3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/nodemask.h>
107 #include <linux/mempolicy.h>
108 #include <linux/mutex.h>
109 #include <linux/rtmutex.h>
111 #include <asm/uaccess.h>
112 #include <asm/cacheflush.h>
113 #include <asm/tlbflush.h>
114 #include <asm/page.h>
117 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
118 * SLAB_RED_ZONE & SLAB_POISON.
119 * 0 for faster, smaller code (especially in the critical paths).
121 * STATS - 1 to collect stats for /proc/slabinfo.
122 * 0 for faster, smaller code (especially in the critical paths).
124 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
127 #ifdef CONFIG_DEBUG_SLAB
130 #define FORCED_DEBUG 1
134 #define FORCED_DEBUG 0
137 /* Shouldn't this be in a header file somewhere? */
138 #define BYTES_PER_WORD sizeof(void *)
140 #ifndef cache_line_size
141 #define cache_line_size() L1_CACHE_BYTES
144 #ifndef ARCH_KMALLOC_MINALIGN
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
150 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
151 * Note that this flag disables some debug features.
153 #define ARCH_KMALLOC_MINALIGN 0
156 #ifndef ARCH_SLAB_MINALIGN
158 * Enforce a minimum alignment for all caches.
159 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
160 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
161 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
162 * some debug features.
164 #define ARCH_SLAB_MINALIGN 0
167 #ifndef ARCH_KMALLOC_FLAGS
168 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171 /* Legal flag mask for kmem_cache_create(). */
173 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
176 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
180 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
181 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
182 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
183 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
189 * Bufctl's are used for linking objs within a slab
192 * This implementation relies on "struct page" for locating the cache &
193 * slab an object belongs to.
194 * This allows the bufctl structure to be small (one int), but limits
195 * the number of objects a slab (not a cache) can contain when off-slab
196 * bufctls are used. The limit is the size of the largest general cache
197 * that does not use off-slab slabs.
198 * For 32bit archs with 4 kB pages, is this 56.
199 * This is not serious, as it is only for large objects, when it is unwise
200 * to have too many per slab.
201 * Note: This limit can be raised by introducing a general cache whose size
202 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 typedef unsigned int kmem_bufctl_t;
206 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
207 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
208 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
209 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 struct list_head list;
220 unsigned long colouroff;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
224 unsigned short nodeid;
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
241 * We assume struct slab_rcu can overlay struct slab when destroying.
244 struct rcu_head head;
245 struct kmem_cache *cachep;
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
257 * The limit is stored in the per-cpu structure to reduce the data cache
264 unsigned int batchcount;
265 unsigned int touched;
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
271 * [0] is for gcc 2.95. It should really be [].
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
279 #define BOOT_CPUCACHE_ENTRIES 1
280 struct arraycache_init {
281 struct array_cache cache;
282 void *entries[BOOT_CPUCACHE_ENTRIES];
286 * The slab lists for all objects.
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
293 unsigned int free_limit;
294 unsigned int colour_next; /* Per-node cache coloring */
295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
298 unsigned long next_reap; /* updated without locking */
299 int free_touched; /* updated without locking */
303 * Need this for bootstrapping a per node allocator.
305 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307 #define CACHE_CACHE 0
309 #define SIZE_L3 (1 + MAX_NUMNODES)
311 static int drain_freelist(struct kmem_cache *cache,
312 struct kmem_list3 *l3, int tofree);
313 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
315 static void enable_cpucache(struct kmem_cache *cachep);
316 static void cache_reap(void *unused);
319 * This function must be completely optimized away if a constant is passed to
320 * it. Mostly the same as what is in linux/slab.h except it returns an index.
322 static __always_inline int index_of(const size_t size)
324 extern void __bad_size(void);
326 if (__builtin_constant_p(size)) {
334 #include "linux/kmalloc_sizes.h"
342 static int slab_early_init = 1;
344 #define INDEX_AC index_of(sizeof(struct arraycache_init))
345 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
347 static void kmem_list3_init(struct kmem_list3 *parent)
349 INIT_LIST_HEAD(&parent->slabs_full);
350 INIT_LIST_HEAD(&parent->slabs_partial);
351 INIT_LIST_HEAD(&parent->slabs_free);
352 parent->shared = NULL;
353 parent->alien = NULL;
354 parent->colour_next = 0;
355 spin_lock_init(&parent->list_lock);
356 parent->free_objects = 0;
357 parent->free_touched = 0;
360 #define MAKE_LIST(cachep, listp, slab, nodeid) \
362 INIT_LIST_HEAD(listp); \
363 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
366 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
368 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
369 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
380 /* 1) per-cpu data, touched during every alloc/free */
381 struct array_cache *array[NR_CPUS];
382 /* 2) Cache tunables. Protected by cache_chain_mutex */
383 unsigned int batchcount;
387 unsigned int buffer_size;
388 /* 3) touched by every alloc & free from the backend */
389 struct kmem_list3 *nodelists[MAX_NUMNODES];
391 unsigned int flags; /* constant flags */
392 unsigned int num; /* # of objs per slab */
394 /* 4) cache_grow/shrink */
395 /* order of pgs per slab (2^n) */
396 unsigned int gfporder;
398 /* force GFP flags, e.g. GFP_DMA */
401 size_t colour; /* cache colouring range */
402 unsigned int colour_off; /* colour offset */
403 struct kmem_cache *slabp_cache;
404 unsigned int slab_size;
405 unsigned int dflags; /* dynamic flags */
407 /* constructor func */
408 void (*ctor) (void *, struct kmem_cache *, unsigned long);
410 /* de-constructor func */
411 void (*dtor) (void *, struct kmem_cache *, unsigned long);
413 /* 5) cache creation/removal */
415 struct list_head next;
419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
428 unsigned long node_overflow;
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
446 #define CFLGS_OFF_SLAB (0x80000000UL)
447 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
449 #define BATCHREFILL_LIMIT 16
451 * Optimization question: fewer reaps means less probability for unnessary
452 * cpucache drain/refill cycles.
454 * OTOH the cpuarrays can contain lots of objects,
455 * which could lock up otherwise freeable slabs.
457 #define REAPTIMEOUT_CPUC (2*HZ)
458 #define REAPTIMEOUT_LIST3 (4*HZ)
461 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
462 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
463 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
464 #define STATS_INC_GROWN(x) ((x)->grown++)
465 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
466 #define STATS_SET_HIGH(x) \
468 if ((x)->num_active > (x)->high_mark) \
469 (x)->high_mark = (x)->num_active; \
471 #define STATS_INC_ERR(x) ((x)->errors++)
472 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
473 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
474 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
475 #define STATS_SET_FREEABLE(x, i) \
477 if ((x)->max_freeable < i) \
478 (x)->max_freeable = i; \
480 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
481 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
482 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
483 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
485 #define STATS_INC_ACTIVE(x) do { } while (0)
486 #define STATS_DEC_ACTIVE(x) do { } while (0)
487 #define STATS_INC_ALLOCED(x) do { } while (0)
488 #define STATS_INC_GROWN(x) do { } while (0)
489 #define STATS_ADD_REAPED(x,y) do { } while (0)
490 #define STATS_SET_HIGH(x) do { } while (0)
491 #define STATS_INC_ERR(x) do { } while (0)
492 #define STATS_INC_NODEALLOCS(x) do { } while (0)
493 #define STATS_INC_NODEFREES(x) do { } while (0)
494 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
495 #define STATS_SET_FREEABLE(x, i) do { } while (0)
496 #define STATS_INC_ALLOCHIT(x) do { } while (0)
497 #define STATS_INC_ALLOCMISS(x) do { } while (0)
498 #define STATS_INC_FREEHIT(x) do { } while (0)
499 #define STATS_INC_FREEMISS(x) do { } while (0)
505 * memory layout of objects:
507 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
508 * the end of an object is aligned with the end of the real
509 * allocation. Catches writes behind the end of the allocation.
510 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
512 * cachep->obj_offset: The real object.
513 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
514 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
515 * [BYTES_PER_WORD long]
517 static int obj_offset(struct kmem_cache *cachep)
519 return cachep->obj_offset;
522 static int obj_size(struct kmem_cache *cachep)
524 return cachep->obj_size;
527 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
529 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
530 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
533 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
535 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
536 if (cachep->flags & SLAB_STORE_USER)
537 return (unsigned long *)(objp + cachep->buffer_size -
539 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
542 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
544 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
545 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
550 #define obj_offset(x) 0
551 #define obj_size(cachep) (cachep->buffer_size)
552 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
553 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
559 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
562 #if defined(CONFIG_LARGE_ALLOCS)
563 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
564 #define MAX_GFP_ORDER 13 /* up to 32Mb */
565 #elif defined(CONFIG_MMU)
566 #define MAX_OBJ_ORDER 5 /* 32 pages */
567 #define MAX_GFP_ORDER 5 /* 32 pages */
569 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
570 #define MAX_GFP_ORDER 8 /* up to 1Mb */
574 * Do not go above this order unless 0 objects fit into the slab.
576 #define BREAK_GFP_ORDER_HI 1
577 #define BREAK_GFP_ORDER_LO 0
578 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
581 * Functions for storing/retrieving the cachep and or slab from the page
582 * allocator. These are used to find the slab an obj belongs to. With kfree(),
583 * these are used to find the cache which an obj belongs to.
585 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
587 page->lru.next = (struct list_head *)cache;
590 static inline struct kmem_cache *page_get_cache(struct page *page)
592 if (unlikely(PageCompound(page)))
593 page = (struct page *)page_private(page);
594 BUG_ON(!PageSlab(page));
595 return (struct kmem_cache *)page->lru.next;
598 static inline void page_set_slab(struct page *page, struct slab *slab)
600 page->lru.prev = (struct list_head *)slab;
603 static inline struct slab *page_get_slab(struct page *page)
605 if (unlikely(PageCompound(page)))
606 page = (struct page *)page_private(page);
607 BUG_ON(!PageSlab(page));
608 return (struct slab *)page->lru.prev;
611 static inline struct kmem_cache *virt_to_cache(const void *obj)
613 struct page *page = virt_to_page(obj);
614 return page_get_cache(page);
617 static inline struct slab *virt_to_slab(const void *obj)
619 struct page *page = virt_to_page(obj);
620 return page_get_slab(page);
623 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
626 return slab->s_mem + cache->buffer_size * idx;
629 static inline unsigned int obj_to_index(struct kmem_cache *cache,
630 struct slab *slab, void *obj)
632 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
636 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 struct cache_sizes malloc_sizes[] = {
639 #define CACHE(x) { .cs_size = (x) },
640 #include <linux/kmalloc_sizes.h>
644 EXPORT_SYMBOL(malloc_sizes);
646 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
652 static struct cache_names __initdata cache_names[] = {
653 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
654 #include <linux/kmalloc_sizes.h>
659 static struct arraycache_init initarray_cache __initdata =
660 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
661 static struct arraycache_init initarray_generic =
662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
664 /* internal cache of cache description objs */
665 static struct kmem_cache cache_cache = {
667 .limit = BOOT_CPUCACHE_ENTRIES,
669 .buffer_size = sizeof(struct kmem_cache),
670 .name = "kmem_cache",
672 .obj_size = sizeof(struct kmem_cache),
676 #define BAD_ALIEN_MAGIC 0x01020304ul
678 #ifdef CONFIG_LOCKDEP
681 * Slab sometimes uses the kmalloc slabs to store the slab headers
682 * for other slabs "off slab".
