags = 0;
2341 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2342 cachep->gfpflags |= GFP_DMA;
2343 cachep->buffer_size = size;
2344 cachep->reciprocal_buffer_size = reciprocal_value(size);
2346 if (flags & CFLGS_OFF_SLAB) {
2347 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2349 * This is a possibility for one of the malloc_sizes caches.
2350 * But since we go off slab only for object size greater than
2351 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2352 * this should not happen at all.
2353 * But leave a BUG_ON for some lucky dude.
2355 BUG_ON(!cachep->slabp_cache);
2357 cachep->ctor = ctor;
2358 cachep->name = name;
2360 if (setup_cpu_cache(cachep)) {
2361 __kmem_cache_destroy(cachep);
2366 /* cache setup completed, link it into the list */
2367 list_add(&cachep->next, &cache_chain);
2369 if (!cachep && (flags & SLAB_PANIC))
2370 panic("kmem_cache_create(): failed to create slab `%s'\n",
2372 mutex_unlock(&cache_chain_mutex);
2375 EXPORT_SYMBOL(kmem_cache_create);
2378 static void check_irq_off(void)
2380 BUG_ON(!irqs_disabled());
2383 static void check_irq_on(void)
2385 BUG_ON(irqs_disabled());
2388 static void check_spinlock_acquired(struct kmem_cache *cachep)
2392 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2396 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2400 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2405 #define check_irq_off() do { } while(0)
2406 #define check_irq_on() do { } while(0)
2407 #define check_spinlock_acquired(x) do { } while(0)
2408 #define check_spinlock_acquired_node(x, y) do { } while(0)
2411 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2412 struct array_cache *ac,
2413 int force, int node);
2415 static void do_drain(void *arg)
2417 struct kmem_cache *cachep = arg;
2418 struct array_cache *ac;
2419 int node = numa_node_id();
2422 ac = cpu_cache_get(cachep);
2423 spin_lock(&cachep->nodelists[node]->list_lock);
2424 free_block(cachep, ac->entry, ac->avail, node);
2425 spin_unlock(&cachep->nodelists[node]->list_lock);
2429 static void drain_cpu_caches(struct kmem_cache *cachep)
2431 struct kmem_list3 *l3;
2434 on_each_cpu(do_drain, cachep, 1, 1);
2436 for_each_online_node(node) {
2437 l3 = cachep->nodelists[node];
2438 if (l3 && l3->alien)
2439 drain_alien_cache(cachep, l3->alien);
2442 for_each_online_node(node) {
2443 l3 = cachep->nodelists[node];
2445 drain_array(cachep, l3, l3->shared, 1, node);
2450 * Remove slabs from the list of free slabs.
2451 * Specify the number of slabs to drain in tofree.
2453 * Returns the actual number of slabs released.
2455 static int drain_freelist(struct kmem_cache *cache,
2456 struct kmem_list3 *l3, int tofree)
2458 struct list_head *p;
2463 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2465 spin_lock_irq(&l3->list_lock);
2466 p = l3->slabs_free.prev;
2467 if (p == &l3->slabs_free) {
2468 spin_unlock_irq(&l3->list_lock);
2472 slabp = list_entry(p, struct slab, list);
2474 BUG_ON(slabp->inuse);
2476 list_del(&slabp->list);
2478 * Safe to drop the lock. The slab is no longer linked
2481 l3->free_objects -= cache->num;
2482 spin_unlock_irq(&l3->list_lock);
2483 slab_destroy(cache, slabp);
2490 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2491 static int __cache_shrink(struct kmem_cache *cachep)
2494 struct kmem_list3 *l3;
2496 drain_cpu_caches(cachep);
2499 for_each_online_node(i) {
2500 l3 = cachep->nodelists[i];
2504 drain_freelist(cachep, l3, l3->free_objects);
2506 ret += !list_empty(&l3->slabs_full) ||
2507 !list_empty(&l3->slabs_partial);
2509 return (ret ? 1 : 0);
2513 * kmem_cache_shrink - Shrink a cache.
2514 * @cachep: The cache to shrink.
2516 * Releases as many slabs as possible for a cache.
2517 * To help debugging, a zero exit status indicates all slabs were released.
2519 int kmem_cache_shrink(struct kmem_cache *cachep)
2522 BUG_ON(!cachep || in_interrupt());
2524 mutex_lock(&cache_chain_mutex);
2525 ret = __cache_shrink(cachep);
2526 mutex_unlock(&cache_chain_mutex);
2529 EXPORT_SYMBOL(kmem_cache_shrink);
2532 * kmem_cache_destroy - delete a cache
2533 * @cachep: the cache to destroy
2535 * Remove a &struct kmem_cache object from the slab cache.
2537 * It is expected this function will be called by a module when it is
2538 * unloaded. This will remove the cache completely, and avoid a duplicate
2539 * cache being allocated each time a module is loaded and unloaded, if the
2540 * module doesn't have persistent in-kernel storage across loads and unloads.
2542 * The cache must be empty before calling this function.
2544 * The caller must guarantee that noone will allocate memory from the cache
2545 * during the kmem_cache_destroy().
2547 void kmem_cache_destroy(struct kmem_cache *cachep)
2549 BUG_ON(!cachep || in_interrupt());
2551 /* Find the cache in the chain of caches. */
2552 mutex_lock(&cache_chain_mutex);
2554 * the chain is never empty, cache_cache is never destroyed
2556 list_del(&cachep->next);
2557 if (__cache_shrink(cachep)) {
2558 slab_error(cachep, "Can't free all objects");
2559 list_add(&cachep->next, &cache_chain);
2560 mutex_unlock(&cache_chain_mutex);
2564 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2567 __kmem_cache_destroy(cachep);
2568 mutex_unlock(&cache_chain_mutex);
2570 EXPORT_SYMBOL(kmem_cache_destroy);
2573 * Get the memory for a slab management obj.
2574 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2575 * always come from malloc_sizes caches. The slab descriptor cannot
2576 * come from the same cache which is getting created because,
2577 * when we are searching for an appropriate cache for these
2578 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2579 * If we are creating a malloc_sizes cache here it would not be visible to
2580 * kmem_find_general_cachep till the initialization is complete.
