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)
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
179 * The limit is stored in the per-cpu structure to reduce the data cache
186 unsigned int batchcount
;
187 unsigned int touched
;
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
197 struct array_cache ac
;
201 * Need this for bootstrapping a per node allocator.
203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
205 #define CACHE_CACHE 0
206 #define SIZE_NODE (MAX_NUMNODES)
208 static int drain_freelist(struct kmem_cache
*cache
,
209 struct kmem_cache_node
*n
, int tofree
);
210 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
211 int node
, struct list_head
*list
);
212 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
213 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
214 static void cache_reap(struct work_struct
*unused
);
216 static int slab_early_init
= 1;
218 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
220 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
222 INIT_LIST_HEAD(&parent
->slabs_full
);
223 INIT_LIST_HEAD(&parent
->slabs_partial
);
224 INIT_LIST_HEAD(&parent
->slabs_free
);
225 parent
->shared
= NULL
;
226 parent
->alien
= NULL
;
227 parent
->colour_next
= 0;
228 spin_lock_init(&parent
->list_lock
);
229 parent
->free_objects
= 0;
230 parent
->free_touched
= 0;
233 #define MAKE_LIST(cachep, listp, slab, nodeid) \
235 INIT_LIST_HEAD(listp); \
236 list_splice(&get_node(cachep, nodeid)->slab, listp); \
239 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
241 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
242 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
243 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
246 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
247 #define CFLGS_OFF_SLAB (0x80000000UL)
248 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
249 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
251 #define BATCHREFILL_LIMIT 16
253 * Optimization question: fewer reaps means less probability for unnessary
254 * cpucache drain/refill cycles.
256 * OTOH the cpuarrays can contain lots of objects,
257 * which could lock up otherwise freeable slabs.
259 #define REAPTIMEOUT_AC (2*HZ)
260 #define REAPTIMEOUT_NODE (4*HZ)
263 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
264 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
265 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
266 #define STATS_INC_GROWN(x) ((x)->grown++)
267 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
268 #define STATS_SET_HIGH(x) \
270 if ((x)->num_active > (x)->high_mark) \
271 (x)->high_mark = (x)->num_active; \
273 #define STATS_INC_ERR(x) ((x)->errors++)
274 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
275 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
276 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
277 #define STATS_SET_FREEABLE(x, i) \
279 if ((x)->max_freeable < i) \
280 (x)->max_freeable = i; \
282 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
283 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
284 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
285 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
287 #define STATS_INC_ACTIVE(x) do { } while (0)
288 #define STATS_DEC_ACTIVE(x) do { } while (0)
289 #define STATS_INC_ALLOCED(x) do { } while (0)
290 #define STATS_INC_GROWN(x) do { } while (0)
291 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
292 #define STATS_SET_HIGH(x) do { } while (0)
293 #define STATS_INC_ERR(x) do { } while (0)
294 #define STATS_INC_NODEALLOCS(x) do { } while (0)
295 #define STATS_INC_NODEFREES(x) do { } while (0)
296 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
297 #define STATS_SET_FREEABLE(x, i) do { } while (0)
298 #define STATS_INC_ALLOCHIT(x) do { } while (0)
299 #define STATS_INC_ALLOCMISS(x) do { } while (0)
300 #define STATS_INC_FREEHIT(x) do { } while (0)
301 #define STATS_INC_FREEMISS(x) do { } while (0)
307 * memory layout of objects:
309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
310 * the end of an object is aligned with the end of the real
311 * allocation. Catches writes behind the end of the allocation.
312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
314 * cachep->obj_offset: The real object.
315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
316 * cachep->size - 1* BYTES_PER_WORD: last caller address
317 * [BYTES_PER_WORD long]
319 static int obj_offset(struct kmem_cache
*cachep
)
321 return cachep
->obj_offset
;
324 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
326 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
327 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
328 sizeof(unsigned long long));
331 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
333 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
334 if (cachep
->flags
& SLAB_STORE_USER
)
335 return (unsigned long long *)(objp
+ cachep
->size
-
336 sizeof(unsigned long long) -
338 return (unsigned long long *) (objp
+ cachep
->size
-
339 sizeof(unsigned long long));
342 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
344 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
345 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
350 #define obj_offset(x) 0
351 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
352 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
353 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
357 #ifdef CONFIG_DEBUG_SLAB_LEAK
359 static inline bool is_store_user_clean(struct kmem_cache
*cachep
)
361 return atomic_read(&cachep
->store_user_clean
) == 1;
364 static inline void set_store_user_clean(struct kmem_cache
*cachep
)
366 atomic_set(&cachep
->store_user_clean
, 1);
369 static inline void set_store_user_dirty(struct kmem_cache
*cachep
)
371 if (is_store_user_clean(cachep
))
372 atomic_set(&cachep
->store_user_clean
, 0);
376 static inline void set_store_user_dirty(struct kmem_cache
*cachep
) {}
381 * Do not go above this order unless 0 objects fit into the slab or
382 * overridden on the command line.
384 #define SLAB_MAX_ORDER_HI 1
385 #define SLAB_MAX_ORDER_LO 0
386 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
387 static bool slab_max_order_set __initdata
;
389 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
391 struct page
*page
= virt_to_head_page(obj
);
392 return page
->slab_cache
;
395 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
398 return page
->s_mem
+ cache
->size
* idx
;
402 * We want to avoid an expensive divide : (offset / cache->size)
403 * Using the fact that size is a constant for a particular cache,
404 * we can replace (offset / cache->size) by
405 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
407 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
408 const struct page
*page
, void *obj
)
410 u32 offset
= (obj
- page
->s_mem
);
411 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
414 #define BOOT_CPUCACHE_ENTRIES 1
415 /* internal cache of cache description objs */
416 static struct kmem_cache kmem_cache_boot
= {
418 .limit
= BOOT_CPUCACHE_ENTRIES
,
420 .size
= sizeof(struct kmem_cache
),
421 .name
= "kmem_cache",
424 #define BAD_ALIEN_MAGIC 0x01020304ul
426 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
428 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
430 return this_cpu_ptr(cachep
->cpu_cache
);
434 * Calculate the number of objects and left-over bytes for a given buffer size.
436 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
437 unsigned long flags
, size_t *left_over
)
440 size_t slab_size
= PAGE_SIZE
<< gfporder
;
443 * The slab management structure can be either off the slab or
444 * on it. For the latter case, the memory allocated for a
447 * - @buffer_size bytes for each object
448 * - One freelist_idx_t for each object
450 * We don't need to consider alignment of freelist because
451 * freelist will be at the end of slab page. The objects will be
452 * at the correct alignment.
454 * If the slab management structure is off the slab, then the
455 * alignment will already be calculated into the size. Because
456 * the slabs are all pages aligned, the objects will be at the
457 * correct alignment when allocated.
459 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
460 num
= slab_size
/ buffer_size
;
461 *left_over
= slab_size
% buffer_size
;
463 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
464 *left_over
= slab_size
%
465 (buffer_size
+ sizeof(freelist_idx_t
));
472 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
474 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
477 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
478 function
, cachep
->name
, msg
);
480 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
485 * By default on NUMA we use alien caches to stage the freeing of
486 * objects allocated from other nodes. This causes massive memory
487 * inefficiencies when using fake NUMA setup to split memory into a
488 * large number of small nodes, so it can be disabled on the command
492 static int use_alien_caches __read_mostly
= 1;
493 static int __init
noaliencache_setup(char *s
)
495 use_alien_caches
= 0;
498 __setup("noaliencache", noaliencache_setup
);
500 static int __init
slab_max_order_setup(char *str
)
502 get_option(&str
, &slab_max_order
);
503 slab_max_order
= slab_max_order
< 0 ? 0 :
504 min(slab_max_order
, MAX_ORDER
- 1);
505 slab_max_order_set
= true;
509 __setup("slab_max_order=", slab_max_order_setup
);
513 * Special reaping functions for NUMA systems called from cache_reap().
514 * These take care of doing round robin flushing of alien caches (containing
515 * objects freed on different nodes from which they were allocated) and the
516 * flushing of remote pcps by calling drain_node_pages.
518 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
520 static void init_reap_node(int cpu
)
524 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
525 if (node
== MAX_NUMNODES
)
526 node
= first_node(node_online_map
);
528 per_cpu(slab_reap_node
, cpu
) = node
;
531 static void next_reap_node(void)
533 int node
= __this_cpu_read(slab_reap_node
);
535 node
= next_node(node
, node_online_map
);
536 if (unlikely(node
>= MAX_NUMNODES
))
537 node
= first_node(node_online_map
);
538 __this_cpu_write(slab_reap_node
, node
);
542 #define init_reap_node(cpu) do { } while (0)
543 #define next_reap_node(void) do { } while (0)
547 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
548 * via the workqueue/eventd.
549 * Add the CPU number into the expiration time to minimize the possibility of
550 * the CPUs getting into lockstep and contending for the global cache chain
553 static void start_cpu_timer(int cpu
)
555 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
558 * When this gets called from do_initcalls via cpucache_init(),
559 * init_workqueues() has already run, so keventd will be setup
562 if (keventd_up() && reap_work
->work
.func
== NULL
) {
564 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
565 schedule_delayed_work_on(cpu
, reap_work
,
566 __round_jiffies_relative(HZ
, cpu
));
570 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
573 * The array_cache structures contain pointers to free object.
574 * However, when such objects are allocated or transferred to another
575 * cache the pointers are not cleared and they could be counted as
576 * valid references during a kmemleak scan. Therefore, kmemleak must
577 * not scan such objects.
579 kmemleak_no_scan(ac
);
583 ac
->batchcount
= batch
;
588 static struct array_cache
*alloc_arraycache(int node
, int entries
,
589 int batchcount
, gfp_t gfp
)
591 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
592 struct array_cache
*ac
= NULL
;
594 ac
= kmalloc_node(memsize
, gfp
, node
);
595 init_arraycache(ac
, entries
, batchcount
);
599 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
600 struct page
*page
, void *objp
)
602 struct kmem_cache_node
*n
;
606 page_node
= page_to_nid(page
);
607 n
= get_node(cachep
, page_node
);
609 spin_lock(&n
->list_lock
);
610 free_block(cachep
, &objp
, 1, page_node
, &list
);
611 spin_unlock(&n
->list_lock
);
613 slabs_destroy(cachep
, &list
);
617 * Transfer objects in one arraycache to another.
618 * Locking must be handled by the caller.
620 * Return the number of entries transferred.