683 * The locking for this is tricky in that it nests within the locks
684 * of all other slabs in a few places; to deal with this special
685 * locking we put on-slab caches into a separate lock-class.
687 * We set lock class for alien array caches which are up during init.
688 * The lock annotation will be lost if all cpus of a node goes down and
689 * then comes back up during hotplug
691 static struct lock_class_key on_slab_l3_key;
692 static struct lock_class_key on_slab_alc_key;
694 static inline void init_lock_keys(void)
698 struct cache_sizes *s = malloc_sizes;
700 while (s->cs_size != ULONG_MAX) {
702 struct array_cache **alc;
704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
705 if (!l3 || OFF_SLAB(s->cs_cachep))
707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
710 * FIXME: This check for BAD_ALIEN_MAGIC
711 * should go away when common slab code is taught to
712 * work even without alien caches.
713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
714 * for alloc_alien_cache,
716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
720 lockdep_set_class(&alc[r]->lock,
728 static inline void init_lock_keys(void)
733 /* Guard access to the cache-chain. */
734 static DEFINE_MUTEX(cache_chain_mutex);
735 static struct list_head cache_chain;
738 * vm_enough_memory() looks at this to determine how many slab-allocated pages
739 * are possibly freeable under pressure
741 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
743 atomic_t slab_reclaim_pages;
746 * chicken and egg problem: delay the per-cpu array allocation
747 * until the general caches are up.
757 * used by boot code to determine if it can use slab based allocator
759 int slab_is_available(void)
761 return g_cpucache_up == FULL;
764 static DEFINE_PER_CPU(struct work_struct, reap_work);
766 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
768 return cachep->array[smp_processor_id()];
771 static inline struct kmem_cache *__find_general_cachep(size_t size,
774 struct cache_sizes *csizep = malloc_sizes;
777 /* This happens if someone tries to call
778 * kmem_cache_create(), or __kmalloc(), before
779 * the generic caches are initialized.
781 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
783 while (size > csizep->cs_size)
787 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
788 * has cs_{dma,}cachep==NULL. Thus no special case
789 * for large kmalloc calls required.
791 if (unlikely(gfpflags & GFP_DMA))
792 return csizep->cs_dmacachep;
793 return csizep->cs_cachep;
796 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
798 return __find_general_cachep(size, gfpflags);
800 EXPORT_SYMBOL(kmem_find_general_cachep);
802 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
804 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
808 * Calculate the number of objects and left-over bytes for a given buffer size.
810 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
811 size_t align, int flags, size_t *left_over,
816 size_t slab_size = PAGE_SIZE << gfporder;
819 * The slab management structure can be either off the slab or
820 * on it. For the latter case, the memory allocated for a
824 * - One kmem_bufctl_t for each object
825 * - Padding to respect alignment of @align
826 * - @buffer_size bytes for each object
828 * If the slab management structure is off the slab, then the
829 * alignment will already be calculated into the size. Because
830 * the slabs are all pages aligned, the objects will be at the
831 * correct alignment when allocated.
833 if (flags & CFLGS_OFF_SLAB) {
835 nr_objs = slab_size / buffer_size;
837 if (nr_objs > SLAB_LIMIT)
838 nr_objs = SLAB_LIMIT;
841 * Ignore padding for the initial guess. The padding
842 * is at most @align-1 bytes, and @buffer_size is at
843 * least @align. In the worst case, this result will
844 * be one greater than the number of objects that fit
845 * into the memory allocation when taking the padding
848 nr_objs = (slab_size - sizeof(struct slab)) /
849 (buffer_size + sizeof(kmem_bufctl_t));
852 * This calculated number will be either the right
853 * amount, or one greater than what we want.
855 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
859 if (nr_objs > SLAB_LIMIT)
860 nr_objs = SLAB_LIMIT;
862 mgmt_size = slab_mgmt_size(nr_objs, align);
865 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
868 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
870 static void __slab_error(const char *function, struct kmem_cache *cachep,
873 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
874 function, cachep->name, msg);
880 * Special reaping functions for NUMA systems called from cache_reap().
881 * These take care of doing round robin flushing of alien caches (containing
882 * objects freed on different nodes from which they were allocated) and the
883 * flushing of remote pcps by calling drain_node_pages.
885 static DEFINE_PER_CPU(unsigned long, reap_node);
887 static void init_reap_node(int cpu)
891 node = next_node(cpu_to_node(cpu), node_online_map);
892 if (node == MAX_NUMNODES)
893 node = first_node(node_online_map);
895 per_cpu(reap_node, cpu) = node;
898 static void next_reap_node(void)
900 int node = __get_cpu_var(reap_node);
903 * Also drain per cpu pages on remote zones
905 if (node != numa_node_id())
906 drain_node_pages(node);
908 node = next_node(node, node_online_map);
909 if (unlikely(node >= MAX_NUMNODES))
910 node = first_node(node_online_map);
911 __get_cpu_var(reap_node) = node;
915 #define init_reap_node(cpu) do { } while (0)
916 #define next_reap_node(void) do { } while (0)
920 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
921 * via the workqueue/eventd.
922 * Add the CPU number into the expiration time to minimize the possibility of
923 * the CPUs getting into lockstep and contending for the global cache chain
926 static void __devinit start_cpu_timer(int cpu)
928 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
931 * When this gets called from do_initcalls via cpucache_init(),
932 * init_workqueues() has already run, so keventd will be setup
935 if (keventd_up() && reap_work->func == NULL) {
937 INIT_WORK(reap_work, cache_reap, NULL);
938 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
942 static struct array_cache *alloc_arraycache(int node, int entries,
945 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
946 struct array_cache *nc = NULL;
948 nc = kmalloc_node(memsize, GFP_KERNEL, node);
952 nc->batchcount = batchcount;
954 spin_lock_init(&nc->lock);
960 * Transfer objects in one arraycache to another.
961 * Locking must be handled by the caller.
963 * Return the number of entries transferred.
965 static int transfer_objects(struct array_cache *to,
966 struct array_cache *from, unsigned int max)
968 /* Figure out how many entries to transfer */
969 int nr = min(min(from->avail, max), to->limit - to->avail);
974 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
984 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
985 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
987 static struct array_cache **alloc_alien_cache(int node, int limit)
989 struct array_cache **ac_ptr;
990 int memsize = sizeof(void *) * MAX_NUMNODES;
995 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
998 if (i == node || !node_online(i)) {
1002 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1004 for (i--; i <= 0; i--)
1014 static void free_alien_cache(struct array_cache **ac_ptr)
1025 static void __drain_alien_cache(struct kmem_cache *cachep,
1026 struct array_cache *ac, int node)
1028 struct kmem_list3 *rl3 = cachep->nodelists[node];
1031 spin_lock(&rl3->list_lock);
1033 * Stuff objects into the remote nodes shared array first.
1034 * That way we could avoid the overhead of putting the objects
1035 * into the free lists and getting them back later.
1038 transfer_objects(rl3->shared, ac, ac->limit);
1040 free_block(cachep, ac->entry, ac->avail, node);
1042 spin_unlock(&rl3->list_lock);
1047 * Called from cache_reap() to regularly drain alien caches round robin.
1049 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1051 int node = __get_cpu_var(reap_node);
1054 struct array_cache *ac = l3->alien[node];
1056 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1057 __drain_alien_cache(cachep, ac, node);
1058 spin_unlock_irq(&ac->lock);
1063 static void drain_alien_cache(struct kmem_cache *cachep,
1064 struct array_cache **alien)
1067 struct array_cache *ac;
1068 unsigned long flags;
1070 for_each_online_node(i) {
1073 spin_lock_irqsave(&ac->lock, flags);
1074 __drain_alien_cache(cachep, ac, i);
1075 spin_unlock_irqrestore(&ac->lock, flags);
1080 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1082 struct slab *slabp = virt_to_slab(objp);
1083 int nodeid = slabp->nodeid;
1084 struct kmem_list3 *l3;
1085 struct array_cache *alien = NULL;
1088 * Make sure we are not freeing a object from another node to the array
1089 * cache on this cpu.
1091 if (likely(slabp->nodeid == numa_node_id()))
1094 l3 = cachep->nodelists[numa_node_id()];
1095 STATS_INC_NODEFREES(cachep);
1096 if (l3->alien && l3->alien[nodeid]) {
1097 alien = l3->alien[nodeid];
1098 spin_lock(&alien->lock);
1099 if (unlikely(alien->avail == alien->limit)) {
1100 STATS_INC_ACOVERFLOW(cachep);
1101 __drain_alien_cache(cachep, alien, nodeid);
1103 alien->entry[alien->avail++] = objp;
1104 spin_unlock(&alien->lock);
1106 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1107 free_block(cachep, &objp, 1, nodeid);
1108 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1115 #define drain_alien_cache(cachep, alien) do { } while (0)
1116 #define reap_alien(cachep, l3) do { } while (0)
1118 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1120 return (struct array_cache **)BAD_ALIEN_MAGIC;
1123 static inline void free_alien_cache(struct array_cache **ac_ptr)
1127 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1134 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1135 unsigned long action, void *hcpu)
1137 long cpu = (long)hcpu;
1138 struct kmem_cache *cachep;
1139 struct kmem_list3 *l3 = NULL;
1140 int node = cpu_to_node(cpu);
1141 int memsize = sizeof(struct kmem_list3);
1144 case CPU_UP_PREPARE:
1145 mutex_lock(&cache_chain_mutex);
1147 * We need to do this right in the beginning since
1148 * alloc_arraycache's are going to use this list.
1149 * kmalloc_node allows us to add the slab to the right
1150 * kmem_list3 and not this cpu's kmem_list3
1153 list_for_each_entry(cachep, &cache_chain, next) {
1155 * Set up the size64 kmemlist for cpu before we can
1156 * begin anything. Make sure some other cpu on this
1157 * node has not already allocated this
1159 if (!cachep->nodelists[node]) {
1160 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1163 kmem_list3_init(l3);
1164 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1165 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1168 * The l3s don't come and go as CPUs come and
1169 * go. cache_chain_mutex is sufficient
1172 cachep->nodelists[node] = l3;
1175 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1176 cachep->nodelists[node]->free_limit =
1177 (1 + nr_cpus_node(node)) *
1178 cachep->batchcount + cachep->num;
1179 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1183 * Now we can go ahead with allocating the shared arrays and
1186 list_for_each_entry(cachep, &cache_chain, next) {
1187 struct array_cache *nc;
1188 struct array_cache *shared;
1189 struct array_cache **alien;
1191 nc = alloc_arraycache(node, cachep->limit,
1192 cachep->batchcount);
1195 shared = alloc_arraycache(node,
1196 cachep->shared * cachep->batchcount,
1201 alien = alloc_alien_cache(node, cachep->limit);
1204 cachep->array[cpu] = nc;
1205 l3 = cachep->nodelists[node];
1208 spin_lock_irq(&l3->list_lock);
1211 * We are serialised from CPU_DEAD or
1212 * CPU_UP_CANCELLED by the cpucontrol lock
1214 l3->shared = shared;
1223 spin_unlock_irq(&l3->list_lock);
1225 free_alien_cache(alien);
1227 mutex_unlock(&cache_chain_mutex);
1230 start_cpu_timer(cpu);
1232 #ifdef CONFIG_HOTPLUG_CPU
1235 * Even if all the cpus of a node are down, we don't free the
1236 * kmem_list3 of any cache. This to avoid a race between
1237 * cpu_down, and a kmalloc allocation from another cpu for
1238 * memory from the node of the cpu going down. The list3
1239 * structure is usually allocated from kmem_cache_create() and
1240 * gets destroyed at kmem_cache_destroy().
