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
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t
;
166 typedef unsigned short freelist_idx_t
;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
172 * true if a page was allocated from pfmemalloc reserves for network-based
175 static bool pfmemalloc_active __read_mostly
;
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
185 * The limit is stored in the per-cpu structure to reduce the data cache
192 unsigned int batchcount
;
193 unsigned int touched
;
196 * Must have this definition in here for the proper
197 * alignment of array_cache. Also simplifies accessing
200 * Entries should not be directly dereferenced as
201 * entries belonging to slabs marked pfmemalloc will
202 * have the lower bits set SLAB_OBJ_PFMEMALLOC
206 #define SLAB_OBJ_PFMEMALLOC 1
207 static inline bool is_obj_pfmemalloc(void *objp
)
209 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
212 static inline void set_obj_pfmemalloc(void **objp
)
214 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
218 static inline void clear_obj_pfmemalloc(void **objp
)
220 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
224 * bootstrap: The caches do not work without cpuarrays anymore, but the
225 * cpuarrays are allocated from the generic caches...
227 #define BOOT_CPUCACHE_ENTRIES 1
228 struct arraycache_init
{
229 struct array_cache cache
;
230 void *entries
[BOOT_CPUCACHE_ENTRIES
];
234 * Need this for bootstrapping a per node allocator.
236 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
237 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
238 #define CACHE_CACHE 0
239 #define SIZE_AC MAX_NUMNODES
240 #define SIZE_NODE (2 * MAX_NUMNODES)
242 static int drain_freelist(struct kmem_cache
*cache
,
243 struct kmem_cache_node
*n
, int tofree
);
244 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
246 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
247 static void cache_reap(struct work_struct
*unused
);
249 static int slab_early_init
= 1;
251 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
252 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
254 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
256 INIT_LIST_HEAD(&parent
->slabs_full
);
257 INIT_LIST_HEAD(&parent
->slabs_partial
);
258 INIT_LIST_HEAD(&parent
->slabs_free
);
259 parent
->shared
= NULL
;
260 parent
->alien
= NULL
;
261 parent
->colour_next
= 0;
262 spin_lock_init(&parent
->list_lock
);
263 parent
->free_objects
= 0;
264 parent
->free_touched
= 0;
267 #define MAKE_LIST(cachep, listp, slab, nodeid) \
269 INIT_LIST_HEAD(listp); \
270 list_splice(&(cachep->node[nodeid]->slab), listp); \
273 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
275 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
276 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
277 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
280 #define CFLGS_OFF_SLAB (0x80000000UL)
281 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
283 #define BATCHREFILL_LIMIT 16
285 * Optimization question: fewer reaps means less probability for unnessary
286 * cpucache drain/refill cycles.
288 * OTOH the cpuarrays can contain lots of objects,
289 * which could lock up otherwise freeable slabs.
291 #define REAPTIMEOUT_AC (2*HZ)
292 #define REAPTIMEOUT_NODE (4*HZ)
295 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
296 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
297 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
298 #define STATS_INC_GROWN(x) ((x)->grown++)
299 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
300 #define STATS_SET_HIGH(x) \
302 if ((x)->num_active > (x)->high_mark) \
303 (x)->high_mark = (x)->num_active; \
305 #define STATS_INC_ERR(x) ((x)->errors++)
306 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
307 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
308 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
309 #define STATS_SET_FREEABLE(x, i) \
311 if ((x)->max_freeable < i) \
312 (x)->max_freeable = i; \
314 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
315 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
316 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
317 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
319 #define STATS_INC_ACTIVE(x) do { } while (0)
320 #define STATS_DEC_ACTIVE(x) do { } while (0)
321 #define STATS_INC_ALLOCED(x) do { } while (0)
322 #define STATS_INC_GROWN(x) do { } while (0)
323 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
324 #define STATS_SET_HIGH(x) do { } while (0)
325 #define STATS_INC_ERR(x) do { } while (0)
326 #define STATS_INC_NODEALLOCS(x) do { } while (0)
327 #define STATS_INC_NODEFREES(x) do { } while (0)
328 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
329 #define STATS_SET_FREEABLE(x, i) do { } while (0)
330 #define STATS_INC_ALLOCHIT(x) do { } while (0)
331 #define STATS_INC_ALLOCMISS(x) do { } while (0)
332 #define STATS_INC_FREEHIT(x) do { } while (0)
333 #define STATS_INC_FREEMISS(x) do { } while (0)
339 * memory layout of objects:
341 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
342 * the end of an object is aligned with the end of the real
343 * allocation. Catches writes behind the end of the allocation.
344 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
346 * cachep->obj_offset: The real object.
347 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
348 * cachep->size - 1* BYTES_PER_WORD: last caller address
349 * [BYTES_PER_WORD long]
351 static int obj_offset(struct kmem_cache
*cachep
)
353 return cachep
->obj_offset
;
356 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
358 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
359 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
360 sizeof(unsigned long long));
363 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
365 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
366 if (cachep
->flags
& SLAB_STORE_USER
)
367 return (unsigned long long *)(objp
+ cachep
->size
-
368 sizeof(unsigned long long) -
370 return (unsigned long long *) (objp
+ cachep
->size
-
371 sizeof(unsigned long long));
374 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
376 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
377 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
382 #define obj_offset(x) 0
383 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
384 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
385 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
389 #define OBJECT_FREE (0)
390 #define OBJECT_ACTIVE (1)
392 #ifdef CONFIG_DEBUG_SLAB_LEAK
394 static void set_obj_status(struct page
*page
, int idx
, int val
)
398 struct kmem_cache
*cachep
= page
->slab_cache
;
400 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
401 status
= (char *)page
->freelist
+ freelist_size
;
405 static inline unsigned int get_obj_status(struct page
*page
, int idx
)
409 struct kmem_cache
*cachep
= page
->slab_cache
;
411 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
412 status
= (char *)page
->freelist
+ freelist_size
;
418 static inline void set_obj_status(struct page
*page
, int idx
, int val
) {}
423 * Do not go above this order unless 0 objects fit into the slab or
424 * overridden on the command line.
426 #define SLAB_MAX_ORDER_HI 1
427 #define SLAB_MAX_ORDER_LO 0
428 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
429 static bool slab_max_order_set __initdata
;
431 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
433 struct page
*page
= virt_to_head_page(obj
);
434 return page
->slab_cache
;
437 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
440 return page
->s_mem
+ cache
->size
* idx
;
444 * We want to avoid an expensive divide : (offset / cache->size)
445 * Using the fact that size is a constant for a particular cache,
446 * we can replace (offset / cache->size) by
447 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
449 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
450 const struct page
*page
, void *obj
)
452 u32 offset
= (obj
- page
->s_mem
);
453 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
456 static struct arraycache_init initarray_generic
=
457 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
459 /* internal cache of cache description objs */
460 static struct kmem_cache kmem_cache_boot
= {
462 .limit
= BOOT_CPUCACHE_ENTRIES
,
464 .size
= sizeof(struct kmem_cache
),
465 .name
= "kmem_cache",
468 #define BAD_ALIEN_MAGIC 0x01020304ul
470 #ifdef CONFIG_LOCKDEP
473 * Slab sometimes uses the kmalloc slabs to store the slab headers
474 * for other slabs "off slab".
475 * The locking for this is tricky in that it nests within the locks
476 * of all other slabs in a few places; to deal with this special
477 * locking we put on-slab caches into a separate lock-class.
479 * We set lock class for alien array caches which are up during init.
480 * The lock annotation will be lost if all cpus of a node goes down and
481 * then comes back up during hotplug
483 static struct lock_class_key on_slab_l3_key
;
484 static struct lock_class_key on_slab_alc_key
;
486 static struct lock_class_key debugobj_l3_key
;
487 static struct lock_class_key debugobj_alc_key
;
489 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
490 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
493 struct array_cache
**alc
;
494 struct kmem_cache_node
*n
;
501 lockdep_set_class(&n
->list_lock
, l3_key
);
504 * FIXME: This check for BAD_ALIEN_MAGIC
505 * should go away when common slab code is taught to
506 * work even without alien caches.
507 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
508 * for alloc_alien_cache,
510 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
514 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
518 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
520 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
523 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
527 for_each_online_node(node
)
528 slab_set_debugobj_lock_classes_node(cachep
, node
);
531 static void init_node_lock_keys(int q
)
538 for (i
= 1; i
<= KMALLOC_SHIFT_HIGH
; i
++) {
539 struct kmem_cache_node
*n
;
540 struct kmem_cache
*cache
= kmalloc_caches
[i
];
546 if (!n
|| OFF_SLAB(cache
))
549 slab_set_lock_classes(cache
, &on_slab_l3_key
,
550 &on_slab_alc_key
, q
);
554 static void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int q
)
556 if (!cachep
->node
[q
])
559 slab_set_lock_classes(cachep
, &on_slab_l3_key
,
560 &on_slab_alc_key
, q
);
563 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
567 VM_BUG_ON(OFF_SLAB(cachep
));
569 on_slab_lock_classes_node(cachep
, node
);
572 static inline void init_lock_keys(void)
577 init_node_lock_keys(node
);
580 static void init_node_lock_keys(int q
)
584 static inline void init_lock_keys(void)
588 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
592 static inline void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int node
)
596 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
600 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
605 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
607 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
609 return cachep
->array
[smp_processor_id()];
612 static size_t calculate_freelist_size(int nr_objs
, size_t align
)
614 size_t freelist_size
;
616 freelist_size
= nr_objs
* sizeof(freelist_idx_t
);
617 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
618 freelist_size
+= nr_objs
* sizeof(char);
621 freelist_size
= ALIGN(freelist_size
, align
);
623 return freelist_size
;
626 static int calculate_nr_objs(size_t slab_size
, size_t buffer_size
,
627 size_t idx_size
, size_t align
)
630 size_t remained_size
;
631 size_t freelist_size
;
634 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
635 extra_space
= sizeof(char);
637 * Ignore padding for the initial guess. The padding
638 * is at most @align-1 bytes, and @buffer_size is at
639 * least @align. In the worst case, this result will
640 * be one greater than the number of objects that fit
641 * into the memory allocation when taking the padding
644 nr_objs
= slab_size
/ (buffer_size
+ idx_size
+ extra_space
);
647 * This calculated number will be either the right
648 * amount, or one greater than what we want.
650 remained_size
= slab_size
- nr_objs
* buffer_size
;
651 freelist_size
= calculate_freelist_size(nr_objs
, align
);
652 if (remained_size
< freelist_size
)
659 * Calculate the number of objects and left-over bytes for a given buffer size.
661 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
662 size_t align
, int flags
, size_t *left_over
,
667 size_t slab_size
= PAGE_SIZE
<< gfporder
;
670 * The slab management structure can be either off the slab or
671 * on it. For the latter case, the memory allocated for a
674 * - One unsigned int for each object
675 * - Padding to respect alignment of @align
676 * - @buffer_size bytes for each object
678 * If the slab management structure is off the slab, then the
679 * alignment will already be calculated into the size. Because
680 * the slabs are all pages aligned, the objects will be at the
681 * correct alignment when allocated.
