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
);
969 * To protect lockless access to n->shared during irq disabled context.
970 * If n->shared isn't NULL in irq disabled context, accessing to it is
971 * guaranteed to be valid until irq is re-enabled, because it will be
972 * freed after synchronize_sched().
980 free_alien_cache(new_alien
);
985 static void cpuup_canceled(long cpu
)
987 struct kmem_cache
*cachep
;
988 struct kmem_cache_node
*n
= NULL
;
989 int node
= cpu_to_mem(cpu
);
990 const struct cpumask
*mask
= cpumask_of_node(node
);
992 list_for_each_entry(cachep
, &slab_caches
, list
) {
993 struct array_cache
*nc
;
994 struct array_cache
*shared
;
995 struct alien_cache
**alien
;
998 n
= get_node(cachep
, node
);
1002 spin_lock_irq(&n
->list_lock
);
1004 /* Free limit for this kmem_cache_node */
1005 n
->free_limit
-= cachep
->batchcount
;
1007 /* cpu is dead; no one can alloc from it. */
1008 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
1010 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1014 if (!cpumask_empty(mask
)) {
1015 spin_unlock_irq(&n
->list_lock
);
1021 free_block(cachep
, shared
->entry
,
1022 shared
->avail
, node
, &list
);
1029 spin_unlock_irq(&n
->list_lock
);
1033 drain_alien_cache(cachep
, alien
);
1034 free_alien_cache(alien
);
1038 slabs_destroy(cachep
, &list
);
1041 * In the previous loop, all the objects were freed to
1042 * the respective cache's slabs, now we can go ahead and
1043 * shrink each nodelist to its limit.
1045 list_for_each_entry(cachep
, &slab_caches
, list
) {
1046 n
= get_node(cachep
, node
);
1049 drain_freelist(cachep
, n
, INT_MAX
);
1053 static int cpuup_prepare(long cpu
)
1055 struct kmem_cache
*cachep
;
1056 int node
= cpu_to_mem(cpu
);
1060 * We need to do this right in the beginning since
1061 * alloc_arraycache's are going to use this list.
1062 * kmalloc_node allows us to add the slab to the right
1063 * kmem_cache_node and not this cpu's kmem_cache_node
1065 err
= init_cache_node_node(node
);
1070 * Now we can go ahead with allocating the shared arrays and
1073 list_for_each_entry(cachep
, &slab_caches
, list
) {
1074 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1081 cpuup_canceled(cpu
);
1085 static int cpuup_callback(struct notifier_block
*nfb
,
1086 unsigned long action
, void *hcpu
)
1088 long cpu
= (long)hcpu
;
1092 case CPU_UP_PREPARE
:
1093 case CPU_UP_PREPARE_FROZEN
:
1094 mutex_lock(&slab_mutex
);
1095 err
= cpuup_prepare(cpu
);
1096 mutex_unlock(&slab_mutex
);
1099 case CPU_ONLINE_FROZEN
:
1100 start_cpu_timer(cpu
);
1102 #ifdef CONFIG_HOTPLUG_CPU
1103 case CPU_DOWN_PREPARE
:
1104 case CPU_DOWN_PREPARE_FROZEN
:
1106 * Shutdown cache reaper. Note that the slab_mutex is
1107 * held so that if cache_reap() is invoked it cannot do
1108 * anything expensive but will only modify reap_work
1109 * and reschedule the timer.
1111 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1112 /* Now the cache_reaper is guaranteed to be not running. */
1113 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1115 case CPU_DOWN_FAILED
:
1116 case CPU_DOWN_FAILED_FROZEN
:
1117 start_cpu_timer(cpu
);
1120 case CPU_DEAD_FROZEN
:
1122 * Even if all the cpus of a node are down, we don't free the
1123 * kmem_cache_node of any cache. This to avoid a race between
1124 * cpu_down, and a kmalloc allocation from another cpu for
1125 * memory from the node of the cpu going down. The node
1126 * structure is usually allocated from kmem_cache_create() and
1127 * gets destroyed at kmem_cache_destroy().
1131 case CPU_UP_CANCELED
:
1132 case CPU_UP_CANCELED_FROZEN
:
1133 mutex_lock(&slab_mutex
);
1134 cpuup_canceled(cpu
);
1135 mutex_unlock(&slab_mutex
);
1138 return notifier_from_errno(err
);
1141 static struct notifier_block cpucache_notifier
= {
1142 &cpuup_callback
, NULL
, 0
1145 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1147 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1148 * Returns -EBUSY if all objects cannot be drained so that the node is not
1151 * Must hold slab_mutex.
1153 static int __meminit
drain_cache_node_node(int node
)
1155 struct kmem_cache
*cachep
;
1158 list_for_each_entry(cachep
, &slab_caches
, list
) {
1159 struct kmem_cache_node
*n
;
1161 n
= get_node(cachep
, node
);
1165 drain_freelist(cachep
, n
, INT_MAX
);
1167 if (!list_empty(&n
->slabs_full
) ||
1168 !list_empty(&n
->slabs_partial
)) {
1176 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1177 unsigned long action
, void *arg
)
1179 struct memory_notify
*mnb
= arg
;
1183 nid
= mnb
->status_change_nid
;
1188 case MEM_GOING_ONLINE
:
1189 mutex_lock(&slab_mutex
);
1190 ret
= init_cache_node_node(nid
);
1191 mutex_unlock(&slab_mutex
);
1193 case MEM_GOING_OFFLINE
:
1194 mutex_lock(&slab_mutex
);
1195 ret
= drain_cache_node_node(nid
);
1196 mutex_unlock(&slab_mutex
);
1200 case MEM_CANCEL_ONLINE
:
1201 case MEM_CANCEL_OFFLINE
:
1205 return notifier_from_errno(ret
);
1207 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1210 * swap the static kmem_cache_node with kmalloced memory
1212 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1215 struct kmem_cache_node
*ptr
;
1217 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1220 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1222 * Do not assume that spinlocks can be initialized via memcpy:
1224 spin_lock_init(&ptr
->list_lock
);
1226 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1227 cachep
->node
[nodeid
] = ptr
;
1231 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1232 * size of kmem_cache_node.
1234 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1238 for_each_online_node(node
) {
1239 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1240 cachep
->node
[node
]->next_reap
= jiffies
+
1242 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1247 * Initialisation. Called after the page allocator have been initialised and
1248 * before smp_init().
1250 void __init
kmem_cache_init(void)
1254 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1255 sizeof(struct rcu_head
));
1256 kmem_cache
= &kmem_cache_boot
;
1258 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1259 use_alien_caches
= 0;
1261 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1262 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1265 * Fragmentation resistance on low memory - only use bigger
1266 * page orders on machines with more than 32MB of memory if
1267 * not overridden on the command line.
1269 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1270 slab_max_order
= SLAB_MAX_ORDER_HI
;
1272 /* Bootstrap is tricky, because several objects are allocated
1273 * from caches that do not exist yet:
1274 * 1) initialize the kmem_cache cache: it contains the struct
1275 * kmem_cache structures of all caches, except kmem_cache itself:
1276 * kmem_cache is statically allocated.
1277 * Initially an __init data area is used for the head array and the
1278 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1279 * array at the end of the bootstrap.
1280 * 2) Create the first kmalloc cache.
1281 * The struct kmem_cache for the new cache is allocated normally.
1282 * An __init data area is used for the head array.
1283 * 3) Create the remaining kmalloc caches, with minimally sized
1285 * 4) Replace the __init data head arrays for kmem_cache and the first
1286 * kmalloc cache with kmalloc allocated arrays.
1287 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1288 * the other cache's with kmalloc allocated memory.
1289 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1292 /* 1) create the kmem_cache */
1295 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1297 create_boot_cache(kmem_cache
, "kmem_cache",
1298 offsetof(struct kmem_cache
, node
) +
1299 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1300 SLAB_HWCACHE_ALIGN
);
1301 list_add(&kmem_cache
->list
, &slab_caches
);
1302 slab_state
= PARTIAL
;
1305 * Initialize the caches that provide memory for the kmem_cache_node
1306 * structures first. Without this, further allocations will bug.
1308 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache("kmalloc-node",
1309 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1310 slab_state
= PARTIAL_NODE
;
1311 setup_kmalloc_cache_index_table();
1313 slab_early_init
= 0;
1315 /* 5) Replace the bootstrap kmem_cache_node */
1319 for_each_online_node(nid
) {
1320 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1322 init_list(kmalloc_caches
[INDEX_NODE
],
1323 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1327 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1330 void __init
kmem_cache_init_late(void)
1332 struct kmem_cache
*cachep
;
1336 /* 6) resize the head arrays to their final sizes */
1337 mutex_lock(&slab_mutex
);
1338 list_for_each_entry(cachep
, &slab_caches
, list
)
1339 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1341 mutex_unlock(&slab_mutex
);
1347 * Register a cpu startup notifier callback that initializes
1348 * cpu_cache_get for all new cpus
1350 register_cpu_notifier(&cpucache_notifier
);
1354 * Register a memory hotplug callback that initializes and frees
1357 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1361 * The reap timers are started later, with a module init call: That part
1362 * of the kernel is not yet operational.
