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
)
527 node
= next_node(cpu_to_mem(cpu
), node_online_map
);
528 if (node
== MAX_NUMNODES
)
529 node
= first_node(node_online_map
);
531 per_cpu(slab_reap_node
, cpu
) = node
;
534 static void next_reap_node(void)
536 int node
= __this_cpu_read(slab_reap_node
);
538 node
= next_node(node
, node_online_map
);
539 if (unlikely(node
>= MAX_NUMNODES
))
540 node
= first_node(node_online_map
);
541 __this_cpu_write(slab_reap_node
, node
);
545 #define init_reap_node(cpu) do { } while (0)
546 #define next_reap_node(void) do { } while (0)
550 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
551 * via the workqueue/eventd.
552 * Add the CPU number into the expiration time to minimize the possibility of
553 * the CPUs getting into lockstep and contending for the global cache chain
556 static void start_cpu_timer(int cpu
)
558 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
561 * When this gets called from do_initcalls via cpucache_init(),
562 * init_workqueues() has already run, so keventd will be setup
565 if (keventd_up() && reap_work
->work
.func
== NULL
) {
567 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
568 schedule_delayed_work_on(cpu
, reap_work
,
569 __round_jiffies_relative(HZ
, cpu
));
573 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
576 * The array_cache structures contain pointers to free object.
577 * However, when such objects are allocated or transferred to another
578 * cache the pointers are not cleared and they could be counted as
579 * valid references during a kmemleak scan. Therefore, kmemleak must
580 * not scan such objects.
582 kmemleak_no_scan(ac
);
586 ac
->batchcount
= batch
;
591 static struct array_cache
*alloc_arraycache(int node
, int entries
,
592 int batchcount
, gfp_t gfp
)
594 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
595 struct array_cache
*ac
= NULL
;
597 ac
= kmalloc_node(memsize
, gfp
, node
);
598 init_arraycache(ac
, entries
, batchcount
);
602 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
603 struct page
*page
, void *objp
)
605 struct kmem_cache_node
*n
;
609 page_node
= page_to_nid(page
);
610 n
= get_node(cachep
, page_node
);
612 spin_lock(&n
->list_lock
);
613 free_block(cachep
, &objp
, 1, page_node
, &list
);
614 spin_unlock(&n
->list_lock
);
616 slabs_destroy(cachep
, &list
);
620 * Transfer objects in one arraycache to another.
621 * Locking must be handled by the caller.
623 * Return the number of entries transferred.
625 static int transfer_objects(struct array_cache
*to
,
626 struct array_cache
*from
, unsigned int max
)
628 /* Figure out how many entries to transfer */
629 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
634 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
644 #define drain_alien_cache(cachep, alien) do { } while (0)
645 #define reap_alien(cachep, n) do { } while (0)
647 static inline struct alien_cache
**alloc_alien_cache(int node
,
648 int limit
, gfp_t gfp
)
653 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
657 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
662 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
668 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
669 gfp_t flags
, int nodeid
)
674 static inline gfp_t
gfp_exact_node(gfp_t flags
)
676 return flags
& ~__GFP_NOFAIL
;
679 #else /* CONFIG_NUMA */
681 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
682 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
684 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
685 int batch
, gfp_t gfp
)
687 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
688 struct alien_cache
*alc
= NULL
;
690 alc
= kmalloc_node(memsize
, gfp
, node
);
691 init_arraycache(&alc
->ac
, entries
, batch
);
692 spin_lock_init(&alc
->lock
);
696 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
698 struct alien_cache
**alc_ptr
;
699 size_t memsize
= sizeof(void *) * nr_node_ids
;
704 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
709 if (i
== node
|| !node_online(i
))
711 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
713 for (i
--; i
>= 0; i
--)
722 static void free_alien_cache(struct alien_cache
**alc_ptr
)
733 static void __drain_alien_cache(struct kmem_cache
*cachep
,
734 struct array_cache
*ac
, int node
,
735 struct list_head
*list
)
737 struct kmem_cache_node
*n
= get_node(cachep
, node
);
740 spin_lock(&n
->list_lock
);
742 * Stuff objects into the remote nodes shared array first.
743 * That way we could avoid the overhead of putting the objects
744 * into the free lists and getting them back later.
747 transfer_objects(n
->shared
, ac
, ac
->limit
);
749 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
751 spin_unlock(&n
->list_lock
);
756 * Called from cache_reap() to regularly drain alien caches round robin.
758 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
760 int node
= __this_cpu_read(slab_reap_node
);
763 struct alien_cache
*alc
= n
->alien
[node
];
764 struct array_cache
*ac
;
768 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
771 __drain_alien_cache(cachep
, ac
, node
, &list
);
772 spin_unlock_irq(&alc
->lock
);
773 slabs_destroy(cachep
, &list
);
779 static void drain_alien_cache(struct kmem_cache
*cachep
,
780 struct alien_cache
**alien
)
783 struct alien_cache
*alc
;
784 struct array_cache
*ac
;
787 for_each_online_node(i
) {
793 spin_lock_irqsave(&alc
->lock
, flags
);
794 __drain_alien_cache(cachep
, ac
, i
, &list
);
795 spin_unlock_irqrestore(&alc
->lock
, flags
);
796 slabs_destroy(cachep
, &list
);
801 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
802 int node
, int page_node
)
804 struct kmem_cache_node
*n
;
805 struct alien_cache
*alien
= NULL
;
806 struct array_cache
*ac
;
809 n
= get_node(cachep
, node
);
810 STATS_INC_NODEFREES(cachep
);
811 if (n
->alien
&& n
->alien
[page_node
]) {
812 alien
= n
->alien
[page_node
];
814 spin_lock(&alien
->lock
);
815 if (unlikely(ac
->avail
== ac
->limit
)) {
816 STATS_INC_ACOVERFLOW(cachep
);
817 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
819 ac
->entry
[ac
->avail
++] = objp
;
820 spin_unlock(&alien
->lock
);
821 slabs_destroy(cachep
, &list
);
823 n
= get_node(cachep
, page_node
);
824 spin_lock(&n
->list_lock
);
825 free_block(cachep
, &objp
, 1, page_node
, &list
);
826 spin_unlock(&n
->list_lock
);
827 slabs_destroy(cachep
, &list
);
832 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
834 int page_node
= page_to_nid(virt_to_page(objp
));
835 int node
= numa_mem_id();
837 * Make sure we are not freeing a object from another node to the array
840 if (likely(node
== page_node
))
843 return __cache_free_alien(cachep
, objp
, node
, page_node
);
847 * Construct gfp mask to allocate from a specific node but do not reclaim or
848 * warn about failures.
850 static inline gfp_t
gfp_exact_node(gfp_t flags
)
852 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
856 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
858 struct kmem_cache_node
*n
;
861 * Set up the kmem_cache_node for cpu before we can
862 * begin anything. Make sure some other cpu on this
863 * node has not already allocated this
865 n
= get_node(cachep
, node
);
867 spin_lock_irq(&n
->list_lock
);
868 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
870 spin_unlock_irq(&n
->list_lock
);
875 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
879 kmem_cache_node_init(n
);
880 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
881 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
884 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
887 * The kmem_cache_nodes don't come and go as CPUs
888 * come and go. slab_mutex is sufficient
891 cachep
->node
[node
] = n
;
897 * Allocates and initializes node for a node on each slab cache, used for
898 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
899 * will be allocated off-node since memory is not yet online for the new node.
900 * When hotplugging memory or a cpu, existing node are not replaced if
903 * Must hold slab_mutex.
905 static int init_cache_node_node(int node
)
908 struct kmem_cache
*cachep
;
910 list_for_each_entry(cachep
, &slab_caches
, list
) {
911 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
919 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
920 int node
, gfp_t gfp
, bool force_change
)
923 struct kmem_cache_node
*n
;
924 struct array_cache
*old_shared
= NULL
;
925 struct array_cache
*new_shared
= NULL
;
926 struct alien_cache
**new_alien
= NULL
;
929 if (use_alien_caches
) {
930 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
935 if (cachep
->shared
) {
936 new_shared
= alloc_arraycache(node
,
937 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
942 ret
= init_cache_node(cachep
, node
, gfp
);
946 n
= get_node(cachep
, node
);
947 spin_lock_irq(&n
->list_lock
);
948 if (n
->shared
&& force_change
) {
949 free_block(cachep
, n
->shared
->entry
,
950 n
->shared
->avail
, node
, &list
);
951 n
->shared
->avail
= 0;
954 if (!n
->shared
|| force_change
) {
955 old_shared
= n
->shared
;
956 n
->shared
= new_shared
;
961 n
->alien
= new_alien
;
965 spin_unlock_irq(&n
->list_lock
);
966 slabs_destroy(cachep
, &list
);
971 free_alien_cache(new_alien
);
976 static void cpuup_canceled(long cpu
)
978 struct kmem_cache
*cachep
;
979 struct kmem_cache_node
*n
= NULL
;
980 int node
= cpu_to_mem(cpu
);
981 const struct cpumask
*mask
= cpumask_of_node(node
);
983 list_for_each_entry(cachep
, &slab_caches
, list
) {
984 struct array_cache
*nc
;
985 struct array_cache
*shared
;
986 struct alien_cache
**alien
;
989 n
= get_node(cachep
, node
);
993 spin_lock_irq(&n
->list_lock
);
995 /* Free limit for this kmem_cache_node */
996 n
->free_limit
-= cachep
->batchcount
;
998 /* cpu is dead; no one can alloc from it. */
999 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
1001 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1005 if (!cpumask_empty(mask
)) {
1006 spin_unlock_irq(&n
->list_lock
);
1012 free_block(cachep
, shared
->entry
,
1013 shared
->avail
, node
, &list
);
1020 spin_unlock_irq(&n
->list_lock
);
1024 drain_alien_cache(cachep
, alien
);
1025 free_alien_cache(alien
);
1029 slabs_destroy(cachep
, &list
);
1032 * In the previous loop, all the objects were freed to
1033 * the respective cache's slabs, now we can go ahead and
1034 * shrink each nodelist to its limit.