622 static int transfer_objects(struct array_cache
*to
,
623 struct array_cache
*from
, unsigned int max
)
625 /* Figure out how many entries to transfer */
626 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
631 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
641 #define drain_alien_cache(cachep, alien) do { } while (0)
642 #define reap_alien(cachep, n) do { } while (0)
644 static inline struct alien_cache
**alloc_alien_cache(int node
,
645 int limit
, gfp_t gfp
)
647 return (struct alien_cache
**)BAD_ALIEN_MAGIC
;
650 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
654 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
659 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
665 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
666 gfp_t flags
, int nodeid
)
671 static inline gfp_t
gfp_exact_node(gfp_t flags
)
676 #else /* CONFIG_NUMA */
678 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
679 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
681 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
682 int batch
, gfp_t gfp
)
684 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
685 struct alien_cache
*alc
= NULL
;
687 alc
= kmalloc_node(memsize
, gfp
, node
);
688 init_arraycache(&alc
->ac
, entries
, batch
);
689 spin_lock_init(&alc
->lock
);
693 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
695 struct alien_cache
**alc_ptr
;
696 size_t memsize
= sizeof(void *) * nr_node_ids
;
701 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
706 if (i
== node
|| !node_online(i
))
708 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
710 for (i
--; i
>= 0; i
--)
719 static void free_alien_cache(struct alien_cache
**alc_ptr
)
730 static void __drain_alien_cache(struct kmem_cache
*cachep
,
731 struct array_cache
*ac
, int node
,
732 struct list_head
*list
)
734 struct kmem_cache_node
*n
= get_node(cachep
, node
);
737 spin_lock(&n
->list_lock
);
739 * Stuff objects into the remote nodes shared array first.
740 * That way we could avoid the overhead of putting the objects
741 * into the free lists and getting them back later.
744 transfer_objects(n
->shared
, ac
, ac
->limit
);
746 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
748 spin_unlock(&n
->list_lock
);
753 * Called from cache_reap() to regularly drain alien caches round robin.
755 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
757 int node
= __this_cpu_read(slab_reap_node
);
760 struct alien_cache
*alc
= n
->alien
[node
];
761 struct array_cache
*ac
;
765 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
768 __drain_alien_cache(cachep
, ac
, node
, &list
);
769 spin_unlock_irq(&alc
->lock
);
770 slabs_destroy(cachep
, &list
);
776 static void drain_alien_cache(struct kmem_cache
*cachep
,
777 struct alien_cache
**alien
)
780 struct alien_cache
*alc
;
781 struct array_cache
*ac
;
784 for_each_online_node(i
) {
790 spin_lock_irqsave(&alc
->lock
, flags
);
791 __drain_alien_cache(cachep
, ac
, i
, &list
);
792 spin_unlock_irqrestore(&alc
->lock
, flags
);
793 slabs_destroy(cachep
, &list
);
798 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
799 int node
, int page_node
)
801 struct kmem_cache_node
*n
;
802 struct alien_cache
*alien
= NULL
;
803 struct array_cache
*ac
;
806 n
= get_node(cachep
, node
);
807 STATS_INC_NODEFREES(cachep
);
808 if (n
->alien
&& n
->alien
[page_node
]) {
809 alien
= n
->alien
[page_node
];
811 spin_lock(&alien
->lock
);
812 if (unlikely(ac
->avail
== ac
->limit
)) {
813 STATS_INC_ACOVERFLOW(cachep
);
814 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
816 ac
->entry
[ac
->avail
++] = objp
;
817 spin_unlock(&alien
->lock
);
818 slabs_destroy(cachep
, &list
);
820 n
= get_node(cachep
, page_node
);
821 spin_lock(&n
->list_lock
);
822 free_block(cachep
, &objp
, 1, page_node
, &list
);
823 spin_unlock(&n
->list_lock
);
824 slabs_destroy(cachep
, &list
);
829 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
831 int page_node
= page_to_nid(virt_to_page(objp
));
832 int node
= numa_mem_id();
834 * Make sure we are not freeing a object from another node to the array
837 if (likely(node
== page_node
))
840 return __cache_free_alien(cachep
, objp
, node
, page_node
);
844 * Construct gfp mask to allocate from a specific node but do not direct reclaim
845 * or warn about failures. kswapd may still wake to reclaim in the background.
847 static inline gfp_t
gfp_exact_node(gfp_t flags
)
849 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~__GFP_DIRECT_RECLAIM
;
854 * Allocates and initializes node for a node on each slab cache, used for
855 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
856 * will be allocated off-node since memory is not yet online for the new node.
857 * When hotplugging memory or a cpu, existing node are not replaced if
860 * Must hold slab_mutex.
862 static int init_cache_node_node(int node
)
864 struct kmem_cache
*cachep
;
865 struct kmem_cache_node
*n
;
866 const size_t memsize
= sizeof(struct kmem_cache_node
);
868 list_for_each_entry(cachep
, &slab_caches
, list
) {
870 * Set up the kmem_cache_node for cpu before we can
871 * begin anything. Make sure some other cpu on this
872 * node has not already allocated this
874 n
= get_node(cachep
, node
);
876 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
879 kmem_cache_node_init(n
);
880 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
881 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
884 * The kmem_cache_nodes don't come and go as CPUs
885 * come and go. slab_mutex is sufficient
888 cachep
->node
[node
] = n
;
891 spin_lock_irq(&n
->list_lock
);
893 (1 + nr_cpus_node(node
)) *
894 cachep
->batchcount
+ cachep
->num
;
895 spin_unlock_irq(&n
->list_lock
);
900 static inline int slabs_tofree(struct kmem_cache
*cachep
,
901 struct kmem_cache_node
*n
)
903 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
906 static void cpuup_canceled(long cpu
)
908 struct kmem_cache
*cachep
;
909 struct kmem_cache_node
*n
= NULL
;
910 int node
= cpu_to_mem(cpu
);
911 const struct cpumask
*mask
= cpumask_of_node(node
);
913 list_for_each_entry(cachep
, &slab_caches
, list
) {
914 struct array_cache
*nc
;
915 struct array_cache
*shared
;
916 struct alien_cache
**alien
;
919 n
= get_node(cachep
, node
);
923 spin_lock_irq(&n
->list_lock
);
925 /* Free limit for this kmem_cache_node */
926 n
->free_limit
-= cachep
->batchcount
;
928 /* cpu is dead; no one can alloc from it. */
929 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
931 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
935 if (!cpumask_empty(mask
)) {
936 spin_unlock_irq(&n
->list_lock
);
942 free_block(cachep
, shared
->entry
,
943 shared
->avail
, node
, &list
);
950 spin_unlock_irq(&n
->list_lock
);
954 drain_alien_cache(cachep
, alien
);
955 free_alien_cache(alien
);
959 slabs_destroy(cachep
, &list
);
962 * In the previous loop, all the objects were freed to
963 * the respective cache's slabs, now we can go ahead and
964 * shrink each nodelist to its limit.
966 list_for_each_entry(cachep
, &slab_caches
, list
) {
967 n
= get_node(cachep
, node
);
970 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
974 static int cpuup_prepare(long cpu
)
976 struct kmem_cache
*cachep
;
977 struct kmem_cache_node
*n
= NULL
;
978 int node
= cpu_to_mem(cpu
);
982 * We need to do this right in the beginning since
983 * alloc_arraycache's are going to use this list.
984 * kmalloc_node allows us to add the slab to the right
985 * kmem_cache_node and not this cpu's kmem_cache_node
987 err
= init_cache_node_node(node
);
992 * Now we can go ahead with allocating the shared arrays and
995 list_for_each_entry(cachep
, &slab_caches
, list
) {
996 struct array_cache
*shared
= NULL
;
997 struct alien_cache
**alien
= NULL
;
999 if (cachep
->shared
) {
1000 shared
= alloc_arraycache(node
,
1001 cachep
->shared
* cachep
->batchcount
,
1002 0xbaadf00d, GFP_KERNEL
);
1006 if (use_alien_caches
) {
1007 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1013 n
= get_node(cachep
, node
);
1016 spin_lock_irq(&n
->list_lock
);
1019 * We are serialised from CPU_DEAD or
1020 * CPU_UP_CANCELLED by the cpucontrol lock
1031 spin_unlock_irq(&n
->list_lock
);
1033 free_alien_cache(alien
);
1038 cpuup_canceled(cpu
);
1042 static int cpuup_callback(struct notifier_block
*nfb
,
1043 unsigned long action
, void *hcpu
)
1045 long cpu
= (long)hcpu
;
1049 case CPU_UP_PREPARE
:
1050 case CPU_UP_PREPARE_FROZEN
:
1051 mutex_lock(&slab_mutex
);
1052 err
= cpuup_prepare(cpu
);
1053 mutex_unlock(&slab_mutex
);
1056 case CPU_ONLINE_FROZEN
:
1057 start_cpu_timer(cpu
);
1059 #ifdef CONFIG_HOTPLUG_CPU
1060 case CPU_DOWN_PREPARE
:
1061 case CPU_DOWN_PREPARE_FROZEN
:
1063 * Shutdown cache reaper. Note that the slab_mutex is
1064 * held so that if cache_reap() is invoked it cannot do
1065 * anything expensive but will only modify reap_work
1066 * and reschedule the timer.
1068 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1069 /* Now the cache_reaper is guaranteed to be not running. */
1070 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1072 case CPU_DOWN_FAILED
:
1073 case CPU_DOWN_FAILED_FROZEN
:
1074 start_cpu_timer(cpu
);
1077 case CPU_DEAD_FROZEN
:
1079 * Even if all the cpus of a node are down, we don't free the
1080 * kmem_cache_node of any cache. This to avoid a race between
1081 * cpu_down, and a kmalloc allocation from another cpu for
1082 * memory from the node of the cpu going down. The node
1083 * structure is usually allocated from kmem_cache_create() and
1084 * gets destroyed at kmem_cache_destroy().
1088 case CPU_UP_CANCELED
:
1089 case CPU_UP_CANCELED_FROZEN
:
1090 mutex_lock(&slab_mutex
);
1091 cpuup_canceled(cpu
);
1092 mutex_unlock(&slab_mutex
);
1095 return notifier_from_errno(err
);
1098 static struct notifier_block cpucache_notifier
= {
1099 &cpuup_callback
, NULL
, 0
1102 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1104 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1105 * Returns -EBUSY if all objects cannot be drained so that the node is not
1108 * Must hold slab_mutex.
1110 static int __meminit
drain_cache_node_node(int node
)
1112 struct kmem_cache
*cachep
;
1115 list_for_each_entry(cachep
, &slab_caches
, list
) {
1116 struct kmem_cache_node
*n
;
1118 n
= get_node(cachep
, node
);
1122 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1124 if (!list_empty(&n
->slabs_full
) ||
1125 !list_empty(&n
->slabs_partial
)) {
1133 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1134 unsigned long action
, void *arg
)
1136 struct memory_notify
*mnb
= arg
;
1140 nid
= mnb
->status_change_nid
;
1145 case MEM_GOING_ONLINE
:
1146 mutex_lock(&slab_mutex
);
1147 ret
= init_cache_node_node(nid
);
1148 mutex_unlock(&slab_mutex
);
1150 case MEM_GOING_OFFLINE
:
1151 mutex_lock(&slab_mutex
);
1152 ret
= drain_cache_node_node(nid
);
1153 mutex_unlock(&slab_mutex
);
1157 case MEM_CANCEL_ONLINE
:
1158 case MEM_CANCEL_OFFLINE
:
1162 return notifier_from_errno(ret
);
1164 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1167 * swap the static kmem_cache_node with kmalloced memory
1169 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1172 struct kmem_cache_node
*ptr
;
1174 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1177 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1179 * Do not assume that spinlocks can be initialized via memcpy:
1181 spin_lock_init(&ptr
->list_lock
);
1183 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1184 cachep
->node
[nodeid
] = ptr
;
1188 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1189 * size of kmem_cache_node.
