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
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly
;
170 * - LIFO ordering, to hand out cache-warm objects from _alloc
171 * - reduce the number of linked list operations
172 * - reduce spinlock operations
174 * The limit is stored in the per-cpu structure to reduce the data cache
181 unsigned int batchcount
;
182 unsigned int touched
;
185 * Must have this definition in here for the proper
186 * alignment of array_cache. Also simplifies accessing
189 * Entries should not be directly dereferenced as
190 * entries belonging to slabs marked pfmemalloc will
191 * have the lower bits set SLAB_OBJ_PFMEMALLOC
195 #define SLAB_OBJ_PFMEMALLOC 1
196 static inline bool is_obj_pfmemalloc(void *objp
)
198 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
201 static inline void set_obj_pfmemalloc(void **objp
)
203 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
207 static inline void clear_obj_pfmemalloc(void **objp
)
209 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
213 * bootstrap: The caches do not work without cpuarrays anymore, but the
214 * cpuarrays are allocated from the generic caches...
216 #define BOOT_CPUCACHE_ENTRIES 1
217 struct arraycache_init
{
218 struct array_cache cache
;
219 void *entries
[BOOT_CPUCACHE_ENTRIES
];
223 * Need this for bootstrapping a per node allocator.
225 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
226 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
227 #define CACHE_CACHE 0
228 #define SIZE_AC MAX_NUMNODES
229 #define SIZE_NODE (2 * MAX_NUMNODES)
231 static int drain_freelist(struct kmem_cache
*cache
,
232 struct kmem_cache_node
*n
, int tofree
);
233 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
235 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
236 static void cache_reap(struct work_struct
*unused
);
238 static int slab_early_init
= 1;
240 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
241 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
243 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
245 INIT_LIST_HEAD(&parent
->slabs_full
);
246 INIT_LIST_HEAD(&parent
->slabs_partial
);
247 INIT_LIST_HEAD(&parent
->slabs_free
);
248 parent
->shared
= NULL
;
249 parent
->alien
= NULL
;
250 parent
->colour_next
= 0;
251 spin_lock_init(&parent
->list_lock
);
252 parent
->free_objects
= 0;
253 parent
->free_touched
= 0;
256 #define MAKE_LIST(cachep, listp, slab, nodeid) \
258 INIT_LIST_HEAD(listp); \
259 list_splice(&(cachep->node[nodeid]->slab), listp); \
262 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
264 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
265 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
266 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
269 #define CFLGS_OFF_SLAB (0x80000000UL)
270 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
272 #define BATCHREFILL_LIMIT 16
274 * Optimization question: fewer reaps means less probability for unnessary
275 * cpucache drain/refill cycles.
277 * OTOH the cpuarrays can contain lots of objects,
278 * which could lock up otherwise freeable slabs.
280 #define REAPTIMEOUT_CPUC (2*HZ)
281 #define REAPTIMEOUT_LIST3 (4*HZ)
284 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
285 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
286 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
287 #define STATS_INC_GROWN(x) ((x)->grown++)
288 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
289 #define STATS_SET_HIGH(x) \
291 if ((x)->num_active > (x)->high_mark) \
292 (x)->high_mark = (x)->num_active; \
294 #define STATS_INC_ERR(x) ((x)->errors++)
295 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
296 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
297 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
298 #define STATS_SET_FREEABLE(x, i) \
300 if ((x)->max_freeable < i) \
301 (x)->max_freeable = i; \
303 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
304 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
305 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
306 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
308 #define STATS_INC_ACTIVE(x) do { } while (0)
309 #define STATS_DEC_ACTIVE(x) do { } while (0)
310 #define STATS_INC_ALLOCED(x) do { } while (0)
311 #define STATS_INC_GROWN(x) do { } while (0)
312 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
313 #define STATS_SET_HIGH(x) do { } while (0)
314 #define STATS_INC_ERR(x) do { } while (0)
315 #define STATS_INC_NODEALLOCS(x) do { } while (0)
316 #define STATS_INC_NODEFREES(x) do { } while (0)
317 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
318 #define STATS_SET_FREEABLE(x, i) do { } while (0)
319 #define STATS_INC_ALLOCHIT(x) do { } while (0)
320 #define STATS_INC_ALLOCMISS(x) do { } while (0)
321 #define STATS_INC_FREEHIT(x) do { } while (0)
322 #define STATS_INC_FREEMISS(x) do { } while (0)
328 * memory layout of objects:
330 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
331 * the end of an object is aligned with the end of the real
332 * allocation. Catches writes behind the end of the allocation.
333 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
335 * cachep->obj_offset: The real object.
336 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
337 * cachep->size - 1* BYTES_PER_WORD: last caller address
338 * [BYTES_PER_WORD long]
340 static int obj_offset(struct kmem_cache
*cachep
)
342 return cachep
->obj_offset
;
345 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
347 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
348 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
349 sizeof(unsigned long long));
352 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
354 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
355 if (cachep
->flags
& SLAB_STORE_USER
)
356 return (unsigned long long *)(objp
+ cachep
->size
-
357 sizeof(unsigned long long) -
359 return (unsigned long long *) (objp
+ cachep
->size
-
360 sizeof(unsigned long long));
363 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
365 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
366 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
371 #define obj_offset(x) 0
372 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
373 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
374 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
378 #define OBJECT_FREE (0)
379 #define OBJECT_ACTIVE (1)
381 #ifdef CONFIG_DEBUG_SLAB_LEAK
383 static void set_obj_status(struct page
*page
, int idx
, int val
)
387 struct kmem_cache
*cachep
= page
->slab_cache
;
389 freelist_size
= cachep
->num
* sizeof(unsigned int);
390 status
= (char *)page
->freelist
+ freelist_size
;
394 static inline unsigned int get_obj_status(struct page
*page
, int idx
)
398 struct kmem_cache
*cachep
= page
->slab_cache
;
400 freelist_size
= cachep
->num
* sizeof(unsigned int);
401 status
= (char *)page
->freelist
+ freelist_size
;
407 static inline void set_obj_status(struct page
*page
, int idx
, int val
) {}
412 * Do not go above this order unless 0 objects fit into the slab or
413 * overridden on the command line.
415 #define SLAB_MAX_ORDER_HI 1
416 #define SLAB_MAX_ORDER_LO 0
417 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
418 static bool slab_max_order_set __initdata
;
420 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
422 struct page
*page
= virt_to_head_page(obj
);
423 return page
->slab_cache
;
426 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
429 return page
->s_mem
+ cache
->size
* idx
;
433 * We want to avoid an expensive divide : (offset / cache->size)
434 * Using the fact that size is a constant for a particular cache,
435 * we can replace (offset / cache->size) by
436 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
438 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
439 const struct page
*page
, void *obj
)
441 u32 offset
= (obj
- page
->s_mem
);
442 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
445 static struct arraycache_init initarray_generic
=
446 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
448 /* internal cache of cache description objs */
449 static struct kmem_cache kmem_cache_boot
= {
451 .limit
= BOOT_CPUCACHE_ENTRIES
,
453 .size
= sizeof(struct kmem_cache
),
454 .name
= "kmem_cache",
457 #define BAD_ALIEN_MAGIC 0x01020304ul
459 #ifdef CONFIG_LOCKDEP
462 * Slab sometimes uses the kmalloc slabs to store the slab headers
463 * for other slabs "off slab".
464 * The locking for this is tricky in that it nests within the locks
465 * of all other slabs in a few places; to deal with this special
466 * locking we put on-slab caches into a separate lock-class.
468 * We set lock class for alien array caches which are up during init.
469 * The lock annotation will be lost if all cpus of a node goes down and
470 * then comes back up during hotplug
472 static struct lock_class_key on_slab_l3_key
;
473 static struct lock_class_key on_slab_alc_key
;
475 static struct lock_class_key debugobj_l3_key
;
476 static struct lock_class_key debugobj_alc_key
;
478 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
479 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
482 struct array_cache
**alc
;
483 struct kmem_cache_node
*n
;
490 lockdep_set_class(&n
->list_lock
, l3_key
);
493 * FIXME: This check for BAD_ALIEN_MAGIC
494 * should go away when common slab code is taught to
495 * work even without alien caches.
496 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
497 * for alloc_alien_cache,
499 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
503 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
507 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
509 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
512 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
516 for_each_online_node(node
)
517 slab_set_debugobj_lock_classes_node(cachep
, node
);
520 static void init_node_lock_keys(int q
)
527 for (i
= 1; i
<= KMALLOC_SHIFT_HIGH
; i
++) {
528 struct kmem_cache_node
*n
;
529 struct kmem_cache
*cache
= kmalloc_caches
[i
];
535 if (!n
|| OFF_SLAB(cache
))
538 slab_set_lock_classes(cache
, &on_slab_l3_key
,
539 &on_slab_alc_key
, q
);
543 static void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int q
)
545 if (!cachep
->node
[q
])
548 slab_set_lock_classes(cachep
, &on_slab_l3_key
,
549 &on_slab_alc_key
, q
);
552 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
556 VM_BUG_ON(OFF_SLAB(cachep
));
558 on_slab_lock_classes_node(cachep
, node
);
561 static inline void init_lock_keys(void)
566 init_node_lock_keys(node
);
569 static void init_node_lock_keys(int q
)
573 static inline void init_lock_keys(void)
577 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
581 static inline void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int node
)
585 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
589 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
594 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
596 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
598 return cachep
->array
[smp_processor_id()];
601 static size_t calculate_freelist_size(int nr_objs
, size_t align
)
603 size_t freelist_size
;
605 freelist_size
= nr_objs
* sizeof(unsigned int);
606 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
607 freelist_size
+= nr_objs
* sizeof(char);
610 freelist_size
= ALIGN(freelist_size
, align
);
612 return freelist_size
;
616 * Calculate the number of objects and left-over bytes for a given buffer size.
618 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
619 size_t align
, int flags
, size_t *left_over
,
624 size_t slab_size
= PAGE_SIZE
<< gfporder
;
627 * The slab management structure can be either off the slab or
628 * on it. For the latter case, the memory allocated for a
631 * - One unsigned int for each object
632 * - Padding to respect alignment of @align
633 * - @buffer_size bytes for each object
635 * If the slab management structure is off the slab, then the
636 * alignment will already be calculated into the size. Because
637 * the slabs are all pages aligned, the objects will be at the
638 * correct alignment when allocated.
640 if (flags
& CFLGS_OFF_SLAB
) {
642 nr_objs
= slab_size
/ buffer_size
;
647 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
648 extra_space
= sizeof(char);
650 * Ignore padding for the initial guess. The padding
651 * is at most @align-1 bytes, and @buffer_size is at
652 * least @align. In the worst case, this result will
653 * be one greater than the number of objects that fit
654 * into the memory allocation when taking the padding
657 nr_objs
= (slab_size
) /
658 (buffer_size
+ sizeof(unsigned int) + extra_space
);
661 * This calculated number will be either the right
662 * amount, or one greater than what we want.
664 if (calculate_freelist_size(nr_objs
, align
) >
665 slab_size
- nr_objs
* buffer_size
)
668 mgmt_size
= calculate_freelist_size(nr_objs
, align
);
671 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
675 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
677 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
680 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
681 function
, cachep
->name
, msg
);
683 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
688 * By default on NUMA we use alien caches to stage the freeing of
689 * objects allocated from other nodes. This causes massive memory
690 * inefficiencies when using fake NUMA setup to split memory into a
691 * large number of small nodes, so it can be disabled on the command
695 static int use_alien_caches __read_mostly
= 1;
696 static int __init
noaliencache_setup(char *s
)
698 use_alien_caches
= 0;
701 __setup("noaliencache", noaliencache_setup
);
703 static int __init
slab_max_order_setup(char *str
)
705 get_option(&str
, &slab_max_order
);
706 slab_max_order
= slab_max_order
< 0 ? 0 :
707 min(slab_max_order
, MAX_ORDER
- 1);
708 slab_max_order_set
= true;
712 __setup("slab_max_order=", slab_max_order_setup
);
716 * Special reaping functions for NUMA systems called from cache_reap().