2581 * Hence we cannot have slabp_cache same as the original cache.
2583 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2584 int colour_off, gfp_t local_flags,
2589 if (OFF_SLAB(cachep)) {
2590 /* Slab management obj is off-slab. */
2591 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2592 local_flags & ~GFP_THISNODE, nodeid);
2596 slabp = objp + colour_off;
2597 colour_off += cachep->slab_size;
2600 slabp->colouroff = colour_off;
2601 slabp->s_mem = objp + colour_off;
2602 slabp->nodeid = nodeid;
2606 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2608 return (kmem_bufctl_t *) (slabp + 1);
2611 static void cache_init_objs(struct kmem_cache *cachep,
2612 struct slab *slabp, unsigned long ctor_flags)
2616 for (i = 0; i < cachep->num; i++) {
2617 void *objp = index_to_obj(cachep, slabp, i);
2619 /* need to poison the objs? */
2620 if (cachep->flags & SLAB_POISON)
2621 poison_obj(cachep, objp, POISON_FREE);
2622 if (cachep->flags & SLAB_STORE_USER)
2623 *dbg_userword(cachep, objp) = NULL;
2625 if (cachep->flags & SLAB_RED_ZONE) {
2626 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2627 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2630 * Constructors are not allowed to allocate memory from the same
2631 * cache which they are a constructor for. Otherwise, deadlock.
2632 * They must also be threaded.
2634 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2635 cachep->ctor(objp + obj_offset(cachep), cachep,
2638 if (cachep->flags & SLAB_RED_ZONE) {
2639 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2640 slab_error(cachep, "constructor overwrote the"
2641 " end of an object");
2642 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2643 slab_error(cachep, "constructor overwrote the"
2644 " start of an object");
2646 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2647 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2648 kernel_map_pages(virt_to_page(objp),
2649 cachep->buffer_size / PAGE_SIZE, 0);
2652 cachep->ctor(objp, cachep, ctor_flags);
2654 slab_bufctl(slabp)[i] = i + 1;
2656 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2660 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2662 if (CONFIG_ZONE_DMA_FLAG) {
2663 if (flags & GFP_DMA)
2664 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2666 BUG_ON(cachep->gfpflags & GFP_DMA);
2670 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2673 void *objp = index_to_obj(cachep, slabp, slabp->free);
2677 next = slab_bufctl(slabp)[slabp->free];
2679 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2680 WARN_ON(slabp->nodeid != nodeid);
2687 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2688 void *objp, int nodeid)
2690 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2693 /* Verify that the slab belongs to the intended node */
2694 WARN_ON(slabp->nodeid != nodeid);
2696 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2697 printk(KERN_ERR "slab: double free detected in cache "
2698 "'%s', objp %p\n", cachep->name, objp);
2702 slab_bufctl(slabp)[objnr] = slabp->free;
2703 slabp->free = objnr;
2708 * Map pages beginning at addr to the given cache and slab. This is required
2709 * for the slab allocator to be able to lookup the cache and slab of a
2710 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2712 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2718 page = virt_to_page(addr);
2721 if (likely(!PageCompound(page)))
2722 nr_pages <<= cache->gfporder;
2725 page_set_cache(page, cache);
2726 page_set_slab(page, slab);
2728 } while (--nr_pages);
2732 * Grow (by 1) the number of slabs within a cache. This is called by
2733 * kmem_cache_alloc() when there are no active objs left in a cache.
2735 static int cache_grow(struct kmem_cache *cachep,
2736 gfp_t flags, int nodeid, void *objp)
2741 unsigned long ctor_flags;
2742 struct kmem_list3 *l3;
2745 * Be lazy and only check for valid flags here, keeping it out of the
2746 * critical path in kmem_cache_alloc().
2748 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
2750 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2751 local_flags = (flags & GFP_LEVEL_MASK);
2752 /* Take the l3 list lock to change the colour_next on this node */
2754 l3 = cachep->nodelists[nodeid];
2755 spin_lock(&l3->list_lock);
2757 /* Get colour for the slab, and cal the next value. */
2758 offset = l3->colour_next;
2760 if (l3->colour_next >= cachep->colour)
2761 l3->colour_next = 0;
2762 spin_unlock(&l3->list_lock);
2764 offset *= cachep->colour_off;
2766 if (local_flags & __GFP_WAIT)
2770 * The test for missing atomic flag is performed here, rather than
2771 * the more obvious place, simply to reduce the critical path length
2772 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2773 * will eventually be caught here (where it matters).
2775 kmem_flagcheck(cachep, flags);
2778 * Get mem for the objs. Attempt to allocate a physical page from
2782 objp = kmem_getpages(cachep, flags, nodeid);
2786 /* Get slab management. */
2787 slabp = alloc_slabmgmt(cachep, objp, offset,
2788 local_flags & ~GFP_THISNODE, nodeid);
2792 slabp->nodeid = nodeid;
2793 slab_map_pages(cachep, slabp, objp);
2795 cache_init_objs(cachep, slabp, ctor_flags);
2797 if (local_flags & __GFP_WAIT)
2798 local_irq_disable();
2800 spin_lock(&l3->list_lock);
2802 /* Make slab active. */
2803 list_add_tail(&slabp->list, &(l3->slabs_free));
2804 STATS_INC_GROWN(cachep);
2805 l3->free_objects += cachep->num;
2806 spin_unlock(&l3->list_lock);
2809 kmem_freepages(cachep, objp);
2811 if (local_flags & __GFP_WAIT)
2812 local_irq_disable();
2819 * Perform extra freeing checks:
2820 * - detect bad pointers.