1191 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1195 for_each_online_node(node
) {
1196 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1197 cachep
->node
[node
]->next_reap
= jiffies
+
1199 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1204 * Initialisation. Called after the page allocator have been initialised and
1205 * before smp_init().
1207 void __init
kmem_cache_init(void)
1211 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1212 sizeof(struct rcu_head
));
1213 kmem_cache
= &kmem_cache_boot
;
1215 if (num_possible_nodes() == 1)
1216 use_alien_caches
= 0;
1218 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1219 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1222 * Fragmentation resistance on low memory - only use bigger
1223 * page orders on machines with more than 32MB of memory if
1224 * not overridden on the command line.
1226 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1227 slab_max_order
= SLAB_MAX_ORDER_HI
;
1229 /* Bootstrap is tricky, because several objects are allocated
1230 * from caches that do not exist yet:
1231 * 1) initialize the kmem_cache cache: it contains the struct
1232 * kmem_cache structures of all caches, except kmem_cache itself:
1233 * kmem_cache is statically allocated.
1234 * Initially an __init data area is used for the head array and the
1235 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1236 * array at the end of the bootstrap.
1237 * 2) Create the first kmalloc cache.
1238 * The struct kmem_cache for the new cache is allocated normally.
1239 * An __init data area is used for the head array.
1240 * 3) Create the remaining kmalloc caches, with minimally sized
1242 * 4) Replace the __init data head arrays for kmem_cache and the first
1243 * kmalloc cache with kmalloc allocated arrays.
1244 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1245 * the other cache's with kmalloc allocated memory.
1246 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1249 /* 1) create the kmem_cache */
1252 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1254 create_boot_cache(kmem_cache
, "kmem_cache",
1255 offsetof(struct kmem_cache
, node
) +
1256 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1257 SLAB_HWCACHE_ALIGN
);
1258 list_add(&kmem_cache
->list
, &slab_caches
);
1259 slab_state
= PARTIAL
;
1262 * Initialize the caches that provide memory for the kmem_cache_node
1263 * structures first. Without this, further allocations will bug.
1265 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1266 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1267 slab_state
= PARTIAL_NODE
;
1268 setup_kmalloc_cache_index_table();
1270 slab_early_init
= 0;
1272 /* 5) Replace the bootstrap kmem_cache_node */
1276 for_each_online_node(nid
) {
1277 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1279 init_list(kmalloc_caches
[INDEX_NODE
],
1280 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1284 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1287 void __init
kmem_cache_init_late(void)
1289 struct kmem_cache
*cachep
;
1293 /* 6) resize the head arrays to their final sizes */
1294 mutex_lock(&slab_mutex
);
1295 list_for_each_entry(cachep
, &slab_caches
, list
)
1296 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1298 mutex_unlock(&slab_mutex
);
1304 * Register a cpu startup notifier callback that initializes
1305 * cpu_cache_get for all new cpus
1307 register_cpu_notifier(&cpucache_notifier
);
1311 * Register a memory hotplug callback that initializes and frees
1314 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1318 * The reap timers are started later, with a module init call: That part
1319 * of the kernel is not yet operational.
1323 static int __init
cpucache_init(void)
1328 * Register the timers that return unneeded pages to the page allocator
1330 for_each_online_cpu(cpu
)
1331 start_cpu_timer(cpu
);
1337 __initcall(cpucache_init
);
1339 static noinline
void
1340 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1343 struct kmem_cache_node
*n
;
1345 unsigned long flags
;
1347 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1348 DEFAULT_RATELIMIT_BURST
);
1350 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1353 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1354 nodeid
, gfpflags
, &gfpflags
);
1355 pr_warn(" cache: %s, object size: %d, order: %d\n",
1356 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1358 for_each_kmem_cache_node(cachep
, node
, n
) {
1359 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1360 unsigned long active_slabs
= 0, num_slabs
= 0;
1362 spin_lock_irqsave(&n
->list_lock
, flags
);
1363 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1364 active_objs
+= cachep
->num
;
1367 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1368 active_objs
+= page
->active
;
1371 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1374 free_objects
+= n
->free_objects
;
1375 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1377 num_slabs
+= active_slabs
;
1378 num_objs
= num_slabs
* cachep
->num
;
1379 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1380 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1387 * Interface to system's page allocator. No need to hold the
1388 * kmem_cache_node ->list_lock.
1390 * If we requested dmaable memory, we will get it. Even if we
1391 * did not request dmaable memory, we might get it, but that
1392 * would be relatively rare and ignorable.
1394 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1400 flags
|= cachep
->allocflags
;
1401 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1402 flags
|= __GFP_RECLAIMABLE
;
1404 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1406 slab_out_of_memory(cachep
, flags
, nodeid
);
1410 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1411 __free_pages(page
, cachep
->gfporder
);
1415 nr_pages
= (1 << cachep
->gfporder
);
1416 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1417 add_zone_page_state(page_zone(page
),
1418 NR_SLAB_RECLAIMABLE
, nr_pages
);
1420 add_zone_page_state(page_zone(page
),
1421 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1423 __SetPageSlab(page
);
1424 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1425 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1426 SetPageSlabPfmemalloc(page
);
1428 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1429 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1432 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1434 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1441 * Interface to system's page release.
1443 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1445 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1447 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1449 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1450 sub_zone_page_state(page_zone(page
),
1451 NR_SLAB_RECLAIMABLE
, nr_freed
);
1453 sub_zone_page_state(page_zone(page
),
1454 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1456 BUG_ON(!PageSlab(page
));
1457 __ClearPageSlabPfmemalloc(page
);
1458 __ClearPageSlab(page
);
1459 page_mapcount_reset(page
);
1460 page
->mapping
= NULL
;
1462 if (current
->reclaim_state
)
1463 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1464 __free_kmem_pages(page
, cachep
->gfporder
);
1467 static void kmem_rcu_free(struct rcu_head
*head
)
1469 struct kmem_cache
*cachep
;
1472 page
= container_of(head
, struct page
, rcu_head
);
1473 cachep
= page
->slab_cache
;
1475 kmem_freepages(cachep
, page
);
1479 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1481 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1482 (cachep
->size
% PAGE_SIZE
) == 0)
1488 #ifdef CONFIG_DEBUG_PAGEALLOC
1489 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1490 unsigned long caller
)
1492 int size
= cachep
->object_size
;
1494 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1496 if (size
< 5 * sizeof(unsigned long))
1499 *addr
++ = 0x12345678;
1501 *addr
++ = smp_processor_id();
1502 size
-= 3 * sizeof(unsigned long);
1504 unsigned long *sptr
= &caller
;
1505 unsigned long svalue
;
1507 while (!kstack_end(sptr
)) {
1509 if (kernel_text_address(svalue
)) {
1511 size
-= sizeof(unsigned long);
1512 if (size
<= sizeof(unsigned long))
1518 *addr
++ = 0x87654321;
1521 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1522 int map
, unsigned long caller
)
1524 if (!is_debug_pagealloc_cache(cachep
))
1528 store_stackinfo(cachep
, objp
, caller
);
1530 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1534 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1535 int map
, unsigned long caller
) {}
1539 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1541 int size
= cachep
->object_size
;
1542 addr
= &((char *)addr
)[obj_offset(cachep
)];
1544 memset(addr
, val
, size
);
1545 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1548 static void dump_line(char *data
, int offset
, int limit
)
1551 unsigned char error
= 0;
1554 printk(KERN_ERR
"%03x: ", offset
);
1555 for (i
= 0; i
< limit
; i
++) {
1556 if (data
[offset
+ i
] != POISON_FREE
) {
1557 error
= data
[offset
+ i
];
1561 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1562 &data
[offset
], limit
, 1);
1564 if (bad_count
== 1) {
1565 error
^= POISON_FREE
;
1566 if (!(error
& (error
- 1))) {
1567 printk(KERN_ERR
"Single bit error detected. Probably "
1570 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1573 printk(KERN_ERR
"Run a memory test tool.\n");
1582 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1587 if (cachep
->flags
& SLAB_RED_ZONE
) {
1588 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1589 *dbg_redzone1(cachep
, objp
),
1590 *dbg_redzone2(cachep
, objp
));
1593 if (cachep
->flags
& SLAB_STORE_USER
) {
1594 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1595 *dbg_userword(cachep
, objp
),
1596 *dbg_userword(cachep
, objp
));
1598 realobj
= (char *)objp
+ obj_offset(cachep
);
1599 size
= cachep
->object_size
;
1600 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1603 if (i
+ limit
> size
)
1605 dump_line(realobj
, i
, limit
);
1609 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1615 if (is_debug_pagealloc_cache(cachep
))
1618 realobj
= (char *)objp
+ obj_offset(cachep
);
1619 size
= cachep
->object_size
;
1621 for (i
= 0; i
< size
; i
++) {
1622 char exp
= POISON_FREE
;
1625 if (realobj
[i
] != exp
) {
1631 "Slab corruption (%s): %s start=%p, len=%d\n",
1632 print_tainted(), cachep
->name
, realobj
, size
);
1633 print_objinfo(cachep
, objp
, 0);
1635 /* Hexdump the affected line */
1638 if (i
+ limit
> size
)
1640 dump_line(realobj
, i
, limit
);
1643 /* Limit to 5 lines */
1649 /* Print some data about the neighboring objects, if they
1652 struct page
*page
= virt_to_head_page(objp
);
1655 objnr
= obj_to_index(cachep
, page
, objp
);
1657 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1658 realobj
= (char *)objp
+ obj_offset(cachep
);
1659 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1661 print_objinfo(cachep
, objp
, 2);
1663 if (objnr
+ 1 < cachep
->num
) {
1664 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1665 realobj
= (char *)objp
+ obj_offset(cachep
);
1666 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1668 print_objinfo(cachep
, objp
, 2);
1675 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1680 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1681 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1685 for (i
= 0; i
< cachep
->num
; i
++) {
1686 void *objp
= index_to_obj(cachep
, page
, i
);
1688 if (cachep
->flags
& SLAB_POISON
) {
1689 check_poison_obj(cachep
, objp
);
1690 slab_kernel_map(cachep
, objp
, 1, 0);
1692 if (cachep
->flags
& SLAB_RED_ZONE
) {
1693 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1694 slab_error(cachep
, "start of a freed object "
1696 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1697 slab_error(cachep
, "end of a freed object "
1703 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1710 * slab_destroy - destroy and release all objects in a slab
1711 * @cachep: cache pointer being destroyed
1712 * @page: page pointer being destroyed
1714 * Destroy all the objs in a slab page, and release the mem back to the system.
