3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly
;
169 * Bufctl's are used for linking objs within a slab
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 typedef unsigned int kmem_bufctl_t
;
186 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
188 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
206 struct rcu_head head
;
207 struct kmem_cache
*cachep
;
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.
221 struct list_head list
;
222 unsigned long colouroff
;
223 void *s_mem
; /* including colour offset */
224 unsigned int inuse
; /* num of objs active in slab */
226 unsigned short nodeid
;
228 struct slab_rcu __slab_cover_slab_rcu
;
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
240 * The limit is stored in the per-cpu structure to reduce the data cache
247 unsigned int batchcount
;
248 unsigned int touched
;
251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
255 * Entries should not be directly dereferenced as
256 * entries belonging to slabs marked pfmemalloc will
257 * have the lower bits set SLAB_OBJ_PFMEMALLOC
261 #define SLAB_OBJ_PFMEMALLOC 1
262 static inline bool is_obj_pfmemalloc(void *objp
)
264 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
267 static inline void set_obj_pfmemalloc(void **objp
)
269 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
273 static inline void clear_obj_pfmemalloc(void **objp
)
275 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init
{
284 struct array_cache cache
;
285 void *entries
[BOOT_CPUCACHE_ENTRIES
];
289 * Need this for bootstrapping a per node allocator.
291 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
292 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
293 #define CACHE_CACHE 0
294 #define SIZE_AC MAX_NUMNODES
295 #define SIZE_NODE (2 * MAX_NUMNODES)
297 static int drain_freelist(struct kmem_cache
*cache
,
298 struct kmem_cache_node
*n
, int tofree
);
299 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
301 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
302 static void cache_reap(struct work_struct
*unused
);
304 static int slab_early_init
= 1;
306 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
307 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
309 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
311 INIT_LIST_HEAD(&parent
->slabs_full
);
312 INIT_LIST_HEAD(&parent
->slabs_partial
);
313 INIT_LIST_HEAD(&parent
->slabs_free
);
314 parent
->shared
= NULL
;
315 parent
->alien
= NULL
;
316 parent
->colour_next
= 0;
317 spin_lock_init(&parent
->list_lock
);
318 parent
->free_objects
= 0;
319 parent
->free_touched
= 0;
322 #define MAKE_LIST(cachep, listp, slab, nodeid) \
324 INIT_LIST_HEAD(listp); \
325 list_splice(&(cachep->node[nodeid]->slab), listp); \
328 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
330 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
331 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
332 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
335 #define CFLGS_OFF_SLAB (0x80000000UL)
336 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
338 #define BATCHREFILL_LIMIT 16
340 * Optimization question: fewer reaps means less probability for unnessary
341 * cpucache drain/refill cycles.
343 * OTOH the cpuarrays can contain lots of objects,
344 * which could lock up otherwise freeable slabs.
346 #define REAPTIMEOUT_CPUC (2*HZ)
347 #define REAPTIMEOUT_LIST3 (4*HZ)
350 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
351 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
352 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
353 #define STATS_INC_GROWN(x) ((x)->grown++)
354 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
355 #define STATS_SET_HIGH(x) \
357 if ((x)->num_active > (x)->high_mark) \
358 (x)->high_mark = (x)->num_active; \
360 #define STATS_INC_ERR(x) ((x)->errors++)
361 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
362 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
363 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
364 #define STATS_SET_FREEABLE(x, i) \
366 if ((x)->max_freeable < i) \
367 (x)->max_freeable = i; \
369 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
370 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
371 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
372 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
374 #define STATS_INC_ACTIVE(x) do { } while (0)
375 #define STATS_DEC_ACTIVE(x) do { } while (0)
376 #define STATS_INC_ALLOCED(x) do { } while (0)
377 #define STATS_INC_GROWN(x) do { } while (0)
378 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
379 #define STATS_SET_HIGH(x) do { } while (0)
380 #define STATS_INC_ERR(x) do { } while (0)
381 #define STATS_INC_NODEALLOCS(x) do { } while (0)
382 #define STATS_INC_NODEFREES(x) do { } while (0)
383 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
384 #define STATS_SET_FREEABLE(x, i) do { } while (0)
385 #define STATS_INC_ALLOCHIT(x) do { } while (0)
386 #define STATS_INC_ALLOCMISS(x) do { } while (0)
387 #define STATS_INC_FREEHIT(x) do { } while (0)
388 #define STATS_INC_FREEMISS(x) do { } while (0)
394 * memory layout of objects:
396 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
397 * the end of an object is aligned with the end of the real
398 * allocation. Catches writes behind the end of the allocation.
399 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
401 * cachep->obj_offset: The real object.
402 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
403 * cachep->size - 1* BYTES_PER_WORD: last caller address
404 * [BYTES_PER_WORD long]
406 static int obj_offset(struct kmem_cache
*cachep
)
408 return cachep
->obj_offset
;
411 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
413 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
414 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
415 sizeof(unsigned long long));
418 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
420 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
421 if (cachep
->flags
& SLAB_STORE_USER
)
422 return (unsigned long long *)(objp
+ cachep
->size
-
423 sizeof(unsigned long long) -
425 return (unsigned long long *) (objp
+ cachep
->size
-
426 sizeof(unsigned long long));
429 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
431 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
432 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
437 #define obj_offset(x) 0
438 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
439 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
440 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
445 * Do not go above this order unless 0 objects fit into the slab or
446 * overridden on the command line.
448 #define SLAB_MAX_ORDER_HI 1
449 #define SLAB_MAX_ORDER_LO 0
450 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
451 static bool slab_max_order_set __initdata
;
453 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
455 struct page
*page
= virt_to_head_page(obj
);
456 return page
->slab_cache
;
459 static inline struct slab
*virt_to_slab(const void *obj
)
461 struct page
*page
= virt_to_head_page(obj
);
463 VM_BUG_ON(!PageSlab(page
));
464 return page
->slab_page
;
467 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
470 return slab
->s_mem
+ cache
->size
* idx
;
474 * We want to avoid an expensive divide : (offset / cache->size)
475 * Using the fact that size is a constant for a particular cache,
476 * we can replace (offset / cache->size) by
477 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
479 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
480 const struct slab
*slab
, void *obj
)
482 u32 offset
= (obj
- slab
->s_mem
);
483 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
486 static struct arraycache_init initarray_generic
=
487 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
489 /* internal cache of cache description objs */
490 static struct kmem_cache kmem_cache_boot
= {
492 .limit
= BOOT_CPUCACHE_ENTRIES
,
494 .size
= sizeof(struct kmem_cache
),
495 .name
= "kmem_cache",
498 #define BAD_ALIEN_MAGIC 0x01020304ul
500 #ifdef CONFIG_LOCKDEP
503 * Slab sometimes uses the kmalloc slabs to store the slab headers
504 * for other slabs "off slab".
505 * The locking for this is tricky in that it nests within the locks
506 * of all other slabs in a few places; to deal with this special
507 * locking we put on-slab caches into a separate lock-class.
509 * We set lock class for alien array caches which are up during init.
510 * The lock annotation will be lost if all cpus of a node goes down and
511 * then comes back up during hotplug
513 static struct lock_class_key on_slab_l3_key
;
514 static struct lock_class_key on_slab_alc_key
;
516 static struct lock_class_key debugobj_l3_key
;
517 static struct lock_class_key debugobj_alc_key
;
519 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
520 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
523 struct array_cache
**alc
;
524 struct kmem_cache_node
*n
;
531 lockdep_set_class(&n
->list_lock
, l3_key
);
534 * FIXME: This check for BAD_ALIEN_MAGIC
535 * should go away when common slab code is taught to
536 * work even without alien caches.
537 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
538 * for alloc_alien_cache,
540 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
544 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
548 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
550 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
553 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
557 for_each_online_node(node
)
558 slab_set_debugobj_lock_classes_node(cachep
, node
);
561 static void init_node_lock_keys(int q
)
568 for (i
= 1; i
<= KMALLOC_SHIFT_HIGH
; i
++) {
569 struct kmem_cache_node
*n
;
570 struct kmem_cache
*cache
= kmalloc_caches
[i
];
576 if (!n
|| OFF_SLAB(cache
))
579 slab_set_lock_classes(cache
, &on_slab_l3_key
,
580 &on_slab_alc_key
, q
);
584 static void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int q
)
586 if (!cachep
->node
[q
])
589 slab_set_lock_classes(cachep
, &on_slab_l3_key
,
590 &on_slab_alc_key
, q
);
593 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
597 VM_BUG_ON(OFF_SLAB(cachep
));
599 on_slab_lock_classes_node(cachep
, node
);
602 static inline void init_lock_keys(void)
607 init_node_lock_keys(node
);
610 static void init_node_lock_keys(int q
)
614 static inline void init_lock_keys(void)
618 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
622 static inline void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int node
)
626 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
630 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
635 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
637 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
639 return cachep
->array
[smp_processor_id()];
642 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
644 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(kmem_bufctl_t
), align
);
648 * Calculate the number of objects and left-over bytes for a given buffer size.
650 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
651 size_t align
, int flags
, size_t *left_over
,
656 size_t slab_size
= PAGE_SIZE
<< gfporder
;
659 * The slab management structure can be either off the slab or
660 * on it. For the latter case, the memory allocated for a
664 * - One kmem_bufctl_t for each object
665 * - Padding to respect alignment of @align
666 * - @buffer_size bytes for each object
668 * If the slab management structure is off the slab, then the
669 * alignment will already be calculated into the size. Because
670 * the slabs are all pages aligned, the objects will be at the
671 * correct alignment when allocated.
673 if (flags
& CFLGS_OFF_SLAB
) {
675 nr_objs
= slab_size
/ buffer_size
;
677 if (nr_objs
> SLAB_LIMIT
)
678 nr_objs
= SLAB_LIMIT
;
681 * Ignore padding for the initial guess. The padding
682 * is at most @align-1 bytes, and @buffer_size is at
683 * least @align. In the worst case, this result will
684 * be one greater than the number of objects that fit
685 * into the memory allocation when taking the padding
688 nr_objs
= (slab_size
- sizeof(struct slab
)) /
689 (buffer_size
+ sizeof(kmem_bufctl_t
));
692 * This calculated number will be either the right
693 * amount, or one greater than what we want.
695 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
699 if (nr_objs
> SLAB_LIMIT
)
700 nr_objs
= SLAB_LIMIT
;
702 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
705 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
709 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
711 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
714 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
715 function
, cachep
->name
, msg
);
717 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
722 * By default on NUMA we use alien caches to stage the freeing of
723 * objects allocated from other nodes. This causes massive memory
724 * inefficiencies when using fake NUMA setup to split memory into a
725 * large number of small nodes, so it can be disabled on the command
729 static int use_alien_caches __read_mostly
= 1;
730 static int __init
noaliencache_setup(char *s
)
732 use_alien_caches
= 0;
735 __setup("noaliencache", noaliencache_setup
);
737 static int __init
slab_max_order_setup(char *str
)
739 get_option(&str
, &slab_max_order
);
740 slab_max_order
= slab_max_order
< 0 ? 0 :
741 min(slab_max_order
, MAX_ORDER
- 1);
742 slab_max_order_set
= true;
746 __setup("slab_max_order=", slab_max_order_setup
);
750 * Special reaping functions for NUMA systems called from cache_reap().
751 * These take care of doing round robin flushing of alien caches (containing
752 * objects freed on different nodes from which they were allocated) and the
753 * flushing of remote pcps by calling drain_node_pages.
755 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
757 static void init_reap_node(int cpu
)
761 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
762 if (node
== MAX_NUMNODES
)
763 node
= first_node(node_online_map
);
765 per_cpu(slab_reap_node
, cpu
) = node
;
768 static void next_reap_node(void)
770 int node
= __this_cpu_read(slab_reap_node
);
772 node
= next_node(node
, node_online_map
);
773 if (unlikely(node
>= MAX_NUMNODES
))
774 node
= first_node(node_online_map
);
775 __this_cpu_write(slab_reap_node
, node
);
779 #define init_reap_node(cpu) do { } while (0)
780 #define next_reap_node(void) do { } while (0)
784 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
785 * via the workqueue/eventd.