683 if (flags
& CFLGS_OFF_SLAB
) {
685 nr_objs
= slab_size
/ buffer_size
;
688 nr_objs
= calculate_nr_objs(slab_size
, buffer_size
,
689 sizeof(freelist_idx_t
), align
);
690 mgmt_size
= calculate_freelist_size(nr_objs
, align
);
693 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
697 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
699 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
702 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
703 function
, cachep
->name
, msg
);
705 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
710 * By default on NUMA we use alien caches to stage the freeing of
711 * objects allocated from other nodes. This causes massive memory
712 * inefficiencies when using fake NUMA setup to split memory into a
713 * large number of small nodes, so it can be disabled on the command
717 static int use_alien_caches __read_mostly
= 1;
718 static int __init
noaliencache_setup(char *s
)
720 use_alien_caches
= 0;
723 __setup("noaliencache", noaliencache_setup
);
725 static int __init
slab_max_order_setup(char *str
)
727 get_option(&str
, &slab_max_order
);
728 slab_max_order
= slab_max_order
< 0 ? 0 :
729 min(slab_max_order
, MAX_ORDER
- 1);
730 slab_max_order_set
= true;
734 __setup("slab_max_order=", slab_max_order_setup
);
738 * Special reaping functions for NUMA systems called from cache_reap().
739 * These take care of doing round robin flushing of alien caches (containing
740 * objects freed on different nodes from which they were allocated) and the
741 * flushing of remote pcps by calling drain_node_pages.
743 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
745 static void init_reap_node(int cpu
)
749 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
750 if (node
== MAX_NUMNODES
)
751 node
= first_node(node_online_map
);
753 per_cpu(slab_reap_node
, cpu
) = node
;
756 static void next_reap_node(void)
758 int node
= __this_cpu_read(slab_reap_node
);
760 node
= next_node(node
, node_online_map
);
761 if (unlikely(node
>= MAX_NUMNODES
))
762 node
= first_node(node_online_map
);
763 __this_cpu_write(slab_reap_node
, node
);
767 #define init_reap_node(cpu) do { } while (0)
768 #define next_reap_node(void) do { } while (0)
772 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
773 * via the workqueue/eventd.
774 * Add the CPU number into the expiration time to minimize the possibility of
775 * the CPUs getting into lockstep and contending for the global cache chain
778 static void start_cpu_timer(int cpu
)
780 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
783 * When this gets called from do_initcalls via cpucache_init(),
784 * init_workqueues() has already run, so keventd will be setup
787 if (keventd_up() && reap_work
->work
.func
== NULL
) {
789 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
790 schedule_delayed_work_on(cpu
, reap_work
,
791 __round_jiffies_relative(HZ
, cpu
));
795 static struct array_cache
*alloc_arraycache(int node
, int entries
,
796 int batchcount
, gfp_t gfp
)
798 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
799 struct array_cache
*nc
= NULL
;
801 nc
= kmalloc_node(memsize
, gfp
, node
);
803 * The array_cache structures contain pointers to free object.
804 * However, when such objects are allocated or transferred to another
805 * cache the pointers are not cleared and they could be counted as
806 * valid references during a kmemleak scan. Therefore, kmemleak must
807 * not scan such objects.
809 kmemleak_no_scan(nc
);
813 nc
->batchcount
= batchcount
;
815 spin_lock_init(&nc
->lock
);
820 static inline bool is_slab_pfmemalloc(struct page
*page
)
822 return PageSlabPfmemalloc(page
);
825 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
826 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
827 struct array_cache
*ac
)
829 struct kmem_cache_node
*n
= cachep
->node
[numa_mem_id()];
833 if (!pfmemalloc_active
)
836 spin_lock_irqsave(&n
->list_lock
, flags
);
837 list_for_each_entry(page
, &n
->slabs_full
, lru
)
838 if (is_slab_pfmemalloc(page
))
841 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
842 if (is_slab_pfmemalloc(page
))
845 list_for_each_entry(page
, &n
->slabs_free
, lru
)
846 if (is_slab_pfmemalloc(page
))
849 pfmemalloc_active
= false;
851 spin_unlock_irqrestore(&n
->list_lock
, flags
);
854 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
855 gfp_t flags
, bool force_refill
)
858 void *objp
= ac
->entry
[--ac
->avail
];
860 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
861 if (unlikely(is_obj_pfmemalloc(objp
))) {
862 struct kmem_cache_node
*n
;
864 if (gfp_pfmemalloc_allowed(flags
)) {
865 clear_obj_pfmemalloc(&objp
);
869 /* The caller cannot use PFMEMALLOC objects, find another one */
870 for (i
= 0; i
< ac
->avail
; i
++) {
871 /* If a !PFMEMALLOC object is found, swap them */
872 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
874 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
875 ac
->entry
[ac
->avail
] = objp
;
881 * If there are empty slabs on the slabs_free list and we are
882 * being forced to refill the cache, mark this one !pfmemalloc.
884 n
= cachep
->node
[numa_mem_id()];
885 if (!list_empty(&n
->slabs_free
) && force_refill
) {
886 struct page
*page
= virt_to_head_page(objp
);
887 ClearPageSlabPfmemalloc(page
);
888 clear_obj_pfmemalloc(&objp
);
889 recheck_pfmemalloc_active(cachep
, ac
);
893 /* No !PFMEMALLOC objects available */
901 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
902 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
906 if (unlikely(sk_memalloc_socks()))
907 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
909 objp
= ac
->entry
[--ac
->avail
];
914 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
917 if (unlikely(pfmemalloc_active
)) {
918 /* Some pfmemalloc slabs exist, check if this is one */
919 struct page
*page
= virt_to_head_page(objp
);
920 if (PageSlabPfmemalloc(page
))
921 set_obj_pfmemalloc(&objp
);
927 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
930 if (unlikely(sk_memalloc_socks()))
931 objp
= __ac_put_obj(cachep
, ac
, objp
);
933 ac
->entry
[ac
->avail
++] = objp
;
937 * Transfer objects in one arraycache to another.
938 * Locking must be handled by the caller.
940 * Return the number of entries transferred.
942 static int transfer_objects(struct array_cache
*to
,
943 struct array_cache
*from
, unsigned int max
)
945 /* Figure out how many entries to transfer */
946 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
951 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
961 #define drain_alien_cache(cachep, alien) do { } while (0)
962 #define reap_alien(cachep, n) do { } while (0)
964 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
966 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
969 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
973 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
978 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
984 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
985 gfp_t flags
, int nodeid
)
990 #else /* CONFIG_NUMA */
992 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
993 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
995 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
997 struct array_cache
**ac_ptr
;
998 int memsize
= sizeof(void *) * nr_node_ids
;
1003 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
1006 if (i
== node
|| !node_online(i
))
1008 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
1010 for (i
--; i
>= 0; i
--)
1020 static void free_alien_cache(struct array_cache
**ac_ptr
)
1031 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1032 struct array_cache
*ac
, int node
)
1034 struct kmem_cache_node
*n
= cachep
->node
[node
];
1037 spin_lock(&n
->list_lock
);
1039 * Stuff objects into the remote nodes shared array first.
1040 * That way we could avoid the overhead of putting the objects
1041 * into the free lists and getting them back later.
1044 transfer_objects(n
->shared
, ac
, ac
->limit
);
1046 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1048 spin_unlock(&n
->list_lock
);
1053 * Called from cache_reap() to regularly drain alien caches round robin.
1055 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
1057 int node
= __this_cpu_read(slab_reap_node
);
1060 struct array_cache
*ac
= n
->alien
[node
];
1062 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1063 __drain_alien_cache(cachep
, ac
, node
);
1064 spin_unlock_irq(&ac
->lock
);
1069 static void drain_alien_cache(struct kmem_cache
*cachep
,
1070 struct array_cache
**alien
)
1073 struct array_cache
*ac
;
1074 unsigned long flags
;
1076 for_each_online_node(i
) {
1079 spin_lock_irqsave(&ac
->lock
, flags
);
1080 __drain_alien_cache(cachep
, ac
, i
);
1081 spin_unlock_irqrestore(&ac
->lock
, flags
);
1086 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1088 int nodeid
= page_to_nid(virt_to_page(objp
));
1089 struct kmem_cache_node
*n
;
1090 struct array_cache
*alien
= NULL
;
1093 node
= numa_mem_id();
1096 * Make sure we are not freeing a object from another node to the array
1097 * cache on this cpu.
1099 if (likely(nodeid
== node
))
1102 n
= cachep
->node
[node
];
1103 STATS_INC_NODEFREES(cachep
);
1104 if (n
->alien
&& n
->alien
[nodeid
]) {
1105 alien
= n
->alien
[nodeid
];
1106 spin_lock(&alien
->lock
);
1107 if (unlikely(alien
->avail
== alien
->limit
)) {
1108 STATS_INC_ACOVERFLOW(cachep
);
1109 __drain_alien_cache(cachep
, alien
, nodeid
);
1111 ac_put_obj(cachep
, alien
, objp
);
1112 spin_unlock(&alien
->lock
);
1114 spin_lock(&(cachep
->node
[nodeid
])->list_lock
);
1115 free_block(cachep
, &objp
, 1, nodeid
);
1116 spin_unlock(&(cachep
->node
[nodeid
])->list_lock
);
1123 * Allocates and initializes node for a node on each slab cache, used for
1124 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1125 * will be allocated off-node since memory is not yet online for the new node.
1126 * When hotplugging memory or a cpu, existing node are not replaced if
1129 * Must hold slab_mutex.
1131 static int init_cache_node_node(int node
)
1133 struct kmem_cache
*cachep
;
1134 struct kmem_cache_node
*n
;
1135 const int memsize
= sizeof(struct kmem_cache_node
);
1137 list_for_each_entry(cachep
, &slab_caches
, list
) {
1139 * Set up the kmem_cache_node for cpu before we can
1140 * begin anything. Make sure some other cpu on this
1141 * node has not already allocated this
1143 if (!cachep
->node
[node
]) {
1144 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1147 kmem_cache_node_init(n
);
1148 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
1149 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1152 * The kmem_cache_nodes don't come and go as CPUs
1153 * come and go. slab_mutex is sufficient
1156 cachep
->node
[node
] = n
;
1159 spin_lock_irq(&cachep
->node
[node
]->list_lock
);
1160 cachep
->node
[node
]->free_limit
=
1161 (1 + nr_cpus_node(node
)) *
1162 cachep
->batchcount
+ cachep
->num
;
1163 spin_unlock_irq(&cachep
->node
[node
]->list_lock
);
1168 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1169 struct kmem_cache_node
*n
)
1171 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1174 static void cpuup_canceled(long cpu
)
1176 struct kmem_cache
*cachep
;
1177 struct kmem_cache_node
*n
= NULL
;
1178 int node
= cpu_to_mem(cpu
);
1179 const struct cpumask
*mask
= cpumask_of_node(node
);
1181 list_for_each_entry(cachep
, &slab_caches
, list
) {
1182 struct array_cache
*nc
;
1183 struct array_cache
*shared
;
1184 struct array_cache
**alien
;
1186 /* cpu is dead; no one can alloc from it. */
1187 nc
= cachep
->array
[cpu
];
1188 cachep
->array
[cpu
] = NULL
;
1189 n
= cachep
->node
[node
];
1192 goto free_array_cache
;
1194 spin_lock_irq(&n
->list_lock
);
1196 /* Free limit for this kmem_cache_node */
1197 n
->free_limit
-= cachep
->batchcount
;
1199 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1201 if (!cpumask_empty(mask
)) {
1202 spin_unlock_irq(&n
->list_lock
);
1203 goto free_array_cache
;
1208 free_block(cachep
, shared
->entry
,
1209 shared
->avail
, node
);
1216 spin_unlock_irq(&n
->list_lock
);
1220 drain_alien_cache(cachep
, alien
);
1221 free_alien_cache(alien
);
1227 * In the previous loop, all the objects were freed to
1228 * the respective cache's slabs, now we can go ahead and
1229 * shrink each nodelist to its limit.
1231 list_for_each_entry(cachep
, &slab_caches
, list
) {
1232 n
= cachep
->node
[node
];
1235 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1239 static int cpuup_prepare(long cpu
)
1241 struct kmem_cache
*cachep
;
1242 struct kmem_cache_node
*n
= NULL
;
1243 int node
= cpu_to_mem(cpu
);
1247 * We need to do this right in the beginning since
1248 * alloc_arraycache's are going to use this list.