1243 case CPU_UP_CANCELED:
1244 mutex_lock(&cache_chain_mutex);
1245 list_for_each_entry(cachep, &cache_chain, next) {
1246 struct array_cache *nc;
1247 struct array_cache *shared;
1248 struct array_cache **alien;
1251 mask = node_to_cpumask(node);
1252 /* cpu is dead; no one can alloc from it. */
1253 nc = cachep->array[cpu];
1254 cachep->array[cpu] = NULL;
1255 l3 = cachep->nodelists[node];
1258 goto free_array_cache;
1260 spin_lock_irq(&l3->list_lock);
1262 /* Free limit for this kmem_list3 */
1263 l3->free_limit -= cachep->batchcount;
1265 free_block(cachep, nc->entry, nc->avail, node);
1267 if (!cpus_empty(mask)) {
1268 spin_unlock_irq(&l3->list_lock);
1269 goto free_array_cache;
1272 shared = l3->shared;
1274 free_block(cachep, l3->shared->entry,
1275 l3->shared->avail, node);
1282 spin_unlock_irq(&l3->list_lock);
1286 drain_alien_cache(cachep, alien);
1287 free_alien_cache(alien);
1293 * In the previous loop, all the objects were freed to
1294 * the respective cache's slabs, now we can go ahead and
1295 * shrink each nodelist to its limit.
1297 list_for_each_entry(cachep, &cache_chain, next) {
1298 l3 = cachep->nodelists[node];
1301 drain_freelist(cachep, l3, l3->free_objects);
1303 mutex_unlock(&cache_chain_mutex);
1309 mutex_unlock(&cache_chain_mutex);
1313 static struct notifier_block __cpuinitdata cpucache_notifier = {
1314 &cpuup_callback, NULL, 0
1318 * swap the static kmem_list3 with kmalloced memory
1320 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1323 struct kmem_list3 *ptr;
1325 BUG_ON(cachep->nodelists[nodeid] != list);
1326 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1329 local_irq_disable();
1330 memcpy(ptr, list, sizeof(struct kmem_list3));
1332 * Do not assume that spinlocks can be initialized via memcpy:
1334 spin_lock_init(&ptr->list_lock);
1336 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1337 cachep->nodelists[nodeid] = ptr;
1342 * Initialisation. Called after the page allocator have been initialised and
1343 * before smp_init().
1345 void __init kmem_cache_init(void)
1348 struct cache_sizes *sizes;
1349 struct cache_names *names;
1353 for (i = 0; i < NUM_INIT_LISTS; i++) {
1354 kmem_list3_init(&initkmem_list3[i]);
1355 if (i < MAX_NUMNODES)
1356 cache_cache.nodelists[i] = NULL;
1360 * Fragmentation resistance on low memory - only use bigger
1361 * page orders on machines with more than 32MB of memory.
1363 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1364 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1366 /* Bootstrap is tricky, because several objects are allocated
1367 * from caches that do not exist yet:
1368 * 1) initialize the cache_cache cache: it contains the struct
1369 * kmem_cache structures of all caches, except cache_cache itself:
1370 * cache_cache is statically allocated.
1371 * Initially an __init data area is used for the head array and the
1372 * kmem_list3 structures, it's replaced with a kmalloc allocated
1373 * array at the end of the bootstrap.
1374 * 2) Create the first kmalloc cache.
1375 * The struct kmem_cache for the new cache is allocated normally.
1376 * An __init data area is used for the head array.
1377 * 3) Create the remaining kmalloc caches, with minimally sized
1379 * 4) Replace the __init data head arrays for cache_cache and the first
1380 * kmalloc cache with kmalloc allocated arrays.
1381 * 5) Replace the __init data for kmem_list3 for cache_cache and
1382 * the other cache's with kmalloc allocated memory.
1383 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1386 /* 1) create the cache_cache */
1387 INIT_LIST_HEAD(&cache_chain);
1388 list_add(&cache_cache.next, &cache_chain);
1389 cache_cache.colour_off = cache_line_size();
1390 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1391 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
1393 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1396 for (order = 0; order < MAX_ORDER; order++) {
1397 cache_estimate(order, cache_cache.buffer_size,
1398 cache_line_size(), 0, &left_over, &cache_cache.num);
1399 if (cache_cache.num)
1402 BUG_ON(!cache_cache.num);
1403 cache_cache.gfporder = order;
1404 cache_cache.colour = left_over / cache_cache.colour_off;
1405 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1406 sizeof(struct slab), cache_line_size());
1408 /* 2+3) create the kmalloc caches */
1409 sizes = malloc_sizes;
1410 names = cache_names;
1413 * Initialize the caches that provide memory for the array cache and the
1414 * kmem_list3 structures first. Without this, further allocations will
1418 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1419 sizes[INDEX_AC].cs_size,
1420 ARCH_KMALLOC_MINALIGN,
1421 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1424 if (INDEX_AC != INDEX_L3) {
1425 sizes[INDEX_L3].cs_cachep =
1426 kmem_cache_create(names[INDEX_L3].name,
1427 sizes[INDEX_L3].cs_size,
1428 ARCH_KMALLOC_MINALIGN,
1429 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1433 slab_early_init = 0;
1435 while (sizes->cs_size != ULONG_MAX) {
1437 * For performance, all the general caches are L1 aligned.
1438 * This should be particularly beneficial on SMP boxes, as it
1439 * eliminates "false sharing".
1440 * Note for systems short on memory removing the alignment will
1441 * allow tighter packing of the smaller caches.
1443 if (!sizes->cs_cachep) {
1444 sizes->cs_cachep = kmem_cache_create(names->name,
1446 ARCH_KMALLOC_MINALIGN,
1447 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1451 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1453 ARCH_KMALLOC_MINALIGN,
1454 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1460 /* 4) Replace the bootstrap head arrays */
1462 struct array_cache *ptr;
1464 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1466 local_irq_disable();
1467 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1468 memcpy(ptr, cpu_cache_get(&cache_cache),
1469 sizeof(struct arraycache_init));
1471 * Do not assume that spinlocks can be initialized via memcpy:
1473 spin_lock_init(&ptr->lock);
1475 cache_cache.array[smp_processor_id()] = ptr;
1478 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1480 local_irq_disable();
1481 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1482 != &initarray_generic.cache);
1483 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1484 sizeof(struct arraycache_init));
1486 * Do not assume that spinlocks can be initialized via memcpy:
1488 spin_lock_init(&ptr->lock);
1490 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1494 /* 5) Replace the bootstrap kmem_list3's */
1497 /* Replace the static kmem_list3 structures for the boot cpu */
1498 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1501 for_each_online_node(node) {
1502 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1503 &initkmem_list3[SIZE_AC + node], node);
1505 if (INDEX_AC != INDEX_L3) {
1506 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1507 &initkmem_list3[SIZE_L3 + node],
1513 /* 6) resize the head arrays to their final sizes */
1515 struct kmem_cache *cachep;
1516 mutex_lock(&cache_chain_mutex);
1517 list_for_each_entry(cachep, &cache_chain, next)
1518 enable_cpucache(cachep);
1519 mutex_unlock(&cache_chain_mutex);
1522 /* Annotate slab for lockdep -- annotate the malloc caches */
1527 g_cpucache_up = FULL;
1530 * Register a cpu startup notifier callback that initializes
1531 * cpu_cache_get for all new cpus
1533 register_cpu_notifier(&cpucache_notifier);
1536 * The reap timers are started later, with a module init call: That part
1537 * of the kernel is not yet operational.
1541 static int __init cpucache_init(void)
1546 * Register the timers that return unneeded pages to the page allocator
1548 for_each_online_cpu(cpu)
1549 start_cpu_timer(cpu);
1552 __initcall(cpucache_init);
1555 * Interface to system's page allocator. No need to hold the cache-lock.
1557 * If we requested dmaable memory, we will get it. Even if we
1558 * did not request dmaable memory, we might get it, but that
1559 * would be relatively rare and ignorable.
1561 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1569 * Nommu uses slab's for process anonymous memory allocations, and thus
1570 * requires __GFP_COMP to properly refcount higher order allocations
1572 flags |= __GFP_COMP;
1574 flags |= cachep->gfpflags;
1576 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1580 nr_pages = (1 << cachep->gfporder);
1581 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1582 atomic_add(nr_pages, &slab_reclaim_pages);
1583 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages);
1584 for (i = 0; i < nr_pages; i++)
1585 __SetPageSlab(page + i);
1586 return page_address(page);
1590 * Interface to system's page release.
1592 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1594 unsigned long i = (1 << cachep->gfporder);
1595 struct page *page = virt_to_page(addr);
1596 const unsigned long nr_freed = i;
1598 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed);
1600 BUG_ON(!PageSlab(page));
1601 __ClearPageSlab(page);
1604 if (current->reclaim_state)
1605 current->reclaim_state->reclaimed_slab += nr_freed;
1606 free_pages((unsigned long)addr, cachep->gfporder);
1607 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1608 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
1611 static void kmem_rcu_free(struct rcu_head *head)
1613 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1614 struct kmem_cache *cachep = slab_rcu->cachep;
1616 kmem_freepages(cachep, slab_rcu->addr);
1617 if (OFF_SLAB(cachep))
1618 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1623 #ifdef CONFIG_DEBUG_PAGEALLOC
1624 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1625 unsigned long caller)
1627 int size = obj_size(cachep);
1629 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1631 if (size < 5 * sizeof(unsigned long))
1634 *addr++ = 0x12345678;
1636 *addr++ = smp_processor_id();
1637 size -= 3 * sizeof(unsigned long);
1639 unsigned long *sptr = &caller;
1640 unsigned long svalue;
1642 while (!kstack_end(sptr)) {
1644 if (kernel_text_address(svalue)) {
1646 size -= sizeof(unsigned long);
1647 if (size <= sizeof(unsigned long))
1653 *addr++ = 0x87654321;
1657 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1659 int size = obj_size(cachep);
1660 addr = &((char *)addr)[obj_offset(cachep)];
1662 memset(addr, val, size);
1663 *(unsigned char *)(addr + size - 1) = POISON_END;
1666 static void dump_line(char *data, int offset, int limit)
1669 unsigned char total = 0, bad_count = 0, errors;
1670 printk(KERN_ERR "%03x:", offset);
1671 for (i = 0; i < limit; i++) {
1672 if (data[offset + i] != POISON_FREE) {
1673 total += data[offset + i];
1676 printk(" %02x", (unsigned char)data[offset + i]);
1680 if (bad_count == 1) {
1681 errors = total ^ POISON_FREE;
1682 if (errors && !(errors & (errors-1))) {
1683 printk(KERN_ERR "Single bit error detected. Probably bad RAM.\n");
1685 printk(KERN_ERR "Run memtest86+ or a similar memory test tool.\n");
1687 printk(KERN_ERR "Run a memory test tool.\n");
1696 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1701 if (cachep->flags & SLAB_RED_ZONE) {
1702 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1703 *dbg_redzone1(cachep, objp),
1704 *dbg_redzone2(cachep, objp));
1707 if (cachep->flags & SLAB_STORE_USER) {
1708 printk(KERN_ERR "Last user: [<%p>]",
1709 *dbg_userword(cachep, objp));
1710 print_symbol("(%s)",
1711 (unsigned long)*dbg_userword(cachep, objp));
1714 realobj = (char *)objp + obj_offset(cachep);
1715 size = obj_size(cachep);
1716 for (i = 0; i < size && lines; i += 16, lines--) {
1719 if (i + limit > size)
1721 dump_line(realobj, i, limit);
1725 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1731 realobj = (char *)objp + obj_offset(cachep);
1732 size = obj_size(cachep);
1734 for (i = 0; i < size; i++) {
1735 char exp = POISON_FREE;
1738 if (realobj[i] != exp) {
1744 "Slab corruption: (%s) start=%p, len=%d\n",
1745 print_tainted(), realobj, size);
1746 print_objinfo(cachep, objp, 0);
1749 /* Hexdump the affected line */
1752 if (i + limit > size)
1754 dump_line(realobj, i, limit);
1757 /* Limit to 5 lines */
1763 /* Print some data about the neighboring objects, if they
1766 struct slab *slabp = virt_to_slab(objp);
1769 objnr = obj_to_index(cachep, slabp, objp);
1771 objp = index_to_obj(cachep, slabp, objnr - 1);
1772 realobj = (char *)objp + obj_offset(cachep);
1773 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1775 print_objinfo(cachep, objp, 2);
1777 if (objnr + 1 < cachep->num) {
1778 objp = index_to_obj(cachep, slabp, objnr + 1);
1779 realobj = (char *)objp + obj_offset(cachep);
1780 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1782 print_objinfo(cachep, objp, 2);
1790 * slab_destroy_objs - destroy a slab and its objects
1791 * @cachep: cache pointer being destroyed
1792 * @slabp: slab pointer being destroyed
1794 * Call the registered destructor for each object in a slab that is being
1797 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1800 for (i = 0; i < cachep->num; i++) {
1801 void *objp = index_to_obj(cachep, slabp, i);
1803 if (cachep->flags & SLAB_POISON) {
1804 #ifdef CONFIG_DEBUG_PAGEALLOC
1805 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1807 kernel_map_pages(virt_to_page(objp),
1808 cachep->buffer_size / PAGE_SIZE, 1);
1810 check_poison_obj(cachep, objp);
1812 check_poison_obj(cachep, objp);
1815 if (cachep->flags & SLAB_RED_ZONE) {
1816 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1817 slab_error(cachep, "start of a freed object "
1819 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1820 slab_error(cachep, "end of a freed object "
1823 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1824 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1828 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1832 for (i = 0; i < cachep->num; i++) {
1833 void *objp = index_to_obj(cachep, slabp, i);
1834 (cachep->dtor) (objp, cachep, 0);
1841 * slab_destroy - destroy and release all objects in a slab
1842 * @cachep: cache pointer being destroyed
1843 * @slabp: slab pointer being destroyed
1845 * Destroy all the objs in a slab, and release the mem back to the system.
