3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t
;
166 typedef unsigned short freelist_idx_t
;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
179 * The limit is stored in the per-cpu structure to reduce the data cache
186 unsigned int batchcount
;
187 unsigned int touched
;
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
197 struct array_cache ac
;
201 * Need this for bootstrapping a per node allocator.
203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
205 #define CACHE_CACHE 0
206 #define SIZE_NODE (MAX_NUMNODES)
208 static int drain_freelist(struct kmem_cache
*cache
,
209 struct kmem_cache_node
*n
, int tofree
);
210 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
211 int node
, struct list_head
*list
);
212 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
213 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
214 static void cache_reap(struct work_struct
*unused
);
216 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
218 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
219 struct kmem_cache_node
*n
, struct page
*page
,
221 static int slab_early_init
= 1;
223 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
225 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
227 INIT_LIST_HEAD(&parent
->slabs_full
);
228 INIT_LIST_HEAD(&parent
->slabs_partial
);
229 INIT_LIST_HEAD(&parent
->slabs_free
);
230 parent
->shared
= NULL
;
231 parent
->alien
= NULL
;
232 parent
->colour_next
= 0;
233 spin_lock_init(&parent
->list_lock
);
234 parent
->free_objects
= 0;
235 parent
->free_touched
= 0;
238 #define MAKE_LIST(cachep, listp, slab, nodeid) \
240 INIT_LIST_HEAD(listp); \
241 list_splice(&get_node(cachep, nodeid)->slab, listp); \
244 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
246 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
247 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
251 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
252 #define CFLGS_OFF_SLAB (0x80000000UL)
253 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
254 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
256 #define BATCHREFILL_LIMIT 16
258 * Optimization question: fewer reaps means less probability for unnessary
259 * cpucache drain/refill cycles.
261 * OTOH the cpuarrays can contain lots of objects,
262 * which could lock up otherwise freeable slabs.
264 #define REAPTIMEOUT_AC (2*HZ)
265 #define REAPTIMEOUT_NODE (4*HZ)
268 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
269 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
270 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
271 #define STATS_INC_GROWN(x) ((x)->grown++)
272 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
273 #define STATS_SET_HIGH(x) \
275 if ((x)->num_active > (x)->high_mark) \
276 (x)->high_mark = (x)->num_active; \
278 #define STATS_INC_ERR(x) ((x)->errors++)
279 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
280 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
281 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
282 #define STATS_SET_FREEABLE(x, i) \
284 if ((x)->max_freeable < i) \
285 (x)->max_freeable = i; \
287 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
288 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
289 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
290 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
292 #define STATS_INC_ACTIVE(x) do { } while (0)
293 #define STATS_DEC_ACTIVE(x) do { } while (0)
294 #define STATS_INC_ALLOCED(x) do { } while (0)
295 #define STATS_INC_GROWN(x) do { } while (0)
296 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
297 #define STATS_SET_HIGH(x) do { } while (0)
298 #define STATS_INC_ERR(x) do { } while (0)
299 #define STATS_INC_NODEALLOCS(x) do { } while (0)
300 #define STATS_INC_NODEFREES(x) do { } while (0)
301 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
302 #define STATS_SET_FREEABLE(x, i) do { } while (0)
303 #define STATS_INC_ALLOCHIT(x) do { } while (0)
304 #define STATS_INC_ALLOCMISS(x) do { } while (0)
305 #define STATS_INC_FREEHIT(x) do { } while (0)
306 #define STATS_INC_FREEMISS(x) do { } while (0)
312 * memory layout of objects:
314 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
315 * the end of an object is aligned with the end of the real
316 * allocation. Catches writes behind the end of the allocation.
317 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
319 * cachep->obj_offset: The real object.
320 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
321 * cachep->size - 1* BYTES_PER_WORD: last caller address
322 * [BYTES_PER_WORD long]
324 static int obj_offset(struct kmem_cache
*cachep
)
326 return cachep
->obj_offset
;
329 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
331 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
332 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
333 sizeof(unsigned long long));
336 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
338 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
339 if (cachep
->flags
& SLAB_STORE_USER
)
340 return (unsigned long long *)(objp
+ cachep
->size
-
341 sizeof(unsigned long long) -
343 return (unsigned long long *) (objp
+ cachep
->size
-
344 sizeof(unsigned long long));
347 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
349 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
350 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
355 #define obj_offset(x) 0
356 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
357 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
358 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
362 #ifdef CONFIG_DEBUG_SLAB_LEAK
364 static inline bool is_store_user_clean(struct kmem_cache
*cachep
)
366 return atomic_read(&cachep
->store_user_clean
) == 1;
369 static inline void set_store_user_clean(struct kmem_cache
*cachep
)
371 atomic_set(&cachep
->store_user_clean
, 1);
374 static inline void set_store_user_dirty(struct kmem_cache
*cachep
)
376 if (is_store_user_clean(cachep
))
377 atomic_set(&cachep
->store_user_clean
, 0);
381 static inline void set_store_user_dirty(struct kmem_cache
*cachep
) {}
386 * Do not go above this order unless 0 objects fit into the slab or
387 * overridden on the command line.
389 #define SLAB_MAX_ORDER_HI 1
390 #define SLAB_MAX_ORDER_LO 0
391 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
392 static bool slab_max_order_set __initdata
;
394 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
396 struct page
*page
= virt_to_head_page(obj
);
397 return page
->slab_cache
;
400 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
403 return page
->s_mem
+ cache
->size
* idx
;
407 * We want to avoid an expensive divide : (offset / cache->size)
408 * Using the fact that size is a constant for a particular cache,
409 * we can replace (offset / cache->size) by
410 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
412 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
413 const struct page
*page
, void *obj
)
415 u32 offset
= (obj
- page
->s_mem
);
416 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
419 #define BOOT_CPUCACHE_ENTRIES 1
420 /* internal cache of cache description objs */
421 static struct kmem_cache kmem_cache_boot
= {
423 .limit
= BOOT_CPUCACHE_ENTRIES
,
425 .size
= sizeof(struct kmem_cache
),
426 .name
= "kmem_cache",
429 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
431 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
433 return this_cpu_ptr(cachep
->cpu_cache
);
437 * Calculate the number of objects and left-over bytes for a given buffer size.
439 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
440 unsigned long flags
, size_t *left_over
)
443 size_t slab_size
= PAGE_SIZE
<< gfporder
;
446 * The slab management structure can be either off the slab or
447 * on it. For the latter case, the memory allocated for a
450 * - @buffer_size bytes for each object
451 * - One freelist_idx_t for each object
453 * We don't need to consider alignment of freelist because
454 * freelist will be at the end of slab page. The objects will be
455 * at the correct alignment.
457 * If the slab management structure is off the slab, then the
458 * alignment will already be calculated into the size. Because
459 * the slabs are all pages aligned, the objects will be at the
460 * correct alignment when allocated.
462 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
463 num
= slab_size
/ buffer_size
;
464 *left_over
= slab_size
% buffer_size
;
466 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
467 *left_over
= slab_size
%
468 (buffer_size
+ sizeof(freelist_idx_t
));
475 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
477 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
480 pr_err("slab error in %s(): cache `%s': %s\n",
481 function
, cachep
->name
, msg
);
483 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
488 * By default on NUMA we use alien caches to stage the freeing of
489 * objects allocated from other nodes. This causes massive memory
490 * inefficiencies when using fake NUMA setup to split memory into a
491 * large number of small nodes, so it can be disabled on the command
495 static int use_alien_caches __read_mostly
= 1;
496 static int __init
noaliencache_setup(char *s
)
498 use_alien_caches
= 0;
501 __setup("noaliencache", noaliencache_setup
);
503 static int __init
slab_max_order_setup(char *str
)
505 get_option(&str
, &slab_max_order
);
506 slab_max_order
= slab_max_order
< 0 ? 0 :
507 min(slab_max_order
, MAX_ORDER
- 1);
508 slab_max_order_set
= true;
512 __setup("slab_max_order=", slab_max_order_setup
);
516 * Special reaping functions for NUMA systems called from cache_reap().
517 * These take care of doing round robin flushing of alien caches (containing
518 * objects freed on different nodes from which they were allocated) and the
519 * flushing of remote pcps by calling drain_node_pages.
521 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
523 static void init_reap_node(int cpu
)
525 per_cpu(slab_reap_node
, cpu
) = next_node_in(cpu_to_mem(cpu
),
529 static void next_reap_node(void)
531 int node
= __this_cpu_read(slab_reap_node
);
533 node
= next_node_in(node
, node_online_map
);
534 __this_cpu_write(slab_reap_node
, node
);
538 #define init_reap_node(cpu) do { } while (0)
539 #define next_reap_node(void) do { } while (0)
543 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
544 * via the workqueue/eventd.
545 * Add the CPU number into the expiration time to minimize the possibility of
546 * the CPUs getting into lockstep and contending for the global cache chain
549 static void start_cpu_timer(int cpu
)
551 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
554 * When this gets called from do_initcalls via cpucache_init(),
555 * init_workqueues() has already run, so keventd will be setup
558 if (keventd_up() && reap_work
->work
.func
== NULL
) {
560 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
561 schedule_delayed_work_on(cpu
, reap_work
,
562 __round_jiffies_relative(HZ
, cpu
));
566 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
569 * The array_cache structures contain pointers to free object.
570 * However, when such objects are allocated or transferred to another
571 * cache the pointers are not cleared and they could be counted as
572 * valid references during a kmemleak scan. Therefore, kmemleak must
573 * not scan such objects.
575 kmemleak_no_scan(ac
);
579 ac
->batchcount
= batch
;
584 static struct array_cache
*alloc_arraycache(int node
, int entries
,
585 int batchcount
, gfp_t gfp
)
587 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
588 struct array_cache
*ac
= NULL
;
590 ac
= kmalloc_node(memsize
, gfp
, node
);
591 init_arraycache(ac
, entries
, batchcount
);
595 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
596 struct page
*page
, void *objp
)
598 struct kmem_cache_node
*n
;
602 page_node
= page_to_nid(page
);
603 n
= get_node(cachep
, page_node
);
605 spin_lock(&n
->list_lock
);
606 free_block(cachep
, &objp
, 1, page_node
, &list
);
607 spin_unlock(&n
->list_lock
);
609 slabs_destroy(cachep
, &list
);
613 * Transfer objects in one arraycache to another.
614 * Locking must be handled by the caller.
616 * Return the number of entries transferred.
618 static int transfer_objects(struct array_cache
*to
,
619 struct array_cache
*from
, unsigned int max
)
621 /* Figure out how many entries to transfer */
622 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
627 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
637 #define drain_alien_cache(cachep, alien) do { } while (0)
638 #define reap_alien(cachep, n) do { } while (0)
640 static inline struct alien_cache
**alloc_alien_cache(int node
,
641 int limit
, gfp_t gfp
)
646 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
650 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
655 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
661 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
662 gfp_t flags
, int nodeid
)
667 static inline gfp_t
gfp_exact_node(gfp_t flags
)
669 return flags
& ~__GFP_NOFAIL
;
672 #else /* CONFIG_NUMA */
674 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
675 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
677 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
678 int batch
, gfp_t gfp
)
680 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
681 struct alien_cache
*alc
= NULL
;
683 alc
= kmalloc_node(memsize
, gfp
, node
);
684 init_arraycache(&alc
->ac
, entries
, batch
);
685 spin_lock_init(&alc
->lock
);
689 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
691 struct alien_cache
**alc_ptr
;
692 size_t memsize
= sizeof(void *) * nr_node_ids
;
697 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
702 if (i
== node
|| !node_online(i
))
704 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
706 for (i
--; i
>= 0; i
--)
715 static void free_alien_cache(struct alien_cache
**alc_ptr
)
726 static void __drain_alien_cache(struct kmem_cache
*cachep
,
727 struct array_cache
*ac
, int node
,
728 struct list_head
*list
)
730 struct kmem_cache_node
*n
= get_node(cachep
, node
);
733 spin_lock(&n
->list_lock
);
735 * Stuff objects into the remote nodes shared array first.
