3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly
;
166 #define SLAB_LIMIT (((unsigned int)(~0U))-1)
171 * Manages the objs in a slab. Placed either at the beginning of mem allocated
172 * for a slab, or allocated from an general cache.
173 * Slabs are chained into three list: fully used, partial, fully free slabs.
177 struct list_head list
;
178 void *s_mem
; /* including colour offset */
179 unsigned int inuse
; /* num of objs active in slab */
188 * - LIFO ordering, to hand out cache-warm objects from _alloc
189 * - reduce the number of linked list operations
190 * - reduce spinlock operations
192 * The limit is stored in the per-cpu structure to reduce the data cache
199 unsigned int batchcount
;
200 unsigned int touched
;
203 * Must have this definition in here for the proper
204 * alignment of array_cache. Also simplifies accessing
207 * Entries should not be directly dereferenced as
208 * entries belonging to slabs marked pfmemalloc will
209 * have the lower bits set SLAB_OBJ_PFMEMALLOC
213 #define SLAB_OBJ_PFMEMALLOC 1
214 static inline bool is_obj_pfmemalloc(void *objp
)
216 return (unsigned long)objp
& SLAB_OBJ_PFMEMALLOC
;
219 static inline void set_obj_pfmemalloc(void **objp
)
221 *objp
= (void *)((unsigned long)*objp
| SLAB_OBJ_PFMEMALLOC
);
225 static inline void clear_obj_pfmemalloc(void **objp
)
227 *objp
= (void *)((unsigned long)*objp
& ~SLAB_OBJ_PFMEMALLOC
);
231 * bootstrap: The caches do not work without cpuarrays anymore, but the
232 * cpuarrays are allocated from the generic caches...
234 #define BOOT_CPUCACHE_ENTRIES 1
235 struct arraycache_init
{
236 struct array_cache cache
;
237 void *entries
[BOOT_CPUCACHE_ENTRIES
];
241 * Need this for bootstrapping a per node allocator.
243 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
244 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
245 #define CACHE_CACHE 0
246 #define SIZE_AC MAX_NUMNODES
247 #define SIZE_NODE (2 * MAX_NUMNODES)
249 static int drain_freelist(struct kmem_cache
*cache
,
250 struct kmem_cache_node
*n
, int tofree
);
251 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
253 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
254 static void cache_reap(struct work_struct
*unused
);
256 static int slab_early_init
= 1;
258 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
259 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
261 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
263 INIT_LIST_HEAD(&parent
->slabs_full
);
264 INIT_LIST_HEAD(&parent
->slabs_partial
);
265 INIT_LIST_HEAD(&parent
->slabs_free
);
266 parent
->shared
= NULL
;
267 parent
->alien
= NULL
;
268 parent
->colour_next
= 0;
269 spin_lock_init(&parent
->list_lock
);
270 parent
->free_objects
= 0;
271 parent
->free_touched
= 0;
274 #define MAKE_LIST(cachep, listp, slab, nodeid) \
276 INIT_LIST_HEAD(listp); \
277 list_splice(&(cachep->node[nodeid]->slab), listp); \
280 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
282 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
283 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
284 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
287 #define CFLGS_OFF_SLAB (0x80000000UL)
288 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
290 #define BATCHREFILL_LIMIT 16
292 * Optimization question: fewer reaps means less probability for unnessary
293 * cpucache drain/refill cycles.
295 * OTOH the cpuarrays can contain lots of objects,
296 * which could lock up otherwise freeable slabs.
298 #define REAPTIMEOUT_CPUC (2*HZ)
299 #define REAPTIMEOUT_LIST3 (4*HZ)
302 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
303 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
304 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
305 #define STATS_INC_GROWN(x) ((x)->grown++)
306 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
307 #define STATS_SET_HIGH(x) \
309 if ((x)->num_active > (x)->high_mark) \
310 (x)->high_mark = (x)->num_active; \
312 #define STATS_INC_ERR(x) ((x)->errors++)
313 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
314 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
315 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
316 #define STATS_SET_FREEABLE(x, i) \
318 if ((x)->max_freeable < i) \
319 (x)->max_freeable = i; \
321 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
322 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
323 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
324 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
326 #define STATS_INC_ACTIVE(x) do { } while (0)
327 #define STATS_DEC_ACTIVE(x) do { } while (0)
328 #define STATS_INC_ALLOCED(x) do { } while (0)
329 #define STATS_INC_GROWN(x) do { } while (0)
330 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
331 #define STATS_SET_HIGH(x) do { } while (0)
332 #define STATS_INC_ERR(x) do { } while (0)
333 #define STATS_INC_NODEALLOCS(x) do { } while (0)
334 #define STATS_INC_NODEFREES(x) do { } while (0)
335 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
336 #define STATS_SET_FREEABLE(x, i) do { } while (0)
337 #define STATS_INC_ALLOCHIT(x) do { } while (0)
338 #define STATS_INC_ALLOCMISS(x) do { } while (0)
339 #define STATS_INC_FREEHIT(x) do { } while (0)
340 #define STATS_INC_FREEMISS(x) do { } while (0)
346 * memory layout of objects:
348 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
349 * the end of an object is aligned with the end of the real
350 * allocation. Catches writes behind the end of the allocation.
351 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
353 * cachep->obj_offset: The real object.
354 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
355 * cachep->size - 1* BYTES_PER_WORD: last caller address
356 * [BYTES_PER_WORD long]
358 static int obj_offset(struct kmem_cache
*cachep
)
360 return cachep
->obj_offset
;
363 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
365 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
366 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
367 sizeof(unsigned long long));
370 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
372 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
373 if (cachep
->flags
& SLAB_STORE_USER
)
374 return (unsigned long long *)(objp
+ cachep
->size
-
375 sizeof(unsigned long long) -
377 return (unsigned long long *) (objp
+ cachep
->size
-
378 sizeof(unsigned long long));
381 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
383 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
384 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
389 #define obj_offset(x) 0
390 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
391 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
392 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
397 * Do not go above this order unless 0 objects fit into the slab or
398 * overridden on the command line.
400 #define SLAB_MAX_ORDER_HI 1
401 #define SLAB_MAX_ORDER_LO 0
402 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
403 static bool slab_max_order_set __initdata
;
405 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
407 struct page
*page
= virt_to_head_page(obj
);
408 return page
->slab_cache
;
411 static inline struct slab
*virt_to_slab(const void *obj
)
413 struct page
*page
= virt_to_head_page(obj
);
415 VM_BUG_ON(!PageSlab(page
));
416 return page
->slab_page
;
419 static inline void *index_to_obj(struct kmem_cache
*cache
, struct slab
*slab
,
422 return slab
->s_mem
+ cache
->size
* idx
;
426 * We want to avoid an expensive divide : (offset / cache->size)
427 * Using the fact that size is a constant for a particular cache,
428 * we can replace (offset / cache->size) by
429 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
431 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
432 const struct slab
*slab
, void *obj
)
434 u32 offset
= (obj
- slab
->s_mem
);
435 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
438 static struct arraycache_init initarray_generic
=
439 { {0, BOOT_CPUCACHE_ENTRIES
, 1, 0} };
441 /* internal cache of cache description objs */
442 static struct kmem_cache kmem_cache_boot
= {
444 .limit
= BOOT_CPUCACHE_ENTRIES
,
446 .size
= sizeof(struct kmem_cache
),
447 .name
= "kmem_cache",
450 #define BAD_ALIEN_MAGIC 0x01020304ul
452 #ifdef CONFIG_LOCKDEP
455 * Slab sometimes uses the kmalloc slabs to store the slab headers
456 * for other slabs "off slab".
457 * The locking for this is tricky in that it nests within the locks
458 * of all other slabs in a few places; to deal with this special
459 * locking we put on-slab caches into a separate lock-class.
461 * We set lock class for alien array caches which are up during init.
462 * The lock annotation will be lost if all cpus of a node goes down and
463 * then comes back up during hotplug
465 static struct lock_class_key on_slab_l3_key
;
466 static struct lock_class_key on_slab_alc_key
;
468 static struct lock_class_key debugobj_l3_key
;
469 static struct lock_class_key debugobj_alc_key
;
471 static void slab_set_lock_classes(struct kmem_cache
*cachep
,
472 struct lock_class_key
*l3_key
, struct lock_class_key
*alc_key
,
475 struct array_cache
**alc
;
476 struct kmem_cache_node
*n
;
483 lockdep_set_class(&n
->list_lock
, l3_key
);
486 * FIXME: This check for BAD_ALIEN_MAGIC
487 * should go away when common slab code is taught to
488 * work even without alien caches.
489 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
490 * for alloc_alien_cache,
492 if (!alc
|| (unsigned long)alc
== BAD_ALIEN_MAGIC
)
496 lockdep_set_class(&alc
[r
]->lock
, alc_key
);
500 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
502 slab_set_lock_classes(cachep
, &debugobj_l3_key
, &debugobj_alc_key
, node
);
505 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
509 for_each_online_node(node
)
510 slab_set_debugobj_lock_classes_node(cachep
, node
);
513 static void init_node_lock_keys(int q
)
520 for (i
= 1; i
<= KMALLOC_SHIFT_HIGH
; i
++) {
521 struct kmem_cache_node
*n
;
522 struct kmem_cache
*cache
= kmalloc_caches
[i
];
528 if (!n
|| OFF_SLAB(cache
))
531 slab_set_lock_classes(cache
, &on_slab_l3_key
,
532 &on_slab_alc_key
, q
);
536 static void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int q
)
538 if (!cachep
->node
[q
])
541 slab_set_lock_classes(cachep
, &on_slab_l3_key
,
542 &on_slab_alc_key
, q
);
545 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
549 VM_BUG_ON(OFF_SLAB(cachep
));
551 on_slab_lock_classes_node(cachep
, node
);
554 static inline void init_lock_keys(void)
559 init_node_lock_keys(node
);
562 static void init_node_lock_keys(int q
)
566 static inline void init_lock_keys(void)
570 static inline void on_slab_lock_classes(struct kmem_cache
*cachep
)
574 static inline void on_slab_lock_classes_node(struct kmem_cache
*cachep
, int node
)
578 static void slab_set_debugobj_lock_classes_node(struct kmem_cache
*cachep
, int node
)
582 static void slab_set_debugobj_lock_classes(struct kmem_cache
*cachep
)
587 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
589 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
591 return cachep
->array
[smp_processor_id()];
594 static size_t slab_mgmt_size(size_t nr_objs
, size_t align
)
596 return ALIGN(sizeof(struct slab
)+nr_objs
*sizeof(unsigned int), align
);
600 * Calculate the number of objects and left-over bytes for a given buffer size.
602 static void cache_estimate(unsigned long gfporder
, size_t buffer_size
,
603 size_t align
, int flags
, size_t *left_over
,
608 size_t slab_size
= PAGE_SIZE
<< gfporder
;
611 * The slab management structure can be either off the slab or
612 * on it. For the latter case, the memory allocated for a
616 * - One unsigned int for each object
617 * - Padding to respect alignment of @align
618 * - @buffer_size bytes for each object
620 * If the slab management structure is off the slab, then the
621 * alignment will already be calculated into the size. Because
622 * the slabs are all pages aligned, the objects will be at the
623 * correct alignment when allocated.
625 if (flags
& CFLGS_OFF_SLAB
) {
627 nr_objs
= slab_size
/ buffer_size
;
629 if (nr_objs
> SLAB_LIMIT
)
630 nr_objs
= SLAB_LIMIT
;
633 * Ignore padding for the initial guess. The padding
634 * is at most @align-1 bytes, and @buffer_size is at
635 * least @align. In the worst case, this result will
636 * be one greater than the number of objects that fit
637 * into the memory allocation when taking the padding
640 nr_objs
= (slab_size
- sizeof(struct slab
)) /
641 (buffer_size
+ sizeof(unsigned int));
644 * This calculated number will be either the right
645 * amount, or one greater than what we want.
647 if (slab_mgmt_size(nr_objs
, align
) + nr_objs
*buffer_size
651 if (nr_objs
> SLAB_LIMIT
)
652 nr_objs
= SLAB_LIMIT
;
654 mgmt_size
= slab_mgmt_size(nr_objs
, align
);
657 *left_over
= slab_size
- nr_objs
*buffer_size
- mgmt_size
;
661 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
663 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
666 printk(KERN_ERR
"slab error in %s(): cache `%s': %s\n",
667 function
, cachep
->name
, msg
);
669 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
674 * By default on NUMA we use alien caches to stage the freeing of
675 * objects allocated from other nodes. This causes massive memory
676 * inefficiencies when using fake NUMA setup to split memory into a
677 * large number of small nodes, so it can be disabled on the command
681 static int use_alien_caches __read_mostly
= 1;
682 static int __init
noaliencache_setup(char *s
)
684 use_alien_caches
= 0;
687 __setup("noaliencache", noaliencache_setup
);
689 static int __init
slab_max_order_setup(char *str
)
691 get_option(&str
, &slab_max_order
);
692 slab_max_order
= slab_max_order
< 0 ? 0 :
693 min(slab_max_order
, MAX_ORDER
- 1);
694 slab_max_order_set
= true;
698 __setup("slab_max_order=", slab_max_order_setup
);
702 * Special reaping functions for NUMA systems called from cache_reap().
