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 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
426 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
428 return this_cpu_ptr(cachep
->cpu_cache
);
432 * Calculate the number of objects and left-over bytes for a given buffer size.
434 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
435 unsigned long flags
, size_t *left_over
)
438 size_t slab_size
= PAGE_SIZE
<< gfporder
;
441 * The slab management structure can be either off the slab or
442 * on it. For the latter case, the memory allocated for a
445 * - @buffer_size bytes for each object
446 * - One freelist_idx_t for each object
448 * We don't need to consider alignment of freelist because
449 * freelist will be at the end of slab page. The objects will be
450 * at the correct alignment.
452 * If the slab management structure is off the slab, then the
453 * alignment will already be calculated into the size. Because
454 * the slabs are all pages aligned, the objects will be at the
455 * correct alignment when allocated.
457 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
458 num
= slab_size
/ buffer_size
;
459 *left_over
= slab_size
% buffer_size
;
461 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
462 *left_over
= slab_size
%
463 (buffer_size
+ sizeof(freelist_idx_t
));
470 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
472 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
475 pr_err("slab error in %s(): cache `%s': %s\n",
476 function
, cachep
->name
, msg
);
478 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
483 * By default on NUMA we use alien caches to stage the freeing of
484 * objects allocated from other nodes. This causes massive memory
485 * inefficiencies when using fake NUMA setup to split memory into a
486 * large number of small nodes, so it can be disabled on the command
490 static int use_alien_caches __read_mostly
= 1;
491 static int __init
noaliencache_setup(char *s
)
493 use_alien_caches
= 0;
496 __setup("noaliencache", noaliencache_setup
);
498 static int __init
slab_max_order_setup(char *str
)
500 get_option(&str
, &slab_max_order
);
501 slab_max_order
= slab_max_order
< 0 ? 0 :
502 min(slab_max_order
, MAX_ORDER
- 1);
503 slab_max_order_set
= true;
507 __setup("slab_max_order=", slab_max_order_setup
);
511 * Special reaping functions for NUMA systems called from cache_reap().
512 * These take care of doing round robin flushing of alien caches (containing
513 * objects freed on different nodes from which they were allocated) and the
514 * flushing of remote pcps by calling drain_node_pages.
516 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
518 static void init_reap_node(int cpu
)
522 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
523 if (node
== MAX_NUMNODES
)
524 node
= first_node(node_online_map
);
526 per_cpu(slab_reap_node
, cpu
) = node
;
529 static void next_reap_node(void)
531 int node
= __this_cpu_read(slab_reap_node
);
533 node
= next_node(node
, node_online_map
);
534 if (unlikely(node
>= MAX_NUMNODES
))
535 node
= first_node(node_online_map
);
536 __this_cpu_write(slab_reap_node
, node
);
540 #define init_reap_node(cpu) do { } while (0)
541 #define next_reap_node(void) do { } while (0)
545 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
546 * via the workqueue/eventd.
547 * Add the CPU number into the expiration time to minimize the possibility of
548 * the CPUs getting into lockstep and contending for the global cache chain
551 static void start_cpu_timer(int cpu
)
553 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
556 * When this gets called from do_initcalls via cpucache_init(),
557 * init_workqueues() has already run, so keventd will be setup
560 if (keventd_up() && reap_work
->work
.func
== NULL
) {
562 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
563 schedule_delayed_work_on(cpu
, reap_work
,
564 __round_jiffies_relative(HZ
, cpu
));
568 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
571 * The array_cache structures contain pointers to free object.
572 * However, when such objects are allocated or transferred to another
573 * cache the pointers are not cleared and they could be counted as
574 * valid references during a kmemleak scan. Therefore, kmemleak must
575 * not scan such objects.
577 kmemleak_no_scan(ac
);
581 ac
->batchcount
= batch
;
586 static struct array_cache
*alloc_arraycache(int node
, int entries
,
587 int batchcount
, gfp_t gfp
)
589 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
590 struct array_cache
*ac
= NULL
;
592 ac
= kmalloc_node(memsize
, gfp
, node
);
593 init_arraycache(ac
, entries
, batchcount
);
597 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
598 struct page
*page
, void *objp
)
600 struct kmem_cache_node
*n
;
604 page_node
= page_to_nid(page
);
605 n
= get_node(cachep
, page_node
);
607 spin_lock(&n
->list_lock
);
608 free_block(cachep
, &objp
, 1, page_node
, &list
);
609 spin_unlock(&n
->list_lock
);
611 slabs_destroy(cachep
, &list
);
615 * Transfer objects in one arraycache to another.
616 * Locking must be handled by the caller.
618 * Return the number of entries transferred.
620 static int transfer_objects(struct array_cache
*to
,
621 struct array_cache
*from
, unsigned int max
)
623 /* Figure out how many entries to transfer */
624 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
629 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
639 #define drain_alien_cache(cachep, alien) do { } while (0)
640 #define reap_alien(cachep, n) do { } while (0)
642 static inline struct alien_cache
**alloc_alien_cache(int node
,
643 int limit
, gfp_t gfp
)
648 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
652 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
657 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
663 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
664 gfp_t flags
, int nodeid
)
669 static inline gfp_t
gfp_exact_node(gfp_t flags
)
671 return flags
& ~__GFP_NOFAIL
;
674 #else /* CONFIG_NUMA */
676 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
677 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
679 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
680 int batch
, gfp_t gfp
)
682 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
683 struct alien_cache
*alc
= NULL
;
685 alc
= kmalloc_node(memsize
, gfp
, node
);
686 init_arraycache(&alc
->ac
, entries
, batch
);
687 spin_lock_init(&alc
->lock
);
691 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
693 struct alien_cache
**alc_ptr
;
694 size_t memsize
= sizeof(void *) * nr_node_ids
;
699 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
704 if (i
== node
|| !node_online(i
))
706 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
708 for (i
--; i
>= 0; i
--)
717 static void free_alien_cache(struct alien_cache
**alc_ptr
)
728 static void __drain_alien_cache(struct kmem_cache
*cachep
,
729 struct array_cache
*ac
, int node
,
730 struct list_head
*list
)
732 struct kmem_cache_node
*n
= get_node(cachep
, node
);
735 spin_lock(&n
->list_lock
);
737 * Stuff objects into the remote nodes shared array first.
738 * That way we could avoid the overhead of putting the objects
739 * into the free lists and getting them back later.
742 transfer_objects(n
->shared
, ac
, ac
->limit
);
744 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
746 spin_unlock(&n
->list_lock
);
751 * Called from cache_reap() to regularly drain alien caches round robin.
753 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
755 int node
= __this_cpu_read(slab_reap_node
);
758 struct alien_cache
*alc
= n
->alien
[node
];
759 struct array_cache
*ac
;
763 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
766 __drain_alien_cache(cachep
, ac
, node
, &list
);
767 spin_unlock_irq(&alc
->lock
);
768 slabs_destroy(cachep
, &list
);
774 static void drain_alien_cache(struct kmem_cache
*cachep
,
775 struct alien_cache
**alien
)
778 struct alien_cache
*alc
;
779 struct array_cache
*ac
;
782 for_each_online_node(i
) {
788 spin_lock_irqsave(&alc
->lock
, flags
);
789 __drain_alien_cache(cachep
, ac
, i
, &list
);
790 spin_unlock_irqrestore(&alc
->lock
, flags
);
791 slabs_destroy(cachep
, &list
);
796 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
797 int node
, int page_node
)
799 struct kmem_cache_node
*n
;
800 struct alien_cache
*alien
= NULL
;
801 struct array_cache
*ac
;
804 n
= get_node(cachep
, node
);
805 STATS_INC_NODEFREES(cachep
);
806 if (n
->alien
&& n
->alien
[page_node
]) {
807 alien
= n
->alien
[page_node
];
809 spin_lock(&alien
->lock
);
810 if (unlikely(ac
->avail
== ac
->limit
)) {
811 STATS_INC_ACOVERFLOW(cachep
);
812 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
814 ac
->entry
[ac
->avail
++] = objp
;
815 spin_unlock(&alien
->lock
);
816 slabs_destroy(cachep
, &list
);
818 n
= get_node(cachep
, page_node
);
819 spin_lock(&n
->list_lock
);
820 free_block(cachep
, &objp
, 1, page_node
, &list
);
821 spin_unlock(&n
->list_lock
);
822 slabs_destroy(cachep
, &list
);
827 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
829 int page_node
= page_to_nid(virt_to_page(objp
));
830 int node
= numa_mem_id();
832 * Make sure we are not freeing a object from another node to the array
835 if (likely(node
== page_node
))
838 return __cache_free_alien(cachep
, objp
, node
, page_node
);
842 * Construct gfp mask to allocate from a specific node but do not reclaim or
843 * warn about failures.
845 static inline gfp_t
gfp_exact_node(gfp_t flags
)
847 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
851 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
853 struct kmem_cache_node
*n
;
856 * Set up the kmem_cache_node for cpu before we can
857 * begin anything. Make sure some other cpu on this
858 * node has not already allocated this
860 n
= get_node(cachep
, node
);
862 spin_lock_irq(&n
->list_lock
);
863 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
865 spin_unlock_irq(&n
->list_lock
);
870 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
874 kmem_cache_node_init(n
);
875 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
876 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
879 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
882 * The kmem_cache_nodes don't come and go as CPUs
883 * come and go. slab_mutex is sufficient
886 cachep
->node
[node
] = n
;
892 * Allocates and initializes node for a node on each slab cache, used for
893 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
894 * will be allocated off-node since memory is not yet online for the new node.
895 * When hotplugging memory or a cpu, existing node are not replaced if
898 * Must hold slab_mutex.
900 static int init_cache_node_node(int node
)
903 struct kmem_cache
*cachep
;
905 list_for_each_entry(cachep
, &slab_caches
, list
) {
906 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
914 static void cpuup_canceled(long cpu
)
916 struct kmem_cache
*cachep
;
917 struct kmem_cache_node
*n
= NULL
;
918 int node
= cpu_to_mem(cpu
);
919 const struct cpumask
*mask
= cpumask_of_node(node
);
921 list_for_each_entry(cachep
, &slab_caches
, list
) {
922 struct array_cache
*nc
;
923 struct array_cache
*shared
;
924 struct alien_cache
**alien
;
927 n
= get_node(cachep
, node
);
931 spin_lock_irq(&n
->list_lock
);
933 /* Free limit for this kmem_cache_node */
934 n
->free_limit
-= cachep
->batchcount
;
936 /* cpu is dead; no one can alloc from it. */
937 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
939 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
943 if (!cpumask_empty(mask
)) {
944 spin_unlock_irq(&n
->list_lock
);
950 free_block(cachep
, shared
->entry
,
951 shared
->avail
, node
, &list
);
958 spin_unlock_irq(&n
->list_lock
);
962 drain_alien_cache(cachep
, alien
);
963 free_alien_cache(alien
);
967 slabs_destroy(cachep
, &list
);
970 * In the previous loop, all the objects were freed to
971 * the respective cache's slabs, now we can go ahead and
972 * shrink each nodelist to its limit.