736 * That way we could avoid the overhead of putting the objects
737 * into the free lists and getting them back later.
740 transfer_objects(n
->shared
, ac
, ac
->limit
);
742 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
744 spin_unlock(&n
->list_lock
);
749 * Called from cache_reap() to regularly drain alien caches round robin.
751 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
753 int node
= __this_cpu_read(slab_reap_node
);
756 struct alien_cache
*alc
= n
->alien
[node
];
757 struct array_cache
*ac
;
761 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
764 __drain_alien_cache(cachep
, ac
, node
, &list
);
765 spin_unlock_irq(&alc
->lock
);
766 slabs_destroy(cachep
, &list
);
772 static void drain_alien_cache(struct kmem_cache
*cachep
,
773 struct alien_cache
**alien
)
776 struct alien_cache
*alc
;
777 struct array_cache
*ac
;
780 for_each_online_node(i
) {
786 spin_lock_irqsave(&alc
->lock
, flags
);
787 __drain_alien_cache(cachep
, ac
, i
, &list
);
788 spin_unlock_irqrestore(&alc
->lock
, flags
);
789 slabs_destroy(cachep
, &list
);
794 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
795 int node
, int page_node
)
797 struct kmem_cache_node
*n
;
798 struct alien_cache
*alien
= NULL
;
799 struct array_cache
*ac
;
802 n
= get_node(cachep
, node
);
803 STATS_INC_NODEFREES(cachep
);
804 if (n
->alien
&& n
->alien
[page_node
]) {
805 alien
= n
->alien
[page_node
];
807 spin_lock(&alien
->lock
);
808 if (unlikely(ac
->avail
== ac
->limit
)) {
809 STATS_INC_ACOVERFLOW(cachep
);
810 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
812 ac
->entry
[ac
->avail
++] = objp
;
813 spin_unlock(&alien
->lock
);
814 slabs_destroy(cachep
, &list
);
816 n
= get_node(cachep
, page_node
);
817 spin_lock(&n
->list_lock
);
818 free_block(cachep
, &objp
, 1, page_node
, &list
);
819 spin_unlock(&n
->list_lock
);
820 slabs_destroy(cachep
, &list
);
825 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
827 int page_node
= page_to_nid(virt_to_page(objp
));
828 int node
= numa_mem_id();
830 * Make sure we are not freeing a object from another node to the array
833 if (likely(node
== page_node
))
836 return __cache_free_alien(cachep
, objp
, node
, page_node
);
840 * Construct gfp mask to allocate from a specific node but do not reclaim or
841 * warn about failures.
843 static inline gfp_t
gfp_exact_node(gfp_t flags
)
845 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
849 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
851 struct kmem_cache_node
*n
;
854 * Set up the kmem_cache_node for cpu before we can
855 * begin anything. Make sure some other cpu on this
856 * node has not already allocated this
858 n
= get_node(cachep
, node
);
860 spin_lock_irq(&n
->list_lock
);
861 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
863 spin_unlock_irq(&n
->list_lock
);
868 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
872 kmem_cache_node_init(n
);
873 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
874 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
877 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
880 * The kmem_cache_nodes don't come and go as CPUs
881 * come and go. slab_mutex is sufficient
884 cachep
->node
[node
] = n
;
890 * Allocates and initializes node for a node on each slab cache, used for
891 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
892 * will be allocated off-node since memory is not yet online for the new node.
893 * When hotplugging memory or a cpu, existing node are not replaced if
896 * Must hold slab_mutex.
898 static int init_cache_node_node(int node
)
901 struct kmem_cache
*cachep
;
903 list_for_each_entry(cachep
, &slab_caches
, list
) {
904 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
912 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
913 int node
, gfp_t gfp
, bool force_change
)
916 struct kmem_cache_node
*n
;
917 struct array_cache
*old_shared
= NULL
;
918 struct array_cache
*new_shared
= NULL
;
919 struct alien_cache
**new_alien
= NULL
;
922 if (use_alien_caches
) {
923 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
928 if (cachep
->shared
) {
929 new_shared
= alloc_arraycache(node
,
930 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
935 ret
= init_cache_node(cachep
, node
, gfp
);
939 n
= get_node(cachep
, node
);
940 spin_lock_irq(&n
->list_lock
);
941 if (n
->shared
&& force_change
) {
942 free_block(cachep
, n
->shared
->entry
,
943 n
->shared
->avail
, node
, &list
);
944 n
->shared
->avail
= 0;
947 if (!n
->shared
|| force_change
) {
948 old_shared
= n
->shared
;
949 n
->shared
= new_shared
;
954 n
->alien
= new_alien
;
958 spin_unlock_irq(&n
->list_lock
);
959 slabs_destroy(cachep
, &list
);
962 * To protect lockless access to n->shared during irq disabled context.
963 * If n->shared isn't NULL in irq disabled context, accessing to it is
964 * guaranteed to be valid until irq is re-enabled, because it will be
965 * freed after synchronize_sched().
973 free_alien_cache(new_alien
);
978 static void cpuup_canceled(long cpu
)
980 struct kmem_cache
*cachep
;
981 struct kmem_cache_node
*n
= NULL
;
982 int node
= cpu_to_mem(cpu
);
983 const struct cpumask
*mask
= cpumask_of_node(node
);
985 list_for_each_entry(cachep
, &slab_caches
, list
) {
986 struct array_cache
*nc
;
987 struct array_cache
*shared
;
988 struct alien_cache
**alien
;
991 n
= get_node(cachep
, node
);
995 spin_lock_irq(&n
->list_lock
);
997 /* Free limit for this kmem_cache_node */
998 n
->free_limit
-= cachep
->batchcount
;
1000 /* cpu is dead; no one can alloc from it. */
1001 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
1003 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1007 if (!cpumask_empty(mask
)) {
1008 spin_unlock_irq(&n
->list_lock
);
1014 free_block(cachep
, shared
->entry
,
1015 shared
->avail
, node
, &list
);
1022 spin_unlock_irq(&n
->list_lock
);
1026 drain_alien_cache(cachep
, alien
);
1027 free_alien_cache(alien
);
1031 slabs_destroy(cachep
, &list
);
1034 * In the previous loop, all the objects were freed to
1035 * the respective cache's slabs, now we can go ahead and
1036 * shrink each nodelist to its limit.
1038 list_for_each_entry(cachep
, &slab_caches
, list
) {
1039 n
= get_node(cachep
, node
);
1042 drain_freelist(cachep
, n
, INT_MAX
);
1046 static int cpuup_prepare(long cpu
)
1048 struct kmem_cache
*cachep
;
1049 int node
= cpu_to_mem(cpu
);
1053 * We need to do this right in the beginning since
1054 * alloc_arraycache's are going to use this list.
1055 * kmalloc_node allows us to add the slab to the right
1056 * kmem_cache_node and not this cpu's kmem_cache_node
1058 err
= init_cache_node_node(node
);
1063 * Now we can go ahead with allocating the shared arrays and
1066 list_for_each_entry(cachep
, &slab_caches
, list
) {
1067 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1074 cpuup_canceled(cpu
);
1078 static int cpuup_callback(struct notifier_block
*nfb
,
1079 unsigned long action
, void *hcpu
)
1081 long cpu
= (long)hcpu
;
1085 case CPU_UP_PREPARE
:
1086 case CPU_UP_PREPARE_FROZEN
:
1087 mutex_lock(&slab_mutex
);
1088 err
= cpuup_prepare(cpu
);
1089 mutex_unlock(&slab_mutex
);
1092 case CPU_ONLINE_FROZEN
:
1093 start_cpu_timer(cpu
);
1095 #ifdef CONFIG_HOTPLUG_CPU
1096 case CPU_DOWN_PREPARE
:
1097 case CPU_DOWN_PREPARE_FROZEN
:
1099 * Shutdown cache reaper. Note that the slab_mutex is
1100 * held so that if cache_reap() is invoked it cannot do
1101 * anything expensive but will only modify reap_work
1102 * and reschedule the timer.
1104 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1105 /* Now the cache_reaper is guaranteed to be not running. */
1106 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1108 case CPU_DOWN_FAILED
:
1109 case CPU_DOWN_FAILED_FROZEN
:
1110 start_cpu_timer(cpu
);
1113 case CPU_DEAD_FROZEN
:
1115 * Even if all the cpus of a node are down, we don't free the
1116 * kmem_cache_node of any cache. This to avoid a race between
1117 * cpu_down, and a kmalloc allocation from another cpu for
1118 * memory from the node of the cpu going down. The node
1119 * structure is usually allocated from kmem_cache_create() and
1120 * gets destroyed at kmem_cache_destroy().
1124 case CPU_UP_CANCELED
:
1125 case CPU_UP_CANCELED_FROZEN
:
1126 mutex_lock(&slab_mutex
);
1127 cpuup_canceled(cpu
);
1128 mutex_unlock(&slab_mutex
);
1131 return notifier_from_errno(err
);
1134 static struct notifier_block cpucache_notifier
= {
1135 &cpuup_callback
, NULL
, 0
1138 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1140 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1141 * Returns -EBUSY if all objects cannot be drained so that the node is not
1144 * Must hold slab_mutex.
1146 static int __meminit
drain_cache_node_node(int node
)
1148 struct kmem_cache
*cachep
;
1151 list_for_each_entry(cachep
, &slab_caches
, list
) {
1152 struct kmem_cache_node
*n
;
1154 n
= get_node(cachep
, node
);
1158 drain_freelist(cachep
, n
, INT_MAX
);
1160 if (!list_empty(&n
->slabs_full
) ||
1161 !list_empty(&n
->slabs_partial
)) {
1169 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1170 unsigned long action
, void *arg
)
1172 struct memory_notify
*mnb
= arg
;
1176 nid
= mnb
->status_change_nid
;
1181 case MEM_GOING_ONLINE
:
1182 mutex_lock(&slab_mutex
);
1183 ret
= init_cache_node_node(nid
);
1184 mutex_unlock(&slab_mutex
);
1186 case MEM_GOING_OFFLINE
:
1187 mutex_lock(&slab_mutex
);
1188 ret
= drain_cache_node_node(nid
);
1189 mutex_unlock(&slab_mutex
);
1193 case MEM_CANCEL_ONLINE
:
1194 case MEM_CANCEL_OFFLINE
:
1198 return notifier_from_errno(ret
);
1200 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1203 * swap the static kmem_cache_node with kmalloced memory
1205 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1208 struct kmem_cache_node
*ptr
;
1210 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1213 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1215 * Do not assume that spinlocks can be initialized via memcpy:
1217 spin_lock_init(&ptr
->list_lock
);
1219 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1220 cachep
->node
[nodeid
] = ptr
;
1224 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1225 * size of kmem_cache_node.
1227 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1231 for_each_online_node(node
) {
1232 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1233 cachep
->node
[node
]->next_reap
= jiffies
+
1235 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1240 * Initialisation. Called after the page allocator have been initialised and
1241 * before smp_init().