1036 list_for_each_entry(cachep
, &slab_caches
, list
) {
1037 n
= get_node(cachep
, node
);
1040 drain_freelist(cachep
, n
, INT_MAX
);
1044 static int cpuup_prepare(long cpu
)
1046 struct kmem_cache
*cachep
;
1047 int node
= cpu_to_mem(cpu
);
1051 * We need to do this right in the beginning since
1052 * alloc_arraycache's are going to use this list.
1053 * kmalloc_node allows us to add the slab to the right
1054 * kmem_cache_node and not this cpu's kmem_cache_node
1056 err
= init_cache_node_node(node
);
1061 * Now we can go ahead with allocating the shared arrays and
1064 list_for_each_entry(cachep
, &slab_caches
, list
) {
1065 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1072 cpuup_canceled(cpu
);
1076 static int cpuup_callback(struct notifier_block
*nfb
,
1077 unsigned long action
, void *hcpu
)
1079 long cpu
= (long)hcpu
;
1083 case CPU_UP_PREPARE
:
1084 case CPU_UP_PREPARE_FROZEN
:
1085 mutex_lock(&slab_mutex
);
1086 err
= cpuup_prepare(cpu
);
1087 mutex_unlock(&slab_mutex
);
1090 case CPU_ONLINE_FROZEN
:
1091 start_cpu_timer(cpu
);
1093 #ifdef CONFIG_HOTPLUG_CPU
1094 case CPU_DOWN_PREPARE
:
1095 case CPU_DOWN_PREPARE_FROZEN
:
1097 * Shutdown cache reaper. Note that the slab_mutex is
1098 * held so that if cache_reap() is invoked it cannot do
1099 * anything expensive but will only modify reap_work
1100 * and reschedule the timer.
1102 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1103 /* Now the cache_reaper is guaranteed to be not running. */
1104 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1106 case CPU_DOWN_FAILED
:
1107 case CPU_DOWN_FAILED_FROZEN
:
1108 start_cpu_timer(cpu
);
1111 case CPU_DEAD_FROZEN
:
1113 * Even if all the cpus of a node are down, we don't free the
1114 * kmem_cache_node of any cache. This to avoid a race between
1115 * cpu_down, and a kmalloc allocation from another cpu for
1116 * memory from the node of the cpu going down. The node
1117 * structure is usually allocated from kmem_cache_create() and
1118 * gets destroyed at kmem_cache_destroy().
1122 case CPU_UP_CANCELED
:
1123 case CPU_UP_CANCELED_FROZEN
:
1124 mutex_lock(&slab_mutex
);
1125 cpuup_canceled(cpu
);
1126 mutex_unlock(&slab_mutex
);
1129 return notifier_from_errno(err
);
1132 static struct notifier_block cpucache_notifier
= {
1133 &cpuup_callback
, NULL
, 0
1136 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1138 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1139 * Returns -EBUSY if all objects cannot be drained so that the node is not
1142 * Must hold slab_mutex.
1144 static int __meminit
drain_cache_node_node(int node
)
1146 struct kmem_cache
*cachep
;
1149 list_for_each_entry(cachep
, &slab_caches
, list
) {
1150 struct kmem_cache_node
*n
;
1152 n
= get_node(cachep
, node
);
1156 drain_freelist(cachep
, n
, INT_MAX
);
1158 if (!list_empty(&n
->slabs_full
) ||
1159 !list_empty(&n
->slabs_partial
)) {
1167 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1168 unsigned long action
, void *arg
)
1170 struct memory_notify
*mnb
= arg
;
1174 nid
= mnb
->status_change_nid
;
1179 case MEM_GOING_ONLINE
:
1180 mutex_lock(&slab_mutex
);
1181 ret
= init_cache_node_node(nid
);
1182 mutex_unlock(&slab_mutex
);
1184 case MEM_GOING_OFFLINE
:
1185 mutex_lock(&slab_mutex
);
1186 ret
= drain_cache_node_node(nid
);
1187 mutex_unlock(&slab_mutex
);
1191 case MEM_CANCEL_ONLINE
:
1192 case MEM_CANCEL_OFFLINE
:
1196 return notifier_from_errno(ret
);
1198 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1201 * swap the static kmem_cache_node with kmalloced memory
1203 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1206 struct kmem_cache_node
*ptr
;
1208 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1211 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1213 * Do not assume that spinlocks can be initialized via memcpy:
1215 spin_lock_init(&ptr
->list_lock
);
1217 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1218 cachep
->node
[nodeid
] = ptr
;
1222 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1223 * size of kmem_cache_node.
1225 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1229 for_each_online_node(node
) {
1230 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1231 cachep
->node
[node
]->next_reap
= jiffies
+
1233 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1238 * Initialisation. Called after the page allocator have been initialised and
1239 * before smp_init().
1241 void __init
kmem_cache_init(void)
1245 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1246 sizeof(struct rcu_head
));
1247 kmem_cache
= &kmem_cache_boot
;
1249 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1250 use_alien_caches
= 0;
1252 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1253 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1256 * Fragmentation resistance on low memory - only use bigger
1257 * page orders on machines with more than 32MB of memory if
1258 * not overridden on the command line.
1260 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1261 slab_max_order
= SLAB_MAX_ORDER_HI
;
1263 /* Bootstrap is tricky, because several objects are allocated
1264 * from caches that do not exist yet:
1265 * 1) initialize the kmem_cache cache: it contains the struct
1266 * kmem_cache structures of all caches, except kmem_cache itself:
1267 * kmem_cache is statically allocated.
1268 * Initially an __init data area is used for the head array and the
1269 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1270 * array at the end of the bootstrap.
1271 * 2) Create the first kmalloc cache.
1272 * The struct kmem_cache for the new cache is allocated normally.
1273 * An __init data area is used for the head array.
1274 * 3) Create the remaining kmalloc caches, with minimally sized
1276 * 4) Replace the __init data head arrays for kmem_cache and the first
1277 * kmalloc cache with kmalloc allocated arrays.
1278 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1279 * the other cache's with kmalloc allocated memory.
1280 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1283 /* 1) create the kmem_cache */
1286 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1288 create_boot_cache(kmem_cache
, "kmem_cache",
1289 offsetof(struct kmem_cache
, node
) +
1290 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1291 SLAB_HWCACHE_ALIGN
);
1292 list_add(&kmem_cache
->list
, &slab_caches
);
1293 slab_state
= PARTIAL
;
1296 * Initialize the caches that provide memory for the kmem_cache_node
1297 * structures first. Without this, further allocations will bug.
1299 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1300 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1301 slab_state
= PARTIAL_NODE
;
1302 setup_kmalloc_cache_index_table();
1304 slab_early_init
= 0;
1306 /* 5) Replace the bootstrap kmem_cache_node */
1310 for_each_online_node(nid
) {
1311 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1313 init_list(kmalloc_caches
[INDEX_NODE
],
1314 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1318 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1321 void __init
kmem_cache_init_late(void)
1323 struct kmem_cache
*cachep
;
1327 /* 6) resize the head arrays to their final sizes */
1328 mutex_lock(&slab_mutex
);
1329 list_for_each_entry(cachep
, &slab_caches
, list
)
1330 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1332 mutex_unlock(&slab_mutex
);
1338 * Register a cpu startup notifier callback that initializes
1339 * cpu_cache_get for all new cpus
1341 register_cpu_notifier(&cpucache_notifier
);
1345 * Register a memory hotplug callback that initializes and frees
1348 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1352 * The reap timers are started later, with a module init call: That part
1353 * of the kernel is not yet operational.
1357 static int __init
cpucache_init(void)
1362 * Register the timers that return unneeded pages to the page allocator
1364 for_each_online_cpu(cpu
)
1365 start_cpu_timer(cpu
);
1371 __initcall(cpucache_init
);
1373 static noinline
void
1374 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1377 struct kmem_cache_node
*n
;
1379 unsigned long flags
;
1381 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1382 DEFAULT_RATELIMIT_BURST
);
1384 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1387 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1388 nodeid
, gfpflags
, &gfpflags
);
1389 pr_warn(" cache: %s, object size: %d, order: %d\n",
1390 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1392 for_each_kmem_cache_node(cachep
, node
, n
) {
1393 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1394 unsigned long active_slabs
= 0, num_slabs
= 0;
1396 spin_lock_irqsave(&n
->list_lock
, flags
);
1397 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1398 active_objs
+= cachep
->num
;
1401 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1402 active_objs
+= page
->active
;
1405 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1408 free_objects
+= n
->free_objects
;
1409 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1411 num_slabs
+= active_slabs
;
1412 num_objs
= num_slabs
* cachep
->num
;
1413 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1414 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1421 * Interface to system's page allocator. No need to hold the
1422 * kmem_cache_node ->list_lock.
1424 * If we requested dmaable memory, we will get it. Even if we
1425 * did not request dmaable memory, we might get it, but that
1426 * would be relatively rare and ignorable.
1428 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1434 flags
|= cachep
->allocflags
;
1435 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1436 flags
|= __GFP_RECLAIMABLE
;
1438 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1440 slab_out_of_memory(cachep
, flags
, nodeid
);
1444 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1445 __free_pages(page
, cachep
->gfporder
);
1449 nr_pages
= (1 << cachep
->gfporder
);
1450 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1451 add_zone_page_state(page_zone(page
),
1452 NR_SLAB_RECLAIMABLE
, nr_pages
);
1454 add_zone_page_state(page_zone(page
),
1455 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1457 __SetPageSlab(page
);
1458 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1459 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1460 SetPageSlabPfmemalloc(page
);
1462 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1463 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1466 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1468 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1475 * Interface to system's page release.
