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 int setup_kmem_cache_node(struct kmem_cache
*cachep
,
915 int node
, gfp_t gfp
, bool force_change
)
918 struct kmem_cache_node
*n
;
919 struct array_cache
*old_shared
= NULL
;
920 struct array_cache
*new_shared
= NULL
;
921 struct alien_cache
**new_alien
= NULL
;
924 if (use_alien_caches
) {
925 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
930 if (cachep
->shared
) {
931 new_shared
= alloc_arraycache(node
,
932 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
937 ret
= init_cache_node(cachep
, node
, gfp
);
941 n
= get_node(cachep
, node
);
942 spin_lock_irq(&n
->list_lock
);
943 if (n
->shared
&& force_change
) {
944 free_block(cachep
, n
->shared
->entry
,
945 n
->shared
->avail
, node
, &list
);
946 n
->shared
->avail
= 0;
949 if (!n
->shared
|| force_change
) {
950 old_shared
= n
->shared
;
951 n
->shared
= new_shared
;
956 n
->alien
= new_alien
;
960 spin_unlock_irq(&n
->list_lock
);
961 slabs_destroy(cachep
, &list
);
966 free_alien_cache(new_alien
);
971 static void cpuup_canceled(long cpu
)
973 struct kmem_cache
*cachep
;
974 struct kmem_cache_node
*n
= NULL
;
975 int node
= cpu_to_mem(cpu
);
976 const struct cpumask
*mask
= cpumask_of_node(node
);
978 list_for_each_entry(cachep
, &slab_caches
, list
) {
979 struct array_cache
*nc
;
980 struct array_cache
*shared
;
981 struct alien_cache
**alien
;
984 n
= get_node(cachep
, node
);
988 spin_lock_irq(&n
->list_lock
);
990 /* Free limit for this kmem_cache_node */
991 n
->free_limit
-= cachep
->batchcount
;
993 /* cpu is dead; no one can alloc from it. */
994 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
996 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1000 if (!cpumask_empty(mask
)) {
1001 spin_unlock_irq(&n
->list_lock
);
1007 free_block(cachep
, shared
->entry
,
1008 shared
->avail
, node
, &list
);
1015 spin_unlock_irq(&n
->list_lock
);
1019 drain_alien_cache(cachep
, alien
);
1020 free_alien_cache(alien
);
1024 slabs_destroy(cachep
, &list
);
1027 * In the previous loop, all the objects were freed to
1028 * the respective cache's slabs, now we can go ahead and
1029 * shrink each nodelist to its limit.
1031 list_for_each_entry(cachep
, &slab_caches
, list
) {
1032 n
= get_node(cachep
, node
);
1035 drain_freelist(cachep
, n
, INT_MAX
);
1039 static int cpuup_prepare(long cpu
)
1041 struct kmem_cache
*cachep
;
1042 int node
= cpu_to_mem(cpu
);
1046 * We need to do this right in the beginning since
1047 * alloc_arraycache's are going to use this list.
1048 * kmalloc_node allows us to add the slab to the right
1049 * kmem_cache_node and not this cpu's kmem_cache_node
1051 err
= init_cache_node_node(node
);
1056 * Now we can go ahead with allocating the shared arrays and
1059 list_for_each_entry(cachep
, &slab_caches
, list
) {
1060 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1067 cpuup_canceled(cpu
);
1071 static int cpuup_callback(struct notifier_block
*nfb
,
1072 unsigned long action
, void *hcpu
)
1074 long cpu
= (long)hcpu
;
1078 case CPU_UP_PREPARE
:
1079 case CPU_UP_PREPARE_FROZEN
:
1080 mutex_lock(&slab_mutex
);
1081 err
= cpuup_prepare(cpu
);
1082 mutex_unlock(&slab_mutex
);
1085 case CPU_ONLINE_FROZEN
:
1086 start_cpu_timer(cpu
);
1088 #ifdef CONFIG_HOTPLUG_CPU
1089 case CPU_DOWN_PREPARE
:
1090 case CPU_DOWN_PREPARE_FROZEN
:
1092 * Shutdown cache reaper. Note that the slab_mutex is
1093 * held so that if cache_reap() is invoked it cannot do
1094 * anything expensive but will only modify reap_work
1095 * and reschedule the timer.
1097 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1098 /* Now the cache_reaper is guaranteed to be not running. */
1099 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1101 case CPU_DOWN_FAILED
:
1102 case CPU_DOWN_FAILED_FROZEN
:
1103 start_cpu_timer(cpu
);
1106 case CPU_DEAD_FROZEN
:
1108 * Even if all the cpus of a node are down, we don't free the
1109 * kmem_cache_node of any cache. This to avoid a race between
1110 * cpu_down, and a kmalloc allocation from another cpu for
1111 * memory from the node of the cpu going down. The node
1112 * structure is usually allocated from kmem_cache_create() and
1113 * gets destroyed at kmem_cache_destroy().
1117 case CPU_UP_CANCELED
:
1118 case CPU_UP_CANCELED_FROZEN
:
1119 mutex_lock(&slab_mutex
);
1120 cpuup_canceled(cpu
);
1121 mutex_unlock(&slab_mutex
);
1124 return notifier_from_errno(err
);
1127 static struct notifier_block cpucache_notifier
= {
1128 &cpuup_callback
, NULL
, 0
1131 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1133 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1134 * Returns -EBUSY if all objects cannot be drained so that the node is not
1137 * Must hold slab_mutex.
1139 static int __meminit
drain_cache_node_node(int node
)
1141 struct kmem_cache
*cachep
;
1144 list_for_each_entry(cachep
, &slab_caches
, list
) {
1145 struct kmem_cache_node
*n
;
1147 n
= get_node(cachep
, node
);
1151 drain_freelist(cachep
, n
, INT_MAX
);
1153 if (!list_empty(&n
->slabs_full
) ||
1154 !list_empty(&n
->slabs_partial
)) {
1162 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1163 unsigned long action
, void *arg
)
1165 struct memory_notify
*mnb
= arg
;
1169 nid
= mnb
->status_change_nid
;
1174 case MEM_GOING_ONLINE
:
1175 mutex_lock(&slab_mutex
);
1176 ret
= init_cache_node_node(nid
);
1177 mutex_unlock(&slab_mutex
);
1179 case MEM_GOING_OFFLINE
:
1180 mutex_lock(&slab_mutex
);
1181 ret
= drain_cache_node_node(nid
);
1182 mutex_unlock(&slab_mutex
);
1186 case MEM_CANCEL_ONLINE
:
1187 case MEM_CANCEL_OFFLINE
:
1191 return notifier_from_errno(ret
);
1193 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1196 * swap the static kmem_cache_node with kmalloced memory
1198 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1201 struct kmem_cache_node
*ptr
;
1203 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1206 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1208 * Do not assume that spinlocks can be initialized via memcpy:
1210 spin_lock_init(&ptr
->list_lock
);
1212 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1213 cachep
->node
[nodeid
] = ptr
;
1217 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1218 * size of kmem_cache_node.
1220 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1224 for_each_online_node(node
) {
1225 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1226 cachep
->node
[node
]->next_reap
= jiffies
+
1228 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1233 * Initialisation. Called after the page allocator have been initialised and
1234 * before smp_init().
1236 void __init
kmem_cache_init(void)
1240 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1241 sizeof(struct rcu_head
));
1242 kmem_cache
= &kmem_cache_boot
;
1244 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1245 use_alien_caches
= 0;
1247 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1248 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1251 * Fragmentation resistance on low memory - only use bigger
1252 * page orders on machines with more than 32MB of memory if
1253 * not overridden on the command line.
1255 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1256 slab_max_order
= SLAB_MAX_ORDER_HI
;
1258 /* Bootstrap is tricky, because several objects are allocated
1259 * from caches that do not exist yet:
1260 * 1) initialize the kmem_cache cache: it contains the struct
1261 * kmem_cache structures of all caches, except kmem_cache itself:
1262 * kmem_cache is statically allocated.
1263 * Initially an __init data area is used for the head array and the
1264 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1265 * array at the end of the bootstrap.
1266 * 2) Create the first kmalloc cache.
1267 * The struct kmem_cache for the new cache is allocated normally.
1268 * An __init data area is used for the head array.
1269 * 3) Create the remaining kmalloc caches, with minimally sized
1271 * 4) Replace the __init data head arrays for kmem_cache and the first
1272 * kmalloc cache with kmalloc allocated arrays.
1273 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1274 * the other cache's with kmalloc allocated memory.
1275 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1278 /* 1) create the kmem_cache */
1281 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1283 create_boot_cache(kmem_cache
, "kmem_cache",
1284 offsetof(struct kmem_cache
, node
) +
1285 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1286 SLAB_HWCACHE_ALIGN
);
1287 list_add(&kmem_cache
->list
, &slab_caches
);
1288 slab_state
= PARTIAL
;
1291 * Initialize the caches that provide memory for the kmem_cache_node
1292 * structures first. Without this, further allocations will bug.