703 * These take care of doing round robin flushing of alien caches (containing
704 * objects freed on different nodes from which they were allocated) and the
705 * flushing of remote pcps by calling drain_node_pages.
707 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
709 static void init_reap_node(int cpu
)
713 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
714 if (node
== MAX_NUMNODES
)
715 node
= first_node(node_online_map
);
717 per_cpu(slab_reap_node
, cpu
) = node
;
720 static void next_reap_node(void)
722 int node
= __this_cpu_read(slab_reap_node
);
724 node
= next_node(node
, node_online_map
);
725 if (unlikely(node
>= MAX_NUMNODES
))
726 node
= first_node(node_online_map
);
727 __this_cpu_write(slab_reap_node
, node
);
731 #define init_reap_node(cpu) do { } while (0)
732 #define next_reap_node(void) do { } while (0)
736 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
737 * via the workqueue/eventd.
738 * Add the CPU number into the expiration time to minimize the possibility of
739 * the CPUs getting into lockstep and contending for the global cache chain
742 static void start_cpu_timer(int cpu
)
744 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
747 * When this gets called from do_initcalls via cpucache_init(),
748 * init_workqueues() has already run, so keventd will be setup
751 if (keventd_up() && reap_work
->work
.func
== NULL
) {
753 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
754 schedule_delayed_work_on(cpu
, reap_work
,
755 __round_jiffies_relative(HZ
, cpu
));
759 static struct array_cache
*alloc_arraycache(int node
, int entries
,
760 int batchcount
, gfp_t gfp
)
762 int memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
763 struct array_cache
*nc
= NULL
;
765 nc
= kmalloc_node(memsize
, gfp
, node
);
767 * The array_cache structures contain pointers to free object.
768 * However, when such objects are allocated or transferred to another
769 * cache the pointers are not cleared and they could be counted as
770 * valid references during a kmemleak scan. Therefore, kmemleak must
771 * not scan such objects.
773 kmemleak_no_scan(nc
);
777 nc
->batchcount
= batchcount
;
779 spin_lock_init(&nc
->lock
);
784 static inline bool is_slab_pfmemalloc(struct slab
*slabp
)
786 struct page
*page
= virt_to_page(slabp
->s_mem
);
788 return PageSlabPfmemalloc(page
);
791 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
792 static void recheck_pfmemalloc_active(struct kmem_cache
*cachep
,
793 struct array_cache
*ac
)
795 struct kmem_cache_node
*n
= cachep
->node
[numa_mem_id()];
799 if (!pfmemalloc_active
)
802 spin_lock_irqsave(&n
->list_lock
, flags
);
803 list_for_each_entry(slabp
, &n
->slabs_full
, list
)
804 if (is_slab_pfmemalloc(slabp
))
807 list_for_each_entry(slabp
, &n
->slabs_partial
, list
)
808 if (is_slab_pfmemalloc(slabp
))
811 list_for_each_entry(slabp
, &n
->slabs_free
, list
)
812 if (is_slab_pfmemalloc(slabp
))
815 pfmemalloc_active
= false;
817 spin_unlock_irqrestore(&n
->list_lock
, flags
);
820 static void *__ac_get_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
821 gfp_t flags
, bool force_refill
)
824 void *objp
= ac
->entry
[--ac
->avail
];
826 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
827 if (unlikely(is_obj_pfmemalloc(objp
))) {
828 struct kmem_cache_node
*n
;
830 if (gfp_pfmemalloc_allowed(flags
)) {
831 clear_obj_pfmemalloc(&objp
);
835 /* The caller cannot use PFMEMALLOC objects, find another one */
836 for (i
= 0; i
< ac
->avail
; i
++) {
837 /* If a !PFMEMALLOC object is found, swap them */
838 if (!is_obj_pfmemalloc(ac
->entry
[i
])) {
840 ac
->entry
[i
] = ac
->entry
[ac
->avail
];
841 ac
->entry
[ac
->avail
] = objp
;
847 * If there are empty slabs on the slabs_free list and we are
848 * being forced to refill the cache, mark this one !pfmemalloc.
850 n
= cachep
->node
[numa_mem_id()];
851 if (!list_empty(&n
->slabs_free
) && force_refill
) {
852 struct slab
*slabp
= virt_to_slab(objp
);
853 ClearPageSlabPfmemalloc(virt_to_head_page(slabp
->s_mem
));
854 clear_obj_pfmemalloc(&objp
);
855 recheck_pfmemalloc_active(cachep
, ac
);
859 /* No !PFMEMALLOC objects available */
867 static inline void *ac_get_obj(struct kmem_cache
*cachep
,
868 struct array_cache
*ac
, gfp_t flags
, bool force_refill
)
872 if (unlikely(sk_memalloc_socks()))
873 objp
= __ac_get_obj(cachep
, ac
, flags
, force_refill
);
875 objp
= ac
->entry
[--ac
->avail
];
880 static void *__ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
883 if (unlikely(pfmemalloc_active
)) {
884 /* Some pfmemalloc slabs exist, check if this is one */
885 struct slab
*slabp
= virt_to_slab(objp
);
886 struct page
*page
= virt_to_head_page(slabp
->s_mem
);
887 if (PageSlabPfmemalloc(page
))
888 set_obj_pfmemalloc(&objp
);
894 static inline void ac_put_obj(struct kmem_cache
*cachep
, struct array_cache
*ac
,
897 if (unlikely(sk_memalloc_socks()))
898 objp
= __ac_put_obj(cachep
, ac
, objp
);
900 ac
->entry
[ac
->avail
++] = objp
;
904 * Transfer objects in one arraycache to another.
905 * Locking must be handled by the caller.
907 * Return the number of entries transferred.
909 static int transfer_objects(struct array_cache
*to
,
910 struct array_cache
*from
, unsigned int max
)
912 /* Figure out how many entries to transfer */
913 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
918 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
928 #define drain_alien_cache(cachep, alien) do { } while (0)
929 #define reap_alien(cachep, n) do { } while (0)
931 static inline struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
933 return (struct array_cache
**)BAD_ALIEN_MAGIC
;
936 static inline void free_alien_cache(struct array_cache
**ac_ptr
)
940 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
945 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
951 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
952 gfp_t flags
, int nodeid
)
957 #else /* CONFIG_NUMA */
959 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
960 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
962 static struct array_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
964 struct array_cache
**ac_ptr
;
965 int memsize
= sizeof(void *) * nr_node_ids
;
970 ac_ptr
= kzalloc_node(memsize
, gfp
, node
);
973 if (i
== node
|| !node_online(i
))
975 ac_ptr
[i
] = alloc_arraycache(node
, limit
, 0xbaadf00d, gfp
);
977 for (i
--; i
>= 0; i
--)
987 static void free_alien_cache(struct array_cache
**ac_ptr
)
998 static void __drain_alien_cache(struct kmem_cache
*cachep
,
999 struct array_cache
*ac
, int node
)
1001 struct kmem_cache_node
*n
= cachep
->node
[node
];
1004 spin_lock(&n
->list_lock
);
1006 * Stuff objects into the remote nodes shared array first.
1007 * That way we could avoid the overhead of putting the objects
1008 * into the free lists and getting them back later.
1011 transfer_objects(n
->shared
, ac
, ac
->limit
);
1013 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
1015 spin_unlock(&n
->list_lock
);
1020 * Called from cache_reap() to regularly drain alien caches round robin.
1022 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
1024 int node
= __this_cpu_read(slab_reap_node
);
1027 struct array_cache
*ac
= n
->alien
[node
];
1029 if (ac
&& ac
->avail
&& spin_trylock_irq(&ac
->lock
)) {
1030 __drain_alien_cache(cachep
, ac
, node
);
1031 spin_unlock_irq(&ac
->lock
);
1036 static void drain_alien_cache(struct kmem_cache
*cachep
,
1037 struct array_cache
**alien
)
1040 struct array_cache
*ac
;
1041 unsigned long flags
;
1043 for_each_online_node(i
) {
1046 spin_lock_irqsave(&ac
->lock
, flags
);
1047 __drain_alien_cache(cachep
, ac
, i
);
1048 spin_unlock_irqrestore(&ac
->lock
, flags
);
1053 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
1055 int nodeid
= page_to_nid(virt_to_page(objp
));
1056 struct kmem_cache_node
*n
;
1057 struct array_cache
*alien
= NULL
;
1060 node
= numa_mem_id();
1063 * Make sure we are not freeing a object from another node to the array
1064 * cache on this cpu.
1066 if (likely(nodeid
== node
))
1069 n
= cachep
->node
[node
];
1070 STATS_INC_NODEFREES(cachep
);
1071 if (n
->alien
&& n
->alien
[nodeid
]) {
1072 alien
= n
->alien
[nodeid
];
1073 spin_lock(&alien
->lock
);
1074 if (unlikely(alien
->avail
== alien
->limit
)) {
1075 STATS_INC_ACOVERFLOW(cachep
);
1076 __drain_alien_cache(cachep
, alien
, nodeid
);
1078 ac_put_obj(cachep
, alien
, objp
);
1079 spin_unlock(&alien
->lock
);
1081 spin_lock(&(cachep
->node
[nodeid
])->list_lock
);
1082 free_block(cachep
, &objp
, 1, nodeid
);
1083 spin_unlock(&(cachep
->node
[nodeid
])->list_lock
);
1090 * Allocates and initializes node for a node on each slab cache, used for
1091 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1092 * will be allocated off-node since memory is not yet online for the new node.
1093 * When hotplugging memory or a cpu, existing node are not replaced if
1096 * Must hold slab_mutex.
1098 static int init_cache_node_node(int node
)
1100 struct kmem_cache
*cachep
;
1101 struct kmem_cache_node
*n
;
1102 const int memsize
= sizeof(struct kmem_cache_node
);
1104 list_for_each_entry(cachep
, &slab_caches
, list
) {
1106 * Set up the size64 kmemlist for cpu before we can
1107 * begin anything. Make sure some other cpu on this
1108 * node has not already allocated this
1110 if (!cachep
->node
[node
]) {
1111 n
= kmalloc_node(memsize
, GFP_KERNEL
, node
);
1114 kmem_cache_node_init(n
);
1115 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
1116 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1119 * The l3s don't come and go as CPUs come and
1120 * go. slab_mutex is sufficient
1123 cachep
->node
[node
] = n
;
1126 spin_lock_irq(&cachep
->node
[node
]->list_lock
);
1127 cachep
->node
[node
]->free_limit
=
1128 (1 + nr_cpus_node(node
)) *
1129 cachep
->batchcount
+ cachep
->num
;
1130 spin_unlock_irq(&cachep
->node
[node
]->list_lock
);
1135 static inline int slabs_tofree(struct kmem_cache
*cachep
,
1136 struct kmem_cache_node
*n
)
1138 return (n
->free_objects
+ cachep
->num
- 1) / cachep
->num
;
1141 static void cpuup_canceled(long cpu
)
1143 struct kmem_cache
*cachep
;
1144 struct kmem_cache_node
*n
= NULL
;
1145 int node
= cpu_to_mem(cpu
);
1146 const struct cpumask
*mask
= cpumask_of_node(node
);
1148 list_for_each_entry(cachep
, &slab_caches
, list
) {
1149 struct array_cache
*nc
;
1150 struct array_cache
*shared
;
1151 struct array_cache
**alien
;
1153 /* cpu is dead; no one can alloc from it. */
1154 nc
= cachep
->array
[cpu
];
1155 cachep
->array
[cpu
] = NULL
;
1156 n
= cachep
->node
[node
];
1159 goto free_array_cache
;
1161 spin_lock_irq(&n
->list_lock
);
1163 /* Free limit for this kmem_cache_node */
1164 n
->free_limit
-= cachep
->batchcount
;
1166 free_block(cachep
, nc
->entry
, nc
->avail
, node
);
1168 if (!cpumask_empty(mask
)) {
1169 spin_unlock_irq(&n
->list_lock
);
1170 goto free_array_cache
;
1175 free_block(cachep
, shared
->entry
,
1176 shared
->avail
, node
);
1183 spin_unlock_irq(&n
->list_lock
);
1187 drain_alien_cache(cachep
, alien
);
1188 free_alien_cache(alien
);
1194 * In the previous loop, all the objects were freed to
1195 * the respective cache's slabs, now we can go ahead and
1196 * shrink each nodelist to its limit.
1198 list_for_each_entry(cachep
, &slab_caches
, list
) {
1199 n
= cachep
->node
[node
];
1202 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1206 static int cpuup_prepare(long cpu
)
1208 struct kmem_cache
*cachep
;
1209 struct kmem_cache_node
*n
= NULL
;
1210 int node
= cpu_to_mem(cpu
);
1214 * We need to do this right in the beginning since
1215 * alloc_arraycache's are going to use this list.