1715 * Before calling the slab page must have been unlinked from the cache. The
1716 * kmem_cache_node ->list_lock is not held/needed.
1718 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1722 freelist
= page
->freelist
;
1723 slab_destroy_debugcheck(cachep
, page
);
1724 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1725 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1727 kmem_freepages(cachep
, page
);
1730 * From now on, we don't use freelist
1731 * although actual page can be freed in rcu context
1733 if (OFF_SLAB(cachep
))
1734 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1737 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1739 struct page
*page
, *n
;
1741 list_for_each_entry_safe(page
, n
, list
, lru
) {
1742 list_del(&page
->lru
);
1743 slab_destroy(cachep
, page
);
1748 * calculate_slab_order - calculate size (page order) of slabs
1749 * @cachep: pointer to the cache that is being created
1750 * @size: size of objects to be created in this cache.
1751 * @flags: slab allocation flags
1753 * Also calculates the number of objects per slab.
1755 * This could be made much more intelligent. For now, try to avoid using
1756 * high order pages for slabs. When the gfp() functions are more friendly
1757 * towards high-order requests, this should be changed.
1759 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1760 size_t size
, unsigned long flags
)
1762 size_t left_over
= 0;
1765 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1769 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1773 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1774 if (num
> SLAB_OBJ_MAX_NUM
)
1777 if (flags
& CFLGS_OFF_SLAB
) {
1778 struct kmem_cache
*freelist_cache
;
1779 size_t freelist_size
;
1781 freelist_size
= num
* sizeof(freelist_idx_t
);
1782 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1783 if (!freelist_cache
)
1787 * Needed to avoid possible looping condition
1790 if (OFF_SLAB(freelist_cache
))
1793 /* check if off slab has enough benefit */
1794 if (freelist_cache
->size
> cachep
->size
/ 2)
1798 /* Found something acceptable - save it away */
1800 cachep
->gfporder
= gfporder
;
1801 left_over
= remainder
;
1804 * A VFS-reclaimable slab tends to have most allocations
1805 * as GFP_NOFS and we really don't want to have to be allocating
1806 * higher-order pages when we are unable to shrink dcache.
1808 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1812 * Large number of objects is good, but very large slabs are
1813 * currently bad for the gfp()s.
1815 if (gfporder
>= slab_max_order
)
1819 * Acceptable internal fragmentation?
1821 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1827 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1828 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1832 struct array_cache __percpu
*cpu_cache
;
1834 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1835 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1840 for_each_possible_cpu(cpu
) {
1841 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1842 entries
, batchcount
);
1848 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1850 if (slab_state
>= FULL
)
1851 return enable_cpucache(cachep
, gfp
);
1853 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1854 if (!cachep
->cpu_cache
)
1857 if (slab_state
== DOWN
) {
1858 /* Creation of first cache (kmem_cache). */
1859 set_up_node(kmem_cache
, CACHE_CACHE
);
1860 } else if (slab_state
== PARTIAL
) {
1861 /* For kmem_cache_node */
1862 set_up_node(cachep
, SIZE_NODE
);
1866 for_each_online_node(node
) {
1867 cachep
->node
[node
] = kmalloc_node(
1868 sizeof(struct kmem_cache_node
), gfp
, node
);
1869 BUG_ON(!cachep
->node
[node
]);
1870 kmem_cache_node_init(cachep
->node
[node
]);
1874 cachep
->node
[numa_mem_id()]->next_reap
=
1875 jiffies
+ REAPTIMEOUT_NODE
+
1876 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1878 cpu_cache_get(cachep
)->avail
= 0;
1879 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1880 cpu_cache_get(cachep
)->batchcount
= 1;
1881 cpu_cache_get(cachep
)->touched
= 0;
1882 cachep
->batchcount
= 1;
1883 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1887 unsigned long kmem_cache_flags(unsigned long object_size
,
1888 unsigned long flags
, const char *name
,
1889 void (*ctor
)(void *))
1895 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1896 unsigned long flags
, void (*ctor
)(void *))
1898 struct kmem_cache
*cachep
;
1900 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1905 * Adjust the object sizes so that we clear
1906 * the complete object on kzalloc.
1908 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1913 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1914 size_t size
, unsigned long flags
)
1920 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1923 left
= calculate_slab_order(cachep
, size
,
1924 flags
| CFLGS_OBJFREELIST_SLAB
);
1928 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1931 cachep
->colour
= left
/ cachep
->colour_off
;
1936 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1937 size_t size
, unsigned long flags
)
1944 * Always use on-slab management when SLAB_NOLEAKTRACE
1945 * to avoid recursive calls into kmemleak.
1947 if (flags
& SLAB_NOLEAKTRACE
)
1951 * Size is large, assume best to place the slab management obj
1952 * off-slab (should allow better packing of objs).
1954 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1959 * If the slab has been placed off-slab, and we have enough space then
1960 * move it on-slab. This is at the expense of any extra colouring.
1962 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1965 cachep
->colour
= left
/ cachep
->colour_off
;
1970 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1971 size_t size
, unsigned long flags
)
1977 left
= calculate_slab_order(cachep
, size
, flags
);
1981 cachep
->colour
= left
/ cachep
->colour_off
;
1987 * __kmem_cache_create - Create a cache.
1988 * @cachep: cache management descriptor
1989 * @flags: SLAB flags
1991 * Returns a ptr to the cache on success, NULL on failure.
1992 * Cannot be called within a int, but can be interrupted.
1993 * The @ctor is run when new pages are allocated by the cache.
1997 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1998 * to catch references to uninitialised memory.
2000 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2001 * for buffer overruns.
2003 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2004 * cacheline. This can be beneficial if you're counting cycles as closely
2008 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2010 size_t ralign
= BYTES_PER_WORD
;
2013 size_t size
= cachep
->size
;
2018 * Enable redzoning and last user accounting, except for caches with
2019 * large objects, if the increased size would increase the object size
2020 * above the next power of two: caches with object sizes just above a
2021 * power of two have a significant amount of internal fragmentation.
2023 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2024 2 * sizeof(unsigned long long)))
2025 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2026 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2027 flags
|= SLAB_POISON
;
2032 * Check that size is in terms of words. This is needed to avoid
2033 * unaligned accesses for some archs when redzoning is used, and makes
2034 * sure any on-slab bufctl's are also correctly aligned.
2036 if (size
& (BYTES_PER_WORD
- 1)) {
2037 size
+= (BYTES_PER_WORD
- 1);
2038 size
&= ~(BYTES_PER_WORD
- 1);
2041 if (flags
& SLAB_RED_ZONE
) {
2042 ralign
= REDZONE_ALIGN
;
2043 /* If redzoning, ensure that the second redzone is suitably
2044 * aligned, by adjusting the object size accordingly. */
2045 size
+= REDZONE_ALIGN
- 1;
2046 size
&= ~(REDZONE_ALIGN
- 1);
2049 /* 3) caller mandated alignment */
2050 if (ralign
< cachep
->align
) {
2051 ralign
= cachep
->align
;
2053 /* disable debug if necessary */
2054 if (ralign
> __alignof__(unsigned long long))
2055 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2059 cachep
->align
= ralign
;
2060 cachep
->colour_off
= cache_line_size();
2061 /* Offset must be a multiple of the alignment. */
2062 if (cachep
->colour_off
< cachep
->align
)
2063 cachep
->colour_off
= cachep
->align
;
2065 if (slab_is_available())
2073 * Both debugging options require word-alignment which is calculated
2076 if (flags
& SLAB_RED_ZONE
) {
2077 /* add space for red zone words */
2078 cachep
->obj_offset
+= sizeof(unsigned long long);
2079 size
+= 2 * sizeof(unsigned long long);
2081 if (flags
& SLAB_STORE_USER
) {
2082 /* user store requires one word storage behind the end of
2083 * the real object. But if the second red zone needs to be
2084 * aligned to 64 bits, we must allow that much space.
2086 if (flags
& SLAB_RED_ZONE
)
2087 size
+= REDZONE_ALIGN
;
2089 size
+= BYTES_PER_WORD
;
2093 size
= ALIGN(size
, cachep
->align
);
2095 * We should restrict the number of objects in a slab to implement
2096 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2098 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2099 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2103 * To activate debug pagealloc, off-slab management is necessary
2104 * requirement. In early phase of initialization, small sized slab
2105 * doesn't get initialized so it would not be possible. So, we need
2106 * to check size >= 256. It guarantees that all necessary small
2107 * sized slab is initialized in current slab initialization sequence.
2109 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2110 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2111 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2112 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2114 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2115 flags
|= CFLGS_OFF_SLAB
;
2116 cachep
->obj_offset
+= tmp_size
- size
;
2124 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2125 flags
|= CFLGS_OBJFREELIST_SLAB
;
2129 if (set_off_slab_cache(cachep
, size
, flags
)) {
2130 flags
|= CFLGS_OFF_SLAB
;
2134 if (set_on_slab_cache(cachep
, size
, flags
))
2140 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2141 cachep
->flags
= flags
;
2142 cachep
->allocflags
= __GFP_COMP
;
2143 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2144 cachep
->allocflags
|= GFP_DMA
;
2145 cachep
->size
= size
;
2146 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2150 * If we're going to use the generic kernel_map_pages()
2151 * poisoning, then it's going to smash the contents of
2152 * the redzone and userword anyhow, so switch them off.
2154 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2155 (cachep
->flags
& SLAB_POISON
) &&
2156 is_debug_pagealloc_cache(cachep
))
2157 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2160 if (OFF_SLAB(cachep
)) {
2161 cachep
->freelist_cache
=
2162 kmalloc_slab(cachep
->freelist_size
, 0u);
2165 err
= setup_cpu_cache(cachep
, gfp
);
2167 __kmem_cache_release(cachep
);
2175 static void check_irq_off(void)
2177 BUG_ON(!irqs_disabled());
2180 static void check_irq_on(void)
2182 BUG_ON(irqs_disabled());
2185 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2189 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2193 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2197 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2202 #define check_irq_off() do { } while(0)
2203 #define check_irq_on() do { } while(0)
2204 #define check_spinlock_acquired(x) do { } while(0)
2205 #define check_spinlock_acquired_node(x, y) do { } while(0)
2208 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2209 struct array_cache
*ac
,
2210 int force
, int node
);
2212 static void do_drain(void *arg
)
2214 struct kmem_cache
*cachep
= arg
;
2215 struct array_cache
*ac
;
2216 int node
= numa_mem_id();
2217 struct kmem_cache_node
*n
;
2221 ac
= cpu_cache_get(cachep
);
2222 n
= get_node(cachep
, node
);
2223 spin_lock(&n
->list_lock
);
2224 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2225 spin_unlock(&n
->list_lock
);
2226 slabs_destroy(cachep
, &list
);
2230 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2232 struct kmem_cache_node
*n
;
2235 on_each_cpu(do_drain
, cachep
, 1);
2237 for_each_kmem_cache_node(cachep
, node
, n
)
2239 drain_alien_cache(cachep
, n
->alien
);
2241 for_each_kmem_cache_node(cachep
, node
, n
)
2242 drain_array(cachep
, n
, n
->shared
, 1, node
);
2246 * Remove slabs from the list of free slabs.