2821 * - POISON/RED_ZONE checking
2823 static void kfree_debugcheck(const void *objp)
2825 if (!virt_addr_valid(objp)) {
2826 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2827 (unsigned long)objp);
2832 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2834 unsigned long long redzone1, redzone2;
2836 redzone1 = *dbg_redzone1(cache, obj);
2837 redzone2 = *dbg_redzone2(cache, obj);
2842 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2845 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2846 slab_error(cache, "double free detected");
2848 slab_error(cache, "memory outside object was overwritten");
2850 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2851 obj, redzone1, redzone2);
2854 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2861 objp -= obj_offset(cachep);
2862 kfree_debugcheck(objp);
2863 page = virt_to_head_page(objp);
2865 slabp = page_get_slab(page);
2867 if (cachep->flags & SLAB_RED_ZONE) {
2868 verify_redzone_free(cachep, objp);
2869 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2870 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2872 if (cachep->flags & SLAB_STORE_USER)
2873 *dbg_userword(cachep, objp) = caller;
2875 objnr = obj_to_index(cachep, slabp, objp);
2877 BUG_ON(objnr >= cachep->num);
2878 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2880 #ifdef CONFIG_DEBUG_SLAB_LEAK
2881 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2883 if (cachep->flags & SLAB_POISON) {
2884 #ifdef CONFIG_DEBUG_PAGEALLOC
2885 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2886 store_stackinfo(cachep, objp, (unsigned long)caller);
2887 kernel_map_pages(virt_to_page(objp),
2888 cachep->buffer_size / PAGE_SIZE, 0);
2890 poison_obj(cachep, objp, POISON_FREE);
2893 poison_obj(cachep, objp, POISON_FREE);
2899 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2904 /* Check slab's freelist to see if this obj is there. */
2905 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2907 if (entries > cachep->num || i >= cachep->num)
2910 if (entries != cachep->num - slabp->inuse) {
2912 printk(KERN_ERR "slab: Internal list corruption detected in "
2913 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2914 cachep->name, cachep->num, slabp, slabp->inuse);
2916 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2919 printk("\n%03x:", i);
2920 printk(" %02x", ((unsigned char *)slabp)[i]);
2927 #define kfree_debugcheck(x) do { } while(0)
2928 #define cache_free_debugcheck(x,objp,z) (objp)
2929 #define check_slabp(x,y) do { } while(0)
2932 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2935 struct kmem_list3 *l3;
2936 struct array_cache *ac;
2939 node = numa_node_id();
2942 ac = cpu_cache_get(cachep);
2944 batchcount = ac->batchcount;
2945 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2947 * If there was little recent activity on this cache, then
2948 * perform only a partial refill. Otherwise we could generate
2951 batchcount = BATCHREFILL_LIMIT;
2953 l3 = cachep->nodelists[node];
2955 BUG_ON(ac->avail > 0 || !l3);
2956 spin_lock(&l3->list_lock);
2958 /* See if we can refill from the shared array */
2959 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2962 while (batchcount > 0) {
2963 struct list_head *entry;
2965 /* Get slab alloc is to come from. */
2966 entry = l3->slabs_partial.next;
2967 if (entry == &l3->slabs_partial) {
2968 l3->free_touched = 1;
2969 entry = l3->slabs_free.next;
2970 if (entry == &l3->slabs_free)
2974 slabp = list_entry(entry, struct slab, list);
2975 check_slabp(cachep, slabp);
2976 check_spinlock_acquired(cachep);
2979 * The slab was either on partial or free list so
2980 * there must be at least one object available for
2983 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2985 while (slabp->inuse < cachep->num && batchcount--) {
2986 STATS_INC_ALLOCED(cachep);
2987 STATS_INC_ACTIVE(cachep);
2988 STATS_SET_HIGH(cachep);
2990 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2993 check_slabp(cachep, slabp);
2995 /* move slabp to correct slabp list: */
2996 list_del(&slabp->list);
2997 if (slabp->free == BUFCTL_END)
2998 list_add(&slabp->list, &l3->slabs_full);
3000 list_add(&slabp->list, &l3->slabs_partial);
3004 l3->free_objects -= ac->avail;
3006 spin_unlock(&l3->list_lock);
3008 if (unlikely(!ac->avail)) {
3010 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3012 /* cache_grow can reenable interrupts, then ac could change. */
3013 ac = cpu_cache_get(cachep);
3014 if (!x && ac->avail == 0) /* no objects in sight? abort */
3017 if (!ac->avail) /* objects refilled by interrupt? */
3021 return ac->entry[--ac->avail];
3024 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3027 might_sleep_if(flags & __GFP_WAIT);
3029 kmem_flagcheck(cachep, flags);
3034 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3035 gfp_t flags, void *objp, void *caller)
3039 if (cachep->flags & SLAB_POISON) {
3040 #ifdef CONFIG_DEBUG_PAGEALLOC
3041 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3042 kernel_map_pages(virt_to_page(objp),
3043 cachep->buffer_size / PAGE_SIZE, 1);
3045 check_poison_obj(cachep, objp);
3047 check_poison_obj(cachep, objp);
3049 poison_obj(cachep, objp, POISON_INUSE);
3051 if (cachep->flags & SLAB_STORE_USER)
3052 *dbg_userword(cachep, objp) = caller;
3054 if (cachep->flags & SLAB_RED_ZONE) {
3055 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3056 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3057 slab_error(cachep, "double free, or memory outside"
3058 " object was overwritten");
3060 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3061 objp, *dbg_redzone1(cachep, objp),
3062 *dbg_redzone2(cachep, objp));
3064 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3065 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3067 #ifdef CONFIG_DEBUG_SLAB_LEAK
3072 slabp = page_get_slab(virt_to_head_page(objp));
3073 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3074 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3077 objp += obj_offset(cachep);
3078 if (cachep->ctor && cachep->flags & SLAB_POISON)
3079 cachep->ctor(objp, cachep, SLAB_CTOR_CONSTRUCTOR);
3080 #if ARCH_SLAB_MINALIGN
3081 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3082 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3083 objp, ARCH_SLAB_MINALIGN);
3089 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3092 #ifdef CONFIG_FAILSLAB
3094 static struct failslab_attr {
3096 struct fault_attr attr;
3098 u32 ignore_gfp_wait;
3099 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3100 struct dentry *ignore_gfp_wait_file;
3104 .attr = FAULT_ATTR_INITIALIZER,
3105 .ignore_gfp_wait = 1,
3108 static int __init setup_failslab(char *str)
3110 return setup_fault_attr(&failslab.attr, str);
3112 __setup("failslab=", setup_failslab);
3114 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3116 if (cachep == &cache_cache)
3118 if (flags & __GFP_NOFAIL)
3120 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3123 return should_fail(&failslab.attr, obj_size(cachep));
3126 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3128 static int __init failslab_debugfs(void)
3130 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3134 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3137 dir = failslab.attr.dentries.dir;
3139 failslab.ignore_gfp_wait_file =
3140 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3141 &failslab.ignore_gfp_wait);
3143 if (!failslab.ignore_gfp_wait_file) {
3145 debugfs_remove(failslab.ignore_gfp_wait_file);
3146 cleanup_fault_attr_dentries(&failslab.attr);
3152 late_initcall(failslab_debugfs);
3154 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3156 #else /* CONFIG_FAILSLAB */
3158 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3163 #endif /* CONFIG_FAILSLAB */
3165 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3168 struct array_cache *ac;
3172 ac = cpu_cache_get(cachep);
3173 if (likely(ac->avail)) {
3174 STATS_INC_ALLOCHIT(cachep);
3176 objp = ac->entry[--ac->avail];
3178 STATS_INC_ALLOCMISS(cachep);
3179 objp = cache_alloc_refill(cachep, flags);
3186 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3188 * If we are in_interrupt, then process context, including cpusets and
3189 * mempolicy, may not apply and should not be used for allocation policy.