1249 * kmalloc_node allows us to add the slab to the right
1250 * kmem_cache_node and not this cpu's kmem_cache_node
1252 err
= init_cache_node_node(node
);
1257 * Now we can go ahead with allocating the shared arrays and
1260 list_for_each_entry(cachep
, &slab_caches
, list
) {
1261 struct array_cache
*nc
;
1262 struct array_cache
*shared
= NULL
;
1263 struct array_cache
**alien
= NULL
;
1265 nc
= alloc_arraycache(node
, cachep
->limit
,
1266 cachep
->batchcount
, GFP_KERNEL
);
1269 if (cachep
->shared
) {
1270 shared
= alloc_arraycache(node
,
1271 cachep
->shared
* cachep
->batchcount
,
1272 0xbaadf00d, GFP_KERNEL
);
1278 if (use_alien_caches
) {
1279 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1286 cachep
->array
[cpu
] = nc
;
1287 n
= cachep
->node
[node
];
1290 spin_lock_irq(&n
->list_lock
);
1293 * We are serialised from CPU_DEAD or
1294 * CPU_UP_CANCELLED by the cpucontrol lock
1305 spin_unlock_irq(&n
->list_lock
);
1307 free_alien_cache(alien
);
1308 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1309 slab_set_debugobj_lock_classes_node(cachep
, node
);
1310 else if (!OFF_SLAB(cachep
) &&
1311 !(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1312 on_slab_lock_classes_node(cachep
, node
);
1314 init_node_lock_keys(node
);
1318 cpuup_canceled(cpu
);
1322 static int cpuup_callback(struct notifier_block
*nfb
,
1323 unsigned long action
, void *hcpu
)
1325 long cpu
= (long)hcpu
;
1329 case CPU_UP_PREPARE
:
1330 case CPU_UP_PREPARE_FROZEN
:
1331 mutex_lock(&slab_mutex
);
1332 err
= cpuup_prepare(cpu
);
1333 mutex_unlock(&slab_mutex
);
1336 case CPU_ONLINE_FROZEN
:
1337 start_cpu_timer(cpu
);
1339 #ifdef CONFIG_HOTPLUG_CPU
1340 case CPU_DOWN_PREPARE
:
1341 case CPU_DOWN_PREPARE_FROZEN
:
1343 * Shutdown cache reaper. Note that the slab_mutex is
1344 * held so that if cache_reap() is invoked it cannot do
1345 * anything expensive but will only modify reap_work
1346 * and reschedule the timer.
1348 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1349 /* Now the cache_reaper is guaranteed to be not running. */
1350 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1352 case CPU_DOWN_FAILED
:
1353 case CPU_DOWN_FAILED_FROZEN
:
1354 start_cpu_timer(cpu
);
1357 case CPU_DEAD_FROZEN
:
1359 * Even if all the cpus of a node are down, we don't free the
1360 * kmem_cache_node of any cache. This to avoid a race between
1361 * cpu_down, and a kmalloc allocation from another cpu for
1362 * memory from the node of the cpu going down. The node
1363 * structure is usually allocated from kmem_cache_create() and
1364 * gets destroyed at kmem_cache_destroy().
1368 case CPU_UP_CANCELED
:
1369 case CPU_UP_CANCELED_FROZEN
:
1370 mutex_lock(&slab_mutex
);
1371 cpuup_canceled(cpu
);
1372 mutex_unlock(&slab_mutex
);
1375 return notifier_from_errno(err
);
1378 static struct notifier_block cpucache_notifier
= {
1379 &cpuup_callback
, NULL
, 0
1382 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1384 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1385 * Returns -EBUSY if all objects cannot be drained so that the node is not
1388 * Must hold slab_mutex.
1390 static int __meminit
drain_cache_node_node(int node
)
1392 struct kmem_cache
*cachep
;
1395 list_for_each_entry(cachep
, &slab_caches
, list
) {
1396 struct kmem_cache_node
*n
;
1398 n
= cachep
->node
[node
];
1402 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1404 if (!list_empty(&n
->slabs_full
) ||
1405 !list_empty(&n
->slabs_partial
)) {
1413 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1414 unsigned long action
, void *arg
)
1416 struct memory_notify
*mnb
= arg
;
1420 nid
= mnb
->status_change_nid
;
1425 case MEM_GOING_ONLINE
:
1426 mutex_lock(&slab_mutex
);
1427 ret
= init_cache_node_node(nid
);
1428 mutex_unlock(&slab_mutex
);
1430 case MEM_GOING_OFFLINE
:
1431 mutex_lock(&slab_mutex
);
1432 ret
= drain_cache_node_node(nid
);
1433 mutex_unlock(&slab_mutex
);
1437 case MEM_CANCEL_ONLINE
:
1438 case MEM_CANCEL_OFFLINE
:
1442 return notifier_from_errno(ret
);
1444 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1447 * swap the static kmem_cache_node with kmalloced memory
1449 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1452 struct kmem_cache_node
*ptr
;
1454 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1457 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1459 * Do not assume that spinlocks can be initialized via memcpy:
1461 spin_lock_init(&ptr
->list_lock
);
1463 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1464 cachep
->node
[nodeid
] = ptr
;
1468 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1469 * size of kmem_cache_node.
1471 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1475 for_each_online_node(node
) {
1476 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1477 cachep
->node
[node
]->next_reap
= jiffies
+
1479 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1484 * The memory after the last cpu cache pointer is used for the
1487 static void setup_node_pointer(struct kmem_cache
*cachep
)
1489 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1493 * Initialisation. Called after the page allocator have been initialised and
1494 * before smp_init().
1496 void __init
kmem_cache_init(void)
1500 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1501 sizeof(struct rcu_head
));
1502 kmem_cache
= &kmem_cache_boot
;
1503 setup_node_pointer(kmem_cache
);
1505 if (num_possible_nodes() == 1)
1506 use_alien_caches
= 0;
1508 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1509 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1511 set_up_node(kmem_cache
, CACHE_CACHE
);
1514 * Fragmentation resistance on low memory - only use bigger
1515 * page orders on machines with more than 32MB of memory if
1516 * not overridden on the command line.
1518 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1519 slab_max_order
= SLAB_MAX_ORDER_HI
;
1521 /* Bootstrap is tricky, because several objects are allocated
1522 * from caches that do not exist yet:
1523 * 1) initialize the kmem_cache cache: it contains the struct
1524 * kmem_cache structures of all caches, except kmem_cache itself:
1525 * kmem_cache is statically allocated.
1526 * Initially an __init data area is used for the head array and the
1527 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1528 * array at the end of the bootstrap.
1529 * 2) Create the first kmalloc cache.
1530 * The struct kmem_cache for the new cache is allocated normally.
1531 * An __init data area is used for the head array.
1532 * 3) Create the remaining kmalloc caches, with minimally sized
1534 * 4) Replace the __init data head arrays for kmem_cache and the first
1535 * kmalloc cache with kmalloc allocated arrays.
1536 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1537 * the other cache's with kmalloc allocated memory.
1538 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1541 /* 1) create the kmem_cache */
1544 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1546 create_boot_cache(kmem_cache
, "kmem_cache",
1547 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1548 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1549 SLAB_HWCACHE_ALIGN
);
1550 list_add(&kmem_cache
->list
, &slab_caches
);
1552 /* 2+3) create the kmalloc caches */
1555 * Initialize the caches that provide memory for the array cache and the
1556 * kmem_cache_node structures first. Without this, further allocations will
1560 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1561 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1563 if (INDEX_AC
!= INDEX_NODE
)
1564 kmalloc_caches
[INDEX_NODE
] =
1565 create_kmalloc_cache("kmalloc-node",
1566 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1568 slab_early_init
= 0;
1570 /* 4) Replace the bootstrap head arrays */
1572 struct array_cache
*ptr
;
1574 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1576 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1577 sizeof(struct arraycache_init
));
1579 * Do not assume that spinlocks can be initialized via memcpy:
1581 spin_lock_init(&ptr
->lock
);
1583 kmem_cache
->array
[smp_processor_id()] = ptr
;
1585 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1587 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1588 != &initarray_generic
.cache
);
1589 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1590 sizeof(struct arraycache_init
));
1592 * Do not assume that spinlocks can be initialized via memcpy:
1594 spin_lock_init(&ptr
->lock
);
1596 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1598 /* 5) Replace the bootstrap kmem_cache_node */
1602 for_each_online_node(nid
) {
1603 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1605 init_list(kmalloc_caches
[INDEX_AC
],
1606 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1608 if (INDEX_AC
!= INDEX_NODE
) {
1609 init_list(kmalloc_caches
[INDEX_NODE
],
1610 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1615 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1618 void __init
kmem_cache_init_late(void)
1620 struct kmem_cache
*cachep
;
1624 /* 6) resize the head arrays to their final sizes */
1625 mutex_lock(&slab_mutex
);
1626 list_for_each_entry(cachep
, &slab_caches
, list
)
1627 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1629 mutex_unlock(&slab_mutex
);
1631 /* Annotate slab for lockdep -- annotate the malloc caches */
1638 * Register a cpu startup notifier callback that initializes
1639 * cpu_cache_get for all new cpus
1641 register_cpu_notifier(&cpucache_notifier
);
1645 * Register a memory hotplug callback that initializes and frees
1648 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1652 * The reap timers are started later, with a module init call: That part
1653 * of the kernel is not yet operational.
1657 static int __init
cpucache_init(void)
1662 * Register the timers that return unneeded pages to the page allocator
1664 for_each_online_cpu(cpu
)
1665 start_cpu_timer(cpu
);
1671 __initcall(cpucache_init
);
1673 static noinline
void
1674 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1677 struct kmem_cache_node
*n
;
1679 unsigned long flags
;
1681 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1682 DEFAULT_RATELIMIT_BURST
);
1684 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1688 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1690 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1691 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1693 for_each_online_node(node
) {
1694 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1695 unsigned long active_slabs
= 0, num_slabs
= 0;
1697 n
= cachep
->node
[node
];
1701 spin_lock_irqsave(&n
->list_lock
, flags
);
1702 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1703 active_objs
+= cachep
->num
;
1706 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1707 active_objs
+= page
->active
;
1710 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1713 free_objects
+= n
->free_objects
;
1714 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1716 num_slabs
+= active_slabs
;
1717 num_objs
= num_slabs
* cachep
->num
;
1719 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1720 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1727 * Interface to system's page allocator. No need to hold the cache-lock.
1729 * If we requested dmaable memory, we will get it. Even if we
1730 * did not request dmaable memory, we might get it, but that
1731 * would be relatively rare and ignorable.
1733 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1739 flags
|= cachep
->allocflags
;
1740 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1741 flags
|= __GFP_RECLAIMABLE
;
1743 if (memcg_charge_slab(cachep
, flags
, cachep
->gfporder
))
1746 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1748 memcg_uncharge_slab(cachep
, cachep
->gfporder
);
1749 slab_out_of_memory(cachep
, flags
, nodeid
);
1753 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1754 if (unlikely(page
->pfmemalloc
))
1755 pfmemalloc_active
= true;
1757 nr_pages
= (1 << cachep
->gfporder
);
1758 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1759 add_zone_page_state(page_zone(page
),
1760 NR_SLAB_RECLAIMABLE
, nr_pages
);
1762 add_zone_page_state(page_zone(page
),
1763 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1764 __SetPageSlab(page
);
1765 if (page
->pfmemalloc
)
1766 SetPageSlabPfmemalloc(page
);
1768 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1769 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1772 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1774 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1781 * Interface to system's page release.
