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 (2 * MAX_NUMNODES)
241 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
242 #define CACHE_CACHE 0
243 #define SIZE_NODE (MAX_NUMNODES)
245 static int drain_freelist(struct kmem_cache
*cache
,
246 struct kmem_cache_node
*n
, int tofree
);
247 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
248 int node
, struct list_head
*list
);
249 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
250 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
251 static void cache_reap(struct work_struct
*unused
);
253 static int slab_early_init
= 1;
255 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
257 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
259 INIT_LIST_HEAD(&parent
->slabs_full
);
260 INIT_LIST_HEAD(&parent
->slabs_partial
);
261 INIT_LIST_HEAD(&parent
->slabs_free
);
262 parent
->shared
= NULL
;
263 parent
->alien
= NULL
;
264 parent
->colour_next
= 0;
265 spin_lock_init(&parent
->list_lock
);
266 parent
->free_objects
= 0;
267 parent
->free_touched
= 0;
270 #define MAKE_LIST(cachep, listp, slab, nodeid) \
272 INIT_LIST_HEAD(listp); \
273 list_splice(&get_node(cachep, nodeid)->slab, listp); \
276 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
278 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
279 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
280 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
283 #define CFLGS_OFF_SLAB (0x80000000UL)
284 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
285 #define OFF_SLAB_MIN_SIZE (max_t(size_t, PAGE_SIZE >> 5, KMALLOC_MIN_SIZE + 1))
287 #define BATCHREFILL_LIMIT 16
289 * Optimization question: fewer reaps means less probability for unnessary
290 * cpucache drain/refill cycles.
292 * OTOH the cpuarrays can contain lots of objects,
293 * which could lock up otherwise freeable slabs.
295 #define REAPTIMEOUT_AC (2*HZ)
296 #define REAPTIMEOUT_NODE (4*HZ)
299 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
300 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
301 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
302 #define STATS_INC_GROWN(x) ((x)->grown++)
303 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
304 #define STATS_SET_HIGH(x) \
306 if ((x)->num_active > (x)->high_mark) \
307 (x)->high_mark = (x)->num_active; \
309 #define STATS_INC_ERR(x) ((x)->errors++)
310 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
311 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
312 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
313 #define STATS_SET_FREEABLE(x, i) \
315 if ((x)->max_freeable < i) \
316 (x)->max_freeable = i; \
318 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
319 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
320 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
321 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
323 #define STATS_INC_ACTIVE(x) do { } while (0)
324 #define STATS_DEC_ACTIVE(x) do { } while (0)
325 #define STATS_INC_ALLOCED(x) do { } while (0)
326 #define STATS_INC_GROWN(x) do { } while (0)
327 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
328 #define STATS_SET_HIGH(x) do { } while (0)
329 #define STATS_INC_ERR(x) do { } while (0)
330 #define STATS_INC_NODEALLOCS(x) do { } while (0)
331 #define STATS_INC_NODEFREES(x) do { } while (0)
332 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
333 #define STATS_SET_FREEABLE(x, i) do { } while (0)
334 #define STATS_INC_ALLOCHIT(x) do { } while (0)
335 #define STATS_INC_ALLOCMISS(x) do { } while (0)
336 #define STATS_INC_FREEHIT(x) do { } while (0)
337 #define STATS_INC_FREEMISS(x) do { } while (0)
343 * memory layout of objects:
345 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
346 * the end of an object is aligned with the end of the real
347 * allocation. Catches writes behind the end of the allocation.
348 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
350 * cachep->obj_offset: The real object.
351 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
352 * cachep->size - 1* BYTES_PER_WORD: last caller address
353 * [BYTES_PER_WORD long]
355 static int obj_offset(struct kmem_cache
*cachep
)
357 return cachep
->obj_offset
;
360 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
362 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
363 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
364 sizeof(unsigned long long));
367 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
369 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
370 if (cachep
->flags
& SLAB_STORE_USER
)
371 return (unsigned long long *)(objp
+ cachep
->size
-
372 sizeof(unsigned long long) -
374 return (unsigned long long *) (objp
+ cachep
->size
-
375 sizeof(unsigned long long));
378 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
380 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
381 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
386 #define obj_offset(x) 0
387 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
388 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
393 #define OBJECT_FREE (0)
394 #define OBJECT_ACTIVE (1)
396 #ifdef CONFIG_DEBUG_SLAB_LEAK
398 static void set_obj_status(struct page
*page
, int idx
, int val
)
402 struct kmem_cache
*cachep
= page
->slab_cache
;
404 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
405 status
= (char *)page
->freelist
+ freelist_size
;
409 static inline unsigned int get_obj_status(struct page
*page
, int idx
)
413 struct kmem_cache
*cachep
= page
->slab_cache
;
415 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
416 status
= (char *)page
->freelist
+ freelist_size
;
422 static inline void set_obj_status(struct page
*page
, int idx
, int val
) {}
427 * Do not go above this order unless 0 objects fit into the slab or
428 * overridden on the command line.
430 #define SLAB_MAX_ORDER_HI 1
431 #define SLAB_MAX_ORDER_LO 0
432 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
433 static bool slab_max_order_set __initdata
;
435 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
437 struct page
*page
= virt_to_head_page(obj
);
438 return page
->slab_cache
;
441 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
444 return page
->s_mem
+ cache
->size
* idx
;
448 * We want to avoid an expensive divide : (offset / cache->size)
449 * Using the fact that size is a constant for a particular cache,
450 * we can replace (offset / cache->size) by
451 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
453 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
454 const struct page
*page
, void *obj
)
456 u32 offset
= (obj
- page
->s_mem
);
457 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
460 /* internal cache of cache description objs */
461 static struct kmem_cache kmem_cache_boot
= {
463 .limit
= BOOT_CPUCACHE_ENTRIES
,
465 .size
= sizeof(struct kmem_cache
),
466 .name
= "kmem_cache",
469 #define BAD_ALIEN_MAGIC 0x01020304ul
471 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
473 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
475 return this_cpu_ptr(cachep
->cpu_cache
);
478 static size_t calculate_freelist_size(int nr_objs
, size_t align
)
480 size_t freelist_size
;
482 freelist_size
= nr_objs
* sizeof(freelist_idx_t
);
483 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
484 freelist_size
+= nr_objs
* sizeof(char);
487 freelist_size
= ALIGN(freelist_size
, align
);
489 return freelist_size
;
492 static int calculate_nr_objs(size_t slab_size
, size_t buffer_size
,
493 size_t idx_size
, size_t align
)
496 size_t remained_size
;
497 size_t freelist_size
;
500 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
501 extra_space
= sizeof(char);
503 * Ignore padding for the initial guess. The padding
504 * is at most @align-1 bytes, and @buffer_size is at
505 * least @align. In the worst case, this result will
506 * be one greater than the number of objects that fit
507 * into the memory allocation when taking the padding
510 nr_objs
= slab_size
/ (buffer_size
+ idx_size
+ extra_space
);
513 * This calculated number will be either the right
514 * amount, or one greater than what we want.
516 remained_size
= slab_size
- nr_objs
* buffer_size
;
517 freelist_size
= calculate_freelist_size(nr_objs
, align
);
518 if (remained_size
< freelist_size
)
525 * Calculate the number of objects and left-over bytes for a given buffer size.
527 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
528 size_t align
, int flags
, size_t *left_over
,
533 size_t slab_size
= PAGE_SIZE
<< gfporder
;
536 * The slab management structure can be either off the slab or
537 * on it. For the latter case, the memory allocated for a
540 * - One unsigned int for each object
541 * - Padding to respect alignment of @align
542 * - @buffer_size bytes for each object
544 * If the slab management structure is off the slab, then the
545 * alignment will already be calculated into the size. Because
546 * the slabs are all pages aligned, the objects will be at the
547 * correct alignment when allocated.
549 if (flags
& CFLGS_OFF_SLAB
) {
551 nr_objs
= slab_size
/ buffer_size
;
554 nr_objs
= calculate_nr_objs(slab_size
, buffer_size
,
555 sizeof(freelist_idx_t
), align
);
556 mgmt_size
= calculate_freelist_size(nr_objs
, align
);
559 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
563 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
565 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
568 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
569 function
, cachep
->name
, msg
);
571 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
576 * By default on NUMA we use alien caches to stage the freeing of
577 * objects allocated from other nodes. This causes massive memory
578 * inefficiencies when using fake NUMA setup to split memory into a
579 * large number of small nodes, so it can be disabled on the command
583 static int use_alien_caches __read_mostly
= 1;
584 static int __init
noaliencache_setup(char *s
)
586 use_alien_caches
= 0;
589 __setup("noaliencache", noaliencache_setup
);
591 static int __init
slab_max_order_setup(char *str
)
593 get_option(&str
, &slab_max_order
);
594 slab_max_order
= slab_max_order
< 0 ? 0 :
595 min(slab_max_order
, MAX_ORDER
- 1);
596 slab_max_order_set
= true;
600 __setup("slab_max_order=", slab_max_order_setup
);
604 * Special reaping functions for NUMA systems called from cache_reap().
605 * These take care of doing round robin flushing of alien caches (containing
606 * objects freed on different nodes from which they were allocated) and the
607 * flushing of remote pcps by calling drain_node_pages.
609 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
611 static void init_reap_node(int cpu
)
615 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
616 if (node
== MAX_NUMNODES
)
617 node
= first_node(node_online_map
);
619 per_cpu(slab_reap_node
, cpu
) = node
;
622 static void next_reap_node(void)
624 int node
= __this_cpu_read(slab_reap_node
);
626 node
= next_node(node
, node_online_map
);
627 if (unlikely(node
>= MAX_NUMNODES
))
628 node
= first_node(node_online_map
);
629 __this_cpu_write(slab_reap_node
, node
);
633 #define init_reap_node(cpu) do { } while (0)
634 #define next_reap_node(void) do { } while (0)
638 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
639 * via the workqueue/eventd.
640 * Add the CPU number into the expiration time to minimize the possibility of
641 * the CPUs getting into lockstep and contending for the global cache chain
644 static void start_cpu_timer(int cpu
)
646 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
649 * When this gets called from do_initcalls via cpucache_init(),
650 * init_workqueues() has already run, so keventd will be setup
653 if (keventd_up() && reap_work
->work
.func
== NULL
) {
655 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
656 schedule_delayed_work_on(cpu
, reap_work
,
657 __round_jiffies_relative(HZ
, cpu
));
661 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
664 * The array_cache structures contain pointers to free object.
665 * However, when such objects are allocated or transferred to another
666 * cache the pointers are not cleared and they could be counted as
667 * valid references during a kmemleak scan. Therefore, kmemleak must
668 * not scan such objects.
