3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t
;
166 typedef unsigned short freelist_idx_t
;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
172 * true if a page was allocated from pfmemalloc reserves for network-based
175 static bool pfmemalloc_active __read_mostly
;
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
185 * The limit is stored in the per-cpu structure to reduce the data cache
192 unsigned int batchcount
;
193 unsigned int touched
;
196 * Must have this definition in here for the proper
197 * alignment of array_cache. Also simplifies accessing
200 * Entries should not be directly dereferenced as
201 * entries belonging to slabs marked pfmemalloc will
202 * have the lower bits set SLAB_OBJ_PFMEMALLOC
206 #define SLAB_OBJ_PFMEMALLOC 1
207 static inline bool is_obj_pfmemalloc(void *objp
)
209 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
212 static inline void set_obj_pfmemalloc(void **objp
)
214 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
218 static inline void clear_obj_pfmemalloc(void **objp
)
220 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
224 * bootstrap: The caches do not work without cpuarrays anymore, but the
225 * cpuarrays are allocated from the generic caches...
227 #define BOOT_CPUCACHE_ENTRIES 1
228 struct arraycache_init
{
229 struct array_cache cache
;
230 void *entries
[BOOT_CPUCACHE_ENTRIES
];
234 * Need this for bootstrapping a per node allocator.
236 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
237 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
238 #define CACHE_CACHE 0
239 #define SIZE_AC MAX_NUMNODES
240 #define SIZE_NODE (2 * MAX_NUMNODES)
242 static int drain_freelist(struct kmem_cache
*cache
,
243 struct kmem_cache_node
*n
, int tofree
);
244 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
246 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
247 static void cache_reap(struct work_struct
*unused
);
249 static int slab_early_init
= 1;
251 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
252 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
254 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
256 INIT_LIST_HEAD(&parent
->slabs_full
);
257 INIT_LIST_HEAD(&parent
->slabs_partial
);
258 INIT_LIST_HEAD(&parent
->slabs_free
);
259 parent
->shared
= NULL
;
260 parent
->alien
= NULL
;
261 parent
->colour_next
= 0;
262 spin_lock_init(&parent
->list_lock
);
263 parent
->free_objects
= 0;
264 parent
->free_touched
= 0;
267 #define MAKE_LIST(cachep, listp, slab, nodeid) \
269 INIT_LIST_HEAD(listp); \
270 list_splice(&(cachep->node[nodeid]->slab), listp); \
273 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
275 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
276 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
277 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
280 #define CFLGS_OFF_SLAB (0x80000000UL)
281 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
283 #define BATCHREFILL_LIMIT 16
285 * Optimization question: fewer reaps means less probability for unnessary
286 * cpucache drain/refill cycles.
288 * OTOH the cpuarrays can contain lots of objects,
289 * which could lock up otherwise freeable slabs.
291 #define REAPTIMEOUT_AC (2*HZ)
292 #define REAPTIMEOUT_NODE (4*HZ)
295 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
296 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
297 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
298 #define STATS_INC_GROWN(x) ((x)->grown++)
299 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
300 #define STATS_SET_HIGH(x) \
302 if ((x)->num_active > (x)->high_mark) \
303 (x)->high_mark = (x)->num_active; \
305 #define STATS_INC_ERR(x) ((x)->errors++)
306 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
307 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
308 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
309 #define STATS_SET_FREEABLE(x, i) \
311 if ((x)->max_freeable < i) \
312 (x)->max_freeable = i; \
314 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
315 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
316 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
317 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
319 #define STATS_INC_ACTIVE(x) do { } while (0)
320 #define STATS_DEC_ACTIVE(x) do { } while (0)
321 #define STATS_INC_ALLOCED(x) do { } while (0)
322 #define STATS_INC_GROWN(x) do { } while (0)
323 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
324 #define STATS_SET_HIGH(x) do { } while (0)
325 #define STATS_INC_ERR(x) do { } while (0)
326 #define STATS_INC_NODEALLOCS(x) do { } while (0)
327 #define STATS_INC_NODEFREES(x) do { } while (0)
328 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
329 #define STATS_SET_FREEABLE(x, i) do { } while (0)
330 #define STATS_INC_ALLOCHIT(x) do { } while (0)
331 #define STATS_INC_ALLOCMISS(x) do { } while (0)
332 #define STATS_INC_FREEHIT(x) do { } while (0)
333 #define STATS_INC_FREEMISS(x) do { } while (0)
339 * memory layout of objects:
341 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
342 * the end of an object is aligned with the end of the real
343 * allocation. Catches writes behind the end of the allocation.
344 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
346 * cachep->obj_offset: The real object.
347 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
348 * cachep->size - 1* BYTES_PER_WORD: last caller address
349 * [BYTES_PER_WORD long]
351 static int obj_offset(struct kmem_cache
*cachep
)
353 return cachep
->obj_offset
;
356 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
358 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
359 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
360 sizeof(unsigned long long));
363 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
365 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
366 if (cachep
->flags
& SLAB_STORE_USER
)
367 return (unsigned long long *)(objp
+ cachep
->size
-
368 sizeof(unsigned long long) -
370 return (unsigned long long *) (objp
+ cachep
->size
-
371 sizeof(unsigned long long));
374 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
376 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
377 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
382 #define obj_offset(x) 0
383 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
384 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
385 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
390 * Do not go above this order unless 0 objects fit into the slab or
391 * overridden on the command line.
393 #define SLAB_MAX_ORDER_HI 1
394 #define SLAB_MAX_ORDER_LO 0
395 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
396 static bool slab_max_order_set __initdata
;
398 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
400 struct page
*page
= virt_to_head_page(obj
);
401 return page
->slab_cache
;
404 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
407 return page
->s_mem
+ cache
->size
* idx
;
411 * We want to avoid an expensive divide : (offset / cache->size)
412 * Using the fact that size is a constant for a particular cache,
413 * we can replace (offset / cache->size) by
414 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
416 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
417 const struct page
*page
, void *obj
)
419 u32 offset
= (obj
- page
->s_mem
);
420 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
423 static struct arraycache_init initarray_generic
=
424 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
426 /* internal cache of cache description objs */
427 static struct kmem_cache kmem_cache_boot
= {
429 .limit
= BOOT_CPUCACHE_ENTRIES
,
431 .size
= sizeof(struct kmem_cache
),
432 .name
= "kmem_cache",
435 #define BAD_ALIEN_MAGIC 0x01020304ul
437 #ifdef CONFIG_LOCKDEP
440 * Slab sometimes uses the kmalloc slabs to store the slab headers
441 * for other slabs "off slab".
442 * The locking for this is tricky in that it nests within the locks
443 * of all other slabs in a few places; to deal with this special
444 * locking we put on-slab caches into a separate lock-class.
446 * We set lock class for alien array caches which are up during init.
447 * The lock annotation will be lost if all cpus of a node goes down and
448 * then comes back up during hotplug
450 static struct lock_class_key on_slab_l3_key
;
451 static struct lock_class_key on_slab_alc_key
;
453 static struct lock_class_key debugobj_l3_key
;
454 static struct lock_class_key debugobj_alc_key
;
456 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
457 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
460 struct array_cache
**alc
;
461 struct kmem_cache_node
*n
;
468 lockdep_set_class(&n
->list_lock
, l3_key
);
471 * FIXME: This check for BAD_ALIEN_MAGIC
472 * should go away when common slab code is taught to
473 * work even without alien caches.
474 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
475 * for alloc_alien_cache,
477 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
481 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
485 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
487 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
490 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
494 for_each_online_node(node
)
495 slab_set_debugobj_lock_classes_node(cachep
, node
);
498 static void init_node_lock_keys(int q
)
505 for (i
= 1; i
<= KMALLOC_SHIFT_HIGH
; i
++) {
506 struct kmem_cache_node
*n
;
507 struct kmem_cache
*cache
= kmalloc_caches
[i
];
513 if (!n
|| OFF_SLAB(cache
))
516 slab_set_lock_classes(cache
, &on_slab_l3_key
,
517 &on_slab_alc_key
, q
);
521 static void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int q
)
523 if (!cachep
->node
[q
])
526 slab_set_lock_classes(cachep
, &on_slab_l3_key
,
527 &on_slab_alc_key
, q
);
530 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
534 VM_BUG_ON(OFF_SLAB(cachep
));
536 on_slab_lock_classes_node(cachep
, node
);
539 static inline void init_lock_keys(void)
544 init_node_lock_keys(node
);
547 static void init_node_lock_keys(int q
)
551 static inline void init_lock_keys(void)
555 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
559 static inline void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int node
)
563 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
567 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
572 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
574 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
576 return cachep
->array
[smp_processor_id()];
579 static int calculate_nr_objs(size_t slab_size
, size_t buffer_size
,
580 size_t idx_size
, size_t align
)
583 size_t freelist_size
;
586 * Ignore padding for the initial guess. The padding
587 * is at most @align-1 bytes, and @buffer_size is at
588 * least @align. In the worst case, this result will
589 * be one greater than the number of objects that fit
590 * into the memory allocation when taking the padding
593 nr_objs
= slab_size
/ (buffer_size
+ idx_size
);
596 * This calculated number will be either the right
597 * amount, or one greater than what we want.
599 freelist_size
= slab_size
- nr_objs
* buffer_size
;
600 if (freelist_size
< ALIGN(nr_objs
* idx_size
, align
))
607 * Calculate the number of objects and left-over bytes for a given buffer size.
609 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
610 size_t align
, int flags
, size_t *left_over
,
615 size_t slab_size
= PAGE_SIZE
<< gfporder
;
618 * The slab management structure can be either off the slab or
619 * on it. For the latter case, the memory allocated for a
622 * - One unsigned int for each object
623 * - Padding to respect alignment of @align
624 * - @buffer_size bytes for each object
626 * If the slab management structure is off the slab, then the
627 * alignment will already be calculated into the size. Because
628 * the slabs are all pages aligned, the objects will be at the
629 * correct alignment when allocated.
631 if (flags
& CFLGS_OFF_SLAB
) {
633 nr_objs
= slab_size
/ buffer_size
;
636 nr_objs
= calculate_nr_objs(slab_size
, buffer_size
,
637 sizeof(freelist_idx_t
), align
);
638 mgmt_size
= ALIGN(nr_objs
* sizeof(freelist_idx_t
), align
);
641 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
645 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
647 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
650 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
651 function
, cachep
->name
, msg
);
653 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
658 * By default on NUMA we use alien caches to stage the freeing of
659 * objects allocated from other nodes. This causes massive memory
660 * inefficiencies when using fake NUMA setup to split memory into a
661 * large number of small nodes, so it can be disabled on the command
665 static int use_alien_caches __read_mostly
= 1;
666 static int __init
noaliencache_setup(char *s
)
668 use_alien_caches
= 0;
671 __setup("noaliencache", noaliencache_setup
);
673 static int __init
slab_max_order_setup(char *str
)
675 get_option(&str
, &slab_max_order
);
676 slab_max_order
= slab_max_order
< 0 ? 0 :
677 min(slab_max_order
, MAX_ORDER
- 1);
678 slab_max_order_set
= true;
682 __setup("slab_max_order=", slab_max_order_setup
);
686 * Special reaping functions for NUMA systems called from cache_reap().
687 * These take care of doing round robin flushing of alien caches (containing
688 * objects freed on different nodes from which they were allocated) and the
689 * flushing of remote pcps by calling drain_node_pages.
