3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t
;
166 typedef unsigned short freelist_idx_t
;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
172 * true if a page was allocated from pfmemalloc reserves for network-based
175 static bool pfmemalloc_active __read_mostly
;
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
185 * The limit is stored in the per-cpu structure to reduce the data cache
192 unsigned int batchcount
;
193 unsigned int touched
;
195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
207 struct array_cache ac
;
210 #define SLAB_OBJ_PFMEMALLOC 1
211 static inline bool is_obj_pfmemalloc(void *objp
)
213 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
216 static inline void set_obj_pfmemalloc(void **objp
)
218 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
222 static inline void clear_obj_pfmemalloc(void **objp
)
224 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
228 * bootstrap: The caches do not work without cpuarrays anymore, but the
229 * cpuarrays are allocated from the generic caches...
231 #define BOOT_CPUCACHE_ENTRIES 1
232 struct arraycache_init
{
233 struct array_cache cache
;
234 void *entries
[BOOT_CPUCACHE_ENTRIES
];
238 * Need this for bootstrapping a per node allocator.
240 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
241 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
242 #define CACHE_CACHE 0
243 #define SIZE_AC MAX_NUMNODES
244 #define SIZE_NODE (2 * MAX_NUMNODES)
246 static int drain_freelist(struct kmem_cache
*cache
,
247 struct kmem_cache_node
*n
, int tofree
);
248 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
249 int node
, struct list_head
*list
);
250 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
251 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
252 static void cache_reap(struct work_struct
*unused
);
254 static int slab_early_init
= 1;
256 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
257 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
259 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
261 INIT_LIST_HEAD(&parent
->slabs_full
);
262 INIT_LIST_HEAD(&parent
->slabs_partial
);
263 INIT_LIST_HEAD(&parent
->slabs_free
);
264 parent
->shared
= NULL
;
265 parent
->alien
= NULL
;
266 parent
->colour_next
= 0;
267 spin_lock_init(&parent
->list_lock
);
268 parent
->free_objects
= 0;
269 parent
->free_touched
= 0;
272 #define MAKE_LIST(cachep, listp, slab, nodeid) \
274 INIT_LIST_HEAD(listp); \
275 list_splice(&get_node(cachep, nodeid)->slab, listp); \
278 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
280 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
281 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
282 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
285 #define CFLGS_OFF_SLAB (0x80000000UL)
286 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
288 #define BATCHREFILL_LIMIT 16
290 * Optimization question: fewer reaps means less probability for unnessary
291 * cpucache drain/refill cycles.
293 * OTOH the cpuarrays can contain lots of objects,
294 * which could lock up otherwise freeable slabs.
296 #define REAPTIMEOUT_AC (2*HZ)
297 #define REAPTIMEOUT_NODE (4*HZ)
300 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
301 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
302 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
303 #define STATS_INC_GROWN(x) ((x)->grown++)
304 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
305 #define STATS_SET_HIGH(x) \
307 if ((x)->num_active > (x)->high_mark) \
308 (x)->high_mark = (x)->num_active; \
310 #define STATS_INC_ERR(x) ((x)->errors++)
311 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
312 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
313 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
314 #define STATS_SET_FREEABLE(x, i) \
316 if ((x)->max_freeable < i) \
317 (x)->max_freeable = i; \
319 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
320 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
321 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
322 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
324 #define STATS_INC_ACTIVE(x) do { } while (0)
325 #define STATS_DEC_ACTIVE(x) do { } while (0)
326 #define STATS_INC_ALLOCED(x) do { } while (0)
327 #define STATS_INC_GROWN(x) do { } while (0)
328 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
329 #define STATS_SET_HIGH(x) do { } while (0)
330 #define STATS_INC_ERR(x) do { } while (0)
331 #define STATS_INC_NODEALLOCS(x) do { } while (0)
332 #define STATS_INC_NODEFREES(x) do { } while (0)
333 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
334 #define STATS_SET_FREEABLE(x, i) do { } while (0)
335 #define STATS_INC_ALLOCHIT(x) do { } while (0)
336 #define STATS_INC_ALLOCMISS(x) do { } while (0)
337 #define STATS_INC_FREEHIT(x) do { } while (0)
338 #define STATS_INC_FREEMISS(x) do { } while (0)
344 * memory layout of objects:
346 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
347 * the end of an object is aligned with the end of the real
348 * allocation. Catches writes behind the end of the allocation.
349 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
351 * cachep->obj_offset: The real object.
352 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
353 * cachep->size - 1* BYTES_PER_WORD: last caller address
354 * [BYTES_PER_WORD long]
356 static int obj_offset(struct kmem_cache
*cachep
)
358 return cachep
->obj_offset
;
361 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
363 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
364 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
365 sizeof(unsigned long long));
368 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
370 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
371 if (cachep
->flags
& SLAB_STORE_USER
)
372 return (unsigned long long *)(objp
+ cachep
->size
-
373 sizeof(unsigned long long) -
375 return (unsigned long long *) (objp
+ cachep
->size
-
376 sizeof(unsigned long long));
379 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
381 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
382 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
387 #define obj_offset(x) 0
388 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
390 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
394 #define OBJECT_FREE (0)
395 #define OBJECT_ACTIVE (1)
397 #ifdef CONFIG_DEBUG_SLAB_LEAK
399 static void set_obj_status(struct page
*page
, int idx
, int val
)
403 struct kmem_cache
*cachep
= page
->slab_cache
;
405 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
406 status
= (char *)page
->freelist
+ freelist_size
;
410 static inline unsigned int get_obj_status(struct page
*page
, int idx
)
414 struct kmem_cache
*cachep
= page
->slab_cache
;
416 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
417 status
= (char *)page
->freelist
+ freelist_size
;
423 static inline void set_obj_status(struct page
*page
, int idx
, int val
) {}
428 * Do not go above this order unless 0 objects fit into the slab or
429 * overridden on the command line.
431 #define SLAB_MAX_ORDER_HI 1
432 #define SLAB_MAX_ORDER_LO 0
433 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
434 static bool slab_max_order_set __initdata
;
436 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
438 struct page
*page
= virt_to_head_page(obj
);
439 return page
->slab_cache
;
442 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
445 return page
->s_mem
+ cache
->size
* idx
;
449 * We want to avoid an expensive divide : (offset / cache->size)
450 * Using the fact that size is a constant for a particular cache,
451 * we can replace (offset / cache->size) by
452 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
454 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
455 const struct page
*page
, void *obj
)
457 u32 offset
= (obj
- page
->s_mem
);
458 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
461 static struct arraycache_init initarray_generic
=
462 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
464 /* internal cache of cache description objs */
465 static struct kmem_cache kmem_cache_boot
= {
467 .limit
= BOOT_CPUCACHE_ENTRIES
,
469 .size
= sizeof(struct kmem_cache
),
470 .name
= "kmem_cache",
473 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
475 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
477 return cachep
->array
[smp_processor_id()];
480 static size_t calculate_freelist_size(int nr_objs
, size_t align
)
482 size_t freelist_size
;
484 freelist_size
= nr_objs
* sizeof(freelist_idx_t
);
485 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
486 freelist_size
+= nr_objs
* sizeof(char);
489 freelist_size
= ALIGN(freelist_size
, align
);
491 return freelist_size
;
494 static int calculate_nr_objs(size_t slab_size
, size_t buffer_size
,
495 size_t idx_size
, size_t align
)
498 size_t remained_size
;
499 size_t freelist_size
;
502 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
503 extra_space
= sizeof(char);
505 * Ignore padding for the initial guess. The padding
506 * is at most @align-1 bytes, and @buffer_size is at
507 * least @align. In the worst case, this result will
508 * be one greater than the number of objects that fit
509 * into the memory allocation when taking the padding
512 nr_objs
= slab_size
/ (buffer_size
+ idx_size
+ extra_space
);
515 * This calculated number will be either the right
516 * amount, or one greater than what we want.
518 remained_size
= slab_size
- nr_objs
* buffer_size
;
519 freelist_size
= calculate_freelist_size(nr_objs
, align
);
520 if (remained_size
< freelist_size
)
527 * Calculate the number of objects and left-over bytes for a given buffer size.
529 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
530 size_t align
, int flags
, size_t *left_over
,
535 size_t slab_size
= PAGE_SIZE
<< gfporder
;
538 * The slab management structure can be either off the slab or
539 * on it. For the latter case, the memory allocated for a
542 * - One unsigned int for each object
543 * - Padding to respect alignment of @align
544 * - @buffer_size bytes for each object
546 * If the slab management structure is off the slab, then the
547 * alignment will already be calculated into the size. Because
548 * the slabs are all pages aligned, the objects will be at the
549 * correct alignment when allocated.
551 if (flags
& CFLGS_OFF_SLAB
) {
553 nr_objs
= slab_size
/ buffer_size
;
556 nr_objs
= calculate_nr_objs(slab_size
, buffer_size
,
557 sizeof(freelist_idx_t
), align
);
558 mgmt_size
= calculate_freelist_size(nr_objs
, align
);
561 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
565 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
567 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
570 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
571 function
, cachep
->name
, msg
);
573 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
578 * By default on NUMA we use alien caches to stage the freeing of
579 * objects allocated from other nodes. This causes massive memory
580 * inefficiencies when using fake NUMA setup to split memory into a
581 * large number of small nodes, so it can be disabled on the command
585 static int use_alien_caches __read_mostly
= 1;
586 static int __init
noaliencache_setup(char *s
)
588 use_alien_caches
= 0;
591 __setup("noaliencache", noaliencache_setup
);
593 static int __init
slab_max_order_setup(char *str
)
595 get_option(&str
, &slab_max_order
);
596 slab_max_order
= slab_max_order
< 0 ? 0 :
597 min(slab_max_order
, MAX_ORDER
- 1);
598 slab_max_order_set
= true;
602 __setup("slab_max_order=", slab_max_order_setup
);
606 * Special reaping functions for NUMA systems called from cache_reap().
607 * These take care of doing round robin flushing of alien caches (containing
608 * objects freed on different nodes from which they were allocated) and the
609 * flushing of remote pcps by calling drain_node_pages.
611 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
613 static void init_reap_node(int cpu
)
617 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
618 if (node
== MAX_NUMNODES
)
619 node
= first_node(node_online_map
);
621 per_cpu(slab_reap_node
, cpu
) = node
;
624 static void next_reap_node(void)
626 int node
= __this_cpu_read(slab_reap_node
);
628 node
= next_node(node
, node_online_map
);
629 if (unlikely(node
>= MAX_NUMNODES
))
630 node
= first_node(node_online_map
);
631 __this_cpu_write(slab_reap_node
, node
);
635 #define init_reap_node(cpu) do { } while (0)
636 #define next_reap_node(void) do { } while (0)
640 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
641 * via the workqueue/eventd.
642 * Add the CPU number into the expiration time to minimize the possibility of
643 * the CPUs getting into lockstep and contending for the global cache chain
646 static void start_cpu_timer(int cpu
)
648 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
651 * When this gets called from do_initcalls via cpucache_init(),
652 * init_workqueues() has already run, so keventd will be setup
655 if (keventd_up() && reap_work
->work
.func
== NULL
) {
657 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
658 schedule_delayed_work_on(cpu
, reap_work
,
659 __round_jiffies_relative(HZ
, cpu
));
663 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
666 * The array_cache structures contain pointers to free object.
667 * However, when such objects are allocated or transferred to another
668 * cache the pointers are not cleared and they could be counted as
669 * valid references during a kmemleak scan. Therefore, kmemleak must
670 * not scan such objects.