670 kmemleak_no_scan(ac
);
674 ac
->batchcount
= batch
;
679 static struct array_cache
*alloc_arraycache(int node
, int entries
,
680 int batchcount
, gfp_t gfp
)
682 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
683 struct array_cache
*ac
= NULL
;
685 ac
= kmalloc_node(memsize
, gfp
, node
);
686 init_arraycache(ac
, entries
, batchcount
);
690 static inline bool is_slab_pfmemalloc(struct page
*page
)
692 return PageSlabPfmemalloc(page
);
695 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
696 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
697 struct array_cache
*ac
)
699 struct kmem_cache_node
*n
= get_node(cachep
, numa_mem_id());
703 if (!pfmemalloc_active
)
706 spin_lock_irqsave(&n
->list_lock
, flags
);
707 list_for_each_entry(page
, &n
->slabs_full
, lru
)
708 if (is_slab_pfmemalloc(page
))
711 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
712 if (is_slab_pfmemalloc(page
))
715 list_for_each_entry(page
, &n
->slabs_free
, lru
)
716 if (is_slab_pfmemalloc(page
))
719 pfmemalloc_active
= false;
721 spin_unlock_irqrestore(&n
->list_lock
, flags
);
724 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
725 gfp_t flags
, bool force_refill
)
728 void *objp
= ac
->entry
[--ac
->avail
];
730 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
731 if (unlikely(is_obj_pfmemalloc(objp
))) {
732 struct kmem_cache_node
*n
;
734 if (gfp_pfmemalloc_allowed(flags
)) {
735 clear_obj_pfmemalloc(&objp
);
739 /* The caller cannot use PFMEMALLOC objects, find another one */
740 for (i
= 0; i
< ac
->avail
; i
++) {
741 /* If a !PFMEMALLOC object is found, swap them */
742 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
744 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
745 ac
->entry
[ac
->avail
] = objp
;
751 * If there are empty slabs on the slabs_free list and we are
752 * being forced to refill the cache, mark this one !pfmemalloc.
754 n
= get_node(cachep
, numa_mem_id());
755 if (!list_empty(&n
->slabs_free
) && force_refill
) {
756 struct page
*page
= virt_to_head_page(objp
);
757 ClearPageSlabPfmemalloc(page
);
758 clear_obj_pfmemalloc(&objp
);
759 recheck_pfmemalloc_active(cachep
, ac
);
763 /* No !PFMEMALLOC objects available */
771 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
772 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
776 if (unlikely(sk_memalloc_socks()))
777 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
779 objp
= ac
->entry
[--ac
->avail
];
784 static noinline
void *__ac_put_obj(struct kmem_cache
*cachep
,
785 struct array_cache
*ac
, void *objp
)
787 if (unlikely(pfmemalloc_active
)) {
788 /* Some pfmemalloc slabs exist, check if this is one */
789 struct page
*page
= virt_to_head_page(objp
);
790 if (PageSlabPfmemalloc(page
))
791 set_obj_pfmemalloc(&objp
);
797 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
800 if (unlikely(sk_memalloc_socks()))
801 objp
= __ac_put_obj(cachep
, ac
, objp
);
803 ac
->entry
[ac
->avail
++] = objp
;
807 * Transfer objects in one arraycache to another.
808 * Locking must be handled by the caller.
810 * Return the number of entries transferred.
812 static int transfer_objects(struct array_cache
*to
,
813 struct array_cache
*from
, unsigned int max
)
815 /* Figure out how many entries to transfer */
816 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
821 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
831 #define drain_alien_cache(cachep, alien) do { } while (0)
832 #define reap_alien(cachep, n) do { } while (0)
834 static inline struct alien_cache
**alloc_alien_cache(int node
,
835 int limit
, gfp_t gfp
)
837 return (struct alien_cache
**)BAD_ALIEN_MAGIC
;
840 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
844 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
849 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
855 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
856 gfp_t flags
, int nodeid
)
861 static inline gfp_t
gfp_exact_node(gfp_t flags
)
866 #else /* CONFIG_NUMA */
868 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
869 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
871 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
872 int batch
, gfp_t gfp
)
874 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
875 struct alien_cache
*alc
= NULL
;
877 alc
= kmalloc_node(memsize
, gfp
, node
);
878 init_arraycache(&alc
->ac
, entries
, batch
);
879 spin_lock_init(&alc
->lock
);
883 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
885 struct alien_cache
**alc_ptr
;
886 size_t memsize
= sizeof(void *) * nr_node_ids
;
891 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
896 if (i
== node
|| !node_online(i
))
898 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
900 for (i
--; i
>= 0; i
--)
909 static void free_alien_cache(struct alien_cache
**alc_ptr
)
920 static void __drain_alien_cache(struct kmem_cache
*cachep
,
921 struct array_cache
*ac
, int node
,
922 struct list_head
*list
)
924 struct kmem_cache_node
*n
= get_node(cachep
, node
);
927 spin_lock(&n
->list_lock
);
929 * Stuff objects into the remote nodes shared array first.
930 * That way we could avoid the overhead of putting the objects
931 * into the free lists and getting them back later.
934 transfer_objects(n
->shared
, ac
, ac
->limit
);
936 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
938 spin_unlock(&n
->list_lock
);
943 * Called from cache_reap() to regularly drain alien caches round robin.
945 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
947 int node
= __this_cpu_read(slab_reap_node
);
950 struct alien_cache
*alc
= n
->alien
[node
];
951 struct array_cache
*ac
;
955 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
958 __drain_alien_cache(cachep
, ac
, node
, &list
);
959 spin_unlock_irq(&alc
->lock
);
960 slabs_destroy(cachep
, &list
);
966 static void drain_alien_cache(struct kmem_cache
*cachep
,
967 struct alien_cache
**alien
)
970 struct alien_cache
*alc
;
971 struct array_cache
*ac
;
974 for_each_online_node(i
) {
980 spin_lock_irqsave(&alc
->lock
, flags
);
981 __drain_alien_cache(cachep
, ac
, i
, &list
);
982 spin_unlock_irqrestore(&alc
->lock
, flags
);
983 slabs_destroy(cachep
, &list
);
988 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
989 int node
, int page_node
)
991 struct kmem_cache_node
*n
;
992 struct alien_cache
*alien
= NULL
;
993 struct array_cache
*ac
;
996 n
= get_node(cachep
, node
);
997 STATS_INC_NODEFREES(cachep
);
998 if (n
->alien
&& n
->alien
[page_node
]) {
999 alien
= n
->alien
[page_node
];
1001 spin_lock(&alien
->lock
);
1002 if (unlikely(ac
->avail
== ac
->limit
)) {
1003 STATS_INC_ACOVERFLOW(cachep
);
1004 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
1006 ac_put_obj(cachep
, ac
, objp
);
1007 spin_unlock(&alien
->lock
);
1008 slabs_destroy(cachep
, &list
);
1010 n
= get_node(cachep
, page_node
);
1011 spin_lock(&n
->list_lock
);
1012 free_block(cachep
, &objp
, 1, page_node
, &list
);
1013 spin_unlock(&n
->list_lock
);
1014 slabs_destroy(cachep
, &list
);
1019 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1021 int page_node
= page_to_nid(virt_to_page(objp
));
1022 int node
= numa_mem_id();
1024 * Make sure we are not freeing a object from another node to the array
1025 * cache on this cpu.
1027 if (likely(node
== page_node
))
1030 return __cache_free_alien(cachep
, objp
, node
, page_node
);
1034 * Construct gfp mask to allocate from a specific node but do not invoke reclaim
1035 * or warn about failures.
1037 static inline gfp_t
gfp_exact_node(gfp_t flags
)
1039 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~__GFP_WAIT
;
1044 * Allocates and initializes node for a node on each slab cache, used for
1045 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1046 * will be allocated off-node since memory is not yet online for the new node.
1047 * When hotplugging memory or a cpu, existing node are not replaced if
1050 * Must hold slab_mutex.
1052 static int init_cache_node_node(int node
)
1054 struct kmem_cache
*cachep
;
1055 struct kmem_cache_node
*n
;
1056 const size_t memsize
= sizeof(struct kmem_cache_node
);
1058 list_for_each_entry(cachep
, &slab_caches
, list
) {
1060 * Set up the kmem_cache_node for cpu before we can
1061 * begin anything. Make sure some other cpu on this
1062 * node has not already allocated this
1064 n
= get_node(cachep
, node
);
1066 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1069 kmem_cache_node_init(n
);
1070 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
1071 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1074 * The kmem_cache_nodes don't come and go as CPUs
1075 * come and go. slab_mutex is sufficient
1078 cachep
->node
[node
] = n
;
1081 spin_lock_irq(&n
->list_lock
);
1083 (1 + nr_cpus_node(node
)) *
1084 cachep
->batchcount
+ cachep
->num
;
1085 spin_unlock_irq(&n
->list_lock
);
1090 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1091 struct kmem_cache_node
*n
)
1093 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1096 static void cpuup_canceled(long cpu
)
1098 struct kmem_cache
*cachep
;
1099 struct kmem_cache_node
*n
= NULL
;
1100 int node
= cpu_to_mem(cpu
);
1101 const struct cpumask
*mask
= cpumask_of_node(node
);
1103 list_for_each_entry(cachep
, &slab_caches
, list
) {
1104 struct array_cache
*nc
;
1105 struct array_cache
*shared
;
1106 struct alien_cache
**alien
;
1109 n
= get_node(cachep
, node
);
1113 spin_lock_irq(&n
->list_lock
);
1115 /* Free limit for this kmem_cache_node */
1116 n
->free_limit
-= cachep
->batchcount
;
1118 /* cpu is dead; no one can alloc from it. */
1119 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
1121 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1125 if (!cpumask_empty(mask
)) {
1126 spin_unlock_irq(&n
->list_lock
);
1132 free_block(cachep
, shared
->entry
,
1133 shared
->avail
, node
, &list
);
1140 spin_unlock_irq(&n
->list_lock
);
1144 drain_alien_cache(cachep
, alien
);
1145 free_alien_cache(alien
);
1149 slabs_destroy(cachep
, &list
);
1152 * In the previous loop, all the objects were freed to
1153 * the respective cache's slabs, now we can go ahead and
1154 * shrink each nodelist to its limit.
1156 list_for_each_entry(cachep
, &slab_caches
, list
) {
1157 n
= get_node(cachep
, node
);
1160 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1164 static int cpuup_prepare(long cpu
)
1166 struct kmem_cache
*cachep
;
1167 struct kmem_cache_node
*n
= NULL
;
1168 int node
= cpu_to_mem(cpu
);
1172 * We need to do this right in the beginning since
1173 * alloc_arraycache's are going to use this list.
1174 * kmalloc_node allows us to add the slab to the right
1175 * kmem_cache_node and not this cpu's kmem_cache_node
1177 err
= init_cache_node_node(node
);
1182 * Now we can go ahead with allocating the shared arrays and
1185 list_for_each_entry(cachep
, &slab_caches
, list
) {
1186 struct array_cache
*shared
= NULL
;
1187 struct alien_cache
**alien
= NULL
;
1189 if (cachep
->shared
) {
1190 shared
= alloc_arraycache(node
,
1191 cachep
->shared
* cachep
->batchcount
,
1192 0xbaadf00d, GFP_KERNEL
);
1196 if (use_alien_caches
) {
1197 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1203 n
= get_node(cachep
, node
);
1206 spin_lock_irq(&n
->list_lock
);
1209 * We are serialised from CPU_DEAD or
1210 * CPU_UP_CANCELLED by the cpucontrol lock
1221 spin_unlock_irq(&n
->list_lock
);
1223 free_alien_cache(alien
);
1228 cpuup_canceled(cpu
);
1232 static int cpuup_callback(struct notifier_block
*nfb
,
1233 unsigned long action
, void *hcpu
)
1235 long cpu
= (long)hcpu
;
1239 case CPU_UP_PREPARE
:
1240 case CPU_UP_PREPARE_FROZEN
:
1241 mutex_lock(&slab_mutex
);
1242 err
= cpuup_prepare(cpu
);
1243 mutex_unlock(&slab_mutex
);
1246 case CPU_ONLINE_FROZEN
:
1247 start_cpu_timer(cpu
);
1249 #ifdef CONFIG_HOTPLUG_CPU
1250 case CPU_DOWN_PREPARE
:
1251 case CPU_DOWN_PREPARE_FROZEN
:
1253 * Shutdown cache reaper. Note that the slab_mutex is
1254 * held so that if cache_reap() is invoked it cannot do
1255 * anything expensive but will only modify reap_work
1256 * and reschedule the timer.
