3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/rtmutex.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 int enable_cpucache(struct kmem_cache
*cachep
);
316 static void cache_reap(struct work_struct
*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)
734 * 1. Guard access to the cache-chain.
735 * 2. Protect sanity of cpu_online_map against cpu hotplug events
737 static DEFINE_MUTEX(cache_chain_mutex
);
738 static struct list_head cache_chain
;
741 * chicken and egg problem: delay the per-cpu array allocation
742 * until the general caches are up.
752 * used by boot code to determine if it can use slab based allocator
754 int slab_is_available(void)
756 return g_cpucache_up
== FULL
;
759 static DEFINE_PER_CPU(struct delayed_work
, reap_work
);
761 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
763 return cachep
->array
[smp_processor_id()];
766 static inline struct kmem_cache
*__find_general_cachep(size_t size
,
769 struct cache_sizes
*csizep
= malloc_sizes
;
772 /* This happens if someone tries to call
773 * kmem_cache_create(), or __kmalloc(), before
774 * the generic caches are initialized.
776 BUG_ON(malloc_sizes
[INDEX_AC
].cs_cachep
== NULL
);
778 while (size
> csizep
->cs_size
)
782 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
783 * has cs_{dma,}cachep==NULL. Thus no special case
784 * for large kmalloc calls required.
786 if (unlikely(gfpflags
& GFP_DMA
))
787 return csizep
->cs_dmacachep
;
788 return csizep
->cs_cachep
;
791 static struct kmem_cache
*kmem_find_general_cachep(size_t size
, gfp_t gfpflags
)
793 return __find_general_cachep(size
, gfpflags
);
796 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
798 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
802 * Calculate the number of objects and left-over bytes for a given buffer size.
804 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
805 size_t align
, int flags
, size_t *left_over
,
810 size_t slab_size
= PAGE_SIZE
<< gfporder
;
813 * The slab management structure can be either off the slab or
814 * on it. For the latter case, the memory allocated for a
818 * - One kmem_bufctl_t for each object
819 * - Padding to respect alignment of @align
820 * - @buffer_size bytes for each object
822 * If the slab management structure is off the slab, then the
823 * alignment will already be calculated into the size. Because
824 * the slabs are all pages aligned, the objects will be at the
825 * correct alignment when allocated.
827 if (flags
& CFLGS_OFF_SLAB
) {
829 nr_objs
= slab_size
/ buffer_size
;
831 if (nr_objs
> SLAB_LIMIT
)
832 nr_objs
= SLAB_LIMIT
;
835 * Ignore padding for the initial guess. The padding
836 * is at most @align-1 bytes, and @buffer_size is at
837 * least @align. In the worst case, this result will
838 * be one greater than the number of objects that fit
839 * into the memory allocation when taking the padding
842 nr_objs
= (slab_size
- sizeof(struct slab
)) /
843 (buffer_size
+ sizeof(kmem_bufctl_t
));
846 * This calculated number will be either the right
847 * amount, or one greater than what we want.
849 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
853 if (nr_objs
> SLAB_LIMIT
)
854 nr_objs
= SLAB_LIMIT
;
856 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
859 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
862 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
864 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
867 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
868 function
, cachep
->name
, msg
);
873 * By default on NUMA we use alien caches to stage the freeing of
874 * objects allocated from other nodes. This causes massive memory
875 * inefficiencies when using fake NUMA setup to split memory into a
876 * large number of small nodes, so it can be disabled on the command
880 static int use_alien_caches __read_mostly
= 1;
881 static int __init
noaliencache_setup(char *s
)
883 use_alien_caches
= 0;
886 __setup("noaliencache", noaliencache_setup
);
890 * Special reaping functions for NUMA systems called from cache_reap().
891 * These take care of doing round robin flushing of alien caches (containing
892 * objects freed on different nodes from which they were allocated) and the
893 * flushing of remote pcps by calling drain_node_pages.
895 static DEFINE_PER_CPU(unsigned long, reap_node
);
897 static void init_reap_node(int cpu
)
901 node
= next_node(cpu_to_node(cpu
), node_online_map
);
902 if (node
== MAX_NUMNODES
)
903 node
= first_node(node_online_map
);
905 per_cpu(reap_node
, cpu
) = node
;
908 static void next_reap_node(void)
910 int node
= __get_cpu_var(reap_node
);
913 * Also drain per cpu pages on remote zones
915 if (node
!= numa_node_id())
916 drain_node_pages(node
);
918 node
= next_node(node
, node_online_map
);
919 if (unlikely(node
>= MAX_NUMNODES
))
920 node
= first_node(node_online_map
);
921 __get_cpu_var(reap_node
) = node
;
925 #define init_reap_node(cpu) do { } while (0)
926 #define next_reap_node(void) do { } while (0)
930 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
931 * via the workqueue/eventd.
932 * Add the CPU number into the expiration time to minimize the possibility of
933 * the CPUs getting into lockstep and contending for the global cache chain
936 static void __devinit
start_cpu_timer(int cpu
)
938 struct delayed_work
*reap_work
= &per_cpu(reap_work
, cpu
);
941 * When this gets called from do_initcalls via cpucache_init(),
942 * init_workqueues() has already run, so keventd will be setup
945 if (keventd_up() && reap_work
->work
.func
== NULL
) {
947 INIT_DELAYED_WORK(reap_work
, cache_reap
);
948 schedule_delayed_work_on(cpu
, reap_work
, HZ
+ 3 * cpu
);
952 static struct array_cache
*alloc_arraycache(int node
, int entries
,
955 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
956 struct array_cache
*nc
= NULL
;
958 nc
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
962 nc
->batchcount
= batchcount
;
964 spin_lock_init(&nc
->lock
);
970 * Transfer objects in one arraycache to another.
971 * Locking must be handled by the caller.
973 * Return the number of entries transferred.
975 static int transfer_objects(struct array_cache
*to
,
976 struct array_cache
*from
, unsigned int max
)
978 /* Figure out how many entries to transfer */
979 int nr
= min(min(from
->avail
, max
), to
->limit
- to
->avail
);
984 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
995 #define drain_alien_cache(cachep, alien) do { } while (0)
996 #define reap_alien(cachep, l3) do { } while (0)
998 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
)
1000 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
1003 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
1007 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1012 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
1018 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1019 gfp_t flags
, int nodeid
)
1024 #else /* CONFIG_NUMA */
1026 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1027 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1029 static struct array_cache
**alloc_alien_cache(int node
, int limit
)
1031 struct array_cache
**ac_ptr
;
1032 int memsize
= sizeof(void *) * MAX_NUMNODES
;
1037 ac_ptr
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1040 if (i
== node
|| !node_online(i
)) {
1044 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d);
1046 for (i
--; i
<= 0; i
--)
1056 static void free_alien_cache(struct array_cache
**ac_ptr
)
1067 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1068 struct array_cache
*ac
, int node
)
1070 struct kmem_list3
*rl3
= cachep
->nodelists
[node
];
1073 spin_lock(&rl3
->list_lock
);
1075 * Stuff objects into the remote nodes shared array first.
1076 * That way we could avoid the overhead of putting the objects
1077 * into the free lists and getting them back later.
1080 transfer_objects(rl3
->shared
, ac
, ac
->limit
);
1082 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1084 spin_unlock(&rl3
->list_lock
);
1089 * Called from cache_reap() to regularly drain alien caches round robin.
1091 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_list3
*l3
)
1093 int node
= __get_cpu_var(reap_node
);
1096 struct array_cache
*ac
= l3
->alien
[node
];
1098 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1099 __drain_alien_cache(cachep
, ac
, node
);
1100 spin_unlock_irq(&ac
->lock
);
1105 static void drain_alien_cache(struct kmem_cache
*cachep
,
1106 struct array_cache
**alien
)
1109 struct array_cache
*ac
;
1110 unsigned long flags
;
1112 for_each_online_node(i
) {
1115 spin_lock_irqsave(&ac
->lock
, flags
);
1116 __drain_alien_cache(cachep
, ac
, i
);
1117 spin_unlock_irqrestore(&ac
->lock
, flags
);
1122 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1124 struct slab
*slabp
= virt_to_slab(objp
);
1125 int nodeid
= slabp
->nodeid
;
1126 struct kmem_list3
*l3
;
1127 struct array_cache
*alien
= NULL
;
1130 node
= numa_node_id();
1133 * Make sure we are not freeing a object from another node to the array
1134 * cache on this cpu.
1136 if (likely(slabp
->nodeid
== node
) || unlikely(!use_alien_caches
))
1139 l3
= cachep
->nodelists
[node
];
1140 STATS_INC_NODEFREES(cachep
);
1141 if (l3
->alien
&& l3
->alien
[nodeid
]) {
1142 alien
= l3
->alien
[nodeid
];
1143 spin_lock(&alien
->lock
);
1144 if (unlikely(alien
->avail
== alien
->limit
)) {
1145 STATS_INC_ACOVERFLOW(cachep
);
1146 __drain_alien_cache(cachep
, alien
, nodeid
);
1148 alien
->entry
[alien
->avail
++] = objp
;
1149 spin_unlock(&alien
->lock
);
1151 spin_lock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1152 free_block(cachep
, &objp
, 1, nodeid
);
1153 spin_unlock(&(cachep
->nodelists
[nodeid
])->list_lock
);
1159 static int __cpuinit
cpuup_callback(struct notifier_block
*nfb
,
1160 unsigned long action
, void *hcpu
)
1162 long cpu
= (long)hcpu
;
1163 struct kmem_cache
*cachep
;
1164 struct kmem_list3
*l3
= NULL
;
1165 int node
= cpu_to_node(cpu
);
1166 int memsize
= sizeof(struct kmem_list3
);
1169 case CPU_UP_PREPARE
:
1170 mutex_lock(&cache_chain_mutex
);
1172 * We need to do this right in the beginning since
1173 * alloc_arraycache's are going to use this list.
1174 * kmalloc_node allows us to add the slab to the right
1175 * kmem_list3 and not this cpu's kmem_list3
1178 list_for_each_entry(cachep
, &cache_chain
, next
) {
1180 * Set up the size64 kmemlist for cpu before we can
1181 * begin anything. Make sure some other cpu on this
1182 * node has not already allocated this
1184 if (!cachep
->nodelists
[node
]) {
1185 l3
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1188 kmem_list3_init(l3
);
1189 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1190 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1193 * The l3s don't come and go as CPUs come and
1194 * go. cache_chain_mutex is sufficient
1197 cachep
->nodelists
[node
] = l3
;
1200 spin_lock_irq(&cachep
->nodelists
[node
]->list_lock
);
1201 cachep
->nodelists
[node
]->free_limit
=
1202 (1 + nr_cpus_node(node
)) *
1203 cachep
->batchcount
+ cachep
->num
;
1204 spin_unlock_irq(&cachep
->nodelists
[node
]->list_lock
);
1208 * Now we can go ahead with allocating the shared arrays and
1211 list_for_each_entry(cachep
, &cache_chain
, next
) {
1212 struct array_cache
*nc
;
1213 struct array_cache
*shared
;
1214 struct array_cache
**alien
= NULL
;
1216 nc
= alloc_arraycache(node
, cachep
->limit
,
1217 cachep
->batchcount
);
1220 shared
= alloc_arraycache(node
,
1221 cachep
->shared
* cachep
->batchcount
,
1226 if (use_alien_caches
) {
1227 alien
= alloc_alien_cache(node
, cachep
->limit
);
1231 cachep
->array
[cpu
] = nc
;
1232 l3
= cachep
->nodelists
[node
];
1235 spin_lock_irq(&l3
->list_lock
);
1238 * We are serialised from CPU_DEAD or
1239 * CPU_UP_CANCELLED by the cpucontrol lock
1241 l3
->shared
= shared
;
1250 spin_unlock_irq(&l3
->list_lock
);
1252 free_alien_cache(alien
);
1256 mutex_unlock(&cache_chain_mutex
);
1257 start_cpu_timer(cpu
);
1259 #ifdef CONFIG_HOTPLUG_CPU
1260 case CPU_DOWN_PREPARE
:
1261 mutex_lock(&cache_chain_mutex
);
1263 case CPU_DOWN_FAILED
:
1264 mutex_unlock(&cache_chain_mutex
);
1268 * Even if all the cpus of a node are down, we don't free the
1269 * kmem_list3 of any cache. This to avoid a race between
1270 * cpu_down, and a kmalloc allocation from another cpu for
1271 * memory from the node of the cpu going down. The list3
1272 * structure is usually allocated from kmem_cache_create() and
1273 * gets destroyed at kmem_cache_destroy().
