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
;
195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
207 struct array_cache ac
;
210 #define SLAB_OBJ_PFMEMALLOC 1
211 static inline bool is_obj_pfmemalloc(void *objp
)
213 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
216 static inline void set_obj_pfmemalloc(void **objp
)
218 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
222 static inline void clear_obj_pfmemalloc(void **objp
)
224 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
228 * bootstrap: The caches do not work without cpuarrays anymore, but the
229 * cpuarrays are allocated from the generic caches...
231 #define BOOT_CPUCACHE_ENTRIES 1
232 struct arraycache_init
{
233 struct array_cache cache
;
234 void *entries
[BOOT_CPUCACHE_ENTRIES
];
238 * Need this for bootstrapping a per node allocator.
240 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
241 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
242 #define CACHE_CACHE 0
243 #define SIZE_AC MAX_NUMNODES
244 #define SIZE_NODE (2 * MAX_NUMNODES)
246 static int drain_freelist(struct kmem_cache
*cache
,
247 struct kmem_cache_node
*n
, int tofree
);
248 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
249 int node
, struct list_head
*list
);
250 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
251 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
252 static void cache_reap(struct work_struct
*unused
);
254 static int slab_early_init
= 1;
256 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
257 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
259 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
261 INIT_LIST_HEAD(&parent
->slabs_full
);
262 INIT_LIST_HEAD(&parent
->slabs_partial
);
263 INIT_LIST_HEAD(&parent
->slabs_free
);
264 parent
->shared
= NULL
;
265 parent
->alien
= NULL
;
266 parent
->colour_next
= 0;
267 spin_lock_init(&parent
->list_lock
);
268 parent
->free_objects
= 0;
269 parent
->free_touched
= 0;
272 #define MAKE_LIST(cachep, listp, slab, nodeid) \
274 INIT_LIST_HEAD(listp); \
275 list_splice(&get_node(cachep, nodeid)->slab, listp); \
278 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
280 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
281 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
282 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
285 #define CFLGS_OFF_SLAB (0x80000000UL)
286 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
288 #define BATCHREFILL_LIMIT 16
290 * Optimization question: fewer reaps means less probability for unnessary
291 * cpucache drain/refill cycles.
293 * OTOH the cpuarrays can contain lots of objects,
294 * which could lock up otherwise freeable slabs.
296 #define REAPTIMEOUT_AC (2*HZ)
297 #define REAPTIMEOUT_NODE (4*HZ)
300 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
301 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
302 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
303 #define STATS_INC_GROWN(x) ((x)->grown++)
304 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
305 #define STATS_SET_HIGH(x) \
307 if ((x)->num_active > (x)->high_mark) \
308 (x)->high_mark = (x)->num_active; \
310 #define STATS_INC_ERR(x) ((x)->errors++)
311 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
312 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
313 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
314 #define STATS_SET_FREEABLE(x, i) \
316 if ((x)->max_freeable < i) \
317 (x)->max_freeable = i; \
319 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
320 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
321 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
322 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
324 #define STATS_INC_ACTIVE(x) do { } while (0)
325 #define STATS_DEC_ACTIVE(x) do { } while (0)
326 #define STATS_INC_ALLOCED(x) do { } while (0)
327 #define STATS_INC_GROWN(x) do { } while (0)
328 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
329 #define STATS_SET_HIGH(x) do { } while (0)
330 #define STATS_INC_ERR(x) do { } while (0)
331 #define STATS_INC_NODEALLOCS(x) do { } while (0)
332 #define STATS_INC_NODEFREES(x) do { } while (0)
333 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
334 #define STATS_SET_FREEABLE(x, i) do { } while (0)
335 #define STATS_INC_ALLOCHIT(x) do { } while (0)
336 #define STATS_INC_ALLOCMISS(x) do { } while (0)
337 #define STATS_INC_FREEHIT(x) do { } while (0)
338 #define STATS_INC_FREEMISS(x) do { } while (0)
344 * memory layout of objects:
346 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
347 * the end of an object is aligned with the end of the real
348 * allocation. Catches writes behind the end of the allocation.
349 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
351 * cachep->obj_offset: The real object.
352 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
353 * cachep->size - 1* BYTES_PER_WORD: last caller address
354 * [BYTES_PER_WORD long]
356 static int obj_offset(struct kmem_cache
*cachep
)
358 return cachep
->obj_offset
;
361 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
363 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
364 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
365 sizeof(unsigned long long));
368 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
370 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
371 if (cachep
->flags
& SLAB_STORE_USER
)
372 return (unsigned long long *)(objp
+ cachep
->size
-
373 sizeof(unsigned long long) -
375 return (unsigned long long *) (objp
+ cachep
->size
-
376 sizeof(unsigned long long));
379 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
381 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
382 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
387 #define obj_offset(x) 0
388 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
390 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
394 #define OBJECT_FREE (0)
395 #define OBJECT_ACTIVE (1)
397 #ifdef CONFIG_DEBUG_SLAB_LEAK
399 static void set_obj_status(struct page
*page
, int idx
, int val
)
403 struct kmem_cache
*cachep
= page
->slab_cache
;
405 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
406 status
= (char *)page
->freelist
+ freelist_size
;
410 static inline unsigned int get_obj_status(struct page
*page
, int idx
)
414 struct kmem_cache
*cachep
= page
->slab_cache
;
416 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
417 status
= (char *)page
->freelist
+ freelist_size
;
423 static inline void set_obj_status(struct page
*page
, int idx
, int val
) {}
428 * Do not go above this order unless 0 objects fit into the slab or
429 * overridden on the command line.
431 #define SLAB_MAX_ORDER_HI 1
432 #define SLAB_MAX_ORDER_LO 0
433 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
434 static bool slab_max_order_set __initdata
;
436 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
438 struct page
*page
= virt_to_head_page(obj
);
439 return page
->slab_cache
;
442 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
445 return page
->s_mem
+ cache
->size
* idx
;
449 * We want to avoid an expensive divide : (offset / cache->size)
450 * Using the fact that size is a constant for a particular cache,
451 * we can replace (offset / cache->size) by
452 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
454 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
455 const struct page
*page
, void *obj
)
457 u32 offset
= (obj
- page
->s_mem
);
458 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
461 static struct arraycache_init initarray_generic
=
462 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
464 /* internal cache of cache description objs */
465 static struct kmem_cache kmem_cache_boot
= {
467 .limit
= BOOT_CPUCACHE_ENTRIES
,
469 .size
= sizeof(struct kmem_cache
),
470 .name
= "kmem_cache",
473 #define BAD_ALIEN_MAGIC 0x01020304ul
475 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
477 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
479 return cachep
->array
[smp_processor_id()];
482 static size_t calculate_freelist_size(int nr_objs
, size_t align
)
484 size_t freelist_size
;
486 freelist_size
= nr_objs
* sizeof(freelist_idx_t
);
487 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
488 freelist_size
+= nr_objs
* sizeof(char);
491 freelist_size
= ALIGN(freelist_size
, align
);
493 return freelist_size
;
496 static int calculate_nr_objs(size_t slab_size
, size_t buffer_size
,
497 size_t idx_size
, size_t align
)
500 size_t remained_size
;
501 size_t freelist_size
;
504 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
505 extra_space
= sizeof(char);
507 * Ignore padding for the initial guess. The padding
508 * is at most @align-1 bytes, and @buffer_size is at
509 * least @align. In the worst case, this result will
510 * be one greater than the number of objects that fit
511 * into the memory allocation when taking the padding
514 nr_objs
= slab_size
/ (buffer_size
+ idx_size
+ extra_space
);
517 * This calculated number will be either the right
518 * amount, or one greater than what we want.
520 remained_size
= slab_size
- nr_objs
* buffer_size
;
521 freelist_size
= calculate_freelist_size(nr_objs
, align
);
522 if (remained_size
< freelist_size
)
529 * Calculate the number of objects and left-over bytes for a given buffer size.
531 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
532 size_t align
, int flags
, size_t *left_over
,
537 size_t slab_size
= PAGE_SIZE
<< gfporder
;
540 * The slab management structure can be either off the slab or
541 * on it. For the latter case, the memory allocated for a
544 * - One unsigned int for each object
545 * - Padding to respect alignment of @align
546 * - @buffer_size bytes for each object
548 * If the slab management structure is off the slab, then the
549 * alignment will already be calculated into the size. Because
550 * the slabs are all pages aligned, the objects will be at the
551 * correct alignment when allocated.
553 if (flags
& CFLGS_OFF_SLAB
) {
555 nr_objs
= slab_size
/ buffer_size
;
558 nr_objs
= calculate_nr_objs(slab_size
, buffer_size
,
559 sizeof(freelist_idx_t
), align
);
560 mgmt_size
= calculate_freelist_size(nr_objs
, align
);
563 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
567 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
569 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
572 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
573 function
, cachep
->name
, msg
);
575 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
580 * By default on NUMA we use alien caches to stage the freeing of
581 * objects allocated from other nodes. This causes massive memory
582 * inefficiencies when using fake NUMA setup to split memory into a
583 * large number of small nodes, so it can be disabled on the command
587 static int use_alien_caches __read_mostly
= 1;
588 static int __init
noaliencache_setup(char *s
)
590 use_alien_caches
= 0;
593 __setup("noaliencache", noaliencache_setup
);
595 static int __init
slab_max_order_setup(char *str
)
597 get_option(&str
, &slab_max_order
);
598 slab_max_order
= slab_max_order
< 0 ? 0 :
599 min(slab_max_order
, MAX_ORDER
- 1);
600 slab_max_order_set
= true;
604 __setup("slab_max_order=", slab_max_order_setup
);
608 * Special reaping functions for NUMA systems called from cache_reap().
609 * These take care of doing round robin flushing of alien caches (containing
610 * objects freed on different nodes from which they were allocated) and the
611 * flushing of remote pcps by calling drain_node_pages.
613 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
615 static void init_reap_node(int cpu
)
619 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
620 if (node
== MAX_NUMNODES
)
621 node
= first_node(node_online_map
);
623 per_cpu(slab_reap_node
, cpu
) = node
;
626 static void next_reap_node(void)
628 int node
= __this_cpu_read(slab_reap_node
);
630 node
= next_node(node
, node_online_map
);
631 if (unlikely(node
>= MAX_NUMNODES
))
632 node
= first_node(node_online_map
);
633 __this_cpu_write(slab_reap_node
, node
);
637 #define init_reap_node(cpu) do { } while (0)
638 #define next_reap_node(void) do { } while (0)
642 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
643 * via the workqueue/eventd.
644 * Add the CPU number into the expiration time to minimize the possibility of
645 * the CPUs getting into lockstep and contending for the global cache chain
648 static void start_cpu_timer(int cpu
)
650 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
653 * When this gets called from do_initcalls via cpucache_init(),
654 * init_workqueues() has already run, so keventd will be setup
657 if (keventd_up() && reap_work
->work
.func
== NULL
) {
659 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
660 schedule_delayed_work_on(cpu
, reap_work
,
661 __round_jiffies_relative(HZ
, cpu
));
665 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
668 * The array_cache structures contain pointers to free object.
669 * However, when such objects are allocated or transferred to another
670 * cache the pointers are not cleared and they could be counted as
671 * valid references during a kmemleak scan. Therefore, kmemleak must
672 * not scan such objects.