1258 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1259 /* Now the cache_reaper is guaranteed to be not running. */
1260 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1262 case CPU_DOWN_FAILED
:
1263 case CPU_DOWN_FAILED_FROZEN
:
1264 start_cpu_timer(cpu
);
1267 case CPU_DEAD_FROZEN
:
1269 * Even if all the cpus of a node are down, we don't free the
1270 * kmem_cache_node of any cache. This to avoid a race between
1271 * cpu_down, and a kmalloc allocation from another cpu for
1272 * memory from the node of the cpu going down. The node
1273 * structure is usually allocated from kmem_cache_create() and
1274 * gets destroyed at kmem_cache_destroy().
1278 case CPU_UP_CANCELED
:
1279 case CPU_UP_CANCELED_FROZEN
:
1280 mutex_lock(&slab_mutex
);
1281 cpuup_canceled(cpu
);
1282 mutex_unlock(&slab_mutex
);
1285 return notifier_from_errno(err
);
1288 static struct notifier_block cpucache_notifier
= {
1289 &cpuup_callback
, NULL
, 0
1292 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1294 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1295 * Returns -EBUSY if all objects cannot be drained so that the node is not
1298 * Must hold slab_mutex.
1300 static int __meminit
drain_cache_node_node(int node
)
1302 struct kmem_cache
*cachep
;
1305 list_for_each_entry(cachep
, &slab_caches
, list
) {
1306 struct kmem_cache_node
*n
;
1308 n
= get_node(cachep
, node
);
1312 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1314 if (!list_empty(&n
->slabs_full
) ||
1315 !list_empty(&n
->slabs_partial
)) {
1323 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1324 unsigned long action
, void *arg
)
1326 struct memory_notify
*mnb
= arg
;
1330 nid
= mnb
->status_change_nid
;
1335 case MEM_GOING_ONLINE
:
1336 mutex_lock(&slab_mutex
);
1337 ret
= init_cache_node_node(nid
);
1338 mutex_unlock(&slab_mutex
);
1340 case MEM_GOING_OFFLINE
:
1341 mutex_lock(&slab_mutex
);
1342 ret
= drain_cache_node_node(nid
);
1343 mutex_unlock(&slab_mutex
);
1347 case MEM_CANCEL_ONLINE
:
1348 case MEM_CANCEL_OFFLINE
:
1352 return notifier_from_errno(ret
);
1354 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1357 * swap the static kmem_cache_node with kmalloced memory
1359 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1362 struct kmem_cache_node
*ptr
;
1364 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1367 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1369 * Do not assume that spinlocks can be initialized via memcpy:
1371 spin_lock_init(&ptr
->list_lock
);
1373 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1374 cachep
->node
[nodeid
] = ptr
;
1378 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1379 * size of kmem_cache_node.
1381 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1385 for_each_online_node(node
) {
1386 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1387 cachep
->node
[node
]->next_reap
= jiffies
+
1389 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1394 * Initialisation. Called after the page allocator have been initialised and
1395 * before smp_init().
1397 void __init
kmem_cache_init(void)
1401 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1402 sizeof(struct rcu_head
));
1403 kmem_cache
= &kmem_cache_boot
;
1405 if (num_possible_nodes() == 1)
1406 use_alien_caches
= 0;
1408 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1409 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1412 * Fragmentation resistance on low memory - only use bigger
1413 * page orders on machines with more than 32MB of memory if
1414 * not overridden on the command line.
1416 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1417 slab_max_order
= SLAB_MAX_ORDER_HI
;
1419 /* Bootstrap is tricky, because several objects are allocated
1420 * from caches that do not exist yet:
1421 * 1) initialize the kmem_cache cache: it contains the struct
1422 * kmem_cache structures of all caches, except kmem_cache itself:
1423 * kmem_cache is statically allocated.
1424 * Initially an __init data area is used for the head array and the
1425 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1426 * array at the end of the bootstrap.
1427 * 2) Create the first kmalloc cache.
1428 * The struct kmem_cache for the new cache is allocated normally.
1429 * An __init data area is used for the head array.
1430 * 3) Create the remaining kmalloc caches, with minimally sized
1432 * 4) Replace the __init data head arrays for kmem_cache and the first
1433 * kmalloc cache with kmalloc allocated arrays.
1434 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1435 * the other cache's with kmalloc allocated memory.
1436 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1439 /* 1) create the kmem_cache */
1442 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1444 create_boot_cache(kmem_cache
, "kmem_cache",
1445 offsetof(struct kmem_cache
, node
) +
1446 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1447 SLAB_HWCACHE_ALIGN
);
1448 list_add(&kmem_cache
->list
, &slab_caches
);
1449 slab_state
= PARTIAL
;
1452 * Initialize the caches that provide memory for the kmem_cache_node
1453 * structures first. Without this, further allocations will bug.
1455 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1456 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1457 slab_state
= PARTIAL_NODE
;
1458 setup_kmalloc_cache_index_table();
1460 slab_early_init
= 0;
1462 /* 5) Replace the bootstrap kmem_cache_node */
1466 for_each_online_node(nid
) {
1467 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1469 init_list(kmalloc_caches
[INDEX_NODE
],
1470 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1474 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1477 void __init
kmem_cache_init_late(void)
1479 struct kmem_cache
*cachep
;
1483 /* 6) resize the head arrays to their final sizes */
1484 mutex_lock(&slab_mutex
);
1485 list_for_each_entry(cachep
, &slab_caches
, list
)
1486 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1488 mutex_unlock(&slab_mutex
);
1494 * Register a cpu startup notifier callback that initializes
1495 * cpu_cache_get for all new cpus
1497 register_cpu_notifier(&cpucache_notifier
);
1501 * Register a memory hotplug callback that initializes and frees
1504 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1508 * The reap timers are started later, with a module init call: That part
1509 * of the kernel is not yet operational.
1513 static int __init
cpucache_init(void)
1518 * Register the timers that return unneeded pages to the page allocator
1520 for_each_online_cpu(cpu
)
1521 start_cpu_timer(cpu
);
1527 __initcall(cpucache_init
);
1529 static noinline
void
1530 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1533 struct kmem_cache_node
*n
;
1535 unsigned long flags
;
1537 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1538 DEFAULT_RATELIMIT_BURST
);
1540 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1544 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1546 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1547 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1549 for_each_kmem_cache_node(cachep
, node
, n
) {
1550 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1551 unsigned long active_slabs
= 0, num_slabs
= 0;
1553 spin_lock_irqsave(&n
->list_lock
, flags
);
1554 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1555 active_objs
+= cachep
->num
;
1558 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1559 active_objs
+= page
->active
;
1562 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1565 free_objects
+= n
->free_objects
;
1566 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1568 num_slabs
+= active_slabs
;
1569 num_objs
= num_slabs
* cachep
->num
;
1571 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1572 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1579 * Interface to system's page allocator. No need to hold the
1580 * kmem_cache_node ->list_lock.
1582 * If we requested dmaable memory, we will get it. Even if we
1583 * did not request dmaable memory, we might get it, but that
1584 * would be relatively rare and ignorable.
1586 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1592 flags
|= cachep
->allocflags
;
1593 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1594 flags
|= __GFP_RECLAIMABLE
;
1596 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1598 slab_out_of_memory(cachep
, flags
, nodeid
);
1602 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1603 __free_pages(page
, cachep
->gfporder
);
1607 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1608 if (page_is_pfmemalloc(page
))
1609 pfmemalloc_active
= true;
1611 nr_pages
= (1 << cachep
->gfporder
);
1612 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1613 add_zone_page_state(page_zone(page
),
1614 NR_SLAB_RECLAIMABLE
, nr_pages
);
1616 add_zone_page_state(page_zone(page
),
1617 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1618 __SetPageSlab(page
);
1619 if (page_is_pfmemalloc(page
))
1620 SetPageSlabPfmemalloc(page
);
1622 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1623 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1626 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1628 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1635 * Interface to system's page release.