1277 case CPU_UP_CANCELED
:
1278 list_for_each_entry(cachep
, &cache_chain
, next
) {
1279 struct array_cache
*nc
;
1280 struct array_cache
*shared
;
1281 struct array_cache
**alien
;
1284 mask
= node_to_cpumask(node
);
1285 /* cpu is dead; no one can alloc from it. */
1286 nc
= cachep
->array
[cpu
];
1287 cachep
->array
[cpu
] = NULL
;
1288 l3
= cachep
->nodelists
[node
];
1291 goto free_array_cache
;
1293 spin_lock_irq(&l3
->list_lock
);
1295 /* Free limit for this kmem_list3 */
1296 l3
->free_limit
-= cachep
->batchcount
;
1298 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1300 if (!cpus_empty(mask
)) {
1301 spin_unlock_irq(&l3
->list_lock
);
1302 goto free_array_cache
;
1305 shared
= l3
->shared
;
1307 free_block(cachep
, l3
->shared
->entry
,
1308 l3
->shared
->avail
, node
);
1315 spin_unlock_irq(&l3
->list_lock
);
1319 drain_alien_cache(cachep
, alien
);
1320 free_alien_cache(alien
);
1326 * In the previous loop, all the objects were freed to
1327 * the respective cache's slabs, now we can go ahead and
1328 * shrink each nodelist to its limit.
1330 list_for_each_entry(cachep
, &cache_chain
, next
) {
1331 l3
= cachep
->nodelists
[node
];
1334 drain_freelist(cachep
, l3
, l3
->free_objects
);
1336 mutex_unlock(&cache_chain_mutex
);
1344 static struct notifier_block __cpuinitdata cpucache_notifier
= {
1345 &cpuup_callback
, NULL
, 0
1349 * swap the static kmem_list3 with kmalloced memory
1351 static void init_list(struct kmem_cache
*cachep
, struct kmem_list3
*list
,
1354 struct kmem_list3
*ptr
;
1356 ptr
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, nodeid
);
1359 local_irq_disable();
1360 memcpy(ptr
, list
, sizeof(struct kmem_list3
));
1362 * Do not assume that spinlocks can be initialized via memcpy:
1364 spin_lock_init(&ptr
->list_lock
);
1366 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1367 cachep
->nodelists
[nodeid
] = ptr
;
1372 * Initialisation. Called after the page allocator have been initialised and
1373 * before smp_init().
1375 void __init
kmem_cache_init(void)
1378 struct cache_sizes
*sizes
;
1379 struct cache_names
*names
;
1384 for (i
= 0; i
< NUM_INIT_LISTS
; i
++) {
1385 kmem_list3_init(&initkmem_list3
[i
]);
1386 if (i
< MAX_NUMNODES
)
1387 cache_cache
.nodelists
[i
] = NULL
;
1391 * Fragmentation resistance on low memory - only use bigger
1392 * page orders on machines with more than 32MB of memory.
1394 if (num_physpages
> (32 << 20) >> PAGE_SHIFT
)
1395 slab_break_gfp_order
= BREAK_GFP_ORDER_HI
;
1397 /* Bootstrap is tricky, because several objects are allocated
1398 * from caches that do not exist yet:
1399 * 1) initialize the cache_cache cache: it contains the struct
1400 * kmem_cache structures of all caches, except cache_cache itself:
1401 * cache_cache is statically allocated.
1402 * Initially an __init data area is used for the head array and the
1403 * kmem_list3 structures, it's replaced with a kmalloc allocated
1404 * array at the end of the bootstrap.
1405 * 2) Create the first kmalloc cache.
1406 * The struct kmem_cache for the new cache is allocated normally.
1407 * An __init data area is used for the head array.
1408 * 3) Create the remaining kmalloc caches, with minimally sized
1410 * 4) Replace the __init data head arrays for cache_cache and the first
1411 * kmalloc cache with kmalloc allocated arrays.
1412 * 5) Replace the __init data for kmem_list3 for cache_cache and
1413 * the other cache's with kmalloc allocated memory.
1414 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1417 node
= numa_node_id();
1419 /* 1) create the cache_cache */
1420 INIT_LIST_HEAD(&cache_chain
);
1421 list_add(&cache_cache
.next
, &cache_chain
);
1422 cache_cache
.colour_off
= cache_line_size();
1423 cache_cache
.array
[smp_processor_id()] = &initarray_cache
.cache
;
1424 cache_cache
.nodelists
[node
] = &initkmem_list3
[CACHE_CACHE
];
1426 cache_cache
.buffer_size
= ALIGN(cache_cache
.buffer_size
,
1429 for (order
= 0; order
< MAX_ORDER
; order
++) {
1430 cache_estimate(order
, cache_cache
.buffer_size
,
1431 cache_line_size(), 0, &left_over
, &cache_cache
.num
);
1432 if (cache_cache
.num
)
1435 BUG_ON(!cache_cache
.num
);
1436 cache_cache
.gfporder
= order
;
1437 cache_cache
.colour
= left_over
/ cache_cache
.colour_off
;
1438 cache_cache
.slab_size
= ALIGN(cache_cache
.num
* sizeof(kmem_bufctl_t
) +
1439 sizeof(struct slab
), cache_line_size());
1441 /* 2+3) create the kmalloc caches */
1442 sizes
= malloc_sizes
;
1443 names
= cache_names
;
1446 * Initialize the caches that provide memory for the array cache and the
1447 * kmem_list3 structures first. Without this, further allocations will
1451 sizes
[INDEX_AC
].cs_cachep
= kmem_cache_create(names
[INDEX_AC
].name
,
1452 sizes
[INDEX_AC
].cs_size
,
1453 ARCH_KMALLOC_MINALIGN
,
1454 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1457 if (INDEX_AC
!= INDEX_L3
) {
1458 sizes
[INDEX_L3
].cs_cachep
=
1459 kmem_cache_create(names
[INDEX_L3
].name
,
1460 sizes
[INDEX_L3
].cs_size
,
1461 ARCH_KMALLOC_MINALIGN
,
1462 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1466 slab_early_init
= 0;
1468 while (sizes
->cs_size
!= ULONG_MAX
) {
1470 * For performance, all the general caches are L1 aligned.
1471 * This should be particularly beneficial on SMP boxes, as it
1472 * eliminates "false sharing".
1473 * Note for systems short on memory removing the alignment will
1474 * allow tighter packing of the smaller caches.
1476 if (!sizes
->cs_cachep
) {
1477 sizes
->cs_cachep
= kmem_cache_create(names
->name
,
1479 ARCH_KMALLOC_MINALIGN
,
1480 ARCH_KMALLOC_FLAGS
|SLAB_PANIC
,
1484 sizes
->cs_dmacachep
= kmem_cache_create(names
->name_dma
,
1486 ARCH_KMALLOC_MINALIGN
,
1487 ARCH_KMALLOC_FLAGS
|SLAB_CACHE_DMA
|
1493 /* 4) Replace the bootstrap head arrays */
1495 struct array_cache
*ptr
;
1497 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1499 local_irq_disable();
1500 BUG_ON(cpu_cache_get(&cache_cache
) != &initarray_cache
.cache
);
1501 memcpy(ptr
, cpu_cache_get(&cache_cache
),
1502 sizeof(struct arraycache_init
));
1504 * Do not assume that spinlocks can be initialized via memcpy:
1506 spin_lock_init(&ptr
->lock
);
1508 cache_cache
.array
[smp_processor_id()] = ptr
;
1511 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
1513 local_irq_disable();
1514 BUG_ON(cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
)
1515 != &initarray_generic
.cache
);
1516 memcpy(ptr
, cpu_cache_get(malloc_sizes
[INDEX_AC
].cs_cachep
),
1517 sizeof(struct arraycache_init
));
1519 * Do not assume that spinlocks can be initialized via memcpy:
1521 spin_lock_init(&ptr
->lock
);
1523 malloc_sizes
[INDEX_AC
].cs_cachep
->array
[smp_processor_id()] =
1527 /* 5) Replace the bootstrap kmem_list3's */
1531 /* Replace the static kmem_list3 structures for the boot cpu */
1532 init_list(&cache_cache
, &initkmem_list3
[CACHE_CACHE
], node
);
1534 for_each_online_node(nid
) {
1535 init_list(malloc_sizes
[INDEX_AC
].cs_cachep
,
1536 &initkmem_list3
[SIZE_AC
+ nid
], nid
);
1538 if (INDEX_AC
!= INDEX_L3
) {
1539 init_list(malloc_sizes
[INDEX_L3
].cs_cachep
,
1540 &initkmem_list3
[SIZE_L3
+ nid
], nid
);
1545 /* 6) resize the head arrays to their final sizes */
1547 struct kmem_cache
*cachep
;
1548 mutex_lock(&cache_chain_mutex
);
1549 list_for_each_entry(cachep
, &cache_chain
, next
)
1550 if (enable_cpucache(cachep
))
1552 mutex_unlock(&cache_chain_mutex
);
1555 /* Annotate slab for lockdep -- annotate the malloc caches */
1560 g_cpucache_up
= FULL
;
1563 * Register a cpu startup notifier callback that initializes
1564 * cpu_cache_get for all new cpus
1566 register_cpu_notifier(&cpucache_notifier
);
1569 * The reap timers are started later, with a module init call: That part
1570 * of the kernel is not yet operational.
1574 static int __init
cpucache_init(void)
1579 * Register the timers that return unneeded pages to the page allocator
1581 for_each_online_cpu(cpu
)
1582 start_cpu_timer(cpu
);
1585 __initcall(cpucache_init
);
1588 * Interface to system's page allocator. No need to hold the cache-lock.
1590 * If we requested dmaable memory, we will get it. Even if we
1591 * did not request dmaable memory, we might get it, but that
1592 * would be relatively rare and ignorable.
1594 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1602 * Nommu uses slab's for process anonymous memory allocations, and thus
1603 * requires __GFP_COMP to properly refcount higher order allocations
1605 flags
|= __GFP_COMP
;
1608 flags
|= cachep
->gfpflags
;
1610 page
= alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1614 nr_pages
= (1 << cachep
->gfporder
);
1615 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1616 add_zone_page_state(page_zone(page
),
1617 NR_SLAB_RECLAIMABLE
, nr_pages
);
1619 add_zone_page_state(page_zone(page
),
1620 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1621 for (i
= 0; i
< nr_pages
; i
++)
1622 __SetPageSlab(page
+ i
);
1623 return page_address(page
);
1627 * Interface to system's page release.