691 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
693 static void init_reap_node(int cpu
)
697 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
698 if (node
== MAX_NUMNODES
)
699 node
= first_node(node_online_map
);
701 per_cpu(slab_reap_node
, cpu
) = node
;
704 static void next_reap_node(void)
706 int node
= __this_cpu_read(slab_reap_node
);
708 node
= next_node(node
, node_online_map
);
709 if (unlikely(node
>= MAX_NUMNODES
))
710 node
= first_node(node_online_map
);
711 __this_cpu_write(slab_reap_node
, node
);
715 #define init_reap_node(cpu) do { } while (0)
716 #define next_reap_node(void) do { } while (0)
720 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
721 * via the workqueue/eventd.
722 * Add the CPU number into the expiration time to minimize the possibility of
723 * the CPUs getting into lockstep and contending for the global cache chain
726 static void start_cpu_timer(int cpu
)
728 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
731 * When this gets called from do_initcalls via cpucache_init(),
732 * init_workqueues() has already run, so keventd will be setup
735 if (keventd_up() && reap_work
->work
.func
== NULL
) {
737 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
738 schedule_delayed_work_on(cpu
, reap_work
,
739 __round_jiffies_relative(HZ
, cpu
));
743 static struct array_cache
*alloc_arraycache(int node
, int entries
,
744 int batchcount
, gfp_t gfp
)
746 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
747 struct array_cache
*nc
= NULL
;
749 nc
= kmalloc_node(memsize
, gfp
, node
);
751 * The array_cache structures contain pointers to free object.
752 * However, when such objects are allocated or transferred to another
753 * cache the pointers are not cleared and they could be counted as
754 * valid references during a kmemleak scan. Therefore, kmemleak must
755 * not scan such objects.
757 kmemleak_no_scan(nc
);
761 nc
->batchcount
= batchcount
;
763 spin_lock_init(&nc
->lock
);
768 static inline bool is_slab_pfmemalloc(struct page
*page
)
770 return PageSlabPfmemalloc(page
);
773 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
774 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
775 struct array_cache
*ac
)
777 struct kmem_cache_node
*n
= cachep
->node
[numa_mem_id()];
781 if (!pfmemalloc_active
)
784 spin_lock_irqsave(&n
->list_lock
, flags
);
785 list_for_each_entry(page
, &n
->slabs_full
, lru
)
786 if (is_slab_pfmemalloc(page
))
789 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
790 if (is_slab_pfmemalloc(page
))
793 list_for_each_entry(page
, &n
->slabs_free
, lru
)
794 if (is_slab_pfmemalloc(page
))
797 pfmemalloc_active
= false;
799 spin_unlock_irqrestore(&n
->list_lock
, flags
);
802 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
803 gfp_t flags
, bool force_refill
)
806 void *objp
= ac
->entry
[--ac
->avail
];
808 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
809 if (unlikely(is_obj_pfmemalloc(objp
))) {
810 struct kmem_cache_node
*n
;
812 if (gfp_pfmemalloc_allowed(flags
)) {
813 clear_obj_pfmemalloc(&objp
);
817 /* The caller cannot use PFMEMALLOC objects, find another one */
818 for (i
= 0; i
< ac
->avail
; i
++) {
819 /* If a !PFMEMALLOC object is found, swap them */
820 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
822 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
823 ac
->entry
[ac
->avail
] = objp
;
829 * If there are empty slabs on the slabs_free list and we are
830 * being forced to refill the cache, mark this one !pfmemalloc.
832 n
= cachep
->node
[numa_mem_id()];
833 if (!list_empty(&n
->slabs_free
) && force_refill
) {
834 struct page
*page
= virt_to_head_page(objp
);
835 ClearPageSlabPfmemalloc(page
);
836 clear_obj_pfmemalloc(&objp
);
837 recheck_pfmemalloc_active(cachep
, ac
);
841 /* No !PFMEMALLOC objects available */
849 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
850 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
854 if (unlikely(sk_memalloc_socks()))
855 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
857 objp
= ac
->entry
[--ac
->avail
];
862 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
865 if (unlikely(pfmemalloc_active
)) {
866 /* Some pfmemalloc slabs exist, check if this is one */
867 struct page
*page
= virt_to_head_page(objp
);
868 if (PageSlabPfmemalloc(page
))
869 set_obj_pfmemalloc(&objp
);
875 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
878 if (unlikely(sk_memalloc_socks()))
879 objp
= __ac_put_obj(cachep
, ac
, objp
);
881 ac
->entry
[ac
->avail
++] = objp
;
885 * Transfer objects in one arraycache to another.
886 * Locking must be handled by the caller.
888 * Return the number of entries transferred.
890 static int transfer_objects(struct array_cache
*to
,
891 struct array_cache
*from
, unsigned int max
)
893 /* Figure out how many entries to transfer */
894 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
899 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
909 #define drain_alien_cache(cachep, alien) do { } while (0)
910 #define reap_alien(cachep, n) do { } while (0)
912 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
914 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
917 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
921 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
926 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
932 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
933 gfp_t flags
, int nodeid
)
938 #else /* CONFIG_NUMA */
940 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
941 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
943 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
945 struct array_cache
**ac_ptr
;
946 int memsize
= sizeof(void *) * nr_node_ids
;
951 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
954 if (i
== node
|| !node_online(i
))
956 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
958 for (i
--; i
>= 0; i
--)
968 static void free_alien_cache(struct array_cache
**ac_ptr
)
979 static void __drain_alien_cache(struct kmem_cache
*cachep
,
980 struct array_cache
*ac
, int node
)
982 struct kmem_cache_node
*n
= cachep
->node
[node
];
985 spin_lock(&n
->list_lock
);
987 * Stuff objects into the remote nodes shared array first.
988 * That way we could avoid the overhead of putting the objects
989 * into the free lists and getting them back later.
992 transfer_objects(n
->shared
, ac
, ac
->limit
);
994 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
996 spin_unlock(&n
->list_lock
);
1001 * Called from cache_reap() to regularly drain alien caches round robin.
1003 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
1005 int node
= __this_cpu_read(slab_reap_node
);
1008 struct array_cache
*ac
= n
->alien
[node
];
1010 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1011 __drain_alien_cache(cachep
, ac
, node
);
1012 spin_unlock_irq(&ac
->lock
);
1017 static void drain_alien_cache(struct kmem_cache
*cachep
,
1018 struct array_cache
**alien
)
1021 struct array_cache
*ac
;
1022 unsigned long flags
;
1024 for_each_online_node(i
) {
1027 spin_lock_irqsave(&ac
->lock
, flags
);
1028 __drain_alien_cache(cachep
, ac
, i
);
1029 spin_unlock_irqrestore(&ac
->lock
, flags
);
1034 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1036 int nodeid
= page_to_nid(virt_to_page(objp
));
1037 struct kmem_cache_node
*n
;
1038 struct array_cache
*alien
= NULL
;
1041 node
= numa_mem_id();
1044 * Make sure we are not freeing a object from another node to the array
1045 * cache on this cpu.
1047 if (likely(nodeid
== node
))
1050 n
= cachep
->node
[node
];
1051 STATS_INC_NODEFREES(cachep
);
1052 if (n
->alien
&& n
->alien
[nodeid
]) {
1053 alien
= n
->alien
[nodeid
];
1054 spin_lock(&alien
->lock
);
1055 if (unlikely(alien
->avail
== alien
->limit
)) {
1056 STATS_INC_ACOVERFLOW(cachep
);
1057 __drain_alien_cache(cachep
, alien
, nodeid
);
1059 ac_put_obj(cachep
, alien
, objp
);
1060 spin_unlock(&alien
->lock
);
1062 spin_lock(&(cachep
->node
[nodeid
])->list_lock
);
1063 free_block(cachep
, &objp
, 1, nodeid
);
1064 spin_unlock(&(cachep
->node
[nodeid
])->list_lock
);
1071 * Allocates and initializes node for a node on each slab cache, used for
1072 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1073 * will be allocated off-node since memory is not yet online for the new node.
1074 * When hotplugging memory or a cpu, existing node are not replaced if
1077 * Must hold slab_mutex.
1079 static int init_cache_node_node(int node
)
1081 struct kmem_cache
*cachep
;
1082 struct kmem_cache_node
*n
;
1083 const int memsize
= sizeof(struct kmem_cache_node
);
1085 list_for_each_entry(cachep
, &slab_caches
, list
) {
1087 * Set up the kmem_cache_node for cpu before we can
1088 * begin anything. Make sure some other cpu on this
1089 * node has not already allocated this
1091 if (!cachep
->node
[node
]) {
1092 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1095 kmem_cache_node_init(n
);
1096 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
1097 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1100 * The kmem_cache_nodes don't come and go as CPUs
1101 * come and go. slab_mutex is sufficient
1104 cachep
->node
[node
] = n
;
1107 spin_lock_irq(&cachep
->node
[node
]->list_lock
);
1108 cachep
->node
[node
]->free_limit
=
1109 (1 + nr_cpus_node(node
)) *
1110 cachep
->batchcount
+ cachep
->num
;
1111 spin_unlock_irq(&cachep
->node
[node
]->list_lock
);
1116 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1117 struct kmem_cache_node
*n
)
1119 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1122 static void cpuup_canceled(long cpu
)
1124 struct kmem_cache
*cachep
;
1125 struct kmem_cache_node
*n
= NULL
;
1126 int node
= cpu_to_mem(cpu
);
1127 const struct cpumask
*mask
= cpumask_of_node(node
);
1129 list_for_each_entry(cachep
, &slab_caches
, list
) {
1130 struct array_cache
*nc
;
1131 struct array_cache
*shared
;
1132 struct array_cache
**alien
;
1134 /* cpu is dead; no one can alloc from it. */
1135 nc
= cachep
->array
[cpu
];
1136 cachep
->array
[cpu
] = NULL
;
1137 n
= cachep
->node
[node
];
1140 goto free_array_cache
;
1142 spin_lock_irq(&n
->list_lock
);
1144 /* Free limit for this kmem_cache_node */
1145 n
->free_limit
-= cachep
->batchcount
;
1147 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1149 if (!cpumask_empty(mask
)) {
1150 spin_unlock_irq(&n
->list_lock
);
1151 goto free_array_cache
;
1156 free_block(cachep
, shared
->entry
,
1157 shared
->avail
, node
);
1164 spin_unlock_irq(&n
->list_lock
);
1168 drain_alien_cache(cachep
, alien
);
1169 free_alien_cache(alien
);
1175 * In the previous loop, all the objects were freed to
1176 * the respective cache's slabs, now we can go ahead and
1177 * shrink each nodelist to its limit.
1179 list_for_each_entry(cachep
, &slab_caches
, list
) {
1180 n
= cachep
->node
[node
];
1183 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1187 static int cpuup_prepare(long cpu
)
1189 struct kmem_cache
*cachep
;
1190 struct kmem_cache_node
*n
= NULL
;
1191 int node
= cpu_to_mem(cpu
);
1195 * We need to do this right in the beginning since
1196 * alloc_arraycache's are going to use this list.