1243 void __init
kmem_cache_init(void)
1247 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1248 sizeof(struct rcu_head
));
1249 kmem_cache
= &kmem_cache_boot
;
1251 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1252 use_alien_caches
= 0;
1254 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1255 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1258 * Fragmentation resistance on low memory - only use bigger
1259 * page orders on machines with more than 32MB of memory if
1260 * not overridden on the command line.
1262 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1263 slab_max_order
= SLAB_MAX_ORDER_HI
;
1265 /* Bootstrap is tricky, because several objects are allocated
1266 * from caches that do not exist yet:
1267 * 1) initialize the kmem_cache cache: it contains the struct
1268 * kmem_cache structures of all caches, except kmem_cache itself:
1269 * kmem_cache is statically allocated.
1270 * Initially an __init data area is used for the head array and the
1271 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1272 * array at the end of the bootstrap.
1273 * 2) Create the first kmalloc cache.
1274 * The struct kmem_cache for the new cache is allocated normally.
1275 * An __init data area is used for the head array.
1276 * 3) Create the remaining kmalloc caches, with minimally sized
1278 * 4) Replace the __init data head arrays for kmem_cache and the first
1279 * kmalloc cache with kmalloc allocated arrays.
1280 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1281 * the other cache's with kmalloc allocated memory.
1282 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1285 /* 1) create the kmem_cache */
1288 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1290 create_boot_cache(kmem_cache
, "kmem_cache",
1291 offsetof(struct kmem_cache
, node
) +
1292 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1293 SLAB_HWCACHE_ALIGN
);
1294 list_add(&kmem_cache
->list
, &slab_caches
);
1295 slab_state
= PARTIAL
;
1298 * Initialize the caches that provide memory for the kmem_cache_node
1299 * structures first. Without this, further allocations will bug.
1301 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1302 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1303 slab_state
= PARTIAL_NODE
;
1304 setup_kmalloc_cache_index_table();
1306 slab_early_init
= 0;
1308 /* 5) Replace the bootstrap kmem_cache_node */
1312 for_each_online_node(nid
) {
1313 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1315 init_list(kmalloc_caches
[INDEX_NODE
],
1316 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1320 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1323 void __init
kmem_cache_init_late(void)
1325 struct kmem_cache
*cachep
;
1329 /* 6) resize the head arrays to their final sizes */
1330 mutex_lock(&slab_mutex
);
1331 list_for_each_entry(cachep
, &slab_caches
, list
)
1332 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1334 mutex_unlock(&slab_mutex
);
1340 * Register a cpu startup notifier callback that initializes
1341 * cpu_cache_get for all new cpus
1343 register_cpu_notifier(&cpucache_notifier
);
1347 * Register a memory hotplug callback that initializes and frees
1350 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1354 * The reap timers are started later, with a module init call: That part
1355 * of the kernel is not yet operational.
1359 static int __init
cpucache_init(void)
1364 * Register the timers that return unneeded pages to the page allocator
1366 for_each_online_cpu(cpu
)
1367 start_cpu_timer(cpu
);
1373 __initcall(cpucache_init
);
1375 static noinline
void
1376 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1379 struct kmem_cache_node
*n
;
1381 unsigned long flags
;
1383 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1384 DEFAULT_RATELIMIT_BURST
);
1386 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1389 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1390 nodeid
, gfpflags
, &gfpflags
);
1391 pr_warn(" cache: %s, object size: %d, order: %d\n",
1392 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1394 for_each_kmem_cache_node(cachep
, node
, n
) {
1395 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1396 unsigned long active_slabs
= 0, num_slabs
= 0;
1398 spin_lock_irqsave(&n
->list_lock
, flags
);
1399 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1400 active_objs
+= cachep
->num
;
1403 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1404 active_objs
+= page
->active
;
1407 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1410 free_objects
+= n
->free_objects
;
1411 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1413 num_slabs
+= active_slabs
;
1414 num_objs
= num_slabs
* cachep
->num
;
1415 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1416 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1423 * Interface to system's page allocator. No need to hold the
1424 * kmem_cache_node ->list_lock.
1426 * If we requested dmaable memory, we will get it. Even if we
1427 * did not request dmaable memory, we might get it, but that
1428 * would be relatively rare and ignorable.
1430 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1436 flags
|= cachep
->allocflags
;
1437 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1438 flags
|= __GFP_RECLAIMABLE
;
1440 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1442 slab_out_of_memory(cachep
, flags
, nodeid
);
1446 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1447 __free_pages(page
, cachep
->gfporder
);
1451 nr_pages
= (1 << cachep
->gfporder
);
1452 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1453 add_zone_page_state(page_zone(page
),
1454 NR_SLAB_RECLAIMABLE
, nr_pages
);
1456 add_zone_page_state(page_zone(page
),
1457 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1459 __SetPageSlab(page
);
1460 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1461 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1462 SetPageSlabPfmemalloc(page
);
1464 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1465 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1468 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1470 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1477 * Interface to system's page release.
1479 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1481 int order
= cachep
->gfporder
;
1482 unsigned long nr_freed
= (1 << order
);
1484 kmemcheck_free_shadow(page
, order
);
1486 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1487 sub_zone_page_state(page_zone(page
),
1488 NR_SLAB_RECLAIMABLE
, nr_freed
);
1490 sub_zone_page_state(page_zone(page
),
1491 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1493 BUG_ON(!PageSlab(page
));
1494 __ClearPageSlabPfmemalloc(page
);
1495 __ClearPageSlab(page
);
1496 page_mapcount_reset(page
);
1497 page
->mapping
= NULL
;
1499 if (current
->reclaim_state
)
1500 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1501 memcg_uncharge_slab(page
, order
, cachep
);
1502 __free_pages(page
, order
);
1505 static void kmem_rcu_free(struct rcu_head
*head
)
1507 struct kmem_cache
*cachep
;
1510 page
= container_of(head
, struct page
, rcu_head
);
1511 cachep
= page
->slab_cache
;
1513 kmem_freepages(cachep
, page
);
1517 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1519 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1520 (cachep
->size
% PAGE_SIZE
) == 0)
1526 #ifdef CONFIG_DEBUG_PAGEALLOC
1527 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1528 unsigned long caller
)
1530 int size
= cachep
->object_size
;
1532 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1534 if (size
< 5 * sizeof(unsigned long))
1537 *addr
++ = 0x12345678;
1539 *addr
++ = smp_processor_id();
1540 size
-= 3 * sizeof(unsigned long);
1542 unsigned long *sptr
= &caller
;
1543 unsigned long svalue
;
1545 while (!kstack_end(sptr
)) {
1547 if (kernel_text_address(svalue
)) {
1549 size
-= sizeof(unsigned long);
1550 if (size
<= sizeof(unsigned long))
1556 *addr
++ = 0x87654321;
1559 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1560 int map
, unsigned long caller
)
1562 if (!is_debug_pagealloc_cache(cachep
))
1566 store_stackinfo(cachep
, objp
, caller
);
1568 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1572 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1573 int map
, unsigned long caller
) {}
1577 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1579 int size
= cachep
->object_size
;
1580 addr
= &((char *)addr
)[obj_offset(cachep
)];
1582 memset(addr
, val
, size
);
1583 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1586 static void dump_line(char *data
, int offset
, int limit
)
1589 unsigned char error
= 0;
1592 pr_err("%03x: ", offset
);
1593 for (i
= 0; i
< limit
; i
++) {
1594 if (data
[offset
+ i
] != POISON_FREE
) {
1595 error
= data
[offset
+ i
];
1599 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1600 &data
[offset
], limit
, 1);
1602 if (bad_count
== 1) {
1603 error
^= POISON_FREE
;
1604 if (!(error
& (error
- 1))) {
1605 pr_err("Single bit error detected. Probably bad RAM.\n");
1607 pr_err("Run memtest86+ or a similar memory test tool.\n");
1609 pr_err("Run a memory test tool.\n");
1618 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1623 if (cachep
->flags
& SLAB_RED_ZONE
) {
1624 pr_err("Redzone: 0x%llx/0x%llx\n",
1625 *dbg_redzone1(cachep
, objp
),
1626 *dbg_redzone2(cachep
, objp
));
1629 if (cachep
->flags
& SLAB_STORE_USER
) {
1630 pr_err("Last user: [<%p>](%pSR)\n",
1631 *dbg_userword(cachep
, objp
),
1632 *dbg_userword(cachep
, objp
));
1634 realobj
= (char *)objp
+ obj_offset(cachep
);
1635 size
= cachep
->object_size
;
1636 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1639 if (i
+ limit
> size
)
1641 dump_line(realobj
, i
, limit
);
1645 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1651 if (is_debug_pagealloc_cache(cachep
))
1654 realobj
= (char *)objp
+ obj_offset(cachep
);
1655 size
= cachep
->object_size
;
1657 for (i
= 0; i
< size
; i
++) {
1658 char exp
= POISON_FREE
;
1661 if (realobj
[i
] != exp
) {
1666 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1667 print_tainted(), cachep
->name
,
1669 print_objinfo(cachep
, objp
, 0);
1671 /* Hexdump the affected line */
1674 if (i
+ limit
> size
)
1676 dump_line(realobj
, i
, limit
);
1679 /* Limit to 5 lines */
1685 /* Print some data about the neighboring objects, if they
1688 struct page
*page
= virt_to_head_page(objp
);
1691 objnr
= obj_to_index(cachep
, page
, objp
);
1693 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1694 realobj
= (char *)objp
+ obj_offset(cachep
);
1695 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1696 print_objinfo(cachep
, objp
, 2);
1698 if (objnr
+ 1 < cachep
->num
) {
1699 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1700 realobj
= (char *)objp
+ obj_offset(cachep
);
1701 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1702 print_objinfo(cachep
, objp
, 2);
1709 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1714 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1715 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1719 for (i
= 0; i
< cachep
->num
; i
++) {
1720 void *objp
= index_to_obj(cachep
, page
, i
);
1722 if (cachep
->flags
& SLAB_POISON
) {
1723 check_poison_obj(cachep
, objp
);
1724 slab_kernel_map(cachep
, objp
, 1, 0);
1726 if (cachep
->flags
& SLAB_RED_ZONE
) {
1727 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1728 slab_error(cachep
, "start of a freed object was overwritten");
1729 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1730 slab_error(cachep
, "end of a freed object was overwritten");
1735 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1742 * slab_destroy - destroy and release all objects in a slab
1743 * @cachep: cache pointer being destroyed
1744 * @page: page pointer being destroyed
1746 * Destroy all the objs in a slab page, and release the mem back to the system.
1747 * Before calling the slab page must have been unlinked from the cache. The
1748 * kmem_cache_node ->list_lock is not held/needed.
1750 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1754 freelist
= page
->freelist
;
1755 slab_destroy_debugcheck(cachep
, page
);
1756 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1757 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1759 kmem_freepages(cachep
, page
);
1762 * From now on, we don't use freelist
1763 * although actual page can be freed in rcu context
1765 if (OFF_SLAB(cachep
))
1766 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1769 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1771 struct page
*page
, *n
;
1773 list_for_each_entry_safe(page
, n
, list
, lru
) {
1774 list_del(&page
->lru
);
1775 slab_destroy(cachep
, page
);
1780 * calculate_slab_order - calculate size (page order) of slabs
1781 * @cachep: pointer to the cache that is being created
1782 * @size: size of objects to be created in this cache.
1783 * @flags: slab allocation flags
1785 * Also calculates the number of objects per slab.