1629 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1631 unsigned long i
= (1 << cachep
->gfporder
);
1632 struct page
*page
= virt_to_page(addr
);
1633 const unsigned long nr_freed
= i
;
1635 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1636 sub_zone_page_state(page_zone(page
),
1637 NR_SLAB_RECLAIMABLE
, nr_freed
);
1639 sub_zone_page_state(page_zone(page
),
1640 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1642 BUG_ON(!PageSlab(page
));
1643 __ClearPageSlab(page
);
1646 if (current
->reclaim_state
)
1647 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1648 free_pages((unsigned long)addr
, cachep
->gfporder
);
1651 static void kmem_rcu_free(struct rcu_head
*head
)
1653 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1654 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1656 kmem_freepages(cachep
, slab_rcu
->addr
);
1657 if (OFF_SLAB(cachep
))
1658 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1663 #ifdef CONFIG_DEBUG_PAGEALLOC
1664 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1665 unsigned long caller
)
1667 int size
= obj_size(cachep
);
1669 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1671 if (size
< 5 * sizeof(unsigned long))
1674 *addr
++ = 0x12345678;
1676 *addr
++ = smp_processor_id();
1677 size
-= 3 * sizeof(unsigned long);
1679 unsigned long *sptr
= &caller
;
1680 unsigned long svalue
;
1682 while (!kstack_end(sptr
)) {
1684 if (kernel_text_address(svalue
)) {
1686 size
-= sizeof(unsigned long);
1687 if (size
<= sizeof(unsigned long))
1693 *addr
++ = 0x87654321;
1697 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1699 int size
= obj_size(cachep
);
1700 addr
= &((char *)addr
)[obj_offset(cachep
)];
1702 memset(addr
, val
, size
);
1703 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1706 static void dump_line(char *data
, int offset
, int limit
)
1709 unsigned char error
= 0;
1712 printk(KERN_ERR
"%03x:", offset
);
1713 for (i
= 0; i
< limit
; i
++) {
1714 if (data
[offset
+ i
] != POISON_FREE
) {
1715 error
= data
[offset
+ i
];
1718 printk(" %02x", (unsigned char)data
[offset
+ i
]);
1722 if (bad_count
== 1) {
1723 error
^= POISON_FREE
;
1724 if (!(error
& (error
- 1))) {
1725 printk(KERN_ERR
"Single bit error detected. Probably "
1728 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1731 printk(KERN_ERR
"Run a memory test tool.\n");
1740 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1745 if (cachep
->flags
& SLAB_RED_ZONE
) {
1746 printk(KERN_ERR
"Redzone: 0x%lx/0x%lx.\n",
1747 *dbg_redzone1(cachep
, objp
),
1748 *dbg_redzone2(cachep
, objp
));
1751 if (cachep
->flags
& SLAB_STORE_USER
) {
1752 printk(KERN_ERR
"Last user: [<%p>]",
1753 *dbg_userword(cachep
, objp
));
1754 print_symbol("(%s)",
1755 (unsigned long)*dbg_userword(cachep
, objp
));
1758 realobj
= (char *)objp
+ obj_offset(cachep
);
1759 size
= obj_size(cachep
);
1760 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1763 if (i
+ limit
> size
)
1765 dump_line(realobj
, i
, limit
);
1769 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1775 realobj
= (char *)objp
+ obj_offset(cachep
);
1776 size
= obj_size(cachep
);
1778 for (i
= 0; i
< size
; i
++) {
1779 char exp
= POISON_FREE
;
1782 if (realobj
[i
] != exp
) {
1788 "Slab corruption: start=%p, len=%d\n",
1790 print_objinfo(cachep
, objp
, 0);
1792 /* Hexdump the affected line */
1795 if (i
+ limit
> size
)
1797 dump_line(realobj
, i
, limit
);
1800 /* Limit to 5 lines */
1806 /* Print some data about the neighboring objects, if they
1809 struct slab
*slabp
= virt_to_slab(objp
);
1812 objnr
= obj_to_index(cachep
, slabp
, objp
);
1814 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1815 realobj
= (char *)objp
+ obj_offset(cachep
);
1816 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1818 print_objinfo(cachep
, objp
, 2);
1820 if (objnr
+ 1 < cachep
->num
) {
1821 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1822 realobj
= (char *)objp
+ obj_offset(cachep
);
1823 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1825 print_objinfo(cachep
, objp
, 2);
1833 * slab_destroy_objs - destroy a slab and its objects
1834 * @cachep: cache pointer being destroyed
1835 * @slabp: slab pointer being destroyed
1837 * Call the registered destructor for each object in a slab that is being
1840 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1843 for (i
= 0; i
< cachep
->num
; i
++) {
1844 void *objp
= index_to_obj(cachep
, slabp
, i
);
1846 if (cachep
->flags
& SLAB_POISON
) {
1847 #ifdef CONFIG_DEBUG_PAGEALLOC
1848 if (cachep
->buffer_size
% PAGE_SIZE
== 0 &&
1850 kernel_map_pages(virt_to_page(objp
),
1851 cachep
->buffer_size
/ PAGE_SIZE
, 1);
1853 check_poison_obj(cachep
, objp
);
1855 check_poison_obj(cachep
, objp
);
1858 if (cachep
->flags
& SLAB_RED_ZONE
) {
1859 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1860 slab_error(cachep
, "start of a freed object "
1862 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1863 slab_error(cachep
, "end of a freed object "
1866 if (cachep
->dtor
&& !(cachep
->flags
& SLAB_POISON
))
1867 (cachep
->dtor
) (objp
+ obj_offset(cachep
), cachep
, 0);
1871 static void slab_destroy_objs(struct kmem_cache
*cachep
, struct slab
*slabp
)
1875 for (i
= 0; i
< cachep
->num
; i
++) {
1876 void *objp
= index_to_obj(cachep
, slabp
, i
);
1877 (cachep
->dtor
) (objp
, cachep
, 0);
1884 * slab_destroy - destroy and release all objects in a slab
1885 * @cachep: cache pointer being destroyed
1886 * @slabp: slab pointer being destroyed
1888 * Destroy all the objs in a slab, and release the mem back to the system.
1889 * Before calling the slab must have been unlinked from the cache. The
1890 * cache-lock is not held/needed.
1892 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1894 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
1896 slab_destroy_objs(cachep
, slabp
);
1897 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1898 struct slab_rcu
*slab_rcu
;
1900 slab_rcu
= (struct slab_rcu
*)slabp
;
1901 slab_rcu
->cachep
= cachep
;
1902 slab_rcu
->addr
= addr
;
1903 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
1905 kmem_freepages(cachep
, addr
);
1906 if (OFF_SLAB(cachep
))
1907 kmem_cache_free(cachep
->slabp_cache
, slabp
);
1912 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1913 * size of kmem_list3.
1915 static void set_up_list3s(struct kmem_cache
*cachep
, int index
)
1919 for_each_online_node(node
) {
1920 cachep
->nodelists
[node
] = &initkmem_list3
[index
+ node
];
1921 cachep
->nodelists
[node
]->next_reap
= jiffies
+
1923 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1927 static void __kmem_cache_destroy(struct kmem_cache
*cachep
)
1930 struct kmem_list3
*l3
;
1932 for_each_online_cpu(i
)
1933 kfree(cachep
->array
[i
]);
1935 /* NUMA: free the list3 structures */
1936 for_each_online_node(i
) {
1937 l3
= cachep
->nodelists
[i
];
1940 free_alien_cache(l3
->alien
);
1944 kmem_cache_free(&cache_cache
, cachep
);
1949 * calculate_slab_order - calculate size (page order) of slabs
1950 * @cachep: pointer to the cache that is being created
1951 * @size: size of objects to be created in this cache.
1952 * @align: required alignment for the objects.
1953 * @flags: slab allocation flags
1955 * Also calculates the number of objects per slab.
1957 * This could be made much more intelligent. For now, try to avoid using
1958 * high order pages for slabs. When the gfp() functions are more friendly
1959 * towards high-order requests, this should be changed.
1961 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1962 size_t size
, size_t align
, unsigned long flags
)
1964 unsigned long offslab_limit
;
1965 size_t left_over
= 0;
1968 for (gfporder
= 0; gfporder
<= MAX_GFP_ORDER
; gfporder
++) {
1972 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1976 if (flags
& CFLGS_OFF_SLAB
) {
1978 * Max number of objs-per-slab for caches which
1979 * use off-slab slabs. Needed to avoid a possible
1980 * looping condition in cache_grow().
1982 offslab_limit
= size
- sizeof(struct slab
);
1983 offslab_limit
/= sizeof(kmem_bufctl_t
);
1985 if (num
> offslab_limit
)
1989 /* Found something acceptable - save it away */
1991 cachep
->gfporder
= gfporder
;
1992 left_over
= remainder
;
1995 * A VFS-reclaimable slab tends to have most allocations
1996 * as GFP_NOFS and we really don't want to have to be allocating
1997 * higher-order pages when we are unable to shrink dcache.
1999 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2003 * Large number of objects is good, but very large slabs are
2004 * currently bad for the gfp()s.
2006 if (gfporder
>= slab_break_gfp_order
)
2010 * Acceptable internal fragmentation?
2012 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2018 static int setup_cpu_cache(struct kmem_cache
*cachep
)
2020 if (g_cpucache_up
== FULL
)
2021 return enable_cpucache(cachep
);
2023 if (g_cpucache_up
== NONE
) {
2025 * Note: the first kmem_cache_create must create the cache
2026 * that's used by kmalloc(24), otherwise the creation of
2027 * further caches will BUG().
2029 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2032 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2033 * the first cache, then we need to set up all its list3s,
2034 * otherwise the creation of further caches will BUG().
2036 set_up_list3s(cachep
, SIZE_AC
);
2037 if (INDEX_AC
== INDEX_L3
)
2038 g_cpucache_up
= PARTIAL_L3
;
2040 g_cpucache_up
= PARTIAL_AC
;
2042 cachep
->array
[smp_processor_id()] =
2043 kmalloc(sizeof(struct arraycache_init
), GFP_KERNEL
);
2045 if (g_cpucache_up
== PARTIAL_AC
) {
2046 set_up_list3s(cachep
, SIZE_L3
);
2047 g_cpucache_up
= PARTIAL_L3
;
2050 for_each_online_node(node
) {
2051 cachep
->nodelists
[node
] =
2052 kmalloc_node(sizeof(struct kmem_list3
),
2054 BUG_ON(!cachep
->nodelists
[node
]);
2055 kmem_list3_init(cachep
->nodelists
[node
]);
2059 cachep
->nodelists
[numa_node_id()]->next_reap
=
2060 jiffies
+ REAPTIMEOUT_LIST3
+
2061 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2063 cpu_cache_get(cachep
)->avail
= 0;
2064 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2065 cpu_cache_get(cachep
)->batchcount
= 1;
2066 cpu_cache_get(cachep
)->touched
= 0;
2067 cachep
->batchcount
= 1;
2068 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2073 * kmem_cache_create - Create a cache.
2074 * @name: A string which is used in /proc/slabinfo to identify this cache.
2075 * @size: The size of objects to be created in this cache.
2076 * @align: The required alignment for the objects.
2077 * @flags: SLAB flags
2078 * @ctor: A constructor for the objects.
2079 * @dtor: A destructor for the objects.
2081 * Returns a ptr to the cache on success, NULL on failure.
2082 * Cannot be called within a int, but can be interrupted.
2083 * The @ctor is run when new pages are allocated by the cache
2084 * and the @dtor is run before the pages are handed back.
2086 * @name must be valid until the cache is destroyed. This implies that
2087 * the module calling this has to destroy the cache before getting unloaded.
2091 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2092 * to catch references to uninitialised memory.
2094 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2095 * for buffer overruns.
2097 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2098 * cacheline. This can be beneficial if you're counting cycles as closely
2102 kmem_cache_create (const char *name
, size_t size
, size_t align
,
2103 unsigned long flags
,
2104 void (*ctor
)(void*, struct kmem_cache
*, unsigned long),
2105 void (*dtor
)(void*, struct kmem_cache
*, unsigned long))
2107 size_t left_over
, slab_size
, ralign
;
2108 struct kmem_cache
*cachep
= NULL
, *pc
;
2111 * Sanity checks... these are all serious usage bugs.