1846 * Before calling the slab must have been unlinked from the cache. The
1847 * cache-lock is not held/needed.
1849 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1851 void *addr = slabp->s_mem - slabp->colouroff;
1853 slab_destroy_objs(cachep, slabp);
1854 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1855 struct slab_rcu *slab_rcu;
1857 slab_rcu = (struct slab_rcu *)slabp;
1858 slab_rcu->cachep = cachep;
1859 slab_rcu->addr = addr;
1860 call_rcu(&slab_rcu->head, kmem_rcu_free);
1862 kmem_freepages(cachep, addr);
1863 if (OFF_SLAB(cachep))
1864 kmem_cache_free(cachep->slabp_cache, slabp);
1869 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1870 * size of kmem_list3.
1872 static void set_up_list3s(struct kmem_cache *cachep, int index)
1876 for_each_online_node(node) {
1877 cachep->nodelists[node] = &initkmem_list3[index + node];
1878 cachep->nodelists[node]->next_reap = jiffies +
1880 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1885 * calculate_slab_order - calculate size (page order) of slabs
1886 * @cachep: pointer to the cache that is being created
1887 * @size: size of objects to be created in this cache.
1888 * @align: required alignment for the objects.
1889 * @flags: slab allocation flags
1891 * Also calculates the number of objects per slab.
1893 * This could be made much more intelligent. For now, try to avoid using
1894 * high order pages for slabs. When the gfp() functions are more friendly
1895 * towards high-order requests, this should be changed.
1897 static size_t calculate_slab_order(struct kmem_cache *cachep,
1898 size_t size, size_t align, unsigned long flags)
1900 unsigned long offslab_limit;
1901 size_t left_over = 0;
1904 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
1908 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1912 if (flags & CFLGS_OFF_SLAB) {
1914 * Max number of objs-per-slab for caches which
1915 * use off-slab slabs. Needed to avoid a possible
1916 * looping condition in cache_grow().
1918 offslab_limit = size - sizeof(struct slab);
1919 offslab_limit /= sizeof(kmem_bufctl_t);
1921 if (num > offslab_limit)
1925 /* Found something acceptable - save it away */
1927 cachep->gfporder = gfporder;
1928 left_over = remainder;
1931 * A VFS-reclaimable slab tends to have most allocations
1932 * as GFP_NOFS and we really don't want to have to be allocating
1933 * higher-order pages when we are unable to shrink dcache.
1935 if (flags & SLAB_RECLAIM_ACCOUNT)
1939 * Large number of objects is good, but very large slabs are
1940 * currently bad for the gfp()s.
1942 if (gfporder >= slab_break_gfp_order)
1946 * Acceptable internal fragmentation?
1948 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1954 static void setup_cpu_cache(struct kmem_cache *cachep)
1956 if (g_cpucache_up == FULL) {
1957 enable_cpucache(cachep);
1960 if (g_cpucache_up == NONE) {
1962 * Note: the first kmem_cache_create must create the cache
1963 * that's used by kmalloc(24), otherwise the creation of
1964 * further caches will BUG().
1966 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1969 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1970 * the first cache, then we need to set up all its list3s,
1971 * otherwise the creation of further caches will BUG().
1973 set_up_list3s(cachep, SIZE_AC);
1974 if (INDEX_AC == INDEX_L3)
1975 g_cpucache_up = PARTIAL_L3;
1977 g_cpucache_up = PARTIAL_AC;
1979 cachep->array[smp_processor_id()] =
1980 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1982 if (g_cpucache_up == PARTIAL_AC) {
1983 set_up_list3s(cachep, SIZE_L3);
1984 g_cpucache_up = PARTIAL_L3;
1987 for_each_online_node(node) {
1988 cachep->nodelists[node] =
1989 kmalloc_node(sizeof(struct kmem_list3),
1991 BUG_ON(!cachep->nodelists[node]);
1992 kmem_list3_init(cachep->nodelists[node]);
1996 cachep->nodelists[numa_node_id()]->next_reap =
1997 jiffies + REAPTIMEOUT_LIST3 +
1998 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2000 cpu_cache_get(cachep)->avail = 0;
2001 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2002 cpu_cache_get(cachep)->batchcount = 1;
2003 cpu_cache_get(cachep)->touched = 0;
2004 cachep->batchcount = 1;
2005 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2009 * kmem_cache_create - Create a cache.
2010 * @name: A string which is used in /proc/slabinfo to identify this cache.
2011 * @size: The size of objects to be created in this cache.
2012 * @align: The required alignment for the objects.
2013 * @flags: SLAB flags
2014 * @ctor: A constructor for the objects.
2015 * @dtor: A destructor for the objects.
2017 * Returns a ptr to the cache on success, NULL on failure.
2018 * Cannot be called within a int, but can be interrupted.
2019 * The @ctor is run when new pages are allocated by the cache
2020 * and the @dtor is run before the pages are handed back.
2022 * @name must be valid until the cache is destroyed. This implies that
2023 * the module calling this has to destroy the cache before getting unloaded.
2027 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2028 * to catch references to uninitialised memory.
2030 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2031 * for buffer overruns.
2033 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2034 * cacheline. This can be beneficial if you're counting cycles as closely
2038 kmem_cache_create (const char *name, size_t size, size_t align,
2039 unsigned long flags,
2040 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2041 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2043 size_t left_over, slab_size, ralign;
2044 struct kmem_cache *cachep = NULL, *pc;
2047 * Sanity checks... these are all serious usage bugs.
2049 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2050 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2051 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2057 * Prevent CPUs from coming and going.
2058 * lock_cpu_hotplug() nests outside cache_chain_mutex
2062 mutex_lock(&cache_chain_mutex);
2064 list_for_each_entry(pc, &cache_chain, next) {
2065 mm_segment_t old_fs = get_fs();
2070 * This happens when the module gets unloaded and doesn't
2071 * destroy its slab cache and no-one else reuses the vmalloc
2072 * area of the module. Print a warning.
2075 res = __get_user(tmp, pc->name);
2078 printk("SLAB: cache with size %d has lost its name\n",
2083 if (!strcmp(pc->name, name)) {
2084 printk("kmem_cache_create: duplicate cache %s\n", name);
2091 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2092 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2093 /* No constructor, but inital state check requested */
2094 printk(KERN_ERR "%s: No con, but init state check "
2095 "requested - %s\n", __FUNCTION__, name);
2096 flags &= ~SLAB_DEBUG_INITIAL;
2100 * Enable redzoning and last user accounting, except for caches with
2101 * large objects, if the increased size would increase the object size
2102 * above the next power of two: caches with object sizes just above a
2103 * power of two have a significant amount of internal fragmentation.
2105 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2106 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2107 if (!(flags & SLAB_DESTROY_BY_RCU))
2108 flags |= SLAB_POISON;
2110 if (flags & SLAB_DESTROY_BY_RCU)
2111 BUG_ON(flags & SLAB_POISON);
2113 if (flags & SLAB_DESTROY_BY_RCU)
2117 * Always checks flags, a caller might be expecting debug support which
2120 BUG_ON(flags & ~CREATE_MASK);
2123 * Check that size is in terms of words. This is needed to avoid
2124 * unaligned accesses for some archs when redzoning is used, and makes
2125 * sure any on-slab bufctl's are also correctly aligned.
2127 if (size & (BYTES_PER_WORD - 1)) {
2128 size += (BYTES_PER_WORD - 1);
2129 size &= ~(BYTES_PER_WORD - 1);
2132 /* calculate the final buffer alignment: */
2134 /* 1) arch recommendation: can be overridden for debug */
2135 if (flags & SLAB_HWCACHE_ALIGN) {
2137 * Default alignment: as specified by the arch code. Except if
2138 * an object is really small, then squeeze multiple objects into
2141 ralign = cache_line_size();
2142 while (size <= ralign / 2)
2145 ralign = BYTES_PER_WORD;
2147 /* 2) arch mandated alignment: disables debug if necessary */
2148 if (ralign < ARCH_SLAB_MINALIGN) {
2149 ralign = ARCH_SLAB_MINALIGN;
2150 if (ralign > BYTES_PER_WORD)
2151 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2153 /* 3) caller mandated alignment: disables debug if necessary */
2154 if (ralign < align) {
2156 if (ralign > BYTES_PER_WORD)
2157 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2160 * 4) Store it. Note that the debug code below can reduce
2161 * the alignment to BYTES_PER_WORD.