1787 * This could be made much more intelligent. For now, try to avoid using
1788 * high order pages for slabs. When the gfp() functions are more friendly
1789 * towards high-order requests, this should be changed.
1791 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1792 size_t size
, unsigned long flags
)
1794 size_t left_over
= 0;
1797 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1801 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1805 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1806 if (num
> SLAB_OBJ_MAX_NUM
)
1809 if (flags
& CFLGS_OFF_SLAB
) {
1810 struct kmem_cache
*freelist_cache
;
1811 size_t freelist_size
;
1813 freelist_size
= num
* sizeof(freelist_idx_t
);
1814 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1815 if (!freelist_cache
)
1819 * Needed to avoid possible looping condition
1820 * in cache_grow_begin()
1822 if (OFF_SLAB(freelist_cache
))
1825 /* check if off slab has enough benefit */
1826 if (freelist_cache
->size
> cachep
->size
/ 2)
1830 /* Found something acceptable - save it away */
1832 cachep
->gfporder
= gfporder
;
1833 left_over
= remainder
;
1836 * A VFS-reclaimable slab tends to have most allocations
1837 * as GFP_NOFS and we really don't want to have to be allocating
1838 * higher-order pages when we are unable to shrink dcache.
1840 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1844 * Large number of objects is good, but very large slabs are
1845 * currently bad for the gfp()s.
1847 if (gfporder
>= slab_max_order
)
1851 * Acceptable internal fragmentation?
1853 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1859 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1860 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1864 struct array_cache __percpu
*cpu_cache
;
1866 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1867 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1872 for_each_possible_cpu(cpu
) {
1873 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1874 entries
, batchcount
);
1880 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1882 if (slab_state
>= FULL
)
1883 return enable_cpucache(cachep
, gfp
);
1885 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1886 if (!cachep
->cpu_cache
)
1889 if (slab_state
== DOWN
) {
1890 /* Creation of first cache (kmem_cache). */
1891 set_up_node(kmem_cache
, CACHE_CACHE
);
1892 } else if (slab_state
== PARTIAL
) {
1893 /* For kmem_cache_node */
1894 set_up_node(cachep
, SIZE_NODE
);
1898 for_each_online_node(node
) {
1899 cachep
->node
[node
] = kmalloc_node(
1900 sizeof(struct kmem_cache_node
), gfp
, node
);
1901 BUG_ON(!cachep
->node
[node
]);
1902 kmem_cache_node_init(cachep
->node
[node
]);
1906 cachep
->node
[numa_mem_id()]->next_reap
=
1907 jiffies
+ REAPTIMEOUT_NODE
+
1908 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1910 cpu_cache_get(cachep
)->avail
= 0;
1911 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1912 cpu_cache_get(cachep
)->batchcount
= 1;
1913 cpu_cache_get(cachep
)->touched
= 0;
1914 cachep
->batchcount
= 1;
1915 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1919 unsigned long kmem_cache_flags(unsigned long object_size
,
1920 unsigned long flags
, const char *name
,
1921 void (*ctor
)(void *))
1927 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1928 unsigned long flags
, void (*ctor
)(void *))
1930 struct kmem_cache
*cachep
;
1932 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1937 * Adjust the object sizes so that we clear
1938 * the complete object on kzalloc.
1940 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1945 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1946 size_t size
, unsigned long flags
)
1952 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1955 left
= calculate_slab_order(cachep
, size
,
1956 flags
| CFLGS_OBJFREELIST_SLAB
);
1960 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1963 cachep
->colour
= left
/ cachep
->colour_off
;
1968 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1969 size_t size
, unsigned long flags
)
1976 * Always use on-slab management when SLAB_NOLEAKTRACE
1977 * to avoid recursive calls into kmemleak.
1979 if (flags
& SLAB_NOLEAKTRACE
)
1983 * Size is large, assume best to place the slab management obj
1984 * off-slab (should allow better packing of objs).
1986 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1991 * If the slab has been placed off-slab, and we have enough space then
1992 * move it on-slab. This is at the expense of any extra colouring.
1994 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1997 cachep
->colour
= left
/ cachep
->colour_off
;
2002 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
2003 size_t size
, unsigned long flags
)
2009 left
= calculate_slab_order(cachep
, size
, flags
);
2013 cachep
->colour
= left
/ cachep
->colour_off
;
2019 * __kmem_cache_create - Create a cache.
2020 * @cachep: cache management descriptor
2021 * @flags: SLAB flags
2023 * Returns a ptr to the cache on success, NULL on failure.
2024 * Cannot be called within a int, but can be interrupted.
2025 * The @ctor is run when new pages are allocated by the cache.
2029 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2030 * to catch references to uninitialised memory.
2032 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2033 * for buffer overruns.
2035 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2036 * cacheline. This can be beneficial if you're counting cycles as closely
2040 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2042 size_t ralign
= BYTES_PER_WORD
;
2045 size_t size
= cachep
->size
;
2050 * Enable redzoning and last user accounting, except for caches with
2051 * large objects, if the increased size would increase the object size
2052 * above the next power of two: caches with object sizes just above a
2053 * power of two have a significant amount of internal fragmentation.
2055 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2056 2 * sizeof(unsigned long long)))
2057 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2058 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2059 flags
|= SLAB_POISON
;
2064 * Check that size is in terms of words. This is needed to avoid
2065 * unaligned accesses for some archs when redzoning is used, and makes
2066 * sure any on-slab bufctl's are also correctly aligned.
2068 if (size
& (BYTES_PER_WORD
- 1)) {
2069 size
+= (BYTES_PER_WORD
- 1);
2070 size
&= ~(BYTES_PER_WORD
- 1);
2073 if (flags
& SLAB_RED_ZONE
) {
2074 ralign
= REDZONE_ALIGN
;
2075 /* If redzoning, ensure that the second redzone is suitably
2076 * aligned, by adjusting the object size accordingly. */
2077 size
+= REDZONE_ALIGN
- 1;
2078 size
&= ~(REDZONE_ALIGN
- 1);
2081 /* 3) caller mandated alignment */
2082 if (ralign
< cachep
->align
) {
2083 ralign
= cachep
->align
;
2085 /* disable debug if necessary */
2086 if (ralign
> __alignof__(unsigned long long))
2087 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2091 cachep
->align
= ralign
;
2092 cachep
->colour_off
= cache_line_size();
2093 /* Offset must be a multiple of the alignment. */
2094 if (cachep
->colour_off
< cachep
->align
)
2095 cachep
->colour_off
= cachep
->align
;
2097 if (slab_is_available())
2105 * Both debugging options require word-alignment which is calculated
2108 if (flags
& SLAB_RED_ZONE
) {
2109 /* add space for red zone words */
2110 cachep
->obj_offset
+= sizeof(unsigned long long);
2111 size
+= 2 * sizeof(unsigned long long);
2113 if (flags
& SLAB_STORE_USER
) {
2114 /* user store requires one word storage behind the end of
2115 * the real object. But if the second red zone needs to be
2116 * aligned to 64 bits, we must allow that much space.
2118 if (flags
& SLAB_RED_ZONE
)
2119 size
+= REDZONE_ALIGN
;
2121 size
+= BYTES_PER_WORD
;
2125 kasan_cache_create(cachep
, &size
, &flags
);
2127 size
= ALIGN(size
, cachep
->align
);
2129 * We should restrict the number of objects in a slab to implement
2130 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2132 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2133 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2137 * To activate debug pagealloc, off-slab management is necessary
2138 * requirement. In early phase of initialization, small sized slab
2139 * doesn't get initialized so it would not be possible. So, we need
2140 * to check size >= 256. It guarantees that all necessary small
2141 * sized slab is initialized in current slab initialization sequence.
2143 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2144 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2145 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2146 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2148 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2149 flags
|= CFLGS_OFF_SLAB
;
2150 cachep
->obj_offset
+= tmp_size
- size
;
2158 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2159 flags
|= CFLGS_OBJFREELIST_SLAB
;
2163 if (set_off_slab_cache(cachep
, size
, flags
)) {
2164 flags
|= CFLGS_OFF_SLAB
;
2168 if (set_on_slab_cache(cachep
, size
, flags
))
2174 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2175 cachep
->flags
= flags
;
2176 cachep
->allocflags
= __GFP_COMP
;
2177 if (flags
& SLAB_CACHE_DMA
)
2178 cachep
->allocflags
|= GFP_DMA
;
2179 cachep
->size
= size
;
2180 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2184 * If we're going to use the generic kernel_map_pages()
2185 * poisoning, then it's going to smash the contents of
2186 * the redzone and userword anyhow, so switch them off.
2188 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2189 (cachep
->flags
& SLAB_POISON
) &&
2190 is_debug_pagealloc_cache(cachep
))
2191 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2194 if (OFF_SLAB(cachep
)) {
2195 cachep
->freelist_cache
=
2196 kmalloc_slab(cachep
->freelist_size
, 0u);
2199 err
= setup_cpu_cache(cachep
, gfp
);
2201 __kmem_cache_release(cachep
);
2209 static void check_irq_off(void)
2211 BUG_ON(!irqs_disabled());
2214 static void check_irq_on(void)
2216 BUG_ON(irqs_disabled());
2219 static void check_mutex_acquired(void)
2221 BUG_ON(!mutex_is_locked(&slab_mutex
));
2224 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2228 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2232 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2236 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2241 #define check_irq_off() do { } while(0)
2242 #define check_irq_on() do { } while(0)
2243 #define check_mutex_acquired() do { } while(0)
2244 #define check_spinlock_acquired(x) do { } while(0)
2245 #define check_spinlock_acquired_node(x, y) do { } while(0)
2248 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2249 int node
, bool free_all
, struct list_head
*list
)
2253 if (!ac
|| !ac
->avail
)
2256 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2257 if (tofree
> ac
->avail
)
2258 tofree
= (ac
->avail
+ 1) / 2;
2260 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2261 ac
->avail
-= tofree
;
2262 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2265 static void do_drain(void *arg
)
2267 struct kmem_cache
*cachep
= arg
;
2268 struct array_cache
*ac
;
2269 int node
= numa_mem_id();
2270 struct kmem_cache_node
*n
;
2274 ac
= cpu_cache_get(cachep
);
2275 n
= get_node(cachep
, node
);
2276 spin_lock(&n
->list_lock
);
2277 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2278 spin_unlock(&n
->list_lock
);
2279 slabs_destroy(cachep
, &list
);
2283 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2285 struct kmem_cache_node
*n
;
2289 on_each_cpu(do_drain
, cachep
, 1);
2291 for_each_kmem_cache_node(cachep
, node
, n
)
2293 drain_alien_cache(cachep
, n
->alien
);
2295 for_each_kmem_cache_node(cachep
, node
, n
) {
2296 spin_lock_irq(&n
->list_lock
);
2297 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2298 spin_unlock_irq(&n
->list_lock
);
2300 slabs_destroy(cachep
, &list
);
2305 * Remove slabs from the list of free slabs.
2306 * Specify the number of slabs to drain in tofree.
2308 * Returns the actual number of slabs released.