3191 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3193 int nid_alloc, nid_here;
3195 if (in_interrupt() || (flags & __GFP_THISNODE))
3197 nid_alloc = nid_here = numa_node_id();
3198 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3199 nid_alloc = cpuset_mem_spread_node();
3200 else if (current->mempolicy)
3201 nid_alloc = slab_node(current->mempolicy);
3202 if (nid_alloc != nid_here)
3203 return ____cache_alloc_node(cachep, flags, nid_alloc);
3208 * Fallback function if there was no memory available and no objects on a
3209 * certain node and fall back is permitted. First we scan all the
3210 * available nodelists for available objects. If that fails then we
3211 * perform an allocation without specifying a node. This allows the page
3212 * allocator to do its reclaim / fallback magic. We then insert the
3213 * slab into the proper nodelist and then allocate from it.
3215 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3217 struct zonelist *zonelist;
3223 if (flags & __GFP_THISNODE)
3226 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3227 ->node_zonelists[gfp_zone(flags)];
3228 local_flags = (flags & GFP_LEVEL_MASK);
3232 * Look through allowed nodes for objects available
3233 * from existing per node queues.
3235 for (z = zonelist->zones; *z && !obj; z++) {
3236 nid = zone_to_nid(*z);
3238 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3239 cache->nodelists[nid] &&
3240 cache->nodelists[nid]->free_objects)
3241 obj = ____cache_alloc_node(cache,
3242 flags | GFP_THISNODE, nid);
3247 * This allocation will be performed within the constraints
3248 * of the current cpuset / memory policy requirements.
3249 * We may trigger various forms of reclaim on the allowed
3250 * set and go into memory reserves if necessary.
3252 if (local_flags & __GFP_WAIT)
3254 kmem_flagcheck(cache, flags);
3255 obj = kmem_getpages(cache, flags, -1);
3256 if (local_flags & __GFP_WAIT)
3257 local_irq_disable();
3260 * Insert into the appropriate per node queues
3262 nid = page_to_nid(virt_to_page(obj));
3263 if (cache_grow(cache, flags, nid, obj)) {
3264 obj = ____cache_alloc_node(cache,
3265 flags | GFP_THISNODE, nid);
3268 * Another processor may allocate the
3269 * objects in the slab since we are
3270 * not holding any locks.
3274 /* cache_grow already freed obj */
3283 * A interface to enable slab creation on nodeid
3285 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3288 struct list_head *entry;
3290 struct kmem_list3 *l3;
3294 l3 = cachep->nodelists[nodeid];
3299 spin_lock(&l3->list_lock);
3300 entry = l3->slabs_partial.next;
3301 if (entry == &l3->slabs_partial) {
3302 l3->free_touched = 1;
3303 entry = l3->slabs_free.next;
3304 if (entry == &l3->slabs_free)
3308 slabp = list_entry(entry, struct slab, list);
3309 check_spinlock_acquired_node(cachep, nodeid);
3310 check_slabp(cachep, slabp);
3312 STATS_INC_NODEALLOCS(cachep);
3313 STATS_INC_ACTIVE(cachep);
3314 STATS_SET_HIGH(cachep);
3316 BUG_ON(slabp->inuse == cachep->num);
3318 obj = slab_get_obj(cachep, slabp, nodeid);
3319 check_slabp(cachep, slabp);
3321 /* move slabp to correct slabp list: */
3322 list_del(&slabp->list);
3324 if (slabp->free == BUFCTL_END)
3325 list_add(&slabp->list, &l3->slabs_full);
3327 list_add(&slabp->list, &l3->slabs_partial);
3329 spin_unlock(&l3->list_lock);
3333 spin_unlock(&l3->list_lock);
3334 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3338 return fallback_alloc(cachep, flags);
3345 * kmem_cache_alloc_node - Allocate an object on the specified node
3346 * @cachep: The cache to allocate from.
3347 * @flags: See kmalloc().
3348 * @nodeid: node number of the target node.
3349 * @caller: return address of caller, used for debug information
3351 * Identical to kmem_cache_alloc but it will allocate memory on the given
3352 * node, which can improve the performance for cpu bound structures.
3354 * Fallback to other node is possible if __GFP_THISNODE is not set.
3356 static __always_inline void *
3357 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3360 unsigned long save_flags;
3363 if (should_failslab(cachep, flags))
3366 cache_alloc_debugcheck_before(cachep, flags);
3367 local_irq_save(save_flags);
3369 if (unlikely(nodeid == -1))
3370 nodeid = numa_node_id();
3372 if (unlikely(!cachep->nodelists[nodeid])) {
3373 /* Node not bootstrapped yet */
3374 ptr = fallback_alloc(cachep, flags);
3378 if (nodeid == numa_node_id()) {
3380 * Use the locally cached objects if possible.
3381 * However ____cache_alloc does not allow fallback
3382 * to other nodes. It may fail while we still have
3383 * objects on other nodes available.