717 * These take care of doing round robin flushing of alien caches (containing
718 * objects freed on different nodes from which they were allocated) and the
719 * flushing of remote pcps by calling drain_node_pages.
721 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
723 static void init_reap_node(int cpu
)
727 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
728 if (node
== MAX_NUMNODES
)
729 node
= first_node(node_online_map
);
731 per_cpu(slab_reap_node
, cpu
) = node
;
734 static void next_reap_node(void)
736 int node
= __this_cpu_read(slab_reap_node
);
738 node
= next_node(node
, node_online_map
);
739 if (unlikely(node
>= MAX_NUMNODES
))
740 node
= first_node(node_online_map
);
741 __this_cpu_write(slab_reap_node
, node
);
745 #define init_reap_node(cpu) do { } while (0)
746 #define next_reap_node(void) do { } while (0)
750 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
751 * via the workqueue/eventd.
752 * Add the CPU number into the expiration time to minimize the possibility of
753 * the CPUs getting into lockstep and contending for the global cache chain
756 static void start_cpu_timer(int cpu
)
758 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
761 * When this gets called from do_initcalls via cpucache_init(),
762 * init_workqueues() has already run, so keventd will be setup
765 if (keventd_up() && reap_work
->work
.func
== NULL
) {
767 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
768 schedule_delayed_work_on(cpu
, reap_work
,
769 __round_jiffies_relative(HZ
, cpu
));
773 static struct array_cache
*alloc_arraycache(int node
, int entries
,
774 int batchcount
, gfp_t gfp
)
776 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
777 struct array_cache
*nc
= NULL
;
779 nc
= kmalloc_node(memsize
, gfp
, node
);
781 * The array_cache structures contain pointers to free object.
782 * However, when such objects are allocated or transferred to another
783 * cache the pointers are not cleared and they could be counted as
784 * valid references during a kmemleak scan. Therefore, kmemleak must
785 * not scan such objects.
787 kmemleak_no_scan(nc
);
791 nc
->batchcount
= batchcount
;
793 spin_lock_init(&nc
->lock
);
798 static inline bool is_slab_pfmemalloc(struct page
*page
)
800 return PageSlabPfmemalloc(page
);
803 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
804 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
805 struct array_cache
*ac
)
807 struct kmem_cache_node
*n
= cachep
->node
[numa_mem_id()];
811 if (!pfmemalloc_active
)
814 spin_lock_irqsave(&n
->list_lock
, flags
);
815 list_for_each_entry(page
, &n
->slabs_full
, lru
)
816 if (is_slab_pfmemalloc(page
))
819 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
820 if (is_slab_pfmemalloc(page
))
823 list_for_each_entry(page
, &n
->slabs_free
, lru
)
824 if (is_slab_pfmemalloc(page
))
827 pfmemalloc_active
= false;
829 spin_unlock_irqrestore(&n
->list_lock
, flags
);
832 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
833 gfp_t flags
, bool force_refill
)
836 void *objp
= ac
->entry
[--ac
->avail
];
838 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
839 if (unlikely(is_obj_pfmemalloc(objp
))) {
840 struct kmem_cache_node
*n
;
842 if (gfp_pfmemalloc_allowed(flags
)) {
843 clear_obj_pfmemalloc(&objp
);
847 /* The caller cannot use PFMEMALLOC objects, find another one */
848 for (i
= 0; i
< ac
->avail
; i
++) {
849 /* If a !PFMEMALLOC object is found, swap them */
850 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
852 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
853 ac
->entry
[ac
->avail
] = objp
;
859 * If there are empty slabs on the slabs_free list and we are
860 * being forced to refill the cache, mark this one !pfmemalloc.
862 n
= cachep
->node
[numa_mem_id()];
863 if (!list_empty(&n
->slabs_free
) && force_refill
) {
864 struct page
*page
= virt_to_head_page(objp
);
865 ClearPageSlabPfmemalloc(page
);
866 clear_obj_pfmemalloc(&objp
);
867 recheck_pfmemalloc_active(cachep
, ac
);
871 /* No !PFMEMALLOC objects available */
879 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
880 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
884 if (unlikely(sk_memalloc_socks()))
885 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
887 objp
= ac
->entry
[--ac
->avail
];
892 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
895 if (unlikely(pfmemalloc_active
)) {
896 /* Some pfmemalloc slabs exist, check if this is one */
897 struct page
*page
= virt_to_head_page(objp
);
898 if (PageSlabPfmemalloc(page
))
899 set_obj_pfmemalloc(&objp
);
905 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
908 if (unlikely(sk_memalloc_socks()))
909 objp
= __ac_put_obj(cachep
, ac
, objp
);
911 ac
->entry
[ac
->avail
++] = objp
;
915 * Transfer objects in one arraycache to another.
916 * Locking must be handled by the caller.
918 * Return the number of entries transferred.
920 static int transfer_objects(struct array_cache
*to
,
921 struct array_cache
*from
, unsigned int max
)
923 /* Figure out how many entries to transfer */
924 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
929 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
939 #define drain_alien_cache(cachep, alien) do { } while (0)
940 #define reap_alien(cachep, n) do { } while (0)
942 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
944 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
947 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
951 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
956 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
962 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
963 gfp_t flags
, int nodeid
)
968 #else /* CONFIG_NUMA */
970 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
971 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
973 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
975 struct array_cache
**ac_ptr
;
976 int memsize
= sizeof(void *) * nr_node_ids
;
981 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
984 if (i
== node
|| !node_online(i
))
986 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
988 for (i
--; i
>= 0; i
--)
998 static void free_alien_cache(struct array_cache
**ac_ptr
)
1009 static void __drain_alien_cache(struct kmem_cache
*cachep
,
1010 struct array_cache
*ac
, int node
)
1012 struct kmem_cache_node
*n
= cachep
->node
[node
];
1015 spin_lock(&n
->list_lock
);
1017 * Stuff objects into the remote nodes shared array first.
1018 * That way we could avoid the overhead of putting the objects
1019 * into the free lists and getting them back later.
1022 transfer_objects(n
->shared
, ac
, ac
->limit
);
1024 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1026 spin_unlock(&n
->list_lock
);
1031 * Called from cache_reap() to regularly drain alien caches round robin.
1033 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
1035 int node
= __this_cpu_read(slab_reap_node
);
1038 struct array_cache
*ac
= n
->alien
[node
];
1040 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1041 __drain_alien_cache(cachep
, ac
, node
);
1042 spin_unlock_irq(&ac
->lock
);
1047 static void drain_alien_cache(struct kmem_cache
*cachep
,
1048 struct array_cache
**alien
)
1051 struct array_cache
*ac
;
1052 unsigned long flags
;
1054 for_each_online_node(i
) {
1057 spin_lock_irqsave(&ac
->lock
, flags
);
1058 __drain_alien_cache(cachep
, ac
, i
);
1059 spin_unlock_irqrestore(&ac
->lock
, flags
);
1064 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1066 int nodeid
= page_to_nid(virt_to_page(objp
));
1067 struct kmem_cache_node
*n
;
1068 struct array_cache
*alien
= NULL
;
1071 node
= numa_mem_id();
1074 * Make sure we are not freeing a object from another node to the array
1075 * cache on this cpu.
1077 if (likely(nodeid
== node
))
1080 n
= cachep
->node
[node
];
1081 STATS_INC_NODEFREES(cachep
);
1082 if (n
->alien
&& n
->alien
[nodeid
]) {
1083 alien
= n
->alien
[nodeid
];
1084 spin_lock(&alien
->lock
);
1085 if (unlikely(alien
->avail
== alien
->limit
)) {
1086 STATS_INC_ACOVERFLOW(cachep
);
1087 __drain_alien_cache(cachep
, alien
, nodeid
);
1089 ac_put_obj(cachep
, alien
, objp
);
1090 spin_unlock(&alien
->lock
);
1092 spin_lock(&(cachep
->node
[nodeid
])->list_lock
);
1093 free_block(cachep
, &objp
, 1, nodeid
);
1094 spin_unlock(&(cachep
->node
[nodeid
])->list_lock
);
1101 * Allocates and initializes node for a node on each slab cache, used for
1102 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1103 * will be allocated off-node since memory is not yet online for the new node.
1104 * When hotplugging memory or a cpu, existing node are not replaced if
1107 * Must hold slab_mutex.
1109 static int init_cache_node_node(int node
)
1111 struct kmem_cache
*cachep
;
1112 struct kmem_cache_node
*n
;
1113 const int memsize
= sizeof(struct kmem_cache_node
);
1115 list_for_each_entry(cachep
, &slab_caches
, list
) {
1117 * Set up the size64 kmemlist for cpu before we can
1118 * begin anything. Make sure some other cpu on this
1119 * node has not already allocated this
1121 if (!cachep
->node
[node
]) {
1122 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1125 kmem_cache_node_init(n
);
1126 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1127 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1130 * The l3s don't come and go as CPUs come and
1131 * go. slab_mutex is sufficient
1134 cachep
->node
[node
] = n
;
1137 spin_lock_irq(&cachep
->node
[node
]->list_lock
);
1138 cachep
->node
[node
]->free_limit
=
1139 (1 + nr_cpus_node(node
)) *
1140 cachep
->batchcount
+ cachep
->num
;
1141 spin_unlock_irq(&cachep
->node
[node
]->list_lock
);
1146 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1147 struct kmem_cache_node
*n
)
1149 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1152 static void cpuup_canceled(long cpu
)
1154 struct kmem_cache
*cachep
;
1155 struct kmem_cache_node
*n
= NULL
;
1156 int node
= cpu_to_mem(cpu
);
1157 const struct cpumask
*mask
= cpumask_of_node(node
);
1159 list_for_each_entry(cachep
, &slab_caches
, list
) {
1160 struct array_cache
*nc
;
1161 struct array_cache
*shared
;
1162 struct array_cache
**alien
;
1164 /* cpu is dead; no one can alloc from it. */
1165 nc
= cachep
->array
[cpu
];
1166 cachep
->array
[cpu
] = NULL
;
1167 n
= cachep
->node
[node
];
1170 goto free_array_cache
;
1172 spin_lock_irq(&n
->list_lock
);
1174 /* Free limit for this kmem_cache_node */
1175 n
->free_limit
-= cachep
->batchcount
;
1177 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1179 if (!cpumask_empty(mask
)) {
1180 spin_unlock_irq(&n
->list_lock
);
1181 goto free_array_cache
;
1186 free_block(cachep
, shared
->entry
,
1187 shared
->avail
, node
);
1194 spin_unlock_irq(&n
->list_lock
);
1198 drain_alien_cache(cachep
, alien
);
1199 free_alien_cache(alien
);
1205 * In the previous loop, all the objects were freed to
1206 * the respective cache's slabs, now we can go ahead and
1207 * shrink each nodelist to its limit.
1209 list_for_each_entry(cachep
, &slab_caches
, list
) {
1210 n
= cachep
->node
[node
];
1213 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1217 static int cpuup_prepare(long cpu
)
1219 struct kmem_cache
*cachep
;
1220 struct kmem_cache_node
*n
= NULL
;
1221 int node
= cpu_to_mem(cpu
);
1225 * We need to do this right in the beginning since
1226 * alloc_arraycache's are going to use this list.