2310 static int drain_freelist(struct kmem_cache
*cache
,
2311 struct kmem_cache_node
*n
, int tofree
)
2313 struct list_head
*p
;
2318 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2320 spin_lock_irq(&n
->list_lock
);
2321 p
= n
->slabs_free
.prev
;
2322 if (p
== &n
->slabs_free
) {
2323 spin_unlock_irq(&n
->list_lock
);
2327 page
= list_entry(p
, struct page
, lru
);
2328 list_del(&page
->lru
);
2330 * Safe to drop the lock. The slab is no longer linked
2333 n
->free_objects
-= cache
->num
;
2334 spin_unlock_irq(&n
->list_lock
);
2335 slab_destroy(cache
, page
);
2342 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2346 struct kmem_cache_node
*n
;
2348 drain_cpu_caches(cachep
);
2351 for_each_kmem_cache_node(cachep
, node
, n
) {
2352 drain_freelist(cachep
, n
, INT_MAX
);
2354 ret
+= !list_empty(&n
->slabs_full
) ||
2355 !list_empty(&n
->slabs_partial
);
2357 return (ret
? 1 : 0);
2360 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2362 return __kmem_cache_shrink(cachep
, false);
2365 void __kmem_cache_release(struct kmem_cache
*cachep
)
2368 struct kmem_cache_node
*n
;
2370 cache_random_seq_destroy(cachep
);
2372 free_percpu(cachep
->cpu_cache
);
2374 /* NUMA: free the node structures */
2375 for_each_kmem_cache_node(cachep
, i
, n
) {
2377 free_alien_cache(n
->alien
);
2379 cachep
->node
[i
] = NULL
;
2384 * Get the memory for a slab management obj.
2386 * For a slab cache when the slab descriptor is off-slab, the
2387 * slab descriptor can't come from the same cache which is being created,
2388 * Because if it is the case, that means we defer the creation of
2389 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2390 * And we eventually call down to __kmem_cache_create(), which
2391 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2392 * This is a "chicken-and-egg" problem.
2394 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2395 * which are all initialized during kmem_cache_init().
2397 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2398 struct page
*page
, int colour_off
,
2399 gfp_t local_flags
, int nodeid
)
2402 void *addr
= page_address(page
);
2404 page
->s_mem
= addr
+ colour_off
;
2407 if (OBJFREELIST_SLAB(cachep
))
2409 else if (OFF_SLAB(cachep
)) {
2410 /* Slab management obj is off-slab. */
2411 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2412 local_flags
, nodeid
);
2416 /* We will use last bytes at the slab for freelist */
2417 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2418 cachep
->freelist_size
;
2424 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2426 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2429 static inline void set_free_obj(struct page
*page
,
2430 unsigned int idx
, freelist_idx_t val
)
2432 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2435 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2440 for (i
= 0; i
< cachep
->num
; i
++) {
2441 void *objp
= index_to_obj(cachep
, page
, i
);
2443 if (cachep
->flags
& SLAB_STORE_USER
)
2444 *dbg_userword(cachep
, objp
) = NULL
;
2446 if (cachep
->flags
& SLAB_RED_ZONE
) {
2447 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2448 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2451 * Constructors are not allowed to allocate memory from the same
2452 * cache which they are a constructor for. Otherwise, deadlock.
2453 * They must also be threaded.
2455 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2456 kasan_unpoison_object_data(cachep
,
2457 objp
+ obj_offset(cachep
));
2458 cachep
->ctor(objp
+ obj_offset(cachep
));
2459 kasan_poison_object_data(
2460 cachep
, objp
+ obj_offset(cachep
));
2463 if (cachep
->flags
& SLAB_RED_ZONE
) {
2464 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2465 slab_error(cachep
, "constructor overwrote the end of an object");
2466 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2467 slab_error(cachep
, "constructor overwrote the start of an object");
2469 /* need to poison the objs? */
2470 if (cachep
->flags
& SLAB_POISON
) {
2471 poison_obj(cachep
, objp
, POISON_FREE
);
2472 slab_kernel_map(cachep
, objp
, 0, 0);
2478 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2479 /* Hold information during a freelist initialization */
2480 union freelist_init_state
{
2487 struct rnd_state rnd_state
;
2491 * Initialize the state based on the randomization methode available.
2492 * return true if the pre-computed list is available, false otherwize.
2494 static bool freelist_state_initialize(union freelist_init_state
*state
,
2495 struct kmem_cache
*cachep
,
2501 /* Use best entropy available to define a random shift */
2502 rand
= get_random_int();
2504 /* Use a random state if the pre-computed list is not available */
2505 if (!cachep
->random_seq
) {
2506 prandom_seed_state(&state
->rnd_state
, rand
);
2509 state
->list
= cachep
->random_seq
;
2510 state
->count
= count
;
2518 /* Get the next entry on the list and randomize it using a random shift */
2519 static freelist_idx_t
next_random_slot(union freelist_init_state
*state
)
2521 return (state
->list
[state
->pos
++] + state
->rand
) % state
->count
;
2524 /* Swap two freelist entries */
2525 static void swap_free_obj(struct page
*page
, unsigned int a
, unsigned int b
)
2527 swap(((freelist_idx_t
*)page
->freelist
)[a
],
2528 ((freelist_idx_t
*)page
->freelist
)[b
]);
2532 * Shuffle the freelist initialization state based on pre-computed lists.
2533 * return true if the list was successfully shuffled, false otherwise.
2535 static bool shuffle_freelist(struct kmem_cache
*cachep
, struct page
*page
)
2537 unsigned int objfreelist
= 0, i
, rand
, count
= cachep
->num
;
2538 union freelist_init_state state
;
2544 precomputed
= freelist_state_initialize(&state
, cachep
, count
);
2546 /* Take a random entry as the objfreelist */
2547 if (OBJFREELIST_SLAB(cachep
)) {
2549 objfreelist
= count
- 1;
2551 objfreelist
= next_random_slot(&state
);
2552 page
->freelist
= index_to_obj(cachep
, page
, objfreelist
) +
2558 * On early boot, generate the list dynamically.
2559 * Later use a pre-computed list for speed.
2562 for (i
= 0; i
< count
; i
++)
2563 set_free_obj(page
, i
, i
);
2565 /* Fisher-Yates shuffle */
2566 for (i
= count
- 1; i
> 0; i
--) {
2567 rand
= prandom_u32_state(&state
.rnd_state
);
2569 swap_free_obj(page
, i
, rand
);
2572 for (i
= 0; i
< count
; i
++)
2573 set_free_obj(page
, i
, next_random_slot(&state
));
2576 if (OBJFREELIST_SLAB(cachep
))
2577 set_free_obj(page
, cachep
->num
- 1, objfreelist
);
2582 static inline bool shuffle_freelist(struct kmem_cache
*cachep
,
2587 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2589 static void cache_init_objs(struct kmem_cache
*cachep
,
2596 cache_init_objs_debug(cachep
, page
);
2598 /* Try to randomize the freelist if enabled */
2599 shuffled
= shuffle_freelist(cachep
, page
);
2601 if (!shuffled
&& OBJFREELIST_SLAB(cachep
)) {
2602 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2606 for (i
= 0; i
< cachep
->num
; i
++) {
2607 /* constructor could break poison info */
2608 if (DEBUG
== 0 && cachep
->ctor
) {
2609 objp
= index_to_obj(cachep
, page
, i
);
2610 kasan_unpoison_object_data(cachep
, objp
);
2612 kasan_poison_object_data(cachep
, objp
);
2616 set_free_obj(page
, i
, i
);
2620 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2624 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2628 if (cachep
->flags
& SLAB_STORE_USER
)
2629 set_store_user_dirty(cachep
);
2635 static void slab_put_obj(struct kmem_cache
*cachep
,
2636 struct page
*page
, void *objp
)
2638 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2642 /* Verify double free bug */
2643 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2644 if (get_free_obj(page
, i
) == objnr
) {
2645 pr_err("slab: double free detected in cache '%s', objp %p\n",
2646 cachep
->name
, objp
);
2652 if (!page
->freelist
)
2653 page
->freelist
= objp
+ obj_offset(cachep
);
2655 set_free_obj(page
, page
->active
, objnr
);
2659 * Map pages beginning at addr to the given cache and slab. This is required
2660 * for the slab allocator to be able to lookup the cache and slab of a
2661 * virtual address for kfree, ksize, and slab debugging.
2663 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2666 page
->slab_cache
= cache
;
2667 page
->freelist
= freelist
;
2671 * Grow (by 1) the number of slabs within a cache. This is called by
2672 * kmem_cache_alloc() when there are no active objs left in a cache.
2674 static struct page
*cache_grow_begin(struct kmem_cache
*cachep
,
2675 gfp_t flags
, int nodeid
)
2681 struct kmem_cache_node
*n
;
2685 * Be lazy and only check for valid flags here, keeping it out of the
2686 * critical path in kmem_cache_alloc().
2688 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2689 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
2690 flags
&= ~GFP_SLAB_BUG_MASK
;
2691 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2692 invalid_mask
, &invalid_mask
, flags
, &flags
);
2695 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2698 if (gfpflags_allow_blocking(local_flags
))
2702 * Get mem for the objs. Attempt to allocate a physical page from
2705 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2709 page_node
= page_to_nid(page
);
2710 n
= get_node(cachep
, page_node
);
2712 /* Get colour for the slab, and cal the next value. */
2714 if (n
->colour_next
>= cachep
->colour
)
2717 offset
= n
->colour_next
;
2718 if (offset
>= cachep
->colour
)
2721 offset
*= cachep
->colour_off
;
2723 /* Get slab management. */
2724 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2725 local_flags
& ~GFP_CONSTRAINT_MASK
, page_node
);
2726 if (OFF_SLAB(cachep
) && !freelist
)
2729 slab_map_pages(cachep
, page
, freelist
);
2731 kasan_poison_slab(page
);
2732 cache_init_objs(cachep
, page
);
2734 if (gfpflags_allow_blocking(local_flags
))
2735 local_irq_disable();
2740 kmem_freepages(cachep
, page
);
2742 if (gfpflags_allow_blocking(local_flags
))
2743 local_irq_disable();
2747 static void cache_grow_end(struct kmem_cache
*cachep
, struct page
*page
)
2749 struct kmem_cache_node
*n
;
2757 INIT_LIST_HEAD(&page
->lru
);
2758 n
= get_node(cachep
, page_to_nid(page
));
2760 spin_lock(&n
->list_lock
);
2762 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2764 fixup_slab_list(cachep
, n
, page
, &list
);
2765 STATS_INC_GROWN(cachep
);
2766 n
->free_objects
+= cachep
->num
- page
->active
;
2767 spin_unlock(&n
->list_lock
);
2769 fixup_objfreelist_debug(cachep
, &list
);
2775 * Perform extra freeing checks:
2776 * - detect bad pointers.