2113 if (!name
|| in_interrupt() || (size
< BYTES_PER_WORD
) ||
2114 (size
> (1 << MAX_OBJ_ORDER
) * PAGE_SIZE
) || (dtor
&& !ctor
)) {
2115 printk(KERN_ERR
"%s: Early error in slab %s\n", __FUNCTION__
,
2121 * We use cache_chain_mutex to ensure a consistent view of
2122 * cpu_online_map as well. Please see cpuup_callback
2124 mutex_lock(&cache_chain_mutex
);
2126 list_for_each_entry(pc
, &cache_chain
, next
) {
2131 * This happens when the module gets unloaded and doesn't
2132 * destroy its slab cache and no-one else reuses the vmalloc
2133 * area of the module. Print a warning.
2135 res
= probe_kernel_address(pc
->name
, tmp
);
2137 printk("SLAB: cache with size %d has lost its name\n",
2142 if (!strcmp(pc
->name
, name
)) {
2143 printk("kmem_cache_create: duplicate cache %s\n", name
);
2150 WARN_ON(strchr(name
, ' ')); /* It confuses parsers */
2151 if ((flags
& SLAB_DEBUG_INITIAL
) && !ctor
) {
2152 /* No constructor, but inital state check requested */
2153 printk(KERN_ERR
"%s: No con, but init state check "
2154 "requested - %s\n", __FUNCTION__
, name
);
2155 flags
&= ~SLAB_DEBUG_INITIAL
;
2159 * Enable redzoning and last user accounting, except for caches with
2160 * large objects, if the increased size would increase the object size
2161 * above the next power of two: caches with object sizes just above a
2162 * power of two have a significant amount of internal fragmentation.
2164 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + 3 * BYTES_PER_WORD
))
2165 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2166 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2167 flags
|= SLAB_POISON
;
2169 if (flags
& SLAB_DESTROY_BY_RCU
)
2170 BUG_ON(flags
& SLAB_POISON
);
2172 if (flags
& SLAB_DESTROY_BY_RCU
)
2176 * Always checks flags, a caller might be expecting debug support which
2179 BUG_ON(flags
& ~CREATE_MASK
);
2182 * Check that size is in terms of words. This is needed to avoid
2183 * unaligned accesses for some archs when redzoning is used, and makes
2184 * sure any on-slab bufctl's are also correctly aligned.
2186 if (size
& (BYTES_PER_WORD
- 1)) {
2187 size
+= (BYTES_PER_WORD
- 1);
2188 size
&= ~(BYTES_PER_WORD
- 1);
2191 /* calculate the final buffer alignment: */
2193 /* 1) arch recommendation: can be overridden for debug */
2194 if (flags
& SLAB_HWCACHE_ALIGN
) {
2196 * Default alignment: as specified by the arch code. Except if
2197 * an object is really small, then squeeze multiple objects into
2200 ralign
= cache_line_size();
2201 while (size
<= ralign
/ 2)
2204 ralign
= BYTES_PER_WORD
;
2208 * Redzoning and user store require word alignment. Note this will be
2209 * overridden by architecture or caller mandated alignment if either
2210 * is greater than BYTES_PER_WORD.
2212 if (flags
& SLAB_RED_ZONE
|| flags
& SLAB_STORE_USER
)
2213 ralign
= BYTES_PER_WORD
;
2215 /* 2) arch mandated alignment */
2216 if (ralign
< ARCH_SLAB_MINALIGN
) {
2217 ralign
= ARCH_SLAB_MINALIGN
;
2219 /* 3) caller mandated alignment */
2220 if (ralign
< align
) {
2223 /* disable debug if necessary */
2224 if (ralign
> BYTES_PER_WORD
)
2225 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2231 /* Get cache's description obj. */
2232 cachep
= kmem_cache_zalloc(&cache_cache
, GFP_KERNEL
);
2237 cachep
->obj_size
= size
;
2240 * Both debugging options require word-alignment which is calculated
2243 if (flags
& SLAB_RED_ZONE
) {
2244 /* add space for red zone words */
2245 cachep
->obj_offset
+= BYTES_PER_WORD
;
2246 size
+= 2 * BYTES_PER_WORD
;
2248 if (flags
& SLAB_STORE_USER
) {
2249 /* user store requires one word storage behind the end of
2252 size
+= BYTES_PER_WORD
;
2254 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2255 if (size
>= malloc_sizes
[INDEX_L3
+ 1].cs_size
2256 && cachep
->obj_size
> cache_line_size() && size
< PAGE_SIZE
) {
2257 cachep
->obj_offset
+= PAGE_SIZE
- size
;
2264 * Determine if the slab management is 'on' or 'off' slab.
2265 * (bootstrapping cannot cope with offslab caches so don't do
2268 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
)
2270 * Size is large, assume best to place the slab management obj
2271 * off-slab (should allow better packing of objs).
2273 flags
|= CFLGS_OFF_SLAB
;
2275 size
= ALIGN(size
, align
);
2277 left_over
= calculate_slab_order(cachep
, size
, align
, flags
);
2280 printk("kmem_cache_create: couldn't create cache %s.\n", name
);
2281 kmem_cache_free(&cache_cache
, cachep
);
2285 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2286 + sizeof(struct slab
), align
);
2289 * If the slab has been placed off-slab, and we have enough space then
2290 * move it on-slab. This is at the expense of any extra colouring.
2292 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2293 flags
&= ~CFLGS_OFF_SLAB
;
2294 left_over
-= slab_size
;
2297 if (flags
& CFLGS_OFF_SLAB
) {
2298 /* really off slab. No need for manual alignment */
2300 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2303 cachep
->colour_off
= cache_line_size();
2304 /* Offset must be a multiple of the alignment. */
2305 if (cachep
->colour_off
< align
)
2306 cachep
->colour_off
= align
;
2307 cachep
->colour
= left_over
/ cachep
->colour_off
;
2308 cachep
->slab_size
= slab_size
;
2309 cachep
->flags
= flags
;
2310 cachep
->gfpflags
= 0;
2311 if (flags
& SLAB_CACHE_DMA
)
2312 cachep
->gfpflags
|= GFP_DMA
;
2313 cachep
->buffer_size
= size
;
2315 if (flags
& CFLGS_OFF_SLAB
) {
2316 cachep
->slabp_cache
= kmem_find_general_cachep(slab_size
, 0u);
2318 * This is a possibility for one of the malloc_sizes caches.
2319 * But since we go off slab only for object size greater than
2320 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2321 * this should not happen at all.
2322 * But leave a BUG_ON for some lucky dude.
2324 BUG_ON(!cachep
->slabp_cache
);
2326 cachep
->ctor
= ctor
;
2327 cachep
->dtor
= dtor
;
2328 cachep
->name
= name
;
2330 if (setup_cpu_cache(cachep
)) {
2331 __kmem_cache_destroy(cachep
);
2336 /* cache setup completed, link it into the list */
2337 list_add(&cachep
->next
, &cache_chain
);
2339 if (!cachep
&& (flags
& SLAB_PANIC
))
2340 panic("kmem_cache_create(): failed to create slab `%s'\n",
2342 mutex_unlock(&cache_chain_mutex
);
2345 EXPORT_SYMBOL(kmem_cache_create
);
2348 static void check_irq_off(void)
2350 BUG_ON(!irqs_disabled());
2353 static void check_irq_on(void)
2355 BUG_ON(irqs_disabled());
2358 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2362 assert_spin_locked(&cachep
->nodelists
[numa_node_id()]->list_lock
);
2366 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2370 assert_spin_locked(&cachep
->nodelists
[node
]->list_lock
);
2375 #define check_irq_off() do { } while(0)
2376 #define check_irq_on() do { } while(0)
2377 #define check_spinlock_acquired(x) do { } while(0)
2378 #define check_spinlock_acquired_node(x, y) do { } while(0)
2381 static void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
2382 struct array_cache
*ac
,
2383 int force
, int node
);
2385 static void do_drain(void *arg
)
2387 struct kmem_cache
*cachep
= arg
;
2388 struct array_cache
*ac
;
2389 int node
= numa_node_id();
2392 ac
= cpu_cache_get(cachep
);
2393 spin_lock(&cachep
->nodelists
[node
]->list_lock
);
2394 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2395 spin_unlock(&cachep
->nodelists
[node
]->list_lock
);
2399 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2401 struct kmem_list3
*l3
;
2404 on_each_cpu(do_drain
, cachep
, 1, 1);
2406 for_each_online_node(node
) {
2407 l3
= cachep
->nodelists
[node
];
2408 if (l3
&& l3
->alien
)
2409 drain_alien_cache(cachep
, l3
->alien
);
2412 for_each_online_node(node
) {
2413 l3
= cachep
->nodelists
[node
];
2415 drain_array(cachep
, l3
, l3
->shared
, 1, node
);
2420 * Remove slabs from the list of free slabs.
2421 * Specify the number of slabs to drain in tofree.
2423 * Returns the actual number of slabs released.
2425 static int drain_freelist(struct kmem_cache
*cache
,
2426 struct kmem_list3
*l3
, int tofree
)
2428 struct list_head
*p
;
2433 while (nr_freed
< tofree
&& !list_empty(&l3
->slabs_free
)) {
2435 spin_lock_irq(&l3
->list_lock
);
2436 p
= l3
->slabs_free
.prev
;
2437 if (p
== &l3
->slabs_free
) {
2438 spin_unlock_irq(&l3
->list_lock
);
2442 slabp
= list_entry(p
, struct slab
, list
);
2444 BUG_ON(slabp
->inuse
);
2446 list_del(&slabp
->list
);
2448 * Safe to drop the lock. The slab is no longer linked
2451 l3
->free_objects
-= cache
->num
;
2452 spin_unlock_irq(&l3
->list_lock
);
2453 slab_destroy(cache
, slabp
);
2460 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2461 static int __cache_shrink(struct kmem_cache
*cachep
)
2464 struct kmem_list3
*l3
;
2466 drain_cpu_caches(cachep
);
2469 for_each_online_node(i
) {
2470 l3
= cachep
->nodelists
[i
];
2474 drain_freelist(cachep
, l3
, l3
->free_objects
);
2476 ret
+= !list_empty(&l3
->slabs_full
) ||
2477 !list_empty(&l3
->slabs_partial
);
2479 return (ret
? 1 : 0);
2483 * kmem_cache_shrink - Shrink a cache.
2484 * @cachep: The cache to shrink.
2486 * Releases as many slabs as possible for a cache.
2487 * To help debugging, a zero exit status indicates all slabs were released.
2489 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2492 BUG_ON(!cachep
|| in_interrupt());
2494 mutex_lock(&cache_chain_mutex
);
2495 ret
= __cache_shrink(cachep
);
2496 mutex_unlock(&cache_chain_mutex
);
2499 EXPORT_SYMBOL(kmem_cache_shrink
);
2502 * kmem_cache_destroy - delete a cache
2503 * @cachep: the cache to destroy
2505 * Remove a struct kmem_cache object from the slab cache.
2507 * It is expected this function will be called by a module when it is
2508 * unloaded. This will remove the cache completely, and avoid a duplicate
2509 * cache being allocated each time a module is loaded and unloaded, if the
2510 * module doesn't have persistent in-kernel storage across loads and unloads.
2512 * The cache must be empty before calling this function.
2514 * The caller must guarantee that noone will allocate memory from the cache
2515 * during the kmem_cache_destroy().