1366 static int __init
cpucache_init(void)
1371 * Register the timers that return unneeded pages to the page allocator
1373 for_each_online_cpu(cpu
)
1374 start_cpu_timer(cpu
);
1380 __initcall(cpucache_init
);
1382 static noinline
void
1383 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1386 struct kmem_cache_node
*n
;
1388 unsigned long flags
;
1390 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1391 DEFAULT_RATELIMIT_BURST
);
1393 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1396 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1397 nodeid
, gfpflags
, &gfpflags
);
1398 pr_warn(" cache: %s, object size: %d, order: %d\n",
1399 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1401 for_each_kmem_cache_node(cachep
, node
, n
) {
1402 unsigned long active_objs
= 0, num_objs
= 0, free_objects
= 0;
1403 unsigned long active_slabs
= 0, num_slabs
= 0;
1405 spin_lock_irqsave(&n
->list_lock
, flags
);
1406 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
1407 active_objs
+= cachep
->num
;
1410 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
1411 active_objs
+= page
->active
;
1414 list_for_each_entry(page
, &n
->slabs_free
, lru
)
1417 free_objects
+= n
->free_objects
;
1418 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1420 num_slabs
+= active_slabs
;
1421 num_objs
= num_slabs
* cachep
->num
;
1422 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1423 node
, active_slabs
, num_slabs
, active_objs
, num_objs
,
1430 * Interface to system's page allocator. No need to hold the
1431 * kmem_cache_node ->list_lock.
1433 * If we requested dmaable memory, we will get it. Even if we
1434 * did not request dmaable memory, we might get it, but that
1435 * would be relatively rare and ignorable.
1437 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1443 flags
|= cachep
->allocflags
;
1444 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1445 flags
|= __GFP_RECLAIMABLE
;
1447 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1449 slab_out_of_memory(cachep
, flags
, nodeid
);
1453 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1454 __free_pages(page
, cachep
->gfporder
);
1458 nr_pages
= (1 << cachep
->gfporder
);
1459 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1460 add_zone_page_state(page_zone(page
),
1461 NR_SLAB_RECLAIMABLE
, nr_pages
);
1463 add_zone_page_state(page_zone(page
),
1464 NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1466 __SetPageSlab(page
);
1467 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1468 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1469 SetPageSlabPfmemalloc(page
);
1471 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1472 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1475 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1477 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1484 * Interface to system's page release.
1486 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1488 int order
= cachep
->gfporder
;
1489 unsigned long nr_freed
= (1 << order
);
1491 kmemcheck_free_shadow(page
, order
);
1493 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1494 sub_zone_page_state(page_zone(page
),
1495 NR_SLAB_RECLAIMABLE
, nr_freed
);
1497 sub_zone_page_state(page_zone(page
),
1498 NR_SLAB_UNRECLAIMABLE
, nr_freed
);
1500 BUG_ON(!PageSlab(page
));
1501 __ClearPageSlabPfmemalloc(page
);
1502 __ClearPageSlab(page
);
1503 page_mapcount_reset(page
);
1504 page
->mapping
= NULL
;
1506 if (current
->reclaim_state
)
1507 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1508 memcg_uncharge_slab(page
, order
, cachep
);
1509 __free_pages(page
, order
);
1512 static void kmem_rcu_free(struct rcu_head
*head
)
1514 struct kmem_cache
*cachep
;
1517 page
= container_of(head
, struct page
, rcu_head
);
1518 cachep
= page
->slab_cache
;
1520 kmem_freepages(cachep
, page
);
1524 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1526 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1527 (cachep
->size
% PAGE_SIZE
) == 0)
1533 #ifdef CONFIG_DEBUG_PAGEALLOC
1534 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1535 unsigned long caller
)
1537 int size
= cachep
->object_size
;
1539 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1541 if (size
< 5 * sizeof(unsigned long))
1544 *addr
++ = 0x12345678;
1546 *addr
++ = smp_processor_id();
1547 size
-= 3 * sizeof(unsigned long);
1549 unsigned long *sptr
= &caller
;
1550 unsigned long svalue
;
1552 while (!kstack_end(sptr
)) {
1554 if (kernel_text_address(svalue
)) {
1556 size
-= sizeof(unsigned long);
1557 if (size
<= sizeof(unsigned long))
1563 *addr
++ = 0x87654321;
1566 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1567 int map
, unsigned long caller
)
1569 if (!is_debug_pagealloc_cache(cachep
))
1573 store_stackinfo(cachep
, objp
, caller
);
1575 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1579 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1580 int map
, unsigned long caller
) {}
1584 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1586 int size
= cachep
->object_size
;
1587 addr
= &((char *)addr
)[obj_offset(cachep
)];
1589 memset(addr
, val
, size
);
1590 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1593 static void dump_line(char *data
, int offset
, int limit
)
1596 unsigned char error
= 0;
1599 pr_err("%03x: ", offset
);
1600 for (i
= 0; i
< limit
; i
++) {
1601 if (data
[offset
+ i
] != POISON_FREE
) {
1602 error
= data
[offset
+ i
];
1606 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1607 &data
[offset
], limit
, 1);
1609 if (bad_count
== 1) {
1610 error
^= POISON_FREE
;
1611 if (!(error
& (error
- 1))) {
1612 pr_err("Single bit error detected. Probably bad RAM.\n");
1614 pr_err("Run memtest86+ or a similar memory test tool.\n");
1616 pr_err("Run a memory test tool.\n");
1625 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1630 if (cachep
->flags
& SLAB_RED_ZONE
) {
1631 pr_err("Redzone: 0x%llx/0x%llx\n",
1632 *dbg_redzone1(cachep
, objp
),
1633 *dbg_redzone2(cachep
, objp
));
1636 if (cachep
->flags
& SLAB_STORE_USER
) {
1637 pr_err("Last user: [<%p>](%pSR)\n",
1638 *dbg_userword(cachep
, objp
),
1639 *dbg_userword(cachep
, objp
));
1641 realobj
= (char *)objp
+ obj_offset(cachep
);
1642 size
= cachep
->object_size
;
1643 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1646 if (i
+ limit
> size
)
1648 dump_line(realobj
, i
, limit
);
1652 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1658 if (is_debug_pagealloc_cache(cachep
))
1661 realobj
= (char *)objp
+ obj_offset(cachep
);
1662 size
= cachep
->object_size
;
1664 for (i
= 0; i
< size
; i
++) {
1665 char exp
= POISON_FREE
;
1668 if (realobj
[i
] != exp
) {
1673 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1674 print_tainted(), cachep
->name
,
1676 print_objinfo(cachep
, objp
, 0);
1678 /* Hexdump the affected line */
1681 if (i
+ limit
> size
)
1683 dump_line(realobj
, i
, limit
);
1686 /* Limit to 5 lines */
1692 /* Print some data about the neighboring objects, if they
1695 struct page
*page
= virt_to_head_page(objp
);
1698 objnr
= obj_to_index(cachep
, page
, objp
);
1700 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1701 realobj
= (char *)objp
+ obj_offset(cachep
);
1702 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1703 print_objinfo(cachep
, objp
, 2);
1705 if (objnr
+ 1 < cachep
->num
) {
1706 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1707 realobj
= (char *)objp
+ obj_offset(cachep
);
1708 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1709 print_objinfo(cachep
, objp
, 2);
1716 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1721 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1722 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1726 for (i
= 0; i
< cachep
->num
; i
++) {
1727 void *objp
= index_to_obj(cachep
, page
, i
);
1729 if (cachep
->flags
& SLAB_POISON
) {
1730 check_poison_obj(cachep
, objp
);
1731 slab_kernel_map(cachep
, objp
, 1, 0);
1733 if (cachep
->flags
& SLAB_RED_ZONE
) {
1734 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1735 slab_error(cachep
, "start of a freed object was overwritten");
1736 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1737 slab_error(cachep
, "end of a freed object was overwritten");
1742 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1749 * slab_destroy - destroy and release all objects in a slab
1750 * @cachep: cache pointer being destroyed
1751 * @page: page pointer being destroyed
1753 * Destroy all the objs in a slab page, and release the mem back to the system.
1754 * Before calling the slab page must have been unlinked from the cache. The
1755 * kmem_cache_node ->list_lock is not held/needed.
1757 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1761 freelist
= page
->freelist
;
1762 slab_destroy_debugcheck(cachep
, page
);
1763 if (unlikely(cachep
->flags
& SLAB_DESTROY_BY_RCU
))
1764 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1766 kmem_freepages(cachep
, page
);
1769 * From now on, we don't use freelist
1770 * although actual page can be freed in rcu context
1772 if (OFF_SLAB(cachep
))
1773 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1776 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1778 struct page
*page
, *n
;
1780 list_for_each_entry_safe(page
, n
, list
, lru
) {
1781 list_del(&page
->lru
);
1782 slab_destroy(cachep
, page
);
1787 * calculate_slab_order - calculate size (page order) of slabs
1788 * @cachep: pointer to the cache that is being created
1789 * @size: size of objects to be created in this cache.
1790 * @flags: slab allocation flags
1792 * Also calculates the number of objects per slab.
1794 * This could be made much more intelligent. For now, try to avoid using
1795 * high order pages for slabs. When the gfp() functions are more friendly
1796 * towards high-order requests, this should be changed.