674 kmemleak_no_scan(ac
);
678 ac
->batchcount
= batch
;
683 static struct array_cache
*alloc_arraycache(int node
, int entries
,
684 int batchcount
, gfp_t gfp
)
686 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
687 struct array_cache
*ac
= NULL
;
689 ac
= kmalloc_node(memsize
, gfp
, node
);
690 init_arraycache(ac
, entries
, batchcount
);
694 static inline bool is_slab_pfmemalloc(struct page
*page
)
696 return PageSlabPfmemalloc(page
);
699 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
700 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
701 struct array_cache
*ac
)
703 struct kmem_cache_node
*n
= get_node(cachep
, numa_mem_id());
707 if (!pfmemalloc_active
)
710 spin_lock_irqsave(&n
->list_lock
, flags
);
711 list_for_each_entry(page
, &n
->slabs_full
, lru
)
712 if (is_slab_pfmemalloc(page
))
715 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
716 if (is_slab_pfmemalloc(page
))
719 list_for_each_entry(page
, &n
->slabs_free
, lru
)
720 if (is_slab_pfmemalloc(page
))
723 pfmemalloc_active
= false;
725 spin_unlock_irqrestore(&n
->list_lock
, flags
);
728 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
729 gfp_t flags
, bool force_refill
)
732 void *objp
= ac
->entry
[--ac
->avail
];
734 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
735 if (unlikely(is_obj_pfmemalloc(objp
))) {
736 struct kmem_cache_node
*n
;
738 if (gfp_pfmemalloc_allowed(flags
)) {
739 clear_obj_pfmemalloc(&objp
);
743 /* The caller cannot use PFMEMALLOC objects, find another one */
744 for (i
= 0; i
< ac
->avail
; i
++) {
745 /* If a !PFMEMALLOC object is found, swap them */
746 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
748 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
749 ac
->entry
[ac
->avail
] = objp
;
755 * If there are empty slabs on the slabs_free list and we are
756 * being forced to refill the cache, mark this one !pfmemalloc.
758 n
= get_node(cachep
, numa_mem_id());
759 if (!list_empty(&n
->slabs_free
) && force_refill
) {
760 struct page
*page
= virt_to_head_page(objp
);
761 ClearPageSlabPfmemalloc(page
);
762 clear_obj_pfmemalloc(&objp
);
763 recheck_pfmemalloc_active(cachep
, ac
);
767 /* No !PFMEMALLOC objects available */
775 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
776 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
780 if (unlikely(sk_memalloc_socks()))
781 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
783 objp
= ac
->entry
[--ac
->avail
];
788 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
791 if (unlikely(pfmemalloc_active
)) {
792 /* Some pfmemalloc slabs exist, check if this is one */
793 struct page
*page
= virt_to_head_page(objp
);
794 if (PageSlabPfmemalloc(page
))
795 set_obj_pfmemalloc(&objp
);
801 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
804 if (unlikely(sk_memalloc_socks()))
805 objp
= __ac_put_obj(cachep
, ac
, objp
);
807 ac
->entry
[ac
->avail
++] = objp
;
811 * Transfer objects in one arraycache to another.
812 * Locking must be handled by the caller.
814 * Return the number of entries transferred.
816 static int transfer_objects(struct array_cache
*to
,
817 struct array_cache
*from
, unsigned int max
)
819 /* Figure out how many entries to transfer */
820 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
825 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
835 #define drain_alien_cache(cachep, alien) do { } while (0)
836 #define reap_alien(cachep, n) do { } while (0)
838 static inline struct alien_cache
**alloc_alien_cache(int node
,
839 int limit
, gfp_t gfp
)
841 return (struct alien_cache
**)BAD_ALIEN_MAGIC
;
844 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
848 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
853 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
859 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
860 gfp_t flags
, int nodeid
)
865 #else /* CONFIG_NUMA */
867 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
868 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
870 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
871 int batch
, gfp_t gfp
)
873 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
874 struct alien_cache
*alc
= NULL
;
876 alc
= kmalloc_node(memsize
, gfp
, node
);
877 init_arraycache(&alc
->ac
, entries
, batch
);
878 spin_lock_init(&alc
->lock
);
882 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
884 struct alien_cache
**alc_ptr
;
885 size_t memsize
= sizeof(void *) * nr_node_ids
;
890 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
895 if (i
== node
|| !node_online(i
))
897 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
899 for (i
--; i
>= 0; i
--)
908 static void free_alien_cache(struct alien_cache
**alc_ptr
)
919 static void __drain_alien_cache(struct kmem_cache
*cachep
,
920 struct array_cache
*ac
, int node
,
921 struct list_head
*list
)
923 struct kmem_cache_node
*n
= get_node(cachep
, node
);
926 spin_lock(&n
->list_lock
);
928 * Stuff objects into the remote nodes shared array first.
929 * That way we could avoid the overhead of putting the objects
930 * into the free lists and getting them back later.
933 transfer_objects(n
->shared
, ac
, ac
->limit
);
935 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
937 spin_unlock(&n
->list_lock
);
942 * Called from cache_reap() to regularly drain alien caches round robin.
944 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
946 int node
= __this_cpu_read(slab_reap_node
);
949 struct alien_cache
*alc
= n
->alien
[node
];
950 struct array_cache
*ac
;
954 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
957 __drain_alien_cache(cachep
, ac
, node
, &list
);
958 spin_unlock_irq(&alc
->lock
);
959 slabs_destroy(cachep
, &list
);
965 static void drain_alien_cache(struct kmem_cache
*cachep
,
966 struct alien_cache
**alien
)
969 struct alien_cache
*alc
;
970 struct array_cache
*ac
;
973 for_each_online_node(i
) {
979 spin_lock_irqsave(&alc
->lock
, flags
);
980 __drain_alien_cache(cachep
, ac
, i
, &list
);
981 spin_unlock_irqrestore(&alc
->lock
, flags
);
982 slabs_destroy(cachep
, &list
);
987 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
989 int nodeid
= page_to_nid(virt_to_page(objp
));
990 struct kmem_cache_node
*n
;
991 struct alien_cache
*alien
= NULL
;
992 struct array_cache
*ac
;
996 node
= numa_mem_id();
999 * Make sure we are not freeing a object from another node to the array
1000 * cache on this cpu.
1002 if (likely(nodeid
== node
))
1005 n
= get_node(cachep
, node
);
1006 STATS_INC_NODEFREES(cachep
);
1007 if (n
->alien
&& n
->alien
[nodeid
]) {
1008 alien
= n
->alien
[nodeid
];
1010 spin_lock(&alien
->lock
);
1011 if (unlikely(ac
->avail
== ac
->limit
)) {
1012 STATS_INC_ACOVERFLOW(cachep
);
1013 __drain_alien_cache(cachep
, ac
, nodeid
, &list
);
1015 ac_put_obj(cachep
, ac
, objp
);
1016 spin_unlock(&alien
->lock
);
1017 slabs_destroy(cachep
, &list
);
1019 n
= get_node(cachep
, nodeid
);
1020 spin_lock(&n
->list_lock
);
1021 free_block(cachep
, &objp
, 1, nodeid
, &list
);
1022 spin_unlock(&n
->list_lock
);
1023 slabs_destroy(cachep
, &list
);
1030 * Allocates and initializes node for a node on each slab cache, used for
1031 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1032 * will be allocated off-node since memory is not yet online for the new node.
1033 * When hotplugging memory or a cpu, existing node are not replaced if
1036 * Must hold slab_mutex.
1038 static int init_cache_node_node(int node
)
1040 struct kmem_cache
*cachep
;
1041 struct kmem_cache_node
*n
;
1042 const size_t memsize
= sizeof(struct kmem_cache_node
);
1044 list_for_each_entry(cachep
, &slab_caches
, list
) {
1046 * Set up the kmem_cache_node for cpu before we can
1047 * begin anything. Make sure some other cpu on this
1048 * node has not already allocated this
1050 n
= get_node(cachep
, node
);
1052 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1055 kmem_cache_node_init(n
);
1056 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
1057 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1060 * The kmem_cache_nodes don't come and go as CPUs
1061 * come and go. slab_mutex is sufficient
1064 cachep
->node
[node
] = n
;
1067 spin_lock_irq(&n
->list_lock
);
1069 (1 + nr_cpus_node(node
)) *
1070 cachep
->batchcount
+ cachep
->num
;
1071 spin_unlock_irq(&n
->list_lock
);
1076 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1077 struct kmem_cache_node
*n
)
1079 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1082 static void cpuup_canceled(long cpu
)
1084 struct kmem_cache
*cachep
;
1085 struct kmem_cache_node
*n
= NULL
;
1086 int node
= cpu_to_mem(cpu
);
1087 const struct cpumask
*mask
= cpumask_of_node(node
);
1089 list_for_each_entry(cachep
, &slab_caches
, list
) {
1090 struct array_cache
*nc
;
1091 struct array_cache
*shared
;
1092 struct alien_cache
**alien
;
1095 /* cpu is dead; no one can alloc from it. */
1096 nc
= cachep
->array
[cpu
];
1097 cachep
->array
[cpu
] = NULL
;
1098 n
= get_node(cachep
, node
);
1101 goto free_array_cache
;
1103 spin_lock_irq(&n
->list_lock
);
1105 /* Free limit for this kmem_cache_node */
1106 n
->free_limit
-= cachep
->batchcount
;
1108 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1110 if (!cpumask_empty(mask
)) {
1111 spin_unlock_irq(&n
->list_lock
);
1112 goto free_array_cache
;
1117 free_block(cachep
, shared
->entry
,
1118 shared
->avail
, node
, &list
);
1125 spin_unlock_irq(&n
->list_lock
);
1129 drain_alien_cache(cachep
, alien
);
1130 free_alien_cache(alien
);
1133 slabs_destroy(cachep
, &list
);
1137 * In the previous loop, all the objects were freed to
1138 * the respective cache's slabs, now we can go ahead and
1139 * shrink each nodelist to its limit.
1141 list_for_each_entry(cachep
, &slab_caches
, list
) {
1142 n
= get_node(cachep
, node
);
1145 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1149 static int cpuup_prepare(long cpu
)
1151 struct kmem_cache
*cachep
;
1152 struct kmem_cache_node
*n
= NULL
;
1153 int node
= cpu_to_mem(cpu
);
1157 * We need to do this right in the beginning since
1158 * alloc_arraycache's are going to use this list.
1159 * kmalloc_node allows us to add the slab to the right
1160 * kmem_cache_node and not this cpu's kmem_cache_node
1162 err
= init_cache_node_node(node
);
1167 * Now we can go ahead with allocating the shared arrays and
1170 list_for_each_entry(cachep
, &slab_caches
, list
) {
1171 struct array_cache
*nc
;
1172 struct array_cache
*shared
= NULL
;
1173 struct alien_cache
**alien
= NULL
;
1175 nc
= alloc_arraycache(node
, cachep
->limit
,
1176 cachep
->batchcount
, GFP_KERNEL
);
1179 if (cachep
->shared
) {
1180 shared
= alloc_arraycache(node
,
1181 cachep
->shared
* cachep
->batchcount
,
1182 0xbaadf00d, GFP_KERNEL
);
1188 if (use_alien_caches
) {
1189 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1196 cachep
->array
[cpu
] = nc
;
1197 n
= get_node(cachep
, node
);
1200 spin_lock_irq(&n
->list_lock
);
1203 * We are serialised from CPU_DEAD or
1204 * CPU_UP_CANCELLED by the cpucontrol lock
1215 spin_unlock_irq(&n
->list_lock
);
1217 free_alien_cache(alien
);
1222 cpuup_canceled(cpu
);
1226 static int cpuup_callback(struct notifier_block
*nfb
,
1227 unsigned long action
, void *hcpu
)
1229 long cpu
= (long)hcpu
;
1233 case CPU_UP_PREPARE
:
1234 case CPU_UP_PREPARE_FROZEN
:
1235 mutex_lock(&slab_mutex
);
1236 err
= cpuup_prepare(cpu
);
1237 mutex_unlock(&slab_mutex
);
1240 case CPU_ONLINE_FROZEN
:
1241 start_cpu_timer(cpu
);
1243 #ifdef CONFIG_HOTPLUG_CPU
1244 case CPU_DOWN_PREPARE
:
1245 case CPU_DOWN_PREPARE_FROZEN
:
1247 * Shutdown cache reaper. Note that the slab_mutex is
1248 * held so that if cache_reap() is invoked it cannot do
1249 * anything expensive but will only modify reap_work
1250 * and reschedule the timer.