1477 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1479 int order
= cachep
->gfporder
;
1480 unsigned long nr_freed
= (1 << order
);
1482 kmemcheck_free_shadow(page
, order
);
1484 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1485 sub_zone_page_state(page_zone(page
),
1486 NR_SLAB_RECLAIMABLE
, nr_freed
);
1488 sub_zone_page_state(page_zone(page
),
1489 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1491 BUG_ON(!PageSlab(page
));
1492 __ClearPageSlabPfmemalloc(page
);
1493 __ClearPageSlab(page
);
1494 page_mapcount_reset(page
);
1495 page
->mapping
= NULL
;
1497 if (current
->reclaim_state
)
1498 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1499 memcg_uncharge_slab(page
, order
, cachep
);
1500 __free_pages(page
, order
);
1503 static void kmem_rcu_free(struct rcu_head
*head
)
1505 struct kmem_cache
*cachep
;
1508 page
= container_of(head
, struct page
, rcu_head
);
1509 cachep
= page
->slab_cache
;
1511 kmem_freepages(cachep
, page
);
1515 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1517 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1518 (cachep
->size
% PAGE_SIZE
) == 0)
1524 #ifdef CONFIG_DEBUG_PAGEALLOC
1525 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1526 unsigned long caller
)
1528 int size
= cachep
->object_size
;
1530 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1532 if (size
< 5 * sizeof(unsigned long))
1535 *addr
++ = 0x12345678;
1537 *addr
++ = smp_processor_id();
1538 size
-= 3 * sizeof(unsigned long);
1540 unsigned long *sptr
= &caller
;
1541 unsigned long svalue
;
1543 while (!kstack_end(sptr
)) {
1545 if (kernel_text_address(svalue
)) {
1547 size
-= sizeof(unsigned long);
1548 if (size
<= sizeof(unsigned long))
1554 *addr
++ = 0x87654321;
1557 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1558 int map
, unsigned long caller
)
1560 if (!is_debug_pagealloc_cache(cachep
))
1564 store_stackinfo(cachep
, objp
, caller
);
1566 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1570 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1571 int map
, unsigned long caller
) {}
1575 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1577 int size
= cachep
->object_size
;
1578 addr
= &((char *)addr
)[obj_offset(cachep
)];
1580 memset(addr
, val
, size
);
1581 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1584 static void dump_line(char *data
, int offset
, int limit
)
1587 unsigned char error
= 0;
1590 pr_err("%03x: ", offset
);
1591 for (i
= 0; i
< limit
; i
++) {
1592 if (data
[offset
+ i
] != POISON_FREE
) {
1593 error
= data
[offset
+ i
];
1597 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1598 &data
[offset
], limit
, 1);
1600 if (bad_count
== 1) {
1601 error
^= POISON_FREE
;
1602 if (!(error
& (error
- 1))) {
1603 pr_err("Single bit error detected. Probably bad RAM.\n");
1605 pr_err("Run memtest86+ or a similar memory test tool.\n");
1607 pr_err("Run a memory test tool.\n");
1616 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1621 if (cachep
->flags
& SLAB_RED_ZONE
) {
1622 pr_err("Redzone: 0x%llx/0x%llx\n",
1623 *dbg_redzone1(cachep
, objp
),
1624 *dbg_redzone2(cachep
, objp
));
1627 if (cachep
->flags
& SLAB_STORE_USER
) {
1628 pr_err("Last user: [<%p>](%pSR)\n",
1629 *dbg_userword(cachep
, objp
),
1630 *dbg_userword(cachep
, objp
));
1632 realobj
= (char *)objp
+ obj_offset(cachep
);
1633 size
= cachep
->object_size
;
1634 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1637 if (i
+ limit
> size
)
1639 dump_line(realobj
, i
, limit
);
1643 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1649 if (is_debug_pagealloc_cache(cachep
))
1652 realobj
= (char *)objp
+ obj_offset(cachep
);
1653 size
= cachep
->object_size
;
1655 for (i
= 0; i
< size
; i
++) {
1656 char exp
= POISON_FREE
;
1659 if (realobj
[i
] != exp
) {
1664 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1665 print_tainted(), cachep
->name
,
1667 print_objinfo(cachep
, objp
, 0);
1669 /* Hexdump the affected line */
1672 if (i
+ limit
> size
)
1674 dump_line(realobj
, i
, limit
);
1677 /* Limit to 5 lines */
1683 /* Print some data about the neighboring objects, if they
1686 struct page
*page
= virt_to_head_page(objp
);
1689 objnr
= obj_to_index(cachep
, page
, objp
);
1691 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1692 realobj
= (char *)objp
+ obj_offset(cachep
);
1693 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1694 print_objinfo(cachep
, objp
, 2);
1696 if (objnr
+ 1 < cachep
->num
) {
1697 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1698 realobj
= (char *)objp
+ obj_offset(cachep
);
1699 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1700 print_objinfo(cachep
, objp
, 2);
1707 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1712 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1713 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1717 for (i
= 0; i
< cachep
->num
; i
++) {
1718 void *objp
= index_to_obj(cachep
, page
, i
);
1720 if (cachep
->flags
& SLAB_POISON
) {
1721 check_poison_obj(cachep
, objp
);
1722 slab_kernel_map(cachep
, objp
, 1, 0);
1724 if (cachep
->flags
& SLAB_RED_ZONE
) {
1725 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1726 slab_error(cachep
, "start of a freed object was overwritten");
1727 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1728 slab_error(cachep
, "end of a freed object was overwritten");
1733 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1740 * slab_destroy - destroy and release all objects in a slab
1741 * @cachep: cache pointer being destroyed
1742 * @page: page pointer being destroyed
1744 * Destroy all the objs in a slab page, and release the mem back to the system.
1745 * Before calling the slab page must have been unlinked from the cache. The
1746 * kmem_cache_node ->list_lock is not held/needed.
1748 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1752 freelist
= page
->freelist
;
1753 slab_destroy_debugcheck(cachep
, page
);
1754 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1755 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1757 kmem_freepages(cachep
, page
);
1760 * From now on, we don't use freelist
1761 * although actual page can be freed in rcu context
1763 if (OFF_SLAB(cachep
))
1764 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1767 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1769 struct page
*page
, *n
;
1771 list_for_each_entry_safe(page
, n
, list
, lru
) {
1772 list_del(&page
->lru
);
1773 slab_destroy(cachep
, page
);
1778 * calculate_slab_order - calculate size (page order) of slabs
1779 * @cachep: pointer to the cache that is being created
1780 * @size: size of objects to be created in this cache.
1781 * @flags: slab allocation flags
1783 * Also calculates the number of objects per slab.
1785 * This could be made much more intelligent. For now, try to avoid using
1786 * high order pages for slabs. When the gfp() functions are more friendly
1787 * towards high-order requests, this should be changed.
1789 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1790 size_t size
, unsigned long flags
)
1792 size_t left_over
= 0;
1795 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1799 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1803 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1804 if (num
> SLAB_OBJ_MAX_NUM
)
1807 if (flags
& CFLGS_OFF_SLAB
) {
1808 struct kmem_cache
*freelist_cache
;
1809 size_t freelist_size
;
1811 freelist_size
= num
* sizeof(freelist_idx_t
);
1812 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1813 if (!freelist_cache
)
1817 * Needed to avoid possible looping condition
1818 * in cache_grow_begin()
1820 if (OFF_SLAB(freelist_cache
))
1823 /* check if off slab has enough benefit */
1824 if (freelist_cache
->size
> cachep
->size
/ 2)
1828 /* Found something acceptable - save it away */
1830 cachep
->gfporder
= gfporder
;
1831 left_over
= remainder
;
1834 * A VFS-reclaimable slab tends to have most allocations
1835 * as GFP_NOFS and we really don't want to have to be allocating
1836 * higher-order pages when we are unable to shrink dcache.
1838 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1842 * Large number of objects is good, but very large slabs are
1843 * currently bad for the gfp()s.
1845 if (gfporder
>= slab_max_order
)
1849 * Acceptable internal fragmentation?
1851 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1857 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1858 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1862 struct array_cache __percpu
*cpu_cache
;
1864 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1865 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1870 for_each_possible_cpu(cpu
) {
1871 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1872 entries
, batchcount
);
1878 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1880 if (slab_state
>= FULL
)
1881 return enable_cpucache(cachep
, gfp
);
1883 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1884 if (!cachep
->cpu_cache
)
1887 if (slab_state
== DOWN
) {
1888 /* Creation of first cache (kmem_cache). */
1889 set_up_node(kmem_cache
, CACHE_CACHE
);
1890 } else if (slab_state
== PARTIAL
) {
1891 /* For kmem_cache_node */
1892 set_up_node(cachep
, SIZE_NODE
);
1896 for_each_online_node(node
) {
1897 cachep
->node
[node
] = kmalloc_node(
1898 sizeof(struct kmem_cache_node
), gfp
, node
);
1899 BUG_ON(!cachep
->node
[node
]);
1900 kmem_cache_node_init(cachep
->node
[node
]);
1904 cachep
->node
[numa_mem_id()]->next_reap
=
1905 jiffies
+ REAPTIMEOUT_NODE
+
1906 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1908 cpu_cache_get(cachep
)->avail
= 0;
1909 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1910 cpu_cache_get(cachep
)->batchcount
= 1;
1911 cpu_cache_get(cachep
)->touched
= 0;
1912 cachep
->batchcount
= 1;
1913 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1917 unsigned long kmem_cache_flags(unsigned long object_size
,
1918 unsigned long flags
, const char *name
,
1919 void (*ctor
)(void *))
1925 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1926 unsigned long flags
, void (*ctor
)(void *))
1928 struct kmem_cache
*cachep
;
1930 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1935 * Adjust the object sizes so that we clear
1936 * the complete object on kzalloc.
1938 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1943 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1944 size_t size
, unsigned long flags
)
1950 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1953 left
= calculate_slab_order(cachep
, size
,
1954 flags
| CFLGS_OBJFREELIST_SLAB
);
1958 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1961 cachep
->colour
= left
/ cachep
->colour_off
;
1966 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1967 size_t size
, unsigned long flags
)
1974 * Always use on-slab management when SLAB_NOLEAKTRACE
1975 * to avoid recursive calls into kmemleak.
1977 if (flags
& SLAB_NOLEAKTRACE
)
1981 * Size is large, assume best to place the slab management obj
1982 * off-slab (should allow better packing of objs).
1984 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1989 * If the slab has been placed off-slab, and we have enough space then
1990 * move it on-slab. This is at the expense of any extra colouring.