672 kmemleak_no_scan(ac
);
676 ac
->batchcount
= batch
;
681 static struct array_cache
*alloc_arraycache(int node
, int entries
,
682 int batchcount
, gfp_t gfp
)
684 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
685 struct array_cache
*ac
= NULL
;
687 ac
= kmalloc_node(memsize
, gfp
, node
);
688 init_arraycache(ac
, entries
, batchcount
);
692 static inline bool is_slab_pfmemalloc(struct page
*page
)
694 return PageSlabPfmemalloc(page
);
697 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
698 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
699 struct array_cache
*ac
)
701 struct kmem_cache_node
*n
= get_node(cachep
, numa_mem_id());
705 if (!pfmemalloc_active
)
708 spin_lock_irqsave(&n
->list_lock
, flags
);
709 list_for_each_entry(page
, &n
->slabs_full
, lru
)
710 if (is_slab_pfmemalloc(page
))
713 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
714 if (is_slab_pfmemalloc(page
))
717 list_for_each_entry(page
, &n
->slabs_free
, lru
)
718 if (is_slab_pfmemalloc(page
))
721 pfmemalloc_active
= false;
723 spin_unlock_irqrestore(&n
->list_lock
, flags
);
726 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
727 gfp_t flags
, bool force_refill
)
730 void *objp
= ac
->entry
[--ac
->avail
];
732 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
733 if (unlikely(is_obj_pfmemalloc(objp
))) {
734 struct kmem_cache_node
*n
;
736 if (gfp_pfmemalloc_allowed(flags
)) {
737 clear_obj_pfmemalloc(&objp
);
741 /* The caller cannot use PFMEMALLOC objects, find another one */
742 for (i
= 0; i
< ac
->avail
; i
++) {
743 /* If a !PFMEMALLOC object is found, swap them */
744 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
746 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
747 ac
->entry
[ac
->avail
] = objp
;
753 * If there are empty slabs on the slabs_free list and we are
754 * being forced to refill the cache, mark this one !pfmemalloc.
756 n
= get_node(cachep
, numa_mem_id());
757 if (!list_empty(&n
->slabs_free
) && force_refill
) {
758 struct page
*page
= virt_to_head_page(objp
);
759 ClearPageSlabPfmemalloc(page
);
760 clear_obj_pfmemalloc(&objp
);
761 recheck_pfmemalloc_active(cachep
, ac
);
765 /* No !PFMEMALLOC objects available */
773 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
774 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
778 if (unlikely(sk_memalloc_socks()))
779 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
781 objp
= ac
->entry
[--ac
->avail
];
786 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
789 if (unlikely(pfmemalloc_active
)) {
790 /* Some pfmemalloc slabs exist, check if this is one */
791 struct page
*page
= virt_to_head_page(objp
);
792 if (PageSlabPfmemalloc(page
))
793 set_obj_pfmemalloc(&objp
);
799 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
802 if (unlikely(sk_memalloc_socks()))
803 objp
= __ac_put_obj(cachep
, ac
, objp
);
805 ac
->entry
[ac
->avail
++] = objp
;
809 * Transfer objects in one arraycache to another.
810 * Locking must be handled by the caller.
812 * Return the number of entries transferred.
814 static int transfer_objects(struct array_cache
*to
,
815 struct array_cache
*from
, unsigned int max
)
817 /* Figure out how many entries to transfer */
818 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
823 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
833 #define drain_alien_cache(cachep, alien) do { } while (0)
834 #define reap_alien(cachep, n) do { } while (0)
836 static inline struct alien_cache
**alloc_alien_cache(int node
,
837 int limit
, gfp_t gfp
)
842 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
846 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
851 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
857 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
858 gfp_t flags
, int nodeid
)
863 #else /* CONFIG_NUMA */
865 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
866 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
868 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
869 int batch
, gfp_t gfp
)
871 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
872 struct alien_cache
*alc
= NULL
;
874 alc
= kmalloc_node(memsize
, gfp
, node
);
875 init_arraycache(&alc
->ac
, entries
, batch
);
876 spin_lock_init(&alc
->lock
);
880 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
882 struct alien_cache
**alc_ptr
;
883 size_t memsize
= sizeof(void *) * nr_node_ids
;
888 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
893 if (i
== node
|| !node_online(i
))
895 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
897 for (i
--; i
>= 0; i
--)
906 static void free_alien_cache(struct alien_cache
**alc_ptr
)
917 static void __drain_alien_cache(struct kmem_cache
*cachep
,
918 struct array_cache
*ac
, int node
,
919 struct list_head
*list
)
921 struct kmem_cache_node
*n
= get_node(cachep
, node
);
924 spin_lock(&n
->list_lock
);
926 * Stuff objects into the remote nodes shared array first.
927 * That way we could avoid the overhead of putting the objects
928 * into the free lists and getting them back later.
931 transfer_objects(n
->shared
, ac
, ac
->limit
);
933 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
935 spin_unlock(&n
->list_lock
);
940 * Called from cache_reap() to regularly drain alien caches round robin.
942 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
944 int node
= __this_cpu_read(slab_reap_node
);
947 struct alien_cache
*alc
= n
->alien
[node
];
948 struct array_cache
*ac
;
952 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
955 __drain_alien_cache(cachep
, ac
, node
, &list
);
956 spin_unlock_irq(&alc
->lock
);
957 slabs_destroy(cachep
, &list
);
963 static void drain_alien_cache(struct kmem_cache
*cachep
,
964 struct alien_cache
**alien
)
967 struct alien_cache
*alc
;
968 struct array_cache
*ac
;
971 for_each_online_node(i
) {
977 spin_lock_irqsave(&alc
->lock
, flags
);
978 __drain_alien_cache(cachep
, ac
, i
, &list
);
979 spin_unlock_irqrestore(&alc
->lock
, flags
);
980 slabs_destroy(cachep
, &list
);
985 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
987 int nodeid
= page_to_nid(virt_to_page(objp
));
988 struct kmem_cache_node
*n
;
989 struct alien_cache
*alien
= NULL
;
990 struct array_cache
*ac
;
994 node
= numa_mem_id();
997 * Make sure we are not freeing a object from another node to the array
1000 if (likely(nodeid
== node
))
1003 n
= get_node(cachep
, node
);
1004 STATS_INC_NODEFREES(cachep
);
1005 if (n
->alien
&& n
->alien
[nodeid
]) {
1006 alien
= n
->alien
[nodeid
];
1008 spin_lock(&alien
->lock
);
1009 if (unlikely(ac
->avail
== ac
->limit
)) {
1010 STATS_INC_ACOVERFLOW(cachep
);
1011 __drain_alien_cache(cachep
, ac
, nodeid
, &list
);
1013 ac_put_obj(cachep
, ac
, objp
);
1014 spin_unlock(&alien
->lock
);
1015 slabs_destroy(cachep
, &list
);
1017 n
= get_node(cachep
, nodeid
);
1018 spin_lock(&n
->list_lock
);
1019 free_block(cachep
, &objp
, 1, nodeid
, &list
);
1020 spin_unlock(&n
->list_lock
);
1021 slabs_destroy(cachep
, &list
);
1028 * Allocates and initializes node for a node on each slab cache, used for
1029 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1030 * will be allocated off-node since memory is not yet online for the new node.
1031 * When hotplugging memory or a cpu, existing node are not replaced if
1034 * Must hold slab_mutex.
1036 static int init_cache_node_node(int node
)
1038 struct kmem_cache
*cachep
;
1039 struct kmem_cache_node
*n
;
1040 const size_t memsize
= sizeof(struct kmem_cache_node
);
1042 list_for_each_entry(cachep
, &slab_caches
, list
) {
1044 * Set up the kmem_cache_node for cpu before we can
1045 * begin anything. Make sure some other cpu on this
1046 * node has not already allocated this
1048 n
= get_node(cachep
, node
);
1050 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1053 kmem_cache_node_init(n
);
1054 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
1055 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1058 * The kmem_cache_nodes don't come and go as CPUs
1059 * come and go. slab_mutex is sufficient
1062 cachep
->node
[node
] = n
;
1065 spin_lock_irq(&n
->list_lock
);
1067 (1 + nr_cpus_node(node
)) *
1068 cachep
->batchcount
+ cachep
->num
;
1069 spin_unlock_irq(&n
->list_lock
);
1074 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1075 struct kmem_cache_node
*n
)
1077 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1080 static void cpuup_canceled(long cpu
)
1082 struct kmem_cache
*cachep
;
1083 struct kmem_cache_node
*n
= NULL
;
1084 int node
= cpu_to_mem(cpu
);
1085 const struct cpumask
*mask
= cpumask_of_node(node
);
1087 list_for_each_entry(cachep
, &slab_caches
, list
) {
1088 struct array_cache
*nc
;
1089 struct array_cache
*shared
;
1090 struct alien_cache
**alien
;
1093 /* cpu is dead; no one can alloc from it. */
1094 nc
= cachep
->array
[cpu
];
1095 cachep
->array
[cpu
] = NULL
;
1096 n
= get_node(cachep
, node
);
1099 goto free_array_cache
;
1101 spin_lock_irq(&n
->list_lock
);
1103 /* Free limit for this kmem_cache_node */
1104 n
->free_limit
-= cachep
->batchcount
;
1106 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1108 if (!cpumask_empty(mask
)) {
1109 spin_unlock_irq(&n
->list_lock
);
1110 goto free_array_cache
;
1115 free_block(cachep
, shared
->entry
,
1116 shared
->avail
, node
, &list
);
1123 spin_unlock_irq(&n
->list_lock
);
1127 drain_alien_cache(cachep
, alien
);
1128 free_alien_cache(alien
);
1131 slabs_destroy(cachep
, &list
);
1135 * In the previous loop, all the objects were freed to
1136 * the respective cache's slabs, now we can go ahead and
1137 * shrink each nodelist to its limit.
1139 list_for_each_entry(cachep
, &slab_caches
, list
) {
1140 n
= get_node(cachep
, node
);
1143 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1147 static int cpuup_prepare(long cpu
)
1149 struct kmem_cache
*cachep
;
1150 struct kmem_cache_node
*n
= NULL
;
1151 int node
= cpu_to_mem(cpu
);
1155 * We need to do this right in the beginning since
1156 * alloc_arraycache's are going to use this list.
1157 * kmalloc_node allows us to add the slab to the right
1158 * kmem_cache_node and not this cpu's kmem_cache_node
1160 err
= init_cache_node_node(node
);
1165 * Now we can go ahead with allocating the shared arrays and
1168 list_for_each_entry(cachep
, &slab_caches
, list
) {
1169 struct array_cache
*nc
;
1170 struct array_cache
*shared
= NULL
;
1171 struct alien_cache
**alien
= NULL
;
1173 nc
= alloc_arraycache(node
, cachep
->limit
,
1174 cachep
->batchcount
, GFP_KERNEL
);
1177 if (cachep
->shared
) {
1178 shared
= alloc_arraycache(node
,
1179 cachep
->shared
* cachep
->batchcount
,
1180 0xbaadf00d, GFP_KERNEL
);
1186 if (use_alien_caches
) {
1187 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1194 cachep
->array
[cpu
] = nc
;
1195 n
= get_node(cachep
, node
);
1198 spin_lock_irq(&n
->list_lock
);
1201 * We are serialised from CPU_DEAD or
1202 * CPU_UP_CANCELLED by the cpucontrol lock
1213 spin_unlock_irq(&n
->list_lock
);
1215 free_alien_cache(alien
);
1220 cpuup_canceled(cpu
);
1224 static int cpuup_callback(struct notifier_block
*nfb
,
1225 unsigned long action
, void *hcpu
)
1227 long cpu
= (long)hcpu
;
1231 case CPU_UP_PREPARE
:
1232 case CPU_UP_PREPARE_FROZEN
:
1233 mutex_lock(&slab_mutex
);
1234 err
= cpuup_prepare(cpu
);
1235 mutex_unlock(&slab_mutex
);
1238 case CPU_ONLINE_FROZEN
:
1239 start_cpu_timer(cpu
);
1241 #ifdef CONFIG_HOTPLUG_CPU
1242 case CPU_DOWN_PREPARE
:
1243 case CPU_DOWN_PREPARE_FROZEN
:
1245 * Shutdown cache reaper. Note that the slab_mutex is
1246 * held so that if cache_reap() is invoked it cannot do
1247 * anything expensive but will only modify reap_work
1248 * and reschedule the timer.