786 * Add the CPU number into the expiration time to minimize the possibility of
787 * the CPUs getting into lockstep and contending for the global cache chain
790 static void start_cpu_timer(int cpu
)
792 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
795 * When this gets called from do_initcalls via cpucache_init(),
796 * init_workqueues() has already run, so keventd will be setup
799 if (keventd_up() && reap_work
->work
.func
== NULL
) {
801 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
802 schedule_delayed_work_on(cpu
, reap_work
,
803 __round_jiffies_relative(HZ
, cpu
));
807 static struct array_cache
*alloc_arraycache(int node
, int entries
,
808 int batchcount
, gfp_t gfp
)
810 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
811 struct array_cache
*nc
= NULL
;
813 nc
= kmalloc_node(memsize
, gfp
, node
);
815 * The array_cache structures contain pointers to free object.
816 * However, when such objects are allocated or transferred to another
817 * cache the pointers are not cleared and they could be counted as
818 * valid references during a kmemleak scan. Therefore, kmemleak must
819 * not scan such objects.
821 kmemleak_no_scan(nc
);
825 nc
->batchcount
= batchcount
;
827 spin_lock_init(&nc
->lock
);
832 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
834 struct page
*page
= virt_to_page(slabp
->s_mem
);
836 return PageSlabPfmemalloc(page
);
839 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
840 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
841 struct array_cache
*ac
)
843 struct kmem_cache_node
*n
= cachep
->node
[numa_mem_id()];
847 if (!pfmemalloc_active
)
850 spin_lock_irqsave(&n
->list_lock
, flags
);
851 list_for_each_entry(slabp
, &n
->slabs_full
, list
)
852 if (is_slab_pfmemalloc(slabp
))
855 list_for_each_entry(slabp
, &n
->slabs_partial
, list
)
856 if (is_slab_pfmemalloc(slabp
))
859 list_for_each_entry(slabp
, &n
->slabs_free
, list
)
860 if (is_slab_pfmemalloc(slabp
))
863 pfmemalloc_active
= false;
865 spin_unlock_irqrestore(&n
->list_lock
, flags
);
868 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
869 gfp_t flags
, bool force_refill
)
872 void *objp
= ac
->entry
[--ac
->avail
];
874 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
875 if (unlikely(is_obj_pfmemalloc(objp
))) {
876 struct kmem_cache_node
*n
;
878 if (gfp_pfmemalloc_allowed(flags
)) {
879 clear_obj_pfmemalloc(&objp
);
883 /* The caller cannot use PFMEMALLOC objects, find another one */
884 for (i
= 0; i
< ac
->avail
; i
++) {
885 /* If a !PFMEMALLOC object is found, swap them */
886 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
888 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
889 ac
->entry
[ac
->avail
] = objp
;
895 * If there are empty slabs on the slabs_free list and we are
896 * being forced to refill the cache, mark this one !pfmemalloc.
898 n
= cachep
->node
[numa_mem_id()];
899 if (!list_empty(&n
->slabs_free
) && force_refill
) {
900 struct slab
*slabp
= virt_to_slab(objp
);
901 ClearPageSlabPfmemalloc(virt_to_head_page(slabp
->s_mem
));
902 clear_obj_pfmemalloc(&objp
);
903 recheck_pfmemalloc_active(cachep
, ac
);
907 /* No !PFMEMALLOC objects available */
915 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
916 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
920 if (unlikely(sk_memalloc_socks()))
921 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
923 objp
= ac
->entry
[--ac
->avail
];
928 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
931 if (unlikely(pfmemalloc_active
)) {
932 /* Some pfmemalloc slabs exist, check if this is one */
933 struct slab
*slabp
= virt_to_slab(objp
);
934 struct page
*page
= virt_to_head_page(slabp
->s_mem
);
935 if (PageSlabPfmemalloc(page
))
936 set_obj_pfmemalloc(&objp
);
942 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
945 if (unlikely(sk_memalloc_socks()))
946 objp
= __ac_put_obj(cachep
, ac
, objp
);
948 ac
->entry
[ac
->avail
++] = objp
;
952 * Transfer objects in one arraycache to another.
953 * Locking must be handled by the caller.
955 * Return the number of entries transferred.
957 static int transfer_objects(struct array_cache
*to
,
958 struct array_cache
*from
, unsigned int max
)
960 /* Figure out how many entries to transfer */
961 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
966 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
976 #define drain_alien_cache(cachep, alien) do { } while (0)
977 #define reap_alien(cachep, n) do { } while (0)
979 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
981 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
984 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
988 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
993 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
999 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
1000 gfp_t flags
, int nodeid
)
1005 #else /* CONFIG_NUMA */
1007 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
1008 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
1010 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
1012 struct array_cache
**ac_ptr
;
1013 int memsize
= sizeof(void *) * nr_node_ids
;
1018 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1021 if (i
== node
|| !node_online(i
))
1023 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1025 for (i
--; i
>= 0; i
--)
1035 static void free_alien_cache(struct array_cache
**ac_ptr
)
1046 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1047 struct array_cache
*ac
, int node
)
1049 struct kmem_cache_node
*n
= cachep
->node
[node
];
1052 spin_lock(&n
->list_lock
);
1054 * Stuff objects into the remote nodes shared array first.
1055 * That way we could avoid the overhead of putting the objects
1056 * into the free lists and getting them back later.
1059 transfer_objects(n
->shared
, ac
, ac
->limit
);
1061 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1063 spin_unlock(&n
->list_lock
);
1068 * Called from cache_reap() to regularly drain alien caches round robin.
1070 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
1072 int node
= __this_cpu_read(slab_reap_node
);
1075 struct array_cache
*ac
= n
->alien
[node
];
1077 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1078 __drain_alien_cache(cachep
, ac
, node
);
1079 spin_unlock_irq(&ac
->lock
);
1084 static void drain_alien_cache(struct kmem_cache
*cachep
,
1085 struct array_cache
**alien
)
1088 struct array_cache
*ac
;
1089 unsigned long flags
;
1091 for_each_online_node(i
) {
1094 spin_lock_irqsave(&ac
->lock
, flags
);
1095 __drain_alien_cache(cachep
, ac
, i
);
1096 spin_unlock_irqrestore(&ac
->lock
, flags
);
1101 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1103 struct slab
*slabp
= virt_to_slab(objp
);
1104 int nodeid
= slabp
->nodeid
;
1105 struct kmem_cache_node
*n
;
1106 struct array_cache
*alien
= NULL
;
1109 node
= numa_mem_id();
1112 * Make sure we are not freeing a object from another node to the array
1113 * cache on this cpu.
1115 if (likely(slabp
->nodeid
== node
))
1118 n
= cachep
->node
[node
];
1119 STATS_INC_NODEFREES(cachep
);
1120 if (n
->alien
&& n
->alien
[nodeid
]) {
1121 alien
= n
->alien
[nodeid
];
1122 spin_lock(&alien
->lock
);
1123 if (unlikely(alien
->avail
== alien
->limit
)) {
1124 STATS_INC_ACOVERFLOW(cachep
);
1125 __drain_alien_cache(cachep
, alien
, nodeid
);
1127 ac_put_obj(cachep
, alien
, objp
);
1128 spin_unlock(&alien
->lock
);
1130 spin_lock(&(cachep
->node
[nodeid
])->list_lock
);
1131 free_block(cachep
, &objp
, 1, nodeid
);
1132 spin_unlock(&(cachep
->node
[nodeid
])->list_lock
);
1139 * Allocates and initializes node for a node on each slab cache, used for
1140 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1141 * will be allocated off-node since memory is not yet online for the new node.
1142 * When hotplugging memory or a cpu, existing node are not replaced if
1145 * Must hold slab_mutex.
1147 static int init_cache_node_node(int node
)
1149 struct kmem_cache
*cachep
;
1150 struct kmem_cache_node
*n
;
1151 const int memsize
= sizeof(struct kmem_cache_node
);
1153 list_for_each_entry(cachep
, &slab_caches
, list
) {
1155 * Set up the size64 kmemlist for cpu before we can
1156 * begin anything. Make sure some other cpu on this
1157 * node has not already allocated this
1159 if (!cachep
->node
[node
]) {
1160 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1163 kmem_cache_node_init(n
);
1164 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1165 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1168 * The l3s don't come and go as CPUs come and
1169 * go. slab_mutex is sufficient
1172 cachep
->node
[node
] = n
;
1175 spin_lock_irq(&cachep
->node
[node
]->list_lock
);
1176 cachep
->node
[node
]->free_limit
=
1177 (1 + nr_cpus_node(node
)) *
1178 cachep
->batchcount
+ cachep
->num
;
1179 spin_unlock_irq(&cachep
->node
[node
]->list_lock
);
1184 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1185 struct kmem_cache_node
*n
)
1187 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1190 static void cpuup_canceled(long cpu
)
1192 struct kmem_cache
*cachep
;
1193 struct kmem_cache_node
*n
= NULL
;
1194 int node
= cpu_to_mem(cpu
);
1195 const struct cpumask
*mask
= cpumask_of_node(node
);
1197 list_for_each_entry(cachep
, &slab_caches
, list
) {
1198 struct array_cache
*nc
;
1199 struct array_cache
*shared
;
1200 struct array_cache
**alien
;
1202 /* cpu is dead; no one can alloc from it. */
1203 nc
= cachep
->array
[cpu
];
1204 cachep
->array
[cpu
] = NULL
;
1205 n
= cachep
->node
[node
];
1208 goto free_array_cache
;
1210 spin_lock_irq(&n
->list_lock
);
1212 /* Free limit for this kmem_cache_node */
1213 n
->free_limit
-= cachep
->batchcount
;
1215 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1217 if (!cpumask_empty(mask
)) {
1218 spin_unlock_irq(&n
->list_lock
);
1219 goto free_array_cache
;
1224 free_block(cachep
, shared
->entry
,
1225 shared
->avail
, node
);
1232 spin_unlock_irq(&n
->list_lock
);
1236 drain_alien_cache(cachep
, alien
);
1237 free_alien_cache(alien
);
1243 * In the previous loop, all the objects were freed to
1244 * the respective cache's slabs, now we can go ahead and
1245 * shrink each nodelist to its limit.
1247 list_for_each_entry(cachep
, &slab_caches
, list
) {
1248 n
= cachep
->node
[node
];
1251 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1255 static int cpuup_prepare(long cpu
)
1257 struct kmem_cache
*cachep
;
1258 struct kmem_cache_node
*n
= NULL
;
1259 int node
= cpu_to_mem(cpu
);
1263 * We need to do this right in the beginning since
1264 * alloc_arraycache's are going to use this list.
1265 * kmalloc_node allows us to add the slab to the right
1266 * kmem_cache_node and not this cpu's kmem_cache_node
1268 err
= init_cache_node_node(node
);
1273 * Now we can go ahead with allocating the shared arrays and
1276 list_for_each_entry(cachep
, &slab_caches
, list
) {
1277 struct array_cache
*nc
;
1278 struct array_cache
*shared
= NULL
;
1279 struct array_cache
**alien
= NULL
;
1281 nc
= alloc_arraycache(node
, cachep
->limit
,
1282 cachep
->batchcount
, GFP_KERNEL
);
1285 if (cachep
->shared
) {
1286 shared
= alloc_arraycache(node
,
1287 cachep
->shared
* cachep
->batchcount
,
1288 0xbaadf00d, GFP_KERNEL
);
1294 if (use_alien_caches
) {
1295 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1302 cachep
->array
[cpu
] = nc
;
1303 n
= cachep
->node
[node
];
1306 spin_lock_irq(&n
->list_lock
);
1309 * We are serialised from CPU_DEAD or
1310 * CPU_UP_CANCELLED by the cpucontrol lock
1321 spin_unlock_irq(&n
->list_lock
);
1323 free_alien_cache(alien
);
1324 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1325 slab_set_debugobj_lock_classes_node(cachep
, node
);
1326 else if (!OFF_SLAB(cachep
) &&
1327 !(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1328 on_slab_lock_classes_node(cachep
, node
);
1330 init_node_lock_keys(node
);
1334 cpuup_canceled(cpu
);
1338 static int cpuup_callback(struct notifier_block
*nfb
,
1339 unsigned long action
, void *hcpu
)
1341 long cpu
= (long)hcpu
;
1345 case CPU_UP_PREPARE
:
1346 case CPU_UP_PREPARE_FROZEN
:
1347 mutex_lock(&slab_mutex
);
1348 err
= cpuup_prepare(cpu
);
1349 mutex_unlock(&slab_mutex
);
1352 case CPU_ONLINE_FROZEN
:
1353 start_cpu_timer(cpu
);
1355 #ifdef CONFIG_HOTPLUG_CPU
1356 case CPU_DOWN_PREPARE
:
1357 case CPU_DOWN_PREPARE_FROZEN
:
1359 * Shutdown cache reaper. Note that the slab_mutex is
1360 * held so that if cache_reap() is invoked it cannot do
1361 * anything expensive but will only modify reap_work
1362 * and reschedule the timer.