1798 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1799 size_t size
, unsigned long flags
)
1801 size_t left_over
= 0;
1804 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1808 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1812 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1813 if (num
> SLAB_OBJ_MAX_NUM
)
1816 if (flags
& CFLGS_OFF_SLAB
) {
1817 struct kmem_cache
*freelist_cache
;
1818 size_t freelist_size
;
1820 freelist_size
= num
* sizeof(freelist_idx_t
);
1821 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1822 if (!freelist_cache
)
1826 * Needed to avoid possible looping condition
1827 * in cache_grow_begin()
1829 if (OFF_SLAB(freelist_cache
))
1832 /* check if off slab has enough benefit */
1833 if (freelist_cache
->size
> cachep
->size
/ 2)
1837 /* Found something acceptable - save it away */
1839 cachep
->gfporder
= gfporder
;
1840 left_over
= remainder
;
1843 * A VFS-reclaimable slab tends to have most allocations
1844 * as GFP_NOFS and we really don't want to have to be allocating
1845 * higher-order pages when we are unable to shrink dcache.
1847 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1851 * Large number of objects is good, but very large slabs are
1852 * currently bad for the gfp()s.
1854 if (gfporder
>= slab_max_order
)
1858 * Acceptable internal fragmentation?
1860 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1866 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1867 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1871 struct array_cache __percpu
*cpu_cache
;
1873 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1874 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1879 for_each_possible_cpu(cpu
) {
1880 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1881 entries
, batchcount
);
1887 static int __init_refok
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1889 if (slab_state
>= FULL
)
1890 return enable_cpucache(cachep
, gfp
);
1892 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1893 if (!cachep
->cpu_cache
)
1896 if (slab_state
== DOWN
) {
1897 /* Creation of first cache (kmem_cache). */
1898 set_up_node(kmem_cache
, CACHE_CACHE
);
1899 } else if (slab_state
== PARTIAL
) {
1900 /* For kmem_cache_node */
1901 set_up_node(cachep
, SIZE_NODE
);
1905 for_each_online_node(node
) {
1906 cachep
->node
[node
] = kmalloc_node(
1907 sizeof(struct kmem_cache_node
), gfp
, node
);
1908 BUG_ON(!cachep
->node
[node
]);
1909 kmem_cache_node_init(cachep
->node
[node
]);
1913 cachep
->node
[numa_mem_id()]->next_reap
=
1914 jiffies
+ REAPTIMEOUT_NODE
+
1915 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1917 cpu_cache_get(cachep
)->avail
= 0;
1918 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1919 cpu_cache_get(cachep
)->batchcount
= 1;
1920 cpu_cache_get(cachep
)->touched
= 0;
1921 cachep
->batchcount
= 1;
1922 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1926 unsigned long kmem_cache_flags(unsigned long object_size
,
1927 unsigned long flags
, const char *name
,
1928 void (*ctor
)(void *))
1934 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1935 unsigned long flags
, void (*ctor
)(void *))
1937 struct kmem_cache
*cachep
;
1939 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1944 * Adjust the object sizes so that we clear
1945 * the complete object on kzalloc.
1947 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1952 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1953 size_t size
, unsigned long flags
)
1959 if (cachep
->ctor
|| flags
& SLAB_DESTROY_BY_RCU
)
1962 left
= calculate_slab_order(cachep
, size
,
1963 flags
| CFLGS_OBJFREELIST_SLAB
);
1967 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1970 cachep
->colour
= left
/ cachep
->colour_off
;
1975 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1976 size_t size
, unsigned long flags
)
1983 * Always use on-slab management when SLAB_NOLEAKTRACE
1984 * to avoid recursive calls into kmemleak.
1986 if (flags
& SLAB_NOLEAKTRACE
)
1990 * Size is large, assume best to place the slab management obj
1991 * off-slab (should allow better packing of objs).
1993 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1998 * If the slab has been placed off-slab, and we have enough space then
1999 * move it on-slab. This is at the expense of any extra colouring.
2001 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
2004 cachep
->colour
= left
/ cachep
->colour_off
;
2009 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
2010 size_t size
, unsigned long flags
)
2016 left
= calculate_slab_order(cachep
, size
, flags
);
2020 cachep
->colour
= left
/ cachep
->colour_off
;
2026 * __kmem_cache_create - Create a cache.
2027 * @cachep: cache management descriptor
2028 * @flags: SLAB flags
2030 * Returns a ptr to the cache on success, NULL on failure.
2031 * Cannot be called within a int, but can be interrupted.
2032 * The @ctor is run when new pages are allocated by the cache.
2036 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2037 * to catch references to uninitialised memory.
2039 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2040 * for buffer overruns.
2042 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2043 * cacheline. This can be beneficial if you're counting cycles as closely
2047 __kmem_cache_create (struct kmem_cache
*cachep
, unsigned long flags
)
2049 size_t ralign
= BYTES_PER_WORD
;
2052 size_t size
= cachep
->size
;
2057 * Enable redzoning and last user accounting, except for caches with
2058 * large objects, if the increased size would increase the object size
2059 * above the next power of two: caches with object sizes just above a
2060 * power of two have a significant amount of internal fragmentation.
2062 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2063 2 * sizeof(unsigned long long)))
2064 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2065 if (!(flags
& SLAB_DESTROY_BY_RCU
))
2066 flags
|= SLAB_POISON
;
2071 * Check that size is in terms of words. This is needed to avoid
2072 * unaligned accesses for some archs when redzoning is used, and makes
2073 * sure any on-slab bufctl's are also correctly aligned.
2075 if (size
& (BYTES_PER_WORD
- 1)) {
2076 size
+= (BYTES_PER_WORD
- 1);
2077 size
&= ~(BYTES_PER_WORD
- 1);
2080 if (flags
& SLAB_RED_ZONE
) {
2081 ralign
= REDZONE_ALIGN
;
2082 /* If redzoning, ensure that the second redzone is suitably
2083 * aligned, by adjusting the object size accordingly. */
2084 size
+= REDZONE_ALIGN
- 1;
2085 size
&= ~(REDZONE_ALIGN
- 1);
2088 /* 3) caller mandated alignment */
2089 if (ralign
< cachep
->align
) {
2090 ralign
= cachep
->align
;
2092 /* disable debug if necessary */
2093 if (ralign
> __alignof__(unsigned long long))
2094 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2098 cachep
->align
= ralign
;
2099 cachep
->colour_off
= cache_line_size();
2100 /* Offset must be a multiple of the alignment. */
2101 if (cachep
->colour_off
< cachep
->align
)
2102 cachep
->colour_off
= cachep
->align
;
2104 if (slab_is_available())
2112 * Both debugging options require word-alignment which is calculated
2115 if (flags
& SLAB_RED_ZONE
) {
2116 /* add space for red zone words */
2117 cachep
->obj_offset
+= sizeof(unsigned long long);
2118 size
+= 2 * sizeof(unsigned long long);
2120 if (flags
& SLAB_STORE_USER
) {
2121 /* user store requires one word storage behind the end of
2122 * the real object. But if the second red zone needs to be
2123 * aligned to 64 bits, we must allow that much space.
2125 if (flags
& SLAB_RED_ZONE
)
2126 size
+= REDZONE_ALIGN
;
2128 size
+= BYTES_PER_WORD
;
2132 kasan_cache_create(cachep
, &size
, &flags
);
2134 size
= ALIGN(size
, cachep
->align
);
2136 * We should restrict the number of objects in a slab to implement
2137 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2139 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2140 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2144 * To activate debug pagealloc, off-slab management is necessary
2145 * requirement. In early phase of initialization, small sized slab
2146 * doesn't get initialized so it would not be possible. So, we need
2147 * to check size >= 256. It guarantees that all necessary small
2148 * sized slab is initialized in current slab initialization sequence.
2150 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2151 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2152 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2153 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2155 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2156 flags
|= CFLGS_OFF_SLAB
;
2157 cachep
->obj_offset
+= tmp_size
- size
;
2165 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2166 flags
|= CFLGS_OBJFREELIST_SLAB
;
2170 if (set_off_slab_cache(cachep
, size
, flags
)) {
2171 flags
|= CFLGS_OFF_SLAB
;
2175 if (set_on_slab_cache(cachep
, size
, flags
))
2181 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2182 cachep
->flags
= flags
;
2183 cachep
->allocflags
= __GFP_COMP
;
2184 if (CONFIG_ZONE_DMA_FLAG
&& (flags
& SLAB_CACHE_DMA
))
2185 cachep
->allocflags
|= GFP_DMA
;
2186 cachep
->size
= size
;
2187 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2191 * If we're going to use the generic kernel_map_pages()
2192 * poisoning, then it's going to smash the contents of
2193 * the redzone and userword anyhow, so switch them off.
