3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/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 *)
140 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
142 #ifndef cache_line_size
143 #define cache_line_size() L1_CACHE_BYTES
146 #ifndef ARCH_KMALLOC_MINALIGN
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
156 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
159 #ifndef ARCH_SLAB_MINALIGN
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
167 #define ARCH_SLAB_MINALIGN 0
170 #ifndef ARCH_KMALLOC_FLAGS
171 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
174 /* Legal flag mask for kmem_cache_create(). */
176 # define CREATE_MASK (SLAB_RED_ZONE | \
177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
183 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
192 * Bufctl's are used for linking objs within a slab
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
208 typedef unsigned int kmem_bufctl_t;
209 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
211 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
227 unsigned short nodeid;
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
244 * We assume struct slab_rcu can overlay struct slab when destroying.
247 struct rcu_head head;
248 struct kmem_cache *cachep;
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
260 * The limit is stored in the per-cpu structure to reduce the data cache
267 unsigned int batchcount;
268 unsigned int touched;
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
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 (3 * MAX_NUMNODES)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
310 #define SIZE_AC MAX_NUMNODES
311 #define SIZE_L3 (2 * 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)(struct kmem_cache *, void *);
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 * Do not go above this order unless 0 objects fit into the slab.
574 #define BREAK_GFP_ORDER_HI 1
575 #define BREAK_GFP_ORDER_LO 0
576 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
579 * Functions for storing/retrieving the cachep and or slab from the page
580 * allocator. These are used to find the slab an obj belongs to. With kfree(),
581 * these are used to find the cache which an obj belongs to.
583 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
585 page->lru.next = (struct list_head *)cache;
588 static inline struct kmem_cache *page_get_cache(struct page *page)
590 page = compound_head(page);
591 BUG_ON(!PageSlab(page));
592 return (struct kmem_cache *)page->lru.next;
595 static inline void page_set_slab(struct page *page, struct slab *slab)
597 page->lru.prev = (struct list_head *)slab;
600 static inline struct slab *page_get_slab(struct page *page)
602 BUG_ON(!PageSlab(page));
603 return (struct slab *)page->lru.prev;
606 static inline struct kmem_cache *virt_to_cache(const void *obj)
608 struct page *page = virt_to_head_page(obj);
609 return page_get_cache(page);
612 static inline struct slab *virt_to_slab(const void *obj)
614 struct page *page = virt_to_head_page(obj);
615 return page_get_slab(page);
618 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
621 return slab->s_mem + cache->buffer_size * idx;
625 * We want to avoid an expensive divide : (offset / cache->buffer_size)
626 * Using the fact that buffer_size is a constant for a particular cache,
627 * we can replace (offset / cache->buffer_size) by
628 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
630 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
631 const struct slab *slab, void *obj)
633 u32 offset = (obj - slab->s_mem);
634 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
638 * These are the default caches for kmalloc. Custom caches can have other sizes.
640 struct cache_sizes malloc_sizes[] = {
641 #define CACHE(x) { .cs_size = (x) },
642 #include <linux/kmalloc_sizes.h>
646 EXPORT_SYMBOL(malloc_sizes);
648 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
654 static struct cache_names __initdata cache_names[] = {
655 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
656 #include <linux/kmalloc_sizes.h>
661 static struct arraycache_init initarray_cache __initdata =
662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
663 static struct arraycache_init initarray_generic =
664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
666 /* internal cache of cache description objs */
667 static struct kmem_cache cache_cache = {
669 .limit = BOOT_CPUCACHE_ENTRIES,
671 .buffer_size = sizeof(struct kmem_cache),
672 .name = "kmem_cache",
675 #define BAD_ALIEN_MAGIC 0x01020304ul
677 #ifdef CONFIG_LOCKDEP
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
686 * We set lock class for alien array caches which are up during init.
687 * The lock annotation will be lost if all cpus of a node goes down and
688 * then comes back up during hotplug
690 static struct lock_class_key on_slab_l3_key;
691 static struct lock_class_key on_slab_alc_key;
693 static inline void init_lock_keys(void)
697 struct cache_sizes *s = malloc_sizes;
699 while (s->cs_size != ULONG_MAX) {
701 struct array_cache **alc;
703 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
704 if (!l3 || OFF_SLAB(s->cs_cachep))
706 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
709 * FIXME: This check for BAD_ALIEN_MAGIC
710 * should go away when common slab code is taught to
711 * work even without alien caches.
712 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
713 * for alloc_alien_cache,
715 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
719 lockdep_set_class(&alc[r]->lock,
727 static inline void init_lock_keys(void)
733 * Guard access to the cache-chain.
735 static DEFINE_MUTEX(cache_chain_mutex);
736 static struct list_head cache_chain;
739 * chicken and egg problem: delay the per-cpu array allocation
740 * until the general caches are up.
750 * used by boot code to determine if it can use slab based allocator
752 int slab_is_available(void)
754 return g_cpucache_up == FULL;
757 static DEFINE_PER_CPU(struct delayed_work, reap_work);
759 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
761 return cachep->array[smp_processor_id()];
764 static inline struct kmem_cache *__find_general_cachep(size_t size,
767 struct cache_sizes *csizep = malloc_sizes;
770 /* This happens if someone tries to call
771 * kmem_cache_create(), or __kmalloc(), before
772 * the generic caches are initialized.
774 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
777 return ZERO_SIZE_PTR;
779 while (size > csizep->cs_size)
783 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
784 * has cs_{dma,}cachep==NULL. Thus no special case
785 * for large kmalloc calls required.
787 #ifdef CONFIG_ZONE_DMA
788 if (unlikely(gfpflags & GFP_DMA))
789 return csizep->cs_dmacachep;
791 return csizep->cs_cachep;
794 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
796 return __find_general_cachep(size, gfpflags);
799 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
801 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
805 * Calculate the number of objects and left-over bytes for a given buffer size.
807 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
808 size_t align, int flags, size_t *left_over,
813 size_t slab_size = PAGE_SIZE << gfporder;
816 * The slab management structure can be either off the slab or
817 * on it. For the latter case, the memory allocated for a
821 * - One kmem_bufctl_t for each object
822 * - Padding to respect alignment of @align
823 * - @buffer_size bytes for each object
825 * If the slab management structure is off the slab, then the
826 * alignment will already be calculated into the size. Because
827 * the slabs are all pages aligned, the objects will be at the
828 * correct alignment when allocated.
830 if (flags & CFLGS_OFF_SLAB) {
832 nr_objs = slab_size / buffer_size;
834 if (nr_objs > SLAB_LIMIT)
835 nr_objs = SLAB_LIMIT;
838 * Ignore padding for the initial guess. The padding
839 * is at most @align-1 bytes, and @buffer_size is at
840 * least @align. In the worst case, this result will
841 * be one greater than the number of objects that fit
842 * into the memory allocation when taking the padding
845 nr_objs = (slab_size - sizeof(struct slab)) /
846 (buffer_size + sizeof(kmem_bufctl_t));
849 * This calculated number will be either the right
850 * amount, or one greater than what we want.
852 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
856 if (nr_objs > SLAB_LIMIT)
857 nr_objs = SLAB_LIMIT;
859 mgmt_size = slab_mgmt_size(nr_objs, align);
862 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
865 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
867 static void __slab_error(const char *function, struct kmem_cache *cachep,
870 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
871 function, cachep->name, msg);
876 * By default on NUMA we use alien caches to stage the freeing of
877 * objects allocated from other nodes. This causes massive memory
878 * inefficiencies when using fake NUMA setup to split memory into a
879 * large number of small nodes, so it can be disabled on the command
883 static int use_alien_caches __read_mostly = 1;
884 static int numa_platform __read_mostly = 1;
885 static int __init noaliencache_setup(char *s)
887 use_alien_caches = 0;
890 __setup("noaliencache", noaliencache_setup);
894 * Special reaping functions for NUMA systems called from cache_reap().
895 * These take care of doing round robin flushing of alien caches (containing
896 * objects freed on different nodes from which they were allocated) and the
897 * flushing of remote pcps by calling drain_node_pages.
899 static DEFINE_PER_CPU(unsigned long, reap_node);
901 static void init_reap_node(int cpu)
905 node = next_node(cpu_to_node(cpu), node_online_map);
906 if (node == MAX_NUMNODES)
907 node = first_node(node_online_map);
909 per_cpu(reap_node, cpu) = node;
912 static void next_reap_node(void)
914 int node = __get_cpu_var(reap_node);
916 node = next_node(node, node_online_map);
917 if (unlikely(node >= MAX_NUMNODES))
918 node = first_node(node_online_map);
919 __get_cpu_var(reap_node) = node;
923 #define init_reap_node(cpu) do { } while (0)
924 #define next_reap_node(void) do { } while (0)
928 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
929 * via the workqueue/eventd.
930 * Add the CPU number into the expiration time to minimize the possibility of
931 * the CPUs getting into lockstep and contending for the global cache chain
934 static void __cpuinit start_cpu_timer(int cpu)
936 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
939 * When this gets called from do_initcalls via cpucache_init(),
940 * init_workqueues() has already run, so keventd will be setup
943 if (keventd_up() && reap_work->work.func == NULL) {
945 INIT_DELAYED_WORK(reap_work, cache_reap);
946 schedule_delayed_work_on(cpu, reap_work,
947 __round_jiffies_relative(HZ, cpu));
951 static struct array_cache *alloc_arraycache(int node, int entries,
954 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
955 struct array_cache *nc = NULL;
957 nc = kmalloc_node(memsize, GFP_KERNEL, node);
961 nc->batchcount = batchcount;
963 spin_lock_init(&nc->lock);
969 * Transfer objects in one arraycache to another.
970 * Locking must be handled by the caller.
972 * Return the number of entries transferred.
974 static int transfer_objects(struct array_cache *to,
975 struct array_cache *from, unsigned int max)
977 /* Figure out how many entries to transfer */
978 int nr = min(min(from->avail, max), to->limit - to->avail);
983 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
994 #define drain_alien_cache(cachep, alien) do { } while (0)
995 #define reap_alien(cachep, l3) do { } while (0)
997 static inline struct array_cache **alloc_alien_cache(int node, int limit)
999 return (struct array_cache **)BAD_ALIEN_MAGIC;
1002 static inline void free_alien_cache(struct array_cache **ac_ptr)
1006 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1011 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1017 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1018 gfp_t flags, int nodeid)
1023 #else /* CONFIG_NUMA */
1025 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1026 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1028 static struct array_cache **alloc_alien_cache(int node, int limit)
1030 struct array_cache **ac_ptr;
1031 int memsize = sizeof(void *) * nr_node_ids;
1036 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1039 if (i == node || !node_online(i)) {
1043 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1045 for (i--; i >= 0; i--)
1055 static void free_alien_cache(struct array_cache **ac_ptr)
1066 static void __drain_alien_cache(struct kmem_cache *cachep,
1067 struct array_cache *ac, int node)
1069 struct kmem_list3 *rl3 = cachep->nodelists[node];
1072 spin_lock(&rl3->list_lock);
1074 * Stuff objects into the remote nodes shared array first.
1075 * That way we could avoid the overhead of putting the objects
1076 * into the free lists and getting them back later.