2247 * Specify the number of slabs to drain in tofree.
2249 * Returns the actual number of slabs released.
2251 static int drain_freelist(struct kmem_cache
*cache
,
2252 struct kmem_cache_node
*n
, int tofree
)
2254 struct list_head
*p
;
2259 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2261 spin_lock_irq(&n
->list_lock
);
2262 p
= n
->slabs_free
.prev
;
2263 if (p
== &n
->slabs_free
) {
2264 spin_unlock_irq(&n
->list_lock
);
2268 page
= list_entry(p
, struct page
, lru
);
2269 list_del(&page
->lru
);
2271 * Safe to drop the lock. The slab is no longer linked
2274 n
->free_objects
-= cache
->num
;
2275 spin_unlock_irq(&n
->list_lock
);
2276 slab_destroy(cache
, page
);
2283 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2287 struct kmem_cache_node
*n
;
2289 drain_cpu_caches(cachep
);
2292 for_each_kmem_cache_node(cachep
, node
, n
) {
2293 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2295 ret
+= !list_empty(&n
->slabs_full
) ||
2296 !list_empty(&n
->slabs_partial
);
2298 return (ret
? 1 : 0);
2301 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2303 return __kmem_cache_shrink(cachep
, false);
2306 void __kmem_cache_release(struct kmem_cache
*cachep
)
2309 struct kmem_cache_node
*n
;
2311 free_percpu(cachep
->cpu_cache
);
2313 /* NUMA: free the node structures */
2314 for_each_kmem_cache_node(cachep
, i
, n
) {
2316 free_alien_cache(n
->alien
);
2318 cachep
->node
[i
] = NULL
;
2323 * Get the memory for a slab management obj.
2325 * For a slab cache when the slab descriptor is off-slab, the
2326 * slab descriptor can't come from the same cache which is being created,
2327 * Because if it is the case, that means we defer the creation of
2328 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2329 * And we eventually call down to __kmem_cache_create(), which
2330 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2331 * This is a "chicken-and-egg" problem.
2333 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2334 * which are all initialized during kmem_cache_init().
2336 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2337 struct page
*page
, int colour_off
,
2338 gfp_t local_flags
, int nodeid
)
2341 void *addr
= page_address(page
);
2343 page
->s_mem
= addr
+ colour_off
;
2346 if (OBJFREELIST_SLAB(cachep
))
2348 else if (OFF_SLAB(cachep
)) {
2349 /* Slab management obj is off-slab. */
2350 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2351 local_flags
, nodeid
);
2355 /* We will use last bytes at the slab for freelist */
2356 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2357 cachep
->freelist_size
;
2363 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2365 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2368 static inline void set_free_obj(struct page
*page
,
2369 unsigned int idx
, freelist_idx_t val
)
2371 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2374 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2379 for (i
= 0; i
< cachep
->num
; i
++) {
2380 void *objp
= index_to_obj(cachep
, page
, i
);
2382 if (cachep
->flags
& SLAB_STORE_USER
)
2383 *dbg_userword(cachep
, objp
) = NULL
;
2385 if (cachep
->flags
& SLAB_RED_ZONE
) {
2386 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2387 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2390 * Constructors are not allowed to allocate memory from the same
2391 * cache which they are a constructor for. Otherwise, deadlock.
2392 * They must also be threaded.
2394 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2395 cachep
->ctor(objp
+ obj_offset(cachep
));
2397 if (cachep
->flags
& SLAB_RED_ZONE
) {
2398 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2399 slab_error(cachep
, "constructor overwrote the"
2400 " end of an object");
2401 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2402 slab_error(cachep
, "constructor overwrote the"
2403 " start of an object");
2405 /* need to poison the objs? */
2406 if (cachep
->flags
& SLAB_POISON
) {
2407 poison_obj(cachep
, objp
, POISON_FREE
);
2408 slab_kernel_map(cachep
, objp
, 0, 0);
2414 static void cache_init_objs(struct kmem_cache
*cachep
,
2419 cache_init_objs_debug(cachep
, page
);
2421 if (OBJFREELIST_SLAB(cachep
)) {
2422 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2426 for (i
= 0; i
< cachep
->num
; i
++) {
2427 /* constructor could break poison info */
2428 if (DEBUG
== 0 && cachep
->ctor
)
2429 cachep
->ctor(index_to_obj(cachep
, page
, i
));
2431 set_free_obj(page
, i
, i
);
2435 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2437 if (CONFIG_ZONE_DMA_FLAG
) {
2438 if (flags
& GFP_DMA
)
2439 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2441 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2445 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2449 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2453 if (cachep
->flags
& SLAB_STORE_USER
)
2454 set_store_user_dirty(cachep
);
2460 static void slab_put_obj(struct kmem_cache
*cachep
,
2461 struct page
*page
, void *objp
)
2463 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2467 /* Verify double free bug */
2468 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2469 if (get_free_obj(page
, i
) == objnr
) {
2470 printk(KERN_ERR
"slab: double free detected in cache "
2471 "'%s', objp %p\n", cachep
->name
, objp
);
2477 if (!page
->freelist
)
2478 page
->freelist
= objp
+ obj_offset(cachep
);
2480 set_free_obj(page
, page
->active
, objnr
);
2484 * Map pages beginning at addr to the given cache and slab. This is required
2485 * for the slab allocator to be able to lookup the cache and slab of a
2486 * virtual address for kfree, ksize, and slab debugging.
2488 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2491 page
->slab_cache
= cache
;
2492 page
->freelist
= freelist
;
2496 * Grow (by 1) the number of slabs within a cache. This is called by
2497 * kmem_cache_alloc() when there are no active objs left in a cache.
2499 static int cache_grow(struct kmem_cache
*cachep
,
2500 gfp_t flags
, int nodeid
, struct page
*page
)
2505 struct kmem_cache_node
*n
;
2508 * Be lazy and only check for valid flags here, keeping it out of the
2509 * critical path in kmem_cache_alloc().
2511 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2512 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2515 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2517 /* Take the node list lock to change the colour_next on this node */
2519 n
= get_node(cachep
, nodeid
);
2520 spin_lock(&n
->list_lock
);
2522 /* Get colour for the slab, and cal the next value. */
2523 offset
= n
->colour_next
;
2525 if (n
->colour_next
>= cachep
->colour
)
2527 spin_unlock(&n
->list_lock
);
2529 offset
*= cachep
->colour_off
;
2531 if (gfpflags_allow_blocking(local_flags
))
2535 * The test for missing atomic flag is performed here, rather than
2536 * the more obvious place, simply to reduce the critical path length
2537 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2538 * will eventually be caught here (where it matters).
2540 kmem_flagcheck(cachep
, flags
);
2543 * Get mem for the objs. Attempt to allocate a physical page from
2547 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2551 /* Get slab management. */
2552 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2553 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2554 if (OFF_SLAB(cachep
) && !freelist
)
2557 slab_map_pages(cachep
, page
, freelist
);
2559 cache_init_objs(cachep
, page
);
2561 if (gfpflags_allow_blocking(local_flags
))
2562 local_irq_disable();
2564 spin_lock(&n
->list_lock
);
2566 /* Make slab active. */
2567 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2568 STATS_INC_GROWN(cachep
);
2569 n
->free_objects
+= cachep
->num
;
2570 spin_unlock(&n
->list_lock
);
2573 kmem_freepages(cachep
, page
);
2575 if (gfpflags_allow_blocking(local_flags
))
2576 local_irq_disable();
2583 * Perform extra freeing checks:
2584 * - detect bad pointers.