1227 * kmalloc_node allows us to add the slab to the right
1228 * kmem_cache_node and not this cpu's kmem_cache_node
1230 err
= init_cache_node_node(node
);
1235 * Now we can go ahead with allocating the shared arrays and
1238 list_for_each_entry(cachep
, &slab_caches
, list
) {
1239 struct array_cache
*nc
;
1240 struct array_cache
*shared
= NULL
;
1241 struct array_cache
**alien
= NULL
;
1243 nc
= alloc_arraycache(node
, cachep
->limit
,
1244 cachep
->batchcount
, GFP_KERNEL
);
1247 if (cachep
->shared
) {
1248 shared
= alloc_arraycache(node
,
1249 cachep
->shared
* cachep
->batchcount
,
1250 0xbaadf00d, GFP_KERNEL
);
1256 if (use_alien_caches
) {
1257 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1264 cachep
->array
[cpu
] = nc
;
1265 n
= cachep
->node
[node
];
1268 spin_lock_irq(&n
->list_lock
);
1271 * We are serialised from CPU_DEAD or
1272 * CPU_UP_CANCELLED by the cpucontrol lock
1283 spin_unlock_irq(&n
->list_lock
);
1285 free_alien_cache(alien
);
1286 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1287 slab_set_debugobj_lock_classes_node(cachep
, node
);
1288 else if (!OFF_SLAB(cachep
) &&
1289 !(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1290 on_slab_lock_classes_node(cachep
, node
);
1292 init_node_lock_keys(node
);
1296 cpuup_canceled(cpu
);
1300 static int cpuup_callback(struct notifier_block
*nfb
,
1301 unsigned long action
, void *hcpu
)
1303 long cpu
= (long)hcpu
;
1307 case CPU_UP_PREPARE
:
1308 case CPU_UP_PREPARE_FROZEN
:
1309 mutex_lock(&slab_mutex
);
1310 err
= cpuup_prepare(cpu
);
1311 mutex_unlock(&slab_mutex
);
1314 case CPU_ONLINE_FROZEN
:
1315 start_cpu_timer(cpu
);
1317 #ifdef CONFIG_HOTPLUG_CPU
1318 case CPU_DOWN_PREPARE
:
1319 case CPU_DOWN_PREPARE_FROZEN
:
1321 * Shutdown cache reaper. Note that the slab_mutex is
1322 * held so that if cache_reap() is invoked it cannot do
1323 * anything expensive but will only modify reap_work
1324 * and reschedule the timer.
1326 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1327 /* Now the cache_reaper is guaranteed to be not running. */
1328 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1330 case CPU_DOWN_FAILED
:
1331 case CPU_DOWN_FAILED_FROZEN
:
1332 start_cpu_timer(cpu
);
1335 case CPU_DEAD_FROZEN
:
1337 * Even if all the cpus of a node are down, we don't free the
1338 * kmem_cache_node of any cache. This to avoid a race between
1339 * cpu_down, and a kmalloc allocation from another cpu for
1340 * memory from the node of the cpu going down. The node
1341 * structure is usually allocated from kmem_cache_create() and
1342 * gets destroyed at kmem_cache_destroy().
1346 case CPU_UP_CANCELED
:
1347 case CPU_UP_CANCELED_FROZEN
:
1348 mutex_lock(&slab_mutex
);
1349 cpuup_canceled(cpu
);
1350 mutex_unlock(&slab_mutex
);
1353 return notifier_from_errno(err
);
1356 static struct notifier_block cpucache_notifier
= {
1357 &cpuup_callback
, NULL
, 0
1360 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1362 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1363 * Returns -EBUSY if all objects cannot be drained so that the node is not
1366 * Must hold slab_mutex.
1368 static int __meminit
drain_cache_node_node(int node
)
1370 struct kmem_cache
*cachep
;
1373 list_for_each_entry(cachep
, &slab_caches
, list
) {
1374 struct kmem_cache_node
*n
;
1376 n
= cachep
->node
[node
];
1380 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1382 if (!list_empty(&n
->slabs_full
) ||
1383 !list_empty(&n
->slabs_partial
)) {
1391 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1392 unsigned long action
, void *arg
)
1394 struct memory_notify
*mnb
= arg
;
1398 nid
= mnb
->status_change_nid
;
1403 case MEM_GOING_ONLINE
:
1404 mutex_lock(&slab_mutex
);
1405 ret
= init_cache_node_node(nid
);
1406 mutex_unlock(&slab_mutex
);
1408 case MEM_GOING_OFFLINE
:
1409 mutex_lock(&slab_mutex
);
1410 ret
= drain_cache_node_node(nid
);
1411 mutex_unlock(&slab_mutex
);
1415 case MEM_CANCEL_ONLINE
:
1416 case MEM_CANCEL_OFFLINE
:
1420 return notifier_from_errno(ret
);
1422 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1425 * swap the static kmem_cache_node with kmalloced memory
1427 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1430 struct kmem_cache_node
*ptr
;
1432 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1435 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1437 * Do not assume that spinlocks can be initialized via memcpy:
1439 spin_lock_init(&ptr
->list_lock
);
1441 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1442 cachep
->node
[nodeid
] = ptr
;
1446 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1447 * size of kmem_cache_node.
1449 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1453 for_each_online_node(node
) {
1454 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1455 cachep
->node
[node
]->next_reap
= jiffies
+
1457 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1462 * The memory after the last cpu cache pointer is used for the
1465 static void setup_node_pointer(struct kmem_cache
*cachep
)
1467 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1471 * Initialisation. Called after the page allocator have been initialised and
1472 * before smp_init().
1474 void __init
kmem_cache_init(void)
1478 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1479 sizeof(struct rcu_head
));
1480 kmem_cache
= &kmem_cache_boot
;
1481 setup_node_pointer(kmem_cache
);
1483 if (num_possible_nodes() == 1)
1484 use_alien_caches
= 0;
1486 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1487 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1489 set_up_node(kmem_cache
, CACHE_CACHE
);
1492 * Fragmentation resistance on low memory - only use bigger
1493 * page orders on machines with more than 32MB of memory if
1494 * not overridden on the command line.
1496 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1497 slab_max_order
= SLAB_MAX_ORDER_HI
;
1499 /* Bootstrap is tricky, because several objects are allocated
1500 * from caches that do not exist yet:
1501 * 1) initialize the kmem_cache cache: it contains the struct
1502 * kmem_cache structures of all caches, except kmem_cache itself:
1503 * kmem_cache is statically allocated.
1504 * Initially an __init data area is used for the head array and the
1505 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1506 * array at the end of the bootstrap.
1507 * 2) Create the first kmalloc cache.
1508 * The struct kmem_cache for the new cache is allocated normally.
1509 * An __init data area is used for the head array.
1510 * 3) Create the remaining kmalloc caches, with minimally sized
1512 * 4) Replace the __init data head arrays for kmem_cache and the first
1513 * kmalloc cache with kmalloc allocated arrays.
1514 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1515 * the other cache's with kmalloc allocated memory.
1516 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1519 /* 1) create the kmem_cache */
1522 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1524 create_boot_cache(kmem_cache
, "kmem_cache",
1525 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1526 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1527 SLAB_HWCACHE_ALIGN
);
1528 list_add(&kmem_cache
->list
, &slab_caches
);
1530 /* 2+3) create the kmalloc caches */
1533 * Initialize the caches that provide memory for the array cache and the
1534 * kmem_cache_node structures first. Without this, further allocations will
1538 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1539 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1541 if (INDEX_AC
!= INDEX_NODE
)
1542 kmalloc_caches
[INDEX_NODE
] =
1543 create_kmalloc_cache("kmalloc-node",
1544 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1546 slab_early_init
= 0;
1548 /* 4) Replace the bootstrap head arrays */
1550 struct array_cache
*ptr
;
1552 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1554 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1555 sizeof(struct arraycache_init
));
1557 * Do not assume that spinlocks can be initialized via memcpy:
1559 spin_lock_init(&ptr
->lock
);
1561 kmem_cache
->array
[smp_processor_id()] = ptr
;
1563 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1565 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1566 != &initarray_generic
.cache
);
1567 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1568 sizeof(struct arraycache_init
));
1570 * Do not assume that spinlocks can be initialized via memcpy:
1572 spin_lock_init(&ptr
->lock
);
1574 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1576 /* 5) Replace the bootstrap kmem_cache_node */
1580 for_each_online_node(nid
) {
1581 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1583 init_list(kmalloc_caches
[INDEX_AC
],
1584 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1586 if (INDEX_AC
!= INDEX_NODE
) {
1587 init_list(kmalloc_caches
[INDEX_NODE
],
1588 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1593 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1596 void __init
kmem_cache_init_late(void)
1598 struct kmem_cache
*cachep
;
1602 /* 6) resize the head arrays to their final sizes */
1603 mutex_lock(&slab_mutex
);
1604 list_for_each_entry(cachep
, &slab_caches
, list
)
1605 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1607 mutex_unlock(&slab_mutex
);
1609 /* Annotate slab for lockdep -- annotate the malloc caches */
1616 * Register a cpu startup notifier callback that initializes
1617 * cpu_cache_get for all new cpus
1619 register_cpu_notifier(&cpucache_notifier
);
1623 * Register a memory hotplug callback that initializes and frees
1626 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1630 * The reap timers are started later, with a module init call: That part
1631 * of the kernel is not yet operational.
1635 static int __init
cpucache_init(void)
1640 * Register the timers that return unneeded pages to the page allocator
1642 for_each_online_cpu(cpu
)
1643 start_cpu_timer(cpu
);
1649 __initcall(cpucache_init
);
1651 static noinline
void
1652 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1654 struct kmem_cache_node
*n
;
1656 unsigned long flags
;
1660 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1662 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1663 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1665 for_each_online_node(node
) {
1666 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1667 unsigned long active_slabs
= 0, num_slabs
= 0;
1669 n
= cachep
->node
[node
];
1673 spin_lock_irqsave(&n
->list_lock
, flags
);
1674 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1675 active_objs
+= cachep
->num
;
1678 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1679 active_objs
+= page
->active
;
1682 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1685 free_objects
+= n
->free_objects
;
1686 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1688 num_slabs
+= active_slabs
;
1689 num_objs
= num_slabs
* cachep
->num
;
1691 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1692 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1698 * Interface to system's page allocator. No need to hold the cache-lock.
1700 * If we requested dmaable memory, we will get it. Even if we
1701 * did not request dmaable memory, we might get it, but that
1702 * would be relatively rare and ignorable.
1704 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1710 flags
|= cachep
->allocflags
;
1711 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1712 flags
|= __GFP_RECLAIMABLE
;
1714 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1716 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1717 slab_out_of_memory(cachep
, flags
, nodeid
);
1721 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1722 if (unlikely(page
->pfmemalloc
))
1723 pfmemalloc_active
= true;
1725 nr_pages
= (1 << cachep
->gfporder
);
1726 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1727 add_zone_page_state(page_zone(page
),
1728 NR_SLAB_RECLAIMABLE
, nr_pages
);
1730 add_zone_page_state(page_zone(page
),
1731 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1732 __SetPageSlab(page
);
1733 if (page
->pfmemalloc
)
1734 SetPageSlabPfmemalloc(page
);
1735 memcg_bind_pages(cachep
, cachep
->gfporder
);
1737 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1738 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1741 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1743 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1750 * Interface to system's page release.