974 list_for_each_entry(cachep
, &slab_caches
, list
) {
975 n
= get_node(cachep
, node
);
978 drain_freelist(cachep
, n
, INT_MAX
);
982 static int cpuup_prepare(long cpu
)
984 struct kmem_cache
*cachep
;
985 struct kmem_cache_node
*n
= NULL
;
986 int node
= cpu_to_mem(cpu
);
990 * We need to do this right in the beginning since
991 * alloc_arraycache's are going to use this list.
992 * kmalloc_node allows us to add the slab to the right
993 * kmem_cache_node and not this cpu's kmem_cache_node
995 err
= init_cache_node_node(node
);
1000 * Now we can go ahead with allocating the shared arrays and
1003 list_for_each_entry(cachep
, &slab_caches
, list
) {
1004 struct array_cache
*shared
= NULL
;
1005 struct alien_cache
**alien
= NULL
;
1007 if (cachep
->shared
) {
1008 shared
= alloc_arraycache(node
,
1009 cachep
->shared
* cachep
->batchcount
,
1010 0xbaadf00d, GFP_KERNEL
);
1014 if (use_alien_caches
) {
1015 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1021 n
= get_node(cachep
, node
);
1024 spin_lock_irq(&n
->list_lock
);
1027 * We are serialised from CPU_DEAD or
1028 * CPU_UP_CANCELLED by the cpucontrol lock
1039 spin_unlock_irq(&n
->list_lock
);
1041 free_alien_cache(alien
);
1046 cpuup_canceled(cpu
);
1050 static int cpuup_callback(struct notifier_block
*nfb
,
1051 unsigned long action
, void *hcpu
)
1053 long cpu
= (long)hcpu
;
1057 case CPU_UP_PREPARE
:
1058 case CPU_UP_PREPARE_FROZEN
:
1059 mutex_lock(&slab_mutex
);
1060 err
= cpuup_prepare(cpu
);
1061 mutex_unlock(&slab_mutex
);
1064 case CPU_ONLINE_FROZEN
:
1065 start_cpu_timer(cpu
);
1067 #ifdef CONFIG_HOTPLUG_CPU
1068 case CPU_DOWN_PREPARE
:
1069 case CPU_DOWN_PREPARE_FROZEN
:
1071 * Shutdown cache reaper. Note that the slab_mutex is
1072 * held so that if cache_reap() is invoked it cannot do
1073 * anything expensive but will only modify reap_work
1074 * and reschedule the timer.
1076 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1077 /* Now the cache_reaper is guaranteed to be not running. */
1078 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1080 case CPU_DOWN_FAILED
:
1081 case CPU_DOWN_FAILED_FROZEN
:
1082 start_cpu_timer(cpu
);
1085 case CPU_DEAD_FROZEN
:
1087 * Even if all the cpus of a node are down, we don't free the
1088 * kmem_cache_node of any cache. This to avoid a race between
1089 * cpu_down, and a kmalloc allocation from another cpu for
1090 * memory from the node of the cpu going down. The node
1091 * structure is usually allocated from kmem_cache_create() and
1092 * gets destroyed at kmem_cache_destroy().
1096 case CPU_UP_CANCELED
:
1097 case CPU_UP_CANCELED_FROZEN
:
1098 mutex_lock(&slab_mutex
);
1099 cpuup_canceled(cpu
);
1100 mutex_unlock(&slab_mutex
);
1103 return notifier_from_errno(err
);
1106 static struct notifier_block cpucache_notifier
= {
1107 &cpuup_callback
, NULL
, 0
1110 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1112 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1113 * Returns -EBUSY if all objects cannot be drained so that the node is not
1116 * Must hold slab_mutex.
1118 static int __meminit
drain_cache_node_node(int node
)
1120 struct kmem_cache
*cachep
;
1123 list_for_each_entry(cachep
, &slab_caches
, list
) {
1124 struct kmem_cache_node
*n
;
1126 n
= get_node(cachep
, node
);
1130 drain_freelist(cachep
, n
, INT_MAX
);
1132 if (!list_empty(&n
->slabs_full
) ||
1133 !list_empty(&n
->slabs_partial
)) {
1141 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1142 unsigned long action
, void *arg
)
1144 struct memory_notify
*mnb
= arg
;
1148 nid
= mnb
->status_change_nid
;
1153 case MEM_GOING_ONLINE
:
1154 mutex_lock(&slab_mutex
);
1155 ret
= init_cache_node_node(nid
);
1156 mutex_unlock(&slab_mutex
);
1158 case MEM_GOING_OFFLINE
:
1159 mutex_lock(&slab_mutex
);
1160 ret
= drain_cache_node_node(nid
);
1161 mutex_unlock(&slab_mutex
);
1165 case MEM_CANCEL_ONLINE
:
1166 case MEM_CANCEL_OFFLINE
:
1170 return notifier_from_errno(ret
);
1172 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1175 * swap the static kmem_cache_node with kmalloced memory
1177 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1180 struct kmem_cache_node
*ptr
;
1182 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1185 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1187 * Do not assume that spinlocks can be initialized via memcpy:
1189 spin_lock_init(&ptr
->list_lock
);
1191 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1192 cachep
->node
[nodeid
] = ptr
;
1196 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1197 * size of kmem_cache_node.
1199 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1203 for_each_online_node(node
) {
1204 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1205 cachep
->node
[node
]->next_reap
= jiffies
+
1207 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1212 * Initialisation. Called after the page allocator have been initialised and
1213 * before smp_init().
1215 void __init
kmem_cache_init(void)
1219 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1220 sizeof(struct rcu_head
));
1221 kmem_cache
= &kmem_cache_boot
;
1223 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1224 use_alien_caches
= 0;
1226 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1227 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1230 * Fragmentation resistance on low memory - only use bigger
1231 * page orders on machines with more than 32MB of memory if
1232 * not overridden on the command line.
1234 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1235 slab_max_order
= SLAB_MAX_ORDER_HI
;
1237 /* Bootstrap is tricky, because several objects are allocated
1238 * from caches that do not exist yet:
1239 * 1) initialize the kmem_cache cache: it contains the struct
1240 * kmem_cache structures of all caches, except kmem_cache itself:
1241 * kmem_cache is statically allocated.
1242 * Initially an __init data area is used for the head array and the
1243 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1244 * array at the end of the bootstrap.
1245 * 2) Create the first kmalloc cache.
1246 * The struct kmem_cache for the new cache is allocated normally.
1247 * An __init data area is used for the head array.
1248 * 3) Create the remaining kmalloc caches, with minimally sized
1250 * 4) Replace the __init data head arrays for kmem_cache and the first
1251 * kmalloc cache with kmalloc allocated arrays.
1252 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1253 * the other cache's with kmalloc allocated memory.
1254 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1257 /* 1) create the kmem_cache */
1260 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1262 create_boot_cache(kmem_cache
, "kmem_cache",
1263 offsetof(struct kmem_cache
, node
) +
1264 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1265 SLAB_HWCACHE_ALIGN
);
1266 list_add(&kmem_cache
->list
, &slab_caches
);
1267 slab_state
= PARTIAL
;
1270 * Initialize the caches that provide memory for the kmem_cache_node
1271 * structures first. Without this, further allocations will bug.
1273 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1274 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1275 slab_state
= PARTIAL_NODE
;
1276 setup_kmalloc_cache_index_table();
1278 slab_early_init
= 0;
1280 /* 5) Replace the bootstrap kmem_cache_node */
1284 for_each_online_node(nid
) {
1285 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1287 init_list(kmalloc_caches
[INDEX_NODE
],
1288 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1292 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1295 void __init
kmem_cache_init_late(void)
1297 struct kmem_cache
*cachep
;
1301 /* 6) resize the head arrays to their final sizes */
1302 mutex_lock(&slab_mutex
);
1303 list_for_each_entry(cachep
, &slab_caches
, list
)
1304 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1306 mutex_unlock(&slab_mutex
);
1312 * Register a cpu startup notifier callback that initializes
1313 * cpu_cache_get for all new cpus
1315 register_cpu_notifier(&cpucache_notifier
);
1319 * Register a memory hotplug callback that initializes and frees
1322 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1326 * The reap timers are started later, with a module init call: That part
1327 * of the kernel is not yet operational.
1331 static int __init
cpucache_init(void)
1336 * Register the timers that return unneeded pages to the page allocator
1338 for_each_online_cpu(cpu
)
1339 start_cpu_timer(cpu
);
1345 __initcall(cpucache_init
);
1347 static noinline
void
1348 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1351 struct kmem_cache_node
*n
;
1353 unsigned long flags
;
1355 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1356 DEFAULT_RATELIMIT_BURST
);
1358 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1361 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1362 nodeid
, gfpflags
, &gfpflags
);
1363 pr_warn(" cache: %s, object size: %d, order: %d\n",
1364 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1366 for_each_kmem_cache_node(cachep
, node
, n
) {
1367 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1368 unsigned long active_slabs
= 0, num_slabs
= 0;
1370 spin_lock_irqsave(&n
->list_lock
, flags
);
1371 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1372 active_objs
+= cachep
->num
;
1375 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1376 active_objs
+= page
->active
;
1379 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1382 free_objects
+= n
->free_objects
;
1383 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1385 num_slabs
+= active_slabs
;
1386 num_objs
= num_slabs
* cachep
->num
;
1387 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1388 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1395 * Interface to system's page allocator. No need to hold the
1396 * kmem_cache_node ->list_lock.
1398 * If we requested dmaable memory, we will get it. Even if we
1399 * did not request dmaable memory, we might get it, but that
1400 * would be relatively rare and ignorable.