1252 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1253 /* Now the cache_reaper is guaranteed to be not running. */
1254 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1256 case CPU_DOWN_FAILED
:
1257 case CPU_DOWN_FAILED_FROZEN
:
1258 start_cpu_timer(cpu
);
1261 case CPU_DEAD_FROZEN
:
1263 * Even if all the cpus of a node are down, we don't free the
1264 * kmem_cache_node of any cache. This to avoid a race between
1265 * cpu_down, and a kmalloc allocation from another cpu for
1266 * memory from the node of the cpu going down. The node
1267 * structure is usually allocated from kmem_cache_create() and
1268 * gets destroyed at kmem_cache_destroy().
1272 case CPU_UP_CANCELED
:
1273 case CPU_UP_CANCELED_FROZEN
:
1274 mutex_lock(&slab_mutex
);
1275 cpuup_canceled(cpu
);
1276 mutex_unlock(&slab_mutex
);
1279 return notifier_from_errno(err
);
1282 static struct notifier_block cpucache_notifier
= {
1283 &cpuup_callback
, NULL
, 0
1286 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1288 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1289 * Returns -EBUSY if all objects cannot be drained so that the node is not
1292 * Must hold slab_mutex.
1294 static int __meminit
drain_cache_node_node(int node
)
1296 struct kmem_cache
*cachep
;
1299 list_for_each_entry(cachep
, &slab_caches
, list
) {
1300 struct kmem_cache_node
*n
;
1302 n
= get_node(cachep
, node
);
1306 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1308 if (!list_empty(&n
->slabs_full
) ||
1309 !list_empty(&n
->slabs_partial
)) {
1317 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1318 unsigned long action
, void *arg
)
1320 struct memory_notify
*mnb
= arg
;
1324 nid
= mnb
->status_change_nid
;
1329 case MEM_GOING_ONLINE
:
1330 mutex_lock(&slab_mutex
);
1331 ret
= init_cache_node_node(nid
);
1332 mutex_unlock(&slab_mutex
);
1334 case MEM_GOING_OFFLINE
:
1335 mutex_lock(&slab_mutex
);
1336 ret
= drain_cache_node_node(nid
);
1337 mutex_unlock(&slab_mutex
);
1341 case MEM_CANCEL_ONLINE
:
1342 case MEM_CANCEL_OFFLINE
:
1346 return notifier_from_errno(ret
);
1348 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1351 * swap the static kmem_cache_node with kmalloced memory
1353 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1356 struct kmem_cache_node
*ptr
;
1358 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1361 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1363 * Do not assume that spinlocks can be initialized via memcpy:
1365 spin_lock_init(&ptr
->list_lock
);
1367 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1368 cachep
->node
[nodeid
] = ptr
;
1372 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1373 * size of kmem_cache_node.
1375 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1379 for_each_online_node(node
) {
1380 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1381 cachep
->node
[node
]->next_reap
= jiffies
+
1383 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1388 * The memory after the last cpu cache pointer is used for the
1391 static void setup_node_pointer(struct kmem_cache
*cachep
)
1393 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1397 * Initialisation. Called after the page allocator have been initialised and
1398 * before smp_init().
1400 void __init
kmem_cache_init(void)
1404 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1405 sizeof(struct rcu_head
));
1406 kmem_cache
= &kmem_cache_boot
;
1407 setup_node_pointer(kmem_cache
);
1409 if (num_possible_nodes() == 1)
1410 use_alien_caches
= 0;
1412 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1413 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1415 set_up_node(kmem_cache
, CACHE_CACHE
);
1418 * Fragmentation resistance on low memory - only use bigger
1419 * page orders on machines with more than 32MB of memory if
1420 * not overridden on the command line.
1422 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1423 slab_max_order
= SLAB_MAX_ORDER_HI
;
1425 /* Bootstrap is tricky, because several objects are allocated
1426 * from caches that do not exist yet:
1427 * 1) initialize the kmem_cache cache: it contains the struct
1428 * kmem_cache structures of all caches, except kmem_cache itself:
1429 * kmem_cache is statically allocated.
1430 * Initially an __init data area is used for the head array and the
1431 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1432 * array at the end of the bootstrap.
1433 * 2) Create the first kmalloc cache.
1434 * The struct kmem_cache for the new cache is allocated normally.
1435 * An __init data area is used for the head array.
1436 * 3) Create the remaining kmalloc caches, with minimally sized
1438 * 4) Replace the __init data head arrays for kmem_cache and the first
1439 * kmalloc cache with kmalloc allocated arrays.
1440 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1441 * the other cache's with kmalloc allocated memory.
1442 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1445 /* 1) create the kmem_cache */
1448 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1450 create_boot_cache(kmem_cache
, "kmem_cache",
1451 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1452 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1453 SLAB_HWCACHE_ALIGN
);
1454 list_add(&kmem_cache
->list
, &slab_caches
);
1456 /* 2+3) create the kmalloc caches */
1459 * Initialize the caches that provide memory for the array cache and the
1460 * kmem_cache_node structures first. Without this, further allocations will
1464 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1465 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1467 if (INDEX_AC
!= INDEX_NODE
)
1468 kmalloc_caches
[INDEX_NODE
] =
1469 create_kmalloc_cache("kmalloc-node",
1470 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1472 slab_early_init
= 0;
1474 /* 4) Replace the bootstrap head arrays */
1476 struct array_cache
*ptr
;
1478 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1480 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1481 sizeof(struct arraycache_init
));
1483 kmem_cache
->array
[smp_processor_id()] = ptr
;
1485 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1487 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1488 != &initarray_generic
.cache
);
1489 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1490 sizeof(struct arraycache_init
));
1492 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1494 /* 5) Replace the bootstrap kmem_cache_node */
1498 for_each_online_node(nid
) {
1499 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1501 init_list(kmalloc_caches
[INDEX_AC
],
1502 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1504 if (INDEX_AC
!= INDEX_NODE
) {
1505 init_list(kmalloc_caches
[INDEX_NODE
],
1506 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1511 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1514 void __init
kmem_cache_init_late(void)
1516 struct kmem_cache
*cachep
;
1520 /* 6) resize the head arrays to their final sizes */
1521 mutex_lock(&slab_mutex
);
1522 list_for_each_entry(cachep
, &slab_caches
, list
)
1523 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1525 mutex_unlock(&slab_mutex
);
1531 * Register a cpu startup notifier callback that initializes
1532 * cpu_cache_get for all new cpus
1534 register_cpu_notifier(&cpucache_notifier
);
1538 * Register a memory hotplug callback that initializes and frees
1541 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1545 * The reap timers are started later, with a module init call: That part
1546 * of the kernel is not yet operational.
1550 static int __init
cpucache_init(void)
1555 * Register the timers that return unneeded pages to the page allocator
1557 for_each_online_cpu(cpu
)
1558 start_cpu_timer(cpu
);
1564 __initcall(cpucache_init
);
1566 static noinline
void
1567 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1570 struct kmem_cache_node
*n
;
1572 unsigned long flags
;
1574 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1575 DEFAULT_RATELIMIT_BURST
);
1577 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1581 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1583 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1584 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1586 for_each_kmem_cache_node(cachep
, node
, n
) {
1587 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1588 unsigned long active_slabs
= 0, num_slabs
= 0;
1590 spin_lock_irqsave(&n
->list_lock
, flags
);
1591 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1592 active_objs
+= cachep
->num
;
1595 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1596 active_objs
+= page
->active
;
1599 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1602 free_objects
+= n
->free_objects
;
1603 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1605 num_slabs
+= active_slabs
;
1606 num_objs
= num_slabs
* cachep
->num
;
1608 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1609 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1616 * Interface to system's page allocator. No need to hold the
1617 * kmem_cache_node ->list_lock.
1619 * If we requested dmaable memory, we will get it. Even if we
1620 * did not request dmaable memory, we might get it, but that
1621 * would be relatively rare and ignorable.
1623 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1629 flags
|= cachep
->allocflags
;
1630 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1631 flags
|= __GFP_RECLAIMABLE
;
1633 if (memcg_charge_slab(cachep
, flags
, cachep
->gfporder
))
1636 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1638 memcg_uncharge_slab(cachep
, cachep
->gfporder
);
1639 slab_out_of_memory(cachep
, flags
, nodeid
);
1643 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1644 if (unlikely(page
->pfmemalloc
))
1645 pfmemalloc_active
= true;
1647 nr_pages
= (1 << cachep
->gfporder
);
1648 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1649 add_zone_page_state(page_zone(page
),
1650 NR_SLAB_RECLAIMABLE
, nr_pages
);
1652 add_zone_page_state(page_zone(page
),
1653 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1654 __SetPageSlab(page
);
1655 if (page
->pfmemalloc
)
1656 SetPageSlabPfmemalloc(page
);
1658 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1659 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1662 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1664 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1671 * Interface to system's page release.