1752 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1754 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1756 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1758 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1759 sub_zone_page_state(page_zone(page
),
1760 NR_SLAB_RECLAIMABLE
, nr_freed
);
1762 sub_zone_page_state(page_zone(page
),
1763 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1765 BUG_ON(!PageSlab(page
));
1766 __ClearPageSlabPfmemalloc(page
);
1767 __ClearPageSlab(page
);
1768 page_mapcount_reset(page
);
1769 page
->mapping
= NULL
;
1771 memcg_release_pages(cachep
, cachep
->gfporder
);
1772 if (current
->reclaim_state
)
1773 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1774 __free_memcg_kmem_pages(page
, cachep
->gfporder
);
1777 static void kmem_rcu_free(struct rcu_head
*head
)
1779 struct kmem_cache
*cachep
;
1782 page
= container_of(head
, struct page
, rcu_head
);
1783 cachep
= page
->slab_cache
;
1785 kmem_freepages(cachep
, page
);
1790 #ifdef CONFIG_DEBUG_PAGEALLOC
1791 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1792 unsigned long caller
)
1794 int size
= cachep
->object_size
;
1796 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1798 if (size
< 5 * sizeof(unsigned long))
1801 *addr
++ = 0x12345678;
1803 *addr
++ = smp_processor_id();
1804 size
-= 3 * sizeof(unsigned long);
1806 unsigned long *sptr
= &caller
;
1807 unsigned long svalue
;
1809 while (!kstack_end(sptr
)) {
1811 if (kernel_text_address(svalue
)) {
1813 size
-= sizeof(unsigned long);
1814 if (size
<= sizeof(unsigned long))
1820 *addr
++ = 0x87654321;
1824 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1826 int size
= cachep
->object_size
;
1827 addr
= &((char *)addr
)[obj_offset(cachep
)];
1829 memset(addr
, val
, size
);
1830 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1833 static void dump_line(char *data
, int offset
, int limit
)
1836 unsigned char error
= 0;
1839 printk(KERN_ERR
"%03x: ", offset
);
1840 for (i
= 0; i
< limit
; i
++) {
1841 if (data
[offset
+ i
] != POISON_FREE
) {
1842 error
= data
[offset
+ i
];
1846 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1847 &data
[offset
], limit
, 1);
1849 if (bad_count
== 1) {
1850 error
^= POISON_FREE
;
1851 if (!(error
& (error
- 1))) {
1852 printk(KERN_ERR
"Single bit error detected. Probably "
1855 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1858 printk(KERN_ERR
"Run a memory test tool.\n");
1867 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1872 if (cachep
->flags
& SLAB_RED_ZONE
) {
1873 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1874 *dbg_redzone1(cachep
, objp
),
1875 *dbg_redzone2(cachep
, objp
));
1878 if (cachep
->flags
& SLAB_STORE_USER
) {
1879 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1880 *dbg_userword(cachep
, objp
),
1881 *dbg_userword(cachep
, objp
));
1883 realobj
= (char *)objp
+ obj_offset(cachep
);
1884 size
= cachep
->object_size
;
1885 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1888 if (i
+ limit
> size
)
1890 dump_line(realobj
, i
, limit
);
1894 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1900 realobj
= (char *)objp
+ obj_offset(cachep
);
1901 size
= cachep
->object_size
;
1903 for (i
= 0; i
< size
; i
++) {
1904 char exp
= POISON_FREE
;
1907 if (realobj
[i
] != exp
) {
1913 "Slab corruption (%s): %s start=%p, len=%d\n",
1914 print_tainted(), cachep
->name
, realobj
, size
);
1915 print_objinfo(cachep
, objp
, 0);
1917 /* Hexdump the affected line */
1920 if (i
+ limit
> size
)
1922 dump_line(realobj
, i
, limit
);
1925 /* Limit to 5 lines */
1931 /* Print some data about the neighboring objects, if they
1934 struct page
*page
= virt_to_head_page(objp
);
1937 objnr
= obj_to_index(cachep
, page
, objp
);
1939 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1940 realobj
= (char *)objp
+ obj_offset(cachep
);
1941 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1943 print_objinfo(cachep
, objp
, 2);
1945 if (objnr
+ 1 < cachep
->num
) {
1946 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1947 realobj
= (char *)objp
+ obj_offset(cachep
);
1948 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1950 print_objinfo(cachep
, objp
, 2);
1957 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1961 for (i
= 0; i
< cachep
->num
; i
++) {
1962 void *objp
= index_to_obj(cachep
, page
, i
);
1964 if (cachep
->flags
& SLAB_POISON
) {
1965 #ifdef CONFIG_DEBUG_PAGEALLOC
1966 if (cachep
->size
% PAGE_SIZE
== 0 &&
1968 kernel_map_pages(virt_to_page(objp
),
1969 cachep
->size
/ PAGE_SIZE
, 1);
1971 check_poison_obj(cachep
, objp
);
1973 check_poison_obj(cachep
, objp
);
1976 if (cachep
->flags
& SLAB_RED_ZONE
) {
1977 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1978 slab_error(cachep
, "start of a freed object "
1980 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1981 slab_error(cachep
, "end of a freed object "
1987 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1994 * slab_destroy - destroy and release all objects in a slab
1995 * @cachep: cache pointer being destroyed
1996 * @page: page pointer being destroyed
1998 * Destroy all the objs in a slab, and release the mem back to the system.
1999 * Before calling the slab must have been unlinked from the cache. The
2000 * cache-lock is not held/needed.
2002 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
2006 freelist
= page
->freelist
;
2007 slab_destroy_debugcheck(cachep
, page
);
2008 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
2009 struct rcu_head
*head
;
2012 * RCU free overloads the RCU head over the LRU.
2013 * slab_page has been overloeaded over the LRU,
2014 * however it is not used from now on so that
2015 * we can use it safely.
2017 head
= (void *)&page
->rcu_head
;
2018 call_rcu(head
, kmem_rcu_free
);
2021 kmem_freepages(cachep
, page
);
2025 * From now on, we don't use freelist
2026 * although actual page can be freed in rcu context
2028 if (OFF_SLAB(cachep
))
2029 kmem_cache_free(cachep
->freelist_cache
, freelist
);
2033 * calculate_slab_order - calculate size (page order) of slabs
2034 * @cachep: pointer to the cache that is being created
2035 * @size: size of objects to be created in this cache.
2036 * @align: required alignment for the objects.
2037 * @flags: slab allocation flags
2039 * Also calculates the number of objects per slab.
2041 * This could be made much more intelligent. For now, try to avoid using
2042 * high order pages for slabs. When the gfp() functions are more friendly
2043 * towards high-order requests, this should be changed.
2045 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2046 size_t size
, size_t align
, unsigned long flags
)
2048 unsigned long offslab_limit
;
2049 size_t left_over
= 0;
2052 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2056 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2060 if (flags
& CFLGS_OFF_SLAB
) {
2061 size_t freelist_size_per_obj
= sizeof(unsigned int);
2063 * Max number of objs-per-slab for caches which
2064 * use off-slab slabs. Needed to avoid a possible
2065 * looping condition in cache_grow().
2067 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK
))
2068 freelist_size_per_obj
+= sizeof(char);
2069 offslab_limit
= size
;
2070 offslab_limit
/= freelist_size_per_obj
;
2072 if (num
> offslab_limit
)
2076 /* Found something acceptable - save it away */
2078 cachep
->gfporder
= gfporder
;
2079 left_over
= remainder
;
2082 * A VFS-reclaimable slab tends to have most allocations
2083 * as GFP_NOFS and we really don't want to have to be allocating
2084 * higher-order pages when we are unable to shrink dcache.
2086 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2090 * Large number of objects is good, but very large slabs are
2091 * currently bad for the gfp()s.
2093 if (gfporder
>= slab_max_order
)
2097 * Acceptable internal fragmentation?
2099 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2105 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2107 if (slab_state
>= FULL
)
2108 return enable_cpucache(cachep
, gfp
);
2110 if (slab_state
== DOWN
) {
2112 * Note: Creation of first cache (kmem_cache).
2113 * The setup_node is taken care
2114 * of by the caller of __kmem_cache_create
2116 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2117 slab_state
= PARTIAL
;
2118 } else if (slab_state
== PARTIAL
) {
2120 * Note: the second kmem_cache_create must create the cache
2121 * that's used by kmalloc(24), otherwise the creation of
2122 * further caches will BUG().
2124 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2127 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2128 * the second cache, then we need to set up all its node/,
2129 * otherwise the creation of further caches will BUG().
2131 set_up_node(cachep
, SIZE_AC
);
2132 if (INDEX_AC
== INDEX_NODE
)
2133 slab_state
= PARTIAL_NODE
;
2135 slab_state
= PARTIAL_ARRAYCACHE
;
2137 /* Remaining boot caches */
2138 cachep
->array
[smp_processor_id()] =
2139 kmalloc(sizeof(struct arraycache_init
), gfp
);
2141 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2142 set_up_node(cachep
, SIZE_NODE
);
2143 slab_state
= PARTIAL_NODE
;
2146 for_each_online_node(node
) {
2147 cachep
->node
[node
] =
2148 kmalloc_node(sizeof(struct kmem_cache_node
),
2150 BUG_ON(!cachep
->node
[node
]);
2151 kmem_cache_node_init(cachep
->node
[node
]);
2155 cachep
->node
[numa_mem_id()]->next_reap
=
2156 jiffies
+ REAPTIMEOUT_LIST3
+
2157 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2159 cpu_cache_get(cachep
)->avail
= 0;
2160 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2161 cpu_cache_get(cachep
)->batchcount
= 1;
2162 cpu_cache_get(cachep
)->touched
= 0;
2163 cachep
->batchcount
= 1;
2164 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2169 * __kmem_cache_create - Create a cache.
2170 * @cachep: cache management descriptor
2171 * @flags: SLAB flags
2173 * Returns a ptr to the cache on success, NULL on failure.
2174 * Cannot be called within a int, but can be interrupted.
2175 * The @ctor is run when new pages are allocated by the cache.
2179 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2180 * to catch references to uninitialised memory.
2182 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2183 * for buffer overruns.
2185 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2186 * cacheline. This can be beneficial if you're counting cycles as closely
2190 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2192 size_t left_over
, freelist_size
;
2193 size_t ralign
= BYTES_PER_WORD
;
2196 size_t size
= cachep
->size
;
2201 * Enable redzoning and last user accounting, except for caches with
2202 * large objects, if the increased size would increase the object size
2203 * above the next power of two: caches with object sizes just above a
2204 * power of two have a significant amount of internal fragmentation.
2206 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2207 2 * sizeof(unsigned long long)))
2208 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2209 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2210 flags
|= SLAB_POISON
;
2212 if (flags
& SLAB_DESTROY_BY_RCU
)
2213 BUG_ON(flags
& SLAB_POISON
);
2217 * Check that size is in terms of words. This is needed to avoid
2218 * unaligned accesses for some archs when redzoning is used, and makes
2219 * sure any on-slab bufctl's are also correctly aligned.
2221 if (size
& (BYTES_PER_WORD
- 1)) {
2222 size
+= (BYTES_PER_WORD
- 1);
2223 size
&= ~(BYTES_PER_WORD
- 1);
2226 if (flags
& SLAB_RED_ZONE
) {
2227 ralign
= REDZONE_ALIGN
;
2228 /* If redzoning, ensure that the second redzone is suitably
2229 * aligned, by adjusting the object size accordingly. */
2230 size
+= REDZONE_ALIGN
- 1;
2231 size
&= ~(REDZONE_ALIGN
- 1);
2234 /* 3) caller mandated alignment */
2235 if (ralign
< cachep
->align
) {
2236 ralign
= cachep
->align
;
2238 /* disable debug if necessary */
2239 if (ralign
> __alignof__(unsigned long long))
2240 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2244 cachep
->align
= ralign
;
2246 if (slab_is_available())
2251 setup_node_pointer(cachep
);
2255 * Both debugging options require word-alignment which is calculated
2258 if (flags
& SLAB_RED_ZONE
) {
2259 /* add space for red zone words */
2260 cachep
->obj_offset
+= sizeof(unsigned long long);
2261 size
+= 2 * sizeof(unsigned long long);
2263 if (flags
& SLAB_STORE_USER
) {
2264 /* user store requires one word storage behind the end of
2265 * the real object. But if the second red zone needs to be
2266 * aligned to 64 bits, we must allow that much space.