1992 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1995 cachep
->colour
= left
/ cachep
->colour_off
;
2000 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
2001 size_t size
, unsigned long flags
)
2007 left
= calculate_slab_order(cachep
, size
, flags
);
2011 cachep
->colour
= left
/ cachep
->colour_off
;
2017 * __kmem_cache_create - Create a cache.
2018 * @cachep: cache management descriptor
2019 * @flags: SLAB flags
2021 * Returns a ptr to the cache on success, NULL on failure.
2022 * Cannot be called within a int, but can be interrupted.
2023 * The @ctor is run when new pages are allocated by the cache.
2027 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2028 * to catch references to uninitialised memory.
2030 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2031 * for buffer overruns.
2033 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2034 * cacheline. This can be beneficial if you're counting cycles as closely
2038 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2040 size_t ralign
= BYTES_PER_WORD
;
2043 size_t size
= cachep
->size
;
2048 * Enable redzoning and last user accounting, except for caches with
2049 * large objects, if the increased size would increase the object size
2050 * above the next power of two: caches with object sizes just above a
2051 * power of two have a significant amount of internal fragmentation.
2053 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2054 2 * sizeof(unsigned long long)))
2055 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2056 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2057 flags
|= SLAB_POISON
;
2062 * Check that size is in terms of words. This is needed to avoid
2063 * unaligned accesses for some archs when redzoning is used, and makes
2064 * sure any on-slab bufctl's are also correctly aligned.
2066 if (size
& (BYTES_PER_WORD
- 1)) {
2067 size
+= (BYTES_PER_WORD
- 1);
2068 size
&= ~(BYTES_PER_WORD
- 1);
2071 if (flags
& SLAB_RED_ZONE
) {
2072 ralign
= REDZONE_ALIGN
;
2073 /* If redzoning, ensure that the second redzone is suitably
2074 * aligned, by adjusting the object size accordingly. */
2075 size
+= REDZONE_ALIGN
- 1;
2076 size
&= ~(REDZONE_ALIGN
- 1);
2079 /* 3) caller mandated alignment */
2080 if (ralign
< cachep
->align
) {
2081 ralign
= cachep
->align
;
2083 /* disable debug if necessary */
2084 if (ralign
> __alignof__(unsigned long long))
2085 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2089 cachep
->align
= ralign
;
2090 cachep
->colour_off
= cache_line_size();
2091 /* Offset must be a multiple of the alignment. */
2092 if (cachep
->colour_off
< cachep
->align
)
2093 cachep
->colour_off
= cachep
->align
;
2095 if (slab_is_available())
2103 * Both debugging options require word-alignment which is calculated
2106 if (flags
& SLAB_RED_ZONE
) {
2107 /* add space for red zone words */
2108 cachep
->obj_offset
+= sizeof(unsigned long long);
2109 size
+= 2 * sizeof(unsigned long long);
2111 if (flags
& SLAB_STORE_USER
) {
2112 /* user store requires one word storage behind the end of
2113 * the real object. But if the second red zone needs to be
2114 * aligned to 64 bits, we must allow that much space.
2116 if (flags
& SLAB_RED_ZONE
)
2117 size
+= REDZONE_ALIGN
;
2119 size
+= BYTES_PER_WORD
;
2123 kasan_cache_create(cachep
, &size
, &flags
);
2125 size
= ALIGN(size
, cachep
->align
);
2127 * We should restrict the number of objects in a slab to implement
2128 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2130 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2131 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2135 * To activate debug pagealloc, off-slab management is necessary
2136 * requirement. In early phase of initialization, small sized slab
2137 * doesn't get initialized so it would not be possible. So, we need
2138 * to check size >= 256. It guarantees that all necessary small
2139 * sized slab is initialized in current slab initialization sequence.
2141 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2142 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2143 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2144 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2146 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2147 flags
|= CFLGS_OFF_SLAB
;
2148 cachep
->obj_offset
+= tmp_size
- size
;
2156 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2157 flags
|= CFLGS_OBJFREELIST_SLAB
;
2161 if (set_off_slab_cache(cachep
, size
, flags
)) {
2162 flags
|= CFLGS_OFF_SLAB
;
2166 if (set_on_slab_cache(cachep
, size
, flags
))
2172 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2173 cachep
->flags
= flags
;
2174 cachep
->allocflags
= __GFP_COMP
;
2175 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2176 cachep
->allocflags
|= GFP_DMA
;
2177 cachep
->size
= size
;
2178 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2182 * If we're going to use the generic kernel_map_pages()
2183 * poisoning, then it's going to smash the contents of
2184 * the redzone and userword anyhow, so switch them off.
2186 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2187 (cachep
->flags
& SLAB_POISON
) &&
2188 is_debug_pagealloc_cache(cachep
))
2189 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2192 if (OFF_SLAB(cachep
)) {
2193 cachep
->freelist_cache
=
2194 kmalloc_slab(cachep
->freelist_size
, 0u);
2197 err
= setup_cpu_cache(cachep
, gfp
);
2199 __kmem_cache_release(cachep
);
2207 static void check_irq_off(void)
2209 BUG_ON(!irqs_disabled());
2212 static void check_irq_on(void)
2214 BUG_ON(irqs_disabled());
2217 static void check_mutex_acquired(void)
2219 BUG_ON(!mutex_is_locked(&slab_mutex
));
2222 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2226 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2230 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2234 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2239 #define check_irq_off() do { } while(0)
2240 #define check_irq_on() do { } while(0)
2241 #define check_mutex_acquired() do { } while(0)
2242 #define check_spinlock_acquired(x) do { } while(0)
2243 #define check_spinlock_acquired_node(x, y) do { } while(0)
2246 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2247 int node
, bool free_all
, struct list_head
*list
)
2251 if (!ac
|| !ac
->avail
)
2254 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2255 if (tofree
> ac
->avail
)
2256 tofree
= (ac
->avail
+ 1) / 2;
2258 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2259 ac
->avail
-= tofree
;
2260 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2263 static void do_drain(void *arg
)
2265 struct kmem_cache
*cachep
= arg
;
2266 struct array_cache
*ac
;
2267 int node
= numa_mem_id();
2268 struct kmem_cache_node
*n
;
2272 ac
= cpu_cache_get(cachep
);
2273 n
= get_node(cachep
, node
);
2274 spin_lock(&n
->list_lock
);
2275 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2276 spin_unlock(&n
->list_lock
);
2277 slabs_destroy(cachep
, &list
);
2281 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2283 struct kmem_cache_node
*n
;
2287 on_each_cpu(do_drain
, cachep
, 1);
2289 for_each_kmem_cache_node(cachep
, node
, n
)
2291 drain_alien_cache(cachep
, n
->alien
);
2293 for_each_kmem_cache_node(cachep
, node
, n
) {
2294 spin_lock_irq(&n
->list_lock
);
2295 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2296 spin_unlock_irq(&n
->list_lock
);
2298 slabs_destroy(cachep
, &list
);
2303 * Remove slabs from the list of free slabs.
2304 * Specify the number of slabs to drain in tofree.
2306 * Returns the actual number of slabs released.
2308 static int drain_freelist(struct kmem_cache
*cache
,
2309 struct kmem_cache_node
*n
, int tofree
)
2311 struct list_head
*p
;
2316 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2318 spin_lock_irq(&n
->list_lock
);
2319 p
= n
->slabs_free
.prev
;
2320 if (p
== &n
->slabs_free
) {
2321 spin_unlock_irq(&n
->list_lock
);
2325 page
= list_entry(p
, struct page
, lru
);
2326 list_del(&page
->lru
);
2328 * Safe to drop the lock. The slab is no longer linked
2331 n
->free_objects
-= cache
->num
;
2332 spin_unlock_irq(&n
->list_lock
);
2333 slab_destroy(cache
, page
);
2340 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2344 struct kmem_cache_node
*n
;
2346 drain_cpu_caches(cachep
);
2349 for_each_kmem_cache_node(cachep
, node
, n
) {
2350 drain_freelist(cachep
, n
, INT_MAX
);
2352 ret
+= !list_empty(&n
->slabs_full
) ||
2353 !list_empty(&n
->slabs_partial
);
2355 return (ret
? 1 : 0);
2358 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2360 return __kmem_cache_shrink(cachep
, false);
2363 void __kmem_cache_release(struct kmem_cache
*cachep
)
2366 struct kmem_cache_node
*n
;
2368 free_percpu(cachep
->cpu_cache
);
2370 /* NUMA: free the node structures */
2371 for_each_kmem_cache_node(cachep
, i
, n
) {
2373 free_alien_cache(n
->alien
);
2375 cachep
->node
[i
] = NULL
;
2380 * Get the memory for a slab management obj.
2382 * For a slab cache when the slab descriptor is off-slab, the
2383 * slab descriptor can't come from the same cache which is being created,
2384 * Because if it is the case, that means we defer the creation of
2385 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2386 * And we eventually call down to __kmem_cache_create(), which
2387 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2388 * This is a "chicken-and-egg" problem.
2390 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2391 * which are all initialized during kmem_cache_init().
2393 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2394 struct page
*page
, int colour_off
,
2395 gfp_t local_flags
, int nodeid
)
2398 void *addr
= page_address(page
);
2400 page
->s_mem
= addr
+ colour_off
;
2403 if (OBJFREELIST_SLAB(cachep
))
2405 else if (OFF_SLAB(cachep
)) {
2406 /* Slab management obj is off-slab. */
2407 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2408 local_flags
, nodeid
);
2412 /* We will use last bytes at the slab for freelist */
2413 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2414 cachep
->freelist_size
;
2420 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2422 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2425 static inline void set_free_obj(struct page
*page
,
2426 unsigned int idx
, freelist_idx_t val
)
2428 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2431 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2436 for (i
= 0; i
< cachep
->num
; i
++) {
2437 void *objp
= index_to_obj(cachep
, page
, i
);
2439 if (cachep
->flags
& SLAB_STORE_USER
)
2440 *dbg_userword(cachep
, objp
) = NULL
;
2442 if (cachep
->flags
& SLAB_RED_ZONE
) {
2443 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2444 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2447 * Constructors are not allowed to allocate memory from the same
2448 * cache which they are a constructor for. Otherwise, deadlock.
