1 // SPDX-License-Identifier: GPL-2.0
4 * Written by Mark Hemment, 1996/97.
5 * (markhe@nextd.demon.co.uk)
7 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
9 * Major cleanup, different bufctl logic, per-cpu arrays
10 * (c) 2000 Manfred Spraul
12 * Cleanup, make the head arrays unconditional, preparation for NUMA
13 * (c) 2002 Manfred Spraul
15 * An implementation of the Slab Allocator as described in outline in;
16 * UNIX Internals: The New Frontiers by Uresh Vahalia
17 * Pub: Prentice Hall ISBN 0-13-101908-2
18 * or with a little more detail in;
19 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
20 * Jeff Bonwick (Sun Microsystems).
21 * Presented at: USENIX Summer 1994 Technical Conference
23 * The memory is organized in caches, one cache for each object type.
24 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
25 * Each cache consists out of many slabs (they are small (usually one
26 * page long) and always contiguous), and each slab contains multiple
27 * initialized objects.
29 * This means, that your constructor is used only for newly allocated
30 * slabs and you must pass objects with the same initializations to
33 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
34 * normal). If you need a special memory type, then must create a new
35 * cache for that memory type.
37 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
38 * full slabs with 0 free objects
40 * empty slabs with no allocated objects
42 * If partial slabs exist, then new allocations come from these slabs,
43 * otherwise from empty slabs or new slabs are allocated.
45 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
46 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
48 * Each cache has a short per-cpu head array, most allocs
49 * and frees go into that array, and if that array overflows, then 1/2
50 * of the entries in the array are given back into the global cache.
51 * The head array is strictly LIFO and should improve the cache hit rates.
52 * On SMP, it additionally reduces the spinlock operations.
54 * The c_cpuarray may not be read with enabled local interrupts -
55 * it's changed with a smp_call_function().
57 * SMP synchronization:
58 * constructors and destructors are called without any locking.
59 * Several members in struct kmem_cache and struct slab never change, they
60 * are accessed without any locking.
61 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
62 * and local interrupts are disabled so slab code is preempt-safe.
63 * The non-constant members are protected with a per-cache irq spinlock.
65 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
66 * in 2000 - many ideas in the current implementation are derived from
69 * Further notes from the original documentation:
71 * 11 April '97. Started multi-threading - markhe
72 * The global cache-chain is protected by the mutex 'slab_mutex'.
73 * The sem is only needed when accessing/extending the cache-chain, which
74 * can never happen inside an interrupt (kmem_cache_create(),
75 * kmem_cache_shrink() and kmem_cache_reap()).
77 * At present, each engine can be growing a cache. This should be blocked.
79 * 15 March 2005. NUMA slab allocator.
80 * Shai Fultheim <shai@scalex86.org>.
81 * Shobhit Dayal <shobhit@calsoftinc.com>
82 * Alok N Kataria <alokk@calsoftinc.com>
83 * Christoph Lameter <christoph@lameter.com>
85 * Modified the slab allocator to be node aware on NUMA systems.
86 * Each node has its own list of partial, free and full slabs.
87 * All object allocations for a node occur from node specific slab lists.
90 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
119 #include <linux/sched/task_stack.h>
121 #include <net/sock.h>
123 #include <asm/cacheflush.h>
124 #include <asm/tlbflush.h>
125 #include <asm/page.h>
127 #include <trace/events/kmem.h>
129 #include "internal.h"
134 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
135 * 0 for faster, smaller code (especially in the critical paths).
137 * STATS - 1 to collect stats for /proc/slabinfo.
138 * 0 for faster, smaller code (especially in the critical paths).
140 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
143 #ifdef CONFIG_DEBUG_SLAB
146 #define FORCED_DEBUG 1
150 #define FORCED_DEBUG 0
153 /* Shouldn't this be in a header file somewhere? */
154 #define BYTES_PER_WORD sizeof(void *)
155 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
157 #ifndef ARCH_KMALLOC_FLAGS
158 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
162 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
164 #if FREELIST_BYTE_INDEX
165 typedef unsigned char freelist_idx_t
;
167 typedef unsigned short freelist_idx_t
;
170 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
176 * - LIFO ordering, to hand out cache-warm objects from _alloc
177 * - reduce the number of linked list operations
178 * - reduce spinlock operations
180 * The limit is stored in the per-cpu structure to reduce the data cache
187 unsigned int batchcount
;
188 unsigned int touched
;
190 * Must have this definition in here for the proper
191 * alignment of array_cache. Also simplifies accessing
198 struct array_cache ac
;
202 * Need this for bootstrapping a per node allocator.
204 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
205 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
206 #define CACHE_CACHE 0
207 #define SIZE_NODE (MAX_NUMNODES)
209 static int drain_freelist(struct kmem_cache
*cache
,
210 struct kmem_cache_node
*n
, int tofree
);
211 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
212 int node
, struct list_head
*list
);
213 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
214 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
215 static void cache_reap(struct work_struct
*unused
);
217 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
219 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
220 struct kmem_cache_node
*n
, struct page
*page
,
222 static int slab_early_init
= 1;
224 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
226 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
228 INIT_LIST_HEAD(&parent
->slabs_full
);
229 INIT_LIST_HEAD(&parent
->slabs_partial
);
230 INIT_LIST_HEAD(&parent
->slabs_free
);
231 parent
->total_slabs
= 0;
232 parent
->free_slabs
= 0;
233 parent
->shared
= NULL
;
234 parent
->alien
= NULL
;
235 parent
->colour_next
= 0;
236 spin_lock_init(&parent
->list_lock
);
237 parent
->free_objects
= 0;
238 parent
->free_touched
= 0;
241 #define MAKE_LIST(cachep, listp, slab, nodeid) \
243 INIT_LIST_HEAD(listp); \
244 list_splice(&get_node(cachep, nodeid)->slab, listp); \
247 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
249 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
250 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
251 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
254 #define CFLGS_OBJFREELIST_SLAB ((slab_flags_t __force)0x40000000U)
255 #define CFLGS_OFF_SLAB ((slab_flags_t __force)0x80000000U)
256 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
257 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
259 #define BATCHREFILL_LIMIT 16
261 * Optimization question: fewer reaps means less probability for unnessary
262 * cpucache drain/refill cycles.
264 * OTOH the cpuarrays can contain lots of objects,
265 * which could lock up otherwise freeable slabs.
267 #define REAPTIMEOUT_AC (2*HZ)
268 #define REAPTIMEOUT_NODE (4*HZ)
271 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
272 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
273 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
274 #define STATS_INC_GROWN(x) ((x)->grown++)
275 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
276 #define STATS_SET_HIGH(x) \
278 if ((x)->num_active > (x)->high_mark) \
279 (x)->high_mark = (x)->num_active; \
281 #define STATS_INC_ERR(x) ((x)->errors++)
282 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
283 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
284 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
285 #define STATS_SET_FREEABLE(x, i) \
287 if ((x)->max_freeable < i) \
288 (x)->max_freeable = i; \
290 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
291 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
292 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
293 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
295 #define STATS_INC_ACTIVE(x) do { } while (0)
296 #define STATS_DEC_ACTIVE(x) do { } while (0)
297 #define STATS_INC_ALLOCED(x) do { } while (0)
298 #define STATS_INC_GROWN(x) do { } while (0)
299 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
300 #define STATS_SET_HIGH(x) do { } while (0)
301 #define STATS_INC_ERR(x) do { } while (0)
302 #define STATS_INC_NODEALLOCS(x) do { } while (0)
303 #define STATS_INC_NODEFREES(x) do { } while (0)
304 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
305 #define STATS_SET_FREEABLE(x, i) do { } while (0)
306 #define STATS_INC_ALLOCHIT(x) do { } while (0)
307 #define STATS_INC_ALLOCMISS(x) do { } while (0)
308 #define STATS_INC_FREEHIT(x) do { } while (0)
309 #define STATS_INC_FREEMISS(x) do { } while (0)
315 * memory layout of objects:
317 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
318 * the end of an object is aligned with the end of the real
319 * allocation. Catches writes behind the end of the allocation.
320 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
322 * cachep->obj_offset: The real object.
323 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
324 * cachep->size - 1* BYTES_PER_WORD: last caller address
325 * [BYTES_PER_WORD long]
327 static int obj_offset(struct kmem_cache
*cachep
)
329 return cachep
->obj_offset
;
332 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
334 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
335 return (unsigned long long*) (objp
+ obj_offset(cachep
) -
336 sizeof(unsigned long long));
339 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
341 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
342 if (cachep
->flags
& SLAB_STORE_USER
)
343 return (unsigned long long *)(objp
+ cachep
->size
-
344 sizeof(unsigned long long) -
346 return (unsigned long long *) (objp
+ cachep
->size
-
347 sizeof(unsigned long long));
350 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
352 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
353 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
358 #define obj_offset(x) 0
359 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
360 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
361 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
365 #ifdef CONFIG_DEBUG_SLAB_LEAK
367 static inline bool is_store_user_clean(struct kmem_cache
*cachep
)
369 return atomic_read(&cachep
->store_user_clean
) == 1;
372 static inline void set_store_user_clean(struct kmem_cache
*cachep
)
374 atomic_set(&cachep
->store_user_clean
, 1);
377 static inline void set_store_user_dirty(struct kmem_cache
*cachep
)
379 if (is_store_user_clean(cachep
))
380 atomic_set(&cachep
->store_user_clean
, 0);
384 static inline void set_store_user_dirty(struct kmem_cache
*cachep
) {}
389 * Do not go above this order unless 0 objects fit into the slab or
390 * overridden on the command line.
392 #define SLAB_MAX_ORDER_HI 1
393 #define SLAB_MAX_ORDER_LO 0
394 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
395 static bool slab_max_order_set __initdata
;
397 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
399 struct page
*page
= virt_to_head_page(obj
);
400 return page
->slab_cache
;
403 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
406 return page
->s_mem
+ cache
->size
* idx
;
409 #define BOOT_CPUCACHE_ENTRIES 1
410 /* internal cache of cache description objs */
411 static struct kmem_cache kmem_cache_boot
= {
413 .limit
= BOOT_CPUCACHE_ENTRIES
,
415 .size
= sizeof(struct kmem_cache
),
416 .name
= "kmem_cache",
419 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
421 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
423 return this_cpu_ptr(cachep
->cpu_cache
);
427 * Calculate the number of objects and left-over bytes for a given buffer size.
429 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
430 slab_flags_t flags
, size_t *left_over
)
433 size_t slab_size
= PAGE_SIZE
<< gfporder
;
436 * The slab management structure can be either off the slab or
437 * on it. For the latter case, the memory allocated for a
440 * - @buffer_size bytes for each object
441 * - One freelist_idx_t for each object
443 * We don't need to consider alignment of freelist because
444 * freelist will be at the end of slab page. The objects will be
445 * at the correct alignment.
447 * If the slab management structure is off the slab, then the
448 * alignment will already be calculated into the size. Because
449 * the slabs are all pages aligned, the objects will be at the
450 * correct alignment when allocated.
452 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
453 num
= slab_size
/ buffer_size
;
454 *left_over
= slab_size
% buffer_size
;
456 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
457 *left_over
= slab_size
%
458 (buffer_size
+ sizeof(freelist_idx_t
));
465 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
467 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
470 pr_err("slab error in %s(): cache `%s': %s\n",
471 function
, cachep
->name
, msg
);
473 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
478 * By default on NUMA we use alien caches to stage the freeing of
479 * objects allocated from other nodes. This causes massive memory
480 * inefficiencies when using fake NUMA setup to split memory into a
481 * large number of small nodes, so it can be disabled on the command
485 static int use_alien_caches __read_mostly
= 1;
486 static int __init
noaliencache_setup(char *s
)
488 use_alien_caches
= 0;
491 __setup("noaliencache", noaliencache_setup
);
493 static int __init
slab_max_order_setup(char *str
)
495 get_option(&str
, &slab_max_order
);
496 slab_max_order
= slab_max_order
< 0 ? 0 :
497 min(slab_max_order
, MAX_ORDER
- 1);
498 slab_max_order_set
= true;
502 __setup("slab_max_order=", slab_max_order_setup
);
506 * Special reaping functions for NUMA systems called from cache_reap().