1197 * kmalloc_node allows us to add the slab to the right
1198 * kmem_cache_node and not this cpu's kmem_cache_node
1200 err
= init_cache_node_node(node
);
1205 * Now we can go ahead with allocating the shared arrays and
1208 list_for_each_entry(cachep
, &slab_caches
, list
) {
1209 struct array_cache
*nc
;
1210 struct array_cache
*shared
= NULL
;
1211 struct array_cache
**alien
= NULL
;
1213 nc
= alloc_arraycache(node
, cachep
->limit
,
1214 cachep
->batchcount
, GFP_KERNEL
);
1217 if (cachep
->shared
) {
1218 shared
= alloc_arraycache(node
,
1219 cachep
->shared
* cachep
->batchcount
,
1220 0xbaadf00d, GFP_KERNEL
);
1226 if (use_alien_caches
) {
1227 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1234 cachep
->array
[cpu
] = nc
;
1235 n
= cachep
->node
[node
];
1238 spin_lock_irq(&n
->list_lock
);
1241 * We are serialised from CPU_DEAD or
1242 * CPU_UP_CANCELLED by the cpucontrol lock
1253 spin_unlock_irq(&n
->list_lock
);
1255 free_alien_cache(alien
);
1256 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1257 slab_set_debugobj_lock_classes_node(cachep
, node
);
1258 else if (!OFF_SLAB(cachep
) &&
1259 !(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1260 on_slab_lock_classes_node(cachep
, node
);
1262 init_node_lock_keys(node
);
1266 cpuup_canceled(cpu
);
1270 static int cpuup_callback(struct notifier_block
*nfb
,
1271 unsigned long action
, void *hcpu
)
1273 long cpu
= (long)hcpu
;
1277 case CPU_UP_PREPARE
:
1278 case CPU_UP_PREPARE_FROZEN
:
1279 mutex_lock(&slab_mutex
);
1280 err
= cpuup_prepare(cpu
);
1281 mutex_unlock(&slab_mutex
);
1284 case CPU_ONLINE_FROZEN
:
1285 start_cpu_timer(cpu
);
1287 #ifdef CONFIG_HOTPLUG_CPU
1288 case CPU_DOWN_PREPARE
:
1289 case CPU_DOWN_PREPARE_FROZEN
:
1291 * Shutdown cache reaper. Note that the slab_mutex is
1292 * held so that if cache_reap() is invoked it cannot do
1293 * anything expensive but will only modify reap_work
1294 * and reschedule the timer.
1296 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1297 /* Now the cache_reaper is guaranteed to be not running. */
1298 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1300 case CPU_DOWN_FAILED
:
1301 case CPU_DOWN_FAILED_FROZEN
:
1302 start_cpu_timer(cpu
);
1305 case CPU_DEAD_FROZEN
:
1307 * Even if all the cpus of a node are down, we don't free the
1308 * kmem_cache_node of any cache. This to avoid a race between
1309 * cpu_down, and a kmalloc allocation from another cpu for
1310 * memory from the node of the cpu going down. The node
1311 * structure is usually allocated from kmem_cache_create() and
1312 * gets destroyed at kmem_cache_destroy().
1316 case CPU_UP_CANCELED
:
1317 case CPU_UP_CANCELED_FROZEN
:
1318 mutex_lock(&slab_mutex
);
1319 cpuup_canceled(cpu
);
1320 mutex_unlock(&slab_mutex
);
1323 return notifier_from_errno(err
);
1326 static struct notifier_block cpucache_notifier
= {
1327 &cpuup_callback
, NULL
, 0
1330 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1332 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1333 * Returns -EBUSY if all objects cannot be drained so that the node is not
1336 * Must hold slab_mutex.
1338 static int __meminit
drain_cache_node_node(int node
)
1340 struct kmem_cache
*cachep
;
1343 list_for_each_entry(cachep
, &slab_caches
, list
) {
1344 struct kmem_cache_node
*n
;
1346 n
= cachep
->node
[node
];
1350 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1352 if (!list_empty(&n
->slabs_full
) ||
1353 !list_empty(&n
->slabs_partial
)) {
1361 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1362 unsigned long action
, void *arg
)
1364 struct memory_notify
*mnb
= arg
;
1368 nid
= mnb
->status_change_nid
;
1373 case MEM_GOING_ONLINE
:
1374 mutex_lock(&slab_mutex
);
1375 ret
= init_cache_node_node(nid
);
1376 mutex_unlock(&slab_mutex
);
1378 case MEM_GOING_OFFLINE
:
1379 mutex_lock(&slab_mutex
);
1380 ret
= drain_cache_node_node(nid
);
1381 mutex_unlock(&slab_mutex
);
1385 case MEM_CANCEL_ONLINE
:
1386 case MEM_CANCEL_OFFLINE
:
1390 return notifier_from_errno(ret
);
1392 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1395 * swap the static kmem_cache_node with kmalloced memory
1397 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1400 struct kmem_cache_node
*ptr
;
1402 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1405 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1407 * Do not assume that spinlocks can be initialized via memcpy:
1409 spin_lock_init(&ptr
->list_lock
);
1411 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1412 cachep
->node
[nodeid
] = ptr
;
1416 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1417 * size of kmem_cache_node.
1419 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1423 for_each_online_node(node
) {
1424 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1425 cachep
->node
[node
]->next_reap
= jiffies
+
1427 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1432 * The memory after the last cpu cache pointer is used for the
1435 static void setup_node_pointer(struct kmem_cache
*cachep
)
1437 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1441 * Initialisation. Called after the page allocator have been initialised and
1442 * before smp_init().
1444 void __init
kmem_cache_init(void)
1448 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1449 sizeof(struct rcu_head
));
1450 kmem_cache
= &kmem_cache_boot
;
1451 setup_node_pointer(kmem_cache
);
1453 if (num_possible_nodes() == 1)
1454 use_alien_caches
= 0;
1456 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1457 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1459 set_up_node(kmem_cache
, CACHE_CACHE
);
1462 * Fragmentation resistance on low memory - only use bigger
1463 * page orders on machines with more than 32MB of memory if
1464 * not overridden on the command line.
1466 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1467 slab_max_order
= SLAB_MAX_ORDER_HI
;
1469 /* Bootstrap is tricky, because several objects are allocated
1470 * from caches that do not exist yet:
1471 * 1) initialize the kmem_cache cache: it contains the struct
1472 * kmem_cache structures of all caches, except kmem_cache itself:
1473 * kmem_cache is statically allocated.
1474 * Initially an __init data area is used for the head array and the
1475 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1476 * array at the end of the bootstrap.
1477 * 2) Create the first kmalloc cache.
1478 * The struct kmem_cache for the new cache is allocated normally.
1479 * An __init data area is used for the head array.
1480 * 3) Create the remaining kmalloc caches, with minimally sized
1482 * 4) Replace the __init data head arrays for kmem_cache and the first
1483 * kmalloc cache with kmalloc allocated arrays.
1484 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1485 * the other cache's with kmalloc allocated memory.
1486 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1489 /* 1) create the kmem_cache */
1492 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1494 create_boot_cache(kmem_cache
, "kmem_cache",
1495 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1496 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1497 SLAB_HWCACHE_ALIGN
);
1498 list_add(&kmem_cache
->list
, &slab_caches
);
1500 /* 2+3) create the kmalloc caches */
1503 * Initialize the caches that provide memory for the array cache and the
1504 * kmem_cache_node structures first. Without this, further allocations will
1508 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1509 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1511 if (INDEX_AC
!= INDEX_NODE
)
1512 kmalloc_caches
[INDEX_NODE
] =
1513 create_kmalloc_cache("kmalloc-node",
1514 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1516 slab_early_init
= 0;
1518 /* 4) Replace the bootstrap head arrays */
1520 struct array_cache
*ptr
;
1522 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1524 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1525 sizeof(struct arraycache_init
));
1527 * Do not assume that spinlocks can be initialized via memcpy:
1529 spin_lock_init(&ptr
->lock
);
1531 kmem_cache
->array
[smp_processor_id()] = ptr
;
1533 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1535 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1536 != &initarray_generic
.cache
);
1537 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1538 sizeof(struct arraycache_init
));
1540 * Do not assume that spinlocks can be initialized via memcpy:
1542 spin_lock_init(&ptr
->lock
);
1544 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1546 /* 5) Replace the bootstrap kmem_cache_node */
1550 for_each_online_node(nid
) {
1551 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1553 init_list(kmalloc_caches
[INDEX_AC
],
1554 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1556 if (INDEX_AC
!= INDEX_NODE
) {
1557 init_list(kmalloc_caches
[INDEX_NODE
],
1558 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1563 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1566 void __init
kmem_cache_init_late(void)
1568 struct kmem_cache
*cachep
;
1572 /* 6) resize the head arrays to their final sizes */
1573 mutex_lock(&slab_mutex
);
1574 list_for_each_entry(cachep
, &slab_caches
, list
)
1575 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1577 mutex_unlock(&slab_mutex
);
1579 /* Annotate slab for lockdep -- annotate the malloc caches */
1586 * Register a cpu startup notifier callback that initializes
1587 * cpu_cache_get for all new cpus
1589 register_cpu_notifier(&cpucache_notifier
);
1593 * Register a memory hotplug callback that initializes and frees
1596 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1600 * The reap timers are started later, with a module init call: That part
1601 * of the kernel is not yet operational.
1605 static int __init
cpucache_init(void)
1610 * Register the timers that return unneeded pages to the page allocator
1612 for_each_online_cpu(cpu
)
1613 start_cpu_timer(cpu
);
1619 __initcall(cpucache_init
);
1621 static noinline
void
1622 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1625 struct kmem_cache_node
*n
;
1627 unsigned long flags
;
1629 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1630 DEFAULT_RATELIMIT_BURST
);
1632 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1636 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1638 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1639 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1641 for_each_online_node(node
) {
1642 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1643 unsigned long active_slabs
= 0, num_slabs
= 0;
1645 n
= cachep
->node
[node
];
1649 spin_lock_irqsave(&n
->list_lock
, flags
);
1650 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1651 active_objs
+= cachep
->num
;
1654 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1655 active_objs
+= page
->active
;
1658 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1661 free_objects
+= n
->free_objects
;
1662 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1664 num_slabs
+= active_slabs
;
1665 num_objs
= num_slabs
* cachep
->num
;
1667 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1668 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1675 * Interface to system's page allocator. No need to hold the cache-lock.
1677 * If we requested dmaable memory, we will get it. Even if we
1678 * did not request dmaable memory, we might get it, but that
1679 * would be relatively rare and ignorable.
1681 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1687 flags
|= cachep
->allocflags
;
1688 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1689 flags
|= __GFP_RECLAIMABLE
;
1691 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1693 slab_out_of_memory(cachep
, flags
, nodeid
);
1697 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1698 if (unlikely(page
->pfmemalloc
))
1699 pfmemalloc_active
= true;
1701 nr_pages
= (1 << cachep
->gfporder
);
1702 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1703 add_zone_page_state(page_zone(page
),
1704 NR_SLAB_RECLAIMABLE
, nr_pages
);
1706 add_zone_page_state(page_zone(page
),
1707 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1708 __SetPageSlab(page
);
1709 if (page
->pfmemalloc
)
1710 SetPageSlabPfmemalloc(page
);
1711 memcg_bind_pages(cachep
, cachep
->gfporder
);
1713 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1714 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1717 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1719 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1726 * Interface to system's page release.
