3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t
;
166 typedef unsigned short freelist_idx_t
;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
179 * The limit is stored in the per-cpu structure to reduce the data cache
186 unsigned int batchcount
;
187 unsigned int touched
;
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
197 struct array_cache ac
;
201 * Need this for bootstrapping a per node allocator.
203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
205 #define CACHE_CACHE 0
206 #define SIZE_NODE (MAX_NUMNODES)
208 static int drain_freelist(struct kmem_cache
*cache
,
209 struct kmem_cache_node
*n
, int tofree
);
210 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
211 int node
, struct list_head
*list
);
212 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
213 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
214 static void cache_reap(struct work_struct
*unused
);
216 static int slab_early_init
= 1;
218 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
220 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
222 INIT_LIST_HEAD(&parent
->slabs_full
);
223 INIT_LIST_HEAD(&parent
->slabs_partial
);
224 INIT_LIST_HEAD(&parent
->slabs_free
);
225 parent
->shared
= NULL
;
226 parent
->alien
= NULL
;
227 parent
->colour_next
= 0;
228 spin_lock_init(&parent
->list_lock
);
229 parent
->free_objects
= 0;
230 parent
->free_touched
= 0;
233 #define MAKE_LIST(cachep, listp, slab, nodeid) \
235 INIT_LIST_HEAD(listp); \
236 list_splice(&get_node(cachep, nodeid)->slab, listp); \
239 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
241 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
242 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
243 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
246 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
247 #define CFLGS_OFF_SLAB (0x80000000UL)
248 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
249 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
251 #define BATCHREFILL_LIMIT 16
253 * Optimization question: fewer reaps means less probability for unnessary
254 * cpucache drain/refill cycles.
256 * OTOH the cpuarrays can contain lots of objects,
257 * which could lock up otherwise freeable slabs.
259 #define REAPTIMEOUT_AC (2*HZ)
260 #define REAPTIMEOUT_NODE (4*HZ)
263 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
264 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
265 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
266 #define STATS_INC_GROWN(x) ((x)->grown++)
267 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
268 #define STATS_SET_HIGH(x) \
270 if ((x)->num_active > (x)->high_mark) \
271 (x)->high_mark = (x)->num_active; \
273 #define STATS_INC_ERR(x) ((x)->errors++)
274 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
275 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
276 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
277 #define STATS_SET_FREEABLE(x, i) \
279 if ((x)->max_freeable < i) \
280 (x)->max_freeable = i; \
282 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
283 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
284 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
285 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
287 #define STATS_INC_ACTIVE(x) do { } while (0)
288 #define STATS_DEC_ACTIVE(x) do { } while (0)
289 #define STATS_INC_ALLOCED(x) do { } while (0)
290 #define STATS_INC_GROWN(x) do { } while (0)
291 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
292 #define STATS_SET_HIGH(x) do { } while (0)
293 #define STATS_INC_ERR(x) do { } while (0)
294 #define STATS_INC_NODEALLOCS(x) do { } while (0)
295 #define STATS_INC_NODEFREES(x) do { } while (0)
296 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
297 #define STATS_SET_FREEABLE(x, i) do { } while (0)
298 #define STATS_INC_ALLOCHIT(x) do { } while (0)
299 #define STATS_INC_ALLOCMISS(x) do { } while (0)
300 #define STATS_INC_FREEHIT(x) do { } while (0)
301 #define STATS_INC_FREEMISS(x) do { } while (0)
307 * memory layout of objects:
309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
310 * the end of an object is aligned with the end of the real
311 * allocation. Catches writes behind the end of the allocation.
312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
314 * cachep->obj_offset: The real object.
315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
316 * cachep->size - 1* BYTES_PER_WORD: last caller address
317 * [BYTES_PER_WORD long]
319 static int obj_offset(struct kmem_cache
*cachep
)
321 return cachep
->obj_offset
;
324 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
326 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
327 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
328 sizeof(unsigned long long));
331 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
333 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
334 if (cachep
->flags
& SLAB_STORE_USER
)
335 return (unsigned long long *)(objp
+ cachep
->size
-
336 sizeof(unsigned long long) -
338 return (unsigned long long *) (objp
+ cachep
->size
-
339 sizeof(unsigned long long));
342 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
344 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
345 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
350 #define obj_offset(x) 0
351 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
352 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
353 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
357 #ifdef CONFIG_DEBUG_SLAB_LEAK
359 static inline bool is_store_user_clean(struct kmem_cache
*cachep
)
361 return atomic_read(&cachep
->store_user_clean
) == 1;
364 static inline void set_store_user_clean(struct kmem_cache
*cachep
)
366 atomic_set(&cachep
->store_user_clean
, 1);
369 static inline void set_store_user_dirty(struct kmem_cache
*cachep
)
371 if (is_store_user_clean(cachep
))
372 atomic_set(&cachep
->store_user_clean
, 0);
376 static inline void set_store_user_dirty(struct kmem_cache
*cachep
) {}
381 * Do not go above this order unless 0 objects fit into the slab or
382 * overridden on the command line.
384 #define SLAB_MAX_ORDER_HI 1
385 #define SLAB_MAX_ORDER_LO 0
386 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
387 static bool slab_max_order_set __initdata
;
389 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
391 struct page
*page
= virt_to_head_page(obj
);
392 return page
->slab_cache
;
395 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
398 return page
->s_mem
+ cache
->size
* idx
;
402 * We want to avoid an expensive divide : (offset / cache->size)
403 * Using the fact that size is a constant for a particular cache,
404 * we can replace (offset / cache->size) by
405 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
407 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
408 const struct page
*page
, void *obj
)
410 u32 offset
= (obj
- page
->s_mem
);
411 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
414 #define BOOT_CPUCACHE_ENTRIES 1
415 /* internal cache of cache description objs */
416 static struct kmem_cache kmem_cache_boot
= {
418 .limit
= BOOT_CPUCACHE_ENTRIES
,
420 .size
= sizeof(struct kmem_cache
),
421 .name
= "kmem_cache",
424 #define BAD_ALIEN_MAGIC 0x01020304ul
426 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
428 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
430 return this_cpu_ptr(cachep
->cpu_cache
);
434 * Calculate the number of objects and left-over bytes for a given buffer size.
436 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
437 unsigned long flags
, size_t *left_over
)
440 size_t slab_size
= PAGE_SIZE
<< gfporder
;
443 * The slab management structure can be either off the slab or
444 * on it. For the latter case, the memory allocated for a
447 * - @buffer_size bytes for each object
448 * - One freelist_idx_t for each object
450 * We don't need to consider alignment of freelist because
451 * freelist will be at the end of slab page. The objects will be
452 * at the correct alignment.
454 * If the slab management structure is off the slab, then the
455 * alignment will already be calculated into the size. Because
456 * the slabs are all pages aligned, the objects will be at the
457 * correct alignment when allocated.
459 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
460 num
= slab_size
/ buffer_size
;
461 *left_over
= slab_size
% buffer_size
;
463 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
464 *left_over
= slab_size
%
465 (buffer_size
+ sizeof(freelist_idx_t
));
472 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
474 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
477 pr_err("slab error in %s(): cache `%s': %s\n",
478 function
, cachep
->name
, msg
);
480 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
485 * By default on NUMA we use alien caches to stage the freeing of
486 * objects allocated from other nodes. This causes massive memory
487 * inefficiencies when using fake NUMA setup to split memory into a
488 * large number of small nodes, so it can be disabled on the command
492 static int use_alien_caches __read_mostly
= 1;
493 static int __init
noaliencache_setup(char *s
)
495 use_alien_caches
= 0;
498 __setup("noaliencache", noaliencache_setup
);
500 static int __init
slab_max_order_setup(char *str
)
502 get_option(&str
, &slab_max_order
);
503 slab_max_order
= slab_max_order
< 0 ? 0 :
504 min(slab_max_order
, MAX_ORDER
- 1);
505 slab_max_order_set
= true;
509 __setup("slab_max_order=", slab_max_order_setup
);
513 * Special reaping functions for NUMA systems called from cache_reap().
514 * These take care of doing round robin flushing of alien caches (containing
515 * objects freed on different nodes from which they were allocated) and the
516 * flushing of remote pcps by calling drain_node_pages.
518 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
520 static void init_reap_node(int cpu
)
524 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
525 if (node
== MAX_NUMNODES
)
526 node
= first_node(node_online_map
);
528 per_cpu(slab_reap_node
, cpu
) = node
;
531 static void next_reap_node(void)
533 int node
= __this_cpu_read(slab_reap_node
);
535 node
= next_node(node
, node_online_map
);
536 if (unlikely(node
>= MAX_NUMNODES
))
537 node
= first_node(node_online_map
);
538 __this_cpu_write(slab_reap_node
, node
);
542 #define init_reap_node(cpu) do { } while (0)
543 #define next_reap_node(void) do { } while (0)
547 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
548 * via the workqueue/eventd.
549 * Add the CPU number into the expiration time to minimize the possibility of
550 * the CPUs getting into lockstep and contending for the global cache chain
553 static void start_cpu_timer(int cpu
)
555 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
558 * When this gets called from do_initcalls via cpucache_init(),
559 * init_workqueues() has already run, so keventd will be setup
562 if (keventd_up() && reap_work
->work
.func
== NULL
) {
564 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
565 schedule_delayed_work_on(cpu
, reap_work
,
566 __round_jiffies_relative(HZ
, cpu
));
570 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
573 * The array_cache structures contain pointers to free object.
574 * However, when such objects are allocated or transferred to another
575 * cache the pointers are not cleared and they could be counted as
576 * valid references during a kmemleak scan. Therefore, kmemleak must
577 * not scan such objects.
579 kmemleak_no_scan(ac
);
583 ac
->batchcount
= batch
;
588 static struct array_cache
*alloc_arraycache(int node
, int entries
,
589 int batchcount
, gfp_t gfp
)
591 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
592 struct array_cache
*ac
= NULL
;
594 ac
= kmalloc_node(memsize
, gfp
, node
);
595 init_arraycache(ac
, entries
, batchcount
);
599 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
600 struct page
*page
, void *objp
)
602 struct kmem_cache_node
*n
;
606 page_node
= page_to_nid(page
);
607 n
= get_node(cachep
, page_node
);
609 spin_lock(&n
->list_lock
);
610 free_block(cachep
, &objp
, 1, page_node
, &list
);
611 spin_unlock(&n
->list_lock
);
613 slabs_destroy(cachep
, &list
);
617 * Transfer objects in one arraycache to another.
618 * Locking must be handled by the caller.
620 * Return the number of entries transferred.