1079 transfer_objects(rl3->shared, ac, ac->limit);
1081 free_block(cachep, ac->entry, ac->avail, node);
1083 spin_unlock(&rl3->list_lock);
1088 * Called from cache_reap() to regularly drain alien caches round robin.
1090 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1092 int node = __get_cpu_var(reap_node);
1095 struct array_cache *ac = l3->alien[node];
1097 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1098 __drain_alien_cache(cachep, ac, node);
1099 spin_unlock_irq(&ac->lock);
1104 static void drain_alien_cache(struct kmem_cache *cachep,
1105 struct array_cache **alien)
1108 struct array_cache *ac;
1109 unsigned long flags;
1111 for_each_online_node(i) {
1114 spin_lock_irqsave(&ac->lock, flags);
1115 __drain_alien_cache(cachep, ac, i);
1116 spin_unlock_irqrestore(&ac->lock, flags);
1121 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1123 struct slab *slabp = virt_to_slab(objp);
1124 int nodeid = slabp->nodeid;
1125 struct kmem_list3 *l3;
1126 struct array_cache *alien = NULL;
1129 node = numa_node_id();
1132 * Make sure we are not freeing a object from another node to the array
1133 * cache on this cpu.
1135 if (likely(slabp->nodeid == node))
1138 l3 = cachep->nodelists[node];
1139 STATS_INC_NODEFREES(cachep);
1140 if (l3->alien && l3->alien[nodeid]) {
1141 alien = l3->alien[nodeid];
1142 spin_lock(&alien->lock);
1143 if (unlikely(alien->avail == alien->limit)) {
1144 STATS_INC_ACOVERFLOW(cachep);
1145 __drain_alien_cache(cachep, alien, nodeid);
1147 alien->entry[alien->avail++] = objp;
1148 spin_unlock(&alien->lock);
1150 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1151 free_block(cachep, &objp, 1, nodeid);
1152 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1158 static void __cpuinit cpuup_canceled(long cpu)
1160 struct kmem_cache *cachep;
1161 struct kmem_list3 *l3 = NULL;
1162 int node = cpu_to_node(cpu);
1164 list_for_each_entry(cachep, &cache_chain, next) {
1165 struct array_cache *nc;
1166 struct array_cache *shared;
1167 struct array_cache **alien;
1170 mask = node_to_cpumask(node);
1171 /* cpu is dead; no one can alloc from it. */
1172 nc = cachep->array[cpu];
1173 cachep->array[cpu] = NULL;
1174 l3 = cachep->nodelists[node];
1177 goto free_array_cache;
1179 spin_lock_irq(&l3->list_lock);
1181 /* Free limit for this kmem_list3 */
1182 l3->free_limit -= cachep->batchcount;
1184 free_block(cachep, nc->entry, nc->avail, node);
1186 if (!cpus_empty(mask)) {
1187 spin_unlock_irq(&l3->list_lock);
1188 goto free_array_cache;
1191 shared = l3->shared;
1193 free_block(cachep, shared->entry,
1194 shared->avail, node);
1201 spin_unlock_irq(&l3->list_lock);
1205 drain_alien_cache(cachep, alien);
1206 free_alien_cache(alien);
1212 * In the previous loop, all the objects were freed to
1213 * the respective cache's slabs, now we can go ahead and
1214 * shrink each nodelist to its limit.
1216 list_for_each_entry(cachep, &cache_chain, next) {
1217 l3 = cachep->nodelists[node];
1220 drain_freelist(cachep, l3, l3->free_objects);
1224 static int __cpuinit cpuup_prepare(long cpu)
1226 struct kmem_cache *cachep;
1227 struct kmem_list3 *l3 = NULL;
1228 int node = cpu_to_node(cpu);
1229 const int memsize = sizeof(struct kmem_list3);
1232 * We need to do this right in the beginning since
1233 * alloc_arraycache's are going to use this list.
1234 * kmalloc_node allows us to add the slab to the right
1235 * kmem_list3 and not this cpu's kmem_list3
1238 list_for_each_entry(cachep, &cache_chain, next) {
1240 * Set up the size64 kmemlist for cpu before we can
1241 * begin anything. Make sure some other cpu on this
1242 * node has not already allocated this
1244 if (!cachep->nodelists[node]) {
1245 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1248 kmem_list3_init(l3);
1249 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1250 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1253 * The l3s don't come and go as CPUs come and
1254 * go. cache_chain_mutex is sufficient
1257 cachep->nodelists[node] = l3;
1260 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1261 cachep->nodelists[node]->free_limit =
1262 (1 + nr_cpus_node(node)) *
1263 cachep->batchcount + cachep->num;
1264 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1268 * Now we can go ahead with allocating the shared arrays and
1271 list_for_each_entry(cachep, &cache_chain, next) {
1272 struct array_cache *nc;
1273 struct array_cache *shared = NULL;
1274 struct array_cache **alien = NULL;
1276 nc = alloc_arraycache(node, cachep->limit,
1277 cachep->batchcount);
1280 if (cachep->shared) {
1281 shared = alloc_arraycache(node,
1282 cachep->shared * cachep->batchcount,
1289 if (use_alien_caches) {
1290 alien = alloc_alien_cache(node, cachep->limit);
1297 cachep->array[cpu] = nc;
1298 l3 = cachep->nodelists[node];
1301 spin_lock_irq(&l3->list_lock);
1304 * We are serialised from CPU_DEAD or
1305 * CPU_UP_CANCELLED by the cpucontrol lock
1307 l3->shared = shared;
1316 spin_unlock_irq(&l3->list_lock);
1318 free_alien_cache(alien);
1322 cpuup_canceled(cpu);
1326 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1327 unsigned long action, void *hcpu)
1329 long cpu = (long)hcpu;
1333 case CPU_UP_PREPARE:
1334 case CPU_UP_PREPARE_FROZEN:
1335 mutex_lock(&cache_chain_mutex);
1336 err = cpuup_prepare(cpu);
1337 mutex_unlock(&cache_chain_mutex);
1340 case CPU_ONLINE_FROZEN:
1341 start_cpu_timer(cpu);
1343 #ifdef CONFIG_HOTPLUG_CPU
1344 case CPU_DOWN_PREPARE:
1345 case CPU_DOWN_PREPARE_FROZEN:
1347 * Shutdown cache reaper. Note that the cache_chain_mutex is
1348 * held so that if cache_reap() is invoked it cannot do
1349 * anything expensive but will only modify reap_work
1350 * and reschedule the timer.
1352 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1353 /* Now the cache_reaper is guaranteed to be not running. */
1354 per_cpu(reap_work, cpu).work.func = NULL;
1356 case CPU_DOWN_FAILED:
1357 case CPU_DOWN_FAILED_FROZEN:
1358 start_cpu_timer(cpu);
1361 case CPU_DEAD_FROZEN:
1363 * Even if all the cpus of a node are down, we don't free the
1364 * kmem_list3 of any cache. This to avoid a race between
1365 * cpu_down, and a kmalloc allocation from another cpu for
1366 * memory from the node of the cpu going down. The list3
1367 * structure is usually allocated from kmem_cache_create() and
1368 * gets destroyed at kmem_cache_destroy().
1372 case CPU_UP_CANCELED:
1373 case CPU_UP_CANCELED_FROZEN:
1374 mutex_lock(&cache_chain_mutex);
1375 cpuup_canceled(cpu);
1376 mutex_unlock(&cache_chain_mutex);
1379 return err ? NOTIFY_BAD : NOTIFY_OK;
1382 static struct notifier_block __cpuinitdata cpucache_notifier = {
1383 &cpuup_callback, NULL, 0
1387 * swap the static kmem_list3 with kmalloced memory
1389 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1392 struct kmem_list3 *ptr;
1394 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1397 local_irq_disable();
1398 memcpy(ptr, list, sizeof(struct kmem_list3));
1400 * Do not assume that spinlocks can be initialized via memcpy:
1402 spin_lock_init(&ptr->list_lock);
1404 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1405 cachep->nodelists[nodeid] = ptr;
1410 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1411 * size of kmem_list3.
1413 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1417 for_each_online_node(node) {
1418 cachep->nodelists[node] = &initkmem_list3[index + node];
1419 cachep->nodelists[node]->next_reap = jiffies +
1421 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1426 * Initialisation. Called after the page allocator have been initialised and
1427 * before smp_init().
1429 void __init kmem_cache_init(void)
1432 struct cache_sizes *sizes;
1433 struct cache_names *names;
1438 if (num_possible_nodes() == 1) {
1439 use_alien_caches = 0;
1443 for (i = 0; i < NUM_INIT_LISTS; i++) {
1444 kmem_list3_init(&initkmem_list3[i]);
1445 if (i < MAX_NUMNODES)
1446 cache_cache.nodelists[i] = NULL;
1448 set_up_list3s(&cache_cache, CACHE_CACHE);
1451 * Fragmentation resistance on low memory - only use bigger
1452 * page orders on machines with more than 32MB of memory.
1454 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1455 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1457 /* Bootstrap is tricky, because several objects are allocated
1458 * from caches that do not exist yet:
1459 * 1) initialize the cache_cache cache: it contains the struct
1460 * kmem_cache structures of all caches, except cache_cache itself:
1461 * cache_cache is statically allocated.
1462 * Initially an __init data area is used for the head array and the
1463 * kmem_list3 structures, it's replaced with a kmalloc allocated
1464 * array at the end of the bootstrap.
1465 * 2) Create the first kmalloc cache.
1466 * The struct kmem_cache for the new cache is allocated normally.
1467 * An __init data area is used for the head array.
1468 * 3) Create the remaining kmalloc caches, with minimally sized
1470 * 4) Replace the __init data head arrays for cache_cache and the first
1471 * kmalloc cache with kmalloc allocated arrays.
1472 * 5) Replace the __init data for kmem_list3 for cache_cache and
1473 * the other cache's with kmalloc allocated memory.
1474 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1477 node = numa_node_id();
1479 /* 1) create the cache_cache */
1480 INIT_LIST_HEAD(&cache_chain);
1481 list_add(&cache_cache.next, &cache_chain);
1482 cache_cache.colour_off = cache_line_size();
1483 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1484 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1487 * struct kmem_cache size depends on nr_node_ids, which
1488 * can be less than MAX_NUMNODES.
1490 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1491 nr_node_ids * sizeof(struct kmem_list3 *);
1493 cache_cache.obj_size = cache_cache.buffer_size;
1495 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1497 cache_cache.reciprocal_buffer_size =
1498 reciprocal_value(cache_cache.buffer_size);
1500 for (order = 0; order < MAX_ORDER; order++) {
1501 cache_estimate(order, cache_cache.buffer_size,
1502 cache_line_size(), 0, &left_over, &cache_cache.num);
1503 if (cache_cache.num)
1506 BUG_ON(!cache_cache.num);
1507 cache_cache.gfporder = order;
1508 cache_cache.colour = left_over / cache_cache.colour_off;
1509 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1510 sizeof(struct slab), cache_line_size());
1512 /* 2+3) create the kmalloc caches */
1513 sizes = malloc_sizes;
1514 names = cache_names;
1517 * Initialize the caches that provide memory for the array cache and the
1518 * kmem_list3 structures first. Without this, further allocations will
1522 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1523 sizes[INDEX_AC].cs_size,
1524 ARCH_KMALLOC_MINALIGN,
1525 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1528 if (INDEX_AC != INDEX_L3) {
1529 sizes[INDEX_L3].cs_cachep =
1530 kmem_cache_create(names[INDEX_L3].name,
1531 sizes[INDEX_L3].cs_size,
1532 ARCH_KMALLOC_MINALIGN,
1533 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1537 slab_early_init = 0;
1539 while (sizes->cs_size != ULONG_MAX) {
1541 * For performance, all the general caches are L1 aligned.