1216 * kmalloc_node allows us to add the slab to the right
1217 * kmem_cache_node and not this cpu's kmem_cache_node
1219 err
= init_cache_node_node(node
);
1224 * Now we can go ahead with allocating the shared arrays and
1227 list_for_each_entry(cachep
, &slab_caches
, list
) {
1228 struct array_cache
*nc
;
1229 struct array_cache
*shared
= NULL
;
1230 struct array_cache
**alien
= NULL
;
1232 nc
= alloc_arraycache(node
, cachep
->limit
,
1233 cachep
->batchcount
, GFP_KERNEL
);
1236 if (cachep
->shared
) {
1237 shared
= alloc_arraycache(node
,
1238 cachep
->shared
* cachep
->batchcount
,
1239 0xbaadf00d, GFP_KERNEL
);
1245 if (use_alien_caches
) {
1246 alien
= alloc_alien_cache(node
, cachep
->limit
, GFP_KERNEL
);
1253 cachep
->array
[cpu
] = nc
;
1254 n
= cachep
->node
[node
];
1257 spin_lock_irq(&n
->list_lock
);
1260 * We are serialised from CPU_DEAD or
1261 * CPU_UP_CANCELLED by the cpucontrol lock
1272 spin_unlock_irq(&n
->list_lock
);
1274 free_alien_cache(alien
);
1275 if (cachep
->flags
& SLAB_DEBUG_OBJECTS
)
1276 slab_set_debugobj_lock_classes_node(cachep
, node
);
1277 else if (!OFF_SLAB(cachep
) &&
1278 !(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1279 on_slab_lock_classes_node(cachep
, node
);
1281 init_node_lock_keys(node
);
1285 cpuup_canceled(cpu
);
1289 static int cpuup_callback(struct notifier_block
*nfb
,
1290 unsigned long action
, void *hcpu
)
1292 long cpu
= (long)hcpu
;
1296 case CPU_UP_PREPARE
:
1297 case CPU_UP_PREPARE_FROZEN
:
1298 mutex_lock(&slab_mutex
);
1299 err
= cpuup_prepare(cpu
);
1300 mutex_unlock(&slab_mutex
);
1303 case CPU_ONLINE_FROZEN
:
1304 start_cpu_timer(cpu
);
1306 #ifdef CONFIG_HOTPLUG_CPU
1307 case CPU_DOWN_PREPARE
:
1308 case CPU_DOWN_PREPARE_FROZEN
:
1310 * Shutdown cache reaper. Note that the slab_mutex is
1311 * held so that if cache_reap() is invoked it cannot do
1312 * anything expensive but will only modify reap_work
1313 * and reschedule the timer.
1315 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1316 /* Now the cache_reaper is guaranteed to be not running. */
1317 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1319 case CPU_DOWN_FAILED
:
1320 case CPU_DOWN_FAILED_FROZEN
:
1321 start_cpu_timer(cpu
);
1324 case CPU_DEAD_FROZEN
:
1326 * Even if all the cpus of a node are down, we don't free the
1327 * kmem_cache_node of any cache. This to avoid a race between
1328 * cpu_down, and a kmalloc allocation from another cpu for
1329 * memory from the node of the cpu going down. The node
1330 * structure is usually allocated from kmem_cache_create() and
1331 * gets destroyed at kmem_cache_destroy().
1335 case CPU_UP_CANCELED
:
1336 case CPU_UP_CANCELED_FROZEN
:
1337 mutex_lock(&slab_mutex
);
1338 cpuup_canceled(cpu
);
1339 mutex_unlock(&slab_mutex
);
1342 return notifier_from_errno(err
);
1345 static struct notifier_block cpucache_notifier
= {
1346 &cpuup_callback
, NULL
, 0
1349 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1351 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1352 * Returns -EBUSY if all objects cannot be drained so that the node is not
1355 * Must hold slab_mutex.
1357 static int __meminit
drain_cache_node_node(int node
)
1359 struct kmem_cache
*cachep
;
1362 list_for_each_entry(cachep
, &slab_caches
, list
) {
1363 struct kmem_cache_node
*n
;
1365 n
= cachep
->node
[node
];
1369 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
1371 if (!list_empty(&n
->slabs_full
) ||
1372 !list_empty(&n
->slabs_partial
)) {
1380 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1381 unsigned long action
, void *arg
)
1383 struct memory_notify
*mnb
= arg
;
1387 nid
= mnb
->status_change_nid
;
1392 case MEM_GOING_ONLINE
:
1393 mutex_lock(&slab_mutex
);
1394 ret
= init_cache_node_node(nid
);
1395 mutex_unlock(&slab_mutex
);
1397 case MEM_GOING_OFFLINE
:
1398 mutex_lock(&slab_mutex
);
1399 ret
= drain_cache_node_node(nid
);
1400 mutex_unlock(&slab_mutex
);
1404 case MEM_CANCEL_ONLINE
:
1405 case MEM_CANCEL_OFFLINE
:
1409 return notifier_from_errno(ret
);
1411 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1414 * swap the static kmem_cache_node with kmalloced memory
1416 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1419 struct kmem_cache_node
*ptr
;
1421 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1424 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1426 * Do not assume that spinlocks can be initialized via memcpy:
1428 spin_lock_init(&ptr
->list_lock
);
1430 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1431 cachep
->node
[nodeid
] = ptr
;
1435 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1436 * size of kmem_cache_node.
1438 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1442 for_each_online_node(node
) {
1443 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1444 cachep
->node
[node
]->next_reap
= jiffies
+
1446 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
1451 * The memory after the last cpu cache pointer is used for the
1454 static void setup_node_pointer(struct kmem_cache
*cachep
)
1456 cachep
->node
= (struct kmem_cache_node
**)&cachep
->array
[nr_cpu_ids
];
1460 * Initialisation. Called after the page allocator have been initialised and
1461 * before smp_init().
1463 void __init
kmem_cache_init(void)
1467 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1468 sizeof(struct rcu_head
));
1469 kmem_cache
= &kmem_cache_boot
;
1470 setup_node_pointer(kmem_cache
);
1472 if (num_possible_nodes() == 1)
1473 use_alien_caches
= 0;
1475 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1476 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1478 set_up_node(kmem_cache
, CACHE_CACHE
);
1481 * Fragmentation resistance on low memory - only use bigger
1482 * page orders on machines with more than 32MB of memory if
1483 * not overridden on the command line.
1485 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1486 slab_max_order
= SLAB_MAX_ORDER_HI
;
1488 /* Bootstrap is tricky, because several objects are allocated
1489 * from caches that do not exist yet:
1490 * 1) initialize the kmem_cache cache: it contains the struct
1491 * kmem_cache structures of all caches, except kmem_cache itself:
1492 * kmem_cache is statically allocated.
1493 * Initially an __init data area is used for the head array and the
1494 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1495 * array at the end of the bootstrap.
1496 * 2) Create the first kmalloc cache.
1497 * The struct kmem_cache for the new cache is allocated normally.
1498 * An __init data area is used for the head array.
1499 * 3) Create the remaining kmalloc caches, with minimally sized
1501 * 4) Replace the __init data head arrays for kmem_cache and the first
1502 * kmalloc cache with kmalloc allocated arrays.
1503 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1504 * the other cache's with kmalloc allocated memory.
1505 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1508 /* 1) create the kmem_cache */
1511 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1513 create_boot_cache(kmem_cache
, "kmem_cache",
1514 offsetof(struct kmem_cache
, array
[nr_cpu_ids
]) +
1515 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1516 SLAB_HWCACHE_ALIGN
);
1517 list_add(&kmem_cache
->list
, &slab_caches
);
1519 /* 2+3) create the kmalloc caches */
1522 * Initialize the caches that provide memory for the array cache and the
1523 * kmem_cache_node structures first. Without this, further allocations will
1527 kmalloc_caches
[INDEX_AC
] = create_kmalloc_cache("kmalloc-ac",
1528 kmalloc_size(INDEX_AC
), ARCH_KMALLOC_FLAGS
);
1530 if (INDEX_AC
!= INDEX_NODE
)
1531 kmalloc_caches
[INDEX_NODE
] =
1532 create_kmalloc_cache("kmalloc-node",
1533 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1535 slab_early_init
= 0;
1537 /* 4) Replace the bootstrap head arrays */
1539 struct array_cache
*ptr
;
1541 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1543 memcpy(ptr
, cpu_cache_get(kmem_cache
),
1544 sizeof(struct arraycache_init
));
1546 * Do not assume that spinlocks can be initialized via memcpy:
1548 spin_lock_init(&ptr
->lock
);
1550 kmem_cache
->array
[smp_processor_id()] = ptr
;
1552 ptr
= kmalloc(sizeof(struct arraycache_init
), GFP_NOWAIT
);
1554 BUG_ON(cpu_cache_get(kmalloc_caches
[INDEX_AC
])
1555 != &initarray_generic
.cache
);
1556 memcpy(ptr
, cpu_cache_get(kmalloc_caches
[INDEX_AC
]),
1557 sizeof(struct arraycache_init
));
1559 * Do not assume that spinlocks can be initialized via memcpy:
1561 spin_lock_init(&ptr
->lock
);
1563 kmalloc_caches
[INDEX_AC
]->array
[smp_processor_id()] = ptr
;
1565 /* 5) Replace the bootstrap kmem_cache_node */
1569 for_each_online_node(nid
) {
1570 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1572 init_list(kmalloc_caches
[INDEX_AC
],
1573 &init_kmem_cache_node
[SIZE_AC
+ nid
], nid
);
1575 if (INDEX_AC
!= INDEX_NODE
) {
1576 init_list(kmalloc_caches
[INDEX_NODE
],
1577 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1582 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1585 void __init
kmem_cache_init_late(void)
1587 struct kmem_cache
*cachep
;
1591 /* 6) resize the head arrays to their final sizes */
1592 mutex_lock(&slab_mutex
);
1593 list_for_each_entry(cachep
, &slab_caches
, list
)
1594 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1596 mutex_unlock(&slab_mutex
);
1598 /* Annotate slab for lockdep -- annotate the malloc caches */
1605 * Register a cpu startup notifier callback that initializes
1606 * cpu_cache_get for all new cpus
1608 register_cpu_notifier(&cpucache_notifier
);
1612 * Register a memory hotplug callback that initializes and frees
1615 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1619 * The reap timers are started later, with a module init call: That part
1620 * of the kernel is not yet operational.
1624 static int __init
cpucache_init(void)
1629 * Register the timers that return unneeded pages to the page allocator
1631 for_each_online_cpu(cpu
)
1632 start_cpu_timer(cpu
);
1638 __initcall(cpucache_init
);
1640 static noinline
void
1641 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1643 struct kmem_cache_node
*n
;
1645 unsigned long flags
;
1649 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1651 printk(KERN_WARNING
" cache: %s, object size: %d, order: %d\n",
1652 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1654 for_each_online_node(node
) {
1655 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1656 unsigned long active_slabs
= 0, num_slabs
= 0;
1658 n
= cachep
->node
[node
];
1662 spin_lock_irqsave(&n
->list_lock
, flags
);
1663 list_for_each_entry(slabp
, &n
->slabs_full
, list
) {
1664 active_objs
+= cachep
->num
;
1667 list_for_each_entry(slabp
, &n
->slabs_partial
, list
) {
1668 active_objs
+= slabp
->inuse
;
1671 list_for_each_entry(slabp
, &n
->slabs_free
, list
)
1674 free_objects
+= n
->free_objects
;
1675 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1677 num_slabs
+= active_slabs
;
1678 num_objs
= num_slabs
* cachep
->num
;
1680 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1681 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1687 * Interface to system's page allocator. No need to hold the cache-lock.
1689 * If we requested dmaable memory, we will get it. Even if we
1690 * did not request dmaable memory, we might get it, but that
1691 * would be relatively rare and ignorable.
1693 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1699 flags
|= cachep
->allocflags
;
1700 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1701 flags
|= __GFP_RECLAIMABLE
;
1703 page
= alloc_pages_exact_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1705 if (!(flags
& __GFP_NOWARN
) && printk_ratelimit())
1706 slab_out_of_memory(cachep
, flags
, nodeid
);
1710 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1711 if (unlikely(page
->pfmemalloc
))
1712 pfmemalloc_active
= true;
1714 nr_pages
= (1 << cachep
->gfporder
);
1715 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1716 add_zone_page_state(page_zone(page
),
1717 NR_SLAB_RECLAIMABLE
, nr_pages
);
1719 add_zone_page_state(page_zone(page
),
1720 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1721 __SetPageSlab(page
);
1722 if (page
->pfmemalloc
)
1723 SetPageSlabPfmemalloc(page
);
1724 memcg_bind_pages(cachep
, cachep
->gfporder
);
1726 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1727 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1730 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1732 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1739 * Interface to system's page release.