2165 /* Get cache's description obj. */
2166 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
2171 cachep->obj_size = size;
2173 if (flags & SLAB_RED_ZONE) {
2174 /* redzoning only works with word aligned caches */
2175 align = BYTES_PER_WORD;
2177 /* add space for red zone words */
2178 cachep->obj_offset += BYTES_PER_WORD;
2179 size += 2 * BYTES_PER_WORD;
2181 if (flags & SLAB_STORE_USER) {
2182 /* user store requires word alignment and
2183 * one word storage behind the end of the real
2186 align = BYTES_PER_WORD;
2187 size += BYTES_PER_WORD;
2189 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2190 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2191 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2192 cachep->obj_offset += PAGE_SIZE - size;
2199 * Determine if the slab management is 'on' or 'off' slab.
2200 * (bootstrapping cannot cope with offslab caches so don't do
2203 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2205 * Size is large, assume best to place the slab management obj
2206 * off-slab (should allow better packing of objs).
2208 flags |= CFLGS_OFF_SLAB;
2210 size = ALIGN(size, align);
2212 left_over = calculate_slab_order(cachep, size, align, flags);
2215 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2216 kmem_cache_free(&cache_cache, cachep);
2220 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2221 + sizeof(struct slab), align);
2224 * If the slab has been placed off-slab, and we have enough space then
2225 * move it on-slab. This is at the expense of any extra colouring.
2227 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2228 flags &= ~CFLGS_OFF_SLAB;
2229 left_over -= slab_size;
2232 if (flags & CFLGS_OFF_SLAB) {
2233 /* really off slab. No need for manual alignment */
2235 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2238 cachep->colour_off = cache_line_size();
2239 /* Offset must be a multiple of the alignment. */
2240 if (cachep->colour_off < align)
2241 cachep->colour_off = align;
2242 cachep->colour = left_over / cachep->colour_off;
2243 cachep->slab_size = slab_size;
2244 cachep->flags = flags;
2245 cachep->gfpflags = 0;
2246 if (flags & SLAB_CACHE_DMA)
2247 cachep->gfpflags |= GFP_DMA;
2248 cachep->buffer_size = size;
2250 if (flags & CFLGS_OFF_SLAB)
2251 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2252 cachep->ctor = ctor;
2253 cachep->dtor = dtor;
2254 cachep->name = name;
2257 setup_cpu_cache(cachep);
2259 /* cache setup completed, link it into the list */
2260 list_add(&cachep->next, &cache_chain);
2262 if (!cachep && (flags & SLAB_PANIC))
2263 panic("kmem_cache_create(): failed to create slab `%s'\n",
2265 mutex_unlock(&cache_chain_mutex);
2266 unlock_cpu_hotplug();
2269 EXPORT_SYMBOL(kmem_cache_create);
2272 static void check_irq_off(void)
2274 BUG_ON(!irqs_disabled());
2277 static void check_irq_on(void)
2279 BUG_ON(irqs_disabled());
2282 static void check_spinlock_acquired(struct kmem_cache *cachep)
2286 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2290 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2294 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2299 #define check_irq_off() do { } while(0)
2300 #define check_irq_on() do { } while(0)
2301 #define check_spinlock_acquired(x) do { } while(0)
2302 #define check_spinlock_acquired_node(x, y) do { } while(0)
2305 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2306 struct array_cache *ac,
2307 int force, int node);
2309 static void do_drain(void *arg)
2311 struct kmem_cache *cachep = arg;
2312 struct array_cache *ac;
2313 int node = numa_node_id();
2316 ac = cpu_cache_get(cachep);
2317 spin_lock(&cachep->nodelists[node]->list_lock);
2318 free_block(cachep, ac->entry, ac->avail, node);
2319 spin_unlock(&cachep->nodelists[node]->list_lock);
2323 static void drain_cpu_caches(struct kmem_cache *cachep)
2325 struct kmem_list3 *l3;
2328 on_each_cpu(do_drain, cachep, 1, 1);
2330 for_each_online_node(node) {
2331 l3 = cachep->nodelists[node];
2332 if (l3 && l3->alien)
2333 drain_alien_cache(cachep, l3->alien);
2336 for_each_online_node(node) {
2337 l3 = cachep->nodelists[node];
2339 drain_array(cachep, l3, l3->shared, 1, node);
2344 * Remove slabs from the list of free slabs.
2345 * Specify the number of slabs to drain in tofree.
2347 * Returns the actual number of slabs released.
2349 static int drain_freelist(struct kmem_cache *cache,
2350 struct kmem_list3 *l3, int tofree)
2352 struct list_head *p;
2357 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2359 spin_lock_irq(&l3->list_lock);
2360 p = l3->slabs_free.prev;
2361 if (p == &l3->slabs_free) {
2362 spin_unlock_irq(&l3->list_lock);
2366 slabp = list_entry(p, struct slab, list);
2368 BUG_ON(slabp->inuse);
2370 list_del(&slabp->list);
2372 * Safe to drop the lock. The slab is no longer linked
2375 l3->free_objects -= cache->num;
2376 spin_unlock_irq(&l3->list_lock);
2377 slab_destroy(cache, slabp);
2384 static int __cache_shrink(struct kmem_cache *cachep)
2387 struct kmem_list3 *l3;
2389 drain_cpu_caches(cachep);
2392 for_each_online_node(i) {
2393 l3 = cachep->nodelists[i];
2397 drain_freelist(cachep, l3, l3->free_objects);
2399 ret += !list_empty(&l3->slabs_full) ||
2400 !list_empty(&l3->slabs_partial);
2402 return (ret ? 1 : 0);
2406 * kmem_cache_shrink - Shrink a cache.
2407 * @cachep: The cache to shrink.
2409 * Releases as many slabs as possible for a cache.
2410 * To help debugging, a zero exit status indicates all slabs were released.
2412 int kmem_cache_shrink(struct kmem_cache *cachep)
2414 BUG_ON(!cachep || in_interrupt());
2416 return __cache_shrink(cachep);
2418 EXPORT_SYMBOL(kmem_cache_shrink);
2421 * kmem_cache_destroy - delete a cache
2422 * @cachep: the cache to destroy
2424 * Remove a struct kmem_cache object from the slab cache.
2425 * Returns 0 on success.
2427 * It is expected this function will be called by a module when it is
2428 * unloaded. This will remove the cache completely, and avoid a duplicate
2429 * cache being allocated each time a module is loaded and unloaded, if the
2430 * module doesn't have persistent in-kernel storage across loads and unloads.
2432 * The cache must be empty before calling this function.
2434 * The caller must guarantee that noone will allocate memory from the cache
2435 * during the kmem_cache_destroy().
2437 int kmem_cache_destroy(struct kmem_cache *cachep)
2440 struct kmem_list3 *l3;
2442 BUG_ON(!cachep || in_interrupt());
2444 /* Don't let CPUs to come and go */
2447 /* Find the cache in the chain of caches. */
2448 mutex_lock(&cache_chain_mutex);
2450 * the chain is never empty, cache_cache is never destroyed
2452 list_del(&cachep->next);
2453 mutex_unlock(&cache_chain_mutex);
2455 if (__cache_shrink(cachep)) {
2456 slab_error(cachep, "Can't free all objects");
2457 mutex_lock(&cache_chain_mutex);
2458 list_add(&cachep->next, &cache_chain);
2459 mutex_unlock(&cache_chain_mutex);
2460 unlock_cpu_hotplug();
2464 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2467 for_each_online_cpu(i)
2468 kfree(cachep->array[i]);
2470 /* NUMA: free the list3 structures */
2471 for_each_online_node(i) {
2472 l3 = cachep->nodelists[i];
2475 free_alien_cache(l3->alien);
2479 kmem_cache_free(&cache_cache, cachep);
2480 unlock_cpu_hotplug();
2483 EXPORT_SYMBOL(kmem_cache_destroy);
2485 /* Get the memory for a slab management obj. */
2486 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2487 int colour_off, gfp_t local_flags,
2492 if (OFF_SLAB(cachep)) {
2493 /* Slab management obj is off-slab. */
2494 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2495 local_flags, nodeid);
2499 slabp = objp + colour_off;
2500 colour_off += cachep->slab_size;
2503 slabp->colouroff = colour_off;
2504 slabp->s_mem = objp + colour_off;
2505 slabp->nodeid = nodeid;
2509 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2511 return (kmem_bufctl_t *) (slabp + 1);
2514 static void cache_init_objs(struct kmem_cache *cachep,
2515 struct slab *slabp, unsigned long ctor_flags)
2519 for (i = 0; i < cachep->num; i++) {
2520 void *objp = index_to_obj(cachep, slabp, i);
2522 /* need to poison the objs? */
2523 if (cachep->flags & SLAB_POISON)
2524 poison_obj(cachep, objp, POISON_FREE);
2525 if (cachep->flags & SLAB_STORE_USER)
2526 *dbg_userword(cachep, objp) = NULL;
2528 if (cachep->flags & SLAB_RED_ZONE) {
2529 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2530 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2533 * Constructors are not allowed to allocate memory from the same
2534 * cache which they are a constructor for. Otherwise, deadlock.
2535 * They must also be threaded.
2537 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2538 cachep->ctor(objp + obj_offset(cachep), cachep,
2541 if (cachep->flags & SLAB_RED_ZONE) {
2542 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2543 slab_error(cachep, "constructor overwrote the"
2544 " end of an object");
2545 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2546 slab_error(cachep, "constructor overwrote the"
2547 " start of an object");
2549 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2550 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2551 kernel_map_pages(virt_to_page(objp),
2552 cachep->buffer_size / PAGE_SIZE, 0);
2555 cachep->ctor(objp, cachep, ctor_flags);
2557 slab_bufctl(slabp)[i] = i + 1;
2559 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2563 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2565 if (flags & SLAB_DMA)
2566 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2568 BUG_ON(cachep->gfpflags & GFP_DMA);
2571 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2574 void *objp = index_to_obj(cachep, slabp, slabp->free);
2578 next = slab_bufctl(slabp)[slabp->free];
2580 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2581 WARN_ON(slabp->nodeid != nodeid);
2588 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2589 void *objp, int nodeid)
2591 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2594 /* Verify that the slab belongs to the intended node */
2595 WARN_ON(slabp->nodeid != nodeid);
2597 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2598 printk(KERN_ERR "slab: double free detected in cache "
2599 "'%s', objp %p\n", cachep->name, objp);
2603 slab_bufctl(slabp)[objnr] = slabp->free;
2604 slabp->free = objnr;
2609 * Map pages beginning at addr to the given cache and slab. This is required
2610 * for the slab allocator to be able to lookup the cache and slab of a
2611 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2613 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2619 page = virt_to_page(addr);
2622 if (likely(!PageCompound(page)))
2623 nr_pages <<= cache->gfporder;
2626 page_set_cache(page, cache);
2627 page_set_slab(page, slab);
2629 } while (--nr_pages);
2633 * Grow (by 1) the number of slabs within a cache. This is called by
2634 * kmem_cache_alloc() when there are no active objs left in a cache.
2636 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
2642 unsigned long ctor_flags;
2643 struct kmem_list3 *l3;
2646 * Be lazy and only check for valid flags here, keeping it out of the
2647 * critical path in kmem_cache_alloc().
2649 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
2650 if (flags & SLAB_NO_GROW)
2653 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2654 local_flags = (flags & SLAB_LEVEL_MASK);
2655 if (!(local_flags & __GFP_WAIT))
2657 * Not allowed to sleep. Need to tell a constructor about
2658 * this - it might need to know...