1783 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1785 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1787 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1789 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1790 sub_zone_page_state(page_zone(page
),
1791 NR_SLAB_RECLAIMABLE
, nr_freed
);
1793 sub_zone_page_state(page_zone(page
),
1794 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1796 BUG_ON(!PageSlab(page
));
1797 __ClearPageSlabPfmemalloc(page
);
1798 __ClearPageSlab(page
);
1799 page_mapcount_reset(page
);
1800 page
->mapping
= NULL
;
1802 if (current
->reclaim_state
)
1803 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1804 __free_pages(page
, cachep
->gfporder
);
1805 memcg_uncharge_slab(cachep
, cachep
->gfporder
);
1808 static void kmem_rcu_free(struct rcu_head
*head
)
1810 struct kmem_cache
*cachep
;
1813 page
= container_of(head
, struct page
, rcu_head
);
1814 cachep
= page
->slab_cache
;
1816 kmem_freepages(cachep
, page
);
1821 #ifdef CONFIG_DEBUG_PAGEALLOC
1822 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1823 unsigned long caller
)
1825 int size
= cachep
->object_size
;
1827 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1829 if (size
< 5 * sizeof(unsigned long))
1832 *addr
++ = 0x12345678;
1834 *addr
++ = smp_processor_id();
1835 size
-= 3 * sizeof(unsigned long);
1837 unsigned long *sptr
= &caller
;
1838 unsigned long svalue
;
1840 while (!kstack_end(sptr
)) {
1842 if (kernel_text_address(svalue
)) {
1844 size
-= sizeof(unsigned long);
1845 if (size
<= sizeof(unsigned long))
1851 *addr
++ = 0x87654321;
1855 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1857 int size
= cachep
->object_size
;
1858 addr
= &((char *)addr
)[obj_offset(cachep
)];
1860 memset(addr
, val
, size
);
1861 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1864 static void dump_line(char *data
, int offset
, int limit
)
1867 unsigned char error
= 0;
1870 printk(KERN_ERR
"%03x: ", offset
);
1871 for (i
= 0; i
< limit
; i
++) {
1872 if (data
[offset
+ i
] != POISON_FREE
) {
1873 error
= data
[offset
+ i
];
1877 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1878 &data
[offset
], limit
, 1);
1880 if (bad_count
== 1) {
1881 error
^= POISON_FREE
;
1882 if (!(error
& (error
- 1))) {
1883 printk(KERN_ERR
"Single bit error detected. Probably "
1886 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1889 printk(KERN_ERR
"Run a memory test tool.\n");
1898 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1903 if (cachep
->flags
& SLAB_RED_ZONE
) {
1904 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1905 *dbg_redzone1(cachep
, objp
),
1906 *dbg_redzone2(cachep
, objp
));
1909 if (cachep
->flags
& SLAB_STORE_USER
) {
1910 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1911 *dbg_userword(cachep
, objp
),
1912 *dbg_userword(cachep
, objp
));
1914 realobj
= (char *)objp
+ obj_offset(cachep
);
1915 size
= cachep
->object_size
;
1916 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1919 if (i
+ limit
> size
)
1921 dump_line(realobj
, i
, limit
);
1925 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1931 realobj
= (char *)objp
+ obj_offset(cachep
);
1932 size
= cachep
->object_size
;
1934 for (i
= 0; i
< size
; i
++) {
1935 char exp
= POISON_FREE
;
1938 if (realobj
[i
] != exp
) {
1944 "Slab corruption (%s): %s start=%p, len=%d\n",
1945 print_tainted(), cachep
->name
, realobj
, size
);
1946 print_objinfo(cachep
, objp
, 0);
1948 /* Hexdump the affected line */
1951 if (i
+ limit
> size
)
1953 dump_line(realobj
, i
, limit
);
1956 /* Limit to 5 lines */
1962 /* Print some data about the neighboring objects, if they
1965 struct page
*page
= virt_to_head_page(objp
);
1968 objnr
= obj_to_index(cachep
, page
, objp
);
1970 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1971 realobj
= (char *)objp
+ obj_offset(cachep
);
1972 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1974 print_objinfo(cachep
, objp
, 2);
1976 if (objnr
+ 1 < cachep
->num
) {
1977 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1978 realobj
= (char *)objp
+ obj_offset(cachep
);
1979 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1981 print_objinfo(cachep
, objp
, 2);
1988 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1992 for (i
= 0; i
< cachep
->num
; i
++) {
1993 void *objp
= index_to_obj(cachep
, page
, i
);
1995 if (cachep
->flags
& SLAB_POISON
) {
1996 #ifdef CONFIG_DEBUG_PAGEALLOC
1997 if (cachep
->size
% PAGE_SIZE
== 0 &&
1999 kernel_map_pages(virt_to_page(objp
),
2000 cachep
->size
/ PAGE_SIZE
, 1);
2002 check_poison_obj(cachep
, objp
);
2004 check_poison_obj(cachep
, objp
);
2007 if (cachep
->flags
& SLAB_RED_ZONE
) {
2008 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2009 slab_error(cachep
, "start of a freed object "
2011 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2012 slab_error(cachep
, "end of a freed object "
2018 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
2025 * slab_destroy - destroy and release all objects in a slab
2026 * @cachep: cache pointer being destroyed
2027 * @page: page pointer being destroyed
2029 * Destroy all the objs in a slab, and release the mem back to the system.
2030 * Before calling the slab must have been unlinked from the cache. The
2031 * cache-lock is not held/needed.
2033 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
2037 freelist
= page
->freelist
;
2038 slab_destroy_debugcheck(cachep
, page
);
2039 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2040 struct rcu_head
*head
;
2043 * RCU free overloads the RCU head over the LRU.
2044 * slab_page has been overloeaded over the LRU,
2045 * however it is not used from now on so that
2046 * we can use it safely.
2048 head
= (void *)&page
->rcu_head
;
2049 call_rcu(head
, kmem_rcu_free
);
2052 kmem_freepages(cachep
, page
);
2056 * From now on, we don't use freelist
2057 * although actual page can be freed in rcu context
2059 if (OFF_SLAB(cachep
))
2060 kmem_cache_free(cachep
->freelist_cache
, freelist
);
2064 * calculate_slab_order - calculate size (page order) of slabs
2065 * @cachep: pointer to the cache that is being created
2066 * @size: size of objects to be created in this cache.
2067 * @align: required alignment for the objects.
2068 * @flags: slab allocation flags
2070 * Also calculates the number of objects per slab.
2072 * This could be made much more intelligent. For now, try to avoid using
2073 * high order pages for slabs. When the gfp() functions are more friendly
2074 * towards high-order requests, this should be changed.
2076 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2077 size_t size
, size_t align
, unsigned long flags
)
2079 unsigned long offslab_limit
;
2080 size_t left_over
= 0;
2083 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2087 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2091 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2092 if (num
> SLAB_OBJ_MAX_NUM
)
2095 if (flags
& CFLGS_OFF_SLAB
) {
2096 size_t freelist_size_per_obj
= sizeof(freelist_idx_t
);
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 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
2103 freelist_size_per_obj
+= sizeof(char);
2104 offslab_limit
= size
;
2105 offslab_limit
/= freelist_size_per_obj
;
2107 if (num
> offslab_limit
)
2111 /* Found something acceptable - save it away */
2113 cachep
->gfporder
= gfporder
;
2114 left_over
= remainder
;
2117 * A VFS-reclaimable slab tends to have most allocations
2118 * as GFP_NOFS and we really don't want to have to be allocating
2119 * higher-order pages when we are unable to shrink dcache.
2121 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2125 * Large number of objects is good, but very large slabs are
2126 * currently bad for the gfp()s.
2128 if (gfporder
>= slab_max_order
)
2132 * Acceptable internal fragmentation?
2134 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2140 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2142 if (slab_state
>= FULL
)
2143 return enable_cpucache(cachep
, gfp
);
2145 if (slab_state
== DOWN
) {
2147 * Note: Creation of first cache (kmem_cache).
2148 * The setup_node is taken care
2149 * of by the caller of __kmem_cache_create
2151 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2152 slab_state
= PARTIAL
;
2153 } else if (slab_state
== PARTIAL
) {
2155 * Note: the second kmem_cache_create must create the cache
2156 * that's used by kmalloc(24), otherwise the creation of
2157 * further caches will BUG().
2159 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2162 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2163 * the second cache, then we need to set up all its node/,
2164 * otherwise the creation of further caches will BUG().
2166 set_up_node(cachep
, SIZE_AC
);
2167 if (INDEX_AC
== INDEX_NODE
)
2168 slab_state
= PARTIAL_NODE
;
2170 slab_state
= PARTIAL_ARRAYCACHE
;
2172 /* Remaining boot caches */
2173 cachep
->array
[smp_processor_id()] =
2174 kmalloc(sizeof(struct arraycache_init
), gfp
);
2176 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2177 set_up_node(cachep
, SIZE_NODE
);
2178 slab_state
= PARTIAL_NODE
;
2181 for_each_online_node(node
) {
2182 cachep
->node
[node
] =
2183 kmalloc_node(sizeof(struct kmem_cache_node
),
2185 BUG_ON(!cachep
->node
[node
]);
2186 kmem_cache_node_init(cachep
->node
[node
]);
2190 cachep
->node
[numa_mem_id()]->next_reap
=
2191 jiffies
+ REAPTIMEOUT_NODE
+
2192 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
2194 cpu_cache_get(cachep
)->avail
= 0;
2195 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2196 cpu_cache_get(cachep
)->batchcount
= 1;
2197 cpu_cache_get(cachep
)->touched
= 0;
2198 cachep
->batchcount
= 1;
2199 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2204 * __kmem_cache_create - Create a cache.
2205 * @cachep: cache management descriptor
2206 * @flags: SLAB flags
2208 * Returns a ptr to the cache on success, NULL on failure.
2209 * Cannot be called within a int, but can be interrupted.
2210 * The @ctor is run when new pages are allocated by the cache.
2214 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2215 * to catch references to uninitialised memory.
2217 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2218 * for buffer overruns.
2220 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2221 * cacheline. This can be beneficial if you're counting cycles as closely
2225 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2227 size_t left_over
, freelist_size
, ralign
;
2230 size_t size
= cachep
->size
;
2235 * Enable redzoning and last user accounting, except for caches with
2236 * large objects, if the increased size would increase the object size
2237 * above the next power of two: caches with object sizes just above a
2238 * power of two have a significant amount of internal fragmentation.
2240 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2241 2 * sizeof(unsigned long long)))
2242 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2243 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2244 flags
|= SLAB_POISON
;
2246 if (flags
& SLAB_DESTROY_BY_RCU
)
2247 BUG_ON(flags
& SLAB_POISON
);
2251 * Check that size is in terms of words. This is needed to avoid
2252 * unaligned accesses for some archs when redzoning is used, and makes
2253 * sure any on-slab bufctl's are also correctly aligned.
2255 if (size
& (BYTES_PER_WORD
- 1)) {
2256 size
+= (BYTES_PER_WORD
- 1);
2257 size
&= ~(BYTES_PER_WORD
- 1);
2261 * Redzoning and user store require word alignment or possibly larger.
2262 * Note this will be overridden by architecture or caller mandated
2263 * alignment if either is greater than BYTES_PER_WORD.