1364 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1365 /* Now the cache_reaper is guaranteed to be not running. */
1366 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1368 case CPU_DOWN_FAILED
:
1369 case CPU_DOWN_FAILED_FROZEN
:
1370 start_cpu_timer(cpu
);
1373 case CPU_DEAD_FROZEN
:
1375 * Even if all the cpus of a node are down, we don't free the
1376 * kmem_cache_node of any cache. This to avoid a race between
1377 * cpu_down, and a kmalloc allocation from another cpu for
1378 * memory from the node of the cpu going down. The node
1379 * structure is usually allocated from kmem_cache_create() and
1380 * gets destroyed at kmem_cache_destroy().
1384 case CPU_UP_CANCELED
:
1385 case CPU_UP_CANCELED_FROZEN
:
1386 mutex_lock(&slab_mutex
);
1387 cpuup_canceled(cpu
);
1388 mutex_unlock(&slab_mutex
);
1391 return notifier_from_errno(err
);
1394 static struct notifier_block cpucache_notifier
= {
1395 &cpuup_callback
, NULL
, 0
1398 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1400 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1401 * Returns -EBUSY if all objects cannot be drained so that the node is not
1404 * Must hold slab_mutex.
1406 static int __meminit
drain_cache_node_node(int node
)
1408 struct kmem_cache
*cachep
;
1411 list_for_each_entry(cachep
, &slab_caches
, list
) {
1412 struct kmem_cache_node
*n
;
1414 n
= cachep
->node
[node
];
1418 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1420 if (!list_empty(&n
->slabs_full
) ||
1421 !list_empty(&n
->slabs_partial
)) {
1429 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1430 unsigned long action
, void *arg
)
1432 struct memory_notify
*mnb
= arg
;
1436 nid
= mnb
->status_change_nid
;
1441 case MEM_GOING_ONLINE
:
1442 mutex_lock(&slab_mutex
);
1443 ret
= init_cache_node_node(nid
);
1444 mutex_unlock(&slab_mutex
);
1446 case MEM_GOING_OFFLINE
:
1447 mutex_lock(&slab_mutex
);
1448 ret
= drain_cache_node_node(nid
);
1449 mutex_unlock(&slab_mutex
);
1453 case MEM_CANCEL_ONLINE
:
1454 case MEM_CANCEL_OFFLINE
:
1458 return notifier_from_errno(ret
);
1460 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1463 * swap the static kmem_cache_node with kmalloced memory
1465 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1468 struct kmem_cache_node
*ptr
;
1470 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1473 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1475 * Do not assume that spinlocks can be initialized via memcpy:
1477 spin_lock_init(&ptr
->list_lock
);
1479 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1480 cachep
->node
[nodeid
] = ptr
;
1484 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1485 * size of kmem_cache_node.
1487 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1491 for_each_online_node(node
) {
1492 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1493 cachep
->node
[node
]->next_reap
= jiffies
+
1495 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1500 * The memory after the last cpu cache pointer is used for the
1503 static void setup_node_pointer(struct kmem_cache
*cachep
)
1505 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1509 * Initialisation. Called after the page allocator have been initialised and
1510 * before smp_init().
1512 void __init
kmem_cache_init(void)
1516 kmem_cache
= &kmem_cache_boot
;
1517 setup_node_pointer(kmem_cache
);
1519 if (num_possible_nodes() == 1)
1520 use_alien_caches
= 0;
1522 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1523 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1525 set_up_node(kmem_cache
, CACHE_CACHE
);
1528 * Fragmentation resistance on low memory - only use bigger
1529 * page orders on machines with more than 32MB of memory if
1530 * not overridden on the command line.
1532 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1533 slab_max_order
= SLAB_MAX_ORDER_HI
;
1535 /* Bootstrap is tricky, because several objects are allocated
1536 * from caches that do not exist yet:
1537 * 1) initialize the kmem_cache cache: it contains the struct
1538 * kmem_cache structures of all caches, except kmem_cache itself:
1539 * kmem_cache is statically allocated.
1540 * Initially an __init data area is used for the head array and the
1541 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1542 * array at the end of the bootstrap.
1543 * 2) Create the first kmalloc cache.
1544 * The struct kmem_cache for the new cache is allocated normally.
1545 * An __init data area is used for the head array.
1546 * 3) Create the remaining kmalloc caches, with minimally sized
1548 * 4) Replace the __init data head arrays for kmem_cache and the first
1549 * kmalloc cache with kmalloc allocated arrays.
1550 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1551 * the other cache's with kmalloc allocated memory.
1552 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1555 /* 1) create the kmem_cache */
1558 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1560 create_boot_cache(kmem_cache
, "kmem_cache",
1561 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1562 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1563 SLAB_HWCACHE_ALIGN
);
1564 list_add(&kmem_cache
->list
, &slab_caches
);
1566 /* 2+3) create the kmalloc caches */
1569 * Initialize the caches that provide memory for the array cache and the
1570 * kmem_cache_node structures first. Without this, further allocations will
1574 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1575 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1577 if (INDEX_AC
!= INDEX_NODE
)
1578 kmalloc_caches
[INDEX_NODE
] =
1579 create_kmalloc_cache("kmalloc-node",
1580 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1582 slab_early_init
= 0;
1584 /* 4) Replace the bootstrap head arrays */
1586 struct array_cache
*ptr
;
1588 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1590 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1591 sizeof(struct arraycache_init
));
1593 * Do not assume that spinlocks can be initialized via memcpy:
1595 spin_lock_init(&ptr
->lock
);
1597 kmem_cache
->array
[smp_processor_id()] = ptr
;
1599 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1601 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1602 != &initarray_generic
.cache
);
1603 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1604 sizeof(struct arraycache_init
));
1606 * Do not assume that spinlocks can be initialized via memcpy:
1608 spin_lock_init(&ptr
->lock
);
1610 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1612 /* 5) Replace the bootstrap kmem_cache_node */
1616 for_each_online_node(nid
) {
1617 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1619 init_list(kmalloc_caches
[INDEX_AC
],
1620 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1622 if (INDEX_AC
!= INDEX_NODE
) {
1623 init_list(kmalloc_caches
[INDEX_NODE
],
1624 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1629 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1632 void __init
kmem_cache_init_late(void)
1634 struct kmem_cache
*cachep
;
1638 /* 6) resize the head arrays to their final sizes */
1639 mutex_lock(&slab_mutex
);
1640 list_for_each_entry(cachep
, &slab_caches
, list
)
1641 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1643 mutex_unlock(&slab_mutex
);
1645 /* Annotate slab for lockdep -- annotate the malloc caches */
1652 * Register a cpu startup notifier callback that initializes
1653 * cpu_cache_get for all new cpus
1655 register_cpu_notifier(&cpucache_notifier
);
1659 * Register a memory hotplug callback that initializes and frees
1662 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1666 * The reap timers are started later, with a module init call: That part
1667 * of the kernel is not yet operational.
1671 static int __init
cpucache_init(void)
1676 * Register the timers that return unneeded pages to the page allocator
1678 for_each_online_cpu(cpu
)
1679 start_cpu_timer(cpu
);
1685 __initcall(cpucache_init
);
1687 static noinline
void
1688 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1690 struct kmem_cache_node
*n
;
1692 unsigned long flags
;
1696 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1698 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1699 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1701 for_each_online_node(node
) {
1702 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1703 unsigned long active_slabs
= 0, num_slabs
= 0;
1705 n
= cachep
->node
[node
];
1709 spin_lock_irqsave(&n
->list_lock
, flags
);
1710 list_for_each_entry(slabp
, &n
->slabs_full
, list
) {
1711 active_objs
+= cachep
->num
;
1714 list_for_each_entry(slabp
, &n
->slabs_partial
, list
) {
1715 active_objs
+= slabp
->inuse
;
1718 list_for_each_entry(slabp
, &n
->slabs_free
, list
)
1721 free_objects
+= n
->free_objects
;
1722 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1724 num_slabs
+= active_slabs
;
1725 num_objs
= num_slabs
* cachep
->num
;
1727 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1728 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1734 * Interface to system's page allocator. No need to hold the cache-lock.
1736 * If we requested dmaable memory, we will get it. Even if we
1737 * did not request dmaable memory, we might get it, but that
1738 * would be relatively rare and ignorable.
1740 static void *kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
1748 * Nommu uses slab's for process anonymous memory allocations, and thus
1749 * requires __GFP_COMP to properly refcount higher order allocations
1751 flags
|= __GFP_COMP
;
1754 flags
|= cachep
->allocflags
;
1755 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1756 flags
|= __GFP_RECLAIMABLE
;
1758 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1760 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1761 slab_out_of_memory(cachep
, flags
, nodeid
);
1765 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1766 if (unlikely(page
->pfmemalloc
))
1767 pfmemalloc_active
= true;
1769 nr_pages
= (1 << cachep
->gfporder
);
1770 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1771 add_zone_page_state(page_zone(page
),
1772 NR_SLAB_RECLAIMABLE
, nr_pages
);
1774 add_zone_page_state(page_zone(page
),
1775 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1776 for (i
= 0; i
< nr_pages
; i
++) {
1777 __SetPageSlab(page
+ i
);
1779 if (page
->pfmemalloc
)
1780 SetPageSlabPfmemalloc(page
);
1782 memcg_bind_pages(cachep
, cachep
->gfporder
);
1784 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1785 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1788 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1790 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1793 return page_address(page
);
1797 * Interface to system's page release.
