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;})
366 * Do not go above this order unless 0 objects fit into the slab or
367 * overridden on the command line.
369 #define SLAB_MAX_ORDER_HI 1
370 #define SLAB_MAX_ORDER_LO 0
371 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
372 static bool slab_max_order_set __initdata
;
374 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
377 return page
->s_mem
+ cache
->size
* idx
;
380 #define BOOT_CPUCACHE_ENTRIES 1
381 /* internal cache of cache description objs */
382 static struct kmem_cache kmem_cache_boot
= {
384 .limit
= BOOT_CPUCACHE_ENTRIES
,
386 .size
= sizeof(struct kmem_cache
),
387 .name
= "kmem_cache",
390 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
392 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
394 return this_cpu_ptr(cachep
->cpu_cache
);
398 * Calculate the number of objects and left-over bytes for a given buffer size.
400 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
401 slab_flags_t flags
, size_t *left_over
)
404 size_t slab_size
= PAGE_SIZE
<< gfporder
;
407 * The slab management structure can be either off the slab or
408 * on it. For the latter case, the memory allocated for a
411 * - @buffer_size bytes for each object
412 * - One freelist_idx_t for each object
414 * We don't need to consider alignment of freelist because
415 * freelist will be at the end of slab page. The objects will be
416 * at the correct alignment.
418 * If the slab management structure is off the slab, then the
419 * alignment will already be calculated into the size. Because
420 * the slabs are all pages aligned, the objects will be at the
421 * correct alignment when allocated.
423 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
424 num
= slab_size
/ buffer_size
;
425 *left_over
= slab_size
% buffer_size
;
427 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
428 *left_over
= slab_size
%
429 (buffer_size
+ sizeof(freelist_idx_t
));
436 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
438 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
441 pr_err("slab error in %s(): cache `%s': %s\n",
442 function
, cachep
->name
, msg
);
444 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
449 * By default on NUMA we use alien caches to stage the freeing of
450 * objects allocated from other nodes. This causes massive memory
451 * inefficiencies when using fake NUMA setup to split memory into a
452 * large number of small nodes, so it can be disabled on the command
456 static int use_alien_caches __read_mostly
= 1;
457 static int __init
noaliencache_setup(char *s
)
459 use_alien_caches
= 0;
462 __setup("noaliencache", noaliencache_setup
);
464 static int __init
slab_max_order_setup(char *str
)
466 get_option(&str
, &slab_max_order
);
467 slab_max_order
= slab_max_order
< 0 ? 0 :
468 min(slab_max_order
, MAX_ORDER
- 1);
469 slab_max_order_set
= true;
473 __setup("slab_max_order=", slab_max_order_setup
);
477 * Special reaping functions for NUMA systems called from cache_reap().
478 * These take care of doing round robin flushing of alien caches (containing
479 * objects freed on different nodes from which they were allocated) and the
480 * flushing of remote pcps by calling drain_node_pages.
482 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
484 static void init_reap_node(int cpu
)
486 per_cpu(slab_reap_node
, cpu
) = next_node_in(cpu_to_mem(cpu
),
490 static void next_reap_node(void)
492 int node
= __this_cpu_read(slab_reap_node
);
494 node
= next_node_in(node
, node_online_map
);
495 __this_cpu_write(slab_reap_node
, node
);
499 #define init_reap_node(cpu) do { } while (0)
500 #define next_reap_node(void) do { } while (0)
504 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
505 * via the workqueue/eventd.
506 * Add the CPU number into the expiration time to minimize the possibility of
507 * the CPUs getting into lockstep and contending for the global cache chain
510 static void start_cpu_timer(int cpu
)
512 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
514 if (reap_work
->work
.func
== NULL
) {
516 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
517 schedule_delayed_work_on(cpu
, reap_work
,
518 __round_jiffies_relative(HZ
, cpu
));
522 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
527 ac
->batchcount
= batch
;
532 static struct array_cache
*alloc_arraycache(int node
, int entries
,
533 int batchcount
, gfp_t gfp
)
535 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
536 struct array_cache
*ac
= NULL
;
538 ac
= kmalloc_node(memsize
, gfp
, node
);
540 * The array_cache structures contain pointers to free object.
541 * However, when such objects are allocated or transferred to another
542 * cache the pointers are not cleared and they could be counted as
543 * valid references during a kmemleak scan. Therefore, kmemleak must
544 * not scan such objects.