1294 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1295 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1296 slab_state
= PARTIAL_NODE
;
1297 setup_kmalloc_cache_index_table();
1299 slab_early_init
= 0;
1301 /* 5) Replace the bootstrap kmem_cache_node */
1305 for_each_online_node(nid
) {
1306 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1308 init_list(kmalloc_caches
[INDEX_NODE
],
1309 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1313 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1316 void __init
kmem_cache_init_late(void)
1318 struct kmem_cache
*cachep
;
1322 /* 6) resize the head arrays to their final sizes */
1323 mutex_lock(&slab_mutex
);
1324 list_for_each_entry(cachep
, &slab_caches
, list
)
1325 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1327 mutex_unlock(&slab_mutex
);
1333 * Register a cpu startup notifier callback that initializes
1334 * cpu_cache_get for all new cpus
1336 register_cpu_notifier(&cpucache_notifier
);
1340 * Register a memory hotplug callback that initializes and frees
1343 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1347 * The reap timers are started later, with a module init call: That part
1348 * of the kernel is not yet operational.
1352 static int __init
cpucache_init(void)
1357 * Register the timers that return unneeded pages to the page allocator
1359 for_each_online_cpu(cpu
)
1360 start_cpu_timer(cpu
);
1366 __initcall(cpucache_init
);
1368 static noinline
void
1369 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1372 struct kmem_cache_node
*n
;
1374 unsigned long flags
;
1376 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1377 DEFAULT_RATELIMIT_BURST
);
1379 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1382 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1383 nodeid
, gfpflags
, &gfpflags
);
1384 pr_warn(" cache: %s, object size: %d, order: %d\n",
1385 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1387 for_each_kmem_cache_node(cachep
, node
, n
) {
1388 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1389 unsigned long active_slabs
= 0, num_slabs
= 0;
1391 spin_lock_irqsave(&n
->list_lock
, flags
);
1392 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1393 active_objs
+= cachep
->num
;
1396 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1397 active_objs
+= page
->active
;
1400 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1403 free_objects
+= n
->free_objects
;
1404 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1406 num_slabs
+= active_slabs
;
1407 num_objs
= num_slabs
* cachep
->num
;
1408 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1409 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1416 * Interface to system's page allocator. No need to hold the
1417 * kmem_cache_node ->list_lock.
1419 * If we requested dmaable memory, we will get it. Even if we
1420 * did not request dmaable memory, we might get it, but that
1421 * would be relatively rare and ignorable.
1423 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1429 flags
|= cachep
->allocflags
;
1430 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1431 flags
|= __GFP_RECLAIMABLE
;
1433 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1435 slab_out_of_memory(cachep
, flags
, nodeid
);
1439 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1440 __free_pages(page
, cachep
->gfporder
);
1444 nr_pages
= (1 << cachep
->gfporder
);
1445 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1446 add_zone_page_state(page_zone(page
),
1447 NR_SLAB_RECLAIMABLE
, nr_pages
);
1449 add_zone_page_state(page_zone(page
),
1450 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1452 __SetPageSlab(page
);
1453 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1454 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1455 SetPageSlabPfmemalloc(page
);
1457 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1458 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1461 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1463 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1470 * Interface to system's page release.
1472 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1474 int order
= cachep
->gfporder
;
1475 unsigned long nr_freed
= (1 << order
);
1477 kmemcheck_free_shadow(page
, order
);
1479 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1480 sub_zone_page_state(page_zone(page
),
1481 NR_SLAB_RECLAIMABLE
, nr_freed
);
1483 sub_zone_page_state(page_zone(page
),
1484 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1486 BUG_ON(!PageSlab(page
));
1487 __ClearPageSlabPfmemalloc(page
);
1488 __ClearPageSlab(page
);
1489 page_mapcount_reset(page
);
1490 page
->mapping
= NULL
;
1492 if (current
->reclaim_state
)
1493 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1494 memcg_uncharge_slab(page
, order
, cachep
);
1495 __free_pages(page
, order
);
1498 static void kmem_rcu_free(struct rcu_head
*head
)
1500 struct kmem_cache
*cachep
;
1503 page
= container_of(head
, struct page
, rcu_head
);
1504 cachep
= page
->slab_cache
;
1506 kmem_freepages(cachep
, page
);
1510 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1512 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1513 (cachep
->size
% PAGE_SIZE
) == 0)
1519 #ifdef CONFIG_DEBUG_PAGEALLOC
1520 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1521 unsigned long caller
)
1523 int size
= cachep
->object_size
;
1525 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1527 if (size
< 5 * sizeof(unsigned long))
1530 *addr
++ = 0x12345678;
1532 *addr
++ = smp_processor_id();
1533 size
-= 3 * sizeof(unsigned long);
1535 unsigned long *sptr
= &caller
;
1536 unsigned long svalue
;
1538 while (!kstack_end(sptr
)) {
1540 if (kernel_text_address(svalue
)) {
1542 size
-= sizeof(unsigned long);
1543 if (size
<= sizeof(unsigned long))
1549 *addr
++ = 0x87654321;
1552 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1553 int map
, unsigned long caller
)
1555 if (!is_debug_pagealloc_cache(cachep
))
1559 store_stackinfo(cachep
, objp
, caller
);
1561 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1565 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1566 int map
, unsigned long caller
) {}
1570 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1572 int size
= cachep
->object_size
;
1573 addr
= &((char *)addr
)[obj_offset(cachep
)];
1575 memset(addr
, val
, size
);
1576 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1579 static void dump_line(char *data
, int offset
, int limit
)
1582 unsigned char error
= 0;
1585 pr_err("%03x: ", offset
);
1586 for (i
= 0; i
< limit
; i
++) {
1587 if (data
[offset
+ i
] != POISON_FREE
) {
1588 error
= data
[offset
+ i
];
1592 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1593 &data
[offset
], limit
, 1);
1595 if (bad_count
== 1) {
1596 error
^= POISON_FREE
;
1597 if (!(error
& (error
- 1))) {
1598 pr_err("Single bit error detected. Probably bad RAM.\n");
1600 pr_err("Run memtest86+ or a similar memory test tool.\n");
1602 pr_err("Run a memory test tool.\n");
1611 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1616 if (cachep
->flags
& SLAB_RED_ZONE
) {
1617 pr_err("Redzone: 0x%llx/0x%llx\n",
1618 *dbg_redzone1(cachep
, objp
),
1619 *dbg_redzone2(cachep
, objp
));
1622 if (cachep
->flags
& SLAB_STORE_USER
) {
1623 pr_err("Last user: [<%p>](%pSR)\n",
1624 *dbg_userword(cachep
, objp
),
1625 *dbg_userword(cachep
, objp
));
1627 realobj
= (char *)objp
+ obj_offset(cachep
);
1628 size
= cachep
->object_size
;
1629 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1632 if (i
+ limit
> size
)
1634 dump_line(realobj
, i
, limit
);
1638 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1644 if (is_debug_pagealloc_cache(cachep
))
1647 realobj
= (char *)objp
+ obj_offset(cachep
);
1648 size
= cachep
->object_size
;
1650 for (i
= 0; i
< size
; i
++) {
1651 char exp
= POISON_FREE
;
1654 if (realobj
[i
] != exp
) {
1659 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1660 print_tainted(), cachep
->name
,
1662 print_objinfo(cachep
, objp
, 0);
1664 /* Hexdump the affected line */
1667 if (i
+ limit
> size
)
1669 dump_line(realobj
, i
, limit
);
1672 /* Limit to 5 lines */
1678 /* Print some data about the neighboring objects, if they
1681 struct page
*page
= virt_to_head_page(objp
);
1684 objnr
= obj_to_index(cachep
, page
, objp
);
1686 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1687 realobj
= (char *)objp
+ obj_offset(cachep
);
1688 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1689 print_objinfo(cachep
, objp
, 2);
1691 if (objnr
+ 1 < cachep
->num
) {
1692 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1693 realobj
= (char *)objp
+ obj_offset(cachep
);
1694 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1695 print_objinfo(cachep
, objp
, 2);
1702 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1707 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1708 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1712 for (i
= 0; i
< cachep
->num
; i
++) {
1713 void *objp
= index_to_obj(cachep
, page
, i
);
1715 if (cachep
->flags
& SLAB_POISON
) {
1716 check_poison_obj(cachep
, objp
);
1717 slab_kernel_map(cachep
, objp
, 1, 0);
1719 if (cachep
->flags
& SLAB_RED_ZONE
) {
1720 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1721 slab_error(cachep
, "start of a freed object was overwritten");
1722 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1723 slab_error(cachep
, "end of a freed object was overwritten");
1728 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1735 * slab_destroy - destroy and release all objects in a slab
1736 * @cachep: cache pointer being destroyed
1737 * @page: page pointer being destroyed
1739 * Destroy all the objs in a slab page, and release the mem back to the system.
1740 * Before calling the slab page must have been unlinked from the cache. The
1741 * kmem_cache_node ->list_lock is not held/needed.