1542 * This should be particularly beneficial on SMP boxes, as it
1543 * eliminates "false sharing".
1544 * Note for systems short on memory removing the alignment will
1545 * allow tighter packing of the smaller caches.
1547 if (!sizes->cs_cachep) {
1548 sizes->cs_cachep = kmem_cache_create(names->name,
1550 ARCH_KMALLOC_MINALIGN,
1551 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1554 #ifdef CONFIG_ZONE_DMA
1555 sizes->cs_dmacachep = kmem_cache_create(
1558 ARCH_KMALLOC_MINALIGN,
1559 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1566 /* 4) Replace the bootstrap head arrays */
1568 struct array_cache *ptr;
1570 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1572 local_irq_disable();
1573 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1574 memcpy(ptr, cpu_cache_get(&cache_cache),
1575 sizeof(struct arraycache_init));
1577 * Do not assume that spinlocks can be initialized via memcpy:
1579 spin_lock_init(&ptr->lock);
1581 cache_cache.array[smp_processor_id()] = ptr;
1584 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1586 local_irq_disable();
1587 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1588 != &initarray_generic.cache);
1589 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1590 sizeof(struct arraycache_init));
1592 * Do not assume that spinlocks can be initialized via memcpy:
1594 spin_lock_init(&ptr->lock);
1596 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1600 /* 5) Replace the bootstrap kmem_list3's */
1604 for_each_online_node(nid) {
1605 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], nid);
1607 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1608 &initkmem_list3[SIZE_AC + nid], nid);
1610 if (INDEX_AC != INDEX_L3) {
1611 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1612 &initkmem_list3[SIZE_L3 + nid], nid);
1617 /* 6) resize the head arrays to their final sizes */
1619 struct kmem_cache *cachep;
1620 mutex_lock(&cache_chain_mutex);
1621 list_for_each_entry(cachep, &cache_chain, next)
1622 if (enable_cpucache(cachep))
1624 mutex_unlock(&cache_chain_mutex);
1627 /* Annotate slab for lockdep -- annotate the malloc caches */
1632 g_cpucache_up = FULL;
1635 * Register a cpu startup notifier callback that initializes
1636 * cpu_cache_get for all new cpus
1638 register_cpu_notifier(&cpucache_notifier);
1641 * The reap timers are started later, with a module init call: That part
1642 * of the kernel is not yet operational.
1646 static int __init cpucache_init(void)
1651 * Register the timers that return unneeded pages to the page allocator
1653 for_each_online_cpu(cpu)
1654 start_cpu_timer(cpu);
1657 __initcall(cpucache_init);
1660 * Interface to system's page allocator. No need to hold the cache-lock.
1662 * If we requested dmaable memory, we will get it. Even if we
1663 * did not request dmaable memory, we might get it, but that
1664 * would be relatively rare and ignorable.
1666 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1674 * Nommu uses slab's for process anonymous memory allocations, and thus
1675 * requires __GFP_COMP to properly refcount higher order allocations
1677 flags |= __GFP_COMP;
1680 flags |= cachep->gfpflags;
1681 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1682 flags |= __GFP_RECLAIMABLE;
1684 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1688 nr_pages = (1 << cachep->gfporder);
1689 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1690 add_zone_page_state(page_zone(page),
1691 NR_SLAB_RECLAIMABLE, nr_pages);
1693 add_zone_page_state(page_zone(page),
1694 NR_SLAB_UNRECLAIMABLE, nr_pages);
1695 for (i = 0; i < nr_pages; i++)
1696 __SetPageSlab(page + i);
1697 return page_address(page);
1701 * Interface to system's page release.
1703 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1705 unsigned long i = (1 << cachep->gfporder);
1706 struct page *page = virt_to_page(addr);
1707 const unsigned long nr_freed = i;
1709 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1710 sub_zone_page_state(page_zone(page),
1711 NR_SLAB_RECLAIMABLE, nr_freed);
1713 sub_zone_page_state(page_zone(page),
1714 NR_SLAB_UNRECLAIMABLE, nr_freed);
1716 BUG_ON(!PageSlab(page));
1717 __ClearPageSlab(page);
1720 if (current->reclaim_state)
1721 current->reclaim_state->reclaimed_slab += nr_freed;
1722 free_pages((unsigned long)addr, cachep->gfporder);
1725 static void kmem_rcu_free(struct rcu_head *head)
1727 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1728 struct kmem_cache *cachep = slab_rcu->cachep;
1730 kmem_freepages(cachep, slab_rcu->addr);
1731 if (OFF_SLAB(cachep))
1732 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1737 #ifdef CONFIG_DEBUG_PAGEALLOC
1738 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1739 unsigned long caller)
1741 int size = obj_size(cachep);
1743 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1745 if (size < 5 * sizeof(unsigned long))
1748 *addr++ = 0x12345678;
1750 *addr++ = smp_processor_id();
1751 size -= 3 * sizeof(unsigned long);
1753 unsigned long *sptr = &caller;
1754 unsigned long svalue;
1756 while (!kstack_end(sptr)) {
1758 if (kernel_text_address(svalue)) {
1760 size -= sizeof(unsigned long);
1761 if (size <= sizeof(unsigned long))
1767 *addr++ = 0x87654321;
1771 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1773 int size = obj_size(cachep);
1774 addr = &((char *)addr)[obj_offset(cachep)];
1776 memset(addr, val, size);
1777 *(unsigned char *)(addr + size - 1) = POISON_END;
1780 static void dump_line(char *data, int offset, int limit)
1783 unsigned char error = 0;
1786 printk(KERN_ERR "%03x:", offset);
1787 for (i = 0; i < limit; i++) {
1788 if (data[offset + i] != POISON_FREE) {
1789 error = data[offset + i];
1792 printk(" %02x", (unsigned char)data[offset + i]);
1796 if (bad_count == 1) {
1797 error ^= POISON_FREE;
1798 if (!(error & (error - 1))) {
1799 printk(KERN_ERR "Single bit error detected. Probably "
1802 printk(KERN_ERR "Run memtest86+ or a similar memory "
1805 printk(KERN_ERR "Run a memory test tool.\n");
1814 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1819 if (cachep->flags & SLAB_RED_ZONE) {
1820 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1821 *dbg_redzone1(cachep, objp),
1822 *dbg_redzone2(cachep, objp));
1825 if (cachep->flags & SLAB_STORE_USER) {
1826 printk(KERN_ERR "Last user: [<%p>]",
1827 *dbg_userword(cachep, objp));
1828 print_symbol("(%s)",
1829 (unsigned long)*dbg_userword(cachep, objp));
1832 realobj = (char *)objp + obj_offset(cachep);
1833 size = obj_size(cachep);
1834 for (i = 0; i < size && lines; i += 16, lines--) {
1837 if (i + limit > size)
1839 dump_line(realobj, i, limit);
1843 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1849 realobj = (char *)objp + obj_offset(cachep);
1850 size = obj_size(cachep);
1852 for (i = 0; i < size; i++) {
1853 char exp = POISON_FREE;
1856 if (realobj[i] != exp) {
1862 "Slab corruption: %s start=%p, len=%d\n",
1863 cachep->name, realobj, size);
1864 print_objinfo(cachep, objp, 0);
1866 /* Hexdump the affected line */
1869 if (i + limit > size)
1871 dump_line(realobj, i, limit);
1874 /* Limit to 5 lines */
1880 /* Print some data about the neighboring objects, if they
1883 struct slab *slabp = virt_to_slab(objp);
1886 objnr = obj_to_index(cachep, slabp, objp);
1888 objp = index_to_obj(cachep, slabp, objnr - 1);
1889 realobj = (char *)objp + obj_offset(cachep);
1890 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1892 print_objinfo(cachep, objp, 2);
1894 if (objnr + 1 < cachep->num) {
1895 objp = index_to_obj(cachep, slabp, objnr + 1);
1896 realobj = (char *)objp + obj_offset(cachep);
1897 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1899 print_objinfo(cachep, objp, 2);
1907 * slab_destroy_objs - destroy a slab and its objects
1908 * @cachep: cache pointer being destroyed
1909 * @slabp: slab pointer being destroyed
1911 * Call the registered destructor for each object in a slab that is being
1914 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1917 for (i = 0; i < cachep->num; i++) {
1918 void *objp = index_to_obj(cachep, slabp, i);
1920 if (cachep->flags & SLAB_POISON) {
1921 #ifdef CONFIG_DEBUG_PAGEALLOC
1922 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1924 kernel_map_pages(virt_to_page(objp),
1925 cachep->buffer_size / PAGE_SIZE, 1);
1927 check_poison_obj(cachep, objp);
1929 check_poison_obj(cachep, objp);
1932 if (cachep->flags & SLAB_RED_ZONE) {
1933 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1934 slab_error(cachep, "start of a freed object "
1936 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1937 slab_error(cachep, "end of a freed object "
1943 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1949 * slab_destroy - destroy and release all objects in a slab
1950 * @cachep: cache pointer being destroyed
1951 * @slabp: slab pointer being destroyed
1953 * Destroy all the objs in a slab, and release the mem back to the system.
1954 * Before calling the slab must have been unlinked from the cache. The
1955 * cache-lock is not held/needed.
1957 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1959 void *addr = slabp->s_mem - slabp->colouroff;
1961 slab_destroy_objs(cachep, slabp);
1962 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1963 struct slab_rcu *slab_rcu;
1965 slab_rcu = (struct slab_rcu *)slabp;
1966 slab_rcu->cachep = cachep;
1967 slab_rcu->addr = addr;
1968 call_rcu(&slab_rcu->head, kmem_rcu_free);
1970 kmem_freepages(cachep, addr);
1971 if (OFF_SLAB(cachep))
1972 kmem_cache_free(cachep->slabp_cache, slabp);
1976 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1979 struct kmem_list3 *l3;
1981 for_each_online_cpu(i)
1982 kfree(cachep->array[i]);
1984 /* NUMA: free the list3 structures */
1985 for_each_online_node(i) {
1986 l3 = cachep->nodelists[i];
1989 free_alien_cache(l3->alien);
1993 kmem_cache_free(&cache_cache, cachep);
1998 * calculate_slab_order - calculate size (page order) of slabs
1999 * @cachep: pointer to the cache that is being created
2000 * @size: size of objects to be created in this cache.
2001 * @align: required alignment for the objects.
2002 * @flags: slab allocation flags
2004 * Also calculates the number of objects per slab.
2006 * This could be made much more intelligent. For now, try to avoid using
2007 * high order pages for slabs. When the gfp() functions are more friendly
2008 * towards high-order requests, this should be changed.