2585 * - POISON/RED_ZONE checking
2587 static void kfree_debugcheck(const void *objp
)
2589 if (!virt_addr_valid(objp
)) {
2590 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2591 (unsigned long)objp
);
2596 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2598 unsigned long long redzone1
, redzone2
;
2600 redzone1
= *dbg_redzone1(cache
, obj
);
2601 redzone2
= *dbg_redzone2(cache
, obj
);
2606 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2609 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2610 slab_error(cache
, "double free detected");
2612 slab_error(cache
, "memory outside object was overwritten");
2614 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2615 obj
, redzone1
, redzone2
);
2618 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2619 unsigned long caller
)
2624 BUG_ON(virt_to_cache(objp
) != cachep
);
2626 objp
-= obj_offset(cachep
);
2627 kfree_debugcheck(objp
);
2628 page
= virt_to_head_page(objp
);
2630 if (cachep
->flags
& SLAB_RED_ZONE
) {
2631 verify_redzone_free(cachep
, objp
);
2632 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2633 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2635 if (cachep
->flags
& SLAB_STORE_USER
) {
2636 set_store_user_dirty(cachep
);
2637 *dbg_userword(cachep
, objp
) = (void *)caller
;
2640 objnr
= obj_to_index(cachep
, page
, objp
);
2642 BUG_ON(objnr
>= cachep
->num
);
2643 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2645 if (cachep
->flags
& SLAB_POISON
) {
2646 poison_obj(cachep
, objp
, POISON_FREE
);
2647 slab_kernel_map(cachep
, objp
, 0, caller
);
2653 #define kfree_debugcheck(x) do { } while(0)
2654 #define cache_free_debugcheck(x,objp,z) (objp)
2657 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2665 objp
= next
- obj_offset(cachep
);
2666 next
= *(void **)next
;
2667 poison_obj(cachep
, objp
, POISON_FREE
);
2672 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2673 struct kmem_cache_node
*n
, struct page
*page
,
2676 /* move slabp to correct slabp list: */
2677 list_del(&page
->lru
);
2678 if (page
->active
== cachep
->num
) {
2679 list_add(&page
->lru
, &n
->slabs_full
);
2680 if (OBJFREELIST_SLAB(cachep
)) {
2682 /* Poisoning will be done without holding the lock */
2683 if (cachep
->flags
& SLAB_POISON
) {
2684 void **objp
= page
->freelist
;
2690 page
->freelist
= NULL
;
2693 list_add(&page
->lru
, &n
->slabs_partial
);
2696 /* Try to find non-pfmemalloc slab if needed */
2697 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2698 struct page
*page
, bool pfmemalloc
)
2706 if (!PageSlabPfmemalloc(page
))
2709 /* No need to keep pfmemalloc slab if we have enough free objects */
2710 if (n
->free_objects
> n
->free_limit
) {
2711 ClearPageSlabPfmemalloc(page
);
2715 /* Move pfmemalloc slab to the end of list to speed up next search */
2716 list_del(&page
->lru
);
2718 list_add_tail(&page
->lru
, &n
->slabs_free
);
2720 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2722 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2723 if (!PageSlabPfmemalloc(page
))
2727 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2728 if (!PageSlabPfmemalloc(page
))
2735 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2739 page
= list_first_entry_or_null(&n
->slabs_partial
,
2742 n
->free_touched
= 1;
2743 page
= list_first_entry_or_null(&n
->slabs_free
,
2747 if (sk_memalloc_socks())
2748 return get_valid_first_slab(n
, page
, pfmemalloc
);
2753 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2754 struct kmem_cache_node
*n
, gfp_t flags
)
2760 if (!gfp_pfmemalloc_allowed(flags
))
2763 spin_lock(&n
->list_lock
);
2764 page
= get_first_slab(n
, true);
2766 spin_unlock(&n
->list_lock
);
2770 obj
= slab_get_obj(cachep
, page
);
2773 fixup_slab_list(cachep
, n
, page
, &list
);
2775 spin_unlock(&n
->list_lock
);
2776 fixup_objfreelist_debug(cachep
, &list
);
2781 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2784 struct kmem_cache_node
*n
;
2785 struct array_cache
*ac
;
2790 node
= numa_mem_id();
2793 ac
= cpu_cache_get(cachep
);
2794 batchcount
= ac
->batchcount
;
2795 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2797 * If there was little recent activity on this cache, then
2798 * perform only a partial refill. Otherwise we could generate
2801 batchcount
= BATCHREFILL_LIMIT
;
2803 n
= get_node(cachep
, node
);
2805 BUG_ON(ac
->avail
> 0 || !n
);
2806 spin_lock(&n
->list_lock
);
2808 /* See if we can refill from the shared array */
2809 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2810 n
->shared
->touched
= 1;
2814 while (batchcount
> 0) {
2816 /* Get slab alloc is to come from. */
2817 page
= get_first_slab(n
, false);
2821 check_spinlock_acquired(cachep
);
2824 * The slab was either on partial or free list so
2825 * there must be at least one object available for
2828 BUG_ON(page
->active
>= cachep
->num
);
2830 while (page
->active
< cachep
->num
&& batchcount
--) {
2831 STATS_INC_ALLOCED(cachep
);
2832 STATS_INC_ACTIVE(cachep
);
2833 STATS_SET_HIGH(cachep
);
2835 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2838 fixup_slab_list(cachep
, n
, page
, &list
);
2842 n
->free_objects
-= ac
->avail
;
2844 spin_unlock(&n
->list_lock
);
2845 fixup_objfreelist_debug(cachep
, &list
);
2847 if (unlikely(!ac
->avail
)) {
2850 /* Check if we can use obj in pfmemalloc slab */
2851 if (sk_memalloc_socks()) {
2852 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2858 x
= cache_grow(cachep
, gfp_exact_node(flags
), node
, NULL
);
2860 /* cache_grow can reenable interrupts, then ac could change. */
2861 ac
= cpu_cache_get(cachep
);
2862 node
= numa_mem_id();
2864 /* no objects in sight? abort */
2865 if (!x
&& ac
->avail
== 0)
2868 if (!ac
->avail
) /* objects refilled by interrupt? */
2873 return ac
->entry
[--ac
->avail
];
2876 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2879 might_sleep_if(gfpflags_allow_blocking(flags
));
2881 kmem_flagcheck(cachep
, flags
);
2886 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2887 gfp_t flags
, void *objp
, unsigned long caller
)
2891 if (cachep
->flags
& SLAB_POISON
) {
2892 check_poison_obj(cachep
, objp
);
2893 slab_kernel_map(cachep
, objp
, 1, 0);
2894 poison_obj(cachep
, objp
, POISON_INUSE
);
2896 if (cachep
->flags
& SLAB_STORE_USER
)
2897 *dbg_userword(cachep
, objp
) = (void *)caller
;
2899 if (cachep
->flags
& SLAB_RED_ZONE
) {
2900 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2901 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2902 slab_error(cachep
, "double free, or memory outside"
2903 " object was overwritten");
2905 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2906 objp
, *dbg_redzone1(cachep
, objp
),
2907 *dbg_redzone2(cachep
, objp
));
2909 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2910 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2913 objp
+= obj_offset(cachep
);
2914 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2916 if (ARCH_SLAB_MINALIGN
&&
2917 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2918 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2919 objp
, (int)ARCH_SLAB_MINALIGN
);
2924 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2927 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2930 struct array_cache
*ac
;
2934 ac
= cpu_cache_get(cachep
);
2935 if (likely(ac
->avail
)) {
2937 objp
= ac
->entry
[--ac
->avail
];
2939 STATS_INC_ALLOCHIT(cachep
);
2943 STATS_INC_ALLOCMISS(cachep
);
2944 objp
= cache_alloc_refill(cachep
, flags
);
2946 * the 'ac' may be updated by cache_alloc_refill(),
2947 * and kmemleak_erase() requires its correct value.
2949 ac
= cpu_cache_get(cachep
);
2953 * To avoid a false negative, if an object that is in one of the
2954 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2955 * treat the array pointers as a reference to the object.
2958 kmemleak_erase(&ac
->entry
[ac
->avail
]);
2964 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2966 * If we are in_interrupt, then process context, including cpusets and
2967 * mempolicy, may not apply and should not be used for allocation policy.
2969 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2971 int nid_alloc
, nid_here
;
2973 if (in_interrupt() || (flags
& __GFP_THISNODE
))
2975 nid_alloc
= nid_here
= numa_mem_id();
2976 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2977 nid_alloc
= cpuset_slab_spread_node();
2978 else if (current
->mempolicy
)
2979 nid_alloc
= mempolicy_slab_node();
2980 if (nid_alloc
!= nid_here
)
2981 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
2986 * Fallback function if there was no memory available and no objects on a
2987 * certain node and fall back is permitted. First we scan all the
2988 * available node for available objects. If that fails then we
2989 * perform an allocation without specifying a node. This allows the page
2990 * allocator to do its reclaim / fallback magic. We then insert the
2991 * slab into the proper nodelist and then allocate from it.
2993 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
2995 struct zonelist
*zonelist
;
2999 enum zone_type high_zoneidx
= gfp_zone(flags
);
3002 unsigned int cpuset_mems_cookie
;
3004 if (flags
& __GFP_THISNODE
)
3007 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3010 cpuset_mems_cookie
= read_mems_allowed_begin();
3011 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3015 * Look through allowed nodes for objects available
3016 * from existing per node queues.
3018 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3019 nid
= zone_to_nid(zone
);
3021 if (cpuset_zone_allowed(zone
, flags
) &&
3022 get_node(cache
, nid
) &&
3023 get_node(cache
, nid
)->free_objects
) {
3024 obj
= ____cache_alloc_node(cache
,
3025 gfp_exact_node(flags
), nid
);
3033 * This allocation will be performed within the constraints
3034 * of the current cpuset / memory policy requirements.
3035 * We may trigger various forms of reclaim on the allowed
3036 * set and go into memory reserves if necessary.
3040 if (gfpflags_allow_blocking(local_flags
))
3042 kmem_flagcheck(cache
, flags
);
3043 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3044 if (gfpflags_allow_blocking(local_flags
))
3045 local_irq_disable();
3048 * Insert into the appropriate per node queues
3050 nid
= page_to_nid(page
);
3051 if (cache_grow(cache
, flags
, nid
, page
)) {
3052 obj
= ____cache_alloc_node(cache
,
3053 gfp_exact_node(flags
), nid
);
3056 * Another processor may allocate the
3057 * objects in the slab since we are
3058 * not holding any locks.
3062 /* cache_grow already freed obj */
3068 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3074 * A interface to enable slab creation on nodeid
3076 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3080 struct kmem_cache_node
*n
;
3085 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3086 n
= get_node(cachep
, nodeid
);
3091 spin_lock(&n
->list_lock
);
3092 page
= get_first_slab(n
, false);
3096 check_spinlock_acquired_node(cachep
, nodeid
);
3098 STATS_INC_NODEALLOCS(cachep
);
3099 STATS_INC_ACTIVE(cachep
);
3100 STATS_SET_HIGH(cachep
);
3102 BUG_ON(page
->active
== cachep
->num
);
3104 obj
= slab_get_obj(cachep
, page
);
3107 fixup_slab_list(cachep
, n
, page
, &list
);
3109 spin_unlock(&n
->list_lock
);
3110 fixup_objfreelist_debug(cachep
, &list
);
3114 spin_unlock(&n
->list_lock
);
3115 x
= cache_grow(cachep
, gfp_exact_node(flags
), nodeid
, NULL
);
3119 return fallback_alloc(cachep
, flags
);
3125 static __always_inline
void *
3126 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3127 unsigned long caller
)
3129 unsigned long save_flags
;
3131 int slab_node
= numa_mem_id();
3133 flags
&= gfp_allowed_mask
;
3134 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3135 if (unlikely(!cachep
))
3138 cache_alloc_debugcheck_before(cachep
, flags
);
3139 local_irq_save(save_flags
);
3141 if (nodeid
== NUMA_NO_NODE
)
3144 if (unlikely(!get_node(cachep
, nodeid
))) {
3145 /* Node not bootstrapped yet */
3146 ptr
= fallback_alloc(cachep
, flags
);
3150 if (nodeid
== slab_node
) {
3152 * Use the locally cached objects if possible.
3153 * However ____cache_alloc does not allow fallback
3154 * to other nodes. It may fail while we still have
3155 * objects on other nodes available.
3157 ptr
= ____cache_alloc(cachep
, flags
);
3161 /* ___cache_alloc_node can fall back to other nodes */
3162 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3164 local_irq_restore(save_flags
);
3165 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3167 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3168 memset(ptr
, 0, cachep
->object_size
);
3170 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3174 static __always_inline
void *
3175 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3179 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3180 objp
= alternate_node_alloc(cache
, flags
);
3184 objp
= ____cache_alloc(cache
, flags
);
3187 * We may just have run out of memory on the local node.
3188 * ____cache_alloc_node() knows how to locate memory on other nodes
3191 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3198 static __always_inline
void *
3199 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3201 return ____cache_alloc(cachep
, flags
);
3204 #endif /* CONFIG_NUMA */
3206 static __always_inline
void *
3207 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3209 unsigned long save_flags
;
3212 flags
&= gfp_allowed_mask
;
3213 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3214 if (unlikely(!cachep
))
3217 cache_alloc_debugcheck_before(cachep
, flags
);
3218 local_irq_save(save_flags
);
3219 objp
= __do_cache_alloc(cachep
, flags
);
3220 local_irq_restore(save_flags
);
3221 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3224 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3225 memset(objp
, 0, cachep
->object_size
);
3227 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3232 * Caller needs to acquire correct kmem_cache_node's list_lock
3233 * @list: List of detached free slabs should be freed by caller
3235 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3236 int nr_objects
, int node
, struct list_head
*list
)
3239 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3241 for (i
= 0; i
< nr_objects
; i
++) {
3247 page
= virt_to_head_page(objp
);
3248 list_del(&page
->lru
);
3249 check_spinlock_acquired_node(cachep
, node
);
3250 slab_put_obj(cachep
, page
, objp
);
3251 STATS_DEC_ACTIVE(cachep
);
3254 /* fixup slab chains */
3255 if (page
->active
== 0) {
3256 if (n
->free_objects
> n
->free_limit
) {
3257 n
->free_objects
-= cachep
->num
;
3258 list_add_tail(&page
->lru
, list
);
3260 list_add(&page
->lru
, &n
->slabs_free
);
3263 /* Unconditionally move a slab to the end of the
3264 * partial list on free - maximum time for the
3265 * other objects to be freed, too.