1728 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1730 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1732 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1734 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1735 sub_zone_page_state(page_zone(page
),
1736 NR_SLAB_RECLAIMABLE
, nr_freed
);
1738 sub_zone_page_state(page_zone(page
),
1739 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1741 BUG_ON(!PageSlab(page
));
1742 __ClearPageSlabPfmemalloc(page
);
1743 __ClearPageSlab(page
);
1744 page_mapcount_reset(page
);
1745 page
->mapping
= NULL
;
1747 memcg_release_pages(cachep
, cachep
->gfporder
);
1748 if (current
->reclaim_state
)
1749 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1750 __free_memcg_kmem_pages(page
, cachep
->gfporder
);
1753 static void kmem_rcu_free(struct rcu_head
*head
)
1755 struct kmem_cache
*cachep
;
1758 page
= container_of(head
, struct page
, rcu_head
);
1759 cachep
= page
->slab_cache
;
1761 kmem_freepages(cachep
, page
);
1766 #ifdef CONFIG_DEBUG_PAGEALLOC
1767 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1768 unsigned long caller
)
1770 int size
= cachep
->object_size
;
1772 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1774 if (size
< 5 * sizeof(unsigned long))
1777 *addr
++ = 0x12345678;
1779 *addr
++ = smp_processor_id();
1780 size
-= 3 * sizeof(unsigned long);
1782 unsigned long *sptr
= &caller
;
1783 unsigned long svalue
;
1785 while (!kstack_end(sptr
)) {
1787 if (kernel_text_address(svalue
)) {
1789 size
-= sizeof(unsigned long);
1790 if (size
<= sizeof(unsigned long))
1796 *addr
++ = 0x87654321;
1800 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1802 int size
= cachep
->object_size
;
1803 addr
= &((char *)addr
)[obj_offset(cachep
)];
1805 memset(addr
, val
, size
);
1806 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1809 static void dump_line(char *data
, int offset
, int limit
)
1812 unsigned char error
= 0;
1815 printk(KERN_ERR
"%03x: ", offset
);
1816 for (i
= 0; i
< limit
; i
++) {
1817 if (data
[offset
+ i
] != POISON_FREE
) {
1818 error
= data
[offset
+ i
];
1822 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1823 &data
[offset
], limit
, 1);
1825 if (bad_count
== 1) {
1826 error
^= POISON_FREE
;
1827 if (!(error
& (error
- 1))) {
1828 printk(KERN_ERR
"Single bit error detected. Probably "
1831 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1834 printk(KERN_ERR
"Run a memory test tool.\n");
1843 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1848 if (cachep
->flags
& SLAB_RED_ZONE
) {
1849 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1850 *dbg_redzone1(cachep
, objp
),
1851 *dbg_redzone2(cachep
, objp
));
1854 if (cachep
->flags
& SLAB_STORE_USER
) {
1855 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1856 *dbg_userword(cachep
, objp
),
1857 *dbg_userword(cachep
, objp
));
1859 realobj
= (char *)objp
+ obj_offset(cachep
);
1860 size
= cachep
->object_size
;
1861 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1864 if (i
+ limit
> size
)
1866 dump_line(realobj
, i
, limit
);
1870 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1876 realobj
= (char *)objp
+ obj_offset(cachep
);
1877 size
= cachep
->object_size
;
1879 for (i
= 0; i
< size
; i
++) {
1880 char exp
= POISON_FREE
;
1883 if (realobj
[i
] != exp
) {
1889 "Slab corruption (%s): %s start=%p, len=%d\n",
1890 print_tainted(), cachep
->name
, realobj
, size
);
1891 print_objinfo(cachep
, objp
, 0);
1893 /* Hexdump the affected line */
1896 if (i
+ limit
> size
)
1898 dump_line(realobj
, i
, limit
);
1901 /* Limit to 5 lines */
1907 /* Print some data about the neighboring objects, if they
1910 struct page
*page
= virt_to_head_page(objp
);
1913 objnr
= obj_to_index(cachep
, page
, objp
);
1915 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1916 realobj
= (char *)objp
+ obj_offset(cachep
);
1917 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1919 print_objinfo(cachep
, objp
, 2);
1921 if (objnr
+ 1 < cachep
->num
) {
1922 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1923 realobj
= (char *)objp
+ obj_offset(cachep
);
1924 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1926 print_objinfo(cachep
, objp
, 2);
1933 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1937 for (i
= 0; i
< cachep
->num
; i
++) {
1938 void *objp
= index_to_obj(cachep
, page
, i
);
1940 if (cachep
->flags
& SLAB_POISON
) {
1941 #ifdef CONFIG_DEBUG_PAGEALLOC
1942 if (cachep
->size
% PAGE_SIZE
== 0 &&
1944 kernel_map_pages(virt_to_page(objp
),
1945 cachep
->size
/ PAGE_SIZE
, 1);
1947 check_poison_obj(cachep
, objp
);
1949 check_poison_obj(cachep
, objp
);
1952 if (cachep
->flags
& SLAB_RED_ZONE
) {
1953 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1954 slab_error(cachep
, "start of a freed object "
1956 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1957 slab_error(cachep
, "end of a freed object "
1963 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1970 * slab_destroy - destroy and release all objects in a slab
1971 * @cachep: cache pointer being destroyed
1972 * @page: page pointer being destroyed
1974 * Destroy all the objs in a slab, and release the mem back to the system.
1975 * Before calling the slab must have been unlinked from the cache. The
1976 * cache-lock is not held/needed.
1978 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1982 freelist
= page
->freelist
;
1983 slab_destroy_debugcheck(cachep
, page
);
1984 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1985 struct rcu_head
*head
;
1988 * RCU free overloads the RCU head over the LRU.
1989 * slab_page has been overloeaded over the LRU,
1990 * however it is not used from now on so that
1991 * we can use it safely.
1993 head
= (void *)&page
->rcu_head
;
1994 call_rcu(head
, kmem_rcu_free
);
1997 kmem_freepages(cachep
, page
);
2001 * From now on, we don't use freelist
2002 * although actual page can be freed in rcu context
2004 if (OFF_SLAB(cachep
))
2005 kmem_cache_free(cachep
->freelist_cache
, freelist
);
2009 * calculate_slab_order - calculate size (page order) of slabs
2010 * @cachep: pointer to the cache that is being created
2011 * @size: size of objects to be created in this cache.
2012 * @align: required alignment for the objects.
2013 * @flags: slab allocation flags
2015 * Also calculates the number of objects per slab.
2017 * This could be made much more intelligent. For now, try to avoid using
2018 * high order pages for slabs. When the gfp() functions are more friendly
2019 * towards high-order requests, this should be changed.
2021 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2022 size_t size
, size_t align
, unsigned long flags
)
2024 unsigned long offslab_limit
;
2025 size_t left_over
= 0;
2028 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2032 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2036 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2037 if (num
> SLAB_OBJ_MAX_NUM
)
2040 if (flags
& CFLGS_OFF_SLAB
) {
2042 * Max number of objs-per-slab for caches which
2043 * use off-slab slabs. Needed to avoid a possible
2044 * looping condition in cache_grow().
2046 offslab_limit
= size
;
2047 offslab_limit
/= sizeof(freelist_idx_t
);
2049 if (num
> offslab_limit
)
2053 /* Found something acceptable - save it away */
2055 cachep
->gfporder
= gfporder
;
2056 left_over
= remainder
;
2059 * A VFS-reclaimable slab tends to have most allocations
2060 * as GFP_NOFS and we really don't want to have to be allocating
2061 * higher-order pages when we are unable to shrink dcache.
2063 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2067 * Large number of objects is good, but very large slabs are
2068 * currently bad for the gfp()s.
2070 if (gfporder
>= slab_max_order
)
2074 * Acceptable internal fragmentation?
2076 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2082 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2084 if (slab_state
>= FULL
)
2085 return enable_cpucache(cachep
, gfp
);
2087 if (slab_state
== DOWN
) {
2089 * Note: Creation of first cache (kmem_cache).
2090 * The setup_node is taken care
2091 * of by the caller of __kmem_cache_create
2093 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2094 slab_state
= PARTIAL
;
2095 } else if (slab_state
== PARTIAL
) {
2097 * Note: the second kmem_cache_create must create the cache
2098 * that's used by kmalloc(24), otherwise the creation of
2099 * further caches will BUG().
2101 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2104 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2105 * the second cache, then we need to set up all its node/,
2106 * otherwise the creation of further caches will BUG().
2108 set_up_node(cachep
, SIZE_AC
);
2109 if (INDEX_AC
== INDEX_NODE
)
2110 slab_state
= PARTIAL_NODE
;
2112 slab_state
= PARTIAL_ARRAYCACHE
;
2114 /* Remaining boot caches */
2115 cachep
->array
[smp_processor_id()] =
2116 kmalloc(sizeof(struct arraycache_init
), gfp
);
2118 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2119 set_up_node(cachep
, SIZE_NODE
);
2120 slab_state
= PARTIAL_NODE
;
2123 for_each_online_node(node
) {
2124 cachep
->node
[node
] =
2125 kmalloc_node(sizeof(struct kmem_cache_node
),
2127 BUG_ON(!cachep
->node
[node
]);
2128 kmem_cache_node_init(cachep
->node
[node
]);
2132 cachep
->node
[numa_mem_id()]->next_reap
=
2133 jiffies
+ REAPTIMEOUT_NODE
+
2134 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
2136 cpu_cache_get(cachep
)->avail
= 0;
2137 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2138 cpu_cache_get(cachep
)->batchcount
= 1;
2139 cpu_cache_get(cachep
)->touched
= 0;
2140 cachep
->batchcount
= 1;
2141 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2146 * __kmem_cache_create - Create a cache.
2147 * @cachep: cache management descriptor
2148 * @flags: SLAB flags
2150 * Returns a ptr to the cache on success, NULL on failure.
2151 * Cannot be called within a int, but can be interrupted.
2152 * The @ctor is run when new pages are allocated by the cache.
2156 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2157 * to catch references to uninitialised memory.
2159 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2160 * for buffer overruns.
2162 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2163 * cacheline. This can be beneficial if you're counting cycles as closely
2167 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2169 size_t left_over
, freelist_size
, ralign
;
2172 size_t size
= cachep
->size
;
2177 * Enable redzoning and last user accounting, except for caches with
2178 * large objects, if the increased size would increase the object size
2179 * above the next power of two: caches with object sizes just above a
2180 * power of two have a significant amount of internal fragmentation.
2182 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2183 2 * sizeof(unsigned long long)))
2184 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2185 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2186 flags
|= SLAB_POISON
;
2188 if (flags
& SLAB_DESTROY_BY_RCU
)
2189 BUG_ON(flags
& SLAB_POISON
);
2193 * Check that size is in terms of words. This is needed to avoid
2194 * unaligned accesses for some archs when redzoning is used, and makes
2195 * sure any on-slab bufctl's are also correctly aligned.
2197 if (size
& (BYTES_PER_WORD
- 1)) {
2198 size
+= (BYTES_PER_WORD
- 1);
2199 size
&= ~(BYTES_PER_WORD
- 1);
2203 * Redzoning and user store require word alignment or possibly larger.
2204 * Note this will be overridden by architecture or caller mandated
2205 * alignment if either is greater than BYTES_PER_WORD.
2207 if (flags
& SLAB_STORE_USER
)
2208 ralign
= BYTES_PER_WORD
;
2210 if (flags
& SLAB_RED_ZONE
) {
2211 ralign
= REDZONE_ALIGN
;
2212 /* If redzoning, ensure that the second redzone is suitably
2213 * aligned, by adjusting the object size accordingly. */
2214 size
+= REDZONE_ALIGN
- 1;
2215 size
&= ~(REDZONE_ALIGN
- 1);
2218 /* 3) caller mandated alignment */
2219 if (ralign
< cachep
->align
) {
2220 ralign
= cachep
->align
;
2222 /* disable debug if necessary */
2223 if (ralign
> __alignof__(unsigned long long))
2224 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2228 cachep
->align
= ralign
;
2230 if (slab_is_available())
2235 setup_node_pointer(cachep
);
2239 * Both debugging options require word-alignment which is calculated
2242 if (flags
& SLAB_RED_ZONE
) {
2243 /* add space for red zone words */
2244 cachep
->obj_offset
+= sizeof(unsigned long long);
2245 size
+= 2 * sizeof(unsigned long long);
2247 if (flags
& SLAB_STORE_USER
) {
2248 /* user store requires one word storage behind the end of
2249 * the real object. But if the second red zone needs to be
2250 * aligned to 64 bits, we must allow that much space.