507 * These take care of doing round robin flushing of alien caches (containing
508 * objects freed on different nodes from which they were allocated) and the
509 * flushing of remote pcps by calling drain_node_pages.
511 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
513 static void init_reap_node(int cpu
)
515 per_cpu(slab_reap_node
, cpu
) = next_node_in(cpu_to_mem(cpu
),
519 static void next_reap_node(void)
521 int node
= __this_cpu_read(slab_reap_node
);
523 node
= next_node_in(node
, node_online_map
);
524 __this_cpu_write(slab_reap_node
, node
);
528 #define init_reap_node(cpu) do { } while (0)
529 #define next_reap_node(void) do { } while (0)
533 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
534 * via the workqueue/eventd.
535 * Add the CPU number into the expiration time to minimize the possibility of
536 * the CPUs getting into lockstep and contending for the global cache chain
539 static void start_cpu_timer(int cpu
)
541 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
543 if (reap_work
->work
.func
== NULL
) {
545 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
546 schedule_delayed_work_on(cpu
, reap_work
,
547 __round_jiffies_relative(HZ
, cpu
));
551 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
556 ac
->batchcount
= batch
;
561 static struct array_cache
*alloc_arraycache(int node
, int entries
,
562 int batchcount
, gfp_t gfp
)
564 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
565 struct array_cache
*ac
= NULL
;
567 ac
= kmalloc_node(memsize
, gfp
, node
);
569 * The array_cache structures contain pointers to free object.
570 * However, when such objects are allocated or transferred to another
571 * cache the pointers are not cleared and they could be counted as
572 * valid references during a kmemleak scan. Therefore, kmemleak must
573 * not scan such objects.
575 kmemleak_no_scan(ac
);
576 init_arraycache(ac
, entries
, batchcount
);
580 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
581 struct page
*page
, void *objp
)
583 struct kmem_cache_node
*n
;
587 page_node
= page_to_nid(page
);
588 n
= get_node(cachep
, page_node
);
590 spin_lock(&n
->list_lock
);
591 free_block(cachep
, &objp
, 1, page_node
, &list
);
592 spin_unlock(&n
->list_lock
);
594 slabs_destroy(cachep
, &list
);
598 * Transfer objects in one arraycache to another.
599 * Locking must be handled by the caller.
601 * Return the number of entries transferred.
603 static int transfer_objects(struct array_cache
*to
,
604 struct array_cache
*from
, unsigned int max
)
606 /* Figure out how many entries to transfer */
607 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
612 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
622 #define drain_alien_cache(cachep, alien) do { } while (0)
623 #define reap_alien(cachep, n) do { } while (0)
625 static inline struct alien_cache
**alloc_alien_cache(int node
,
626 int limit
, gfp_t gfp
)
631 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
635 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
640 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
646 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
647 gfp_t flags
, int nodeid
)
652 static inline gfp_t
gfp_exact_node(gfp_t flags
)
654 return flags
& ~__GFP_NOFAIL
;
657 #else /* CONFIG_NUMA */
659 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
660 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
662 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
663 int batch
, gfp_t gfp
)
665 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
666 struct alien_cache
*alc
= NULL
;
668 alc
= kmalloc_node(memsize
, gfp
, node
);
670 kmemleak_no_scan(alc
);
671 init_arraycache(&alc
->ac
, entries
, batch
);
672 spin_lock_init(&alc
->lock
);
677 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
679 struct alien_cache
**alc_ptr
;
684 alc_ptr
= kcalloc_node(nr_node_ids
, sizeof(void *), gfp
, node
);
689 if (i
== node
|| !node_online(i
))
691 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
693 for (i
--; i
>= 0; i
--)
702 static void free_alien_cache(struct alien_cache
**alc_ptr
)
713 static void __drain_alien_cache(struct kmem_cache
*cachep
,
714 struct array_cache
*ac
, int node
,
715 struct list_head
*list
)
717 struct kmem_cache_node
*n
= get_node(cachep
, node
);
720 spin_lock(&n
->list_lock
);
722 * Stuff objects into the remote nodes shared array first.
723 * That way we could avoid the overhead of putting the objects
724 * into the free lists and getting them back later.
727 transfer_objects(n
->shared
, ac
, ac
->limit
);
729 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
731 spin_unlock(&n
->list_lock
);
736 * Called from cache_reap() to regularly drain alien caches round robin.
738 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
740 int node
= __this_cpu_read(slab_reap_node
);
743 struct alien_cache
*alc
= n
->alien
[node
];
744 struct array_cache
*ac
;
748 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
751 __drain_alien_cache(cachep
, ac
, node
, &list
);
752 spin_unlock_irq(&alc
->lock
);
753 slabs_destroy(cachep
, &list
);
759 static void drain_alien_cache(struct kmem_cache
*cachep
,
760 struct alien_cache
**alien
)
763 struct alien_cache
*alc
;
764 struct array_cache
*ac
;
767 for_each_online_node(i
) {
773 spin_lock_irqsave(&alc
->lock
, flags
);
774 __drain_alien_cache(cachep
, ac
, i
, &list
);
775 spin_unlock_irqrestore(&alc
->lock
, flags
);
776 slabs_destroy(cachep
, &list
);
781 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
782 int node
, int page_node
)
784 struct kmem_cache_node
*n
;
785 struct alien_cache
*alien
= NULL
;
786 struct array_cache
*ac
;
789 n
= get_node(cachep
, node
);
790 STATS_INC_NODEFREES(cachep
);
791 if (n
->alien
&& n
->alien
[page_node
]) {
792 alien
= n
->alien
[page_node
];
794 spin_lock(&alien
->lock
);
795 if (unlikely(ac
->avail
== ac
->limit
)) {
796 STATS_INC_ACOVERFLOW(cachep
);
797 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
799 ac
->entry
[ac
->avail
++] = objp
;
800 spin_unlock(&alien
->lock
);
801 slabs_destroy(cachep
, &list
);
803 n
= get_node(cachep
, page_node
);
804 spin_lock(&n
->list_lock
);
805 free_block(cachep
, &objp
, 1, page_node
, &list
);
806 spin_unlock(&n
->list_lock
);
807 slabs_destroy(cachep
, &list
);
812 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
814 int page_node
= page_to_nid(virt_to_page(objp
));
815 int node
= numa_mem_id();
817 * Make sure we are not freeing a object from another node to the array
820 if (likely(node
== page_node
))
823 return __cache_free_alien(cachep
, objp
, node
, page_node
);
827 * Construct gfp mask to allocate from a specific node but do not reclaim or
828 * warn about failures.
830 static inline gfp_t
gfp_exact_node(gfp_t flags
)
832 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
836 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
838 struct kmem_cache_node
*n
;
841 * Set up the kmem_cache_node for cpu before we can
842 * begin anything. Make sure some other cpu on this
843 * node has not already allocated this
845 n
= get_node(cachep
, node
);
847 spin_lock_irq(&n
->list_lock
);
848 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
850 spin_unlock_irq(&n
->list_lock
);
855 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
859 kmem_cache_node_init(n
);
860 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
861 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
864 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
867 * The kmem_cache_nodes don't come and go as CPUs
868 * come and go. slab_mutex is sufficient
871 cachep
->node
[node
] = n
;
876 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
878 * Allocates and initializes node for a node on each slab cache, used for
879 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
880 * will be allocated off-node since memory is not yet online for the new node.
881 * When hotplugging memory or a cpu, existing node are not replaced if
884 * Must hold slab_mutex.
886 static int init_cache_node_node(int node
)
889 struct kmem_cache
*cachep
;
891 list_for_each_entry(cachep
, &slab_caches
, list
) {
892 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
901 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
902 int node
, gfp_t gfp
, bool force_change
)
905 struct kmem_cache_node
*n
;
906 struct array_cache
*old_shared
= NULL
;
907 struct array_cache
*new_shared
= NULL
;
908 struct alien_cache
**new_alien
= NULL
;
911 if (use_alien_caches
) {
912 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
917 if (cachep
->shared
) {
918 new_shared
= alloc_arraycache(node
,
919 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
924 ret
= init_cache_node(cachep
, node
, gfp
);
928 n
= get_node(cachep
, node
);
929 spin_lock_irq(&n
->list_lock
);
930 if (n
->shared
&& force_change
) {
931 free_block(cachep
, n
->shared
->entry
,
932 n
->shared
->avail
, node
, &list
);
933 n
->shared
->avail
= 0;
936 if (!n
->shared
|| force_change
) {
937 old_shared
= n
->shared
;
938 n
->shared
= new_shared
;
943 n
->alien
= new_alien
;
947 spin_unlock_irq(&n
->list_lock
);
948 slabs_destroy(cachep
, &list
);
951 * To protect lockless access to n->shared during irq disabled context.
952 * If n->shared isn't NULL in irq disabled context, accessing to it is
953 * guaranteed to be valid until irq is re-enabled, because it will be
954 * freed after synchronize_rcu().
956 if (old_shared
&& force_change
)
962 free_alien_cache(new_alien
);
969 static void cpuup_canceled(long cpu
)
971 struct kmem_cache
*cachep
;
972 struct kmem_cache_node
*n
= NULL
;
973 int node
= cpu_to_mem(cpu
);
974 const struct cpumask
*mask
= cpumask_of_node(node
);
976 list_for_each_entry(cachep
, &slab_caches
, list
) {
977 struct array_cache
*nc
;
978 struct array_cache
*shared
;
979 struct alien_cache
**alien
;
982 n
= get_node(cachep
, node
);
986 spin_lock_irq(&n
->list_lock
);
988 /* Free limit for this kmem_cache_node */
989 n
->free_limit
-= cachep
->batchcount
;
991 /* cpu is dead; no one can alloc from it. */
992 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
993 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
996 if (!cpumask_empty(mask
)) {
997 spin_unlock_irq(&n
->list_lock
);
1003 free_block(cachep
, shared
->entry
,
1004 shared
->avail
, node
, &list
);
1011 spin_unlock_irq(&n
->list_lock
);
1015 drain_alien_cache(cachep
, alien
);
1016 free_alien_cache(alien
);
1020 slabs_destroy(cachep
, &list
);
1023 * In the previous loop, all the objects were freed to
1024 * the respective cache's slabs, now we can go ahead and
1025 * shrink each nodelist to its limit.
1027 list_for_each_entry(cachep
, &slab_caches
, list
) {
1028 n
= get_node(cachep
, node
);
1031 drain_freelist(cachep
, n
, INT_MAX
);
1035 static int cpuup_prepare(long cpu
)
1037 struct kmem_cache
*cachep
;
1038 int node
= cpu_to_mem(cpu
);
1042 * We need to do this right in the beginning since
1043 * alloc_arraycache's are going to use this list.
1044 * kmalloc_node allows us to add the slab to the right
1045 * kmem_cache_node and not this cpu's kmem_cache_node
1047 err
= init_cache_node_node(node
);
1052 * Now we can go ahead with allocating the shared arrays and
1055 list_for_each_entry(cachep
, &slab_caches
, list
) {
1056 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1063 cpuup_canceled(cpu
);
1067 int slab_prepare_cpu(unsigned int cpu
)
1071 mutex_lock(&slab_mutex
);
1072 err
= cpuup_prepare(cpu
);
1073 mutex_unlock(&slab_mutex
);
1078 * This is called for a failed online attempt and for a successful
1081 * Even if all the cpus of a node are down, we don't free the
1082 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1083 * a kmalloc allocation from another cpu for memory from the node of
1084 * the cpu going down. The list3 structure is usually allocated from
1085 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1087 int slab_dead_cpu(unsigned int cpu
)
1089 mutex_lock(&slab_mutex
);
1090 cpuup_canceled(cpu
);
1091 mutex_unlock(&slab_mutex
);
1096 static int slab_online_cpu(unsigned int cpu
)
1098 start_cpu_timer(cpu
);
1102 static int slab_offline_cpu(unsigned int cpu
)
1105 * Shutdown cache reaper. Note that the slab_mutex is held so
1106 * that if cache_reap() is invoked it cannot do anything
1107 * expensive but will only modify reap_work and reschedule the
1110 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1111 /* Now the cache_reaper is guaranteed to be not running. */
1112 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1116 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1118 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1119 * Returns -EBUSY if all objects cannot be drained so that the node is not
1122 * Must hold slab_mutex.