1250 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1251 /* Now the cache_reaper is guaranteed to be not running. */
1252 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1254 case CPU_DOWN_FAILED
:
1255 case CPU_DOWN_FAILED_FROZEN
:
1256 start_cpu_timer(cpu
);
1259 case CPU_DEAD_FROZEN
:
1261 * Even if all the cpus of a node are down, we don't free the
1262 * kmem_cache_node of any cache. This to avoid a race between
1263 * cpu_down, and a kmalloc allocation from another cpu for
1264 * memory from the node of the cpu going down. The node
1265 * structure is usually allocated from kmem_cache_create() and
1266 * gets destroyed at kmem_cache_destroy().
1270 case CPU_UP_CANCELED
:
1271 case CPU_UP_CANCELED_FROZEN
:
1272 mutex_lock(&slab_mutex
);
1273 cpuup_canceled(cpu
);
1274 mutex_unlock(&slab_mutex
);
1277 return notifier_from_errno(err
);
1280 static struct notifier_block cpucache_notifier
= {
1281 &cpuup_callback
, NULL
, 0
1284 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1286 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1287 * Returns -EBUSY if all objects cannot be drained so that the node is not
1290 * Must hold slab_mutex.
1292 static int __meminit
drain_cache_node_node(int node
)
1294 struct kmem_cache
*cachep
;
1297 list_for_each_entry(cachep
, &slab_caches
, list
) {
1298 struct kmem_cache_node
*n
;
1300 n
= get_node(cachep
, node
);
1304 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1306 if (!list_empty(&n
->slabs_full
) ||
1307 !list_empty(&n
->slabs_partial
)) {
1315 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1316 unsigned long action
, void *arg
)
1318 struct memory_notify
*mnb
= arg
;
1322 nid
= mnb
->status_change_nid
;
1327 case MEM_GOING_ONLINE
:
1328 mutex_lock(&slab_mutex
);
1329 ret
= init_cache_node_node(nid
);
1330 mutex_unlock(&slab_mutex
);
1332 case MEM_GOING_OFFLINE
:
1333 mutex_lock(&slab_mutex
);
1334 ret
= drain_cache_node_node(nid
);
1335 mutex_unlock(&slab_mutex
);
1339 case MEM_CANCEL_ONLINE
:
1340 case MEM_CANCEL_OFFLINE
:
1344 return notifier_from_errno(ret
);
1346 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1349 * swap the static kmem_cache_node with kmalloced memory
1351 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1354 struct kmem_cache_node
*ptr
;
1356 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1359 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1361 * Do not assume that spinlocks can be initialized via memcpy:
1363 spin_lock_init(&ptr
->list_lock
);
1365 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1366 cachep
->node
[nodeid
] = ptr
;
1370 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1371 * size of kmem_cache_node.
1373 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1377 for_each_online_node(node
) {
1378 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1379 cachep
->node
[node
]->next_reap
= jiffies
+
1381 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1386 * The memory after the last cpu cache pointer is used for the
1389 static void setup_node_pointer(struct kmem_cache
*cachep
)
1391 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1395 * Initialisation. Called after the page allocator have been initialised and
1396 * before smp_init().
1398 void __init
kmem_cache_init(void)
1402 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1403 sizeof(struct rcu_head
));
1404 kmem_cache
= &kmem_cache_boot
;
1405 setup_node_pointer(kmem_cache
);
1407 if (num_possible_nodes() == 1)
1408 use_alien_caches
= 0;
1410 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1411 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1413 set_up_node(kmem_cache
, CACHE_CACHE
);
1416 * Fragmentation resistance on low memory - only use bigger
1417 * page orders on machines with more than 32MB of memory if
1418 * not overridden on the command line.
1420 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1421 slab_max_order
= SLAB_MAX_ORDER_HI
;
1423 /* Bootstrap is tricky, because several objects are allocated
1424 * from caches that do not exist yet:
1425 * 1) initialize the kmem_cache cache: it contains the struct
1426 * kmem_cache structures of all caches, except kmem_cache itself:
1427 * kmem_cache is statically allocated.
1428 * Initially an __init data area is used for the head array and the
1429 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1430 * array at the end of the bootstrap.
1431 * 2) Create the first kmalloc cache.
1432 * The struct kmem_cache for the new cache is allocated normally.
1433 * An __init data area is used for the head array.
1434 * 3) Create the remaining kmalloc caches, with minimally sized
1436 * 4) Replace the __init data head arrays for kmem_cache and the first
1437 * kmalloc cache with kmalloc allocated arrays.
1438 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1439 * the other cache's with kmalloc allocated memory.
1440 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1443 /* 1) create the kmem_cache */
1446 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1448 create_boot_cache(kmem_cache
, "kmem_cache",
1449 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1450 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1451 SLAB_HWCACHE_ALIGN
);
1452 list_add(&kmem_cache
->list
, &slab_caches
);
1454 /* 2+3) create the kmalloc caches */
1457 * Initialize the caches that provide memory for the array cache and the
1458 * kmem_cache_node structures first. Without this, further allocations will
1462 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1463 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1465 if (INDEX_AC
!= INDEX_NODE
)
1466 kmalloc_caches
[INDEX_NODE
] =
1467 create_kmalloc_cache("kmalloc-node",
1468 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1470 slab_early_init
= 0;
1472 /* 4) Replace the bootstrap head arrays */
1474 struct array_cache
*ptr
;
1476 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1478 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1479 sizeof(struct arraycache_init
));
1481 kmem_cache
->array
[smp_processor_id()] = ptr
;
1483 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1485 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1486 != &initarray_generic
.cache
);
1487 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1488 sizeof(struct arraycache_init
));
1490 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1492 /* 5) Replace the bootstrap kmem_cache_node */
1496 for_each_online_node(nid
) {
1497 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1499 init_list(kmalloc_caches
[INDEX_AC
],
1500 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1502 if (INDEX_AC
!= INDEX_NODE
) {
1503 init_list(kmalloc_caches
[INDEX_NODE
],
1504 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1509 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1512 void __init
kmem_cache_init_late(void)
1514 struct kmem_cache
*cachep
;
1518 /* 6) resize the head arrays to their final sizes */
1519 mutex_lock(&slab_mutex
);
1520 list_for_each_entry(cachep
, &slab_caches
, list
)
1521 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1523 mutex_unlock(&slab_mutex
);
1529 * Register a cpu startup notifier callback that initializes
1530 * cpu_cache_get for all new cpus
1532 register_cpu_notifier(&cpucache_notifier
);
1536 * Register a memory hotplug callback that initializes and frees
1539 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1543 * The reap timers are started later, with a module init call: That part
1544 * of the kernel is not yet operational.
1548 static int __init
cpucache_init(void)
1553 * Register the timers that return unneeded pages to the page allocator
1555 for_each_online_cpu(cpu
)
1556 start_cpu_timer(cpu
);
1562 __initcall(cpucache_init
);
1564 static noinline
void
1565 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1568 struct kmem_cache_node
*n
;
1570 unsigned long flags
;
1572 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1573 DEFAULT_RATELIMIT_BURST
);
1575 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1579 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1581 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1582 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1584 for_each_kmem_cache_node(cachep
, node
, n
) {
1585 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1586 unsigned long active_slabs
= 0, num_slabs
= 0;
1588 spin_lock_irqsave(&n
->list_lock
, flags
);
1589 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1590 active_objs
+= cachep
->num
;
1593 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1594 active_objs
+= page
->active
;
1597 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1600 free_objects
+= n
->free_objects
;
1601 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1603 num_slabs
+= active_slabs
;
1604 num_objs
= num_slabs
* cachep
->num
;
1606 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1607 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1614 * Interface to system's page allocator. No need to hold the
1615 * kmem_cache_node ->list_lock.
1617 * If we requested dmaable memory, we will get it. Even if we
1618 * did not request dmaable memory, we might get it, but that
1619 * would be relatively rare and ignorable.
1621 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1627 flags
|= cachep
->allocflags
;
1628 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1629 flags
|= __GFP_RECLAIMABLE
;
1631 if (memcg_charge_slab(cachep
, flags
, cachep
->gfporder
))
1634 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1636 memcg_uncharge_slab(cachep
, cachep
->gfporder
);
1637 slab_out_of_memory(cachep
, flags
, nodeid
);
1641 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1642 if (unlikely(page
->pfmemalloc
))
1643 pfmemalloc_active
= true;
1645 nr_pages
= (1 << cachep
->gfporder
);
1646 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1647 add_zone_page_state(page_zone(page
),
1648 NR_SLAB_RECLAIMABLE
, nr_pages
);
1650 add_zone_page_state(page_zone(page
),
1651 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1652 __SetPageSlab(page
);
1653 if (page
->pfmemalloc
)
1654 SetPageSlabPfmemalloc(page
);
1656 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1657 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1660 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1662 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1669 * Interface to system's page release.