2010 static size_t calculate_slab_order(struct kmem_cache *cachep,
2011 size_t size, size_t align, unsigned long flags)
2013 unsigned long offslab_limit;
2014 size_t left_over = 0;
2017 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2021 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2025 if (flags & CFLGS_OFF_SLAB) {
2027 * Max number of objs-per-slab for caches which
2028 * use off-slab slabs. Needed to avoid a possible
2029 * looping condition in cache_grow().
2031 offslab_limit = size - sizeof(struct slab);
2032 offslab_limit /= sizeof(kmem_bufctl_t);
2034 if (num > offslab_limit)
2038 /* Found something acceptable - save it away */
2040 cachep->gfporder = gfporder;
2041 left_over = remainder;
2044 * A VFS-reclaimable slab tends to have most allocations
2045 * as GFP_NOFS and we really don't want to have to be allocating
2046 * higher-order pages when we are unable to shrink dcache.
2048 if (flags & SLAB_RECLAIM_ACCOUNT)
2052 * Large number of objects is good, but very large slabs are
2053 * currently bad for the gfp()s.
2055 if (gfporder >= slab_break_gfp_order)
2059 * Acceptable internal fragmentation?
2061 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2067 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
2069 if (g_cpucache_up == FULL)
2070 return enable_cpucache(cachep);
2072 if (g_cpucache_up == NONE) {
2074 * Note: the first kmem_cache_create must create the cache
2075 * that's used by kmalloc(24), otherwise the creation of
2076 * further caches will BUG().
2078 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2081 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2082 * the first cache, then we need to set up all its list3s,
2083 * otherwise the creation of further caches will BUG().
2085 set_up_list3s(cachep, SIZE_AC);
2086 if (INDEX_AC == INDEX_L3)
2087 g_cpucache_up = PARTIAL_L3;
2089 g_cpucache_up = PARTIAL_AC;
2091 cachep->array[smp_processor_id()] =
2092 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2094 if (g_cpucache_up == PARTIAL_AC) {
2095 set_up_list3s(cachep, SIZE_L3);
2096 g_cpucache_up = PARTIAL_L3;
2099 for_each_online_node(node) {
2100 cachep->nodelists[node] =
2101 kmalloc_node(sizeof(struct kmem_list3),
2103 BUG_ON(!cachep->nodelists[node]);
2104 kmem_list3_init(cachep->nodelists[node]);
2108 cachep->nodelists[numa_node_id()]->next_reap =
2109 jiffies + REAPTIMEOUT_LIST3 +
2110 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2112 cpu_cache_get(cachep)->avail = 0;
2113 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2114 cpu_cache_get(cachep)->batchcount = 1;
2115 cpu_cache_get(cachep)->touched = 0;
2116 cachep->batchcount = 1;
2117 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2122 * kmem_cache_create - Create a cache.
2123 * @name: A string which is used in /proc/slabinfo to identify this cache.
2124 * @size: The size of objects to be created in this cache.
2125 * @align: The required alignment for the objects.
2126 * @flags: SLAB flags
2127 * @ctor: A constructor for the objects.
2129 * Returns a ptr to the cache on success, NULL on failure.
2130 * Cannot be called within a int, but can be interrupted.
2131 * The @ctor is run when new pages are allocated by the cache.
2133 * @name must be valid until the cache is destroyed. This implies that
2134 * the module calling this has to destroy the cache before getting unloaded.
2138 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2139 * to catch references to uninitialised memory.
2141 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2142 * for buffer overruns.
2144 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2145 * cacheline. This can be beneficial if you're counting cycles as closely
2149 kmem_cache_create (const char *name, size_t size, size_t align,
2150 unsigned long flags,
2151 void (*ctor)(struct kmem_cache *, void *))
2153 size_t left_over, slab_size, ralign;
2154 struct kmem_cache *cachep = NULL, *pc;
2157 * Sanity checks... these are all serious usage bugs.
2159 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2160 size > KMALLOC_MAX_SIZE) {
2161 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2167 * We use cache_chain_mutex to ensure a consistent view of
2168 * cpu_online_map as well. Please see cpuup_callback
2171 mutex_lock(&cache_chain_mutex);
2173 list_for_each_entry(pc, &cache_chain, next) {
2178 * This happens when the module gets unloaded and doesn't
2179 * destroy its slab cache and no-one else reuses the vmalloc
2180 * area of the module. Print a warning.
2182 res = probe_kernel_address(pc->name, tmp);
2185 "SLAB: cache with size %d has lost its name\n",
2190 if (!strcmp(pc->name, name)) {
2192 "kmem_cache_create: duplicate cache %s\n", name);
2199 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2202 * Enable redzoning and last user accounting, except for caches with
2203 * large objects, if the increased size would increase the object size
2204 * above the next power of two: caches with object sizes just above a
2205 * power of two have a significant amount of internal fragmentation.
2207 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2208 2 * sizeof(unsigned long long)))
2209 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2210 if (!(flags & SLAB_DESTROY_BY_RCU))
2211 flags |= SLAB_POISON;
2213 if (flags & SLAB_DESTROY_BY_RCU)
2214 BUG_ON(flags & SLAB_POISON);
2217 * Always checks flags, a caller might be expecting debug support which
2220 BUG_ON(flags & ~CREATE_MASK);
2223 * Check that size is in terms of words. This is needed to avoid
2224 * unaligned accesses for some archs when redzoning is used, and makes
2225 * sure any on-slab bufctl's are also correctly aligned.
2227 if (size & (BYTES_PER_WORD - 1)) {
2228 size += (BYTES_PER_WORD - 1);
2229 size &= ~(BYTES_PER_WORD - 1);
2232 /* calculate the final buffer alignment: */
2234 /* 1) arch recommendation: can be overridden for debug */
2235 if (flags & SLAB_HWCACHE_ALIGN) {
2237 * Default alignment: as specified by the arch code. Except if
2238 * an object is really small, then squeeze multiple objects into
2241 ralign = cache_line_size();
2242 while (size <= ralign / 2)
2245 ralign = BYTES_PER_WORD;
2249 * Redzoning and user store require word alignment or possibly larger.
2250 * Note this will be overridden by architecture or caller mandated
2251 * alignment if either is greater than BYTES_PER_WORD.
2253 if (flags & SLAB_STORE_USER)
2254 ralign = BYTES_PER_WORD;
2256 if (flags & SLAB_RED_ZONE) {
2257 ralign = REDZONE_ALIGN;
2258 /* If redzoning, ensure that the second redzone is suitably
2259 * aligned, by adjusting the object size accordingly. */
2260 size += REDZONE_ALIGN - 1;
2261 size &= ~(REDZONE_ALIGN - 1);
2264 /* 2) arch mandated alignment */
2265 if (ralign < ARCH_SLAB_MINALIGN) {
2266 ralign = ARCH_SLAB_MINALIGN;
2268 /* 3) caller mandated alignment */
2269 if (ralign < align) {
2272 /* disable debug if necessary */
2273 if (ralign > __alignof__(unsigned long long))
2274 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2280 /* Get cache's description obj. */
2281 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2286 cachep->obj_size = size;
2289 * Both debugging options require word-alignment which is calculated
2292 if (flags & SLAB_RED_ZONE) {
2293 /* add space for red zone words */
2294 cachep->obj_offset += sizeof(unsigned long long);
2295 size += 2 * sizeof(unsigned long long);
2297 if (flags & SLAB_STORE_USER) {
2298 /* user store requires one word storage behind the end of
2299 * the real object. But if the second red zone needs to be
2300 * aligned to 64 bits, we must allow that much space.
2302 if (flags & SLAB_RED_ZONE)
2303 size += REDZONE_ALIGN;
2305 size += BYTES_PER_WORD;
2307 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2308 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2309 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2310 cachep->obj_offset += PAGE_SIZE - size;
2317 * Determine if the slab management is 'on' or 'off' slab.
2318 * (bootstrapping cannot cope with offslab caches so don't do
2321 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2323 * Size is large, assume best to place the slab management obj
2324 * off-slab (should allow better packing of objs).
2326 flags |= CFLGS_OFF_SLAB;
2328 size = ALIGN(size, align);
2330 left_over = calculate_slab_order(cachep, size, align, flags);
2334 "kmem_cache_create: couldn't create cache %s.\n", name);
2335 kmem_cache_free(&cache_cache, cachep);
2339 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2340 + sizeof(struct slab), align);
2343 * If the slab has been placed off-slab, and we have enough space then
2344 * move it on-slab. This is at the expense of any extra colouring.
2346 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2347 flags &= ~CFLGS_OFF_SLAB;
2348 left_over -= slab_size;
2351 if (flags & CFLGS_OFF_SLAB) {
2352 /* really off slab. No need for manual alignment */
2354 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2357 cachep->colour_off = cache_line_size();
2358 /* Offset must be a multiple of the alignment. */
2359 if (cachep->colour_off < align)
2360 cachep->colour_off = align;
2361 cachep->colour = left_over / cachep->colour_off;
2362 cachep->slab_size = slab_size;
2363 cachep->flags = flags;
2364 cachep->gfpflags = 0;
2365 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2366 cachep->gfpflags |= GFP_DMA;
2367 cachep->buffer_size = size;
2368 cachep->reciprocal_buffer_size = reciprocal_value(size);
2370 if (flags & CFLGS_OFF_SLAB) {
2371 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2373 * This is a possibility for one of the malloc_sizes caches.
2374 * But since we go off slab only for object size greater than
2375 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2376 * this should not happen at all.
2377 * But leave a BUG_ON for some lucky dude.
2379 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2381 cachep->ctor = ctor;
2382 cachep->name = name;
2384 if (setup_cpu_cache(cachep)) {
2385 __kmem_cache_destroy(cachep);
2390 /* cache setup completed, link it into the list */
2391 list_add(&cachep->next, &cache_chain);
2393 if (!cachep && (flags & SLAB_PANIC))
2394 panic("kmem_cache_create(): failed to create slab `%s'\n",
2396 mutex_unlock(&cache_chain_mutex);
2400 EXPORT_SYMBOL(kmem_cache_create);
2403 static void check_irq_off(void)
2405 BUG_ON(!irqs_disabled());
2408 static void check_irq_on(void)
2410 BUG_ON(irqs_disabled());
2413 static void check_spinlock_acquired(struct kmem_cache *cachep)
2417 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2421 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2425 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2430 #define check_irq_off() do { } while(0)
2431 #define check_irq_on() do { } while(0)
2432 #define check_spinlock_acquired(x) do { } while(0)
2433 #define check_spinlock_acquired_node(x, y) do { } while(0)
2436 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2437 struct array_cache *ac,
2438 int force, int node);
2440 static void do_drain(void *arg)
2442 struct kmem_cache *cachep = arg;
2443 struct array_cache *ac;
2444 int node = numa_node_id();
2447 ac = cpu_cache_get(cachep);
2448 spin_lock(&cachep->nodelists[node]->list_lock);
2449 free_block(cachep, ac->entry, ac->avail, node);
2450 spin_unlock(&cachep->nodelists[node]->list_lock);
2454 static void drain_cpu_caches(struct kmem_cache *cachep)
2456 struct kmem_list3 *l3;
2459 on_each_cpu(do_drain, cachep, 1, 1);
2461 for_each_online_node(node) {
2462 l3 = cachep->nodelists[node];
2463 if (l3 && l3->alien)
2464 drain_alien_cache(cachep, l3->alien);
2467 for_each_online_node(node) {
2468 l3 = cachep->nodelists[node];
2470 drain_array(cachep, l3, l3->shared, 1, node);
2475 * Remove slabs from the list of free slabs.