1124 static int __meminit
drain_cache_node_node(int node
)
1126 struct kmem_cache
*cachep
;
1129 list_for_each_entry(cachep
, &slab_caches
, list
) {
1130 struct kmem_cache_node
*n
;
1132 n
= get_node(cachep
, node
);
1136 drain_freelist(cachep
, n
, INT_MAX
);
1138 if (!list_empty(&n
->slabs_full
) ||
1139 !list_empty(&n
->slabs_partial
)) {
1147 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1148 unsigned long action
, void *arg
)
1150 struct memory_notify
*mnb
= arg
;
1154 nid
= mnb
->status_change_nid
;
1159 case MEM_GOING_ONLINE
:
1160 mutex_lock(&slab_mutex
);
1161 ret
= init_cache_node_node(nid
);
1162 mutex_unlock(&slab_mutex
);
1164 case MEM_GOING_OFFLINE
:
1165 mutex_lock(&slab_mutex
);
1166 ret
= drain_cache_node_node(nid
);
1167 mutex_unlock(&slab_mutex
);
1171 case MEM_CANCEL_ONLINE
:
1172 case MEM_CANCEL_OFFLINE
:
1176 return notifier_from_errno(ret
);
1178 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1181 * swap the static kmem_cache_node with kmalloced memory
1183 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1186 struct kmem_cache_node
*ptr
;
1188 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1191 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1193 * Do not assume that spinlocks can be initialized via memcpy:
1195 spin_lock_init(&ptr
->list_lock
);
1197 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1198 cachep
->node
[nodeid
] = ptr
;
1202 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1203 * size of kmem_cache_node.
1205 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1209 for_each_online_node(node
) {
1210 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1211 cachep
->node
[node
]->next_reap
= jiffies
+
1213 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1218 * Initialisation. Called after the page allocator have been initialised and
1219 * before smp_init().
1221 void __init
kmem_cache_init(void)
1225 kmem_cache
= &kmem_cache_boot
;
1227 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1228 use_alien_caches
= 0;
1230 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1231 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1234 * Fragmentation resistance on low memory - only use bigger
1235 * page orders on machines with more than 32MB of memory if
1236 * not overridden on the command line.
1238 if (!slab_max_order_set
&& totalram_pages() > (32 << 20) >> PAGE_SHIFT
)
1239 slab_max_order
= SLAB_MAX_ORDER_HI
;
1241 /* Bootstrap is tricky, because several objects are allocated
1242 * from caches that do not exist yet:
1243 * 1) initialize the kmem_cache cache: it contains the struct
1244 * kmem_cache structures of all caches, except kmem_cache itself:
1245 * kmem_cache is statically allocated.
1246 * Initially an __init data area is used for the head array and the
1247 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1248 * array at the end of the bootstrap.
1249 * 2) Create the first kmalloc cache.
1250 * The struct kmem_cache for the new cache is allocated normally.
1251 * An __init data area is used for the head array.
1252 * 3) Create the remaining kmalloc caches, with minimally sized
1254 * 4) Replace the __init data head arrays for kmem_cache and the first
1255 * kmalloc cache with kmalloc allocated arrays.
1256 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1257 * the other cache's with kmalloc allocated memory.
1258 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1261 /* 1) create the kmem_cache */
1264 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1266 create_boot_cache(kmem_cache
, "kmem_cache",
1267 offsetof(struct kmem_cache
, node
) +
1268 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1269 SLAB_HWCACHE_ALIGN
, 0, 0);
1270 list_add(&kmem_cache
->list
, &slab_caches
);
1271 memcg_link_cache(kmem_cache
);
1272 slab_state
= PARTIAL
;
1275 * Initialize the caches that provide memory for the kmem_cache_node
1276 * structures first. Without this, further allocations will bug.
1278 kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
] = create_kmalloc_cache(
1279 kmalloc_info
[INDEX_NODE
].name
,
1280 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
,
1281 0, kmalloc_size(INDEX_NODE
));
1282 slab_state
= PARTIAL_NODE
;
1283 setup_kmalloc_cache_index_table();
1285 slab_early_init
= 0;
1287 /* 5) Replace the bootstrap kmem_cache_node */
1291 for_each_online_node(nid
) {
1292 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1294 init_list(kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
],
1295 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1299 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1302 void __init
kmem_cache_init_late(void)
1304 struct kmem_cache
*cachep
;
1306 /* 6) resize the head arrays to their final sizes */
1307 mutex_lock(&slab_mutex
);
1308 list_for_each_entry(cachep
, &slab_caches
, list
)
1309 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1311 mutex_unlock(&slab_mutex
);
1318 * Register a memory hotplug callback that initializes and frees
1321 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1325 * The reap timers are started later, with a module init call: That part
1326 * of the kernel is not yet operational.
1330 static int __init
cpucache_init(void)
1335 * Register the timers that return unneeded pages to the page allocator
1337 ret
= cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "SLAB online",
1338 slab_online_cpu
, slab_offline_cpu
);
1343 __initcall(cpucache_init
);
1345 static noinline
void
1346 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1349 struct kmem_cache_node
*n
;
1350 unsigned long flags
;
1352 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1353 DEFAULT_RATELIMIT_BURST
);
1355 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1358 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1359 nodeid
, gfpflags
, &gfpflags
);
1360 pr_warn(" cache: %s, object size: %d, order: %d\n",
1361 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1363 for_each_kmem_cache_node(cachep
, node
, n
) {
1364 unsigned long total_slabs
, free_slabs
, free_objs
;
1366 spin_lock_irqsave(&n
->list_lock
, flags
);
1367 total_slabs
= n
->total_slabs
;
1368 free_slabs
= n
->free_slabs
;
1369 free_objs
= n
->free_objects
;
1370 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1372 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1373 node
, total_slabs
- free_slabs
, total_slabs
,
1374 (total_slabs
* cachep
->num
) - free_objs
,
1375 total_slabs
* cachep
->num
);
1381 * Interface to system's page allocator. No need to hold the
1382 * kmem_cache_node ->list_lock.
1384 * If we requested dmaable memory, we will get it. Even if we
1385 * did not request dmaable memory, we might get it, but that
1386 * would be relatively rare and ignorable.
1388 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1394 flags
|= cachep
->allocflags
;
1396 page
= __alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1398 slab_out_of_memory(cachep
, flags
, nodeid
);
1402 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1403 __free_pages(page
, cachep
->gfporder
);
1407 nr_pages
= (1 << cachep
->gfporder
);
1408 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1409 mod_lruvec_page_state(page
, NR_SLAB_RECLAIMABLE
, nr_pages
);
1411 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1413 __SetPageSlab(page
);
1414 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1415 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1416 SetPageSlabPfmemalloc(page
);
1422 * Interface to system's page release.
1424 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1426 int order
= cachep
->gfporder
;
1427 unsigned long nr_freed
= (1 << order
);
1429 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1430 mod_lruvec_page_state(page
, NR_SLAB_RECLAIMABLE
, -nr_freed
);
1432 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE
, -nr_freed
);
1434 BUG_ON(!PageSlab(page
));
1435 __ClearPageSlabPfmemalloc(page
);
1436 __ClearPageSlab(page
);
1437 page_mapcount_reset(page
);
1438 page
->mapping
= NULL
;
1440 if (current
->reclaim_state
)
1441 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1442 memcg_uncharge_slab(page
, order
, cachep
);
1443 __free_pages(page
, order
);
1446 static void kmem_rcu_free(struct rcu_head
*head
)
1448 struct kmem_cache
*cachep
;
1451 page
= container_of(head
, struct page
, rcu_head
);
1452 cachep
= page
->slab_cache
;
1454 kmem_freepages(cachep
, page
);
1458 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1460 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1461 (cachep
->size
% PAGE_SIZE
) == 0)
1467 #ifdef CONFIG_DEBUG_PAGEALLOC
1468 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
, int map
)
1470 if (!is_debug_pagealloc_cache(cachep
))
1473 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1477 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1482 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1484 int size
= cachep
->object_size
;
1485 addr
= &((char *)addr
)[obj_offset(cachep
)];
1487 memset(addr
, val
, size
);
1488 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1491 static void dump_line(char *data
, int offset
, int limit
)
1494 unsigned char error
= 0;
1497 pr_err("%03x: ", offset
);
1498 for (i
= 0; i
< limit
; i
++) {
1499 if (data
[offset
+ i
] != POISON_FREE
) {
1500 error
= data
[offset
+ i
];
1504 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1505 &data
[offset
], limit
, 1);
1507 if (bad_count
== 1) {
1508 error
^= POISON_FREE
;
1509 if (!(error
& (error
- 1))) {
1510 pr_err("Single bit error detected. Probably bad RAM.\n");
1512 pr_err("Run memtest86+ or a similar memory test tool.\n");
1514 pr_err("Run a memory test tool.\n");
1523 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1528 if (cachep
->flags
& SLAB_RED_ZONE
) {
1529 pr_err("Redzone: 0x%llx/0x%llx\n",
1530 *dbg_redzone1(cachep
, objp
),
1531 *dbg_redzone2(cachep
, objp
));
1534 if (cachep
->flags
& SLAB_STORE_USER
)
1535 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep
, objp
));
1536 realobj
= (char *)objp
+ obj_offset(cachep
);
1537 size
= cachep
->object_size
;
1538 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1541 if (i
+ limit
> size
)
1543 dump_line(realobj
, i
, limit
);
1547 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1553 if (is_debug_pagealloc_cache(cachep
))
1556 realobj
= (char *)objp
+ obj_offset(cachep
);
1557 size
= cachep
->object_size
;
1559 for (i
= 0; i
< size
; i
++) {
1560 char exp
= POISON_FREE
;
1563 if (realobj
[i
] != exp
) {
1568 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1569 print_tainted(), cachep
->name
,
1571 print_objinfo(cachep
, objp
, 0);
1573 /* Hexdump the affected line */
1576 if (i
+ limit
> size
)
1578 dump_line(realobj
, i
, limit
);
1581 /* Limit to 5 lines */
1587 /* Print some data about the neighboring objects, if they
1590 struct page
*page
= virt_to_head_page(objp
);
1593 objnr
= obj_to_index(cachep
, page
, objp
);
1595 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1596 realobj
= (char *)objp
+ obj_offset(cachep
);
1597 pr_err("Prev obj: start=%px, len=%d\n", realobj
, size
);
1598 print_objinfo(cachep
, objp
, 2);
1600 if (objnr
+ 1 < cachep
->num
) {
1601 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1602 realobj
= (char *)objp
+ obj_offset(cachep
);
1603 pr_err("Next obj: start=%px, len=%d\n", realobj
, size
);
1604 print_objinfo(cachep
, objp
, 2);
1611 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1616 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1617 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1621 for (i
= 0; i
< cachep
->num
; i
++) {
1622 void *objp
= index_to_obj(cachep
, page
, i
);
1624 if (cachep
->flags
& SLAB_POISON
) {
1625 check_poison_obj(cachep
, objp
);
1626 slab_kernel_map(cachep
, objp
, 1);
1628 if (cachep
->flags
& SLAB_RED_ZONE
) {
1629 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1630 slab_error(cachep
, "start of a freed object was overwritten");
1631 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1632 slab_error(cachep
, "end of a freed object was overwritten");
1637 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1644 * slab_destroy - destroy and release all objects in a slab
1645 * @cachep: cache pointer being destroyed
1646 * @page: page pointer being destroyed
1648 * Destroy all the objs in a slab page, and release the mem back to the system.
1649 * Before calling the slab page must have been unlinked from the cache. The
1650 * kmem_cache_node ->list_lock is not held/needed.
1652 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1656 freelist
= page
->freelist
;
1657 slab_destroy_debugcheck(cachep
, page
);
1658 if (unlikely(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
1659 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1661 kmem_freepages(cachep
, page
);
1664 * From now on, we don't use freelist
1665 * although actual page can be freed in rcu context
1667 if (OFF_SLAB(cachep
))
1668 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1671 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1673 struct page
*page
, *n
;
1675 list_for_each_entry_safe(page
, n
, list
, slab_list
) {
1676 list_del(&page
->slab_list
);
1677 slab_destroy(cachep
, page
);
1682 * calculate_slab_order - calculate size (page order) of slabs
1683 * @cachep: pointer to the cache that is being created
1684 * @size: size of objects to be created in this cache.