2252 if (flags
& SLAB_RED_ZONE
)
2253 size
+= REDZONE_ALIGN
;
2255 size
+= BYTES_PER_WORD
;
2257 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2258 if (size
>= kmalloc_size(INDEX_NODE
+ 1)
2259 && cachep
->object_size
> cache_line_size()
2260 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2261 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2268 * Determine if the slab management is 'on' or 'off' slab.
2269 * (bootstrapping cannot cope with offslab caches so don't do
2270 * it too early on. Always use on-slab management when
2271 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2273 if ((size
>= (PAGE_SIZE
>> 5)) && !slab_early_init
&&
2274 !(flags
& SLAB_NOLEAKTRACE
))
2276 * Size is large, assume best to place the slab management obj
2277 * off-slab (should allow better packing of objs).
2279 flags
|= CFLGS_OFF_SLAB
;
2281 size
= ALIGN(size
, cachep
->align
);
2283 * We should restrict the number of objects in a slab to implement
2284 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2286 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2287 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2289 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2295 ALIGN(cachep
->num
* sizeof(freelist_idx_t
), cachep
->align
);
2298 * If the slab has been placed off-slab, and we have enough space then
2299 * move it on-slab. This is at the expense of any extra colouring.
2301 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= freelist_size
) {
2302 flags
&= ~CFLGS_OFF_SLAB
;
2303 left_over
-= freelist_size
;
2306 if (flags
& CFLGS_OFF_SLAB
) {
2307 /* really off slab. No need for manual alignment */
2308 freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2310 #ifdef CONFIG_PAGE_POISONING
2311 /* If we're going to use the generic kernel_map_pages()
2312 * poisoning, then it's going to smash the contents of
2313 * the redzone and userword anyhow, so switch them off.
2315 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2316 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2320 cachep
->colour_off
= cache_line_size();
2321 /* Offset must be a multiple of the alignment. */
2322 if (cachep
->colour_off
< cachep
->align
)
2323 cachep
->colour_off
= cachep
->align
;
2324 cachep
->colour
= left_over
/ cachep
->colour_off
;
2325 cachep
->freelist_size
= freelist_size
;
2326 cachep
->flags
= flags
;
2327 cachep
->allocflags
= __GFP_COMP
;
2328 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2329 cachep
->allocflags
|= GFP_DMA
;
2330 cachep
->size
= size
;
2331 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2333 if (flags
& CFLGS_OFF_SLAB
) {
2334 cachep
->freelist_cache
= kmalloc_slab(freelist_size
, 0u);
2336 * This is a possibility for one of the kmalloc_{dma,}_caches.
2337 * But since we go off slab only for object size greater than
2338 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2339 * in ascending order,this should not happen at all.
2340 * But leave a BUG_ON for some lucky dude.
2342 BUG_ON(ZERO_OR_NULL_PTR(cachep
->freelist_cache
));
2345 err
= setup_cpu_cache(cachep
, gfp
);
2347 __kmem_cache_shutdown(cachep
);
2351 if (flags
& SLAB_DEBUG_OBJECTS
) {
2353 * Would deadlock through slab_destroy()->call_rcu()->
2354 * debug_object_activate()->kmem_cache_alloc().
2356 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2358 slab_set_debugobj_lock_classes(cachep
);
2359 } else if (!OFF_SLAB(cachep
) && !(flags
& SLAB_DESTROY_BY_RCU
))
2360 on_slab_lock_classes(cachep
);
2366 static void check_irq_off(void)
2368 BUG_ON(!irqs_disabled());
2371 static void check_irq_on(void)
2373 BUG_ON(irqs_disabled());
2376 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2380 assert_spin_locked(&cachep
->node
[numa_mem_id()]->list_lock
);
2384 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2388 assert_spin_locked(&cachep
->node
[node
]->list_lock
);
2393 #define check_irq_off() do { } while(0)
2394 #define check_irq_on() do { } while(0)
2395 #define check_spinlock_acquired(x) do { } while(0)
2396 #define check_spinlock_acquired_node(x, y) do { } while(0)
2399 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2400 struct array_cache
*ac
,
2401 int force
, int node
);
2403 static void do_drain(void *arg
)
2405 struct kmem_cache
*cachep
= arg
;
2406 struct array_cache
*ac
;
2407 int node
= numa_mem_id();
2410 ac
= cpu_cache_get(cachep
);
2411 spin_lock(&cachep
->node
[node
]->list_lock
);
2412 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2413 spin_unlock(&cachep
->node
[node
]->list_lock
);
2417 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2419 struct kmem_cache_node
*n
;
2422 on_each_cpu(do_drain
, cachep
, 1);
2424 for_each_online_node(node
) {
2425 n
= cachep
->node
[node
];
2427 drain_alien_cache(cachep
, n
->alien
);
2430 for_each_online_node(node
) {
2431 n
= cachep
->node
[node
];
2433 drain_array(cachep
, n
, n
->shared
, 1, node
);
2438 * Remove slabs from the list of free slabs.
2439 * Specify the number of slabs to drain in tofree.
2441 * Returns the actual number of slabs released.
2443 static int drain_freelist(struct kmem_cache
*cache
,
2444 struct kmem_cache_node
*n
, int tofree
)
2446 struct list_head
*p
;
2451 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2453 spin_lock_irq(&n
->list_lock
);
2454 p
= n
->slabs_free
.prev
;
2455 if (p
== &n
->slabs_free
) {
2456 spin_unlock_irq(&n
->list_lock
);
2460 page
= list_entry(p
, struct page
, lru
);
2462 BUG_ON(page
->active
);
2464 list_del(&page
->lru
);
2466 * Safe to drop the lock. The slab is no longer linked
2469 n
->free_objects
-= cache
->num
;
2470 spin_unlock_irq(&n
->list_lock
);
2471 slab_destroy(cache
, page
);
2478 /* Called with slab_mutex held to protect against cpu hotplug */
2479 static int __cache_shrink(struct kmem_cache
*cachep
)
2482 struct kmem_cache_node
*n
;
2484 drain_cpu_caches(cachep
);
2487 for_each_online_node(i
) {
2488 n
= cachep
->node
[i
];
2492 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2494 ret
+= !list_empty(&n
->slabs_full
) ||
2495 !list_empty(&n
->slabs_partial
);
2497 return (ret
? 1 : 0);
2501 * kmem_cache_shrink - Shrink a cache.
2502 * @cachep: The cache to shrink.
2504 * Releases as many slabs as possible for a cache.
2505 * To help debugging, a zero exit status indicates all slabs were released.
2507 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2510 BUG_ON(!cachep
|| in_interrupt());
2513 mutex_lock(&slab_mutex
);
2514 ret
= __cache_shrink(cachep
);
2515 mutex_unlock(&slab_mutex
);
2519 EXPORT_SYMBOL(kmem_cache_shrink
);
2521 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2524 struct kmem_cache_node
*n
;
2525 int rc
= __cache_shrink(cachep
);
2530 for_each_online_cpu(i
)
2531 kfree(cachep
->array
[i
]);
2533 /* NUMA: free the node structures */
2534 for_each_online_node(i
) {
2535 n
= cachep
->node
[i
];
2538 free_alien_cache(n
->alien
);
2546 * Get the memory for a slab management obj.
2548 * For a slab cache when the slab descriptor is off-slab, the
2549 * slab descriptor can't come from the same cache which is being created,
2550 * Because if it is the case, that means we defer the creation of
2551 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2552 * And we eventually call down to __kmem_cache_create(), which
2553 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2554 * This is a "chicken-and-egg" problem.
2556 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2557 * which are all initialized during kmem_cache_init().
2559 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2560 struct page
*page
, int colour_off
,
2561 gfp_t local_flags
, int nodeid
)
2564 void *addr
= page_address(page
);
2566 if (OFF_SLAB(cachep
)) {
2567 /* Slab management obj is off-slab. */
2568 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2569 local_flags
, nodeid
);
2573 freelist
= addr
+ colour_off
;
2574 colour_off
+= cachep
->freelist_size
;
2577 page
->s_mem
= addr
+ colour_off
;
2581 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2583 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2586 static inline void set_free_obj(struct page
*page
,
2587 unsigned int idx
, freelist_idx_t val
)
2589 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2592 static void cache_init_objs(struct kmem_cache
*cachep
,
2597 for (i
= 0; i
< cachep
->num
; i
++) {
2598 void *objp
= index_to_obj(cachep
, page
, i
);
2600 /* need to poison the objs? */
2601 if (cachep
->flags
& SLAB_POISON
)
2602 poison_obj(cachep
, objp
, POISON_FREE
);
2603 if (cachep
->flags
& SLAB_STORE_USER
)
2604 *dbg_userword(cachep
, objp
) = NULL
;
2606 if (cachep
->flags
& SLAB_RED_ZONE
) {
2607 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2608 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2611 * Constructors are not allowed to allocate memory from the same
2612 * cache which they are a constructor for. Otherwise, deadlock.
2613 * They must also be threaded.
2615 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2616 cachep
->ctor(objp
+ obj_offset(cachep
));
2618 if (cachep
->flags
& SLAB_RED_ZONE
) {
2619 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2620 slab_error(cachep
, "constructor overwrote the"
2621 " end of an object");
2622 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2623 slab_error(cachep
, "constructor overwrote the"
2624 " start of an object");
2626 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2627 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2628 kernel_map_pages(virt_to_page(objp
),
2629 cachep
->size
/ PAGE_SIZE
, 0);
2634 set_free_obj(page
, i
, i
);
2638 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2640 if (CONFIG_ZONE_DMA_FLAG
) {
2641 if (flags
& GFP_DMA
)
2642 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2644 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2648 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
,
2653 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2656 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2662 static void slab_put_obj(struct kmem_cache
*cachep
, struct page
*page
,
2663 void *objp
, int nodeid
)
2665 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2669 /* Verify that the slab belongs to the intended node */
2670 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2672 /* Verify double free bug */
2673 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2674 if (get_free_obj(page
, i
) == objnr
) {
2675 printk(KERN_ERR
"slab: double free detected in cache "
2676 "'%s', objp %p\n", cachep
->name
, objp
);
2682 set_free_obj(page
, page
->active
, objnr
);
2686 * Map pages beginning at addr to the given cache and slab. This is required
2687 * for the slab allocator to be able to lookup the cache and slab of a
2688 * virtual address for kfree, ksize, and slab debugging.
2690 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2693 page
->slab_cache
= cache
;
2694 page
->freelist
= freelist
;
2698 * Grow (by 1) the number of slabs within a cache. This is called by
2699 * kmem_cache_alloc() when there are no active objs left in a cache.
2701 static int cache_grow(struct kmem_cache
*cachep
,
2702 gfp_t flags
, int nodeid
, struct page
*page
)
2707 struct kmem_cache_node
*n
;
2710 * Be lazy and only check for valid flags here, keeping it out of the
2711 * critical path in kmem_cache_alloc().
2713 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2714 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2716 /* Take the node list lock to change the colour_next on this node */
2718 n
= cachep
->node
[nodeid
];
2719 spin_lock(&n
->list_lock
);
2721 /* Get colour for the slab, and cal the next value. */
2722 offset
= n
->colour_next
;
2724 if (n
->colour_next
>= cachep
->colour
)
2726 spin_unlock(&n
->list_lock
);
2728 offset
*= cachep
->colour_off
;
2730 if (local_flags
& __GFP_WAIT
)
2734 * The test for missing atomic flag is performed here, rather than
2735 * the more obvious place, simply to reduce the critical path length
2736 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2737 * will eventually be caught here (where it matters).