622 static int transfer_objects(struct array_cache
*to
,
623 struct array_cache
*from
, unsigned int max
)
625 /* Figure out how many entries to transfer */
626 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
631 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
641 #define drain_alien_cache(cachep, alien) do { } while (0)
642 #define reap_alien(cachep, n) do { } while (0)
644 static inline struct alien_cache
**alloc_alien_cache(int node
,
645 int limit
, gfp_t gfp
)
647 return (struct alien_cache
**)BAD_ALIEN_MAGIC
;
650 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
654 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
659 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
665 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
666 gfp_t flags
, int nodeid
)
671 static inline gfp_t
gfp_exact_node(gfp_t flags
)
673 return flags
& ~__GFP_NOFAIL
;
676 #else /* CONFIG_NUMA */
678 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
679 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
681 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
682 int batch
, gfp_t gfp
)
684 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
685 struct alien_cache
*alc
= NULL
;
687 alc
= kmalloc_node(memsize
, gfp
, node
);
688 init_arraycache(&alc
->ac
, entries
, batch
);
689 spin_lock_init(&alc
->lock
);
693 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
695 struct alien_cache
**alc_ptr
;
696 size_t memsize
= sizeof(void *) * nr_node_ids
;
701 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
706 if (i
== node
|| !node_online(i
))
708 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
710 for (i
--; i
>= 0; i
--)
719 static void free_alien_cache(struct alien_cache
**alc_ptr
)
730 static void __drain_alien_cache(struct kmem_cache
*cachep
,
731 struct array_cache
*ac
, int node
,
732 struct list_head
*list
)
734 struct kmem_cache_node
*n
= get_node(cachep
, node
);
737 spin_lock(&n
->list_lock
);
739 * Stuff objects into the remote nodes shared array first.
740 * That way we could avoid the overhead of putting the objects
741 * into the free lists and getting them back later.
744 transfer_objects(n
->shared
, ac
, ac
->limit
);
746 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
748 spin_unlock(&n
->list_lock
);
753 * Called from cache_reap() to regularly drain alien caches round robin.
755 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
757 int node
= __this_cpu_read(slab_reap_node
);
760 struct alien_cache
*alc
= n
->alien
[node
];
761 struct array_cache
*ac
;
765 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
768 __drain_alien_cache(cachep
, ac
, node
, &list
);
769 spin_unlock_irq(&alc
->lock
);
770 slabs_destroy(cachep
, &list
);
776 static void drain_alien_cache(struct kmem_cache
*cachep
,
777 struct alien_cache
**alien
)
780 struct alien_cache
*alc
;
781 struct array_cache
*ac
;
784 for_each_online_node(i
) {
790 spin_lock_irqsave(&alc
->lock
, flags
);
791 __drain_alien_cache(cachep
, ac
, i
, &list
);
792 spin_unlock_irqrestore(&alc
->lock
, flags
);
793 slabs_destroy(cachep
, &list
);
798 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
799 int node
, int page_node
)
801 struct kmem_cache_node
*n
;
802 struct alien_cache
*alien
= NULL
;
803 struct array_cache
*ac
;
806 n
= get_node(cachep
, node
);
807 STATS_INC_NODEFREES(cachep
);
808 if (n
->alien
&& n
->alien
[page_node
]) {
809 alien
= n
->alien
[page_node
];
811 spin_lock(&alien
->lock
);
812 if (unlikely(ac
->avail
== ac
->limit
)) {
813 STATS_INC_ACOVERFLOW(cachep
);
814 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
816 ac
->entry
[ac
->avail
++] = objp
;
817 spin_unlock(&alien
->lock
);
818 slabs_destroy(cachep
, &list
);
820 n
= get_node(cachep
, page_node
);
821 spin_lock(&n
->list_lock
);
822 free_block(cachep
, &objp
, 1, page_node
, &list
);
823 spin_unlock(&n
->list_lock
);
824 slabs_destroy(cachep
, &list
);
829 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
831 int page_node
= page_to_nid(virt_to_page(objp
));
832 int node
= numa_mem_id();
834 * Make sure we are not freeing a object from another node to the array
837 if (likely(node
== page_node
))
840 return __cache_free_alien(cachep
, objp
, node
, page_node
);
844 * Construct gfp mask to allocate from a specific node but do not reclaim or
845 * warn about failures.
847 static inline gfp_t
gfp_exact_node(gfp_t flags
)
849 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
854 * Allocates and initializes node for a node on each slab cache, used for
855 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
856 * will be allocated off-node since memory is not yet online for the new node.
857 * When hotplugging memory or a cpu, existing node are not replaced if
860 * Must hold slab_mutex.
862 static int init_cache_node_node(int node
)
864 struct kmem_cache
*cachep
;
865 struct kmem_cache_node
*n
;
866 const size_t memsize
= sizeof(struct kmem_cache_node
);
868 list_for_each_entry(cachep
, &slab_caches
, list
) {
870 * Set up the kmem_cache_node for cpu before we can
871 * begin anything. Make sure some other cpu on this
872 * node has not already allocated this
874 n
= get_node(cachep
, node
);
876 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
879 kmem_cache_node_init(n
);
880 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
881 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
884 * The kmem_cache_nodes don't come and go as CPUs
885 * come and go. slab_mutex is sufficient
888 cachep
->node
[node
] = n
;
891 spin_lock_irq(&n
->list_lock
);
893 (1 + nr_cpus_node(node
)) *
894 cachep
->batchcount
+ cachep
->num
;
895 spin_unlock_irq(&n
->list_lock
);
900 static inline int slabs_tofree(struct kmem_cache
*cachep
,
901 struct kmem_cache_node
*n
)
903 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
906 static void cpuup_canceled(long cpu
)
908 struct kmem_cache
*cachep
;
909 struct kmem_cache_node
*n
= NULL
;
910 int node
= cpu_to_mem(cpu
);
911 const struct cpumask
*mask
= cpumask_of_node(node
);
913 list_for_each_entry(cachep
, &slab_caches
, list
) {
914 struct array_cache
*nc
;
915 struct array_cache
*shared
;
916 struct alien_cache
**alien
;
919 n
= get_node(cachep
, node
);
923 spin_lock_irq(&n
->list_lock
);
925 /* Free limit for this kmem_cache_node */
926 n
->free_limit
-= cachep
->batchcount
;
928 /* cpu is dead; no one can alloc from it. */
929 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
931 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
935 if (!cpumask_empty(mask
)) {
936 spin_unlock_irq(&n
->list_lock
);
942 free_block(cachep
, shared
->entry
,
943 shared
->avail
, node
, &list
);
950 spin_unlock_irq(&n
->list_lock
);
954 drain_alien_cache(cachep
, alien
);
955 free_alien_cache(alien
);
959 slabs_destroy(cachep
, &list
);
962 * In the previous loop, all the objects were freed to
963 * the respective cache's slabs, now we can go ahead and
964 * shrink each nodelist to its limit.
966 list_for_each_entry(cachep
, &slab_caches
, list
) {
967 n
= get_node(cachep
, node
);
970 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
974 static int cpuup_prepare(long cpu
)
976 struct kmem_cache
*cachep
;
977 struct kmem_cache_node
*n
= NULL
;
978 int node
= cpu_to_mem(cpu
);
982 * We need to do this right in the beginning since
983 * alloc_arraycache's are going to use this list.
984 * kmalloc_node allows us to add the slab to the right
985 * kmem_cache_node and not this cpu's kmem_cache_node
987 err
= init_cache_node_node(node
);
992 * Now we can go ahead with allocating the shared arrays and
995 list_for_each_entry(cachep
, &slab_caches
, list
) {
996 struct array_cache
*shared
= NULL
;
997 struct alien_cache
**alien
= NULL
;
999 if (cachep
->shared
) {
1000 shared
= alloc_arraycache(node
,
1001 cachep
->shared
* cachep
->batchcount
,
1002 0xbaadf00d, GFP_KERNEL
);
1006 if (use_alien_caches
) {
1007 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1013 n
= get_node(cachep
, node
);
1016 spin_lock_irq(&n
->list_lock
);
1019 * We are serialised from CPU_DEAD or
1020 * CPU_UP_CANCELLED by the cpucontrol lock
1031 spin_unlock_irq(&n
->list_lock
);
1033 free_alien_cache(alien
);
1038 cpuup_canceled(cpu
);
1042 static int cpuup_callback(struct notifier_block
*nfb
,
1043 unsigned long action
, void *hcpu
)
1045 long cpu
= (long)hcpu
;
1049 case CPU_UP_PREPARE
:
1050 case CPU_UP_PREPARE_FROZEN
:
1051 mutex_lock(&slab_mutex
);
1052 err
= cpuup_prepare(cpu
);
1053 mutex_unlock(&slab_mutex
);
1056 case CPU_ONLINE_FROZEN
:
1057 start_cpu_timer(cpu
);
1059 #ifdef CONFIG_HOTPLUG_CPU
1060 case CPU_DOWN_PREPARE
:
1061 case CPU_DOWN_PREPARE_FROZEN
:
1063 * Shutdown cache reaper. Note that the slab_mutex is
1064 * held so that if cache_reap() is invoked it cannot do
1065 * anything expensive but will only modify reap_work
1066 * and reschedule the timer.
1068 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1069 /* Now the cache_reaper is guaranteed to be not running. */
1070 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1072 case CPU_DOWN_FAILED
:
1073 case CPU_DOWN_FAILED_FROZEN
:
1074 start_cpu_timer(cpu
);
1077 case CPU_DEAD_FROZEN
:
1079 * Even if all the cpus of a node are down, we don't free the
1080 * kmem_cache_node of any cache. This to avoid a race between
1081 * cpu_down, and a kmalloc allocation from another cpu for
1082 * memory from the node of the cpu going down. The node
1083 * structure is usually allocated from kmem_cache_create() and
1084 * gets destroyed at kmem_cache_destroy().
1088 case CPU_UP_CANCELED
:
1089 case CPU_UP_CANCELED_FROZEN
:
1090 mutex_lock(&slab_mutex
);
1091 cpuup_canceled(cpu
);
1092 mutex_unlock(&slab_mutex
);
1095 return notifier_from_errno(err
);
1098 static struct notifier_block cpucache_notifier
= {
1099 &cpuup_callback
, NULL
, 0
1102 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1104 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1105 * Returns -EBUSY if all objects cannot be drained so that the node is not
1108 * Must hold slab_mutex.
1110 static int __meminit
drain_cache_node_node(int node
)
1112 struct kmem_cache
*cachep
;
1115 list_for_each_entry(cachep
, &slab_caches
, list
) {
1116 struct kmem_cache_node
*n
;
1118 n
= get_node(cachep
, node
);
1122 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1124 if (!list_empty(&n
->slabs_full
) ||
1125 !list_empty(&n
->slabs_partial
)) {
1133 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1134 unsigned long action
, void *arg
)
1136 struct memory_notify
*mnb
= arg
;
1140 nid
= mnb
->status_change_nid
;
1145 case MEM_GOING_ONLINE
:
1146 mutex_lock(&slab_mutex
);
1147 ret
= init_cache_node_node(nid
);
1148 mutex_unlock(&slab_mutex
);
1150 case MEM_GOING_OFFLINE
:
1151 mutex_lock(&slab_mutex
);
1152 ret
= drain_cache_node_node(nid
);
1153 mutex_unlock(&slab_mutex
);
1157 case MEM_CANCEL_ONLINE
:
1158 case MEM_CANCEL_OFFLINE
:
1162 return notifier_from_errno(ret
);
1164 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1167 * swap the static kmem_cache_node with kmalloced memory
1169 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1172 struct kmem_cache_node
*ptr
;
1174 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1177 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1179 * Do not assume that spinlocks can be initialized via memcpy:
1181 spin_lock_init(&ptr
->list_lock
);
1183 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1184 cachep
->node
[nodeid
] = ptr
;
1188 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1189 * size of kmem_cache_node.