2660 ctor_flags |= SLAB_CTOR_ATOMIC;
2662 /* Take the l3 list lock to change the colour_next on this node */
2664 l3 = cachep->nodelists[nodeid];
2665 spin_lock(&l3->list_lock);
2667 /* Get colour for the slab, and cal the next value. */
2668 offset = l3->colour_next;
2670 if (l3->colour_next >= cachep->colour)
2671 l3->colour_next = 0;
2672 spin_unlock(&l3->list_lock);
2674 offset *= cachep->colour_off;
2676 if (local_flags & __GFP_WAIT)
2680 * The test for missing atomic flag is performed here, rather than
2681 * the more obvious place, simply to reduce the critical path length
2682 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2683 * will eventually be caught here (where it matters).
2685 kmem_flagcheck(cachep, flags);
2688 * Get mem for the objs. Attempt to allocate a physical page from
2691 objp = kmem_getpages(cachep, flags, nodeid);
2695 /* Get slab management. */
2696 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
2700 slabp->nodeid = nodeid;
2701 slab_map_pages(cachep, slabp, objp);
2703 cache_init_objs(cachep, slabp, ctor_flags);
2705 if (local_flags & __GFP_WAIT)
2706 local_irq_disable();
2708 spin_lock(&l3->list_lock);
2710 /* Make slab active. */
2711 list_add_tail(&slabp->list, &(l3->slabs_free));
2712 STATS_INC_GROWN(cachep);
2713 l3->free_objects += cachep->num;
2714 spin_unlock(&l3->list_lock);
2717 kmem_freepages(cachep, objp);
2719 if (local_flags & __GFP_WAIT)
2720 local_irq_disable();
2727 * Perform extra freeing checks:
2728 * - detect bad pointers.
2729 * - POISON/RED_ZONE checking
2730 * - destructor calls, for caches with POISON+dtor
2732 static void kfree_debugcheck(const void *objp)
2736 if (!virt_addr_valid(objp)) {
2737 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2738 (unsigned long)objp);
2741 page = virt_to_page(objp);
2742 if (!PageSlab(page)) {
2743 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2744 (unsigned long)objp);
2749 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2751 unsigned long redzone1, redzone2;
2753 redzone1 = *dbg_redzone1(cache, obj);
2754 redzone2 = *dbg_redzone2(cache, obj);
2759 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2762 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2763 slab_error(cache, "double free detected");
2765 slab_error(cache, "memory outside object was overwritten");
2767 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2768 obj, redzone1, redzone2);
2771 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2778 objp -= obj_offset(cachep);
2779 kfree_debugcheck(objp);
2780 page = virt_to_page(objp);
2782 slabp = page_get_slab(page);
2784 if (cachep->flags & SLAB_RED_ZONE) {
2785 verify_redzone_free(cachep, objp);
2786 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2787 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2789 if (cachep->flags & SLAB_STORE_USER)
2790 *dbg_userword(cachep, objp) = caller;
2792 objnr = obj_to_index(cachep, slabp, objp);
2794 BUG_ON(objnr >= cachep->num);
2795 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2797 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2799 * Need to call the slab's constructor so the caller can
2800 * perform a verify of its state (debugging). Called without
2801 * the cache-lock held.
2803 cachep->ctor(objp + obj_offset(cachep),
2804 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2806 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2807 /* we want to cache poison the object,
2808 * call the destruction callback
2810 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2812 #ifdef CONFIG_DEBUG_SLAB_LEAK
2813 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2815 if (cachep->flags & SLAB_POISON) {
2816 #ifdef CONFIG_DEBUG_PAGEALLOC
2817 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2818 store_stackinfo(cachep, objp, (unsigned long)caller);
2819 kernel_map_pages(virt_to_page(objp),
2820 cachep->buffer_size / PAGE_SIZE, 0);
2822 poison_obj(cachep, objp, POISON_FREE);
2825 poison_obj(cachep, objp, POISON_FREE);
2831 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2836 /* Check slab's freelist to see if this obj is there. */
2837 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2839 if (entries > cachep->num || i >= cachep->num)
2842 if (entries != cachep->num - slabp->inuse) {
2844 printk(KERN_ERR "slab: Internal list corruption detected in "
2845 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2846 cachep->name, cachep->num, slabp, slabp->inuse);
2848 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2851 printk("\n%03x:", i);
2852 printk(" %02x", ((unsigned char *)slabp)[i]);
2859 #define kfree_debugcheck(x) do { } while(0)
2860 #define cache_free_debugcheck(x,objp,z) (objp)
2861 #define check_slabp(x,y) do { } while(0)
2864 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2867 struct kmem_list3 *l3;
2868 struct array_cache *ac;
2871 ac = cpu_cache_get(cachep);
2873 batchcount = ac->batchcount;
2874 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2876 * If there was little recent activity on this cache, then
2877 * perform only a partial refill. Otherwise we could generate
2880 batchcount = BATCHREFILL_LIMIT;
2882 l3 = cachep->nodelists[numa_node_id()];
2884 BUG_ON(ac->avail > 0 || !l3);
2885 spin_lock(&l3->list_lock);
2887 /* See if we can refill from the shared array */
2888 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2891 while (batchcount > 0) {
2892 struct list_head *entry;
2894 /* Get slab alloc is to come from. */
2895 entry = l3->slabs_partial.next;
2896 if (entry == &l3->slabs_partial) {
2897 l3->free_touched = 1;
2898 entry = l3->slabs_free.next;
2899 if (entry == &l3->slabs_free)
2903 slabp = list_entry(entry, struct slab, list);
2904 check_slabp(cachep, slabp);
2905 check_spinlock_acquired(cachep);
2906 while (slabp->inuse < cachep->num && batchcount--) {
2907 STATS_INC_ALLOCED(cachep);
2908 STATS_INC_ACTIVE(cachep);
2909 STATS_SET_HIGH(cachep);
2911 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2914 check_slabp(cachep, slabp);
2916 /* move slabp to correct slabp list: */
2917 list_del(&slabp->list);
2918 if (slabp->free == BUFCTL_END)
2919 list_add(&slabp->list, &l3->slabs_full);
2921 list_add(&slabp->list, &l3->slabs_partial);
2925 l3->free_objects -= ac->avail;
2927 spin_unlock(&l3->list_lock);
2929 if (unlikely(!ac->avail)) {
2931 x = cache_grow(cachep, flags, numa_node_id());
2933 /* cache_grow can reenable interrupts, then ac could change. */
2934 ac = cpu_cache_get(cachep);
2935 if (!x && ac->avail == 0) /* no objects in sight? abort */
2938 if (!ac->avail) /* objects refilled by interrupt? */
2942 return ac->entry[--ac->avail];
2945 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2948 might_sleep_if(flags & __GFP_WAIT);
2950 kmem_flagcheck(cachep, flags);
2955 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2956 gfp_t flags, void *objp, void *caller)
2960 if (cachep->flags & SLAB_POISON) {
2961 #ifdef CONFIG_DEBUG_PAGEALLOC
2962 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2963 kernel_map_pages(virt_to_page(objp),
2964 cachep->buffer_size / PAGE_SIZE, 1);
2966 check_poison_obj(cachep, objp);
2968 check_poison_obj(cachep, objp);
2970 poison_obj(cachep, objp, POISON_INUSE);
2972 if (cachep->flags & SLAB_STORE_USER)
2973 *dbg_userword(cachep, objp) = caller;
2975 if (cachep->flags & SLAB_RED_ZONE) {
2976 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2977 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2978 slab_error(cachep, "double free, or memory outside"
2979 " object was overwritten");
2981 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2982 objp, *dbg_redzone1(cachep, objp),
2983 *dbg_redzone2(cachep, objp));
2985 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2986 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2988 #ifdef CONFIG_DEBUG_SLAB_LEAK
2993 slabp = page_get_slab(virt_to_page(objp));
2994 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2995 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2998 objp += obj_offset(cachep);
2999 if (cachep->ctor && cachep->flags & SLAB_POISON) {
3000 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
3002 if (!(flags & __GFP_WAIT))
3003 ctor_flags |= SLAB_CTOR_ATOMIC;
3005 cachep->ctor(objp, cachep, ctor_flags);
3010 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3013 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3016 struct array_cache *ac;
3019 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3020 objp = alternate_node_alloc(cachep, flags);
3027 ac = cpu_cache_get(cachep);
3028 if (likely(ac->avail)) {
3029 STATS_INC_ALLOCHIT(cachep);
3031 objp = ac->entry[--ac->avail];
3033 STATS_INC_ALLOCMISS(cachep);
3034 objp = cache_alloc_refill(cachep, flags);
3039 static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3040 gfp_t flags, void *caller)
3042 unsigned long save_flags;
3045 cache_alloc_debugcheck_before(cachep, flags);
3047 local_irq_save(save_flags);
3048 objp = ____cache_alloc(cachep, flags);
3049 local_irq_restore(save_flags);
3050 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
3058 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3060 * If we are in_interrupt, then process context, including cpusets and
3061 * mempolicy, may not apply and should not be used for allocation policy.
3063 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3065 int nid_alloc, nid_here;
3069 nid_alloc = nid_here = numa_node_id();
3070 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3071 nid_alloc = cpuset_mem_spread_node();
3072 else if (current->mempolicy)
3073 nid_alloc = slab_node(current->mempolicy);
3074 if (nid_alloc != nid_here)
3075 return __cache_alloc_node(cachep, flags, nid_alloc);
3080 * A interface to enable slab creation on nodeid
3082 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3085 struct list_head *entry;
3087 struct kmem_list3 *l3;
3091 l3 = cachep->nodelists[nodeid];
3096 spin_lock(&l3->list_lock);
3097 entry = l3->slabs_partial.next;
3098 if (entry == &l3->slabs_partial) {
3099 l3->free_touched = 1;
3100 entry = l3->slabs_free.next;
3101 if (entry == &l3->slabs_free)
3105 slabp = list_entry(entry, struct slab, list);
3106 check_spinlock_acquired_node(cachep, nodeid);
3107 check_slabp(cachep, slabp);
3109 STATS_INC_NODEALLOCS(cachep);
3110 STATS_INC_ACTIVE(cachep);
3111 STATS_SET_HIGH(cachep);
3113 BUG_ON(slabp->inuse == cachep->num);
3115 obj = slab_get_obj(cachep, slabp, nodeid);
3116 check_slabp(cachep, slabp);
3118 /* move slabp to correct slabp list: */
3119 list_del(&slabp->list);
3121 if (slabp->free == BUFCTL_END)
3122 list_add(&slabp->list, &l3->slabs_full);
3124 list_add(&slabp->list, &l3->slabs_partial);
3126 spin_unlock(&l3->list_lock);
3130 spin_unlock(&l3->list_lock);
3131 x = cache_grow(cachep, flags, nodeid);
3143 * Caller needs to acquire correct kmem_list's list_lock
3145 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3149 struct kmem_list3 *l3;
3151 for (i = 0; i < nr_objects; i++) {
3152 void *objp = objpp[i];
3155 slabp = virt_to_slab(objp);
3156 l3 = cachep->nodelists[node];
3157 list_del(&slabp->list);
3158 check_spinlock_acquired_node(cachep, node);
3159 check_slabp(cachep, slabp);
3160 slab_put_obj(cachep, slabp, objp, node);
3161 STATS_DEC_ACTIVE(cachep);
3163 check_slabp(cachep, slabp);
3165 /* fixup slab chains */
3166 if (slabp->inuse == 0) {
3167 if (l3->free_objects > l3->free_limit) {
3168 l3->free_objects -= cachep->num;
3169 slab_destroy(cachep, slabp);
3171 list_add(&slabp->list, &l3->slabs_free);
3174 /* Unconditionally move a slab to the end of the
3175 * partial list on free - maximum time for the
3176 * other objects to be freed, too.