1741 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1743 const unsigned long nr_freed
= (1 << cachep
->gfporder
);
1745 kmemcheck_free_shadow(page
, cachep
->gfporder
);
1747 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1748 sub_zone_page_state(page_zone(page
),
1749 NR_SLAB_RECLAIMABLE
, nr_freed
);
1751 sub_zone_page_state(page_zone(page
),
1752 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1754 BUG_ON(!PageSlab(page
));
1755 __ClearPageSlabPfmemalloc(page
);
1756 __ClearPageSlab(page
);
1758 memcg_release_pages(cachep
, cachep
->gfporder
);
1759 if (current
->reclaim_state
)
1760 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1761 __free_memcg_kmem_pages(page
, cachep
->gfporder
);
1764 static void kmem_rcu_free(struct rcu_head
*head
)
1766 struct kmem_cache
*cachep
;
1769 page
= container_of(head
, struct page
, rcu_head
);
1770 cachep
= page
->slab_cache
;
1772 kmem_freepages(cachep
, page
);
1777 #ifdef CONFIG_DEBUG_PAGEALLOC
1778 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1779 unsigned long caller
)
1781 int size
= cachep
->object_size
;
1783 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1785 if (size
< 5 * sizeof(unsigned long))
1788 *addr
++ = 0x12345678;
1790 *addr
++ = smp_processor_id();
1791 size
-= 3 * sizeof(unsigned long);
1793 unsigned long *sptr
= &caller
;
1794 unsigned long svalue
;
1796 while (!kstack_end(sptr
)) {
1798 if (kernel_text_address(svalue
)) {
1800 size
-= sizeof(unsigned long);
1801 if (size
<= sizeof(unsigned long))
1807 *addr
++ = 0x87654321;
1811 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1813 int size
= cachep
->object_size
;
1814 addr
= &((char *)addr
)[obj_offset(cachep
)];
1816 memset(addr
, val
, size
);
1817 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1820 static void dump_line(char *data
, int offset
, int limit
)
1823 unsigned char error
= 0;
1826 printk(KERN_ERR
"%03x: ", offset
);
1827 for (i
= 0; i
< limit
; i
++) {
1828 if (data
[offset
+ i
] != POISON_FREE
) {
1829 error
= data
[offset
+ i
];
1833 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1834 &data
[offset
], limit
, 1);
1836 if (bad_count
== 1) {
1837 error
^= POISON_FREE
;
1838 if (!(error
& (error
- 1))) {
1839 printk(KERN_ERR
"Single bit error detected. Probably "
1842 printk(KERN_ERR
"Run memtest86+ or a similar memory "
1845 printk(KERN_ERR
"Run a memory test tool.\n");
1854 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1859 if (cachep
->flags
& SLAB_RED_ZONE
) {
1860 printk(KERN_ERR
"Redzone: 0x%llx/0x%llx.\n",
1861 *dbg_redzone1(cachep
, objp
),
1862 *dbg_redzone2(cachep
, objp
));
1865 if (cachep
->flags
& SLAB_STORE_USER
) {
1866 printk(KERN_ERR
"Last user: [<%p>](%pSR)\n",
1867 *dbg_userword(cachep
, objp
),
1868 *dbg_userword(cachep
, objp
));
1870 realobj
= (char *)objp
+ obj_offset(cachep
);
1871 size
= cachep
->object_size
;
1872 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1875 if (i
+ limit
> size
)
1877 dump_line(realobj
, i
, limit
);
1881 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1887 realobj
= (char *)objp
+ obj_offset(cachep
);
1888 size
= cachep
->object_size
;
1890 for (i
= 0; i
< size
; i
++) {
1891 char exp
= POISON_FREE
;
1894 if (realobj
[i
] != exp
) {
1900 "Slab corruption (%s): %s start=%p, len=%d\n",
1901 print_tainted(), cachep
->name
, realobj
, size
);
1902 print_objinfo(cachep
, objp
, 0);
1904 /* Hexdump the affected line */
1907 if (i
+ limit
> size
)
1909 dump_line(realobj
, i
, limit
);
1912 /* Limit to 5 lines */
1918 /* Print some data about the neighboring objects, if they
1921 struct slab
*slabp
= virt_to_slab(objp
);
1924 objnr
= obj_to_index(cachep
, slabp
, objp
);
1926 objp
= index_to_obj(cachep
, slabp
, objnr
- 1);
1927 realobj
= (char *)objp
+ obj_offset(cachep
);
1928 printk(KERN_ERR
"Prev obj: start=%p, len=%d\n",
1930 print_objinfo(cachep
, objp
, 2);
1932 if (objnr
+ 1 < cachep
->num
) {
1933 objp
= index_to_obj(cachep
, slabp
, objnr
+ 1);
1934 realobj
= (char *)objp
+ obj_offset(cachep
);
1935 printk(KERN_ERR
"Next obj: start=%p, len=%d\n",
1937 print_objinfo(cachep
, objp
, 2);
1944 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1947 for (i
= 0; i
< cachep
->num
; i
++) {
1948 void *objp
= index_to_obj(cachep
, slabp
, i
);
1950 if (cachep
->flags
& SLAB_POISON
) {
1951 #ifdef CONFIG_DEBUG_PAGEALLOC
1952 if (cachep
->size
% PAGE_SIZE
== 0 &&
1954 kernel_map_pages(virt_to_page(objp
),
1955 cachep
->size
/ PAGE_SIZE
, 1);
1957 check_poison_obj(cachep
, objp
);
1959 check_poison_obj(cachep
, objp
);
1962 if (cachep
->flags
& SLAB_RED_ZONE
) {
1963 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1964 slab_error(cachep
, "start of a freed object "
1966 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1967 slab_error(cachep
, "end of a freed object "
1973 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
, struct slab
*slabp
)
1979 * slab_destroy - destroy and release all objects in a slab
1980 * @cachep: cache pointer being destroyed
1981 * @slabp: slab pointer being destroyed
1983 * Destroy all the objs in a slab, and release the mem back to the system.
1984 * Before calling the slab must have been unlinked from the cache. The
1985 * cache-lock is not held/needed.
1987 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slabp
)
1989 struct page
*page
= virt_to_head_page(slabp
->s_mem
);
1991 slab_destroy_debugcheck(cachep
, slabp
);
1992 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
)) {
1993 struct rcu_head
*head
;
1996 * RCU free overloads the RCU head over the LRU.
1997 * slab_page has been overloeaded over the LRU,
1998 * however it is not used from now on so that
1999 * we can use it safely.
2001 head
= (void *)&page
->rcu_head
;
2002 call_rcu(head
, kmem_rcu_free
);
2005 kmem_freepages(cachep
, page
);
2009 * From now on, we don't use slab management
2010 * although actual page can be freed in rcu context
2012 if (OFF_SLAB(cachep
))
2013 kmem_cache_free(cachep
->slabp_cache
, slabp
);
2017 * calculate_slab_order - calculate size (page order) of slabs
2018 * @cachep: pointer to the cache that is being created
2019 * @size: size of objects to be created in this cache.
2020 * @align: required alignment for the objects.
2021 * @flags: slab allocation flags
2023 * Also calculates the number of objects per slab.
2025 * This could be made much more intelligent. For now, try to avoid using
2026 * high order pages for slabs. When the gfp() functions are more friendly
2027 * towards high-order requests, this should be changed.
2029 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
2030 size_t size
, size_t align
, unsigned long flags
)
2032 unsigned long offslab_limit
;
2033 size_t left_over
= 0;
2036 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
2040 cache_estimate(gfporder
, size
, align
, flags
, &remainder
, &num
);
2044 if (flags
& CFLGS_OFF_SLAB
) {
2046 * Max number of objs-per-slab for caches which
2047 * use off-slab slabs. Needed to avoid a possible
2048 * looping condition in cache_grow().
2050 offslab_limit
= size
- sizeof(struct slab
);
2051 offslab_limit
/= sizeof(unsigned int);
2053 if (num
> offslab_limit
)
2057 /* Found something acceptable - save it away */
2059 cachep
->gfporder
= gfporder
;
2060 left_over
= remainder
;
2063 * A VFS-reclaimable slab tends to have most allocations
2064 * as GFP_NOFS and we really don't want to have to be allocating
2065 * higher-order pages when we are unable to shrink dcache.
2067 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2071 * Large number of objects is good, but very large slabs are
2072 * currently bad for the gfp()s.
2074 if (gfporder
>= slab_max_order
)
2078 * Acceptable internal fragmentation?
2080 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
2086 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2088 if (slab_state
>= FULL
)
2089 return enable_cpucache(cachep
, gfp
);
2091 if (slab_state
== DOWN
) {
2093 * Note: Creation of first cache (kmem_cache).
2094 * The setup_node is taken care
2095 * of by the caller of __kmem_cache_create
2097 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2098 slab_state
= PARTIAL
;
2099 } else if (slab_state
== PARTIAL
) {
2101 * Note: the second kmem_cache_create must create the cache
2102 * that's used by kmalloc(24), otherwise the creation of
2103 * further caches will BUG().
2105 cachep
->array
[smp_processor_id()] = &initarray_generic
.cache
;
2108 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2109 * the second cache, then we need to set up all its node/,
2110 * otherwise the creation of further caches will BUG().
2112 set_up_node(cachep
, SIZE_AC
);
2113 if (INDEX_AC
== INDEX_NODE
)
2114 slab_state
= PARTIAL_NODE
;
2116 slab_state
= PARTIAL_ARRAYCACHE
;
2118 /* Remaining boot caches */
2119 cachep
->array
[smp_processor_id()] =
2120 kmalloc(sizeof(struct arraycache_init
), gfp
);
2122 if (slab_state
== PARTIAL_ARRAYCACHE
) {
2123 set_up_node(cachep
, SIZE_NODE
);
2124 slab_state
= PARTIAL_NODE
;
2127 for_each_online_node(node
) {
2128 cachep
->node
[node
] =
2129 kmalloc_node(sizeof(struct kmem_cache_node
),
2131 BUG_ON(!cachep
->node
[node
]);
2132 kmem_cache_node_init(cachep
->node
[node
]);
2136 cachep
->node
[numa_mem_id()]->next_reap
=
2137 jiffies
+ REAPTIMEOUT_LIST3
+
2138 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
2140 cpu_cache_get(cachep
)->avail
= 0;
2141 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
2142 cpu_cache_get(cachep
)->batchcount
= 1;
2143 cpu_cache_get(cachep
)->touched
= 0;
2144 cachep
->batchcount
= 1;
2145 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
2150 * __kmem_cache_create - Create a cache.
2151 * @cachep: cache management descriptor
2152 * @flags: SLAB flags
2154 * Returns a ptr to the cache on success, NULL on failure.
2155 * Cannot be called within a int, but can be interrupted.
2156 * The @ctor is run when new pages are allocated by the cache.
2160 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2161 * to catch references to uninitialised memory.
2163 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2164 * for buffer overruns.
2166 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2167 * cacheline. This can be beneficial if you're counting cycles as closely
2171 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2173 size_t left_over
, slab_size
, ralign
;
2176 size_t size
= cachep
->size
;
2181 * Enable redzoning and last user accounting, except for caches with
2182 * large objects, if the increased size would increase the object size
2183 * above the next power of two: caches with object sizes just above a
2184 * power of two have a significant amount of internal fragmentation.
2186 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2187 2 * sizeof(unsigned long long)))
2188 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2189 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2190 flags
|= SLAB_POISON
;
2192 if (flags
& SLAB_DESTROY_BY_RCU
)
2193 BUG_ON(flags
& SLAB_POISON
);
2197 * Check that size is in terms of words. This is needed to avoid
2198 * unaligned accesses for some archs when redzoning is used, and makes
2199 * sure any on-slab bufctl's are also correctly aligned.
2201 if (size
& (BYTES_PER_WORD
- 1)) {
2202 size
+= (BYTES_PER_WORD
- 1);
2203 size
&= ~(BYTES_PER_WORD
- 1);
2207 * Redzoning and user store require word alignment or possibly larger.
2208 * Note this will be overridden by architecture or caller mandated
2209 * alignment if either is greater than BYTES_PER_WORD.
2211 if (flags
& SLAB_STORE_USER
)
2212 ralign
= BYTES_PER_WORD
;
2214 if (flags
& SLAB_RED_ZONE
) {
2215 ralign
= REDZONE_ALIGN
;
2216 /* If redzoning, ensure that the second redzone is suitably
2217 * aligned, by adjusting the object size accordingly. */
2218 size
+= REDZONE_ALIGN
- 1;
2219 size
&= ~(REDZONE_ALIGN
- 1);
2222 /* 3) caller mandated alignment */
2223 if (ralign
< cachep
->align
) {
2224 ralign
= cachep
->align
;
2226 /* disable debug if necessary */
2227 if (ralign
> __alignof__(unsigned long long))
2228 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2232 cachep
->align
= ralign
;
2234 if (slab_is_available())
2239 setup_node_pointer(cachep
);
2243 * Both debugging options require word-alignment which is calculated
2246 if (flags
& SLAB_RED_ZONE
) {
2247 /* add space for red zone words */
2248 cachep
->obj_offset
+= sizeof(unsigned long long);
2249 size
+= 2 * sizeof(unsigned long long);
2251 if (flags
& SLAB_STORE_USER
) {
2252 /* user store requires one word storage behind the end of
2253 * the real object. But if the second red zone needs to be
2254 * aligned to 64 bits, we must allow that much space.