546 kmemleak_no_scan(ac
);
547 init_arraycache(ac
, entries
, batchcount
);
551 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
552 struct page
*page
, void *objp
)
554 struct kmem_cache_node
*n
;
558 page_node
= page_to_nid(page
);
559 n
= get_node(cachep
, page_node
);
561 spin_lock(&n
->list_lock
);
562 free_block(cachep
, &objp
, 1, page_node
, &list
);
563 spin_unlock(&n
->list_lock
);
565 slabs_destroy(cachep
, &list
);
569 * Transfer objects in one arraycache to another.
570 * Locking must be handled by the caller.
572 * Return the number of entries transferred.
574 static int transfer_objects(struct array_cache
*to
,
575 struct array_cache
*from
, unsigned int max
)
577 /* Figure out how many entries to transfer */
578 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
583 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
593 #define drain_alien_cache(cachep, alien) do { } while (0)
594 #define reap_alien(cachep, n) do { } while (0)
596 static inline struct alien_cache
**alloc_alien_cache(int node
,
597 int limit
, gfp_t gfp
)
602 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
606 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
611 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
617 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
618 gfp_t flags
, int nodeid
)
623 static inline gfp_t
gfp_exact_node(gfp_t flags
)
625 return flags
& ~__GFP_NOFAIL
;
628 #else /* CONFIG_NUMA */
630 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
631 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
633 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
634 int batch
, gfp_t gfp
)
636 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
637 struct alien_cache
*alc
= NULL
;
639 alc
= kmalloc_node(memsize
, gfp
, node
);
641 kmemleak_no_scan(alc
);
642 init_arraycache(&alc
->ac
, entries
, batch
);
643 spin_lock_init(&alc
->lock
);
648 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
650 struct alien_cache
**alc_ptr
;
655 alc_ptr
= kcalloc_node(nr_node_ids
, sizeof(void *), gfp
, node
);
660 if (i
== node
|| !node_online(i
))
662 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
664 for (i
--; i
>= 0; i
--)
673 static void free_alien_cache(struct alien_cache
**alc_ptr
)
684 static void __drain_alien_cache(struct kmem_cache
*cachep
,
685 struct array_cache
*ac
, int node
,
686 struct list_head
*list
)
688 struct kmem_cache_node
*n
= get_node(cachep
, node
);
691 spin_lock(&n
->list_lock
);
693 * Stuff objects into the remote nodes shared array first.
694 * That way we could avoid the overhead of putting the objects
695 * into the free lists and getting them back later.
698 transfer_objects(n
->shared
, ac
, ac
->limit
);
700 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
702 spin_unlock(&n
->list_lock
);
707 * Called from cache_reap() to regularly drain alien caches round robin.
709 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
711 int node
= __this_cpu_read(slab_reap_node
);
714 struct alien_cache
*alc
= n
->alien
[node
];
715 struct array_cache
*ac
;
719 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
722 __drain_alien_cache(cachep
, ac
, node
, &list
);
723 spin_unlock_irq(&alc
->lock
);
724 slabs_destroy(cachep
, &list
);
730 static void drain_alien_cache(struct kmem_cache
*cachep
,
731 struct alien_cache
**alien
)
734 struct alien_cache
*alc
;
735 struct array_cache
*ac
;
738 for_each_online_node(i
) {
744 spin_lock_irqsave(&alc
->lock
, flags
);
745 __drain_alien_cache(cachep
, ac
, i
, &list
);
746 spin_unlock_irqrestore(&alc
->lock
, flags
);
747 slabs_destroy(cachep
, &list
);
752 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
753 int node
, int page_node
)
755 struct kmem_cache_node
*n
;
756 struct alien_cache
*alien
= NULL
;
757 struct array_cache
*ac
;
760 n
= get_node(cachep
, node
);
761 STATS_INC_NODEFREES(cachep
);
762 if (n
->alien
&& n
->alien
[page_node
]) {
763 alien
= n
->alien
[page_node
];
765 spin_lock(&alien
->lock
);
766 if (unlikely(ac
->avail
== ac
->limit
)) {
767 STATS_INC_ACOVERFLOW(cachep
);
768 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
770 ac
->entry
[ac
->avail
++] = objp
;
771 spin_unlock(&alien
->lock
);
772 slabs_destroy(cachep
, &list
);
774 n
= get_node(cachep
, page_node
);
775 spin_lock(&n
->list_lock
);
776 free_block(cachep
, &objp
, 1, page_node
, &list
);
777 spin_unlock(&n
->list_lock
);
778 slabs_destroy(cachep
, &list
);
783 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
785 int page_node
= page_to_nid(virt_to_page(objp
));
786 int node
= numa_mem_id();
788 * Make sure we are not freeing a object from another node to the array
791 if (likely(node
== page_node
))
794 return __cache_free_alien(cachep
, objp
, node
, page_node
);
798 * Construct gfp mask to allocate from a specific node but do not reclaim or
799 * warn about failures.