1673 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1675 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1677 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1679 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1680 sub_zone_page_state(page_zone(page
),
1681 NR_SLAB_RECLAIMABLE
, nr_freed
);
1683 sub_zone_page_state(page_zone(page
),
1684 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1686 BUG_ON(!PageSlab(page
));
1687 __ClearPageSlabPfmemalloc(page
);
1688 __ClearPageSlab(page
);
1689 page_mapcount_reset(page
);
1690 page
->mapping
= NULL
;
1692 if (current
->reclaim_state
)
1693 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1694 __free_pages(page
, cachep
->gfporder
);
1695 memcg_uncharge_slab(cachep
, cachep
->gfporder
);
1698 static void kmem_rcu_free(struct rcu_head
*head
)
1700 struct kmem_cache
*cachep
;
1703 page
= container_of(head
, struct page
, rcu_head
);
1704 cachep
= page
->slab_cache
;
1706 kmem_freepages(cachep
, page
);
1711 #ifdef CONFIG_DEBUG_PAGEALLOC
1712 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1713 unsigned long caller
)
1715 int size
= cachep
->object_size
;
1717 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1719 if (size
< 5 * sizeof(unsigned long))
1722 *addr
++ = 0x12345678;
1724 *addr
++ = smp_processor_id();
1725 size
-= 3 * sizeof(unsigned long);
1727 unsigned long *sptr
= &caller
;
1728 unsigned long svalue
;
1730 while (!kstack_end(sptr
)) {
1732 if (kernel_text_address(svalue
)) {
1734 size
-= sizeof(unsigned long);
1735 if (size
<= sizeof(unsigned long))
1741 *addr
++ = 0x87654321;
1745 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1747 int size
= cachep
->object_size
;
1748 addr
= &((char *)addr
)[obj_offset(cachep
)];
1750 memset(addr
, val
, size
);
1751 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1754 static void dump_line(char *data
, int offset
, int limit
)
1757 unsigned char error
= 0;
1760 printk(KERN_ERR
"%03x: ", offset
);
1761 for (i
= 0; i
< limit
; i
++) {
1762 if (data
[offset
+ i
] != POISON_FREE
) {
1763 error
= data
[offset
+ i
];
1767 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1768 &data
[offset
], limit
, 1);
1770 if (bad_count
== 1) {
1771 error
^= POISON_FREE
;
1772 if (!(error
& (error
- 1))) {
1773 printk(KERN_ERR
"Single bit error detected. Probably "
1776 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1779 printk(KERN_ERR
"Run a memory test tool.\n");
1788 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1793 if (cachep
->flags
& SLAB_RED_ZONE
) {
1794 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1795 *dbg_redzone1(cachep
, objp
),
1796 *dbg_redzone2(cachep
, objp
));
1799 if (cachep
->flags
& SLAB_STORE_USER
) {
1800 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1801 *dbg_userword(cachep
, objp
),
1802 *dbg_userword(cachep
, objp
));
1804 realobj
= (char *)objp
+ obj_offset(cachep
);
1805 size
= cachep
->object_size
;
1806 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1809 if (i
+ limit
> size
)
1811 dump_line(realobj
, i
, limit
);
1815 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1821 realobj
= (char *)objp
+ obj_offset(cachep
);
1822 size
= cachep
->object_size
;
1824 for (i
= 0; i
< size
; i
++) {
1825 char exp
= POISON_FREE
;
1828 if (realobj
[i
] != exp
) {
1834 "Slab corruption (%s): %s start=%p, len=%d\n",
1835 print_tainted(), cachep
->name
, realobj
, size
);
1836 print_objinfo(cachep
, objp
, 0);
1838 /* Hexdump the affected line */
1841 if (i
+ limit
> size
)
1843 dump_line(realobj
, i
, limit
);
1846 /* Limit to 5 lines */
1852 /* Print some data about the neighboring objects, if they
1855 struct page
*page
= virt_to_head_page(objp
);
1858 objnr
= obj_to_index(cachep
, page
, objp
);
1860 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1861 realobj
= (char *)objp
+ obj_offset(cachep
);
1862 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1864 print_objinfo(cachep
, objp
, 2);
1866 if (objnr
+ 1 < cachep
->num
) {
1867 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1868 realobj
= (char *)objp
+ obj_offset(cachep
);
1869 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1871 print_objinfo(cachep
, objp
, 2);
1878 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1882 for (i
= 0; i
< cachep
->num
; i
++) {
1883 void *objp
= index_to_obj(cachep
, page
, i
);
1885 if (cachep
->flags
& SLAB_POISON
) {
1886 #ifdef CONFIG_DEBUG_PAGEALLOC
1887 if (cachep
->size
% PAGE_SIZE
== 0 &&
1889 kernel_map_pages(virt_to_page(objp
),
1890 cachep
->size
/ PAGE_SIZE
, 1);
1892 check_poison_obj(cachep
, objp
);
1894 check_poison_obj(cachep
, objp
);
1897 if (cachep
->flags
& SLAB_RED_ZONE
) {
1898 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1899 slab_error(cachep
, "start of a freed object "
1901 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1902 slab_error(cachep
, "end of a freed object "
1908 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1915 * slab_destroy - destroy and release all objects in a slab
1916 * @cachep: cache pointer being destroyed
1917 * @page: page pointer being destroyed
1919 * Destroy all the objs in a slab page, and release the mem back to the system.
1920 * Before calling the slab page must have been unlinked from the cache. The
1921 * kmem_cache_node ->list_lock is not held/needed.
1923 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1927 freelist
= page
->freelist
;
1928 slab_destroy_debugcheck(cachep
, page
);
1929 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1930 struct rcu_head
*head
;
1933 * RCU free overloads the RCU head over the LRU.
1934 * slab_page has been overloeaded over the LRU,
1935 * however it is not used from now on so that
1936 * we can use it safely.
1938 head
= (void *)&page
->rcu_head
;
1939 call_rcu(head
, kmem_rcu_free
);
1942 kmem_freepages(cachep
, page
);
1946 * From now on, we don't use freelist
1947 * although actual page can be freed in rcu context
1949 if (OFF_SLAB(cachep
))
1950 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1953 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1955 struct page
*page
, *n
;
1957 list_for_each_entry_safe(page
, n
, list
, lru
) {
1958 list_del(&page
->lru
);
1959 slab_destroy(cachep
, page
);
1964 * calculate_slab_order - calculate size (page order) of slabs
1965 * @cachep: pointer to the cache that is being created
1966 * @size: size of objects to be created in this cache.
1967 * @align: required alignment for the objects.
1968 * @flags: slab allocation flags
1970 * Also calculates the number of objects per slab.
1972 * This could be made much more intelligent. For now, try to avoid using
1973 * high order pages for slabs. When the gfp() functions are more friendly
1974 * towards high-order requests, this should be changed.
1976 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1977 size_t size
, size_t align
, unsigned long flags
)
1979 unsigned long offslab_limit
;
1980 size_t left_over
= 0;
1983 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1987 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1991 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1992 if (num
> SLAB_OBJ_MAX_NUM
)
1995 if (flags
& CFLGS_OFF_SLAB
) {
1996 size_t freelist_size_per_obj
= sizeof(freelist_idx_t
);
1998 * Max number of objs-per-slab for caches which
1999 * use off-slab slabs. Needed to avoid a possible
2000 * looping condition in cache_grow().
2002 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
2003 freelist_size_per_obj
+= sizeof(char);
2004 offslab_limit
= size
;
2005 offslab_limit
/= freelist_size_per_obj
;
2007 if (num
> offslab_limit
)
2011 /* Found something acceptable - save it away */
2013 cachep
->gfporder
= gfporder
;
2014 left_over
= remainder
;
2017 * A VFS-reclaimable slab tends to have most allocations
2018 * as GFP_NOFS and we really don't want to have to be allocating
2019 * higher-order pages when we are unable to shrink dcache.
2021 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2025 * Large number of objects is good, but very large slabs are
2026 * currently bad for the gfp()s.
2028 if (gfporder
>= slab_max_order
)
2032 * Acceptable internal fragmentation?
2034 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2040 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2042 if (slab_state
>= FULL
)
2043 return enable_cpucache(cachep
, gfp
);
2045 if (slab_state
== DOWN
) {
2047 * Note: Creation of first cache (kmem_cache).
2048 * The setup_node is taken care
2049 * of by the caller of __kmem_cache_create
2051 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2052 slab_state
= PARTIAL
;
2053 } else if (slab_state
== PARTIAL
) {
2055 * Note: the second kmem_cache_create must create the cache
2056 * that's used by kmalloc(24), otherwise the creation of
2057 * further caches will BUG().
2059 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2062 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2063 * the second cache, then we need to set up all its node/,
2064 * otherwise the creation of further caches will BUG().
2066 set_up_node(cachep
, SIZE_AC
);
2067 if (INDEX_AC
== INDEX_NODE
)
2068 slab_state
= PARTIAL_NODE
;
2070 slab_state
= PARTIAL_ARRAYCACHE
;
2072 /* Remaining boot caches */
2073 cachep
->array
[smp_processor_id()] =
2074 kmalloc(sizeof(struct arraycache_init
), gfp
);
2076 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2077 set_up_node(cachep
, SIZE_NODE
);
2078 slab_state
= PARTIAL_NODE
;
2081 for_each_online_node(node
) {
2082 cachep
->node
[node
] =
2083 kmalloc_node(sizeof(struct kmem_cache_node
),
2085 BUG_ON(!cachep
->node
[node
]);
2086 kmem_cache_node_init(cachep
->node
[node
]);
2090 cachep
->node
[numa_mem_id()]->next_reap
=
2091 jiffies
+ REAPTIMEOUT_NODE
+
2092 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
2094 cpu_cache_get(cachep
)->avail
= 0;
2095 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2096 cpu_cache_get(cachep
)->batchcount
= 1;
2097 cpu_cache_get(cachep
)->touched
= 0;
2098 cachep
->batchcount
= 1;
2099 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2104 * __kmem_cache_create - Create a cache.
2105 * @cachep: cache management descriptor
2106 * @flags: SLAB flags
2108 * Returns a ptr to the cache on success, NULL on failure.
2109 * Cannot be called within a int, but can be interrupted.
2110 * The @ctor is run when new pages are allocated by the cache.
2114 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2115 * to catch references to uninitialised memory.
2117 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2118 * for buffer overruns.
2120 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2121 * cacheline. This can be beneficial if you're counting cycles as closely
2125 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2127 size_t left_over
, freelist_size
, ralign
;
2130 size_t size
= cachep
->size
;
2135 * Enable redzoning and last user accounting, except for caches with
2136 * large objects, if the increased size would increase the object size
2137 * above the next power of two: caches with object sizes just above a
2138 * power of two have a significant amount of internal fragmentation.
2140 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2141 2 * sizeof(unsigned long long)))
2142 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2143 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2144 flags
|= SLAB_POISON
;
2146 if (flags
& SLAB_DESTROY_BY_RCU
)
2147 BUG_ON(flags
& SLAB_POISON
);
2151 * Check that size is in terms of words. This is needed to avoid
2152 * unaligned accesses for some archs when redzoning is used, and makes
2153 * sure any on-slab bufctl's are also correctly aligned.
2155 if (size
& (BYTES_PER_WORD
- 1)) {
2156 size
+= (BYTES_PER_WORD
- 1);
2157 size
&= ~(BYTES_PER_WORD
- 1);
2161 * Redzoning and user store require word alignment or possibly larger.
2162 * Note this will be overridden by architecture or caller mandated
2163 * alignment if either is greater than BYTES_PER_WORD.
2165 if (flags
& SLAB_STORE_USER
)
2166 ralign
= BYTES_PER_WORD
;
2168 if (flags
& SLAB_RED_ZONE
) {
2169 ralign
= REDZONE_ALIGN
;
2170 /* If redzoning, ensure that the second redzone is suitably
2171 * aligned, by adjusting the object size accordingly. */
2172 size
+= REDZONE_ALIGN
- 1;
2173 size
&= ~(REDZONE_ALIGN
- 1);
2176 /* 3) caller mandated alignment */
2177 if (ralign
< cachep
->align
) {
2178 ralign
= cachep
->align
;
2180 /* disable debug if necessary */
2181 if (ralign
> __alignof__(unsigned long long))
2182 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2186 cachep
->align
= ralign
;
2188 if (slab_is_available())
2193 setup_node_pointer(cachep
);
2197 * Both debugging options require word-alignment which is calculated
2200 if (flags
& SLAB_RED_ZONE
) {
2201 /* add space for red zone words */
2202 cachep
->obj_offset
+= sizeof(unsigned long long);
2203 size
+= 2 * sizeof(unsigned long long);
2205 if (flags
& SLAB_STORE_USER
) {
2206 /* user store requires one word storage behind the end of
2207 * the real object. But if the second red zone needs to be
2208 * aligned to 64 bits, we must allow that much space.
2210 if (flags
& SLAB_RED_ZONE
)
2211 size
+= REDZONE_ALIGN
;
2213 size
+= BYTES_PER_WORD
;
2215 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2216 if (size
>= kmalloc_size(INDEX_NODE
+ 1)
2217 && cachep
->object_size
> cache_line_size()
2218 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2219 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2226 * Determine if the slab management is 'on' or 'off' slab.