2739 kmem_flagcheck(cachep
, flags
);
2742 * Get mem for the objs. Attempt to allocate a physical page from
2746 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2750 /* Get slab management. */
2751 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2752 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2756 slab_map_pages(cachep
, page
, freelist
);
2758 cache_init_objs(cachep
, page
);
2760 if (local_flags
& __GFP_WAIT
)
2761 local_irq_disable();
2763 spin_lock(&n
->list_lock
);
2765 /* Make slab active. */
2766 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2767 STATS_INC_GROWN(cachep
);
2768 n
->free_objects
+= cachep
->num
;
2769 spin_unlock(&n
->list_lock
);
2772 kmem_freepages(cachep
, page
);
2774 if (local_flags
& __GFP_WAIT
)
2775 local_irq_disable();
2782 * Perform extra freeing checks:
2783 * - detect bad pointers.
2784 * - POISON/RED_ZONE checking
2786 static void kfree_debugcheck(const void *objp
)
2788 if (!virt_addr_valid(objp
)) {
2789 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2790 (unsigned long)objp
);
2795 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2797 unsigned long long redzone1
, redzone2
;
2799 redzone1
= *dbg_redzone1(cache
, obj
);
2800 redzone2
= *dbg_redzone2(cache
, obj
);
2805 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2808 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2809 slab_error(cache
, "double free detected");
2811 slab_error(cache
, "memory outside object was overwritten");
2813 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2814 obj
, redzone1
, redzone2
);
2817 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2818 unsigned long caller
)
2823 BUG_ON(virt_to_cache(objp
) != cachep
);
2825 objp
-= obj_offset(cachep
);
2826 kfree_debugcheck(objp
);
2827 page
= virt_to_head_page(objp
);
2829 if (cachep
->flags
& SLAB_RED_ZONE
) {
2830 verify_redzone_free(cachep
, objp
);
2831 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2832 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2834 if (cachep
->flags
& SLAB_STORE_USER
)
2835 *dbg_userword(cachep
, objp
) = (void *)caller
;
2837 objnr
= obj_to_index(cachep
, page
, objp
);
2839 BUG_ON(objnr
>= cachep
->num
);
2840 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2842 if (cachep
->flags
& SLAB_POISON
) {
2843 #ifdef CONFIG_DEBUG_PAGEALLOC
2844 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2845 store_stackinfo(cachep
, objp
, caller
);
2846 kernel_map_pages(virt_to_page(objp
),
2847 cachep
->size
/ PAGE_SIZE
, 0);
2849 poison_obj(cachep
, objp
, POISON_FREE
);
2852 poison_obj(cachep
, objp
, POISON_FREE
);
2859 #define kfree_debugcheck(x) do { } while(0)
2860 #define cache_free_debugcheck(x,objp,z) (objp)
2863 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2867 struct kmem_cache_node
*n
;
2868 struct array_cache
*ac
;
2872 node
= numa_mem_id();
2873 if (unlikely(force_refill
))
2876 ac
= cpu_cache_get(cachep
);
2877 batchcount
= ac
->batchcount
;
2878 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2880 * If there was little recent activity on this cache, then
2881 * perform only a partial refill. Otherwise we could generate
2884 batchcount
= BATCHREFILL_LIMIT
;
2886 n
= cachep
->node
[node
];
2888 BUG_ON(ac
->avail
> 0 || !n
);
2889 spin_lock(&n
->list_lock
);
2891 /* See if we can refill from the shared array */
2892 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2893 n
->shared
->touched
= 1;
2897 while (batchcount
> 0) {
2898 struct list_head
*entry
;
2900 /* Get slab alloc is to come from. */
2901 entry
= n
->slabs_partial
.next
;
2902 if (entry
== &n
->slabs_partial
) {
2903 n
->free_touched
= 1;
2904 entry
= n
->slabs_free
.next
;
2905 if (entry
== &n
->slabs_free
)
2909 page
= list_entry(entry
, struct page
, lru
);
2910 check_spinlock_acquired(cachep
);
2913 * The slab was either on partial or free list so
2914 * there must be at least one object available for
2917 BUG_ON(page
->active
>= cachep
->num
);
2919 while (page
->active
< cachep
->num
&& batchcount
--) {
2920 STATS_INC_ALLOCED(cachep
);
2921 STATS_INC_ACTIVE(cachep
);
2922 STATS_SET_HIGH(cachep
);
2924 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, page
,
2928 /* move slabp to correct slabp list: */
2929 list_del(&page
->lru
);
2930 if (page
->active
== cachep
->num
)
2931 list_add(&page
->lru
, &n
->slabs_full
);
2933 list_add(&page
->lru
, &n
->slabs_partial
);
2937 n
->free_objects
-= ac
->avail
;
2939 spin_unlock(&n
->list_lock
);
2941 if (unlikely(!ac
->avail
)) {
2944 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2946 /* cache_grow can reenable interrupts, then ac could change. */
2947 ac
= cpu_cache_get(cachep
);
2948 node
= numa_mem_id();
2950 /* no objects in sight? abort */
2951 if (!x
&& (ac
->avail
== 0 || force_refill
))
2954 if (!ac
->avail
) /* objects refilled by interrupt? */
2959 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2962 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2965 might_sleep_if(flags
& __GFP_WAIT
);
2967 kmem_flagcheck(cachep
, flags
);
2972 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2973 gfp_t flags
, void *objp
, unsigned long caller
)
2977 if (cachep
->flags
& SLAB_POISON
) {
2978 #ifdef CONFIG_DEBUG_PAGEALLOC
2979 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2980 kernel_map_pages(virt_to_page(objp
),
2981 cachep
->size
/ PAGE_SIZE
, 1);
2983 check_poison_obj(cachep
, objp
);
2985 check_poison_obj(cachep
, objp
);
2987 poison_obj(cachep
, objp
, POISON_INUSE
);
2989 if (cachep
->flags
& SLAB_STORE_USER
)
2990 *dbg_userword(cachep
, objp
) = (void *)caller
;
2992 if (cachep
->flags
& SLAB_RED_ZONE
) {
2993 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2994 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2995 slab_error(cachep
, "double free, or memory outside"
2996 " object was overwritten");
2998 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2999 objp
, *dbg_redzone1(cachep
, objp
),
3000 *dbg_redzone2(cachep
, objp
));
3002 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3003 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3005 objp
+= obj_offset(cachep
);
3006 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3008 if (ARCH_SLAB_MINALIGN
&&
3009 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3010 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3011 objp
, (int)ARCH_SLAB_MINALIGN
);
3016 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3019 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3021 if (cachep
== kmem_cache
)
3024 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3027 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3030 struct array_cache
*ac
;
3031 bool force_refill
= false;
3035 ac
= cpu_cache_get(cachep
);
3036 if (likely(ac
->avail
)) {
3038 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3041 * Allow for the possibility all avail objects are not allowed
3042 * by the current flags
3045 STATS_INC_ALLOCHIT(cachep
);
3048 force_refill
= true;
3051 STATS_INC_ALLOCMISS(cachep
);
3052 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3054 * the 'ac' may be updated by cache_alloc_refill(),
3055 * and kmemleak_erase() requires its correct value.
3057 ac
= cpu_cache_get(cachep
);
3061 * To avoid a false negative, if an object that is in one of the
3062 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3063 * treat the array pointers as a reference to the object.
3066 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3072 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
3074 * If we are in_interrupt, then process context, including cpusets and
3075 * mempolicy, may not apply and should not be used for allocation policy.
3077 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3079 int nid_alloc
, nid_here
;
3081 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3083 nid_alloc
= nid_here
= numa_mem_id();
3084 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3085 nid_alloc
= cpuset_slab_spread_node();
3086 else if (current
->mempolicy
)
3087 nid_alloc
= mempolicy_slab_node();
3088 if (nid_alloc
!= nid_here
)
3089 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3094 * Fallback function if there was no memory available and no objects on a
3095 * certain node and fall back is permitted. First we scan all the
3096 * available node for available objects. If that fails then we
3097 * perform an allocation without specifying a node. This allows the page
3098 * allocator to do its reclaim / fallback magic. We then insert the
3099 * slab into the proper nodelist and then allocate from it.
3101 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3103 struct zonelist
*zonelist
;
3107 enum zone_type high_zoneidx
= gfp_zone(flags
);
3110 unsigned int cpuset_mems_cookie
;
3112 if (flags
& __GFP_THISNODE
)
3115 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3118 cpuset_mems_cookie
= read_mems_allowed_begin();
3119 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3123 * Look through allowed nodes for objects available
3124 * from existing per node queues.
3126 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3127 nid
= zone_to_nid(zone
);
3129 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3131 cache
->node
[nid
]->free_objects
) {
3132 obj
= ____cache_alloc_node(cache
,
3133 flags
| GFP_THISNODE
, nid
);
3141 * This allocation will be performed within the constraints
3142 * of the current cpuset / memory policy requirements.
3143 * We may trigger various forms of reclaim on the allowed
3144 * set and go into memory reserves if necessary.
3148 if (local_flags
& __GFP_WAIT
)
3150 kmem_flagcheck(cache
, flags
);
3151 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3152 if (local_flags
& __GFP_WAIT
)
3153 local_irq_disable();
3156 * Insert into the appropriate per node queues
3158 nid
= page_to_nid(page
);
3159 if (cache_grow(cache
, flags
, nid
, page
)) {
3160 obj
= ____cache_alloc_node(cache
,
3161 flags
| GFP_THISNODE
, nid
);
3164 * Another processor may allocate the
3165 * objects in the slab since we are
3166 * not holding any locks.
3170 /* cache_grow already freed obj */
3176 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3182 * A interface to enable slab creation on nodeid
3184 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3187 struct list_head
*entry
;
3189 struct kmem_cache_node
*n
;
3193 VM_BUG_ON(nodeid
> num_online_nodes());
3194 n
= cachep
->node
[nodeid
];
3199 spin_lock(&n
->list_lock
);
3200 entry
= n
->slabs_partial
.next
;
3201 if (entry
== &n
->slabs_partial
) {
3202 n
->free_touched
= 1;
3203 entry
= n
->slabs_free
.next
;
3204 if (entry
== &n
->slabs_free
)
3208 page
= list_entry(entry
, struct page
, lru
);
3209 check_spinlock_acquired_node(cachep
, nodeid
);
3211 STATS_INC_NODEALLOCS(cachep
);
3212 STATS_INC_ACTIVE(cachep
);
3213 STATS_SET_HIGH(cachep
);
3215 BUG_ON(page
->active
== cachep
->num
);
3217 obj
= slab_get_obj(cachep
, page
, nodeid
);
3219 /* move slabp to correct slabp list: */
3220 list_del(&page
->lru
);
3222 if (page
->active
== cachep
->num
)
3223 list_add(&page
->lru
, &n
->slabs_full
);
3225 list_add(&page
->lru
, &n
->slabs_partial
);
3227 spin_unlock(&n
->list_lock
);
3231 spin_unlock(&n
->list_lock
);
3232 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3236 return fallback_alloc(cachep
, flags
);
3242 static __always_inline
void *
3243 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3244 unsigned long caller
)
3246 unsigned long save_flags
;
3248 int slab_node
= numa_mem_id();
3250 flags
&= gfp_allowed_mask
;
3252 lockdep_trace_alloc(flags
);
3254 if (slab_should_failslab(cachep
, flags
))
3257 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3259 cache_alloc_debugcheck_before(cachep
, flags
);
3260 local_irq_save(save_flags
);
3262 if (nodeid
== NUMA_NO_NODE
)
3265 if (unlikely(!cachep
->node
[nodeid
])) {
3266 /* Node not bootstrapped yet */
3267 ptr
= fallback_alloc(cachep
, flags
);
3271 if (nodeid
== slab_node
) {
3273 * Use the locally cached objects if possible.