801 static inline gfp_t
gfp_exact_node(gfp_t flags
)
803 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
807 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
809 struct kmem_cache_node
*n
;
812 * Set up the kmem_cache_node for cpu before we can
813 * begin anything. Make sure some other cpu on this
814 * node has not already allocated this
816 n
= get_node(cachep
, node
);
818 spin_lock_irq(&n
->list_lock
);
819 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
821 spin_unlock_irq(&n
->list_lock
);
826 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
830 kmem_cache_node_init(n
);
831 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
832 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
835 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
838 * The kmem_cache_nodes don't come and go as CPUs
839 * come and go. slab_mutex is sufficient
842 cachep
->node
[node
] = n
;
847 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
849 * Allocates and initializes node for a node on each slab cache, used for
850 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
851 * will be allocated off-node since memory is not yet online for the new node.
852 * When hotplugging memory or a cpu, existing node are not replaced if
855 * Must hold slab_mutex.
857 static int init_cache_node_node(int node
)
860 struct kmem_cache
*cachep
;
862 list_for_each_entry(cachep
, &slab_caches
, list
) {
863 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
872 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
873 int node
, gfp_t gfp
, bool force_change
)
876 struct kmem_cache_node
*n
;
877 struct array_cache
*old_shared
= NULL
;
878 struct array_cache
*new_shared
= NULL
;
879 struct alien_cache
**new_alien
= NULL
;
882 if (use_alien_caches
) {
883 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
888 if (cachep
->shared
) {
889 new_shared
= alloc_arraycache(node
,
890 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
895 ret
= init_cache_node(cachep
, node
, gfp
);
899 n
= get_node(cachep
, node
);
900 spin_lock_irq(&n
->list_lock
);
901 if (n
->shared
&& force_change
) {
902 free_block(cachep
, n
->shared
->entry
,
903 n
->shared
->avail
, node
, &list
);
904 n
->shared
->avail
= 0;
907 if (!n
->shared
|| force_change
) {
908 old_shared
= n
->shared
;
909 n
->shared
= new_shared
;
914 n
->alien
= new_alien
;
918 spin_unlock_irq(&n
->list_lock
);
919 slabs_destroy(cachep
, &list
);
922 * To protect lockless access to n->shared during irq disabled context.
923 * If n->shared isn't NULL in irq disabled context, accessing to it is
924 * guaranteed to be valid until irq is re-enabled, because it will be
925 * freed after synchronize_rcu().
927 if (old_shared
&& force_change
)
933 free_alien_cache(new_alien
);
940 static void cpuup_canceled(long cpu
)
942 struct kmem_cache
*cachep
;
943 struct kmem_cache_node
*n
= NULL
;
944 int node
= cpu_to_mem(cpu
);
945 const struct cpumask
*mask
= cpumask_of_node(node
);
947 list_for_each_entry(cachep
, &slab_caches
, list
) {
948 struct array_cache
*nc
;
949 struct array_cache
*shared
;
950 struct alien_cache
**alien
;
953 n
= get_node(cachep
, node
);
957 spin_lock_irq(&n
->list_lock
);
959 /* Free limit for this kmem_cache_node */
960 n
->free_limit
-= cachep
->batchcount
;
962 /* cpu is dead; no one can alloc from it. */
963 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
964 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
967 if (!cpumask_empty(mask
)) {
968 spin_unlock_irq(&n
->list_lock
);
974 free_block(cachep
, shared
->entry
,
975 shared
->avail
, node
, &list
);
982 spin_unlock_irq(&n
->list_lock
);
986 drain_alien_cache(cachep
, alien
);
987 free_alien_cache(alien
);
991 slabs_destroy(cachep
, &list
);
994 * In the previous loop, all the objects were freed to
995 * the respective cache's slabs, now we can go ahead and
996 * shrink each nodelist to its limit.