2227 * (bootstrapping cannot cope with offslab caches so don't do
2228 * it too early on. Always use on-slab management when
2229 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2231 if ((size
>= (PAGE_SIZE
>> 5)) && !slab_early_init
&&
2232 !(flags
& SLAB_NOLEAKTRACE
))
2234 * Size is large, assume best to place the slab management obj
2235 * off-slab (should allow better packing of objs).
2237 flags
|= CFLGS_OFF_SLAB
;
2239 size
= ALIGN(size
, cachep
->align
);
2241 * We should restrict the number of objects in a slab to implement
2242 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2244 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2245 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2247 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2252 freelist_size
= calculate_freelist_size(cachep
->num
, cachep
->align
);
2255 * If the slab has been placed off-slab, and we have enough space then
2256 * move it on-slab. This is at the expense of any extra colouring.
2258 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= freelist_size
) {
2259 flags
&= ~CFLGS_OFF_SLAB
;
2260 left_over
-= freelist_size
;
2263 if (flags
& CFLGS_OFF_SLAB
) {
2264 /* really off slab. No need for manual alignment */
2265 freelist_size
= calculate_freelist_size(cachep
->num
, 0);
2267 #ifdef CONFIG_PAGE_POISONING
2268 /* If we're going to use the generic kernel_map_pages()
2269 * poisoning, then it's going to smash the contents of
2270 * the redzone and userword anyhow, so switch them off.
2272 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2273 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2277 cachep
->colour_off
= cache_line_size();
2278 /* Offset must be a multiple of the alignment. */
2279 if (cachep
->colour_off
< cachep
->align
)
2280 cachep
->colour_off
= cachep
->align
;
2281 cachep
->colour
= left_over
/ cachep
->colour_off
;
2282 cachep
->freelist_size
= freelist_size
;
2283 cachep
->flags
= flags
;
2284 cachep
->allocflags
= __GFP_COMP
;
2285 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2286 cachep
->allocflags
|= GFP_DMA
;
2287 cachep
->size
= size
;
2288 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2290 if (flags
& CFLGS_OFF_SLAB
) {
2291 cachep
->freelist_cache
= kmalloc_slab(freelist_size
, 0u);
2293 * This is a possibility for one of the kmalloc_{dma,}_caches.
2294 * But since we go off slab only for object size greater than
2295 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2296 * in ascending order,this should not happen at all.
2297 * But leave a BUG_ON for some lucky dude.
2299 BUG_ON(ZERO_OR_NULL_PTR(cachep
->freelist_cache
));
2302 err
= setup_cpu_cache(cachep
, gfp
);
2304 __kmem_cache_shutdown(cachep
);
2312 static void check_irq_off(void)
2314 BUG_ON(!irqs_disabled());
2317 static void check_irq_on(void)
2319 BUG_ON(irqs_disabled());
2322 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2326 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2330 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2334 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2339 #define check_irq_off() do { } while(0)
2340 #define check_irq_on() do { } while(0)
2341 #define check_spinlock_acquired(x) do { } while(0)
2342 #define check_spinlock_acquired_node(x, y) do { } while(0)
2345 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2346 struct array_cache
*ac
,
2347 int force
, int node
);
2349 static void do_drain(void *arg
)
2351 struct kmem_cache
*cachep
= arg
;
2352 struct array_cache
*ac
;
2353 int node
= numa_mem_id();
2354 struct kmem_cache_node
*n
;
2358 ac
= cpu_cache_get(cachep
);
2359 n
= get_node(cachep
, node
);
2360 spin_lock(&n
->list_lock
);
2361 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2362 spin_unlock(&n
->list_lock
);
2363 slabs_destroy(cachep
, &list
);
2367 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2369 struct kmem_cache_node
*n
;
2372 on_each_cpu(do_drain
, cachep
, 1);
2374 for_each_kmem_cache_node(cachep
, node
, n
)
2376 drain_alien_cache(cachep
, n
->alien
);
2378 for_each_kmem_cache_node(cachep
, node
, n
)
2379 drain_array(cachep
, n
, n
->shared
, 1, node
);
2383 * Remove slabs from the list of free slabs.
2384 * Specify the number of slabs to drain in tofree.
2386 * Returns the actual number of slabs released.
2388 static int drain_freelist(struct kmem_cache
*cache
,
2389 struct kmem_cache_node
*n
, int tofree
)
2391 struct list_head
*p
;
2396 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2398 spin_lock_irq(&n
->list_lock
);
2399 p
= n
->slabs_free
.prev
;
2400 if (p
== &n
->slabs_free
) {
2401 spin_unlock_irq(&n
->list_lock
);
2405 page
= list_entry(p
, struct page
, lru
);
2407 BUG_ON(page
->active
);
2409 list_del(&page
->lru
);
2411 * Safe to drop the lock. The slab is no longer linked
2414 n
->free_objects
-= cache
->num
;
2415 spin_unlock_irq(&n
->list_lock
);
2416 slab_destroy(cache
, page
);
2423 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2427 struct kmem_cache_node
*n
;
2429 drain_cpu_caches(cachep
);
2432 for_each_kmem_cache_node(cachep
, node
, n
) {
2433 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2435 ret
+= !list_empty(&n
->slabs_full
) ||
2436 !list_empty(&n
->slabs_partial
);
2438 return (ret
? 1 : 0);
2441 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2444 struct kmem_cache_node
*n
;
2445 int rc
= __kmem_cache_shrink(cachep
);
2450 for_each_online_cpu(i
)
2451 kfree(cachep
->array
[i
]);
2453 /* NUMA: free the node structures */
2454 for_each_kmem_cache_node(cachep
, i
, n
) {
2456 free_alien_cache(n
->alien
);
2458 cachep
->node
[i
] = NULL
;
2464 * Get the memory for a slab management obj.
2466 * For a slab cache when the slab descriptor is off-slab, the
2467 * slab descriptor can't come from the same cache which is being created,
2468 * Because if it is the case, that means we defer the creation of
2469 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2470 * And we eventually call down to __kmem_cache_create(), which
2471 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2472 * This is a "chicken-and-egg" problem.
2474 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2475 * which are all initialized during kmem_cache_init().
2477 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2478 struct page
*page
, int colour_off
,
2479 gfp_t local_flags
, int nodeid
)
2482 void *addr
= page_address(page
);
2484 if (OFF_SLAB(cachep
)) {
2485 /* Slab management obj is off-slab. */
2486 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2487 local_flags
, nodeid
);
2491 freelist
= addr
+ colour_off
;
2492 colour_off
+= cachep
->freelist_size
;
2495 page
->s_mem
= addr
+ colour_off
;
2499 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2501 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2504 static inline void set_free_obj(struct page
*page
,
2505 unsigned int idx
, freelist_idx_t val
)
2507 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2510 static void cache_init_objs(struct kmem_cache
*cachep
,
2515 for (i
= 0; i
< cachep
->num
; i
++) {
2516 void *objp
= index_to_obj(cachep
, page
, i
);
2518 /* need to poison the objs? */
2519 if (cachep
->flags
& SLAB_POISON
)
2520 poison_obj(cachep
, objp
, POISON_FREE
);
2521 if (cachep
->flags
& SLAB_STORE_USER
)
2522 *dbg_userword(cachep
, objp
) = NULL
;
2524 if (cachep
->flags
& SLAB_RED_ZONE
) {
2525 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2526 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2529 * Constructors are not allowed to allocate memory from the same
2530 * cache which they are a constructor for. Otherwise, deadlock.
2531 * They must also be threaded.
2533 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2534 cachep
->ctor(objp
+ obj_offset(cachep
));
2536 if (cachep
->flags
& SLAB_RED_ZONE
) {
2537 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2538 slab_error(cachep
, "constructor overwrote the"
2539 " end of an object");
2540 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2541 slab_error(cachep
, "constructor overwrote the"
2542 " start of an object");
2544 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2545 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2546 kernel_map_pages(virt_to_page(objp
),
2547 cachep
->size
/ PAGE_SIZE
, 0);
2552 set_obj_status(page
, i
, OBJECT_FREE
);
2553 set_free_obj(page
, i
, i
);
2557 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2559 if (CONFIG_ZONE_DMA_FLAG
) {
2560 if (flags
& GFP_DMA
)
2561 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2563 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2567 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
,
2572 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2575 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2581 static void slab_put_obj(struct kmem_cache
*cachep
, struct page
*page
,
2582 void *objp
, int nodeid
)
2584 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2588 /* Verify that the slab belongs to the intended node */
2589 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2591 /* Verify double free bug */
2592 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2593 if (get_free_obj(page
, i
) == objnr
) {
2594 printk(KERN_ERR
"slab: double free detected in cache "
2595 "'%s', objp %p\n", cachep
->name
, objp
);
2601 set_free_obj(page
, page
->active
, objnr
);
2605 * Map pages beginning at addr to the given cache and slab. This is required
2606 * for the slab allocator to be able to lookup the cache and slab of a
2607 * virtual address for kfree, ksize, and slab debugging.
2609 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2612 page
->slab_cache
= cache
;
2613 page
->freelist
= freelist
;
2617 * Grow (by 1) the number of slabs within a cache. This is called by
2618 * kmem_cache_alloc() when there are no active objs left in a cache.
2620 static int cache_grow(struct kmem_cache
*cachep
,
2621 gfp_t flags
, int nodeid
, struct page
*page
)
2626 struct kmem_cache_node
*n
;
2629 * Be lazy and only check for valid flags here, keeping it out of the
2630 * critical path in kmem_cache_alloc().
2632 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2633 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2635 /* Take the node list lock to change the colour_next on this node */
2637 n
= get_node(cachep
, nodeid
);
2638 spin_lock(&n
->list_lock
);
2640 /* Get colour for the slab, and cal the next value. */
2641 offset
= n
->colour_next
;
2643 if (n
->colour_next
>= cachep
->colour
)
2645 spin_unlock(&n
->list_lock
);
2647 offset
*= cachep
->colour_off
;
2649 if (local_flags
& __GFP_WAIT
)
2653 * The test for missing atomic flag is performed here, rather than
2654 * the more obvious place, simply to reduce the critical path length
2655 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2656 * will eventually be caught here (where it matters).
2658 kmem_flagcheck(cachep
, flags
);
2661 * Get mem for the objs. Attempt to allocate a physical page from
2665 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2669 /* Get slab management. */
2670 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2671 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2675 slab_map_pages(cachep
, page
, freelist
);
2677 cache_init_objs(cachep
, page
);
2679 if (local_flags
& __GFP_WAIT
)
2680 local_irq_disable();
2682 spin_lock(&n
->list_lock
);
2684 /* Make slab active. */
2685 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2686 STATS_INC_GROWN(cachep
);
2687 n
->free_objects
+= cachep
->num
;
2688 spin_unlock(&n
->list_lock
);
2691 kmem_freepages(cachep
, page
);
2693 if (local_flags
& __GFP_WAIT
)
2694 local_irq_disable();
2701 * Perform extra freeing checks:
2702 * - detect bad pointers.