1671 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1673 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1675 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1677 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1678 sub_zone_page_state(page_zone(page
),
1679 NR_SLAB_RECLAIMABLE
, nr_freed
);
1681 sub_zone_page_state(page_zone(page
),
1682 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1684 BUG_ON(!PageSlab(page
));
1685 __ClearPageSlabPfmemalloc(page
);
1686 __ClearPageSlab(page
);
1687 page_mapcount_reset(page
);
1688 page
->mapping
= NULL
;
1690 if (current
->reclaim_state
)
1691 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1692 __free_pages(page
, cachep
->gfporder
);
1693 memcg_uncharge_slab(cachep
, cachep
->gfporder
);
1696 static void kmem_rcu_free(struct rcu_head
*head
)
1698 struct kmem_cache
*cachep
;
1701 page
= container_of(head
, struct page
, rcu_head
);
1702 cachep
= page
->slab_cache
;
1704 kmem_freepages(cachep
, page
);
1709 #ifdef CONFIG_DEBUG_PAGEALLOC
1710 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1711 unsigned long caller
)
1713 int size
= cachep
->object_size
;
1715 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1717 if (size
< 5 * sizeof(unsigned long))
1720 *addr
++ = 0x12345678;
1722 *addr
++ = smp_processor_id();
1723 size
-= 3 * sizeof(unsigned long);
1725 unsigned long *sptr
= &caller
;
1726 unsigned long svalue
;
1728 while (!kstack_end(sptr
)) {
1730 if (kernel_text_address(svalue
)) {
1732 size
-= sizeof(unsigned long);
1733 if (size
<= sizeof(unsigned long))
1739 *addr
++ = 0x87654321;
1743 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1745 int size
= cachep
->object_size
;
1746 addr
= &((char *)addr
)[obj_offset(cachep
)];
1748 memset(addr
, val
, size
);
1749 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1752 static void dump_line(char *data
, int offset
, int limit
)
1755 unsigned char error
= 0;
1758 printk(KERN_ERR
"%03x: ", offset
);
1759 for (i
= 0; i
< limit
; i
++) {
1760 if (data
[offset
+ i
] != POISON_FREE
) {
1761 error
= data
[offset
+ i
];
1765 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1766 &data
[offset
], limit
, 1);
1768 if (bad_count
== 1) {
1769 error
^= POISON_FREE
;
1770 if (!(error
& (error
- 1))) {
1771 printk(KERN_ERR
"Single bit error detected. Probably "
1774 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1777 printk(KERN_ERR
"Run a memory test tool.\n");
1786 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1791 if (cachep
->flags
& SLAB_RED_ZONE
) {
1792 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1793 *dbg_redzone1(cachep
, objp
),
1794 *dbg_redzone2(cachep
, objp
));
1797 if (cachep
->flags
& SLAB_STORE_USER
) {
1798 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1799 *dbg_userword(cachep
, objp
),
1800 *dbg_userword(cachep
, objp
));
1802 realobj
= (char *)objp
+ obj_offset(cachep
);
1803 size
= cachep
->object_size
;
1804 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1807 if (i
+ limit
> size
)
1809 dump_line(realobj
, i
, limit
);
1813 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1819 realobj
= (char *)objp
+ obj_offset(cachep
);
1820 size
= cachep
->object_size
;
1822 for (i
= 0; i
< size
; i
++) {
1823 char exp
= POISON_FREE
;
1826 if (realobj
[i
] != exp
) {
1832 "Slab corruption (%s): %s start=%p, len=%d\n",
1833 print_tainted(), cachep
->name
, realobj
, size
);
1834 print_objinfo(cachep
, objp
, 0);
1836 /* Hexdump the affected line */
1839 if (i
+ limit
> size
)
1841 dump_line(realobj
, i
, limit
);
1844 /* Limit to 5 lines */
1850 /* Print some data about the neighboring objects, if they
1853 struct page
*page
= virt_to_head_page(objp
);
1856 objnr
= obj_to_index(cachep
, page
, objp
);
1858 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1859 realobj
= (char *)objp
+ obj_offset(cachep
);
1860 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1862 print_objinfo(cachep
, objp
, 2);
1864 if (objnr
+ 1 < cachep
->num
) {
1865 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1866 realobj
= (char *)objp
+ obj_offset(cachep
);
1867 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1869 print_objinfo(cachep
, objp
, 2);
1876 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1880 for (i
= 0; i
< cachep
->num
; i
++) {
1881 void *objp
= index_to_obj(cachep
, page
, i
);
1883 if (cachep
->flags
& SLAB_POISON
) {
1884 #ifdef CONFIG_DEBUG_PAGEALLOC
1885 if (cachep
->size
% PAGE_SIZE
== 0 &&
1887 kernel_map_pages(virt_to_page(objp
),
1888 cachep
->size
/ PAGE_SIZE
, 1);
1890 check_poison_obj(cachep
, objp
);
1892 check_poison_obj(cachep
, objp
);
1895 if (cachep
->flags
& SLAB_RED_ZONE
) {
1896 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1897 slab_error(cachep
, "start of a freed object "
1899 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1900 slab_error(cachep
, "end of a freed object "
1906 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1913 * slab_destroy - destroy and release all objects in a slab
1914 * @cachep: cache pointer being destroyed
1915 * @page: page pointer being destroyed
1917 * Destroy all the objs in a slab page, and release the mem back to the system.
1918 * Before calling the slab page must have been unlinked from the cache. The
1919 * kmem_cache_node ->list_lock is not held/needed.
1921 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1925 freelist
= page
->freelist
;
1926 slab_destroy_debugcheck(cachep
, page
);
1927 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1928 struct rcu_head
*head
;
1931 * RCU free overloads the RCU head over the LRU.
1932 * slab_page has been overloeaded over the LRU,
1933 * however it is not used from now on so that
1934 * we can use it safely.
1936 head
= (void *)&page
->rcu_head
;
1937 call_rcu(head
, kmem_rcu_free
);
1940 kmem_freepages(cachep
, page
);
1944 * From now on, we don't use freelist
1945 * although actual page can be freed in rcu context
1947 if (OFF_SLAB(cachep
))
1948 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1951 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1953 struct page
*page
, *n
;
1955 list_for_each_entry_safe(page
, n
, list
, lru
) {
1956 list_del(&page
->lru
);
1957 slab_destroy(cachep
, page
);
1962 * calculate_slab_order - calculate size (page order) of slabs
1963 * @cachep: pointer to the cache that is being created
1964 * @size: size of objects to be created in this cache.
1965 * @align: required alignment for the objects.
1966 * @flags: slab allocation flags
1968 * Also calculates the number of objects per slab.
1970 * This could be made much more intelligent. For now, try to avoid using
1971 * high order pages for slabs. When the gfp() functions are more friendly
1972 * towards high-order requests, this should be changed.
1974 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1975 size_t size
, size_t align
, unsigned long flags
)
1977 unsigned long offslab_limit
;
1978 size_t left_over
= 0;
1981 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1985 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
1989 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1990 if (num
> SLAB_OBJ_MAX_NUM
)
1993 if (flags
& CFLGS_OFF_SLAB
) {
1994 size_t freelist_size_per_obj
= sizeof(freelist_idx_t
);
1996 * Max number of objs-per-slab for caches which
1997 * use off-slab slabs. Needed to avoid a possible
1998 * looping condition in cache_grow().
2000 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
2001 freelist_size_per_obj
+= sizeof(char);
2002 offslab_limit
= size
;
2003 offslab_limit
/= freelist_size_per_obj
;
2005 if (num
> offslab_limit
)
2009 /* Found something acceptable - save it away */
2011 cachep
->gfporder
= gfporder
;
2012 left_over
= remainder
;
2015 * A VFS-reclaimable slab tends to have most allocations
2016 * as GFP_NOFS and we really don't want to have to be allocating
2017 * higher-order pages when we are unable to shrink dcache.
2019 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2023 * Large number of objects is good, but very large slabs are
2024 * currently bad for the gfp()s.
2026 if (gfporder
>= slab_max_order
)
2030 * Acceptable internal fragmentation?
2032 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2038 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2040 if (slab_state
>= FULL
)
2041 return enable_cpucache(cachep
, gfp
);
2043 if (slab_state
== DOWN
) {
2045 * Note: Creation of first cache (kmem_cache).
2046 * The setup_node is taken care
2047 * of by the caller of __kmem_cache_create
2049 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2050 slab_state
= PARTIAL
;
2051 } else if (slab_state
== PARTIAL
) {
2053 * Note: the second kmem_cache_create must create the cache
2054 * that's used by kmalloc(24), otherwise the creation of
2055 * further caches will BUG().
2057 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2060 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2061 * the second cache, then we need to set up all its node/,
2062 * otherwise the creation of further caches will BUG().
2064 set_up_node(cachep
, SIZE_AC
);
2065 if (INDEX_AC
== INDEX_NODE
)
2066 slab_state
= PARTIAL_NODE
;
2068 slab_state
= PARTIAL_ARRAYCACHE
;
2070 /* Remaining boot caches */
2071 cachep
->array
[smp_processor_id()] =
2072 kmalloc(sizeof(struct arraycache_init
), gfp
);
2074 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2075 set_up_node(cachep
, SIZE_NODE
);
2076 slab_state
= PARTIAL_NODE
;
2079 for_each_online_node(node
) {
2080 cachep
->node
[node
] =
2081 kmalloc_node(sizeof(struct kmem_cache_node
),
2083 BUG_ON(!cachep
->node
[node
]);
2084 kmem_cache_node_init(cachep
->node
[node
]);
2088 cachep
->node
[numa_mem_id()]->next_reap
=
2089 jiffies
+ REAPTIMEOUT_NODE
+
2090 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
2092 cpu_cache_get(cachep
)->avail
= 0;
2093 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2094 cpu_cache_get(cachep
)->batchcount
= 1;
2095 cpu_cache_get(cachep
)->touched
= 0;
2096 cachep
->batchcount
= 1;
2097 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2102 * __kmem_cache_create - Create a cache.
2103 * @cachep: cache management descriptor
2104 * @flags: SLAB flags
2106 * Returns a ptr to the cache on success, NULL on failure.
2107 * Cannot be called within a int, but can be interrupted.
2108 * The @ctor is run when new pages are allocated by the cache.
2112 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2113 * to catch references to uninitialised memory.
2115 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2116 * for buffer overruns.
2118 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2119 * cacheline. This can be beneficial if you're counting cycles as closely
2123 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2125 size_t left_over
, freelist_size
, ralign
;
2128 size_t size
= cachep
->size
;
2133 * Enable redzoning and last user accounting, except for caches with
2134 * large objects, if the increased size would increase the object size
2135 * above the next power of two: caches with object sizes just above a
2136 * power of two have a significant amount of internal fragmentation.
2138 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2139 2 * sizeof(unsigned long long)))
2140 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2141 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2142 flags
|= SLAB_POISON
;
2144 if (flags
& SLAB_DESTROY_BY_RCU
)
2145 BUG_ON(flags
& SLAB_POISON
);
2149 * Check that size is in terms of words. This is needed to avoid
2150 * unaligned accesses for some archs when redzoning is used, and makes
2151 * sure any on-slab bufctl's are also correctly aligned.
2153 if (size
& (BYTES_PER_WORD
- 1)) {
2154 size
+= (BYTES_PER_WORD
- 1);
2155 size
&= ~(BYTES_PER_WORD
- 1);
2159 * Redzoning and user store require word alignment or possibly larger.
2160 * Note this will be overridden by architecture or caller mandated
2161 * alignment if either is greater than BYTES_PER_WORD.
2163 if (flags
& SLAB_STORE_USER
)
2164 ralign
= BYTES_PER_WORD
;
2166 if (flags
& SLAB_RED_ZONE
) {
2167 ralign
= REDZONE_ALIGN
;
2168 /* If redzoning, ensure that the second redzone is suitably
2169 * aligned, by adjusting the object size accordingly. */
2170 size
+= REDZONE_ALIGN
- 1;
2171 size
&= ~(REDZONE_ALIGN
- 1);
2174 /* 3) caller mandated alignment */
2175 if (ralign
< cachep
->align
) {
2176 ralign
= cachep
->align
;
2178 /* disable debug if necessary */
2179 if (ralign
> __alignof__(unsigned long long))
2180 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2184 cachep
->align
= ralign
;
2186 if (slab_is_available())
2191 setup_node_pointer(cachep
);
2195 * Both debugging options require word-alignment which is calculated
2198 if (flags
& SLAB_RED_ZONE
) {
2199 /* add space for red zone words */
2200 cachep
->obj_offset
+= sizeof(unsigned long long);
2201 size
+= 2 * sizeof(unsigned long long);
2203 if (flags
& SLAB_STORE_USER
) {
2204 /* user store requires one word storage behind the end of
2205 * the real object. But if the second red zone needs to be
2206 * aligned to 64 bits, we must allow that much space.