2265 if (flags
& SLAB_STORE_USER
)
2266 ralign
= BYTES_PER_WORD
;
2268 if (flags
& SLAB_RED_ZONE
) {
2269 ralign
= REDZONE_ALIGN
;
2270 /* If redzoning, ensure that the second redzone is suitably
2271 * aligned, by adjusting the object size accordingly. */
2272 size
+= REDZONE_ALIGN
- 1;
2273 size
&= ~(REDZONE_ALIGN
- 1);
2276 /* 3) caller mandated alignment */
2277 if (ralign
< cachep
->align
) {
2278 ralign
= cachep
->align
;
2280 /* disable debug if necessary */
2281 if (ralign
> __alignof__(unsigned long long))
2282 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2286 cachep
->align
= ralign
;
2288 if (slab_is_available())
2293 setup_node_pointer(cachep
);
2297 * Both debugging options require word-alignment which is calculated
2300 if (flags
& SLAB_RED_ZONE
) {
2301 /* add space for red zone words */
2302 cachep
->obj_offset
+= sizeof(unsigned long long);
2303 size
+= 2 * sizeof(unsigned long long);
2305 if (flags
& SLAB_STORE_USER
) {
2306 /* user store requires one word storage behind the end of
2307 * the real object. But if the second red zone needs to be
2308 * aligned to 64 bits, we must allow that much space.
2310 if (flags
& SLAB_RED_ZONE
)
2311 size
+= REDZONE_ALIGN
;
2313 size
+= BYTES_PER_WORD
;
2315 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2316 if (size
>= kmalloc_size(INDEX_NODE
+ 1)
2317 && cachep
->object_size
> cache_line_size()
2318 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2319 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2326 * Determine if the slab management is 'on' or 'off' slab.
2327 * (bootstrapping cannot cope with offslab caches so don't do
2328 * it too early on. Always use on-slab management when
2329 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2331 if ((size
>= (PAGE_SIZE
>> 5)) && !slab_early_init
&&
2332 !(flags
& SLAB_NOLEAKTRACE
))
2334 * Size is large, assume best to place the slab management obj
2335 * off-slab (should allow better packing of objs).
2337 flags
|= CFLGS_OFF_SLAB
;
2339 size
= ALIGN(size
, cachep
->align
);
2341 * We should restrict the number of objects in a slab to implement
2342 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2344 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2345 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2347 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2352 freelist_size
= calculate_freelist_size(cachep
->num
, cachep
->align
);
2355 * If the slab has been placed off-slab, and we have enough space then
2356 * move it on-slab. This is at the expense of any extra colouring.
2358 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= freelist_size
) {
2359 flags
&= ~CFLGS_OFF_SLAB
;
2360 left_over
-= freelist_size
;
2363 if (flags
& CFLGS_OFF_SLAB
) {
2364 /* really off slab. No need for manual alignment */
2365 freelist_size
= calculate_freelist_size(cachep
->num
, 0);
2367 #ifdef CONFIG_PAGE_POISONING
2368 /* If we're going to use the generic kernel_map_pages()
2369 * poisoning, then it's going to smash the contents of
2370 * the redzone and userword anyhow, so switch them off.
2372 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2373 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2377 cachep
->colour_off
= cache_line_size();
2378 /* Offset must be a multiple of the alignment. */
2379 if (cachep
->colour_off
< cachep
->align
)
2380 cachep
->colour_off
= cachep
->align
;
2381 cachep
->colour
= left_over
/ cachep
->colour_off
;
2382 cachep
->freelist_size
= freelist_size
;
2383 cachep
->flags
= flags
;
2384 cachep
->allocflags
= __GFP_COMP
;
2385 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2386 cachep
->allocflags
|= GFP_DMA
;
2387 cachep
->size
= size
;
2388 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2390 if (flags
& CFLGS_OFF_SLAB
) {
2391 cachep
->freelist_cache
= kmalloc_slab(freelist_size
, 0u);
2393 * This is a possibility for one of the kmalloc_{dma,}_caches.
2394 * But since we go off slab only for object size greater than
2395 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2396 * in ascending order,this should not happen at all.
2397 * But leave a BUG_ON for some lucky dude.
2399 BUG_ON(ZERO_OR_NULL_PTR(cachep
->freelist_cache
));
2402 err
= setup_cpu_cache(cachep
, gfp
);
2404 __kmem_cache_shutdown(cachep
);
2408 if (flags
& SLAB_DEBUG_OBJECTS
) {
2410 * Would deadlock through slab_destroy()->call_rcu()->
2411 * debug_object_activate()->kmem_cache_alloc().
2413 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2415 slab_set_debugobj_lock_classes(cachep
);
2416 } else if (!OFF_SLAB(cachep
) && !(flags
& SLAB_DESTROY_BY_RCU
))
2417 on_slab_lock_classes(cachep
);
2423 static void check_irq_off(void)
2425 BUG_ON(!irqs_disabled());
2428 static void check_irq_on(void)
2430 BUG_ON(irqs_disabled());
2433 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2437 assert_spin_locked(&cachep
->node
[numa_mem_id()]->list_lock
);
2441 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2445 assert_spin_locked(&cachep
->node
[node
]->list_lock
);
2450 #define check_irq_off() do { } while(0)
2451 #define check_irq_on() do { } while(0)
2452 #define check_spinlock_acquired(x) do { } while(0)
2453 #define check_spinlock_acquired_node(x, y) do { } while(0)
2456 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2457 struct array_cache
*ac
,
2458 int force
, int node
);
2460 static void do_drain(void *arg
)
2462 struct kmem_cache
*cachep
= arg
;
2463 struct array_cache
*ac
;
2464 int node
= numa_mem_id();
2467 ac
= cpu_cache_get(cachep
);
2468 spin_lock(&cachep
->node
[node
]->list_lock
);
2469 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2470 spin_unlock(&cachep
->node
[node
]->list_lock
);
2474 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2476 struct kmem_cache_node
*n
;
2479 on_each_cpu(do_drain
, cachep
, 1);
2481 for_each_online_node(node
) {
2482 n
= cachep
->node
[node
];
2484 drain_alien_cache(cachep
, n
->alien
);
2487 for_each_online_node(node
) {
2488 n
= cachep
->node
[node
];
2490 drain_array(cachep
, n
, n
->shared
, 1, node
);
2495 * Remove slabs from the list of free slabs.
2496 * Specify the number of slabs to drain in tofree.
2498 * Returns the actual number of slabs released.
2500 static int drain_freelist(struct kmem_cache
*cache
,
2501 struct kmem_cache_node
*n
, int tofree
)
2503 struct list_head
*p
;
2508 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2510 spin_lock_irq(&n
->list_lock
);
2511 p
= n
->slabs_free
.prev
;
2512 if (p
== &n
->slabs_free
) {
2513 spin_unlock_irq(&n
->list_lock
);
2517 page
= list_entry(p
, struct page
, lru
);
2519 BUG_ON(page
->active
);
2521 list_del(&page
->lru
);
2523 * Safe to drop the lock. The slab is no longer linked
2526 n
->free_objects
-= cache
->num
;
2527 spin_unlock_irq(&n
->list_lock
);
2528 slab_destroy(cache
, page
);
2535 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2538 struct kmem_cache_node
*n
;
2540 drain_cpu_caches(cachep
);
2543 for_each_online_node(i
) {
2544 n
= cachep
->node
[i
];
2548 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2550 ret
+= !list_empty(&n
->slabs_full
) ||
2551 !list_empty(&n
->slabs_partial
);
2553 return (ret
? 1 : 0);
2556 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2559 struct kmem_cache_node
*n
;
2560 int rc
= __kmem_cache_shrink(cachep
);
2565 for_each_online_cpu(i
)
2566 kfree(cachep
->array
[i
]);
2568 /* NUMA: free the node structures */
2569 for_each_online_node(i
) {
2570 n
= cachep
->node
[i
];
2573 free_alien_cache(n
->alien
);
2581 * Get the memory for a slab management obj.
2583 * For a slab cache when the slab descriptor is off-slab, the
2584 * slab descriptor can't come from the same cache which is being created,
2585 * Because if it is the case, that means we defer the creation of
2586 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2587 * And we eventually call down to __kmem_cache_create(), which
2588 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2589 * This is a "chicken-and-egg" problem.
2591 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2592 * which are all initialized during kmem_cache_init().
2594 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2595 struct page
*page
, int colour_off
,
2596 gfp_t local_flags
, int nodeid
)
2599 void *addr
= page_address(page
);
2601 if (OFF_SLAB(cachep
)) {
2602 /* Slab management obj is off-slab. */
2603 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2604 local_flags
, nodeid
);
2608 freelist
= addr
+ colour_off
;
2609 colour_off
+= cachep
->freelist_size
;
2612 page
->s_mem
= addr
+ colour_off
;
2616 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2618 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2621 static inline void set_free_obj(struct page
*page
,
2622 unsigned int idx
, freelist_idx_t val
)
2624 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2627 static void cache_init_objs(struct kmem_cache
*cachep
,
2632 for (i
= 0; i
< cachep
->num
; i
++) {
2633 void *objp
= index_to_obj(cachep
, page
, i
);
2635 /* need to poison the objs? */
2636 if (cachep
->flags
& SLAB_POISON
)
2637 poison_obj(cachep
, objp
, POISON_FREE
);
2638 if (cachep
->flags
& SLAB_STORE_USER
)
2639 *dbg_userword(cachep
, objp
) = NULL
;
2641 if (cachep
->flags
& SLAB_RED_ZONE
) {
2642 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2643 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2646 * Constructors are not allowed to allocate memory from the same
2647 * cache which they are a constructor for. Otherwise, deadlock.
2648 * They must also be threaded.
2650 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2651 cachep
->ctor(objp
+ obj_offset(cachep
));
2653 if (cachep
->flags
& SLAB_RED_ZONE
) {
2654 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2655 slab_error(cachep
, "constructor overwrote the"
2656 " end of an object");
2657 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2658 slab_error(cachep
, "constructor overwrote the"
2659 " start of an object");
2661 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2662 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2663 kernel_map_pages(virt_to_page(objp
),
2664 cachep
->size
/ PAGE_SIZE
, 0);
2669 set_obj_status(page
, i
, OBJECT_FREE
);
2670 set_free_obj(page
, i
, i
);
2674 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2676 if (CONFIG_ZONE_DMA_FLAG
) {
2677 if (flags
& GFP_DMA
)
2678 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2680 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2684 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
,
2689 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2692 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2698 static void slab_put_obj(struct kmem_cache
*cachep
, struct page
*page
,
2699 void *objp
, int nodeid
)
2701 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2705 /* Verify that the slab belongs to the intended node */
2706 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2708 /* Verify double free bug */
2709 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2710 if (get_free_obj(page
, i
) == objnr
) {
2711 printk(KERN_ERR
"slab: double free detected in cache "
2712 "'%s', objp %p\n", cachep
->name
, objp
);
2718 set_free_obj(page
, page
->active
, objnr
);
2722 * Map pages beginning at addr to the given cache and slab. This is required
2723 * for the slab allocator to be able to lookup the cache and slab of a
2724 * virtual address for kfree, ksize, and slab debugging.
2726 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2729 page
->slab_cache
= cache
;
2730 page
->freelist
= freelist
;
2734 * Grow (by 1) the number of slabs within a cache. This is called by
2735 * kmem_cache_alloc() when there are no active objs left in a cache.
2737 static int cache_grow(struct kmem_cache
*cachep
,
2738 gfp_t flags
, int nodeid
, struct page
*page
)
2743 struct kmem_cache_node
*n
;
2746 * Be lazy and only check for valid flags here, keeping it out of the
2747 * critical path in kmem_cache_alloc().
2749 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2750 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2752 /* Take the node list lock to change the colour_next on this node */
2754 n
= cachep
->node
[nodeid
];
2755 spin_lock(&n
->list_lock
);
2757 /* Get colour for the slab, and cal the next value. */
2758 offset
= n
->colour_next
;
2760 if (n
->colour_next
>= cachep
->colour
)
2762 spin_unlock(&n
->list_lock
);
2764 offset
*= cachep
->colour_off
;
2766 if (local_flags
& __GFP_WAIT
)
2770 * The test for missing atomic flag is performed here, rather than
2771 * the more obvious place, simply to reduce the critical path length
2772 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2773 * will eventually be caught here (where it matters).