3274 * However ____cache_alloc does not allow fallback
3275 * to other nodes. It may fail while we still have
3276 * objects on other nodes available.
3278 ptr
= ____cache_alloc(cachep
, flags
);
3282 /* ___cache_alloc_node can fall back to other nodes */
3283 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3285 local_irq_restore(save_flags
);
3286 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3287 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3291 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3292 if (unlikely(flags
& __GFP_ZERO
))
3293 memset(ptr
, 0, cachep
->object_size
);
3299 static __always_inline
void *
3300 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3304 if (current
->mempolicy
|| unlikely(current
->flags
& PF_SPREAD_SLAB
)) {
3305 objp
= alternate_node_alloc(cache
, flags
);
3309 objp
= ____cache_alloc(cache
, flags
);
3312 * We may just have run out of memory on the local node.
3313 * ____cache_alloc_node() knows how to locate memory on other nodes
3316 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3323 static __always_inline
void *
3324 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3326 return ____cache_alloc(cachep
, flags
);
3329 #endif /* CONFIG_NUMA */
3331 static __always_inline
void *
3332 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3334 unsigned long save_flags
;
3337 flags
&= gfp_allowed_mask
;
3339 lockdep_trace_alloc(flags
);
3341 if (slab_should_failslab(cachep
, flags
))
3344 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3346 cache_alloc_debugcheck_before(cachep
, flags
);
3347 local_irq_save(save_flags
);
3348 objp
= __do_cache_alloc(cachep
, flags
);
3349 local_irq_restore(save_flags
);
3350 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3351 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3356 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3357 if (unlikely(flags
& __GFP_ZERO
))
3358 memset(objp
, 0, cachep
->object_size
);
3365 * Caller needs to acquire correct kmem_cache_node's list_lock
3367 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3371 struct kmem_cache_node
*n
;
3373 for (i
= 0; i
< nr_objects
; i
++) {
3377 clear_obj_pfmemalloc(&objpp
[i
]);
3380 page
= virt_to_head_page(objp
);
3381 n
= cachep
->node
[node
];
3382 list_del(&page
->lru
);
3383 check_spinlock_acquired_node(cachep
, node
);
3384 slab_put_obj(cachep
, page
, objp
, node
);
3385 STATS_DEC_ACTIVE(cachep
);
3388 /* fixup slab chains */
3389 if (page
->active
== 0) {
3390 if (n
->free_objects
> n
->free_limit
) {
3391 n
->free_objects
-= cachep
->num
;
3392 /* No need to drop any previously held
3393 * lock here, even if we have a off-slab slab
3394 * descriptor it is guaranteed to come from
3395 * a different cache, refer to comments before
3398 slab_destroy(cachep
, page
);
3400 list_add(&page
->lru
, &n
->slabs_free
);
3403 /* Unconditionally move a slab to the end of the
3404 * partial list on free - maximum time for the
3405 * other objects to be freed, too.
3407 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3412 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3415 struct kmem_cache_node
*n
;
3416 int node
= numa_mem_id();
3418 batchcount
= ac
->batchcount
;
3420 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3423 n
= cachep
->node
[node
];
3424 spin_lock(&n
->list_lock
);
3426 struct array_cache
*shared_array
= n
->shared
;
3427 int max
= shared_array
->limit
- shared_array
->avail
;
3429 if (batchcount
> max
)
3431 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3432 ac
->entry
, sizeof(void *) * batchcount
);
3433 shared_array
->avail
+= batchcount
;
3438 free_block(cachep
, ac
->entry
, batchcount
, node
);
3443 struct list_head
*p
;
3445 p
= n
->slabs_free
.next
;
3446 while (p
!= &(n
->slabs_free
)) {
3449 page
= list_entry(p
, struct page
, lru
);
3450 BUG_ON(page
->active
);
3455 STATS_SET_FREEABLE(cachep
, i
);
3458 spin_unlock(&n
->list_lock
);
3459 ac
->avail
-= batchcount
;
3460 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3464 * Release an obj back to its cache. If the obj has a constructed state, it must
3465 * be in this state _before_ it is released. Called with disabled ints.
3467 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3468 unsigned long caller
)
3470 struct array_cache
*ac
= cpu_cache_get(cachep
);
3473 kmemleak_free_recursive(objp
, cachep
->flags
);
3474 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3476 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3479 * Skip calling cache_free_alien() when the platform is not numa.
3480 * This will avoid cache misses that happen while accessing slabp (which
3481 * is per page memory reference) to get nodeid. Instead use a global
3482 * variable to skip the call, which is mostly likely to be present in
3485 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3488 if (likely(ac
->avail
< ac
->limit
)) {
3489 STATS_INC_FREEHIT(cachep
);
3491 STATS_INC_FREEMISS(cachep
);
3492 cache_flusharray(cachep
, ac
);
3495 ac_put_obj(cachep
, ac
, objp
);
3499 * kmem_cache_alloc - Allocate an object
3500 * @cachep: The cache to allocate from.
3501 * @flags: See kmalloc().
3503 * Allocate an object from this cache. The flags are only relevant
3504 * if the cache has no available objects.
3506 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3508 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3510 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3511 cachep
->object_size
, cachep
->size
, flags
);
3515 EXPORT_SYMBOL(kmem_cache_alloc
);
3517 #ifdef CONFIG_TRACING
3519 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3523 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3525 trace_kmalloc(_RET_IP_
, ret
,
3526 size
, cachep
->size
, flags
);
3529 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3534 * kmem_cache_alloc_node - Allocate an object on the specified node
3535 * @cachep: The cache to allocate from.
3536 * @flags: See kmalloc().
3537 * @nodeid: node number of the target node.
3539 * Identical to kmem_cache_alloc but it will allocate memory on the given
3540 * node, which can improve the performance for cpu bound structures.
3542 * Fallback to other node is possible if __GFP_THISNODE is not set.
3544 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3546 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3548 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3549 cachep
->object_size
, cachep
->size
,
3554 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3556 #ifdef CONFIG_TRACING
3557 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3564 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3566 trace_kmalloc_node(_RET_IP_
, ret
,
3571 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3574 static __always_inline
void *
3575 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3577 struct kmem_cache
*cachep
;
3579 cachep
= kmalloc_slab(size
, flags
);
3580 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3582 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3585 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3586 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3588 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3590 EXPORT_SYMBOL(__kmalloc_node
);
3592 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3593 int node
, unsigned long caller
)
3595 return __do_kmalloc_node(size
, flags
, node
, caller
);
3597 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3599 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3601 return __do_kmalloc_node(size
, flags
, node
, 0);
3603 EXPORT_SYMBOL(__kmalloc_node
);
3604 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3605 #endif /* CONFIG_NUMA */
3608 * __do_kmalloc - allocate memory
3609 * @size: how many bytes of memory are required.
3610 * @flags: the type of memory to allocate (see kmalloc).
3611 * @caller: function caller for debug tracking of the caller
3613 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3614 unsigned long caller
)
3616 struct kmem_cache
*cachep
;
3619 cachep
= kmalloc_slab(size
, flags
);
3620 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3622 ret
= slab_alloc(cachep
, flags
, caller
);
3624 trace_kmalloc(caller
, ret
,
3625 size
, cachep
->size
, flags
);
3631 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3632 void *__kmalloc(size_t size
, gfp_t flags
)
3634 return __do_kmalloc(size
, flags
, _RET_IP_
);
3636 EXPORT_SYMBOL(__kmalloc
);
3638 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3640 return __do_kmalloc(size
, flags
, caller
);
3642 EXPORT_SYMBOL(__kmalloc_track_caller
);
3645 void *__kmalloc(size_t size
, gfp_t flags
)
3647 return __do_kmalloc(size
, flags
, 0);
3649 EXPORT_SYMBOL(__kmalloc
);
3653 * kmem_cache_free - Deallocate an object
3654 * @cachep: The cache the allocation was from.
3655 * @objp: The previously allocated object.
3657 * Free an object which was previously allocated from this
3660 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3662 unsigned long flags
;
3663 cachep
= cache_from_obj(cachep
, objp
);
3667 local_irq_save(flags
);
3668 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3669 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3670 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3671 __cache_free(cachep
, objp
, _RET_IP_
);
3672 local_irq_restore(flags
);
3674 trace_kmem_cache_free(_RET_IP_
, objp
);
3676 EXPORT_SYMBOL(kmem_cache_free
);
3679 * kfree - free previously allocated memory
3680 * @objp: pointer returned by kmalloc.
3682 * If @objp is NULL, no operation is performed.
3684 * Don't free memory not originally allocated by kmalloc()
3685 * or you will run into trouble.
3687 void kfree(const void *objp
)
3689 struct kmem_cache
*c
;
3690 unsigned long flags
;
3692 trace_kfree(_RET_IP_
, objp
);
3694 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3696 local_irq_save(flags
);
3697 kfree_debugcheck(objp
);
3698 c
= virt_to_cache(objp
);
3699 debug_check_no_locks_freed(objp
, c
->object_size
);
3701 debug_check_no_obj_freed(objp
, c
->object_size
);
3702 __cache_free(c
, (void *)objp
, _RET_IP_
);
3703 local_irq_restore(flags
);
3705 EXPORT_SYMBOL(kfree
);
3708 * This initializes kmem_cache_node or resizes various caches for all nodes.
3710 static int alloc_kmem_cache_node(struct kmem_cache
*cachep
, gfp_t gfp
)
3713 struct kmem_cache_node
*n
;
3714 struct array_cache
*new_shared
;
3715 struct array_cache
**new_alien
= NULL
;
3717 for_each_online_node(node
) {
3719 if (use_alien_caches
) {
3720 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3726 if (cachep
->shared
) {
3727 new_shared
= alloc_arraycache(node
,
3728 cachep
->shared
*cachep
->batchcount
,
3731 free_alien_cache(new_alien
);
3736 n
= cachep
->node
[node
];
3738 struct array_cache
*shared
= n
->shared
;
3740 spin_lock_irq(&n
->list_lock
);
3743 free_block(cachep
, shared
->entry
,
3744 shared
->avail
, node
);
3746 n
->shared
= new_shared
;
3748 n
->alien
= new_alien
;
3751 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3752 cachep
->batchcount
+ cachep
->num
;
3753 spin_unlock_irq(&n
->list_lock
);
3755 free_alien_cache(new_alien
);
3758 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3760 free_alien_cache(new_alien
);
3765 kmem_cache_node_init(n
);
3766 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
3767 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
3768 n
->shared
= new_shared
;
3769 n
->alien
= new_alien
;
3770 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3771 cachep
->batchcount
+ cachep
->num
;
3772 cachep
->node
[node
] = n
;
3777 if (!cachep
->list
.next
) {
3778 /* Cache is not active yet. Roll back what we did */
3781 if (cachep
->node
[node
]) {
3782 n
= cachep
->node
[node
];
3785 free_alien_cache(n
->alien
);
3787 cachep
->node
[node
] = NULL
;
3795 struct ccupdate_struct
{
3796 struct kmem_cache
*cachep
;
3797 struct array_cache
*new[0];
3800 static void do_ccupdate_local(void *info
)
3802 struct ccupdate_struct
*new = info
;
3803 struct array_cache
*old
;
3806 old
= cpu_cache_get(new->cachep
);
3808 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3809 new->new[smp_processor_id()] = old
;
3812 /* Always called with the slab_mutex held */
3813 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3814 int batchcount
, int shared
, gfp_t gfp
)
3816 struct ccupdate_struct
*new;
3819 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3824 for_each_online_cpu(i
) {
3825 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3828 for (i
--; i
>= 0; i
--)
3834 new->cachep
= cachep
;
3836 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3839 cachep
->batchcount
= batchcount
;
3840 cachep
->limit
= limit
;
3841 cachep
->shared
= shared
;
3843 for_each_online_cpu(i
) {
3844 struct array_cache
*ccold
= new->new[i
];
3847 spin_lock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3848 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
3849 spin_unlock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3853 return alloc_kmem_cache_node(cachep
, gfp
);
3856 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3857 int batchcount
, int shared
, gfp_t gfp
)
3860 struct kmem_cache
*c
= NULL
;
3863 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3865 if (slab_state
< FULL
)
3868 if ((ret
< 0) || !is_root_cache(cachep
))
3871 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3872 for_each_memcg_cache_index(i
) {
3873 c
= cache_from_memcg_idx(cachep
, i
);
3875 /* return value determined by the parent cache only */
3876 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3882 /* Called with slab_mutex held always */
3883 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3890 if (!is_root_cache(cachep
)) {
3891 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3892 limit
= root
->limit
;
3893 shared
= root
->shared
;
3894 batchcount
= root
->batchcount
;
3897 if (limit
&& shared
&& batchcount
)
3900 * The head array serves three purposes:
3901 * - create a LIFO ordering, i.e. return objects that are cache-warm
3902 * - reduce the number of spinlock operations.