2256 if (flags
& SLAB_RED_ZONE
)
2257 size
+= REDZONE_ALIGN
;
2259 size
+= BYTES_PER_WORD
;
2261 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2262 if (size
>= kmalloc_size(INDEX_NODE
+ 1)
2263 && cachep
->object_size
> cache_line_size()
2264 && ALIGN(size
, cachep
->align
) < PAGE_SIZE
) {
2265 cachep
->obj_offset
+= PAGE_SIZE
- ALIGN(size
, cachep
->align
);
2272 * Determine if the slab management is 'on' or 'off' slab.
2273 * (bootstrapping cannot cope with offslab caches so don't do
2274 * it too early on. Always use on-slab management when
2275 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2277 if ((size
>= (PAGE_SIZE
>> 3)) && !slab_early_init
&&
2278 !(flags
& SLAB_NOLEAKTRACE
))
2280 * Size is large, assume best to place the slab management obj
2281 * off-slab (should allow better packing of objs).
2283 flags
|= CFLGS_OFF_SLAB
;
2285 size
= ALIGN(size
, cachep
->align
);
2287 left_over
= calculate_slab_order(cachep
, size
, cachep
->align
, flags
);
2292 slab_size
= ALIGN(cachep
->num
* sizeof(unsigned int)
2293 + sizeof(struct slab
), cachep
->align
);
2296 * If the slab has been placed off-slab, and we have enough space then
2297 * move it on-slab. This is at the expense of any extra colouring.
2299 if (flags
& CFLGS_OFF_SLAB
&& left_over
>= slab_size
) {
2300 flags
&= ~CFLGS_OFF_SLAB
;
2301 left_over
-= slab_size
;
2304 if (flags
& CFLGS_OFF_SLAB
) {
2305 /* really off slab. No need for manual alignment */
2307 cachep
->num
* sizeof(unsigned int) + sizeof(struct slab
);
2309 #ifdef CONFIG_PAGE_POISONING
2310 /* If we're going to use the generic kernel_map_pages()
2311 * poisoning, then it's going to smash the contents of
2312 * the redzone and userword anyhow, so switch them off.
2314 if (size
% PAGE_SIZE
== 0 && flags
& SLAB_POISON
)
2315 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2319 cachep
->colour_off
= cache_line_size();
2320 /* Offset must be a multiple of the alignment. */
2321 if (cachep
->colour_off
< cachep
->align
)
2322 cachep
->colour_off
= cachep
->align
;
2323 cachep
->colour
= left_over
/ cachep
->colour_off
;
2324 cachep
->slab_size
= slab_size
;
2325 cachep
->flags
= flags
;
2326 cachep
->allocflags
= __GFP_COMP
;
2327 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2328 cachep
->allocflags
|= GFP_DMA
;
2329 cachep
->size
= size
;
2330 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2332 if (flags
& CFLGS_OFF_SLAB
) {
2333 cachep
->slabp_cache
= kmalloc_slab(slab_size
, 0u);
2335 * This is a possibility for one of the malloc_sizes caches.
2336 * But since we go off slab only for object size greater than
2337 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2338 * this should not happen at all.
2339 * But leave a BUG_ON for some lucky dude.
2341 BUG_ON(ZERO_OR_NULL_PTR(cachep
->slabp_cache
));
2344 err
= setup_cpu_cache(cachep
, gfp
);
2346 __kmem_cache_shutdown(cachep
);
2350 if (flags
& SLAB_DEBUG_OBJECTS
) {
2352 * Would deadlock through slab_destroy()->call_rcu()->
2353 * debug_object_activate()->kmem_cache_alloc().
2355 WARN_ON_ONCE(flags
& SLAB_DESTROY_BY_RCU
);
2357 slab_set_debugobj_lock_classes(cachep
);
2358 } else if (!OFF_SLAB(cachep
) && !(flags
& SLAB_DESTROY_BY_RCU
))
2359 on_slab_lock_classes(cachep
);
2365 static void check_irq_off(void)
2367 BUG_ON(!irqs_disabled());
2370 static void check_irq_on(void)
2372 BUG_ON(irqs_disabled());
2375 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2379 assert_spin_locked(&cachep
->node
[numa_mem_id()]->list_lock
);
2383 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2387 assert_spin_locked(&cachep
->node
[node
]->list_lock
);
2392 #define check_irq_off() do { } while(0)
2393 #define check_irq_on() do { } while(0)
2394 #define check_spinlock_acquired(x) do { } while(0)
2395 #define check_spinlock_acquired_node(x, y) do { } while(0)
2398 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
2399 struct array_cache
*ac
,
2400 int force
, int node
);
2402 static void do_drain(void *arg
)
2404 struct kmem_cache
*cachep
= arg
;
2405 struct array_cache
*ac
;
2406 int node
= numa_mem_id();
2409 ac
= cpu_cache_get(cachep
);
2410 spin_lock(&cachep
->node
[node
]->list_lock
);
2411 free_block(cachep
, ac
->entry
, ac
->avail
, node
);
2412 spin_unlock(&cachep
->node
[node
]->list_lock
);
2416 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2418 struct kmem_cache_node
*n
;
2421 on_each_cpu(do_drain
, cachep
, 1);
2423 for_each_online_node(node
) {
2424 n
= cachep
->node
[node
];
2426 drain_alien_cache(cachep
, n
->alien
);
2429 for_each_online_node(node
) {
2430 n
= cachep
->node
[node
];
2432 drain_array(cachep
, n
, n
->shared
, 1, node
);
2437 * Remove slabs from the list of free slabs.
2438 * Specify the number of slabs to drain in tofree.
2440 * Returns the actual number of slabs released.
2442 static int drain_freelist(struct kmem_cache
*cache
,
2443 struct kmem_cache_node
*n
, int tofree
)
2445 struct list_head
*p
;
2450 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2452 spin_lock_irq(&n
->list_lock
);
2453 p
= n
->slabs_free
.prev
;
2454 if (p
== &n
->slabs_free
) {
2455 spin_unlock_irq(&n
->list_lock
);
2459 slabp
= list_entry(p
, struct slab
, list
);
2461 BUG_ON(slabp
->inuse
);
2463 list_del(&slabp
->list
);
2465 * Safe to drop the lock. The slab is no longer linked
2468 n
->free_objects
-= cache
->num
;
2469 spin_unlock_irq(&n
->list_lock
);
2470 slab_destroy(cache
, slabp
);
2477 /* Called with slab_mutex held to protect against cpu hotplug */
2478 static int __cache_shrink(struct kmem_cache
*cachep
)
2481 struct kmem_cache_node
*n
;
2483 drain_cpu_caches(cachep
);
2486 for_each_online_node(i
) {
2487 n
= cachep
->node
[i
];
2491 drain_freelist(cachep
, n
, slabs_tofree(cachep
, n
));
2493 ret
+= !list_empty(&n
->slabs_full
) ||
2494 !list_empty(&n
->slabs_partial
);
2496 return (ret
? 1 : 0);
2500 * kmem_cache_shrink - Shrink a cache.
2501 * @cachep: The cache to shrink.
2503 * Releases as many slabs as possible for a cache.
2504 * To help debugging, a zero exit status indicates all slabs were released.
2506 int kmem_cache_shrink(struct kmem_cache
*cachep
)
2509 BUG_ON(!cachep
|| in_interrupt());
2512 mutex_lock(&slab_mutex
);
2513 ret
= __cache_shrink(cachep
);
2514 mutex_unlock(&slab_mutex
);
2518 EXPORT_SYMBOL(kmem_cache_shrink
);
2520 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2523 struct kmem_cache_node
*n
;
2524 int rc
= __cache_shrink(cachep
);
2529 for_each_online_cpu(i
)
2530 kfree(cachep
->array
[i
]);
2532 /* NUMA: free the node structures */
2533 for_each_online_node(i
) {
2534 n
= cachep
->node
[i
];
2537 free_alien_cache(n
->alien
);
2545 * Get the memory for a slab management obj.
2546 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2547 * always come from malloc_sizes caches. The slab descriptor cannot
2548 * come from the same cache which is getting created because,
2549 * when we are searching for an appropriate cache for these
2550 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2551 * If we are creating a malloc_sizes cache here it would not be visible to
2552 * kmem_find_general_cachep till the initialization is complete.
2553 * Hence we cannot have slabp_cache same as the original cache.
2555 static struct slab
*alloc_slabmgmt(struct kmem_cache
*cachep
,
2556 struct page
*page
, int colour_off
,
2557 gfp_t local_flags
, int nodeid
)
2560 void *addr
= page_address(page
);
2562 if (OFF_SLAB(cachep
)) {
2563 /* Slab management obj is off-slab. */
2564 slabp
= kmem_cache_alloc_node(cachep
->slabp_cache
,
2565 local_flags
, nodeid
);
2567 * If the first object in the slab is leaked (it's allocated
2568 * but no one has a reference to it), we want to make sure
2569 * kmemleak does not treat the ->s_mem pointer as a reference
2570 * to the object. Otherwise we will not report the leak.
2572 kmemleak_scan_area(&slabp
->list
, sizeof(struct list_head
),
2577 slabp
= addr
+ colour_off
;
2578 colour_off
+= cachep
->slab_size
;
2581 slabp
->s_mem
= addr
+ colour_off
;
2586 static inline unsigned int *slab_bufctl(struct slab
*slabp
)
2588 return (unsigned int *) (slabp
+ 1);
2591 static void cache_init_objs(struct kmem_cache
*cachep
,
2596 for (i
= 0; i
< cachep
->num
; i
++) {
2597 void *objp
= index_to_obj(cachep
, slabp
, i
);
2599 /* need to poison the objs? */
2600 if (cachep
->flags
& SLAB_POISON
)
2601 poison_obj(cachep
, objp
, POISON_FREE
);
2602 if (cachep
->flags
& SLAB_STORE_USER
)
2603 *dbg_userword(cachep
, objp
) = NULL
;
2605 if (cachep
->flags
& SLAB_RED_ZONE
) {
2606 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2607 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2610 * Constructors are not allowed to allocate memory from the same
2611 * cache which they are a constructor for. Otherwise, deadlock.
2612 * They must also be threaded.
2614 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
))
2615 cachep
->ctor(objp
+ obj_offset(cachep
));
2617 if (cachep
->flags
& SLAB_RED_ZONE
) {
2618 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2619 slab_error(cachep
, "constructor overwrote the"
2620 " end of an object");
2621 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2622 slab_error(cachep
, "constructor overwrote the"
2623 " start of an object");
2625 if ((cachep
->size
% PAGE_SIZE
) == 0 &&
2626 OFF_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
)
2627 kernel_map_pages(virt_to_page(objp
),
2628 cachep
->size
/ PAGE_SIZE
, 0);
2633 slab_bufctl(slabp
)[i
] = i
;
2637 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2639 if (CONFIG_ZONE_DMA_FLAG
) {
2640 if (flags
& GFP_DMA
)
2641 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2643 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2647 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2653 objp
= index_to_obj(cachep
, slabp
, slab_bufctl(slabp
)[slabp
->free
]);
2655 WARN_ON(page_to_nid(virt_to_page(objp
)) != nodeid
);
2662 static void slab_put_obj(struct kmem_cache
*cachep
, struct slab
*slabp
,
2663 void *objp
, int nodeid
)
2665 unsigned int objnr
= obj_to_index(cachep
, slabp
, 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
= slabp
->free
; i
< cachep
->num
; i
++) {
2674 if (slab_bufctl(slabp
)[i
] == objnr
) {
2675 printk(KERN_ERR
"slab: double free detected in cache "
2676 "'%s', objp %p\n", cachep
->name
, objp
);
2682 slab_bufctl(slabp
)[slabp
->free
] = objnr
;
2687 * Map pages beginning at addr to the given cache and slab. This is required
2688 * for the slab allocator to be able to lookup the cache and slab of a
2689 * virtual address for kfree, ksize, and slab debugging.
2691 static void slab_map_pages(struct kmem_cache
*cache
, struct slab
*slab
,
2694 page
->slab_cache
= cache
;
2695 page
->slab_page
= slab
;
2699 * Grow (by 1) the number of slabs within a cache. This is called by
2700 * kmem_cache_alloc() when there are no active objs left in a cache.
2702 static int cache_grow(struct kmem_cache
*cachep
,
2703 gfp_t flags
, int nodeid
, struct page
*page
)
2708 struct kmem_cache_node
*n
;
2711 * Be lazy and only check for valid flags here, keeping it out of the
2712 * critical path in kmem_cache_alloc().
2714 BUG_ON(flags
& GFP_SLAB_BUG_MASK
);
2715 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2717 /* Take the node list lock to change the colour_next on this node */
2719 n
= cachep
->node
[nodeid
];
2720 spin_lock(&n
->list_lock
);
2722 /* Get colour for the slab, and cal the next value. */
2723 offset
= n
->colour_next
;
2725 if (n
->colour_next
>= cachep
->colour
)
2727 spin_unlock(&n
->list_lock
);
2729 offset
*= cachep
->colour_off
;
2731 if (local_flags
& __GFP_WAIT
)
2735 * The test for missing atomic flag is performed here, rather than
2736 * the more obvious place, simply to reduce the critical path length
2737 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2738 * will eventually be caught here (where it matters).