2476 * Specify the number of slabs to drain in tofree.
2478 * Returns the actual number of slabs released.
2480 static int drain_freelist(struct kmem_cache *cache,
2481 struct kmem_list3 *l3, int tofree)
2483 struct list_head *p;
2488 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2490 spin_lock_irq(&l3->list_lock);
2491 p = l3->slabs_free.prev;
2492 if (p == &l3->slabs_free) {
2493 spin_unlock_irq(&l3->list_lock);
2497 slabp = list_entry(p, struct slab, list);
2499 BUG_ON(slabp->inuse);
2501 list_del(&slabp->list);
2503 * Safe to drop the lock. The slab is no longer linked
2506 l3->free_objects -= cache->num;
2507 spin_unlock_irq(&l3->list_lock);
2508 slab_destroy(cache, slabp);
2515 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2516 static int __cache_shrink(struct kmem_cache *cachep)
2519 struct kmem_list3 *l3;
2521 drain_cpu_caches(cachep);
2524 for_each_online_node(i) {
2525 l3 = cachep->nodelists[i];
2529 drain_freelist(cachep, l3, l3->free_objects);
2531 ret += !list_empty(&l3->slabs_full) ||
2532 !list_empty(&l3->slabs_partial);
2534 return (ret ? 1 : 0);
2538 * kmem_cache_shrink - Shrink a cache.
2539 * @cachep: The cache to shrink.
2541 * Releases as many slabs as possible for a cache.
2542 * To help debugging, a zero exit status indicates all slabs were released.
2544 int kmem_cache_shrink(struct kmem_cache *cachep)
2547 BUG_ON(!cachep || in_interrupt());
2550 mutex_lock(&cache_chain_mutex);
2551 ret = __cache_shrink(cachep);
2552 mutex_unlock(&cache_chain_mutex);
2556 EXPORT_SYMBOL(kmem_cache_shrink);
2559 * kmem_cache_destroy - delete a cache
2560 * @cachep: the cache to destroy
2562 * Remove a &struct kmem_cache object from the slab cache.
2564 * It is expected this function will be called by a module when it is
2565 * unloaded. This will remove the cache completely, and avoid a duplicate
2566 * cache being allocated each time a module is loaded and unloaded, if the
2567 * module doesn't have persistent in-kernel storage across loads and unloads.
2569 * The cache must be empty before calling this function.
2571 * The caller must guarantee that noone will allocate memory from the cache
2572 * during the kmem_cache_destroy().
2574 void kmem_cache_destroy(struct kmem_cache *cachep)
2576 BUG_ON(!cachep || in_interrupt());
2578 /* Find the cache in the chain of caches. */
2580 mutex_lock(&cache_chain_mutex);
2582 * the chain is never empty, cache_cache is never destroyed
2584 list_del(&cachep->next);
2585 if (__cache_shrink(cachep)) {
2586 slab_error(cachep, "Can't free all objects");
2587 list_add(&cachep->next, &cache_chain);
2588 mutex_unlock(&cache_chain_mutex);
2593 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2596 __kmem_cache_destroy(cachep);
2597 mutex_unlock(&cache_chain_mutex);
2600 EXPORT_SYMBOL(kmem_cache_destroy);
2603 * Get the memory for a slab management obj.
2604 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2605 * always come from malloc_sizes caches. The slab descriptor cannot
2606 * come from the same cache which is getting created because,
2607 * when we are searching for an appropriate cache for these
2608 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2609 * If we are creating a malloc_sizes cache here it would not be visible to
2610 * kmem_find_general_cachep till the initialization is complete.
2611 * Hence we cannot have slabp_cache same as the original cache.
2613 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2614 int colour_off, gfp_t local_flags,
2619 if (OFF_SLAB(cachep)) {
2620 /* Slab management obj is off-slab. */
2621 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2622 local_flags & ~GFP_THISNODE, nodeid);
2626 slabp = objp + colour_off;
2627 colour_off += cachep->slab_size;
2630 slabp->colouroff = colour_off;
2631 slabp->s_mem = objp + colour_off;
2632 slabp->nodeid = nodeid;
2636 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2638 return (kmem_bufctl_t *) (slabp + 1);
2641 static void cache_init_objs(struct kmem_cache *cachep,
2646 for (i = 0; i < cachep->num; i++) {
2647 void *objp = index_to_obj(cachep, slabp, i);
2649 /* need to poison the objs? */
2650 if (cachep->flags & SLAB_POISON)
2651 poison_obj(cachep, objp, POISON_FREE);
2652 if (cachep->flags & SLAB_STORE_USER)
2653 *dbg_userword(cachep, objp) = NULL;
2655 if (cachep->flags & SLAB_RED_ZONE) {
2656 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2657 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2660 * Constructors are not allowed to allocate memory from the same
2661 * cache which they are a constructor for. Otherwise, deadlock.
2662 * They must also be threaded.
2664 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2665 cachep->ctor(cachep, objp + obj_offset(cachep));
2667 if (cachep->flags & SLAB_RED_ZONE) {
2668 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2669 slab_error(cachep, "constructor overwrote the"
2670 " end of an object");
2671 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2672 slab_error(cachep, "constructor overwrote the"
2673 " start of an object");
2675 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2676 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2677 kernel_map_pages(virt_to_page(objp),
2678 cachep->buffer_size / PAGE_SIZE, 0);
2681 cachep->ctor(cachep, objp);
2683 slab_bufctl(slabp)[i] = i + 1;
2685 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2689 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2691 if (CONFIG_ZONE_DMA_FLAG) {
2692 if (flags & GFP_DMA)
2693 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2695 BUG_ON(cachep->gfpflags & GFP_DMA);
2699 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2702 void *objp = index_to_obj(cachep, slabp, slabp->free);
2706 next = slab_bufctl(slabp)[slabp->free];
2708 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2709 WARN_ON(slabp->nodeid != nodeid);
2716 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2717 void *objp, int nodeid)
2719 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2722 /* Verify that the slab belongs to the intended node */
2723 WARN_ON(slabp->nodeid != nodeid);
2725 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2726 printk(KERN_ERR "slab: double free detected in cache "
2727 "'%s', objp %p\n", cachep->name, objp);
2731 slab_bufctl(slabp)[objnr] = slabp->free;
2732 slabp->free = objnr;
2737 * Map pages beginning at addr to the given cache and slab. This is required
2738 * for the slab allocator to be able to lookup the cache and slab of a
2739 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2741 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2747 page = virt_to_page(addr);
2750 if (likely(!PageCompound(page)))
2751 nr_pages <<= cache->gfporder;
2754 page_set_cache(page, cache);
2755 page_set_slab(page, slab);
2757 } while (--nr_pages);
2761 * Grow (by 1) the number of slabs within a cache. This is called by
2762 * kmem_cache_alloc() when there are no active objs left in a cache.
2764 static int cache_grow(struct kmem_cache *cachep,
2765 gfp_t flags, int nodeid, void *objp)
2770 struct kmem_list3 *l3;
2773 * Be lazy and only check for valid flags here, keeping it out of the
2774 * critical path in kmem_cache_alloc().
2776 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2777 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2779 /* Take the l3 list lock to change the colour_next on this node */
2781 l3 = cachep->nodelists[nodeid];
2782 spin_lock(&l3->list_lock);
2784 /* Get colour for the slab, and cal the next value. */
2785 offset = l3->colour_next;
2787 if (l3->colour_next >= cachep->colour)
2788 l3->colour_next = 0;
2789 spin_unlock(&l3->list_lock);
2791 offset *= cachep->colour_off;
2793 if (local_flags & __GFP_WAIT)
2797 * The test for missing atomic flag is performed here, rather than
2798 * the more obvious place, simply to reduce the critical path length
2799 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2800 * will eventually be caught here (where it matters).
2802 kmem_flagcheck(cachep, flags);
2805 * Get mem for the objs. Attempt to allocate a physical page from
2809 objp = kmem_getpages(cachep, local_flags, nodeid);
2813 /* Get slab management. */
2814 slabp = alloc_slabmgmt(cachep, objp, offset,
2815 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2819 slabp->nodeid = nodeid;
2820 slab_map_pages(cachep, slabp, objp);
2822 cache_init_objs(cachep, slabp);
2824 if (local_flags & __GFP_WAIT)
2825 local_irq_disable();
2827 spin_lock(&l3->list_lock);
2829 /* Make slab active. */
2830 list_add_tail(&slabp->list, &(l3->slabs_free));
2831 STATS_INC_GROWN(cachep);
2832 l3->free_objects += cachep->num;
2833 spin_unlock(&l3->list_lock);
2836 kmem_freepages(cachep, objp);
2838 if (local_flags & __GFP_WAIT)
2839 local_irq_disable();
2846 * Perform extra freeing checks:
2847 * - detect bad pointers.