1402 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1408 flags
|= cachep
->allocflags
;
1409 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1410 flags
|= __GFP_RECLAIMABLE
;
1412 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1414 slab_out_of_memory(cachep
, flags
, nodeid
);
1418 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1419 __free_pages(page
, cachep
->gfporder
);
1423 nr_pages
= (1 << cachep
->gfporder
);
1424 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1425 add_zone_page_state(page_zone(page
),
1426 NR_SLAB_RECLAIMABLE
, nr_pages
);
1428 add_zone_page_state(page_zone(page
),
1429 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1431 __SetPageSlab(page
);
1432 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1433 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1434 SetPageSlabPfmemalloc(page
);
1436 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1437 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1440 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1442 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1449 * Interface to system's page release.
1451 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1453 int order
= cachep
->gfporder
;
1454 unsigned long nr_freed
= (1 << order
);
1456 kmemcheck_free_shadow(page
, order
);
1458 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1459 sub_zone_page_state(page_zone(page
),
1460 NR_SLAB_RECLAIMABLE
, nr_freed
);
1462 sub_zone_page_state(page_zone(page
),
1463 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1465 BUG_ON(!PageSlab(page
));
1466 __ClearPageSlabPfmemalloc(page
);
1467 __ClearPageSlab(page
);
1468 page_mapcount_reset(page
);
1469 page
->mapping
= NULL
;
1471 if (current
->reclaim_state
)
1472 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1473 memcg_uncharge_slab(page
, order
, cachep
);
1474 __free_pages(page
, order
);
1477 static void kmem_rcu_free(struct rcu_head
*head
)
1479 struct kmem_cache
*cachep
;
1482 page
= container_of(head
, struct page
, rcu_head
);
1483 cachep
= page
->slab_cache
;
1485 kmem_freepages(cachep
, page
);
1489 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1491 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1492 (cachep
->size
% PAGE_SIZE
) == 0)
1498 #ifdef CONFIG_DEBUG_PAGEALLOC
1499 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1500 unsigned long caller
)
1502 int size
= cachep
->object_size
;
1504 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1506 if (size
< 5 * sizeof(unsigned long))
1509 *addr
++ = 0x12345678;
1511 *addr
++ = smp_processor_id();
1512 size
-= 3 * sizeof(unsigned long);
1514 unsigned long *sptr
= &caller
;
1515 unsigned long svalue
;
1517 while (!kstack_end(sptr
)) {
1519 if (kernel_text_address(svalue
)) {
1521 size
-= sizeof(unsigned long);
1522 if (size
<= sizeof(unsigned long))
1528 *addr
++ = 0x87654321;
1531 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1532 int map
, unsigned long caller
)
1534 if (!is_debug_pagealloc_cache(cachep
))
1538 store_stackinfo(cachep
, objp
, caller
);
1540 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1544 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1545 int map
, unsigned long caller
) {}
1549 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1551 int size
= cachep
->object_size
;
1552 addr
= &((char *)addr
)[obj_offset(cachep
)];
1554 memset(addr
, val
, size
);
1555 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1558 static void dump_line(char *data
, int offset
, int limit
)
1561 unsigned char error
= 0;
1564 pr_err("%03x: ", offset
);
1565 for (i
= 0; i
< limit
; i
++) {
1566 if (data
[offset
+ i
] != POISON_FREE
) {
1567 error
= data
[offset
+ i
];
1571 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1572 &data
[offset
], limit
, 1);
1574 if (bad_count
== 1) {
1575 error
^= POISON_FREE
;
1576 if (!(error
& (error
- 1))) {
1577 pr_err("Single bit error detected. Probably bad RAM.\n");
1579 pr_err("Run memtest86+ or a similar memory test tool.\n");
1581 pr_err("Run a memory test tool.\n");
1590 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1595 if (cachep
->flags
& SLAB_RED_ZONE
) {
1596 pr_err("Redzone: 0x%llx/0x%llx\n",
1597 *dbg_redzone1(cachep
, objp
),
1598 *dbg_redzone2(cachep
, objp
));
1601 if (cachep
->flags
& SLAB_STORE_USER
) {
1602 pr_err("Last user: [<%p>](%pSR)\n",
1603 *dbg_userword(cachep
, objp
),
1604 *dbg_userword(cachep
, objp
));
1606 realobj
= (char *)objp
+ obj_offset(cachep
);
1607 size
= cachep
->object_size
;
1608 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1611 if (i
+ limit
> size
)
1613 dump_line(realobj
, i
, limit
);
1617 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1623 if (is_debug_pagealloc_cache(cachep
))
1626 realobj
= (char *)objp
+ obj_offset(cachep
);
1627 size
= cachep
->object_size
;
1629 for (i
= 0; i
< size
; i
++) {
1630 char exp
= POISON_FREE
;
1633 if (realobj
[i
] != exp
) {
1638 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1639 print_tainted(), cachep
->name
,
1641 print_objinfo(cachep
, objp
, 0);
1643 /* Hexdump the affected line */
1646 if (i
+ limit
> size
)
1648 dump_line(realobj
, i
, limit
);
1651 /* Limit to 5 lines */
1657 /* Print some data about the neighboring objects, if they
1660 struct page
*page
= virt_to_head_page(objp
);
1663 objnr
= obj_to_index(cachep
, page
, objp
);
1665 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1666 realobj
= (char *)objp
+ obj_offset(cachep
);
1667 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1668 print_objinfo(cachep
, objp
, 2);
1670 if (objnr
+ 1 < cachep
->num
) {
1671 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1672 realobj
= (char *)objp
+ obj_offset(cachep
);
1673 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1674 print_objinfo(cachep
, objp
, 2);
1681 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1686 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1687 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1691 for (i
= 0; i
< cachep
->num
; i
++) {
1692 void *objp
= index_to_obj(cachep
, page
, i
);
1694 if (cachep
->flags
& SLAB_POISON
) {
1695 check_poison_obj(cachep
, objp
);
1696 slab_kernel_map(cachep
, objp
, 1, 0);
1698 if (cachep
->flags
& SLAB_RED_ZONE
) {
1699 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1700 slab_error(cachep
, "start of a freed object was overwritten");
1701 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1702 slab_error(cachep
, "end of a freed object was overwritten");
1707 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1714 * slab_destroy - destroy and release all objects in a slab
1715 * @cachep: cache pointer being destroyed
1716 * @page: page pointer being destroyed
1718 * Destroy all the objs in a slab page, and release the mem back to the system.
1719 * Before calling the slab page must have been unlinked from the cache. The
1720 * kmem_cache_node ->list_lock is not held/needed.
1722 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1726 freelist
= page
->freelist
;
1727 slab_destroy_debugcheck(cachep
, page
);
1728 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1729 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1731 kmem_freepages(cachep
, page
);
1734 * From now on, we don't use freelist
1735 * although actual page can be freed in rcu context
1737 if (OFF_SLAB(cachep
))
1738 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1741 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1743 struct page
*page
, *n
;
1745 list_for_each_entry_safe(page
, n
, list
, lru
) {
1746 list_del(&page
->lru
);
1747 slab_destroy(cachep
, page
);
1752 * calculate_slab_order - calculate size (page order) of slabs
1753 * @cachep: pointer to the cache that is being created
1754 * @size: size of objects to be created in this cache.
1755 * @flags: slab allocation flags
1757 * Also calculates the number of objects per slab.
1759 * This could be made much more intelligent. For now, try to avoid using
1760 * high order pages for slabs. When the gfp() functions are more friendly
1761 * towards high-order requests, this should be changed.
1763 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1764 size_t size
, unsigned long flags
)
1766 size_t left_over
= 0;
1769 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1773 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1777 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1778 if (num
> SLAB_OBJ_MAX_NUM
)
1781 if (flags
& CFLGS_OFF_SLAB
) {
1782 struct kmem_cache
*freelist_cache
;
1783 size_t freelist_size
;
1785 freelist_size
= num
* sizeof(freelist_idx_t
);
1786 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1787 if (!freelist_cache
)
1791 * Needed to avoid possible looping condition
1794 if (OFF_SLAB(freelist_cache
))
1797 /* check if off slab has enough benefit */
1798 if (freelist_cache
->size
> cachep
->size
/ 2)
1802 /* Found something acceptable - save it away */
1804 cachep
->gfporder
= gfporder
;
1805 left_over
= remainder
;
1808 * A VFS-reclaimable slab tends to have most allocations
1809 * as GFP_NOFS and we really don't want to have to be allocating
1810 * higher-order pages when we are unable to shrink dcache.
1812 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1816 * Large number of objects is good, but very large slabs are
1817 * currently bad for the gfp()s.
1819 if (gfporder
>= slab_max_order
)
1823 * Acceptable internal fragmentation?
1825 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1831 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1832 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1836 struct array_cache __percpu
*cpu_cache
;
1838 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1839 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1844 for_each_possible_cpu(cpu
) {
1845 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1846 entries
, batchcount
);
1852 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1854 if (slab_state
>= FULL
)
1855 return enable_cpucache(cachep
, gfp
);
1857 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1858 if (!cachep
->cpu_cache
)
1861 if (slab_state
== DOWN
) {
1862 /* Creation of first cache (kmem_cache). */
1863 set_up_node(kmem_cache
, CACHE_CACHE
);
1864 } else if (slab_state
== PARTIAL
) {
1865 /* For kmem_cache_node */
1866 set_up_node(cachep
, SIZE_NODE
);
1870 for_each_online_node(node
) {
1871 cachep
->node
[node
] = kmalloc_node(
1872 sizeof(struct kmem_cache_node
), gfp
, node
);
1873 BUG_ON(!cachep
->node
[node
]);
1874 kmem_cache_node_init(cachep
->node
[node
]);
1878 cachep
->node
[numa_mem_id()]->next_reap
=
1879 jiffies
+ REAPTIMEOUT_NODE
+
1880 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1882 cpu_cache_get(cachep
)->avail
= 0;
1883 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1884 cpu_cache_get(cachep
)->batchcount
= 1;
1885 cpu_cache_get(cachep
)->touched
= 0;
1886 cachep
->batchcount
= 1;
1887 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1891 unsigned long kmem_cache_flags(unsigned long object_size
,
1892 unsigned long flags
, const char *name
,
1893 void (*ctor
)(void *))
1899 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1900 unsigned long flags
, void (*ctor
)(void *))
1902 struct kmem_cache
*cachep
;
1904 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1909 * Adjust the object sizes so that we clear
1910 * the complete object on kzalloc.