1191 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1195 for_each_online_node(node
) {
1196 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1197 cachep
->node
[node
]->next_reap
= jiffies
+
1199 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1204 * Initialisation. Called after the page allocator have been initialised and
1205 * before smp_init().
1207 void __init
kmem_cache_init(void)
1211 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1212 sizeof(struct rcu_head
));
1213 kmem_cache
= &kmem_cache_boot
;
1215 if (num_possible_nodes() == 1)
1216 use_alien_caches
= 0;
1218 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1219 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1222 * Fragmentation resistance on low memory - only use bigger
1223 * page orders on machines with more than 32MB of memory if
1224 * not overridden on the command line.
1226 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1227 slab_max_order
= SLAB_MAX_ORDER_HI
;
1229 /* Bootstrap is tricky, because several objects are allocated
1230 * from caches that do not exist yet:
1231 * 1) initialize the kmem_cache cache: it contains the struct
1232 * kmem_cache structures of all caches, except kmem_cache itself:
1233 * kmem_cache is statically allocated.
1234 * Initially an __init data area is used for the head array and the
1235 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1236 * array at the end of the bootstrap.
1237 * 2) Create the first kmalloc cache.
1238 * The struct kmem_cache for the new cache is allocated normally.
1239 * An __init data area is used for the head array.
1240 * 3) Create the remaining kmalloc caches, with minimally sized
1242 * 4) Replace the __init data head arrays for kmem_cache and the first
1243 * kmalloc cache with kmalloc allocated arrays.
1244 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1245 * the other cache's with kmalloc allocated memory.
1246 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1249 /* 1) create the kmem_cache */
1252 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1254 create_boot_cache(kmem_cache
, "kmem_cache",
1255 offsetof(struct kmem_cache
, node
) +
1256 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1257 SLAB_HWCACHE_ALIGN
);
1258 list_add(&kmem_cache
->list
, &slab_caches
);
1259 slab_state
= PARTIAL
;
1262 * Initialize the caches that provide memory for the kmem_cache_node
1263 * structures first. Without this, further allocations will bug.
1265 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1266 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1267 slab_state
= PARTIAL_NODE
;
1268 setup_kmalloc_cache_index_table();
1270 slab_early_init
= 0;
1272 /* 5) Replace the bootstrap kmem_cache_node */
1276 for_each_online_node(nid
) {
1277 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1279 init_list(kmalloc_caches
[INDEX_NODE
],
1280 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1284 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1287 void __init
kmem_cache_init_late(void)
1289 struct kmem_cache
*cachep
;
1293 /* 6) resize the head arrays to their final sizes */
1294 mutex_lock(&slab_mutex
);
1295 list_for_each_entry(cachep
, &slab_caches
, list
)
1296 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1298 mutex_unlock(&slab_mutex
);
1304 * Register a cpu startup notifier callback that initializes
1305 * cpu_cache_get for all new cpus
1307 register_cpu_notifier(&cpucache_notifier
);
1311 * Register a memory hotplug callback that initializes and frees
1314 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1318 * The reap timers are started later, with a module init call: That part
1319 * of the kernel is not yet operational.
1323 static int __init
cpucache_init(void)
1328 * Register the timers that return unneeded pages to the page allocator
1330 for_each_online_cpu(cpu
)
1331 start_cpu_timer(cpu
);
1337 __initcall(cpucache_init
);
1339 static noinline
void
1340 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1343 struct kmem_cache_node
*n
;
1345 unsigned long flags
;
1347 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1348 DEFAULT_RATELIMIT_BURST
);
1350 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1353 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1354 nodeid
, gfpflags
, &gfpflags
);
1355 pr_warn(" cache: %s, object size: %d, order: %d\n",
1356 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1358 for_each_kmem_cache_node(cachep
, node
, n
) {
1359 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1360 unsigned long active_slabs
= 0, num_slabs
= 0;
1362 spin_lock_irqsave(&n
->list_lock
, flags
);
1363 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1364 active_objs
+= cachep
->num
;
1367 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1368 active_objs
+= page
->active
;
1371 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1374 free_objects
+= n
->free_objects
;
1375 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1377 num_slabs
+= active_slabs
;
1378 num_objs
= num_slabs
* cachep
->num
;
1379 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1380 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1387 * Interface to system's page allocator. No need to hold the
1388 * kmem_cache_node ->list_lock.
1390 * If we requested dmaable memory, we will get it. Even if we
1391 * did not request dmaable memory, we might get it, but that
1392 * would be relatively rare and ignorable.
1394 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1400 flags
|= cachep
->allocflags
;
1401 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1402 flags
|= __GFP_RECLAIMABLE
;
1404 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1406 slab_out_of_memory(cachep
, flags
, nodeid
);
1410 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1411 __free_pages(page
, cachep
->gfporder
);
1415 nr_pages
= (1 << cachep
->gfporder
);
1416 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1417 add_zone_page_state(page_zone(page
),
1418 NR_SLAB_RECLAIMABLE
, nr_pages
);
1420 add_zone_page_state(page_zone(page
),
1421 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1423 __SetPageSlab(page
);
1424 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1425 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1426 SetPageSlabPfmemalloc(page
);
1428 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1429 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1432 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1434 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1441 * Interface to system's page release.
1443 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1445 int order
= cachep
->gfporder
;
1446 unsigned long nr_freed
= (1 << order
);
1448 kmemcheck_free_shadow(page
, order
);
1450 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1451 sub_zone_page_state(page_zone(page
),
1452 NR_SLAB_RECLAIMABLE
, nr_freed
);
1454 sub_zone_page_state(page_zone(page
),
1455 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1457 BUG_ON(!PageSlab(page
));
1458 __ClearPageSlabPfmemalloc(page
);
1459 __ClearPageSlab(page
);
1460 page_mapcount_reset(page
);
1461 page
->mapping
= NULL
;
1463 if (current
->reclaim_state
)
1464 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1465 memcg_uncharge_slab(page
, order
, cachep
);
1466 __free_pages(page
, order
);
1469 static void kmem_rcu_free(struct rcu_head
*head
)
1471 struct kmem_cache
*cachep
;
1474 page
= container_of(head
, struct page
, rcu_head
);
1475 cachep
= page
->slab_cache
;
1477 kmem_freepages(cachep
, page
);
1481 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1483 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1484 (cachep
->size
% PAGE_SIZE
) == 0)
1490 #ifdef CONFIG_DEBUG_PAGEALLOC
1491 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1492 unsigned long caller
)
1494 int size
= cachep
->object_size
;
1496 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1498 if (size
< 5 * sizeof(unsigned long))
1501 *addr
++ = 0x12345678;
1503 *addr
++ = smp_processor_id();
1504 size
-= 3 * sizeof(unsigned long);
1506 unsigned long *sptr
= &caller
;
1507 unsigned long svalue
;
1509 while (!kstack_end(sptr
)) {
1511 if (kernel_text_address(svalue
)) {
1513 size
-= sizeof(unsigned long);
1514 if (size
<= sizeof(unsigned long))
1520 *addr
++ = 0x87654321;
1523 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1524 int map
, unsigned long caller
)
1526 if (!is_debug_pagealloc_cache(cachep
))
1530 store_stackinfo(cachep
, objp
, caller
);
1532 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1536 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1537 int map
, unsigned long caller
) {}
1541 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1543 int size
= cachep
->object_size
;
1544 addr
= &((char *)addr
)[obj_offset(cachep
)];
1546 memset(addr
, val
, size
);
1547 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1550 static void dump_line(char *data
, int offset
, int limit
)
1553 unsigned char error
= 0;
1556 pr_err("%03x: ", offset
);
1557 for (i
= 0; i
< limit
; i
++) {
1558 if (data
[offset
+ i
] != POISON_FREE
) {
1559 error
= data
[offset
+ i
];
1563 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1564 &data
[offset
], limit
, 1);
1566 if (bad_count
== 1) {
1567 error
^= POISON_FREE
;
1568 if (!(error
& (error
- 1))) {
1569 pr_err("Single bit error detected. Probably bad RAM.\n");
1571 pr_err("Run memtest86+ or a similar memory test tool.\n");
1573 pr_err("Run a memory test tool.\n");
1582 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1587 if (cachep
->flags
& SLAB_RED_ZONE
) {
1588 pr_err("Redzone: 0x%llx/0x%llx\n",
1589 *dbg_redzone1(cachep
, objp
),
1590 *dbg_redzone2(cachep
, objp
));
1593 if (cachep
->flags
& SLAB_STORE_USER
) {
1594 pr_err("Last user: [<%p>](%pSR)\n",
1595 *dbg_userword(cachep
, objp
),
1596 *dbg_userword(cachep
, objp
));
1598 realobj
= (char *)objp
+ obj_offset(cachep
);
1599 size
= cachep
->object_size
;
1600 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1603 if (i
+ limit
> size
)
1605 dump_line(realobj
, i
, limit
);
1609 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1615 if (is_debug_pagealloc_cache(cachep
))
1618 realobj
= (char *)objp
+ obj_offset(cachep
);
1619 size
= cachep
->object_size
;
1621 for (i
= 0; i
< size
; i
++) {
1622 char exp
= POISON_FREE
;
1625 if (realobj
[i
] != exp
) {
1630 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1631 print_tainted(), cachep
->name
,
1633 print_objinfo(cachep
, objp
, 0);
1635 /* Hexdump the affected line */
1638 if (i
+ limit
> size
)
1640 dump_line(realobj
, i
, limit
);
1643 /* Limit to 5 lines */
1649 /* Print some data about the neighboring objects, if they
1652 struct page
*page
= virt_to_head_page(objp
);
1655 objnr
= obj_to_index(cachep
, page
, objp
);
1657 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1658 realobj
= (char *)objp
+ obj_offset(cachep
);
1659 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1660 print_objinfo(cachep
, objp
, 2);
1662 if (objnr
+ 1 < cachep
->num
) {
1663 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1664 realobj
= (char *)objp
+ obj_offset(cachep
);
1665 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1666 print_objinfo(cachep
, objp
, 2);
1673 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1678 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1679 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1683 for (i
= 0; i
< cachep
->num
; i
++) {
1684 void *objp
= index_to_obj(cachep
, page
, i
);
1686 if (cachep
->flags
& SLAB_POISON
) {
1687 check_poison_obj(cachep
, objp
);
1688 slab_kernel_map(cachep
, objp
, 1, 0);
1690 if (cachep
->flags
& SLAB_RED_ZONE
) {
1691 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1692 slab_error(cachep
, "start of a freed object was overwritten");
1693 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1694 slab_error(cachep
, "end of a freed object was overwritten");
1699 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1706 * slab_destroy - destroy and release all objects in a slab
1707 * @cachep: cache pointer being destroyed
1708 * @page: page pointer being destroyed
1710 * Destroy all the objs in a slab page, and release the mem back to the system.