1637 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1639 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1641 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1643 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1644 sub_zone_page_state(page_zone(page
),
1645 NR_SLAB_RECLAIMABLE
, nr_freed
);
1647 sub_zone_page_state(page_zone(page
),
1648 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1650 BUG_ON(!PageSlab(page
));
1651 __ClearPageSlabPfmemalloc(page
);
1652 __ClearPageSlab(page
);
1653 page_mapcount_reset(page
);
1654 page
->mapping
= NULL
;
1656 if (current
->reclaim_state
)
1657 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1658 __free_kmem_pages(page
, cachep
->gfporder
);
1661 static void kmem_rcu_free(struct rcu_head
*head
)
1663 struct kmem_cache
*cachep
;
1666 page
= container_of(head
, struct page
, rcu_head
);
1667 cachep
= page
->slab_cache
;
1669 kmem_freepages(cachep
, page
);
1674 #ifdef CONFIG_DEBUG_PAGEALLOC
1675 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1676 unsigned long caller
)
1678 int size
= cachep
->object_size
;
1680 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1682 if (size
< 5 * sizeof(unsigned long))
1685 *addr
++ = 0x12345678;
1687 *addr
++ = smp_processor_id();
1688 size
-= 3 * sizeof(unsigned long);
1690 unsigned long *sptr
= &caller
;
1691 unsigned long svalue
;
1693 while (!kstack_end(sptr
)) {
1695 if (kernel_text_address(svalue
)) {
1697 size
-= sizeof(unsigned long);
1698 if (size
<= sizeof(unsigned long))
1704 *addr
++ = 0x87654321;
1708 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1710 int size
= cachep
->object_size
;
1711 addr
= &((char *)addr
)[obj_offset(cachep
)];
1713 memset(addr
, val
, size
);
1714 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1717 static void dump_line(char *data
, int offset
, int limit
)
1720 unsigned char error
= 0;
1723 printk(KERN_ERR
"%03x: ", offset
);
1724 for (i
= 0; i
< limit
; i
++) {
1725 if (data
[offset
+ i
] != POISON_FREE
) {
1726 error
= data
[offset
+ i
];
1730 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1731 &data
[offset
], limit
, 1);
1733 if (bad_count
== 1) {
1734 error
^= POISON_FREE
;
1735 if (!(error
& (error
- 1))) {
1736 printk(KERN_ERR
"Single bit error detected. Probably "
1739 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1742 printk(KERN_ERR
"Run a memory test tool.\n");
1751 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1756 if (cachep
->flags
& SLAB_RED_ZONE
) {
1757 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1758 *dbg_redzone1(cachep
, objp
),
1759 *dbg_redzone2(cachep
, objp
));
1762 if (cachep
->flags
& SLAB_STORE_USER
) {
1763 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1764 *dbg_userword(cachep
, objp
),
1765 *dbg_userword(cachep
, objp
));
1767 realobj
= (char *)objp
+ obj_offset(cachep
);
1768 size
= cachep
->object_size
;
1769 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1772 if (i
+ limit
> size
)
1774 dump_line(realobj
, i
, limit
);
1778 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1784 realobj
= (char *)objp
+ obj_offset(cachep
);
1785 size
= cachep
->object_size
;
1787 for (i
= 0; i
< size
; i
++) {
1788 char exp
= POISON_FREE
;
1791 if (realobj
[i
] != exp
) {
1797 "Slab corruption (%s): %s start=%p, len=%d\n",
1798 print_tainted(), cachep
->name
, realobj
, size
);
1799 print_objinfo(cachep
, objp
, 0);
1801 /* Hexdump the affected line */
1804 if (i
+ limit
> size
)
1806 dump_line(realobj
, i
, limit
);
1809 /* Limit to 5 lines */
1815 /* Print some data about the neighboring objects, if they
1818 struct page
*page
= virt_to_head_page(objp
);
1821 objnr
= obj_to_index(cachep
, page
, objp
);
1823 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1824 realobj
= (char *)objp
+ obj_offset(cachep
);
1825 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1827 print_objinfo(cachep
, objp
, 2);
1829 if (objnr
+ 1 < cachep
->num
) {
1830 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1831 realobj
= (char *)objp
+ obj_offset(cachep
);
1832 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1834 print_objinfo(cachep
, objp
, 2);
1841 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1845 for (i
= 0; i
< cachep
->num
; i
++) {
1846 void *objp
= index_to_obj(cachep
, page
, i
);
1848 if (cachep
->flags
& SLAB_POISON
) {
1849 #ifdef CONFIG_DEBUG_PAGEALLOC
1850 if (cachep
->size
% PAGE_SIZE
== 0 &&
1852 kernel_map_pages(virt_to_page(objp
),
1853 cachep
->size
/ PAGE_SIZE
, 1);
1855 check_poison_obj(cachep
, objp
);
1857 check_poison_obj(cachep
, objp
);
1860 if (cachep
->flags
& SLAB_RED_ZONE
) {
1861 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1862 slab_error(cachep
, "start of a freed object "
1864 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1865 slab_error(cachep
, "end of a freed object "
1871 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1878 * slab_destroy - destroy and release all objects in a slab
1879 * @cachep: cache pointer being destroyed
1880 * @page: page pointer being destroyed
1882 * Destroy all the objs in a slab page, and release the mem back to the system.
1883 * Before calling the slab page must have been unlinked from the cache. The
1884 * kmem_cache_node ->list_lock is not held/needed.
1886 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1890 freelist
= page
->freelist
;
1891 slab_destroy_debugcheck(cachep
, page
);
1892 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1893 struct rcu_head
*head
;
1896 * RCU free overloads the RCU head over the LRU.
1897 * slab_page has been overloeaded over the LRU,
1898 * however it is not used from now on so that
1899 * we can use it safely.
1901 head
= (void *)&page
->rcu_head
;
1902 call_rcu(head
, kmem_rcu_free
);
1905 kmem_freepages(cachep
, page
);
1909 * From now on, we don't use freelist
1910 * although actual page can be freed in rcu context
1912 if (OFF_SLAB(cachep
))
1913 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1916 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1918 struct page
*page
, *n
;
1920 list_for_each_entry_safe(page
, n
, list
, lru
) {
1921 list_del(&page
->lru
);
1922 slab_destroy(cachep
, page
);
1927 * calculate_slab_order - calculate size (page order) of slabs
1928 * @cachep: pointer to the cache that is being created
1929 * @size: size of objects to be created in this cache.
1930 * @align: required alignment for the objects.
1931 * @flags: slab allocation flags
1933 * Also calculates the number of objects per slab.
1935 * This could be made much more intelligent. For now, try to avoid using
1936 * high order pages for slabs. When the gfp() functions are more friendly
1937 * towards high-order requests, this should be changed.
1939 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1940 size_t size
, size_t align
, unsigned long flags
)
1942 unsigned long offslab_limit
;
1943 size_t left_over
= 0;
1946 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1950 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1954 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1955 if (num
> SLAB_OBJ_MAX_NUM
)
1958 if (flags
& CFLGS_OFF_SLAB
) {
1959 size_t freelist_size_per_obj
= sizeof(freelist_idx_t
);
1961 * Max number of objs-per-slab for caches which
1962 * use off-slab slabs. Needed to avoid a possible
1963 * looping condition in cache_grow().
1965 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
1966 freelist_size_per_obj
+= sizeof(char);
1967 offslab_limit
= size
;
1968 offslab_limit
/= freelist_size_per_obj
;
1970 if (num
> offslab_limit
)
1974 /* Found something acceptable - save it away */
1976 cachep
->gfporder
= gfporder
;
1977 left_over
= remainder
;
1980 * A VFS-reclaimable slab tends to have most allocations
1981 * as GFP_NOFS and we really don't want to have to be allocating
1982 * higher-order pages when we are unable to shrink dcache.
1984 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1988 * Large number of objects is good, but very large slabs are
1989 * currently bad for the gfp()s.
1991 if (gfporder
>= slab_max_order
)
1995 * Acceptable internal fragmentation?
1997 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2003 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
2004 struct kmem_cache
*cachep
, int entries
, int batchcount
)
2008 struct array_cache __percpu
*cpu_cache
;
2010 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
2011 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
2016 for_each_possible_cpu(cpu
) {
2017 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
2018 entries
, batchcount
);
2024 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2026 if (slab_state
>= FULL
)
2027 return enable_cpucache(cachep
, gfp
);
2029 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
2030 if (!cachep
->cpu_cache
)
2033 if (slab_state
== DOWN
) {
2034 /* Creation of first cache (kmem_cache). */
2035 set_up_node(kmem_cache
, CACHE_CACHE
);
2036 } else if (slab_state
== PARTIAL
) {
2037 /* For kmem_cache_node */
2038 set_up_node(cachep
, SIZE_NODE
);
2042 for_each_online_node(node
) {
2043 cachep
->node
[node
] = kmalloc_node(
2044 sizeof(struct kmem_cache_node
), gfp
, node
);
2045 BUG_ON(!cachep
->node
[node
]);
2046 kmem_cache_node_init(cachep
->node
[node
]);
2050 cachep
->node
[numa_mem_id()]->next_reap
=
2051 jiffies
+ REAPTIMEOUT_NODE
+
2052 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
2054 cpu_cache_get(cachep
)->avail
= 0;
2055 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2056 cpu_cache_get(cachep
)->batchcount
= 1;
2057 cpu_cache_get(cachep
)->touched
= 0;
2058 cachep
->batchcount
= 1;
2059 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2063 unsigned long kmem_cache_flags(unsigned long object_size
,
2064 unsigned long flags
, const char *name
,
2065 void (*ctor
)(void *))
2071 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
2072 unsigned long flags
, void (*ctor
)(void *))
2074 struct kmem_cache
*cachep
;
2076 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
2081 * Adjust the object sizes so that we clear
2082 * the complete object on kzalloc.
2084 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
2090 * __kmem_cache_create - Create a cache.
2091 * @cachep: cache management descriptor
2092 * @flags: SLAB flags
2094 * Returns a ptr to the cache on success, NULL on failure.
2095 * Cannot be called within a int, but can be interrupted.
2096 * The @ctor is run when new pages are allocated by the cache.
2100 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2101 * to catch references to uninitialised memory.
2103 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2104 * for buffer overruns.
2106 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2107 * cacheline. This can be beneficial if you're counting cycles as closely
2111 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2113 size_t left_over
, freelist_size
;
2114 size_t ralign
= BYTES_PER_WORD
;
2117 size_t size
= cachep
->size
;
2122 * Enable redzoning and last user accounting, except for caches with
2123 * large objects, if the increased size would increase the object size
2124 * above the next power of two: caches with object sizes just above a
2125 * power of two have a significant amount of internal fragmentation.
2127 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2128 2 * sizeof(unsigned long long)))
2129 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2130 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2131 flags
|= SLAB_POISON
;
2133 if (flags
& SLAB_DESTROY_BY_RCU
)
2134 BUG_ON(flags
& SLAB_POISON
);
2138 * Check that size is in terms of words. This is needed to avoid
2139 * unaligned accesses for some archs when redzoning is used, and makes
2140 * sure any on-slab bufctl's are also correctly aligned.
2142 if (size
& (BYTES_PER_WORD
- 1)) {
2143 size
+= (BYTES_PER_WORD
- 1);
2144 size
&= ~(BYTES_PER_WORD
- 1);
2147 if (flags
& SLAB_RED_ZONE
) {
2148 ralign
= REDZONE_ALIGN
;
2149 /* If redzoning, ensure that the second redzone is suitably
2150 * aligned, by adjusting the object size accordingly. */
2151 size
+= REDZONE_ALIGN
- 1;
2152 size
&= ~(REDZONE_ALIGN
- 1);
2155 /* 3) caller mandated alignment */
2156 if (ralign
< cachep
->align
) {
2157 ralign
= cachep
->align
;
2159 /* disable debug if necessary */
2160 if (ralign
> __alignof__(unsigned long long))
2161 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2165 cachep
->align
= ralign
;
2167 if (slab_is_available())
2175 * Both debugging options require word-alignment which is calculated
2178 if (flags
& SLAB_RED_ZONE
) {
2179 /* add space for red zone words */
2180 cachep
->obj_offset
+= sizeof(unsigned long long);
2181 size
+= 2 * sizeof(unsigned long long);
2183 if (flags
& SLAB_STORE_USER
) {
2184 /* user store requires one word storage behind the end of
2185 * the real object. But if the second red zone needs to be
2186 * aligned to 64 bits, we must allow that much space.
2188 if (flags
& SLAB_RED_ZONE
)
2189 size
+= REDZONE_ALIGN
;
2191 size
+= BYTES_PER_WORD
;
2193 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2195 * To activate debug pagealloc, off-slab management is necessary
2196 * requirement. In early phase of initialization, small sized slab
2197 * doesn't get initialized so it would not be possible. So, we need
2198 * to check size >= 256. It guarantees that all necessary small
2199 * sized slab is initialized in current slab initialization sequence.
2201 if (!slab_early_init
&& size
>= kmalloc_size(INDEX_NODE
) &&
2202 size
>= 256 && cachep
->object_size
> cache_line_size() &&
2203 ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2204 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2211 * Determine if the slab management is 'on' or 'off' slab.
2212 * (bootstrapping cannot cope with offslab caches so don't do
2213 * it too early on. Always use on-slab management when
2214 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2216 if (size
>= OFF_SLAB_MIN_SIZE
&& !slab_early_init
&&
2217 !(flags
& SLAB_NOLEAKTRACE
))
2219 * Size is large, assume best to place the slab management obj
2220 * off-slab (should allow better packing of objs).