1912 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1917 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1918 size_t size
, unsigned long flags
)
1924 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1927 left
= calculate_slab_order(cachep
, size
,
1928 flags
| CFLGS_OBJFREELIST_SLAB
);
1932 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1935 cachep
->colour
= left
/ cachep
->colour_off
;
1940 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1941 size_t size
, unsigned long flags
)
1948 * Always use on-slab management when SLAB_NOLEAKTRACE
1949 * to avoid recursive calls into kmemleak.
1951 if (flags
& SLAB_NOLEAKTRACE
)
1955 * Size is large, assume best to place the slab management obj
1956 * off-slab (should allow better packing of objs).
1958 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1963 * If the slab has been placed off-slab, and we have enough space then
1964 * move it on-slab. This is at the expense of any extra colouring.
1966 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1969 cachep
->colour
= left
/ cachep
->colour_off
;
1974 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1975 size_t size
, unsigned long flags
)
1981 left
= calculate_slab_order(cachep
, size
, flags
);
1985 cachep
->colour
= left
/ cachep
->colour_off
;
1991 * __kmem_cache_create - Create a cache.
1992 * @cachep: cache management descriptor
1993 * @flags: SLAB flags
1995 * Returns a ptr to the cache on success, NULL on failure.
1996 * Cannot be called within a int, but can be interrupted.
1997 * The @ctor is run when new pages are allocated by the cache.
2001 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2002 * to catch references to uninitialised memory.
2004 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2005 * for buffer overruns.
2007 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2008 * cacheline. This can be beneficial if you're counting cycles as closely
2012 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2014 size_t ralign
= BYTES_PER_WORD
;
2017 size_t size
= cachep
->size
;
2022 * Enable redzoning and last user accounting, except for caches with
2023 * large objects, if the increased size would increase the object size
2024 * above the next power of two: caches with object sizes just above a
2025 * power of two have a significant amount of internal fragmentation.
2027 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2028 2 * sizeof(unsigned long long)))
2029 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2030 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2031 flags
|= SLAB_POISON
;
2036 * Check that size is in terms of words. This is needed to avoid
2037 * unaligned accesses for some archs when redzoning is used, and makes
2038 * sure any on-slab bufctl's are also correctly aligned.
2040 if (size
& (BYTES_PER_WORD
- 1)) {
2041 size
+= (BYTES_PER_WORD
- 1);
2042 size
&= ~(BYTES_PER_WORD
- 1);
2045 if (flags
& SLAB_RED_ZONE
) {
2046 ralign
= REDZONE_ALIGN
;
2047 /* If redzoning, ensure that the second redzone is suitably
2048 * aligned, by adjusting the object size accordingly. */
2049 size
+= REDZONE_ALIGN
- 1;
2050 size
&= ~(REDZONE_ALIGN
- 1);
2053 /* 3) caller mandated alignment */
2054 if (ralign
< cachep
->align
) {
2055 ralign
= cachep
->align
;
2057 /* disable debug if necessary */
2058 if (ralign
> __alignof__(unsigned long long))
2059 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2063 cachep
->align
= ralign
;
2064 cachep
->colour_off
= cache_line_size();
2065 /* Offset must be a multiple of the alignment. */
2066 if (cachep
->colour_off
< cachep
->align
)
2067 cachep
->colour_off
= cachep
->align
;
2069 if (slab_is_available())
2077 * Both debugging options require word-alignment which is calculated
2080 if (flags
& SLAB_RED_ZONE
) {
2081 /* add space for red zone words */
2082 cachep
->obj_offset
+= sizeof(unsigned long long);
2083 size
+= 2 * sizeof(unsigned long long);
2085 if (flags
& SLAB_STORE_USER
) {
2086 /* user store requires one word storage behind the end of
2087 * the real object. But if the second red zone needs to be
2088 * aligned to 64 bits, we must allow that much space.
2090 if (flags
& SLAB_RED_ZONE
)
2091 size
+= REDZONE_ALIGN
;
2093 size
+= BYTES_PER_WORD
;
2097 kasan_cache_create(cachep
, &size
, &flags
);
2099 size
= ALIGN(size
, cachep
->align
);
2101 * We should restrict the number of objects in a slab to implement
2102 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2104 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2105 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2109 * To activate debug pagealloc, off-slab management is necessary
2110 * requirement. In early phase of initialization, small sized slab
2111 * doesn't get initialized so it would not be possible. So, we need
2112 * to check size >= 256. It guarantees that all necessary small
2113 * sized slab is initialized in current slab initialization sequence.
2115 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2116 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2117 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2118 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2120 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2121 flags
|= CFLGS_OFF_SLAB
;
2122 cachep
->obj_offset
+= tmp_size
- size
;
2130 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2131 flags
|= CFLGS_OBJFREELIST_SLAB
;
2135 if (set_off_slab_cache(cachep
, size
, flags
)) {
2136 flags
|= CFLGS_OFF_SLAB
;
2140 if (set_on_slab_cache(cachep
, size
, flags
))
2146 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2147 cachep
->flags
= flags
;
2148 cachep
->allocflags
= __GFP_COMP
;
2149 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2150 cachep
->allocflags
|= GFP_DMA
;
2151 cachep
->size
= size
;
2152 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2156 * If we're going to use the generic kernel_map_pages()
2157 * poisoning, then it's going to smash the contents of
2158 * the redzone and userword anyhow, so switch them off.
2160 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2161 (cachep
->flags
& SLAB_POISON
) &&
2162 is_debug_pagealloc_cache(cachep
))
2163 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2166 if (OFF_SLAB(cachep
)) {
2167 cachep
->freelist_cache
=
2168 kmalloc_slab(cachep
->freelist_size
, 0u);
2171 err
= setup_cpu_cache(cachep
, gfp
);
2173 __kmem_cache_release(cachep
);
2181 static void check_irq_off(void)
2183 BUG_ON(!irqs_disabled());
2186 static void check_irq_on(void)
2188 BUG_ON(irqs_disabled());
2191 static void check_mutex_acquired(void)
2193 BUG_ON(!mutex_is_locked(&slab_mutex
));
2196 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2200 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2204 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2208 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2213 #define check_irq_off() do { } while(0)
2214 #define check_irq_on() do { } while(0)
2215 #define check_mutex_acquired() do { } while(0)
2216 #define check_spinlock_acquired(x) do { } while(0)
2217 #define check_spinlock_acquired_node(x, y) do { } while(0)
2220 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2221 int node
, bool free_all
, struct list_head
*list
)
2225 if (!ac
|| !ac
->avail
)
2228 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2229 if (tofree
> ac
->avail
)
2230 tofree
= (ac
->avail
+ 1) / 2;
2232 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2233 ac
->avail
-= tofree
;
2234 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2237 static void do_drain(void *arg
)
2239 struct kmem_cache
*cachep
= arg
;
2240 struct array_cache
*ac
;
2241 int node
= numa_mem_id();
2242 struct kmem_cache_node
*n
;
2246 ac
= cpu_cache_get(cachep
);
2247 n
= get_node(cachep
, node
);
2248 spin_lock(&n
->list_lock
);
2249 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2250 spin_unlock(&n
->list_lock
);
2251 slabs_destroy(cachep
, &list
);
2255 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2257 struct kmem_cache_node
*n
;
2261 on_each_cpu(do_drain
, cachep
, 1);
2263 for_each_kmem_cache_node(cachep
, node
, n
)
2265 drain_alien_cache(cachep
, n
->alien
);
2267 for_each_kmem_cache_node(cachep
, node
, n
) {
2268 spin_lock_irq(&n
->list_lock
);
2269 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2270 spin_unlock_irq(&n
->list_lock
);
2272 slabs_destroy(cachep
, &list
);
2277 * Remove slabs from the list of free slabs.
2278 * Specify the number of slabs to drain in tofree.
2280 * Returns the actual number of slabs released.
2282 static int drain_freelist(struct kmem_cache
*cache
,
2283 struct kmem_cache_node
*n
, int tofree
)
2285 struct list_head
*p
;
2290 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2292 spin_lock_irq(&n
->list_lock
);
2293 p
= n
->slabs_free
.prev
;
2294 if (p
== &n
->slabs_free
) {
2295 spin_unlock_irq(&n
->list_lock
);
2299 page
= list_entry(p
, struct page
, lru
);
2300 list_del(&page
->lru
);
2302 * Safe to drop the lock. The slab is no longer linked
2305 n
->free_objects
-= cache
->num
;
2306 spin_unlock_irq(&n
->list_lock
);
2307 slab_destroy(cache
, page
);
2314 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2318 struct kmem_cache_node
*n
;
2320 drain_cpu_caches(cachep
);
2323 for_each_kmem_cache_node(cachep
, node
, n
) {
2324 drain_freelist(cachep
, n
, INT_MAX
);
2326 ret
+= !list_empty(&n
->slabs_full
) ||
2327 !list_empty(&n
->slabs_partial
);
2329 return (ret
? 1 : 0);
2332 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2334 return __kmem_cache_shrink(cachep
, false);
2337 void __kmem_cache_release(struct kmem_cache
*cachep
)
2340 struct kmem_cache_node
*n
;
2342 free_percpu(cachep
->cpu_cache
);
2344 /* NUMA: free the node structures */
2345 for_each_kmem_cache_node(cachep
, i
, n
) {
2347 free_alien_cache(n
->alien
);
2349 cachep
->node
[i
] = NULL
;
2354 * Get the memory for a slab management obj.
2356 * For a slab cache when the slab descriptor is off-slab, the
2357 * slab descriptor can't come from the same cache which is being created,
2358 * Because if it is the case, that means we defer the creation of
2359 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2360 * And we eventually call down to __kmem_cache_create(), which
2361 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2362 * This is a "chicken-and-egg" problem.
2364 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2365 * which are all initialized during kmem_cache_init().
2367 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2368 struct page
*page
, int colour_off
,
2369 gfp_t local_flags
, int nodeid
)
2372 void *addr
= page_address(page
);
2374 page
->s_mem
= addr
+ colour_off
;
2377 if (OBJFREELIST_SLAB(cachep
))
2379 else if (OFF_SLAB(cachep
)) {
2380 /* Slab management obj is off-slab. */
2381 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2382 local_flags
, nodeid
);
2386 /* We will use last bytes at the slab for freelist */
2387 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2388 cachep
->freelist_size
;
2394 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2396 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2399 static inline void set_free_obj(struct page
*page
,
2400 unsigned int idx
, freelist_idx_t val
)
2402 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2405 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2410 for (i
= 0; i
< cachep
->num
; i
++) {
2411 void *objp
= index_to_obj(cachep
, page
, i
);
2413 if (cachep
->flags
& SLAB_STORE_USER
)
2414 *dbg_userword(cachep
, objp
) = NULL
;
2416 if (cachep
->flags
& SLAB_RED_ZONE
) {
2417 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2418 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2421 * Constructors are not allowed to allocate memory from the same
2422 * cache which they are a constructor for. Otherwise, deadlock.