1743 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1747 freelist
= page
->freelist
;
1748 slab_destroy_debugcheck(cachep
, page
);
1749 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1750 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1752 kmem_freepages(cachep
, page
);
1755 * From now on, we don't use freelist
1756 * although actual page can be freed in rcu context
1758 if (OFF_SLAB(cachep
))
1759 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1762 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1764 struct page
*page
, *n
;
1766 list_for_each_entry_safe(page
, n
, list
, lru
) {
1767 list_del(&page
->lru
);
1768 slab_destroy(cachep
, page
);
1773 * calculate_slab_order - calculate size (page order) of slabs
1774 * @cachep: pointer to the cache that is being created
1775 * @size: size of objects to be created in this cache.
1776 * @flags: slab allocation flags
1778 * Also calculates the number of objects per slab.
1780 * This could be made much more intelligent. For now, try to avoid using
1781 * high order pages for slabs. When the gfp() functions are more friendly
1782 * towards high-order requests, this should be changed.
1784 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1785 size_t size
, unsigned long flags
)
1787 size_t left_over
= 0;
1790 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1794 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1798 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1799 if (num
> SLAB_OBJ_MAX_NUM
)
1802 if (flags
& CFLGS_OFF_SLAB
) {
1803 struct kmem_cache
*freelist_cache
;
1804 size_t freelist_size
;
1806 freelist_size
= num
* sizeof(freelist_idx_t
);
1807 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1808 if (!freelist_cache
)
1812 * Needed to avoid possible looping condition
1815 if (OFF_SLAB(freelist_cache
))
1818 /* check if off slab has enough benefit */
1819 if (freelist_cache
->size
> cachep
->size
/ 2)
1823 /* Found something acceptable - save it away */
1825 cachep
->gfporder
= gfporder
;
1826 left_over
= remainder
;
1829 * A VFS-reclaimable slab tends to have most allocations
1830 * as GFP_NOFS and we really don't want to have to be allocating
1831 * higher-order pages when we are unable to shrink dcache.
1833 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1837 * Large number of objects is good, but very large slabs are
1838 * currently bad for the gfp()s.
1840 if (gfporder
>= slab_max_order
)
1844 * Acceptable internal fragmentation?
1846 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1852 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1853 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1857 struct array_cache __percpu
*cpu_cache
;
1859 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1860 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1865 for_each_possible_cpu(cpu
) {
1866 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1867 entries
, batchcount
);
1873 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1875 if (slab_state
>= FULL
)
1876 return enable_cpucache(cachep
, gfp
);
1878 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1879 if (!cachep
->cpu_cache
)
1882 if (slab_state
== DOWN
) {
1883 /* Creation of first cache (kmem_cache). */
1884 set_up_node(kmem_cache
, CACHE_CACHE
);
1885 } else if (slab_state
== PARTIAL
) {
1886 /* For kmem_cache_node */
1887 set_up_node(cachep
, SIZE_NODE
);
1891 for_each_online_node(node
) {
1892 cachep
->node
[node
] = kmalloc_node(
1893 sizeof(struct kmem_cache_node
), gfp
, node
);
1894 BUG_ON(!cachep
->node
[node
]);
1895 kmem_cache_node_init(cachep
->node
[node
]);
1899 cachep
->node
[numa_mem_id()]->next_reap
=
1900 jiffies
+ REAPTIMEOUT_NODE
+
1901 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1903 cpu_cache_get(cachep
)->avail
= 0;
1904 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1905 cpu_cache_get(cachep
)->batchcount
= 1;
1906 cpu_cache_get(cachep
)->touched
= 0;
1907 cachep
->batchcount
= 1;
1908 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1912 unsigned long kmem_cache_flags(unsigned long object_size
,
1913 unsigned long flags
, const char *name
,
1914 void (*ctor
)(void *))
1920 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1921 unsigned long flags
, void (*ctor
)(void *))
1923 struct kmem_cache
*cachep
;
1925 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1930 * Adjust the object sizes so that we clear
1931 * the complete object on kzalloc.
1933 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1938 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1939 size_t size
, unsigned long flags
)
1945 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1948 left
= calculate_slab_order(cachep
, size
,
1949 flags
| CFLGS_OBJFREELIST_SLAB
);
1953 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1956 cachep
->colour
= left
/ cachep
->colour_off
;
1961 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1962 size_t size
, unsigned long flags
)
1969 * Always use on-slab management when SLAB_NOLEAKTRACE
1970 * to avoid recursive calls into kmemleak.
1972 if (flags
& SLAB_NOLEAKTRACE
)
1976 * Size is large, assume best to place the slab management obj
1977 * off-slab (should allow better packing of objs).
1979 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1984 * If the slab has been placed off-slab, and we have enough space then
1985 * move it on-slab. This is at the expense of any extra colouring.
1987 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1990 cachep
->colour
= left
/ cachep
->colour_off
;
1995 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1996 size_t size
, unsigned long flags
)
2002 left
= calculate_slab_order(cachep
, size
, flags
);
2006 cachep
->colour
= left
/ cachep
->colour_off
;
2012 * __kmem_cache_create - Create a cache.
2013 * @cachep: cache management descriptor
2014 * @flags: SLAB flags
2016 * Returns a ptr to the cache on success, NULL on failure.
2017 * Cannot be called within a int, but can be interrupted.
2018 * The @ctor is run when new pages are allocated by the cache.
2022 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2023 * to catch references to uninitialised memory.
2025 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2026 * for buffer overruns.
2028 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2029 * cacheline. This can be beneficial if you're counting cycles as closely
2033 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2035 size_t ralign
= BYTES_PER_WORD
;
2038 size_t size
= cachep
->size
;
2043 * Enable redzoning and last user accounting, except for caches with
2044 * large objects, if the increased size would increase the object size
2045 * above the next power of two: caches with object sizes just above a
2046 * power of two have a significant amount of internal fragmentation.
2048 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2049 2 * sizeof(unsigned long long)))
2050 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2051 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2052 flags
|= SLAB_POISON
;
2057 * Check that size is in terms of words. This is needed to avoid
2058 * unaligned accesses for some archs when redzoning is used, and makes
2059 * sure any on-slab bufctl's are also correctly aligned.
2061 if (size
& (BYTES_PER_WORD
- 1)) {
2062 size
+= (BYTES_PER_WORD
- 1);
2063 size
&= ~(BYTES_PER_WORD
- 1);
2066 if (flags
& SLAB_RED_ZONE
) {
2067 ralign
= REDZONE_ALIGN
;
2068 /* If redzoning, ensure that the second redzone is suitably
2069 * aligned, by adjusting the object size accordingly. */
2070 size
+= REDZONE_ALIGN
- 1;
2071 size
&= ~(REDZONE_ALIGN
- 1);
2074 /* 3) caller mandated alignment */
2075 if (ralign
< cachep
->align
) {
2076 ralign
= cachep
->align
;
2078 /* disable debug if necessary */
2079 if (ralign
> __alignof__(unsigned long long))
2080 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2084 cachep
->align
= ralign
;
2085 cachep
->colour_off
= cache_line_size();
2086 /* Offset must be a multiple of the alignment. */
2087 if (cachep
->colour_off
< cachep
->align
)
2088 cachep
->colour_off
= cachep
->align
;
2090 if (slab_is_available())
2098 * Both debugging options require word-alignment which is calculated
2101 if (flags
& SLAB_RED_ZONE
) {
2102 /* add space for red zone words */
2103 cachep
->obj_offset
+= sizeof(unsigned long long);
2104 size
+= 2 * sizeof(unsigned long long);
2106 if (flags
& SLAB_STORE_USER
) {
2107 /* user store requires one word storage behind the end of
2108 * the real object. But if the second red zone needs to be
2109 * aligned to 64 bits, we must allow that much space.
2111 if (flags
& SLAB_RED_ZONE
)
2112 size
+= REDZONE_ALIGN
;
2114 size
+= BYTES_PER_WORD
;
2118 kasan_cache_create(cachep
, &size
, &flags
);
2120 size
= ALIGN(size
, cachep
->align
);
2122 * We should restrict the number of objects in a slab to implement
2123 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2125 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2126 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2130 * To activate debug pagealloc, off-slab management is necessary
2131 * requirement. In early phase of initialization, small sized slab
2132 * doesn't get initialized so it would not be possible. So, we need
2133 * to check size >= 256. It guarantees that all necessary small
2134 * sized slab is initialized in current slab initialization sequence.
2136 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2137 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2138 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2139 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2141 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2142 flags
|= CFLGS_OFF_SLAB
;
2143 cachep
->obj_offset
+= tmp_size
- size
;
2151 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2152 flags
|= CFLGS_OBJFREELIST_SLAB
;
2156 if (set_off_slab_cache(cachep
, size
, flags
)) {
2157 flags
|= CFLGS_OFF_SLAB
;
2161 if (set_on_slab_cache(cachep
, size
, flags
))
2167 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2168 cachep
->flags
= flags
;
2169 cachep
->allocflags
= __GFP_COMP
;
2170 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2171 cachep
->allocflags
|= GFP_DMA
;
2172 cachep
->size
= size
;
2173 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2177 * If we're going to use the generic kernel_map_pages()
2178 * poisoning, then it's going to smash the contents of
2179 * the redzone and userword anyhow, so switch them off.