3385 ptr = ____cache_alloc(cachep, flags);
3389 /* ___cache_alloc_node can fall back to other nodes */
3390 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3392 local_irq_restore(save_flags);
3393 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3398 static __always_inline void *
3399 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3403 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3404 objp = alternate_node_alloc(cache, flags);
3408 objp = ____cache_alloc(cache, flags);
3411 * We may just have run out of memory on the local node.
3412 * ____cache_alloc_node() knows how to locate memory on other nodes
3415 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3422 static __always_inline void *
3423 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3425 return ____cache_alloc(cachep, flags);
3428 #endif /* CONFIG_NUMA */
3430 static __always_inline void *
3431 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3433 unsigned long save_flags;
3436 if (should_failslab(cachep, flags))
3439 cache_alloc_debugcheck_before(cachep, flags);
3440 local_irq_save(save_flags);
3441 objp = __do_cache_alloc(cachep, flags);
3442 local_irq_restore(save_flags);
3443 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3450 * Caller needs to acquire correct kmem_list's list_lock
3452 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3456 struct kmem_list3 *l3;
3458 for (i = 0; i < nr_objects; i++) {
3459 void *objp = objpp[i];
3462 slabp = virt_to_slab(objp);
3463 l3 = cachep->nodelists[node];
3464 list_del(&slabp->list);
3465 check_spinlock_acquired_node(cachep, node);
3466 check_slabp(cachep, slabp);
3467 slab_put_obj(cachep, slabp, objp, node);
3468 STATS_DEC_ACTIVE(cachep);
3470 check_slabp(cachep, slabp);
3472 /* fixup slab chains */
3473 if (slabp->inuse == 0) {
3474 if (l3->free_objects > l3->free_limit) {
3475 l3->free_objects -= cachep->num;
3476 /* No need to drop any previously held
3477 * lock here, even if we have a off-slab slab
3478 * descriptor it is guaranteed to come from
3479 * a different cache, refer to comments before
3482 slab_destroy(cachep, slabp);
3484 list_add(&slabp->list, &l3->slabs_free);
3487 /* Unconditionally move a slab to the end of the
3488 * partial list on free - maximum time for the
3489 * other objects to be freed, too.
3491 list_add_tail(&slabp->list, &l3->slabs_partial);
3496 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3499 struct kmem_list3 *l3;
3500 int node = numa_node_id();
3502 batchcount = ac->batchcount;
3504 BUG_ON(!batchcount || batchcount > ac->avail);
3507 l3 = cachep->nodelists[node];
3508 spin_lock(&l3->list_lock);
3510 struct array_cache *shared_array = l3->shared;
3511 int max = shared_array->limit - shared_array->avail;
3513 if (batchcount > max)
3515 memcpy(&(shared_array->entry[shared_array->avail]),
3516 ac->entry, sizeof(void *) * batchcount);
3517 shared_array->avail += batchcount;
3522 free_block(cachep, ac->entry, batchcount, node);
3527 struct list_head *p;
3529 p = l3->slabs_free.next;
3530 while (p != &(l3->slabs_free)) {
3533 slabp = list_entry(p, struct slab, list);
3534 BUG_ON(slabp->inuse);
3539 STATS_SET_FREEABLE(cachep, i);
3542 spin_unlock(&l3->list_lock);
3543 ac->avail -= batchcount;
3544 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3548 * Release an obj back to its cache. If the obj has a constructed state, it must
3549 * be in this state _before_ it is released. Called with disabled ints.
3551 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3553 struct array_cache *ac = cpu_cache_get(cachep);
3556 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3558 if (use_alien_caches && cache_free_alien(cachep, objp))
3561 if (likely(ac->avail < ac->limit)) {
3562 STATS_INC_FREEHIT(cachep);
3563 ac->entry[ac->avail++] = objp;
3566 STATS_INC_FREEMISS(cachep);
3567 cache_flusharray(cachep, ac);
3568 ac->entry[ac->avail++] = objp;
3573 * kmem_cache_alloc - Allocate an object
3574 * @cachep: The cache to allocate from.
3575 * @flags: See kmalloc().
3577 * Allocate an object from this cache. The flags are only relevant
3578 * if the cache has no available objects.
3580 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3582 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3584 EXPORT_SYMBOL(kmem_cache_alloc);
3587 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3588 * @cache: The cache to allocate from.
3589 * @flags: See kmalloc().
3591 * Allocate an object from this cache and set the allocated memory to zero.
3592 * The flags are only relevant if the cache has no available objects.
3594 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3596 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3598 memset(ret, 0, obj_size(cache));
3601 EXPORT_SYMBOL(kmem_cache_zalloc);
3604 * kmem_ptr_validate - check if an untrusted pointer might
3606 * @cachep: the cache we're checking against
3607 * @ptr: pointer to validate
3609 * This verifies that the untrusted pointer looks sane:
3610 * it is _not_ a guarantee that the pointer is actually
3611 * part of the slab cache in question, but it at least
3612 * validates that the pointer can be dereferenced and
3613 * looks half-way sane.
3615 * Currently only used for dentry validation.
3617 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3619 unsigned long addr = (unsigned long)ptr;
3620 unsigned long min_addr = PAGE_OFFSET;
3621 unsigned long align_mask = BYTES_PER_WORD - 1;
3622 unsigned long size = cachep->buffer_size;
3625 if (unlikely(addr < min_addr))
3627 if (unlikely(addr > (unsigned long)high_memory - size))
3629 if (unlikely(addr & align_mask))
3631 if (unlikely(!kern_addr_valid(addr)))
3633 if (unlikely(!kern_addr_valid(addr + size - 1)))
3635 page = virt_to_page(ptr);
3636 if (unlikely(!PageSlab(page)))
3638 if (unlikely(page_get_cache(page) != cachep))
3646 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3648 return __cache_alloc_node(cachep, flags, nodeid,
3649 __builtin_return_address(0));
3651 EXPORT_SYMBOL(kmem_cache_alloc_node);
3653 static __always_inline void *
3654 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3656 struct kmem_cache *cachep;
3658 cachep = kmem_find_general_cachep(size, flags);
3659 if (unlikely(cachep == NULL))
3661 return kmem_cache_alloc_node(cachep, flags, node);
3664 #ifdef CONFIG_DEBUG_SLAB
3665 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3667 return __do_kmalloc_node(size, flags, node,
3668 __builtin_return_address(0));
3670 EXPORT_SYMBOL(__kmalloc_node);
3672 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3673 int node, void *caller)
3675 return __do_kmalloc_node(size, flags, node, caller);
3677 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3679 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3681 return __do_kmalloc_node(size, flags, node, NULL);
3683 EXPORT_SYMBOL(__kmalloc_node);
3684 #endif /* CONFIG_DEBUG_SLAB */
3685 #endif /* CONFIG_NUMA */
3688 * __do_kmalloc - allocate memory
3689 * @size: how many bytes of memory are required.