2517 void kmem_cache_destroy(struct kmem_cache
*cachep
)
2519 BUG_ON(!cachep
|| in_interrupt());
2521 /* Find the cache in the chain of caches. */
2522 mutex_lock(&cache_chain_mutex
);
2524 * the chain is never empty, cache_cache is never destroyed
2526 list_del(&cachep
->next
);
2527 if (__cache_shrink(cachep
)) {
2528 slab_error(cachep
, "Can't free all objects");
2529 list_add(&cachep
->next
, &cache_chain
);
2530 mutex_unlock(&cache_chain_mutex
);
2534 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
2537 __kmem_cache_destroy(cachep
);
2538 mutex_unlock(&cache_chain_mutex
);
2540 EXPORT_SYMBOL(kmem_cache_destroy
);
2543 * Get the memory for a slab management obj.
2544 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2545 * always come from malloc_sizes caches. The slab descriptor cannot
2546 * come from the same cache which is getting created because,
2547 * when we are searching for an appropriate cache for these
2548 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2549 * If we are creating a malloc_sizes cache here it would not be visible to
2550 * kmem_find_general_cachep till the initialization is complete.
2551 * Hence we cannot have slabp_cache same as the original cache.
2553 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2554 int colour_off
, gfp_t local_flags
,
2559 if (OFF_SLAB(cachep
)) {
2560 /* Slab management obj is off-slab. */
2561 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2562 local_flags
& ~GFP_THISNODE
, nodeid
);
2566 slabp
= objp
+ colour_off
;
2567 colour_off
+= cachep
->slab_size
;
2570 slabp
->colouroff
= colour_off
;
2571 slabp
->s_mem
= objp
+ colour_off
;
2572 slabp
->nodeid
= nodeid
;
2576 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2578 return (kmem_bufctl_t
*) (slabp
+ 1);
2581 static void cache_init_objs(struct kmem_cache
*cachep
,
2582 struct slab
*slabp
, unsigned long ctor_flags
)
2586 for (i
= 0; i
< cachep
->num
; i
++) {
2587 void *objp
= index_to_obj(cachep
, slabp
, i
);
2589 /* need to poison the objs? */
2590 if (cachep
->flags
& SLAB_POISON
)
2591 poison_obj(cachep
, objp
, POISON_FREE
);
2592 if (cachep
->flags
& SLAB_STORE_USER
)
2593 *dbg_userword(cachep
, objp
) = NULL
;
2595 if (cachep
->flags
& SLAB_RED_ZONE
) {
2596 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2597 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2600 * Constructors are not allowed to allocate memory from the same
2601 * cache which they are a constructor for. Otherwise, deadlock.
2602 * They must also be threaded.
2604 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2605 cachep
->ctor(objp
+ obj_offset(cachep
), cachep
,
2608 if (cachep
->flags
& SLAB_RED_ZONE
) {
2609 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2610 slab_error(cachep
, "constructor overwrote the"
2611 " end of an object");
2612 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2613 slab_error(cachep
, "constructor overwrote the"
2614 " start of an object");
2616 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 &&
2617 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2618 kernel_map_pages(virt_to_page(objp
),
2619 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2622 cachep
->ctor(objp
, cachep
, ctor_flags
);
2624 slab_bufctl(slabp
)[i
] = i
+ 1;
2626 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2630 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2632 if (flags
& GFP_DMA
)
2633 BUG_ON(!(cachep
->gfpflags
& GFP_DMA
));
2635 BUG_ON(cachep
->gfpflags
& GFP_DMA
);
2638 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2641 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2645 next
= slab_bufctl(slabp
)[slabp
->free
];
2647 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2648 WARN_ON(slabp
->nodeid
!= nodeid
);
2655 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2656 void *objp
, int nodeid
)
2658 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2661 /* Verify that the slab belongs to the intended node */
2662 WARN_ON(slabp
->nodeid
!= nodeid
);
2664 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2665 printk(KERN_ERR
"slab: double free detected in cache "
2666 "'%s', objp %p\n", cachep
->name
, objp
);
2670 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2671 slabp
->free
= objnr
;
2676 * Map pages beginning at addr to the given cache and slab. This is required
2677 * for the slab allocator to be able to lookup the cache and slab of a
2678 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2680 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2686 page
= virt_to_page(addr
);
2689 if (likely(!PageCompound(page
)))
2690 nr_pages
<<= cache
->gfporder
;
2693 page_set_cache(page
, cache
);
2694 page_set_slab(page
, slab
);
2696 } while (--nr_pages
);
2700 * Grow (by 1) the number of slabs within a cache. This is called by
2701 * kmem_cache_alloc() when there are no active objs left in a cache.
2703 static int cache_grow(struct kmem_cache
*cachep
,
2704 gfp_t flags
, int nodeid
, void *objp
)
2709 unsigned long ctor_flags
;
2710 struct kmem_list3
*l3
;
2713 * Be lazy and only check for valid flags here, keeping it out of the
2714 * critical path in kmem_cache_alloc().
2716 BUG_ON(flags
& ~(GFP_DMA
| GFP_LEVEL_MASK
| __GFP_NO_GROW
));
2717 if (flags
& __GFP_NO_GROW
)
2720 ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
2721 local_flags
= (flags
& GFP_LEVEL_MASK
);
2722 if (!(local_flags
& __GFP_WAIT
))
2724 * Not allowed to sleep. Need to tell a constructor about
2725 * this - it might need to know...
2727 ctor_flags
|= SLAB_CTOR_ATOMIC
;
2729 /* Take the l3 list lock to change the colour_next on this node */
2731 l3
= cachep
->nodelists
[nodeid
];
2732 spin_lock(&l3
->list_lock
);
2734 /* Get colour for the slab, and cal the next value. */
2735 offset
= l3
->colour_next
;
2737 if (l3
->colour_next
>= cachep
->colour
)
2738 l3
->colour_next
= 0;
2739 spin_unlock(&l3
->list_lock
);
2741 offset
*= cachep
->colour_off
;
2743 if (local_flags
& __GFP_WAIT
)
2747 * The test for missing atomic flag is performed here, rather than
2748 * the more obvious place, simply to reduce the critical path length
2749 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2750 * will eventually be caught here (where it matters).
2752 kmem_flagcheck(cachep
, flags
);
2755 * Get mem for the objs. Attempt to allocate a physical page from
2759 objp
= kmem_getpages(cachep
, flags
, nodeid
);
2763 /* Get slab management. */
2764 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2765 local_flags
& ~GFP_THISNODE
, nodeid
);
2769 slabp
->nodeid
= nodeid
;
2770 slab_map_pages(cachep
, slabp
, objp
);
2772 cache_init_objs(cachep
, slabp
, ctor_flags
);
2774 if (local_flags
& __GFP_WAIT
)
2775 local_irq_disable();
2777 spin_lock(&l3
->list_lock
);
2779 /* Make slab active. */
2780 list_add_tail(&slabp
->list
, &(l3
->slabs_free
));
2781 STATS_INC_GROWN(cachep
);
2782 l3
->free_objects
+= cachep
->num
;
2783 spin_unlock(&l3
->list_lock
);
2786 kmem_freepages(cachep
, objp
);
2788 if (local_flags
& __GFP_WAIT
)
2789 local_irq_disable();
2796 * Perform extra freeing checks:
2797 * - detect bad pointers.
2798 * - POISON/RED_ZONE checking
2799 * - destructor calls, for caches with POISON+dtor
2801 static void kfree_debugcheck(const void *objp
)
2805 if (!virt_addr_valid(objp
)) {
2806 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2807 (unsigned long)objp
);
2810 page
= virt_to_page(objp
);
2811 if (!PageSlab(page
)) {
2812 printk(KERN_ERR
"kfree_debugcheck: bad ptr %lxh.\n",
2813 (unsigned long)objp
);
2818 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2820 unsigned long redzone1
, redzone2
;
2822 redzone1
= *dbg_redzone1(cache
, obj
);
2823 redzone2
= *dbg_redzone2(cache
, obj
);
2828 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2831 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2832 slab_error(cache
, "double free detected");
2834 slab_error(cache
, "memory outside object was overwritten");
2836 printk(KERN_ERR
"%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2837 obj
, redzone1
, redzone2
);
2840 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2847 objp
-= obj_offset(cachep
);
2848 kfree_debugcheck(objp
);
2849 page
= virt_to_page(objp
);
2851 slabp
= page_get_slab(page
);
2853 if (cachep
->flags
& SLAB_RED_ZONE
) {
2854 verify_redzone_free(cachep
, objp
);
2855 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2856 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2858 if (cachep
->flags
& SLAB_STORE_USER
)
2859 *dbg_userword(cachep
, objp
) = caller
;
2861 objnr
= obj_to_index(cachep
, slabp
, objp
);
2863 BUG_ON(objnr
>= cachep
->num
);
2864 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2866 if (cachep
->flags
& SLAB_DEBUG_INITIAL
) {
2868 * Need to call the slab's constructor so the caller can
2869 * perform a verify of its state (debugging). Called without
2870 * the cache-lock held.