2181 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2182 (cachep
->flags
& SLAB_POISON
) &&
2183 is_debug_pagealloc_cache(cachep
))
2184 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2187 if (OFF_SLAB(cachep
)) {
2188 cachep
->freelist_cache
=
2189 kmalloc_slab(cachep
->freelist_size
, 0u);
2192 err
= setup_cpu_cache(cachep
, gfp
);
2194 __kmem_cache_release(cachep
);
2202 static void check_irq_off(void)
2204 BUG_ON(!irqs_disabled());
2207 static void check_irq_on(void)
2209 BUG_ON(irqs_disabled());
2212 static void check_mutex_acquired(void)
2214 BUG_ON(!mutex_is_locked(&slab_mutex
));
2217 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2221 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2225 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2229 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2234 #define check_irq_off() do { } while(0)
2235 #define check_irq_on() do { } while(0)
2236 #define check_mutex_acquired() do { } while(0)
2237 #define check_spinlock_acquired(x) do { } while(0)
2238 #define check_spinlock_acquired_node(x, y) do { } while(0)
2241 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2242 int node
, bool free_all
, struct list_head
*list
)
2246 if (!ac
|| !ac
->avail
)
2249 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2250 if (tofree
> ac
->avail
)
2251 tofree
= (ac
->avail
+ 1) / 2;
2253 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2254 ac
->avail
-= tofree
;
2255 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2258 static void do_drain(void *arg
)
2260 struct kmem_cache
*cachep
= arg
;
2261 struct array_cache
*ac
;
2262 int node
= numa_mem_id();
2263 struct kmem_cache_node
*n
;
2267 ac
= cpu_cache_get(cachep
);
2268 n
= get_node(cachep
, node
);
2269 spin_lock(&n
->list_lock
);
2270 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2271 spin_unlock(&n
->list_lock
);
2272 slabs_destroy(cachep
, &list
);
2276 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2278 struct kmem_cache_node
*n
;
2282 on_each_cpu(do_drain
, cachep
, 1);
2284 for_each_kmem_cache_node(cachep
, node
, n
)
2286 drain_alien_cache(cachep
, n
->alien
);
2288 for_each_kmem_cache_node(cachep
, node
, n
) {
2289 spin_lock_irq(&n
->list_lock
);
2290 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2291 spin_unlock_irq(&n
->list_lock
);
2293 slabs_destroy(cachep
, &list
);
2298 * Remove slabs from the list of free slabs.
2299 * Specify the number of slabs to drain in tofree.
2301 * Returns the actual number of slabs released.
2303 static int drain_freelist(struct kmem_cache
*cache
,
2304 struct kmem_cache_node
*n
, int tofree
)
2306 struct list_head
*p
;
2311 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2313 spin_lock_irq(&n
->list_lock
);
2314 p
= n
->slabs_free
.prev
;
2315 if (p
== &n
->slabs_free
) {
2316 spin_unlock_irq(&n
->list_lock
);
2320 page
= list_entry(p
, struct page
, lru
);
2321 list_del(&page
->lru
);
2323 * Safe to drop the lock. The slab is no longer linked
2326 n
->free_objects
-= cache
->num
;
2327 spin_unlock_irq(&n
->list_lock
);
2328 slab_destroy(cache
, page
);
2335 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2339 struct kmem_cache_node
*n
;
2341 drain_cpu_caches(cachep
);
2344 for_each_kmem_cache_node(cachep
, node
, n
) {
2345 drain_freelist(cachep
, n
, INT_MAX
);
2347 ret
+= !list_empty(&n
->slabs_full
) ||
2348 !list_empty(&n
->slabs_partial
);
2350 return (ret
? 1 : 0);
2353 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2355 return __kmem_cache_shrink(cachep
, false);
2358 void __kmem_cache_release(struct kmem_cache
*cachep
)
2361 struct kmem_cache_node
*n
;
2363 free_percpu(cachep
->cpu_cache
);
2365 /* NUMA: free the node structures */
2366 for_each_kmem_cache_node(cachep
, i
, n
) {
2368 free_alien_cache(n
->alien
);
2370 cachep
->node
[i
] = NULL
;
2375 * Get the memory for a slab management obj.
2377 * For a slab cache when the slab descriptor is off-slab, the
2378 * slab descriptor can't come from the same cache which is being created,
2379 * Because if it is the case, that means we defer the creation of
2380 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2381 * And we eventually call down to __kmem_cache_create(), which
2382 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2383 * This is a "chicken-and-egg" problem.
2385 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2386 * which are all initialized during kmem_cache_init().
2388 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2389 struct page
*page
, int colour_off
,
2390 gfp_t local_flags
, int nodeid
)
2393 void *addr
= page_address(page
);
2395 page
->s_mem
= addr
+ colour_off
;
2398 if (OBJFREELIST_SLAB(cachep
))
2400 else if (OFF_SLAB(cachep
)) {
2401 /* Slab management obj is off-slab. */
2402 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2403 local_flags
, nodeid
);
2407 /* We will use last bytes at the slab for freelist */
2408 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2409 cachep
->freelist_size
;
2415 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2417 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2420 static inline void set_free_obj(struct page
*page
,
2421 unsigned int idx
, freelist_idx_t val
)
2423 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2426 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2431 for (i
= 0; i
< cachep
->num
; i
++) {
2432 void *objp
= index_to_obj(cachep
, page
, i
);
2434 if (cachep
->flags
& SLAB_STORE_USER
)
2435 *dbg_userword(cachep
, objp
) = NULL
;
2437 if (cachep
->flags
& SLAB_RED_ZONE
) {
2438 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2439 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2442 * Constructors are not allowed to allocate memory from the same
2443 * cache which they are a constructor for. Otherwise, deadlock.
2444 * They must also be threaded.
2446 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2447 kasan_unpoison_object_data(cachep
,
2448 objp
+ obj_offset(cachep
));
2449 cachep
->ctor(objp
+ obj_offset(cachep
));
2450 kasan_poison_object_data(
2451 cachep
, objp
+ obj_offset(cachep
));
2454 if (cachep
->flags
& SLAB_RED_ZONE
) {
2455 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2456 slab_error(cachep
, "constructor overwrote the end of an object");
2457 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2458 slab_error(cachep
, "constructor overwrote the start of an object");
2460 /* need to poison the objs? */
2461 if (cachep
->flags
& SLAB_POISON
) {
2462 poison_obj(cachep
, objp
, POISON_FREE
);
2463 slab_kernel_map(cachep
, objp
, 0, 0);
2469 static void cache_init_objs(struct kmem_cache
*cachep
,
2475 cache_init_objs_debug(cachep
, page
);
2477 if (OBJFREELIST_SLAB(cachep
)) {
2478 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2482 for (i
= 0; i
< cachep
->num
; i
++) {
2483 /* constructor could break poison info */
2484 if (DEBUG
== 0 && cachep
->ctor
) {
2485 objp
= index_to_obj(cachep
, page
, i
);
2486 kasan_unpoison_object_data(cachep
, objp
);
2488 kasan_poison_object_data(cachep
, objp
);
2491 set_free_obj(page
, i
, i
);
2495 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2497 if (CONFIG_ZONE_DMA_FLAG
) {
2498 if (flags
& GFP_DMA
)
2499 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2501 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2505 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2509 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2513 if (cachep
->flags
& SLAB_STORE_USER
)
2514 set_store_user_dirty(cachep
);
2520 static void slab_put_obj(struct kmem_cache
*cachep
,
2521 struct page
*page
, void *objp
)
2523 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2527 /* Verify double free bug */
2528 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2529 if (get_free_obj(page
, i
) == objnr
) {
2530 pr_err("slab: double free detected in cache '%s', objp %p\n",
2531 cachep
->name
, objp
);
2537 if (!page
->freelist
)
2538 page
->freelist
= objp
+ obj_offset(cachep
);
2540 set_free_obj(page
, page
->active
, objnr
);
2544 * Map pages beginning at addr to the given cache and slab. This is required
2545 * for the slab allocator to be able to lookup the cache and slab of a
2546 * virtual address for kfree, ksize, and slab debugging.
2548 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2551 page
->slab_cache
= cache
;
2552 page
->freelist
= freelist
;
2556 * Grow (by 1) the number of slabs within a cache. This is called by
2557 * kmem_cache_alloc() when there are no active objs left in a cache.
2559 static int cache_grow(struct kmem_cache
*cachep
,
2560 gfp_t flags
, int nodeid
, struct page
*page
)
2565 struct kmem_cache_node
*n
;
2568 * Be lazy and only check for valid flags here, keeping it out of the
2569 * critical path in kmem_cache_alloc().
2571 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2572 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2575 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2577 /* Take the node list lock to change the colour_next on this node */
2579 n
= get_node(cachep
, nodeid
);
2580 spin_lock(&n
->list_lock
);
2582 /* Get colour for the slab, and cal the next value. */
2583 offset
= n
->colour_next
;
2585 if (n
->colour_next
>= cachep
->colour
)
2587 spin_unlock(&n
->list_lock
);
2589 offset
*= cachep
->colour_off
;
2591 if (gfpflags_allow_blocking(local_flags
))
2595 * The test for missing atomic flag is performed here, rather than
2596 * the more obvious place, simply to reduce the critical path length
2597 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2598 * will eventually be caught here (where it matters).