2208 if (flags
& SLAB_RED_ZONE
)
2209 size
+= REDZONE_ALIGN
;
2211 size
+= BYTES_PER_WORD
;
2213 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2214 if (size
>= kmalloc_size(INDEX_NODE
+ 1)
2215 && cachep
->object_size
> cache_line_size()
2216 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2217 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2224 * Determine if the slab management is 'on' or 'off' slab.
2225 * (bootstrapping cannot cope with offslab caches so don't do
2226 * it too early on. Always use on-slab management when
2227 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2229 if ((size
>= (PAGE_SIZE
>> 5)) && !slab_early_init
&&
2230 !(flags
& SLAB_NOLEAKTRACE
))
2232 * Size is large, assume best to place the slab management obj
2233 * off-slab (should allow better packing of objs).
2235 flags
|= CFLGS_OFF_SLAB
;
2237 size
= ALIGN(size
, cachep
->align
);
2239 * We should restrict the number of objects in a slab to implement
2240 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2242 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2243 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2245 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2250 freelist_size
= calculate_freelist_size(cachep
->num
, cachep
->align
);
2253 * If the slab has been placed off-slab, and we have enough space then
2254 * move it on-slab. This is at the expense of any extra colouring.
2256 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= freelist_size
) {
2257 flags
&= ~CFLGS_OFF_SLAB
;
2258 left_over
-= freelist_size
;
2261 if (flags
& CFLGS_OFF_SLAB
) {
2262 /* really off slab. No need for manual alignment */
2263 freelist_size
= calculate_freelist_size(cachep
->num
, 0);
2265 #ifdef CONFIG_PAGE_POISONING
2266 /* If we're going to use the generic kernel_map_pages()
2267 * poisoning, then it's going to smash the contents of
2268 * the redzone and userword anyhow, so switch them off.
2270 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2271 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2275 cachep
->colour_off
= cache_line_size();
2276 /* Offset must be a multiple of the alignment. */
2277 if (cachep
->colour_off
< cachep
->align
)
2278 cachep
->colour_off
= cachep
->align
;
2279 cachep
->colour
= left_over
/ cachep
->colour_off
;
2280 cachep
->freelist_size
= freelist_size
;
2281 cachep
->flags
= flags
;
2282 cachep
->allocflags
= __GFP_COMP
;
2283 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2284 cachep
->allocflags
|= GFP_DMA
;
2285 cachep
->size
= size
;
2286 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2288 if (flags
& CFLGS_OFF_SLAB
) {
2289 cachep
->freelist_cache
= kmalloc_slab(freelist_size
, 0u);
2291 * This is a possibility for one of the kmalloc_{dma,}_caches.
2292 * But since we go off slab only for object size greater than
2293 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2294 * in ascending order,this should not happen at all.
2295 * But leave a BUG_ON for some lucky dude.
2297 BUG_ON(ZERO_OR_NULL_PTR(cachep
->freelist_cache
));
2300 err
= setup_cpu_cache(cachep
, gfp
);
2302 __kmem_cache_shutdown(cachep
);
2310 static void check_irq_off(void)
2312 BUG_ON(!irqs_disabled());
2315 static void check_irq_on(void)
2317 BUG_ON(irqs_disabled());
2320 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2324 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2328 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2332 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2337 #define check_irq_off() do { } while(0)
2338 #define check_irq_on() do { } while(0)
2339 #define check_spinlock_acquired(x) do { } while(0)
2340 #define check_spinlock_acquired_node(x, y) do { } while(0)
2343 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2344 struct array_cache
*ac
,
2345 int force
, int node
);
2347 static void do_drain(void *arg
)
2349 struct kmem_cache
*cachep
= arg
;
2350 struct array_cache
*ac
;
2351 int node
= numa_mem_id();
2352 struct kmem_cache_node
*n
;
2356 ac
= cpu_cache_get(cachep
);
2357 n
= get_node(cachep
, node
);
2358 spin_lock(&n
->list_lock
);
2359 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2360 spin_unlock(&n
->list_lock
);
2361 slabs_destroy(cachep
, &list
);
2365 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2367 struct kmem_cache_node
*n
;
2370 on_each_cpu(do_drain
, cachep
, 1);
2372 for_each_kmem_cache_node(cachep
, node
, n
)
2374 drain_alien_cache(cachep
, n
->alien
);
2376 for_each_kmem_cache_node(cachep
, node
, n
)
2377 drain_array(cachep
, n
, n
->shared
, 1, node
);
2381 * Remove slabs from the list of free slabs.
2382 * Specify the number of slabs to drain in tofree.
2384 * Returns the actual number of slabs released.
2386 static int drain_freelist(struct kmem_cache
*cache
,
2387 struct kmem_cache_node
*n
, int tofree
)
2389 struct list_head
*p
;
2394 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2396 spin_lock_irq(&n
->list_lock
);
2397 p
= n
->slabs_free
.prev
;
2398 if (p
== &n
->slabs_free
) {
2399 spin_unlock_irq(&n
->list_lock
);
2403 page
= list_entry(p
, struct page
, lru
);
2405 BUG_ON(page
->active
);
2407 list_del(&page
->lru
);
2409 * Safe to drop the lock. The slab is no longer linked
2412 n
->free_objects
-= cache
->num
;
2413 spin_unlock_irq(&n
->list_lock
);
2414 slab_destroy(cache
, page
);
2421 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2425 struct kmem_cache_node
*n
;
2427 drain_cpu_caches(cachep
);
2430 for_each_kmem_cache_node(cachep
, node
, n
) {
2431 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2433 ret
+= !list_empty(&n
->slabs_full
) ||
2434 !list_empty(&n
->slabs_partial
);
2436 return (ret
? 1 : 0);
2439 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2442 struct kmem_cache_node
*n
;
2443 int rc
= __kmem_cache_shrink(cachep
);
2448 for_each_online_cpu(i
)
2449 kfree(cachep
->array
[i
]);
2451 /* NUMA: free the node structures */
2452 for_each_kmem_cache_node(cachep
, i
, n
) {
2454 free_alien_cache(n
->alien
);
2456 cachep
->node
[i
] = NULL
;
2462 * Get the memory for a slab management obj.
2464 * For a slab cache when the slab descriptor is off-slab, the
2465 * slab descriptor can't come from the same cache which is being created,
2466 * Because if it is the case, that means we defer the creation of
2467 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2468 * And we eventually call down to __kmem_cache_create(), which
2469 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2470 * This is a "chicken-and-egg" problem.
2472 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2473 * which are all initialized during kmem_cache_init().
2475 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2476 struct page
*page
, int colour_off
,
2477 gfp_t local_flags
, int nodeid
)
2480 void *addr
= page_address(page
);
2482 if (OFF_SLAB(cachep
)) {
2483 /* Slab management obj is off-slab. */
2484 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2485 local_flags
, nodeid
);
2489 freelist
= addr
+ colour_off
;
2490 colour_off
+= cachep
->freelist_size
;
2493 page
->s_mem
= addr
+ colour_off
;
2497 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2499 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2502 static inline void set_free_obj(struct page
*page
,
2503 unsigned int idx
, freelist_idx_t val
)
2505 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2508 static void cache_init_objs(struct kmem_cache
*cachep
,
2513 for (i
= 0; i
< cachep
->num
; i
++) {
2514 void *objp
= index_to_obj(cachep
, page
, i
);
2516 /* need to poison the objs? */
2517 if (cachep
->flags
& SLAB_POISON
)
2518 poison_obj(cachep
, objp
, POISON_FREE
);
2519 if (cachep
->flags
& SLAB_STORE_USER
)
2520 *dbg_userword(cachep
, objp
) = NULL
;
2522 if (cachep
->flags
& SLAB_RED_ZONE
) {
2523 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2524 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2527 * Constructors are not allowed to allocate memory from the same
2528 * cache which they are a constructor for. Otherwise, deadlock.
2529 * They must also be threaded.
2531 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2532 cachep
->ctor(objp
+ obj_offset(cachep
));
2534 if (cachep
->flags
& SLAB_RED_ZONE
) {
2535 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2536 slab_error(cachep
, "constructor overwrote the"
2537 " end of an object");
2538 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2539 slab_error(cachep
, "constructor overwrote the"
2540 " start of an object");
2542 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2543 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2544 kernel_map_pages(virt_to_page(objp
),
2545 cachep
->size
/ PAGE_SIZE
, 0);
2550 set_obj_status(page
, i
, OBJECT_FREE
);
2551 set_free_obj(page
, i
, i
);
2555 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2557 if (CONFIG_ZONE_DMA_FLAG
) {
2558 if (flags
& GFP_DMA
)
2559 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2561 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2565 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
,
2570 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2573 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2579 static void slab_put_obj(struct kmem_cache
*cachep
, struct page
*page
,
2580 void *objp
, int nodeid
)
2582 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2586 /* Verify that the slab belongs to the intended node */
2587 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2589 /* Verify double free bug */
2590 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2591 if (get_free_obj(page
, i
) == objnr
) {
2592 printk(KERN_ERR
"slab: double free detected in cache "
2593 "'%s', objp %p\n", cachep
->name
, objp
);
2599 set_free_obj(page
, page
->active
, objnr
);
2603 * Map pages beginning at addr to the given cache and slab. This is required
2604 * for the slab allocator to be able to lookup the cache and slab of a
2605 * virtual address for kfree, ksize, and slab debugging.
2607 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2610 page
->slab_cache
= cache
;
2611 page
->freelist
= freelist
;
2615 * Grow (by 1) the number of slabs within a cache. This is called by
2616 * kmem_cache_alloc() when there are no active objs left in a cache.
2618 static int cache_grow(struct kmem_cache
*cachep
,
2619 gfp_t flags
, int nodeid
, struct page
*page
)
2624 struct kmem_cache_node
*n
;
2627 * Be lazy and only check for valid flags here, keeping it out of the
2628 * critical path in kmem_cache_alloc().
2630 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2631 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2633 /* Take the node list lock to change the colour_next on this node */
2635 n
= get_node(cachep
, nodeid
);
2636 spin_lock(&n
->list_lock
);
2638 /* Get colour for the slab, and cal the next value. */
2639 offset
= n
->colour_next
;
2641 if (n
->colour_next
>= cachep
->colour
)
2643 spin_unlock(&n
->list_lock
);
2645 offset
*= cachep
->colour_off
;
2647 if (local_flags
& __GFP_WAIT
)
2651 * The test for missing atomic flag is performed here, rather than
2652 * the more obvious place, simply to reduce the critical path length
2653 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2654 * will eventually be caught here (where it matters).
2656 kmem_flagcheck(cachep
, flags
);
2659 * Get mem for the objs. Attempt to allocate a physical page from
2663 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2667 /* Get slab management. */
2668 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2669 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2673 slab_map_pages(cachep
, page
, freelist
);
2675 cache_init_objs(cachep
, page
);
2677 if (local_flags
& __GFP_WAIT
)
2678 local_irq_disable();
2680 spin_lock(&n
->list_lock
);
2682 /* Make slab active. */
2683 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2684 STATS_INC_GROWN(cachep
);
2685 n
->free_objects
+= cachep
->num
;
2686 spin_unlock(&n
->list_lock
);
2689 kmem_freepages(cachep
, page
);
2691 if (local_flags
& __GFP_WAIT
)
2692 local_irq_disable();
2699 * Perform extra freeing checks:
2700 * - detect bad pointers.