2222 flags
|= CFLGS_OFF_SLAB
;
2224 size
= ALIGN(size
, cachep
->align
);
2226 * We should restrict the number of objects in a slab to implement
2227 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2229 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2230 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2232 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2237 freelist_size
= calculate_freelist_size(cachep
->num
, cachep
->align
);
2240 * If the slab has been placed off-slab, and we have enough space then
2241 * move it on-slab. This is at the expense of any extra colouring.
2243 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= freelist_size
) {
2244 flags
&= ~CFLGS_OFF_SLAB
;
2245 left_over
-= freelist_size
;
2248 if (flags
& CFLGS_OFF_SLAB
) {
2249 /* really off slab. No need for manual alignment */
2250 freelist_size
= calculate_freelist_size(cachep
->num
, 0);
2252 #ifdef CONFIG_PAGE_POISONING
2253 /* If we're going to use the generic kernel_map_pages()
2254 * poisoning, then it's going to smash the contents of
2255 * the redzone and userword anyhow, so switch them off.
2257 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2258 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2262 cachep
->colour_off
= cache_line_size();
2263 /* Offset must be a multiple of the alignment. */
2264 if (cachep
->colour_off
< cachep
->align
)
2265 cachep
->colour_off
= cachep
->align
;
2266 cachep
->colour
= left_over
/ cachep
->colour_off
;
2267 cachep
->freelist_size
= freelist_size
;
2268 cachep
->flags
= flags
;
2269 cachep
->allocflags
= __GFP_COMP
;
2270 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2271 cachep
->allocflags
|= GFP_DMA
;
2272 cachep
->size
= size
;
2273 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2275 if (flags
& CFLGS_OFF_SLAB
) {
2276 cachep
->freelist_cache
= kmalloc_slab(freelist_size
, 0u);
2278 * This is a possibility for one of the kmalloc_{dma,}_caches.
2279 * But since we go off slab only for object size greater than
2280 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
2281 * in ascending order,this should not happen at all.
2282 * But leave a BUG_ON for some lucky dude.
2284 BUG_ON(ZERO_OR_NULL_PTR(cachep
->freelist_cache
));
2287 err
= setup_cpu_cache(cachep
, gfp
);
2289 __kmem_cache_shutdown(cachep
);
2297 static void check_irq_off(void)
2299 BUG_ON(!irqs_disabled());
2302 static void check_irq_on(void)
2304 BUG_ON(irqs_disabled());
2307 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2311 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2315 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2319 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2324 #define check_irq_off() do { } while(0)
2325 #define check_irq_on() do { } while(0)
2326 #define check_spinlock_acquired(x) do { } while(0)
2327 #define check_spinlock_acquired_node(x, y) do { } while(0)
2330 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2331 struct array_cache
*ac
,
2332 int force
, int node
);
2334 static void do_drain(void *arg
)
2336 struct kmem_cache
*cachep
= arg
;
2337 struct array_cache
*ac
;
2338 int node
= numa_mem_id();
2339 struct kmem_cache_node
*n
;
2343 ac
= cpu_cache_get(cachep
);
2344 n
= get_node(cachep
, node
);
2345 spin_lock(&n
->list_lock
);
2346 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2347 spin_unlock(&n
->list_lock
);
2348 slabs_destroy(cachep
, &list
);
2352 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2354 struct kmem_cache_node
*n
;
2357 on_each_cpu(do_drain
, cachep
, 1);
2359 for_each_kmem_cache_node(cachep
, node
, n
)
2361 drain_alien_cache(cachep
, n
->alien
);
2363 for_each_kmem_cache_node(cachep
, node
, n
)
2364 drain_array(cachep
, n
, n
->shared
, 1, node
);
2368 * Remove slabs from the list of free slabs.
2369 * Specify the number of slabs to drain in tofree.
2371 * Returns the actual number of slabs released.
2373 static int drain_freelist(struct kmem_cache
*cache
,
2374 struct kmem_cache_node
*n
, int tofree
)
2376 struct list_head
*p
;
2381 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2383 spin_lock_irq(&n
->list_lock
);
2384 p
= n
->slabs_free
.prev
;
2385 if (p
== &n
->slabs_free
) {
2386 spin_unlock_irq(&n
->list_lock
);
2390 page
= list_entry(p
, struct page
, lru
);
2392 BUG_ON(page
->active
);
2394 list_del(&page
->lru
);
2396 * Safe to drop the lock. The slab is no longer linked
2399 n
->free_objects
-= cache
->num
;
2400 spin_unlock_irq(&n
->list_lock
);
2401 slab_destroy(cache
, page
);
2408 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2412 struct kmem_cache_node
*n
;
2414 drain_cpu_caches(cachep
);
2417 for_each_kmem_cache_node(cachep
, node
, n
) {
2418 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2420 ret
+= !list_empty(&n
->slabs_full
) ||
2421 !list_empty(&n
->slabs_partial
);
2423 return (ret
? 1 : 0);
2426 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2429 struct kmem_cache_node
*n
;
2430 int rc
= __kmem_cache_shrink(cachep
, false);
2435 free_percpu(cachep
->cpu_cache
);
2437 /* NUMA: free the node structures */
2438 for_each_kmem_cache_node(cachep
, i
, n
) {
2440 free_alien_cache(n
->alien
);
2442 cachep
->node
[i
] = NULL
;
2448 * Get the memory for a slab management obj.
2450 * For a slab cache when the slab descriptor is off-slab, the
2451 * slab descriptor can't come from the same cache which is being created,
2452 * Because if it is the case, that means we defer the creation of
2453 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2454 * And we eventually call down to __kmem_cache_create(), which
2455 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2456 * This is a "chicken-and-egg" problem.
2458 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2459 * which are all initialized during kmem_cache_init().
2461 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2462 struct page
*page
, int colour_off
,
2463 gfp_t local_flags
, int nodeid
)
2466 void *addr
= page_address(page
);
2468 if (OFF_SLAB(cachep
)) {
2469 /* Slab management obj is off-slab. */
2470 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2471 local_flags
, nodeid
);
2475 freelist
= addr
+ colour_off
;
2476 colour_off
+= cachep
->freelist_size
;
2479 page
->s_mem
= addr
+ colour_off
;
2483 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2485 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2488 static inline void set_free_obj(struct page
*page
,
2489 unsigned int idx
, freelist_idx_t val
)
2491 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2494 static void cache_init_objs(struct kmem_cache
*cachep
,
2499 for (i
= 0; i
< cachep
->num
; i
++) {
2500 void *objp
= index_to_obj(cachep
, page
, i
);
2502 /* need to poison the objs? */
2503 if (cachep
->flags
& SLAB_POISON
)
2504 poison_obj(cachep
, objp
, POISON_FREE
);
2505 if (cachep
->flags
& SLAB_STORE_USER
)
2506 *dbg_userword(cachep
, objp
) = NULL
;
2508 if (cachep
->flags
& SLAB_RED_ZONE
) {
2509 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2510 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2513 * Constructors are not allowed to allocate memory from the same
2514 * cache which they are a constructor for. Otherwise, deadlock.
2515 * They must also be threaded.
2517 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2518 cachep
->ctor(objp
+ obj_offset(cachep
));
2520 if (cachep
->flags
& SLAB_RED_ZONE
) {
2521 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2522 slab_error(cachep
, "constructor overwrote the"
2523 " end of an object");
2524 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2525 slab_error(cachep
, "constructor overwrote the"
2526 " start of an object");
2528 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2529 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2530 kernel_map_pages(virt_to_page(objp
),
2531 cachep
->size
/ PAGE_SIZE
, 0);
2536 set_obj_status(page
, i
, OBJECT_FREE
);
2537 set_free_obj(page
, i
, i
);
2541 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2543 if (CONFIG_ZONE_DMA_FLAG
) {
2544 if (flags
& GFP_DMA
)
2545 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2547 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2551 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
,
2556 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2559 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2565 static void slab_put_obj(struct kmem_cache
*cachep
, struct page
*page
,
2566 void *objp
, int nodeid
)
2568 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2572 /* Verify that the slab belongs to the intended node */
2573 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2575 /* Verify double free bug */
2576 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2577 if (get_free_obj(page
, i
) == objnr
) {
2578 printk(KERN_ERR
"slab: double free detected in cache "
2579 "'%s', objp %p\n", cachep
->name
, objp
);
2585 set_free_obj(page
, page
->active
, objnr
);
2589 * Map pages beginning at addr to the given cache and slab. This is required
2590 * for the slab allocator to be able to lookup the cache and slab of a
2591 * virtual address for kfree, ksize, and slab debugging.
2593 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2596 page
->slab_cache
= cache
;
2597 page
->freelist
= freelist
;
2601 * Grow (by 1) the number of slabs within a cache. This is called by
2602 * kmem_cache_alloc() when there are no active objs left in a cache.
2604 static int cache_grow(struct kmem_cache
*cachep
,
2605 gfp_t flags
, int nodeid
, struct page
*page
)
2610 struct kmem_cache_node
*n
;
2613 * Be lazy and only check for valid flags here, keeping it out of the
2614 * critical path in kmem_cache_alloc().
2616 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2617 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2620 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2622 /* Take the node list lock to change the colour_next on this node */
2624 n
= get_node(cachep
, nodeid
);
2625 spin_lock(&n
->list_lock
);
2627 /* Get colour for the slab, and cal the next value. */
2628 offset
= n
->colour_next
;
2630 if (n
->colour_next
>= cachep
->colour
)
2632 spin_unlock(&n
->list_lock
);
2634 offset
*= cachep
->colour_off
;
2636 if (local_flags
& __GFP_WAIT
)
2640 * The test for missing atomic flag is performed here, rather than
2641 * the more obvious place, simply to reduce the critical path length
2642 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2643 * will eventually be caught here (where it matters).
2645 kmem_flagcheck(cachep
, flags
);
2648 * Get mem for the objs. Attempt to allocate a physical page from
2652 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2656 /* Get slab management. */
2657 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2658 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2662 slab_map_pages(cachep
, page
, freelist
);
2664 cache_init_objs(cachep
, page
);
2666 if (local_flags
& __GFP_WAIT
)
2667 local_irq_disable();
2669 spin_lock(&n
->list_lock
);
2671 /* Make slab active. */
2672 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2673 STATS_INC_GROWN(cachep
);
2674 n
->free_objects
+= cachep
->num
;
2675 spin_unlock(&n
->list_lock
);
2678 kmem_freepages(cachep
, page
);
2680 if (local_flags
& __GFP_WAIT
)
2681 local_irq_disable();
2688 * Perform extra freeing checks:
2689 * - detect bad pointers.