1799 static void kmem_freepages(struct kmem_cache
*cachep
, void *addr
)
1801 unsigned long i
= (1 << cachep
->gfporder
);
1802 struct page
*page
= virt_to_page(addr
);
1803 const unsigned long nr_freed
= i
;
1805 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1807 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1808 sub_zone_page_state(page_zone(page
),
1809 NR_SLAB_RECLAIMABLE
, nr_freed
);
1811 sub_zone_page_state(page_zone(page
),
1812 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1814 __ClearPageSlabPfmemalloc(page
);
1816 BUG_ON(!PageSlab(page
));
1817 __ClearPageSlab(page
);
1821 memcg_release_pages(cachep
, cachep
->gfporder
);
1822 if (current
->reclaim_state
)
1823 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1824 free_memcg_kmem_pages((unsigned long)addr
, cachep
->gfporder
);
1827 static void kmem_rcu_free(struct rcu_head
*head
)
1829 struct slab_rcu
*slab_rcu
= (struct slab_rcu
*)head
;
1830 struct kmem_cache
*cachep
= slab_rcu
->cachep
;
1832 kmem_freepages(cachep
, slab_rcu
->addr
);
1833 if (OFF_SLAB(cachep
))
1834 kmem_cache_free(cachep
->slabp_cache
, slab_rcu
);
1839 #ifdef CONFIG_DEBUG_PAGEALLOC
1840 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1841 unsigned long caller
)
1843 int size
= cachep
->object_size
;
1845 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1847 if (size
< 5 * sizeof(unsigned long))
1850 *addr
++ = 0x12345678;
1852 *addr
++ = smp_processor_id();
1853 size
-= 3 * sizeof(unsigned long);
1855 unsigned long *sptr
= &caller
;
1856 unsigned long svalue
;
1858 while (!kstack_end(sptr
)) {
1860 if (kernel_text_address(svalue
)) {
1862 size
-= sizeof(unsigned long);
1863 if (size
<= sizeof(unsigned long))
1869 *addr
++ = 0x87654321;
1873 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1875 int size
= cachep
->object_size
;
1876 addr
= &((char *)addr
)[obj_offset(cachep
)];
1878 memset(addr
, val
, size
);
1879 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1882 static void dump_line(char *data
, int offset
, int limit
)
1885 unsigned char error
= 0;
1888 printk(KERN_ERR
"%03x: ", offset
);
1889 for (i
= 0; i
< limit
; i
++) {
1890 if (data
[offset
+ i
] != POISON_FREE
) {
1891 error
= data
[offset
+ i
];
1895 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1896 &data
[offset
], limit
, 1);
1898 if (bad_count
== 1) {
1899 error
^= POISON_FREE
;
1900 if (!(error
& (error
- 1))) {
1901 printk(KERN_ERR
"Single bit error detected. Probably "
1904 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1907 printk(KERN_ERR
"Run a memory test tool.\n");
1916 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1921 if (cachep
->flags
& SLAB_RED_ZONE
) {
1922 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1923 *dbg_redzone1(cachep
, objp
),
1924 *dbg_redzone2(cachep
, objp
));
1927 if (cachep
->flags
& SLAB_STORE_USER
) {
1928 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1929 *dbg_userword(cachep
, objp
),
1930 *dbg_userword(cachep
, objp
));
1932 realobj
= (char *)objp
+ obj_offset(cachep
);
1933 size
= cachep
->object_size
;
1934 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1937 if (i
+ limit
> size
)
1939 dump_line(realobj
, i
, limit
);
1943 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1949 realobj
= (char *)objp
+ obj_offset(cachep
);
1950 size
= cachep
->object_size
;
1952 for (i
= 0; i
< size
; i
++) {
1953 char exp
= POISON_FREE
;
1956 if (realobj
[i
] != exp
) {
1962 "Slab corruption (%s): %s start=%p, len=%d\n",
1963 print_tainted(), cachep
->name
, realobj
, size
);
1964 print_objinfo(cachep
, objp
, 0);
1966 /* Hexdump the affected line */
1969 if (i
+ limit
> size
)
1971 dump_line(realobj
, i
, limit
);
1974 /* Limit to 5 lines */
1980 /* Print some data about the neighboring objects, if they
1983 struct slab
*slabp
= virt_to_slab(objp
);
1986 objnr
= obj_to_index(cachep
, slabp
, objp
);
1988 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1989 realobj
= (char *)objp
+ obj_offset(cachep
);
1990 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1992 print_objinfo(cachep
, objp
, 2);
1994 if (objnr
+ 1 < cachep
->num
) {
1995 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1996 realobj
= (char *)objp
+ obj_offset(cachep
);
1997 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1999 print_objinfo(cachep
, objp
, 2);
2006 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2009 for (i
= 0; i
< cachep
->num
; i
++) {
2010 void *objp
= index_to_obj(cachep
, slabp
, i
);
2012 if (cachep
->flags
& SLAB_POISON
) {
2013 #ifdef CONFIG_DEBUG_PAGEALLOC
2014 if (cachep
->size
% PAGE_SIZE
== 0 &&
2016 kernel_map_pages(virt_to_page(objp
),
2017 cachep
->size
/ PAGE_SIZE
, 1);
2019 check_poison_obj(cachep
, objp
);
2021 check_poison_obj(cachep
, objp
);
2024 if (cachep
->flags
& SLAB_RED_ZONE
) {
2025 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2026 slab_error(cachep
, "start of a freed object "
2028 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2029 slab_error(cachep
, "end of a freed object "
2035 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
2041 * slab_destroy - destroy and release all objects in a slab
2042 * @cachep: cache pointer being destroyed
2043 * @slabp: slab pointer being destroyed
2045 * Destroy all the objs in a slab, and release the mem back to the system.
2046 * Before calling the slab must have been unlinked from the cache. The
2047 * cache-lock is not held/needed.
2049 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
2051 void *addr
= slabp
->s_mem
- slabp
->colouroff
;
2053 slab_destroy_debugcheck(cachep
, slabp
);
2054 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2055 struct slab_rcu
*slab_rcu
;
2057 slab_rcu
= (struct slab_rcu
*)slabp
;
2058 slab_rcu
->cachep
= cachep
;
2059 slab_rcu
->addr
= addr
;
2060 call_rcu(&slab_rcu
->head
, kmem_rcu_free
);
2062 kmem_freepages(cachep
, addr
);
2063 if (OFF_SLAB(cachep
))
2064 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2069 * calculate_slab_order - calculate size (page order) of slabs
2070 * @cachep: pointer to the cache that is being created
2071 * @size: size of objects to be created in this cache.
2072 * @align: required alignment for the objects.
2073 * @flags: slab allocation flags
2075 * Also calculates the number of objects per slab.
2077 * This could be made much more intelligent. For now, try to avoid using
2078 * high order pages for slabs. When the gfp() functions are more friendly
2079 * towards high-order requests, this should be changed.
2081 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2082 size_t size
, size_t align
, unsigned long flags
)
2084 unsigned long offslab_limit
;
2085 size_t left_over
= 0;
2088 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2092 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2096 if (flags
& CFLGS_OFF_SLAB
) {
2098 * Max number of objs-per-slab for caches which
2099 * use off-slab slabs. Needed to avoid a possible
2100 * looping condition in cache_grow().
2102 offslab_limit
= size
- sizeof(struct slab
);
2103 offslab_limit
/= sizeof(kmem_bufctl_t
);
2105 if (num
> offslab_limit
)
2109 /* Found something acceptable - save it away */
2111 cachep
->gfporder
= gfporder
;
2112 left_over
= remainder
;
2115 * A VFS-reclaimable slab tends to have most allocations
2116 * as GFP_NOFS and we really don't want to have to be allocating
2117 * higher-order pages when we are unable to shrink dcache.
2119 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2123 * Large number of objects is good, but very large slabs are
2124 * currently bad for the gfp()s.
2126 if (gfporder
>= slab_max_order
)
2130 * Acceptable internal fragmentation?
2132 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2138 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2140 if (slab_state
>= FULL
)
2141 return enable_cpucache(cachep
, gfp
);
2143 if (slab_state
== DOWN
) {
2145 * Note: Creation of first cache (kmem_cache).
2146 * The setup_node is taken care
2147 * of by the caller of __kmem_cache_create
2149 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2150 slab_state
= PARTIAL
;
2151 } else if (slab_state
== PARTIAL
) {
2153 * Note: the second kmem_cache_create must create the cache
2154 * that's used by kmalloc(24), otherwise the creation of
2155 * further caches will BUG().
2157 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2160 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2161 * the second cache, then we need to set up all its node/,
2162 * otherwise the creation of further caches will BUG().
2164 set_up_node(cachep
, SIZE_AC
);
2165 if (INDEX_AC
== INDEX_NODE
)
2166 slab_state
= PARTIAL_NODE
;
2168 slab_state
= PARTIAL_ARRAYCACHE
;
2170 /* Remaining boot caches */
2171 cachep
->array
[smp_processor_id()] =
2172 kmalloc(sizeof(struct arraycache_init
), gfp
);
2174 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2175 set_up_node(cachep
, SIZE_NODE
);
2176 slab_state
= PARTIAL_NODE
;
2179 for_each_online_node(node
) {
2180 cachep
->node
[node
] =
2181 kmalloc_node(sizeof(struct kmem_cache_node
),
2183 BUG_ON(!cachep
->node
[node
]);
2184 kmem_cache_node_init(cachep
->node
[node
]);
2188 cachep
->node
[numa_mem_id()]->next_reap
=
2189 jiffies
+ REAPTIMEOUT_LIST3
+
2190 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2192 cpu_cache_get(cachep
)->avail
= 0;
2193 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2194 cpu_cache_get(cachep
)->batchcount
= 1;
2195 cpu_cache_get(cachep
)->touched
= 0;
2196 cachep
->batchcount
= 1;
2197 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2202 * __kmem_cache_create - Create a cache.
2203 * @cachep: cache management descriptor
2204 * @flags: SLAB flags
2206 * Returns a ptr to the cache on success, NULL on failure.
2207 * Cannot be called within a int, but can be interrupted.
2208 * The @ctor is run when new pages are allocated by the cache.
2212 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2213 * to catch references to uninitialised memory.
2215 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2216 * for buffer overruns.
2218 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2219 * cacheline. This can be beneficial if you're counting cycles as closely
2223 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2225 size_t left_over
, slab_size
, ralign
;
2228 size_t size
= cachep
->size
;
2233 * Enable redzoning and last user accounting, except for caches with
2234 * large objects, if the increased size would increase the object size
2235 * above the next power of two: caches with object sizes just above a
2236 * power of two have a significant amount of internal fragmentation.
2238 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2239 2 * sizeof(unsigned long long)))
2240 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2241 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2242 flags
|= SLAB_POISON
;
2244 if (flags
& SLAB_DESTROY_BY_RCU
)
2245 BUG_ON(flags
& SLAB_POISON
);
2249 * Check that size is in terms of words. This is needed to avoid
2250 * unaligned accesses for some archs when redzoning is used, and makes
2251 * sure any on-slab bufctl's are also correctly aligned.
2253 if (size
& (BYTES_PER_WORD
- 1)) {
2254 size
+= (BYTES_PER_WORD
- 1);
2255 size
&= ~(BYTES_PER_WORD
- 1);
2259 * Redzoning and user store require word alignment or possibly larger.
2260 * Note this will be overridden by architecture or caller mandated
2261 * alignment if either is greater than BYTES_PER_WORD.
2263 if (flags
& SLAB_STORE_USER
)
2264 ralign
= BYTES_PER_WORD
;
2266 if (flags
& SLAB_RED_ZONE
) {
2267 ralign
= REDZONE_ALIGN
;
2268 /* If redzoning, ensure that the second redzone is suitably
2269 * aligned, by adjusting the object size accordingly. */
2270 size
+= REDZONE_ALIGN
- 1;
2271 size
&= ~(REDZONE_ALIGN
- 1);
2274 /* 3) caller mandated alignment */
2275 if (ralign
< cachep
->align
) {
2276 ralign
= cachep
->align
;
2278 /* disable debug if necessary */
2279 if (ralign
> __alignof__(unsigned long long))
2280 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2284 cachep
->align
= ralign
;
2286 if (slab_is_available())
2291 setup_node_pointer(cachep
);
2295 * Both debugging options require word-alignment which is calculated
2298 if (flags
& SLAB_RED_ZONE
) {
2299 /* add space for red zone words */
2300 cachep
->obj_offset
+= sizeof(unsigned long long);
2301 size
+= 2 * sizeof(unsigned long long);
2303 if (flags
& SLAB_STORE_USER
) {
2304 /* user store requires one word storage behind the end of
2305 * the real object. But if the second red zone needs to be
2306 * aligned to 64 bits, we must allow that much space.
2308 if (flags
& SLAB_RED_ZONE
)
2309 size
+= REDZONE_ALIGN
;
2311 size
+= BYTES_PER_WORD
;
2313 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2314 if (size
>= kmalloc_size(INDEX_NODE
+ 1)
2315 && cachep
->object_size
> cache_line_size()
2316 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2317 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2324 * Determine if the slab management is 'on' or 'off' slab.
2325 * (bootstrapping cannot cope with offslab caches so don't do
2326 * it too early on. Always use on-slab management when
2327 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2329 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2330 !(flags
& SLAB_NOLEAKTRACE
))
2332 * Size is large, assume best to place the slab management obj
2333 * off-slab (should allow better packing of objs).
2335 flags
|= CFLGS_OFF_SLAB
;
2337 size
= ALIGN(size
, cachep
->align
);
2339 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2344 slab_size
= ALIGN(cachep
->num
* sizeof(kmem_bufctl_t
)
2345 + sizeof(struct slab
), cachep
->align
);
2348 * If the slab has been placed off-slab, and we have enough space then
2349 * move it on-slab. This is at the expense of any extra colouring.