1711 * Before calling the slab page must have been unlinked from the cache. The
1712 * kmem_cache_node ->list_lock is not held/needed.
1714 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1718 freelist
= page
->freelist
;
1719 slab_destroy_debugcheck(cachep
, page
);
1720 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1721 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1723 kmem_freepages(cachep
, page
);
1726 * From now on, we don't use freelist
1727 * although actual page can be freed in rcu context
1729 if (OFF_SLAB(cachep
))
1730 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1733 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1735 struct page
*page
, *n
;
1737 list_for_each_entry_safe(page
, n
, list
, lru
) {
1738 list_del(&page
->lru
);
1739 slab_destroy(cachep
, page
);
1744 * calculate_slab_order - calculate size (page order) of slabs
1745 * @cachep: pointer to the cache that is being created
1746 * @size: size of objects to be created in this cache.
1747 * @flags: slab allocation flags
1749 * Also calculates the number of objects per slab.
1751 * This could be made much more intelligent. For now, try to avoid using
1752 * high order pages for slabs. When the gfp() functions are more friendly
1753 * towards high-order requests, this should be changed.
1755 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1756 size_t size
, unsigned long flags
)
1758 size_t left_over
= 0;
1761 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1765 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1769 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1770 if (num
> SLAB_OBJ_MAX_NUM
)
1773 if (flags
& CFLGS_OFF_SLAB
) {
1774 struct kmem_cache
*freelist_cache
;
1775 size_t freelist_size
;
1777 freelist_size
= num
* sizeof(freelist_idx_t
);
1778 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1779 if (!freelist_cache
)
1783 * Needed to avoid possible looping condition
1786 if (OFF_SLAB(freelist_cache
))
1789 /* check if off slab has enough benefit */
1790 if (freelist_cache
->size
> cachep
->size
/ 2)
1794 /* Found something acceptable - save it away */
1796 cachep
->gfporder
= gfporder
;
1797 left_over
= remainder
;
1800 * A VFS-reclaimable slab tends to have most allocations
1801 * as GFP_NOFS and we really don't want to have to be allocating
1802 * higher-order pages when we are unable to shrink dcache.
1804 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1808 * Large number of objects is good, but very large slabs are
1809 * currently bad for the gfp()s.
1811 if (gfporder
>= slab_max_order
)
1815 * Acceptable internal fragmentation?
1817 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1823 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1824 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1828 struct array_cache __percpu
*cpu_cache
;
1830 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1831 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1836 for_each_possible_cpu(cpu
) {
1837 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1838 entries
, batchcount
);
1844 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1846 if (slab_state
>= FULL
)
1847 return enable_cpucache(cachep
, gfp
);
1849 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1850 if (!cachep
->cpu_cache
)
1853 if (slab_state
== DOWN
) {
1854 /* Creation of first cache (kmem_cache). */
1855 set_up_node(kmem_cache
, CACHE_CACHE
);
1856 } else if (slab_state
== PARTIAL
) {
1857 /* For kmem_cache_node */
1858 set_up_node(cachep
, SIZE_NODE
);
1862 for_each_online_node(node
) {
1863 cachep
->node
[node
] = kmalloc_node(
1864 sizeof(struct kmem_cache_node
), gfp
, node
);
1865 BUG_ON(!cachep
->node
[node
]);
1866 kmem_cache_node_init(cachep
->node
[node
]);
1870 cachep
->node
[numa_mem_id()]->next_reap
=
1871 jiffies
+ REAPTIMEOUT_NODE
+
1872 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1874 cpu_cache_get(cachep
)->avail
= 0;
1875 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1876 cpu_cache_get(cachep
)->batchcount
= 1;
1877 cpu_cache_get(cachep
)->touched
= 0;
1878 cachep
->batchcount
= 1;
1879 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1883 unsigned long kmem_cache_flags(unsigned long object_size
,
1884 unsigned long flags
, const char *name
,
1885 void (*ctor
)(void *))
1891 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1892 unsigned long flags
, void (*ctor
)(void *))
1894 struct kmem_cache
*cachep
;
1896 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1901 * Adjust the object sizes so that we clear
1902 * the complete object on kzalloc.
1904 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1909 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1910 size_t size
, unsigned long flags
)
1916 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1919 left
= calculate_slab_order(cachep
, size
,
1920 flags
| CFLGS_OBJFREELIST_SLAB
);
1924 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1927 cachep
->colour
= left
/ cachep
->colour_off
;
1932 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1933 size_t size
, unsigned long flags
)
1940 * Always use on-slab management when SLAB_NOLEAKTRACE
1941 * to avoid recursive calls into kmemleak.
1943 if (flags
& SLAB_NOLEAKTRACE
)
1947 * Size is large, assume best to place the slab management obj
1948 * off-slab (should allow better packing of objs).
1950 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1955 * If the slab has been placed off-slab, and we have enough space then
1956 * move it on-slab. This is at the expense of any extra colouring.
1958 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1961 cachep
->colour
= left
/ cachep
->colour_off
;
1966 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1967 size_t size
, unsigned long flags
)
1973 left
= calculate_slab_order(cachep
, size
, flags
);
1977 cachep
->colour
= left
/ cachep
->colour_off
;
1983 * __kmem_cache_create - Create a cache.
1984 * @cachep: cache management descriptor
1985 * @flags: SLAB flags
1987 * Returns a ptr to the cache on success, NULL on failure.
1988 * Cannot be called within a int, but can be interrupted.
1989 * The @ctor is run when new pages are allocated by the cache.
1993 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1994 * to catch references to uninitialised memory.
1996 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1997 * for buffer overruns.
1999 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2000 * cacheline. This can be beneficial if you're counting cycles as closely
2004 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2006 size_t ralign
= BYTES_PER_WORD
;
2009 size_t size
= cachep
->size
;
2014 * Enable redzoning and last user accounting, except for caches with
2015 * large objects, if the increased size would increase the object size
2016 * above the next power of two: caches with object sizes just above a
2017 * power of two have a significant amount of internal fragmentation.
2019 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2020 2 * sizeof(unsigned long long)))
2021 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2022 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2023 flags
|= SLAB_POISON
;
2028 * Check that size is in terms of words. This is needed to avoid
2029 * unaligned accesses for some archs when redzoning is used, and makes
2030 * sure any on-slab bufctl's are also correctly aligned.
2032 if (size
& (BYTES_PER_WORD
- 1)) {
2033 size
+= (BYTES_PER_WORD
- 1);
2034 size
&= ~(BYTES_PER_WORD
- 1);
2037 if (flags
& SLAB_RED_ZONE
) {
2038 ralign
= REDZONE_ALIGN
;
2039 /* If redzoning, ensure that the second redzone is suitably
2040 * aligned, by adjusting the object size accordingly. */
2041 size
+= REDZONE_ALIGN
- 1;
2042 size
&= ~(REDZONE_ALIGN
- 1);
2045 /* 3) caller mandated alignment */
2046 if (ralign
< cachep
->align
) {
2047 ralign
= cachep
->align
;
2049 /* disable debug if necessary */
2050 if (ralign
> __alignof__(unsigned long long))
2051 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2055 cachep
->align
= ralign
;
2056 cachep
->colour_off
= cache_line_size();
2057 /* Offset must be a multiple of the alignment. */
2058 if (cachep
->colour_off
< cachep
->align
)
2059 cachep
->colour_off
= cachep
->align
;
2061 if (slab_is_available())
2069 * Both debugging options require word-alignment which is calculated
2072 if (flags
& SLAB_RED_ZONE
) {
2073 /* add space for red zone words */
2074 cachep
->obj_offset
+= sizeof(unsigned long long);
2075 size
+= 2 * sizeof(unsigned long long);
2077 if (flags
& SLAB_STORE_USER
) {
2078 /* user store requires one word storage behind the end of
2079 * the real object. But if the second red zone needs to be
2080 * aligned to 64 bits, we must allow that much space.
2082 if (flags
& SLAB_RED_ZONE
)
2083 size
+= REDZONE_ALIGN
;
2085 size
+= BYTES_PER_WORD
;
2089 size
= ALIGN(size
, cachep
->align
);
2091 * We should restrict the number of objects in a slab to implement
2092 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2094 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2095 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2099 * To activate debug pagealloc, off-slab management is necessary
2100 * requirement. In early phase of initialization, small sized slab
2101 * doesn't get initialized so it would not be possible. So, we need
2102 * to check size >= 256. It guarantees that all necessary small
2103 * sized slab is initialized in current slab initialization sequence.
2105 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2106 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2107 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2108 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2110 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2111 flags
|= CFLGS_OFF_SLAB
;
2112 cachep
->obj_offset
+= tmp_size
- size
;
2120 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2121 flags
|= CFLGS_OBJFREELIST_SLAB
;
2125 if (set_off_slab_cache(cachep
, size
, flags
)) {
2126 flags
|= CFLGS_OFF_SLAB
;
2130 if (set_on_slab_cache(cachep
, size
, flags
))
2136 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2137 cachep
->flags
= flags
;
2138 cachep
->allocflags
= __GFP_COMP
;
2139 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2140 cachep
->allocflags
|= GFP_DMA
;
2141 cachep
->size
= size
;
2142 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2146 * If we're going to use the generic kernel_map_pages()
2147 * poisoning, then it's going to smash the contents of
2148 * the redzone and userword anyhow, so switch them off.
2150 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2151 (cachep
->flags
& SLAB_POISON
) &&
2152 is_debug_pagealloc_cache(cachep
))
2153 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2156 if (OFF_SLAB(cachep
)) {
2157 cachep
->freelist_cache
=
2158 kmalloc_slab(cachep
->freelist_size
, 0u);
2161 err
= setup_cpu_cache(cachep
, gfp
);
2163 __kmem_cache_release(cachep
);
2171 static void check_irq_off(void)
2173 BUG_ON(!irqs_disabled());
2176 static void check_irq_on(void)
2178 BUG_ON(irqs_disabled());
2181 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2185 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2189 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2193 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2198 #define check_irq_off() do { } while(0)
2199 #define check_irq_on() do { } while(0)
2200 #define check_spinlock_acquired(x) do { } while(0)
2201 #define check_spinlock_acquired_node(x, y) do { } while(0)
2204 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2205 struct array_cache
*ac
,
2206 int force
, int node
);
2208 static void do_drain(void *arg
)
2210 struct kmem_cache
*cachep
= arg
;
2211 struct array_cache
*ac
;
2212 int node
= numa_mem_id();
2213 struct kmem_cache_node
*n
;
2217 ac
= cpu_cache_get(cachep
);
2218 n
= get_node(cachep
, node
);
2219 spin_lock(&n
->list_lock
);
2220 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2221 spin_unlock(&n
->list_lock
);
2222 slabs_destroy(cachep
, &list
);
2226 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2228 struct kmem_cache_node
*n
;
2231 on_each_cpu(do_drain
, cachep
, 1);
2233 for_each_kmem_cache_node(cachep
, node
, n
)
2235 drain_alien_cache(cachep
, n
->alien
);
2237 for_each_kmem_cache_node(cachep
, node
, n
)
2238 drain_array(cachep
, n
, n
->shared
, 1, node
);
2242 * Remove slabs from the list of free slabs.