2423 * They must also be threaded.
2425 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2426 kasan_unpoison_object_data(cachep
,
2427 objp
+ obj_offset(cachep
));
2428 cachep
->ctor(objp
+ obj_offset(cachep
));
2429 kasan_poison_object_data(
2430 cachep
, objp
+ obj_offset(cachep
));
2433 if (cachep
->flags
& SLAB_RED_ZONE
) {
2434 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2435 slab_error(cachep
, "constructor overwrote the end of an object");
2436 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2437 slab_error(cachep
, "constructor overwrote the start of an object");
2439 /* need to poison the objs? */
2440 if (cachep
->flags
& SLAB_POISON
) {
2441 poison_obj(cachep
, objp
, POISON_FREE
);
2442 slab_kernel_map(cachep
, objp
, 0, 0);
2448 static void cache_init_objs(struct kmem_cache
*cachep
,
2454 cache_init_objs_debug(cachep
, page
);
2456 if (OBJFREELIST_SLAB(cachep
)) {
2457 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2461 for (i
= 0; i
< cachep
->num
; i
++) {
2462 /* constructor could break poison info */
2463 if (DEBUG
== 0 && cachep
->ctor
) {
2464 objp
= index_to_obj(cachep
, page
, i
);
2465 kasan_unpoison_object_data(cachep
, objp
);
2467 kasan_poison_object_data(cachep
, objp
);
2470 set_free_obj(page
, i
, i
);
2474 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2476 if (CONFIG_ZONE_DMA_FLAG
) {
2477 if (flags
& GFP_DMA
)
2478 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2480 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2484 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2488 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2492 if (cachep
->flags
& SLAB_STORE_USER
)
2493 set_store_user_dirty(cachep
);
2499 static void slab_put_obj(struct kmem_cache
*cachep
,
2500 struct page
*page
, void *objp
)
2502 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2506 /* Verify double free bug */
2507 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2508 if (get_free_obj(page
, i
) == objnr
) {
2509 pr_err("slab: double free detected in cache '%s', objp %p\n",
2510 cachep
->name
, objp
);
2516 if (!page
->freelist
)
2517 page
->freelist
= objp
+ obj_offset(cachep
);
2519 set_free_obj(page
, page
->active
, objnr
);
2523 * Map pages beginning at addr to the given cache and slab. This is required
2524 * for the slab allocator to be able to lookup the cache and slab of a
2525 * virtual address for kfree, ksize, and slab debugging.
2527 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2530 page
->slab_cache
= cache
;
2531 page
->freelist
= freelist
;
2535 * Grow (by 1) the number of slabs within a cache. This is called by
2536 * kmem_cache_alloc() when there are no active objs left in a cache.
2538 static int cache_grow(struct kmem_cache
*cachep
,
2539 gfp_t flags
, int nodeid
, struct page
*page
)
2544 struct kmem_cache_node
*n
;
2547 * Be lazy and only check for valid flags here, keeping it out of the
2548 * critical path in kmem_cache_alloc().
2550 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2551 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2554 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2556 /* Take the node list lock to change the colour_next on this node */
2558 n
= get_node(cachep
, nodeid
);
2559 spin_lock(&n
->list_lock
);
2561 /* Get colour for the slab, and cal the next value. */
2562 offset
= n
->colour_next
;
2564 if (n
->colour_next
>= cachep
->colour
)
2566 spin_unlock(&n
->list_lock
);
2568 offset
*= cachep
->colour_off
;
2570 if (gfpflags_allow_blocking(local_flags
))
2574 * The test for missing atomic flag is performed here, rather than
2575 * the more obvious place, simply to reduce the critical path length
2576 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2577 * will eventually be caught here (where it matters).
2579 kmem_flagcheck(cachep
, flags
);
2582 * Get mem for the objs. Attempt to allocate a physical page from
2586 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2590 /* Get slab management. */
2591 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2592 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2593 if (OFF_SLAB(cachep
) && !freelist
)
2596 slab_map_pages(cachep
, page
, freelist
);
2598 kasan_poison_slab(page
);
2599 cache_init_objs(cachep
, page
);
2601 if (gfpflags_allow_blocking(local_flags
))
2602 local_irq_disable();
2604 spin_lock(&n
->list_lock
);
2606 /* Make slab active. */
2607 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2608 STATS_INC_GROWN(cachep
);
2609 n
->free_objects
+= cachep
->num
;
2610 spin_unlock(&n
->list_lock
);
2613 kmem_freepages(cachep
, page
);
2615 if (gfpflags_allow_blocking(local_flags
))
2616 local_irq_disable();
2623 * Perform extra freeing checks:
2624 * - detect bad pointers.
2625 * - POISON/RED_ZONE checking
2627 static void kfree_debugcheck(const void *objp
)
2629 if (!virt_addr_valid(objp
)) {
2630 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2631 (unsigned long)objp
);
2636 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2638 unsigned long long redzone1
, redzone2
;
2640 redzone1
= *dbg_redzone1(cache
, obj
);
2641 redzone2
= *dbg_redzone2(cache
, obj
);
2646 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2649 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2650 slab_error(cache
, "double free detected");
2652 slab_error(cache
, "memory outside object was overwritten");
2654 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2655 obj
, redzone1
, redzone2
);
2658 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2659 unsigned long caller
)
2664 BUG_ON(virt_to_cache(objp
) != cachep
);
2666 objp
-= obj_offset(cachep
);
2667 kfree_debugcheck(objp
);
2668 page
= virt_to_head_page(objp
);
2670 if (cachep
->flags
& SLAB_RED_ZONE
) {
2671 verify_redzone_free(cachep
, objp
);
2672 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2673 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2675 if (cachep
->flags
& SLAB_STORE_USER
) {
2676 set_store_user_dirty(cachep
);
2677 *dbg_userword(cachep
, objp
) = (void *)caller
;
2680 objnr
= obj_to_index(cachep
, page
, objp
);
2682 BUG_ON(objnr
>= cachep
->num
);
2683 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2685 if (cachep
->flags
& SLAB_POISON
) {
2686 poison_obj(cachep
, objp
, POISON_FREE
);
2687 slab_kernel_map(cachep
, objp
, 0, caller
);
2693 #define kfree_debugcheck(x) do { } while(0)
2694 #define cache_free_debugcheck(x,objp,z) (objp)
2697 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2705 objp
= next
- obj_offset(cachep
);
2706 next
= *(void **)next
;
2707 poison_obj(cachep
, objp
, POISON_FREE
);
2712 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2713 struct kmem_cache_node
*n
, struct page
*page
,
2716 /* move slabp to correct slabp list: */
2717 list_del(&page
->lru
);
2718 if (page
->active
== cachep
->num
) {
2719 list_add(&page
->lru
, &n
->slabs_full
);
2720 if (OBJFREELIST_SLAB(cachep
)) {
2722 /* Poisoning will be done without holding the lock */
2723 if (cachep
->flags
& SLAB_POISON
) {
2724 void **objp
= page
->freelist
;
2730 page
->freelist
= NULL
;
2733 list_add(&page
->lru
, &n
->slabs_partial
);
2736 /* Try to find non-pfmemalloc slab if needed */
2737 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2738 struct page
*page
, bool pfmemalloc
)
2746 if (!PageSlabPfmemalloc(page
))
2749 /* No need to keep pfmemalloc slab if we have enough free objects */
2750 if (n
->free_objects
> n
->free_limit
) {
2751 ClearPageSlabPfmemalloc(page
);
2755 /* Move pfmemalloc slab to the end of list to speed up next search */
2756 list_del(&page
->lru
);
2758 list_add_tail(&page
->lru
, &n
->slabs_free
);
2760 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2762 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2763 if (!PageSlabPfmemalloc(page
))
2767 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2768 if (!PageSlabPfmemalloc(page
))
2775 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2779 page
= list_first_entry_or_null(&n
->slabs_partial
,
2782 n
->free_touched
= 1;
2783 page
= list_first_entry_or_null(&n
->slabs_free
,
2787 if (sk_memalloc_socks())
2788 return get_valid_first_slab(n
, page
, pfmemalloc
);
2793 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2794 struct kmem_cache_node
*n
, gfp_t flags
)
2800 if (!gfp_pfmemalloc_allowed(flags
))
2803 spin_lock(&n
->list_lock
);
2804 page
= get_first_slab(n
, true);
2806 spin_unlock(&n
->list_lock
);
2810 obj
= slab_get_obj(cachep
, page
);
2813 fixup_slab_list(cachep
, n
, page
, &list
);
2815 spin_unlock(&n
->list_lock
);
2816 fixup_objfreelist_debug(cachep
, &list
);
2821 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2824 struct kmem_cache_node
*n
;
2825 struct array_cache
*ac
;
2830 node
= numa_mem_id();
2833 ac
= cpu_cache_get(cachep
);
2834 batchcount
= ac
->batchcount
;
2835 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2837 * If there was little recent activity on this cache, then
2838 * perform only a partial refill. Otherwise we could generate
2841 batchcount
= BATCHREFILL_LIMIT
;
2843 n
= get_node(cachep
, node
);
2845 BUG_ON(ac
->avail
> 0 || !n
);
2846 spin_lock(&n
->list_lock
);
2848 /* See if we can refill from the shared array */
2849 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2850 n
->shared
->touched
= 1;
2854 while (batchcount
> 0) {
2856 /* Get slab alloc is to come from. */
2857 page
= get_first_slab(n
, false);
2861 check_spinlock_acquired(cachep
);
2864 * The slab was either on partial or free list so
2865 * there must be at least one object available for
2868 BUG_ON(page
->active
>= cachep
->num
);
2870 while (page
->active
< cachep
->num
&& batchcount
--) {
2871 STATS_INC_ALLOCED(cachep
);
2872 STATS_INC_ACTIVE(cachep
);
2873 STATS_SET_HIGH(cachep
);
2875 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2878 fixup_slab_list(cachep
, n
, page
, &list
);
2882 n
->free_objects
-= ac
->avail
;
2884 spin_unlock(&n
->list_lock
);
2885 fixup_objfreelist_debug(cachep
, &list
);
2887 if (unlikely(!ac
->avail
)) {
2890 /* Check if we can use obj in pfmemalloc slab */
2891 if (sk_memalloc_socks()) {
2892 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2898 x
= cache_grow(cachep
, gfp_exact_node(flags
), node
, NULL
);
2900 /* cache_grow can reenable interrupts, then ac could change. */
2901 ac
= cpu_cache_get(cachep
);
2902 node
= numa_mem_id();
2904 /* no objects in sight? abort */
2905 if (!x
&& ac
->avail
== 0)
2908 if (!ac
->avail
) /* objects refilled by interrupt? */
2913 return ac
->entry
[--ac
->avail
];
2916 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2919 might_sleep_if(gfpflags_allow_blocking(flags
));
2921 kmem_flagcheck(cachep
, flags
);
2926 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2927 gfp_t flags
, void *objp
, unsigned long caller
)
2931 if (cachep
->flags
& SLAB_POISON
) {
2932 check_poison_obj(cachep
, objp
);
2933 slab_kernel_map(cachep
, objp
, 1, 0);
2934 poison_obj(cachep
, objp
, POISON_INUSE
);
2936 if (cachep
->flags
& SLAB_STORE_USER
)
2937 *dbg_userword(cachep
, objp
) = (void *)caller
;
2939 if (cachep
->flags
& SLAB_RED_ZONE
) {
2940 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2941 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2942 slab_error(cachep
, "double free, or memory outside object was overwritten");
2943 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2944 objp
, *dbg_redzone1(cachep
, objp
),
2945 *dbg_redzone2(cachep
, objp
));
2947 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2948 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2951 objp
+= obj_offset(cachep
);
2952 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2954 if (ARCH_SLAB_MINALIGN
&&
2955 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2956 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2957 objp
, (int)ARCH_SLAB_MINALIGN
);
2962 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2965 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2968 struct array_cache
*ac
;
2972 ac
= cpu_cache_get(cachep
);
2973 if (likely(ac
->avail
)) {
2975 objp
= ac
->entry
[--ac
->avail
];
2977 STATS_INC_ALLOCHIT(cachep
);
2981 STATS_INC_ALLOCMISS(cachep
);
2982 objp
= cache_alloc_refill(cachep
, flags
);
2984 * the 'ac' may be updated by cache_alloc_refill(),
2985 * and kmemleak_erase() requires its correct value.