2449 * They must also be threaded.
2451 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2452 kasan_unpoison_object_data(cachep
,
2453 objp
+ obj_offset(cachep
));
2454 cachep
->ctor(objp
+ obj_offset(cachep
));
2455 kasan_poison_object_data(
2456 cachep
, objp
+ obj_offset(cachep
));
2459 if (cachep
->flags
& SLAB_RED_ZONE
) {
2460 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2461 slab_error(cachep
, "constructor overwrote the end of an object");
2462 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2463 slab_error(cachep
, "constructor overwrote the start of an object");
2465 /* need to poison the objs? */
2466 if (cachep
->flags
& SLAB_POISON
) {
2467 poison_obj(cachep
, objp
, POISON_FREE
);
2468 slab_kernel_map(cachep
, objp
, 0, 0);
2474 static void cache_init_objs(struct kmem_cache
*cachep
,
2480 cache_init_objs_debug(cachep
, page
);
2482 if (OBJFREELIST_SLAB(cachep
)) {
2483 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2487 for (i
= 0; i
< cachep
->num
; i
++) {
2488 /* constructor could break poison info */
2489 if (DEBUG
== 0 && cachep
->ctor
) {
2490 objp
= index_to_obj(cachep
, page
, i
);
2491 kasan_unpoison_object_data(cachep
, objp
);
2493 kasan_poison_object_data(cachep
, objp
);
2496 set_free_obj(page
, i
, i
);
2500 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2502 if (CONFIG_ZONE_DMA_FLAG
) {
2503 if (flags
& GFP_DMA
)
2504 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2506 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2510 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2514 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2518 if (cachep
->flags
& SLAB_STORE_USER
)
2519 set_store_user_dirty(cachep
);
2525 static void slab_put_obj(struct kmem_cache
*cachep
,
2526 struct page
*page
, void *objp
)
2528 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2532 /* Verify double free bug */
2533 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2534 if (get_free_obj(page
, i
) == objnr
) {
2535 pr_err("slab: double free detected in cache '%s', objp %p\n",
2536 cachep
->name
, objp
);
2542 if (!page
->freelist
)
2543 page
->freelist
= objp
+ obj_offset(cachep
);
2545 set_free_obj(page
, page
->active
, objnr
);
2549 * Map pages beginning at addr to the given cache and slab. This is required
2550 * for the slab allocator to be able to lookup the cache and slab of a
2551 * virtual address for kfree, ksize, and slab debugging.
2553 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2556 page
->slab_cache
= cache
;
2557 page
->freelist
= freelist
;
2561 * Grow (by 1) the number of slabs within a cache. This is called by
2562 * kmem_cache_alloc() when there are no active objs left in a cache.
2564 static struct page
*cache_grow_begin(struct kmem_cache
*cachep
,
2565 gfp_t flags
, int nodeid
)
2571 struct kmem_cache_node
*n
;
2575 * Be lazy and only check for valid flags here, keeping it out of the
2576 * critical path in kmem_cache_alloc().
2578 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2579 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2582 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2585 if (gfpflags_allow_blocking(local_flags
))
2589 * The test for missing atomic flag is performed here, rather than
2590 * the more obvious place, simply to reduce the critical path length
2591 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2592 * will eventually be caught here (where it matters).
2594 kmem_flagcheck(cachep
, flags
);
2597 * Get mem for the objs. Attempt to allocate a physical page from
2600 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2604 page_node
= page_to_nid(page
);
2605 n
= get_node(cachep
, page_node
);
2607 /* Get colour for the slab, and cal the next value. */
2609 if (n
->colour_next
>= cachep
->colour
)
2612 offset
= n
->colour_next
;
2613 if (offset
>= cachep
->colour
)
2616 offset
*= cachep
->colour_off
;
2618 /* Get slab management. */
2619 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2620 local_flags
& ~GFP_CONSTRAINT_MASK
, page_node
);
2621 if (OFF_SLAB(cachep
) && !freelist
)
2624 slab_map_pages(cachep
, page
, freelist
);
2626 kasan_poison_slab(page
);
2627 cache_init_objs(cachep
, page
);
2629 if (gfpflags_allow_blocking(local_flags
))
2630 local_irq_disable();
2635 kmem_freepages(cachep
, page
);
2637 if (gfpflags_allow_blocking(local_flags
))
2638 local_irq_disable();
2642 static void cache_grow_end(struct kmem_cache
*cachep
, struct page
*page
)
2644 struct kmem_cache_node
*n
;
2652 INIT_LIST_HEAD(&page
->lru
);
2653 n
= get_node(cachep
, page_to_nid(page
));
2655 spin_lock(&n
->list_lock
);
2657 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2659 fixup_slab_list(cachep
, n
, page
, &list
);
2660 STATS_INC_GROWN(cachep
);
2661 n
->free_objects
+= cachep
->num
- page
->active
;
2662 spin_unlock(&n
->list_lock
);
2664 fixup_objfreelist_debug(cachep
, &list
);
2670 * Perform extra freeing checks:
2671 * - detect bad pointers.
2672 * - POISON/RED_ZONE checking
2674 static void kfree_debugcheck(const void *objp
)
2676 if (!virt_addr_valid(objp
)) {
2677 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2678 (unsigned long)objp
);
2683 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2685 unsigned long long redzone1
, redzone2
;
2687 redzone1
= *dbg_redzone1(cache
, obj
);
2688 redzone2
= *dbg_redzone2(cache
, obj
);
2693 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2696 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2697 slab_error(cache
, "double free detected");
2699 slab_error(cache
, "memory outside object was overwritten");
2701 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2702 obj
, redzone1
, redzone2
);
2705 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2706 unsigned long caller
)
2711 BUG_ON(virt_to_cache(objp
) != cachep
);
2713 objp
-= obj_offset(cachep
);
2714 kfree_debugcheck(objp
);
2715 page
= virt_to_head_page(objp
);
2717 if (cachep
->flags
& SLAB_RED_ZONE
) {
2718 verify_redzone_free(cachep
, objp
);
2719 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2720 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2722 if (cachep
->flags
& SLAB_STORE_USER
) {
2723 set_store_user_dirty(cachep
);
2724 *dbg_userword(cachep
, objp
) = (void *)caller
;
2727 objnr
= obj_to_index(cachep
, page
, objp
);
2729 BUG_ON(objnr
>= cachep
->num
);
2730 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2732 if (cachep
->flags
& SLAB_POISON
) {
2733 poison_obj(cachep
, objp
, POISON_FREE
);
2734 slab_kernel_map(cachep
, objp
, 0, caller
);
2740 #define kfree_debugcheck(x) do { } while(0)
2741 #define cache_free_debugcheck(x,objp,z) (objp)
2744 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2752 objp
= next
- obj_offset(cachep
);
2753 next
= *(void **)next
;
2754 poison_obj(cachep
, objp
, POISON_FREE
);
2759 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2760 struct kmem_cache_node
*n
, struct page
*page
,
2763 /* move slabp to correct slabp list: */
2764 list_del(&page
->lru
);
2765 if (page
->active
== cachep
->num
) {
2766 list_add(&page
->lru
, &n
->slabs_full
);
2767 if (OBJFREELIST_SLAB(cachep
)) {
2769 /* Poisoning will be done without holding the lock */
2770 if (cachep
->flags
& SLAB_POISON
) {
2771 void **objp
= page
->freelist
;
2777 page
->freelist
= NULL
;
2780 list_add(&page
->lru
, &n
->slabs_partial
);
2783 /* Try to find non-pfmemalloc slab if needed */
2784 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2785 struct page
*page
, bool pfmemalloc
)
2793 if (!PageSlabPfmemalloc(page
))
2796 /* No need to keep pfmemalloc slab if we have enough free objects */
2797 if (n
->free_objects
> n
->free_limit
) {
2798 ClearPageSlabPfmemalloc(page
);
2802 /* Move pfmemalloc slab to the end of list to speed up next search */
2803 list_del(&page
->lru
);
2805 list_add_tail(&page
->lru
, &n
->slabs_free
);
2807 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2809 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2810 if (!PageSlabPfmemalloc(page
))
2814 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2815 if (!PageSlabPfmemalloc(page
))
2822 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2826 page
= list_first_entry_or_null(&n
->slabs_partial
,
2829 n
->free_touched
= 1;
2830 page
= list_first_entry_or_null(&n
->slabs_free
,
2834 if (sk_memalloc_socks())
2835 return get_valid_first_slab(n
, page
, pfmemalloc
);
2840 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2841 struct kmem_cache_node
*n
, gfp_t flags
)
2847 if (!gfp_pfmemalloc_allowed(flags
))
2850 spin_lock(&n
->list_lock
);
2851 page
= get_first_slab(n
, true);
2853 spin_unlock(&n
->list_lock
);
2857 obj
= slab_get_obj(cachep
, page
);
2860 fixup_slab_list(cachep
, n
, page
, &list
);
2862 spin_unlock(&n
->list_lock
);
2863 fixup_objfreelist_debug(cachep
, &list
);
2869 * Slab list should be fixed up by fixup_slab_list() for existing slab
2870 * or cache_grow_end() for new slab
2872 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2873 struct array_cache
*ac
, struct page
*page
, int batchcount
)
2876 * There must be at least one object available for
2879 BUG_ON(page
->active
>= cachep
->num
);
2881 while (page
->active
< cachep
->num
&& batchcount
--) {
2882 STATS_INC_ALLOCED(cachep
);
2883 STATS_INC_ACTIVE(cachep
);
2884 STATS_SET_HIGH(cachep
);
2886 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2892 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2895 struct kmem_cache_node
*n
;
2896 struct array_cache
*ac
;
2902 node
= numa_mem_id();
2904 ac
= cpu_cache_get(cachep
);
2905 batchcount
= ac
->batchcount
;
2906 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2908 * If there was little recent activity on this cache, then
2909 * perform only a partial refill. Otherwise we could generate
2912 batchcount
= BATCHREFILL_LIMIT
;
2914 n
= get_node(cachep
, node
);
2916 BUG_ON(ac
->avail
> 0 || !n
);
2917 spin_lock(&n
->list_lock
);
2919 /* See if we can refill from the shared array */
2920 if (n
->shared
&& transfer_objects(ac
, n
->shared
, batchcount
)) {
2921 n
->shared
->touched
= 1;
2925 while (batchcount
> 0) {
2926 /* Get slab alloc is to come from. */
2927 page
= get_first_slab(n
, false);
2931 check_spinlock_acquired(cachep
);
2933 batchcount
= alloc_block(cachep
, ac
, page
, batchcount
);
2934 fixup_slab_list(cachep
, n
, page
, &list
);
2938 n
->free_objects
-= ac
->avail
;
2940 spin_unlock(&n
->list_lock
);
2941 fixup_objfreelist_debug(cachep
, &list
);
2943 if (unlikely(!ac
->avail
)) {
2944 /* Check if we can use obj in pfmemalloc slab */
2945 if (sk_memalloc_socks()) {
2946 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2952 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
2955 * cache_grow_begin() can reenable interrupts,
2956 * then ac could change.