2351 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2352 flags
&= ~CFLGS_OFF_SLAB
;
2353 left_over
-= slab_size
;
2356 if (flags
& CFLGS_OFF_SLAB
) {
2357 /* really off slab. No need for manual alignment */
2359 cachep
->num
* sizeof(kmem_bufctl_t
) + sizeof(struct slab
);
2361 #ifdef CONFIG_PAGE_POISONING
2362 /* If we're going to use the generic kernel_map_pages()
2363 * poisoning, then it's going to smash the contents of
2364 * the redzone and userword anyhow, so switch them off.
2366 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2367 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2371 cachep
->colour_off
= cache_line_size();
2372 /* Offset must be a multiple of the alignment. */
2373 if (cachep
->colour_off
< cachep
->align
)
2374 cachep
->colour_off
= cachep
->align
;
2375 cachep
->colour
= left_over
/ cachep
->colour_off
;
2376 cachep
->slab_size
= slab_size
;
2377 cachep
->flags
= flags
;
2378 cachep
->allocflags
= 0;
2379 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2380 cachep
->allocflags
|= GFP_DMA
;
2381 cachep
->size
= size
;
2382 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2384 if (flags
& CFLGS_OFF_SLAB
) {
2385 cachep
->slabp_cache
= kmalloc_slab(slab_size
, 0u);
2387 * This is a possibility for one of the malloc_sizes caches.
2388 * But since we go off slab only for object size greater than
2389 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2390 * this should not happen at all.
2391 * But leave a BUG_ON for some lucky dude.
2393 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2396 err
= setup_cpu_cache(cachep
, gfp
);
2398 __kmem_cache_shutdown(cachep
);
2402 if (flags
& SLAB_DEBUG_OBJECTS
) {
2404 * Would deadlock through slab_destroy()->call_rcu()->
2405 * debug_object_activate()->kmem_cache_alloc().
2407 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2409 slab_set_debugobj_lock_classes(cachep
);
2410 } else if (!OFF_SLAB(cachep
) && !(flags
& SLAB_DESTROY_BY_RCU
))
2411 on_slab_lock_classes(cachep
);
2417 static void check_irq_off(void)
2419 BUG_ON(!irqs_disabled());
2422 static void check_irq_on(void)
2424 BUG_ON(irqs_disabled());
2427 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2431 assert_spin_locked(&cachep
->node
[numa_mem_id()]->list_lock
);
2435 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2439 assert_spin_locked(&cachep
->node
[node
]->list_lock
);
2444 #define check_irq_off() do { } while(0)
2445 #define check_irq_on() do { } while(0)
2446 #define check_spinlock_acquired(x) do { } while(0)
2447 #define check_spinlock_acquired_node(x, y) do { } while(0)
2450 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2451 struct array_cache
*ac
,
2452 int force
, int node
);
2454 static void do_drain(void *arg
)
2456 struct kmem_cache
*cachep
= arg
;
2457 struct array_cache
*ac
;
2458 int node
= numa_mem_id();
2461 ac
= cpu_cache_get(cachep
);
2462 spin_lock(&cachep
->node
[node
]->list_lock
);
2463 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2464 spin_unlock(&cachep
->node
[node
]->list_lock
);
2468 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2470 struct kmem_cache_node
*n
;
2473 on_each_cpu(do_drain
, cachep
, 1);
2475 for_each_online_node(node
) {
2476 n
= cachep
->node
[node
];
2478 drain_alien_cache(cachep
, n
->alien
);
2481 for_each_online_node(node
) {
2482 n
= cachep
->node
[node
];
2484 drain_array(cachep
, n
, n
->shared
, 1, node
);
2489 * Remove slabs from the list of free slabs.
2490 * Specify the number of slabs to drain in tofree.
2492 * Returns the actual number of slabs released.
2494 static int drain_freelist(struct kmem_cache
*cache
,
2495 struct kmem_cache_node
*n
, int tofree
)
2497 struct list_head
*p
;
2502 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2504 spin_lock_irq(&n
->list_lock
);
2505 p
= n
->slabs_free
.prev
;
2506 if (p
== &n
->slabs_free
) {
2507 spin_unlock_irq(&n
->list_lock
);
2511 slabp
= list_entry(p
, struct slab
, list
);
2513 BUG_ON(slabp
->inuse
);
2515 list_del(&slabp
->list
);
2517 * Safe to drop the lock. The slab is no longer linked
2520 n
->free_objects
-= cache
->num
;
2521 spin_unlock_irq(&n
->list_lock
);
2522 slab_destroy(cache
, slabp
);
2529 /* Called with slab_mutex held to protect against cpu hotplug */
2530 static int __cache_shrink(struct kmem_cache
*cachep
)
2533 struct kmem_cache_node
*n
;
2535 drain_cpu_caches(cachep
);
2538 for_each_online_node(i
) {
2539 n
= cachep
->node
[i
];
2543 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2545 ret
+= !list_empty(&n
->slabs_full
) ||
2546 !list_empty(&n
->slabs_partial
);
2548 return (ret
? 1 : 0);
2552 * kmem_cache_shrink - Shrink a cache.
2553 * @cachep: The cache to shrink.
2555 * Releases as many slabs as possible for a cache.
2556 * To help debugging, a zero exit status indicates all slabs were released.
2558 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2561 BUG_ON(!cachep
|| in_interrupt());
2564 mutex_lock(&slab_mutex
);
2565 ret
= __cache_shrink(cachep
);
2566 mutex_unlock(&slab_mutex
);
2570 EXPORT_SYMBOL(kmem_cache_shrink
);
2572 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2575 struct kmem_cache_node
*n
;
2576 int rc
= __cache_shrink(cachep
);
2581 for_each_online_cpu(i
)
2582 kfree(cachep
->array
[i
]);
2584 /* NUMA: free the node structures */
2585 for_each_online_node(i
) {
2586 n
= cachep
->node
[i
];
2589 free_alien_cache(n
->alien
);
2597 * Get the memory for a slab management obj.
2598 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2599 * always come from malloc_sizes caches. The slab descriptor cannot
2600 * come from the same cache which is getting created because,
2601 * when we are searching for an appropriate cache for these
2602 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2603 * If we are creating a malloc_sizes cache here it would not be visible to
2604 * kmem_find_general_cachep till the initialization is complete.
2605 * Hence we cannot have slabp_cache same as the original cache.
2607 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
, void *objp
,
2608 int colour_off
, gfp_t local_flags
,
2613 if (OFF_SLAB(cachep
)) {
2614 /* Slab management obj is off-slab. */
2615 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2616 local_flags
, nodeid
);
2618 * If the first object in the slab is leaked (it's allocated
2619 * but no one has a reference to it), we want to make sure
2620 * kmemleak does not treat the ->s_mem pointer as a reference
2621 * to the object. Otherwise we will not report the leak.
2623 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2628 slabp
= objp
+ colour_off
;
2629 colour_off
+= cachep
->slab_size
;
2632 slabp
->colouroff
= colour_off
;
2633 slabp
->s_mem
= objp
+ colour_off
;
2634 slabp
->nodeid
= nodeid
;
2639 static inline kmem_bufctl_t
*slab_bufctl(struct slab
*slabp
)
2641 return (kmem_bufctl_t
*) (slabp
+ 1);
2644 static void cache_init_objs(struct kmem_cache
*cachep
,
2649 for (i
= 0; i
< cachep
->num
; i
++) {
2650 void *objp
= index_to_obj(cachep
, slabp
, i
);
2652 /* need to poison the objs? */
2653 if (cachep
->flags
& SLAB_POISON
)
2654 poison_obj(cachep
, objp
, POISON_FREE
);
2655 if (cachep
->flags
& SLAB_STORE_USER
)
2656 *dbg_userword(cachep
, objp
) = NULL
;
2658 if (cachep
->flags
& SLAB_RED_ZONE
) {
2659 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2660 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2663 * Constructors are not allowed to allocate memory from the same
2664 * cache which they are a constructor for. Otherwise, deadlock.
2665 * They must also be threaded.
2667 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2668 cachep
->ctor(objp
+ obj_offset(cachep
));
2670 if (cachep
->flags
& SLAB_RED_ZONE
) {
2671 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2672 slab_error(cachep
, "constructor overwrote the"
2673 " end of an object");
2674 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2675 slab_error(cachep
, "constructor overwrote the"
2676 " start of an object");
2678 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2679 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2680 kernel_map_pages(virt_to_page(objp
),
2681 cachep
->size
/ PAGE_SIZE
, 0);
2686 slab_bufctl(slabp
)[i
] = i
+ 1;
2688 slab_bufctl(slabp
)[i
- 1] = BUFCTL_END
;
2691 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2693 if (CONFIG_ZONE_DMA_FLAG
) {
2694 if (flags
& GFP_DMA
)
2695 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2697 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2701 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2704 void *objp
= index_to_obj(cachep
, slabp
, slabp
->free
);
2708 next
= slab_bufctl(slabp
)[slabp
->free
];
2710 slab_bufctl(slabp
)[slabp
->free
] = BUFCTL_FREE
;
2711 WARN_ON(slabp
->nodeid
!= nodeid
);
2718 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2719 void *objp
, int nodeid
)
2721 unsigned int objnr
= obj_to_index(cachep
, slabp
, objp
);
2724 /* Verify that the slab belongs to the intended node */
2725 WARN_ON(slabp
->nodeid
!= nodeid
);
2727 if (slab_bufctl(slabp
)[objnr
] + 1 <= SLAB_LIMIT
+ 1) {
2728 printk(KERN_ERR
"slab: double free detected in cache "
2729 "'%s', objp %p\n", cachep
->name
, objp
);
2733 slab_bufctl(slabp
)[objnr
] = slabp
->free
;
2734 slabp
->free
= objnr
;
2739 * Map pages beginning at addr to the given cache and slab. This is required
2740 * for the slab allocator to be able to lookup the cache and slab of a
2741 * virtual address for kfree, ksize, and slab debugging.
2743 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2749 page
= virt_to_page(addr
);
2752 if (likely(!PageCompound(page
)))
2753 nr_pages
<<= cache
->gfporder
;
2756 page
->slab_cache
= cache
;
2757 page
->slab_page
= slab
;
2759 } while (--nr_pages
);
2763 * Grow (by 1) the number of slabs within a cache. This is called by
2764 * kmem_cache_alloc() when there are no active objs left in a cache.
2766 static int cache_grow(struct kmem_cache
*cachep
,
2767 gfp_t flags
, int nodeid
, void *objp
)
2772 struct kmem_cache_node
*n
;
2775 * Be lazy and only check for valid flags here, keeping it out of the
2776 * critical path in kmem_cache_alloc().
2778 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2779 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2781 /* Take the node list lock to change the colour_next on this node */
2783 n
= cachep
->node
[nodeid
];
2784 spin_lock(&n
->list_lock
);
2786 /* Get colour for the slab, and cal the next value. */
2787 offset
= n
->colour_next
;
2789 if (n
->colour_next
>= cachep
->colour
)
2791 spin_unlock(&n
->list_lock
);
2793 offset
*= cachep
->colour_off
;
2795 if (local_flags
& __GFP_WAIT
)
2799 * The test for missing atomic flag is performed here, rather than
2800 * the more obvious place, simply to reduce the critical path length
2801 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2802 * will eventually be caught here (where it matters).
2804 kmem_flagcheck(cachep
, flags
);
2807 * Get mem for the objs. Attempt to allocate a physical page from
2811 objp
= kmem_getpages(cachep
, local_flags
, nodeid
);
2815 /* Get slab management. */
2816 slabp
= alloc_slabmgmt(cachep
, objp
, offset
,
2817 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2821 slab_map_pages(cachep
, slabp
, objp
);
2823 cache_init_objs(cachep
, slabp
);
2825 if (local_flags
& __GFP_WAIT
)
2826 local_irq_disable();
2828 spin_lock(&n
->list_lock
);
2830 /* Make slab active. */
2831 list_add_tail(&slabp
->list
, &(n
->slabs_free
));
2832 STATS_INC_GROWN(cachep
);
2833 n
->free_objects
+= cachep
->num
;
2834 spin_unlock(&n
->list_lock
);
2837 kmem_freepages(cachep
, objp
);
2839 if (local_flags
& __GFP_WAIT
)
2840 local_irq_disable();
2847 * Perform extra freeing checks:
2848 * - detect bad pointers.