3690 * @flags: the type of memory to allocate (see kmalloc).
3691 * @caller: function caller for debug tracking of the caller
3693 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3696 struct kmem_cache *cachep;
3698 /* If you want to save a few bytes .text space: replace
3700 * Then kmalloc uses the uninlined functions instead of the inline
3703 cachep = __find_general_cachep(size, flags);
3704 if (unlikely(cachep == NULL))
3706 return __cache_alloc(cachep, flags, caller);
3710 #ifdef CONFIG_DEBUG_SLAB
3711 void *__kmalloc(size_t size, gfp_t flags)
3713 return __do_kmalloc(size, flags, __builtin_return_address(0));
3715 EXPORT_SYMBOL(__kmalloc);
3717 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3719 return __do_kmalloc(size, flags, caller);
3721 EXPORT_SYMBOL(__kmalloc_track_caller);
3724 void *__kmalloc(size_t size, gfp_t flags)
3726 return __do_kmalloc(size, flags, NULL);
3728 EXPORT_SYMBOL(__kmalloc);
3732 * krealloc - reallocate memory. The contents will remain unchanged.
3733 * @p: object to reallocate memory for.
3734 * @new_size: how many bytes of memory are required.
3735 * @flags: the type of memory to allocate.
3737 * The contents of the object pointed to are preserved up to the
3738 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3739 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3740 * %NULL pointer, the object pointed to is freed.
3742 void *krealloc(const void *p, size_t new_size, gfp_t flags)
3744 struct kmem_cache *cache, *new_cache;
3748 return kmalloc_track_caller(new_size, flags);
3750 if (unlikely(!new_size)) {
3755 cache = virt_to_cache(p);
3756 new_cache = __find_general_cachep(new_size, flags);
3759 * If new size fits in the current cache, bail out.
3761 if (likely(cache == new_cache))
3765 * We are on the slow-path here so do not use __cache_alloc
3766 * because it bloats kernel text.
3768 ret = kmalloc_track_caller(new_size, flags);
3770 memcpy(ret, p, min(new_size, ksize(p)));
3775 EXPORT_SYMBOL(krealloc);
3778 * kmem_cache_free - Deallocate an object
3779 * @cachep: The cache the allocation was from.
3780 * @objp: The previously allocated object.
3782 * Free an object which was previously allocated from this
3785 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3787 unsigned long flags;
3789 BUG_ON(virt_to_cache(objp) != cachep);
3791 local_irq_save(flags);
3792 debug_check_no_locks_freed(objp, obj_size(cachep));
3793 __cache_free(cachep, objp);
3794 local_irq_restore(flags);
3796 EXPORT_SYMBOL(kmem_cache_free);
3799 * kfree - free previously allocated memory
3800 * @objp: pointer returned by kmalloc.
3802 * If @objp is NULL, no operation is performed.
3804 * Don't free memory not originally allocated by kmalloc()
3805 * or you will run into trouble.
3807 void kfree(const void *objp)
3809 struct kmem_cache *c;
3810 unsigned long flags;
3812 if (unlikely(!objp))
3814 local_irq_save(flags);
3815 kfree_debugcheck(objp);
3816 c = virt_to_cache(objp);
3817 debug_check_no_locks_freed(objp, obj_size(c));
3818 __cache_free(c, (void *)objp);
3819 local_irq_restore(flags);
3821 EXPORT_SYMBOL(kfree);
3823 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3825 return obj_size(cachep);
3827 EXPORT_SYMBOL(kmem_cache_size);
3829 const char *kmem_cache_name(struct kmem_cache *cachep)
3831 return cachep->name;
3833 EXPORT_SYMBOL_GPL(kmem_cache_name);
3836 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3838 static int alloc_kmemlist(struct kmem_cache *cachep)
3841 struct kmem_list3 *l3;
3842 struct array_cache *new_shared;
3843 struct array_cache **new_alien = NULL;
3845 for_each_online_node(node) {
3847 if (use_alien_caches) {
3848 new_alien = alloc_alien_cache(node, cachep->limit);
3854 if (cachep->shared) {
3855 new_shared = alloc_arraycache(node,
3856 cachep->shared*cachep->batchcount,
3859 free_alien_cache(new_alien);
3864 l3 = cachep->nodelists[node];
3866 struct array_cache *shared = l3->shared;
3868 spin_lock_irq(&l3->list_lock);
3871 free_block(cachep, shared->entry,
3872 shared->avail, node);
3874 l3->shared = new_shared;
3876 l3->alien = new_alien;
3879 l3->free_limit = (1 + nr_cpus_node(node)) *
3880 cachep->batchcount + cachep->num;
3881 spin_unlock_irq(&l3->list_lock);
3883 free_alien_cache(new_alien);
3886 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3888 free_alien_cache(new_alien);
3893 kmem_list3_init(l3);
3894 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3895 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3896 l3->shared = new_shared;
3897 l3->alien = new_alien;
3898 l3->free_limit = (1 + nr_cpus_node(node)) *
3899 cachep->batchcount + cachep->num;
3900 cachep->nodelists[node] = l3;
3905 if (!cachep->next.next) {
3906 /* Cache is not active yet. Roll back what we did */
3909 if (cachep->nodelists[node]) {
3910 l3 = cachep->nodelists[node];
3913 free_alien_cache(l3->alien);
3915 cachep->nodelists[node] = NULL;
3923 struct ccupdate_struct {
3924 struct kmem_cache *cachep;
3925 struct array_cache *new[NR_CPUS];
3928 static void do_ccupdate_local(void *info)
3930 struct ccupdate_struct *new = info;
3931 struct array_cache *old;
3934 old = cpu_cache_get(new->cachep);
3936 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3937 new->new[smp_processor_id()] = old;
3940 /* Always called with the cache_chain_mutex held */
3941 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3942 int batchcount, int shared)
3944 struct ccupdate_struct *new;
3947 new = kzalloc(sizeof(*new), GFP_KERNEL);
3951 for_each_online_cpu(i) {
3952 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3955 for (i--; i >= 0; i--)
3961 new->cachep = cachep;
3963 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3966 cachep->batchcount = batchcount;
3967 cachep->limit = limit;
3968 cachep->shared = shared;
3970 for_each_online_cpu(i) {
3971 struct array_cache *ccold = new->new[i];
3974 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3975 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3976 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3980 return alloc_kmemlist(cachep);
3983 /* Called with cache_chain_mutex held always */
3984 static int enable_cpucache(struct kmem_cache *cachep)
3990 * The head array serves three purposes:
3991 * - create a LIFO ordering, i.e. return objects that are cache-warm
3992 * - reduce the number of spinlock operations.