2740 kmem_flagcheck(cachep
, flags
);
2743 * Get mem for the objs. Attempt to allocate a physical page from
2747 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2751 /* Get slab management. */
2752 slabp
= alloc_slabmgmt(cachep
, page
, offset
,
2753 local_flags
& ~GFP_CONSTRAINT_MASK
, nodeid
);
2757 slab_map_pages(cachep
, slabp
, page
);
2759 cache_init_objs(cachep
, slabp
);
2761 if (local_flags
& __GFP_WAIT
)
2762 local_irq_disable();
2764 spin_lock(&n
->list_lock
);
2766 /* Make slab active. */
2767 list_add_tail(&slabp
->list
, &(n
->slabs_free
));
2768 STATS_INC_GROWN(cachep
);
2769 n
->free_objects
+= cachep
->num
;
2770 spin_unlock(&n
->list_lock
);
2773 kmem_freepages(cachep
, page
);
2775 if (local_flags
& __GFP_WAIT
)
2776 local_irq_disable();
2783 * Perform extra freeing checks:
2784 * - detect bad pointers.
2785 * - POISON/RED_ZONE checking
2787 static void kfree_debugcheck(const void *objp
)
2789 if (!virt_addr_valid(objp
)) {
2790 printk(KERN_ERR
"kfree_debugcheck: out of range ptr %lxh.\n",
2791 (unsigned long)objp
);
2796 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2798 unsigned long long redzone1
, redzone2
;
2800 redzone1
= *dbg_redzone1(cache
, obj
);
2801 redzone2
= *dbg_redzone2(cache
, obj
);
2806 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2809 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2810 slab_error(cache
, "double free detected");
2812 slab_error(cache
, "memory outside object was overwritten");
2814 printk(KERN_ERR
"%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2815 obj
, redzone1
, redzone2
);
2818 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2819 unsigned long caller
)
2824 BUG_ON(virt_to_cache(objp
) != cachep
);
2826 objp
-= obj_offset(cachep
);
2827 kfree_debugcheck(objp
);
2828 slabp
= virt_to_slab(objp
);
2830 if (cachep
->flags
& SLAB_RED_ZONE
) {
2831 verify_redzone_free(cachep
, objp
);
2832 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2833 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2835 if (cachep
->flags
& SLAB_STORE_USER
)
2836 *dbg_userword(cachep
, objp
) = (void *)caller
;
2838 objnr
= obj_to_index(cachep
, slabp
, objp
);
2840 BUG_ON(objnr
>= cachep
->num
);
2841 BUG_ON(objp
!= index_to_obj(cachep
, slabp
, objnr
));
2843 if (cachep
->flags
& SLAB_POISON
) {
2844 #ifdef CONFIG_DEBUG_PAGEALLOC
2845 if ((cachep
->size
% PAGE_SIZE
)==0 && OFF_SLAB(cachep
)) {
2846 store_stackinfo(cachep
, objp
, caller
);
2847 kernel_map_pages(virt_to_page(objp
),
2848 cachep
->size
/ PAGE_SIZE
, 0);
2850 poison_obj(cachep
, objp
, POISON_FREE
);
2853 poison_obj(cachep
, objp
, POISON_FREE
);
2860 #define kfree_debugcheck(x) do { } while(0)
2861 #define cache_free_debugcheck(x,objp,z) (objp)
2864 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
,
2868 struct kmem_cache_node
*n
;
2869 struct array_cache
*ac
;
2873 node
= numa_mem_id();
2874 if (unlikely(force_refill
))
2877 ac
= cpu_cache_get(cachep
);
2878 batchcount
= ac
->batchcount
;
2879 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2881 * If there was little recent activity on this cache, then
2882 * perform only a partial refill. Otherwise we could generate
2885 batchcount
= BATCHREFILL_LIMIT
;
2887 n
= cachep
->node
[node
];
2889 BUG_ON(ac
->avail
> 0 || !n
);
2890 spin_lock(&n
->list_lock
);
2892 /* See if we can refill from the shared array */
2893 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2894 n
->shared
->touched
= 1;
2898 while (batchcount
> 0) {
2899 struct list_head
*entry
;
2901 /* Get slab alloc is to come from. */
2902 entry
= n
->slabs_partial
.next
;
2903 if (entry
== &n
->slabs_partial
) {
2904 n
->free_touched
= 1;
2905 entry
= n
->slabs_free
.next
;
2906 if (entry
== &n
->slabs_free
)
2910 slabp
= list_entry(entry
, struct slab
, list
);
2911 check_spinlock_acquired(cachep
);
2914 * The slab was either on partial or free list so
2915 * there must be at least one object available for
2918 BUG_ON(slabp
->inuse
>= cachep
->num
);
2920 while (slabp
->inuse
< cachep
->num
&& batchcount
--) {
2921 STATS_INC_ALLOCED(cachep
);
2922 STATS_INC_ACTIVE(cachep
);
2923 STATS_SET_HIGH(cachep
);
2925 ac_put_obj(cachep
, ac
, slab_get_obj(cachep
, slabp
,
2929 /* move slabp to correct slabp list: */
2930 list_del(&slabp
->list
);
2931 if (slabp
->free
== cachep
->num
)
2932 list_add(&slabp
->list
, &n
->slabs_full
);
2934 list_add(&slabp
->list
, &n
->slabs_partial
);
2938 n
->free_objects
-= ac
->avail
;
2940 spin_unlock(&n
->list_lock
);
2942 if (unlikely(!ac
->avail
)) {
2945 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, node
, NULL
);
2947 /* cache_grow can reenable interrupts, then ac could change. */
2948 ac
= cpu_cache_get(cachep
);
2949 node
= numa_mem_id();
2951 /* no objects in sight? abort */
2952 if (!x
&& (ac
->avail
== 0 || force_refill
))
2955 if (!ac
->avail
) /* objects refilled by interrupt? */
2960 return ac_get_obj(cachep
, ac
, flags
, force_refill
);
2963 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2966 might_sleep_if(flags
& __GFP_WAIT
);
2968 kmem_flagcheck(cachep
, flags
);
2973 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2974 gfp_t flags
, void *objp
, unsigned long caller
)
2978 if (cachep
->flags
& SLAB_POISON
) {
2979 #ifdef CONFIG_DEBUG_PAGEALLOC
2980 if ((cachep
->size
% PAGE_SIZE
) == 0 && OFF_SLAB(cachep
))
2981 kernel_map_pages(virt_to_page(objp
),
2982 cachep
->size
/ PAGE_SIZE
, 1);
2984 check_poison_obj(cachep
, objp
);
2986 check_poison_obj(cachep
, objp
);
2988 poison_obj(cachep
, objp
, POISON_INUSE
);
2990 if (cachep
->flags
& SLAB_STORE_USER
)
2991 *dbg_userword(cachep
, objp
) = (void *)caller
;
2993 if (cachep
->flags
& SLAB_RED_ZONE
) {
2994 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2995 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2996 slab_error(cachep
, "double free, or memory outside"
2997 " object was overwritten");
2999 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3000 objp
, *dbg_redzone1(cachep
, objp
),
3001 *dbg_redzone2(cachep
, objp
));
3003 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3004 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3006 objp
+= obj_offset(cachep
);
3007 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3009 if (ARCH_SLAB_MINALIGN
&&
3010 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3011 printk(KERN_ERR
"0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3012 objp
, (int)ARCH_SLAB_MINALIGN
);
3017 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3020 static bool slab_should_failslab(struct kmem_cache
*cachep
, gfp_t flags
)
3022 if (cachep
== kmem_cache
)
3025 return should_failslab(cachep
->object_size
, flags
, cachep
->flags
);
3028 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3031 struct array_cache
*ac
;
3032 bool force_refill
= false;
3036 ac
= cpu_cache_get(cachep
);
3037 if (likely(ac
->avail
)) {
3039 objp
= ac_get_obj(cachep
, ac
, flags
, false);
3042 * Allow for the possibility all avail objects are not allowed
3043 * by the current flags
3046 STATS_INC_ALLOCHIT(cachep
);
3049 force_refill
= true;
3052 STATS_INC_ALLOCMISS(cachep
);
3053 objp
= cache_alloc_refill(cachep
, flags
, force_refill
);
3055 * the 'ac' may be updated by cache_alloc_refill(),
3056 * and kmemleak_erase() requires its correct value.
3058 ac
= cpu_cache_get(cachep
);
3062 * To avoid a false negative, if an object that is in one of the
3063 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3064 * treat the array pointers as a reference to the object.
3067 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3073 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3075 * If we are in_interrupt, then process context, including cpusets and
3076 * mempolicy, may not apply and should not be used for allocation policy.
3078 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3080 int nid_alloc
, nid_here
;
3082 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3084 nid_alloc
= nid_here
= numa_mem_id();
3085 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3086 nid_alloc
= cpuset_slab_spread_node();
3087 else if (current
->mempolicy
)
3088 nid_alloc
= slab_node();
3089 if (nid_alloc
!= nid_here
)
3090 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3095 * Fallback function if there was no memory available and no objects on a
3096 * certain node and fall back is permitted. First we scan all the
3097 * available node for available objects. If that fails then we
3098 * perform an allocation without specifying a node. This allows the page
3099 * allocator to do its reclaim / fallback magic. We then insert the
3100 * slab into the proper nodelist and then allocate from it.
3102 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3104 struct zonelist
*zonelist
;
3108 enum zone_type high_zoneidx
= gfp_zone(flags
);
3111 unsigned int cpuset_mems_cookie
;
3113 if (flags
& __GFP_THISNODE
)
3116 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
3119 cpuset_mems_cookie
= get_mems_allowed();
3120 zonelist
= node_zonelist(slab_node(), flags
);
3124 * Look through allowed nodes for objects available
3125 * from existing per node queues.
3127 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3128 nid
= zone_to_nid(zone
);
3130 if (cpuset_zone_allowed_hardwall(zone
, flags
) &&
3132 cache
->node
[nid
]->free_objects
) {
3133 obj
= ____cache_alloc_node(cache
,
3134 flags
| GFP_THISNODE
, nid
);
3142 * This allocation will be performed within the constraints
3143 * of the current cpuset / memory policy requirements.
3144 * We may trigger various forms of reclaim on the allowed
3145 * set and go into memory reserves if necessary.
3149 if (local_flags
& __GFP_WAIT
)
3151 kmem_flagcheck(cache
, flags
);
3152 page
= kmem_getpages(cache
, local_flags
, numa_mem_id());
3153 if (local_flags
& __GFP_WAIT
)
3154 local_irq_disable();
3157 * Insert into the appropriate per node queues
3159 nid
= page_to_nid(page
);
3160 if (cache_grow(cache
, flags
, nid
, page
)) {
3161 obj
= ____cache_alloc_node(cache
,
3162 flags
| GFP_THISNODE
, nid
);
3165 * Another processor may allocate the
3166 * objects in the slab since we are
3167 * not holding any locks.
3171 /* cache_grow already freed obj */
3177 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !obj
))
3183 * A interface to enable slab creation on nodeid
3185 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3188 struct list_head
*entry
;
3190 struct kmem_cache_node
*n
;
3194 VM_BUG_ON(nodeid
> num_online_nodes());
3195 n
= cachep
->node
[nodeid
];
3200 spin_lock(&n
->list_lock
);
3201 entry
= n
->slabs_partial
.next
;
3202 if (entry
== &n
->slabs_partial
) {
3203 n
->free_touched
= 1;
3204 entry
= n
->slabs_free
.next
;
3205 if (entry
== &n
->slabs_free
)
3209 slabp
= list_entry(entry
, struct slab
, list
);
3210 check_spinlock_acquired_node(cachep
, nodeid
);
3212 STATS_INC_NODEALLOCS(cachep
);
3213 STATS_INC_ACTIVE(cachep
);
3214 STATS_SET_HIGH(cachep
);
3216 BUG_ON(slabp
->inuse
== cachep
->num
);
3218 obj
= slab_get_obj(cachep
, slabp
, nodeid
);
3220 /* move slabp to correct slabp list: */
3221 list_del(&slabp
->list
);
3223 if (slabp
->free
== cachep
->num
)
3224 list_add(&slabp
->list
, &n
->slabs_full
);
3226 list_add(&slabp
->list
, &n
->slabs_partial
);
3228 spin_unlock(&n
->list_lock
);
3232 spin_unlock(&n
->list_lock
);
3233 x
= cache_grow(cachep
, flags
| GFP_THISNODE
, nodeid
, NULL
);
3237 return fallback_alloc(cachep
, flags
);
3243 static __always_inline
void *
3244 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3245 unsigned long caller
)
3247 unsigned long save_flags
;
3249 int slab_node
= numa_mem_id();
3251 flags
&= gfp_allowed_mask
;
3253 lockdep_trace_alloc(flags
);
3255 if (slab_should_failslab(cachep
, flags
))
3258 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3260 cache_alloc_debugcheck_before(cachep
, flags
);
3261 local_irq_save(save_flags
);
3263 if (nodeid
== NUMA_NO_NODE
)
3266 if (unlikely(!cachep
->node
[nodeid
])) {
3267 /* Node not bootstrapped yet */
3268 ptr
= fallback_alloc(cachep
, flags
);
3272 if (nodeid
== slab_node
) {
3274 * Use the locally cached objects if possible.