2777 * - POISON/RED_ZONE checking
2779 static void kfree_debugcheck(const void *objp
)
2781 if (!virt_addr_valid(objp
)) {
2782 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2783 (unsigned long)objp
);
2788 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2790 unsigned long long redzone1
, redzone2
;
2792 redzone1
= *dbg_redzone1(cache
, obj
);
2793 redzone2
= *dbg_redzone2(cache
, obj
);
2798 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2801 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2802 slab_error(cache
, "double free detected");
2804 slab_error(cache
, "memory outside object was overwritten");
2806 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2807 obj
, redzone1
, redzone2
);
2810 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2811 unsigned long caller
)
2816 BUG_ON(virt_to_cache(objp
) != cachep
);
2818 objp
-= obj_offset(cachep
);
2819 kfree_debugcheck(objp
);
2820 page
= virt_to_head_page(objp
);
2822 if (cachep
->flags
& SLAB_RED_ZONE
) {
2823 verify_redzone_free(cachep
, objp
);
2824 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2825 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2827 if (cachep
->flags
& SLAB_STORE_USER
) {
2828 set_store_user_dirty(cachep
);
2829 *dbg_userword(cachep
, objp
) = (void *)caller
;
2832 objnr
= obj_to_index(cachep
, page
, objp
);
2834 BUG_ON(objnr
>= cachep
->num
);
2835 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2837 if (cachep
->flags
& SLAB_POISON
) {
2838 poison_obj(cachep
, objp
, POISON_FREE
);
2839 slab_kernel_map(cachep
, objp
, 0, caller
);
2845 #define kfree_debugcheck(x) do { } while(0)
2846 #define cache_free_debugcheck(x,objp,z) (objp)
2849 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2857 objp
= next
- obj_offset(cachep
);
2858 next
= *(void **)next
;
2859 poison_obj(cachep
, objp
, POISON_FREE
);
2864 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2865 struct kmem_cache_node
*n
, struct page
*page
,
2868 /* move slabp to correct slabp list: */
2869 list_del(&page
->lru
);
2870 if (page
->active
== cachep
->num
) {
2871 list_add(&page
->lru
, &n
->slabs_full
);
2872 if (OBJFREELIST_SLAB(cachep
)) {
2874 /* Poisoning will be done without holding the lock */
2875 if (cachep
->flags
& SLAB_POISON
) {
2876 void **objp
= page
->freelist
;
2882 page
->freelist
= NULL
;
2885 list_add(&page
->lru
, &n
->slabs_partial
);
2888 /* Try to find non-pfmemalloc slab if needed */
2889 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2890 struct page
*page
, bool pfmemalloc
)
2898 if (!PageSlabPfmemalloc(page
))
2901 /* No need to keep pfmemalloc slab if we have enough free objects */
2902 if (n
->free_objects
> n
->free_limit
) {
2903 ClearPageSlabPfmemalloc(page
);
2907 /* Move pfmemalloc slab to the end of list to speed up next search */
2908 list_del(&page
->lru
);
2910 list_add_tail(&page
->lru
, &n
->slabs_free
);
2912 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2914 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2915 if (!PageSlabPfmemalloc(page
))
2919 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2920 if (!PageSlabPfmemalloc(page
))
2927 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2931 page
= list_first_entry_or_null(&n
->slabs_partial
,
2934 n
->free_touched
= 1;
2935 page
= list_first_entry_or_null(&n
->slabs_free
,
2939 if (sk_memalloc_socks())
2940 return get_valid_first_slab(n
, page
, pfmemalloc
);
2945 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2946 struct kmem_cache_node
*n
, gfp_t flags
)
2952 if (!gfp_pfmemalloc_allowed(flags
))
2955 spin_lock(&n
->list_lock
);
2956 page
= get_first_slab(n
, true);
2958 spin_unlock(&n
->list_lock
);
2962 obj
= slab_get_obj(cachep
, page
);
2965 fixup_slab_list(cachep
, n
, page
, &list
);
2967 spin_unlock(&n
->list_lock
);
2968 fixup_objfreelist_debug(cachep
, &list
);
2974 * Slab list should be fixed up by fixup_slab_list() for existing slab
2975 * or cache_grow_end() for new slab
2977 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2978 struct array_cache
*ac
, struct page
*page
, int batchcount
)
2981 * There must be at least one object available for
2984 BUG_ON(page
->active
>= cachep
->num
);
2986 while (page
->active
< cachep
->num
&& batchcount
--) {
2987 STATS_INC_ALLOCED(cachep
);
2988 STATS_INC_ACTIVE(cachep
);
2989 STATS_SET_HIGH(cachep
);
2991 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2997 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
3000 struct kmem_cache_node
*n
;
3001 struct array_cache
*ac
, *shared
;
3007 node
= numa_mem_id();
3009 ac
= cpu_cache_get(cachep
);
3010 batchcount
= ac
->batchcount
;
3011 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3013 * If there was little recent activity on this cache, then
3014 * perform only a partial refill. Otherwise we could generate
3017 batchcount
= BATCHREFILL_LIMIT
;
3019 n
= get_node(cachep
, node
);
3021 BUG_ON(ac
->avail
> 0 || !n
);
3022 shared
= READ_ONCE(n
->shared
);
3023 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
3026 spin_lock(&n
->list_lock
);
3027 shared
= READ_ONCE(n
->shared
);
3029 /* See if we can refill from the shared array */
3030 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
3031 shared
->touched
= 1;
3035 while (batchcount
> 0) {
3036 /* Get slab alloc is to come from. */
3037 page
= get_first_slab(n
, false);
3041 check_spinlock_acquired(cachep
);
3043 batchcount
= alloc_block(cachep
, ac
, page
, batchcount
);
3044 fixup_slab_list(cachep
, n
, page
, &list
);
3048 n
->free_objects
-= ac
->avail
;
3050 spin_unlock(&n
->list_lock
);
3051 fixup_objfreelist_debug(cachep
, &list
);
3054 if (unlikely(!ac
->avail
)) {
3055 /* Check if we can use obj in pfmemalloc slab */
3056 if (sk_memalloc_socks()) {
3057 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
3063 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
3066 * cache_grow_begin() can reenable interrupts,
3067 * then ac could change.
3069 ac
= cpu_cache_get(cachep
);
3070 if (!ac
->avail
&& page
)
3071 alloc_block(cachep
, ac
, page
, batchcount
);
3072 cache_grow_end(cachep
, page
);
3079 return ac
->entry
[--ac
->avail
];
3082 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3085 might_sleep_if(gfpflags_allow_blocking(flags
));
3089 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3090 gfp_t flags
, void *objp
, unsigned long caller
)
3094 if (cachep
->flags
& SLAB_POISON
) {
3095 check_poison_obj(cachep
, objp
);
3096 slab_kernel_map(cachep
, objp
, 1, 0);
3097 poison_obj(cachep
, objp
, POISON_INUSE
);
3099 if (cachep
->flags
& SLAB_STORE_USER
)
3100 *dbg_userword(cachep
, objp
) = (void *)caller
;
3102 if (cachep
->flags
& SLAB_RED_ZONE
) {
3103 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3104 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3105 slab_error(cachep
, "double free, or memory outside object was overwritten");
3106 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3107 objp
, *dbg_redzone1(cachep
, objp
),
3108 *dbg_redzone2(cachep
, objp
));
3110 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3111 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3114 objp
+= obj_offset(cachep
);
3115 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3117 if (ARCH_SLAB_MINALIGN
&&
3118 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3119 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3120 objp
, (int)ARCH_SLAB_MINALIGN
);
3125 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3128 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3131 struct array_cache
*ac
;
3135 ac
= cpu_cache_get(cachep
);
3136 if (likely(ac
->avail
)) {
3138 objp
= ac
->entry
[--ac
->avail
];
3140 STATS_INC_ALLOCHIT(cachep
);
3144 STATS_INC_ALLOCMISS(cachep
);
3145 objp
= cache_alloc_refill(cachep
, flags
);
3147 * the 'ac' may be updated by cache_alloc_refill(),
3148 * and kmemleak_erase() requires its correct value.
3150 ac
= cpu_cache_get(cachep
);
3154 * To avoid a false negative, if an object that is in one of the
3155 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3156 * treat the array pointers as a reference to the object.
3159 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3165 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3167 * If we are in_interrupt, then process context, including cpusets and
3168 * mempolicy, may not apply and should not be used for allocation policy.
3170 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3172 int nid_alloc
, nid_here
;
3174 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3176 nid_alloc
= nid_here
= numa_mem_id();
3177 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3178 nid_alloc
= cpuset_slab_spread_node();
3179 else if (current
->mempolicy
)
3180 nid_alloc
= mempolicy_slab_node();
3181 if (nid_alloc
!= nid_here
)
3182 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3187 * Fallback function if there was no memory available and no objects on a
3188 * certain node and fall back is permitted. First we scan all the
3189 * available node for available objects. If that fails then we
3190 * perform an allocation without specifying a node. This allows the page
3191 * allocator to do its reclaim / fallback magic. We then insert the
3192 * slab into the proper nodelist and then allocate from it.
3194 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3196 struct zonelist
*zonelist
;
3199 enum zone_type high_zoneidx
= gfp_zone(flags
);
3203 unsigned int cpuset_mems_cookie
;
3205 if (flags
& __GFP_THISNODE
)
3209 cpuset_mems_cookie
= read_mems_allowed_begin();
3210 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3214 * Look through allowed nodes for objects available
3215 * from existing per node queues.
3217 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3218 nid
= zone_to_nid(zone
);
3220 if (cpuset_zone_allowed(zone
, flags
) &&
3221 get_node(cache
, nid
) &&
3222 get_node(cache
, nid
)->free_objects
) {
3223 obj
= ____cache_alloc_node(cache
,
3224 gfp_exact_node(flags
), nid
);
3232 * This allocation will be performed within the constraints
3233 * of the current cpuset / memory policy requirements.
3234 * We may trigger various forms of reclaim on the allowed
3235 * set and go into memory reserves if necessary.
3237 page
= cache_grow_begin(cache
, flags
, numa_mem_id());
3238 cache_grow_end(cache
, page
);
3240 nid
= page_to_nid(page
);
3241 obj
= ____cache_alloc_node(cache
,
3242 gfp_exact_node(flags
), nid
);
3245 * Another processor may allocate the objects in
3246 * the slab since we are not holding any locks.
3253 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3259 * A interface to enable slab creation on nodeid
3261 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3265 struct kmem_cache_node
*n
;
3269 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3270 n
= get_node(cachep
, nodeid
);
3274 spin_lock(&n
->list_lock
);
3275 page
= get_first_slab(n
, false);
3279 check_spinlock_acquired_node(cachep
, nodeid
);
3281 STATS_INC_NODEALLOCS(cachep
);
3282 STATS_INC_ACTIVE(cachep
);
3283 STATS_SET_HIGH(cachep
);
3285 BUG_ON(page
->active
== cachep
->num
);
3287 obj
= slab_get_obj(cachep
, page
);
3290 fixup_slab_list(cachep
, n
, page
, &list
);
3292 spin_unlock(&n
->list_lock
);
3293 fixup_objfreelist_debug(cachep
, &list
);
3297 spin_unlock(&n
->list_lock
);
3298 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3300 /* This slab isn't counted yet so don't update free_objects */
3301 obj
= slab_get_obj(cachep
, page
);
3303 cache_grow_end(cachep
, page
);
3305 return obj
? obj
: fallback_alloc(cachep
, flags
);
3308 static __always_inline
void *
3309 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3310 unsigned long caller
)
3312 unsigned long save_flags
;
3314 int slab_node
= numa_mem_id();
3316 flags
&= gfp_allowed_mask
;
3317 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3318 if (unlikely(!cachep
))
3321 cache_alloc_debugcheck_before(cachep
, flags
);
3322 local_irq_save(save_flags
);
3324 if (nodeid
== NUMA_NO_NODE
)
3327 if (unlikely(!get_node(cachep
, nodeid
))) {
3328 /* Node not bootstrapped yet */
3329 ptr
= fallback_alloc(cachep
, flags
);
3333 if (nodeid
== slab_node
) {
3335 * Use the locally cached objects if possible.
3336 * However ____cache_alloc does not allow fallback
3337 * to other nodes. It may fail while we still have
3338 * objects on other nodes available.
3340 ptr
= ____cache_alloc(cachep
, flags
);
3344 /* ___cache_alloc_node can fall back to other nodes */
3345 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3347 local_irq_restore(save_flags
);
3348 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3350 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3351 memset(ptr
, 0, cachep
->object_size
);
3353 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3357 static __always_inline
void *
3358 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3362 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3363 objp
= alternate_node_alloc(cache
, flags
);
3367 objp
= ____cache_alloc(cache
, flags
);
3370 * We may just have run out of memory on the local node.