3993 * - reduce the number of linked list operations on the slab and
3994 * bufctl chains: array operations are cheaper.
3995 * The numbers are guessed, we should auto-tune as described by
3998 if (cachep->buffer_size > 131072)
4000 else if (cachep->buffer_size > PAGE_SIZE)
4002 else if (cachep->buffer_size > 1024)
4004 else if (cachep->buffer_size > 256)
4010 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4011 * allocation behaviour: Most allocs on one cpu, most free operations
4012 * on another cpu. For these cases, an efficient object passing between
4013 * cpus is necessary. This is provided by a shared array. The array
4014 * replaces Bonwick's magazine layer.
4015 * On uniprocessor, it's functionally equivalent (but less efficient)
4016 * to a larger limit. Thus disabled by default.
4019 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4024 * With debugging enabled, large batchcount lead to excessively long
4025 * periods with disabled local interrupts. Limit the batchcount
4030 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4032 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4033 cachep->name, -err);
4038 * Drain an array if it contains any elements taking the l3 lock only if
4039 * necessary. Note that the l3 listlock also protects the array_cache
4040 * if drain_array() is used on the shared array.
4042 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4043 struct array_cache *ac, int force, int node)
4047 if (!ac || !ac->avail)
4049 if (ac->touched && !force) {
4052 spin_lock_irq(&l3->list_lock);
4054 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4055 if (tofree > ac->avail)
4056 tofree = (ac->avail + 1) / 2;
4057 free_block(cachep, ac->entry, tofree, node);
4058 ac->avail -= tofree;
4059 memmove(ac->entry, &(ac->entry[tofree]),
4060 sizeof(void *) * ac->avail);
4062 spin_unlock_irq(&l3->list_lock);
4067 * cache_reap - Reclaim memory from caches.
4068 * @w: work descriptor
4070 * Called from workqueue/eventd every few seconds.
4072 * - clear the per-cpu caches for this CPU.
4073 * - return freeable pages to the main free memory pool.
4075 * If we cannot acquire the cache chain mutex then just give up - we'll try
4076 * again on the next iteration.
4078 static void cache_reap(struct work_struct *w)
4080 struct kmem_cache *searchp;
4081 struct kmem_list3 *l3;
4082 int node = numa_node_id();
4083 struct delayed_work *work =
4084 container_of(w, struct delayed_work, work);
4086 if (!mutex_trylock(&cache_chain_mutex))
4087 /* Give up. Setup the next iteration. */
4090 list_for_each_entry(searchp, &cache_chain, next) {
4094 * We only take the l3 lock if absolutely necessary and we
4095 * have established with reasonable certainty that
4096 * we can do some work if the lock was obtained.
4098 l3 = searchp->nodelists[node];
4100 reap_alien(searchp, l3);
4102 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4105 * These are racy checks but it does not matter
4106 * if we skip one check or scan twice.
4108 if (time_after(l3->next_reap, jiffies))
4111 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4113 drain_array(searchp, l3, l3->shared, 0, node);
4115 if (l3->free_touched)
4116 l3->free_touched = 0;
4120 freed = drain_freelist(searchp, l3, (l3->free_limit +
4121 5 * searchp->num - 1) / (5 * searchp->num));
4122 STATS_ADD_REAPED(searchp, freed);
4128 mutex_unlock(&cache_chain_mutex);
4131 /* Set up the next iteration */
4132 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4135 #ifdef CONFIG_PROC_FS
4137 static void print_slabinfo_header(struct seq_file *m)
4140 * Output format version, so at least we can change it
4141 * without _too_ many complaints.
4144 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4146 seq_puts(m, "slabinfo - version: 2.1\n");
4148 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4149 "<objperslab> <pagesperslab>");
4150 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4151 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4153 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4154 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4155 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4160 static void *s_start(struct seq_file *m, loff_t *pos)
4163 struct list_head *p;
4165 mutex_lock(&cache_chain_mutex);
4167 print_slabinfo_header(m);
4168 p = cache_chain.next;
4171 if (p == &cache_chain)
4174 return list_entry(p, struct kmem_cache, next);
4177 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4179 struct kmem_cache *cachep = p;
4181 return cachep->next.next == &cache_chain ?