2703 * - POISON/RED_ZONE checking
2705 static void kfree_debugcheck(const void *objp
)
2707 if (!virt_addr_valid(objp
)) {
2708 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2709 (unsigned long)objp
);
2714 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2716 unsigned long long redzone1
, redzone2
;
2718 redzone1
= *dbg_redzone1(cache
, obj
);
2719 redzone2
= *dbg_redzone2(cache
, obj
);
2724 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2727 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2728 slab_error(cache
, "double free detected");
2730 slab_error(cache
, "memory outside object was overwritten");
2732 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2733 obj
, redzone1
, redzone2
);
2736 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2737 unsigned long caller
)
2742 BUG_ON(virt_to_cache(objp
) != cachep
);
2744 objp
-= obj_offset(cachep
);
2745 kfree_debugcheck(objp
);
2746 page
= virt_to_head_page(objp
);
2748 if (cachep
->flags
& SLAB_RED_ZONE
) {
2749 verify_redzone_free(cachep
, objp
);
2750 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2751 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2753 if (cachep
->flags
& SLAB_STORE_USER
)
2754 *dbg_userword(cachep
, objp
) = (void *)caller
;
2756 objnr
= obj_to_index(cachep
, page
, objp
);
2758 BUG_ON(objnr
>= cachep
->num
);
2759 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2761 set_obj_status(page
, objnr
, OBJECT_FREE
);
2762 if (cachep
->flags
& SLAB_POISON
) {
2763 #ifdef CONFIG_DEBUG_PAGEALLOC
2764 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2765 store_stackinfo(cachep
, objp
, caller
);
2766 kernel_map_pages(virt_to_page(objp
),
2767 cachep
->size
/ PAGE_SIZE
, 0);
2769 poison_obj(cachep
, objp
, POISON_FREE
);
2772 poison_obj(cachep
, objp
, POISON_FREE
);
2779 #define kfree_debugcheck(x) do { } while(0)
2780 #define cache_free_debugcheck(x,objp,z) (objp)
2783 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2787 struct kmem_cache_node
*n
;
2788 struct array_cache
*ac
;
2792 node
= numa_mem_id();
2793 if (unlikely(force_refill
))
2796 ac
= cpu_cache_get(cachep
);
2797 batchcount
= ac
->batchcount
;
2798 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2800 * If there was little recent activity on this cache, then
2801 * perform only a partial refill. Otherwise we could generate
2804 batchcount
= BATCHREFILL_LIMIT
;
2806 n
= get_node(cachep
, node
);
2808 BUG_ON(ac
->avail
> 0 || !n
);
2809 spin_lock(&n
->list_lock
);
2811 /* See if we can refill from the shared array */
2812 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2813 n
->shared
->touched
= 1;
2817 while (batchcount
> 0) {
2818 struct list_head
*entry
;
2820 /* Get slab alloc is to come from. */
2821 entry
= n
->slabs_partial
.next
;
2822 if (entry
== &n
->slabs_partial
) {
2823 n
->free_touched
= 1;
2824 entry
= n
->slabs_free
.next
;
2825 if (entry
== &n
->slabs_free
)
2829 page
= list_entry(entry
, struct page
, lru
);
2830 check_spinlock_acquired(cachep
);
2833 * The slab was either on partial or free list so
2834 * there must be at least one object available for
2837 BUG_ON(page
->active
>= cachep
->num
);
2839 while (page
->active
< cachep
->num
&& batchcount
--) {
2840 STATS_INC_ALLOCED(cachep
);
2841 STATS_INC_ACTIVE(cachep
);
2842 STATS_SET_HIGH(cachep
);
2844 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, page
,
2848 /* move slabp to correct slabp list: */
2849 list_del(&page
->lru
);
2850 if (page
->active
== cachep
->num
)
2851 list_add(&page
->lru
, &n
->slabs_full
);
2853 list_add(&page
->lru
, &n
->slabs_partial
);
2857 n
->free_objects
-= ac
->avail
;
2859 spin_unlock(&n
->list_lock
);
2861 if (unlikely(!ac
->avail
)) {
2864 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2866 /* cache_grow can reenable interrupts, then ac could change. */
2867 ac
= cpu_cache_get(cachep
);
2868 node
= numa_mem_id();
2870 /* no objects in sight? abort */
2871 if (!x
&& (ac
->avail
== 0 || force_refill
))
2874 if (!ac
->avail
) /* objects refilled by interrupt? */
2879 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2882 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2885 might_sleep_if(flags
& __GFP_WAIT
);
2887 kmem_flagcheck(cachep
, flags
);
2892 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2893 gfp_t flags
, void *objp
, unsigned long caller
)
2899 if (cachep
->flags
& SLAB_POISON
) {
2900 #ifdef CONFIG_DEBUG_PAGEALLOC
2901 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2902 kernel_map_pages(virt_to_page(objp
),
2903 cachep
->size
/ PAGE_SIZE
, 1);
2905 check_poison_obj(cachep
, objp
);
2907 check_poison_obj(cachep
, objp
);
2909 poison_obj(cachep
, objp
, POISON_INUSE
);
2911 if (cachep
->flags
& SLAB_STORE_USER
)
2912 *dbg_userword(cachep
, objp
) = (void *)caller
;
2914 if (cachep
->flags
& SLAB_RED_ZONE
) {
2915 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2916 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2917 slab_error(cachep
, "double free, or memory outside"
2918 " object was overwritten");
2920 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2921 objp
, *dbg_redzone1(cachep
, objp
),
2922 *dbg_redzone2(cachep
, objp
));
2924 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2925 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2928 page
= virt_to_head_page(objp
);
2929 set_obj_status(page
, obj_to_index(cachep
, page
, objp
), OBJECT_ACTIVE
);
2930 objp
+= obj_offset(cachep
);
2931 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2933 if (ARCH_SLAB_MINALIGN
&&
2934 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2935 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2936 objp
, (int)ARCH_SLAB_MINALIGN
);
2941 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2944 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
2946 if (unlikely(cachep
== kmem_cache
))
2949 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
2952 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2955 struct array_cache
*ac
;
2956 bool force_refill
= false;
2960 ac
= cpu_cache_get(cachep
);
2961 if (likely(ac
->avail
)) {
2963 objp
= ac_get_obj(cachep
, ac
, flags
, false);
2966 * Allow for the possibility all avail objects are not allowed
2967 * by the current flags
2970 STATS_INC_ALLOCHIT(cachep
);
2973 force_refill
= true;
2976 STATS_INC_ALLOCMISS(cachep
);
2977 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
2979 * the 'ac' may be updated by cache_alloc_refill(),
2980 * and kmemleak_erase() requires its correct value.
2982 ac
= cpu_cache_get(cachep
);
2986 * To avoid a false negative, if an object that is in one of the
2987 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2988 * treat the array pointers as a reference to the object.
2991 kmemleak_erase(&ac
->entry
[ac
->avail
]);
2997 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
2999 * If we are in_interrupt, then process context, including cpusets and
3000 * mempolicy, may not apply and should not be used for allocation policy.
3002 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3004 int nid_alloc
, nid_here
;
3006 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3008 nid_alloc
= nid_here
= numa_mem_id();
3009 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3010 nid_alloc
= cpuset_slab_spread_node();
3011 else if (current
->mempolicy
)
3012 nid_alloc
= mempolicy_slab_node();
3013 if (nid_alloc
!= nid_here
)
3014 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3019 * Fallback function if there was no memory available and no objects on a
3020 * certain node and fall back is permitted. First we scan all the
3021 * available node for available objects. If that fails then we
3022 * perform an allocation without specifying a node. This allows the page
3023 * allocator to do its reclaim / fallback magic. We then insert the
3024 * slab into the proper nodelist and then allocate from it.
3026 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3028 struct zonelist
*zonelist
;
3032 enum zone_type high_zoneidx
= gfp_zone(flags
);
3035 unsigned int cpuset_mems_cookie
;
3037 if (flags
& __GFP_THISNODE
)
3040 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3043 cpuset_mems_cookie
= read_mems_allowed_begin();
3044 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3048 * Look through allowed nodes for objects available
3049 * from existing per node queues.
3051 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3052 nid
= zone_to_nid(zone
);
3054 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3055 get_node(cache
, nid
) &&
3056 get_node(cache
, nid
)->free_objects
) {
3057 obj
= ____cache_alloc_node(cache
,
3058 flags
| GFP_THISNODE
, nid
);
3066 * This allocation will be performed within the constraints
3067 * of the current cpuset / memory policy requirements.
3068 * We may trigger various forms of reclaim on the allowed
3069 * set and go into memory reserves if necessary.
3073 if (local_flags
& __GFP_WAIT
)
3075 kmem_flagcheck(cache
, flags
);
3076 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3077 if (local_flags
& __GFP_WAIT
)
3078 local_irq_disable();
3081 * Insert into the appropriate per node queues
3083 nid
= page_to_nid(page
);
3084 if (cache_grow(cache
, flags
, nid
, page
)) {
3085 obj
= ____cache_alloc_node(cache
,
3086 flags
| GFP_THISNODE
, nid
);
3089 * Another processor may allocate the
3090 * objects in the slab since we are
3091 * not holding any locks.
3095 /* cache_grow already freed obj */
3101 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3107 * A interface to enable slab creation on nodeid
3109 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3112 struct list_head
*entry
;
3114 struct kmem_cache_node
*n
;
3118 VM_BUG_ON(nodeid
> num_online_nodes());
3119 n
= get_node(cachep
, nodeid
);
3124 spin_lock(&n
->list_lock
);
3125 entry
= n
->slabs_partial
.next
;
3126 if (entry
== &n
->slabs_partial
) {
3127 n
->free_touched
= 1;
3128 entry
= n
->slabs_free
.next
;
3129 if (entry
== &n
->slabs_free
)
3133 page
= list_entry(entry
, struct page
, lru
);
3134 check_spinlock_acquired_node(cachep
, nodeid
);
3136 STATS_INC_NODEALLOCS(cachep
);
3137 STATS_INC_ACTIVE(cachep
);
3138 STATS_SET_HIGH(cachep
);
3140 BUG_ON(page
->active
== cachep
->num
);
3142 obj
= slab_get_obj(cachep
, page
, nodeid
);
3144 /* move slabp to correct slabp list: */
3145 list_del(&page
->lru
);
3147 if (page
->active
== cachep
->num
)
3148 list_add(&page
->lru
, &n
->slabs_full
);
3150 list_add(&page
->lru
, &n
->slabs_partial
);
3152 spin_unlock(&n
->list_lock
);
3156 spin_unlock(&n
->list_lock
);
3157 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3161 return fallback_alloc(cachep
, flags
);
3167 static __always_inline
void *
3168 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3169 unsigned long caller
)
3171 unsigned long save_flags
;
3173 int slab_node
= numa_mem_id();
3175 flags
&= gfp_allowed_mask
;
3177 lockdep_trace_alloc(flags
);
3179 if (slab_should_failslab(cachep
, flags
))
3182 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3184 cache_alloc_debugcheck_before(cachep
, flags
);
3185 local_irq_save(save_flags
);
3187 if (nodeid
== NUMA_NO_NODE
)
3190 if (unlikely(!get_node(cachep
, nodeid
))) {
3191 /* Node not bootstrapped yet */
3192 ptr
= fallback_alloc(cachep
, flags
);
3196 if (nodeid
== slab_node
) {
3198 * Use the locally cached objects if possible.
3199 * However ____cache_alloc does not allow fallback
3200 * to other nodes. It may fail while we still have
3201 * objects on other nodes available.