2243 * Specify the number of slabs to drain in tofree.
2245 * Returns the actual number of slabs released.
2247 static int drain_freelist(struct kmem_cache
*cache
,
2248 struct kmem_cache_node
*n
, int tofree
)
2250 struct list_head
*p
;
2255 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2257 spin_lock_irq(&n
->list_lock
);
2258 p
= n
->slabs_free
.prev
;
2259 if (p
== &n
->slabs_free
) {
2260 spin_unlock_irq(&n
->list_lock
);
2264 page
= list_entry(p
, struct page
, lru
);
2265 list_del(&page
->lru
);
2267 * Safe to drop the lock. The slab is no longer linked
2270 n
->free_objects
-= cache
->num
;
2271 spin_unlock_irq(&n
->list_lock
);
2272 slab_destroy(cache
, page
);
2279 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2283 struct kmem_cache_node
*n
;
2285 drain_cpu_caches(cachep
);
2288 for_each_kmem_cache_node(cachep
, node
, n
) {
2289 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2291 ret
+= !list_empty(&n
->slabs_full
) ||
2292 !list_empty(&n
->slabs_partial
);
2294 return (ret
? 1 : 0);
2297 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2299 return __kmem_cache_shrink(cachep
, false);
2302 void __kmem_cache_release(struct kmem_cache
*cachep
)
2305 struct kmem_cache_node
*n
;
2307 free_percpu(cachep
->cpu_cache
);
2309 /* NUMA: free the node structures */
2310 for_each_kmem_cache_node(cachep
, i
, n
) {
2312 free_alien_cache(n
->alien
);
2314 cachep
->node
[i
] = NULL
;
2319 * Get the memory for a slab management obj.
2321 * For a slab cache when the slab descriptor is off-slab, the
2322 * slab descriptor can't come from the same cache which is being created,
2323 * Because if it is the case, that means we defer the creation of
2324 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2325 * And we eventually call down to __kmem_cache_create(), which
2326 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2327 * This is a "chicken-and-egg" problem.
2329 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2330 * which are all initialized during kmem_cache_init().
2332 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2333 struct page
*page
, int colour_off
,
2334 gfp_t local_flags
, int nodeid
)
2337 void *addr
= page_address(page
);
2339 page
->s_mem
= addr
+ colour_off
;
2342 if (OBJFREELIST_SLAB(cachep
))
2344 else if (OFF_SLAB(cachep
)) {
2345 /* Slab management obj is off-slab. */
2346 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2347 local_flags
, nodeid
);
2351 /* We will use last bytes at the slab for freelist */
2352 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2353 cachep
->freelist_size
;
2359 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2361 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2364 static inline void set_free_obj(struct page
*page
,
2365 unsigned int idx
, freelist_idx_t val
)
2367 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2370 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2375 for (i
= 0; i
< cachep
->num
; i
++) {
2376 void *objp
= index_to_obj(cachep
, page
, i
);
2378 if (cachep
->flags
& SLAB_STORE_USER
)
2379 *dbg_userword(cachep
, objp
) = NULL
;
2381 if (cachep
->flags
& SLAB_RED_ZONE
) {
2382 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2383 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2386 * Constructors are not allowed to allocate memory from the same
2387 * cache which they are a constructor for. Otherwise, deadlock.
2388 * They must also be threaded.
2390 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2391 cachep
->ctor(objp
+ obj_offset(cachep
));
2393 if (cachep
->flags
& SLAB_RED_ZONE
) {
2394 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2395 slab_error(cachep
, "constructor overwrote the end of an object");
2396 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2397 slab_error(cachep
, "constructor overwrote the start of an object");
2399 /* need to poison the objs? */
2400 if (cachep
->flags
& SLAB_POISON
) {
2401 poison_obj(cachep
, objp
, POISON_FREE
);
2402 slab_kernel_map(cachep
, objp
, 0, 0);
2408 static void cache_init_objs(struct kmem_cache
*cachep
,
2413 cache_init_objs_debug(cachep
, page
);
2415 if (OBJFREELIST_SLAB(cachep
)) {
2416 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2420 for (i
= 0; i
< cachep
->num
; i
++) {
2421 /* constructor could break poison info */
2422 if (DEBUG
== 0 && cachep
->ctor
)
2423 cachep
->ctor(index_to_obj(cachep
, page
, i
));
2425 set_free_obj(page
, i
, i
);
2429 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2431 if (CONFIG_ZONE_DMA_FLAG
) {
2432 if (flags
& GFP_DMA
)
2433 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2435 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2439 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2443 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2447 if (cachep
->flags
& SLAB_STORE_USER
)
2448 set_store_user_dirty(cachep
);
2454 static void slab_put_obj(struct kmem_cache
*cachep
,
2455 struct page
*page
, void *objp
)
2457 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2461 /* Verify double free bug */
2462 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2463 if (get_free_obj(page
, i
) == objnr
) {
2464 pr_err("slab: double free detected in cache '%s', objp %p\n",
2465 cachep
->name
, objp
);
2471 if (!page
->freelist
)
2472 page
->freelist
= objp
+ obj_offset(cachep
);
2474 set_free_obj(page
, page
->active
, objnr
);
2478 * Map pages beginning at addr to the given cache and slab. This is required
2479 * for the slab allocator to be able to lookup the cache and slab of a
2480 * virtual address for kfree, ksize, and slab debugging.
2482 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2485 page
->slab_cache
= cache
;
2486 page
->freelist
= freelist
;
2490 * Grow (by 1) the number of slabs within a cache. This is called by
2491 * kmem_cache_alloc() when there are no active objs left in a cache.
2493 static int cache_grow(struct kmem_cache
*cachep
,
2494 gfp_t flags
, int nodeid
, struct page
*page
)
2499 struct kmem_cache_node
*n
;
2502 * Be lazy and only check for valid flags here, keeping it out of the
2503 * critical path in kmem_cache_alloc().
2505 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2506 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2509 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2511 /* Take the node list lock to change the colour_next on this node */
2513 n
= get_node(cachep
, nodeid
);
2514 spin_lock(&n
->list_lock
);
2516 /* Get colour for the slab, and cal the next value. */
2517 offset
= n
->colour_next
;
2519 if (n
->colour_next
>= cachep
->colour
)
2521 spin_unlock(&n
->list_lock
);
2523 offset
*= cachep
->colour_off
;
2525 if (gfpflags_allow_blocking(local_flags
))
2529 * The test for missing atomic flag is performed here, rather than
2530 * the more obvious place, simply to reduce the critical path length
2531 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2532 * will eventually be caught here (where it matters).
2534 kmem_flagcheck(cachep
, flags
);
2537 * Get mem for the objs. Attempt to allocate a physical page from
2541 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2545 /* Get slab management. */
2546 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2547 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2548 if (OFF_SLAB(cachep
) && !freelist
)
2551 slab_map_pages(cachep
, page
, freelist
);
2553 cache_init_objs(cachep
, page
);
2555 if (gfpflags_allow_blocking(local_flags
))
2556 local_irq_disable();
2558 spin_lock(&n
->list_lock
);
2560 /* Make slab active. */
2561 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2562 STATS_INC_GROWN(cachep
);
2563 n
->free_objects
+= cachep
->num
;
2564 spin_unlock(&n
->list_lock
);
2567 kmem_freepages(cachep
, page
);
2569 if (gfpflags_allow_blocking(local_flags
))
2570 local_irq_disable();
2577 * Perform extra freeing checks:
2578 * - detect bad pointers.