3178 list_add_tail(&slabp->list, &l3->slabs_partial);
3183 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3186 struct kmem_list3 *l3;
3187 int node = numa_node_id();
3189 batchcount = ac->batchcount;
3191 BUG_ON(!batchcount || batchcount > ac->avail);
3194 l3 = cachep->nodelists[node];
3195 spin_lock(&l3->list_lock);
3197 struct array_cache *shared_array = l3->shared;
3198 int max = shared_array->limit - shared_array->avail;
3200 if (batchcount > max)
3202 memcpy(&(shared_array->entry[shared_array->avail]),
3203 ac->entry, sizeof(void *) * batchcount);
3204 shared_array->avail += batchcount;
3209 free_block(cachep, ac->entry, batchcount, node);
3214 struct list_head *p;
3216 p = l3->slabs_free.next;
3217 while (p != &(l3->slabs_free)) {
3220 slabp = list_entry(p, struct slab, list);
3221 BUG_ON(slabp->inuse);
3226 STATS_SET_FREEABLE(cachep, i);
3229 spin_unlock(&l3->list_lock);
3230 ac->avail -= batchcount;
3231 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3235 * Release an obj back to its cache. If the obj has a constructed state, it must
3236 * be in this state _before_ it is released. Called with disabled ints.
3238 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3240 struct array_cache *ac = cpu_cache_get(cachep);
3243 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3245 if (cache_free_alien(cachep, objp))
3248 if (likely(ac->avail < ac->limit)) {
3249 STATS_INC_FREEHIT(cachep);
3250 ac->entry[ac->avail++] = objp;
3253 STATS_INC_FREEMISS(cachep);
3254 cache_flusharray(cachep, ac);
3255 ac->entry[ac->avail++] = objp;
3260 * kmem_cache_alloc - Allocate an object
3261 * @cachep: The cache to allocate from.
3262 * @flags: See kmalloc().
3264 * Allocate an object from this cache. The flags are only relevant
3265 * if the cache has no available objects.
3267 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3269 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3271 EXPORT_SYMBOL(kmem_cache_alloc);
3274 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3275 * @cache: The cache to allocate from.
3276 * @flags: See kmalloc().
3278 * Allocate an object from this cache and set the allocated memory to zero.
3279 * The flags are only relevant if the cache has no available objects.
3281 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3283 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3285 memset(ret, 0, obj_size(cache));
3288 EXPORT_SYMBOL(kmem_cache_zalloc);
3291 * kmem_ptr_validate - check if an untrusted pointer might
3293 * @cachep: the cache we're checking against
3294 * @ptr: pointer to validate
3296 * This verifies that the untrusted pointer looks sane:
3297 * it is _not_ a guarantee that the pointer is actually
3298 * part of the slab cache in question, but it at least
3299 * validates that the pointer can be dereferenced and
3300 * looks half-way sane.
3302 * Currently only used for dentry validation.
3304 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
3306 unsigned long addr = (unsigned long)ptr;
3307 unsigned long min_addr = PAGE_OFFSET;
3308 unsigned long align_mask = BYTES_PER_WORD - 1;
3309 unsigned long size = cachep->buffer_size;
3312 if (unlikely(addr < min_addr))
3314 if (unlikely(addr > (unsigned long)high_memory - size))
3316 if (unlikely(addr & align_mask))
3318 if (unlikely(!kern_addr_valid(addr)))
3320 if (unlikely(!kern_addr_valid(addr + size - 1)))
3322 page = virt_to_page(ptr);
3323 if (unlikely(!PageSlab(page)))
3325 if (unlikely(page_get_cache(page) != cachep))
3334 * kmem_cache_alloc_node - Allocate an object on the specified node
3335 * @cachep: The cache to allocate from.
3336 * @flags: See kmalloc().
3337 * @nodeid: node number of the target node.
3339 * Identical to kmem_cache_alloc, except that this function is slow
3340 * and can sleep. And it will allocate memory on the given node, which
3341 * can improve the performance for cpu bound structures.
3342 * New and improved: it will now make sure that the object gets
3343 * put on the correct node list so that there is no false sharing.
3345 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3347 unsigned long save_flags;
3350 cache_alloc_debugcheck_before(cachep, flags);
3351 local_irq_save(save_flags);
3353 if (nodeid == -1 || nodeid == numa_node_id() ||
3354 !cachep->nodelists[nodeid])
3355 ptr = ____cache_alloc(cachep, flags);
3357 ptr = __cache_alloc_node(cachep, flags, nodeid);
3358 local_irq_restore(save_flags);
3360 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3361 __builtin_return_address(0));
3365 EXPORT_SYMBOL(kmem_cache_alloc_node);
3367 void *kmalloc_node(size_t size, gfp_t flags, int node)
3369 struct kmem_cache *cachep;
3371 cachep = kmem_find_general_cachep(size, flags);
3372 if (unlikely(cachep == NULL))
3374 return kmem_cache_alloc_node(cachep, flags, node);
3376 EXPORT_SYMBOL(kmalloc_node);
3380 * __do_kmalloc - allocate memory
3381 * @size: how many bytes of memory are required.
3382 * @flags: the type of memory to allocate (see kmalloc).
3383 * @caller: function caller for debug tracking of the caller
3385 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3388 struct kmem_cache *cachep;
3390 /* If you want to save a few bytes .text space: replace
3392 * Then kmalloc uses the uninlined functions instead of the inline
3395 cachep = __find_general_cachep(size, flags);
3396 if (unlikely(cachep == NULL))
3398 return __cache_alloc(cachep, flags, caller);
3402 void *__kmalloc(size_t size, gfp_t flags)
3404 #ifndef CONFIG_DEBUG_SLAB
3405 return __do_kmalloc(size, flags, NULL);
3407 return __do_kmalloc(size, flags, __builtin_return_address(0));
3410 EXPORT_SYMBOL(__kmalloc);
3412 #ifdef CONFIG_DEBUG_SLAB
3413 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3415 return __do_kmalloc(size, flags, caller);
3417 EXPORT_SYMBOL(__kmalloc_track_caller);
3422 * __alloc_percpu - allocate one copy of the object for every present
3423 * cpu in the system, zeroing them.
3424 * Objects should be dereferenced using the per_cpu_ptr macro only.
3426 * @size: how many bytes of memory are required.
3428 void *__alloc_percpu(size_t size)
3431 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
3437 * Cannot use for_each_online_cpu since a cpu may come online
3438 * and we have no way of figuring out how to fix the array
3439 * that we have allocated then....
3441 for_each_possible_cpu(i) {
3442 int node = cpu_to_node(i);
3444 if (node_online(node))
3445 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3447 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
3449 if (!pdata->ptrs[i])
3451 memset(pdata->ptrs[i], 0, size);
3454 /* Catch derefs w/o wrappers */
3455 return (void *)(~(unsigned long)pdata);
3459 if (!cpu_possible(i))
3461 kfree(pdata->ptrs[i]);
3466 EXPORT_SYMBOL(__alloc_percpu);
3470 * kmem_cache_free - Deallocate an object
3471 * @cachep: The cache the allocation was from.
3472 * @objp: The previously allocated object.
3474 * Free an object which was previously allocated from this
3477 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3479 unsigned long flags;
3481 BUG_ON(virt_to_cache(objp) != cachep);
3483 local_irq_save(flags);
3484 __cache_free(cachep, objp);
3485 local_irq_restore(flags);
3487 EXPORT_SYMBOL(kmem_cache_free);
3490 * kfree - free previously allocated memory
3491 * @objp: pointer returned by kmalloc.
3493 * If @objp is NULL, no operation is performed.
3495 * Don't free memory not originally allocated by kmalloc()
3496 * or you will run into trouble.
3498 void kfree(const void *objp)
3500 struct kmem_cache *c;
3501 unsigned long flags;
3503 if (unlikely(!objp))
3505 local_irq_save(flags);
3506 kfree_debugcheck(objp);
3507 c = virt_to_cache(objp);
3508 debug_check_no_locks_freed(objp, obj_size(c));
3509 __cache_free(c, (void *)objp);
3510 local_irq_restore(flags);
3512 EXPORT_SYMBOL(kfree);
3516 * free_percpu - free previously allocated percpu memory
3517 * @objp: pointer returned by alloc_percpu.
3519 * Don't free memory not originally allocated by alloc_percpu()
3520 * The complemented objp is to check for that.
3522 void free_percpu(const void *objp)
3525 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
3528 * We allocate for all cpus so we cannot use for online cpu here.
3530 for_each_possible_cpu(i)
3534 EXPORT_SYMBOL(free_percpu);
3537 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3539 return obj_size(cachep);
3541 EXPORT_SYMBOL(kmem_cache_size);
3543 const char *kmem_cache_name(struct kmem_cache *cachep)
3545 return cachep->name;
3547 EXPORT_SYMBOL_GPL(kmem_cache_name);
3550 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3552 static int alloc_kmemlist(struct kmem_cache *cachep)
3555 struct kmem_list3 *l3;
3556 struct array_cache *new_shared;
3557 struct array_cache **new_alien;
3559 for_each_online_node(node) {
3561 new_alien = alloc_alien_cache(node, cachep->limit);
3565 new_shared = alloc_arraycache(node,
3566 cachep->shared*cachep->batchcount,
3569 free_alien_cache(new_alien);
3573 l3 = cachep->nodelists[node];
3575 struct array_cache *shared = l3->shared;
3577 spin_lock_irq(&l3->list_lock);
3580 free_block(cachep, shared->entry,
3581 shared->avail, node);
3583 l3->shared = new_shared;
3585 l3->alien = new_alien;
3588 l3->free_limit = (1 + nr_cpus_node(node)) *
3589 cachep->batchcount + cachep->num;
3590 spin_unlock_irq(&l3->list_lock);
3592 free_alien_cache(new_alien);
3595 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3597 free_alien_cache(new_alien);
3602 kmem_list3_init(l3);
3603 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3604 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3605 l3->shared = new_shared;
3606 l3->alien = new_alien;
3607 l3->free_limit = (1 + nr_cpus_node(node)) *
3608 cachep->batchcount + cachep->num;
3609 cachep->nodelists[node] = l3;
3614 if (!cachep->next.next) {
3615 /* Cache is not active yet. Roll back what we did */
3618 if (cachep->nodelists[node]) {
3619 l3 = cachep->nodelists[node];
3622 free_alien_cache(l3->alien);
3624 cachep->nodelists[node] = NULL;
3632 struct ccupdate_struct {
3633 struct kmem_cache *cachep;
3634 struct array_cache *new[NR_CPUS];
3637 static void do_ccupdate_local(void *info)
3639 struct ccupdate_struct *new = info;
3640 struct array_cache *old;
3643 old = cpu_cache_get(new->cachep);
3645 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3646 new->new[smp_processor_id()] = old;
3649 /* Always called with the cache_chain_mutex held */
3650 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3651 int batchcount, int shared)
3653 struct ccupdate_struct new;
3656 memset(&new.new, 0, sizeof(new.new));
3657 for_each_online_cpu(i) {
3658 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3661 for (i--; i >= 0; i--)
3666 new.cachep = cachep;
3668 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
3671 cachep->batchcount = batchcount;
3672 cachep->limit = limit;
3673 cachep->shared = shared;
3675 for_each_online_cpu(i) {
3676 struct array_cache *ccold = new.new[i];
3679 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3680 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3681 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3685 err = alloc_kmemlist(cachep);
3687 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3688 cachep->name, -err);
3694 /* Called with cache_chain_mutex held always */
3695 static void enable_cpucache(struct kmem_cache *cachep)
3701 * The head array serves three purposes:
3702 * - create a LIFO ordering, i.e. return objects that are cache-warm
3703 * - reduce the number of spinlock operations.