2268 if (flags
& SLAB_RED_ZONE
)
2269 size
+= REDZONE_ALIGN
;
2271 size
+= BYTES_PER_WORD
;
2273 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2275 * To activate debug pagealloc, off-slab management is necessary
2276 * requirement. In early phase of initialization, small sized slab
2277 * doesn't get initialized so it would not be possible. So, we need
2278 * to check size >= 256. It guarantees that all necessary small
2279 * sized slab is initialized in current slab initialization sequence.
2281 if (!slab_early_init
&& size
>= kmalloc_size(INDEX_NODE
) &&
2282 size
>= 256 && cachep
->object_size
> cache_line_size() &&
2283 ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2284 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2291 * Determine if the slab management is 'on' or 'off' slab.
2292 * (bootstrapping cannot cope with offslab caches so don't do
2293 * it too early on. Always use on-slab management when
2294 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2296 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2297 !(flags
& SLAB_NOLEAKTRACE
))
2299 * Size is large, assume best to place the slab management obj
2300 * off-slab (should allow better packing of objs).
2302 flags
|= CFLGS_OFF_SLAB
;
2304 size
= ALIGN(size
, cachep
->align
);
2306 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2311 freelist_size
= calculate_freelist_size(cachep
->num
, cachep
->align
);
2314 * If the slab has been placed off-slab, and we have enough space then
2315 * move it on-slab. This is at the expense of any extra colouring.
2317 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= freelist_size
) {
2318 flags
&= ~CFLGS_OFF_SLAB
;
2319 left_over
-= freelist_size
;
2322 if (flags
& CFLGS_OFF_SLAB
) {
2323 /* really off slab. No need for manual alignment */
2324 freelist_size
= calculate_freelist_size(cachep
->num
, 0);
2326 #ifdef CONFIG_PAGE_POISONING
2327 /* If we're going to use the generic kernel_map_pages()
2328 * poisoning, then it's going to smash the contents of
2329 * the redzone and userword anyhow, so switch them off.
2331 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2332 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2336 cachep
->colour_off
= cache_line_size();
2337 /* Offset must be a multiple of the alignment. */
2338 if (cachep
->colour_off
< cachep
->align
)
2339 cachep
->colour_off
= cachep
->align
;
2340 cachep
->colour
= left_over
/ cachep
->colour_off
;
2341 cachep
->freelist_size
= freelist_size
;
2342 cachep
->flags
= flags
;
2343 cachep
->allocflags
= __GFP_COMP
;
2344 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2345 cachep
->allocflags
|= GFP_DMA
;
2346 cachep
->size
= size
;
2347 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2349 if (flags
& CFLGS_OFF_SLAB
) {
2350 cachep
->freelist_cache
= kmalloc_slab(freelist_size
, 0u);
2352 * This is a possibility for one of the malloc_sizes caches.
2353 * But since we go off slab only for object size greater than
2354 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2355 * this should not happen at all.
2356 * But leave a BUG_ON for some lucky dude.
2358 BUG_ON(ZERO_OR_NULL_PTR(cachep
->freelist_cache
));
2361 err
= setup_cpu_cache(cachep
, gfp
);
2363 __kmem_cache_shutdown(cachep
);
2367 if (flags
& SLAB_DEBUG_OBJECTS
) {
2369 * Would deadlock through slab_destroy()->call_rcu()->
2370 * debug_object_activate()->kmem_cache_alloc().
2372 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2374 slab_set_debugobj_lock_classes(cachep
);
2375 } else if (!OFF_SLAB(cachep
) && !(flags
& SLAB_DESTROY_BY_RCU
))
2376 on_slab_lock_classes(cachep
);
2382 static void check_irq_off(void)
2384 BUG_ON(!irqs_disabled());
2387 static void check_irq_on(void)
2389 BUG_ON(irqs_disabled());
2392 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2396 assert_spin_locked(&cachep
->node
[numa_mem_id()]->list_lock
);
2400 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2404 assert_spin_locked(&cachep
->node
[node
]->list_lock
);
2409 #define check_irq_off() do { } while(0)
2410 #define check_irq_on() do { } while(0)
2411 #define check_spinlock_acquired(x) do { } while(0)
2412 #define check_spinlock_acquired_node(x, y) do { } while(0)
2415 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2416 struct array_cache
*ac
,
2417 int force
, int node
);
2419 static void do_drain(void *arg
)
2421 struct kmem_cache
*cachep
= arg
;
2422 struct array_cache
*ac
;
2423 int node
= numa_mem_id();
2426 ac
= cpu_cache_get(cachep
);
2427 spin_lock(&cachep
->node
[node
]->list_lock
);
2428 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2429 spin_unlock(&cachep
->node
[node
]->list_lock
);
2433 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2435 struct kmem_cache_node
*n
;
2438 on_each_cpu(do_drain
, cachep
, 1);
2440 for_each_online_node(node
) {
2441 n
= cachep
->node
[node
];
2443 drain_alien_cache(cachep
, n
->alien
);
2446 for_each_online_node(node
) {
2447 n
= cachep
->node
[node
];
2449 drain_array(cachep
, n
, n
->shared
, 1, node
);
2454 * Remove slabs from the list of free slabs.
2455 * Specify the number of slabs to drain in tofree.
2457 * Returns the actual number of slabs released.
2459 static int drain_freelist(struct kmem_cache
*cache
,
2460 struct kmem_cache_node
*n
, int tofree
)
2462 struct list_head
*p
;
2467 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2469 spin_lock_irq(&n
->list_lock
);
2470 p
= n
->slabs_free
.prev
;
2471 if (p
== &n
->slabs_free
) {
2472 spin_unlock_irq(&n
->list_lock
);
2476 page
= list_entry(p
, struct page
, lru
);
2478 BUG_ON(page
->active
);
2480 list_del(&page
->lru
);
2482 * Safe to drop the lock. The slab is no longer linked
2485 n
->free_objects
-= cache
->num
;
2486 spin_unlock_irq(&n
->list_lock
);
2487 slab_destroy(cache
, page
);
2494 /* Called with slab_mutex held to protect against cpu hotplug */
2495 static int __cache_shrink(struct kmem_cache
*cachep
)
2498 struct kmem_cache_node
*n
;
2500 drain_cpu_caches(cachep
);
2503 for_each_online_node(i
) {
2504 n
= cachep
->node
[i
];
2508 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2510 ret
+= !list_empty(&n
->slabs_full
) ||
2511 !list_empty(&n
->slabs_partial
);
2513 return (ret
? 1 : 0);
2517 * kmem_cache_shrink - Shrink a cache.
2518 * @cachep: The cache to shrink.
2520 * Releases as many slabs as possible for a cache.
2521 * To help debugging, a zero exit status indicates all slabs were released.
2523 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2526 BUG_ON(!cachep
|| in_interrupt());
2529 mutex_lock(&slab_mutex
);
2530 ret
= __cache_shrink(cachep
);
2531 mutex_unlock(&slab_mutex
);
2535 EXPORT_SYMBOL(kmem_cache_shrink
);
2537 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2540 struct kmem_cache_node
*n
;
2541 int rc
= __cache_shrink(cachep
);
2546 for_each_online_cpu(i
)
2547 kfree(cachep
->array
[i
]);
2549 /* NUMA: free the node structures */
2550 for_each_online_node(i
) {
2551 n
= cachep
->node
[i
];
2554 free_alien_cache(n
->alien
);
2562 * Get the memory for a slab management obj.
2563 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2564 * always come from malloc_sizes caches. The slab descriptor cannot
2565 * come from the same cache which is getting created because,
2566 * when we are searching for an appropriate cache for these
2567 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2568 * If we are creating a malloc_sizes cache here it would not be visible to
2569 * kmem_find_general_cachep till the initialization is complete.
2570 * Hence we cannot have freelist_cache same as the original cache.
2572 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2573 struct page
*page
, int colour_off
,
2574 gfp_t local_flags
, int nodeid
)
2577 void *addr
= page_address(page
);
2579 if (OFF_SLAB(cachep
)) {
2580 /* Slab management obj is off-slab. */
2581 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2582 local_flags
, nodeid
);
2586 freelist
= addr
+ colour_off
;
2587 colour_off
+= cachep
->freelist_size
;
2590 page
->s_mem
= addr
+ colour_off
;
2594 static inline unsigned int *slab_freelist(struct page
*page
)
2596 return (unsigned int *)(page
->freelist
);
2599 static void cache_init_objs(struct kmem_cache
*cachep
,
2604 for (i
= 0; i
< cachep
->num
; i
++) {
2605 void *objp
= index_to_obj(cachep
, page
, i
);
2607 /* need to poison the objs? */
2608 if (cachep
->flags
& SLAB_POISON
)
2609 poison_obj(cachep
, objp
, POISON_FREE
);
2610 if (cachep
->flags
& SLAB_STORE_USER
)
2611 *dbg_userword(cachep
, objp
) = NULL
;
2613 if (cachep
->flags
& SLAB_RED_ZONE
) {
2614 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2615 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2618 * Constructors are not allowed to allocate memory from the same
2619 * cache which they are a constructor for. Otherwise, deadlock.
2620 * They must also be threaded.
2622 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2623 cachep
->ctor(objp
+ obj_offset(cachep
));
2625 if (cachep
->flags
& SLAB_RED_ZONE
) {
2626 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2627 slab_error(cachep
, "constructor overwrote the"
2628 " end of an object");
2629 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2630 slab_error(cachep
, "constructor overwrote the"
2631 " start of an object");
2633 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2634 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2635 kernel_map_pages(virt_to_page(objp
),
2636 cachep
->size
/ PAGE_SIZE
, 0);
2641 set_obj_status(page
, i
, OBJECT_FREE
);
2642 slab_freelist(page
)[i
] = i
;
2646 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2648 if (CONFIG_ZONE_DMA_FLAG
) {
2649 if (flags
& GFP_DMA
)
2650 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2652 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2656 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
,
2661 objp
= index_to_obj(cachep
, page
, slab_freelist(page
)[page
->active
]);
2664 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2670 static void slab_put_obj(struct kmem_cache
*cachep
, struct page
*page
,
2671 void *objp
, int nodeid
)
2673 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2677 /* Verify that the slab belongs to the intended node */
2678 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2680 /* Verify double free bug */
2681 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2682 if (slab_freelist(page
)[i
] == objnr
) {
2683 printk(KERN_ERR
"slab: double free detected in cache "
2684 "'%s', objp %p\n", cachep
->name
, objp
);
2690 slab_freelist(page
)[page
->active
] = objnr
;
2694 * Map pages beginning at addr to the given cache and slab. This is required
2695 * for the slab allocator to be able to lookup the cache and slab of a
2696 * virtual address for kfree, ksize, and slab debugging.
2698 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2701 page
->slab_cache
= cache
;
2702 page
->freelist
= freelist
;
2706 * Grow (by 1) the number of slabs within a cache. This is called by
2707 * kmem_cache_alloc() when there are no active objs left in a cache.
2709 static int cache_grow(struct kmem_cache
*cachep
,
2710 gfp_t flags
, int nodeid
, struct page
*page
)
2715 struct kmem_cache_node
*n
;
2718 * Be lazy and only check for valid flags here, keeping it out of the
2719 * critical path in kmem_cache_alloc().
2721 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2722 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2724 /* Take the node list lock to change the colour_next on this node */
2726 n
= cachep
->node
[nodeid
];
2727 spin_lock(&n
->list_lock
);
2729 /* Get colour for the slab, and cal the next value. */
2730 offset
= n
->colour_next
;
2732 if (n
->colour_next
>= cachep
->colour
)
2734 spin_unlock(&n
->list_lock
);
2736 offset
*= cachep
->colour_off
;
2738 if (local_flags
& __GFP_WAIT
)
2742 * The test for missing atomic flag is performed here, rather than
2743 * the more obvious place, simply to reduce the critical path length
2744 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2745 * will eventually be caught here (where it matters).