2579 * - POISON/RED_ZONE checking
2581 static void kfree_debugcheck(const void *objp
)
2583 if (!virt_addr_valid(objp
)) {
2584 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2585 (unsigned long)objp
);
2590 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2592 unsigned long long redzone1
, redzone2
;
2594 redzone1
= *dbg_redzone1(cache
, obj
);
2595 redzone2
= *dbg_redzone2(cache
, obj
);
2600 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2603 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2604 slab_error(cache
, "double free detected");
2606 slab_error(cache
, "memory outside object was overwritten");
2608 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2609 obj
, redzone1
, redzone2
);
2612 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2613 unsigned long caller
)
2618 BUG_ON(virt_to_cache(objp
) != cachep
);
2620 objp
-= obj_offset(cachep
);
2621 kfree_debugcheck(objp
);
2622 page
= virt_to_head_page(objp
);
2624 if (cachep
->flags
& SLAB_RED_ZONE
) {
2625 verify_redzone_free(cachep
, objp
);
2626 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2627 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2629 if (cachep
->flags
& SLAB_STORE_USER
) {
2630 set_store_user_dirty(cachep
);
2631 *dbg_userword(cachep
, objp
) = (void *)caller
;
2634 objnr
= obj_to_index(cachep
, page
, objp
);
2636 BUG_ON(objnr
>= cachep
->num
);
2637 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2639 if (cachep
->flags
& SLAB_POISON
) {
2640 poison_obj(cachep
, objp
, POISON_FREE
);
2641 slab_kernel_map(cachep
, objp
, 0, caller
);
2647 #define kfree_debugcheck(x) do { } while(0)
2648 #define cache_free_debugcheck(x,objp,z) (objp)
2651 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2659 objp
= next
- obj_offset(cachep
);
2660 next
= *(void **)next
;
2661 poison_obj(cachep
, objp
, POISON_FREE
);
2666 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2667 struct kmem_cache_node
*n
, struct page
*page
,
2670 /* move slabp to correct slabp list: */
2671 list_del(&page
->lru
);
2672 if (page
->active
== cachep
->num
) {
2673 list_add(&page
->lru
, &n
->slabs_full
);
2674 if (OBJFREELIST_SLAB(cachep
)) {
2676 /* Poisoning will be done without holding the lock */
2677 if (cachep
->flags
& SLAB_POISON
) {
2678 void **objp
= page
->freelist
;
2684 page
->freelist
= NULL
;
2687 list_add(&page
->lru
, &n
->slabs_partial
);
2690 /* Try to find non-pfmemalloc slab if needed */
2691 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2692 struct page
*page
, bool pfmemalloc
)
2700 if (!PageSlabPfmemalloc(page
))
2703 /* No need to keep pfmemalloc slab if we have enough free objects */
2704 if (n
->free_objects
> n
->free_limit
) {
2705 ClearPageSlabPfmemalloc(page
);
2709 /* Move pfmemalloc slab to the end of list to speed up next search */
2710 list_del(&page
->lru
);
2712 list_add_tail(&page
->lru
, &n
->slabs_free
);
2714 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2716 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2717 if (!PageSlabPfmemalloc(page
))
2721 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2722 if (!PageSlabPfmemalloc(page
))
2729 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2733 page
= list_first_entry_or_null(&n
->slabs_partial
,
2736 n
->free_touched
= 1;
2737 page
= list_first_entry_or_null(&n
->slabs_free
,
2741 if (sk_memalloc_socks())
2742 return get_valid_first_slab(n
, page
, pfmemalloc
);
2747 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2748 struct kmem_cache_node
*n
, gfp_t flags
)
2754 if (!gfp_pfmemalloc_allowed(flags
))
2757 spin_lock(&n
->list_lock
);
2758 page
= get_first_slab(n
, true);
2760 spin_unlock(&n
->list_lock
);
2764 obj
= slab_get_obj(cachep
, page
);
2767 fixup_slab_list(cachep
, n
, page
, &list
);
2769 spin_unlock(&n
->list_lock
);
2770 fixup_objfreelist_debug(cachep
, &list
);
2775 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2778 struct kmem_cache_node
*n
;
2779 struct array_cache
*ac
;
2784 node
= numa_mem_id();
2787 ac
= cpu_cache_get(cachep
);
2788 batchcount
= ac
->batchcount
;
2789 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2791 * If there was little recent activity on this cache, then
2792 * perform only a partial refill. Otherwise we could generate
2795 batchcount
= BATCHREFILL_LIMIT
;
2797 n
= get_node(cachep
, node
);
2799 BUG_ON(ac
->avail
> 0 || !n
);
2800 spin_lock(&n
->list_lock
);
2802 /* See if we can refill from the shared array */
2803 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2804 n
->shared
->touched
= 1;
2808 while (batchcount
> 0) {
2810 /* Get slab alloc is to come from. */
2811 page
= get_first_slab(n
, false);
2815 check_spinlock_acquired(cachep
);
2818 * The slab was either on partial or free list so
2819 * there must be at least one object available for
2822 BUG_ON(page
->active
>= cachep
->num
);
2824 while (page
->active
< cachep
->num
&& batchcount
--) {
2825 STATS_INC_ALLOCED(cachep
);
2826 STATS_INC_ACTIVE(cachep
);
2827 STATS_SET_HIGH(cachep
);
2829 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2832 fixup_slab_list(cachep
, n
, page
, &list
);
2836 n
->free_objects
-= ac
->avail
;
2838 spin_unlock(&n
->list_lock
);
2839 fixup_objfreelist_debug(cachep
, &list
);
2841 if (unlikely(!ac
->avail
)) {
2844 /* Check if we can use obj in pfmemalloc slab */
2845 if (sk_memalloc_socks()) {
2846 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2852 x
= cache_grow(cachep
, gfp_exact_node(flags
), node
, NULL
);
2854 /* cache_grow can reenable interrupts, then ac could change. */
2855 ac
= cpu_cache_get(cachep
);
2856 node
= numa_mem_id();
2858 /* no objects in sight? abort */
2859 if (!x
&& ac
->avail
== 0)
2862 if (!ac
->avail
) /* objects refilled by interrupt? */
2867 return ac
->entry
[--ac
->avail
];
2870 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2873 might_sleep_if(gfpflags_allow_blocking(flags
));
2875 kmem_flagcheck(cachep
, flags
);
2880 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2881 gfp_t flags
, void *objp
, unsigned long caller
)
2885 if (cachep
->flags
& SLAB_POISON
) {
2886 check_poison_obj(cachep
, objp
);
2887 slab_kernel_map(cachep
, objp
, 1, 0);
2888 poison_obj(cachep
, objp
, POISON_INUSE
);
2890 if (cachep
->flags
& SLAB_STORE_USER
)
2891 *dbg_userword(cachep
, objp
) = (void *)caller
;
2893 if (cachep
->flags
& SLAB_RED_ZONE
) {
2894 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2895 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2896 slab_error(cachep
, "double free, or memory outside object was overwritten");
2897 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2898 objp
, *dbg_redzone1(cachep
, objp
),
2899 *dbg_redzone2(cachep
, objp
));
2901 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2902 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2905 objp
+= obj_offset(cachep
);
2906 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2908 if (ARCH_SLAB_MINALIGN
&&
2909 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
2910 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2911 objp
, (int)ARCH_SLAB_MINALIGN
);
2916 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2919 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2922 struct array_cache
*ac
;
2926 ac
= cpu_cache_get(cachep
);
2927 if (likely(ac
->avail
)) {
2929 objp
= ac
->entry
[--ac
->avail
];
2931 STATS_INC_ALLOCHIT(cachep
);
2935 STATS_INC_ALLOCMISS(cachep
);
2936 objp
= cache_alloc_refill(cachep
, flags
);
2938 * the 'ac' may be updated by cache_alloc_refill(),
2939 * and kmemleak_erase() requires its correct value.
2941 ac
= cpu_cache_get(cachep
);
2945 * To avoid a false negative, if an object that is in one of the
2946 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2947 * treat the array pointers as a reference to the object.
2950 kmemleak_erase(&ac
->entry
[ac
->avail
]);
2956 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2958 * If we are in_interrupt, then process context, including cpusets and
2959 * mempolicy, may not apply and should not be used for allocation policy.
2961 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2963 int nid_alloc
, nid_here
;
2965 if (in_interrupt() || (flags
& __GFP_THISNODE
))
2967 nid_alloc
= nid_here
= numa_mem_id();
2968 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
2969 nid_alloc
= cpuset_slab_spread_node();
2970 else if (current
->mempolicy
)
2971 nid_alloc
= mempolicy_slab_node();
2972 if (nid_alloc
!= nid_here
)
2973 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
2978 * Fallback function if there was no memory available and no objects on a
2979 * certain node and fall back is permitted. First we scan all the
2980 * available node for available objects. If that fails then we
2981 * perform an allocation without specifying a node. This allows the page
2982 * allocator to do its reclaim / fallback magic. We then insert the
2983 * slab into the proper nodelist and then allocate from it.
2985 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
2987 struct zonelist
*zonelist
;
2991 enum zone_type high_zoneidx
= gfp_zone(flags
);
2994 unsigned int cpuset_mems_cookie
;
2996 if (flags
& __GFP_THISNODE
)
2999 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3002 cpuset_mems_cookie
= read_mems_allowed_begin();
3003 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3007 * Look through allowed nodes for objects available
3008 * from existing per node queues.
3010 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3011 nid
= zone_to_nid(zone
);
3013 if (cpuset_zone_allowed(zone
, flags
) &&
3014 get_node(cache
, nid
) &&
3015 get_node(cache
, nid
)->free_objects
) {
3016 obj
= ____cache_alloc_node(cache
,
3017 gfp_exact_node(flags
), nid
);
3025 * This allocation will be performed within the constraints
3026 * of the current cpuset / memory policy requirements.
3027 * We may trigger various forms of reclaim on the allowed
3028 * set and go into memory reserves if necessary.
3032 if (gfpflags_allow_blocking(local_flags
))
3034 kmem_flagcheck(cache
, flags
);
3035 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3036 if (gfpflags_allow_blocking(local_flags
))
3037 local_irq_disable();
3040 * Insert into the appropriate per node queues
3042 nid
= page_to_nid(page
);
3043 if (cache_grow(cache
, flags
, nid
, page
)) {
3044 obj
= ____cache_alloc_node(cache
,
3045 gfp_exact_node(flags
), nid
);
3048 * Another processor may allocate the
3049 * objects in the slab since we are
3050 * not holding any locks.
3054 /* cache_grow already freed obj */
3060 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3066 * A interface to enable slab creation on nodeid
3068 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3072 struct kmem_cache_node
*n
;
3077 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3078 n
= get_node(cachep
, nodeid
);
3083 spin_lock(&n
->list_lock
);
3084 page
= get_first_slab(n
, false);
3088 check_spinlock_acquired_node(cachep
, nodeid
);
3090 STATS_INC_NODEALLOCS(cachep
);
3091 STATS_INC_ACTIVE(cachep
);
3092 STATS_SET_HIGH(cachep
);
3094 BUG_ON(page
->active
== cachep
->num
);
3096 obj
= slab_get_obj(cachep
, page
);
3099 fixup_slab_list(cachep
, n
, page
, &list
);
3101 spin_unlock(&n
->list_lock
);
3102 fixup_objfreelist_debug(cachep
, &list
);
3106 spin_unlock(&n
->list_lock
);
3107 x
= cache_grow(cachep
, gfp_exact_node(flags
), nodeid
, NULL
);
3111 return fallback_alloc(cachep
, flags
);
3117 static __always_inline
void *
3118 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3119 unsigned long caller
)
3121 unsigned long save_flags
;
3123 int slab_node
= numa_mem_id();
3125 flags
&= gfp_allowed_mask
;
3126 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3127 if (unlikely(!cachep
))
3130 cache_alloc_debugcheck_before(cachep
, flags
);
3131 local_irq_save(save_flags
);
3133 if (nodeid
== NUMA_NO_NODE
)
3136 if (unlikely(!get_node(cachep
, nodeid
))) {
3137 /* Node not bootstrapped yet */
3138 ptr
= fallback_alloc(cachep
, flags
);
3142 if (nodeid
== slab_node
) {
3144 * Use the locally cached objects if possible.
3145 * However ____cache_alloc does not allow fallback
3146 * to other nodes. It may fail while we still have
3147 * objects on other nodes available.
3149 ptr
= ____cache_alloc(cachep
, flags
);
3153 /* ___cache_alloc_node can fall back to other nodes */
3154 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3156 local_irq_restore(save_flags
);
3157 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3159 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3160 memset(ptr
, 0, cachep
->object_size
);
3162 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3166 static __always_inline
void *
3167 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3171 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3172 objp
= alternate_node_alloc(cache
, flags
);
3176 objp
= ____cache_alloc(cache
, flags
);
3179 * We may just have run out of memory on the local node.
3180 * ____cache_alloc_node() knows how to locate memory on other nodes
3183 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3190 static __always_inline
void *
3191 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3193 return ____cache_alloc(cachep
, flags
);
3196 #endif /* CONFIG_NUMA */
3198 static __always_inline
void *
3199 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3201 unsigned long save_flags
;
3204 flags
&= gfp_allowed_mask
;
3205 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3206 if (unlikely(!cachep
))
3209 cache_alloc_debugcheck_before(cachep
, flags
);
3210 local_irq_save(save_flags
);
3211 objp
= __do_cache_alloc(cachep
, flags
);
3212 local_irq_restore(save_flags
);
3213 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3216 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3217 memset(objp
, 0, cachep
->object_size
);
3219 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3224 * Caller needs to acquire correct kmem_cache_node's list_lock
3225 * @list: List of detached free slabs should be freed by caller
3227 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3228 int nr_objects
, int node
, struct list_head
*list
)
3231 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3233 for (i
= 0; i
< nr_objects
; i
++) {
3239 page
= virt_to_head_page(objp
);
3240 list_del(&page
->lru
);
3241 check_spinlock_acquired_node(cachep
, node
);
3242 slab_put_obj(cachep
, page
, objp
);
3243 STATS_DEC_ACTIVE(cachep
);
3246 /* fixup slab chains */
3247 if (page
->active
== 0) {
3248 if (n
->free_objects
> n
->free_limit
) {
3249 n
->free_objects
-= cachep
->num
;
3250 list_add_tail(&page
->lru
, list
);
3252 list_add(&page
->lru
, &n
->slabs_free
);
3255 /* Unconditionally move a slab to the end of the
3256 * partial list on free - maximum time for the
3257 * other objects to be freed, too.