1685 * @flags: slab allocation flags
1687 * Also calculates the number of objects per slab.
1689 * This could be made much more intelligent. For now, try to avoid using
1690 * high order pages for slabs. When the gfp() functions are more friendly
1691 * towards high-order requests, this should be changed.
1693 * Return: number of left-over bytes in a slab
1695 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1696 size_t size
, slab_flags_t flags
)
1698 size_t left_over
= 0;
1701 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1705 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1709 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1710 if (num
> SLAB_OBJ_MAX_NUM
)
1713 if (flags
& CFLGS_OFF_SLAB
) {
1714 struct kmem_cache
*freelist_cache
;
1715 size_t freelist_size
;
1717 freelist_size
= num
* sizeof(freelist_idx_t
);
1718 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1719 if (!freelist_cache
)
1723 * Needed to avoid possible looping condition
1724 * in cache_grow_begin()
1726 if (OFF_SLAB(freelist_cache
))
1729 /* check if off slab has enough benefit */
1730 if (freelist_cache
->size
> cachep
->size
/ 2)
1734 /* Found something acceptable - save it away */
1736 cachep
->gfporder
= gfporder
;
1737 left_over
= remainder
;
1740 * A VFS-reclaimable slab tends to have most allocations
1741 * as GFP_NOFS and we really don't want to have to be allocating
1742 * higher-order pages when we are unable to shrink dcache.
1744 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1748 * Large number of objects is good, but very large slabs are
1749 * currently bad for the gfp()s.
1751 if (gfporder
>= slab_max_order
)
1755 * Acceptable internal fragmentation?
1757 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1763 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1764 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1768 struct array_cache __percpu
*cpu_cache
;
1770 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1771 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1776 for_each_possible_cpu(cpu
) {
1777 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1778 entries
, batchcount
);
1784 static int __ref
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1786 if (slab_state
>= FULL
)
1787 return enable_cpucache(cachep
, gfp
);
1789 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1790 if (!cachep
->cpu_cache
)
1793 if (slab_state
== DOWN
) {
1794 /* Creation of first cache (kmem_cache). */
1795 set_up_node(kmem_cache
, CACHE_CACHE
);
1796 } else if (slab_state
== PARTIAL
) {
1797 /* For kmem_cache_node */
1798 set_up_node(cachep
, SIZE_NODE
);
1802 for_each_online_node(node
) {
1803 cachep
->node
[node
] = kmalloc_node(
1804 sizeof(struct kmem_cache_node
), gfp
, node
);
1805 BUG_ON(!cachep
->node
[node
]);
1806 kmem_cache_node_init(cachep
->node
[node
]);
1810 cachep
->node
[numa_mem_id()]->next_reap
=
1811 jiffies
+ REAPTIMEOUT_NODE
+
1812 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1814 cpu_cache_get(cachep
)->avail
= 0;
1815 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1816 cpu_cache_get(cachep
)->batchcount
= 1;
1817 cpu_cache_get(cachep
)->touched
= 0;
1818 cachep
->batchcount
= 1;
1819 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1823 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1824 slab_flags_t flags
, const char *name
,
1825 void (*ctor
)(void *))
1831 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
1832 slab_flags_t flags
, void (*ctor
)(void *))
1834 struct kmem_cache
*cachep
;
1836 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1841 * Adjust the object sizes so that we clear
1842 * the complete object on kzalloc.
1844 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1849 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1850 size_t size
, slab_flags_t flags
)
1856 if (cachep
->ctor
|| flags
& SLAB_TYPESAFE_BY_RCU
)
1859 left
= calculate_slab_order(cachep
, size
,
1860 flags
| CFLGS_OBJFREELIST_SLAB
);
1864 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1867 cachep
->colour
= left
/ cachep
->colour_off
;
1872 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1873 size_t size
, slab_flags_t flags
)
1880 * Always use on-slab management when SLAB_NOLEAKTRACE
1881 * to avoid recursive calls into kmemleak.
1883 if (flags
& SLAB_NOLEAKTRACE
)
1887 * Size is large, assume best to place the slab management obj
1888 * off-slab (should allow better packing of objs).
1890 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1895 * If the slab has been placed off-slab, and we have enough space then
1896 * move it on-slab. This is at the expense of any extra colouring.
1898 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1901 cachep
->colour
= left
/ cachep
->colour_off
;
1906 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1907 size_t size
, slab_flags_t flags
)
1913 left
= calculate_slab_order(cachep
, size
, flags
);
1917 cachep
->colour
= left
/ cachep
->colour_off
;
1923 * __kmem_cache_create - Create a cache.
1924 * @cachep: cache management descriptor
1925 * @flags: SLAB flags
1927 * Returns a ptr to the cache on success, NULL on failure.
1928 * Cannot be called within a int, but can be interrupted.
1929 * The @ctor is run when new pages are allocated by the cache.
1933 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1934 * to catch references to uninitialised memory.
1936 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1937 * for buffer overruns.
1939 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1940 * cacheline. This can be beneficial if you're counting cycles as closely
1943 * Return: a pointer to the created cache or %NULL in case of error
1945 int __kmem_cache_create(struct kmem_cache
*cachep
, slab_flags_t flags
)
1947 size_t ralign
= BYTES_PER_WORD
;
1950 unsigned int size
= cachep
->size
;
1955 * Enable redzoning and last user accounting, except for caches with
1956 * large objects, if the increased size would increase the object size
1957 * above the next power of two: caches with object sizes just above a
1958 * power of two have a significant amount of internal fragmentation.
1960 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
1961 2 * sizeof(unsigned long long)))
1962 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1963 if (!(flags
& SLAB_TYPESAFE_BY_RCU
))
1964 flags
|= SLAB_POISON
;
1969 * Check that size is in terms of words. This is needed to avoid
1970 * unaligned accesses for some archs when redzoning is used, and makes
1971 * sure any on-slab bufctl's are also correctly aligned.
1973 size
= ALIGN(size
, BYTES_PER_WORD
);
1975 if (flags
& SLAB_RED_ZONE
) {
1976 ralign
= REDZONE_ALIGN
;
1977 /* If redzoning, ensure that the second redzone is suitably
1978 * aligned, by adjusting the object size accordingly. */
1979 size
= ALIGN(size
, REDZONE_ALIGN
);
1982 /* 3) caller mandated alignment */
1983 if (ralign
< cachep
->align
) {
1984 ralign
= cachep
->align
;
1986 /* disable debug if necessary */
1987 if (ralign
> __alignof__(unsigned long long))
1988 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1992 cachep
->align
= ralign
;
1993 cachep
->colour_off
= cache_line_size();
1994 /* Offset must be a multiple of the alignment. */
1995 if (cachep
->colour_off
< cachep
->align
)
1996 cachep
->colour_off
= cachep
->align
;
1998 if (slab_is_available())
2006 * Both debugging options require word-alignment which is calculated
2009 if (flags
& SLAB_RED_ZONE
) {
2010 /* add space for red zone words */
2011 cachep
->obj_offset
+= sizeof(unsigned long long);
2012 size
+= 2 * sizeof(unsigned long long);
2014 if (flags
& SLAB_STORE_USER
) {
2015 /* user store requires one word storage behind the end of
2016 * the real object. But if the second red zone needs to be
2017 * aligned to 64 bits, we must allow that much space.
2019 if (flags
& SLAB_RED_ZONE
)
2020 size
+= REDZONE_ALIGN
;
2022 size
+= BYTES_PER_WORD
;
2026 kasan_cache_create(cachep
, &size
, &flags
);
2028 size
= ALIGN(size
, cachep
->align
);
2030 * We should restrict the number of objects in a slab to implement
2031 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2033 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2034 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2038 * To activate debug pagealloc, off-slab management is necessary
2039 * requirement. In early phase of initialization, small sized slab
2040 * doesn't get initialized so it would not be possible. So, we need
2041 * to check size >= 256. It guarantees that all necessary small
2042 * sized slab is initialized in current slab initialization sequence.
2044 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2045 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2046 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2047 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2049 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2050 flags
|= CFLGS_OFF_SLAB
;
2051 cachep
->obj_offset
+= tmp_size
- size
;
2059 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2060 flags
|= CFLGS_OBJFREELIST_SLAB
;
2064 if (set_off_slab_cache(cachep
, size
, flags
)) {
2065 flags
|= CFLGS_OFF_SLAB
;
2069 if (set_on_slab_cache(cachep
, size
, flags
))
2075 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2076 cachep
->flags
= flags
;
2077 cachep
->allocflags
= __GFP_COMP
;
2078 if (flags
& SLAB_CACHE_DMA
)
2079 cachep
->allocflags
|= GFP_DMA
;
2080 if (flags
& SLAB_CACHE_DMA32
)
2081 cachep
->allocflags
|= GFP_DMA32
;
2082 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2083 cachep
->allocflags
|= __GFP_RECLAIMABLE
;
2084 cachep
->size
= size
;
2085 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2089 * If we're going to use the generic kernel_map_pages()
2090 * poisoning, then it's going to smash the contents of
2091 * the redzone and userword anyhow, so switch them off.
2093 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2094 (cachep
->flags
& SLAB_POISON
) &&
2095 is_debug_pagealloc_cache(cachep
))
2096 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2099 if (OFF_SLAB(cachep
)) {
2100 cachep
->freelist_cache
=
2101 kmalloc_slab(cachep
->freelist_size
, 0u);
2104 err
= setup_cpu_cache(cachep
, gfp
);
2106 __kmem_cache_release(cachep
);
2114 static void check_irq_off(void)
2116 BUG_ON(!irqs_disabled());
2119 static void check_irq_on(void)
2121 BUG_ON(irqs_disabled());
2124 static void check_mutex_acquired(void)
2126 BUG_ON(!mutex_is_locked(&slab_mutex
));
2129 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2133 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2137 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2141 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2146 #define check_irq_off() do { } while(0)
2147 #define check_irq_on() do { } while(0)
2148 #define check_mutex_acquired() do { } while(0)
2149 #define check_spinlock_acquired(x) do { } while(0)
2150 #define check_spinlock_acquired_node(x, y) do { } while(0)
2153 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2154 int node
, bool free_all
, struct list_head
*list
)
2158 if (!ac
|| !ac
->avail
)
2161 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2162 if (tofree
> ac
->avail
)
2163 tofree
= (ac
->avail
+ 1) / 2;
2165 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2166 ac
->avail
-= tofree
;
2167 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2170 static void do_drain(void *arg
)
2172 struct kmem_cache
*cachep
= arg
;
2173 struct array_cache
*ac
;
2174 int node
= numa_mem_id();
2175 struct kmem_cache_node
*n
;
2179 ac
= cpu_cache_get(cachep
);
2180 n
= get_node(cachep
, node
);
2181 spin_lock(&n
->list_lock
);
2182 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2183 spin_unlock(&n
->list_lock
);
2184 slabs_destroy(cachep
, &list
);
2188 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2190 struct kmem_cache_node
*n
;
2194 on_each_cpu(do_drain
, cachep
, 1);
2196 for_each_kmem_cache_node(cachep
, node
, n
)
2198 drain_alien_cache(cachep
, n
->alien
);
2200 for_each_kmem_cache_node(cachep
, node
, n
) {
2201 spin_lock_irq(&n
->list_lock
);
2202 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2203 spin_unlock_irq(&n
->list_lock
);
2205 slabs_destroy(cachep
, &list
);
2210 * Remove slabs from the list of free slabs.
2211 * Specify the number of slabs to drain in tofree.
2213 * Returns the actual number of slabs released.