2987 ac
= cpu_cache_get(cachep
);
2991 * To avoid a false negative, if an object that is in one of the
2992 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2993 * treat the array pointers as a reference to the object.
2996 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3002 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3004 * If we are in_interrupt, then process context, including cpusets and
3005 * mempolicy, may not apply and should not be used for allocation policy.
3007 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3009 int nid_alloc
, nid_here
;
3011 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3013 nid_alloc
= nid_here
= numa_mem_id();
3014 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3015 nid_alloc
= cpuset_slab_spread_node();
3016 else if (current
->mempolicy
)
3017 nid_alloc
= mempolicy_slab_node();
3018 if (nid_alloc
!= nid_here
)
3019 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3024 * Fallback function if there was no memory available and no objects on a
3025 * certain node and fall back is permitted. First we scan all the
3026 * available node for available objects. If that fails then we
3027 * perform an allocation without specifying a node. This allows the page
3028 * allocator to do its reclaim / fallback magic. We then insert the
3029 * slab into the proper nodelist and then allocate from it.
3031 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3033 struct zonelist
*zonelist
;
3037 enum zone_type high_zoneidx
= gfp_zone(flags
);
3040 unsigned int cpuset_mems_cookie
;
3042 if (flags
& __GFP_THISNODE
)
3045 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3048 cpuset_mems_cookie
= read_mems_allowed_begin();
3049 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3053 * Look through allowed nodes for objects available
3054 * from existing per node queues.
3056 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3057 nid
= zone_to_nid(zone
);
3059 if (cpuset_zone_allowed(zone
, flags
) &&
3060 get_node(cache
, nid
) &&
3061 get_node(cache
, nid
)->free_objects
) {
3062 obj
= ____cache_alloc_node(cache
,
3063 gfp_exact_node(flags
), nid
);
3071 * This allocation will be performed within the constraints
3072 * of the current cpuset / memory policy requirements.
3073 * We may trigger various forms of reclaim on the allowed
3074 * set and go into memory reserves if necessary.
3078 if (gfpflags_allow_blocking(local_flags
))
3080 kmem_flagcheck(cache
, flags
);
3081 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3082 if (gfpflags_allow_blocking(local_flags
))
3083 local_irq_disable();
3086 * Insert into the appropriate per node queues
3088 nid
= page_to_nid(page
);
3089 if (cache_grow(cache
, flags
, nid
, page
)) {
3090 obj
= ____cache_alloc_node(cache
,
3091 gfp_exact_node(flags
), nid
);
3094 * Another processor may allocate the
3095 * objects in the slab since we are
3096 * not holding any locks.
3100 /* cache_grow already freed obj */
3106 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3112 * A interface to enable slab creation on nodeid
3114 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3118 struct kmem_cache_node
*n
;
3123 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3124 n
= get_node(cachep
, nodeid
);
3129 spin_lock(&n
->list_lock
);
3130 page
= get_first_slab(n
, false);
3134 check_spinlock_acquired_node(cachep
, nodeid
);
3136 STATS_INC_NODEALLOCS(cachep
);
3137 STATS_INC_ACTIVE(cachep
);
3138 STATS_SET_HIGH(cachep
);
3140 BUG_ON(page
->active
== cachep
->num
);
3142 obj
= slab_get_obj(cachep
, page
);
3145 fixup_slab_list(cachep
, n
, page
, &list
);
3147 spin_unlock(&n
->list_lock
);
3148 fixup_objfreelist_debug(cachep
, &list
);
3152 spin_unlock(&n
->list_lock
);
3153 x
= cache_grow(cachep
, gfp_exact_node(flags
), nodeid
, NULL
);
3157 return fallback_alloc(cachep
, flags
);
3163 static __always_inline
void *
3164 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3165 unsigned long caller
)
3167 unsigned long save_flags
;
3169 int slab_node
= numa_mem_id();
3171 flags
&= gfp_allowed_mask
;
3172 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3173 if (unlikely(!cachep
))
3176 cache_alloc_debugcheck_before(cachep
, flags
);
3177 local_irq_save(save_flags
);
3179 if (nodeid
== NUMA_NO_NODE
)
3182 if (unlikely(!get_node(cachep
, nodeid
))) {
3183 /* Node not bootstrapped yet */
3184 ptr
= fallback_alloc(cachep
, flags
);
3188 if (nodeid
== slab_node
) {
3190 * Use the locally cached objects if possible.
3191 * However ____cache_alloc does not allow fallback
3192 * to other nodes. It may fail while we still have
3193 * objects on other nodes available.
3195 ptr
= ____cache_alloc(cachep
, flags
);
3199 /* ___cache_alloc_node can fall back to other nodes */
3200 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3202 local_irq_restore(save_flags
);
3203 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3205 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3206 memset(ptr
, 0, cachep
->object_size
);
3208 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3212 static __always_inline
void *
3213 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3217 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3218 objp
= alternate_node_alloc(cache
, flags
);
3222 objp
= ____cache_alloc(cache
, flags
);
3225 * We may just have run out of memory on the local node.
3226 * ____cache_alloc_node() knows how to locate memory on other nodes
3229 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3236 static __always_inline
void *
3237 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3239 return ____cache_alloc(cachep
, flags
);
3242 #endif /* CONFIG_NUMA */
3244 static __always_inline
void *
3245 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3247 unsigned long save_flags
;
3250 flags
&= gfp_allowed_mask
;
3251 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3252 if (unlikely(!cachep
))
3255 cache_alloc_debugcheck_before(cachep
, flags
);
3256 local_irq_save(save_flags
);
3257 objp
= __do_cache_alloc(cachep
, flags
);
3258 local_irq_restore(save_flags
);
3259 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3262 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3263 memset(objp
, 0, cachep
->object_size
);
3265 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3270 * Caller needs to acquire correct kmem_cache_node's list_lock
3271 * @list: List of detached free slabs should be freed by caller
3273 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3274 int nr_objects
, int node
, struct list_head
*list
)
3277 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3279 for (i
= 0; i
< nr_objects
; i
++) {
3285 page
= virt_to_head_page(objp
);
3286 list_del(&page
->lru
);
3287 check_spinlock_acquired_node(cachep
, node
);
3288 slab_put_obj(cachep
, page
, objp
);
3289 STATS_DEC_ACTIVE(cachep
);
3292 /* fixup slab chains */
3293 if (page
->active
== 0) {
3294 if (n
->free_objects
> n
->free_limit
) {
3295 n
->free_objects
-= cachep
->num
;
3296 list_add_tail(&page
->lru
, list
);
3298 list_add(&page
->lru
, &n
->slabs_free
);
3301 /* Unconditionally move a slab to the end of the
3302 * partial list on free - maximum time for the
3303 * other objects to be freed, too.
3305 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3310 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3313 struct kmem_cache_node
*n
;
3314 int node
= numa_mem_id();
3317 batchcount
= ac
->batchcount
;
3320 n
= get_node(cachep
, node
);
3321 spin_lock(&n
->list_lock
);
3323 struct array_cache
*shared_array
= n
->shared
;
3324 int max
= shared_array
->limit
- shared_array
->avail
;
3326 if (batchcount
> max
)
3328 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3329 ac
->entry
, sizeof(void *) * batchcount
);
3330 shared_array
->avail
+= batchcount
;
3335 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3342 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3343 BUG_ON(page
->active
);
3347 STATS_SET_FREEABLE(cachep
, i
);
3350 spin_unlock(&n
->list_lock
);
3351 slabs_destroy(cachep
, &list
);
3352 ac
->avail
-= batchcount
;
3353 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3357 * Release an obj back to its cache. If the obj has a constructed state, it must
3358 * be in this state _before_ it is released. Called with disabled ints.
3360 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3361 unsigned long caller
)
3363 struct array_cache
*ac
= cpu_cache_get(cachep
);
3365 kasan_slab_free(cachep
, objp
);
3368 kmemleak_free_recursive(objp
, cachep
->flags
);
3369 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3371 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3374 * Skip calling cache_free_alien() when the platform is not numa.