2600 kmem_flagcheck(cachep
, flags
);
2603 * Get mem for the objs. Attempt to allocate a physical page from
2607 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2611 /* Get slab management. */
2612 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2613 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2614 if (OFF_SLAB(cachep
) && !freelist
)
2617 slab_map_pages(cachep
, page
, freelist
);
2619 kasan_poison_slab(page
);
2620 cache_init_objs(cachep
, page
);
2622 if (gfpflags_allow_blocking(local_flags
))
2623 local_irq_disable();
2625 spin_lock(&n
->list_lock
);
2627 /* Make slab active. */
2628 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2629 STATS_INC_GROWN(cachep
);
2630 n
->free_objects
+= cachep
->num
;
2631 spin_unlock(&n
->list_lock
);
2634 kmem_freepages(cachep
, page
);
2636 if (gfpflags_allow_blocking(local_flags
))
2637 local_irq_disable();
2644 * Perform extra freeing checks:
2645 * - detect bad pointers.
2646 * - POISON/RED_ZONE checking
2648 static void kfree_debugcheck(const void *objp
)
2650 if (!virt_addr_valid(objp
)) {
2651 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2652 (unsigned long)objp
);
2657 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2659 unsigned long long redzone1
, redzone2
;
2661 redzone1
= *dbg_redzone1(cache
, obj
);
2662 redzone2
= *dbg_redzone2(cache
, obj
);
2667 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2670 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2671 slab_error(cache
, "double free detected");
2673 slab_error(cache
, "memory outside object was overwritten");
2675 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2676 obj
, redzone1
, redzone2
);
2679 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2680 unsigned long caller
)
2685 BUG_ON(virt_to_cache(objp
) != cachep
);
2687 objp
-= obj_offset(cachep
);
2688 kfree_debugcheck(objp
);
2689 page
= virt_to_head_page(objp
);
2691 if (cachep
->flags
& SLAB_RED_ZONE
) {
2692 verify_redzone_free(cachep
, objp
);
2693 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2694 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2696 if (cachep
->flags
& SLAB_STORE_USER
) {
2697 set_store_user_dirty(cachep
);
2698 *dbg_userword(cachep
, objp
) = (void *)caller
;
2701 objnr
= obj_to_index(cachep
, page
, objp
);
2703 BUG_ON(objnr
>= cachep
->num
);
2704 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2706 if (cachep
->flags
& SLAB_POISON
) {
2707 poison_obj(cachep
, objp
, POISON_FREE
);
2708 slab_kernel_map(cachep
, objp
, 0, caller
);
2714 #define kfree_debugcheck(x) do { } while(0)
2715 #define cache_free_debugcheck(x,objp,z) (objp)
2718 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2726 objp
= next
- obj_offset(cachep
);
2727 next
= *(void **)next
;
2728 poison_obj(cachep
, objp
, POISON_FREE
);
2733 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2734 struct kmem_cache_node
*n
, struct page
*page
,
2737 /* move slabp to correct slabp list: */
2738 list_del(&page
->lru
);
2739 if (page
->active
== cachep
->num
) {
2740 list_add(&page
->lru
, &n
->slabs_full
);
2741 if (OBJFREELIST_SLAB(cachep
)) {
2743 /* Poisoning will be done without holding the lock */
2744 if (cachep
->flags
& SLAB_POISON
) {
2745 void **objp
= page
->freelist
;
2751 page
->freelist
= NULL
;
2754 list_add(&page
->lru
, &n
->slabs_partial
);
2757 /* Try to find non-pfmemalloc slab if needed */
2758 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2759 struct page
*page
, bool pfmemalloc
)
2767 if (!PageSlabPfmemalloc(page
))
2770 /* No need to keep pfmemalloc slab if we have enough free objects */
2771 if (n
->free_objects
> n
->free_limit
) {
2772 ClearPageSlabPfmemalloc(page
);
2776 /* Move pfmemalloc slab to the end of list to speed up next search */
2777 list_del(&page
->lru
);
2779 list_add_tail(&page
->lru
, &n
->slabs_free
);
2781 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2783 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2784 if (!PageSlabPfmemalloc(page
))
2788 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2789 if (!PageSlabPfmemalloc(page
))
2796 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2800 page
= list_first_entry_or_null(&n
->slabs_partial
,
2803 n
->free_touched
= 1;
2804 page
= list_first_entry_or_null(&n
->slabs_free
,
2808 if (sk_memalloc_socks())
2809 return get_valid_first_slab(n
, page
, pfmemalloc
);
2814 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2815 struct kmem_cache_node
*n
, gfp_t flags
)
2821 if (!gfp_pfmemalloc_allowed(flags
))
2824 spin_lock(&n
->list_lock
);
2825 page
= get_first_slab(n
, true);
2827 spin_unlock(&n
->list_lock
);
2831 obj
= slab_get_obj(cachep
, page
);
2834 fixup_slab_list(cachep
, n
, page
, &list
);
2836 spin_unlock(&n
->list_lock
);
2837 fixup_objfreelist_debug(cachep
, &list
);
2842 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2845 struct kmem_cache_node
*n
;
2846 struct array_cache
*ac
;
2851 node
= numa_mem_id();
2854 ac
= cpu_cache_get(cachep
);
2855 batchcount
= ac
->batchcount
;
2856 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2858 * If there was little recent activity on this cache, then
2859 * perform only a partial refill. Otherwise we could generate
2862 batchcount
= BATCHREFILL_LIMIT
;
2864 n
= get_node(cachep
, node
);
2866 BUG_ON(ac
->avail
> 0 || !n
);
2867 spin_lock(&n
->list_lock
);
2869 /* See if we can refill from the shared array */
2870 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2871 n
->shared
->touched
= 1;
2875 while (batchcount
> 0) {
2877 /* Get slab alloc is to come from. */
2878 page
= get_first_slab(n
, false);
2882 check_spinlock_acquired(cachep
);
2885 * The slab was either on partial or free list so
2886 * there must be at least one object available for
2889 BUG_ON(page
->active
>= cachep
->num
);
2891 while (page
->active
< cachep
->num
&& batchcount
--) {
2892 STATS_INC_ALLOCED(cachep
);
2893 STATS_INC_ACTIVE(cachep
);
2894 STATS_SET_HIGH(cachep
);
2896 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2899 fixup_slab_list(cachep
, n
, page
, &list
);
2903 n
->free_objects
-= ac
->avail
;
2905 spin_unlock(&n
->list_lock
);
2906 fixup_objfreelist_debug(cachep
, &list
);
2908 if (unlikely(!ac
->avail
)) {
2911 /* Check if we can use obj in pfmemalloc slab */
2912 if (sk_memalloc_socks()) {
2913 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2919 x
= cache_grow(cachep
, gfp_exact_node(flags
), node
, NULL
);
2921 /* cache_grow can reenable interrupts, then ac could change. */
2922 ac
= cpu_cache_get(cachep
);
2923 node
= numa_mem_id();
2925 /* no objects in sight? abort */
2926 if (!x
&& ac
->avail
== 0)
2929 if (!ac
->avail
) /* objects refilled by interrupt? */
2934 return ac
->entry
[--ac
->avail
];
2937 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2940 might_sleep_if(gfpflags_allow_blocking(flags
));
2942 kmem_flagcheck(cachep
, flags
);
2947 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2948 gfp_t flags
, void *objp
, unsigned long caller
)
2952 if (cachep
->flags
& SLAB_POISON
) {
2953 check_poison_obj(cachep
, objp
);
2954 slab_kernel_map(cachep
, objp
, 1, 0);
2955 poison_obj(cachep
, objp
, POISON_INUSE
);
2957 if (cachep
->flags
& SLAB_STORE_USER
)
2958 *dbg_userword(cachep
, objp
) = (void *)caller
;
2960 if (cachep
->flags
& SLAB_RED_ZONE
) {
2961 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2962 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2963 slab_error(cachep
, "double free, or memory outside object was overwritten");
2964 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2965 objp
, *dbg_redzone1(cachep
, objp
),
2966 *dbg_redzone2(cachep
, objp
));
2968 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2969 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2972 objp
+= obj_offset(cachep
);
2973 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2975 if (ARCH_SLAB_MINALIGN
&&
2976 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2977 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2978 objp
, (int)ARCH_SLAB_MINALIGN
);
2983 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2986 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2989 struct array_cache
*ac
;
2993 ac
= cpu_cache_get(cachep
);
2994 if (likely(ac
->avail
)) {
2996 objp
= ac
->entry
[--ac
->avail
];
2998 STATS_INC_ALLOCHIT(cachep
);
3002 STATS_INC_ALLOCMISS(cachep
);
3003 objp
= cache_alloc_refill(cachep
, flags
);
3005 * the 'ac' may be updated by cache_alloc_refill(),
3006 * and kmemleak_erase() requires its correct value.
3008 ac
= cpu_cache_get(cachep
);
3012 * To avoid a false negative, if an object that is in one of the
3013 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3014 * treat the array pointers as a reference to the object.
3017 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3023 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3025 * If we are in_interrupt, then process context, including cpusets and
3026 * mempolicy, may not apply and should not be used for allocation policy.