2701 * - POISON/RED_ZONE checking
2703 static void kfree_debugcheck(const void *objp
)
2705 if (!virt_addr_valid(objp
)) {
2706 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2707 (unsigned long)objp
);
2712 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2714 unsigned long long redzone1
, redzone2
;
2716 redzone1
= *dbg_redzone1(cache
, obj
);
2717 redzone2
= *dbg_redzone2(cache
, obj
);
2722 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2725 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2726 slab_error(cache
, "double free detected");
2728 slab_error(cache
, "memory outside object was overwritten");
2730 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2731 obj
, redzone1
, redzone2
);
2734 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2735 unsigned long caller
)
2740 BUG_ON(virt_to_cache(objp
) != cachep
);
2742 objp
-= obj_offset(cachep
);
2743 kfree_debugcheck(objp
);
2744 page
= virt_to_head_page(objp
);
2746 if (cachep
->flags
& SLAB_RED_ZONE
) {
2747 verify_redzone_free(cachep
, objp
);
2748 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2749 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2751 if (cachep
->flags
& SLAB_STORE_USER
)
2752 *dbg_userword(cachep
, objp
) = (void *)caller
;
2754 objnr
= obj_to_index(cachep
, page
, objp
);
2756 BUG_ON(objnr
>= cachep
->num
);
2757 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2759 set_obj_status(page
, objnr
, OBJECT_FREE
);
2760 if (cachep
->flags
& SLAB_POISON
) {
2761 #ifdef CONFIG_DEBUG_PAGEALLOC
2762 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2763 store_stackinfo(cachep
, objp
, caller
);
2764 kernel_map_pages(virt_to_page(objp
),
2765 cachep
->size
/ PAGE_SIZE
, 0);
2767 poison_obj(cachep
, objp
, POISON_FREE
);
2770 poison_obj(cachep
, objp
, POISON_FREE
);
2777 #define kfree_debugcheck(x) do { } while(0)
2778 #define cache_free_debugcheck(x,objp,z) (objp)
2781 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2785 struct kmem_cache_node
*n
;
2786 struct array_cache
*ac
;
2790 node
= numa_mem_id();
2791 if (unlikely(force_refill
))
2794 ac
= cpu_cache_get(cachep
);
2795 batchcount
= ac
->batchcount
;
2796 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2798 * If there was little recent activity on this cache, then
2799 * perform only a partial refill. Otherwise we could generate
2802 batchcount
= BATCHREFILL_LIMIT
;
2804 n
= get_node(cachep
, node
);
2806 BUG_ON(ac
->avail
> 0 || !n
);
2807 spin_lock(&n
->list_lock
);
2809 /* See if we can refill from the shared array */
2810 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2811 n
->shared
->touched
= 1;
2815 while (batchcount
> 0) {
2816 struct list_head
*entry
;
2818 /* Get slab alloc is to come from. */
2819 entry
= n
->slabs_partial
.next
;
2820 if (entry
== &n
->slabs_partial
) {
2821 n
->free_touched
= 1;
2822 entry
= n
->slabs_free
.next
;
2823 if (entry
== &n
->slabs_free
)
2827 page
= list_entry(entry
, struct page
, lru
);
2828 check_spinlock_acquired(cachep
);
2831 * The slab was either on partial or free list so
2832 * there must be at least one object available for
2835 BUG_ON(page
->active
>= cachep
->num
);
2837 while (page
->active
< cachep
->num
&& batchcount
--) {
2838 STATS_INC_ALLOCED(cachep
);
2839 STATS_INC_ACTIVE(cachep
);
2840 STATS_SET_HIGH(cachep
);
2842 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, page
,
2846 /* move slabp to correct slabp list: */
2847 list_del(&page
->lru
);
2848 if (page
->active
== cachep
->num
)
2849 list_add(&page
->lru
, &n
->slabs_full
);
2851 list_add(&page
->lru
, &n
->slabs_partial
);
2855 n
->free_objects
-= ac
->avail
;
2857 spin_unlock(&n
->list_lock
);
2859 if (unlikely(!ac
->avail
)) {
2862 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2864 /* cache_grow can reenable interrupts, then ac could change. */
2865 ac
= cpu_cache_get(cachep
);
2866 node
= numa_mem_id();
2868 /* no objects in sight? abort */
2869 if (!x
&& (ac
->avail
== 0 || force_refill
))
2872 if (!ac
->avail
) /* objects refilled by interrupt? */
2877 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2880 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2883 might_sleep_if(flags
& __GFP_WAIT
);
2885 kmem_flagcheck(cachep
, flags
);
2890 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2891 gfp_t flags
, void *objp
, unsigned long caller
)
2897 if (cachep
->flags
& SLAB_POISON
) {
2898 #ifdef CONFIG_DEBUG_PAGEALLOC
2899 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2900 kernel_map_pages(virt_to_page(objp
),
2901 cachep
->size
/ PAGE_SIZE
, 1);
2903 check_poison_obj(cachep
, objp
);
2905 check_poison_obj(cachep
, objp
);
2907 poison_obj(cachep
, objp
, POISON_INUSE
);
2909 if (cachep
->flags
& SLAB_STORE_USER
)
2910 *dbg_userword(cachep
, objp
) = (void *)caller
;
2912 if (cachep
->flags
& SLAB_RED_ZONE
) {
2913 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2914 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2915 slab_error(cachep
, "double free, or memory outside"
2916 " object was overwritten");
2918 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2919 objp
, *dbg_redzone1(cachep
, objp
),
2920 *dbg_redzone2(cachep
, objp
));
2922 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2923 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2926 page
= virt_to_head_page(objp
);
2927 set_obj_status(page
, obj_to_index(cachep
, page
, objp
), OBJECT_ACTIVE
);
2928 objp
+= obj_offset(cachep
);
2929 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2931 if (ARCH_SLAB_MINALIGN
&&
2932 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2933 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2934 objp
, (int)ARCH_SLAB_MINALIGN
);
2939 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2942 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
2944 if (unlikely(cachep
== kmem_cache
))
2947 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
2950 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2953 struct array_cache
*ac
;
2954 bool force_refill
= false;
2958 ac
= cpu_cache_get(cachep
);
2959 if (likely(ac
->avail
)) {
2961 objp
= ac_get_obj(cachep
, ac
, flags
, false);
2964 * Allow for the possibility all avail objects are not allowed
2965 * by the current flags
2968 STATS_INC_ALLOCHIT(cachep
);
2971 force_refill
= true;
2974 STATS_INC_ALLOCMISS(cachep
);
2975 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
2977 * the 'ac' may be updated by cache_alloc_refill(),
2978 * and kmemleak_erase() requires its correct value.
2980 ac
= cpu_cache_get(cachep
);
2984 * To avoid a false negative, if an object that is in one of the
2985 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2986 * treat the array pointers as a reference to the object.
2989 kmemleak_erase(&ac
->entry
[ac
->avail
]);
2995 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
2997 * If we are in_interrupt, then process context, including cpusets and
2998 * mempolicy, may not apply and should not be used for allocation policy.
3000 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3002 int nid_alloc
, nid_here
;
3004 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3006 nid_alloc
= nid_here
= numa_mem_id();
3007 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3008 nid_alloc
= cpuset_slab_spread_node();
3009 else if (current
->mempolicy
)
3010 nid_alloc
= mempolicy_slab_node();
3011 if (nid_alloc
!= nid_here
)
3012 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3017 * Fallback function if there was no memory available and no objects on a
3018 * certain node and fall back is permitted. First we scan all the
3019 * available node for available objects. If that fails then we
3020 * perform an allocation without specifying a node. This allows the page
3021 * allocator to do its reclaim / fallback magic. We then insert the
3022 * slab into the proper nodelist and then allocate from it.
3024 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3026 struct zonelist
*zonelist
;
3030 enum zone_type high_zoneidx
= gfp_zone(flags
);
3033 unsigned int cpuset_mems_cookie
;
3035 if (flags
& __GFP_THISNODE
)
3038 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3041 cpuset_mems_cookie
= read_mems_allowed_begin();
3042 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3046 * Look through allowed nodes for objects available
3047 * from existing per node queues.
3049 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3050 nid
= zone_to_nid(zone
);
3052 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3053 get_node(cache
, nid
) &&
3054 get_node(cache
, nid
)->free_objects
) {
3055 obj
= ____cache_alloc_node(cache
,
3056 flags
| GFP_THISNODE
, nid
);
3064 * This allocation will be performed within the constraints
3065 * of the current cpuset / memory policy requirements.
3066 * We may trigger various forms of reclaim on the allowed
3067 * set and go into memory reserves if necessary.
3071 if (local_flags
& __GFP_WAIT
)
3073 kmem_flagcheck(cache
, flags
);
3074 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3075 if (local_flags
& __GFP_WAIT
)
3076 local_irq_disable();
3079 * Insert into the appropriate per node queues
3081 nid
= page_to_nid(page
);
3082 if (cache_grow(cache
, flags
, nid
, page
)) {
3083 obj
= ____cache_alloc_node(cache
,
3084 flags
| GFP_THISNODE
, nid
);
3087 * Another processor may allocate the
3088 * objects in the slab since we are
3089 * not holding any locks.
3093 /* cache_grow already freed obj */
3099 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3105 * A interface to enable slab creation on nodeid
3107 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3110 struct list_head
*entry
;
3112 struct kmem_cache_node
*n
;
3116 VM_BUG_ON(nodeid
> num_online_nodes());
3117 n
= get_node(cachep
, nodeid
);
3122 spin_lock(&n
->list_lock
);
3123 entry
= n
->slabs_partial
.next
;
3124 if (entry
== &n
->slabs_partial
) {
3125 n
->free_touched
= 1;
3126 entry
= n
->slabs_free
.next
;
3127 if (entry
== &n
->slabs_free
)
3131 page
= list_entry(entry
, struct page
, lru
);
3132 check_spinlock_acquired_node(cachep
, nodeid
);
3134 STATS_INC_NODEALLOCS(cachep
);
3135 STATS_INC_ACTIVE(cachep
);
3136 STATS_SET_HIGH(cachep
);
3138 BUG_ON(page
->active
== cachep
->num
);
3140 obj
= slab_get_obj(cachep
, page
, nodeid
);
3142 /* move slabp to correct slabp list: */
3143 list_del(&page
->lru
);
3145 if (page
->active
== cachep
->num
)
3146 list_add(&page
->lru
, &n
->slabs_full
);
3148 list_add(&page
->lru
, &n
->slabs_partial
);
3150 spin_unlock(&n
->list_lock
);
3154 spin_unlock(&n
->list_lock
);
3155 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3159 return fallback_alloc(cachep
, flags
);
3165 static __always_inline
void *
3166 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3167 unsigned long caller
)
3169 unsigned long save_flags
;
3171 int slab_node
= numa_mem_id();
3173 flags
&= gfp_allowed_mask
;
3175 lockdep_trace_alloc(flags
);
3177 if (slab_should_failslab(cachep
, flags
))
3180 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3182 cache_alloc_debugcheck_before(cachep
, flags
);
3183 local_irq_save(save_flags
);
3185 if (nodeid
== NUMA_NO_NODE
)
3188 if (unlikely(!get_node(cachep
, nodeid
))) {
3189 /* Node not bootstrapped yet */
3190 ptr
= fallback_alloc(cachep
, flags
);
3194 if (nodeid
== slab_node
) {
3196 * Use the locally cached objects if possible.
3197 * However ____cache_alloc does not allow fallback
3198 * to other nodes. It may fail while we still have
3199 * objects on other nodes available.