3375 * This will avoid cache misses that happen while accessing slabp (which
3376 * is per page memory reference) to get nodeid. Instead use a global
3377 * variable to skip the call, which is mostly likely to be present in
3380 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3383 if (ac
->avail
< ac
->limit
) {
3384 STATS_INC_FREEHIT(cachep
);
3386 STATS_INC_FREEMISS(cachep
);
3387 cache_flusharray(cachep
, ac
);
3390 if (sk_memalloc_socks()) {
3391 struct page
*page
= virt_to_head_page(objp
);
3393 if (unlikely(PageSlabPfmemalloc(page
))) {
3394 cache_free_pfmemalloc(cachep
, page
, objp
);
3399 ac
->entry
[ac
->avail
++] = objp
;
3403 * kmem_cache_alloc - Allocate an object
3404 * @cachep: The cache to allocate from.
3405 * @flags: See kmalloc().
3407 * Allocate an object from this cache. The flags are only relevant
3408 * if the cache has no available objects.
3410 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3412 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3414 kasan_slab_alloc(cachep
, ret
, flags
);
3415 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3416 cachep
->object_size
, cachep
->size
, flags
);
3420 EXPORT_SYMBOL(kmem_cache_alloc
);
3422 static __always_inline
void
3423 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3424 size_t size
, void **p
, unsigned long caller
)
3428 for (i
= 0; i
< size
; i
++)
3429 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3432 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3437 s
= slab_pre_alloc_hook(s
, flags
);
3441 cache_alloc_debugcheck_before(s
, flags
);
3443 local_irq_disable();
3444 for (i
= 0; i
< size
; i
++) {
3445 void *objp
= __do_cache_alloc(s
, flags
);
3447 if (unlikely(!objp
))
3453 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3455 /* Clear memory outside IRQ disabled section */
3456 if (unlikely(flags
& __GFP_ZERO
))
3457 for (i
= 0; i
< size
; i
++)
3458 memset(p
[i
], 0, s
->object_size
);
3460 slab_post_alloc_hook(s
, flags
, size
, p
);
3461 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3465 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3466 slab_post_alloc_hook(s
, flags
, i
, p
);
3467 __kmem_cache_free_bulk(s
, i
, p
);
3470 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3472 #ifdef CONFIG_TRACING
3474 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3478 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3480 kasan_kmalloc(cachep
, ret
, size
, flags
);
3481 trace_kmalloc(_RET_IP_
, ret
,
3482 size
, cachep
->size
, flags
);
3485 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3490 * kmem_cache_alloc_node - Allocate an object on the specified node
3491 * @cachep: The cache to allocate from.
3492 * @flags: See kmalloc().
3493 * @nodeid: node number of the target node.
3495 * Identical to kmem_cache_alloc but it will allocate memory on the given
3496 * node, which can improve the performance for cpu bound structures.
3498 * Fallback to other node is possible if __GFP_THISNODE is not set.
3500 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3502 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3504 kasan_slab_alloc(cachep
, ret
, flags
);
3505 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3506 cachep
->object_size
, cachep
->size
,
3511 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3513 #ifdef CONFIG_TRACING
3514 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3521 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3523 kasan_kmalloc(cachep
, ret
, size
, flags
);
3524 trace_kmalloc_node(_RET_IP_
, ret
,
3529 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3532 static __always_inline
void *
3533 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3535 struct kmem_cache
*cachep
;
3538 cachep
= kmalloc_slab(size
, flags
);
3539 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3541 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3542 kasan_kmalloc(cachep
, ret
, size
, flags
);
3547 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3549 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3551 EXPORT_SYMBOL(__kmalloc_node
);
3553 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3554 int node
, unsigned long caller
)
3556 return __do_kmalloc_node(size
, flags
, node
, caller
);
3558 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3559 #endif /* CONFIG_NUMA */
3562 * __do_kmalloc - allocate memory
3563 * @size: how many bytes of memory are required.
3564 * @flags: the type of memory to allocate (see kmalloc).
3565 * @caller: function caller for debug tracking of the caller
3567 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3568 unsigned long caller
)
3570 struct kmem_cache
*cachep
;
3573 cachep
= kmalloc_slab(size
, flags
);
3574 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3576 ret
= slab_alloc(cachep
, flags
, caller
);
3578 kasan_kmalloc(cachep
, ret
, size
, flags
);
3579 trace_kmalloc(caller
, ret
,
3580 size
, cachep
->size
, flags
);
3585 void *__kmalloc(size_t size
, gfp_t flags
)
3587 return __do_kmalloc(size
, flags
, _RET_IP_
);
3589 EXPORT_SYMBOL(__kmalloc
);
3591 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3593 return __do_kmalloc(size
, flags
, caller
);
3595 EXPORT_SYMBOL(__kmalloc_track_caller
);
3598 * kmem_cache_free - Deallocate an object
3599 * @cachep: The cache the allocation was from.
3600 * @objp: The previously allocated object.
3602 * Free an object which was previously allocated from this
3605 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3607 unsigned long flags
;
3608 cachep
= cache_from_obj(cachep
, objp
);
3612 local_irq_save(flags
);
3613 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3614 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3615 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3616 __cache_free(cachep
, objp
, _RET_IP_
);
3617 local_irq_restore(flags
);
3619 trace_kmem_cache_free(_RET_IP_
, objp
);
3621 EXPORT_SYMBOL(kmem_cache_free
);
3623 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3625 struct kmem_cache
*s
;
3628 local_irq_disable();
3629 for (i
= 0; i
< size
; i
++) {
3632 if (!orig_s
) /* called via kfree_bulk */
3633 s
= virt_to_cache(objp
);
3635 s
= cache_from_obj(orig_s
, objp
);
3637 debug_check_no_locks_freed(objp
, s
->object_size
);
3638 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3639 debug_check_no_obj_freed(objp
, s
->object_size
);
3641 __cache_free(s
, objp
, _RET_IP_
);
3645 /* FIXME: add tracing */
3647 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3650 * kfree - free previously allocated memory
3651 * @objp: pointer returned by kmalloc.
3653 * If @objp is NULL, no operation is performed.
3655 * Don't free memory not originally allocated by kmalloc()
3656 * or you will run into trouble.
3658 void kfree(const void *objp
)
3660 struct kmem_cache
*c
;
3661 unsigned long flags
;
3663 trace_kfree(_RET_IP_
, objp
);
3665 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3667 local_irq_save(flags
);
3668 kfree_debugcheck(objp
);
3669 c
= virt_to_cache(objp
);
3670 debug_check_no_locks_freed(objp
, c
->object_size
);
3672 debug_check_no_obj_freed(objp
, c
->object_size
);
3673 __cache_free(c
, (void *)objp
, _RET_IP_
);
3674 local_irq_restore(flags
);
3676 EXPORT_SYMBOL(kfree
);
3679 * This initializes kmem_cache_node or resizes various caches for all nodes.
3681 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3684 struct kmem_cache_node
*n
;
3685 struct array_cache
*new_shared
;
3686 struct alien_cache
**new_alien
= NULL
;
3688 for_each_online_node(node
) {
3690 if (use_alien_caches
) {
3691 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3697 if (cachep
->shared
) {
3698 new_shared
= alloc_arraycache(node
,
3699 cachep
->shared
*cachep
->batchcount
,
3702 free_alien_cache(new_alien
);
3707 n
= get_node(cachep
, node
);
3709 struct array_cache
*shared
= n
->shared
;
3712 spin_lock_irq(&n
->list_lock
);
3715 free_block(cachep
, shared
->entry
,
3716 shared
->avail
, node
, &list
);
3718 n
->shared
= new_shared
;
3720 n
->alien
= new_alien
;
3723 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3724 cachep
->batchcount
+ cachep
->num
;
3725 spin_unlock_irq(&n
->list_lock
);
3726 slabs_destroy(cachep
, &list
);
3728 free_alien_cache(new_alien
);
3731 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3733 free_alien_cache(new_alien
);
3738 kmem_cache_node_init(n
);
3739 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3740 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3741 n
->shared
= new_shared
;
3742 n
->alien
= new_alien
;
3743 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3744 cachep
->batchcount
+ cachep
->num
;
3745 cachep
->node
[node
] = n
;
3750 if (!cachep
->list
.next
) {
3751 /* Cache is not active yet. Roll back what we did */
3754 n
= get_node(cachep
, node
);
3757 free_alien_cache(n
->alien
);
3759 cachep
->node
[node
] = NULL
;
3767 /* Always called with the slab_mutex held */
3768 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3769 int batchcount
, int shared
, gfp_t gfp
)
3771 struct array_cache __percpu
*cpu_cache
, *prev
;
3774 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3778 prev
= cachep
->cpu_cache
;
3779 cachep
->cpu_cache
= cpu_cache
;
3780 kick_all_cpus_sync();
3783 cachep
->batchcount
= batchcount
;
3784 cachep
->limit
= limit
;
3785 cachep
->shared
= shared
;
3790 for_each_online_cpu(cpu
) {
3793 struct kmem_cache_node
*n
;
3794 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3796 node
= cpu_to_mem(cpu
);
3797 n
= get_node(cachep
, node
);
3798 spin_lock_irq(&n
->list_lock
);
3799 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3800 spin_unlock_irq(&n
->list_lock
);
3801 slabs_destroy(cachep
, &list
);
3806 return alloc_kmem_cache_node(cachep
, gfp
);
3809 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3810 int batchcount
, int shared
, gfp_t gfp
)
3813 struct kmem_cache
*c
;
3815 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3817 if (slab_state
< FULL
)
3820 if ((ret
< 0) || !is_root_cache(cachep
))
3823 lockdep_assert_held(&slab_mutex
);
3824 for_each_memcg_cache(c
, cachep
) {
3825 /* return value determined by the root cache only */
3826 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3832 /* Called with slab_mutex held always */
3833 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3840 if (!is_root_cache(cachep
)) {
3841 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3842 limit
= root
->limit
;
3843 shared
= root
->shared
;
3844 batchcount
= root
->batchcount
;
3847 if (limit
&& shared
&& batchcount
)
3850 * The head array serves three purposes:
3851 * - create a LIFO ordering, i.e. return objects that are cache-warm
3852 * - reduce the number of spinlock operations.
3853 * - reduce the number of linked list operations on the slab and
3854 * bufctl chains: array operations are cheaper.