2849 * - POISON/RED_ZONE checking
2851 static void kfree_debugcheck(const void *objp
)
2853 if (!virt_addr_valid(objp
)) {
2854 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2855 (unsigned long)objp
);
2860 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2862 unsigned long long redzone1
, redzone2
;
2864 redzone1
= *dbg_redzone1(cache
, obj
);
2865 redzone2
= *dbg_redzone2(cache
, obj
);
2870 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2873 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2874 slab_error(cache
, "double free detected");
2876 slab_error(cache
, "memory outside object was overwritten");
2878 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2879 obj
, redzone1
, redzone2
);
2882 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2883 unsigned long caller
)
2889 BUG_ON(virt_to_cache(objp
) != cachep
);
2891 objp
-= obj_offset(cachep
);
2892 kfree_debugcheck(objp
);
2893 page
= virt_to_head_page(objp
);
2895 slabp
= page
->slab_page
;
2897 if (cachep
->flags
& SLAB_RED_ZONE
) {
2898 verify_redzone_free(cachep
, objp
);
2899 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2900 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2902 if (cachep
->flags
& SLAB_STORE_USER
)
2903 *dbg_userword(cachep
, objp
) = (void *)caller
;
2905 objnr
= obj_to_index(cachep
, slabp
, objp
);
2907 BUG_ON(objnr
>= cachep
->num
);
2908 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2910 #ifdef CONFIG_DEBUG_SLAB_LEAK
2911 slab_bufctl(slabp
)[objnr
] = BUFCTL_FREE
;
2913 if (cachep
->flags
& SLAB_POISON
) {
2914 #ifdef CONFIG_DEBUG_PAGEALLOC
2915 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2916 store_stackinfo(cachep
, objp
, caller
);
2917 kernel_map_pages(virt_to_page(objp
),
2918 cachep
->size
/ PAGE_SIZE
, 0);
2920 poison_obj(cachep
, objp
, POISON_FREE
);
2923 poison_obj(cachep
, objp
, POISON_FREE
);
2929 static void check_slabp(struct kmem_cache
*cachep
, struct slab
*slabp
)
2934 /* Check slab's freelist to see if this obj is there. */
2935 for (i
= slabp
->free
; i
!= BUFCTL_END
; i
= slab_bufctl(slabp
)[i
]) {
2937 if (entries
> cachep
->num
|| i
>= cachep
->num
)
2940 if (entries
!= cachep
->num
- slabp
->inuse
) {
2942 printk(KERN_ERR
"slab: Internal list corruption detected in "
2943 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2944 cachep
->name
, cachep
->num
, slabp
, slabp
->inuse
,
2946 print_hex_dump(KERN_ERR
, "", DUMP_PREFIX_OFFSET
, 16, 1, slabp
,
2947 sizeof(*slabp
) + cachep
->num
* sizeof(kmem_bufctl_t
),
2953 #define kfree_debugcheck(x) do { } while(0)
2954 #define cache_free_debugcheck(x,objp,z) (objp)
2955 #define check_slabp(x,y) do { } while(0)
2958 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2962 struct kmem_cache_node
*n
;
2963 struct array_cache
*ac
;
2967 node
= numa_mem_id();
2968 if (unlikely(force_refill
))
2971 ac
= cpu_cache_get(cachep
);
2972 batchcount
= ac
->batchcount
;
2973 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2975 * If there was little recent activity on this cache, then
2976 * perform only a partial refill. Otherwise we could generate
2979 batchcount
= BATCHREFILL_LIMIT
;
2981 n
= cachep
->node
[node
];
2983 BUG_ON(ac
->avail
> 0 || !n
);
2984 spin_lock(&n
->list_lock
);
2986 /* See if we can refill from the shared array */
2987 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2988 n
->shared
->touched
= 1;
2992 while (batchcount
> 0) {
2993 struct list_head
*entry
;
2995 /* Get slab alloc is to come from. */
2996 entry
= n
->slabs_partial
.next
;
2997 if (entry
== &n
->slabs_partial
) {
2998 n
->free_touched
= 1;
2999 entry
= n
->slabs_free
.next
;
3000 if (entry
== &n
->slabs_free
)
3004 slabp
= list_entry(entry
, struct slab
, list
);
3005 check_slabp(cachep
, slabp
);
3006 check_spinlock_acquired(cachep
);
3009 * The slab was either on partial or free list so
3010 * there must be at least one object available for
3013 BUG_ON(slabp
->inuse
>= cachep
->num
);
3015 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
3016 STATS_INC_ALLOCED(cachep
);
3017 STATS_INC_ACTIVE(cachep
);
3018 STATS_SET_HIGH(cachep
);
3020 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
3023 check_slabp(cachep
, slabp
);
3025 /* move slabp to correct slabp list: */
3026 list_del(&slabp
->list
);
3027 if (slabp
->free
== BUFCTL_END
)
3028 list_add(&slabp
->list
, &n
->slabs_full
);
3030 list_add(&slabp
->list
, &n
->slabs_partial
);
3034 n
->free_objects
-= ac
->avail
;
3036 spin_unlock(&n
->list_lock
);
3038 if (unlikely(!ac
->avail
)) {
3041 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
3043 /* cache_grow can reenable interrupts, then ac could change. */
3044 ac
= cpu_cache_get(cachep
);
3045 node
= numa_mem_id();
3047 /* no objects in sight? abort */
3048 if (!x
&& (ac
->avail
== 0 || force_refill
))
3051 if (!ac
->avail
) /* objects refilled by interrupt? */
3056 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
3059 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3062 might_sleep_if(flags
& __GFP_WAIT
);
3064 kmem_flagcheck(cachep
, flags
);
3069 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3070 gfp_t flags
, void *objp
, unsigned long caller
)
3074 if (cachep
->flags
& SLAB_POISON
) {
3075 #ifdef CONFIG_DEBUG_PAGEALLOC
3076 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3077 kernel_map_pages(virt_to_page(objp
),
3078 cachep
->size
/ PAGE_SIZE
, 1);
3080 check_poison_obj(cachep
, objp
);
3082 check_poison_obj(cachep
, objp
);
3084 poison_obj(cachep
, objp
, POISON_INUSE
);
3086 if (cachep
->flags
& SLAB_STORE_USER
)
3087 *dbg_userword(cachep
, objp
) = (void *)caller
;
3089 if (cachep
->flags
& SLAB_RED_ZONE
) {
3090 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3091 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3092 slab_error(cachep
, "double free, or memory outside"
3093 " object was overwritten");
3095 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3096 objp
, *dbg_redzone1(cachep
, objp
),
3097 *dbg_redzone2(cachep
, objp
));
3099 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3100 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3102 #ifdef CONFIG_DEBUG_SLAB_LEAK
3107 slabp
= virt_to_head_page(objp
)->slab_page
;
3108 objnr
= (unsigned)(objp
- slabp
->s_mem
) / cachep
->size
;
3109 slab_bufctl(slabp
)[objnr
] = BUFCTL_ACTIVE
;
3112 objp
+= obj_offset(cachep
);
3113 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3115 if (ARCH_SLAB_MINALIGN
&&
3116 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3117 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3118 objp
, (int)ARCH_SLAB_MINALIGN
);
3123 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3126 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3128 if (cachep
== kmem_cache
)
3131 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3134 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3137 struct array_cache
*ac
;
3138 bool force_refill
= false;
3142 ac
= cpu_cache_get(cachep
);
3143 if (likely(ac
->avail
)) {
3145 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3148 * Allow for the possibility all avail objects are not allowed
3149 * by the current flags
3152 STATS_INC_ALLOCHIT(cachep
);
3155 force_refill
= true;
3158 STATS_INC_ALLOCMISS(cachep
);
3159 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3161 * the 'ac' may be updated by cache_alloc_refill(),
3162 * and kmemleak_erase() requires its correct value.
3164 ac
= cpu_cache_get(cachep
);
3168 * To avoid a false negative, if an object that is in one of the
3169 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3170 * treat the array pointers as a reference to the object.
3173 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3179 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3181 * If we are in_interrupt, then process context, including cpusets and
3182 * mempolicy, may not apply and should not be used for allocation policy.
3184 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3186 int nid_alloc
, nid_here
;
3188 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3190 nid_alloc
= nid_here
= numa_mem_id();
3191 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3192 nid_alloc
= cpuset_slab_spread_node();
3193 else if (current
->mempolicy
)
3194 nid_alloc
= slab_node();
3195 if (nid_alloc
!= nid_here
)
3196 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3201 * Fallback function if there was no memory available and no objects on a
3202 * certain node and fall back is permitted. First we scan all the
3203 * available node for available objects. If that fails then we
3204 * perform an allocation without specifying a node. This allows the page
3205 * allocator to do its reclaim / fallback magic. We then insert the
3206 * slab into the proper nodelist and then allocate from it.
3208 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3210 struct zonelist
*zonelist
;
3214 enum zone_type high_zoneidx
= gfp_zone(flags
);
3217 unsigned int cpuset_mems_cookie
;
3219 if (flags
& __GFP_THISNODE
)
3222 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3225 cpuset_mems_cookie
= get_mems_allowed();
3226 zonelist
= node_zonelist(slab_node(), flags
);
3230 * Look through allowed nodes for objects available
3231 * from existing per node queues.
3233 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3234 nid
= zone_to_nid(zone
);
3236 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3238 cache
->node
[nid
]->free_objects
) {
3239 obj
= ____cache_alloc_node(cache
,
3240 flags
| GFP_THISNODE
, nid
);
3248 * This allocation will be performed within the constraints
3249 * of the current cpuset / memory policy requirements.
3250 * We may trigger various forms of reclaim on the allowed
3251 * set and go into memory reserves if necessary.
3253 if (local_flags
& __GFP_WAIT
)
3255 kmem_flagcheck(cache
, flags
);
3256 obj
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3257 if (local_flags
& __GFP_WAIT
)
3258 local_irq_disable();
3261 * Insert into the appropriate per node queues
3263 nid
= page_to_nid(virt_to_page(obj
));
3264 if (cache_grow(cache
, flags
, nid
, obj
)) {
3265 obj
= ____cache_alloc_node(cache
,
3266 flags
| GFP_THISNODE
, nid
);
3269 * Another processor may allocate the
3270 * objects in the slab since we are
3271 * not holding any locks.
3275 /* cache_grow already freed obj */
3281 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3287 * A interface to enable slab creation on nodeid
3289 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3292 struct list_head
*entry
;
3294 struct kmem_cache_node
*n
;
3298 VM_BUG_ON(nodeid
> num_online_nodes());
3299 n
= cachep
->node
[nodeid
];
3304 spin_lock(&n
->list_lock
);
3305 entry
= n
->slabs_partial
.next
;
3306 if (entry
== &n
->slabs_partial
) {
3307 n
->free_touched
= 1;
3308 entry
= n
->slabs_free
.next
;
3309 if (entry
== &n
->slabs_free
)
3313 slabp
= list_entry(entry
, struct slab
, list
);
3314 check_spinlock_acquired_node(cachep
, nodeid
);
3315 check_slabp(cachep
, slabp
);
3317 STATS_INC_NODEALLOCS(cachep
);
3318 STATS_INC_ACTIVE(cachep
);
3319 STATS_SET_HIGH(cachep
);
3321 BUG_ON(slabp
->inuse
== cachep
->num
);
3323 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3324 check_slabp(cachep
, slabp
);
3326 /* move slabp to correct slabp list: */
3327 list_del(&slabp
->list
);
3329 if (slabp
->free
== BUFCTL_END
)
3330 list_add(&slabp
->list
, &n
->slabs_full
);
3332 list_add(&slabp
->list
, &n
->slabs_partial
);
3334 spin_unlock(&n
->list_lock
);
3338 spin_unlock(&n
->list_lock
);
3339 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3343 return fallback_alloc(cachep
, flags
);
3349 static __always_inline
void *
3350 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3351 unsigned long caller
)
3353 unsigned long save_flags
;
3355 int slab_node
= numa_mem_id();
3357 flags
&= gfp_allowed_mask
;
3359 lockdep_trace_alloc(flags
);
3361 if (slab_should_failslab(cachep
, flags
))
3364 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3366 cache_alloc_debugcheck_before(cachep
, flags
);
3367 local_irq_save(save_flags
);
3369 if (nodeid
== NUMA_NO_NODE
)
3372 if (unlikely(!cachep
->node
[nodeid
])) {
3373 /* Node not bootstrapped yet */
3374 ptr
= fallback_alloc(cachep
, flags
);
3378 if (nodeid
== slab_node
) {
3380 * Use the locally cached objects if possible.