2775 kmem_flagcheck(cachep
, flags
);
2778 * Get mem for the objs. Attempt to allocate a physical page from
2782 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2786 /* Get slab management. */
2787 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2788 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2792 slab_map_pages(cachep
, page
, freelist
);
2794 cache_init_objs(cachep
, page
);
2796 if (local_flags
& __GFP_WAIT
)
2797 local_irq_disable();
2799 spin_lock(&n
->list_lock
);
2801 /* Make slab active. */
2802 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2803 STATS_INC_GROWN(cachep
);
2804 n
->free_objects
+= cachep
->num
;
2805 spin_unlock(&n
->list_lock
);
2808 kmem_freepages(cachep
, page
);
2810 if (local_flags
& __GFP_WAIT
)
2811 local_irq_disable();
2818 * Perform extra freeing checks:
2819 * - detect bad pointers.
2820 * - POISON/RED_ZONE checking
2822 static void kfree_debugcheck(const void *objp
)
2824 if (!virt_addr_valid(objp
)) {
2825 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2826 (unsigned long)objp
);
2831 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2833 unsigned long long redzone1
, redzone2
;
2835 redzone1
= *dbg_redzone1(cache
, obj
);
2836 redzone2
= *dbg_redzone2(cache
, obj
);
2841 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2844 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2845 slab_error(cache
, "double free detected");
2847 slab_error(cache
, "memory outside object was overwritten");
2849 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2850 obj
, redzone1
, redzone2
);
2853 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2854 unsigned long caller
)
2859 BUG_ON(virt_to_cache(objp
) != cachep
);
2861 objp
-= obj_offset(cachep
);
2862 kfree_debugcheck(objp
);
2863 page
= virt_to_head_page(objp
);
2865 if (cachep
->flags
& SLAB_RED_ZONE
) {
2866 verify_redzone_free(cachep
, objp
);
2867 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2868 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2870 if (cachep
->flags
& SLAB_STORE_USER
)
2871 *dbg_userword(cachep
, objp
) = (void *)caller
;
2873 objnr
= obj_to_index(cachep
, page
, objp
);
2875 BUG_ON(objnr
>= cachep
->num
);
2876 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2878 set_obj_status(page
, objnr
, OBJECT_FREE
);
2879 if (cachep
->flags
& SLAB_POISON
) {
2880 #ifdef CONFIG_DEBUG_PAGEALLOC
2881 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2882 store_stackinfo(cachep
, objp
, caller
);
2883 kernel_map_pages(virt_to_page(objp
),
2884 cachep
->size
/ PAGE_SIZE
, 0);
2886 poison_obj(cachep
, objp
, POISON_FREE
);
2889 poison_obj(cachep
, objp
, POISON_FREE
);
2896 #define kfree_debugcheck(x) do { } while(0)
2897 #define cache_free_debugcheck(x,objp,z) (objp)
2900 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2904 struct kmem_cache_node
*n
;
2905 struct array_cache
*ac
;
2909 node
= numa_mem_id();
2910 if (unlikely(force_refill
))
2913 ac
= cpu_cache_get(cachep
);
2914 batchcount
= ac
->batchcount
;
2915 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2917 * If there was little recent activity on this cache, then
2918 * perform only a partial refill. Otherwise we could generate
2921 batchcount
= BATCHREFILL_LIMIT
;
2923 n
= cachep
->node
[node
];
2925 BUG_ON(ac
->avail
> 0 || !n
);
2926 spin_lock(&n
->list_lock
);
2928 /* See if we can refill from the shared array */
2929 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2930 n
->shared
->touched
= 1;
2934 while (batchcount
> 0) {
2935 struct list_head
*entry
;
2937 /* Get slab alloc is to come from. */
2938 entry
= n
->slabs_partial
.next
;
2939 if (entry
== &n
->slabs_partial
) {
2940 n
->free_touched
= 1;
2941 entry
= n
->slabs_free
.next
;
2942 if (entry
== &n
->slabs_free
)
2946 page
= list_entry(entry
, struct page
, lru
);
2947 check_spinlock_acquired(cachep
);
2950 * The slab was either on partial or free list so
2951 * there must be at least one object available for
2954 BUG_ON(page
->active
>= cachep
->num
);
2956 while (page
->active
< cachep
->num
&& batchcount
--) {
2957 STATS_INC_ALLOCED(cachep
);
2958 STATS_INC_ACTIVE(cachep
);
2959 STATS_SET_HIGH(cachep
);
2961 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, page
,
2965 /* move slabp to correct slabp list: */
2966 list_del(&page
->lru
);
2967 if (page
->active
== cachep
->num
)
2968 list_add(&page
->lru
, &n
->slabs_full
);
2970 list_add(&page
->lru
, &n
->slabs_partial
);
2974 n
->free_objects
-= ac
->avail
;
2976 spin_unlock(&n
->list_lock
);
2978 if (unlikely(!ac
->avail
)) {
2981 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2983 /* cache_grow can reenable interrupts, then ac could change. */
2984 ac
= cpu_cache_get(cachep
);
2985 node
= numa_mem_id();
2987 /* no objects in sight? abort */
2988 if (!x
&& (ac
->avail
== 0 || force_refill
))
2991 if (!ac
->avail
) /* objects refilled by interrupt? */
2996 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2999 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3002 might_sleep_if(flags
& __GFP_WAIT
);
3004 kmem_flagcheck(cachep
, flags
);
3009 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3010 gfp_t flags
, void *objp
, unsigned long caller
)
3016 if (cachep
->flags
& SLAB_POISON
) {
3017 #ifdef CONFIG_DEBUG_PAGEALLOC
3018 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
3019 kernel_map_pages(virt_to_page(objp
),
3020 cachep
->size
/ PAGE_SIZE
, 1);
3022 check_poison_obj(cachep
, objp
);
3024 check_poison_obj(cachep
, objp
);
3026 poison_obj(cachep
, objp
, POISON_INUSE
);
3028 if (cachep
->flags
& SLAB_STORE_USER
)
3029 *dbg_userword(cachep
, objp
) = (void *)caller
;
3031 if (cachep
->flags
& SLAB_RED_ZONE
) {
3032 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3033 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3034 slab_error(cachep
, "double free, or memory outside"
3035 " object was overwritten");
3037 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3038 objp
, *dbg_redzone1(cachep
, objp
),
3039 *dbg_redzone2(cachep
, objp
));
3041 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3042 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3045 page
= virt_to_head_page(objp
);
3046 set_obj_status(page
, obj_to_index(cachep
, page
, objp
), OBJECT_ACTIVE
);
3047 objp
+= obj_offset(cachep
);
3048 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3050 if (ARCH_SLAB_MINALIGN
&&
3051 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3052 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3053 objp
, (int)ARCH_SLAB_MINALIGN
);
3058 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3061 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3063 if (cachep
== kmem_cache
)
3066 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3069 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3072 struct array_cache
*ac
;
3073 bool force_refill
= false;
3077 ac
= cpu_cache_get(cachep
);
3078 if (likely(ac
->avail
)) {
3080 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3083 * Allow for the possibility all avail objects are not allowed
3084 * by the current flags
3087 STATS_INC_ALLOCHIT(cachep
);
3090 force_refill
= true;
3093 STATS_INC_ALLOCMISS(cachep
);
3094 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3096 * the 'ac' may be updated by cache_alloc_refill(),
3097 * and kmemleak_erase() requires its correct value.
3099 ac
= cpu_cache_get(cachep
);
3103 * To avoid a false negative, if an object that is in one of the
3104 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3105 * treat the array pointers as a reference to the object.
3108 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3114 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
3116 * If we are in_interrupt, then process context, including cpusets and
3117 * mempolicy, may not apply and should not be used for allocation policy.
3119 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3121 int nid_alloc
, nid_here
;
3123 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3125 nid_alloc
= nid_here
= numa_mem_id();
3126 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3127 nid_alloc
= cpuset_slab_spread_node();
3128 else if (current
->mempolicy
)
3129 nid_alloc
= mempolicy_slab_node();
3130 if (nid_alloc
!= nid_here
)
3131 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3136 * Fallback function if there was no memory available and no objects on a
3137 * certain node and fall back is permitted. First we scan all the
3138 * available node for available objects. If that fails then we
3139 * perform an allocation without specifying a node. This allows the page
3140 * allocator to do its reclaim / fallback magic. We then insert the
3141 * slab into the proper nodelist and then allocate from it.
3143 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3145 struct zonelist
*zonelist
;
3149 enum zone_type high_zoneidx
= gfp_zone(flags
);
3152 unsigned int cpuset_mems_cookie
;
3154 if (flags
& __GFP_THISNODE
)
3157 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3160 cpuset_mems_cookie
= read_mems_allowed_begin();
3161 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3165 * Look through allowed nodes for objects available
3166 * from existing per node queues.
3168 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3169 nid
= zone_to_nid(zone
);
3171 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3173 cache
->node
[nid
]->free_objects
) {
3174 obj
= ____cache_alloc_node(cache
,
3175 flags
| GFP_THISNODE
, nid
);
3183 * This allocation will be performed within the constraints
3184 * of the current cpuset / memory policy requirements.
3185 * We may trigger various forms of reclaim on the allowed
3186 * set and go into memory reserves if necessary.
3190 if (local_flags
& __GFP_WAIT
)
3192 kmem_flagcheck(cache
, flags
);
3193 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3194 if (local_flags
& __GFP_WAIT
)
3195 local_irq_disable();
3198 * Insert into the appropriate per node queues
3200 nid
= page_to_nid(page
);
3201 if (cache_grow(cache
, flags
, nid
, page
)) {
3202 obj
= ____cache_alloc_node(cache
,
3203 flags
| GFP_THISNODE
, nid
);
3206 * Another processor may allocate the
3207 * objects in the slab since we are
3208 * not holding any locks.
3212 /* cache_grow already freed obj */
3218 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3224 * A interface to enable slab creation on nodeid
3226 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3229 struct list_head
*entry
;
3231 struct kmem_cache_node
*n
;
3235 VM_BUG_ON(nodeid
> num_online_nodes());
3236 n
= cachep
->node
[nodeid
];
3241 spin_lock(&n
->list_lock
);
3242 entry
= n
->slabs_partial
.next
;
3243 if (entry
== &n
->slabs_partial
) {
3244 n
->free_touched
= 1;
3245 entry
= n
->slabs_free
.next
;
3246 if (entry
== &n
->slabs_free
)
3250 page
= list_entry(entry
, struct page
, lru
);
3251 check_spinlock_acquired_node(cachep
, nodeid
);
3253 STATS_INC_NODEALLOCS(cachep
);
3254 STATS_INC_ACTIVE(cachep
);
3255 STATS_SET_HIGH(cachep
);
3257 BUG_ON(page
->active
== cachep
->num
);
3259 obj
= slab_get_obj(cachep
, page
, nodeid
);
3261 /* move slabp to correct slabp list: */
3262 list_del(&page
->lru
);
3264 if (page
->active
== cachep
->num
)
3265 list_add(&page
->lru
, &n
->slabs_full
);
3267 list_add(&page
->lru
, &n
->slabs_partial
);
3269 spin_unlock(&n
->list_lock
);
3273 spin_unlock(&n
->list_lock
);
3274 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3278 return fallback_alloc(cachep
, flags
);
3284 static __always_inline
void *
3285 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3286 unsigned long caller
)
3288 unsigned long save_flags
;
3290 int slab_node
= numa_mem_id();
3292 flags
&= gfp_allowed_mask
;
3294 lockdep_trace_alloc(flags
);
3296 if (slab_should_failslab(cachep
, flags
))
3299 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3301 cache_alloc_debugcheck_before(cachep
, flags
);
3302 local_irq_save(save_flags
);
3304 if (nodeid
== NUMA_NO_NODE
)
3307 if (unlikely(!cachep
->node
[nodeid
])) {
3308 /* Node not bootstrapped yet */
3309 ptr
= fallback_alloc(cachep
, flags
);
3313 if (nodeid
== slab_node
) {
3315 * Use the locally cached objects if possible.