3903 * - reduce the number of linked list operations on the slab and
3904 * bufctl chains: array operations are cheaper.
3905 * The numbers are guessed, we should auto-tune as described by
3908 if (cachep
->size
> 131072)
3910 else if (cachep
->size
> PAGE_SIZE
)
3912 else if (cachep
->size
> 1024)
3914 else if (cachep
->size
> 256)
3920 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3921 * allocation behaviour: Most allocs on one cpu, most free operations
3922 * on another cpu. For these cases, an efficient object passing between
3923 * cpus is necessary. This is provided by a shared array. The array
3924 * replaces Bonwick's magazine layer.
3925 * On uniprocessor, it's functionally equivalent (but less efficient)
3926 * to a larger limit. Thus disabled by default.
3929 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3934 * With debugging enabled, large batchcount lead to excessively long
3935 * periods with disabled local interrupts. Limit the batchcount
3940 batchcount
= (limit
+ 1) / 2;
3942 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3944 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3945 cachep
->name
, -err
);
3950 * Drain an array if it contains any elements taking the node lock only if
3951 * necessary. Note that the node listlock also protects the array_cache
3952 * if drain_array() is used on the shared array.
3954 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3955 struct array_cache
*ac
, int force
, int node
)
3959 if (!ac
|| !ac
->avail
)
3961 if (ac
->touched
&& !force
) {
3964 spin_lock_irq(&n
->list_lock
);
3966 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3967 if (tofree
> ac
->avail
)
3968 tofree
= (ac
->avail
+ 1) / 2;
3969 free_block(cachep
, ac
->entry
, tofree
, node
);
3970 ac
->avail
-= tofree
;
3971 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3972 sizeof(void *) * ac
->avail
);
3974 spin_unlock_irq(&n
->list_lock
);
3979 * cache_reap - Reclaim memory from caches.
3980 * @w: work descriptor
3982 * Called from workqueue/eventd every few seconds.
3984 * - clear the per-cpu caches for this CPU.
3985 * - return freeable pages to the main free memory pool.
3987 * If we cannot acquire the cache chain mutex then just give up - we'll try
3988 * again on the next iteration.
3990 static void cache_reap(struct work_struct
*w
)
3992 struct kmem_cache
*searchp
;
3993 struct kmem_cache_node
*n
;
3994 int node
= numa_mem_id();
3995 struct delayed_work
*work
= to_delayed_work(w
);
3997 if (!mutex_trylock(&slab_mutex
))
3998 /* Give up. Setup the next iteration. */
4001 list_for_each_entry(searchp
, &slab_caches
, list
) {
4005 * We only take the node lock if absolutely necessary and we
4006 * have established with reasonable certainty that
4007 * we can do some work if the lock was obtained.
4009 n
= searchp
->node
[node
];
4011 reap_alien(searchp
, n
);
4013 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
4016 * These are racy checks but it does not matter
4017 * if we skip one check or scan twice.
4019 if (time_after(n
->next_reap
, jiffies
))
4022 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
4024 drain_array(searchp
, n
, n
->shared
, 0, node
);
4026 if (n
->free_touched
)
4027 n
->free_touched
= 0;
4031 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4032 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4033 STATS_ADD_REAPED(searchp
, freed
);
4039 mutex_unlock(&slab_mutex
);
4042 /* Set up the next iteration */
4043 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
4046 #ifdef CONFIG_SLABINFO
4047 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4050 unsigned long active_objs
;
4051 unsigned long num_objs
;
4052 unsigned long active_slabs
= 0;
4053 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4057 struct kmem_cache_node
*n
;
4061 for_each_online_node(node
) {
4062 n
= cachep
->node
[node
];
4067 spin_lock_irq(&n
->list_lock
);
4069 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4070 if (page
->active
!= cachep
->num
&& !error
)
4071 error
= "slabs_full accounting error";
4072 active_objs
+= cachep
->num
;
4075 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4076 if (page
->active
== cachep
->num
&& !error
)
4077 error
= "slabs_partial accounting error";
4078 if (!page
->active
&& !error
)
4079 error
= "slabs_partial accounting error";
4080 active_objs
+= page
->active
;
4083 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4084 if (page
->active
&& !error
)
4085 error
= "slabs_free accounting error";
4088 free_objects
+= n
->free_objects
;
4090 shared_avail
+= n
->shared
->avail
;
4092 spin_unlock_irq(&n
->list_lock
);
4094 num_slabs
+= active_slabs
;
4095 num_objs
= num_slabs
* cachep
->num
;
4096 if (num_objs
- active_objs
!= free_objects
&& !error
)
4097 error
= "free_objects accounting error";
4099 name
= cachep
->name
;
4101 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4103 sinfo
->active_objs
= active_objs
;
4104 sinfo
->num_objs
= num_objs
;
4105 sinfo
->active_slabs
= active_slabs
;
4106 sinfo
->num_slabs
= num_slabs
;
4107 sinfo
->shared_avail
= shared_avail
;
4108 sinfo
->limit
= cachep
->limit
;
4109 sinfo
->batchcount
= cachep
->batchcount
;
4110 sinfo
->shared
= cachep
->shared
;
4111 sinfo
->objects_per_slab
= cachep
->num
;
4112 sinfo
->cache_order
= cachep
->gfporder
;
4115 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4119 unsigned long high
= cachep
->high_mark
;
4120 unsigned long allocs
= cachep
->num_allocations
;
4121 unsigned long grown
= cachep
->grown
;
4122 unsigned long reaped
= cachep
->reaped
;
4123 unsigned long errors
= cachep
->errors
;
4124 unsigned long max_freeable
= cachep
->max_freeable
;
4125 unsigned long node_allocs
= cachep
->node_allocs
;
4126 unsigned long node_frees
= cachep
->node_frees
;
4127 unsigned long overflows
= cachep
->node_overflow
;
4129 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4130 "%4lu %4lu %4lu %4lu %4lu",
4131 allocs
, high
, grown
,
4132 reaped
, errors
, max_freeable
, node_allocs
,
4133 node_frees
, overflows
);
4137 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4138 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4139 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4140 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4142 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4143 allochit
, allocmiss
, freehit
, freemiss
);
4148 #define MAX_SLABINFO_WRITE 128
4150 * slabinfo_write - Tuning for the slab allocator
4152 * @buffer: user buffer
4153 * @count: data length
4156 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4157 size_t count
, loff_t
*ppos
)
4159 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4160 int limit
, batchcount
, shared
, res
;
4161 struct kmem_cache
*cachep
;
4163 if (count
> MAX_SLABINFO_WRITE
)
4165 if (copy_from_user(&kbuf
, buffer
, count
))
4167 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4169 tmp
= strchr(kbuf
, ' ');
4174 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4177 /* Find the cache in the chain of caches. */
4178 mutex_lock(&slab_mutex
);
4180 list_for_each_entry(cachep
, &slab_caches
, list
) {
4181 if (!strcmp(cachep
->name
, kbuf
)) {
4182 if (limit
< 1 || batchcount
< 1 ||
4183 batchcount
> limit
|| shared
< 0) {
4186 res
= do_tune_cpucache(cachep
, limit
,
4193 mutex_unlock(&slab_mutex
);
4199 #ifdef CONFIG_DEBUG_SLAB_LEAK
4201 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4203 mutex_lock(&slab_mutex
);
4204 return seq_list_start(&slab_caches
, *pos
);
4207 static inline int add_caller(unsigned long *n
, unsigned long v
)
4217 unsigned long *q
= p
+ 2 * i
;
4231 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4237 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4245 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4248 for (j
= page
->active
; j
< c
->num
; j
++) {
4249 /* Skip freed item */
4250 if (get_free_obj(page
, j
) == i
) {
4258 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4263 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4265 #ifdef CONFIG_KALLSYMS
4266 unsigned long offset
, size
;
4267 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4269 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4270 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4272 seq_printf(m
, " [%s]", modname
);
4276 seq_printf(m
, "%p", (void *)address
);
4279 static int leaks_show(struct seq_file
*m
, void *p
)
4281 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4283 struct kmem_cache_node
*n
;
4285 unsigned long *x
= m
->private;
4289 if (!(cachep
->flags
& SLAB_STORE_USER
))
4291 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4294 /* OK, we can do it */
4298 for_each_online_node(node
) {
4299 n
= cachep
->node
[node
];
4304 spin_lock_irq(&n
->list_lock
);
4306 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4307 handle_slab(x
, cachep
, page
);
4308 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4309 handle_slab(x
, cachep
, page
);
4310 spin_unlock_irq(&n
->list_lock
);
4312 name
= cachep
->name
;
4314 /* Increase the buffer size */
4315 mutex_unlock(&slab_mutex
);
4316 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4318 /* Too bad, we are really out */
4320 mutex_lock(&slab_mutex
);
4323 *(unsigned long *)m
->private = x
[0] * 2;
4325 mutex_lock(&slab_mutex
);
4326 /* Now make sure this entry will be retried */
4330 for (i
= 0; i
< x
[1]; i
++) {
4331 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4332 show_symbol(m
, x
[2*i
+2]);
4339 static const struct seq_operations slabstats_op
= {
4340 .start
= leaks_start
,
4346 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4348 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4351 ret
= seq_open(file
, &slabstats_op
);
4353 struct seq_file
*m
= file
->private_data
;
4354 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4363 static const struct file_operations proc_slabstats_operations
= {
4364 .open
= slabstats_open
,
4366 .llseek
= seq_lseek
,
4367 .release
= seq_release_private
,
4371 static int __init
slab_proc_init(void)
4373 #ifdef CONFIG_DEBUG_SLAB_LEAK
4374 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4378 module_init(slab_proc_init
);
4382 * ksize - get the actual amount of memory allocated for a given object
4383 * @objp: Pointer to the object
4385 * kmalloc may internally round up allocations and return more memory
4386 * than requested. ksize() can be used to determine the actual amount of
4387 * memory allocated. The caller may use this additional memory, even though
4388 * a smaller amount of memory was initially specified with the kmalloc call.
4389 * The caller must guarantee that objp points to a valid object previously
4390 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4391 * must not be freed during the duration of the call.
4393 size_t ksize(const void *objp
)
4396 if (unlikely(objp
== ZERO_SIZE_PTR
))
4399 return virt_to_cache(objp
)->object_size
;
4401 EXPORT_SYMBOL(ksize
);