2195 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2196 (cachep
->flags
& SLAB_POISON
) &&
2197 is_debug_pagealloc_cache(cachep
))
2198 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2201 if (OFF_SLAB(cachep
)) {
2202 cachep
->freelist_cache
=
2203 kmalloc_slab(cachep
->freelist_size
, 0u);
2206 err
= setup_cpu_cache(cachep
, gfp
);
2208 __kmem_cache_release(cachep
);
2216 static void check_irq_off(void)
2218 BUG_ON(!irqs_disabled());
2221 static void check_irq_on(void)
2223 BUG_ON(irqs_disabled());
2226 static void check_mutex_acquired(void)
2228 BUG_ON(!mutex_is_locked(&slab_mutex
));
2231 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2235 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2239 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2243 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2248 #define check_irq_off() do { } while(0)
2249 #define check_irq_on() do { } while(0)
2250 #define check_mutex_acquired() do { } while(0)
2251 #define check_spinlock_acquired(x) do { } while(0)
2252 #define check_spinlock_acquired_node(x, y) do { } while(0)
2255 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2256 int node
, bool free_all
, struct list_head
*list
)
2260 if (!ac
|| !ac
->avail
)
2263 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2264 if (tofree
> ac
->avail
)
2265 tofree
= (ac
->avail
+ 1) / 2;
2267 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2268 ac
->avail
-= tofree
;
2269 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2272 static void do_drain(void *arg
)
2274 struct kmem_cache
*cachep
= arg
;
2275 struct array_cache
*ac
;
2276 int node
= numa_mem_id();
2277 struct kmem_cache_node
*n
;
2281 ac
= cpu_cache_get(cachep
);
2282 n
= get_node(cachep
, node
);
2283 spin_lock(&n
->list_lock
);
2284 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2285 spin_unlock(&n
->list_lock
);
2286 slabs_destroy(cachep
, &list
);
2290 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2292 struct kmem_cache_node
*n
;
2296 on_each_cpu(do_drain
, cachep
, 1);
2298 for_each_kmem_cache_node(cachep
, node
, n
)
2300 drain_alien_cache(cachep
, n
->alien
);
2302 for_each_kmem_cache_node(cachep
, node
, n
) {
2303 spin_lock_irq(&n
->list_lock
);
2304 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2305 spin_unlock_irq(&n
->list_lock
);
2307 slabs_destroy(cachep
, &list
);
2312 * Remove slabs from the list of free slabs.
2313 * Specify the number of slabs to drain in tofree.
2315 * Returns the actual number of slabs released.
2317 static int drain_freelist(struct kmem_cache
*cache
,
2318 struct kmem_cache_node
*n
, int tofree
)
2320 struct list_head
*p
;
2325 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2327 spin_lock_irq(&n
->list_lock
);
2328 p
= n
->slabs_free
.prev
;
2329 if (p
== &n
->slabs_free
) {
2330 spin_unlock_irq(&n
->list_lock
);
2334 page
= list_entry(p
, struct page
, lru
);
2335 list_del(&page
->lru
);
2337 * Safe to drop the lock. The slab is no longer linked
2340 n
->free_objects
-= cache
->num
;
2341 spin_unlock_irq(&n
->list_lock
);
2342 slab_destroy(cache
, page
);
2349 int __kmem_cache_shrink(struct kmem_cache
*cachep
, bool deactivate
)
2353 struct kmem_cache_node
*n
;
2355 drain_cpu_caches(cachep
);
2358 for_each_kmem_cache_node(cachep
, node
, n
) {
2359 drain_freelist(cachep
, n
, INT_MAX
);
2361 ret
+= !list_empty(&n
->slabs_full
) ||
2362 !list_empty(&n
->slabs_partial
);
2364 return (ret
? 1 : 0);
2367 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2369 return __kmem_cache_shrink(cachep
, false);
2372 void __kmem_cache_release(struct kmem_cache
*cachep
)
2375 struct kmem_cache_node
*n
;
2377 free_percpu(cachep
->cpu_cache
);
2379 /* NUMA: free the node structures */
2380 for_each_kmem_cache_node(cachep
, i
, n
) {
2382 free_alien_cache(n
->alien
);
2384 cachep
->node
[i
] = NULL
;
2389 * Get the memory for a slab management obj.
2391 * For a slab cache when the slab descriptor is off-slab, the
2392 * slab descriptor can't come from the same cache which is being created,
2393 * Because if it is the case, that means we defer the creation of
2394 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2395 * And we eventually call down to __kmem_cache_create(), which
2396 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2397 * This is a "chicken-and-egg" problem.
2399 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2400 * which are all initialized during kmem_cache_init().
2402 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2403 struct page
*page
, int colour_off
,
2404 gfp_t local_flags
, int nodeid
)
2407 void *addr
= page_address(page
);
2409 page
->s_mem
= addr
+ colour_off
;
2412 if (OBJFREELIST_SLAB(cachep
))
2414 else if (OFF_SLAB(cachep
)) {
2415 /* Slab management obj is off-slab. */
2416 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2417 local_flags
, nodeid
);
2421 /* We will use last bytes at the slab for freelist */
2422 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2423 cachep
->freelist_size
;
2429 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2431 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2434 static inline void set_free_obj(struct page
*page
,
2435 unsigned int idx
, freelist_idx_t val
)
2437 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2440 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2445 for (i
= 0; i
< cachep
->num
; i
++) {
2446 void *objp
= index_to_obj(cachep
, page
, i
);
2448 if (cachep
->flags
& SLAB_STORE_USER
)
2449 *dbg_userword(cachep
, objp
) = NULL
;
2451 if (cachep
->flags
& SLAB_RED_ZONE
) {
2452 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2453 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2456 * Constructors are not allowed to allocate memory from the same
2457 * cache which they are a constructor for. Otherwise, deadlock.
2458 * They must also be threaded.
2460 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2461 kasan_unpoison_object_data(cachep
,
2462 objp
+ obj_offset(cachep
));
2463 cachep
->ctor(objp
+ obj_offset(cachep
));
2464 kasan_poison_object_data(
2465 cachep
, objp
+ obj_offset(cachep
));
2468 if (cachep
->flags
& SLAB_RED_ZONE
) {
2469 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2470 slab_error(cachep
, "constructor overwrote the end of an object");
2471 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2472 slab_error(cachep
, "constructor overwrote the start of an object");
2474 /* need to poison the objs? */
2475 if (cachep
->flags
& SLAB_POISON
) {
2476 poison_obj(cachep
, objp
, POISON_FREE
);
2477 slab_kernel_map(cachep
, objp
, 0, 0);
2483 static void cache_init_objs(struct kmem_cache
*cachep
,
2489 cache_init_objs_debug(cachep
, page
);
2491 if (OBJFREELIST_SLAB(cachep
)) {
2492 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2496 for (i
= 0; i
< cachep
->num
; i
++) {
2497 /* constructor could break poison info */
2498 if (DEBUG
== 0 && cachep
->ctor
) {
2499 objp
= index_to_obj(cachep
, page
, i
);
2500 kasan_unpoison_object_data(cachep
, objp
);
2502 kasan_poison_object_data(cachep
, objp
);
2505 set_free_obj(page
, i
, i
);
2509 static void kmem_flagcheck(struct kmem_cache
*cachep
, gfp_t flags
)
2511 if (CONFIG_ZONE_DMA_FLAG
) {
2512 if (flags
& GFP_DMA
)
2513 BUG_ON(!(cachep
->allocflags
& GFP_DMA
));
2515 BUG_ON(cachep
->allocflags
& GFP_DMA
);
2519 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2523 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2527 if (cachep
->flags
& SLAB_STORE_USER
)
2528 set_store_user_dirty(cachep
);
2534 static void slab_put_obj(struct kmem_cache
*cachep
,
2535 struct page
*page
, void *objp
)
2537 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2541 /* Verify double free bug */
2542 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2543 if (get_free_obj(page
, i
) == objnr
) {
2544 pr_err("slab: double free detected in cache '%s', objp %p\n",
2545 cachep
->name
, objp
);
2551 if (!page
->freelist
)
2552 page
->freelist
= objp
+ obj_offset(cachep
);
2554 set_free_obj(page
, page
->active
, objnr
);
2558 * Map pages beginning at addr to the given cache and slab. This is required
2559 * for the slab allocator to be able to lookup the cache and slab of a
2560 * virtual address for kfree, ksize, and slab debugging.
2562 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2565 page
->slab_cache
= cache
;
2566 page
->freelist
= freelist
;
2570 * Grow (by 1) the number of slabs within a cache. This is called by
2571 * kmem_cache_alloc() when there are no active objs left in a cache.
2573 static struct page
*cache_grow_begin(struct kmem_cache
*cachep
,
2574 gfp_t flags
, int nodeid
)
2580 struct kmem_cache_node
*n
;
2584 * Be lazy and only check for valid flags here, keeping it out of the
2585 * critical path in kmem_cache_alloc().
2587 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2588 pr_emerg("gfp: %u\n", flags
& GFP_SLAB_BUG_MASK
);
2591 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2594 if (gfpflags_allow_blocking(local_flags
))
2598 * The test for missing atomic flag is performed here, rather than
2599 * the more obvious place, simply to reduce the critical path length
2600 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2601 * will eventually be caught here (where it matters).