2872 cachep
->ctor(objp
+ obj_offset(cachep
),
2873 cachep
, SLAB_CTOR_CONSTRUCTOR
| SLAB_CTOR_VERIFY
);
2875 if (cachep
->flags
& SLAB_POISON
&& cachep
->dtor
) {
2876 /* we want to cache poison the object,
2877 * call the destruction callback
2879 cachep
->dtor(objp
+ obj_offset(cachep
), cachep
, 0);
2881 #ifdef CONFIG_DEBUG_SLAB_LEAK
2882 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2884 if (cachep
->flags
& SLAB_POISON
) {
2885 #ifdef CONFIG_DEBUG_PAGEALLOC
2886 if ((cachep
->buffer_size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2887 store_stackinfo(cachep
, objp
, (unsigned long)caller
);
2888 kernel_map_pages(virt_to_page(objp
),
2889 cachep
->buffer_size
/ PAGE_SIZE
, 0);
2891 poison_obj(cachep
, objp
, POISON_FREE
);
2894 poison_obj(cachep
, objp
, POISON_FREE
);
2900 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2905 /* Check slab's freelist to see if this obj is there. */
2906 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2908 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2911 if (entries
!= cachep
->num
- slabp
->inuse
) {
2913 printk(KERN_ERR
"slab: Internal list corruption detected in "
2914 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2915 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
);
2917 i
< sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
);
2920 printk("\n%03x:", i
);
2921 printk(" %02x", ((unsigned char *)slabp
)[i
]);
2928 #define kfree_debugcheck(x) do { } while(0)
2929 #define cache_free_debugcheck(x,objp,z) (objp)
2930 #define check_slabp(x,y) do { } while(0)
2933 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2936 struct kmem_list3
*l3
;
2937 struct array_cache
*ac
;
2940 node
= numa_node_id();
2943 ac
= cpu_cache_get(cachep
);
2945 batchcount
= ac
->batchcount
;
2946 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2948 * If there was little recent activity on this cache, then
2949 * perform only a partial refill. Otherwise we could generate
2952 batchcount
= BATCHREFILL_LIMIT
;
2954 l3
= cachep
->nodelists
[node
];
2956 BUG_ON(ac
->avail
> 0 || !l3
);
2957 spin_lock(&l3
->list_lock
);
2959 /* See if we can refill from the shared array */
2960 if (l3
->shared
&& transfer_objects(ac
, l3
->shared
, batchcount
))
2963 while (batchcount
> 0) {
2964 struct list_head
*entry
;
2966 /* Get slab alloc is to come from. */
2967 entry
= l3
->slabs_partial
.next
;
2968 if (entry
== &l3
->slabs_partial
) {
2969 l3
->free_touched
= 1;
2970 entry
= l3
->slabs_free
.next
;
2971 if (entry
== &l3
->slabs_free
)
2975 slabp
= list_entry(entry
, struct slab
, list
);
2976 check_slabp(cachep
, slabp
);
2977 check_spinlock_acquired(cachep
);
2978 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2979 STATS_INC_ALLOCED(cachep
);
2980 STATS_INC_ACTIVE(cachep
);
2981 STATS_SET_HIGH(cachep
);
2983 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slabp
,
2986 check_slabp(cachep
, slabp
);
2988 /* move slabp to correct slabp list: */
2989 list_del(&slabp
->list
);
2990 if (slabp
->free
== BUFCTL_END
)
2991 list_add(&slabp
->list
, &l3
->slabs_full
);
2993 list_add(&slabp
->list
, &l3
->slabs_partial
);
2997 l3
->free_objects
-= ac
->avail
;
2999 spin_unlock(&l3
->list_lock
);
3001 if (unlikely(!ac
->avail
)) {
3003 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3005 /* cache_grow can reenable interrupts, then ac could change. */
3006 ac
= cpu_cache_get(cachep
);
3007 if (!x
&& ac
->avail
== 0) /* no objects in sight? abort */
3010 if (!ac
->avail
) /* objects refilled by interrupt? */
3014 return ac
->entry
[--ac
->avail
];
3017 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3020 might_sleep_if(flags
& __GFP_WAIT
);
3022 kmem_flagcheck(cachep
, flags
);
3027 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3028 gfp_t flags
, void *objp
, void *caller
)
3032 if (cachep
->flags
& SLAB_POISON
) {
3033 #ifdef CONFIG_DEBUG_PAGEALLOC
3034 if ((cachep
->buffer_size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3035 kernel_map_pages(virt_to_page(objp
),
3036 cachep
->buffer_size
/ PAGE_SIZE
, 1);
3038 check_poison_obj(cachep
, objp
);
3040 check_poison_obj(cachep
, objp
);
3042 poison_obj(cachep
, objp
, POISON_INUSE
);
3044 if (cachep
->flags
& SLAB_STORE_USER
)
3045 *dbg_userword(cachep
, objp
) = caller
;
3047 if (cachep
->flags
& SLAB_RED_ZONE
) {
3048 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3049 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3050 slab_error(cachep
, "double free, or memory outside"
3051 " object was overwritten");
3053 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3054 objp
, *dbg_redzone1(cachep
, objp
),
3055 *dbg_redzone2(cachep
, objp
));
3057 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3058 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3060 #ifdef CONFIG_DEBUG_SLAB_LEAK
3065 slabp
= page_get_slab(virt_to_page(objp
));
3066 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->buffer_size
;
3067 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3070 objp
+= obj_offset(cachep
);
3071 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
) {
3072 unsigned long ctor_flags
= SLAB_CTOR_CONSTRUCTOR
;
3074 if (!(flags
& __GFP_WAIT
))
3075 ctor_flags
|= SLAB_CTOR_ATOMIC
;
3077 cachep
->ctor(objp
, cachep
, ctor_flags
);
3079 #if ARCH_SLAB_MINALIGN
3080 if ((u32
)objp
& (ARCH_SLAB_MINALIGN
-1)) {
3081 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3082 objp
, ARCH_SLAB_MINALIGN
);
3088 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3091 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3094 struct array_cache
*ac
;
3097 ac
= cpu_cache_get(cachep
);
3098 if (likely(ac
->avail
)) {
3099 STATS_INC_ALLOCHIT(cachep
);
3101 objp
= ac
->entry
[--ac
->avail
];
3103 STATS_INC_ALLOCMISS(cachep
);
3104 objp
= cache_alloc_refill(cachep
, flags
);
3109 static __always_inline
void *__cache_alloc(struct kmem_cache
*cachep
,
3110 gfp_t flags
, void *caller
)
3112 unsigned long save_flags
;
3115 cache_alloc_debugcheck_before(cachep
, flags
);
3117 local_irq_save(save_flags
);
3119 if (unlikely(NUMA_BUILD
&&
3120 current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
)))
3121 objp
= alternate_node_alloc(cachep
, flags
);
3124 objp
= ____cache_alloc(cachep
, flags
);
3126 * We may just have run out of memory on the local node.
3127 * ____cache_alloc_node() knows how to locate memory on other nodes
3129 if (NUMA_BUILD
&& !objp
)
3130 objp
= ____cache_alloc_node(cachep
, flags
, numa_node_id());
3131 local_irq_restore(save_flags
);
3132 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
,
3140 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3142 * If we are in_interrupt, then process context, including cpusets and
3143 * mempolicy, may not apply and should not be used for allocation policy.
3145 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3147 int nid_alloc
, nid_here
;
3149 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3151 nid_alloc
= nid_here
= numa_node_id();
3152 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3153 nid_alloc
= cpuset_mem_spread_node();
3154 else if (current
->mempolicy
)
3155 nid_alloc
= slab_node(current
->mempolicy
);
3156 if (nid_alloc
!= nid_here
)
3157 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3162 * Fallback function if there was no memory available and no objects on a
3163 * certain node and fall back is permitted. First we scan all the
3164 * available nodelists for available objects. If that fails then we
3165 * perform an allocation without specifying a node. This allows the page
3166 * allocator to do its reclaim / fallback magic. We then insert the
3167 * slab into the proper nodelist and then allocate from it.
3169 void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3171 struct zonelist
*zonelist
= &NODE_DATA(slab_node(current
->mempolicy
))
3172 ->node_zonelists
[gfp_zone(flags
)];
3179 * Look through allowed nodes for objects available
3180 * from existing per node queues.
3182 for (z
= zonelist
->zones
; *z
&& !obj
; z
++) {
3183 nid
= zone_to_nid(*z
);
3185 if (cpuset_zone_allowed(*z
, flags
) &&
3186 cache
->nodelists
[nid
] &&
3187 cache
->nodelists
[nid
]->free_objects
)
3188 obj
= ____cache_alloc_node(cache
,
3189 flags
| GFP_THISNODE
, nid
);
3194 * This allocation will be performed within the constraints
3195 * of the current cpuset / memory policy requirements.
3196 * We may trigger various forms of reclaim on the allowed
3197 * set and go into memory reserves if necessary.
3199 obj
= kmem_getpages(cache
, flags
, -1);
3202 * Insert into the appropriate per node queues
3204 nid
= page_to_nid(virt_to_page(obj
));
3205 if (cache_grow(cache
, flags
, nid
, obj
)) {
3206 obj
= ____cache_alloc_node(cache
,
3207 flags
| GFP_THISNODE
, nid
);
3210 * Another processor may allocate the
3211 * objects in the slab since we are
3212 * not holding any locks.
3216 kmem_freepages(cache
, obj
);
3225 * A interface to enable slab creation on nodeid
3227 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3230 struct list_head
*entry
;
3232 struct kmem_list3
*l3
;
3236 l3
= cachep
->nodelists
[nodeid
];
3241 spin_lock(&l3
->list_lock
);
3242 entry
= l3
->slabs_partial
.next
;
3243 if (entry
== &l3
->slabs_partial
) {
3244 l3
->free_touched
= 1;
3245 entry
= l3
->slabs_free
.next
;
3246 if (entry
== &l3
->slabs_free
)
3250 slabp
= list_entry(entry
, struct slab
, list
);
3251 check_spinlock_acquired_node(cachep
, nodeid
);
3252 check_slabp(cachep
, slabp
);
3254 STATS_INC_NODEALLOCS(cachep
);
3255 STATS_INC_ACTIVE(cachep
);
3256 STATS_SET_HIGH(cachep
);
3258 BUG_ON(slabp
->inuse
== cachep
->num
);
3260 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3261 check_slabp(cachep
, slabp
);
3263 /* move slabp to correct slabp list: */
3264 list_del(&slabp
->list
);
3266 if (slabp
->free
== BUFCTL_END
)
3267 list_add(&slabp
->list
, &l3
->slabs_full
);
3269 list_add(&slabp
->list
, &l3
->slabs_partial
);
3271 spin_unlock(&l3
->list_lock
);
3275 spin_unlock(&l3
->list_lock
);
3276 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3280 if (!(flags
& __GFP_THISNODE
))
3281 /* Unable to grow the cache. Fall back to other nodes. */
3282 return fallback_alloc(cachep
, flags
);
3292 * Caller needs to acquire correct kmem_list's list_lock
3294 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3298 struct kmem_list3
*l3
;
3300 for (i
= 0; i
< nr_objects
; i
++) {
3301 void *objp
= objpp
[i
];
3304 slabp
= virt_to_slab(objp
);
3305 l3
= cachep
->nodelists
[node
];
3306 list_del(&slabp
->list
);
3307 check_spinlock_acquired_node(cachep
, node
);
3308 check_slabp(cachep
, slabp
);
3309 slab_put_obj(cachep
, slabp
, objp
, node
);
3310 STATS_DEC_ACTIVE(cachep
);
3312 check_slabp(cachep
, slabp
);
3314 /* fixup slab chains */
3315 if (slabp
->inuse
== 0) {
3316 if (l3
->free_objects
> l3
->free_limit
) {
3317 l3
->free_objects
-= cachep
->num
;
3318 /* No need to drop any previously held
3319 * lock here, even if we have a off-slab slab
3320 * descriptor it is guaranteed to come from
3321 * a different cache, refer to comments before
3324 slab_destroy(cachep
, slabp
);
3326 list_add(&slabp
->list
, &l3
->slabs_free
);
3329 /* Unconditionally move a slab to the end of the
3330 * partial list on free - maximum time for the
3331 * other objects to be freed, too.
3333 list_add_tail(&slabp
->list
, &l3
->slabs_partial
);
3338 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3341 struct kmem_list3
*l3
;
3342 int node
= numa_node_id();
3344 batchcount
= ac
->batchcount
;
3346 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3349 l3
= cachep
->nodelists
[node
];
3350 spin_lock(&l3
->list_lock
);
3352 struct array_cache
*shared_array
= l3
->shared
;
3353 int max
= shared_array
->limit
- shared_array
->avail
;
3355 if (batchcount
> max
)
3357 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3358 ac
->entry
, sizeof(void *) * batchcount
);
3359 shared_array
->avail
+= batchcount
;
3364 free_block(cachep
, ac
->entry
, batchcount
, node
);
3369 struct list_head
*p
;
3371 p
= l3
->slabs_free
.next
;
3372 while (p
!= &(l3
->slabs_free
)) {
3375 slabp
= list_entry(p
, struct slab
, list
);
3376 BUG_ON(slabp
->inuse
);
3381 STATS_SET_FREEABLE(cachep
, i
);
3384 spin_unlock(&l3
->list_lock
);
3385 ac
->avail
-= batchcount
;
3386 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3390 * Release an obj back to its cache. If the obj has a constructed state, it must
3391 * be in this state _before_ it is released. Called with disabled ints.
3393 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
)
3395 struct array_cache
*ac
= cpu_cache_get(cachep
);
3398 objp
= cache_free_debugcheck(cachep
, objp
, __builtin_return_address(0));
3400 if (cache_free_alien(cachep
, objp
))
3403 if (likely(ac
->avail
< ac
->limit
)) {
3404 STATS_INC_FREEHIT(cachep
);
3405 ac
->entry
[ac
->avail
++] = objp
;
3408 STATS_INC_FREEMISS(cachep
);
3409 cache_flusharray(cachep
, ac
);
3410 ac
->entry
[ac
->avail
++] = objp
;
3415 * kmem_cache_alloc - Allocate an object
3416 * @cachep: The cache to allocate from.
3417 * @flags: See kmalloc().
3419 * Allocate an object from this cache. The flags are only relevant
3420 * if the cache has no available objects.
3422 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3424 return __cache_alloc(cachep
, flags
, __builtin_return_address(0));
3426 EXPORT_SYMBOL(kmem_cache_alloc
);
3429 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3430 * @cache: The cache to allocate from.