3316 * However ____cache_alloc does not allow fallback
3317 * to other nodes. It may fail while we still have
3318 * objects on other nodes available.
3320 ptr
= ____cache_alloc(cachep
, flags
);
3324 /* ___cache_alloc_node can fall back to other nodes */
3325 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3327 local_irq_restore(save_flags
);
3328 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3329 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3333 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3334 if (unlikely(flags
& __GFP_ZERO
))
3335 memset(ptr
, 0, cachep
->object_size
);
3341 static __always_inline
void *
3342 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3346 if (current
->mempolicy
|| unlikely(current
->flags
& PF_SPREAD_SLAB
)) {
3347 objp
= alternate_node_alloc(cache
, flags
);
3351 objp
= ____cache_alloc(cache
, flags
);
3354 * We may just have run out of memory on the local node.
3355 * ____cache_alloc_node() knows how to locate memory on other nodes
3358 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3365 static __always_inline
void *
3366 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3368 return ____cache_alloc(cachep
, flags
);
3371 #endif /* CONFIG_NUMA */
3373 static __always_inline
void *
3374 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3376 unsigned long save_flags
;
3379 flags
&= gfp_allowed_mask
;
3381 lockdep_trace_alloc(flags
);
3383 if (slab_should_failslab(cachep
, flags
))
3386 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3388 cache_alloc_debugcheck_before(cachep
, flags
);
3389 local_irq_save(save_flags
);
3390 objp
= __do_cache_alloc(cachep
, flags
);
3391 local_irq_restore(save_flags
);
3392 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3393 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3398 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3399 if (unlikely(flags
& __GFP_ZERO
))
3400 memset(objp
, 0, cachep
->object_size
);
3407 * Caller needs to acquire correct kmem_cache_node's list_lock
3409 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3413 struct kmem_cache_node
*n
;
3415 for (i
= 0; i
< nr_objects
; i
++) {
3419 clear_obj_pfmemalloc(&objpp
[i
]);
3422 page
= virt_to_head_page(objp
);
3423 n
= cachep
->node
[node
];
3424 list_del(&page
->lru
);
3425 check_spinlock_acquired_node(cachep
, node
);
3426 slab_put_obj(cachep
, page
, objp
, node
);
3427 STATS_DEC_ACTIVE(cachep
);
3430 /* fixup slab chains */
3431 if (page
->active
== 0) {
3432 if (n
->free_objects
> n
->free_limit
) {
3433 n
->free_objects
-= cachep
->num
;
3434 /* No need to drop any previously held
3435 * lock here, even if we have a off-slab slab
3436 * descriptor it is guaranteed to come from
3437 * a different cache, refer to comments before
3440 slab_destroy(cachep
, page
);
3442 list_add(&page
->lru
, &n
->slabs_free
);
3445 /* Unconditionally move a slab to the end of the
3446 * partial list on free - maximum time for the
3447 * other objects to be freed, too.
3449 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3454 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3457 struct kmem_cache_node
*n
;
3458 int node
= numa_mem_id();
3460 batchcount
= ac
->batchcount
;
3462 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3465 n
= cachep
->node
[node
];
3466 spin_lock(&n
->list_lock
);
3468 struct array_cache
*shared_array
= n
->shared
;
3469 int max
= shared_array
->limit
- shared_array
->avail
;
3471 if (batchcount
> max
)
3473 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3474 ac
->entry
, sizeof(void *) * batchcount
);
3475 shared_array
->avail
+= batchcount
;
3480 free_block(cachep
, ac
->entry
, batchcount
, node
);
3485 struct list_head
*p
;
3487 p
= n
->slabs_free
.next
;
3488 while (p
!= &(n
->slabs_free
)) {
3491 page
= list_entry(p
, struct page
, lru
);
3492 BUG_ON(page
->active
);
3497 STATS_SET_FREEABLE(cachep
, i
);
3500 spin_unlock(&n
->list_lock
);
3501 ac
->avail
-= batchcount
;
3502 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3506 * Release an obj back to its cache. If the obj has a constructed state, it must
3507 * be in this state _before_ it is released. Called with disabled ints.
3509 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3510 unsigned long caller
)
3512 struct array_cache
*ac
= cpu_cache_get(cachep
);
3515 kmemleak_free_recursive(objp
, cachep
->flags
);
3516 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3518 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3521 * Skip calling cache_free_alien() when the platform is not numa.
3522 * This will avoid cache misses that happen while accessing slabp (which
3523 * is per page memory reference) to get nodeid. Instead use a global
3524 * variable to skip the call, which is mostly likely to be present in
3527 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3530 if (likely(ac
->avail
< ac
->limit
)) {
3531 STATS_INC_FREEHIT(cachep
);
3533 STATS_INC_FREEMISS(cachep
);
3534 cache_flusharray(cachep
, ac
);
3537 ac_put_obj(cachep
, ac
, objp
);
3541 * kmem_cache_alloc - Allocate an object
3542 * @cachep: The cache to allocate from.
3543 * @flags: See kmalloc().
3545 * Allocate an object from this cache. The flags are only relevant
3546 * if the cache has no available objects.
3548 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3550 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3552 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3553 cachep
->object_size
, cachep
->size
, flags
);
3557 EXPORT_SYMBOL(kmem_cache_alloc
);
3559 #ifdef CONFIG_TRACING
3561 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3565 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3567 trace_kmalloc(_RET_IP_
, ret
,
3568 size
, cachep
->size
, flags
);
3571 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3576 * kmem_cache_alloc_node - Allocate an object on the specified node
3577 * @cachep: The cache to allocate from.
3578 * @flags: See kmalloc().
3579 * @nodeid: node number of the target node.
3581 * Identical to kmem_cache_alloc but it will allocate memory on the given
3582 * node, which can improve the performance for cpu bound structures.
3584 * Fallback to other node is possible if __GFP_THISNODE is not set.
3586 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3588 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3590 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3591 cachep
->object_size
, cachep
->size
,
3596 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3598 #ifdef CONFIG_TRACING
3599 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3606 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3608 trace_kmalloc_node(_RET_IP_
, ret
,
3613 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3616 static __always_inline
void *
3617 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3619 struct kmem_cache
*cachep
;
3621 cachep
= kmalloc_slab(size
, flags
);
3622 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3624 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3627 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3628 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3630 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3632 EXPORT_SYMBOL(__kmalloc_node
);
3634 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3635 int node
, unsigned long caller
)
3637 return __do_kmalloc_node(size
, flags
, node
, caller
);
3639 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3641 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3643 return __do_kmalloc_node(size
, flags
, node
, 0);
3645 EXPORT_SYMBOL(__kmalloc_node
);
3646 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3647 #endif /* CONFIG_NUMA */
3650 * __do_kmalloc - allocate memory
3651 * @size: how many bytes of memory are required.
3652 * @flags: the type of memory to allocate (see kmalloc).
3653 * @caller: function caller for debug tracking of the caller
3655 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3656 unsigned long caller
)
3658 struct kmem_cache
*cachep
;
3661 cachep
= kmalloc_slab(size
, flags
);
3662 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3664 ret
= slab_alloc(cachep
, flags
, caller
);
3666 trace_kmalloc(caller
, ret
,
3667 size
, cachep
->size
, flags
);
3673 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3674 void *__kmalloc(size_t size
, gfp_t flags
)
3676 return __do_kmalloc(size
, flags
, _RET_IP_
);
3678 EXPORT_SYMBOL(__kmalloc
);
3680 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3682 return __do_kmalloc(size
, flags
, caller
);
3684 EXPORT_SYMBOL(__kmalloc_track_caller
);
3687 void *__kmalloc(size_t size
, gfp_t flags
)
3689 return __do_kmalloc(size
, flags
, 0);
3691 EXPORT_SYMBOL(__kmalloc
);
3695 * kmem_cache_free - Deallocate an object
3696 * @cachep: The cache the allocation was from.
3697 * @objp: The previously allocated object.
3699 * Free an object which was previously allocated from this
3702 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3704 unsigned long flags
;
3705 cachep
= cache_from_obj(cachep
, objp
);
3709 local_irq_save(flags
);
3710 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3711 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3712 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3713 __cache_free(cachep
, objp
, _RET_IP_
);
3714 local_irq_restore(flags
);
3716 trace_kmem_cache_free(_RET_IP_
, objp
);
3718 EXPORT_SYMBOL(kmem_cache_free
);
3721 * kfree - free previously allocated memory
3722 * @objp: pointer returned by kmalloc.
3724 * If @objp is NULL, no operation is performed.
3726 * Don't free memory not originally allocated by kmalloc()
3727 * or you will run into trouble.
3729 void kfree(const void *objp
)
3731 struct kmem_cache
*c
;
3732 unsigned long flags
;
3734 trace_kfree(_RET_IP_
, objp
);
3736 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3738 local_irq_save(flags
);
3739 kfree_debugcheck(objp
);
3740 c
= virt_to_cache(objp
);
3741 debug_check_no_locks_freed(objp
, c
->object_size
);
3743 debug_check_no_obj_freed(objp
, c
->object_size
);
3744 __cache_free(c
, (void *)objp
, _RET_IP_
);
3745 local_irq_restore(flags
);
3747 EXPORT_SYMBOL(kfree
);
3750 * This initializes kmem_cache_node or resizes various caches for all nodes.
3752 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3755 struct kmem_cache_node
*n
;
3756 struct array_cache
*new_shared
;
3757 struct array_cache
**new_alien
= NULL
;
3759 for_each_online_node(node
) {
3761 if (use_alien_caches
) {
3762 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3768 if (cachep
->shared
) {
3769 new_shared
= alloc_arraycache(node
,
3770 cachep
->shared
*cachep
->batchcount
,
3773 free_alien_cache(new_alien
);
3778 n
= cachep
->node
[node
];
3780 struct array_cache
*shared
= n
->shared
;
3782 spin_lock_irq(&n
->list_lock
);
3785 free_block(cachep
, shared
->entry
,
3786 shared
->avail
, node
);
3788 n
->shared
= new_shared
;
3790 n
->alien
= new_alien
;
3793 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3794 cachep
->batchcount
+ cachep
->num
;
3795 spin_unlock_irq(&n
->list_lock
);
3797 free_alien_cache(new_alien
);
3800 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3802 free_alien_cache(new_alien
);
3807 kmem_cache_node_init(n
);
3808 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3809 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3810 n
->shared
= new_shared
;
3811 n
->alien
= new_alien
;
3812 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3813 cachep
->batchcount
+ cachep
->num
;
3814 cachep
->node
[node
] = n
;
3819 if (!cachep
->list
.next
) {
3820 /* Cache is not active yet. Roll back what we did */
3823 if (cachep
->node
[node
]) {
3824 n
= cachep
->node
[node
];
3827 free_alien_cache(n
->alien
);
3829 cachep
->node
[node
] = NULL
;
3837 struct ccupdate_struct
{
3838 struct kmem_cache
*cachep
;
3839 struct array_cache
*new[0];
3842 static void do_ccupdate_local(void *info
)
3844 struct ccupdate_struct
*new = info
;
3845 struct array_cache
*old
;
3848 old
= cpu_cache_get(new->cachep
);
3850 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3851 new->new[smp_processor_id()] = old
;
3854 /* Always called with the slab_mutex held */
3855 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3856 int batchcount
, int shared
, gfp_t gfp
)
3858 struct ccupdate_struct
*new;
3861 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3866 for_each_online_cpu(i
) {
3867 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3870 for (i
--; i
>= 0; i
--)
3876 new->cachep
= cachep
;
3878 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3881 cachep
->batchcount
= batchcount
;
3882 cachep
->limit
= limit
;
3883 cachep
->shared
= shared
;
3885 for_each_online_cpu(i
) {
3886 struct array_cache
*ccold
= new->new[i
];
3889 spin_lock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3890 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
3891 spin_unlock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3895 return alloc_kmem_cache_node(cachep
, gfp
);
3898 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3899 int batchcount
, int shared
, gfp_t gfp
)
3902 struct kmem_cache
*c
= NULL
;
3905 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3907 if (slab_state
< FULL
)
3910 if ((ret
< 0) || !is_root_cache(cachep
))
3913 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3914 for_each_memcg_cache_index(i
) {
3915 c
= cache_from_memcg_idx(cachep
, i
);
3917 /* return value determined by the parent cache only */
3918 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3924 /* Called with slab_mutex held always */
3925 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3932 if (!is_root_cache(cachep
)) {
3933 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3934 limit
= root
->limit
;
3935 shared
= root
->shared
;
3936 batchcount
= root
->batchcount
;
3939 if (limit
&& shared
&& batchcount
)
3942 * The head array serves three purposes:
3943 * - create a LIFO ordering, i.e. return objects that are cache-warm
3944 * - reduce the number of spinlock operations.