2848 * - POISON/RED_ZONE checking
2850 static void kfree_debugcheck(const void *objp)
2852 if (!virt_addr_valid(objp)) {
2853 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2854 (unsigned long)objp);
2859 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2861 unsigned long long redzone1, redzone2;
2863 redzone1 = *dbg_redzone1(cache, obj);
2864 redzone2 = *dbg_redzone2(cache, obj);
2869 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2872 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2873 slab_error(cache, "double free detected");
2875 slab_error(cache, "memory outside object was overwritten");
2877 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2878 obj, redzone1, redzone2);
2881 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2888 BUG_ON(virt_to_cache(objp) != cachep);
2890 objp -= obj_offset(cachep);
2891 kfree_debugcheck(objp);
2892 page = virt_to_head_page(objp);
2894 slabp = page_get_slab(page);
2896 if (cachep->flags & SLAB_RED_ZONE) {
2897 verify_redzone_free(cachep, objp);
2898 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2899 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2901 if (cachep->flags & SLAB_STORE_USER)
2902 *dbg_userword(cachep, objp) = caller;
2904 objnr = obj_to_index(cachep, slabp, objp);
2906 BUG_ON(objnr >= cachep->num);
2907 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2909 #ifdef CONFIG_DEBUG_SLAB_LEAK
2910 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2912 if (cachep->flags & SLAB_POISON) {
2913 #ifdef CONFIG_DEBUG_PAGEALLOC
2914 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2915 store_stackinfo(cachep, objp, (unsigned long)caller);
2916 kernel_map_pages(virt_to_page(objp),
2917 cachep->buffer_size / PAGE_SIZE, 0);
2919 poison_obj(cachep, objp, POISON_FREE);
2922 poison_obj(cachep, objp, POISON_FREE);
2928 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2933 /* Check slab's freelist to see if this obj is there. */
2934 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2936 if (entries > cachep->num || i >= cachep->num)
2939 if (entries != cachep->num - slabp->inuse) {
2941 printk(KERN_ERR "slab: Internal list corruption detected in "
2942 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2943 cachep->name, cachep->num, slabp, slabp->inuse);
2945 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2948 printk("\n%03x:", i);
2949 printk(" %02x", ((unsigned char *)slabp)[i]);
2956 #define kfree_debugcheck(x) do { } while(0)
2957 #define cache_free_debugcheck(x,objp,z) (objp)
2958 #define check_slabp(x,y) do { } while(0)
2961 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2964 struct kmem_list3 *l3;
2965 struct array_cache *ac;
2968 node = numa_node_id();
2971 ac = cpu_cache_get(cachep);
2973 batchcount = ac->batchcount;
2974 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2976 * If there was little recent activity on this cache, then
2977 * perform only a partial refill. Otherwise we could generate
2980 batchcount = BATCHREFILL_LIMIT;
2982 l3 = cachep->nodelists[node];
2984 BUG_ON(ac->avail > 0 || !l3);
2985 spin_lock(&l3->list_lock);
2987 /* See if we can refill from the shared array */
2988 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2991 while (batchcount > 0) {
2992 struct list_head *entry;
2994 /* Get slab alloc is to come from. */
2995 entry = l3->slabs_partial.next;
2996 if (entry == &l3->slabs_partial) {
2997 l3->free_touched = 1;
2998 entry = l3->slabs_free.next;
2999 if (entry == &l3->slabs_free)
3003 slabp = list_entry(entry, struct slab, list);
3004 check_slabp(cachep, slabp);
3005 check_spinlock_acquired(cachep);
3008 * The slab was either on partial or free list so
3009 * there must be at least one object available for
3012 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3014 while (slabp->inuse < cachep->num && batchcount--) {
3015 STATS_INC_ALLOCED(cachep);
3016 STATS_INC_ACTIVE(cachep);
3017 STATS_SET_HIGH(cachep);
3019 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3022 check_slabp(cachep, slabp);
3024 /* move slabp to correct slabp list: */
3025 list_del(&slabp->list);
3026 if (slabp->free == BUFCTL_END)
3027 list_add(&slabp->list, &l3->slabs_full);
3029 list_add(&slabp->list, &l3->slabs_partial);
3033 l3->free_objects -= ac->avail;
3035 spin_unlock(&l3->list_lock);
3037 if (unlikely(!ac->avail)) {
3039 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3041 /* cache_grow can reenable interrupts, then ac could change. */
3042 ac = cpu_cache_get(cachep);
3043 if (!x && ac->avail == 0) /* no objects in sight? abort */
3046 if (!ac->avail) /* objects refilled by interrupt? */
3050 return ac->entry[--ac->avail];
3053 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3056 might_sleep_if(flags & __GFP_WAIT);
3058 kmem_flagcheck(cachep, flags);
3063 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3064 gfp_t flags, void *objp, void *caller)
3068 if (cachep->flags & SLAB_POISON) {
3069 #ifdef CONFIG_DEBUG_PAGEALLOC
3070 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3071 kernel_map_pages(virt_to_page(objp),
3072 cachep->buffer_size / PAGE_SIZE, 1);
3074 check_poison_obj(cachep, objp);
3076 check_poison_obj(cachep, objp);
3078 poison_obj(cachep, objp, POISON_INUSE);
3080 if (cachep->flags & SLAB_STORE_USER)
3081 *dbg_userword(cachep, objp) = caller;
3083 if (cachep->flags & SLAB_RED_ZONE) {
3084 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3085 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3086 slab_error(cachep, "double free, or memory outside"
3087 " object was overwritten");
3089 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3090 objp, *dbg_redzone1(cachep, objp),
3091 *dbg_redzone2(cachep, objp));
3093 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3094 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3096 #ifdef CONFIG_DEBUG_SLAB_LEAK
3101 slabp = page_get_slab(virt_to_head_page(objp));
3102 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3103 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3106 objp += obj_offset(cachep);
3107 if (cachep->ctor && cachep->flags & SLAB_POISON)
3108 cachep->ctor(cachep, objp);
3109 #if ARCH_SLAB_MINALIGN
3110 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3111 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3112 objp, ARCH_SLAB_MINALIGN);
3118 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3121 #ifdef CONFIG_FAILSLAB
3123 static struct failslab_attr {
3125 struct fault_attr attr;
3127 u32 ignore_gfp_wait;
3128 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3129 struct dentry *ignore_gfp_wait_file;
3133 .attr = FAULT_ATTR_INITIALIZER,
3134 .ignore_gfp_wait = 1,
3137 static int __init setup_failslab(char *str)
3139 return setup_fault_attr(&failslab.attr, str);
3141 __setup("failslab=", setup_failslab);
3143 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3145 if (cachep == &cache_cache)
3147 if (flags & __GFP_NOFAIL)
3149 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3152 return should_fail(&failslab.attr, obj_size(cachep));
3155 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3157 static int __init failslab_debugfs(void)
3159 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3163 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3166 dir = failslab.attr.dentries.dir;
3168 failslab.ignore_gfp_wait_file =
3169 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3170 &failslab.ignore_gfp_wait);
3172 if (!failslab.ignore_gfp_wait_file) {
3174 debugfs_remove(failslab.ignore_gfp_wait_file);
3175 cleanup_fault_attr_dentries(&failslab.attr);
3181 late_initcall(failslab_debugfs);
3183 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3185 #else /* CONFIG_FAILSLAB */
3187 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3192 #endif /* CONFIG_FAILSLAB */
3194 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3197 struct array_cache *ac;
3201 ac = cpu_cache_get(cachep);
3202 if (likely(ac->avail)) {
3203 STATS_INC_ALLOCHIT(cachep);
3205 objp = ac->entry[--ac->avail];
3207 STATS_INC_ALLOCMISS(cachep);
3208 objp = cache_alloc_refill(cachep, flags);
3215 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3217 * If we are in_interrupt, then process context, including cpusets and
3218 * mempolicy, may not apply and should not be used for allocation policy.
3220 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3222 int nid_alloc, nid_here;
3224 if (in_interrupt() || (flags & __GFP_THISNODE))
3226 nid_alloc = nid_here = numa_node_id();
3227 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3228 nid_alloc = cpuset_mem_spread_node();
3229 else if (current->mempolicy)
3230 nid_alloc = slab_node(current->mempolicy);
3231 if (nid_alloc != nid_here)
3232 return ____cache_alloc_node(cachep, flags, nid_alloc);
3237 * Fallback function if there was no memory available and no objects on a
3238 * certain node and fall back is permitted. First we scan all the
3239 * available nodelists for available objects. If that fails then we
3240 * perform an allocation without specifying a node. This allows the page
3241 * allocator to do its reclaim / fallback magic. We then insert the
3242 * slab into the proper nodelist and then allocate from it.
3244 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3246 struct zonelist *zonelist;
3252 if (flags & __GFP_THISNODE)
3255 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3256 ->node_zonelists[gfp_zone(flags)];
3257 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3261 * Look through allowed nodes for objects available
3262 * from existing per node queues.
3264 for (z = zonelist->zones; *z && !obj; z++) {
3265 nid = zone_to_nid(*z);
3267 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3268 cache->nodelists[nid] &&
3269 cache->nodelists[nid]->free_objects)
3270 obj = ____cache_alloc_node(cache,
3271 flags | GFP_THISNODE, nid);
3276 * This allocation will be performed within the constraints
3277 * of the current cpuset / memory policy requirements.
3278 * We may trigger various forms of reclaim on the allowed
3279 * set and go into memory reserves if necessary.
3281 if (local_flags & __GFP_WAIT)
3283 kmem_flagcheck(cache, flags);
3284 obj = kmem_getpages(cache, flags, -1);
3285 if (local_flags & __GFP_WAIT)
3286 local_irq_disable();
3289 * Insert into the appropriate per node queues
3291 nid = page_to_nid(virt_to_page(obj));
3292 if (cache_grow(cache, flags, nid, obj)) {
3293 obj = ____cache_alloc_node(cache,
3294 flags | GFP_THISNODE, nid);
3297 * Another processor may allocate the
3298 * objects in the slab since we are
3299 * not holding any locks.
3303 /* cache_grow already freed obj */
3312 * A interface to enable slab creation on nodeid
3314 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3317 struct list_head *entry;
3319 struct kmem_list3 *l3;
3323 l3 = cachep->nodelists[nodeid];
3328 spin_lock(&l3->list_lock);
3329 entry = l3->slabs_partial.next;
3330 if (entry == &l3->slabs_partial) {
3331 l3->free_touched = 1;
3332 entry = l3->slabs_free.next;
3333 if (entry == &l3->slabs_free)
3337 slabp = list_entry(entry, struct slab, list);
3338 check_spinlock_acquired_node(cachep, nodeid);
3339 check_slabp(cachep, slabp);
3341 STATS_INC_NODEALLOCS(cachep);
3342 STATS_INC_ACTIVE(cachep);
3343 STATS_SET_HIGH(cachep);
3345 BUG_ON(slabp->inuse == cachep->num);
3347 obj = slab_get_obj(cachep, slabp, nodeid);
3348 check_slabp(cachep, slabp);
3350 /* move slabp to correct slabp list: */
3351 list_del(&slabp->list);
3353 if (slabp->free == BUFCTL_END)
3354 list_add(&slabp->list, &l3->slabs_full);
3356 list_add(&slabp->list, &l3->slabs_partial);
3358 spin_unlock(&l3->list_lock);
3362 spin_unlock(&l3->list_lock);
3363 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3367 return fallback_alloc(cachep, flags);
3374 * kmem_cache_alloc_node - Allocate an object on the specified node
3375 * @cachep: The cache to allocate from.
3376 * @flags: See kmalloc().
3377 * @nodeid: node number of the target node.
3378 * @caller: return address of caller, used for debug information
3380 * Identical to kmem_cache_alloc but it will allocate memory on the given
3381 * node, which can improve the performance for cpu bound structures.
3383 * Fallback to other node is possible if __GFP_THISNODE is not set.
3385 static __always_inline void *
3386 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3389 unsigned long save_flags;
3392 if (should_failslab(cachep, flags))
3395 cache_alloc_debugcheck_before(cachep, flags);
3396 local_irq_save(save_flags);
3398 if (unlikely(nodeid == -1))
3399 nodeid = numa_node_id();
3401 if (unlikely(!cachep->nodelists[nodeid])) {
3402 /* Node not bootstrapped yet */
3403 ptr = fallback_alloc(cachep, flags);
3407 if (nodeid == numa_node_id()) {
3409 * Use the locally cached objects if possible.
3410 * However ____cache_alloc does not allow fallback
3411 * to other nodes. It may fail while we still have
3412 * objects on other nodes available.
3414 ptr = ____cache_alloc(cachep, flags);
3418 /* ___cache_alloc_node can fall back to other nodes */
3419 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3421 local_irq_restore(save_flags);
3422 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3424 if (unlikely((flags & __GFP_ZERO) && ptr))
3425 memset(ptr, 0, obj_size(cachep));
3430 static __always_inline void *
3431 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3435 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3436 objp = alternate_node_alloc(cache, flags);
3440 objp = ____cache_alloc(cache, flags);
3443 * We may just have run out of memory on the local node.