3371 * ____cache_alloc_node() knows how to locate memory on other nodes
3374 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3381 static __always_inline
void *
3382 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3384 return ____cache_alloc(cachep
, flags
);
3387 #endif /* CONFIG_NUMA */
3389 static __always_inline
void *
3390 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3392 unsigned long save_flags
;
3395 flags
&= gfp_allowed_mask
;
3396 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3397 if (unlikely(!cachep
))
3400 cache_alloc_debugcheck_before(cachep
, flags
);
3401 local_irq_save(save_flags
);
3402 objp
= __do_cache_alloc(cachep
, flags
);
3403 local_irq_restore(save_flags
);
3404 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3407 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3408 memset(objp
, 0, cachep
->object_size
);
3410 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3415 * Caller needs to acquire correct kmem_cache_node's list_lock
3416 * @list: List of detached free slabs should be freed by caller
3418 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3419 int nr_objects
, int node
, struct list_head
*list
)
3422 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3425 n
->free_objects
+= nr_objects
;
3427 for (i
= 0; i
< nr_objects
; i
++) {
3433 page
= virt_to_head_page(objp
);
3434 list_del(&page
->lru
);
3435 check_spinlock_acquired_node(cachep
, node
);
3436 slab_put_obj(cachep
, page
, objp
);
3437 STATS_DEC_ACTIVE(cachep
);
3439 /* fixup slab chains */
3440 if (page
->active
== 0)
3441 list_add(&page
->lru
, &n
->slabs_free
);
3443 /* Unconditionally move a slab to the end of the
3444 * partial list on free - maximum time for the
3445 * other objects to be freed, too.
3447 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3451 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3452 n
->free_objects
-= cachep
->num
;
3454 page
= list_last_entry(&n
->slabs_free
, struct page
, lru
);
3455 list_move(&page
->lru
, list
);
3459 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3462 struct kmem_cache_node
*n
;
3463 int node
= numa_mem_id();
3466 batchcount
= ac
->batchcount
;
3469 n
= get_node(cachep
, node
);
3470 spin_lock(&n
->list_lock
);
3472 struct array_cache
*shared_array
= n
->shared
;
3473 int max
= shared_array
->limit
- shared_array
->avail
;
3475 if (batchcount
> max
)
3477 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3478 ac
->entry
, sizeof(void *) * batchcount
);
3479 shared_array
->avail
+= batchcount
;
3484 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3491 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3492 BUG_ON(page
->active
);
3496 STATS_SET_FREEABLE(cachep
, i
);
3499 spin_unlock(&n
->list_lock
);
3500 slabs_destroy(cachep
, &list
);
3501 ac
->avail
-= batchcount
;
3502 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3506 * Release an obj back to its cache. If the obj has a constructed state, it must
3507 * be in this state _before_ it is released. Called with disabled ints.
3509 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3510 unsigned long caller
)
3512 /* Put the object into the quarantine, don't touch it for now. */
3513 if (kasan_slab_free(cachep
, objp
))
3516 ___cache_free(cachep
, objp
, caller
);
3519 void ___cache_free(struct kmem_cache
*cachep
, void *objp
,
3520 unsigned long caller
)
3522 struct array_cache
*ac
= cpu_cache_get(cachep
);
3525 kmemleak_free_recursive(objp
, cachep
->flags
);
3526 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3528 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3531 * Skip calling cache_free_alien() when the platform is not numa.
3532 * This will avoid cache misses that happen while accessing slabp (which
3533 * is per page memory reference) to get nodeid. Instead use a global
3534 * variable to skip the call, which is mostly likely to be present in
3537 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3540 if (ac
->avail
< ac
->limit
) {
3541 STATS_INC_FREEHIT(cachep
);
3543 STATS_INC_FREEMISS(cachep
);
3544 cache_flusharray(cachep
, ac
);
3547 if (sk_memalloc_socks()) {
3548 struct page
*page
= virt_to_head_page(objp
);
3550 if (unlikely(PageSlabPfmemalloc(page
))) {
3551 cache_free_pfmemalloc(cachep
, page
, objp
);
3556 ac
->entry
[ac
->avail
++] = objp
;
3560 * kmem_cache_alloc - Allocate an object
3561 * @cachep: The cache to allocate from.
3562 * @flags: See kmalloc().
3564 * Allocate an object from this cache. The flags are only relevant
3565 * if the cache has no available objects.
3567 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3569 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3571 kasan_slab_alloc(cachep
, ret
, flags
);
3572 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3573 cachep
->object_size
, cachep
->size
, flags
);
3577 EXPORT_SYMBOL(kmem_cache_alloc
);
3579 static __always_inline
void
3580 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3581 size_t size
, void **p
, unsigned long caller
)
3585 for (i
= 0; i
< size
; i
++)
3586 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3589 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3594 s
= slab_pre_alloc_hook(s
, flags
);
3598 cache_alloc_debugcheck_before(s
, flags
);
3600 local_irq_disable();
3601 for (i
= 0; i
< size
; i
++) {
3602 void *objp
= __do_cache_alloc(s
, flags
);
3604 if (unlikely(!objp
))
3610 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3612 /* Clear memory outside IRQ disabled section */
3613 if (unlikely(flags
& __GFP_ZERO
))
3614 for (i
= 0; i
< size
; i
++)
3615 memset(p
[i
], 0, s
->object_size
);
3617 slab_post_alloc_hook(s
, flags
, size
, p
);
3618 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3622 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3623 slab_post_alloc_hook(s
, flags
, i
, p
);
3624 __kmem_cache_free_bulk(s
, i
, p
);
3627 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3629 #ifdef CONFIG_TRACING
3631 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3635 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3637 kasan_kmalloc(cachep
, ret
, size
, flags
);
3638 trace_kmalloc(_RET_IP_
, ret
,
3639 size
, cachep
->size
, flags
);
3642 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3647 * kmem_cache_alloc_node - Allocate an object on the specified node
3648 * @cachep: The cache to allocate from.
3649 * @flags: See kmalloc().
3650 * @nodeid: node number of the target node.
3652 * Identical to kmem_cache_alloc but it will allocate memory on the given
3653 * node, which can improve the performance for cpu bound structures.
3655 * Fallback to other node is possible if __GFP_THISNODE is not set.
3657 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3659 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3661 kasan_slab_alloc(cachep
, ret
, flags
);
3662 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3663 cachep
->object_size
, cachep
->size
,
3668 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3670 #ifdef CONFIG_TRACING
3671 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3678 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3680 kasan_kmalloc(cachep
, ret
, size
, flags
);
3681 trace_kmalloc_node(_RET_IP_
, ret
,
3686 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3689 static __always_inline
void *
3690 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3692 struct kmem_cache
*cachep
;
3695 cachep
= kmalloc_slab(size
, flags
);
3696 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3698 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3699 kasan_kmalloc(cachep
, ret
, size
, flags
);
3704 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3706 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3708 EXPORT_SYMBOL(__kmalloc_node
);
3710 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3711 int node
, unsigned long caller
)
3713 return __do_kmalloc_node(size
, flags
, node
, caller
);
3715 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3716 #endif /* CONFIG_NUMA */
3719 * __do_kmalloc - allocate memory
3720 * @size: how many bytes of memory are required.
3721 * @flags: the type of memory to allocate (see kmalloc).
3722 * @caller: function caller for debug tracking of the caller
3724 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3725 unsigned long caller
)
3727 struct kmem_cache
*cachep
;
3730 cachep
= kmalloc_slab(size
, flags
);
3731 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3733 ret
= slab_alloc(cachep
, flags
, caller
);
3735 kasan_kmalloc(cachep
, ret
, size
, flags
);
3736 trace_kmalloc(caller
, ret
,
3737 size
, cachep
->size
, flags
);
3742 void *__kmalloc(size_t size
, gfp_t flags
)
3744 return __do_kmalloc(size
, flags
, _RET_IP_
);
3746 EXPORT_SYMBOL(__kmalloc
);
3748 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3750 return __do_kmalloc(size
, flags
, caller
);
3752 EXPORT_SYMBOL(__kmalloc_track_caller
);
3755 * kmem_cache_free - Deallocate an object
3756 * @cachep: The cache the allocation was from.
3757 * @objp: The previously allocated object.
3759 * Free an object which was previously allocated from this
3762 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3764 unsigned long flags
;
3765 cachep
= cache_from_obj(cachep
, objp
);
3769 local_irq_save(flags
);
3770 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3771 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3772 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3773 __cache_free(cachep
, objp
, _RET_IP_
);
3774 local_irq_restore(flags
);
3776 trace_kmem_cache_free(_RET_IP_
, objp
);
3778 EXPORT_SYMBOL(kmem_cache_free
);
3780 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3782 struct kmem_cache
*s
;
3785 local_irq_disable();
3786 for (i
= 0; i
< size
; i
++) {
3789 if (!orig_s
) /* called via kfree_bulk */
3790 s
= virt_to_cache(objp
);
3792 s
= cache_from_obj(orig_s
, objp
);
3794 debug_check_no_locks_freed(objp
, s
->object_size
);
3795 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3796 debug_check_no_obj_freed(objp
, s
->object_size
);
3798 __cache_free(s
, objp
, _RET_IP_
);
3802 /* FIXME: add tracing */
3804 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3807 * kfree - free previously allocated memory
3808 * @objp: pointer returned by kmalloc.
3810 * If @objp is NULL, no operation is performed.
3812 * Don't free memory not originally allocated by kmalloc()
3813 * or you will run into trouble.
3815 void kfree(const void *objp
)
3817 struct kmem_cache
*c
;
3818 unsigned long flags
;
3820 trace_kfree(_RET_IP_
, objp
);
3822 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3824 local_irq_save(flags
);
3825 kfree_debugcheck(objp
);
3826 c
= virt_to_cache(objp
);
3827 debug_check_no_locks_freed(objp
, c
->object_size
);
3829 debug_check_no_obj_freed(objp
, c
->object_size
);
3830 __cache_free(c
, (void *)objp
, _RET_IP_
);
3831 local_irq_restore(flags
);
3833 EXPORT_SYMBOL(kfree
);
3836 * This initializes kmem_cache_node or resizes various caches for all nodes.
3838 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3842 struct kmem_cache_node
*n
;
3844 for_each_online_node(node
) {
3845 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3854 if (!cachep
->list
.next
) {
3855 /* Cache is not active yet. Roll back what we did */
3858 n
= get_node(cachep
, node
);
3861 free_alien_cache(n
->alien
);
3863 cachep
->node
[node
] = NULL
;
3871 /* Always called with the slab_mutex held */
3872 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3873 int batchcount
, int shared
, gfp_t gfp
)
3875 struct array_cache __percpu
*cpu_cache
, *prev
;
3878 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3882 prev
= cachep
->cpu_cache
;
3883 cachep
->cpu_cache
= cpu_cache
;
3884 kick_all_cpus_sync();
3887 cachep
->batchcount
= batchcount
;
3888 cachep
->limit
= limit
;
3889 cachep
->shared
= shared
;
3894 for_each_online_cpu(cpu
) {
3897 struct kmem_cache_node
*n
;
3898 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3900 node
= cpu_to_mem(cpu
);
3901 n
= get_node(cachep
, node
);
3902 spin_lock_irq(&n
->list_lock
);
3903 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3904 spin_unlock_irq(&n
->list_lock
);
3905 slabs_destroy(cachep
, &list
);
3910 return setup_kmem_cache_nodes(cachep
, gfp
);
3913 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3914 int batchcount
, int shared
, gfp_t gfp
)
3917 struct kmem_cache
*c
;
3919 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3921 if (slab_state
< FULL
)
3924 if ((ret
< 0) || !is_root_cache(cachep
))
3927 lockdep_assert_held(&slab_mutex
);
3928 for_each_memcg_cache(c
, cachep
) {
3929 /* return value determined by the root cache only */
3930 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3936 /* Called with slab_mutex held always */
3937 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3944 err
= cache_random_seq_create(cachep
, cachep
->num
, gfp
);
3948 if (!is_root_cache(cachep
)) {
3949 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3950 limit
= root
->limit
;
3951 shared
= root
->shared
;
3952 batchcount
= root
->batchcount
;
3955 if (limit
&& shared
&& batchcount
)
3958 * The head array serves three purposes:
3959 * - create a LIFO ordering, i.e. return objects that are cache-warm
3960 * - reduce the number of spinlock operations.