3855 * The numbers are guessed, we should auto-tune as described by
3858 if (cachep
->size
> 131072)
3860 else if (cachep
->size
> PAGE_SIZE
)
3862 else if (cachep
->size
> 1024)
3864 else if (cachep
->size
> 256)
3870 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3871 * allocation behaviour: Most allocs on one cpu, most free operations
3872 * on another cpu. For these cases, an efficient object passing between
3873 * cpus is necessary. This is provided by a shared array. The array
3874 * replaces Bonwick's magazine layer.
3875 * On uniprocessor, it's functionally equivalent (but less efficient)
3876 * to a larger limit. Thus disabled by default.
3879 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3884 * With debugging enabled, large batchcount lead to excessively long
3885 * periods with disabled local interrupts. Limit the batchcount
3890 batchcount
= (limit
+ 1) / 2;
3892 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3894 pr_err("enable_cpucache failed for %s, error %d\n",
3895 cachep
->name
, -err
);
3900 * Drain an array if it contains any elements taking the node lock only if
3901 * necessary. Note that the node listlock also protects the array_cache
3902 * if drain_array() is used on the shared array.
3904 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3905 struct array_cache
*ac
, int node
)
3909 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3910 check_mutex_acquired();
3912 if (!ac
|| !ac
->avail
)
3920 spin_lock_irq(&n
->list_lock
);
3921 drain_array_locked(cachep
, ac
, node
, false, &list
);
3922 spin_unlock_irq(&n
->list_lock
);
3924 slabs_destroy(cachep
, &list
);
3928 * cache_reap - Reclaim memory from caches.
3929 * @w: work descriptor
3931 * Called from workqueue/eventd every few seconds.
3933 * - clear the per-cpu caches for this CPU.
3934 * - return freeable pages to the main free memory pool.
3936 * If we cannot acquire the cache chain mutex then just give up - we'll try
3937 * again on the next iteration.
3939 static void cache_reap(struct work_struct
*w
)
3941 struct kmem_cache
*searchp
;
3942 struct kmem_cache_node
*n
;
3943 int node
= numa_mem_id();
3944 struct delayed_work
*work
= to_delayed_work(w
);
3946 if (!mutex_trylock(&slab_mutex
))
3947 /* Give up. Setup the next iteration. */
3950 list_for_each_entry(searchp
, &slab_caches
, list
) {
3954 * We only take the node lock if absolutely necessary and we
3955 * have established with reasonable certainty that
3956 * we can do some work if the lock was obtained.
3958 n
= get_node(searchp
, node
);
3960 reap_alien(searchp
, n
);
3962 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3965 * These are racy checks but it does not matter
3966 * if we skip one check or scan twice.
3968 if (time_after(n
->next_reap
, jiffies
))
3971 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3973 drain_array(searchp
, n
, n
->shared
, node
);
3975 if (n
->free_touched
)
3976 n
->free_touched
= 0;
3980 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3981 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3982 STATS_ADD_REAPED(searchp
, freed
);
3988 mutex_unlock(&slab_mutex
);
3991 /* Set up the next iteration */
3992 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3995 #ifdef CONFIG_SLABINFO
3996 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3999 unsigned long active_objs
;
4000 unsigned long num_objs
;
4001 unsigned long active_slabs
= 0;
4002 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4006 struct kmem_cache_node
*n
;
4010 for_each_kmem_cache_node(cachep
, node
, n
) {
4013 spin_lock_irq(&n
->list_lock
);
4015 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4016 if (page
->active
!= cachep
->num
&& !error
)
4017 error
= "slabs_full accounting error";
4018 active_objs
+= cachep
->num
;
4021 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4022 if (page
->active
== cachep
->num
&& !error
)
4023 error
= "slabs_partial accounting error";
4024 if (!page
->active
&& !error
)
4025 error
= "slabs_partial accounting error";
4026 active_objs
+= page
->active
;
4029 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4030 if (page
->active
&& !error
)
4031 error
= "slabs_free accounting error";
4034 free_objects
+= n
->free_objects
;
4036 shared_avail
+= n
->shared
->avail
;
4038 spin_unlock_irq(&n
->list_lock
);
4040 num_slabs
+= active_slabs
;
4041 num_objs
= num_slabs
* cachep
->num
;
4042 if (num_objs
- active_objs
!= free_objects
&& !error
)
4043 error
= "free_objects accounting error";
4045 name
= cachep
->name
;
4047 pr_err("slab: cache %s error: %s\n", name
, error
);
4049 sinfo
->active_objs
= active_objs
;
4050 sinfo
->num_objs
= num_objs
;
4051 sinfo
->active_slabs
= active_slabs
;
4052 sinfo
->num_slabs
= num_slabs
;
4053 sinfo
->shared_avail
= shared_avail
;
4054 sinfo
->limit
= cachep
->limit
;
4055 sinfo
->batchcount
= cachep
->batchcount
;
4056 sinfo
->shared
= cachep
->shared
;
4057 sinfo
->objects_per_slab
= cachep
->num
;
4058 sinfo
->cache_order
= cachep
->gfporder
;
4061 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4065 unsigned long high
= cachep
->high_mark
;
4066 unsigned long allocs
= cachep
->num_allocations
;
4067 unsigned long grown
= cachep
->grown
;
4068 unsigned long reaped
= cachep
->reaped
;
4069 unsigned long errors
= cachep
->errors
;
4070 unsigned long max_freeable
= cachep
->max_freeable
;
4071 unsigned long node_allocs
= cachep
->node_allocs
;
4072 unsigned long node_frees
= cachep
->node_frees
;
4073 unsigned long overflows
= cachep
->node_overflow
;
4075 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4076 allocs
, high
, grown
,
4077 reaped
, errors
, max_freeable
, node_allocs
,
4078 node_frees
, overflows
);
4082 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4083 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4084 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4085 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4087 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4088 allochit
, allocmiss
, freehit
, freemiss
);
4093 #define MAX_SLABINFO_WRITE 128
4095 * slabinfo_write - Tuning for the slab allocator
4097 * @buffer: user buffer
4098 * @count: data length
4101 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4102 size_t count
, loff_t
*ppos
)
4104 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4105 int limit
, batchcount
, shared
, res
;
4106 struct kmem_cache
*cachep
;
4108 if (count
> MAX_SLABINFO_WRITE
)
4110 if (copy_from_user(&kbuf
, buffer
, count
))
4112 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4114 tmp
= strchr(kbuf
, ' ');
4119 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4122 /* Find the cache in the chain of caches. */
4123 mutex_lock(&slab_mutex
);
4125 list_for_each_entry(cachep
, &slab_caches
, list
) {
4126 if (!strcmp(cachep
->name
, kbuf
)) {
4127 if (limit
< 1 || batchcount
< 1 ||
4128 batchcount
> limit
|| shared
< 0) {
4131 res
= do_tune_cpucache(cachep
, limit
,
4138 mutex_unlock(&slab_mutex
);
4144 #ifdef CONFIG_DEBUG_SLAB_LEAK
4146 static inline int add_caller(unsigned long *n
, unsigned long v
)
4156 unsigned long *q
= p
+ 2 * i
;
4170 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4176 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4185 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4188 for (j
= page
->active
; j
< c
->num
; j
++) {
4189 if (get_free_obj(page
, j
) == i
) {
4199 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4200 * mapping is established when actual object allocation and
4201 * we could mistakenly access the unmapped object in the cpu
4204 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4207 if (!add_caller(n
, v
))
4212 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4214 #ifdef CONFIG_KALLSYMS
4215 unsigned long offset
, size
;
4216 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4218 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4219 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4221 seq_printf(m
, " [%s]", modname
);
4225 seq_printf(m
, "%p", (void *)address
);
4228 static int leaks_show(struct seq_file
*m
, void *p
)
4230 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4232 struct kmem_cache_node
*n
;
4234 unsigned long *x
= m
->private;
4238 if (!(cachep
->flags
& SLAB_STORE_USER
))
4240 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4244 * Set store_user_clean and start to grab stored user information
4245 * for all objects on this cache. If some alloc/free requests comes
4246 * during the processing, information would be wrong so restart
4250 set_store_user_clean(cachep
);
4251 drain_cpu_caches(cachep
);
4255 for_each_kmem_cache_node(cachep
, node
, n
) {
4258 spin_lock_irq(&n
->list_lock
);
4260 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4261 handle_slab(x
, cachep
, page
);
4262 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4263 handle_slab(x
, cachep
, page
);
4264 spin_unlock_irq(&n
->list_lock
);
4266 } while (!is_store_user_clean(cachep
));
4268 name
= cachep
->name
;
4270 /* Increase the buffer size */
4271 mutex_unlock(&slab_mutex
);
4272 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4274 /* Too bad, we are really out */
4276 mutex_lock(&slab_mutex
);
4279 *(unsigned long *)m
->private = x
[0] * 2;
4281 mutex_lock(&slab_mutex
);
4282 /* Now make sure this entry will be retried */
4286 for (i
= 0; i
< x
[1]; i
++) {
4287 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4288 show_symbol(m
, x
[2*i
+2]);
4295 static const struct seq_operations slabstats_op
= {
4296 .start
= slab_start
,
4302 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4306 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4310 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4315 static const struct file_operations proc_slabstats_operations
= {
4316 .open
= slabstats_open
,
4318 .llseek
= seq_lseek
,
4319 .release
= seq_release_private
,
4323 static int __init
slab_proc_init(void)
4325 #ifdef CONFIG_DEBUG_SLAB_LEAK
4326 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4330 module_init(slab_proc_init
);
4334 * ksize - get the actual amount of memory allocated for a given object
4335 * @objp: Pointer to the object
4337 * kmalloc may internally round up allocations and return more memory
4338 * than requested. ksize() can be used to determine the actual amount of
4339 * memory allocated. The caller may use this additional memory, even though
4340 * a smaller amount of memory was initially specified with the kmalloc call.
4341 * The caller must guarantee that objp points to a valid object previously
4342 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4343 * must not be freed during the duration of the call.
4345 size_t ksize(const void *objp
)
4350 if (unlikely(objp
== ZERO_SIZE_PTR
))
4353 size
= virt_to_cache(objp
)->object_size
;
4354 /* We assume that ksize callers could use the whole allocated area,
4355 * so we need to unpoison this area.
4357 kasan_krealloc(objp
, size
, GFP_NOWAIT
);
4361 EXPORT_SYMBOL(ksize
);