3259 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3264 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3267 struct kmem_cache_node
*n
;
3268 int node
= numa_mem_id();
3271 batchcount
= ac
->batchcount
;
3274 n
= get_node(cachep
, node
);
3275 spin_lock(&n
->list_lock
);
3277 struct array_cache
*shared_array
= n
->shared
;
3278 int max
= shared_array
->limit
- shared_array
->avail
;
3280 if (batchcount
> max
)
3282 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3283 ac
->entry
, sizeof(void *) * batchcount
);
3284 shared_array
->avail
+= batchcount
;
3289 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3296 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3297 BUG_ON(page
->active
);
3301 STATS_SET_FREEABLE(cachep
, i
);
3304 spin_unlock(&n
->list_lock
);
3305 slabs_destroy(cachep
, &list
);
3306 ac
->avail
-= batchcount
;
3307 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3311 * Release an obj back to its cache. If the obj has a constructed state, it must
3312 * be in this state _before_ it is released. Called with disabled ints.
3314 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3315 unsigned long caller
)
3317 struct array_cache
*ac
= cpu_cache_get(cachep
);
3320 kmemleak_free_recursive(objp
, cachep
->flags
);
3321 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3323 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3326 * Skip calling cache_free_alien() when the platform is not numa.
3327 * This will avoid cache misses that happen while accessing slabp (which
3328 * is per page memory reference) to get nodeid. Instead use a global
3329 * variable to skip the call, which is mostly likely to be present in
3332 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3335 if (ac
->avail
< ac
->limit
) {
3336 STATS_INC_FREEHIT(cachep
);
3338 STATS_INC_FREEMISS(cachep
);
3339 cache_flusharray(cachep
, ac
);
3342 if (sk_memalloc_socks()) {
3343 struct page
*page
= virt_to_head_page(objp
);
3345 if (unlikely(PageSlabPfmemalloc(page
))) {
3346 cache_free_pfmemalloc(cachep
, page
, objp
);
3351 ac
->entry
[ac
->avail
++] = objp
;
3355 * kmem_cache_alloc - Allocate an object
3356 * @cachep: The cache to allocate from.
3357 * @flags: See kmalloc().
3359 * Allocate an object from this cache. The flags are only relevant
3360 * if the cache has no available objects.
3362 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3364 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3366 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3367 cachep
->object_size
, cachep
->size
, flags
);
3371 EXPORT_SYMBOL(kmem_cache_alloc
);
3373 static __always_inline
void
3374 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3375 size_t size
, void **p
, unsigned long caller
)
3379 for (i
= 0; i
< size
; i
++)
3380 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3383 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3388 s
= slab_pre_alloc_hook(s
, flags
);
3392 cache_alloc_debugcheck_before(s
, flags
);
3394 local_irq_disable();
3395 for (i
= 0; i
< size
; i
++) {
3396 void *objp
= __do_cache_alloc(s
, flags
);
3398 if (unlikely(!objp
))
3404 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3406 /* Clear memory outside IRQ disabled section */
3407 if (unlikely(flags
& __GFP_ZERO
))
3408 for (i
= 0; i
< size
; i
++)
3409 memset(p
[i
], 0, s
->object_size
);
3411 slab_post_alloc_hook(s
, flags
, size
, p
);
3412 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3416 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3417 slab_post_alloc_hook(s
, flags
, i
, p
);
3418 __kmem_cache_free_bulk(s
, i
, p
);
3421 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3423 #ifdef CONFIG_TRACING
3425 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3429 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3431 trace_kmalloc(_RET_IP_
, ret
,
3432 size
, cachep
->size
, flags
);
3435 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3440 * kmem_cache_alloc_node - Allocate an object on the specified node
3441 * @cachep: The cache to allocate from.
3442 * @flags: See kmalloc().
3443 * @nodeid: node number of the target node.
3445 * Identical to kmem_cache_alloc but it will allocate memory on the given
3446 * node, which can improve the performance for cpu bound structures.
3448 * Fallback to other node is possible if __GFP_THISNODE is not set.
3450 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3452 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3454 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3455 cachep
->object_size
, cachep
->size
,
3460 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3462 #ifdef CONFIG_TRACING
3463 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3470 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3472 trace_kmalloc_node(_RET_IP_
, ret
,
3477 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3480 static __always_inline
void *
3481 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3483 struct kmem_cache
*cachep
;
3485 cachep
= kmalloc_slab(size
, flags
);
3486 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3488 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3491 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3493 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3495 EXPORT_SYMBOL(__kmalloc_node
);
3497 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3498 int node
, unsigned long caller
)
3500 return __do_kmalloc_node(size
, flags
, node
, caller
);
3502 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3503 #endif /* CONFIG_NUMA */
3506 * __do_kmalloc - allocate memory
3507 * @size: how many bytes of memory are required.
3508 * @flags: the type of memory to allocate (see kmalloc).
3509 * @caller: function caller for debug tracking of the caller
3511 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3512 unsigned long caller
)
3514 struct kmem_cache
*cachep
;
3517 cachep
= kmalloc_slab(size
, flags
);
3518 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3520 ret
= slab_alloc(cachep
, flags
, caller
);
3522 trace_kmalloc(caller
, ret
,
3523 size
, cachep
->size
, flags
);
3528 void *__kmalloc(size_t size
, gfp_t flags
)
3530 return __do_kmalloc(size
, flags
, _RET_IP_
);
3532 EXPORT_SYMBOL(__kmalloc
);
3534 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3536 return __do_kmalloc(size
, flags
, caller
);
3538 EXPORT_SYMBOL(__kmalloc_track_caller
);
3541 * kmem_cache_free - Deallocate an object
3542 * @cachep: The cache the allocation was from.
3543 * @objp: The previously allocated object.
3545 * Free an object which was previously allocated from this
3548 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3550 unsigned long flags
;
3551 cachep
= cache_from_obj(cachep
, objp
);
3555 local_irq_save(flags
);
3556 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3557 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3558 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3559 __cache_free(cachep
, objp
, _RET_IP_
);
3560 local_irq_restore(flags
);
3562 trace_kmem_cache_free(_RET_IP_
, objp
);
3564 EXPORT_SYMBOL(kmem_cache_free
);
3566 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3568 struct kmem_cache
*s
;
3571 local_irq_disable();
3572 for (i
= 0; i
< size
; i
++) {
3575 if (!orig_s
) /* called via kfree_bulk */
3576 s
= virt_to_cache(objp
);
3578 s
= cache_from_obj(orig_s
, objp
);
3580 debug_check_no_locks_freed(objp
, s
->object_size
);
3581 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3582 debug_check_no_obj_freed(objp
, s
->object_size
);
3584 __cache_free(s
, objp
, _RET_IP_
);
3588 /* FIXME: add tracing */
3590 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3593 * kfree - free previously allocated memory
3594 * @objp: pointer returned by kmalloc.
3596 * If @objp is NULL, no operation is performed.
3598 * Don't free memory not originally allocated by kmalloc()
3599 * or you will run into trouble.
3601 void kfree(const void *objp
)
3603 struct kmem_cache
*c
;
3604 unsigned long flags
;
3606 trace_kfree(_RET_IP_
, objp
);
3608 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3610 local_irq_save(flags
);
3611 kfree_debugcheck(objp
);
3612 c
= virt_to_cache(objp
);
3613 debug_check_no_locks_freed(objp
, c
->object_size
);
3615 debug_check_no_obj_freed(objp
, c
->object_size
);
3616 __cache_free(c
, (void *)objp
, _RET_IP_
);
3617 local_irq_restore(flags
);
3619 EXPORT_SYMBOL(kfree
);
3622 * This initializes kmem_cache_node or resizes various caches for all nodes.
3624 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3627 struct kmem_cache_node
*n
;
3628 struct array_cache
*new_shared
;
3629 struct alien_cache
**new_alien
= NULL
;
3631 for_each_online_node(node
) {
3633 if (use_alien_caches
) {
3634 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3640 if (cachep
->shared
) {
3641 new_shared
= alloc_arraycache(node
,
3642 cachep
->shared
*cachep
->batchcount
,
3645 free_alien_cache(new_alien
);
3650 n
= get_node(cachep
, node
);
3652 struct array_cache
*shared
= n
->shared
;
3655 spin_lock_irq(&n
->list_lock
);
3658 free_block(cachep
, shared
->entry
,
3659 shared
->avail
, node
, &list
);
3661 n
->shared
= new_shared
;
3663 n
->alien
= new_alien
;
3666 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3667 cachep
->batchcount
+ cachep
->num
;
3668 spin_unlock_irq(&n
->list_lock
);
3669 slabs_destroy(cachep
, &list
);
3671 free_alien_cache(new_alien
);
3674 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3676 free_alien_cache(new_alien
);
3681 kmem_cache_node_init(n
);
3682 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3683 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3684 n
->shared
= new_shared
;
3685 n
->alien
= new_alien
;
3686 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3687 cachep
->batchcount
+ cachep
->num
;
3688 cachep
->node
[node
] = n
;
3693 if (!cachep
->list
.next
) {
3694 /* Cache is not active yet. Roll back what we did */
3697 n
= get_node(cachep
, node
);
3700 free_alien_cache(n
->alien
);
3702 cachep
->node
[node
] = NULL
;
3710 /* Always called with the slab_mutex held */
3711 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3712 int batchcount
, int shared
, gfp_t gfp
)
3714 struct array_cache __percpu
*cpu_cache
, *prev
;
3717 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3721 prev
= cachep
->cpu_cache
;
3722 cachep
->cpu_cache
= cpu_cache
;
3723 kick_all_cpus_sync();
3726 cachep
->batchcount
= batchcount
;
3727 cachep
->limit
= limit
;
3728 cachep
->shared
= shared
;
3733 for_each_online_cpu(cpu
) {
3736 struct kmem_cache_node
*n
;
3737 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3739 node
= cpu_to_mem(cpu
);
3740 n
= get_node(cachep
, node
);
3741 spin_lock_irq(&n
->list_lock
);
3742 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3743 spin_unlock_irq(&n
->list_lock
);
3744 slabs_destroy(cachep
, &list
);
3749 return alloc_kmem_cache_node(cachep
, gfp
);
3752 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3753 int batchcount
, int shared
, gfp_t gfp
)
3756 struct kmem_cache
*c
;
3758 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3760 if (slab_state
< FULL
)
3763 if ((ret
< 0) || !is_root_cache(cachep
))
3766 lockdep_assert_held(&slab_mutex
);
3767 for_each_memcg_cache(c
, cachep
) {
3768 /* return value determined by the root cache only */
3769 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3775 /* Called with slab_mutex held always */
3776 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3783 if (!is_root_cache(cachep
)) {
3784 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3785 limit
= root
->limit
;
3786 shared
= root
->shared
;
3787 batchcount
= root
->batchcount
;
3790 if (limit
&& shared
&& batchcount
)
3793 * The head array serves three purposes:
3794 * - create a LIFO ordering, i.e. return objects that are cache-warm
3795 * - reduce the number of spinlock operations.