2215 static int drain_freelist(struct kmem_cache
*cache
,
2216 struct kmem_cache_node
*n
, int tofree
)
2218 struct list_head
*p
;
2223 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2225 spin_lock_irq(&n
->list_lock
);
2226 p
= n
->slabs_free
.prev
;
2227 if (p
== &n
->slabs_free
) {
2228 spin_unlock_irq(&n
->list_lock
);
2232 page
= list_entry(p
, struct page
, slab_list
);
2233 list_del(&page
->slab_list
);
2237 * Safe to drop the lock. The slab is no longer linked
2240 n
->free_objects
-= cache
->num
;
2241 spin_unlock_irq(&n
->list_lock
);
2242 slab_destroy(cache
, page
);
2249 bool __kmem_cache_empty(struct kmem_cache
*s
)
2252 struct kmem_cache_node
*n
;
2254 for_each_kmem_cache_node(s
, node
, n
)
2255 if (!list_empty(&n
->slabs_full
) ||
2256 !list_empty(&n
->slabs_partial
))
2261 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2265 struct kmem_cache_node
*n
;
2267 drain_cpu_caches(cachep
);
2270 for_each_kmem_cache_node(cachep
, node
, n
) {
2271 drain_freelist(cachep
, n
, INT_MAX
);
2273 ret
+= !list_empty(&n
->slabs_full
) ||
2274 !list_empty(&n
->slabs_partial
);
2276 return (ret
? 1 : 0);
2280 void __kmemcg_cache_deactivate(struct kmem_cache
*cachep
)
2282 __kmem_cache_shrink(cachep
);
2286 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2288 return __kmem_cache_shrink(cachep
);
2291 void __kmem_cache_release(struct kmem_cache
*cachep
)
2294 struct kmem_cache_node
*n
;
2296 cache_random_seq_destroy(cachep
);
2298 free_percpu(cachep
->cpu_cache
);
2300 /* NUMA: free the node structures */
2301 for_each_kmem_cache_node(cachep
, i
, n
) {
2303 free_alien_cache(n
->alien
);
2305 cachep
->node
[i
] = NULL
;
2310 * Get the memory for a slab management obj.
2312 * For a slab cache when the slab descriptor is off-slab, the
2313 * slab descriptor can't come from the same cache which is being created,
2314 * Because if it is the case, that means we defer the creation of
2315 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2316 * And we eventually call down to __kmem_cache_create(), which
2317 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2318 * This is a "chicken-and-egg" problem.
2320 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2321 * which are all initialized during kmem_cache_init().
2323 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2324 struct page
*page
, int colour_off
,
2325 gfp_t local_flags
, int nodeid
)
2328 void *addr
= page_address(page
);
2330 page
->s_mem
= addr
+ colour_off
;
2333 if (OBJFREELIST_SLAB(cachep
))
2335 else if (OFF_SLAB(cachep
)) {
2336 /* Slab management obj is off-slab. */
2337 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2338 local_flags
, nodeid
);
2342 /* We will use last bytes at the slab for freelist */
2343 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2344 cachep
->freelist_size
;
2350 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2352 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2355 static inline void set_free_obj(struct page
*page
,
2356 unsigned int idx
, freelist_idx_t val
)
2358 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2361 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2366 for (i
= 0; i
< cachep
->num
; i
++) {
2367 void *objp
= index_to_obj(cachep
, page
, i
);
2369 if (cachep
->flags
& SLAB_STORE_USER
)
2370 *dbg_userword(cachep
, objp
) = NULL
;
2372 if (cachep
->flags
& SLAB_RED_ZONE
) {
2373 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2374 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2377 * Constructors are not allowed to allocate memory from the same
2378 * cache which they are a constructor for. Otherwise, deadlock.
2379 * They must also be threaded.
2381 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2382 kasan_unpoison_object_data(cachep
,
2383 objp
+ obj_offset(cachep
));
2384 cachep
->ctor(objp
+ obj_offset(cachep
));
2385 kasan_poison_object_data(
2386 cachep
, objp
+ obj_offset(cachep
));
2389 if (cachep
->flags
& SLAB_RED_ZONE
) {
2390 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2391 slab_error(cachep
, "constructor overwrote the end of an object");
2392 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2393 slab_error(cachep
, "constructor overwrote the start of an object");
2395 /* need to poison the objs? */
2396 if (cachep
->flags
& SLAB_POISON
) {
2397 poison_obj(cachep
, objp
, POISON_FREE
);
2398 slab_kernel_map(cachep
, objp
, 0);
2404 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2405 /* Hold information during a freelist initialization */
2406 union freelist_init_state
{
2412 struct rnd_state rnd_state
;
2416 * Initialize the state based on the randomization methode available.
2417 * return true if the pre-computed list is available, false otherwize.
2419 static bool freelist_state_initialize(union freelist_init_state
*state
,
2420 struct kmem_cache
*cachep
,
2426 /* Use best entropy available to define a random shift */
2427 rand
= get_random_int();
2429 /* Use a random state if the pre-computed list is not available */
2430 if (!cachep
->random_seq
) {
2431 prandom_seed_state(&state
->rnd_state
, rand
);
2434 state
->list
= cachep
->random_seq
;
2435 state
->count
= count
;
2436 state
->pos
= rand
% count
;
2442 /* Get the next entry on the list and randomize it using a random shift */
2443 static freelist_idx_t
next_random_slot(union freelist_init_state
*state
)
2445 if (state
->pos
>= state
->count
)
2447 return state
->list
[state
->pos
++];
2450 /* Swap two freelist entries */
2451 static void swap_free_obj(struct page
*page
, unsigned int a
, unsigned int b
)
2453 swap(((freelist_idx_t
*)page
->freelist
)[a
],
2454 ((freelist_idx_t
*)page
->freelist
)[b
]);
2458 * Shuffle the freelist initialization state based on pre-computed lists.
2459 * return true if the list was successfully shuffled, false otherwise.
2461 static bool shuffle_freelist(struct kmem_cache
*cachep
, struct page
*page
)
2463 unsigned int objfreelist
= 0, i
, rand
, count
= cachep
->num
;
2464 union freelist_init_state state
;
2470 precomputed
= freelist_state_initialize(&state
, cachep
, count
);
2472 /* Take a random entry as the objfreelist */
2473 if (OBJFREELIST_SLAB(cachep
)) {
2475 objfreelist
= count
- 1;
2477 objfreelist
= next_random_slot(&state
);
2478 page
->freelist
= index_to_obj(cachep
, page
, objfreelist
) +
2484 * On early boot, generate the list dynamically.
2485 * Later use a pre-computed list for speed.
2488 for (i
= 0; i
< count
; i
++)
2489 set_free_obj(page
, i
, i
);
2491 /* Fisher-Yates shuffle */
2492 for (i
= count
- 1; i
> 0; i
--) {
2493 rand
= prandom_u32_state(&state
.rnd_state
);
2495 swap_free_obj(page
, i
, rand
);
2498 for (i
= 0; i
< count
; i
++)
2499 set_free_obj(page
, i
, next_random_slot(&state
));
2502 if (OBJFREELIST_SLAB(cachep
))
2503 set_free_obj(page
, cachep
->num
- 1, objfreelist
);
2508 static inline bool shuffle_freelist(struct kmem_cache
*cachep
,
2513 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2515 static void cache_init_objs(struct kmem_cache
*cachep
,
2522 cache_init_objs_debug(cachep
, page
);
2524 /* Try to randomize the freelist if enabled */
2525 shuffled
= shuffle_freelist(cachep
, page
);
2527 if (!shuffled
&& OBJFREELIST_SLAB(cachep
)) {
2528 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2532 for (i
= 0; i
< cachep
->num
; i
++) {
2533 objp
= index_to_obj(cachep
, page
, i
);
2534 objp
= kasan_init_slab_obj(cachep
, objp
);
2536 /* constructor could break poison info */
2537 if (DEBUG
== 0 && cachep
->ctor
) {
2538 kasan_unpoison_object_data(cachep
, objp
);
2540 kasan_poison_object_data(cachep
, objp
);
2544 set_free_obj(page
, i
, i
);
2548 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2552 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2556 if (cachep
->flags
& SLAB_STORE_USER
)
2557 set_store_user_dirty(cachep
);
2563 static void slab_put_obj(struct kmem_cache
*cachep
,
2564 struct page
*page
, void *objp
)
2566 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2570 /* Verify double free bug */
2571 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2572 if (get_free_obj(page
, i
) == objnr
) {
2573 pr_err("slab: double free detected in cache '%s', objp %px\n",
2574 cachep
->name
, objp
);
2580 if (!page
->freelist
)
2581 page
->freelist
= objp
+ obj_offset(cachep
);
2583 set_free_obj(page
, page
->active
, objnr
);
2587 * Map pages beginning at addr to the given cache and slab. This is required
2588 * for the slab allocator to be able to lookup the cache and slab of a
2589 * virtual address for kfree, ksize, and slab debugging.
2591 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2594 page
->slab_cache
= cache
;
2595 page
->freelist
= freelist
;
2599 * Grow (by 1) the number of slabs within a cache. This is called by
2600 * kmem_cache_alloc() when there are no active objs left in a cache.
2602 static struct page
*cache_grow_begin(struct kmem_cache
*cachep
,
2603 gfp_t flags
, int nodeid
)
2609 struct kmem_cache_node
*n
;
2613 * Be lazy and only check for valid flags here, keeping it out of the
2614 * critical path in kmem_cache_alloc().
2616 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2617 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
2618 flags
&= ~GFP_SLAB_BUG_MASK
;
2619 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2620 invalid_mask
, &invalid_mask
, flags
, &flags
);
2623 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2624 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2627 if (gfpflags_allow_blocking(local_flags
))
2631 * Get mem for the objs. Attempt to allocate a physical page from
2634 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2638 page_node
= page_to_nid(page
);
2639 n
= get_node(cachep
, page_node
);
2641 /* Get colour for the slab, and cal the next value. */
2643 if (n
->colour_next
>= cachep
->colour
)
2646 offset
= n
->colour_next
;
2647 if (offset
>= cachep
->colour
)
2650 offset
*= cachep
->colour_off
;
2653 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2654 * page_address() in the latter returns a non-tagged pointer,
2655 * as it should be for slab pages.
2657 kasan_poison_slab(page
);
2659 /* Get slab management. */
2660 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2661 local_flags
& ~GFP_CONSTRAINT_MASK
, page_node
);
2662 if (OFF_SLAB(cachep
) && !freelist
)
2665 slab_map_pages(cachep
, page
, freelist
);
2667 cache_init_objs(cachep
, page
);
2669 if (gfpflags_allow_blocking(local_flags
))
2670 local_irq_disable();
2675 kmem_freepages(cachep
, page
);
2677 if (gfpflags_allow_blocking(local_flags
))
2678 local_irq_disable();
2682 static void cache_grow_end(struct kmem_cache
*cachep
, struct page
*page
)
2684 struct kmem_cache_node
*n
;
2692 INIT_LIST_HEAD(&page
->slab_list
);
2693 n
= get_node(cachep
, page_to_nid(page
));
2695 spin_lock(&n
->list_lock
);
2697 if (!page
->active
) {
2698 list_add_tail(&page
->slab_list
, &n
->slabs_free
);
2701 fixup_slab_list(cachep
, n
, page
, &list
);
2703 STATS_INC_GROWN(cachep
);
2704 n
->free_objects
+= cachep
->num
- page
->active
;
2705 spin_unlock(&n
->list_lock
);
2707 fixup_objfreelist_debug(cachep
, &list
);
2713 * Perform extra freeing checks:
2714 * - detect bad pointers.