3381 * However ____cache_alloc does not allow fallback
3382 * to other nodes. It may fail while we still have
3383 * objects on other nodes available.
3385 ptr
= ____cache_alloc(cachep
, flags
);
3389 /* ___cache_alloc_node can fall back to other nodes */
3390 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3392 local_irq_restore(save_flags
);
3393 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3394 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3398 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3400 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3401 memset(ptr
, 0, cachep
->object_size
);
3406 static __always_inline
void *
3407 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3411 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3412 objp
= alternate_node_alloc(cache
, flags
);
3416 objp
= ____cache_alloc(cache
, flags
);
3419 * We may just have run out of memory on the local node.
3420 * ____cache_alloc_node() knows how to locate memory on other nodes
3423 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3430 static __always_inline
void *
3431 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3433 return ____cache_alloc(cachep
, flags
);
3436 #endif /* CONFIG_NUMA */
3438 static __always_inline
void *
3439 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3441 unsigned long save_flags
;
3444 flags
&= gfp_allowed_mask
;
3446 lockdep_trace_alloc(flags
);
3448 if (slab_should_failslab(cachep
, flags
))
3451 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3453 cache_alloc_debugcheck_before(cachep
, flags
);
3454 local_irq_save(save_flags
);
3455 objp
= __do_cache_alloc(cachep
, flags
);
3456 local_irq_restore(save_flags
);
3457 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3458 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3463 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3465 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3466 memset(objp
, 0, cachep
->object_size
);
3472 * Caller needs to acquire correct kmem_list's list_lock
3474 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3478 struct kmem_cache_node
*n
;
3480 for (i
= 0; i
< nr_objects
; i
++) {
3484 clear_obj_pfmemalloc(&objpp
[i
]);
3487 slabp
= virt_to_slab(objp
);
3488 n
= cachep
->node
[node
];
3489 list_del(&slabp
->list
);
3490 check_spinlock_acquired_node(cachep
, node
);
3491 check_slabp(cachep
, slabp
);
3492 slab_put_obj(cachep
, slabp
, objp
, node
);
3493 STATS_DEC_ACTIVE(cachep
);
3495 check_slabp(cachep
, slabp
);
3497 /* fixup slab chains */
3498 if (slabp
->inuse
== 0) {
3499 if (n
->free_objects
> n
->free_limit
) {
3500 n
->free_objects
-= cachep
->num
;
3501 /* No need to drop any previously held
3502 * lock here, even if we have a off-slab slab
3503 * descriptor it is guaranteed to come from
3504 * a different cache, refer to comments before
3507 slab_destroy(cachep
, slabp
);
3509 list_add(&slabp
->list
, &n
->slabs_free
);
3512 /* Unconditionally move a slab to the end of the
3513 * partial list on free - maximum time for the
3514 * other objects to be freed, too.
3516 list_add_tail(&slabp
->list
, &n
->slabs_partial
);
3521 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3524 struct kmem_cache_node
*n
;
3525 int node
= numa_mem_id();
3527 batchcount
= ac
->batchcount
;
3529 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3532 n
= cachep
->node
[node
];
3533 spin_lock(&n
->list_lock
);
3535 struct array_cache
*shared_array
= n
->shared
;
3536 int max
= shared_array
->limit
- shared_array
->avail
;
3538 if (batchcount
> max
)
3540 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3541 ac
->entry
, sizeof(void *) * batchcount
);
3542 shared_array
->avail
+= batchcount
;
3547 free_block(cachep
, ac
->entry
, batchcount
, node
);
3552 struct list_head
*p
;
3554 p
= n
->slabs_free
.next
;
3555 while (p
!= &(n
->slabs_free
)) {
3558 slabp
= list_entry(p
, struct slab
, list
);
3559 BUG_ON(slabp
->inuse
);
3564 STATS_SET_FREEABLE(cachep
, i
);
3567 spin_unlock(&n
->list_lock
);
3568 ac
->avail
-= batchcount
;
3569 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3573 * Release an obj back to its cache. If the obj has a constructed state, it must
3574 * be in this state _before_ it is released. Called with disabled ints.
3576 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3577 unsigned long caller
)
3579 struct array_cache
*ac
= cpu_cache_get(cachep
);
3582 kmemleak_free_recursive(objp
, cachep
->flags
);
3583 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3585 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3588 * Skip calling cache_free_alien() when the platform is not numa.
3589 * This will avoid cache misses that happen while accessing slabp (which
3590 * is per page memory reference) to get nodeid. Instead use a global
3591 * variable to skip the call, which is mostly likely to be present in
3594 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3597 if (likely(ac
->avail
< ac
->limit
)) {
3598 STATS_INC_FREEHIT(cachep
);
3600 STATS_INC_FREEMISS(cachep
);
3601 cache_flusharray(cachep
, ac
);
3604 ac_put_obj(cachep
, ac
, objp
);
3608 * kmem_cache_alloc - Allocate an object
3609 * @cachep: The cache to allocate from.
3610 * @flags: See kmalloc().
3612 * Allocate an object from this cache. The flags are only relevant
3613 * if the cache has no available objects.
3615 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3617 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3619 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3620 cachep
->object_size
, cachep
->size
, flags
);
3624 EXPORT_SYMBOL(kmem_cache_alloc
);
3626 #ifdef CONFIG_TRACING
3628 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3632 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3634 trace_kmalloc(_RET_IP_
, ret
,
3635 size
, cachep
->size
, flags
);
3638 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3643 * kmem_cache_alloc_node - Allocate an object on the specified node
3644 * @cachep: The cache to allocate from.
3645 * @flags: See kmalloc().
3646 * @nodeid: node number of the target node.
3648 * Identical to kmem_cache_alloc but it will allocate memory on the given
3649 * node, which can improve the performance for cpu bound structures.
3651 * Fallback to other node is possible if __GFP_THISNODE is not set.
3653 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3655 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3657 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3658 cachep
->object_size
, cachep
->size
,
3663 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3665 #ifdef CONFIG_TRACING
3666 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3673 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3675 trace_kmalloc_node(_RET_IP_
, ret
,
3680 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3683 static __always_inline
void *
3684 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3686 struct kmem_cache
*cachep
;
3688 cachep
= kmalloc_slab(size
, flags
);
3689 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3691 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3694 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3695 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3697 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3699 EXPORT_SYMBOL(__kmalloc_node
);
3701 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3702 int node
, unsigned long caller
)
3704 return __do_kmalloc_node(size
, flags
, node
, caller
);
3706 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3708 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3710 return __do_kmalloc_node(size
, flags
, node
, 0);
3712 EXPORT_SYMBOL(__kmalloc_node
);
3713 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3714 #endif /* CONFIG_NUMA */
3717 * __do_kmalloc - allocate memory
3718 * @size: how many bytes of memory are required.
3719 * @flags: the type of memory to allocate (see kmalloc).
3720 * @caller: function caller for debug tracking of the caller
3722 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3723 unsigned long caller
)
3725 struct kmem_cache
*cachep
;
3728 /* If you want to save a few bytes .text space: replace
3730 * Then kmalloc uses the uninlined functions instead of the inline
3733 cachep
= kmalloc_slab(size
, flags
);
3734 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3736 ret
= slab_alloc(cachep
, flags
, caller
);
3738 trace_kmalloc(caller
, ret
,
3739 size
, cachep
->size
, flags
);
3745 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3746 void *__kmalloc(size_t size
, gfp_t flags
)
3748 return __do_kmalloc(size
, flags
, _RET_IP_
);
3750 EXPORT_SYMBOL(__kmalloc
);
3752 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3754 return __do_kmalloc(size
, flags
, caller
);
3756 EXPORT_SYMBOL(__kmalloc_track_caller
);
3759 void *__kmalloc(size_t size
, gfp_t flags
)
3761 return __do_kmalloc(size
, flags
, 0);
3763 EXPORT_SYMBOL(__kmalloc
);
3767 * kmem_cache_free - Deallocate an object
3768 * @cachep: The cache the allocation was from.
3769 * @objp: The previously allocated object.
3771 * Free an object which was previously allocated from this
3774 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3776 unsigned long flags
;
3777 cachep
= cache_from_obj(cachep
, objp
);
3781 local_irq_save(flags
);
3782 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3783 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3784 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3785 __cache_free(cachep
, objp
, _RET_IP_
);
3786 local_irq_restore(flags
);
3788 trace_kmem_cache_free(_RET_IP_
, objp
);
3790 EXPORT_SYMBOL(kmem_cache_free
);
3793 * kfree - free previously allocated memory
3794 * @objp: pointer returned by kmalloc.
3796 * If @objp is NULL, no operation is performed.
3798 * Don't free memory not originally allocated by kmalloc()
3799 * or you will run into trouble.
3801 void kfree(const void *objp
)
3803 struct kmem_cache
*c
;
3804 unsigned long flags
;
3806 trace_kfree(_RET_IP_
, objp
);
3808 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3810 local_irq_save(flags
);
3811 kfree_debugcheck(objp
);
3812 c
= virt_to_cache(objp
);
3813 debug_check_no_locks_freed(objp
, c
->object_size
);
3815 debug_check_no_obj_freed(objp
, c
->object_size
);
3816 __cache_free(c
, (void *)objp
, _RET_IP_
);
3817 local_irq_restore(flags
);
3819 EXPORT_SYMBOL(kfree
);
3822 * This initializes kmem_cache_node or resizes various caches for all nodes.
3824 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3827 struct kmem_cache_node
*n
;
3828 struct array_cache
*new_shared
;
3829 struct array_cache
**new_alien
= NULL
;
3831 for_each_online_node(node
) {
3833 if (use_alien_caches
) {
3834 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3840 if (cachep
->shared
) {
3841 new_shared
= alloc_arraycache(node
,
3842 cachep
->shared
*cachep
->batchcount
,
3845 free_alien_cache(new_alien
);
3850 n
= cachep
->node
[node
];
3852 struct array_cache
*shared
= n
->shared
;
3854 spin_lock_irq(&n
->list_lock
);
3857 free_block(cachep
, shared
->entry
,
3858 shared
->avail
, node
);
3860 n
->shared
= new_shared
;
3862 n
->alien
= new_alien
;
3865 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3866 cachep
->batchcount
+ cachep
->num
;
3867 spin_unlock_irq(&n
->list_lock
);
3869 free_alien_cache(new_alien
);
3872 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3874 free_alien_cache(new_alien
);
3879 kmem_cache_node_init(n
);
3880 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3881 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3882 n
->shared
= new_shared
;
3883 n
->alien
= new_alien
;
3884 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3885 cachep
->batchcount
+ cachep
->num
;
3886 cachep
->node
[node
] = n
;
3891 if (!cachep
->list
.next
) {
3892 /* Cache is not active yet. Roll back what we did */
3895 if (cachep
->node
[node
]) {
3896 n
= cachep
->node
[node
];
3899 free_alien_cache(n
->alien
);
3901 cachep
->node
[node
] = NULL
;
3909 struct ccupdate_struct
{
3910 struct kmem_cache
*cachep
;
3911 struct array_cache
*new[0];
3914 static void do_ccupdate_local(void *info
)
3916 struct ccupdate_struct
*new = info
;
3917 struct array_cache
*old
;
3920 old
= cpu_cache_get(new->cachep
);
3922 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3923 new->new[smp_processor_id()] = old
;
3926 /* Always called with the slab_mutex held */
3927 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3928 int batchcount
, int shared
, gfp_t gfp
)
3930 struct ccupdate_struct
*new;
3933 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3938 for_each_online_cpu(i
) {
3939 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3942 for (i
--; i
>= 0; i
--)
3948 new->cachep
= cachep
;
3950 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3953 cachep
->batchcount
= batchcount
;
3954 cachep
->limit
= limit
;
3955 cachep
->shared
= shared
;
3957 for_each_online_cpu(i
) {
3958 struct array_cache
*ccold
= new->new[i
];
3961 spin_lock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3962 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
3963 spin_unlock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3967 return alloc_kmemlist(cachep
, gfp
);
3970 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3971 int batchcount
, int shared
, gfp_t gfp
)
3974 struct kmem_cache
*c
= NULL
;
3977 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3979 if (slab_state
< FULL
)
3982 if ((ret
< 0) || !is_root_cache(cachep
))
3985 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3986 for_each_memcg_cache_index(i
) {
3987 c
= cache_from_memcg(cachep
, i
);
3989 /* return value determined by the parent cache only */
3990 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3996 /* Called with slab_mutex held always */
3997 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
4004 if (!is_root_cache(cachep
)) {
4005 struct kmem_cache
*root
= memcg_root_cache(cachep
);
4006 limit
= root
->limit
;
4007 shared
= root
->shared
;
4008 batchcount
= root
->batchcount
;
4011 if (limit
&& shared
&& batchcount
)
4014 * The head array serves three purposes:
4015 * - create a LIFO ordering, i.e. return objects that are cache-warm
4016 * - reduce the number of spinlock operations.