4182 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
4185 static void s_stop(struct seq_file *m, void *p)
4187 mutex_unlock(&cache_chain_mutex);
4190 static int s_show(struct seq_file *m, void *p)
4192 struct kmem_cache *cachep = p;
4194 unsigned long active_objs;
4195 unsigned long num_objs;
4196 unsigned long active_slabs = 0;
4197 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4201 struct kmem_list3 *l3;
4205 for_each_online_node(node) {
4206 l3 = cachep->nodelists[node];
4211 spin_lock_irq(&l3->list_lock);
4213 list_for_each_entry(slabp, &l3->slabs_full, list) {
4214 if (slabp->inuse != cachep->num && !error)
4215 error = "slabs_full accounting error";
4216 active_objs += cachep->num;
4219 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4220 if (slabp->inuse == cachep->num && !error)
4221 error = "slabs_partial inuse accounting error";
4222 if (!slabp->inuse && !error)
4223 error = "slabs_partial/inuse accounting error";
4224 active_objs += slabp->inuse;
4227 list_for_each_entry(slabp, &l3->slabs_free, list) {
4228 if (slabp->inuse && !error)
4229 error = "slabs_free/inuse accounting error";
4232 free_objects += l3->free_objects;
4234 shared_avail += l3->shared->avail;
4236 spin_unlock_irq(&l3->list_lock);
4238 num_slabs += active_slabs;
4239 num_objs = num_slabs * cachep->num;
4240 if (num_objs - active_objs != free_objects && !error)
4241 error = "free_objects accounting error";
4243 name = cachep->name;
4245 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4247 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4248 name, active_objs, num_objs, cachep->buffer_size,
4249 cachep->num, (1 << cachep->gfporder));
4250 seq_printf(m, " : tunables %4u %4u %4u",
4251 cachep->limit, cachep->batchcount, cachep->shared);
4252 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4253 active_slabs, num_slabs, shared_avail);
4256 unsigned long high = cachep->high_mark;
4257 unsigned long allocs = cachep->num_allocations;
4258 unsigned long grown = cachep->grown;
4259 unsigned long reaped = cachep->reaped;
4260 unsigned long errors = cachep->errors;
4261 unsigned long max_freeable = cachep->max_freeable;
4262 unsigned long node_allocs = cachep->node_allocs;
4263 unsigned long node_frees = cachep->node_frees;
4264 unsigned long overflows = cachep->node_overflow;
4266 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4267 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4268 reaped, errors, max_freeable, node_allocs,
4269 node_frees, overflows);
4273 unsigned long allochit = atomic_read(&cachep->allochit);
4274 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4275 unsigned long freehit = atomic_read(&cachep->freehit);
4276 unsigned long freemiss = atomic_read(&cachep->freemiss);
4278 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4279 allochit, allocmiss, freehit, freemiss);
4287 * slabinfo_op - iterator that generates /proc/slabinfo
4296 * num-pages-per-slab
4297 * + further values on SMP and with statistics enabled
4300 const struct seq_operations slabinfo_op = {
4307 #define MAX_SLABINFO_WRITE 128
4309 * slabinfo_write - Tuning for the slab allocator
4311 * @buffer: user buffer
4312 * @count: data length
4315 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4316 size_t count, loff_t *ppos)
4318 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4319 int limit, batchcount, shared, res;
4320 struct kmem_cache *cachep;
4322 if (count > MAX_SLABINFO_WRITE)
4324 if (copy_from_user(&kbuf, buffer, count))
4326 kbuf[MAX_SLABINFO_WRITE] = '\0';
4328 tmp = strchr(kbuf, ' ');
4333 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4336 /* Find the cache in the chain of caches. */
4337 mutex_lock(&cache_chain_mutex);
4339 list_for_each_entry(cachep, &cache_chain, next) {
4340 if (!strcmp(cachep->name, kbuf)) {
4341 if (limit < 1 || batchcount < 1 ||
4342 batchcount > limit || shared < 0) {
4345 res = do_tune_cpucache(cachep, limit,
4346 batchcount, shared);
4351 mutex_unlock(&cache_chain_mutex);
4357 #ifdef CONFIG_DEBUG_SLAB_LEAK
4359 static void *leaks_start(struct seq_file *m, loff_t *pos)
4362 struct list_head *p;
4364 mutex_lock(&cache_chain_mutex);
4365 p = cache_chain.next;
4368 if (p == &cache_chain)
4371 return list_entry(p, struct kmem_cache, next);
4374 static inline int add_caller(unsigned long *n, unsigned long v)
4384 unsigned long *q = p + 2 * i;
4398 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4404 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4410 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4411 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4413 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4418 static void show_symbol(struct seq_file *m, unsigned long address)
4420 #ifdef CONFIG_KALLSYMS
4421 unsigned long offset, size;
4422 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1];
4424 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4425 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4427 seq_printf(m, " [%s]", modname);
4431 seq_printf(m, "%p", (void *)address);
4434 static int leaks_show(struct seq_file *m, void *p)
4436 struct kmem_cache *cachep = p;
4438 struct kmem_list3 *l3;
4440 unsigned long *n = m->private;
4444 if (!(cachep->flags & SLAB_STORE_USER))
4446 if (!(cachep->flags & SLAB_RED_ZONE))
4449 /* OK, we can do it */
4453 for_each_online_node(node) {
4454 l3 = cachep->nodelists[node];
4459 spin_lock_irq(&l3->list_lock);
4461 list_for_each_entry(slabp, &l3->slabs_full, list)
4462 handle_slab(n, cachep, slabp);
4463 list_for_each_entry(slabp, &l3->slabs_partial, list)
4464 handle_slab(n, cachep, slabp);
4465 spin_unlock_irq(&l3->list_lock);
4467 name = cachep->name;
4469 /* Increase the buffer size */
4470 mutex_unlock(&cache_chain_mutex);
4471 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4473 /* Too bad, we are really out */
4475 mutex_lock(&cache_chain_mutex);
4478 *(unsigned long *)m->private = n[0] * 2;
4480 mutex_lock(&cache_chain_mutex);
4481 /* Now make sure this entry will be retried */
4485 for (i = 0; i < n[1]; i++) {
4486 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4487 show_symbol(m, n[2*i+2]);
4494 const struct seq_operations slabstats_op = {
4495 .start = leaks_start,
4504 * ksize - get the actual amount of memory allocated for a given object
4505 * @objp: Pointer to the object
4507 * kmalloc may internally round up allocations and return more memory
4508 * than requested. ksize() can be used to determine the actual amount of
4509 * memory allocated. The caller may use this additional memory, even though
4510 * a smaller amount of memory was initially specified with the kmalloc call.
4511 * The caller must guarantee that objp points to a valid object previously
4512 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4513 * must not be freed during the duration of the call.
4515 size_t ksize(const void *objp)
4517 if (unlikely(objp == NULL))
4520 return obj_size(virt_to_cache(objp));