2715 * - POISON/RED_ZONE checking
2717 static void kfree_debugcheck(const void *objp
)
2719 if (!virt_addr_valid(objp
)) {
2720 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2721 (unsigned long)objp
);
2726 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2728 unsigned long long redzone1
, redzone2
;
2730 redzone1
= *dbg_redzone1(cache
, obj
);
2731 redzone2
= *dbg_redzone2(cache
, obj
);
2736 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2739 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2740 slab_error(cache
, "double free detected");
2742 slab_error(cache
, "memory outside object was overwritten");
2744 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2745 obj
, redzone1
, redzone2
);
2748 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2749 unsigned long caller
)
2754 BUG_ON(virt_to_cache(objp
) != cachep
);
2756 objp
-= obj_offset(cachep
);
2757 kfree_debugcheck(objp
);
2758 page
= virt_to_head_page(objp
);
2760 if (cachep
->flags
& SLAB_RED_ZONE
) {
2761 verify_redzone_free(cachep
, objp
);
2762 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2763 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2765 if (cachep
->flags
& SLAB_STORE_USER
) {
2766 set_store_user_dirty(cachep
);
2767 *dbg_userword(cachep
, objp
) = (void *)caller
;
2770 objnr
= obj_to_index(cachep
, page
, objp
);
2772 BUG_ON(objnr
>= cachep
->num
);
2773 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2775 if (cachep
->flags
& SLAB_POISON
) {
2776 poison_obj(cachep
, objp
, POISON_FREE
);
2777 slab_kernel_map(cachep
, objp
, 0);
2783 #define kfree_debugcheck(x) do { } while(0)
2784 #define cache_free_debugcheck(x,objp,z) (objp)
2787 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2795 objp
= next
- obj_offset(cachep
);
2796 next
= *(void **)next
;
2797 poison_obj(cachep
, objp
, POISON_FREE
);
2802 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2803 struct kmem_cache_node
*n
, struct page
*page
,
2806 /* move slabp to correct slabp list: */
2807 list_del(&page
->slab_list
);
2808 if (page
->active
== cachep
->num
) {
2809 list_add(&page
->slab_list
, &n
->slabs_full
);
2810 if (OBJFREELIST_SLAB(cachep
)) {
2812 /* Poisoning will be done without holding the lock */
2813 if (cachep
->flags
& SLAB_POISON
) {
2814 void **objp
= page
->freelist
;
2820 page
->freelist
= NULL
;
2823 list_add(&page
->slab_list
, &n
->slabs_partial
);
2826 /* Try to find non-pfmemalloc slab if needed */
2827 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2828 struct page
*page
, bool pfmemalloc
)
2836 if (!PageSlabPfmemalloc(page
))
2839 /* No need to keep pfmemalloc slab if we have enough free objects */
2840 if (n
->free_objects
> n
->free_limit
) {
2841 ClearPageSlabPfmemalloc(page
);
2845 /* Move pfmemalloc slab to the end of list to speed up next search */
2846 list_del(&page
->slab_list
);
2847 if (!page
->active
) {
2848 list_add_tail(&page
->slab_list
, &n
->slabs_free
);
2851 list_add_tail(&page
->slab_list
, &n
->slabs_partial
);
2853 list_for_each_entry(page
, &n
->slabs_partial
, slab_list
) {
2854 if (!PageSlabPfmemalloc(page
))
2858 n
->free_touched
= 1;
2859 list_for_each_entry(page
, &n
->slabs_free
, slab_list
) {
2860 if (!PageSlabPfmemalloc(page
)) {
2869 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2873 assert_spin_locked(&n
->list_lock
);
2874 page
= list_first_entry_or_null(&n
->slabs_partial
, struct page
,
2877 n
->free_touched
= 1;
2878 page
= list_first_entry_or_null(&n
->slabs_free
, struct page
,
2884 if (sk_memalloc_socks())
2885 page
= get_valid_first_slab(n
, page
, pfmemalloc
);
2890 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2891 struct kmem_cache_node
*n
, gfp_t flags
)
2897 if (!gfp_pfmemalloc_allowed(flags
))
2900 spin_lock(&n
->list_lock
);
2901 page
= get_first_slab(n
, true);
2903 spin_unlock(&n
->list_lock
);
2907 obj
= slab_get_obj(cachep
, page
);
2910 fixup_slab_list(cachep
, n
, page
, &list
);
2912 spin_unlock(&n
->list_lock
);
2913 fixup_objfreelist_debug(cachep
, &list
);
2919 * Slab list should be fixed up by fixup_slab_list() for existing slab
2920 * or cache_grow_end() for new slab
2922 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2923 struct array_cache
*ac
, struct page
*page
, int batchcount
)
2926 * There must be at least one object available for
2929 BUG_ON(page
->active
>= cachep
->num
);
2931 while (page
->active
< cachep
->num
&& batchcount
--) {
2932 STATS_INC_ALLOCED(cachep
);
2933 STATS_INC_ACTIVE(cachep
);
2934 STATS_SET_HIGH(cachep
);
2936 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2942 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2945 struct kmem_cache_node
*n
;
2946 struct array_cache
*ac
, *shared
;
2952 node
= numa_mem_id();
2954 ac
= cpu_cache_get(cachep
);
2955 batchcount
= ac
->batchcount
;
2956 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2958 * If there was little recent activity on this cache, then
2959 * perform only a partial refill. Otherwise we could generate
2962 batchcount
= BATCHREFILL_LIMIT
;
2964 n
= get_node(cachep
, node
);
2966 BUG_ON(ac
->avail
> 0 || !n
);
2967 shared
= READ_ONCE(n
->shared
);
2968 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
2971 spin_lock(&n
->list_lock
);
2972 shared
= READ_ONCE(n
->shared
);
2974 /* See if we can refill from the shared array */
2975 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
2976 shared
->touched
= 1;
2980 while (batchcount
> 0) {
2981 /* Get slab alloc is to come from. */
2982 page
= get_first_slab(n
, false);
2986 check_spinlock_acquired(cachep
);
2988 batchcount
= alloc_block(cachep
, ac
, page
, batchcount
);
2989 fixup_slab_list(cachep
, n
, page
, &list
);
2993 n
->free_objects
-= ac
->avail
;
2995 spin_unlock(&n
->list_lock
);
2996 fixup_objfreelist_debug(cachep
, &list
);
2999 if (unlikely(!ac
->avail
)) {
3000 /* Check if we can use obj in pfmemalloc slab */
3001 if (sk_memalloc_socks()) {
3002 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
3008 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
3011 * cache_grow_begin() can reenable interrupts,
3012 * then ac could change.
3014 ac
= cpu_cache_get(cachep
);
3015 if (!ac
->avail
&& page
)
3016 alloc_block(cachep
, ac
, page
, batchcount
);
3017 cache_grow_end(cachep
, page
);
3024 return ac
->entry
[--ac
->avail
];
3027 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3030 might_sleep_if(gfpflags_allow_blocking(flags
));
3034 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3035 gfp_t flags
, void *objp
, unsigned long caller
)
3037 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
3040 if (cachep
->flags
& SLAB_POISON
) {
3041 check_poison_obj(cachep
, objp
);
3042 slab_kernel_map(cachep
, objp
, 1);
3043 poison_obj(cachep
, objp
, POISON_INUSE
);
3045 if (cachep
->flags
& SLAB_STORE_USER
)
3046 *dbg_userword(cachep
, objp
) = (void *)caller
;
3048 if (cachep
->flags
& SLAB_RED_ZONE
) {
3049 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3050 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3051 slab_error(cachep
, "double free, or memory outside object was overwritten");
3052 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3053 objp
, *dbg_redzone1(cachep
, objp
),
3054 *dbg_redzone2(cachep
, objp
));
3056 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3057 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3060 objp
+= obj_offset(cachep
);
3061 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3063 if (ARCH_SLAB_MINALIGN
&&
3064 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3065 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3066 objp
, (int)ARCH_SLAB_MINALIGN
);
3071 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3074 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3077 struct array_cache
*ac
;
3081 ac
= cpu_cache_get(cachep
);
3082 if (likely(ac
->avail
)) {
3084 objp
= ac
->entry
[--ac
->avail
];
3086 STATS_INC_ALLOCHIT(cachep
);
3090 STATS_INC_ALLOCMISS(cachep
);
3091 objp
= cache_alloc_refill(cachep
, flags
);
3093 * the 'ac' may be updated by cache_alloc_refill(),
3094 * and kmemleak_erase() requires its correct value.
3096 ac
= cpu_cache_get(cachep
);
3100 * To avoid a false negative, if an object that is in one of the
3101 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3102 * treat the array pointers as a reference to the object.
3105 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3111 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3113 * If we are in_interrupt, then process context, including cpusets and
3114 * mempolicy, may not apply and should not be used for allocation policy.
3116 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3118 int nid_alloc
, nid_here
;
3120 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3122 nid_alloc
= nid_here
= numa_mem_id();
3123 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3124 nid_alloc
= cpuset_slab_spread_node();
3125 else if (current
->mempolicy
)
3126 nid_alloc
= mempolicy_slab_node();
3127 if (nid_alloc
!= nid_here
)
3128 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3133 * Fallback function if there was no memory available and no objects on a
3134 * certain node and fall back is permitted. First we scan all the
3135 * available node for available objects. If that fails then we
3136 * perform an allocation without specifying a node. This allows the page
3137 * allocator to do its reclaim / fallback magic. We then insert the
3138 * slab into the proper nodelist and then allocate from it.
3140 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3142 struct zonelist
*zonelist
;
3145 enum zone_type high_zoneidx
= gfp_zone(flags
);
3149 unsigned int cpuset_mems_cookie
;
3151 if (flags
& __GFP_THISNODE
)
3155 cpuset_mems_cookie
= read_mems_allowed_begin();
3156 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3160 * Look through allowed nodes for objects available
3161 * from existing per node queues.
3163 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3164 nid
= zone_to_nid(zone
);
3166 if (cpuset_zone_allowed(zone
, flags
) &&
3167 get_node(cache
, nid
) &&
3168 get_node(cache
, nid
)->free_objects
) {
3169 obj
= ____cache_alloc_node(cache
,
3170 gfp_exact_node(flags
), nid
);
3178 * This allocation will be performed within the constraints
3179 * of the current cpuset / memory policy requirements.
3180 * We may trigger various forms of reclaim on the allowed
3181 * set and go into memory reserves if necessary.
3183 page
= cache_grow_begin(cache
, flags
, numa_mem_id());
3184 cache_grow_end(cache
, page
);
3186 nid
= page_to_nid(page
);
3187 obj
= ____cache_alloc_node(cache
,
3188 gfp_exact_node(flags
), nid
);
3191 * Another processor may allocate the objects in
3192 * the slab since we are not holding any locks.
3199 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3205 * A interface to enable slab creation on nodeid
3207 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3211 struct kmem_cache_node
*n
;
3215 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3216 n
= get_node(cachep
, nodeid
);
3220 spin_lock(&n
->list_lock
);
3221 page
= get_first_slab(n
, false);
3225 check_spinlock_acquired_node(cachep
, nodeid
);
3227 STATS_INC_NODEALLOCS(cachep
);
3228 STATS_INC_ACTIVE(cachep
);
3229 STATS_SET_HIGH(cachep
);
3231 BUG_ON(page
->active
== cachep
->num
);
3233 obj
= slab_get_obj(cachep
, page
);
3236 fixup_slab_list(cachep
, n
, page
, &list
);
3238 spin_unlock(&n
->list_lock
);
3239 fixup_objfreelist_debug(cachep
, &list
);
3243 spin_unlock(&n
->list_lock
);
3244 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3246 /* This slab isn't counted yet so don't update free_objects */
3247 obj
= slab_get_obj(cachep
, page
);
3249 cache_grow_end(cachep
, page
);
3251 return obj
? obj
: fallback_alloc(cachep
, flags
);
3254 static __always_inline
void *
3255 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3256 unsigned long caller
)
3258 unsigned long save_flags
;
3260 int slab_node
= numa_mem_id();
3262 flags
&= gfp_allowed_mask
;
3263 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3264 if (unlikely(!cachep
))
3267 cache_alloc_debugcheck_before(cachep
, flags
);
3268 local_irq_save(save_flags
);
3270 if (nodeid
== NUMA_NO_NODE
)
3273 if (unlikely(!get_node(cachep
, nodeid
))) {
3274 /* Node not bootstrapped yet */
3275 ptr
= fallback_alloc(cachep
, flags
);
3279 if (nodeid
== slab_node
) {
3281 * Use the locally cached objects if possible.
3282 * However ____cache_alloc does not allow fallback
3283 * to other nodes. It may fail while we still have
3284 * objects on other nodes available.
3286 ptr
= ____cache_alloc(cachep
, flags
);
3290 /* ___cache_alloc_node can fall back to other nodes */
3291 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3293 local_irq_restore(save_flags
);
3294 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3296 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3297 memset(ptr
, 0, cachep
->object_size
);
3299 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3303 static __always_inline
void *
3304 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3308 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3309 objp
= alternate_node_alloc(cache
, flags
);
3313 objp
= ____cache_alloc(cache
, flags
);
3316 * We may just have run out of memory on the local node.