3444 * ____cache_alloc_node() knows how to locate memory on other nodes
3447 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3454 static __always_inline void *
3455 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3457 return ____cache_alloc(cachep, flags);
3460 #endif /* CONFIG_NUMA */
3462 static __always_inline void *
3463 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3465 unsigned long save_flags;
3468 if (should_failslab(cachep, flags))
3471 cache_alloc_debugcheck_before(cachep, flags);
3472 local_irq_save(save_flags);
3473 objp = __do_cache_alloc(cachep, flags);
3474 local_irq_restore(save_flags);
3475 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3478 if (unlikely((flags & __GFP_ZERO) && objp))
3479 memset(objp, 0, obj_size(cachep));
3485 * Caller needs to acquire correct kmem_list's list_lock
3487 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3491 struct kmem_list3 *l3;
3493 for (i = 0; i < nr_objects; i++) {
3494 void *objp = objpp[i];
3497 slabp = virt_to_slab(objp);
3498 l3 = cachep->nodelists[node];
3499 list_del(&slabp->list);
3500 check_spinlock_acquired_node(cachep, node);
3501 check_slabp(cachep, slabp);
3502 slab_put_obj(cachep, slabp, objp, node);
3503 STATS_DEC_ACTIVE(cachep);
3505 check_slabp(cachep, slabp);
3507 /* fixup slab chains */
3508 if (slabp->inuse == 0) {
3509 if (l3->free_objects > l3->free_limit) {
3510 l3->free_objects -= cachep->num;
3511 /* No need to drop any previously held
3512 * lock here, even if we have a off-slab slab
3513 * descriptor it is guaranteed to come from
3514 * a different cache, refer to comments before
3517 slab_destroy(cachep, slabp);
3519 list_add(&slabp->list, &l3->slabs_free);
3522 /* Unconditionally move a slab to the end of the
3523 * partial list on free - maximum time for the
3524 * other objects to be freed, too.
3526 list_add_tail(&slabp->list, &l3->slabs_partial);
3531 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3534 struct kmem_list3 *l3;
3535 int node = numa_node_id();
3537 batchcount = ac->batchcount;
3539 BUG_ON(!batchcount || batchcount > ac->avail);
3542 l3 = cachep->nodelists[node];
3543 spin_lock(&l3->list_lock);
3545 struct array_cache *shared_array = l3->shared;
3546 int max = shared_array->limit - shared_array->avail;
3548 if (batchcount > max)
3550 memcpy(&(shared_array->entry[shared_array->avail]),
3551 ac->entry, sizeof(void *) * batchcount);
3552 shared_array->avail += batchcount;
3557 free_block(cachep, ac->entry, batchcount, node);
3562 struct list_head *p;
3564 p = l3->slabs_free.next;
3565 while (p != &(l3->slabs_free)) {
3568 slabp = list_entry(p, struct slab, list);
3569 BUG_ON(slabp->inuse);
3574 STATS_SET_FREEABLE(cachep, i);
3577 spin_unlock(&l3->list_lock);
3578 ac->avail -= batchcount;
3579 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3583 * Release an obj back to its cache. If the obj has a constructed state, it must
3584 * be in this state _before_ it is released. Called with disabled ints.
3586 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3588 struct array_cache *ac = cpu_cache_get(cachep);
3591 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3594 * Skip calling cache_free_alien() when the platform is not numa.
3595 * This will avoid cache misses that happen while accessing slabp (which
3596 * is per page memory reference) to get nodeid. Instead use a global
3597 * variable to skip the call, which is mostly likely to be present in
3600 if (numa_platform && cache_free_alien(cachep, objp))
3603 if (likely(ac->avail < ac->limit)) {
3604 STATS_INC_FREEHIT(cachep);
3605 ac->entry[ac->avail++] = objp;
3608 STATS_INC_FREEMISS(cachep);
3609 cache_flusharray(cachep, ac);
3610 ac->entry[ac->avail++] = objp;
3615 * kmem_cache_alloc - Allocate an object
3616 * @cachep: The cache to allocate from.
3617 * @flags: See kmalloc().
3619 * Allocate an object from this cache. The flags are only relevant
3620 * if the cache has no available objects.
3622 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3624 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3626 EXPORT_SYMBOL(kmem_cache_alloc);
3629 * kmem_ptr_validate - check if an untrusted pointer might
3631 * @cachep: the cache we're checking against
3632 * @ptr: pointer to validate
3634 * This verifies that the untrusted pointer looks sane:
3635 * it is _not_ a guarantee that the pointer is actually
3636 * part of the slab cache in question, but it at least
3637 * validates that the pointer can be dereferenced and
3638 * looks half-way sane.
3640 * Currently only used for dentry validation.
3642 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3644 unsigned long addr = (unsigned long)ptr;
3645 unsigned long min_addr = PAGE_OFFSET;
3646 unsigned long align_mask = BYTES_PER_WORD - 1;
3647 unsigned long size = cachep->buffer_size;
3650 if (unlikely(addr < min_addr))
3652 if (unlikely(addr > (unsigned long)high_memory - size))
3654 if (unlikely(addr & align_mask))
3656 if (unlikely(!kern_addr_valid(addr)))
3658 if (unlikely(!kern_addr_valid(addr + size - 1)))
3660 page = virt_to_page(ptr);
3661 if (unlikely(!PageSlab(page)))
3663 if (unlikely(page_get_cache(page) != cachep))
3671 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3673 return __cache_alloc_node(cachep, flags, nodeid,
3674 __builtin_return_address(0));
3676 EXPORT_SYMBOL(kmem_cache_alloc_node);
3678 static __always_inline void *
3679 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3681 struct kmem_cache *cachep;
3683 cachep = kmem_find_general_cachep(size, flags);
3684 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3686 return kmem_cache_alloc_node(cachep, flags, node);
3689 #ifdef CONFIG_DEBUG_SLAB
3690 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3692 return __do_kmalloc_node(size, flags, node,
3693 __builtin_return_address(0));
3695 EXPORT_SYMBOL(__kmalloc_node);
3697 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3698 int node, void *caller)
3700 return __do_kmalloc_node(size, flags, node, caller);
3702 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3704 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3706 return __do_kmalloc_node(size, flags, node, NULL);
3708 EXPORT_SYMBOL(__kmalloc_node);
3709 #endif /* CONFIG_DEBUG_SLAB */
3710 #endif /* CONFIG_NUMA */
3713 * __do_kmalloc - allocate memory
3714 * @size: how many bytes of memory are required.
3715 * @flags: the type of memory to allocate (see kmalloc).
3716 * @caller: function caller for debug tracking of the caller
3718 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3721 struct kmem_cache *cachep;
3723 /* If you want to save a few bytes .text space: replace
3725 * Then kmalloc uses the uninlined functions instead of the inline
3728 cachep = __find_general_cachep(size, flags);
3729 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3731 return __cache_alloc(cachep, flags, caller);
3735 #ifdef CONFIG_DEBUG_SLAB
3736 void *__kmalloc(size_t size, gfp_t flags)
3738 return __do_kmalloc(size, flags, __builtin_return_address(0));
3740 EXPORT_SYMBOL(__kmalloc);
3742 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3744 return __do_kmalloc(size, flags, caller);
3746 EXPORT_SYMBOL(__kmalloc_track_caller);
3749 void *__kmalloc(size_t size, gfp_t flags)
3751 return __do_kmalloc(size, flags, NULL);
3753 EXPORT_SYMBOL(__kmalloc);
3757 * kmem_cache_free - Deallocate an object
3758 * @cachep: The cache the allocation was from.
3759 * @objp: The previously allocated object.
3761 * Free an object which was previously allocated from this
3764 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3766 unsigned long flags;
3768 local_irq_save(flags);
3769 debug_check_no_locks_freed(objp, obj_size(cachep));
3770 __cache_free(cachep, objp);
3771 local_irq_restore(flags);
3773 EXPORT_SYMBOL(kmem_cache_free);
3776 * kfree - free previously allocated memory
3777 * @objp: pointer returned by kmalloc.
3779 * If @objp is NULL, no operation is performed.
3781 * Don't free memory not originally allocated by kmalloc()
3782 * or you will run into trouble.
3784 void kfree(const void *objp)
3786 struct kmem_cache *c;
3787 unsigned long flags;
3789 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3791 local_irq_save(flags);
3792 kfree_debugcheck(objp);
3793 c = virt_to_cache(objp);
3794 debug_check_no_locks_freed(objp, obj_size(c));
3795 __cache_free(c, (void *)objp);
3796 local_irq_restore(flags);
3798 EXPORT_SYMBOL(kfree);
3800 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3802 return obj_size(cachep);
3804 EXPORT_SYMBOL(kmem_cache_size);
3806 const char *kmem_cache_name(struct kmem_cache *cachep)
3808 return cachep->name;
3810 EXPORT_SYMBOL_GPL(kmem_cache_name);
3813 * This initializes kmem_list3 or resizes various caches for all nodes.
3815 static int alloc_kmemlist(struct kmem_cache *cachep)
3818 struct kmem_list3 *l3;
3819 struct array_cache *new_shared;
3820 struct array_cache **new_alien = NULL;
3822 for_each_online_node(node) {
3824 if (use_alien_caches) {
3825 new_alien = alloc_alien_cache(node, cachep->limit);
3831 if (cachep->shared) {
3832 new_shared = alloc_arraycache(node,
3833 cachep->shared*cachep->batchcount,
3836 free_alien_cache(new_alien);
3841 l3 = cachep->nodelists[node];
3843 struct array_cache *shared = l3->shared;
3845 spin_lock_irq(&l3->list_lock);
3848 free_block(cachep, shared->entry,
3849 shared->avail, node);
3851 l3->shared = new_shared;
3853 l3->alien = new_alien;
3856 l3->free_limit = (1 + nr_cpus_node(node)) *
3857 cachep->batchcount + cachep->num;
3858 spin_unlock_irq(&l3->list_lock);
3860 free_alien_cache(new_alien);
3863 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3865 free_alien_cache(new_alien);
3870 kmem_list3_init(l3);
3871 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3872 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3873 l3->shared = new_shared;
3874 l3->alien = new_alien;
3875 l3->free_limit = (1 + nr_cpus_node(node)) *
3876 cachep->batchcount + cachep->num;
3877 cachep->nodelists[node] = l3;
3882 if (!cachep->next.next) {
3883 /* Cache is not active yet. Roll back what we did */
3886 if (cachep->nodelists[node]) {
3887 l3 = cachep->nodelists[node];
3890 free_alien_cache(l3->alien);
3892 cachep->nodelists[node] = NULL;
3900 struct ccupdate_struct {
3901 struct kmem_cache *cachep;
3902 struct array_cache *new[NR_CPUS];
3905 static void do_ccupdate_local(void *info)
3907 struct ccupdate_struct *new = info;
3908 struct array_cache *old;
3911 old = cpu_cache_get(new->cachep);
3913 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3914 new->new[smp_processor_id()] = old;
3917 /* Always called with the cache_chain_mutex held */
3918 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3919 int batchcount, int shared)
3921 struct ccupdate_struct *new;
3924 new = kzalloc(sizeof(*new), GFP_KERNEL);
3928 for_each_online_cpu(i) {
3929 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3932 for (i--; i >= 0; i--)
3938 new->cachep = cachep;
3940 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3943 cachep->batchcount = batchcount;
3944 cachep->limit = limit;
3945 cachep->shared = shared;
3947 for_each_online_cpu(i) {
3948 struct array_cache *ccold = new->new[i];
3951 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3952 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3953 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3957 return alloc_kmemlist(cachep);
3960 /* Called with cache_chain_mutex held always */
3961 static int enable_cpucache(struct kmem_cache *cachep)
3967 * The head array serves three purposes:
3968 * - create a LIFO ordering, i.e. return objects that are cache-warm
3969 * - reduce the number of spinlock operations.