2958 ac
= cpu_cache_get(cachep
);
2959 if (!ac
->avail
&& page
)
2960 alloc_block(cachep
, ac
, page
, batchcount
);
2961 cache_grow_end(cachep
, page
);
2968 return ac
->entry
[--ac
->avail
];
2971 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2974 might_sleep_if(gfpflags_allow_blocking(flags
));
2976 kmem_flagcheck(cachep
, flags
);
2981 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2982 gfp_t flags
, void *objp
, unsigned long caller
)
2986 if (cachep
->flags
& SLAB_POISON
) {
2987 check_poison_obj(cachep
, objp
);
2988 slab_kernel_map(cachep
, objp
, 1, 0);
2989 poison_obj(cachep
, objp
, POISON_INUSE
);
2991 if (cachep
->flags
& SLAB_STORE_USER
)
2992 *dbg_userword(cachep
, objp
) = (void *)caller
;
2994 if (cachep
->flags
& SLAB_RED_ZONE
) {
2995 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2996 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2997 slab_error(cachep
, "double free, or memory outside object was overwritten");
2998 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2999 objp
, *dbg_redzone1(cachep
, objp
),
3000 *dbg_redzone2(cachep
, objp
));
3002 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3003 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3006 objp
+= obj_offset(cachep
);
3007 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3009 if (ARCH_SLAB_MINALIGN
&&
3010 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3011 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3012 objp
, (int)ARCH_SLAB_MINALIGN
);
3017 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3020 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3023 struct array_cache
*ac
;
3027 ac
= cpu_cache_get(cachep
);
3028 if (likely(ac
->avail
)) {
3030 objp
= ac
->entry
[--ac
->avail
];
3032 STATS_INC_ALLOCHIT(cachep
);
3036 STATS_INC_ALLOCMISS(cachep
);
3037 objp
= cache_alloc_refill(cachep
, flags
);
3039 * the 'ac' may be updated by cache_alloc_refill(),
3040 * and kmemleak_erase() requires its correct value.
3042 ac
= cpu_cache_get(cachep
);
3046 * To avoid a false negative, if an object that is in one of the
3047 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3048 * treat the array pointers as a reference to the object.
3051 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3057 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3059 * If we are in_interrupt, then process context, including cpusets and
3060 * mempolicy, may not apply and should not be used for allocation policy.
3062 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3064 int nid_alloc
, nid_here
;
3066 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3068 nid_alloc
= nid_here
= numa_mem_id();
3069 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3070 nid_alloc
= cpuset_slab_spread_node();
3071 else if (current
->mempolicy
)
3072 nid_alloc
= mempolicy_slab_node();
3073 if (nid_alloc
!= nid_here
)
3074 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3079 * Fallback function if there was no memory available and no objects on a
3080 * certain node and fall back is permitted. First we scan all the
3081 * available node for available objects. If that fails then we
3082 * perform an allocation without specifying a node. This allows the page
3083 * allocator to do its reclaim / fallback magic. We then insert the
3084 * slab into the proper nodelist and then allocate from it.
3086 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3088 struct zonelist
*zonelist
;
3091 enum zone_type high_zoneidx
= gfp_zone(flags
);
3095 unsigned int cpuset_mems_cookie
;
3097 if (flags
& __GFP_THISNODE
)
3101 cpuset_mems_cookie
= read_mems_allowed_begin();
3102 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3106 * Look through allowed nodes for objects available
3107 * from existing per node queues.
3109 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3110 nid
= zone_to_nid(zone
);
3112 if (cpuset_zone_allowed(zone
, flags
) &&
3113 get_node(cache
, nid
) &&
3114 get_node(cache
, nid
)->free_objects
) {
3115 obj
= ____cache_alloc_node(cache
,
3116 gfp_exact_node(flags
), nid
);
3124 * This allocation will be performed within the constraints
3125 * of the current cpuset / memory policy requirements.
3126 * We may trigger various forms of reclaim on the allowed
3127 * set and go into memory reserves if necessary.
3129 page
= cache_grow_begin(cache
, flags
, numa_mem_id());
3130 cache_grow_end(cache
, page
);
3132 nid
= page_to_nid(page
);
3133 obj
= ____cache_alloc_node(cache
,
3134 gfp_exact_node(flags
), nid
);
3137 * Another processor may allocate the objects in
3138 * the slab since we are not holding any locks.
3145 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3151 * A interface to enable slab creation on nodeid
3153 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3157 struct kmem_cache_node
*n
;
3161 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3162 n
= get_node(cachep
, nodeid
);
3166 spin_lock(&n
->list_lock
);
3167 page
= get_first_slab(n
, false);
3171 check_spinlock_acquired_node(cachep
, nodeid
);
3173 STATS_INC_NODEALLOCS(cachep
);
3174 STATS_INC_ACTIVE(cachep
);
3175 STATS_SET_HIGH(cachep
);
3177 BUG_ON(page
->active
== cachep
->num
);
3179 obj
= slab_get_obj(cachep
, page
);
3182 fixup_slab_list(cachep
, n
, page
, &list
);
3184 spin_unlock(&n
->list_lock
);
3185 fixup_objfreelist_debug(cachep
, &list
);
3189 spin_unlock(&n
->list_lock
);
3190 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3192 /* This slab isn't counted yet so don't update free_objects */
3193 obj
= slab_get_obj(cachep
, page
);
3195 cache_grow_end(cachep
, page
);
3197 return obj
? obj
: fallback_alloc(cachep
, flags
);
3200 static __always_inline
void *
3201 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3202 unsigned long caller
)
3204 unsigned long save_flags
;
3206 int slab_node
= numa_mem_id();
3208 flags
&= gfp_allowed_mask
;
3209 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3210 if (unlikely(!cachep
))
3213 cache_alloc_debugcheck_before(cachep
, flags
);
3214 local_irq_save(save_flags
);
3216 if (nodeid
== NUMA_NO_NODE
)
3219 if (unlikely(!get_node(cachep
, nodeid
))) {
3220 /* Node not bootstrapped yet */
3221 ptr
= fallback_alloc(cachep
, flags
);
3225 if (nodeid
== slab_node
) {
3227 * Use the locally cached objects if possible.
3228 * However ____cache_alloc does not allow fallback
3229 * to other nodes. It may fail while we still have
3230 * objects on other nodes available.
3232 ptr
= ____cache_alloc(cachep
, flags
);
3236 /* ___cache_alloc_node can fall back to other nodes */
3237 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3239 local_irq_restore(save_flags
);
3240 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3242 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3243 memset(ptr
, 0, cachep
->object_size
);
3245 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3249 static __always_inline
void *
3250 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3254 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3255 objp
= alternate_node_alloc(cache
, flags
);
3259 objp
= ____cache_alloc(cache
, flags
);
3262 * We may just have run out of memory on the local node.
3263 * ____cache_alloc_node() knows how to locate memory on other nodes
3266 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3273 static __always_inline
void *
3274 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3276 return ____cache_alloc(cachep
, flags
);
3279 #endif /* CONFIG_NUMA */
3281 static __always_inline
void *
3282 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3284 unsigned long save_flags
;
3287 flags
&= gfp_allowed_mask
;
3288 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3289 if (unlikely(!cachep
))
3292 cache_alloc_debugcheck_before(cachep
, flags
);
3293 local_irq_save(save_flags
);
3294 objp
= __do_cache_alloc(cachep
, flags
);
3295 local_irq_restore(save_flags
);
3296 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3299 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3300 memset(objp
, 0, cachep
->object_size
);
3302 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3307 * Caller needs to acquire correct kmem_cache_node's list_lock
3308 * @list: List of detached free slabs should be freed by caller
3310 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3311 int nr_objects
, int node
, struct list_head
*list
)
3314 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3317 n
->free_objects
+= nr_objects
;
3319 for (i
= 0; i
< nr_objects
; i
++) {
3325 page
= virt_to_head_page(objp
);
3326 list_del(&page
->lru
);
3327 check_spinlock_acquired_node(cachep
, node
);
3328 slab_put_obj(cachep
, page
, objp
);
3329 STATS_DEC_ACTIVE(cachep
);
3331 /* fixup slab chains */
3332 if (page
->active
== 0)
3333 list_add(&page
->lru
, &n
->slabs_free
);
3335 /* Unconditionally move a slab to the end of the
3336 * partial list on free - maximum time for the
3337 * other objects to be freed, too.
3339 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3343 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3344 n
->free_objects
-= cachep
->num
;
3346 page
= list_last_entry(&n
->slabs_free
, struct page
, lru
);
3347 list_del(&page
->lru
);
3348 list_add(&page
->lru
, list
);
3352 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3355 struct kmem_cache_node
*n
;
3356 int node
= numa_mem_id();
3359 batchcount
= ac
->batchcount
;
3362 n
= get_node(cachep
, node
);
3363 spin_lock(&n
->list_lock
);
3365 struct array_cache
*shared_array
= n
->shared
;
3366 int max
= shared_array
->limit
- shared_array
->avail
;
3368 if (batchcount
> max
)
3370 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3371 ac
->entry
, sizeof(void *) * batchcount
);
3372 shared_array
->avail
+= batchcount
;
3377 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3384 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3385 BUG_ON(page
->active
);
3389 STATS_SET_FREEABLE(cachep
, i
);
3392 spin_unlock(&n
->list_lock
);
3393 slabs_destroy(cachep
, &list
);
3394 ac
->avail
-= batchcount
;
3395 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3399 * Release an obj back to its cache. If the obj has a constructed state, it must
3400 * be in this state _before_ it is released. Called with disabled ints.