3796 * - reduce the number of linked list operations on the slab and
3797 * bufctl chains: array operations are cheaper.
3798 * The numbers are guessed, we should auto-tune as described by
3801 if (cachep
->size
> 131072)
3803 else if (cachep
->size
> PAGE_SIZE
)
3805 else if (cachep
->size
> 1024)
3807 else if (cachep
->size
> 256)
3813 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3814 * allocation behaviour: Most allocs on one cpu, most free operations
3815 * on another cpu. For these cases, an efficient object passing between
3816 * cpus is necessary. This is provided by a shared array. The array
3817 * replaces Bonwick's magazine layer.
3818 * On uniprocessor, it's functionally equivalent (but less efficient)
3819 * to a larger limit. Thus disabled by default.
3822 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3827 * With debugging enabled, large batchcount lead to excessively long
3828 * periods with disabled local interrupts. Limit the batchcount
3833 batchcount
= (limit
+ 1) / 2;
3835 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3837 pr_err("enable_cpucache failed for %s, error %d\n",
3838 cachep
->name
, -err
);
3843 * Drain an array if it contains any elements taking the node lock only if
3844 * necessary. Note that the node listlock also protects the array_cache
3845 * if drain_array() is used on the shared array.
3847 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3848 struct array_cache
*ac
, int force
, int node
)
3853 if (!ac
|| !ac
->avail
)
3855 if (ac
->touched
&& !force
) {
3858 spin_lock_irq(&n
->list_lock
);
3860 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3861 if (tofree
> ac
->avail
)
3862 tofree
= (ac
->avail
+ 1) / 2;
3863 free_block(cachep
, ac
->entry
, tofree
, node
, &list
);
3864 ac
->avail
-= tofree
;
3865 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3866 sizeof(void *) * ac
->avail
);
3868 spin_unlock_irq(&n
->list_lock
);
3869 slabs_destroy(cachep
, &list
);
3874 * cache_reap - Reclaim memory from caches.
3875 * @w: work descriptor
3877 * Called from workqueue/eventd every few seconds.
3879 * - clear the per-cpu caches for this CPU.
3880 * - return freeable pages to the main free memory pool.
3882 * If we cannot acquire the cache chain mutex then just give up - we'll try
3883 * again on the next iteration.
3885 static void cache_reap(struct work_struct
*w
)
3887 struct kmem_cache
*searchp
;
3888 struct kmem_cache_node
*n
;
3889 int node
= numa_mem_id();
3890 struct delayed_work
*work
= to_delayed_work(w
);
3892 if (!mutex_trylock(&slab_mutex
))
3893 /* Give up. Setup the next iteration. */
3896 list_for_each_entry(searchp
, &slab_caches
, list
) {
3900 * We only take the node lock if absolutely necessary and we
3901 * have established with reasonable certainty that
3902 * we can do some work if the lock was obtained.
3904 n
= get_node(searchp
, node
);
3906 reap_alien(searchp
, n
);
3908 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
3911 * These are racy checks but it does not matter
3912 * if we skip one check or scan twice.
3914 if (time_after(n
->next_reap
, jiffies
))
3917 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3919 drain_array(searchp
, n
, n
->shared
, 0, node
);
3921 if (n
->free_touched
)
3922 n
->free_touched
= 0;
3926 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3927 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3928 STATS_ADD_REAPED(searchp
, freed
);
3934 mutex_unlock(&slab_mutex
);
3937 /* Set up the next iteration */
3938 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3941 #ifdef CONFIG_SLABINFO
3942 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3945 unsigned long active_objs
;
3946 unsigned long num_objs
;
3947 unsigned long active_slabs
= 0;
3948 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3952 struct kmem_cache_node
*n
;
3956 for_each_kmem_cache_node(cachep
, node
, n
) {
3959 spin_lock_irq(&n
->list_lock
);
3961 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
3962 if (page
->active
!= cachep
->num
&& !error
)
3963 error
= "slabs_full accounting error";
3964 active_objs
+= cachep
->num
;
3967 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
3968 if (page
->active
== cachep
->num
&& !error
)
3969 error
= "slabs_partial accounting error";
3970 if (!page
->active
&& !error
)
3971 error
= "slabs_partial accounting error";
3972 active_objs
+= page
->active
;
3975 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3976 if (page
->active
&& !error
)
3977 error
= "slabs_free accounting error";
3980 free_objects
+= n
->free_objects
;
3982 shared_avail
+= n
->shared
->avail
;
3984 spin_unlock_irq(&n
->list_lock
);
3986 num_slabs
+= active_slabs
;
3987 num_objs
= num_slabs
* cachep
->num
;
3988 if (num_objs
- active_objs
!= free_objects
&& !error
)
3989 error
= "free_objects accounting error";
3991 name
= cachep
->name
;
3993 pr_err("slab: cache %s error: %s\n", name
, error
);
3995 sinfo
->active_objs
= active_objs
;
3996 sinfo
->num_objs
= num_objs
;
3997 sinfo
->active_slabs
= active_slabs
;
3998 sinfo
->num_slabs
= num_slabs
;
3999 sinfo
->shared_avail
= shared_avail
;
4000 sinfo
->limit
= cachep
->limit
;
4001 sinfo
->batchcount
= cachep
->batchcount
;
4002 sinfo
->shared
= cachep
->shared
;
4003 sinfo
->objects_per_slab
= cachep
->num
;
4004 sinfo
->cache_order
= cachep
->gfporder
;
4007 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4011 unsigned long high
= cachep
->high_mark
;
4012 unsigned long allocs
= cachep
->num_allocations
;
4013 unsigned long grown
= cachep
->grown
;
4014 unsigned long reaped
= cachep
->reaped
;
4015 unsigned long errors
= cachep
->errors
;
4016 unsigned long max_freeable
= cachep
->max_freeable
;
4017 unsigned long node_allocs
= cachep
->node_allocs
;
4018 unsigned long node_frees
= cachep
->node_frees
;
4019 unsigned long overflows
= cachep
->node_overflow
;
4021 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4022 allocs
, high
, grown
,
4023 reaped
, errors
, max_freeable
, node_allocs
,
4024 node_frees
, overflows
);
4028 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4029 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4030 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4031 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4033 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4034 allochit
, allocmiss
, freehit
, freemiss
);
4039 #define MAX_SLABINFO_WRITE 128
4041 * slabinfo_write - Tuning for the slab allocator
4043 * @buffer: user buffer
4044 * @count: data length
4047 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4048 size_t count
, loff_t
*ppos
)
4050 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4051 int limit
, batchcount
, shared
, res
;
4052 struct kmem_cache
*cachep
;
4054 if (count
> MAX_SLABINFO_WRITE
)
4056 if (copy_from_user(&kbuf
, buffer
, count
))
4058 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4060 tmp
= strchr(kbuf
, ' ');
4065 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4068 /* Find the cache in the chain of caches. */
4069 mutex_lock(&slab_mutex
);
4071 list_for_each_entry(cachep
, &slab_caches
, list
) {
4072 if (!strcmp(cachep
->name
, kbuf
)) {
4073 if (limit
< 1 || batchcount
< 1 ||
4074 batchcount
> limit
|| shared
< 0) {
4077 res
= do_tune_cpucache(cachep
, limit
,
4084 mutex_unlock(&slab_mutex
);
4090 #ifdef CONFIG_DEBUG_SLAB_LEAK
4092 static inline int add_caller(unsigned long *n
, unsigned long v
)
4102 unsigned long *q
= p
+ 2 * i
;
4116 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4122 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4131 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4134 for (j
= page
->active
; j
< c
->num
; j
++) {
4135 if (get_free_obj(page
, j
) == i
) {
4145 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4146 * mapping is established when actual object allocation and
4147 * we could mistakenly access the unmapped object in the cpu
4150 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4153 if (!add_caller(n
, v
))
4158 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4160 #ifdef CONFIG_KALLSYMS
4161 unsigned long offset
, size
;
4162 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4164 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4165 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4167 seq_printf(m
, " [%s]", modname
);
4171 seq_printf(m
, "%p", (void *)address
);
4174 static int leaks_show(struct seq_file
*m
, void *p
)
4176 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4178 struct kmem_cache_node
*n
;
4180 unsigned long *x
= m
->private;
4184 if (!(cachep
->flags
& SLAB_STORE_USER
))
4186 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4190 * Set store_user_clean and start to grab stored user information
4191 * for all objects on this cache. If some alloc/free requests comes
4192 * during the processing, information would be wrong so restart
4196 set_store_user_clean(cachep
);
4197 drain_cpu_caches(cachep
);
4201 for_each_kmem_cache_node(cachep
, node
, n
) {
4204 spin_lock_irq(&n
->list_lock
);
4206 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4207 handle_slab(x
, cachep
, page
);
4208 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4209 handle_slab(x
, cachep
, page
);
4210 spin_unlock_irq(&n
->list_lock
);
4212 } while (!is_store_user_clean(cachep
));
4214 name
= cachep
->name
;
4216 /* Increase the buffer size */
4217 mutex_unlock(&slab_mutex
);
4218 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4220 /* Too bad, we are really out */
4222 mutex_lock(&slab_mutex
);
4225 *(unsigned long *)m
->private = x
[0] * 2;
4227 mutex_lock(&slab_mutex
);
4228 /* Now make sure this entry will be retried */
4232 for (i
= 0; i
< x
[1]; i
++) {
4233 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4234 show_symbol(m
, x
[2*i
+2]);
4241 static const struct seq_operations slabstats_op
= {
4242 .start
= slab_start
,
4248 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4252 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4256 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4261 static const struct file_operations proc_slabstats_operations
= {
4262 .open
= slabstats_open
,
4264 .llseek
= seq_lseek
,
4265 .release
= seq_release_private
,
4269 static int __init
slab_proc_init(void)
4271 #ifdef CONFIG_DEBUG_SLAB_LEAK
4272 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4276 module_init(slab_proc_init
);
4280 * ksize - get the actual amount of memory allocated for a given object
4281 * @objp: Pointer to the object
4283 * kmalloc may internally round up allocations and return more memory
4284 * than requested. ksize() can be used to determine the actual amount of
4285 * memory allocated. The caller may use this additional memory, even though
4286 * a smaller amount of memory was initially specified with the kmalloc call.
4287 * The caller must guarantee that objp points to a valid object previously
4288 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4289 * must not be freed during the duration of the call.
4291 size_t ksize(const void *objp
)
4294 if (unlikely(objp
== ZERO_SIZE_PTR
))
4297 return virt_to_cache(objp
)->object_size
;
4299 EXPORT_SYMBOL(ksize
);