2747 kmem_flagcheck(cachep
, flags
);
2750 * Get mem for the objs. Attempt to allocate a physical page from
2754 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2758 /* Get slab management. */
2759 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2760 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2764 slab_map_pages(cachep
, page
, freelist
);
2766 cache_init_objs(cachep
, page
);
2768 if (local_flags
& __GFP_WAIT
)
2769 local_irq_disable();
2771 spin_lock(&n
->list_lock
);
2773 /* Make slab active. */
2774 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2775 STATS_INC_GROWN(cachep
);
2776 n
->free_objects
+= cachep
->num
;
2777 spin_unlock(&n
->list_lock
);
2780 kmem_freepages(cachep
, page
);
2782 if (local_flags
& __GFP_WAIT
)
2783 local_irq_disable();
2790 * Perform extra freeing checks:
2791 * - detect bad pointers.
2792 * - POISON/RED_ZONE checking
2794 static void kfree_debugcheck(const void *objp
)
2796 if (!virt_addr_valid(objp
)) {
2797 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2798 (unsigned long)objp
);
2803 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2805 unsigned long long redzone1
, redzone2
;
2807 redzone1
= *dbg_redzone1(cache
, obj
);
2808 redzone2
= *dbg_redzone2(cache
, obj
);
2813 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2816 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2817 slab_error(cache
, "double free detected");
2819 slab_error(cache
, "memory outside object was overwritten");
2821 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2822 obj
, redzone1
, redzone2
);
2825 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2826 unsigned long caller
)
2831 BUG_ON(virt_to_cache(objp
) != cachep
);
2833 objp
-= obj_offset(cachep
);
2834 kfree_debugcheck(objp
);
2835 page
= virt_to_head_page(objp
);
2837 if (cachep
->flags
& SLAB_RED_ZONE
) {
2838 verify_redzone_free(cachep
, objp
);
2839 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2840 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2842 if (cachep
->flags
& SLAB_STORE_USER
)
2843 *dbg_userword(cachep
, objp
) = (void *)caller
;
2845 objnr
= obj_to_index(cachep
, page
, objp
);
2847 BUG_ON(objnr
>= cachep
->num
);
2848 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2850 set_obj_status(page
, objnr
, OBJECT_FREE
);
2851 if (cachep
->flags
& SLAB_POISON
) {
2852 #ifdef CONFIG_DEBUG_PAGEALLOC
2853 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2854 store_stackinfo(cachep
, objp
, caller
);
2855 kernel_map_pages(virt_to_page(objp
),
2856 cachep
->size
/ PAGE_SIZE
, 0);
2858 poison_obj(cachep
, objp
, POISON_FREE
);
2861 poison_obj(cachep
, objp
, POISON_FREE
);
2868 #define kfree_debugcheck(x) do { } while(0)
2869 #define cache_free_debugcheck(x,objp,z) (objp)
2872 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2876 struct kmem_cache_node
*n
;
2877 struct array_cache
*ac
;
2881 node
= numa_mem_id();
2882 if (unlikely(force_refill
))
2885 ac
= cpu_cache_get(cachep
);
2886 batchcount
= ac
->batchcount
;
2887 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2889 * If there was little recent activity on this cache, then
2890 * perform only a partial refill. Otherwise we could generate
2893 batchcount
= BATCHREFILL_LIMIT
;
2895 n
= cachep
->node
[node
];
2897 BUG_ON(ac
->avail
> 0 || !n
);
2898 spin_lock(&n
->list_lock
);
2900 /* See if we can refill from the shared array */
2901 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2902 n
->shared
->touched
= 1;
2906 while (batchcount
> 0) {
2907 struct list_head
*entry
;
2909 /* Get slab alloc is to come from. */
2910 entry
= n
->slabs_partial
.next
;
2911 if (entry
== &n
->slabs_partial
) {
2912 n
->free_touched
= 1;
2913 entry
= n
->slabs_free
.next
;
2914 if (entry
== &n
->slabs_free
)
2918 page
= list_entry(entry
, struct page
, lru
);
2919 check_spinlock_acquired(cachep
);
2922 * The slab was either on partial or free list so
2923 * there must be at least one object available for
2926 BUG_ON(page
->active
>= cachep
->num
);
2928 while (page
->active
< cachep
->num
&& batchcount
--) {
2929 STATS_INC_ALLOCED(cachep
);
2930 STATS_INC_ACTIVE(cachep
);
2931 STATS_SET_HIGH(cachep
);
2933 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, page
,
2937 /* move slabp to correct slabp list: */
2938 list_del(&page
->lru
);
2939 if (page
->active
== cachep
->num
)
2940 list_add(&page
->list
, &n
->slabs_full
);
2942 list_add(&page
->list
, &n
->slabs_partial
);
2946 n
->free_objects
-= ac
->avail
;
2948 spin_unlock(&n
->list_lock
);
2950 if (unlikely(!ac
->avail
)) {
2953 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2955 /* cache_grow can reenable interrupts, then ac could change. */
2956 ac
= cpu_cache_get(cachep
);
2957 node
= numa_mem_id();
2959 /* no objects in sight? abort */
2960 if (!x
&& (ac
->avail
== 0 || force_refill
))
2963 if (!ac
->avail
) /* objects refilled by interrupt? */
2968 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2971 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2974 might_sleep_if(flags
& __GFP_WAIT
);
2976 kmem_flagcheck(cachep
, flags
);
2981 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2982 gfp_t flags
, void *objp
, unsigned long caller
)
2988 if (cachep
->flags
& SLAB_POISON
) {
2989 #ifdef CONFIG_DEBUG_PAGEALLOC
2990 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2991 kernel_map_pages(virt_to_page(objp
),
2992 cachep
->size
/ PAGE_SIZE
, 1);
2994 check_poison_obj(cachep
, objp
);
2996 check_poison_obj(cachep
, objp
);
2998 poison_obj(cachep
, objp
, POISON_INUSE
);
3000 if (cachep
->flags
& SLAB_STORE_USER
)
3001 *dbg_userword(cachep
, objp
) = (void *)caller
;
3003 if (cachep
->flags
& SLAB_RED_ZONE
) {
3004 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3005 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3006 slab_error(cachep
, "double free, or memory outside"
3007 " object was overwritten");
3009 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3010 objp
, *dbg_redzone1(cachep
, objp
),
3011 *dbg_redzone2(cachep
, objp
));
3013 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3014 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3017 page
= virt_to_head_page(objp
);
3018 set_obj_status(page
, obj_to_index(cachep
, page
, objp
), OBJECT_ACTIVE
);
3019 objp
+= obj_offset(cachep
);
3020 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3022 if (ARCH_SLAB_MINALIGN
&&
3023 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3024 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3025 objp
, (int)ARCH_SLAB_MINALIGN
);
3030 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3033 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3035 if (cachep
== kmem_cache
)
3038 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3041 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3044 struct array_cache
*ac
;
3045 bool force_refill
= false;
3049 ac
= cpu_cache_get(cachep
);
3050 if (likely(ac
->avail
)) {
3052 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3055 * Allow for the possibility all avail objects are not allowed
3056 * by the current flags
3059 STATS_INC_ALLOCHIT(cachep
);
3062 force_refill
= true;
3065 STATS_INC_ALLOCMISS(cachep
);
3066 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3068 * the 'ac' may be updated by cache_alloc_refill(),
3069 * and kmemleak_erase() requires its correct value.
3071 ac
= cpu_cache_get(cachep
);
3075 * To avoid a false negative, if an object that is in one of the
3076 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3077 * treat the array pointers as a reference to the object.
3080 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3086 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3088 * If we are in_interrupt, then process context, including cpusets and
3089 * mempolicy, may not apply and should not be used for allocation policy.
3091 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3093 int nid_alloc
, nid_here
;
3095 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3097 nid_alloc
= nid_here
= numa_mem_id();
3098 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3099 nid_alloc
= cpuset_slab_spread_node();
3100 else if (current
->mempolicy
)
3101 nid_alloc
= slab_node();
3102 if (nid_alloc
!= nid_here
)
3103 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3108 * Fallback function if there was no memory available and no objects on a
3109 * certain node and fall back is permitted. First we scan all the
3110 * available node for available objects. If that fails then we
3111 * perform an allocation without specifying a node. This allows the page
3112 * allocator to do its reclaim / fallback magic. We then insert the
3113 * slab into the proper nodelist and then allocate from it.
3115 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3117 struct zonelist
*zonelist
;
3121 enum zone_type high_zoneidx
= gfp_zone(flags
);
3124 unsigned int cpuset_mems_cookie
;
3126 if (flags
& __GFP_THISNODE
)
3129 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3132 cpuset_mems_cookie
= read_mems_allowed_begin();
3133 zonelist
= node_zonelist(slab_node(), flags
);
3137 * Look through allowed nodes for objects available
3138 * from existing per node queues.
3140 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3141 nid
= zone_to_nid(zone
);
3143 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3145 cache
->node
[nid
]->free_objects
) {
3146 obj
= ____cache_alloc_node(cache
,
3147 flags
| GFP_THISNODE
, nid
);
3155 * This allocation will be performed within the constraints
3156 * of the current cpuset / memory policy requirements.
3157 * We may trigger various forms of reclaim on the allowed
3158 * set and go into memory reserves if necessary.
3162 if (local_flags
& __GFP_WAIT
)
3164 kmem_flagcheck(cache
, flags
);
3165 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3166 if (local_flags
& __GFP_WAIT
)
3167 local_irq_disable();
3170 * Insert into the appropriate per node queues
3172 nid
= page_to_nid(page
);
3173 if (cache_grow(cache
, flags
, nid
, page
)) {
3174 obj
= ____cache_alloc_node(cache
,
3175 flags
| GFP_THISNODE
, nid
);
3178 * Another processor may allocate the
3179 * objects in the slab since we are
3180 * not holding any locks.
3184 /* cache_grow already freed obj */
3190 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3196 * A interface to enable slab creation on nodeid
3198 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3201 struct list_head
*entry
;
3203 struct kmem_cache_node
*n
;
3207 VM_BUG_ON(nodeid
> num_online_nodes());
3208 n
= cachep
->node
[nodeid
];
3213 spin_lock(&n
->list_lock
);
3214 entry
= n
->slabs_partial
.next
;
3215 if (entry
== &n
->slabs_partial
) {
3216 n
->free_touched
= 1;
3217 entry
= n
->slabs_free
.next
;
3218 if (entry
== &n
->slabs_free
)
3222 page
= list_entry(entry
, struct page
, lru
);
3223 check_spinlock_acquired_node(cachep
, nodeid
);
3225 STATS_INC_NODEALLOCS(cachep
);
3226 STATS_INC_ACTIVE(cachep
);
3227 STATS_SET_HIGH(cachep
);
3229 BUG_ON(page
->active
== cachep
->num
);
3231 obj
= slab_get_obj(cachep
, page
, nodeid
);
3233 /* move slabp to correct slabp list: */
3234 list_del(&page
->lru
);
3236 if (page
->active
== cachep
->num
)
3237 list_add(&page
->lru
, &n
->slabs_full
);
3239 list_add(&page
->lru
, &n
->slabs_partial
);
3241 spin_unlock(&n
->list_lock
);
3245 spin_unlock(&n
->list_lock
);
3246 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3250 return fallback_alloc(cachep
, flags
);
3256 static __always_inline
void *
3257 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3258 unsigned long caller
)
3260 unsigned long save_flags
;
3262 int slab_node
= numa_mem_id();
3264 flags
&= gfp_allowed_mask
;
3266 lockdep_trace_alloc(flags
);
3268 if (slab_should_failslab(cachep
, flags
))
3271 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3273 cache_alloc_debugcheck_before(cachep
, flags
);
3274 local_irq_save(save_flags
);
3276 if (nodeid
== NUMA_NO_NODE
)
3279 if (unlikely(!cachep
->node
[nodeid
])) {
3280 /* Node not bootstrapped yet */
3281 ptr
= fallback_alloc(cachep
, flags
);
3285 if (nodeid
== slab_node
) {
3287 * Use the locally cached objects if possible.