3317 * ____cache_alloc_node() knows how to locate memory on other nodes
3320 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3327 static __always_inline
void *
3328 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3330 return ____cache_alloc(cachep
, flags
);
3333 #endif /* CONFIG_NUMA */
3335 static __always_inline
void *
3336 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3338 unsigned long save_flags
;
3341 flags
&= gfp_allowed_mask
;
3342 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3343 if (unlikely(!cachep
))
3346 cache_alloc_debugcheck_before(cachep
, flags
);
3347 local_irq_save(save_flags
);
3348 objp
= __do_cache_alloc(cachep
, flags
);
3349 local_irq_restore(save_flags
);
3350 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3353 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3354 memset(objp
, 0, cachep
->object_size
);
3356 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3361 * Caller needs to acquire correct kmem_cache_node's list_lock
3362 * @list: List of detached free slabs should be freed by caller
3364 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3365 int nr_objects
, int node
, struct list_head
*list
)
3368 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3371 n
->free_objects
+= nr_objects
;
3373 for (i
= 0; i
< nr_objects
; i
++) {
3379 page
= virt_to_head_page(objp
);
3380 list_del(&page
->slab_list
);
3381 check_spinlock_acquired_node(cachep
, node
);
3382 slab_put_obj(cachep
, page
, objp
);
3383 STATS_DEC_ACTIVE(cachep
);
3385 /* fixup slab chains */
3386 if (page
->active
== 0) {
3387 list_add(&page
->slab_list
, &n
->slabs_free
);
3390 /* Unconditionally move a slab to the end of the
3391 * partial list on free - maximum time for the
3392 * other objects to be freed, too.
3394 list_add_tail(&page
->slab_list
, &n
->slabs_partial
);
3398 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3399 n
->free_objects
-= cachep
->num
;
3401 page
= list_last_entry(&n
->slabs_free
, struct page
, slab_list
);
3402 list_move(&page
->slab_list
, list
);
3408 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3411 struct kmem_cache_node
*n
;
3412 int node
= numa_mem_id();
3415 batchcount
= ac
->batchcount
;
3418 n
= get_node(cachep
, node
);
3419 spin_lock(&n
->list_lock
);
3421 struct array_cache
*shared_array
= n
->shared
;
3422 int max
= shared_array
->limit
- shared_array
->avail
;
3424 if (batchcount
> max
)
3426 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3427 ac
->entry
, sizeof(void *) * batchcount
);
3428 shared_array
->avail
+= batchcount
;
3433 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3440 list_for_each_entry(page
, &n
->slabs_free
, slab_list
) {
3441 BUG_ON(page
->active
);
3445 STATS_SET_FREEABLE(cachep
, i
);
3448 spin_unlock(&n
->list_lock
);
3449 slabs_destroy(cachep
, &list
);
3450 ac
->avail
-= batchcount
;
3451 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3455 * Release an obj back to its cache. If the obj has a constructed state, it must
3456 * be in this state _before_ it is released. Called with disabled ints.
3458 static __always_inline
void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3459 unsigned long caller
)
3461 /* Put the object into the quarantine, don't touch it for now. */
3462 if (kasan_slab_free(cachep
, objp
, _RET_IP_
))
3465 ___cache_free(cachep
, objp
, caller
);
3468 void ___cache_free(struct kmem_cache
*cachep
, void *objp
,
3469 unsigned long caller
)
3471 struct array_cache
*ac
= cpu_cache_get(cachep
);
3474 kmemleak_free_recursive(objp
, cachep
->flags
);
3475 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3478 * Skip calling cache_free_alien() when the platform is not numa.
3479 * This will avoid cache misses that happen while accessing slabp (which
3480 * is per page memory reference) to get nodeid. Instead use a global
3481 * variable to skip the call, which is mostly likely to be present in
3484 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3487 if (ac
->avail
< ac
->limit
) {
3488 STATS_INC_FREEHIT(cachep
);
3490 STATS_INC_FREEMISS(cachep
);
3491 cache_flusharray(cachep
, ac
);
3494 if (sk_memalloc_socks()) {
3495 struct page
*page
= virt_to_head_page(objp
);
3497 if (unlikely(PageSlabPfmemalloc(page
))) {
3498 cache_free_pfmemalloc(cachep
, page
, objp
);
3503 ac
->entry
[ac
->avail
++] = objp
;
3507 * kmem_cache_alloc - Allocate an object
3508 * @cachep: The cache to allocate from.
3509 * @flags: See kmalloc().
3511 * Allocate an object from this cache. The flags are only relevant
3512 * if the cache has no available objects.
3514 * Return: pointer to the new object or %NULL in case of error
3516 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3518 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3520 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3521 cachep
->object_size
, cachep
->size
, flags
);
3525 EXPORT_SYMBOL(kmem_cache_alloc
);
3527 static __always_inline
void
3528 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3529 size_t size
, void **p
, unsigned long caller
)
3533 for (i
= 0; i
< size
; i
++)
3534 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3537 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3542 s
= slab_pre_alloc_hook(s
, flags
);
3546 cache_alloc_debugcheck_before(s
, flags
);
3548 local_irq_disable();
3549 for (i
= 0; i
< size
; i
++) {
3550 void *objp
= __do_cache_alloc(s
, flags
);
3552 if (unlikely(!objp
))
3558 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3560 /* Clear memory outside IRQ disabled section */
3561 if (unlikely(flags
& __GFP_ZERO
))
3562 for (i
= 0; i
< size
; i
++)
3563 memset(p
[i
], 0, s
->object_size
);
3565 slab_post_alloc_hook(s
, flags
, size
, p
);
3566 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3570 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3571 slab_post_alloc_hook(s
, flags
, i
, p
);
3572 __kmem_cache_free_bulk(s
, i
, p
);
3575 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3577 #ifdef CONFIG_TRACING
3579 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3583 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3585 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3586 trace_kmalloc(_RET_IP_
, ret
,
3587 size
, cachep
->size
, flags
);
3590 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3595 * kmem_cache_alloc_node - Allocate an object on the specified node
3596 * @cachep: The cache to allocate from.
3597 * @flags: See kmalloc().
3598 * @nodeid: node number of the target node.
3600 * Identical to kmem_cache_alloc but it will allocate memory on the given
3601 * node, which can improve the performance for cpu bound structures.
3603 * Fallback to other node is possible if __GFP_THISNODE is not set.
3605 * Return: pointer to the new object or %NULL in case of error
3607 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3609 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3611 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3612 cachep
->object_size
, cachep
->size
,
3617 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3619 #ifdef CONFIG_TRACING
3620 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3627 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3629 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3630 trace_kmalloc_node(_RET_IP_
, ret
,
3635 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3638 static __always_inline
void *
3639 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3641 struct kmem_cache
*cachep
;
3644 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3646 cachep
= kmalloc_slab(size
, flags
);
3647 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3649 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3650 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3655 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3657 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3659 EXPORT_SYMBOL(__kmalloc_node
);
3661 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3662 int node
, unsigned long caller
)
3664 return __do_kmalloc_node(size
, flags
, node
, caller
);
3666 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3667 #endif /* CONFIG_NUMA */
3670 * __do_kmalloc - allocate memory
3671 * @size: how many bytes of memory are required.
3672 * @flags: the type of memory to allocate (see kmalloc).
3673 * @caller: function caller for debug tracking of the caller
3675 * Return: pointer to the allocated memory or %NULL in case of error
3677 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3678 unsigned long caller
)
3680 struct kmem_cache
*cachep
;
3683 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3685 cachep
= kmalloc_slab(size
, flags
);
3686 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3688 ret
= slab_alloc(cachep
, flags
, caller
);
3690 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3691 trace_kmalloc(caller
, ret
,
3692 size
, cachep
->size
, flags
);
3697 void *__kmalloc(size_t size
, gfp_t flags
)
3699 return __do_kmalloc(size
, flags
, _RET_IP_
);
3701 EXPORT_SYMBOL(__kmalloc
);
3703 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3705 return __do_kmalloc(size
, flags
, caller
);
3707 EXPORT_SYMBOL(__kmalloc_track_caller
);
3710 * kmem_cache_free - Deallocate an object
3711 * @cachep: The cache the allocation was from.
3712 * @objp: The previously allocated object.
3714 * Free an object which was previously allocated from this
3717 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3719 unsigned long flags
;
3720 cachep
= cache_from_obj(cachep
, objp
);
3724 local_irq_save(flags
);
3725 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3726 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3727 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3728 __cache_free(cachep
, objp
, _RET_IP_
);
3729 local_irq_restore(flags
);
3731 trace_kmem_cache_free(_RET_IP_
, objp
);
3733 EXPORT_SYMBOL(kmem_cache_free
);
3735 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3737 struct kmem_cache
*s
;
3740 local_irq_disable();
3741 for (i
= 0; i
< size
; i
++) {
3744 if (!orig_s
) /* called via kfree_bulk */
3745 s
= virt_to_cache(objp
);
3747 s
= cache_from_obj(orig_s
, objp
);
3749 debug_check_no_locks_freed(objp
, s
->object_size
);
3750 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3751 debug_check_no_obj_freed(objp
, s
->object_size
);
3753 __cache_free(s
, objp
, _RET_IP_
);
3757 /* FIXME: add tracing */
3759 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3762 * kfree - free previously allocated memory
3763 * @objp: pointer returned by kmalloc.
3765 * If @objp is NULL, no operation is performed.
3767 * Don't free memory not originally allocated by kmalloc()
3768 * or you will run into trouble.
3770 void kfree(const void *objp
)
3772 struct kmem_cache
*c
;
3773 unsigned long flags
;
3775 trace_kfree(_RET_IP_
, objp
);
3777 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3779 local_irq_save(flags
);
3780 kfree_debugcheck(objp
);
3781 c
= virt_to_cache(objp
);
3782 debug_check_no_locks_freed(objp
, c
->object_size
);
3784 debug_check_no_obj_freed(objp
, c
->object_size
);
3785 __cache_free(c
, (void *)objp
, _RET_IP_
);
3786 local_irq_restore(flags
);
3788 EXPORT_SYMBOL(kfree
);
3791 * This initializes kmem_cache_node or resizes various caches for all nodes.
3793 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3797 struct kmem_cache_node
*n
;
3799 for_each_online_node(node
) {
3800 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3809 if (!cachep
->list
.next
) {
3810 /* Cache is not active yet. Roll back what we did */
3813 n
= get_node(cachep
, node
);
3816 free_alien_cache(n
->alien
);
3818 cachep
->node
[node
] = NULL
;
3826 /* Always called with the slab_mutex held */
3827 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3828 int batchcount
, int shared
, gfp_t gfp
)
3830 struct array_cache __percpu
*cpu_cache
, *prev
;
3833 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3837 prev
= cachep
->cpu_cache
;
3838 cachep
->cpu_cache
= cpu_cache
;
3840 * Without a previous cpu_cache there's no need to synchronize remote
3841 * cpus, so skip the IPIs.
3844 kick_all_cpus_sync();
3847 cachep
->batchcount
= batchcount
;
3848 cachep
->limit
= limit
;
3849 cachep
->shared
= shared
;
3854 for_each_online_cpu(cpu
) {
3857 struct kmem_cache_node
*n
;
3858 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3860 node
= cpu_to_mem(cpu
);
3861 n
= get_node(cachep
, node
);
3862 spin_lock_irq(&n
->list_lock
);
3863 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3864 spin_unlock_irq(&n
->list_lock
);
3865 slabs_destroy(cachep
, &list
);
3870 return setup_kmem_cache_nodes(cachep
, gfp
);
3873 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3874 int batchcount
, int shared
, gfp_t gfp
)
3877 struct kmem_cache
*c
;
3879 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3881 if (slab_state
< FULL
)
3884 if ((ret
< 0) || !is_root_cache(cachep
))
3887 lockdep_assert_held(&slab_mutex
);
3888 for_each_memcg_cache(c
, cachep
) {
3889 /* return value determined by the root cache only */
3890 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3896 /* Called with slab_mutex held always */
3897 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3904 err
= cache_random_seq_create(cachep
, cachep
->num
, gfp
);
3908 if (!is_root_cache(cachep
)) {
3909 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3910 limit
= root
->limit
;
3911 shared
= root
->shared
;
3912 batchcount
= root
->batchcount
;
3915 if (limit
&& shared
&& batchcount
)
3918 * The head array serves three purposes:
3919 * - create a LIFO ordering, i.e. return objects that are cache-warm
3920 * - reduce the number of spinlock operations.