4017 * - reduce the number of linked list operations on the slab and
4018 * bufctl chains: array operations are cheaper.
4019 * The numbers are guessed, we should auto-tune as described by
4022 if (cachep
->size
> 131072)
4024 else if (cachep
->size
> PAGE_SIZE
)
4026 else if (cachep
->size
> 1024)
4028 else if (cachep
->size
> 256)
4034 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4035 * allocation behaviour: Most allocs on one cpu, most free operations
4036 * on another cpu. For these cases, an efficient object passing between
4037 * cpus is necessary. This is provided by a shared array. The array
4038 * replaces Bonwick's magazine layer.
4039 * On uniprocessor, it's functionally equivalent (but less efficient)
4040 * to a larger limit. Thus disabled by default.
4043 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
4048 * With debugging enabled, large batchcount lead to excessively long
4049 * periods with disabled local interrupts. Limit the batchcount
4054 batchcount
= (limit
+ 1) / 2;
4056 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
4058 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
4059 cachep
->name
, -err
);
4064 * Drain an array if it contains any elements taking the node lock only if
4065 * necessary. Note that the node listlock also protects the array_cache
4066 * if drain_array() is used on the shared array.
4068 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
4069 struct array_cache
*ac
, int force
, int node
)
4073 if (!ac
|| !ac
->avail
)
4075 if (ac
->touched
&& !force
) {
4078 spin_lock_irq(&n
->list_lock
);
4080 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4081 if (tofree
> ac
->avail
)
4082 tofree
= (ac
->avail
+ 1) / 2;
4083 free_block(cachep
, ac
->entry
, tofree
, node
);
4084 ac
->avail
-= tofree
;
4085 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4086 sizeof(void *) * ac
->avail
);
4088 spin_unlock_irq(&n
->list_lock
);
4093 * cache_reap - Reclaim memory from caches.
4094 * @w: work descriptor
4096 * Called from workqueue/eventd every few seconds.
4098 * - clear the per-cpu caches for this CPU.
4099 * - return freeable pages to the main free memory pool.
4101 * If we cannot acquire the cache chain mutex then just give up - we'll try
4102 * again on the next iteration.
4104 static void cache_reap(struct work_struct
*w
)
4106 struct kmem_cache
*searchp
;
4107 struct kmem_cache_node
*n
;
4108 int node
= numa_mem_id();
4109 struct delayed_work
*work
= to_delayed_work(w
);
4111 if (!mutex_trylock(&slab_mutex
))
4112 /* Give up. Setup the next iteration. */
4115 list_for_each_entry(searchp
, &slab_caches
, list
) {
4119 * We only take the node lock if absolutely necessary and we
4120 * have established with reasonable certainty that
4121 * we can do some work if the lock was obtained.
4123 n
= searchp
->node
[node
];
4125 reap_alien(searchp
, n
);
4127 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
4130 * These are racy checks but it does not matter
4131 * if we skip one check or scan twice.
4133 if (time_after(n
->next_reap
, jiffies
))
4136 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4138 drain_array(searchp
, n
, n
->shared
, 0, node
);
4140 if (n
->free_touched
)
4141 n
->free_touched
= 0;
4145 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4146 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4147 STATS_ADD_REAPED(searchp
, freed
);
4153 mutex_unlock(&slab_mutex
);
4156 /* Set up the next iteration */
4157 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4160 #ifdef CONFIG_SLABINFO
4161 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4164 unsigned long active_objs
;
4165 unsigned long num_objs
;
4166 unsigned long active_slabs
= 0;
4167 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4171 struct kmem_cache_node
*n
;
4175 for_each_online_node(node
) {
4176 n
= cachep
->node
[node
];
4181 spin_lock_irq(&n
->list_lock
);
4183 list_for_each_entry(slabp
, &n
->slabs_full
, list
) {
4184 if (slabp
->inuse
!= cachep
->num
&& !error
)
4185 error
= "slabs_full accounting error";
4186 active_objs
+= cachep
->num
;
4189 list_for_each_entry(slabp
, &n
->slabs_partial
, list
) {
4190 if (slabp
->inuse
== cachep
->num
&& !error
)
4191 error
= "slabs_partial inuse accounting error";
4192 if (!slabp
->inuse
&& !error
)
4193 error
= "slabs_partial/inuse accounting error";
4194 active_objs
+= slabp
->inuse
;
4197 list_for_each_entry(slabp
, &n
->slabs_free
, list
) {
4198 if (slabp
->inuse
&& !error
)
4199 error
= "slabs_free/inuse accounting error";
4202 free_objects
+= n
->free_objects
;
4204 shared_avail
+= n
->shared
->avail
;
4206 spin_unlock_irq(&n
->list_lock
);
4208 num_slabs
+= active_slabs
;
4209 num_objs
= num_slabs
* cachep
->num
;
4210 if (num_objs
- active_objs
!= free_objects
&& !error
)
4211 error
= "free_objects accounting error";
4213 name
= cachep
->name
;
4215 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4217 sinfo
->active_objs
= active_objs
;
4218 sinfo
->num_objs
= num_objs
;
4219 sinfo
->active_slabs
= active_slabs
;
4220 sinfo
->num_slabs
= num_slabs
;
4221 sinfo
->shared_avail
= shared_avail
;
4222 sinfo
->limit
= cachep
->limit
;
4223 sinfo
->batchcount
= cachep
->batchcount
;
4224 sinfo
->shared
= cachep
->shared
;
4225 sinfo
->objects_per_slab
= cachep
->num
;
4226 sinfo
->cache_order
= cachep
->gfporder
;
4229 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4233 unsigned long high
= cachep
->high_mark
;
4234 unsigned long allocs
= cachep
->num_allocations
;
4235 unsigned long grown
= cachep
->grown
;
4236 unsigned long reaped
= cachep
->reaped
;
4237 unsigned long errors
= cachep
->errors
;
4238 unsigned long max_freeable
= cachep
->max_freeable
;
4239 unsigned long node_allocs
= cachep
->node_allocs
;
4240 unsigned long node_frees
= cachep
->node_frees
;
4241 unsigned long overflows
= cachep
->node_overflow
;
4243 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4244 "%4lu %4lu %4lu %4lu %4lu",
4245 allocs
, high
, grown
,
4246 reaped
, errors
, max_freeable
, node_allocs
,
4247 node_frees
, overflows
);
4251 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4252 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4253 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4254 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4256 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4257 allochit
, allocmiss
, freehit
, freemiss
);
4262 #define MAX_SLABINFO_WRITE 128
4264 * slabinfo_write - Tuning for the slab allocator
4266 * @buffer: user buffer
4267 * @count: data length
4270 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4271 size_t count
, loff_t
*ppos
)
4273 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4274 int limit
, batchcount
, shared
, res
;
4275 struct kmem_cache
*cachep
;
4277 if (count
> MAX_SLABINFO_WRITE
)
4279 if (copy_from_user(&kbuf
, buffer
, count
))
4281 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4283 tmp
= strchr(kbuf
, ' ');
4288 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4291 /* Find the cache in the chain of caches. */
4292 mutex_lock(&slab_mutex
);
4294 list_for_each_entry(cachep
, &slab_caches
, list
) {
4295 if (!strcmp(cachep
->name
, kbuf
)) {
4296 if (limit
< 1 || batchcount
< 1 ||
4297 batchcount
> limit
|| shared
< 0) {
4300 res
= do_tune_cpucache(cachep
, limit
,
4307 mutex_unlock(&slab_mutex
);
4313 #ifdef CONFIG_DEBUG_SLAB_LEAK
4315 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4317 mutex_lock(&slab_mutex
);
4318 return seq_list_start(&slab_caches
, *pos
);
4321 static inline int add_caller(unsigned long *n
, unsigned long v
)
4331 unsigned long *q
= p
+ 2 * i
;
4345 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4351 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4357 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4358 if (slab_bufctl(s
)[i
] != BUFCTL_ACTIVE
)
4360 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4365 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4367 #ifdef CONFIG_KALLSYMS
4368 unsigned long offset
, size
;
4369 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4371 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4372 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4374 seq_printf(m
, " [%s]", modname
);
4378 seq_printf(m
, "%p", (void *)address
);
4381 static int leaks_show(struct seq_file
*m
, void *p
)
4383 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4385 struct kmem_cache_node
*n
;
4387 unsigned long *x
= m
->private;
4391 if (!(cachep
->flags
& SLAB_STORE_USER
))
4393 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4396 /* OK, we can do it */
4400 for_each_online_node(node
) {
4401 n
= cachep
->node
[node
];
4406 spin_lock_irq(&n
->list_lock
);
4408 list_for_each_entry(slabp
, &n
->slabs_full
, list
)
4409 handle_slab(x
, cachep
, slabp
);
4410 list_for_each_entry(slabp
, &n
->slabs_partial
, list
)
4411 handle_slab(x
, cachep
, slabp
);
4412 spin_unlock_irq(&n
->list_lock
);
4414 name
= cachep
->name
;
4416 /* Increase the buffer size */
4417 mutex_unlock(&slab_mutex
);
4418 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4420 /* Too bad, we are really out */
4422 mutex_lock(&slab_mutex
);
4425 *(unsigned long *)m
->private = x
[0] * 2;
4427 mutex_lock(&slab_mutex
);
4428 /* Now make sure this entry will be retried */
4432 for (i
= 0; i
< x
[1]; i
++) {
4433 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4434 show_symbol(m
, x
[2*i
+2]);
4441 static const struct seq_operations slabstats_op
= {
4442 .start
= leaks_start
,
4448 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4450 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4453 ret
= seq_open(file
, &slabstats_op
);
4455 struct seq_file
*m
= file
->private_data
;
4456 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4465 static const struct file_operations proc_slabstats_operations
= {
4466 .open
= slabstats_open
,
4468 .llseek
= seq_lseek
,
4469 .release
= seq_release_private
,
4473 static int __init
slab_proc_init(void)
4475 #ifdef CONFIG_DEBUG_SLAB_LEAK
4476 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4480 module_init(slab_proc_init
);
4484 * ksize - get the actual amount of memory allocated for a given object
4485 * @objp: Pointer to the object
4487 * kmalloc may internally round up allocations and return more memory
4488 * than requested. ksize() can be used to determine the actual amount of
4489 * memory allocated. The caller may use this additional memory, even though
4490 * a smaller amount of memory was initially specified with the kmalloc call.
4491 * The caller must guarantee that objp points to a valid object previously
4492 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4493 * must not be freed during the duration of the call.
4495 size_t ksize(const void *objp
)
4498 if (unlikely(objp
== ZERO_SIZE_PTR
))
4501 return virt_to_cache(objp
)->object_size
;
4503 EXPORT_SYMBOL(ksize
);