3961 * - reduce the number of linked list operations on the slab and
3962 * bufctl chains: array operations are cheaper.
3963 * The numbers are guessed, we should auto-tune as described by
3966 if (cachep
->size
> 131072)
3968 else if (cachep
->size
> PAGE_SIZE
)
3970 else if (cachep
->size
> 1024)
3972 else if (cachep
->size
> 256)
3978 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3979 * allocation behaviour: Most allocs on one cpu, most free operations
3980 * on another cpu. For these cases, an efficient object passing between
3981 * cpus is necessary. This is provided by a shared array. The array
3982 * replaces Bonwick's magazine layer.
3983 * On uniprocessor, it's functionally equivalent (but less efficient)
3984 * to a larger limit. Thus disabled by default.
3987 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3992 * With debugging enabled, large batchcount lead to excessively long
3993 * periods with disabled local interrupts. Limit the batchcount
3998 batchcount
= (limit
+ 1) / 2;
4000 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
4003 pr_err("enable_cpucache failed for %s, error %d\n",
4004 cachep
->name
, -err
);
4009 * Drain an array if it contains any elements taking the node lock only if
4010 * necessary. Note that the node listlock also protects the array_cache
4011 * if drain_array() is used on the shared array.
4013 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
4014 struct array_cache
*ac
, int node
)
4018 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4019 check_mutex_acquired();
4021 if (!ac
|| !ac
->avail
)
4029 spin_lock_irq(&n
->list_lock
);
4030 drain_array_locked(cachep
, ac
, node
, false, &list
);
4031 spin_unlock_irq(&n
->list_lock
);
4033 slabs_destroy(cachep
, &list
);
4037 * cache_reap - Reclaim memory from caches.
4038 * @w: work descriptor
4040 * Called from workqueue/eventd every few seconds.
4042 * - clear the per-cpu caches for this CPU.
4043 * - return freeable pages to the main free memory pool.
4045 * If we cannot acquire the cache chain mutex then just give up - we'll try
4046 * again on the next iteration.
4048 static void cache_reap(struct work_struct
*w
)
4050 struct kmem_cache
*searchp
;
4051 struct kmem_cache_node
*n
;
4052 int node
= numa_mem_id();
4053 struct delayed_work
*work
= to_delayed_work(w
);
4055 if (!mutex_trylock(&slab_mutex
))
4056 /* Give up. Setup the next iteration. */
4059 list_for_each_entry(searchp
, &slab_caches
, list
) {
4063 * We only take the node lock if absolutely necessary and we
4064 * have established with reasonable certainty that
4065 * we can do some work if the lock was obtained.
4067 n
= get_node(searchp
, node
);
4069 reap_alien(searchp
, n
);
4071 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
4074 * These are racy checks but it does not matter
4075 * if we skip one check or scan twice.
4077 if (time_after(n
->next_reap
, jiffies
))
4080 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
4082 drain_array(searchp
, n
, n
->shared
, node
);
4084 if (n
->free_touched
)
4085 n
->free_touched
= 0;
4089 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4090 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4091 STATS_ADD_REAPED(searchp
, freed
);
4097 mutex_unlock(&slab_mutex
);
4100 /* Set up the next iteration */
4101 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
4104 #ifdef CONFIG_SLABINFO
4105 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4108 unsigned long active_objs
;
4109 unsigned long num_objs
;
4110 unsigned long active_slabs
= 0;
4111 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4115 struct kmem_cache_node
*n
;
4119 for_each_kmem_cache_node(cachep
, node
, n
) {
4122 spin_lock_irq(&n
->list_lock
);
4124 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4125 if (page
->active
!= cachep
->num
&& !error
)
4126 error
= "slabs_full accounting error";
4127 active_objs
+= cachep
->num
;
4130 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4131 if (page
->active
== cachep
->num
&& !error
)
4132 error
= "slabs_partial accounting error";
4133 if (!page
->active
&& !error
)
4134 error
= "slabs_partial accounting error";
4135 active_objs
+= page
->active
;
4138 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4139 if (page
->active
&& !error
)
4140 error
= "slabs_free accounting error";
4143 free_objects
+= n
->free_objects
;
4145 shared_avail
+= n
->shared
->avail
;
4147 spin_unlock_irq(&n
->list_lock
);
4149 num_slabs
+= active_slabs
;
4150 num_objs
= num_slabs
* cachep
->num
;
4151 if (num_objs
- active_objs
!= free_objects
&& !error
)
4152 error
= "free_objects accounting error";
4154 name
= cachep
->name
;
4156 pr_err("slab: cache %s error: %s\n", name
, error
);
4158 sinfo
->active_objs
= active_objs
;
4159 sinfo
->num_objs
= num_objs
;
4160 sinfo
->active_slabs
= active_slabs
;
4161 sinfo
->num_slabs
= num_slabs
;
4162 sinfo
->shared_avail
= shared_avail
;
4163 sinfo
->limit
= cachep
->limit
;
4164 sinfo
->batchcount
= cachep
->batchcount
;
4165 sinfo
->shared
= cachep
->shared
;
4166 sinfo
->objects_per_slab
= cachep
->num
;
4167 sinfo
->cache_order
= cachep
->gfporder
;
4170 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4174 unsigned long high
= cachep
->high_mark
;
4175 unsigned long allocs
= cachep
->num_allocations
;
4176 unsigned long grown
= cachep
->grown
;
4177 unsigned long reaped
= cachep
->reaped
;
4178 unsigned long errors
= cachep
->errors
;
4179 unsigned long max_freeable
= cachep
->max_freeable
;
4180 unsigned long node_allocs
= cachep
->node_allocs
;
4181 unsigned long node_frees
= cachep
->node_frees
;
4182 unsigned long overflows
= cachep
->node_overflow
;
4184 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4185 allocs
, high
, grown
,
4186 reaped
, errors
, max_freeable
, node_allocs
,
4187 node_frees
, overflows
);
4191 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4192 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4193 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4194 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4196 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4197 allochit
, allocmiss
, freehit
, freemiss
);
4202 #define MAX_SLABINFO_WRITE 128
4204 * slabinfo_write - Tuning for the slab allocator
4206 * @buffer: user buffer
4207 * @count: data length
4210 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4211 size_t count
, loff_t
*ppos
)
4213 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4214 int limit
, batchcount
, shared
, res
;
4215 struct kmem_cache
*cachep
;
4217 if (count
> MAX_SLABINFO_WRITE
)
4219 if (copy_from_user(&kbuf
, buffer
, count
))
4221 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4223 tmp
= strchr(kbuf
, ' ');
4228 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4231 /* Find the cache in the chain of caches. */
4232 mutex_lock(&slab_mutex
);
4234 list_for_each_entry(cachep
, &slab_caches
, list
) {
4235 if (!strcmp(cachep
->name
, kbuf
)) {
4236 if (limit
< 1 || batchcount
< 1 ||
4237 batchcount
> limit
|| shared
< 0) {
4240 res
= do_tune_cpucache(cachep
, limit
,
4247 mutex_unlock(&slab_mutex
);
4253 #ifdef CONFIG_DEBUG_SLAB_LEAK
4255 static inline int add_caller(unsigned long *n
, unsigned long v
)
4265 unsigned long *q
= p
+ 2 * i
;
4279 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4285 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4294 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4297 for (j
= page
->active
; j
< c
->num
; j
++) {
4298 if (get_free_obj(page
, j
) == i
) {
4308 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4309 * mapping is established when actual object allocation and
4310 * we could mistakenly access the unmapped object in the cpu
4313 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4316 if (!add_caller(n
, v
))
4321 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4323 #ifdef CONFIG_KALLSYMS
4324 unsigned long offset
, size
;
4325 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4327 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4328 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4330 seq_printf(m
, " [%s]", modname
);
4334 seq_printf(m
, "%p", (void *)address
);
4337 static int leaks_show(struct seq_file
*m
, void *p
)
4339 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4341 struct kmem_cache_node
*n
;
4343 unsigned long *x
= m
->private;
4347 if (!(cachep
->flags
& SLAB_STORE_USER
))
4349 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4353 * Set store_user_clean and start to grab stored user information
4354 * for all objects on this cache. If some alloc/free requests comes
4355 * during the processing, information would be wrong so restart
4359 set_store_user_clean(cachep
);
4360 drain_cpu_caches(cachep
);
4364 for_each_kmem_cache_node(cachep
, node
, n
) {
4367 spin_lock_irq(&n
->list_lock
);
4369 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4370 handle_slab(x
, cachep
, page
);
4371 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4372 handle_slab(x
, cachep
, page
);
4373 spin_unlock_irq(&n
->list_lock
);
4375 } while (!is_store_user_clean(cachep
));
4377 name
= cachep
->name
;
4379 /* Increase the buffer size */
4380 mutex_unlock(&slab_mutex
);
4381 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4383 /* Too bad, we are really out */
4385 mutex_lock(&slab_mutex
);
4388 *(unsigned long *)m
->private = x
[0] * 2;
4390 mutex_lock(&slab_mutex
);
4391 /* Now make sure this entry will be retried */
4395 for (i
= 0; i
< x
[1]; i
++) {
4396 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4397 show_symbol(m
, x
[2*i
+2]);
4404 static const struct seq_operations slabstats_op
= {
4405 .start
= slab_start
,
4411 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4415 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4419 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4424 static const struct file_operations proc_slabstats_operations
= {
4425 .open
= slabstats_open
,
4427 .llseek
= seq_lseek
,
4428 .release
= seq_release_private
,
4432 static int __init
slab_proc_init(void)
4434 #ifdef CONFIG_DEBUG_SLAB_LEAK
4435 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4439 module_init(slab_proc_init
);
4443 * ksize - get the actual amount of memory allocated for a given object
4444 * @objp: Pointer to the object
4446 * kmalloc may internally round up allocations and return more memory
4447 * than requested. ksize() can be used to determine the actual amount of
4448 * memory allocated. The caller may use this additional memory, even though
4449 * a smaller amount of memory was initially specified with the kmalloc call.
4450 * The caller must guarantee that objp points to a valid object previously
4451 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4452 * must not be freed during the duration of the call.
4454 size_t ksize(const void *objp
)
4459 if (unlikely(objp
== ZERO_SIZE_PTR
))
4462 size
= virt_to_cache(objp
)->object_size
;
4463 /* We assume that ksize callers could use the whole allocated area,
4464 * so we need to unpoison this area.
4466 kasan_unpoison_shadow(objp
, size
);
4470 EXPORT_SYMBOL(ksize
);