3201 ptr
= ____cache_alloc(cachep
, flags
);
3205 /* ___cache_alloc_node can fall back to other nodes */
3206 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3208 local_irq_restore(save_flags
);
3209 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3210 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3214 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3215 if (unlikely(flags
& __GFP_ZERO
))
3216 memset(ptr
, 0, cachep
->object_size
);
3222 static __always_inline
void *
3223 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3227 if (current
->mempolicy
|| unlikely(current
->flags
& PF_SPREAD_SLAB
)) {
3228 objp
= alternate_node_alloc(cache
, flags
);
3232 objp
= ____cache_alloc(cache
, flags
);
3235 * We may just have run out of memory on the local node.
3236 * ____cache_alloc_node() knows how to locate memory on other nodes
3239 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3246 static __always_inline
void *
3247 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3249 return ____cache_alloc(cachep
, flags
);
3252 #endif /* CONFIG_NUMA */
3254 static __always_inline
void *
3255 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3257 unsigned long save_flags
;
3260 flags
&= gfp_allowed_mask
;
3262 lockdep_trace_alloc(flags
);
3264 if (slab_should_failslab(cachep
, flags
))
3267 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3269 cache_alloc_debugcheck_before(cachep
, flags
);
3270 local_irq_save(save_flags
);
3271 objp
= __do_cache_alloc(cachep
, flags
);
3272 local_irq_restore(save_flags
);
3273 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3274 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3279 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3280 if (unlikely(flags
& __GFP_ZERO
))
3281 memset(objp
, 0, cachep
->object_size
);
3288 * Caller needs to acquire correct kmem_cache_node's list_lock
3289 * @list: List of detached free slabs should be freed by caller
3291 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3292 int nr_objects
, int node
, struct list_head
*list
)
3295 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3297 for (i
= 0; i
< nr_objects
; i
++) {
3301 clear_obj_pfmemalloc(&objpp
[i
]);
3304 page
= virt_to_head_page(objp
);
3305 list_del(&page
->lru
);
3306 check_spinlock_acquired_node(cachep
, node
);
3307 slab_put_obj(cachep
, page
, objp
, node
);
3308 STATS_DEC_ACTIVE(cachep
);
3311 /* fixup slab chains */
3312 if (page
->active
== 0) {
3313 if (n
->free_objects
> n
->free_limit
) {
3314 n
->free_objects
-= cachep
->num
;
3315 list_add_tail(&page
->lru
, list
);
3317 list_add(&page
->lru
, &n
->slabs_free
);
3320 /* Unconditionally move a slab to the end of the
3321 * partial list on free - maximum time for the
3322 * other objects to be freed, too.
3324 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3329 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3332 struct kmem_cache_node
*n
;
3333 int node
= numa_mem_id();
3336 batchcount
= ac
->batchcount
;
3338 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3341 n
= get_node(cachep
, node
);
3342 spin_lock(&n
->list_lock
);
3344 struct array_cache
*shared_array
= n
->shared
;
3345 int max
= shared_array
->limit
- shared_array
->avail
;
3347 if (batchcount
> max
)
3349 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3350 ac
->entry
, sizeof(void *) * batchcount
);
3351 shared_array
->avail
+= batchcount
;
3356 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3361 struct list_head
*p
;
3363 p
= n
->slabs_free
.next
;
3364 while (p
!= &(n
->slabs_free
)) {
3367 page
= list_entry(p
, struct page
, lru
);
3368 BUG_ON(page
->active
);
3373 STATS_SET_FREEABLE(cachep
, i
);
3376 spin_unlock(&n
->list_lock
);
3377 slabs_destroy(cachep
, &list
);
3378 ac
->avail
-= batchcount
;
3379 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3383 * Release an obj back to its cache. If the obj has a constructed state, it must
3384 * be in this state _before_ it is released. Called with disabled ints.
3386 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3387 unsigned long caller
)
3389 struct array_cache
*ac
= cpu_cache_get(cachep
);
3392 kmemleak_free_recursive(objp
, cachep
->flags
);
3393 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3395 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3398 * Skip calling cache_free_alien() when the platform is not numa.
3399 * This will avoid cache misses that happen while accessing slabp (which
3400 * is per page memory reference) to get nodeid. Instead use a global
3401 * variable to skip the call, which is mostly likely to be present in
3404 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3407 if (likely(ac
->avail
< ac
->limit
)) {
3408 STATS_INC_FREEHIT(cachep
);
3410 STATS_INC_FREEMISS(cachep
);
3411 cache_flusharray(cachep
, ac
);
3414 ac_put_obj(cachep
, ac
, objp
);
3418 * kmem_cache_alloc - Allocate an object
3419 * @cachep: The cache to allocate from.
3420 * @flags: See kmalloc().
3422 * Allocate an object from this cache. The flags are only relevant
3423 * if the cache has no available objects.
3425 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3427 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3429 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3430 cachep
->object_size
, cachep
->size
, flags
);
3434 EXPORT_SYMBOL(kmem_cache_alloc
);
3436 #ifdef CONFIG_TRACING
3438 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3442 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3444 trace_kmalloc(_RET_IP_
, ret
,
3445 size
, cachep
->size
, flags
);
3448 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3453 * kmem_cache_alloc_node - Allocate an object on the specified node
3454 * @cachep: The cache to allocate from.
3455 * @flags: See kmalloc().
3456 * @nodeid: node number of the target node.
3458 * Identical to kmem_cache_alloc but it will allocate memory on the given
3459 * node, which can improve the performance for cpu bound structures.
3461 * Fallback to other node is possible if __GFP_THISNODE is not set.
3463 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3465 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3467 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3468 cachep
->object_size
, cachep
->size
,
3473 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3475 #ifdef CONFIG_TRACING
3476 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3483 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3485 trace_kmalloc_node(_RET_IP_
, ret
,
3490 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3493 static __always_inline
void *
3494 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3496 struct kmem_cache
*cachep
;
3498 cachep
= kmalloc_slab(size
, flags
);
3499 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3501 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3504 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3505 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3507 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3509 EXPORT_SYMBOL(__kmalloc_node
);
3511 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3512 int node
, unsigned long caller
)
3514 return __do_kmalloc_node(size
, flags
, node
, caller
);
3516 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3518 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3520 return __do_kmalloc_node(size
, flags
, node
, 0);
3522 EXPORT_SYMBOL(__kmalloc_node
);
3523 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3524 #endif /* CONFIG_NUMA */
3527 * __do_kmalloc - allocate memory
3528 * @size: how many bytes of memory are required.
3529 * @flags: the type of memory to allocate (see kmalloc).
3530 * @caller: function caller for debug tracking of the caller
3532 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3533 unsigned long caller
)
3535 struct kmem_cache
*cachep
;
3538 cachep
= kmalloc_slab(size
, flags
);
3539 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3541 ret
= slab_alloc(cachep
, flags
, caller
);
3543 trace_kmalloc(caller
, ret
,
3544 size
, cachep
->size
, flags
);
3550 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3551 void *__kmalloc(size_t size
, gfp_t flags
)
3553 return __do_kmalloc(size
, flags
, _RET_IP_
);
3555 EXPORT_SYMBOL(__kmalloc
);
3557 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3559 return __do_kmalloc(size
, flags
, caller
);
3561 EXPORT_SYMBOL(__kmalloc_track_caller
);
3564 void *__kmalloc(size_t size
, gfp_t flags
)
3566 return __do_kmalloc(size
, flags
, 0);
3568 EXPORT_SYMBOL(__kmalloc
);
3572 * kmem_cache_free - Deallocate an object
3573 * @cachep: The cache the allocation was from.
3574 * @objp: The previously allocated object.
3576 * Free an object which was previously allocated from this
3579 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3581 unsigned long flags
;
3582 cachep
= cache_from_obj(cachep
, objp
);
3586 local_irq_save(flags
);
3587 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3588 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3589 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3590 __cache_free(cachep
, objp
, _RET_IP_
);
3591 local_irq_restore(flags
);
3593 trace_kmem_cache_free(_RET_IP_
, objp
);
3595 EXPORT_SYMBOL(kmem_cache_free
);
3598 * kfree - free previously allocated memory
3599 * @objp: pointer returned by kmalloc.
3601 * If @objp is NULL, no operation is performed.
3603 * Don't free memory not originally allocated by kmalloc()
3604 * or you will run into trouble.
3606 void kfree(const void *objp
)
3608 struct kmem_cache
*c
;
3609 unsigned long flags
;
3611 trace_kfree(_RET_IP_
, objp
);
3613 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3615 local_irq_save(flags
);
3616 kfree_debugcheck(objp
);
3617 c
= virt_to_cache(objp
);
3618 debug_check_no_locks_freed(objp
, c
->object_size
);
3620 debug_check_no_obj_freed(objp
, c
->object_size
);
3621 __cache_free(c
, (void *)objp
, _RET_IP_
);
3622 local_irq_restore(flags
);
3624 EXPORT_SYMBOL(kfree
);
3627 * This initializes kmem_cache_node or resizes various caches for all nodes.
3629 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3632 struct kmem_cache_node
*n
;
3633 struct array_cache
*new_shared
;
3634 struct alien_cache
**new_alien
= NULL
;
3636 for_each_online_node(node
) {
3638 if (use_alien_caches
) {
3639 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3645 if (cachep
->shared
) {
3646 new_shared
= alloc_arraycache(node
,
3647 cachep
->shared
*cachep
->batchcount
,
3650 free_alien_cache(new_alien
);
3655 n
= get_node(cachep
, node
);
3657 struct array_cache
*shared
= n
->shared
;
3660 spin_lock_irq(&n
->list_lock
);
3663 free_block(cachep
, shared
->entry
,
3664 shared
->avail
, node
, &list
);
3666 n
->shared
= new_shared
;
3668 n
->alien
= new_alien
;
3671 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3672 cachep
->batchcount
+ cachep
->num
;
3673 spin_unlock_irq(&n
->list_lock
);
3674 slabs_destroy(cachep
, &list
);
3676 free_alien_cache(new_alien
);
3679 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3681 free_alien_cache(new_alien
);
3686 kmem_cache_node_init(n
);
3687 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3688 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3689 n
->shared
= new_shared
;
3690 n
->alien
= new_alien
;
3691 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3692 cachep
->batchcount
+ cachep
->num
;
3693 cachep
->node
[node
] = n
;
3698 if (!cachep
->list
.next
) {
3699 /* Cache is not active yet. Roll back what we did */
3702 n
= get_node(cachep
, node
);
3705 free_alien_cache(n
->alien
);
3707 cachep
->node
[node
] = NULL
;
3715 struct ccupdate_struct
{
3716 struct kmem_cache
*cachep
;
3717 struct array_cache
*new[0];
3720 static void do_ccupdate_local(void *info
)
3722 struct ccupdate_struct
*new = info
;
3723 struct array_cache
*old
;
3726 old
= cpu_cache_get(new->cachep
);
3728 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3729 new->new[smp_processor_id()] = old
;
3732 /* Always called with the slab_mutex held */
3733 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3734 int batchcount
, int shared
, gfp_t gfp
)
3736 struct ccupdate_struct
*new;
3739 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3744 for_each_online_cpu(i
) {
3745 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3748 for (i
--; i
>= 0; i
--)
3754 new->cachep
= cachep
;
3756 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3759 cachep
->batchcount
= batchcount
;
3760 cachep
->limit
= limit
;
3761 cachep
->shared
= shared
;
3763 for_each_online_cpu(i
) {
3765 struct array_cache
*ccold
= new->new[i
];
3767 struct kmem_cache_node
*n
;
3772 node
= cpu_to_mem(i
);
3773 n
= get_node(cachep
, node
);
3774 spin_lock_irq(&n
->list_lock
);
3775 free_block(cachep
, ccold
->entry
, ccold
->avail
, node
, &list
);
3776 spin_unlock_irq(&n
->list_lock
);
3777 slabs_destroy(cachep
, &list
);
3781 return alloc_kmem_cache_node(cachep
, gfp
);
3784 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3785 int batchcount
, int shared
, gfp_t gfp
)
3788 struct kmem_cache
*c
= NULL
;
3791 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3793 if (slab_state
< FULL
)
3796 if ((ret
< 0) || !is_root_cache(cachep
))
3799 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3800 for_each_memcg_cache_index(i
) {
3801 c
= cache_from_memcg_idx(cachep
, i
);
3803 /* return value determined by the parent cache only */
3804 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3810 /* Called with slab_mutex held always */
3811 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3818 if (!is_root_cache(cachep
)) {
3819 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3820 limit
= root
->limit
;
3821 shared
= root
->shared
;
3822 batchcount
= root
->batchcount
;
3825 if (limit
&& shared
&& batchcount
)
3828 * The head array serves three purposes:
3829 * - create a LIFO ordering, i.e. return objects that are cache-warm
3830 * - reduce the number of spinlock operations.