3203 ptr
= ____cache_alloc(cachep
, flags
);
3207 /* ___cache_alloc_node can fall back to other nodes */
3208 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3210 local_irq_restore(save_flags
);
3211 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3212 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3216 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3217 if (unlikely(flags
& __GFP_ZERO
))
3218 memset(ptr
, 0, cachep
->object_size
);
3224 static __always_inline
void *
3225 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3229 if (current
->mempolicy
|| unlikely(current
->flags
& PF_SPREAD_SLAB
)) {
3230 objp
= alternate_node_alloc(cache
, flags
);
3234 objp
= ____cache_alloc(cache
, flags
);
3237 * We may just have run out of memory on the local node.
3238 * ____cache_alloc_node() knows how to locate memory on other nodes
3241 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3248 static __always_inline
void *
3249 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3251 return ____cache_alloc(cachep
, flags
);
3254 #endif /* CONFIG_NUMA */
3256 static __always_inline
void *
3257 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3259 unsigned long save_flags
;
3262 flags
&= gfp_allowed_mask
;
3264 lockdep_trace_alloc(flags
);
3266 if (slab_should_failslab(cachep
, flags
))
3269 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3271 cache_alloc_debugcheck_before(cachep
, flags
);
3272 local_irq_save(save_flags
);
3273 objp
= __do_cache_alloc(cachep
, flags
);
3274 local_irq_restore(save_flags
);
3275 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3276 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3281 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3282 if (unlikely(flags
& __GFP_ZERO
))
3283 memset(objp
, 0, cachep
->object_size
);
3290 * Caller needs to acquire correct kmem_cache_node's list_lock
3291 * @list: List of detached free slabs should be freed by caller
3293 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3294 int nr_objects
, int node
, struct list_head
*list
)
3297 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3299 for (i
= 0; i
< nr_objects
; i
++) {
3303 clear_obj_pfmemalloc(&objpp
[i
]);
3306 page
= virt_to_head_page(objp
);
3307 list_del(&page
->lru
);
3308 check_spinlock_acquired_node(cachep
, node
);
3309 slab_put_obj(cachep
, page
, objp
, node
);
3310 STATS_DEC_ACTIVE(cachep
);
3313 /* fixup slab chains */
3314 if (page
->active
== 0) {
3315 if (n
->free_objects
> n
->free_limit
) {
3316 n
->free_objects
-= cachep
->num
;
3317 list_add_tail(&page
->lru
, list
);
3319 list_add(&page
->lru
, &n
->slabs_free
);
3322 /* Unconditionally move a slab to the end of the
3323 * partial list on free - maximum time for the
3324 * other objects to be freed, too.
3326 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3331 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3334 struct kmem_cache_node
*n
;
3335 int node
= numa_mem_id();
3338 batchcount
= ac
->batchcount
;
3340 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3343 n
= get_node(cachep
, node
);
3344 spin_lock(&n
->list_lock
);
3346 struct array_cache
*shared_array
= n
->shared
;
3347 int max
= shared_array
->limit
- shared_array
->avail
;
3349 if (batchcount
> max
)
3351 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3352 ac
->entry
, sizeof(void *) * batchcount
);
3353 shared_array
->avail
+= batchcount
;
3358 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3363 struct list_head
*p
;
3365 p
= n
->slabs_free
.next
;
3366 while (p
!= &(n
->slabs_free
)) {
3369 page
= list_entry(p
, struct page
, lru
);
3370 BUG_ON(page
->active
);
3375 STATS_SET_FREEABLE(cachep
, i
);
3378 spin_unlock(&n
->list_lock
);
3379 slabs_destroy(cachep
, &list
);
3380 ac
->avail
-= batchcount
;
3381 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3385 * Release an obj back to its cache. If the obj has a constructed state, it must
3386 * be in this state _before_ it is released. Called with disabled ints.
3388 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3389 unsigned long caller
)
3391 struct array_cache
*ac
= cpu_cache_get(cachep
);
3394 kmemleak_free_recursive(objp
, cachep
->flags
);
3395 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3397 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3400 * Skip calling cache_free_alien() when the platform is not numa.
3401 * This will avoid cache misses that happen while accessing slabp (which
3402 * is per page memory reference) to get nodeid. Instead use a global
3403 * variable to skip the call, which is mostly likely to be present in
3406 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3409 if (likely(ac
->avail
< ac
->limit
)) {
3410 STATS_INC_FREEHIT(cachep
);
3412 STATS_INC_FREEMISS(cachep
);
3413 cache_flusharray(cachep
, ac
);
3416 ac_put_obj(cachep
, ac
, objp
);
3420 * kmem_cache_alloc - Allocate an object
3421 * @cachep: The cache to allocate from.
3422 * @flags: See kmalloc().
3424 * Allocate an object from this cache. The flags are only relevant
3425 * if the cache has no available objects.
3427 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3429 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3431 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3432 cachep
->object_size
, cachep
->size
, flags
);
3436 EXPORT_SYMBOL(kmem_cache_alloc
);
3438 #ifdef CONFIG_TRACING
3440 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3444 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3446 trace_kmalloc(_RET_IP_
, ret
,
3447 size
, cachep
->size
, flags
);
3450 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3455 * kmem_cache_alloc_node - Allocate an object on the specified node
3456 * @cachep: The cache to allocate from.
3457 * @flags: See kmalloc().
3458 * @nodeid: node number of the target node.
3460 * Identical to kmem_cache_alloc but it will allocate memory on the given
3461 * node, which can improve the performance for cpu bound structures.
3463 * Fallback to other node is possible if __GFP_THISNODE is not set.
3465 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3467 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3469 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3470 cachep
->object_size
, cachep
->size
,
3475 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3477 #ifdef CONFIG_TRACING
3478 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3485 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3487 trace_kmalloc_node(_RET_IP_
, ret
,
3492 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3495 static __always_inline
void *
3496 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3498 struct kmem_cache
*cachep
;
3500 cachep
= kmalloc_slab(size
, flags
);
3501 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3503 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3506 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3507 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3509 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3511 EXPORT_SYMBOL(__kmalloc_node
);
3513 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3514 int node
, unsigned long caller
)
3516 return __do_kmalloc_node(size
, flags
, node
, caller
);
3518 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3520 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3522 return __do_kmalloc_node(size
, flags
, node
, 0);
3524 EXPORT_SYMBOL(__kmalloc_node
);
3525 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3526 #endif /* CONFIG_NUMA */
3529 * __do_kmalloc - allocate memory
3530 * @size: how many bytes of memory are required.
3531 * @flags: the type of memory to allocate (see kmalloc).
3532 * @caller: function caller for debug tracking of the caller
3534 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3535 unsigned long caller
)
3537 struct kmem_cache
*cachep
;
3540 cachep
= kmalloc_slab(size
, flags
);
3541 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3543 ret
= slab_alloc(cachep
, flags
, caller
);
3545 trace_kmalloc(caller
, ret
,
3546 size
, cachep
->size
, flags
);
3552 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3553 void *__kmalloc(size_t size
, gfp_t flags
)
3555 return __do_kmalloc(size
, flags
, _RET_IP_
);
3557 EXPORT_SYMBOL(__kmalloc
);
3559 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3561 return __do_kmalloc(size
, flags
, caller
);
3563 EXPORT_SYMBOL(__kmalloc_track_caller
);
3566 void *__kmalloc(size_t size
, gfp_t flags
)
3568 return __do_kmalloc(size
, flags
, 0);
3570 EXPORT_SYMBOL(__kmalloc
);
3574 * kmem_cache_free - Deallocate an object
3575 * @cachep: The cache the allocation was from.
3576 * @objp: The previously allocated object.
3578 * Free an object which was previously allocated from this
3581 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3583 unsigned long flags
;
3584 cachep
= cache_from_obj(cachep
, objp
);
3588 local_irq_save(flags
);
3589 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3590 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3591 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3592 __cache_free(cachep
, objp
, _RET_IP_
);
3593 local_irq_restore(flags
);
3595 trace_kmem_cache_free(_RET_IP_
, objp
);
3597 EXPORT_SYMBOL(kmem_cache_free
);
3600 * kfree - free previously allocated memory
3601 * @objp: pointer returned by kmalloc.
3603 * If @objp is NULL, no operation is performed.
3605 * Don't free memory not originally allocated by kmalloc()
3606 * or you will run into trouble.
3608 void kfree(const void *objp
)
3610 struct kmem_cache
*c
;
3611 unsigned long flags
;
3613 trace_kfree(_RET_IP_
, objp
);
3615 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3617 local_irq_save(flags
);
3618 kfree_debugcheck(objp
);
3619 c
= virt_to_cache(objp
);
3620 debug_check_no_locks_freed(objp
, c
->object_size
);
3622 debug_check_no_obj_freed(objp
, c
->object_size
);
3623 __cache_free(c
, (void *)objp
, _RET_IP_
);
3624 local_irq_restore(flags
);
3626 EXPORT_SYMBOL(kfree
);
3629 * This initializes kmem_cache_node or resizes various caches for all nodes.
3631 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3634 struct kmem_cache_node
*n
;
3635 struct array_cache
*new_shared
;
3636 struct alien_cache
**new_alien
= NULL
;
3638 for_each_online_node(node
) {
3640 if (use_alien_caches
) {
3641 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3647 if (cachep
->shared
) {
3648 new_shared
= alloc_arraycache(node
,
3649 cachep
->shared
*cachep
->batchcount
,
3652 free_alien_cache(new_alien
);
3657 n
= get_node(cachep
, node
);
3659 struct array_cache
*shared
= n
->shared
;
3662 spin_lock_irq(&n
->list_lock
);
3665 free_block(cachep
, shared
->entry
,
3666 shared
->avail
, node
, &list
);
3668 n
->shared
= new_shared
;
3670 n
->alien
= new_alien
;
3673 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3674 cachep
->batchcount
+ cachep
->num
;
3675 spin_unlock_irq(&n
->list_lock
);
3676 slabs_destroy(cachep
, &list
);
3678 free_alien_cache(new_alien
);
3681 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3683 free_alien_cache(new_alien
);
3688 kmem_cache_node_init(n
);
3689 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3690 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3691 n
->shared
= new_shared
;
3692 n
->alien
= new_alien
;
3693 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3694 cachep
->batchcount
+ cachep
->num
;
3695 cachep
->node
[node
] = n
;
3700 if (!cachep
->list
.next
) {
3701 /* Cache is not active yet. Roll back what we did */
3704 n
= get_node(cachep
, node
);
3707 free_alien_cache(n
->alien
);
3709 cachep
->node
[node
] = NULL
;
3717 struct ccupdate_struct
{
3718 struct kmem_cache
*cachep
;
3719 struct array_cache
*new[0];
3722 static void do_ccupdate_local(void *info
)
3724 struct ccupdate_struct
*new = info
;
3725 struct array_cache
*old
;
3728 old
= cpu_cache_get(new->cachep
);
3730 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3731 new->new[smp_processor_id()] = old
;
3734 /* Always called with the slab_mutex held */
3735 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3736 int batchcount
, int shared
, gfp_t gfp
)
3738 struct ccupdate_struct
*new;
3741 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3746 for_each_online_cpu(i
) {
3747 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3750 for (i
--; i
>= 0; i
--)
3756 new->cachep
= cachep
;
3758 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3761 cachep
->batchcount
= batchcount
;
3762 cachep
->limit
= limit
;
3763 cachep
->shared
= shared
;
3765 for_each_online_cpu(i
) {
3767 struct array_cache
*ccold
= new->new[i
];
3769 struct kmem_cache_node
*n
;
3774 node
= cpu_to_mem(i
);
3775 n
= get_node(cachep
, node
);
3776 spin_lock_irq(&n
->list_lock
);
3777 free_block(cachep
, ccold
->entry
, ccold
->avail
, node
, &list
);
3778 spin_unlock_irq(&n
->list_lock
);
3779 slabs_destroy(cachep
, &list
);
3783 return alloc_kmem_cache_node(cachep
, gfp
);
3786 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3787 int batchcount
, int shared
, gfp_t gfp
)
3790 struct kmem_cache
*c
= NULL
;
3793 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3795 if (slab_state
< FULL
)
3798 if ((ret
< 0) || !is_root_cache(cachep
))
3801 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3802 for_each_memcg_cache_index(i
) {
3803 c
= cache_from_memcg_idx(cachep
, i
);
3805 /* return value determined by the parent cache only */
3806 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3812 /* Called with slab_mutex held always */
3813 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3820 if (!is_root_cache(cachep
)) {
3821 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3822 limit
= root
->limit
;
3823 shared
= root
->shared
;
3824 batchcount
= root
->batchcount
;
3827 if (limit
&& shared
&& batchcount
)
3830 * The head array serves three purposes:
3831 * - create a LIFO ordering, i.e. return objects that are cache-warm
3832 * - reduce the number of spinlock operations.