3402 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3403 unsigned long caller
)
3405 struct array_cache
*ac
= cpu_cache_get(cachep
);
3407 kasan_slab_free(cachep
, objp
);
3410 kmemleak_free_recursive(objp
, cachep
->flags
);
3411 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3413 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3416 * Skip calling cache_free_alien() when the platform is not numa.
3417 * This will avoid cache misses that happen while accessing slabp (which
3418 * is per page memory reference) to get nodeid. Instead use a global
3419 * variable to skip the call, which is mostly likely to be present in
3422 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3425 if (ac
->avail
< ac
->limit
) {
3426 STATS_INC_FREEHIT(cachep
);
3428 STATS_INC_FREEMISS(cachep
);
3429 cache_flusharray(cachep
, ac
);
3432 if (sk_memalloc_socks()) {
3433 struct page
*page
= virt_to_head_page(objp
);
3435 if (unlikely(PageSlabPfmemalloc(page
))) {
3436 cache_free_pfmemalloc(cachep
, page
, objp
);
3441 ac
->entry
[ac
->avail
++] = objp
;
3445 * kmem_cache_alloc - Allocate an object
3446 * @cachep: The cache to allocate from.
3447 * @flags: See kmalloc().
3449 * Allocate an object from this cache. The flags are only relevant
3450 * if the cache has no available objects.
3452 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3454 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3456 kasan_slab_alloc(cachep
, ret
, flags
);
3457 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3458 cachep
->object_size
, cachep
->size
, flags
);
3462 EXPORT_SYMBOL(kmem_cache_alloc
);
3464 static __always_inline
void
3465 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3466 size_t size
, void **p
, unsigned long caller
)
3470 for (i
= 0; i
< size
; i
++)
3471 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3474 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3479 s
= slab_pre_alloc_hook(s
, flags
);
3483 cache_alloc_debugcheck_before(s
, flags
);
3485 local_irq_disable();
3486 for (i
= 0; i
< size
; i
++) {
3487 void *objp
= __do_cache_alloc(s
, flags
);
3489 if (unlikely(!objp
))
3495 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3497 /* Clear memory outside IRQ disabled section */
3498 if (unlikely(flags
& __GFP_ZERO
))
3499 for (i
= 0; i
< size
; i
++)
3500 memset(p
[i
], 0, s
->object_size
);
3502 slab_post_alloc_hook(s
, flags
, size
, p
);
3503 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3507 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3508 slab_post_alloc_hook(s
, flags
, i
, p
);
3509 __kmem_cache_free_bulk(s
, i
, p
);
3512 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3514 #ifdef CONFIG_TRACING
3516 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3520 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3522 kasan_kmalloc(cachep
, ret
, size
, flags
);
3523 trace_kmalloc(_RET_IP_
, ret
,
3524 size
, cachep
->size
, flags
);
3527 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3532 * kmem_cache_alloc_node - Allocate an object on the specified node
3533 * @cachep: The cache to allocate from.
3534 * @flags: See kmalloc().
3535 * @nodeid: node number of the target node.
3537 * Identical to kmem_cache_alloc but it will allocate memory on the given
3538 * node, which can improve the performance for cpu bound structures.
3540 * Fallback to other node is possible if __GFP_THISNODE is not set.
3542 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3544 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3546 kasan_slab_alloc(cachep
, ret
, flags
);
3547 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3548 cachep
->object_size
, cachep
->size
,
3553 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3555 #ifdef CONFIG_TRACING
3556 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3563 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3565 kasan_kmalloc(cachep
, ret
, size
, flags
);
3566 trace_kmalloc_node(_RET_IP_
, ret
,
3571 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3574 static __always_inline
void *
3575 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3577 struct kmem_cache
*cachep
;
3580 cachep
= kmalloc_slab(size
, flags
);
3581 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3583 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3584 kasan_kmalloc(cachep
, ret
, size
, flags
);
3589 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3591 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3593 EXPORT_SYMBOL(__kmalloc_node
);
3595 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3596 int node
, unsigned long caller
)
3598 return __do_kmalloc_node(size
, flags
, node
, caller
);
3600 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3601 #endif /* CONFIG_NUMA */
3604 * __do_kmalloc - allocate memory
3605 * @size: how many bytes of memory are required.
3606 * @flags: the type of memory to allocate (see kmalloc).
3607 * @caller: function caller for debug tracking of the caller
3609 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3610 unsigned long caller
)
3612 struct kmem_cache
*cachep
;
3615 cachep
= kmalloc_slab(size
, flags
);
3616 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3618 ret
= slab_alloc(cachep
, flags
, caller
);
3620 kasan_kmalloc(cachep
, ret
, size
, flags
);
3621 trace_kmalloc(caller
, ret
,
3622 size
, cachep
->size
, flags
);
3627 void *__kmalloc(size_t size
, gfp_t flags
)
3629 return __do_kmalloc(size
, flags
, _RET_IP_
);
3631 EXPORT_SYMBOL(__kmalloc
);
3633 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3635 return __do_kmalloc(size
, flags
, caller
);
3637 EXPORT_SYMBOL(__kmalloc_track_caller
);
3640 * kmem_cache_free - Deallocate an object
3641 * @cachep: The cache the allocation was from.
3642 * @objp: The previously allocated object.
3644 * Free an object which was previously allocated from this
3647 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3649 unsigned long flags
;
3650 cachep
= cache_from_obj(cachep
, objp
);
3654 local_irq_save(flags
);
3655 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3656 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3657 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3658 __cache_free(cachep
, objp
, _RET_IP_
);
3659 local_irq_restore(flags
);
3661 trace_kmem_cache_free(_RET_IP_
, objp
);
3663 EXPORT_SYMBOL(kmem_cache_free
);
3665 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3667 struct kmem_cache
*s
;
3670 local_irq_disable();
3671 for (i
= 0; i
< size
; i
++) {
3674 if (!orig_s
) /* called via kfree_bulk */
3675 s
= virt_to_cache(objp
);
3677 s
= cache_from_obj(orig_s
, objp
);
3679 debug_check_no_locks_freed(objp
, s
->object_size
);
3680 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3681 debug_check_no_obj_freed(objp
, s
->object_size
);
3683 __cache_free(s
, objp
, _RET_IP_
);
3687 /* FIXME: add tracing */
3689 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3692 * kfree - free previously allocated memory
3693 * @objp: pointer returned by kmalloc.
3695 * If @objp is NULL, no operation is performed.
3697 * Don't free memory not originally allocated by kmalloc()
3698 * or you will run into trouble.
3700 void kfree(const void *objp
)
3702 struct kmem_cache
*c
;
3703 unsigned long flags
;
3705 trace_kfree(_RET_IP_
, objp
);
3707 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3709 local_irq_save(flags
);
3710 kfree_debugcheck(objp
);
3711 c
= virt_to_cache(objp
);
3712 debug_check_no_locks_freed(objp
, c
->object_size
);
3714 debug_check_no_obj_freed(objp
, c
->object_size
);
3715 __cache_free(c
, (void *)objp
, _RET_IP_
);
3716 local_irq_restore(flags
);
3718 EXPORT_SYMBOL(kfree
);
3721 * This initializes kmem_cache_node or resizes various caches for all nodes.
3723 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3727 struct kmem_cache_node
*n
;
3729 for_each_online_node(node
) {
3730 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3739 if (!cachep
->list
.next
) {
3740 /* Cache is not active yet. Roll back what we did */
3743 n
= get_node(cachep
, node
);
3746 free_alien_cache(n
->alien
);
3748 cachep
->node
[node
] = NULL
;
3756 /* Always called with the slab_mutex held */
3757 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3758 int batchcount
, int shared
, gfp_t gfp
)
3760 struct array_cache __percpu
*cpu_cache
, *prev
;
3763 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3767 prev
= cachep
->cpu_cache
;
3768 cachep
->cpu_cache
= cpu_cache
;
3769 kick_all_cpus_sync();
3772 cachep
->batchcount
= batchcount
;
3773 cachep
->limit
= limit
;
3774 cachep
->shared
= shared
;
3779 for_each_online_cpu(cpu
) {
3782 struct kmem_cache_node
*n
;
3783 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3785 node
= cpu_to_mem(cpu
);
3786 n
= get_node(cachep
, node
);
3787 spin_lock_irq(&n
->list_lock
);
3788 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3789 spin_unlock_irq(&n
->list_lock
);
3790 slabs_destroy(cachep
, &list
);
3795 return setup_kmem_cache_nodes(cachep
, gfp
);
3798 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3799 int batchcount
, int shared
, gfp_t gfp
)
3802 struct kmem_cache
*c
;
3804 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3806 if (slab_state
< FULL
)
3809 if ((ret
< 0) || !is_root_cache(cachep
))
3812 lockdep_assert_held(&slab_mutex
);
3813 for_each_memcg_cache(c
, cachep
) {
3814 /* return value determined by the root cache only */
3815 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3821 /* Called with slab_mutex held always */
3822 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3829 if (!is_root_cache(cachep
)) {
3830 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3831 limit
= root
->limit
;
3832 shared
= root
->shared
;
3833 batchcount
= root
->batchcount
;
3836 if (limit
&& shared
&& batchcount
)
3839 * The head array serves three purposes:
3840 * - create a LIFO ordering, i.e. return objects that are cache-warm
3841 * - reduce the number of spinlock operations.
3842 * - reduce the number of linked list operations on the slab and
3843 * bufctl chains: array operations are cheaper.
3844 * The numbers are guessed, we should auto-tune as described by
3847 if (cachep
->size
> 131072)
3849 else if (cachep
->size
> PAGE_SIZE
)
3851 else if (cachep
->size
> 1024)
3853 else if (cachep
->size
> 256)
3859 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3860 * allocation behaviour: Most allocs on one cpu, most free operations
3861 * on another cpu. For these cases, an efficient object passing between
3862 * cpus is necessary. This is provided by a shared array. The array
3863 * replaces Bonwick's magazine layer.