3431 * @flags: See kmalloc().
3433 * Allocate an object from this cache and set the allocated memory to zero.
3434 * The flags are only relevant if the cache has no available objects.
3436 void *kmem_cache_zalloc(struct kmem_cache
*cache
, gfp_t flags
)
3438 void *ret
= __cache_alloc(cache
, flags
, __builtin_return_address(0));
3440 memset(ret
, 0, obj_size(cache
));
3443 EXPORT_SYMBOL(kmem_cache_zalloc
);
3446 * kmem_ptr_validate - check if an untrusted pointer might
3448 * @cachep: the cache we're checking against
3449 * @ptr: pointer to validate
3451 * This verifies that the untrusted pointer looks sane:
3452 * it is _not_ a guarantee that the pointer is actually
3453 * part of the slab cache in question, but it at least
3454 * validates that the pointer can be dereferenced and
3455 * looks half-way sane.
3457 * Currently only used for dentry validation.
3459 int fastcall
kmem_ptr_validate(struct kmem_cache
*cachep
, void *ptr
)
3461 unsigned long addr
= (unsigned long)ptr
;
3462 unsigned long min_addr
= PAGE_OFFSET
;
3463 unsigned long align_mask
= BYTES_PER_WORD
- 1;
3464 unsigned long size
= cachep
->buffer_size
;
3467 if (unlikely(addr
< min_addr
))
3469 if (unlikely(addr
> (unsigned long)high_memory
- size
))
3471 if (unlikely(addr
& align_mask
))
3473 if (unlikely(!kern_addr_valid(addr
)))
3475 if (unlikely(!kern_addr_valid(addr
+ size
- 1)))
3477 page
= virt_to_page(ptr
);
3478 if (unlikely(!PageSlab(page
)))
3480 if (unlikely(page_get_cache(page
) != cachep
))
3489 * kmem_cache_alloc_node - Allocate an object on the specified node
3490 * @cachep: The cache to allocate from.
3491 * @flags: See kmalloc().
3492 * @nodeid: node number of the target node.
3494 * Identical to kmem_cache_alloc but it will allocate memory on the given
3495 * node, which can improve the performance for cpu bound structures.
3497 * Fallback to other node is possible if __GFP_THISNODE is not set.
3499 static __always_inline
void *
3500 __cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3501 int nodeid
, void *caller
)
3503 unsigned long save_flags
;
3506 cache_alloc_debugcheck_before(cachep
, flags
);
3507 local_irq_save(save_flags
);
3509 if (unlikely(nodeid
== -1))
3510 nodeid
= numa_node_id();
3512 if (likely(cachep
->nodelists
[nodeid
])) {
3513 if (nodeid
== numa_node_id()) {
3515 * Use the locally cached objects if possible.
3516 * However ____cache_alloc does not allow fallback
3517 * to other nodes. It may fail while we still have
3518 * objects on other nodes available.
3520 ptr
= ____cache_alloc(cachep
, flags
);
3523 /* ___cache_alloc_node can fall back to other nodes */
3524 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3527 /* Node not bootstrapped yet */
3528 if (!(flags
& __GFP_THISNODE
))
3529 ptr
= fallback_alloc(cachep
, flags
);
3532 local_irq_restore(save_flags
);
3533 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3538 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3540 return __cache_alloc_node(cachep
, flags
, nodeid
,
3541 __builtin_return_address(0));
3543 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3545 static __always_inline
void *
3546 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, void *caller
)
3548 struct kmem_cache
*cachep
;
3550 cachep
= kmem_find_general_cachep(size
, flags
);
3551 if (unlikely(cachep
== NULL
))
3553 return kmem_cache_alloc_node(cachep
, flags
, node
);
3556 #ifdef CONFIG_DEBUG_SLAB
3557 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3559 return __do_kmalloc_node(size
, flags
, node
,
3560 __builtin_return_address(0));
3562 EXPORT_SYMBOL(__kmalloc_node
);
3564 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3565 int node
, void *caller
)
3567 return __do_kmalloc_node(size
, flags
, node
, caller
);
3569 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3571 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3573 return __do_kmalloc_node(size
, flags
, node
, NULL
);
3575 EXPORT_SYMBOL(__kmalloc_node
);
3576 #endif /* CONFIG_DEBUG_SLAB */
3577 #endif /* CONFIG_NUMA */
3580 * __do_kmalloc - allocate memory
3581 * @size: how many bytes of memory are required.
3582 * @flags: the type of memory to allocate (see kmalloc).
3583 * @caller: function caller for debug tracking of the caller
3585 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3588 struct kmem_cache
*cachep
;
3590 /* If you want to save a few bytes .text space: replace
3592 * Then kmalloc uses the uninlined functions instead of the inline
3595 cachep
= __find_general_cachep(size
, flags
);
3596 if (unlikely(cachep
== NULL
))
3598 return __cache_alloc(cachep
, flags
, caller
);
3602 #ifdef CONFIG_DEBUG_SLAB
3603 void *__kmalloc(size_t size
, gfp_t flags
)
3605 return __do_kmalloc(size
, flags
, __builtin_return_address(0));
3607 EXPORT_SYMBOL(__kmalloc
);
3609 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, void *caller
)
3611 return __do_kmalloc(size
, flags
, caller
);
3613 EXPORT_SYMBOL(__kmalloc_track_caller
);
3616 void *__kmalloc(size_t size
, gfp_t flags
)
3618 return __do_kmalloc(size
, flags
, NULL
);
3620 EXPORT_SYMBOL(__kmalloc
);
3624 * kmem_cache_free - Deallocate an object
3625 * @cachep: The cache the allocation was from.
3626 * @objp: The previously allocated object.
3628 * Free an object which was previously allocated from this
3631 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3633 unsigned long flags
;
3635 BUG_ON(virt_to_cache(objp
) != cachep
);
3637 local_irq_save(flags
);
3638 __cache_free(cachep
, objp
);
3639 local_irq_restore(flags
);
3641 EXPORT_SYMBOL(kmem_cache_free
);
3644 * kfree - free previously allocated memory
3645 * @objp: pointer returned by kmalloc.
3647 * If @objp is NULL, no operation is performed.
3649 * Don't free memory not originally allocated by kmalloc()
3650 * or you will run into trouble.
3652 void kfree(const void *objp
)
3654 struct kmem_cache
*c
;
3655 unsigned long flags
;
3657 if (unlikely(!objp
))
3659 local_irq_save(flags
);
3660 kfree_debugcheck(objp
);
3661 c
= virt_to_cache(objp
);
3662 debug_check_no_locks_freed(objp
, obj_size(c
));
3663 __cache_free(c
, (void *)objp
);
3664 local_irq_restore(flags
);
3666 EXPORT_SYMBOL(kfree
);
3668 unsigned int kmem_cache_size(struct kmem_cache
*cachep
)
3670 return obj_size(cachep
);
3672 EXPORT_SYMBOL(kmem_cache_size
);
3674 const char *kmem_cache_name(struct kmem_cache
*cachep
)
3676 return cachep
->name
;
3678 EXPORT_SYMBOL_GPL(kmem_cache_name
);
3681 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3683 static int alloc_kmemlist(struct kmem_cache
*cachep
)
3686 struct kmem_list3
*l3
;
3687 struct array_cache
*new_shared
;
3688 struct array_cache
**new_alien
= NULL
;
3690 for_each_online_node(node
) {
3692 if (use_alien_caches
) {
3693 new_alien
= alloc_alien_cache(node
, cachep
->limit
);
3698 new_shared
= alloc_arraycache(node
,
3699 cachep
->shared
*cachep
->batchcount
,
3702 free_alien_cache(new_alien
);
3706 l3
= cachep
->nodelists
[node
];
3708 struct array_cache
*shared
= l3
->shared
;
3710 spin_lock_irq(&l3
->list_lock
);
3713 free_block(cachep
, shared
->entry
,
3714 shared
->avail
, node
);
3716 l3
->shared
= new_shared
;
3718 l3
->alien
= new_alien
;
3721 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3722 cachep
->batchcount
+ cachep
->num
;
3723 spin_unlock_irq(&l3
->list_lock
);
3725 free_alien_cache(new_alien
);
3728 l3
= kmalloc_node(sizeof(struct kmem_list3
), GFP_KERNEL
, node
);
3730 free_alien_cache(new_alien
);
3735 kmem_list3_init(l3
);
3736 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3737 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3738 l3
->shared
= new_shared
;
3739 l3
->alien
= new_alien
;
3740 l3
->free_limit
= (1 + nr_cpus_node(node
)) *
3741 cachep
->batchcount
+ cachep
->num
;
3742 cachep
->nodelists
[node
] = l3
;
3747 if (!cachep
->next
.next
) {
3748 /* Cache is not active yet. Roll back what we did */
3751 if (cachep
->nodelists
[node
]) {
3752 l3
= cachep
->nodelists
[node
];
3755 free_alien_cache(l3
->alien
);
3757 cachep
->nodelists
[node
] = NULL
;
3765 struct ccupdate_struct
{
3766 struct kmem_cache
*cachep
;
3767 struct array_cache
*new[NR_CPUS
];
3770 static void do_ccupdate_local(void *info
)
3772 struct ccupdate_struct
*new = info
;
3773 struct array_cache
*old
;
3776 old
= cpu_cache_get(new->cachep
);
3778 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3779 new->new[smp_processor_id()] = old
;
3782 /* Always called with the cache_chain_mutex held */
3783 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3784 int batchcount
, int shared
)
3786 struct ccupdate_struct
*new;
3789 new = kzalloc(sizeof(*new), GFP_KERNEL
);
3793 for_each_online_cpu(i
) {
3794 new->new[i
] = alloc_arraycache(cpu_to_node(i
), limit
,
3797 for (i
--; i
>= 0; i
--)
3803 new->cachep
= cachep
;
3805 on_each_cpu(do_ccupdate_local
, (void *)new, 1, 1);
3808 cachep
->batchcount
= batchcount
;
3809 cachep
->limit
= limit
;
3810 cachep
->shared
= shared
;
3812 for_each_online_cpu(i
) {
3813 struct array_cache
*ccold
= new->new[i
];
3816 spin_lock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3817 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_node(i
));
3818 spin_unlock_irq(&cachep
->nodelists
[cpu_to_node(i
)]->list_lock
);
3822 return alloc_kmemlist(cachep
);
3825 /* Called with cache_chain_mutex held always */
3826 static int enable_cpucache(struct kmem_cache
*cachep
)
3832 * The head array serves three purposes:
3833 * - create a LIFO ordering, i.e. return objects that are cache-warm
3834 * - reduce the number of spinlock operations.
3835 * - reduce the number of linked list operations on the slab and
3836 * bufctl chains: array operations are cheaper.
3837 * The numbers are guessed, we should auto-tune as described by
3840 if (cachep
->buffer_size
> 131072)
3842 else if (cachep
->buffer_size
> PAGE_SIZE
)
3844 else if (cachep
->buffer_size
> 1024)
3846 else if (cachep
->buffer_size
> 256)
3852 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3853 * allocation behaviour: Most allocs on one cpu, most free operations
3854 * on another cpu. For these cases, an efficient object passing between
3855 * cpus is necessary. This is provided by a shared array. The array
3856 * replaces Bonwick's magazine layer.
3857 * On uniprocessor, it's functionally equivalent (but less efficient)
3858 * to a larger limit. Thus disabled by default.
3862 if (cachep
->buffer_size
<= PAGE_SIZE
)
3868 * With debugging enabled, large batchcount lead to excessively long
3869 * periods with disabled local interrupts. Limit the batchcount
3874 err
= do_tune_cpucache(cachep
, limit
, (limit
+ 1) / 2, shared
);
3876 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3877 cachep
->name
, -err
);
3882 * Drain an array if it contains any elements taking the l3 lock only if
3883 * necessary. Note that the l3 listlock also protects the array_cache
3884 * if drain_array() is used on the shared array.