3288 * However ____cache_alloc does not allow fallback
3289 * to other nodes. It may fail while we still have
3290 * objects on other nodes available.
3292 ptr
= ____cache_alloc(cachep
, flags
);
3296 /* ___cache_alloc_node can fall back to other nodes */
3297 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3299 local_irq_restore(save_flags
);
3300 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3301 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3305 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3307 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3308 memset(ptr
, 0, cachep
->object_size
);
3313 static __always_inline
void *
3314 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3318 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3319 objp
= alternate_node_alloc(cache
, flags
);
3323 objp
= ____cache_alloc(cache
, flags
);
3326 * We may just have run out of memory on the local node.
3327 * ____cache_alloc_node() knows how to locate memory on other nodes
3330 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3337 static __always_inline
void *
3338 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3340 return ____cache_alloc(cachep
, flags
);
3343 #endif /* CONFIG_NUMA */
3345 static __always_inline
void *
3346 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3348 unsigned long save_flags
;
3351 flags
&= gfp_allowed_mask
;
3353 lockdep_trace_alloc(flags
);
3355 if (slab_should_failslab(cachep
, flags
))
3358 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3360 cache_alloc_debugcheck_before(cachep
, flags
);
3361 local_irq_save(save_flags
);
3362 objp
= __do_cache_alloc(cachep
, flags
);
3363 local_irq_restore(save_flags
);
3364 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3365 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3370 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3372 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3373 memset(objp
, 0, cachep
->object_size
);
3379 * Caller needs to acquire correct kmem_list's list_lock
3381 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3385 struct kmem_cache_node
*n
;
3387 for (i
= 0; i
< nr_objects
; i
++) {
3391 clear_obj_pfmemalloc(&objpp
[i
]);
3394 page
= virt_to_head_page(objp
);
3395 n
= cachep
->node
[node
];
3396 list_del(&page
->lru
);
3397 check_spinlock_acquired_node(cachep
, node
);
3398 slab_put_obj(cachep
, page
, objp
, node
);
3399 STATS_DEC_ACTIVE(cachep
);
3402 /* fixup slab chains */
3403 if (page
->active
== 0) {
3404 if (n
->free_objects
> n
->free_limit
) {
3405 n
->free_objects
-= cachep
->num
;
3406 /* No need to drop any previously held
3407 * lock here, even if we have a off-slab slab
3408 * descriptor it is guaranteed to come from
3409 * a different cache, refer to comments before
3412 slab_destroy(cachep
, page
);
3414 list_add(&page
->lru
, &n
->slabs_free
);
3417 /* Unconditionally move a slab to the end of the
3418 * partial list on free - maximum time for the
3419 * other objects to be freed, too.
3421 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3426 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3429 struct kmem_cache_node
*n
;
3430 int node
= numa_mem_id();
3432 batchcount
= ac
->batchcount
;
3434 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3437 n
= cachep
->node
[node
];
3438 spin_lock(&n
->list_lock
);
3440 struct array_cache
*shared_array
= n
->shared
;
3441 int max
= shared_array
->limit
- shared_array
->avail
;
3443 if (batchcount
> max
)
3445 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3446 ac
->entry
, sizeof(void *) * batchcount
);
3447 shared_array
->avail
+= batchcount
;
3452 free_block(cachep
, ac
->entry
, batchcount
, node
);
3457 struct list_head
*p
;
3459 p
= n
->slabs_free
.next
;
3460 while (p
!= &(n
->slabs_free
)) {
3463 page
= list_entry(p
, struct page
, lru
);
3464 BUG_ON(page
->active
);
3469 STATS_SET_FREEABLE(cachep
, i
);
3472 spin_unlock(&n
->list_lock
);
3473 ac
->avail
-= batchcount
;
3474 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3478 * Release an obj back to its cache. If the obj has a constructed state, it must
3479 * be in this state _before_ it is released. Called with disabled ints.
3481 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3482 unsigned long caller
)
3484 struct array_cache
*ac
= cpu_cache_get(cachep
);
3487 kmemleak_free_recursive(objp
, cachep
->flags
);
3488 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3490 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3493 * Skip calling cache_free_alien() when the platform is not numa.
3494 * This will avoid cache misses that happen while accessing slabp (which
3495 * is per page memory reference) to get nodeid. Instead use a global
3496 * variable to skip the call, which is mostly likely to be present in
3499 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3502 if (likely(ac
->avail
< ac
->limit
)) {
3503 STATS_INC_FREEHIT(cachep
);
3505 STATS_INC_FREEMISS(cachep
);
3506 cache_flusharray(cachep
, ac
);
3509 ac_put_obj(cachep
, ac
, objp
);
3513 * kmem_cache_alloc - Allocate an object
3514 * @cachep: The cache to allocate from.
3515 * @flags: See kmalloc().
3517 * Allocate an object from this cache. The flags are only relevant
3518 * if the cache has no available objects.
3520 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3522 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3524 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3525 cachep
->object_size
, cachep
->size
, flags
);
3529 EXPORT_SYMBOL(kmem_cache_alloc
);
3531 #ifdef CONFIG_TRACING
3533 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3537 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3539 trace_kmalloc(_RET_IP_
, ret
,
3540 size
, cachep
->size
, flags
);
3543 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3548 * kmem_cache_alloc_node - Allocate an object on the specified node
3549 * @cachep: The cache to allocate from.
3550 * @flags: See kmalloc().
3551 * @nodeid: node number of the target node.
3553 * Identical to kmem_cache_alloc but it will allocate memory on the given
3554 * node, which can improve the performance for cpu bound structures.
3556 * Fallback to other node is possible if __GFP_THISNODE is not set.
3558 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3560 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3562 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3563 cachep
->object_size
, cachep
->size
,
3568 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3570 #ifdef CONFIG_TRACING
3571 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3578 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3580 trace_kmalloc_node(_RET_IP_
, ret
,
3585 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3588 static __always_inline
void *
3589 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3591 struct kmem_cache
*cachep
;
3593 cachep
= kmalloc_slab(size
, flags
);
3594 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3596 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3599 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3600 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3602 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3604 EXPORT_SYMBOL(__kmalloc_node
);
3606 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3607 int node
, unsigned long caller
)
3609 return __do_kmalloc_node(size
, flags
, node
, caller
);
3611 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3613 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3615 return __do_kmalloc_node(size
, flags
, node
, 0);
3617 EXPORT_SYMBOL(__kmalloc_node
);
3618 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3619 #endif /* CONFIG_NUMA */
3622 * __do_kmalloc - allocate memory
3623 * @size: how many bytes of memory are required.
3624 * @flags: the type of memory to allocate (see kmalloc).
3625 * @caller: function caller for debug tracking of the caller
3627 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3628 unsigned long caller
)
3630 struct kmem_cache
*cachep
;
3633 /* If you want to save a few bytes .text space: replace
3635 * Then kmalloc uses the uninlined functions instead of the inline
3638 cachep
= kmalloc_slab(size
, flags
);
3639 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3641 ret
= slab_alloc(cachep
, flags
, caller
);
3643 trace_kmalloc(caller
, ret
,
3644 size
, cachep
->size
, flags
);
3650 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3651 void *__kmalloc(size_t size
, gfp_t flags
)
3653 return __do_kmalloc(size
, flags
, _RET_IP_
);
3655 EXPORT_SYMBOL(__kmalloc
);
3657 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3659 return __do_kmalloc(size
, flags
, caller
);
3661 EXPORT_SYMBOL(__kmalloc_track_caller
);
3664 void *__kmalloc(size_t size
, gfp_t flags
)
3666 return __do_kmalloc(size
, flags
, 0);
3668 EXPORT_SYMBOL(__kmalloc
);
3672 * kmem_cache_free - Deallocate an object
3673 * @cachep: The cache the allocation was from.
3674 * @objp: The previously allocated object.
3676 * Free an object which was previously allocated from this
3679 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3681 unsigned long flags
;
3682 cachep
= cache_from_obj(cachep
, objp
);
3686 local_irq_save(flags
);
3687 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3688 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3689 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3690 __cache_free(cachep
, objp
, _RET_IP_
);
3691 local_irq_restore(flags
);
3693 trace_kmem_cache_free(_RET_IP_
, objp
);
3695 EXPORT_SYMBOL(kmem_cache_free
);
3698 * kfree - free previously allocated memory
3699 * @objp: pointer returned by kmalloc.
3701 * If @objp is NULL, no operation is performed.
3703 * Don't free memory not originally allocated by kmalloc()
3704 * or you will run into trouble.
3706 void kfree(const void *objp
)
3708 struct kmem_cache
*c
;
3709 unsigned long flags
;
3711 trace_kfree(_RET_IP_
, objp
);
3713 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3715 local_irq_save(flags
);
3716 kfree_debugcheck(objp
);
3717 c
= virt_to_cache(objp
);
3718 debug_check_no_locks_freed(objp
, c
->object_size
);
3720 debug_check_no_obj_freed(objp
, c
->object_size
);
3721 __cache_free(c
, (void *)objp
, _RET_IP_
);
3722 local_irq_restore(flags
);
3724 EXPORT_SYMBOL(kfree
);
3727 * This initializes kmem_cache_node or resizes various caches for all nodes.
3729 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3732 struct kmem_cache_node
*n
;
3733 struct array_cache
*new_shared
;
3734 struct array_cache
**new_alien
= NULL
;
3736 for_each_online_node(node
) {
3738 if (use_alien_caches
) {
3739 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3745 if (cachep
->shared
) {
3746 new_shared
= alloc_arraycache(node
,
3747 cachep
->shared
*cachep
->batchcount
,
3750 free_alien_cache(new_alien
);
3755 n
= cachep
->node
[node
];
3757 struct array_cache
*shared
= n
->shared
;
3759 spin_lock_irq(&n
->list_lock
);
3762 free_block(cachep
, shared
->entry
,
3763 shared
->avail
, node
);
3765 n
->shared
= new_shared
;
3767 n
->alien
= new_alien
;
3770 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3771 cachep
->batchcount
+ cachep
->num
;
3772 spin_unlock_irq(&n
->list_lock
);
3774 free_alien_cache(new_alien
);
3777 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3779 free_alien_cache(new_alien
);
3784 kmem_cache_node_init(n
);
3785 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3786 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3787 n
->shared
= new_shared
;
3788 n
->alien
= new_alien
;
3789 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3790 cachep
->batchcount
+ cachep
->num
;
3791 cachep
->node
[node
] = n
;
3796 if (!cachep
->list
.next
) {
3797 /* Cache is not active yet. Roll back what we did */
3800 if (cachep
->node
[node
]) {
3801 n
= cachep
->node
[node
];
3804 free_alien_cache(n
->alien
);
3806 cachep
->node
[node
] = NULL
;
3814 struct ccupdate_struct
{
3815 struct kmem_cache
*cachep
;
3816 struct array_cache
*new[0];
3819 static void do_ccupdate_local(void *info
)
3821 struct ccupdate_struct
*new = info
;
3822 struct array_cache
*old
;
3825 old
= cpu_cache_get(new->cachep
);
3827 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3828 new->new[smp_processor_id()] = old
;
3831 /* Always called with the slab_mutex held */
3832 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3833 int batchcount
, int shared
, gfp_t gfp
)
3835 struct ccupdate_struct
*new;
3838 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3843 for_each_online_cpu(i
) {
3844 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3847 for (i
--; i
>= 0; i
--)
3853 new->cachep
= cachep
;
3855 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3858 cachep
->batchcount
= batchcount
;
3859 cachep
->limit
= limit
;
3860 cachep
->shared
= shared
;
3862 for_each_online_cpu(i
) {
3863 struct array_cache
*ccold
= new->new[i
];
3866 spin_lock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3867 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
3868 spin_unlock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3872 return alloc_kmemlist(cachep
, gfp
);
3875 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3876 int batchcount
, int shared
, gfp_t gfp
)
3879 struct kmem_cache
*c
= NULL
;
3882 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3884 if (slab_state
< FULL
)
3887 if ((ret
< 0) || !is_root_cache(cachep
))
3890 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3891 for_each_memcg_cache_index(i
) {
3892 c
= cache_from_memcg_idx(cachep
, i
);
3894 /* return value determined by the parent cache only */
3895 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3901 /* Called with slab_mutex held always */
3902 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3909 if (!is_root_cache(cachep
)) {
3910 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3911 limit
= root
->limit
;
3912 shared
= root
->shared
;
3913 batchcount
= root
->batchcount
;
3916 if (limit
&& shared
&& batchcount
)
3919 * The head array serves three purposes:
3920 * - create a LIFO ordering, i.e. return objects that are cache-warm
3921 * - reduce the number of spinlock operations.