3864 * On uniprocessor, it's functionally equivalent (but less efficient)
3865 * to a larger limit. Thus disabled by default.
3868 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3873 * With debugging enabled, large batchcount lead to excessively long
3874 * periods with disabled local interrupts. Limit the batchcount
3879 batchcount
= (limit
+ 1) / 2;
3881 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3883 pr_err("enable_cpucache failed for %s, error %d\n",
3884 cachep
->name
, -err
);
3889 * Drain an array if it contains any elements taking the node lock only if
3890 * necessary. Note that the node listlock also protects the array_cache
3891 * if drain_array() is used on the shared array.
3893 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3894 struct array_cache
*ac
, int node
)
3898 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3899 check_mutex_acquired();
3901 if (!ac
|| !ac
->avail
)
3909 spin_lock_irq(&n
->list_lock
);
3910 drain_array_locked(cachep
, ac
, node
, false, &list
);
3911 spin_unlock_irq(&n
->list_lock
);
3913 slabs_destroy(cachep
, &list
);
3917 * cache_reap - Reclaim memory from caches.
3918 * @w: work descriptor
3920 * Called from workqueue/eventd every few seconds.
3922 * - clear the per-cpu caches for this CPU.
3923 * - return freeable pages to the main free memory pool.
3925 * If we cannot acquire the cache chain mutex then just give up - we'll try
3926 * again on the next iteration.
3928 static void cache_reap(struct work_struct
*w
)
3930 struct kmem_cache
*searchp
;
3931 struct kmem_cache_node
*n
;
3932 int node
= numa_mem_id();
3933 struct delayed_work
*work
= to_delayed_work(w
);
3935 if (!mutex_trylock(&slab_mutex
))
3936 /* Give up. Setup the next iteration. */
3939 list_for_each_entry(searchp
, &slab_caches
, list
) {
3943 * We only take the node lock if absolutely necessary and we
3944 * have established with reasonable certainty that
3945 * we can do some work if the lock was obtained.
3947 n
= get_node(searchp
, node
);
3949 reap_alien(searchp
, n
);
3951 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3954 * These are racy checks but it does not matter
3955 * if we skip one check or scan twice.
3957 if (time_after(n
->next_reap
, jiffies
))
3960 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3962 drain_array(searchp
, n
, n
->shared
, node
);
3964 if (n
->free_touched
)
3965 n
->free_touched
= 0;
3969 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3970 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3971 STATS_ADD_REAPED(searchp
, freed
);
3977 mutex_unlock(&slab_mutex
);
3980 /* Set up the next iteration */
3981 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3984 #ifdef CONFIG_SLABINFO
3985 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3988 unsigned long active_objs
;
3989 unsigned long num_objs
;
3990 unsigned long active_slabs
= 0;
3991 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
3995 struct kmem_cache_node
*n
;
3999 for_each_kmem_cache_node(cachep
, node
, n
) {
4002 spin_lock_irq(&n
->list_lock
);
4004 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4005 if (page
->active
!= cachep
->num
&& !error
)
4006 error
= "slabs_full accounting error";
4007 active_objs
+= cachep
->num
;
4010 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4011 if (page
->active
== cachep
->num
&& !error
)
4012 error
= "slabs_partial accounting error";
4013 if (!page
->active
&& !error
)
4014 error
= "slabs_partial accounting error";
4015 active_objs
+= page
->active
;
4018 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4019 if (page
->active
&& !error
)
4020 error
= "slabs_free accounting error";
4023 free_objects
+= n
->free_objects
;
4025 shared_avail
+= n
->shared
->avail
;
4027 spin_unlock_irq(&n
->list_lock
);
4029 num_slabs
+= active_slabs
;
4030 num_objs
= num_slabs
* cachep
->num
;
4031 if (num_objs
- active_objs
!= free_objects
&& !error
)
4032 error
= "free_objects accounting error";
4034 name
= cachep
->name
;
4036 pr_err("slab: cache %s error: %s\n", name
, error
);
4038 sinfo
->active_objs
= active_objs
;
4039 sinfo
->num_objs
= num_objs
;
4040 sinfo
->active_slabs
= active_slabs
;
4041 sinfo
->num_slabs
= num_slabs
;
4042 sinfo
->shared_avail
= shared_avail
;
4043 sinfo
->limit
= cachep
->limit
;
4044 sinfo
->batchcount
= cachep
->batchcount
;
4045 sinfo
->shared
= cachep
->shared
;
4046 sinfo
->objects_per_slab
= cachep
->num
;
4047 sinfo
->cache_order
= cachep
->gfporder
;
4050 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4054 unsigned long high
= cachep
->high_mark
;
4055 unsigned long allocs
= cachep
->num_allocations
;
4056 unsigned long grown
= cachep
->grown
;
4057 unsigned long reaped
= cachep
->reaped
;
4058 unsigned long errors
= cachep
->errors
;
4059 unsigned long max_freeable
= cachep
->max_freeable
;
4060 unsigned long node_allocs
= cachep
->node_allocs
;
4061 unsigned long node_frees
= cachep
->node_frees
;
4062 unsigned long overflows
= cachep
->node_overflow
;
4064 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4065 allocs
, high
, grown
,
4066 reaped
, errors
, max_freeable
, node_allocs
,
4067 node_frees
, overflows
);
4071 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4072 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4073 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4074 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4076 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4077 allochit
, allocmiss
, freehit
, freemiss
);
4082 #define MAX_SLABINFO_WRITE 128
4084 * slabinfo_write - Tuning for the slab allocator
4086 * @buffer: user buffer
4087 * @count: data length
4090 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4091 size_t count
, loff_t
*ppos
)
4093 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4094 int limit
, batchcount
, shared
, res
;
4095 struct kmem_cache
*cachep
;
4097 if (count
> MAX_SLABINFO_WRITE
)
4099 if (copy_from_user(&kbuf
, buffer
, count
))
4101 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4103 tmp
= strchr(kbuf
, ' ');
4108 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4111 /* Find the cache in the chain of caches. */
4112 mutex_lock(&slab_mutex
);
4114 list_for_each_entry(cachep
, &slab_caches
, list
) {
4115 if (!strcmp(cachep
->name
, kbuf
)) {
4116 if (limit
< 1 || batchcount
< 1 ||
4117 batchcount
> limit
|| shared
< 0) {
4120 res
= do_tune_cpucache(cachep
, limit
,
4127 mutex_unlock(&slab_mutex
);
4133 #ifdef CONFIG_DEBUG_SLAB_LEAK
4135 static inline int add_caller(unsigned long *n
, unsigned long v
)
4145 unsigned long *q
= p
+ 2 * i
;
4159 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4165 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4174 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4177 for (j
= page
->active
; j
< c
->num
; j
++) {
4178 if (get_free_obj(page
, j
) == i
) {
4188 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4189 * mapping is established when actual object allocation and
4190 * we could mistakenly access the unmapped object in the cpu
4193 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4196 if (!add_caller(n
, v
))
4201 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4203 #ifdef CONFIG_KALLSYMS
4204 unsigned long offset
, size
;
4205 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4207 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4208 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4210 seq_printf(m
, " [%s]", modname
);
4214 seq_printf(m
, "%p", (void *)address
);
4217 static int leaks_show(struct seq_file
*m
, void *p
)
4219 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4221 struct kmem_cache_node
*n
;
4223 unsigned long *x
= m
->private;
4227 if (!(cachep
->flags
& SLAB_STORE_USER
))
4229 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4233 * Set store_user_clean and start to grab stored user information
4234 * for all objects on this cache. If some alloc/free requests comes
4235 * during the processing, information would be wrong so restart
4239 set_store_user_clean(cachep
);
4240 drain_cpu_caches(cachep
);
4244 for_each_kmem_cache_node(cachep
, node
, n
) {
4247 spin_lock_irq(&n
->list_lock
);
4249 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4250 handle_slab(x
, cachep
, page
);
4251 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4252 handle_slab(x
, cachep
, page
);
4253 spin_unlock_irq(&n
->list_lock
);
4255 } while (!is_store_user_clean(cachep
));
4257 name
= cachep
->name
;
4259 /* Increase the buffer size */
4260 mutex_unlock(&slab_mutex
);
4261 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4263 /* Too bad, we are really out */
4265 mutex_lock(&slab_mutex
);
4268 *(unsigned long *)m
->private = x
[0] * 2;
4270 mutex_lock(&slab_mutex
);
4271 /* Now make sure this entry will be retried */
4275 for (i
= 0; i
< x
[1]; i
++) {
4276 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4277 show_symbol(m
, x
[2*i
+2]);
4284 static const struct seq_operations slabstats_op
= {
4285 .start
= slab_start
,
4291 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4295 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4299 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4304 static const struct file_operations proc_slabstats_operations
= {
4305 .open
= slabstats_open
,
4307 .llseek
= seq_lseek
,
4308 .release
= seq_release_private
,
4312 static int __init
slab_proc_init(void)
4314 #ifdef CONFIG_DEBUG_SLAB_LEAK
4315 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4319 module_init(slab_proc_init
);
4323 * ksize - get the actual amount of memory allocated for a given object
4324 * @objp: Pointer to the object
4326 * kmalloc may internally round up allocations and return more memory
4327 * than requested. ksize() can be used to determine the actual amount of
4328 * memory allocated. The caller may use this additional memory, even though
4329 * a smaller amount of memory was initially specified with the kmalloc call.
4330 * The caller must guarantee that objp points to a valid object previously
4331 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4332 * must not be freed during the duration of the call.
4334 size_t ksize(const void *objp
)
4339 if (unlikely(objp
== ZERO_SIZE_PTR
))
4342 size
= virt_to_cache(objp
)->object_size
;
4343 /* We assume that ksize callers could use the whole allocated area,
4344 * so we need to unpoison this area.
4346 kasan_krealloc(objp
, size
, GFP_NOWAIT
);
4350 EXPORT_SYMBOL(ksize
);