2690 * - POISON/RED_ZONE checking
2692 static void kfree_debugcheck(const void *objp
)
2694 if (!virt_addr_valid(objp
)) {
2695 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2696 (unsigned long)objp
);
2701 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2703 unsigned long long redzone1
, redzone2
;
2705 redzone1
= *dbg_redzone1(cache
, obj
);
2706 redzone2
= *dbg_redzone2(cache
, obj
);
2711 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2714 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2715 slab_error(cache
, "double free detected");
2717 slab_error(cache
, "memory outside object was overwritten");
2719 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2720 obj
, redzone1
, redzone2
);
2723 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2724 unsigned long caller
)
2729 BUG_ON(virt_to_cache(objp
) != cachep
);
2731 objp
-= obj_offset(cachep
);
2732 kfree_debugcheck(objp
);
2733 page
= virt_to_head_page(objp
);
2735 if (cachep
->flags
& SLAB_RED_ZONE
) {
2736 verify_redzone_free(cachep
, objp
);
2737 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2738 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2740 if (cachep
->flags
& SLAB_STORE_USER
)
2741 *dbg_userword(cachep
, objp
) = (void *)caller
;
2743 objnr
= obj_to_index(cachep
, page
, objp
);
2745 BUG_ON(objnr
>= cachep
->num
);
2746 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2748 set_obj_status(page
, objnr
, OBJECT_FREE
);
2749 if (cachep
->flags
& SLAB_POISON
) {
2750 #ifdef CONFIG_DEBUG_PAGEALLOC
2751 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2752 store_stackinfo(cachep
, objp
, caller
);
2753 kernel_map_pages(virt_to_page(objp
),
2754 cachep
->size
/ PAGE_SIZE
, 0);
2756 poison_obj(cachep
, objp
, POISON_FREE
);
2759 poison_obj(cachep
, objp
, POISON_FREE
);
2766 #define kfree_debugcheck(x) do { } while(0)
2767 #define cache_free_debugcheck(x,objp,z) (objp)
2770 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2774 struct kmem_cache_node
*n
;
2775 struct array_cache
*ac
;
2779 node
= numa_mem_id();
2780 if (unlikely(force_refill
))
2783 ac
= cpu_cache_get(cachep
);
2784 batchcount
= ac
->batchcount
;
2785 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2787 * If there was little recent activity on this cache, then
2788 * perform only a partial refill. Otherwise we could generate
2791 batchcount
= BATCHREFILL_LIMIT
;
2793 n
= get_node(cachep
, node
);
2795 BUG_ON(ac
->avail
> 0 || !n
);
2796 spin_lock(&n
->list_lock
);
2798 /* See if we can refill from the shared array */
2799 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2800 n
->shared
->touched
= 1;
2804 while (batchcount
> 0) {
2805 struct list_head
*entry
;
2807 /* Get slab alloc is to come from. */
2808 entry
= n
->slabs_partial
.next
;
2809 if (entry
== &n
->slabs_partial
) {
2810 n
->free_touched
= 1;
2811 entry
= n
->slabs_free
.next
;
2812 if (entry
== &n
->slabs_free
)
2816 page
= list_entry(entry
, struct page
, lru
);
2817 check_spinlock_acquired(cachep
);
2820 * The slab was either on partial or free list so
2821 * there must be at least one object available for
2824 BUG_ON(page
->active
>= cachep
->num
);
2826 while (page
->active
< cachep
->num
&& batchcount
--) {
2827 STATS_INC_ALLOCED(cachep
);
2828 STATS_INC_ACTIVE(cachep
);
2829 STATS_SET_HIGH(cachep
);
2831 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, page
,
2835 /* move slabp to correct slabp list: */
2836 list_del(&page
->lru
);
2837 if (page
->active
== cachep
->num
)
2838 list_add(&page
->lru
, &n
->slabs_full
);
2840 list_add(&page
->lru
, &n
->slabs_partial
);
2844 n
->free_objects
-= ac
->avail
;
2846 spin_unlock(&n
->list_lock
);
2848 if (unlikely(!ac
->avail
)) {
2851 x
= cache_grow(cachep
, gfp_exact_node(flags
), node
, NULL
);
2853 /* cache_grow can reenable interrupts, then ac could change. */
2854 ac
= cpu_cache_get(cachep
);
2855 node
= numa_mem_id();
2857 /* no objects in sight? abort */
2858 if (!x
&& (ac
->avail
== 0 || force_refill
))
2861 if (!ac
->avail
) /* objects refilled by interrupt? */
2866 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2869 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2872 might_sleep_if(flags
& __GFP_WAIT
);
2874 kmem_flagcheck(cachep
, flags
);
2879 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2880 gfp_t flags
, void *objp
, unsigned long caller
)
2886 if (cachep
->flags
& SLAB_POISON
) {
2887 #ifdef CONFIG_DEBUG_PAGEALLOC
2888 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2889 kernel_map_pages(virt_to_page(objp
),
2890 cachep
->size
/ PAGE_SIZE
, 1);
2892 check_poison_obj(cachep
, objp
);
2894 check_poison_obj(cachep
, objp
);
2896 poison_obj(cachep
, objp
, POISON_INUSE
);
2898 if (cachep
->flags
& SLAB_STORE_USER
)
2899 *dbg_userword(cachep
, objp
) = (void *)caller
;
2901 if (cachep
->flags
& SLAB_RED_ZONE
) {
2902 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2903 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2904 slab_error(cachep
, "double free, or memory outside"
2905 " object was overwritten");
2907 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2908 objp
, *dbg_redzone1(cachep
, objp
),
2909 *dbg_redzone2(cachep
, objp
));
2911 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2912 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2915 page
= virt_to_head_page(objp
);
2916 set_obj_status(page
, obj_to_index(cachep
, page
, objp
), OBJECT_ACTIVE
);
2917 objp
+= obj_offset(cachep
);
2918 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2920 if (ARCH_SLAB_MINALIGN
&&
2921 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2922 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2923 objp
, (int)ARCH_SLAB_MINALIGN
);
2928 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2931 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
2933 if (unlikely(cachep
== kmem_cache
))
2936 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
2939 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2942 struct array_cache
*ac
;
2943 bool force_refill
= false;
2947 ac
= cpu_cache_get(cachep
);
2948 if (likely(ac
->avail
)) {
2950 objp
= ac_get_obj(cachep
, ac
, flags
, false);
2953 * Allow for the possibility all avail objects are not allowed
2954 * by the current flags
2957 STATS_INC_ALLOCHIT(cachep
);
2960 force_refill
= true;
2963 STATS_INC_ALLOCMISS(cachep
);
2964 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
2966 * the 'ac' may be updated by cache_alloc_refill(),
2967 * and kmemleak_erase() requires its correct value.
2969 ac
= cpu_cache_get(cachep
);
2973 * To avoid a false negative, if an object that is in one of the
2974 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2975 * treat the array pointers as a reference to the object.
2978 kmemleak_erase(&ac
->entry
[ac
->avail
]);
2984 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2986 * If we are in_interrupt, then process context, including cpusets and
2987 * mempolicy, may not apply and should not be used for allocation policy.
2989 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2991 int nid_alloc
, nid_here
;
2993 if (in_interrupt() || (flags
& __GFP_THISNODE
))
2995 nid_alloc
= nid_here
= numa_mem_id();
2996 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2997 nid_alloc
= cpuset_slab_spread_node();
2998 else if (current
->mempolicy
)
2999 nid_alloc
= mempolicy_slab_node();
3000 if (nid_alloc
!= nid_here
)
3001 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3006 * Fallback function if there was no memory available and no objects on a
3007 * certain node and fall back is permitted. First we scan all the
3008 * available node for available objects. If that fails then we
3009 * perform an allocation without specifying a node. This allows the page
3010 * allocator to do its reclaim / fallback magic. We then insert the
3011 * slab into the proper nodelist and then allocate from it.
3013 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3015 struct zonelist
*zonelist
;
3019 enum zone_type high_zoneidx
= gfp_zone(flags
);
3022 unsigned int cpuset_mems_cookie
;
3024 if (flags
& __GFP_THISNODE
)
3027 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3030 cpuset_mems_cookie
= read_mems_allowed_begin();
3031 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3035 * Look through allowed nodes for objects available
3036 * from existing per node queues.
3038 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3039 nid
= zone_to_nid(zone
);
3041 if (cpuset_zone_allowed(zone
, flags
) &&
3042 get_node(cache
, nid
) &&
3043 get_node(cache
, nid
)->free_objects
) {
3044 obj
= ____cache_alloc_node(cache
,
3045 gfp_exact_node(flags
), nid
);
3053 * This allocation will be performed within the constraints
3054 * of the current cpuset / memory policy requirements.
3055 * We may trigger various forms of reclaim on the allowed
3056 * set and go into memory reserves if necessary.
3060 if (local_flags
& __GFP_WAIT
)
3062 kmem_flagcheck(cache
, flags
);
3063 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3064 if (local_flags
& __GFP_WAIT
)
3065 local_irq_disable();
3068 * Insert into the appropriate per node queues
3070 nid
= page_to_nid(page
);
3071 if (cache_grow(cache
, flags
, nid
, page
)) {
3072 obj
= ____cache_alloc_node(cache
,
3073 gfp_exact_node(flags
), nid
);
3076 * Another processor may allocate the
3077 * objects in the slab since we are
3078 * not holding any locks.
3082 /* cache_grow already freed obj */
3088 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3094 * A interface to enable slab creation on nodeid
3096 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3099 struct list_head
*entry
;
3101 struct kmem_cache_node
*n
;
3105 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3106 n
= get_node(cachep
, nodeid
);
3111 spin_lock(&n
->list_lock
);
3112 entry
= n
->slabs_partial
.next
;
3113 if (entry
== &n
->slabs_partial
) {
3114 n
->free_touched
= 1;
3115 entry
= n
->slabs_free
.next
;
3116 if (entry
== &n
->slabs_free
)
3120 page
= list_entry(entry
, struct page
, lru
);
3121 check_spinlock_acquired_node(cachep
, nodeid
);
3123 STATS_INC_NODEALLOCS(cachep
);
3124 STATS_INC_ACTIVE(cachep
);
3125 STATS_SET_HIGH(cachep
);
3127 BUG_ON(page
->active
== cachep
->num
);
3129 obj
= slab_get_obj(cachep
, page
, nodeid
);
3131 /* move slabp to correct slabp list: */
3132 list_del(&page
->lru
);
3134 if (page
->active
== cachep
->num
)
3135 list_add(&page
->lru
, &n
->slabs_full
);
3137 list_add(&page
->lru
, &n
->slabs_partial
);
3139 spin_unlock(&n
->list_lock
);
3143 spin_unlock(&n
->list_lock
);
3144 x
= cache_grow(cachep
, gfp_exact_node(flags
), nodeid
, NULL
);
3148 return fallback_alloc(cachep
, flags
);
3154 static __always_inline
void *
3155 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3156 unsigned long caller
)
3158 unsigned long save_flags
;
3160 int slab_node
= numa_mem_id();
3162 flags
&= gfp_allowed_mask
;
3164 lockdep_trace_alloc(flags
);
3166 if (slab_should_failslab(cachep
, flags
))
3169 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3171 cache_alloc_debugcheck_before(cachep
, flags
);
3172 local_irq_save(save_flags
);
3174 if (nodeid
== NUMA_NO_NODE
)
3177 if (unlikely(!get_node(cachep
, nodeid
))) {
3178 /* Node not bootstrapped yet */
3179 ptr
= fallback_alloc(cachep
, flags
);
3183 if (nodeid
== slab_node
) {
3185 * Use the locally cached objects if possible.
3186 * However ____cache_alloc does not allow fallback
3187 * to other nodes. It may fail while we still have
3188 * objects on other nodes available.