3267 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3272 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3275 struct kmem_cache_node
*n
;
3276 int node
= numa_mem_id();
3279 batchcount
= ac
->batchcount
;
3282 n
= get_node(cachep
, node
);
3283 spin_lock(&n
->list_lock
);
3285 struct array_cache
*shared_array
= n
->shared
;
3286 int max
= shared_array
->limit
- shared_array
->avail
;
3288 if (batchcount
> max
)
3290 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3291 ac
->entry
, sizeof(void *) * batchcount
);
3292 shared_array
->avail
+= batchcount
;
3297 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3304 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3305 BUG_ON(page
->active
);
3309 STATS_SET_FREEABLE(cachep
, i
);
3312 spin_unlock(&n
->list_lock
);
3313 slabs_destroy(cachep
, &list
);
3314 ac
->avail
-= batchcount
;
3315 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3319 * Release an obj back to its cache. If the obj has a constructed state, it must
3320 * be in this state _before_ it is released. Called with disabled ints.
3322 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3323 unsigned long caller
)
3325 struct array_cache
*ac
= cpu_cache_get(cachep
);
3328 kmemleak_free_recursive(objp
, cachep
->flags
);
3329 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3331 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3334 * Skip calling cache_free_alien() when the platform is not numa.
3335 * This will avoid cache misses that happen while accessing slabp (which
3336 * is per page memory reference) to get nodeid. Instead use a global
3337 * variable to skip the call, which is mostly likely to be present in
3340 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3343 if (ac
->avail
< ac
->limit
) {
3344 STATS_INC_FREEHIT(cachep
);
3346 STATS_INC_FREEMISS(cachep
);
3347 cache_flusharray(cachep
, ac
);
3350 if (sk_memalloc_socks()) {
3351 struct page
*page
= virt_to_head_page(objp
);
3353 if (unlikely(PageSlabPfmemalloc(page
))) {
3354 cache_free_pfmemalloc(cachep
, page
, objp
);
3359 ac
->entry
[ac
->avail
++] = objp
;
3363 * kmem_cache_alloc - Allocate an object
3364 * @cachep: The cache to allocate from.
3365 * @flags: See kmalloc().
3367 * Allocate an object from this cache. The flags are only relevant
3368 * if the cache has no available objects.
3370 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3372 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3374 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3375 cachep
->object_size
, cachep
->size
, flags
);
3379 EXPORT_SYMBOL(kmem_cache_alloc
);
3381 static __always_inline
void
3382 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3383 size_t size
, void **p
, unsigned long caller
)
3387 for (i
= 0; i
< size
; i
++)
3388 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3391 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3396 s
= slab_pre_alloc_hook(s
, flags
);
3400 cache_alloc_debugcheck_before(s
, flags
);
3402 local_irq_disable();
3403 for (i
= 0; i
< size
; i
++) {
3404 void *objp
= __do_cache_alloc(s
, flags
);
3406 if (unlikely(!objp
))
3412 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3414 /* Clear memory outside IRQ disabled section */
3415 if (unlikely(flags
& __GFP_ZERO
))
3416 for (i
= 0; i
< size
; i
++)
3417 memset(p
[i
], 0, s
->object_size
);
3419 slab_post_alloc_hook(s
, flags
, size
, p
);
3420 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3424 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3425 slab_post_alloc_hook(s
, flags
, i
, p
);
3426 __kmem_cache_free_bulk(s
, i
, p
);
3429 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3431 #ifdef CONFIG_TRACING
3433 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3437 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3439 trace_kmalloc(_RET_IP_
, ret
,
3440 size
, cachep
->size
, flags
);
3443 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3448 * kmem_cache_alloc_node - Allocate an object on the specified node
3449 * @cachep: The cache to allocate from.
3450 * @flags: See kmalloc().
3451 * @nodeid: node number of the target node.
3453 * Identical to kmem_cache_alloc but it will allocate memory on the given
3454 * node, which can improve the performance for cpu bound structures.
3456 * Fallback to other node is possible if __GFP_THISNODE is not set.
3458 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3460 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3462 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3463 cachep
->object_size
, cachep
->size
,
3468 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3470 #ifdef CONFIG_TRACING
3471 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3478 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3480 trace_kmalloc_node(_RET_IP_
, ret
,
3485 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3488 static __always_inline
void *
3489 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3491 struct kmem_cache
*cachep
;
3493 cachep
= kmalloc_slab(size
, flags
);
3494 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3496 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3499 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3501 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3503 EXPORT_SYMBOL(__kmalloc_node
);
3505 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3506 int node
, unsigned long caller
)
3508 return __do_kmalloc_node(size
, flags
, node
, caller
);
3510 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3511 #endif /* CONFIG_NUMA */
3514 * __do_kmalloc - allocate memory
3515 * @size: how many bytes of memory are required.
3516 * @flags: the type of memory to allocate (see kmalloc).
3517 * @caller: function caller for debug tracking of the caller
3519 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3520 unsigned long caller
)
3522 struct kmem_cache
*cachep
;
3525 cachep
= kmalloc_slab(size
, flags
);
3526 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3528 ret
= slab_alloc(cachep
, flags
, caller
);
3530 trace_kmalloc(caller
, ret
,
3531 size
, cachep
->size
, flags
);
3536 void *__kmalloc(size_t size
, gfp_t flags
)
3538 return __do_kmalloc(size
, flags
, _RET_IP_
);
3540 EXPORT_SYMBOL(__kmalloc
);
3542 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3544 return __do_kmalloc(size
, flags
, caller
);
3546 EXPORT_SYMBOL(__kmalloc_track_caller
);
3549 * kmem_cache_free - Deallocate an object
3550 * @cachep: The cache the allocation was from.
3551 * @objp: The previously allocated object.
3553 * Free an object which was previously allocated from this
3556 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3558 unsigned long flags
;
3559 cachep
= cache_from_obj(cachep
, objp
);
3563 local_irq_save(flags
);
3564 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3565 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3566 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3567 __cache_free(cachep
, objp
, _RET_IP_
);
3568 local_irq_restore(flags
);
3570 trace_kmem_cache_free(_RET_IP_
, objp
);
3572 EXPORT_SYMBOL(kmem_cache_free
);
3574 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3576 struct kmem_cache
*s
;
3579 local_irq_disable();
3580 for (i
= 0; i
< size
; i
++) {
3583 if (!orig_s
) /* called via kfree_bulk */
3584 s
= virt_to_cache(objp
);
3586 s
= cache_from_obj(orig_s
, objp
);
3588 debug_check_no_locks_freed(objp
, s
->object_size
);
3589 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3590 debug_check_no_obj_freed(objp
, s
->object_size
);
3592 __cache_free(s
, objp
, _RET_IP_
);
3596 /* FIXME: add tracing */
3598 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3601 * kfree - free previously allocated memory
3602 * @objp: pointer returned by kmalloc.
3604 * If @objp is NULL, no operation is performed.
3606 * Don't free memory not originally allocated by kmalloc()
3607 * or you will run into trouble.
3609 void kfree(const void *objp
)
3611 struct kmem_cache
*c
;
3612 unsigned long flags
;
3614 trace_kfree(_RET_IP_
, objp
);
3616 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3618 local_irq_save(flags
);
3619 kfree_debugcheck(objp
);
3620 c
= virt_to_cache(objp
);
3621 debug_check_no_locks_freed(objp
, c
->object_size
);
3623 debug_check_no_obj_freed(objp
, c
->object_size
);
3624 __cache_free(c
, (void *)objp
, _RET_IP_
);
3625 local_irq_restore(flags
);
3627 EXPORT_SYMBOL(kfree
);
3630 * This initializes kmem_cache_node or resizes various caches for all nodes.
3632 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3635 struct kmem_cache_node
*n
;
3636 struct array_cache
*new_shared
;
3637 struct alien_cache
**new_alien
= NULL
;
3639 for_each_online_node(node
) {
3641 if (use_alien_caches
) {
3642 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3648 if (cachep
->shared
) {
3649 new_shared
= alloc_arraycache(node
,
3650 cachep
->shared
*cachep
->batchcount
,
3653 free_alien_cache(new_alien
);
3658 n
= get_node(cachep
, node
);
3660 struct array_cache
*shared
= n
->shared
;
3663 spin_lock_irq(&n
->list_lock
);
3666 free_block(cachep
, shared
->entry
,
3667 shared
->avail
, node
, &list
);
3669 n
->shared
= new_shared
;
3671 n
->alien
= new_alien
;
3674 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3675 cachep
->batchcount
+ cachep
->num
;
3676 spin_unlock_irq(&n
->list_lock
);
3677 slabs_destroy(cachep
, &list
);
3679 free_alien_cache(new_alien
);
3682 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3684 free_alien_cache(new_alien
);
3689 kmem_cache_node_init(n
);
3690 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3691 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3692 n
->shared
= new_shared
;
3693 n
->alien
= new_alien
;
3694 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3695 cachep
->batchcount
+ cachep
->num
;
3696 cachep
->node
[node
] = n
;
3701 if (!cachep
->list
.next
) {
3702 /* Cache is not active yet. Roll back what we did */
3705 n
= get_node(cachep
, node
);
3708 free_alien_cache(n
->alien
);
3710 cachep
->node
[node
] = NULL
;
3718 /* Always called with the slab_mutex held */
3719 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3720 int batchcount
, int shared
, gfp_t gfp
)
3722 struct array_cache __percpu
*cpu_cache
, *prev
;
3725 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3729 prev
= cachep
->cpu_cache
;
3730 cachep
->cpu_cache
= cpu_cache
;
3731 kick_all_cpus_sync();
3734 cachep
->batchcount
= batchcount
;
3735 cachep
->limit
= limit
;
3736 cachep
->shared
= shared
;
3741 for_each_online_cpu(cpu
) {
3744 struct kmem_cache_node
*n
;
3745 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3747 node
= cpu_to_mem(cpu
);
3748 n
= get_node(cachep
, node
);
3749 spin_lock_irq(&n
->list_lock
);
3750 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3751 spin_unlock_irq(&n
->list_lock
);
3752 slabs_destroy(cachep
, &list
);
3757 return alloc_kmem_cache_node(cachep
, gfp
);
3760 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3761 int batchcount
, int shared
, gfp_t gfp
)
3764 struct kmem_cache
*c
;
3766 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3768 if (slab_state
< FULL
)
3771 if ((ret
< 0) || !is_root_cache(cachep
))
3774 lockdep_assert_held(&slab_mutex
);
3775 for_each_memcg_cache(c
, cachep
) {
3776 /* return value determined by the root cache only */
3777 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3783 /* Called with slab_mutex held always */
3784 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3791 if (!is_root_cache(cachep
)) {
3792 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3793 limit
= root
->limit
;
3794 shared
= root
->shared
;
3795 batchcount
= root
->batchcount
;
3798 if (limit
&& shared
&& batchcount
)
3801 * The head array serves three purposes:
3802 * - create a LIFO ordering, i.e. return objects that are cache-warm
3803 * - reduce the number of spinlock operations.