3028 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3030 int nid_alloc
, nid_here
;
3032 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3034 nid_alloc
= nid_here
= numa_mem_id();
3035 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3036 nid_alloc
= cpuset_slab_spread_node();
3037 else if (current
->mempolicy
)
3038 nid_alloc
= mempolicy_slab_node();
3039 if (nid_alloc
!= nid_here
)
3040 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3045 * Fallback function if there was no memory available and no objects on a
3046 * certain node and fall back is permitted. First we scan all the
3047 * available node for available objects. If that fails then we
3048 * perform an allocation without specifying a node. This allows the page
3049 * allocator to do its reclaim / fallback magic. We then insert the
3050 * slab into the proper nodelist and then allocate from it.
3052 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3054 struct zonelist
*zonelist
;
3058 enum zone_type high_zoneidx
= gfp_zone(flags
);
3061 unsigned int cpuset_mems_cookie
;
3063 if (flags
& __GFP_THISNODE
)
3066 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3069 cpuset_mems_cookie
= read_mems_allowed_begin();
3070 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3074 * Look through allowed nodes for objects available
3075 * from existing per node queues.
3077 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3078 nid
= zone_to_nid(zone
);
3080 if (cpuset_zone_allowed(zone
, flags
) &&
3081 get_node(cache
, nid
) &&
3082 get_node(cache
, nid
)->free_objects
) {
3083 obj
= ____cache_alloc_node(cache
,
3084 gfp_exact_node(flags
), nid
);
3092 * This allocation will be performed within the constraints
3093 * of the current cpuset / memory policy requirements.
3094 * We may trigger various forms of reclaim on the allowed
3095 * set and go into memory reserves if necessary.
3099 if (gfpflags_allow_blocking(local_flags
))
3101 kmem_flagcheck(cache
, flags
);
3102 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3103 if (gfpflags_allow_blocking(local_flags
))
3104 local_irq_disable();
3107 * Insert into the appropriate per node queues
3109 nid
= page_to_nid(page
);
3110 if (cache_grow(cache
, flags
, nid
, page
)) {
3111 obj
= ____cache_alloc_node(cache
,
3112 gfp_exact_node(flags
), nid
);
3115 * Another processor may allocate the
3116 * objects in the slab since we are
3117 * not holding any locks.
3121 /* cache_grow already freed obj */
3127 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3133 * A interface to enable slab creation on nodeid
3135 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3139 struct kmem_cache_node
*n
;
3144 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3145 n
= get_node(cachep
, nodeid
);
3150 spin_lock(&n
->list_lock
);
3151 page
= get_first_slab(n
, false);
3155 check_spinlock_acquired_node(cachep
, nodeid
);
3157 STATS_INC_NODEALLOCS(cachep
);
3158 STATS_INC_ACTIVE(cachep
);
3159 STATS_SET_HIGH(cachep
);
3161 BUG_ON(page
->active
== cachep
->num
);
3163 obj
= slab_get_obj(cachep
, page
);
3166 fixup_slab_list(cachep
, n
, page
, &list
);
3168 spin_unlock(&n
->list_lock
);
3169 fixup_objfreelist_debug(cachep
, &list
);
3173 spin_unlock(&n
->list_lock
);
3174 x
= cache_grow(cachep
, gfp_exact_node(flags
), nodeid
, NULL
);
3178 return fallback_alloc(cachep
, flags
);
3184 static __always_inline
void *
3185 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3186 unsigned long caller
)
3188 unsigned long save_flags
;
3190 int slab_node
= numa_mem_id();
3192 flags
&= gfp_allowed_mask
;
3193 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3194 if (unlikely(!cachep
))
3197 cache_alloc_debugcheck_before(cachep
, flags
);
3198 local_irq_save(save_flags
);
3200 if (nodeid
== NUMA_NO_NODE
)
3203 if (unlikely(!get_node(cachep
, nodeid
))) {
3204 /* Node not bootstrapped yet */
3205 ptr
= fallback_alloc(cachep
, flags
);
3209 if (nodeid
== slab_node
) {
3211 * Use the locally cached objects if possible.
3212 * However ____cache_alloc does not allow fallback
3213 * to other nodes. It may fail while we still have
3214 * objects on other nodes available.
3216 ptr
= ____cache_alloc(cachep
, flags
);
3220 /* ___cache_alloc_node can fall back to other nodes */
3221 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3223 local_irq_restore(save_flags
);
3224 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3226 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3227 memset(ptr
, 0, cachep
->object_size
);
3229 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3233 static __always_inline
void *
3234 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3238 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3239 objp
= alternate_node_alloc(cache
, flags
);
3243 objp
= ____cache_alloc(cache
, flags
);
3246 * We may just have run out of memory on the local node.
3247 * ____cache_alloc_node() knows how to locate memory on other nodes
3250 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3257 static __always_inline
void *
3258 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3260 return ____cache_alloc(cachep
, flags
);
3263 #endif /* CONFIG_NUMA */
3265 static __always_inline
void *
3266 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3268 unsigned long save_flags
;
3271 flags
&= gfp_allowed_mask
;
3272 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3273 if (unlikely(!cachep
))
3276 cache_alloc_debugcheck_before(cachep
, flags
);
3277 local_irq_save(save_flags
);
3278 objp
= __do_cache_alloc(cachep
, flags
);
3279 local_irq_restore(save_flags
);
3280 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3283 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3284 memset(objp
, 0, cachep
->object_size
);
3286 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3291 * Caller needs to acquire correct kmem_cache_node's list_lock
3292 * @list: List of detached free slabs should be freed by caller
3294 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3295 int nr_objects
, int node
, struct list_head
*list
)
3298 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3301 n
->free_objects
+= nr_objects
;
3303 for (i
= 0; i
< nr_objects
; i
++) {
3309 page
= virt_to_head_page(objp
);
3310 list_del(&page
->lru
);
3311 check_spinlock_acquired_node(cachep
, node
);
3312 slab_put_obj(cachep
, page
, objp
);
3313 STATS_DEC_ACTIVE(cachep
);
3315 /* fixup slab chains */
3316 if (page
->active
== 0)
3317 list_add(&page
->lru
, &n
->slabs_free
);
3319 /* Unconditionally move a slab to the end of the
3320 * partial list on free - maximum time for the
3321 * other objects to be freed, too.
3323 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3327 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3328 n
->free_objects
-= cachep
->num
;
3330 page
= list_last_entry(&n
->slabs_free
, struct page
, lru
);
3331 list_del(&page
->lru
);
3332 list_add(&page
->lru
, list
);
3336 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3339 struct kmem_cache_node
*n
;
3340 int node
= numa_mem_id();
3343 batchcount
= ac
->batchcount
;
3346 n
= get_node(cachep
, node
);
3347 spin_lock(&n
->list_lock
);
3349 struct array_cache
*shared_array
= n
->shared
;
3350 int max
= shared_array
->limit
- shared_array
->avail
;
3352 if (batchcount
> max
)
3354 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3355 ac
->entry
, sizeof(void *) * batchcount
);
3356 shared_array
->avail
+= batchcount
;
3361 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3368 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3369 BUG_ON(page
->active
);
3373 STATS_SET_FREEABLE(cachep
, i
);
3376 spin_unlock(&n
->list_lock
);
3377 slabs_destroy(cachep
, &list
);
3378 ac
->avail
-= batchcount
;
3379 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3383 * Release an obj back to its cache. If the obj has a constructed state, it must
3384 * be in this state _before_ it is released. Called with disabled ints.
3386 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3387 unsigned long caller
)
3389 struct array_cache
*ac
= cpu_cache_get(cachep
);
3391 kasan_slab_free(cachep
, objp
);
3394 kmemleak_free_recursive(objp
, cachep
->flags
);
3395 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3397 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3400 * Skip calling cache_free_alien() when the platform is not numa.
3401 * This will avoid cache misses that happen while accessing slabp (which
3402 * is per page memory reference) to get nodeid. Instead use a global
3403 * variable to skip the call, which is mostly likely to be present in
3406 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3409 if (ac
->avail
< ac
->limit
) {
3410 STATS_INC_FREEHIT(cachep
);
3412 STATS_INC_FREEMISS(cachep
);
3413 cache_flusharray(cachep
, ac
);
3416 if (sk_memalloc_socks()) {
3417 struct page
*page
= virt_to_head_page(objp
);
3419 if (unlikely(PageSlabPfmemalloc(page
))) {
3420 cache_free_pfmemalloc(cachep
, page
, objp
);
3425 ac
->entry
[ac
->avail
++] = objp
;
3429 * kmem_cache_alloc - Allocate an object
3430 * @cachep: The cache to allocate from.
3431 * @flags: See kmalloc().
3433 * Allocate an object from this cache. The flags are only relevant
3434 * if the cache has no available objects.