3886 void drain_array(struct kmem_cache
*cachep
, struct kmem_list3
*l3
,
3887 struct array_cache
*ac
, int force
, int node
)
3891 if (!ac
|| !ac
->avail
)
3893 if (ac
->touched
&& !force
) {
3896 spin_lock_irq(&l3
->list_lock
);
3898 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3899 if (tofree
> ac
->avail
)
3900 tofree
= (ac
->avail
+ 1) / 2;
3901 free_block(cachep
, ac
->entry
, tofree
, node
);
3902 ac
->avail
-= tofree
;
3903 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3904 sizeof(void *) * ac
->avail
);
3906 spin_unlock_irq(&l3
->list_lock
);
3911 * cache_reap - Reclaim memory from caches.
3912 * @unused: unused parameter
3914 * Called from workqueue/eventd every few seconds.
3916 * - clear the per-cpu caches for this CPU.
3917 * - return freeable pages to the main free memory pool.
3919 * If we cannot acquire the cache chain mutex then just give up - we'll try
3920 * again on the next iteration.
3922 static void cache_reap(struct work_struct
*unused
)
3924 struct kmem_cache
*searchp
;
3925 struct kmem_list3
*l3
;
3926 int node
= numa_node_id();
3928 if (!mutex_trylock(&cache_chain_mutex
)) {
3929 /* Give up. Setup the next iteration. */
3930 schedule_delayed_work(&__get_cpu_var(reap_work
),
3935 list_for_each_entry(searchp
, &cache_chain
, next
) {
3939 * We only take the l3 lock if absolutely necessary and we
3940 * have established with reasonable certainty that
3941 * we can do some work if the lock was obtained.
3943 l3
= searchp
->nodelists
[node
];
3945 reap_alien(searchp
, l3
);
3947 drain_array(searchp
, l3
, cpu_cache_get(searchp
), 0, node
);
3950 * These are racy checks but it does not matter
3951 * if we skip one check or scan twice.
3953 if (time_after(l3
->next_reap
, jiffies
))
3956 l3
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
3958 drain_array(searchp
, l3
, l3
->shared
, 0, node
);
3960 if (l3
->free_touched
)
3961 l3
->free_touched
= 0;
3965 freed
= drain_freelist(searchp
, l3
, (l3
->free_limit
+
3966 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3967 STATS_ADD_REAPED(searchp
, freed
);
3973 mutex_unlock(&cache_chain_mutex
);
3975 refresh_cpu_vm_stats(smp_processor_id());
3976 /* Set up the next iteration */
3977 schedule_delayed_work(&__get_cpu_var(reap_work
), REAPTIMEOUT_CPUC
);
3980 #ifdef CONFIG_PROC_FS
3982 static void print_slabinfo_header(struct seq_file
*m
)
3985 * Output format version, so at least we can change it
3986 * without _too_ many complaints.
3989 seq_puts(m
, "slabinfo - version: 2.1 (statistics)\n");
3991 seq_puts(m
, "slabinfo - version: 2.1\n");
3993 seq_puts(m
, "# name <active_objs> <num_objs> <objsize> "
3994 "<objperslab> <pagesperslab>");
3995 seq_puts(m
, " : tunables <limit> <batchcount> <sharedfactor>");
3996 seq_puts(m
, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3998 seq_puts(m
, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3999 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4000 seq_puts(m
, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4005 static void *s_start(struct seq_file
*m
, loff_t
*pos
)
4008 struct list_head
*p
;
4010 mutex_lock(&cache_chain_mutex
);
4012 print_slabinfo_header(m
);
4013 p
= cache_chain
.next
;
4016 if (p
== &cache_chain
)
4019 return list_entry(p
, struct kmem_cache
, next
);
4022 static void *s_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
4024 struct kmem_cache
*cachep
= p
;
4026 return cachep
->next
.next
== &cache_chain
?
4027 NULL
: list_entry(cachep
->next
.next
, struct kmem_cache
, next
);
4030 static void s_stop(struct seq_file
*m
, void *p
)
4032 mutex_unlock(&cache_chain_mutex
);
4035 static int s_show(struct seq_file
*m
, void *p
)
4037 struct kmem_cache
*cachep
= p
;
4039 unsigned long active_objs
;
4040 unsigned long num_objs
;
4041 unsigned long active_slabs
= 0;
4042 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4046 struct kmem_list3
*l3
;
4050 for_each_online_node(node
) {
4051 l3
= cachep
->nodelists
[node
];
4056 spin_lock_irq(&l3
->list_lock
);
4058 list_for_each_entry(slabp
, &l3
->slabs_full
, list
) {
4059 if (slabp
->inuse
!= cachep
->num
&& !error
)
4060 error
= "slabs_full accounting error";
4061 active_objs
+= cachep
->num
;
4064 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
) {
4065 if (slabp
->inuse
== cachep
->num
&& !error
)
4066 error
= "slabs_partial inuse accounting error";
4067 if (!slabp
->inuse
&& !error
)
4068 error
= "slabs_partial/inuse accounting error";
4069 active_objs
+= slabp
->inuse
;
4072 list_for_each_entry(slabp
, &l3
->slabs_free
, list
) {
4073 if (slabp
->inuse
&& !error
)
4074 error
= "slabs_free/inuse accounting error";
4077 free_objects
+= l3
->free_objects
;
4079 shared_avail
+= l3
->shared
->avail
;
4081 spin_unlock_irq(&l3
->list_lock
);
4083 num_slabs
+= active_slabs
;
4084 num_objs
= num_slabs
* cachep
->num
;
4085 if (num_objs
- active_objs
!= free_objects
&& !error
)
4086 error
= "free_objects accounting error";
4088 name
= cachep
->name
;
4090 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4092 seq_printf(m
, "%-17s %6lu %6lu %6u %4u %4d",
4093 name
, active_objs
, num_objs
, cachep
->buffer_size
,
4094 cachep
->num
, (1 << cachep
->gfporder
));
4095 seq_printf(m
, " : tunables %4u %4u %4u",
4096 cachep
->limit
, cachep
->batchcount
, cachep
->shared
);
4097 seq_printf(m
, " : slabdata %6lu %6lu %6lu",
4098 active_slabs
, num_slabs
, shared_avail
);
4101 unsigned long high
= cachep
->high_mark
;
4102 unsigned long allocs
= cachep
->num_allocations
;
4103 unsigned long grown
= cachep
->grown
;
4104 unsigned long reaped
= cachep
->reaped
;
4105 unsigned long errors
= cachep
->errors
;
4106 unsigned long max_freeable
= cachep
->max_freeable
;
4107 unsigned long node_allocs
= cachep
->node_allocs
;
4108 unsigned long node_frees
= cachep
->node_frees
;
4109 unsigned long overflows
= cachep
->node_overflow
;
4111 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu \
4112 %4lu %4lu %4lu %4lu %4lu", allocs
, high
, grown
,
4113 reaped
, errors
, max_freeable
, node_allocs
,
4114 node_frees
, overflows
);
4118 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4119 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4120 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4121 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4123 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4124 allochit
, allocmiss
, freehit
, freemiss
);
4132 * slabinfo_op - iterator that generates /proc/slabinfo
4141 * num-pages-per-slab
4142 * + further values on SMP and with statistics enabled
4145 const struct seq_operations slabinfo_op
= {
4152 #define MAX_SLABINFO_WRITE 128
4154 * slabinfo_write - Tuning for the slab allocator
4156 * @buffer: user buffer
4157 * @count: data length
4160 ssize_t
slabinfo_write(struct file
*file
, const char __user
* buffer
,
4161 size_t count
, loff_t
*ppos
)
4163 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4164 int limit
, batchcount
, shared
, res
;
4165 struct kmem_cache
*cachep
;
4167 if (count
> MAX_SLABINFO_WRITE
)
4169 if (copy_from_user(&kbuf
, buffer
, count
))
4171 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4173 tmp
= strchr(kbuf
, ' ');
4178 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4181 /* Find the cache in the chain of caches. */
4182 mutex_lock(&cache_chain_mutex
);
4184 list_for_each_entry(cachep
, &cache_chain
, next
) {
4185 if (!strcmp(cachep
->name
, kbuf
)) {
4186 if (limit
< 1 || batchcount
< 1 ||
4187 batchcount
> limit
|| shared
< 0) {
4190 res
= do_tune_cpucache(cachep
, limit
,
4191 batchcount
, shared
);
4196 mutex_unlock(&cache_chain_mutex
);
4202 #ifdef CONFIG_DEBUG_SLAB_LEAK
4204 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4207 struct list_head
*p
;
4209 mutex_lock(&cache_chain_mutex
);
4210 p
= cache_chain
.next
;
4213 if (p
== &cache_chain
)
4216 return list_entry(p
, struct kmem_cache
, next
);
4219 static inline int add_caller(unsigned long *n
, unsigned long v
)
4229 unsigned long *q
= p
+ 2 * i
;
4243 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4249 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4255 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->buffer_size
) {
4256 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4258 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4263 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4265 #ifdef CONFIG_KALLSYMS
4268 unsigned long offset
, size
;
4269 char namebuf
[KSYM_NAME_LEN
+1];
4271 name
= kallsyms_lookup(address
, &size
, &offset
, &modname
, namebuf
);
4274 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4276 seq_printf(m
, " [%s]", modname
);
4280 seq_printf(m
, "%p", (void *)address
);
4283 static int leaks_show(struct seq_file
*m
, void *p
)
4285 struct kmem_cache
*cachep
= p
;
4287 struct kmem_list3
*l3
;
4289 unsigned long *n
= m
->private;
4293 if (!(cachep
->flags
& SLAB_STORE_USER
))
4295 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4298 /* OK, we can do it */
4302 for_each_online_node(node
) {
4303 l3
= cachep
->nodelists
[node
];
4308 spin_lock_irq(&l3
->list_lock
);
4310 list_for_each_entry(slabp
, &l3
->slabs_full
, list
)
4311 handle_slab(n
, cachep
, slabp
);
4312 list_for_each_entry(slabp
, &l3
->slabs_partial
, list
)
4313 handle_slab(n
, cachep
, slabp
);
4314 spin_unlock_irq(&l3
->list_lock
);
4316 name
= cachep
->name
;
4318 /* Increase the buffer size */
4319 mutex_unlock(&cache_chain_mutex
);
4320 m
->private = kzalloc(n
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4322 /* Too bad, we are really out */
4324 mutex_lock(&cache_chain_mutex
);
4327 *(unsigned long *)m
->private = n
[0] * 2;
4329 mutex_lock(&cache_chain_mutex
);
4330 /* Now make sure this entry will be retried */
4334 for (i
= 0; i
< n
[1]; i
++) {
4335 seq_printf(m
, "%s: %lu ", name
, n
[2*i
+3]);
4336 show_symbol(m
, n
[2*i
+2]);
4343 const struct seq_operations slabstats_op
= {
4344 .start
= leaks_start
,
4353 * ksize - get the actual amount of memory allocated for a given object
4354 * @objp: Pointer to the object
4356 * kmalloc may internally round up allocations and return more memory
4357 * than requested. ksize() can be used to determine the actual amount of
4358 * memory allocated. The caller may use this additional memory, even though
4359 * a smaller amount of memory was initially specified with the kmalloc call.
4360 * The caller must guarantee that objp points to a valid object previously
4361 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4362 * must not be freed during the duration of the call.
4364 unsigned int ksize(const void *objp
)
4366 if (unlikely(objp
== NULL
))
4369 return obj_size(virt_to_cache(objp
));