3945 * - reduce the number of linked list operations on the slab and
3946 * bufctl chains: array operations are cheaper.
3947 * The numbers are guessed, we should auto-tune as described by
3950 if (cachep
->size
> 131072)
3952 else if (cachep
->size
> PAGE_SIZE
)
3954 else if (cachep
->size
> 1024)
3956 else if (cachep
->size
> 256)
3962 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3963 * allocation behaviour: Most allocs on one cpu, most free operations
3964 * on another cpu. For these cases, an efficient object passing between
3965 * cpus is necessary. This is provided by a shared array. The array
3966 * replaces Bonwick's magazine layer.
3967 * On uniprocessor, it's functionally equivalent (but less efficient)
3968 * to a larger limit. Thus disabled by default.
3971 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3976 * With debugging enabled, large batchcount lead to excessively long
3977 * periods with disabled local interrupts. Limit the batchcount
3982 batchcount
= (limit
+ 1) / 2;
3984 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3986 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3987 cachep
->name
, -err
);
3992 * Drain an array if it contains any elements taking the node lock only if
3993 * necessary. Note that the node listlock also protects the array_cache
3994 * if drain_array() is used on the shared array.
3996 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3997 struct array_cache
*ac
, int force
, int node
)
4001 if (!ac
|| !ac
->avail
)
4003 if (ac
->touched
&& !force
) {
4006 spin_lock_irq(&n
->list_lock
);
4008 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
4009 if (tofree
> ac
->avail
)
4010 tofree
= (ac
->avail
+ 1) / 2;
4011 free_block(cachep
, ac
->entry
, tofree
, node
);
4012 ac
->avail
-= tofree
;
4013 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
4014 sizeof(void *) * ac
->avail
);
4016 spin_unlock_irq(&n
->list_lock
);
4021 * cache_reap - Reclaim memory from caches.
4022 * @w: work descriptor
4024 * Called from workqueue/eventd every few seconds.
4026 * - clear the per-cpu caches for this CPU.
4027 * - return freeable pages to the main free memory pool.
4029 * If we cannot acquire the cache chain mutex then just give up - we'll try
4030 * again on the next iteration.
4032 static void cache_reap(struct work_struct
*w
)
4034 struct kmem_cache
*searchp
;
4035 struct kmem_cache_node
*n
;
4036 int node
= numa_mem_id();
4037 struct delayed_work
*work
= to_delayed_work(w
);
4039 if (!mutex_trylock(&slab_mutex
))
4040 /* Give up. Setup the next iteration. */
4043 list_for_each_entry(searchp
, &slab_caches
, list
) {
4047 * We only take the node lock if absolutely necessary and we
4048 * have established with reasonable certainty that
4049 * we can do some work if the lock was obtained.
4051 n
= searchp
->node
[node
];
4053 reap_alien(searchp
, n
);
4055 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
4058 * These are racy checks but it does not matter
4059 * if we skip one check or scan twice.
4061 if (time_after(n
->next_reap
, jiffies
))
4064 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
4066 drain_array(searchp
, n
, n
->shared
, 0, node
);
4068 if (n
->free_touched
)
4069 n
->free_touched
= 0;
4073 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4074 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4075 STATS_ADD_REAPED(searchp
, freed
);
4081 mutex_unlock(&slab_mutex
);
4084 /* Set up the next iteration */
4085 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
4088 #ifdef CONFIG_SLABINFO
4089 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4092 unsigned long active_objs
;
4093 unsigned long num_objs
;
4094 unsigned long active_slabs
= 0;
4095 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4099 struct kmem_cache_node
*n
;
4103 for_each_online_node(node
) {
4104 n
= cachep
->node
[node
];
4109 spin_lock_irq(&n
->list_lock
);
4111 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4112 if (page
->active
!= cachep
->num
&& !error
)
4113 error
= "slabs_full accounting error";
4114 active_objs
+= cachep
->num
;
4117 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4118 if (page
->active
== cachep
->num
&& !error
)
4119 error
= "slabs_partial accounting error";
4120 if (!page
->active
&& !error
)
4121 error
= "slabs_partial accounting error";
4122 active_objs
+= page
->active
;
4125 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4126 if (page
->active
&& !error
)
4127 error
= "slabs_free accounting error";
4130 free_objects
+= n
->free_objects
;
4132 shared_avail
+= n
->shared
->avail
;
4134 spin_unlock_irq(&n
->list_lock
);
4136 num_slabs
+= active_slabs
;
4137 num_objs
= num_slabs
* cachep
->num
;
4138 if (num_objs
- active_objs
!= free_objects
&& !error
)
4139 error
= "free_objects accounting error";
4141 name
= cachep
->name
;
4143 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4145 sinfo
->active_objs
= active_objs
;
4146 sinfo
->num_objs
= num_objs
;
4147 sinfo
->active_slabs
= active_slabs
;
4148 sinfo
->num_slabs
= num_slabs
;
4149 sinfo
->shared_avail
= shared_avail
;
4150 sinfo
->limit
= cachep
->limit
;
4151 sinfo
->batchcount
= cachep
->batchcount
;
4152 sinfo
->shared
= cachep
->shared
;
4153 sinfo
->objects_per_slab
= cachep
->num
;
4154 sinfo
->cache_order
= cachep
->gfporder
;
4157 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4161 unsigned long high
= cachep
->high_mark
;
4162 unsigned long allocs
= cachep
->num_allocations
;
4163 unsigned long grown
= cachep
->grown
;
4164 unsigned long reaped
= cachep
->reaped
;
4165 unsigned long errors
= cachep
->errors
;
4166 unsigned long max_freeable
= cachep
->max_freeable
;
4167 unsigned long node_allocs
= cachep
->node_allocs
;
4168 unsigned long node_frees
= cachep
->node_frees
;
4169 unsigned long overflows
= cachep
->node_overflow
;
4171 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4172 "%4lu %4lu %4lu %4lu %4lu",
4173 allocs
, high
, grown
,
4174 reaped
, errors
, max_freeable
, node_allocs
,
4175 node_frees
, overflows
);
4179 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4180 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4181 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4182 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4184 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4185 allochit
, allocmiss
, freehit
, freemiss
);
4190 #define MAX_SLABINFO_WRITE 128
4192 * slabinfo_write - Tuning for the slab allocator
4194 * @buffer: user buffer
4195 * @count: data length
4198 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4199 size_t count
, loff_t
*ppos
)
4201 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4202 int limit
, batchcount
, shared
, res
;
4203 struct kmem_cache
*cachep
;
4205 if (count
> MAX_SLABINFO_WRITE
)
4207 if (copy_from_user(&kbuf
, buffer
, count
))
4209 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4211 tmp
= strchr(kbuf
, ' ');
4216 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4219 /* Find the cache in the chain of caches. */
4220 mutex_lock(&slab_mutex
);
4222 list_for_each_entry(cachep
, &slab_caches
, list
) {
4223 if (!strcmp(cachep
->name
, kbuf
)) {
4224 if (limit
< 1 || batchcount
< 1 ||
4225 batchcount
> limit
|| shared
< 0) {
4228 res
= do_tune_cpucache(cachep
, limit
,
4235 mutex_unlock(&slab_mutex
);
4241 #ifdef CONFIG_DEBUG_SLAB_LEAK
4243 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4245 mutex_lock(&slab_mutex
);
4246 return seq_list_start(&slab_caches
, *pos
);
4249 static inline int add_caller(unsigned long *n
, unsigned long v
)
4259 unsigned long *q
= p
+ 2 * i
;
4273 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4279 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4287 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4288 if (get_obj_status(page
, i
) != OBJECT_ACTIVE
)
4291 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4296 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4298 #ifdef CONFIG_KALLSYMS
4299 unsigned long offset
, size
;
4300 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4302 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4303 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4305 seq_printf(m
, " [%s]", modname
);
4309 seq_printf(m
, "%p", (void *)address
);
4312 static int leaks_show(struct seq_file
*m
, void *p
)
4314 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4316 struct kmem_cache_node
*n
;
4318 unsigned long *x
= m
->private;
4322 if (!(cachep
->flags
& SLAB_STORE_USER
))
4324 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4327 /* OK, we can do it */
4331 for_each_online_node(node
) {
4332 n
= cachep
->node
[node
];
4337 spin_lock_irq(&n
->list_lock
);
4339 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4340 handle_slab(x
, cachep
, page
);
4341 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4342 handle_slab(x
, cachep
, page
);
4343 spin_unlock_irq(&n
->list_lock
);
4345 name
= cachep
->name
;
4347 /* Increase the buffer size */
4348 mutex_unlock(&slab_mutex
);
4349 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4351 /* Too bad, we are really out */
4353 mutex_lock(&slab_mutex
);
4356 *(unsigned long *)m
->private = x
[0] * 2;
4358 mutex_lock(&slab_mutex
);
4359 /* Now make sure this entry will be retried */
4363 for (i
= 0; i
< x
[1]; i
++) {
4364 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4365 show_symbol(m
, x
[2*i
+2]);
4372 static const struct seq_operations slabstats_op
= {
4373 .start
= leaks_start
,
4379 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4381 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4384 ret
= seq_open(file
, &slabstats_op
);
4386 struct seq_file
*m
= file
->private_data
;
4387 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4396 static const struct file_operations proc_slabstats_operations
= {
4397 .open
= slabstats_open
,
4399 .llseek
= seq_lseek
,
4400 .release
= seq_release_private
,
4404 static int __init
slab_proc_init(void)
4406 #ifdef CONFIG_DEBUG_SLAB_LEAK
4407 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4411 module_init(slab_proc_init
);
4415 * ksize - get the actual amount of memory allocated for a given object
4416 * @objp: Pointer to the object
4418 * kmalloc may internally round up allocations and return more memory
4419 * than requested. ksize() can be used to determine the actual amount of
4420 * memory allocated. The caller may use this additional memory, even though
4421 * a smaller amount of memory was initially specified with the kmalloc call.
4422 * The caller must guarantee that objp points to a valid object previously
4423 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4424 * must not be freed during the duration of the call.
4426 size_t ksize(const void *objp
)
4429 if (unlikely(objp
== ZERO_SIZE_PTR
))
4432 return virt_to_cache(objp
)->object_size
;
4434 EXPORT_SYMBOL(ksize
);