3921 * - reduce the number of linked list operations on the slab and
3922 * bufctl chains: array operations are cheaper.
3923 * The numbers are guessed, we should auto-tune as described by
3926 if (cachep
->size
> 131072)
3928 else if (cachep
->size
> PAGE_SIZE
)
3930 else if (cachep
->size
> 1024)
3932 else if (cachep
->size
> 256)
3938 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3939 * allocation behaviour: Most allocs on one cpu, most free operations
3940 * on another cpu. For these cases, an efficient object passing between
3941 * cpus is necessary. This is provided by a shared array. The array
3942 * replaces Bonwick's magazine layer.
3943 * On uniprocessor, it's functionally equivalent (but less efficient)
3944 * to a larger limit. Thus disabled by default.
3947 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3952 * With debugging enabled, large batchcount lead to excessively long
3953 * periods with disabled local interrupts. Limit the batchcount
3958 batchcount
= (limit
+ 1) / 2;
3960 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3963 pr_err("enable_cpucache failed for %s, error %d\n",
3964 cachep
->name
, -err
);
3969 * Drain an array if it contains any elements taking the node lock only if
3970 * necessary. Note that the node listlock also protects the array_cache
3971 * if drain_array() is used on the shared array.
3973 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3974 struct array_cache
*ac
, int node
)
3978 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3979 check_mutex_acquired();
3981 if (!ac
|| !ac
->avail
)
3989 spin_lock_irq(&n
->list_lock
);
3990 drain_array_locked(cachep
, ac
, node
, false, &list
);
3991 spin_unlock_irq(&n
->list_lock
);
3993 slabs_destroy(cachep
, &list
);
3997 * cache_reap - Reclaim memory from caches.
3998 * @w: work descriptor
4000 * Called from workqueue/eventd every few seconds.
4002 * - clear the per-cpu caches for this CPU.
4003 * - return freeable pages to the main free memory pool.
4005 * If we cannot acquire the cache chain mutex then just give up - we'll try
4006 * again on the next iteration.
4008 static void cache_reap(struct work_struct
*w
)
4010 struct kmem_cache
*searchp
;
4011 struct kmem_cache_node
*n
;
4012 int node
= numa_mem_id();
4013 struct delayed_work
*work
= to_delayed_work(w
);
4015 if (!mutex_trylock(&slab_mutex
))
4016 /* Give up. Setup the next iteration. */
4019 list_for_each_entry(searchp
, &slab_caches
, list
) {
4023 * We only take the node lock if absolutely necessary and we
4024 * have established with reasonable certainty that
4025 * we can do some work if the lock was obtained.
4027 n
= get_node(searchp
, node
);
4029 reap_alien(searchp
, n
);
4031 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
4034 * These are racy checks but it does not matter
4035 * if we skip one check or scan twice.
4037 if (time_after(n
->next_reap
, jiffies
))
4040 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
4042 drain_array(searchp
, n
, n
->shared
, node
);
4044 if (n
->free_touched
)
4045 n
->free_touched
= 0;
4049 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4050 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4051 STATS_ADD_REAPED(searchp
, freed
);
4057 mutex_unlock(&slab_mutex
);
4060 /* Set up the next iteration */
4061 schedule_delayed_work_on(smp_processor_id(), work
,
4062 round_jiffies_relative(REAPTIMEOUT_AC
));
4065 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4067 unsigned long active_objs
, num_objs
, active_slabs
;
4068 unsigned long total_slabs
= 0, free_objs
= 0, shared_avail
= 0;
4069 unsigned long free_slabs
= 0;
4071 struct kmem_cache_node
*n
;
4073 for_each_kmem_cache_node(cachep
, node
, n
) {
4075 spin_lock_irq(&n
->list_lock
);
4077 total_slabs
+= n
->total_slabs
;
4078 free_slabs
+= n
->free_slabs
;
4079 free_objs
+= n
->free_objects
;
4082 shared_avail
+= n
->shared
->avail
;
4084 spin_unlock_irq(&n
->list_lock
);
4086 num_objs
= total_slabs
* cachep
->num
;
4087 active_slabs
= total_slabs
- free_slabs
;
4088 active_objs
= num_objs
- free_objs
;
4090 sinfo
->active_objs
= active_objs
;
4091 sinfo
->num_objs
= num_objs
;
4092 sinfo
->active_slabs
= active_slabs
;
4093 sinfo
->num_slabs
= total_slabs
;
4094 sinfo
->shared_avail
= shared_avail
;
4095 sinfo
->limit
= cachep
->limit
;
4096 sinfo
->batchcount
= cachep
->batchcount
;
4097 sinfo
->shared
= cachep
->shared
;
4098 sinfo
->objects_per_slab
= cachep
->num
;
4099 sinfo
->cache_order
= cachep
->gfporder
;
4102 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4106 unsigned long high
= cachep
->high_mark
;
4107 unsigned long allocs
= cachep
->num_allocations
;
4108 unsigned long grown
= cachep
->grown
;
4109 unsigned long reaped
= cachep
->reaped
;
4110 unsigned long errors
= cachep
->errors
;
4111 unsigned long max_freeable
= cachep
->max_freeable
;
4112 unsigned long node_allocs
= cachep
->node_allocs
;
4113 unsigned long node_frees
= cachep
->node_frees
;
4114 unsigned long overflows
= cachep
->node_overflow
;
4116 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4117 allocs
, high
, grown
,
4118 reaped
, errors
, max_freeable
, node_allocs
,
4119 node_frees
, overflows
);
4123 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4124 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4125 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4126 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4128 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4129 allochit
, allocmiss
, freehit
, freemiss
);
4134 #define MAX_SLABINFO_WRITE 128
4136 * slabinfo_write - Tuning for the slab allocator
4138 * @buffer: user buffer
4139 * @count: data length
4142 * Return: %0 on success, negative error code otherwise.
4144 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4145 size_t count
, loff_t
*ppos
)
4147 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4148 int limit
, batchcount
, shared
, res
;
4149 struct kmem_cache
*cachep
;
4151 if (count
> MAX_SLABINFO_WRITE
)
4153 if (copy_from_user(&kbuf
, buffer
, count
))
4155 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4157 tmp
= strchr(kbuf
, ' ');
4162 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4165 /* Find the cache in the chain of caches. */
4166 mutex_lock(&slab_mutex
);
4168 list_for_each_entry(cachep
, &slab_caches
, list
) {
4169 if (!strcmp(cachep
->name
, kbuf
)) {
4170 if (limit
< 1 || batchcount
< 1 ||
4171 batchcount
> limit
|| shared
< 0) {
4174 res
= do_tune_cpucache(cachep
, limit
,
4181 mutex_unlock(&slab_mutex
);
4187 #ifdef CONFIG_DEBUG_SLAB_LEAK
4189 static inline int add_caller(unsigned long *n
, unsigned long v
)
4199 unsigned long *q
= p
+ 2 * i
;
4213 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4219 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4228 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4231 for (j
= page
->active
; j
< c
->num
; j
++) {
4232 if (get_free_obj(page
, j
) == i
) {
4242 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4243 * mapping is established when actual object allocation and
4244 * we could mistakenly access the unmapped object in the cpu
4247 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4250 if (!add_caller(n
, v
))
4255 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4257 #ifdef CONFIG_KALLSYMS
4258 unsigned long offset
, size
;
4259 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4261 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4262 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4264 seq_printf(m
, " [%s]", modname
);
4268 seq_printf(m
, "%px", (void *)address
);
4271 static int leaks_show(struct seq_file
*m
, void *p
)
4273 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
,
4276 struct kmem_cache_node
*n
;
4278 unsigned long *x
= m
->private;
4282 if (!(cachep
->flags
& SLAB_STORE_USER
))
4284 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4288 * Set store_user_clean and start to grab stored user information
4289 * for all objects on this cache. If some alloc/free requests comes
4290 * during the processing, information would be wrong so restart
4294 drain_cpu_caches(cachep
);
4296 * drain_cpu_caches() could make kmemleak_object and
4297 * debug_objects_cache dirty, so reset afterwards.
4299 set_store_user_clean(cachep
);
4303 for_each_kmem_cache_node(cachep
, node
, n
) {
4306 spin_lock_irq(&n
->list_lock
);
4308 list_for_each_entry(page
, &n
->slabs_full
, slab_list
)
4309 handle_slab(x
, cachep
, page
);
4310 list_for_each_entry(page
, &n
->slabs_partial
, slab_list
)
4311 handle_slab(x
, cachep
, page
);
4312 spin_unlock_irq(&n
->list_lock
);
4314 } while (!is_store_user_clean(cachep
));
4316 name
= cachep
->name
;
4318 /* Increase the buffer size */
4319 mutex_unlock(&slab_mutex
);
4320 m
->private = kcalloc(x
[0] * 4, sizeof(unsigned long),
4323 /* Too bad, we are really out */
4325 mutex_lock(&slab_mutex
);
4328 *(unsigned long *)m
->private = x
[0] * 2;
4330 mutex_lock(&slab_mutex
);
4331 /* Now make sure this entry will be retried */
4335 for (i
= 0; i
< x
[1]; i
++) {
4336 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4337 show_symbol(m
, x
[2*i
+2]);
4344 static const struct seq_operations slabstats_op
= {
4345 .start
= slab_start
,
4351 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4355 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4359 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4364 static const struct file_operations proc_slabstats_operations
= {
4365 .open
= slabstats_open
,
4367 .llseek
= seq_lseek
,
4368 .release
= seq_release_private
,
4372 static int __init
slab_proc_init(void)
4374 #ifdef CONFIG_DEBUG_SLAB_LEAK
4375 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4379 module_init(slab_proc_init
);
4381 #ifdef CONFIG_HARDENED_USERCOPY
4383 * Rejects incorrectly sized objects and objects that are to be copied
4384 * to/from userspace but do not fall entirely within the containing slab
4385 * cache's usercopy region.
4387 * Returns NULL if check passes, otherwise const char * to name of cache
4388 * to indicate an error.
4390 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4393 struct kmem_cache
*cachep
;
4395 unsigned long offset
;
4397 ptr
= kasan_reset_tag(ptr
);
4399 /* Find and validate object. */
4400 cachep
= page
->slab_cache
;
4401 objnr
= obj_to_index(cachep
, page
, (void *)ptr
);
4402 BUG_ON(objnr
>= cachep
->num
);
4404 /* Find offset within object. */
4405 offset
= ptr
- index_to_obj(cachep
, page
, objnr
) - obj_offset(cachep
);
4407 /* Allow address range falling entirely within usercopy region. */
4408 if (offset
>= cachep
->useroffset
&&
4409 offset
- cachep
->useroffset
<= cachep
->usersize
&&
4410 n
<= cachep
->useroffset
- offset
+ cachep
->usersize
)
4414 * If the copy is still within the allocated object, produce
4415 * a warning instead of rejecting the copy. This is intended
4416 * to be a temporary method to find any missing usercopy
4419 if (usercopy_fallback
&&
4420 offset
<= cachep
->object_size
&&
4421 n
<= cachep
->object_size
- offset
) {
4422 usercopy_warn("SLAB object", cachep
->name
, to_user
, offset
, n
);
4426 usercopy_abort("SLAB object", cachep
->name
, to_user
, offset
, n
);
4428 #endif /* CONFIG_HARDENED_USERCOPY */
4431 * ksize - get the actual amount of memory allocated for a given object
4432 * @objp: Pointer to the object
4434 * kmalloc may internally round up allocations and return more memory
4435 * than requested. ksize() can be used to determine the actual amount of
4436 * memory allocated. The caller may use this additional memory, even though
4437 * a smaller amount of memory was initially specified with the kmalloc call.
4438 * The caller must guarantee that objp points to a valid object previously
4439 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4440 * must not be freed during the duration of the call.
4442 * Return: size of the actual memory used by @objp in bytes
4444 size_t ksize(const void *objp
)
4449 if (unlikely(objp
== ZERO_SIZE_PTR
))
4452 size
= virt_to_cache(objp
)->object_size
;
4453 /* We assume that ksize callers could use the whole allocated area,
4454 * so we need to unpoison this area.
4456 kasan_unpoison_shadow(objp
, size
);
4460 EXPORT_SYMBOL(ksize
);