3804 * - reduce the number of linked list operations on the slab and
3805 * bufctl chains: array operations are cheaper.
3806 * The numbers are guessed, we should auto-tune as described by
3809 if (cachep
->size
> 131072)
3811 else if (cachep
->size
> PAGE_SIZE
)
3813 else if (cachep
->size
> 1024)
3815 else if (cachep
->size
> 256)
3821 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3822 * allocation behaviour: Most allocs on one cpu, most free operations
3823 * on another cpu. For these cases, an efficient object passing between
3824 * cpus is necessary. This is provided by a shared array. The array
3825 * replaces Bonwick's magazine layer.
3826 * On uniprocessor, it's functionally equivalent (but less efficient)
3827 * to a larger limit. Thus disabled by default.
3830 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3835 * With debugging enabled, large batchcount lead to excessively long
3836 * periods with disabled local interrupts. Limit the batchcount
3841 batchcount
= (limit
+ 1) / 2;
3843 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3845 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3846 cachep
->name
, -err
);
3851 * Drain an array if it contains any elements taking the node lock only if
3852 * necessary. Note that the node listlock also protects the array_cache
3853 * if drain_array() is used on the shared array.
3855 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3856 struct array_cache
*ac
, int force
, int node
)
3861 if (!ac
|| !ac
->avail
)
3863 if (ac
->touched
&& !force
) {
3866 spin_lock_irq(&n
->list_lock
);
3868 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3869 if (tofree
> ac
->avail
)
3870 tofree
= (ac
->avail
+ 1) / 2;
3871 free_block(cachep
, ac
->entry
, tofree
, node
, &list
);
3872 ac
->avail
-= tofree
;
3873 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3874 sizeof(void *) * ac
->avail
);
3876 spin_unlock_irq(&n
->list_lock
);
3877 slabs_destroy(cachep
, &list
);
3882 * cache_reap - Reclaim memory from caches.
3883 * @w: work descriptor
3885 * Called from workqueue/eventd every few seconds.
3887 * - clear the per-cpu caches for this CPU.
3888 * - return freeable pages to the main free memory pool.
3890 * If we cannot acquire the cache chain mutex then just give up - we'll try
3891 * again on the next iteration.
3893 static void cache_reap(struct work_struct
*w
)
3895 struct kmem_cache
*searchp
;
3896 struct kmem_cache_node
*n
;
3897 int node
= numa_mem_id();
3898 struct delayed_work
*work
= to_delayed_work(w
);
3900 if (!mutex_trylock(&slab_mutex
))
3901 /* Give up. Setup the next iteration. */
3904 list_for_each_entry(searchp
, &slab_caches
, list
) {
3908 * We only take the node lock if absolutely necessary and we
3909 * have established with reasonable certainty that
3910 * we can do some work if the lock was obtained.
3912 n
= get_node(searchp
, node
);
3914 reap_alien(searchp
, n
);
3916 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
3919 * These are racy checks but it does not matter
3920 * if we skip one check or scan twice.
3922 if (time_after(n
->next_reap
, jiffies
))
3925 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3927 drain_array(searchp
, n
, n
->shared
, 0, node
);
3929 if (n
->free_touched
)
3930 n
->free_touched
= 0;
3934 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3935 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3936 STATS_ADD_REAPED(searchp
, freed
);
3942 mutex_unlock(&slab_mutex
);
3945 /* Set up the next iteration */
3946 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3949 #ifdef CONFIG_SLABINFO
3950 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3953 unsigned long active_objs
;
3954 unsigned long num_objs
;
3955 unsigned long active_slabs
= 0;
3956 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3960 struct kmem_cache_node
*n
;
3964 for_each_kmem_cache_node(cachep
, node
, n
) {
3967 spin_lock_irq(&n
->list_lock
);
3969 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
3970 if (page
->active
!= cachep
->num
&& !error
)
3971 error
= "slabs_full accounting error";
3972 active_objs
+= cachep
->num
;
3975 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
3976 if (page
->active
== cachep
->num
&& !error
)
3977 error
= "slabs_partial accounting error";
3978 if (!page
->active
&& !error
)
3979 error
= "slabs_partial accounting error";
3980 active_objs
+= page
->active
;
3983 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3984 if (page
->active
&& !error
)
3985 error
= "slabs_free accounting error";
3988 free_objects
+= n
->free_objects
;
3990 shared_avail
+= n
->shared
->avail
;
3992 spin_unlock_irq(&n
->list_lock
);
3994 num_slabs
+= active_slabs
;
3995 num_objs
= num_slabs
* cachep
->num
;
3996 if (num_objs
- active_objs
!= free_objects
&& !error
)
3997 error
= "free_objects accounting error";
3999 name
= cachep
->name
;
4001 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4003 sinfo
->active_objs
= active_objs
;
4004 sinfo
->num_objs
= num_objs
;
4005 sinfo
->active_slabs
= active_slabs
;
4006 sinfo
->num_slabs
= num_slabs
;
4007 sinfo
->shared_avail
= shared_avail
;
4008 sinfo
->limit
= cachep
->limit
;
4009 sinfo
->batchcount
= cachep
->batchcount
;
4010 sinfo
->shared
= cachep
->shared
;
4011 sinfo
->objects_per_slab
= cachep
->num
;
4012 sinfo
->cache_order
= cachep
->gfporder
;
4015 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4019 unsigned long high
= cachep
->high_mark
;
4020 unsigned long allocs
= cachep
->num_allocations
;
4021 unsigned long grown
= cachep
->grown
;
4022 unsigned long reaped
= cachep
->reaped
;
4023 unsigned long errors
= cachep
->errors
;
4024 unsigned long max_freeable
= cachep
->max_freeable
;
4025 unsigned long node_allocs
= cachep
->node_allocs
;
4026 unsigned long node_frees
= cachep
->node_frees
;
4027 unsigned long overflows
= cachep
->node_overflow
;
4029 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4030 "%4lu %4lu %4lu %4lu %4lu",
4031 allocs
, high
, grown
,
4032 reaped
, errors
, max_freeable
, node_allocs
,
4033 node_frees
, overflows
);
4037 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4038 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4039 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4040 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4042 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4043 allochit
, allocmiss
, freehit
, freemiss
);
4048 #define MAX_SLABINFO_WRITE 128
4050 * slabinfo_write - Tuning for the slab allocator
4052 * @buffer: user buffer
4053 * @count: data length
4056 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4057 size_t count
, loff_t
*ppos
)
4059 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4060 int limit
, batchcount
, shared
, res
;
4061 struct kmem_cache
*cachep
;
4063 if (count
> MAX_SLABINFO_WRITE
)
4065 if (copy_from_user(&kbuf
, buffer
, count
))
4067 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4069 tmp
= strchr(kbuf
, ' ');
4074 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4077 /* Find the cache in the chain of caches. */
4078 mutex_lock(&slab_mutex
);
4080 list_for_each_entry(cachep
, &slab_caches
, list
) {
4081 if (!strcmp(cachep
->name
, kbuf
)) {
4082 if (limit
< 1 || batchcount
< 1 ||
4083 batchcount
> limit
|| shared
< 0) {
4086 res
= do_tune_cpucache(cachep
, limit
,
4093 mutex_unlock(&slab_mutex
);
4099 #ifdef CONFIG_DEBUG_SLAB_LEAK
4101 static inline int add_caller(unsigned long *n
, unsigned long v
)
4111 unsigned long *q
= p
+ 2 * i
;
4125 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4131 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4140 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4143 for (j
= page
->active
; j
< c
->num
; j
++) {
4144 if (get_free_obj(page
, j
) == i
) {
4154 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4155 * mapping is established when actual object allocation and
4156 * we could mistakenly access the unmapped object in the cpu
4159 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4162 if (!add_caller(n
, v
))
4167 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4169 #ifdef CONFIG_KALLSYMS
4170 unsigned long offset
, size
;
4171 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4173 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4174 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4176 seq_printf(m
, " [%s]", modname
);
4180 seq_printf(m
, "%p", (void *)address
);
4183 static int leaks_show(struct seq_file
*m
, void *p
)
4185 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4187 struct kmem_cache_node
*n
;
4189 unsigned long *x
= m
->private;
4193 if (!(cachep
->flags
& SLAB_STORE_USER
))
4195 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4199 * Set store_user_clean and start to grab stored user information
4200 * for all objects on this cache. If some alloc/free requests comes
4201 * during the processing, information would be wrong so restart
4205 set_store_user_clean(cachep
);
4206 drain_cpu_caches(cachep
);
4210 for_each_kmem_cache_node(cachep
, node
, n
) {
4213 spin_lock_irq(&n
->list_lock
);
4215 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4216 handle_slab(x
, cachep
, page
);
4217 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4218 handle_slab(x
, cachep
, page
);
4219 spin_unlock_irq(&n
->list_lock
);
4221 } while (!is_store_user_clean(cachep
));
4223 name
= cachep
->name
;
4225 /* Increase the buffer size */
4226 mutex_unlock(&slab_mutex
);
4227 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4229 /* Too bad, we are really out */
4231 mutex_lock(&slab_mutex
);
4234 *(unsigned long *)m
->private = x
[0] * 2;
4236 mutex_lock(&slab_mutex
);
4237 /* Now make sure this entry will be retried */
4241 for (i
= 0; i
< x
[1]; i
++) {
4242 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4243 show_symbol(m
, x
[2*i
+2]);
4250 static const struct seq_operations slabstats_op
= {
4251 .start
= slab_start
,
4257 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4261 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4265 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4270 static const struct file_operations proc_slabstats_operations
= {
4271 .open
= slabstats_open
,
4273 .llseek
= seq_lseek
,
4274 .release
= seq_release_private
,
4278 static int __init
slab_proc_init(void)
4280 #ifdef CONFIG_DEBUG_SLAB_LEAK
4281 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4285 module_init(slab_proc_init
);
4289 * ksize - get the actual amount of memory allocated for a given object
4290 * @objp: Pointer to the object
4292 * kmalloc may internally round up allocations and return more memory
4293 * than requested. ksize() can be used to determine the actual amount of
4294 * memory allocated. The caller may use this additional memory, even though
4295 * a smaller amount of memory was initially specified with the kmalloc call.
4296 * The caller must guarantee that objp points to a valid object previously
4297 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4298 * must not be freed during the duration of the call.
4300 size_t ksize(const void *objp
)
4303 if (unlikely(objp
== ZERO_SIZE_PTR
))
4306 return virt_to_cache(objp
)->object_size
;
4308 EXPORT_SYMBOL(ksize
);