3704 * - reduce the number of linked list operations on the slab and
3705 * bufctl chains: array operations are cheaper.
3706 * The numbers are guessed, we should auto-tune as described by
3709 if (cachep->buffer_size > 131072)
3711 else if (cachep->buffer_size > PAGE_SIZE)
3713 else if (cachep->buffer_size > 1024)
3715 else if (cachep->buffer_size > 256)
3721 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3722 * allocation behaviour: Most allocs on one cpu, most free operations
3723 * on another cpu. For these cases, an efficient object passing between
3724 * cpus is necessary. This is provided by a shared array. The array
3725 * replaces Bonwick's magazine layer.
3726 * On uniprocessor, it's functionally equivalent (but less efficient)
3727 * to a larger limit. Thus disabled by default.
3731 if (cachep->buffer_size <= PAGE_SIZE)
3737 * With debugging enabled, large batchcount lead to excessively long
3738 * periods with disabled local interrupts. Limit the batchcount
3743 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3745 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3746 cachep->name, -err);
3750 * Drain an array if it contains any elements taking the l3 lock only if
3751 * necessary. Note that the l3 listlock also protects the array_cache
3752 * if drain_array() is used on the shared array.
3754 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3755 struct array_cache *ac, int force, int node)
3759 if (!ac || !ac->avail)
3761 if (ac->touched && !force) {
3764 spin_lock_irq(&l3->list_lock);
3766 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3767 if (tofree > ac->avail)
3768 tofree = (ac->avail + 1) / 2;
3769 free_block(cachep, ac->entry, tofree, node);
3770 ac->avail -= tofree;
3771 memmove(ac->entry, &(ac->entry[tofree]),
3772 sizeof(void *) * ac->avail);
3774 spin_unlock_irq(&l3->list_lock);
3779 * cache_reap - Reclaim memory from caches.
3780 * @unused: unused parameter
3782 * Called from workqueue/eventd every few seconds.
3784 * - clear the per-cpu caches for this CPU.
3785 * - return freeable pages to the main free memory pool.
3787 * If we cannot acquire the cache chain mutex then just give up - we'll try
3788 * again on the next iteration.
3790 static void cache_reap(void *unused)
3792 struct kmem_cache *searchp;
3793 struct kmem_list3 *l3;
3794 int node = numa_node_id();
3796 if (!mutex_trylock(&cache_chain_mutex)) {
3797 /* Give up. Setup the next iteration. */
3798 schedule_delayed_work(&__get_cpu_var(reap_work),
3803 list_for_each_entry(searchp, &cache_chain, next) {
3807 * We only take the l3 lock if absolutely necessary and we
3808 * have established with reasonable certainty that
3809 * we can do some work if the lock was obtained.
3811 l3 = searchp->nodelists[node];
3813 reap_alien(searchp, l3);
3815 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
3818 * These are racy checks but it does not matter
3819 * if we skip one check or scan twice.
3821 if (time_after(l3->next_reap, jiffies))
3824 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
3826 drain_array(searchp, l3, l3->shared, 0, node);
3828 if (l3->free_touched)
3829 l3->free_touched = 0;
3833 freed = drain_freelist(searchp, l3, (l3->free_limit +
3834 5 * searchp->num - 1) / (5 * searchp->num));
3835 STATS_ADD_REAPED(searchp, freed);
3841 mutex_unlock(&cache_chain_mutex);
3843 refresh_cpu_vm_stats(smp_processor_id());
3844 /* Set up the next iteration */
3845 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
3848 #ifdef CONFIG_PROC_FS
3850 static void print_slabinfo_header(struct seq_file *m)
3853 * Output format version, so at least we can change it
3854 * without _too_ many complaints.
3857 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3859 seq_puts(m, "slabinfo - version: 2.1\n");
3861 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3862 "<objperslab> <pagesperslab>");
3863 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3864 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3866 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3867 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
3868 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3873 static void *s_start(struct seq_file *m, loff_t *pos)
3876 struct list_head *p;
3878 mutex_lock(&cache_chain_mutex);
3880 print_slabinfo_header(m);
3881 p = cache_chain.next;
3884 if (p == &cache_chain)
3887 return list_entry(p, struct kmem_cache, next);
3890 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3892 struct kmem_cache *cachep = p;
3894 return cachep->next.next == &cache_chain ?
3895 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
3898 static void s_stop(struct seq_file *m, void *p)
3900 mutex_unlock(&cache_chain_mutex);
3903 static int s_show(struct seq_file *m, void *p)
3905 struct kmem_cache *cachep = p;
3907 unsigned long active_objs;
3908 unsigned long num_objs;
3909 unsigned long active_slabs = 0;
3910 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3914 struct kmem_list3 *l3;
3918 for_each_online_node(node) {
3919 l3 = cachep->nodelists[node];
3924 spin_lock_irq(&l3->list_lock);
3926 list_for_each_entry(slabp, &l3->slabs_full, list) {
3927 if (slabp->inuse != cachep->num && !error)
3928 error = "slabs_full accounting error";
3929 active_objs += cachep->num;
3932 list_for_each_entry(slabp, &l3->slabs_partial, list) {
3933 if (slabp->inuse == cachep->num && !error)
3934 error = "slabs_partial inuse accounting error";
3935 if (!slabp->inuse && !error)
3936 error = "slabs_partial/inuse accounting error";
3937 active_objs += slabp->inuse;
3940 list_for_each_entry(slabp, &l3->slabs_free, list) {
3941 if (slabp->inuse && !error)
3942 error = "slabs_free/inuse accounting error";
3945 free_objects += l3->free_objects;
3947 shared_avail += l3->shared->avail;
3949 spin_unlock_irq(&l3->list_lock);
3951 num_slabs += active_slabs;
3952 num_objs = num_slabs * cachep->num;
3953 if (num_objs - active_objs != free_objects && !error)
3954 error = "free_objects accounting error";
3956 name = cachep->name;
3958 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3960 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3961 name, active_objs, num_objs, cachep->buffer_size,
3962 cachep->num, (1 << cachep->gfporder));
3963 seq_printf(m, " : tunables %4u %4u %4u",
3964 cachep->limit, cachep->batchcount, cachep->shared);
3965 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3966 active_slabs, num_slabs, shared_avail);
3969 unsigned long high = cachep->high_mark;
3970 unsigned long allocs = cachep->num_allocations;
3971 unsigned long grown = cachep->grown;
3972 unsigned long reaped = cachep->reaped;
3973 unsigned long errors = cachep->errors;
3974 unsigned long max_freeable = cachep->max_freeable;
3975 unsigned long node_allocs = cachep->node_allocs;
3976 unsigned long node_frees = cachep->node_frees;
3977 unsigned long overflows = cachep->node_overflow;
3979 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3980 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
3981 reaped, errors, max_freeable, node_allocs,
3982 node_frees, overflows);
3986 unsigned long allochit = atomic_read(&cachep->allochit);
3987 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3988 unsigned long freehit = atomic_read(&cachep->freehit);
3989 unsigned long freemiss = atomic_read(&cachep->freemiss);
3991 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3992 allochit, allocmiss, freehit, freemiss);
4000 * slabinfo_op - iterator that generates /proc/slabinfo
4009 * num-pages-per-slab
4010 * + further values on SMP and with statistics enabled
4013 struct seq_operations slabinfo_op = {
4020 #define MAX_SLABINFO_WRITE 128
4022 * slabinfo_write - Tuning for the slab allocator
4024 * @buffer: user buffer
4025 * @count: data length
4028 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4029 size_t count, loff_t *ppos)
4031 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4032 int limit, batchcount, shared, res;
4033 struct kmem_cache *cachep;
4035 if (count > MAX_SLABINFO_WRITE)
4037 if (copy_from_user(&kbuf, buffer, count))
4039 kbuf[MAX_SLABINFO_WRITE] = '\0';
4041 tmp = strchr(kbuf, ' ');
4046 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4049 /* Find the cache in the chain of caches. */
4050 mutex_lock(&cache_chain_mutex);
4052 list_for_each_entry(cachep, &cache_chain, next) {
4053 if (!strcmp(cachep->name, kbuf)) {
4054 if (limit < 1 || batchcount < 1 ||
4055 batchcount > limit || shared < 0) {
4058 res = do_tune_cpucache(cachep, limit,
4059 batchcount, shared);
4064 mutex_unlock(&cache_chain_mutex);
4070 #ifdef CONFIG_DEBUG_SLAB_LEAK
4072 static void *leaks_start(struct seq_file *m, loff_t *pos)
4075 struct list_head *p;
4077 mutex_lock(&cache_chain_mutex);
4078 p = cache_chain.next;
4081 if (p == &cache_chain)
4084 return list_entry(p, struct kmem_cache, next);
4087 static inline int add_caller(unsigned long *n, unsigned long v)
4097 unsigned long *q = p + 2 * i;
4111 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4117 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4123 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4124 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4126 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4131 static void show_symbol(struct seq_file *m, unsigned long address)
4133 #ifdef CONFIG_KALLSYMS
4136 unsigned long offset, size;
4137 char namebuf[KSYM_NAME_LEN+1];
4139 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4142 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4144 seq_printf(m, " [%s]", modname);
4148 seq_printf(m, "%p", (void *)address);
4151 static int leaks_show(struct seq_file *m, void *p)
4153 struct kmem_cache *cachep = p;
4155 struct kmem_list3 *l3;
4157 unsigned long *n = m->private;
4161 if (!(cachep->flags & SLAB_STORE_USER))
4163 if (!(cachep->flags & SLAB_RED_ZONE))
4166 /* OK, we can do it */
4170 for_each_online_node(node) {
4171 l3 = cachep->nodelists[node];
4176 spin_lock_irq(&l3->list_lock);
4178 list_for_each_entry(slabp, &l3->slabs_full, list)
4179 handle_slab(n, cachep, slabp);
4180 list_for_each_entry(slabp, &l3->slabs_partial, list)
4181 handle_slab(n, cachep, slabp);
4182 spin_unlock_irq(&l3->list_lock);
4184 name = cachep->name;
4186 /* Increase the buffer size */
4187 mutex_unlock(&cache_chain_mutex);
4188 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4190 /* Too bad, we are really out */
4192 mutex_lock(&cache_chain_mutex);
4195 *(unsigned long *)m->private = n[0] * 2;
4197 mutex_lock(&cache_chain_mutex);
4198 /* Now make sure this entry will be retried */
4202 for (i = 0; i < n[1]; i++) {
4203 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4204 show_symbol(m, n[2*i+2]);
4210 struct seq_operations slabstats_op = {
4211 .start = leaks_start,
4220 * ksize - get the actual amount of memory allocated for a given object
4221 * @objp: Pointer to the object
4223 * kmalloc may internally round up allocations and return more memory
4224 * than requested. ksize() can be used to determine the actual amount of
4225 * memory allocated. The caller may use this additional memory, even though
4226 * a smaller amount of memory was initially specified with the kmalloc call.
4227 * The caller must guarantee that objp points to a valid object previously
4228 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4229 * must not be freed during the duration of the call.
4231 unsigned int ksize(const void *objp)
4233 if (unlikely(objp == NULL))
4236 return obj_size(virt_to_cache(objp));