3922 * - reduce the number of linked list operations on the slab and
3923 * bufctl chains: array operations are cheaper.
3924 * The numbers are guessed, we should auto-tune as described by
3927 if (cachep
->size
> 131072)
3929 else if (cachep
->size
> PAGE_SIZE
)
3931 else if (cachep
->size
> 1024)
3933 else if (cachep
->size
> 256)
3939 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3940 * allocation behaviour: Most allocs on one cpu, most free operations
3941 * on another cpu. For these cases, an efficient object passing between
3942 * cpus is necessary. This is provided by a shared array. The array
3943 * replaces Bonwick's magazine layer.
3944 * On uniprocessor, it's functionally equivalent (but less efficient)
3945 * to a larger limit. Thus disabled by default.
3948 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3953 * With debugging enabled, large batchcount lead to excessively long
3954 * periods with disabled local interrupts. Limit the batchcount
3959 batchcount
= (limit
+ 1) / 2;
3961 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3963 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3964 cachep
->name
, -err
);
3969 * Drain an array if it contains any elements taking the node lock only if
3970 * necessary. Note that the node listlock also protects the array_cache
3971 * if drain_array() is used on the shared array.
3973 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3974 struct array_cache
*ac
, int force
, int node
)
3978 if (!ac
|| !ac
->avail
)
3980 if (ac
->touched
&& !force
) {
3983 spin_lock_irq(&n
->list_lock
);
3985 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3986 if (tofree
> ac
->avail
)
3987 tofree
= (ac
->avail
+ 1) / 2;
3988 free_block(cachep
, ac
->entry
, tofree
, node
);
3989 ac
->avail
-= tofree
;
3990 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3991 sizeof(void *) * ac
->avail
);
3993 spin_unlock_irq(&n
->list_lock
);
3998 * cache_reap - Reclaim memory from caches.
3999 * @w: work descriptor
4001 * Called from workqueue/eventd every few seconds.
4003 * - clear the per-cpu caches for this CPU.
4004 * - return freeable pages to the main free memory pool.
4006 * If we cannot acquire the cache chain mutex then just give up - we'll try
4007 * again on the next iteration.
4009 static void cache_reap(struct work_struct
*w
)
4011 struct kmem_cache
*searchp
;
4012 struct kmem_cache_node
*n
;
4013 int node
= numa_mem_id();
4014 struct delayed_work
*work
= to_delayed_work(w
);
4016 if (!mutex_trylock(&slab_mutex
))
4017 /* Give up. Setup the next iteration. */
4020 list_for_each_entry(searchp
, &slab_caches
, list
) {
4024 * We only take the node lock if absolutely necessary and we
4025 * have established with reasonable certainty that
4026 * we can do some work if the lock was obtained.
4028 n
= searchp
->node
[node
];
4030 reap_alien(searchp
, n
);
4032 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
4035 * These are racy checks but it does not matter
4036 * if we skip one check or scan twice.
4038 if (time_after(n
->next_reap
, jiffies
))
4041 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4043 drain_array(searchp
, n
, n
->shared
, 0, node
);
4045 if (n
->free_touched
)
4046 n
->free_touched
= 0;
4050 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4051 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4052 STATS_ADD_REAPED(searchp
, freed
);
4058 mutex_unlock(&slab_mutex
);
4061 /* Set up the next iteration */
4062 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4065 #ifdef CONFIG_SLABINFO
4066 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4069 unsigned long active_objs
;
4070 unsigned long num_objs
;
4071 unsigned long active_slabs
= 0;
4072 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4076 struct kmem_cache_node
*n
;
4080 for_each_online_node(node
) {
4081 n
= cachep
->node
[node
];
4086 spin_lock_irq(&n
->list_lock
);
4088 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4089 if (page
->active
!= cachep
->num
&& !error
)
4090 error
= "slabs_full accounting error";
4091 active_objs
+= cachep
->num
;
4094 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4095 if (page
->active
== cachep
->num
&& !error
)
4096 error
= "slabs_partial accounting error";
4097 if (!page
->active
&& !error
)
4098 error
= "slabs_partial accounting error";
4099 active_objs
+= page
->active
;
4102 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4103 if (page
->active
&& !error
)
4104 error
= "slabs_free accounting error";
4107 free_objects
+= n
->free_objects
;
4109 shared_avail
+= n
->shared
->avail
;
4111 spin_unlock_irq(&n
->list_lock
);
4113 num_slabs
+= active_slabs
;
4114 num_objs
= num_slabs
* cachep
->num
;
4115 if (num_objs
- active_objs
!= free_objects
&& !error
)
4116 error
= "free_objects accounting error";
4118 name
= cachep
->name
;
4120 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4122 sinfo
->active_objs
= active_objs
;
4123 sinfo
->num_objs
= num_objs
;
4124 sinfo
->active_slabs
= active_slabs
;
4125 sinfo
->num_slabs
= num_slabs
;
4126 sinfo
->shared_avail
= shared_avail
;
4127 sinfo
->limit
= cachep
->limit
;
4128 sinfo
->batchcount
= cachep
->batchcount
;
4129 sinfo
->shared
= cachep
->shared
;
4130 sinfo
->objects_per_slab
= cachep
->num
;
4131 sinfo
->cache_order
= cachep
->gfporder
;
4134 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4138 unsigned long high
= cachep
->high_mark
;
4139 unsigned long allocs
= cachep
->num_allocations
;
4140 unsigned long grown
= cachep
->grown
;
4141 unsigned long reaped
= cachep
->reaped
;
4142 unsigned long errors
= cachep
->errors
;
4143 unsigned long max_freeable
= cachep
->max_freeable
;
4144 unsigned long node_allocs
= cachep
->node_allocs
;
4145 unsigned long node_frees
= cachep
->node_frees
;
4146 unsigned long overflows
= cachep
->node_overflow
;
4148 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4149 "%4lu %4lu %4lu %4lu %4lu",
4150 allocs
, high
, grown
,
4151 reaped
, errors
, max_freeable
, node_allocs
,
4152 node_frees
, overflows
);
4156 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4157 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4158 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4159 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4161 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4162 allochit
, allocmiss
, freehit
, freemiss
);
4167 #define MAX_SLABINFO_WRITE 128
4169 * slabinfo_write - Tuning for the slab allocator
4171 * @buffer: user buffer
4172 * @count: data length
4175 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4176 size_t count
, loff_t
*ppos
)
4178 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4179 int limit
, batchcount
, shared
, res
;
4180 struct kmem_cache
*cachep
;
4182 if (count
> MAX_SLABINFO_WRITE
)
4184 if (copy_from_user(&kbuf
, buffer
, count
))
4186 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4188 tmp
= strchr(kbuf
, ' ');
4193 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4196 /* Find the cache in the chain of caches. */
4197 mutex_lock(&slab_mutex
);
4199 list_for_each_entry(cachep
, &slab_caches
, list
) {
4200 if (!strcmp(cachep
->name
, kbuf
)) {
4201 if (limit
< 1 || batchcount
< 1 ||
4202 batchcount
> limit
|| shared
< 0) {
4205 res
= do_tune_cpucache(cachep
, limit
,
4212 mutex_unlock(&slab_mutex
);
4218 #ifdef CONFIG_DEBUG_SLAB_LEAK
4220 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4222 mutex_lock(&slab_mutex
);
4223 return seq_list_start(&slab_caches
, *pos
);
4226 static inline int add_caller(unsigned long *n
, unsigned long v
)
4236 unsigned long *q
= p
+ 2 * i
;
4250 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4256 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4264 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4265 if (get_obj_status(page
, i
) != OBJECT_ACTIVE
)
4268 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4273 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4275 #ifdef CONFIG_KALLSYMS
4276 unsigned long offset
, size
;
4277 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4279 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4280 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4282 seq_printf(m
, " [%s]", modname
);
4286 seq_printf(m
, "%p", (void *)address
);
4289 static int leaks_show(struct seq_file
*m
, void *p
)
4291 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4293 struct kmem_cache_node
*n
;
4295 unsigned long *x
= m
->private;
4299 if (!(cachep
->flags
& SLAB_STORE_USER
))
4301 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4304 /* OK, we can do it */
4308 for_each_online_node(node
) {
4309 n
= cachep
->node
[node
];
4314 spin_lock_irq(&n
->list_lock
);
4316 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4317 handle_slab(x
, cachep
, page
);
4318 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4319 handle_slab(x
, cachep
, page
);
4320 spin_unlock_irq(&n
->list_lock
);
4322 name
= cachep
->name
;
4324 /* Increase the buffer size */
4325 mutex_unlock(&slab_mutex
);
4326 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4328 /* Too bad, we are really out */
4330 mutex_lock(&slab_mutex
);
4333 *(unsigned long *)m
->private = x
[0] * 2;
4335 mutex_lock(&slab_mutex
);
4336 /* Now make sure this entry will be retried */
4340 for (i
= 0; i
< x
[1]; i
++) {
4341 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4342 show_symbol(m
, x
[2*i
+2]);
4349 static const struct seq_operations slabstats_op
= {
4350 .start
= leaks_start
,
4356 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4358 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4361 ret
= seq_open(file
, &slabstats_op
);
4363 struct seq_file
*m
= file
->private_data
;
4364 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4373 static const struct file_operations proc_slabstats_operations
= {
4374 .open
= slabstats_open
,
4376 .llseek
= seq_lseek
,
4377 .release
= seq_release_private
,
4381 static int __init
slab_proc_init(void)
4383 #ifdef CONFIG_DEBUG_SLAB_LEAK
4384 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4388 module_init(slab_proc_init
);
4392 * ksize - get the actual amount of memory allocated for a given object
4393 * @objp: Pointer to the object
4395 * kmalloc may internally round up allocations and return more memory
4396 * than requested. ksize() can be used to determine the actual amount of
4397 * memory allocated. The caller may use this additional memory, even though
4398 * a smaller amount of memory was initially specified with the kmalloc call.
4399 * The caller must guarantee that objp points to a valid object previously
4400 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4401 * must not be freed during the duration of the call.
4403 size_t ksize(const void *objp
)
4406 if (unlikely(objp
== ZERO_SIZE_PTR
))
4409 return virt_to_cache(objp
)->object_size
;
4411 EXPORT_SYMBOL(ksize
);