2603 kmem_flagcheck(cachep
, flags
);
2606 * Get mem for the objs. Attempt to allocate a physical page from
2609 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2613 page_node
= page_to_nid(page
);
2614 n
= get_node(cachep
, page_node
);
2616 /* Get colour for the slab, and cal the next value. */
2618 if (n
->colour_next
>= cachep
->colour
)
2621 offset
= n
->colour_next
;
2622 if (offset
>= cachep
->colour
)
2625 offset
*= cachep
->colour_off
;
2627 /* Get slab management. */
2628 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2629 local_flags
& ~GFP_CONSTRAINT_MASK
, page_node
);
2630 if (OFF_SLAB(cachep
) && !freelist
)
2633 slab_map_pages(cachep
, page
, freelist
);
2635 kasan_poison_slab(page
);
2636 cache_init_objs(cachep
, page
);
2638 if (gfpflags_allow_blocking(local_flags
))
2639 local_irq_disable();
2644 kmem_freepages(cachep
, page
);
2646 if (gfpflags_allow_blocking(local_flags
))
2647 local_irq_disable();
2651 static void cache_grow_end(struct kmem_cache
*cachep
, struct page
*page
)
2653 struct kmem_cache_node
*n
;
2661 INIT_LIST_HEAD(&page
->lru
);
2662 n
= get_node(cachep
, page_to_nid(page
));
2664 spin_lock(&n
->list_lock
);
2666 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2668 fixup_slab_list(cachep
, n
, page
, &list
);
2669 STATS_INC_GROWN(cachep
);
2670 n
->free_objects
+= cachep
->num
- page
->active
;
2671 spin_unlock(&n
->list_lock
);
2673 fixup_objfreelist_debug(cachep
, &list
);
2679 * Perform extra freeing checks:
2680 * - detect bad pointers.
2681 * - POISON/RED_ZONE checking
2683 static void kfree_debugcheck(const void *objp
)
2685 if (!virt_addr_valid(objp
)) {
2686 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2687 (unsigned long)objp
);
2692 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2694 unsigned long long redzone1
, redzone2
;
2696 redzone1
= *dbg_redzone1(cache
, obj
);
2697 redzone2
= *dbg_redzone2(cache
, obj
);
2702 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2705 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2706 slab_error(cache
, "double free detected");
2708 slab_error(cache
, "memory outside object was overwritten");
2710 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2711 obj
, redzone1
, redzone2
);
2714 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2715 unsigned long caller
)
2720 BUG_ON(virt_to_cache(objp
) != cachep
);
2722 objp
-= obj_offset(cachep
);
2723 kfree_debugcheck(objp
);
2724 page
= virt_to_head_page(objp
);
2726 if (cachep
->flags
& SLAB_RED_ZONE
) {
2727 verify_redzone_free(cachep
, objp
);
2728 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2729 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2731 if (cachep
->flags
& SLAB_STORE_USER
) {
2732 set_store_user_dirty(cachep
);
2733 *dbg_userword(cachep
, objp
) = (void *)caller
;
2736 objnr
= obj_to_index(cachep
, page
, objp
);
2738 BUG_ON(objnr
>= cachep
->num
);
2739 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2741 if (cachep
->flags
& SLAB_POISON
) {
2742 poison_obj(cachep
, objp
, POISON_FREE
);
2743 slab_kernel_map(cachep
, objp
, 0, caller
);
2749 #define kfree_debugcheck(x) do { } while(0)
2750 #define cache_free_debugcheck(x,objp,z) (objp)
2753 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2761 objp
= next
- obj_offset(cachep
);
2762 next
= *(void **)next
;
2763 poison_obj(cachep
, objp
, POISON_FREE
);
2768 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2769 struct kmem_cache_node
*n
, struct page
*page
,
2772 /* move slabp to correct slabp list: */
2773 list_del(&page
->lru
);
2774 if (page
->active
== cachep
->num
) {
2775 list_add(&page
->lru
, &n
->slabs_full
);
2776 if (OBJFREELIST_SLAB(cachep
)) {
2778 /* Poisoning will be done without holding the lock */
2779 if (cachep
->flags
& SLAB_POISON
) {
2780 void **objp
= page
->freelist
;
2786 page
->freelist
= NULL
;
2789 list_add(&page
->lru
, &n
->slabs_partial
);
2792 /* Try to find non-pfmemalloc slab if needed */
2793 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2794 struct page
*page
, bool pfmemalloc
)
2802 if (!PageSlabPfmemalloc(page
))
2805 /* No need to keep pfmemalloc slab if we have enough free objects */
2806 if (n
->free_objects
> n
->free_limit
) {
2807 ClearPageSlabPfmemalloc(page
);
2811 /* Move pfmemalloc slab to the end of list to speed up next search */
2812 list_del(&page
->lru
);
2814 list_add_tail(&page
->lru
, &n
->slabs_free
);
2816 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2818 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2819 if (!PageSlabPfmemalloc(page
))
2823 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2824 if (!PageSlabPfmemalloc(page
))
2831 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2835 page
= list_first_entry_or_null(&n
->slabs_partial
,
2838 n
->free_touched
= 1;
2839 page
= list_first_entry_or_null(&n
->slabs_free
,
2843 if (sk_memalloc_socks())
2844 return get_valid_first_slab(n
, page
, pfmemalloc
);
2849 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2850 struct kmem_cache_node
*n
, gfp_t flags
)
2856 if (!gfp_pfmemalloc_allowed(flags
))
2859 spin_lock(&n
->list_lock
);
2860 page
= get_first_slab(n
, true);
2862 spin_unlock(&n
->list_lock
);
2866 obj
= slab_get_obj(cachep
, page
);
2869 fixup_slab_list(cachep
, n
, page
, &list
);
2871 spin_unlock(&n
->list_lock
);
2872 fixup_objfreelist_debug(cachep
, &list
);
2878 * Slab list should be fixed up by fixup_slab_list() for existing slab
2879 * or cache_grow_end() for new slab
2881 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2882 struct array_cache
*ac
, struct page
*page
, int batchcount
)
2885 * There must be at least one object available for
2888 BUG_ON(page
->active
>= cachep
->num
);
2890 while (page
->active
< cachep
->num
&& batchcount
--) {
2891 STATS_INC_ALLOCED(cachep
);
2892 STATS_INC_ACTIVE(cachep
);
2893 STATS_SET_HIGH(cachep
);
2895 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2901 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2904 struct kmem_cache_node
*n
;
2905 struct array_cache
*ac
, *shared
;
2911 node
= numa_mem_id();
2913 ac
= cpu_cache_get(cachep
);
2914 batchcount
= ac
->batchcount
;
2915 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2917 * If there was little recent activity on this cache, then
2918 * perform only a partial refill. Otherwise we could generate
2921 batchcount
= BATCHREFILL_LIMIT
;
2923 n
= get_node(cachep
, node
);
2925 BUG_ON(ac
->avail
> 0 || !n
);
2926 shared
= READ_ONCE(n
->shared
);
2927 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
2930 spin_lock(&n
->list_lock
);
2931 shared
= READ_ONCE(n
->shared
);
2933 /* See if we can refill from the shared array */
2934 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
2935 shared
->touched
= 1;
2939 while (batchcount
> 0) {
2940 /* Get slab alloc is to come from. */
2941 page
= get_first_slab(n
, false);
2945 check_spinlock_acquired(cachep
);
2947 batchcount
= alloc_block(cachep
, ac
, page
, batchcount
);
2948 fixup_slab_list(cachep
, n
, page
, &list
);
2952 n
->free_objects
-= ac
->avail
;
2954 spin_unlock(&n
->list_lock
);
2955 fixup_objfreelist_debug(cachep
, &list
);
2958 if (unlikely(!ac
->avail
)) {
2959 /* Check if we can use obj in pfmemalloc slab */
2960 if (sk_memalloc_socks()) {
2961 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2967 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
2970 * cache_grow_begin() can reenable interrupts,
2971 * then ac could change.
2973 ac
= cpu_cache_get(cachep
);
2974 if (!ac
->avail
&& page
)
2975 alloc_block(cachep
, ac
, page
, batchcount
);
2976 cache_grow_end(cachep
, page
);
2983 return ac
->entry
[--ac
->avail
];
2986 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2989 might_sleep_if(gfpflags_allow_blocking(flags
));
2991 kmem_flagcheck(cachep
, flags
);
2996 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2997 gfp_t flags
, void *objp
, unsigned long caller
)
3001 if (cachep
->flags
& SLAB_POISON
) {
3002 check_poison_obj(cachep
, objp
);
3003 slab_kernel_map(cachep
, objp
, 1, 0);
3004 poison_obj(cachep
, objp
, POISON_INUSE
);
3006 if (cachep
->flags
& SLAB_STORE_USER
)
3007 *dbg_userword(cachep
, objp
) = (void *)caller
;
3009 if (cachep
->flags
& SLAB_RED_ZONE
) {
3010 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3011 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3012 slab_error(cachep
, "double free, or memory outside object was overwritten");
3013 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3014 objp
, *dbg_redzone1(cachep
, objp
),
3015 *dbg_redzone2(cachep
, objp
));
3017 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3018 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3021 objp
+= obj_offset(cachep
);
3022 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3024 if (ARCH_SLAB_MINALIGN
&&
3025 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3026 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3027 objp
, (int)ARCH_SLAB_MINALIGN
);
3032 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3035 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3038 struct array_cache
*ac
;
3042 ac
= cpu_cache_get(cachep
);
3043 if (likely(ac
->avail
)) {
3045 objp
= ac
->entry
[--ac
->avail
];
3047 STATS_INC_ALLOCHIT(cachep
);
3051 STATS_INC_ALLOCMISS(cachep
);
3052 objp
= cache_alloc_refill(cachep
, flags
);
3054 * the 'ac' may be updated by cache_alloc_refill(),
3055 * and kmemleak_erase() requires its correct value.