3190 ptr
= ____cache_alloc(cachep
, flags
);
3194 /* ___cache_alloc_node can fall back to other nodes */
3195 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3197 local_irq_restore(save_flags
);
3198 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3199 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3203 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3204 if (unlikely(flags
& __GFP_ZERO
))
3205 memset(ptr
, 0, cachep
->object_size
);
3208 memcg_kmem_put_cache(cachep
);
3212 static __always_inline
void *
3213 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3217 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3218 objp
= alternate_node_alloc(cache
, flags
);
3222 objp
= ____cache_alloc(cache
, flags
);
3225 * We may just have run out of memory on the local node.
3226 * ____cache_alloc_node() knows how to locate memory on other nodes
3229 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3236 static __always_inline
void *
3237 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3239 return ____cache_alloc(cachep
, flags
);
3242 #endif /* CONFIG_NUMA */
3244 static __always_inline
void *
3245 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3247 unsigned long save_flags
;
3250 flags
&= gfp_allowed_mask
;
3252 lockdep_trace_alloc(flags
);
3254 if (slab_should_failslab(cachep
, flags
))
3257 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3259 cache_alloc_debugcheck_before(cachep
, flags
);
3260 local_irq_save(save_flags
);
3261 objp
= __do_cache_alloc(cachep
, flags
);
3262 local_irq_restore(save_flags
);
3263 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3264 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3269 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3270 if (unlikely(flags
& __GFP_ZERO
))
3271 memset(objp
, 0, cachep
->object_size
);
3274 memcg_kmem_put_cache(cachep
);
3279 * Caller needs to acquire correct kmem_cache_node's list_lock
3280 * @list: List of detached free slabs should be freed by caller
3282 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3283 int nr_objects
, int node
, struct list_head
*list
)
3286 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3288 for (i
= 0; i
< nr_objects
; i
++) {
3292 clear_obj_pfmemalloc(&objpp
[i
]);
3295 page
= virt_to_head_page(objp
);
3296 list_del(&page
->lru
);
3297 check_spinlock_acquired_node(cachep
, node
);
3298 slab_put_obj(cachep
, page
, objp
, node
);
3299 STATS_DEC_ACTIVE(cachep
);
3302 /* fixup slab chains */
3303 if (page
->active
== 0) {
3304 if (n
->free_objects
> n
->free_limit
) {
3305 n
->free_objects
-= cachep
->num
;
3306 list_add_tail(&page
->lru
, list
);
3308 list_add(&page
->lru
, &n
->slabs_free
);
3311 /* Unconditionally move a slab to the end of the
3312 * partial list on free - maximum time for the
3313 * other objects to be freed, too.
3315 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3320 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3323 struct kmem_cache_node
*n
;
3324 int node
= numa_mem_id();
3327 batchcount
= ac
->batchcount
;
3329 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3332 n
= get_node(cachep
, node
);
3333 spin_lock(&n
->list_lock
);
3335 struct array_cache
*shared_array
= n
->shared
;
3336 int max
= shared_array
->limit
- shared_array
->avail
;
3338 if (batchcount
> max
)
3340 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3341 ac
->entry
, sizeof(void *) * batchcount
);
3342 shared_array
->avail
+= batchcount
;
3347 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3352 struct list_head
*p
;
3354 p
= n
->slabs_free
.next
;
3355 while (p
!= &(n
->slabs_free
)) {
3358 page
= list_entry(p
, struct page
, lru
);
3359 BUG_ON(page
->active
);
3364 STATS_SET_FREEABLE(cachep
, i
);
3367 spin_unlock(&n
->list_lock
);
3368 slabs_destroy(cachep
, &list
);
3369 ac
->avail
-= batchcount
;
3370 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3374 * Release an obj back to its cache. If the obj has a constructed state, it must
3375 * be in this state _before_ it is released. Called with disabled ints.
3377 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3378 unsigned long caller
)
3380 struct array_cache
*ac
= cpu_cache_get(cachep
);
3383 kmemleak_free_recursive(objp
, cachep
->flags
);
3384 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3386 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3389 * Skip calling cache_free_alien() when the platform is not numa.
3390 * This will avoid cache misses that happen while accessing slabp (which
3391 * is per page memory reference) to get nodeid. Instead use a global
3392 * variable to skip the call, which is mostly likely to be present in
3395 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3398 if (ac
->avail
< ac
->limit
) {
3399 STATS_INC_FREEHIT(cachep
);
3401 STATS_INC_FREEMISS(cachep
);
3402 cache_flusharray(cachep
, ac
);
3405 ac_put_obj(cachep
, ac
, objp
);
3409 * kmem_cache_alloc - Allocate an object
3410 * @cachep: The cache to allocate from.
3411 * @flags: See kmalloc().
3413 * Allocate an object from this cache. The flags are only relevant
3414 * if the cache has no available objects.
3416 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3418 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3420 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3421 cachep
->object_size
, cachep
->size
, flags
);
3425 EXPORT_SYMBOL(kmem_cache_alloc
);
3427 void kmem_cache_free_bulk(struct kmem_cache
*s
, size_t size
, void **p
)
3429 __kmem_cache_free_bulk(s
, size
, p
);
3431 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3433 bool kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3436 return __kmem_cache_alloc_bulk(s
, flags
, size
, p
);
3438 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3440 #ifdef CONFIG_TRACING
3442 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3446 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3448 trace_kmalloc(_RET_IP_
, ret
,
3449 size
, cachep
->size
, flags
);
3452 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3457 * kmem_cache_alloc_node - Allocate an object on the specified node
3458 * @cachep: The cache to allocate from.
3459 * @flags: See kmalloc().
3460 * @nodeid: node number of the target node.
3462 * Identical to kmem_cache_alloc but it will allocate memory on the given
3463 * node, which can improve the performance for cpu bound structures.
3465 * Fallback to other node is possible if __GFP_THISNODE is not set.
3467 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3469 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3471 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3472 cachep
->object_size
, cachep
->size
,
3477 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3479 #ifdef CONFIG_TRACING
3480 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3487 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3489 trace_kmalloc_node(_RET_IP_
, ret
,
3494 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3497 static __always_inline
void *
3498 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3500 struct kmem_cache
*cachep
;
3502 cachep
= kmalloc_slab(size
, flags
);
3503 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3505 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3508 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3510 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3512 EXPORT_SYMBOL(__kmalloc_node
);
3514 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3515 int node
, unsigned long caller
)
3517 return __do_kmalloc_node(size
, flags
, node
, caller
);
3519 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3520 #endif /* CONFIG_NUMA */
3523 * __do_kmalloc - allocate memory
3524 * @size: how many bytes of memory are required.
3525 * @flags: the type of memory to allocate (see kmalloc).
3526 * @caller: function caller for debug tracking of the caller
3528 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3529 unsigned long caller
)
3531 struct kmem_cache
*cachep
;
3534 cachep
= kmalloc_slab(size
, flags
);
3535 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3537 ret
= slab_alloc(cachep
, flags
, caller
);
3539 trace_kmalloc(caller
, ret
,
3540 size
, cachep
->size
, flags
);
3545 void *__kmalloc(size_t size
, gfp_t flags
)
3547 return __do_kmalloc(size
, flags
, _RET_IP_
);
3549 EXPORT_SYMBOL(__kmalloc
);
3551 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3553 return __do_kmalloc(size
, flags
, caller
);
3555 EXPORT_SYMBOL(__kmalloc_track_caller
);
3558 * kmem_cache_free - Deallocate an object
3559 * @cachep: The cache the allocation was from.
3560 * @objp: The previously allocated object.
3562 * Free an object which was previously allocated from this
3565 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3567 unsigned long flags
;
3568 cachep
= cache_from_obj(cachep
, objp
);
3572 local_irq_save(flags
);
3573 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3574 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3575 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3576 __cache_free(cachep
, objp
, _RET_IP_
);
3577 local_irq_restore(flags
);
3579 trace_kmem_cache_free(_RET_IP_
, objp
);
3581 EXPORT_SYMBOL(kmem_cache_free
);
3584 * kfree - free previously allocated memory
3585 * @objp: pointer returned by kmalloc.
3587 * If @objp is NULL, no operation is performed.
3589 * Don't free memory not originally allocated by kmalloc()
3590 * or you will run into trouble.
3592 void kfree(const void *objp
)
3594 struct kmem_cache
*c
;
3595 unsigned long flags
;
3597 trace_kfree(_RET_IP_
, objp
);
3599 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3601 local_irq_save(flags
);
3602 kfree_debugcheck(objp
);
3603 c
= virt_to_cache(objp
);
3604 debug_check_no_locks_freed(objp
, c
->object_size
);
3606 debug_check_no_obj_freed(objp
, c
->object_size
);
3607 __cache_free(c
, (void *)objp
, _RET_IP_
);
3608 local_irq_restore(flags
);
3610 EXPORT_SYMBOL(kfree
);
3613 * This initializes kmem_cache_node or resizes various caches for all nodes.
3615 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3618 struct kmem_cache_node
*n
;
3619 struct array_cache
*new_shared
;
3620 struct alien_cache
**new_alien
= NULL
;
3622 for_each_online_node(node
) {
3624 if (use_alien_caches
) {
3625 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3631 if (cachep
->shared
) {
3632 new_shared
= alloc_arraycache(node
,
3633 cachep
->shared
*cachep
->batchcount
,
3636 free_alien_cache(new_alien
);
3641 n
= get_node(cachep
, node
);
3643 struct array_cache
*shared
= n
->shared
;
3646 spin_lock_irq(&n
->list_lock
);
3649 free_block(cachep
, shared
->entry
,
3650 shared
->avail
, node
, &list
);
3652 n
->shared
= new_shared
;
3654 n
->alien
= new_alien
;
3657 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3658 cachep
->batchcount
+ cachep
->num
;
3659 spin_unlock_irq(&n
->list_lock
);
3660 slabs_destroy(cachep
, &list
);
3662 free_alien_cache(new_alien
);
3665 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3667 free_alien_cache(new_alien
);
3672 kmem_cache_node_init(n
);
3673 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3674 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3675 n
->shared
= new_shared
;
3676 n
->alien
= new_alien
;
3677 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3678 cachep
->batchcount
+ cachep
->num
;
3679 cachep
->node
[node
] = n
;
3684 if (!cachep
->list
.next
) {
3685 /* Cache is not active yet. Roll back what we did */
3688 n
= get_node(cachep
, node
);
3691 free_alien_cache(n
->alien
);
3693 cachep
->node
[node
] = NULL
;
3701 /* Always called with the slab_mutex held */
3702 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3703 int batchcount
, int shared
, gfp_t gfp
)
3705 struct array_cache __percpu
*cpu_cache
, *prev
;
3708 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3712 prev
= cachep
->cpu_cache
;
3713 cachep
->cpu_cache
= cpu_cache
;
3714 kick_all_cpus_sync();
3717 cachep
->batchcount
= batchcount
;
3718 cachep
->limit
= limit
;
3719 cachep
->shared
= shared
;
3724 for_each_online_cpu(cpu
) {
3727 struct kmem_cache_node
*n
;
3728 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3730 node
= cpu_to_mem(cpu
);
3731 n
= get_node(cachep
, node
);
3732 spin_lock_irq(&n
->list_lock
);
3733 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3734 spin_unlock_irq(&n
->list_lock
);
3735 slabs_destroy(cachep
, &list
);
3740 return alloc_kmem_cache_node(cachep
, gfp
);
3743 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3744 int batchcount
, int shared
, gfp_t gfp
)
3747 struct kmem_cache
*c
;
3749 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3751 if (slab_state
< FULL
)
3754 if ((ret
< 0) || !is_root_cache(cachep
))
3757 lockdep_assert_held(&slab_mutex
);
3758 for_each_memcg_cache(c
, cachep
) {
3759 /* return value determined by the root cache only */
3760 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3766 /* Called with slab_mutex held always */
3767 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3774 if (!is_root_cache(cachep
)) {
3775 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3776 limit
= root
->limit
;
3777 shared
= root
->shared
;
3778 batchcount
= root
->batchcount
;
3781 if (limit
&& shared
&& batchcount
)
3784 * The head array serves three purposes:
3785 * - create a LIFO ordering, i.e. return objects that are cache-warm
3786 * - reduce the number of spinlock operations.