3833 * - reduce the number of linked list operations on the slab and
3834 * bufctl chains: array operations are cheaper.
3835 * The numbers are guessed, we should auto-tune as described by
3838 if (cachep
->size
> 131072)
3840 else if (cachep
->size
> PAGE_SIZE
)
3842 else if (cachep
->size
> 1024)
3844 else if (cachep
->size
> 256)
3850 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3851 * allocation behaviour: Most allocs on one cpu, most free operations
3852 * on another cpu. For these cases, an efficient object passing between
3853 * cpus is necessary. This is provided by a shared array. The array
3854 * replaces Bonwick's magazine layer.
3855 * On uniprocessor, it's functionally equivalent (but less efficient)
3856 * to a larger limit. Thus disabled by default.
3859 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3864 * With debugging enabled, large batchcount lead to excessively long
3865 * periods with disabled local interrupts. Limit the batchcount
3870 batchcount
= (limit
+ 1) / 2;
3872 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3874 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3875 cachep
->name
, -err
);
3880 * Drain an array if it contains any elements taking the node lock only if
3881 * necessary. Note that the node listlock also protects the array_cache
3882 * if drain_array() is used on the shared array.
3884 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3885 struct array_cache
*ac
, int force
, int node
)
3890 if (!ac
|| !ac
->avail
)
3892 if (ac
->touched
&& !force
) {
3895 spin_lock_irq(&n
->list_lock
);
3897 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3898 if (tofree
> ac
->avail
)
3899 tofree
= (ac
->avail
+ 1) / 2;
3900 free_block(cachep
, ac
->entry
, tofree
, node
, &list
);
3901 ac
->avail
-= tofree
;
3902 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3903 sizeof(void *) * ac
->avail
);
3905 spin_unlock_irq(&n
->list_lock
);
3906 slabs_destroy(cachep
, &list
);
3911 * cache_reap - Reclaim memory from caches.
3912 * @w: work descriptor
3914 * Called from workqueue/eventd every few seconds.
3916 * - clear the per-cpu caches for this CPU.
3917 * - return freeable pages to the main free memory pool.
3919 * If we cannot acquire the cache chain mutex then just give up - we'll try
3920 * again on the next iteration.
3922 static void cache_reap(struct work_struct
*w
)
3924 struct kmem_cache
*searchp
;
3925 struct kmem_cache_node
*n
;
3926 int node
= numa_mem_id();
3927 struct delayed_work
*work
= to_delayed_work(w
);
3929 if (!mutex_trylock(&slab_mutex
))
3930 /* Give up. Setup the next iteration. */
3933 list_for_each_entry(searchp
, &slab_caches
, list
) {
3937 * We only take the node lock if absolutely necessary and we
3938 * have established with reasonable certainty that
3939 * we can do some work if the lock was obtained.
3941 n
= get_node(searchp
, node
);
3943 reap_alien(searchp
, n
);
3945 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
3948 * These are racy checks but it does not matter
3949 * if we skip one check or scan twice.
3951 if (time_after(n
->next_reap
, jiffies
))
3954 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3956 drain_array(searchp
, n
, n
->shared
, 0, node
);
3958 if (n
->free_touched
)
3959 n
->free_touched
= 0;
3963 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3964 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3965 STATS_ADD_REAPED(searchp
, freed
);
3971 mutex_unlock(&slab_mutex
);
3974 /* Set up the next iteration */
3975 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3978 #ifdef CONFIG_SLABINFO
3979 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3982 unsigned long active_objs
;
3983 unsigned long num_objs
;
3984 unsigned long active_slabs
= 0;
3985 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3989 struct kmem_cache_node
*n
;
3993 for_each_kmem_cache_node(cachep
, node
, n
) {
3996 spin_lock_irq(&n
->list_lock
);
3998 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
3999 if (page
->active
!= cachep
->num
&& !error
)
4000 error
= "slabs_full accounting error";
4001 active_objs
+= cachep
->num
;
4004 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4005 if (page
->active
== cachep
->num
&& !error
)
4006 error
= "slabs_partial accounting error";
4007 if (!page
->active
&& !error
)
4008 error
= "slabs_partial accounting error";
4009 active_objs
+= page
->active
;
4012 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4013 if (page
->active
&& !error
)
4014 error
= "slabs_free accounting error";
4017 free_objects
+= n
->free_objects
;
4019 shared_avail
+= n
->shared
->avail
;
4021 spin_unlock_irq(&n
->list_lock
);
4023 num_slabs
+= active_slabs
;
4024 num_objs
= num_slabs
* cachep
->num
;
4025 if (num_objs
- active_objs
!= free_objects
&& !error
)
4026 error
= "free_objects accounting error";
4028 name
= cachep
->name
;
4030 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4032 sinfo
->active_objs
= active_objs
;
4033 sinfo
->num_objs
= num_objs
;
4034 sinfo
->active_slabs
= active_slabs
;
4035 sinfo
->num_slabs
= num_slabs
;
4036 sinfo
->shared_avail
= shared_avail
;
4037 sinfo
->limit
= cachep
->limit
;
4038 sinfo
->batchcount
= cachep
->batchcount
;
4039 sinfo
->shared
= cachep
->shared
;
4040 sinfo
->objects_per_slab
= cachep
->num
;
4041 sinfo
->cache_order
= cachep
->gfporder
;
4044 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4048 unsigned long high
= cachep
->high_mark
;
4049 unsigned long allocs
= cachep
->num_allocations
;
4050 unsigned long grown
= cachep
->grown
;
4051 unsigned long reaped
= cachep
->reaped
;
4052 unsigned long errors
= cachep
->errors
;
4053 unsigned long max_freeable
= cachep
->max_freeable
;
4054 unsigned long node_allocs
= cachep
->node_allocs
;
4055 unsigned long node_frees
= cachep
->node_frees
;
4056 unsigned long overflows
= cachep
->node_overflow
;
4058 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4059 "%4lu %4lu %4lu %4lu %4lu",
4060 allocs
, high
, grown
,
4061 reaped
, errors
, max_freeable
, node_allocs
,
4062 node_frees
, overflows
);
4066 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4067 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4068 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4069 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4071 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4072 allochit
, allocmiss
, freehit
, freemiss
);
4077 #define MAX_SLABINFO_WRITE 128
4079 * slabinfo_write - Tuning for the slab allocator
4081 * @buffer: user buffer
4082 * @count: data length
4085 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4086 size_t count
, loff_t
*ppos
)
4088 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4089 int limit
, batchcount
, shared
, res
;
4090 struct kmem_cache
*cachep
;
4092 if (count
> MAX_SLABINFO_WRITE
)
4094 if (copy_from_user(&kbuf
, buffer
, count
))
4096 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4098 tmp
= strchr(kbuf
, ' ');
4103 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4106 /* Find the cache in the chain of caches. */
4107 mutex_lock(&slab_mutex
);
4109 list_for_each_entry(cachep
, &slab_caches
, list
) {
4110 if (!strcmp(cachep
->name
, kbuf
)) {
4111 if (limit
< 1 || batchcount
< 1 ||
4112 batchcount
> limit
|| shared
< 0) {
4115 res
= do_tune_cpucache(cachep
, limit
,
4122 mutex_unlock(&slab_mutex
);
4128 #ifdef CONFIG_DEBUG_SLAB_LEAK
4130 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4132 mutex_lock(&slab_mutex
);
4133 return seq_list_start(&slab_caches
, *pos
);
4136 static inline int add_caller(unsigned long *n
, unsigned long v
)
4146 unsigned long *q
= p
+ 2 * i
;
4160 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4166 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4174 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4175 if (get_obj_status(page
, i
) != OBJECT_ACTIVE
)
4178 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4183 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4185 #ifdef CONFIG_KALLSYMS
4186 unsigned long offset
, size
;
4187 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4189 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4190 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4192 seq_printf(m
, " [%s]", modname
);
4196 seq_printf(m
, "%p", (void *)address
);
4199 static int leaks_show(struct seq_file
*m
, void *p
)
4201 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4203 struct kmem_cache_node
*n
;
4205 unsigned long *x
= m
->private;
4209 if (!(cachep
->flags
& SLAB_STORE_USER
))
4211 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4214 /* OK, we can do it */
4218 for_each_kmem_cache_node(cachep
, node
, n
) {
4221 spin_lock_irq(&n
->list_lock
);
4223 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4224 handle_slab(x
, cachep
, page
);
4225 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4226 handle_slab(x
, cachep
, page
);
4227 spin_unlock_irq(&n
->list_lock
);
4229 name
= cachep
->name
;
4231 /* Increase the buffer size */
4232 mutex_unlock(&slab_mutex
);
4233 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4235 /* Too bad, we are really out */
4237 mutex_lock(&slab_mutex
);
4240 *(unsigned long *)m
->private = x
[0] * 2;
4242 mutex_lock(&slab_mutex
);
4243 /* Now make sure this entry will be retried */
4247 for (i
= 0; i
< x
[1]; i
++) {
4248 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4249 show_symbol(m
, x
[2*i
+2]);
4256 static const struct seq_operations slabstats_op
= {
4257 .start
= leaks_start
,
4263 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4265 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4268 ret
= seq_open(file
, &slabstats_op
);
4270 struct seq_file
*m
= file
->private_data
;
4271 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4280 static const struct file_operations proc_slabstats_operations
= {
4281 .open
= slabstats_open
,
4283 .llseek
= seq_lseek
,
4284 .release
= seq_release_private
,
4288 static int __init
slab_proc_init(void)
4290 #ifdef CONFIG_DEBUG_SLAB_LEAK
4291 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4295 module_init(slab_proc_init
);
4299 * ksize - get the actual amount of memory allocated for a given object
4300 * @objp: Pointer to the object
4302 * kmalloc may internally round up allocations and return more memory
4303 * than requested. ksize() can be used to determine the actual amount of
4304 * memory allocated. The caller may use this additional memory, even though
4305 * a smaller amount of memory was initially specified with the kmalloc call.
4306 * The caller must guarantee that objp points to a valid object previously
4307 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4308 * must not be freed during the duration of the call.
4310 size_t ksize(const void *objp
)
4313 if (unlikely(objp
== ZERO_SIZE_PTR
))
4316 return virt_to_cache(objp
)->object_size
;
4318 EXPORT_SYMBOL(ksize
);