3057 ac
= cpu_cache_get(cachep
);
3061 * To avoid a false negative, if an object that is in one of the
3062 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3063 * treat the array pointers as a reference to the object.
3066 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3072 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3074 * If we are in_interrupt, then process context, including cpusets and
3075 * mempolicy, may not apply and should not be used for allocation policy.
3077 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3079 int nid_alloc
, nid_here
;
3081 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3083 nid_alloc
= nid_here
= numa_mem_id();
3084 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3085 nid_alloc
= cpuset_slab_spread_node();
3086 else if (current
->mempolicy
)
3087 nid_alloc
= mempolicy_slab_node();
3088 if (nid_alloc
!= nid_here
)
3089 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3094 * Fallback function if there was no memory available and no objects on a
3095 * certain node and fall back is permitted. First we scan all the
3096 * available node for available objects. If that fails then we
3097 * perform an allocation without specifying a node. This allows the page
3098 * allocator to do its reclaim / fallback magic. We then insert the
3099 * slab into the proper nodelist and then allocate from it.
3101 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3103 struct zonelist
*zonelist
;
3106 enum zone_type high_zoneidx
= gfp_zone(flags
);
3110 unsigned int cpuset_mems_cookie
;
3112 if (flags
& __GFP_THISNODE
)
3116 cpuset_mems_cookie
= read_mems_allowed_begin();
3117 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3121 * Look through allowed nodes for objects available
3122 * from existing per node queues.
3124 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3125 nid
= zone_to_nid(zone
);
3127 if (cpuset_zone_allowed(zone
, flags
) &&
3128 get_node(cache
, nid
) &&
3129 get_node(cache
, nid
)->free_objects
) {
3130 obj
= ____cache_alloc_node(cache
,
3131 gfp_exact_node(flags
), nid
);
3139 * This allocation will be performed within the constraints
3140 * of the current cpuset / memory policy requirements.
3141 * We may trigger various forms of reclaim on the allowed
3142 * set and go into memory reserves if necessary.
3144 page
= cache_grow_begin(cache
, flags
, numa_mem_id());
3145 cache_grow_end(cache
, page
);
3147 nid
= page_to_nid(page
);
3148 obj
= ____cache_alloc_node(cache
,
3149 gfp_exact_node(flags
), nid
);
3152 * Another processor may allocate the objects in
3153 * the slab since we are not holding any locks.
3160 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3166 * A interface to enable slab creation on nodeid
3168 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3172 struct kmem_cache_node
*n
;
3176 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3177 n
= get_node(cachep
, nodeid
);
3181 spin_lock(&n
->list_lock
);
3182 page
= get_first_slab(n
, false);
3186 check_spinlock_acquired_node(cachep
, nodeid
);
3188 STATS_INC_NODEALLOCS(cachep
);
3189 STATS_INC_ACTIVE(cachep
);
3190 STATS_SET_HIGH(cachep
);
3192 BUG_ON(page
->active
== cachep
->num
);
3194 obj
= slab_get_obj(cachep
, page
);
3197 fixup_slab_list(cachep
, n
, page
, &list
);
3199 spin_unlock(&n
->list_lock
);
3200 fixup_objfreelist_debug(cachep
, &list
);
3204 spin_unlock(&n
->list_lock
);
3205 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3207 /* This slab isn't counted yet so don't update free_objects */
3208 obj
= slab_get_obj(cachep
, page
);
3210 cache_grow_end(cachep
, page
);
3212 return obj
? obj
: fallback_alloc(cachep
, flags
);
3215 static __always_inline
void *
3216 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3217 unsigned long caller
)
3219 unsigned long save_flags
;
3221 int slab_node
= numa_mem_id();
3223 flags
&= gfp_allowed_mask
;
3224 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3225 if (unlikely(!cachep
))
3228 cache_alloc_debugcheck_before(cachep
, flags
);
3229 local_irq_save(save_flags
);
3231 if (nodeid
== NUMA_NO_NODE
)
3234 if (unlikely(!get_node(cachep
, nodeid
))) {
3235 /* Node not bootstrapped yet */
3236 ptr
= fallback_alloc(cachep
, flags
);
3240 if (nodeid
== slab_node
) {
3242 * Use the locally cached objects if possible.
3243 * However ____cache_alloc does not allow fallback
3244 * to other nodes. It may fail while we still have
3245 * objects on other nodes available.
3247 ptr
= ____cache_alloc(cachep
, flags
);
3251 /* ___cache_alloc_node can fall back to other nodes */
3252 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3254 local_irq_restore(save_flags
);
3255 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3257 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3258 memset(ptr
, 0, cachep
->object_size
);
3260 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3264 static __always_inline
void *
3265 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3269 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3270 objp
= alternate_node_alloc(cache
, flags
);
3274 objp
= ____cache_alloc(cache
, flags
);
3277 * We may just have run out of memory on the local node.
3278 * ____cache_alloc_node() knows how to locate memory on other nodes
3281 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3288 static __always_inline
void *
3289 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3291 return ____cache_alloc(cachep
, flags
);
3294 #endif /* CONFIG_NUMA */
3296 static __always_inline
void *
3297 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3299 unsigned long save_flags
;
3302 flags
&= gfp_allowed_mask
;
3303 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3304 if (unlikely(!cachep
))
3307 cache_alloc_debugcheck_before(cachep
, flags
);
3308 local_irq_save(save_flags
);
3309 objp
= __do_cache_alloc(cachep
, flags
);
3310 local_irq_restore(save_flags
);
3311 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3314 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3315 memset(objp
, 0, cachep
->object_size
);
3317 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3322 * Caller needs to acquire correct kmem_cache_node's list_lock
3323 * @list: List of detached free slabs should be freed by caller
3325 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3326 int nr_objects
, int node
, struct list_head
*list
)
3329 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3332 n
->free_objects
+= nr_objects
;
3334 for (i
= 0; i
< nr_objects
; i
++) {
3340 page
= virt_to_head_page(objp
);
3341 list_del(&page
->lru
);
3342 check_spinlock_acquired_node(cachep
, node
);
3343 slab_put_obj(cachep
, page
, objp
);
3344 STATS_DEC_ACTIVE(cachep
);
3346 /* fixup slab chains */
3347 if (page
->active
== 0)
3348 list_add(&page
->lru
, &n
->slabs_free
);
3350 /* Unconditionally move a slab to the end of the
3351 * partial list on free - maximum time for the
3352 * other objects to be freed, too.
3354 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3358 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3359 n
->free_objects
-= cachep
->num
;
3361 page
= list_last_entry(&n
->slabs_free
, struct page
, lru
);
3362 list_del(&page
->lru
);
3363 list_add(&page
->lru
, list
);
3367 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3370 struct kmem_cache_node
*n
;
3371 int node
= numa_mem_id();
3374 batchcount
= ac
->batchcount
;
3377 n
= get_node(cachep
, node
);
3378 spin_lock(&n
->list_lock
);
3380 struct array_cache
*shared_array
= n
->shared
;
3381 int max
= shared_array
->limit
- shared_array
->avail
;
3383 if (batchcount
> max
)
3385 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3386 ac
->entry
, sizeof(void *) * batchcount
);
3387 shared_array
->avail
+= batchcount
;
3392 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3399 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3400 BUG_ON(page
->active
);
3404 STATS_SET_FREEABLE(cachep
, i
);
3407 spin_unlock(&n
->list_lock
);
3408 slabs_destroy(cachep
, &list
);
3409 ac
->avail
-= batchcount
;
3410 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3414 * Release an obj back to its cache. If the obj has a constructed state, it must
3415 * be in this state _before_ it is released. Called with disabled ints.
3417 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3418 unsigned long caller
)
3420 struct array_cache
*ac
= cpu_cache_get(cachep
);
3422 kasan_slab_free(cachep
, objp
);
3425 kmemleak_free_recursive(objp
, cachep
->flags
);
3426 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3428 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3431 * Skip calling cache_free_alien() when the platform is not numa.
3432 * This will avoid cache misses that happen while accessing slabp (which
3433 * is per page memory reference) to get nodeid. Instead use a global
3434 * variable to skip the call, which is mostly likely to be present in
3437 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3440 if (ac
->avail
< ac
->limit
) {
3441 STATS_INC_FREEHIT(cachep
);
3443 STATS_INC_FREEMISS(cachep
);
3444 cache_flusharray(cachep
, ac
);
3447 if (sk_memalloc_socks()) {
3448 struct page
*page
= virt_to_head_page(objp
);
3450 if (unlikely(PageSlabPfmemalloc(page
))) {
3451 cache_free_pfmemalloc(cachep
, page
, objp
);
3456 ac
->entry
[ac
->avail
++] = objp
;
3460 * kmem_cache_alloc - Allocate an object
3461 * @cachep: The cache to allocate from.
3462 * @flags: See kmalloc().
3464 * Allocate an object from this cache. The flags are only relevant
3465 * if the cache has no available objects.