3275 * However ____cache_alloc does not allow fallback
3276 * to other nodes. It may fail while we still have
3277 * objects on other nodes available.
3279 ptr
= ____cache_alloc(cachep
, flags
);
3283 /* ___cache_alloc_node can fall back to other nodes */
3284 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3286 local_irq_restore(save_flags
);
3287 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3288 kmemleak_alloc_recursive(ptr
, cachep
->object_size
, 1, cachep
->flags
,
3292 kmemcheck_slab_alloc(cachep
, flags
, ptr
, cachep
->object_size
);
3294 if (unlikely((flags
& __GFP_ZERO
) && ptr
))
3295 memset(ptr
, 0, cachep
->object_size
);
3300 static __always_inline
void *
3301 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3305 if (unlikely(current
->flags
& (PF_SPREAD_SLAB
| PF_MEMPOLICY
))) {
3306 objp
= alternate_node_alloc(cache
, flags
);
3310 objp
= ____cache_alloc(cache
, flags
);
3313 * We may just have run out of memory on the local node.
3314 * ____cache_alloc_node() knows how to locate memory on other nodes
3317 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3324 static __always_inline
void *
3325 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3327 return ____cache_alloc(cachep
, flags
);
3330 #endif /* CONFIG_NUMA */
3332 static __always_inline
void *
3333 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3335 unsigned long save_flags
;
3338 flags
&= gfp_allowed_mask
;
3340 lockdep_trace_alloc(flags
);
3342 if (slab_should_failslab(cachep
, flags
))
3345 cachep
= memcg_kmem_get_cache(cachep
, flags
);
3347 cache_alloc_debugcheck_before(cachep
, flags
);
3348 local_irq_save(save_flags
);
3349 objp
= __do_cache_alloc(cachep
, flags
);
3350 local_irq_restore(save_flags
);
3351 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3352 kmemleak_alloc_recursive(objp
, cachep
->object_size
, 1, cachep
->flags
,
3357 kmemcheck_slab_alloc(cachep
, flags
, objp
, cachep
->object_size
);
3359 if (unlikely((flags
& __GFP_ZERO
) && objp
))
3360 memset(objp
, 0, cachep
->object_size
);
3366 * Caller needs to acquire correct kmem_list's list_lock
3368 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int nr_objects
,
3372 struct kmem_cache_node
*n
;
3374 for (i
= 0; i
< nr_objects
; i
++) {
3378 clear_obj_pfmemalloc(&objpp
[i
]);
3381 slabp
= virt_to_slab(objp
);
3382 n
= cachep
->node
[node
];
3383 list_del(&slabp
->list
);
3384 check_spinlock_acquired_node(cachep
, node
);
3385 slab_put_obj(cachep
, slabp
, objp
, node
);
3386 STATS_DEC_ACTIVE(cachep
);
3389 /* fixup slab chains */
3390 if (slabp
->inuse
== 0) {
3391 if (n
->free_objects
> n
->free_limit
) {
3392 n
->free_objects
-= cachep
->num
;
3393 /* No need to drop any previously held
3394 * lock here, even if we have a off-slab slab
3395 * descriptor it is guaranteed to come from
3396 * a different cache, refer to comments before
3399 slab_destroy(cachep
, slabp
);
3401 list_add(&slabp
->list
, &n
->slabs_free
);
3404 /* Unconditionally move a slab to the end of the
3405 * partial list on free - maximum time for the
3406 * other objects to be freed, too.
3408 list_add_tail(&slabp
->list
, &n
->slabs_partial
);
3413 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3416 struct kmem_cache_node
*n
;
3417 int node
= numa_mem_id();
3419 batchcount
= ac
->batchcount
;
3421 BUG_ON(!batchcount
|| batchcount
> ac
->avail
);
3424 n
= cachep
->node
[node
];
3425 spin_lock(&n
->list_lock
);
3427 struct array_cache
*shared_array
= n
->shared
;
3428 int max
= shared_array
->limit
- shared_array
->avail
;
3430 if (batchcount
> max
)
3432 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3433 ac
->entry
, sizeof(void *) * batchcount
);
3434 shared_array
->avail
+= batchcount
;
3439 free_block(cachep
, ac
->entry
, batchcount
, node
);
3444 struct list_head
*p
;
3446 p
= n
->slabs_free
.next
;
3447 while (p
!= &(n
->slabs_free
)) {
3450 slabp
= list_entry(p
, struct slab
, list
);
3451 BUG_ON(slabp
->inuse
);
3456 STATS_SET_FREEABLE(cachep
, i
);
3459 spin_unlock(&n
->list_lock
);
3460 ac
->avail
-= batchcount
;
3461 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3465 * Release an obj back to its cache. If the obj has a constructed state, it must
3466 * be in this state _before_ it is released. Called with disabled ints.
3468 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3469 unsigned long caller
)
3471 struct array_cache
*ac
= cpu_cache_get(cachep
);
3474 kmemleak_free_recursive(objp
, cachep
->flags
);
3475 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3477 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3480 * Skip calling cache_free_alien() when the platform is not numa.
3481 * This will avoid cache misses that happen while accessing slabp (which
3482 * is per page memory reference) to get nodeid. Instead use a global
3483 * variable to skip the call, which is mostly likely to be present in
3486 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3489 if (likely(ac
->avail
< ac
->limit
)) {
3490 STATS_INC_FREEHIT(cachep
);
3492 STATS_INC_FREEMISS(cachep
);
3493 cache_flusharray(cachep
, ac
);
3496 ac_put_obj(cachep
, ac
, objp
);
3500 * kmem_cache_alloc - Allocate an object
3501 * @cachep: The cache to allocate from.
3502 * @flags: See kmalloc().
3504 * Allocate an object from this cache. The flags are only relevant
3505 * if the cache has no available objects.
3507 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3509 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3511 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3512 cachep
->object_size
, cachep
->size
, flags
);
3516 EXPORT_SYMBOL(kmem_cache_alloc
);
3518 #ifdef CONFIG_TRACING
3520 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3524 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3526 trace_kmalloc(_RET_IP_
, ret
,
3527 size
, cachep
->size
, flags
);
3530 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3535 * kmem_cache_alloc_node - Allocate an object on the specified node
3536 * @cachep: The cache to allocate from.
3537 * @flags: See kmalloc().
3538 * @nodeid: node number of the target node.
3540 * Identical to kmem_cache_alloc but it will allocate memory on the given
3541 * node, which can improve the performance for cpu bound structures.
3543 * Fallback to other node is possible if __GFP_THISNODE is not set.
3545 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3547 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3549 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3550 cachep
->object_size
, cachep
->size
,
3555 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3557 #ifdef CONFIG_TRACING
3558 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3565 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3567 trace_kmalloc_node(_RET_IP_
, ret
,
3572 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3575 static __always_inline
void *
3576 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3578 struct kmem_cache
*cachep
;
3580 cachep
= kmalloc_slab(size
, flags
);
3581 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3583 return kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3586 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3587 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3589 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3591 EXPORT_SYMBOL(__kmalloc_node
);
3593 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3594 int node
, unsigned long caller
)
3596 return __do_kmalloc_node(size
, flags
, node
, caller
);
3598 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3600 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3602 return __do_kmalloc_node(size
, flags
, node
, 0);
3604 EXPORT_SYMBOL(__kmalloc_node
);
3605 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3606 #endif /* CONFIG_NUMA */
3609 * __do_kmalloc - allocate memory
3610 * @size: how many bytes of memory are required.
3611 * @flags: the type of memory to allocate (see kmalloc).
3612 * @caller: function caller for debug tracking of the caller
3614 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3615 unsigned long caller
)
3617 struct kmem_cache
*cachep
;
3620 /* If you want to save a few bytes .text space: replace
3622 * Then kmalloc uses the uninlined functions instead of the inline
3625 cachep
= kmalloc_slab(size
, flags
);
3626 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3628 ret
= slab_alloc(cachep
, flags
, caller
);
3630 trace_kmalloc(caller
, ret
,
3631 size
, cachep
->size
, flags
);
3637 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3638 void *__kmalloc(size_t size
, gfp_t flags
)
3640 return __do_kmalloc(size
, flags
, _RET_IP_
);
3642 EXPORT_SYMBOL(__kmalloc
);
3644 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3646 return __do_kmalloc(size
, flags
, caller
);
3648 EXPORT_SYMBOL(__kmalloc_track_caller
);
3651 void *__kmalloc(size_t size
, gfp_t flags
)
3653 return __do_kmalloc(size
, flags
, 0);
3655 EXPORT_SYMBOL(__kmalloc
);
3659 * kmem_cache_free - Deallocate an object
3660 * @cachep: The cache the allocation was from.
3661 * @objp: The previously allocated object.
3663 * Free an object which was previously allocated from this
3666 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3668 unsigned long flags
;
3669 cachep
= cache_from_obj(cachep
, objp
);
3673 local_irq_save(flags
);
3674 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3675 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3676 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3677 __cache_free(cachep
, objp
, _RET_IP_
);
3678 local_irq_restore(flags
);
3680 trace_kmem_cache_free(_RET_IP_
, objp
);
3682 EXPORT_SYMBOL(kmem_cache_free
);
3685 * kfree - free previously allocated memory
3686 * @objp: pointer returned by kmalloc.
3688 * If @objp is NULL, no operation is performed.
3690 * Don't free memory not originally allocated by kmalloc()
3691 * or you will run into trouble.
3693 void kfree(const void *objp
)
3695 struct kmem_cache
*c
;
3696 unsigned long flags
;
3698 trace_kfree(_RET_IP_
, objp
);
3700 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3702 local_irq_save(flags
);
3703 kfree_debugcheck(objp
);
3704 c
= virt_to_cache(objp
);
3705 debug_check_no_locks_freed(objp
, c
->object_size
);
3707 debug_check_no_obj_freed(objp
, c
->object_size
);
3708 __cache_free(c
, (void *)objp
, _RET_IP_
);
3709 local_irq_restore(flags
);
3711 EXPORT_SYMBOL(kfree
);
3714 * This initializes kmem_cache_node or resizes various caches for all nodes.
3716 static int alloc_kmemlist(struct kmem_cache
*cachep
, gfp_t gfp
)
3719 struct kmem_cache_node
*n
;
3720 struct array_cache
*new_shared
;
3721 struct array_cache
**new_alien
= NULL
;
3723 for_each_online_node(node
) {
3725 if (use_alien_caches
) {
3726 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
3732 if (cachep
->shared
) {
3733 new_shared
= alloc_arraycache(node
,
3734 cachep
->shared
*cachep
->batchcount
,
3737 free_alien_cache(new_alien
);
3742 n
= cachep
->node
[node
];
3744 struct array_cache
*shared
= n
->shared
;
3746 spin_lock_irq(&n
->list_lock
);
3749 free_block(cachep
, shared
->entry
,
3750 shared
->avail
, node
);
3752 n
->shared
= new_shared
;
3754 n
->alien
= new_alien
;
3757 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3758 cachep
->batchcount
+ cachep
->num
;
3759 spin_unlock_irq(&n
->list_lock
);
3761 free_alien_cache(new_alien
);
3764 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
3766 free_alien_cache(new_alien
);
3771 kmem_cache_node_init(n
);
3772 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
+
3773 ((unsigned long)cachep
) % REAPTIMEOUT_LIST3
;
3774 n
->shared
= new_shared
;
3775 n
->alien
= new_alien
;
3776 n
->free_limit
= (1 + nr_cpus_node(node
)) *
3777 cachep
->batchcount
+ cachep
->num
;
3778 cachep
->node
[node
] = n
;
3783 if (!cachep
->list
.next
) {
3784 /* Cache is not active yet. Roll back what we did */
3787 if (cachep
->node
[node
]) {
3788 n
= cachep
->node
[node
];
3791 free_alien_cache(n
->alien
);
3793 cachep
->node
[node
] = NULL
;
3801 struct ccupdate_struct
{
3802 struct kmem_cache
*cachep
;
3803 struct array_cache
*new[0];
3806 static void do_ccupdate_local(void *info
)
3808 struct ccupdate_struct
*new = info
;
3809 struct array_cache
*old
;
3812 old
= cpu_cache_get(new->cachep
);
3814 new->cachep
->array
[smp_processor_id()] = new->new[smp_processor_id()];
3815 new->new[smp_processor_id()] = old
;
3818 /* Always called with the slab_mutex held */
3819 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3820 int batchcount
, int shared
, gfp_t gfp
)
3822 struct ccupdate_struct
*new;
3825 new = kzalloc(sizeof(*new) + nr_cpu_ids
* sizeof(struct array_cache
*),
3830 for_each_online_cpu(i
) {
3831 new->new[i
] = alloc_arraycache(cpu_to_mem(i
), limit
,
3834 for (i
--; i
>= 0; i
--)
3840 new->cachep
= cachep
;
3842 on_each_cpu(do_ccupdate_local
, (void *)new, 1);
3845 cachep
->batchcount
= batchcount
;
3846 cachep
->limit
= limit
;
3847 cachep
->shared
= shared
;
3849 for_each_online_cpu(i
) {
3850 struct array_cache
*ccold
= new->new[i
];
3853 spin_lock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3854 free_block(cachep
, ccold
->entry
, ccold
->avail
, cpu_to_mem(i
));
3855 spin_unlock_irq(&cachep
->node
[cpu_to_mem(i
)]->list_lock
);
3859 return alloc_kmemlist(cachep
, gfp
);
3862 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3863 int batchcount
, int shared
, gfp_t gfp
)
3866 struct kmem_cache
*c
= NULL
;
3869 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3871 if (slab_state
< FULL
)
3874 if ((ret
< 0) || !is_root_cache(cachep
))
3877 VM_BUG_ON(!mutex_is_locked(&slab_mutex
));
3878 for_each_memcg_cache_index(i
) {
3879 c
= cache_from_memcg(cachep
, i
);
3881 /* return value determined by the parent cache only */
3882 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3888 /* Called with slab_mutex held always */
3889 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3896 if (!is_root_cache(cachep
)) {
3897 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3898 limit
= root
->limit
;
3899 shared
= root
->shared
;
3900 batchcount
= root
->batchcount
;
3903 if (limit
&& shared
&& batchcount
)
3906 * The head array serves three purposes:
3907 * - create a LIFO ordering, i.e. return objects that are cache-warm
3908 * - reduce the number of spinlock operations.