3436 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3438 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3440 kasan_slab_alloc(cachep
, ret
, flags
);
3441 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3442 cachep
->object_size
, cachep
->size
, flags
);
3446 EXPORT_SYMBOL(kmem_cache_alloc
);
3448 static __always_inline
void
3449 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3450 size_t size
, void **p
, unsigned long caller
)
3454 for (i
= 0; i
< size
; i
++)
3455 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3458 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3463 s
= slab_pre_alloc_hook(s
, flags
);
3467 cache_alloc_debugcheck_before(s
, flags
);
3469 local_irq_disable();
3470 for (i
= 0; i
< size
; i
++) {
3471 void *objp
= __do_cache_alloc(s
, flags
);
3473 if (unlikely(!objp
))
3479 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3481 /* Clear memory outside IRQ disabled section */
3482 if (unlikely(flags
& __GFP_ZERO
))
3483 for (i
= 0; i
< size
; i
++)
3484 memset(p
[i
], 0, s
->object_size
);
3486 slab_post_alloc_hook(s
, flags
, size
, p
);
3487 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3491 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3492 slab_post_alloc_hook(s
, flags
, i
, p
);
3493 __kmem_cache_free_bulk(s
, i
, p
);
3496 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3498 #ifdef CONFIG_TRACING
3500 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3504 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3506 kasan_kmalloc(cachep
, ret
, size
, flags
);
3507 trace_kmalloc(_RET_IP_
, ret
,
3508 size
, cachep
->size
, flags
);
3511 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3516 * kmem_cache_alloc_node - Allocate an object on the specified node
3517 * @cachep: The cache to allocate from.
3518 * @flags: See kmalloc().
3519 * @nodeid: node number of the target node.
3521 * Identical to kmem_cache_alloc but it will allocate memory on the given
3522 * node, which can improve the performance for cpu bound structures.
3524 * Fallback to other node is possible if __GFP_THISNODE is not set.
3526 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3528 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3530 kasan_slab_alloc(cachep
, ret
, flags
);
3531 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3532 cachep
->object_size
, cachep
->size
,
3537 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3539 #ifdef CONFIG_TRACING
3540 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3547 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3549 kasan_kmalloc(cachep
, ret
, size
, flags
);
3550 trace_kmalloc_node(_RET_IP_
, ret
,
3555 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3558 static __always_inline
void *
3559 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3561 struct kmem_cache
*cachep
;
3564 cachep
= kmalloc_slab(size
, flags
);
3565 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3567 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3568 kasan_kmalloc(cachep
, ret
, size
, flags
);
3573 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3575 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3577 EXPORT_SYMBOL(__kmalloc_node
);
3579 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3580 int node
, unsigned long caller
)
3582 return __do_kmalloc_node(size
, flags
, node
, caller
);
3584 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3585 #endif /* CONFIG_NUMA */
3588 * __do_kmalloc - allocate memory
3589 * @size: how many bytes of memory are required.
3590 * @flags: the type of memory to allocate (see kmalloc).
3591 * @caller: function caller for debug tracking of the caller
3593 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3594 unsigned long caller
)
3596 struct kmem_cache
*cachep
;
3599 cachep
= kmalloc_slab(size
, flags
);
3600 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3602 ret
= slab_alloc(cachep
, flags
, caller
);
3604 kasan_kmalloc(cachep
, ret
, size
, flags
);
3605 trace_kmalloc(caller
, ret
,
3606 size
, cachep
->size
, flags
);
3611 void *__kmalloc(size_t size
, gfp_t flags
)
3613 return __do_kmalloc(size
, flags
, _RET_IP_
);
3615 EXPORT_SYMBOL(__kmalloc
);
3617 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3619 return __do_kmalloc(size
, flags
, caller
);
3621 EXPORT_SYMBOL(__kmalloc_track_caller
);
3624 * kmem_cache_free - Deallocate an object
3625 * @cachep: The cache the allocation was from.
3626 * @objp: The previously allocated object.
3628 * Free an object which was previously allocated from this
3631 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3633 unsigned long flags
;
3634 cachep
= cache_from_obj(cachep
, objp
);
3638 local_irq_save(flags
);
3639 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3640 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3641 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3642 __cache_free(cachep
, objp
, _RET_IP_
);
3643 local_irq_restore(flags
);
3645 trace_kmem_cache_free(_RET_IP_
, objp
);
3647 EXPORT_SYMBOL(kmem_cache_free
);
3649 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3651 struct kmem_cache
*s
;
3654 local_irq_disable();
3655 for (i
= 0; i
< size
; i
++) {
3658 if (!orig_s
) /* called via kfree_bulk */
3659 s
= virt_to_cache(objp
);
3661 s
= cache_from_obj(orig_s
, objp
);
3663 debug_check_no_locks_freed(objp
, s
->object_size
);
3664 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3665 debug_check_no_obj_freed(objp
, s
->object_size
);
3667 __cache_free(s
, objp
, _RET_IP_
);
3671 /* FIXME: add tracing */
3673 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3676 * kfree - free previously allocated memory
3677 * @objp: pointer returned by kmalloc.
3679 * If @objp is NULL, no operation is performed.
3681 * Don't free memory not originally allocated by kmalloc()
3682 * or you will run into trouble.
3684 void kfree(const void *objp
)
3686 struct kmem_cache
*c
;
3687 unsigned long flags
;
3689 trace_kfree(_RET_IP_
, objp
);
3691 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3693 local_irq_save(flags
);
3694 kfree_debugcheck(objp
);
3695 c
= virt_to_cache(objp
);
3696 debug_check_no_locks_freed(objp
, c
->object_size
);
3698 debug_check_no_obj_freed(objp
, c
->object_size
);
3699 __cache_free(c
, (void *)objp
, _RET_IP_
);
3700 local_irq_restore(flags
);
3702 EXPORT_SYMBOL(kfree
);
3705 * This initializes kmem_cache_node or resizes various caches for all nodes.
3707 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3711 struct kmem_cache_node
*n
;
3713 for_each_online_node(node
) {
3714 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3723 if (!cachep
->list
.next
) {
3724 /* Cache is not active yet. Roll back what we did */
3727 n
= get_node(cachep
, node
);
3730 free_alien_cache(n
->alien
);
3732 cachep
->node
[node
] = NULL
;
3740 /* Always called with the slab_mutex held */
3741 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3742 int batchcount
, int shared
, gfp_t gfp
)
3744 struct array_cache __percpu
*cpu_cache
, *prev
;
3747 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3751 prev
= cachep
->cpu_cache
;
3752 cachep
->cpu_cache
= cpu_cache
;
3753 kick_all_cpus_sync();
3756 cachep
->batchcount
= batchcount
;
3757 cachep
->limit
= limit
;
3758 cachep
->shared
= shared
;
3763 for_each_online_cpu(cpu
) {
3766 struct kmem_cache_node
*n
;
3767 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3769 node
= cpu_to_mem(cpu
);
3770 n
= get_node(cachep
, node
);
3771 spin_lock_irq(&n
->list_lock
);
3772 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3773 spin_unlock_irq(&n
->list_lock
);
3774 slabs_destroy(cachep
, &list
);
3779 return setup_kmem_cache_nodes(cachep
, gfp
);
3782 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3783 int batchcount
, int shared
, gfp_t gfp
)
3786 struct kmem_cache
*c
;
3788 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3790 if (slab_state
< FULL
)
3793 if ((ret
< 0) || !is_root_cache(cachep
))
3796 lockdep_assert_held(&slab_mutex
);
3797 for_each_memcg_cache(c
, cachep
) {
3798 /* return value determined by the root cache only */
3799 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3805 /* Called with slab_mutex held always */
3806 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3813 if (!is_root_cache(cachep
)) {
3814 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3815 limit
= root
->limit
;
3816 shared
= root
->shared
;
3817 batchcount
= root
->batchcount
;
3820 if (limit
&& shared
&& batchcount
)
3823 * The head array serves three purposes:
3824 * - create a LIFO ordering, i.e. return objects that are cache-warm
3825 * - reduce the number of spinlock operations.
3826 * - reduce the number of linked list operations on the slab and
3827 * bufctl chains: array operations are cheaper.
3828 * The numbers are guessed, we should auto-tune as described by
3831 if (cachep
->size
> 131072)
3833 else if (cachep
->size
> PAGE_SIZE
)
3835 else if (cachep
->size
> 1024)
3837 else if (cachep
->size
> 256)
3843 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3844 * allocation behaviour: Most allocs on one cpu, most free operations
3845 * on another cpu. For these cases, an efficient object passing between
3846 * cpus is necessary. This is provided by a shared array. The array
3847 * replaces Bonwick's magazine layer.
3848 * On uniprocessor, it's functionally equivalent (but less efficient)
3849 * to a larger limit. Thus disabled by default.
3852 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3857 * With debugging enabled, large batchcount lead to excessively long
3858 * periods with disabled local interrupts. Limit the batchcount
3863 batchcount
= (limit
+ 1) / 2;
3865 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3867 pr_err("enable_cpucache failed for %s, error %d\n",
3868 cachep
->name
, -err
);
3873 * Drain an array if it contains any elements taking the node lock only if
3874 * necessary. Note that the node listlock also protects the array_cache
3875 * if drain_array() is used on the shared array.