3970 * - reduce the number of linked list operations on the slab and
3971 * bufctl chains: array operations are cheaper.
3972 * The numbers are guessed, we should auto-tune as described by
3975 if (cachep->buffer_size > 131072)
3977 else if (cachep->buffer_size > PAGE_SIZE)
3979 else if (cachep->buffer_size > 1024)
3981 else if (cachep->buffer_size > 256)
3987 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3988 * allocation behaviour: Most allocs on one cpu, most free operations
3989 * on another cpu. For these cases, an efficient object passing between
3990 * cpus is necessary. This is provided by a shared array. The array
3991 * replaces Bonwick's magazine layer.
3992 * On uniprocessor, it's functionally equivalent (but less efficient)
3993 * to a larger limit. Thus disabled by default.
3996 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4001 * With debugging enabled, large batchcount lead to excessively long
4002 * periods with disabled local interrupts. Limit the batchcount
4007 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4009 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4010 cachep->name, -err);
4015 * Drain an array if it contains any elements taking the l3 lock only if
4016 * necessary. Note that the l3 listlock also protects the array_cache
4017 * if drain_array() is used on the shared array.
4019 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4020 struct array_cache *ac, int force, int node)
4024 if (!ac || !ac->avail)
4026 if (ac->touched && !force) {
4029 spin_lock_irq(&l3->list_lock);
4031 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4032 if (tofree > ac->avail)
4033 tofree = (ac->avail + 1) / 2;
4034 free_block(cachep, ac->entry, tofree, node);
4035 ac->avail -= tofree;
4036 memmove(ac->entry, &(ac->entry[tofree]),
4037 sizeof(void *) * ac->avail);
4039 spin_unlock_irq(&l3->list_lock);
4044 * cache_reap - Reclaim memory from caches.
4045 * @w: work descriptor
4047 * Called from workqueue/eventd every few seconds.
4049 * - clear the per-cpu caches for this CPU.
4050 * - return freeable pages to the main free memory pool.
4052 * If we cannot acquire the cache chain mutex then just give up - we'll try
4053 * again on the next iteration.
4055 static void cache_reap(struct work_struct *w)
4057 struct kmem_cache *searchp;
4058 struct kmem_list3 *l3;
4059 int node = numa_node_id();
4060 struct delayed_work *work =
4061 container_of(w, struct delayed_work, work);
4063 if (!mutex_trylock(&cache_chain_mutex))
4064 /* Give up. Setup the next iteration. */
4067 list_for_each_entry(searchp, &cache_chain, next) {
4071 * We only take the l3 lock if absolutely necessary and we
4072 * have established with reasonable certainty that
4073 * we can do some work if the lock was obtained.
4075 l3 = searchp->nodelists[node];
4077 reap_alien(searchp, l3);
4079 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4082 * These are racy checks but it does not matter
4083 * if we skip one check or scan twice.
4085 if (time_after(l3->next_reap, jiffies))
4088 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4090 drain_array(searchp, l3, l3->shared, 0, node);
4092 if (l3->free_touched)
4093 l3->free_touched = 0;
4097 freed = drain_freelist(searchp, l3, (l3->free_limit +
4098 5 * searchp->num - 1) / (5 * searchp->num));
4099 STATS_ADD_REAPED(searchp, freed);
4105 mutex_unlock(&cache_chain_mutex);
4108 /* Set up the next iteration */
4109 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4112 #ifdef CONFIG_SLABINFO
4114 static void print_slabinfo_header(struct seq_file *m)
4117 * Output format version, so at least we can change it
4118 * without _too_ many complaints.
4121 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4123 seq_puts(m, "slabinfo - version: 2.1\n");
4125 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4126 "<objperslab> <pagesperslab>");
4127 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4128 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4130 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4131 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4132 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4137 static void *s_start(struct seq_file *m, loff_t *pos)
4141 mutex_lock(&cache_chain_mutex);
4143 print_slabinfo_header(m);
4145 return seq_list_start(&cache_chain, *pos);
4148 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4150 return seq_list_next(p, &cache_chain, pos);
4153 static void s_stop(struct seq_file *m, void *p)
4155 mutex_unlock(&cache_chain_mutex);
4158 static int s_show(struct seq_file *m, void *p)
4160 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4162 unsigned long active_objs;
4163 unsigned long num_objs;
4164 unsigned long active_slabs = 0;
4165 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4169 struct kmem_list3 *l3;
4173 for_each_online_node(node) {
4174 l3 = cachep->nodelists[node];
4179 spin_lock_irq(&l3->list_lock);
4181 list_for_each_entry(slabp, &l3->slabs_full, list) {
4182 if (slabp->inuse != cachep->num && !error)
4183 error = "slabs_full accounting error";
4184 active_objs += cachep->num;
4187 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4188 if (slabp->inuse == cachep->num && !error)
4189 error = "slabs_partial inuse accounting error";
4190 if (!slabp->inuse && !error)
4191 error = "slabs_partial/inuse accounting error";
4192 active_objs += slabp->inuse;
4195 list_for_each_entry(slabp, &l3->slabs_free, list) {
4196 if (slabp->inuse && !error)
4197 error = "slabs_free/inuse accounting error";
4200 free_objects += l3->free_objects;
4202 shared_avail += l3->shared->avail;
4204 spin_unlock_irq(&l3->list_lock);
4206 num_slabs += active_slabs;
4207 num_objs = num_slabs * cachep->num;
4208 if (num_objs - active_objs != free_objects && !error)
4209 error = "free_objects accounting error";
4211 name = cachep->name;
4213 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4215 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4216 name, active_objs, num_objs, cachep->buffer_size,
4217 cachep->num, (1 << cachep->gfporder));
4218 seq_printf(m, " : tunables %4u %4u %4u",
4219 cachep->limit, cachep->batchcount, cachep->shared);
4220 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4221 active_slabs, num_slabs, shared_avail);
4224 unsigned long high = cachep->high_mark;
4225 unsigned long allocs = cachep->num_allocations;
4226 unsigned long grown = cachep->grown;
4227 unsigned long reaped = cachep->reaped;
4228 unsigned long errors = cachep->errors;
4229 unsigned long max_freeable = cachep->max_freeable;
4230 unsigned long node_allocs = cachep->node_allocs;
4231 unsigned long node_frees = cachep->node_frees;
4232 unsigned long overflows = cachep->node_overflow;
4234 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4235 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4236 reaped, errors, max_freeable, node_allocs,
4237 node_frees, overflows);
4241 unsigned long allochit = atomic_read(&cachep->allochit);
4242 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4243 unsigned long freehit = atomic_read(&cachep->freehit);
4244 unsigned long freemiss = atomic_read(&cachep->freemiss);
4246 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4247 allochit, allocmiss, freehit, freemiss);
4255 * slabinfo_op - iterator that generates /proc/slabinfo
4264 * num-pages-per-slab
4265 * + further values on SMP and with statistics enabled
4268 const struct seq_operations slabinfo_op = {
4275 #define MAX_SLABINFO_WRITE 128
4277 * slabinfo_write - Tuning for the slab allocator
4279 * @buffer: user buffer
4280 * @count: data length
4283 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4284 size_t count, loff_t *ppos)
4286 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4287 int limit, batchcount, shared, res;
4288 struct kmem_cache *cachep;
4290 if (count > MAX_SLABINFO_WRITE)
4292 if (copy_from_user(&kbuf, buffer, count))
4294 kbuf[MAX_SLABINFO_WRITE] = '\0';
4296 tmp = strchr(kbuf, ' ');
4301 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4304 /* Find the cache in the chain of caches. */
4305 mutex_lock(&cache_chain_mutex);
4307 list_for_each_entry(cachep, &cache_chain, next) {
4308 if (!strcmp(cachep->name, kbuf)) {
4309 if (limit < 1 || batchcount < 1 ||
4310 batchcount > limit || shared < 0) {
4313 res = do_tune_cpucache(cachep, limit,
4314 batchcount, shared);
4319 mutex_unlock(&cache_chain_mutex);
4325 #ifdef CONFIG_DEBUG_SLAB_LEAK
4327 static void *leaks_start(struct seq_file *m, loff_t *pos)
4329 mutex_lock(&cache_chain_mutex);
4330 return seq_list_start(&cache_chain, *pos);
4333 static inline int add_caller(unsigned long *n, unsigned long v)
4343 unsigned long *q = p + 2 * i;
4357 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4363 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4369 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4370 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4372 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4377 static void show_symbol(struct seq_file *m, unsigned long address)
4379 #ifdef CONFIG_KALLSYMS
4380 unsigned long offset, size;
4381 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4383 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4384 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4386 seq_printf(m, " [%s]", modname);
4390 seq_printf(m, "%p", (void *)address);
4393 static int leaks_show(struct seq_file *m, void *p)
4395 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4397 struct kmem_list3 *l3;
4399 unsigned long *n = m->private;
4403 if (!(cachep->flags & SLAB_STORE_USER))
4405 if (!(cachep->flags & SLAB_RED_ZONE))
4408 /* OK, we can do it */
4412 for_each_online_node(node) {
4413 l3 = cachep->nodelists[node];
4418 spin_lock_irq(&l3->list_lock);
4420 list_for_each_entry(slabp, &l3->slabs_full, list)
4421 handle_slab(n, cachep, slabp);
4422 list_for_each_entry(slabp, &l3->slabs_partial, list)
4423 handle_slab(n, cachep, slabp);
4424 spin_unlock_irq(&l3->list_lock);
4426 name = cachep->name;
4428 /* Increase the buffer size */
4429 mutex_unlock(&cache_chain_mutex);
4430 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4432 /* Too bad, we are really out */
4434 mutex_lock(&cache_chain_mutex);
4437 *(unsigned long *)m->private = n[0] * 2;
4439 mutex_lock(&cache_chain_mutex);
4440 /* Now make sure this entry will be retried */
4444 for (i = 0; i < n[1]; i++) {
4445 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4446 show_symbol(m, n[2*i+2]);
4453 const struct seq_operations slabstats_op = {
4454 .start = leaks_start,
4463 * ksize - get the actual amount of memory allocated for a given object
4464 * @objp: Pointer to the object
4466 * kmalloc may internally round up allocations and return more memory
4467 * than requested. ksize() can be used to determine the actual amount of
4468 * memory allocated. The caller may use this additional memory, even though
4469 * a smaller amount of memory was initially specified with the kmalloc call.
4470 * The caller must guarantee that objp points to a valid object previously
4471 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4472 * must not be freed during the duration of the call.
4474 size_t ksize(const void *objp)
4477 if (unlikely(objp == ZERO_SIZE_PTR))
4480 return obj_size(virt_to_cache(objp));
4482 EXPORT_SYMBOL(ksize);