3467 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3469 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3471 kasan_slab_alloc(cachep
, ret
, flags
);
3472 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3473 cachep
->object_size
, cachep
->size
, flags
);
3477 EXPORT_SYMBOL(kmem_cache_alloc
);
3479 static __always_inline
void
3480 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3481 size_t size
, void **p
, unsigned long caller
)
3485 for (i
= 0; i
< size
; i
++)
3486 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3489 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3494 s
= slab_pre_alloc_hook(s
, flags
);
3498 cache_alloc_debugcheck_before(s
, flags
);
3500 local_irq_disable();
3501 for (i
= 0; i
< size
; i
++) {
3502 void *objp
= __do_cache_alloc(s
, flags
);
3504 if (unlikely(!objp
))
3510 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3512 /* Clear memory outside IRQ disabled section */
3513 if (unlikely(flags
& __GFP_ZERO
))
3514 for (i
= 0; i
< size
; i
++)
3515 memset(p
[i
], 0, s
->object_size
);
3517 slab_post_alloc_hook(s
, flags
, size
, p
);
3518 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3522 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3523 slab_post_alloc_hook(s
, flags
, i
, p
);
3524 __kmem_cache_free_bulk(s
, i
, p
);
3527 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3529 #ifdef CONFIG_TRACING
3531 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3535 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3537 kasan_kmalloc(cachep
, ret
, size
, flags
);
3538 trace_kmalloc(_RET_IP_
, ret
,
3539 size
, cachep
->size
, flags
);
3542 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3547 * kmem_cache_alloc_node - Allocate an object on the specified node
3548 * @cachep: The cache to allocate from.
3549 * @flags: See kmalloc().
3550 * @nodeid: node number of the target node.
3552 * Identical to kmem_cache_alloc but it will allocate memory on the given
3553 * node, which can improve the performance for cpu bound structures.
3555 * Fallback to other node is possible if __GFP_THISNODE is not set.
3557 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3559 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3561 kasan_slab_alloc(cachep
, ret
, flags
);
3562 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3563 cachep
->object_size
, cachep
->size
,
3568 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3570 #ifdef CONFIG_TRACING
3571 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3578 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3580 kasan_kmalloc(cachep
, ret
, size
, flags
);
3581 trace_kmalloc_node(_RET_IP_
, ret
,
3586 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3589 static __always_inline
void *
3590 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3592 struct kmem_cache
*cachep
;
3595 cachep
= kmalloc_slab(size
, flags
);
3596 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3598 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3599 kasan_kmalloc(cachep
, ret
, size
, flags
);
3604 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3606 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3608 EXPORT_SYMBOL(__kmalloc_node
);
3610 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3611 int node
, unsigned long caller
)
3613 return __do_kmalloc_node(size
, flags
, node
, caller
);
3615 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3616 #endif /* CONFIG_NUMA */
3619 * __do_kmalloc - allocate memory
3620 * @size: how many bytes of memory are required.
3621 * @flags: the type of memory to allocate (see kmalloc).
3622 * @caller: function caller for debug tracking of the caller
3624 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3625 unsigned long caller
)
3627 struct kmem_cache
*cachep
;
3630 cachep
= kmalloc_slab(size
, flags
);
3631 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3633 ret
= slab_alloc(cachep
, flags
, caller
);
3635 kasan_kmalloc(cachep
, ret
, size
, flags
);
3636 trace_kmalloc(caller
, ret
,
3637 size
, cachep
->size
, flags
);
3642 void *__kmalloc(size_t size
, gfp_t flags
)
3644 return __do_kmalloc(size
, flags
, _RET_IP_
);
3646 EXPORT_SYMBOL(__kmalloc
);
3648 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3650 return __do_kmalloc(size
, flags
, caller
);
3652 EXPORT_SYMBOL(__kmalloc_track_caller
);
3655 * kmem_cache_free - Deallocate an object
3656 * @cachep: The cache the allocation was from.
3657 * @objp: The previously allocated object.
3659 * Free an object which was previously allocated from this
3662 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3664 unsigned long flags
;
3665 cachep
= cache_from_obj(cachep
, objp
);
3669 local_irq_save(flags
);
3670 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3671 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3672 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3673 __cache_free(cachep
, objp
, _RET_IP_
);
3674 local_irq_restore(flags
);
3676 trace_kmem_cache_free(_RET_IP_
, objp
);
3678 EXPORT_SYMBOL(kmem_cache_free
);
3680 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3682 struct kmem_cache
*s
;
3685 local_irq_disable();
3686 for (i
= 0; i
< size
; i
++) {
3689 if (!orig_s
) /* called via kfree_bulk */
3690 s
= virt_to_cache(objp
);
3692 s
= cache_from_obj(orig_s
, objp
);
3694 debug_check_no_locks_freed(objp
, s
->object_size
);
3695 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3696 debug_check_no_obj_freed(objp
, s
->object_size
);
3698 __cache_free(s
, objp
, _RET_IP_
);
3702 /* FIXME: add tracing */
3704 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3707 * kfree - free previously allocated memory
3708 * @objp: pointer returned by kmalloc.
3710 * If @objp is NULL, no operation is performed.
3712 * Don't free memory not originally allocated by kmalloc()
3713 * or you will run into trouble.
3715 void kfree(const void *objp
)
3717 struct kmem_cache
*c
;
3718 unsigned long flags
;
3720 trace_kfree(_RET_IP_
, objp
);
3722 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3724 local_irq_save(flags
);
3725 kfree_debugcheck(objp
);
3726 c
= virt_to_cache(objp
);
3727 debug_check_no_locks_freed(objp
, c
->object_size
);
3729 debug_check_no_obj_freed(objp
, c
->object_size
);
3730 __cache_free(c
, (void *)objp
, _RET_IP_
);
3731 local_irq_restore(flags
);
3733 EXPORT_SYMBOL(kfree
);
3736 * This initializes kmem_cache_node or resizes various caches for all nodes.
3738 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3742 struct kmem_cache_node
*n
;
3744 for_each_online_node(node
) {
3745 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3754 if (!cachep
->list
.next
) {
3755 /* Cache is not active yet. Roll back what we did */
3758 n
= get_node(cachep
, node
);
3761 free_alien_cache(n
->alien
);
3763 cachep
->node
[node
] = NULL
;
3771 /* Always called with the slab_mutex held */
3772 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3773 int batchcount
, int shared
, gfp_t gfp
)
3775 struct array_cache __percpu
*cpu_cache
, *prev
;
3778 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3782 prev
= cachep
->cpu_cache
;
3783 cachep
->cpu_cache
= cpu_cache
;
3784 kick_all_cpus_sync();
3787 cachep
->batchcount
= batchcount
;
3788 cachep
->limit
= limit
;
3789 cachep
->shared
= shared
;
3794 for_each_online_cpu(cpu
) {
3797 struct kmem_cache_node
*n
;
3798 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3800 node
= cpu_to_mem(cpu
);
3801 n
= get_node(cachep
, node
);
3802 spin_lock_irq(&n
->list_lock
);
3803 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3804 spin_unlock_irq(&n
->list_lock
);
3805 slabs_destroy(cachep
, &list
);
3810 return setup_kmem_cache_nodes(cachep
, gfp
);
3813 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3814 int batchcount
, int shared
, gfp_t gfp
)
3817 struct kmem_cache
*c
;
3819 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3821 if (slab_state
< FULL
)
3824 if ((ret
< 0) || !is_root_cache(cachep
))
3827 lockdep_assert_held(&slab_mutex
);
3828 for_each_memcg_cache(c
, cachep
) {
3829 /* return value determined by the root cache only */
3830 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3836 /* Called with slab_mutex held always */
3837 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3844 if (!is_root_cache(cachep
)) {
3845 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3846 limit
= root
->limit
;
3847 shared
= root
->shared
;
3848 batchcount
= root
->batchcount
;
3851 if (limit
&& shared
&& batchcount
)
3854 * The head array serves three purposes:
3855 * - create a LIFO ordering, i.e. return objects that are cache-warm
3856 * - reduce the number of spinlock operations.
3857 * - reduce the number of linked list operations on the slab and
3858 * bufctl chains: array operations are cheaper.
3859 * The numbers are guessed, we should auto-tune as described by
3862 if (cachep
->size
> 131072)
3864 else if (cachep
->size
> PAGE_SIZE
)
3866 else if (cachep
->size
> 1024)
3868 else if (cachep
->size
> 256)
3874 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3875 * allocation behaviour: Most allocs on one cpu, most free operations
3876 * on another cpu. For these cases, an efficient object passing between
3877 * cpus is necessary. This is provided by a shared array. The array
3878 * replaces Bonwick's magazine layer.
3879 * On uniprocessor, it's functionally equivalent (but less efficient)
3880 * to a larger limit. Thus disabled by default.
3883 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3888 * With debugging enabled, large batchcount lead to excessively long
3889 * periods with disabled local interrupts. Limit the batchcount
3894 batchcount
= (limit
+ 1) / 2;
3896 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3898 pr_err("enable_cpucache failed for %s, error %d\n",
3899 cachep
->name
, -err
);
3904 * Drain an array if it contains any elements taking the node lock only if
3905 * necessary. Note that the node listlock also protects the array_cache
3906 * if drain_array() is used on the shared array.