3831 * - reduce the number of linked list operations on the slab and
3832 * bufctl chains: array operations are cheaper.
3833 * The numbers are guessed, we should auto-tune as described by
3836 if (cachep
->size
> 131072)
3838 else if (cachep
->size
> PAGE_SIZE
)
3840 else if (cachep
->size
> 1024)
3842 else if (cachep
->size
> 256)
3848 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3849 * allocation behaviour: Most allocs on one cpu, most free operations
3850 * on another cpu. For these cases, an efficient object passing between
3851 * cpus is necessary. This is provided by a shared array. The array
3852 * replaces Bonwick's magazine layer.
3853 * On uniprocessor, it's functionally equivalent (but less efficient)
3854 * to a larger limit. Thus disabled by default.
3857 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3862 * With debugging enabled, large batchcount lead to excessively long
3863 * periods with disabled local interrupts. Limit the batchcount
3868 batchcount
= (limit
+ 1) / 2;
3870 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3872 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3873 cachep
->name
, -err
);
3878 * Drain an array if it contains any elements taking the node lock only if
3879 * necessary. Note that the node listlock also protects the array_cache
3880 * if drain_array() is used on the shared array.
3882 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3883 struct array_cache
*ac
, int force
, int node
)
3888 if (!ac
|| !ac
->avail
)
3890 if (ac
->touched
&& !force
) {
3893 spin_lock_irq(&n
->list_lock
);
3895 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3896 if (tofree
> ac
->avail
)
3897 tofree
= (ac
->avail
+ 1) / 2;
3898 free_block(cachep
, ac
->entry
, tofree
, node
, &list
);
3899 ac
->avail
-= tofree
;
3900 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3901 sizeof(void *) * ac
->avail
);
3903 spin_unlock_irq(&n
->list_lock
);
3904 slabs_destroy(cachep
, &list
);
3909 * cache_reap - Reclaim memory from caches.
3910 * @w: work descriptor
3912 * Called from workqueue/eventd every few seconds.
3914 * - clear the per-cpu caches for this CPU.
3915 * - return freeable pages to the main free memory pool.
3917 * If we cannot acquire the cache chain mutex then just give up - we'll try
3918 * again on the next iteration.
3920 static void cache_reap(struct work_struct
*w
)
3922 struct kmem_cache
*searchp
;
3923 struct kmem_cache_node
*n
;
3924 int node
= numa_mem_id();
3925 struct delayed_work
*work
= to_delayed_work(w
);
3927 if (!mutex_trylock(&slab_mutex
))
3928 /* Give up. Setup the next iteration. */
3931 list_for_each_entry(searchp
, &slab_caches
, list
) {
3935 * We only take the node lock if absolutely necessary and we
3936 * have established with reasonable certainty that
3937 * we can do some work if the lock was obtained.
3939 n
= get_node(searchp
, node
);
3941 reap_alien(searchp
, n
);
3943 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
3946 * These are racy checks but it does not matter
3947 * if we skip one check or scan twice.
3949 if (time_after(n
->next_reap
, jiffies
))
3952 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3954 drain_array(searchp
, n
, n
->shared
, 0, node
);
3956 if (n
->free_touched
)
3957 n
->free_touched
= 0;
3961 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3962 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3963 STATS_ADD_REAPED(searchp
, freed
);
3969 mutex_unlock(&slab_mutex
);
3972 /* Set up the next iteration */
3973 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3976 #ifdef CONFIG_SLABINFO
3977 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3980 unsigned long active_objs
;
3981 unsigned long num_objs
;
3982 unsigned long active_slabs
= 0;
3983 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3987 struct kmem_cache_node
*n
;
3991 for_each_kmem_cache_node(cachep
, node
, n
) {
3994 spin_lock_irq(&n
->list_lock
);
3996 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
3997 if (page
->active
!= cachep
->num
&& !error
)
3998 error
= "slabs_full accounting error";
3999 active_objs
+= cachep
->num
;
4002 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4003 if (page
->active
== cachep
->num
&& !error
)
4004 error
= "slabs_partial accounting error";
4005 if (!page
->active
&& !error
)
4006 error
= "slabs_partial accounting error";
4007 active_objs
+= page
->active
;
4010 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4011 if (page
->active
&& !error
)
4012 error
= "slabs_free accounting error";
4015 free_objects
+= n
->free_objects
;
4017 shared_avail
+= n
->shared
->avail
;
4019 spin_unlock_irq(&n
->list_lock
);
4021 num_slabs
+= active_slabs
;
4022 num_objs
= num_slabs
* cachep
->num
;
4023 if (num_objs
- active_objs
!= free_objects
&& !error
)
4024 error
= "free_objects accounting error";
4026 name
= cachep
->name
;
4028 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4030 sinfo
->active_objs
= active_objs
;
4031 sinfo
->num_objs
= num_objs
;
4032 sinfo
->active_slabs
= active_slabs
;
4033 sinfo
->num_slabs
= num_slabs
;
4034 sinfo
->shared_avail
= shared_avail
;
4035 sinfo
->limit
= cachep
->limit
;
4036 sinfo
->batchcount
= cachep
->batchcount
;
4037 sinfo
->shared
= cachep
->shared
;
4038 sinfo
->objects_per_slab
= cachep
->num
;
4039 sinfo
->cache_order
= cachep
->gfporder
;
4042 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4046 unsigned long high
= cachep
->high_mark
;
4047 unsigned long allocs
= cachep
->num_allocations
;
4048 unsigned long grown
= cachep
->grown
;
4049 unsigned long reaped
= cachep
->reaped
;
4050 unsigned long errors
= cachep
->errors
;
4051 unsigned long max_freeable
= cachep
->max_freeable
;
4052 unsigned long node_allocs
= cachep
->node_allocs
;
4053 unsigned long node_frees
= cachep
->node_frees
;
4054 unsigned long overflows
= cachep
->node_overflow
;
4056 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4057 "%4lu %4lu %4lu %4lu %4lu",
4058 allocs
, high
, grown
,
4059 reaped
, errors
, max_freeable
, node_allocs
,
4060 node_frees
, overflows
);
4064 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4065 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4066 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4067 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4069 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4070 allochit
, allocmiss
, freehit
, freemiss
);
4075 #define MAX_SLABINFO_WRITE 128
4077 * slabinfo_write - Tuning for the slab allocator
4079 * @buffer: user buffer
4080 * @count: data length
4083 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4084 size_t count
, loff_t
*ppos
)
4086 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4087 int limit
, batchcount
, shared
, res
;
4088 struct kmem_cache
*cachep
;
4090 if (count
> MAX_SLABINFO_WRITE
)
4092 if (copy_from_user(&kbuf
, buffer
, count
))
4094 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4096 tmp
= strchr(kbuf
, ' ');
4101 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4104 /* Find the cache in the chain of caches. */
4105 mutex_lock(&slab_mutex
);
4107 list_for_each_entry(cachep
, &slab_caches
, list
) {
4108 if (!strcmp(cachep
->name
, kbuf
)) {
4109 if (limit
< 1 || batchcount
< 1 ||
4110 batchcount
> limit
|| shared
< 0) {
4113 res
= do_tune_cpucache(cachep
, limit
,
4120 mutex_unlock(&slab_mutex
);
4126 #ifdef CONFIG_DEBUG_SLAB_LEAK
4128 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4130 mutex_lock(&slab_mutex
);
4131 return seq_list_start(&slab_caches
, *pos
);
4134 static inline int add_caller(unsigned long *n
, unsigned long v
)
4144 unsigned long *q
= p
+ 2 * i
;
4158 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4164 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4172 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4173 if (get_obj_status(page
, i
) != OBJECT_ACTIVE
)
4176 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4181 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4183 #ifdef CONFIG_KALLSYMS
4184 unsigned long offset
, size
;
4185 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4187 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4188 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4190 seq_printf(m
, " [%s]", modname
);
4194 seq_printf(m
, "%p", (void *)address
);
4197 static int leaks_show(struct seq_file
*m
, void *p
)
4199 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4201 struct kmem_cache_node
*n
;
4203 unsigned long *x
= m
->private;
4207 if (!(cachep
->flags
& SLAB_STORE_USER
))
4209 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4212 /* OK, we can do it */
4216 for_each_kmem_cache_node(cachep
, node
, n
) {
4219 spin_lock_irq(&n
->list_lock
);
4221 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4222 handle_slab(x
, cachep
, page
);
4223 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4224 handle_slab(x
, cachep
, page
);
4225 spin_unlock_irq(&n
->list_lock
);
4227 name
= cachep
->name
;
4229 /* Increase the buffer size */
4230 mutex_unlock(&slab_mutex
);
4231 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4233 /* Too bad, we are really out */
4235 mutex_lock(&slab_mutex
);
4238 *(unsigned long *)m
->private = x
[0] * 2;
4240 mutex_lock(&slab_mutex
);
4241 /* Now make sure this entry will be retried */
4245 for (i
= 0; i
< x
[1]; i
++) {
4246 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4247 show_symbol(m
, x
[2*i
+2]);
4254 static const struct seq_operations slabstats_op
= {
4255 .start
= leaks_start
,
4261 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4263 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4266 ret
= seq_open(file
, &slabstats_op
);
4268 struct seq_file
*m
= file
->private_data
;
4269 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4278 static const struct file_operations proc_slabstats_operations
= {
4279 .open
= slabstats_open
,
4281 .llseek
= seq_lseek
,
4282 .release
= seq_release_private
,
4286 static int __init
slab_proc_init(void)
4288 #ifdef CONFIG_DEBUG_SLAB_LEAK
4289 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4293 module_init(slab_proc_init
);
4297 * ksize - get the actual amount of memory allocated for a given object
4298 * @objp: Pointer to the object
4300 * kmalloc may internally round up allocations and return more memory
4301 * than requested. ksize() can be used to determine the actual amount of
4302 * memory allocated. The caller may use this additional memory, even though
4303 * a smaller amount of memory was initially specified with the kmalloc call.
4304 * The caller must guarantee that objp points to a valid object previously
4305 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4306 * must not be freed during the duration of the call.
4308 size_t ksize(const void *objp
)
4311 if (unlikely(objp
== ZERO_SIZE_PTR
))
4314 return virt_to_cache(objp
)->object_size
;
4316 EXPORT_SYMBOL(ksize
);