3909 * - reduce the number of linked list operations on the slab and
3910 * bufctl chains: array operations are cheaper.
3911 * The numbers are guessed, we should auto-tune as described by
3914 if (cachep
->size
> 131072)
3916 else if (cachep
->size
> PAGE_SIZE
)
3918 else if (cachep
->size
> 1024)
3920 else if (cachep
->size
> 256)
3926 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3927 * allocation behaviour: Most allocs on one cpu, most free operations
3928 * on another cpu. For these cases, an efficient object passing between
3929 * cpus is necessary. This is provided by a shared array. The array
3930 * replaces Bonwick's magazine layer.
3931 * On uniprocessor, it's functionally equivalent (but less efficient)
3932 * to a larger limit. Thus disabled by default.
3935 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3940 * With debugging enabled, large batchcount lead to excessively long
3941 * periods with disabled local interrupts. Limit the batchcount
3946 batchcount
= (limit
+ 1) / 2;
3948 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3950 printk(KERN_ERR
"enable_cpucache failed for %s, error %d.\n",
3951 cachep
->name
, -err
);
3956 * Drain an array if it contains any elements taking the node lock only if
3957 * necessary. Note that the node listlock also protects the array_cache
3958 * if drain_array() is used on the shared array.
3960 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3961 struct array_cache
*ac
, int force
, int node
)
3965 if (!ac
|| !ac
->avail
)
3967 if (ac
->touched
&& !force
) {
3970 spin_lock_irq(&n
->list_lock
);
3972 tofree
= force
? ac
->avail
: (ac
->limit
+ 4) / 5;
3973 if (tofree
> ac
->avail
)
3974 tofree
= (ac
->avail
+ 1) / 2;
3975 free_block(cachep
, ac
->entry
, tofree
, node
);
3976 ac
->avail
-= tofree
;
3977 memmove(ac
->entry
, &(ac
->entry
[tofree
]),
3978 sizeof(void *) * ac
->avail
);
3980 spin_unlock_irq(&n
->list_lock
);
3985 * cache_reap - Reclaim memory from caches.
3986 * @w: work descriptor
3988 * Called from workqueue/eventd every few seconds.
3990 * - clear the per-cpu caches for this CPU.
3991 * - return freeable pages to the main free memory pool.
3993 * If we cannot acquire the cache chain mutex then just give up - we'll try
3994 * again on the next iteration.
3996 static void cache_reap(struct work_struct
*w
)
3998 struct kmem_cache
*searchp
;
3999 struct kmem_cache_node
*n
;
4000 int node
= numa_mem_id();
4001 struct delayed_work
*work
= to_delayed_work(w
);
4003 if (!mutex_trylock(&slab_mutex
))
4004 /* Give up. Setup the next iteration. */
4007 list_for_each_entry(searchp
, &slab_caches
, list
) {
4011 * We only take the node lock if absolutely necessary and we
4012 * have established with reasonable certainty that
4013 * we can do some work if the lock was obtained.
4015 n
= searchp
->node
[node
];
4017 reap_alien(searchp
, n
);
4019 drain_array(searchp
, n
, cpu_cache_get(searchp
), 0, node
);
4022 * These are racy checks but it does not matter
4023 * if we skip one check or scan twice.
4025 if (time_after(n
->next_reap
, jiffies
))
4028 n
->next_reap
= jiffies
+ REAPTIMEOUT_LIST3
;
4030 drain_array(searchp
, n
, n
->shared
, 0, node
);
4032 if (n
->free_touched
)
4033 n
->free_touched
= 0;
4037 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4038 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4039 STATS_ADD_REAPED(searchp
, freed
);
4045 mutex_unlock(&slab_mutex
);
4048 /* Set up the next iteration */
4049 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_CPUC
));
4052 #ifdef CONFIG_SLABINFO
4053 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4056 unsigned long active_objs
;
4057 unsigned long num_objs
;
4058 unsigned long active_slabs
= 0;
4059 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4063 struct kmem_cache_node
*n
;
4067 for_each_online_node(node
) {
4068 n
= cachep
->node
[node
];
4073 spin_lock_irq(&n
->list_lock
);
4075 list_for_each_entry(slabp
, &n
->slabs_full
, list
) {
4076 if (slabp
->inuse
!= cachep
->num
&& !error
)
4077 error
= "slabs_full accounting error";
4078 active_objs
+= cachep
->num
;
4081 list_for_each_entry(slabp
, &n
->slabs_partial
, list
) {
4082 if (slabp
->inuse
== cachep
->num
&& !error
)
4083 error
= "slabs_partial inuse accounting error";
4084 if (!slabp
->inuse
&& !error
)
4085 error
= "slabs_partial/inuse accounting error";
4086 active_objs
+= slabp
->inuse
;
4089 list_for_each_entry(slabp
, &n
->slabs_free
, list
) {
4090 if (slabp
->inuse
&& !error
)
4091 error
= "slabs_free/inuse accounting error";
4094 free_objects
+= n
->free_objects
;
4096 shared_avail
+= n
->shared
->avail
;
4098 spin_unlock_irq(&n
->list_lock
);
4100 num_slabs
+= active_slabs
;
4101 num_objs
= num_slabs
* cachep
->num
;
4102 if (num_objs
- active_objs
!= free_objects
&& !error
)
4103 error
= "free_objects accounting error";
4105 name
= cachep
->name
;
4107 printk(KERN_ERR
"slab: cache %s error: %s\n", name
, error
);
4109 sinfo
->active_objs
= active_objs
;
4110 sinfo
->num_objs
= num_objs
;
4111 sinfo
->active_slabs
= active_slabs
;
4112 sinfo
->num_slabs
= num_slabs
;
4113 sinfo
->shared_avail
= shared_avail
;
4114 sinfo
->limit
= cachep
->limit
;
4115 sinfo
->batchcount
= cachep
->batchcount
;
4116 sinfo
->shared
= cachep
->shared
;
4117 sinfo
->objects_per_slab
= cachep
->num
;
4118 sinfo
->cache_order
= cachep
->gfporder
;
4121 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4125 unsigned long high
= cachep
->high_mark
;
4126 unsigned long allocs
= cachep
->num_allocations
;
4127 unsigned long grown
= cachep
->grown
;
4128 unsigned long reaped
= cachep
->reaped
;
4129 unsigned long errors
= cachep
->errors
;
4130 unsigned long max_freeable
= cachep
->max_freeable
;
4131 unsigned long node_allocs
= cachep
->node_allocs
;
4132 unsigned long node_frees
= cachep
->node_frees
;
4133 unsigned long overflows
= cachep
->node_overflow
;
4135 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu "
4136 "%4lu %4lu %4lu %4lu %4lu",
4137 allocs
, high
, grown
,
4138 reaped
, errors
, max_freeable
, node_allocs
,
4139 node_frees
, overflows
);
4143 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4144 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4145 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4146 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4148 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4149 allochit
, allocmiss
, freehit
, freemiss
);
4154 #define MAX_SLABINFO_WRITE 128
4156 * slabinfo_write - Tuning for the slab allocator
4158 * @buffer: user buffer
4159 * @count: data length
4162 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4163 size_t count
, loff_t
*ppos
)
4165 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4166 int limit
, batchcount
, shared
, res
;
4167 struct kmem_cache
*cachep
;
4169 if (count
> MAX_SLABINFO_WRITE
)
4171 if (copy_from_user(&kbuf
, buffer
, count
))
4173 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4175 tmp
= strchr(kbuf
, ' ');
4180 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4183 /* Find the cache in the chain of caches. */
4184 mutex_lock(&slab_mutex
);
4186 list_for_each_entry(cachep
, &slab_caches
, list
) {
4187 if (!strcmp(cachep
->name
, kbuf
)) {
4188 if (limit
< 1 || batchcount
< 1 ||
4189 batchcount
> limit
|| shared
< 0) {
4192 res
= do_tune_cpucache(cachep
, limit
,
4199 mutex_unlock(&slab_mutex
);
4205 #ifdef CONFIG_DEBUG_SLAB_LEAK
4207 static void *leaks_start(struct seq_file
*m
, loff_t
*pos
)
4209 mutex_lock(&slab_mutex
);
4210 return seq_list_start(&slab_caches
, *pos
);
4213 static inline int add_caller(unsigned long *n
, unsigned long v
)
4223 unsigned long *q
= p
+ 2 * i
;
4237 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4243 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
, struct slab
*s
)
4250 for (i
= 0, p
= s
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4253 for (j
= s
->free
; j
< c
->num
; j
++) {
4254 /* Skip freed item */
4255 if (slab_bufctl(s
)[j
] == i
) {
4263 if (!add_caller(n
, (unsigned long)*dbg_userword(c
, p
)))
4268 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4270 #ifdef CONFIG_KALLSYMS
4271 unsigned long offset
, size
;
4272 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4274 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4275 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4277 seq_printf(m
, " [%s]", modname
);
4281 seq_printf(m
, "%p", (void *)address
);
4284 static int leaks_show(struct seq_file
*m
, void *p
)
4286 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4288 struct kmem_cache_node
*n
;
4290 unsigned long *x
= m
->private;
4294 if (!(cachep
->flags
& SLAB_STORE_USER
))
4296 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4299 /* OK, we can do it */
4303 for_each_online_node(node
) {
4304 n
= cachep
->node
[node
];
4309 spin_lock_irq(&n
->list_lock
);
4311 list_for_each_entry(slabp
, &n
->slabs_full
, list
)
4312 handle_slab(x
, cachep
, slabp
);
4313 list_for_each_entry(slabp
, &n
->slabs_partial
, list
)
4314 handle_slab(x
, cachep
, slabp
);
4315 spin_unlock_irq(&n
->list_lock
);
4317 name
= cachep
->name
;
4319 /* Increase the buffer size */
4320 mutex_unlock(&slab_mutex
);
4321 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4323 /* Too bad, we are really out */
4325 mutex_lock(&slab_mutex
);
4328 *(unsigned long *)m
->private = x
[0] * 2;
4330 mutex_lock(&slab_mutex
);
4331 /* Now make sure this entry will be retried */
4335 for (i
= 0; i
< x
[1]; i
++) {
4336 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4337 show_symbol(m
, x
[2*i
+2]);
4344 static const struct seq_operations slabstats_op
= {
4345 .start
= leaks_start
,
4351 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4353 unsigned long *n
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
4356 ret
= seq_open(file
, &slabstats_op
);
4358 struct seq_file
*m
= file
->private_data
;
4359 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4368 static const struct file_operations proc_slabstats_operations
= {
4369 .open
= slabstats_open
,
4371 .llseek
= seq_lseek
,
4372 .release
= seq_release_private
,
4376 static int __init
slab_proc_init(void)
4378 #ifdef CONFIG_DEBUG_SLAB_LEAK
4379 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4383 module_init(slab_proc_init
);
4387 * ksize - get the actual amount of memory allocated for a given object
4388 * @objp: Pointer to the object
4390 * kmalloc may internally round up allocations and return more memory
4391 * than requested. ksize() can be used to determine the actual amount of
4392 * memory allocated. The caller may use this additional memory, even though
4393 * a smaller amount of memory was initially specified with the kmalloc call.
4394 * The caller must guarantee that objp points to a valid object previously
4395 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4396 * must not be freed during the duration of the call.
4398 size_t ksize(const void *objp
)
4401 if (unlikely(objp
== ZERO_SIZE_PTR
))
4404 return virt_to_cache(objp
)->object_size
;
4406 EXPORT_SYMBOL(ksize
);