3877 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3878 struct array_cache
*ac
, int node
)
3882 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3883 check_mutex_acquired();
3885 if (!ac
|| !ac
->avail
)
3893 spin_lock_irq(&n
->list_lock
);
3894 drain_array_locked(cachep
, ac
, node
, false, &list
);
3895 spin_unlock_irq(&n
->list_lock
);
3897 slabs_destroy(cachep
, &list
);
3901 * cache_reap - Reclaim memory from caches.
3902 * @w: work descriptor
3904 * Called from workqueue/eventd every few seconds.
3906 * - clear the per-cpu caches for this CPU.
3907 * - return freeable pages to the main free memory pool.
3909 * If we cannot acquire the cache chain mutex then just give up - we'll try
3910 * again on the next iteration.
3912 static void cache_reap(struct work_struct
*w
)
3914 struct kmem_cache
*searchp
;
3915 struct kmem_cache_node
*n
;
3916 int node
= numa_mem_id();
3917 struct delayed_work
*work
= to_delayed_work(w
);
3919 if (!mutex_trylock(&slab_mutex
))
3920 /* Give up. Setup the next iteration. */
3923 list_for_each_entry(searchp
, &slab_caches
, list
) {
3927 * We only take the node lock if absolutely necessary and we
3928 * have established with reasonable certainty that
3929 * we can do some work if the lock was obtained.
3931 n
= get_node(searchp
, node
);
3933 reap_alien(searchp
, n
);
3935 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3938 * These are racy checks but it does not matter
3939 * if we skip one check or scan twice.
3941 if (time_after(n
->next_reap
, jiffies
))
3944 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3946 drain_array(searchp
, n
, n
->shared
, node
);
3948 if (n
->free_touched
)
3949 n
->free_touched
= 0;
3953 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3954 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3955 STATS_ADD_REAPED(searchp
, freed
);
3961 mutex_unlock(&slab_mutex
);
3964 /* Set up the next iteration */
3965 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3968 #ifdef CONFIG_SLABINFO
3969 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3972 unsigned long active_objs
;
3973 unsigned long num_objs
;
3974 unsigned long active_slabs
= 0;
3975 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3979 struct kmem_cache_node
*n
;
3983 for_each_kmem_cache_node(cachep
, node
, n
) {
3986 spin_lock_irq(&n
->list_lock
);
3988 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
3989 if (page
->active
!= cachep
->num
&& !error
)
3990 error
= "slabs_full accounting error";
3991 active_objs
+= cachep
->num
;
3994 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
3995 if (page
->active
== cachep
->num
&& !error
)
3996 error
= "slabs_partial accounting error";
3997 if (!page
->active
&& !error
)
3998 error
= "slabs_partial accounting error";
3999 active_objs
+= page
->active
;
4002 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4003 if (page
->active
&& !error
)
4004 error
= "slabs_free accounting error";
4007 free_objects
+= n
->free_objects
;
4009 shared_avail
+= n
->shared
->avail
;
4011 spin_unlock_irq(&n
->list_lock
);
4013 num_slabs
+= active_slabs
;
4014 num_objs
= num_slabs
* cachep
->num
;
4015 if (num_objs
- active_objs
!= free_objects
&& !error
)
4016 error
= "free_objects accounting error";
4018 name
= cachep
->name
;
4020 pr_err("slab: cache %s error: %s\n", name
, error
);
4022 sinfo
->active_objs
= active_objs
;
4023 sinfo
->num_objs
= num_objs
;
4024 sinfo
->active_slabs
= active_slabs
;
4025 sinfo
->num_slabs
= num_slabs
;
4026 sinfo
->shared_avail
= shared_avail
;
4027 sinfo
->limit
= cachep
->limit
;
4028 sinfo
->batchcount
= cachep
->batchcount
;
4029 sinfo
->shared
= cachep
->shared
;
4030 sinfo
->objects_per_slab
= cachep
->num
;
4031 sinfo
->cache_order
= cachep
->gfporder
;
4034 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4038 unsigned long high
= cachep
->high_mark
;
4039 unsigned long allocs
= cachep
->num_allocations
;
4040 unsigned long grown
= cachep
->grown
;
4041 unsigned long reaped
= cachep
->reaped
;
4042 unsigned long errors
= cachep
->errors
;
4043 unsigned long max_freeable
= cachep
->max_freeable
;
4044 unsigned long node_allocs
= cachep
->node_allocs
;
4045 unsigned long node_frees
= cachep
->node_frees
;
4046 unsigned long overflows
= cachep
->node_overflow
;
4048 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4049 allocs
, high
, grown
,
4050 reaped
, errors
, max_freeable
, node_allocs
,
4051 node_frees
, overflows
);
4055 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4056 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4057 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4058 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4060 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4061 allochit
, allocmiss
, freehit
, freemiss
);
4066 #define MAX_SLABINFO_WRITE 128
4068 * slabinfo_write - Tuning for the slab allocator
4070 * @buffer: user buffer
4071 * @count: data length
4074 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4075 size_t count
, loff_t
*ppos
)
4077 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4078 int limit
, batchcount
, shared
, res
;
4079 struct kmem_cache
*cachep
;
4081 if (count
> MAX_SLABINFO_WRITE
)
4083 if (copy_from_user(&kbuf
, buffer
, count
))
4085 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4087 tmp
= strchr(kbuf
, ' ');
4092 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4095 /* Find the cache in the chain of caches. */
4096 mutex_lock(&slab_mutex
);
4098 list_for_each_entry(cachep
, &slab_caches
, list
) {
4099 if (!strcmp(cachep
->name
, kbuf
)) {
4100 if (limit
< 1 || batchcount
< 1 ||
4101 batchcount
> limit
|| shared
< 0) {
4104 res
= do_tune_cpucache(cachep
, limit
,
4111 mutex_unlock(&slab_mutex
);
4117 #ifdef CONFIG_DEBUG_SLAB_LEAK
4119 static inline int add_caller(unsigned long *n
, unsigned long v
)
4129 unsigned long *q
= p
+ 2 * i
;
4143 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4149 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4158 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4161 for (j
= page
->active
; j
< c
->num
; j
++) {
4162 if (get_free_obj(page
, j
) == i
) {
4172 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4173 * mapping is established when actual object allocation and
4174 * we could mistakenly access the unmapped object in the cpu
4177 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4180 if (!add_caller(n
, v
))
4185 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4187 #ifdef CONFIG_KALLSYMS
4188 unsigned long offset
, size
;
4189 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4191 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4192 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4194 seq_printf(m
, " [%s]", modname
);
4198 seq_printf(m
, "%p", (void *)address
);
4201 static int leaks_show(struct seq_file
*m
, void *p
)
4203 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4205 struct kmem_cache_node
*n
;
4207 unsigned long *x
= m
->private;
4211 if (!(cachep
->flags
& SLAB_STORE_USER
))
4213 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4217 * Set store_user_clean and start to grab stored user information
4218 * for all objects on this cache. If some alloc/free requests comes
4219 * during the processing, information would be wrong so restart
4223 set_store_user_clean(cachep
);
4224 drain_cpu_caches(cachep
);
4228 for_each_kmem_cache_node(cachep
, node
, n
) {
4231 spin_lock_irq(&n
->list_lock
);
4233 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4234 handle_slab(x
, cachep
, page
);
4235 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4236 handle_slab(x
, cachep
, page
);
4237 spin_unlock_irq(&n
->list_lock
);
4239 } while (!is_store_user_clean(cachep
));
4241 name
= cachep
->name
;
4243 /* Increase the buffer size */
4244 mutex_unlock(&slab_mutex
);
4245 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4247 /* Too bad, we are really out */
4249 mutex_lock(&slab_mutex
);
4252 *(unsigned long *)m
->private = x
[0] * 2;
4254 mutex_lock(&slab_mutex
);
4255 /* Now make sure this entry will be retried */
4259 for (i
= 0; i
< x
[1]; i
++) {
4260 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4261 show_symbol(m
, x
[2*i
+2]);
4268 static const struct seq_operations slabstats_op
= {
4269 .start
= slab_start
,
4275 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4279 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4283 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4288 static const struct file_operations proc_slabstats_operations
= {
4289 .open
= slabstats_open
,
4291 .llseek
= seq_lseek
,
4292 .release
= seq_release_private
,
4296 static int __init
slab_proc_init(void)
4298 #ifdef CONFIG_DEBUG_SLAB_LEAK
4299 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4303 module_init(slab_proc_init
);
4307 * ksize - get the actual amount of memory allocated for a given object
4308 * @objp: Pointer to the object
4310 * kmalloc may internally round up allocations and return more memory
4311 * than requested. ksize() can be used to determine the actual amount of
4312 * memory allocated. The caller may use this additional memory, even though
4313 * a smaller amount of memory was initially specified with the kmalloc call.
4314 * The caller must guarantee that objp points to a valid object previously
4315 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4316 * must not be freed during the duration of the call.
4318 size_t ksize(const void *objp
)
4323 if (unlikely(objp
== ZERO_SIZE_PTR
))
4326 size
= virt_to_cache(objp
)->object_size
;
4327 /* We assume that ksize callers could use the whole allocated area,
4328 * so we need to unpoison this area.
4330 kasan_krealloc(objp
, size
, GFP_NOWAIT
);
4334 EXPORT_SYMBOL(ksize
);