3787 * - reduce the number of linked list operations on the slab and
3788 * bufctl chains: array operations are cheaper.
3789 * The numbers are guessed, we should auto-tune as described by
3792 if (cachep
->size
> 131072)
3794 else if (cachep
->size
> PAGE_SIZE
)
3796 else if (cachep
->size
> 1024)
3798 else if (cachep
->size
> 256)
3804 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3805 * allocation behaviour: Most allocs on one cpu, most free operations
3806 * on another cpu. For these cases, an efficient object passing between
3807 * cpus is necessary. This is provided by a shared array. The array
3808 * replaces Bonwick's magazine layer.
3809 * On uniprocessor, it's functionally equivalent (but less efficient)
3810 * to a larger limit. Thus disabled by default.
3813 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3818 * With debugging enabled, large batchcount lead to excessively long
3819 * periods with disabled local interrupts. Limit the batchcount
3824 batchcount
= (limit
+ 1) / 2;
3826 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3828 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3829 cachep
->name
, -err
);
3834 * Drain an array if it contains any elements taking the node lock only if
3835 * necessary. Note that the node listlock also protects the array_cache
3836 * if drain_array() is used on the shared array.
3838 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3839 struct array_cache
*ac
, int force
, int node
)
3844 if (!ac
|| !ac
->avail
)
3846 if (ac
->touched
&& !force
) {
3849 spin_lock_irq(&n
->list_lock
);
3851 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3852 if (tofree
> ac
->avail
)
3853 tofree
= (ac
->avail
+ 1) / 2;
3854 free_block(cachep
, ac
->entry
, tofree
, node
, &list
);
3855 ac
->avail
-= tofree
;
3856 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3857 sizeof(void *) * ac
->avail
);
3859 spin_unlock_irq(&n
->list_lock
);
3860 slabs_destroy(cachep
, &list
);
3865 * cache_reap - Reclaim memory from caches.
3866 * @w: work descriptor
3868 * Called from workqueue/eventd every few seconds.
3870 * - clear the per-cpu caches for this CPU.
3871 * - return freeable pages to the main free memory pool.
3873 * If we cannot acquire the cache chain mutex then just give up - we'll try
3874 * again on the next iteration.
3876 static void cache_reap(struct work_struct
*w
)
3878 struct kmem_cache
*searchp
;
3879 struct kmem_cache_node
*n
;
3880 int node
= numa_mem_id();
3881 struct delayed_work
*work
= to_delayed_work(w
);
3883 if (!mutex_trylock(&slab_mutex
))
3884 /* Give up. Setup the next iteration. */
3887 list_for_each_entry(searchp
, &slab_caches
, list
) {
3891 * We only take the node lock if absolutely necessary and we
3892 * have established with reasonable certainty that
3893 * we can do some work if the lock was obtained.
3895 n
= get_node(searchp
, node
);
3897 reap_alien(searchp
, n
);
3899 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
3902 * These are racy checks but it does not matter
3903 * if we skip one check or scan twice.
3905 if (time_after(n
->next_reap
, jiffies
))
3908 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3910 drain_array(searchp
, n
, n
->shared
, 0, node
);
3912 if (n
->free_touched
)
3913 n
->free_touched
= 0;
3917 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3918 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3919 STATS_ADD_REAPED(searchp
, freed
);
3925 mutex_unlock(&slab_mutex
);
3928 /* Set up the next iteration */
3929 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3932 #ifdef CONFIG_SLABINFO
3933 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3936 unsigned long active_objs
;
3937 unsigned long num_objs
;
3938 unsigned long active_slabs
= 0;
3939 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3943 struct kmem_cache_node
*n
;
3947 for_each_kmem_cache_node(cachep
, node
, n
) {
3950 spin_lock_irq(&n
->list_lock
);
3952 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
3953 if (page
->active
!= cachep
->num
&& !error
)
3954 error
= "slabs_full accounting error";
3955 active_objs
+= cachep
->num
;
3958 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
3959 if (page
->active
== cachep
->num
&& !error
)
3960 error
= "slabs_partial accounting error";
3961 if (!page
->active
&& !error
)
3962 error
= "slabs_partial accounting error";
3963 active_objs
+= page
->active
;
3966 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3967 if (page
->active
&& !error
)
3968 error
= "slabs_free accounting error";
3971 free_objects
+= n
->free_objects
;
3973 shared_avail
+= n
->shared
->avail
;
3975 spin_unlock_irq(&n
->list_lock
);
3977 num_slabs
+= active_slabs
;
3978 num_objs
= num_slabs
* cachep
->num
;
3979 if (num_objs
- active_objs
!= free_objects
&& !error
)
3980 error
= "free_objects accounting error";
3982 name
= cachep
->name
;
3984 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
3986 sinfo
->active_objs
= active_objs
;
3987 sinfo
->num_objs
= num_objs
;
3988 sinfo
->active_slabs
= active_slabs
;
3989 sinfo
->num_slabs
= num_slabs
;
3990 sinfo
->shared_avail
= shared_avail
;
3991 sinfo
->limit
= cachep
->limit
;
3992 sinfo
->batchcount
= cachep
->batchcount
;
3993 sinfo
->shared
= cachep
->shared
;
3994 sinfo
->objects_per_slab
= cachep
->num
;
3995 sinfo
->cache_order
= cachep
->gfporder
;
3998 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4002 unsigned long high
= cachep
->high_mark
;
4003 unsigned long allocs
= cachep
->num_allocations
;
4004 unsigned long grown
= cachep
->grown
;
4005 unsigned long reaped
= cachep
->reaped
;
4006 unsigned long errors
= cachep
->errors
;
4007 unsigned long max_freeable
= cachep
->max_freeable
;
4008 unsigned long node_allocs
= cachep
->node_allocs
;
4009 unsigned long node_frees
= cachep
->node_frees
;
4010 unsigned long overflows
= cachep
->node_overflow
;
4012 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4013 "%4lu %4lu %4lu %4lu %4lu",
4014 allocs
, high
, grown
,
4015 reaped
, errors
, max_freeable
, node_allocs
,
4016 node_frees
, overflows
);
4020 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4021 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4022 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4023 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4025 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4026 allochit
, allocmiss
, freehit
, freemiss
);
4031 #define MAX_SLABINFO_WRITE 128
4033 * slabinfo_write - Tuning for the slab allocator
4035 * @buffer: user buffer
4036 * @count: data length
4039 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4040 size_t count
, loff_t
*ppos
)
4042 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4043 int limit
, batchcount
, shared
, res
;
4044 struct kmem_cache
*cachep
;
4046 if (count
> MAX_SLABINFO_WRITE
)
4048 if (copy_from_user(&kbuf
, buffer
, count
))
4050 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4052 tmp
= strchr(kbuf
, ' ');
4057 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4060 /* Find the cache in the chain of caches. */
4061 mutex_lock(&slab_mutex
);
4063 list_for_each_entry(cachep
, &slab_caches
, list
) {
4064 if (!strcmp(cachep
->name
, kbuf
)) {
4065 if (limit
< 1 || batchcount
< 1 ||
4066 batchcount
> limit
|| shared
< 0) {
4069 res
= do_tune_cpucache(cachep
, limit
,
4076 mutex_unlock(&slab_mutex
);
4082 #ifdef CONFIG_DEBUG_SLAB_LEAK
4084 static inline int add_caller(unsigned long *n
, unsigned long v
)
4094 unsigned long *q
= p
+ 2 * i
;
4108 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4114 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4122 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4123 if (get_obj_status(page
, i
) != OBJECT_ACTIVE
)
4126 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4131 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4133 #ifdef CONFIG_KALLSYMS
4134 unsigned long offset
, size
;
4135 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4137 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4138 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4140 seq_printf(m
, " [%s]", modname
);
4144 seq_printf(m
, "%p", (void *)address
);
4147 static int leaks_show(struct seq_file
*m
, void *p
)
4149 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4151 struct kmem_cache_node
*n
;
4153 unsigned long *x
= m
->private;
4157 if (!(cachep
->flags
& SLAB_STORE_USER
))
4159 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4162 /* OK, we can do it */
4166 for_each_kmem_cache_node(cachep
, node
, n
) {
4169 spin_lock_irq(&n
->list_lock
);
4171 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4172 handle_slab(x
, cachep
, page
);
4173 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4174 handle_slab(x
, cachep
, page
);
4175 spin_unlock_irq(&n
->list_lock
);
4177 name
= cachep
->name
;
4179 /* Increase the buffer size */
4180 mutex_unlock(&slab_mutex
);
4181 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4183 /* Too bad, we are really out */
4185 mutex_lock(&slab_mutex
);
4188 *(unsigned long *)m
->private = x
[0] * 2;
4190 mutex_lock(&slab_mutex
);
4191 /* Now make sure this entry will be retried */
4195 for (i
= 0; i
< x
[1]; i
++) {
4196 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4197 show_symbol(m
, x
[2*i
+2]);
4204 static const struct seq_operations slabstats_op
= {
4205 .start
= slab_start
,
4211 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4215 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4219 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4224 static const struct file_operations proc_slabstats_operations
= {
4225 .open
= slabstats_open
,
4227 .llseek
= seq_lseek
,
4228 .release
= seq_release_private
,
4232 static int __init
slab_proc_init(void)
4234 #ifdef CONFIG_DEBUG_SLAB_LEAK
4235 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4239 module_init(slab_proc_init
);
4243 * ksize - get the actual amount of memory allocated for a given object
4244 * @objp: Pointer to the object
4246 * kmalloc may internally round up allocations and return more memory
4247 * than requested. ksize() can be used to determine the actual amount of
4248 * memory allocated. The caller may use this additional memory, even though
4249 * a smaller amount of memory was initially specified with the kmalloc call.
4250 * The caller must guarantee that objp points to a valid object previously
4251 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4252 * must not be freed during the duration of the call.
4254 size_t ksize(const void *objp
)
4257 if (unlikely(objp
== ZERO_SIZE_PTR
))
4260 return virt_to_cache(objp
)->object_size
;
4262 EXPORT_SYMBOL(ksize
);