3908 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3909 struct array_cache
*ac
, int node
)
3913 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3914 check_mutex_acquired();
3916 if (!ac
|| !ac
->avail
)
3924 spin_lock_irq(&n
->list_lock
);
3925 drain_array_locked(cachep
, ac
, node
, false, &list
);
3926 spin_unlock_irq(&n
->list_lock
);
3928 slabs_destroy(cachep
, &list
);
3932 * cache_reap - Reclaim memory from caches.
3933 * @w: work descriptor
3935 * Called from workqueue/eventd every few seconds.
3937 * - clear the per-cpu caches for this CPU.
3938 * - return freeable pages to the main free memory pool.
3940 * If we cannot acquire the cache chain mutex then just give up - we'll try
3941 * again on the next iteration.
3943 static void cache_reap(struct work_struct
*w
)
3945 struct kmem_cache
*searchp
;
3946 struct kmem_cache_node
*n
;
3947 int node
= numa_mem_id();
3948 struct delayed_work
*work
= to_delayed_work(w
);
3950 if (!mutex_trylock(&slab_mutex
))
3951 /* Give up. Setup the next iteration. */
3954 list_for_each_entry(searchp
, &slab_caches
, list
) {
3958 * We only take the node lock if absolutely necessary and we
3959 * have established with reasonable certainty that
3960 * we can do some work if the lock was obtained.
3962 n
= get_node(searchp
, node
);
3964 reap_alien(searchp
, n
);
3966 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3969 * These are racy checks but it does not matter
3970 * if we skip one check or scan twice.
3972 if (time_after(n
->next_reap
, jiffies
))
3975 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3977 drain_array(searchp
, n
, n
->shared
, node
);
3979 if (n
->free_touched
)
3980 n
->free_touched
= 0;
3984 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3985 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3986 STATS_ADD_REAPED(searchp
, freed
);
3992 mutex_unlock(&slab_mutex
);
3995 /* Set up the next iteration */
3996 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
3999 #ifdef CONFIG_SLABINFO
4000 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4003 unsigned long active_objs
;
4004 unsigned long num_objs
;
4005 unsigned long active_slabs
= 0;
4006 unsigned long num_slabs
, free_objects
= 0, shared_avail
= 0;
4010 struct kmem_cache_node
*n
;
4014 for_each_kmem_cache_node(cachep
, node
, n
) {
4017 spin_lock_irq(&n
->list_lock
);
4019 list_for_each_entry(page
, &n
->slabs_full
, lru
) {
4020 if (page
->active
!= cachep
->num
&& !error
)
4021 error
= "slabs_full accounting error";
4022 active_objs
+= cachep
->num
;
4025 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
4026 if (page
->active
== cachep
->num
&& !error
)
4027 error
= "slabs_partial accounting error";
4028 if (!page
->active
&& !error
)
4029 error
= "slabs_partial accounting error";
4030 active_objs
+= page
->active
;
4033 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
4034 if (page
->active
&& !error
)
4035 error
= "slabs_free accounting error";
4038 free_objects
+= n
->free_objects
;
4040 shared_avail
+= n
->shared
->avail
;
4042 spin_unlock_irq(&n
->list_lock
);
4044 num_slabs
+= active_slabs
;
4045 num_objs
= num_slabs
* cachep
->num
;
4046 if (num_objs
- active_objs
!= free_objects
&& !error
)
4047 error
= "free_objects accounting error";
4049 name
= cachep
->name
;
4051 pr_err("slab: cache %s error: %s\n", name
, error
);
4053 sinfo
->active_objs
= active_objs
;
4054 sinfo
->num_objs
= num_objs
;
4055 sinfo
->active_slabs
= active_slabs
;
4056 sinfo
->num_slabs
= num_slabs
;
4057 sinfo
->shared_avail
= shared_avail
;
4058 sinfo
->limit
= cachep
->limit
;
4059 sinfo
->batchcount
= cachep
->batchcount
;
4060 sinfo
->shared
= cachep
->shared
;
4061 sinfo
->objects_per_slab
= cachep
->num
;
4062 sinfo
->cache_order
= cachep
->gfporder
;
4065 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4069 unsigned long high
= cachep
->high_mark
;
4070 unsigned long allocs
= cachep
->num_allocations
;
4071 unsigned long grown
= cachep
->grown
;
4072 unsigned long reaped
= cachep
->reaped
;
4073 unsigned long errors
= cachep
->errors
;
4074 unsigned long max_freeable
= cachep
->max_freeable
;
4075 unsigned long node_allocs
= cachep
->node_allocs
;
4076 unsigned long node_frees
= cachep
->node_frees
;
4077 unsigned long overflows
= cachep
->node_overflow
;
4079 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4080 allocs
, high
, grown
,
4081 reaped
, errors
, max_freeable
, node_allocs
,
4082 node_frees
, overflows
);
4086 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4087 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4088 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4089 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4091 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4092 allochit
, allocmiss
, freehit
, freemiss
);
4097 #define MAX_SLABINFO_WRITE 128
4099 * slabinfo_write - Tuning for the slab allocator
4101 * @buffer: user buffer
4102 * @count: data length
4105 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4106 size_t count
, loff_t
*ppos
)
4108 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4109 int limit
, batchcount
, shared
, res
;
4110 struct kmem_cache
*cachep
;
4112 if (count
> MAX_SLABINFO_WRITE
)
4114 if (copy_from_user(&kbuf
, buffer
, count
))
4116 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4118 tmp
= strchr(kbuf
, ' ');
4123 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4126 /* Find the cache in the chain of caches. */
4127 mutex_lock(&slab_mutex
);
4129 list_for_each_entry(cachep
, &slab_caches
, list
) {
4130 if (!strcmp(cachep
->name
, kbuf
)) {
4131 if (limit
< 1 || batchcount
< 1 ||
4132 batchcount
> limit
|| shared
< 0) {
4135 res
= do_tune_cpucache(cachep
, limit
,
4142 mutex_unlock(&slab_mutex
);
4148 #ifdef CONFIG_DEBUG_SLAB_LEAK
4150 static inline int add_caller(unsigned long *n
, unsigned long v
)
4160 unsigned long *q
= p
+ 2 * i
;
4174 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4180 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4189 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4192 for (j
= page
->active
; j
< c
->num
; j
++) {
4193 if (get_free_obj(page
, j
) == i
) {
4203 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4204 * mapping is established when actual object allocation and
4205 * we could mistakenly access the unmapped object in the cpu
4208 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4211 if (!add_caller(n
, v
))
4216 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4218 #ifdef CONFIG_KALLSYMS
4219 unsigned long offset
, size
;
4220 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4222 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4223 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4225 seq_printf(m
, " [%s]", modname
);
4229 seq_printf(m
, "%p", (void *)address
);
4232 static int leaks_show(struct seq_file
*m
, void *p
)
4234 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4236 struct kmem_cache_node
*n
;
4238 unsigned long *x
= m
->private;
4242 if (!(cachep
->flags
& SLAB_STORE_USER
))
4244 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4248 * Set store_user_clean and start to grab stored user information
4249 * for all objects on this cache. If some alloc/free requests comes
4250 * during the processing, information would be wrong so restart
4254 set_store_user_clean(cachep
);
4255 drain_cpu_caches(cachep
);
4259 for_each_kmem_cache_node(cachep
, node
, n
) {
4262 spin_lock_irq(&n
->list_lock
);
4264 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4265 handle_slab(x
, cachep
, page
);
4266 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4267 handle_slab(x
, cachep
, page
);
4268 spin_unlock_irq(&n
->list_lock
);
4270 } while (!is_store_user_clean(cachep
));
4272 name
= cachep
->name
;
4274 /* Increase the buffer size */
4275 mutex_unlock(&slab_mutex
);
4276 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4278 /* Too bad, we are really out */
4280 mutex_lock(&slab_mutex
);
4283 *(unsigned long *)m
->private = x
[0] * 2;
4285 mutex_lock(&slab_mutex
);
4286 /* Now make sure this entry will be retried */
4290 for (i
= 0; i
< x
[1]; i
++) {
4291 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4292 show_symbol(m
, x
[2*i
+2]);
4299 static const struct seq_operations slabstats_op
= {
4300 .start
= slab_start
,
4306 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4310 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4314 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4319 static const struct file_operations proc_slabstats_operations
= {
4320 .open
= slabstats_open
,
4322 .llseek
= seq_lseek
,
4323 .release
= seq_release_private
,
4327 static int __init
slab_proc_init(void)
4329 #ifdef CONFIG_DEBUG_SLAB_LEAK
4330 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4334 module_init(slab_proc_init
);
4338 * ksize - get the actual amount of memory allocated for a given object
4339 * @objp: Pointer to the object
4341 * kmalloc may internally round up allocations and return more memory
4342 * than requested. ksize() can be used to determine the actual amount of
4343 * memory allocated. The caller may use this additional memory, even though
4344 * a smaller amount of memory was initially specified with the kmalloc call.
4345 * The caller must guarantee that objp points to a valid object previously
4346 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4347 * must not be freed during the duration of the call.
4349 size_t ksize(const void *objp
)
4354 if (unlikely(objp
== ZERO_SIZE_PTR
))
4357 size
= virt_to_cache(objp
)->object_size
;
4358 /* We assume that ksize callers could use the whole allocated area,
4359 * so we need to unpoison this area.
4361 kasan_krealloc(objp
, size
, GFP_NOWAIT
);
4365 EXPORT_SYMBOL(ksize
);