998 list_for_each_entry(cachep
, &slab_caches
, list
) {
999 n
= get_node(cachep
, node
);
1002 drain_freelist(cachep
, n
, INT_MAX
);
1006 static int cpuup_prepare(long cpu
)
1008 struct kmem_cache
*cachep
;
1009 int node
= cpu_to_mem(cpu
);
1013 * We need to do this right in the beginning since
1014 * alloc_arraycache's are going to use this list.
1015 * kmalloc_node allows us to add the slab to the right
1016 * kmem_cache_node and not this cpu's kmem_cache_node
1018 err
= init_cache_node_node(node
);
1023 * Now we can go ahead with allocating the shared arrays and
1026 list_for_each_entry(cachep
, &slab_caches
, list
) {
1027 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1034 cpuup_canceled(cpu
);
1038 int slab_prepare_cpu(unsigned int cpu
)
1042 mutex_lock(&slab_mutex
);
1043 err
= cpuup_prepare(cpu
);
1044 mutex_unlock(&slab_mutex
);
1049 * This is called for a failed online attempt and for a successful
1052 * Even if all the cpus of a node are down, we don't free the
1053 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1054 * a kmalloc allocation from another cpu for memory from the node of
1055 * the cpu going down. The list3 structure is usually allocated from
1056 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1058 int slab_dead_cpu(unsigned int cpu
)
1060 mutex_lock(&slab_mutex
);
1061 cpuup_canceled(cpu
);
1062 mutex_unlock(&slab_mutex
);
1067 static int slab_online_cpu(unsigned int cpu
)
1069 start_cpu_timer(cpu
);
1073 static int slab_offline_cpu(unsigned int cpu
)
1076 * Shutdown cache reaper. Note that the slab_mutex is held so
1077 * that if cache_reap() is invoked it cannot do anything
1078 * expensive but will only modify reap_work and reschedule the
1081 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1082 /* Now the cache_reaper is guaranteed to be not running. */
1083 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1087 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1089 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1090 * Returns -EBUSY if all objects cannot be drained so that the node is not
1093 * Must hold slab_mutex.
1095 static int __meminit
drain_cache_node_node(int node
)
1097 struct kmem_cache
*cachep
;
1100 list_for_each_entry(cachep
, &slab_caches
, list
) {
1101 struct kmem_cache_node
*n
;
1103 n
= get_node(cachep
, node
);
1107 drain_freelist(cachep
, n
, INT_MAX
);
1109 if (!list_empty(&n
->slabs_full
) ||
1110 !list_empty(&n
->slabs_partial
)) {
1118 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1119 unsigned long action
, void *arg
)
1121 struct memory_notify
*mnb
= arg
;
1125 nid
= mnb
->status_change_nid
;
1130 case MEM_GOING_ONLINE
:
1131 mutex_lock(&slab_mutex
);
1132 ret
= init_cache_node_node(nid
);
1133 mutex_unlock(&slab_mutex
);
1135 case MEM_GOING_OFFLINE
:
1136 mutex_lock(&slab_mutex
);
1137 ret
= drain_cache_node_node(nid
);
1138 mutex_unlock(&slab_mutex
);
1142 case MEM_CANCEL_ONLINE
:
1143 case MEM_CANCEL_OFFLINE
:
1147 return notifier_from_errno(ret
);
1149 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1152 * swap the static kmem_cache_node with kmalloced memory
1154 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1157 struct kmem_cache_node
*ptr
;
1159 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1162 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1164 * Do not assume that spinlocks can be initialized via memcpy:
1166 spin_lock_init(&ptr
->list_lock
);
1168 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1169 cachep
->node
[nodeid
] = ptr
;
1173 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1174 * size of kmem_cache_node.
1176 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1180 for_each_online_node(node
) {
1181 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1182 cachep
->node
[node
]->next_reap
= jiffies
+
1184 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1189 * Initialisation. Called after the page allocator have been initialised and
1190 * before smp_init().
1192 void __init
kmem_cache_init(void)
1196 kmem_cache
= &kmem_cache_boot
;
1198 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1199 use_alien_caches
= 0;
1201 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1202 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1205 * Fragmentation resistance on low memory - only use bigger
1206 * page orders on machines with more than 32MB of memory if
1207 * not overridden on the command line.
1209 if (!slab_max_order_set
&& totalram_pages() > (32 << 20) >> PAGE_SHIFT
)
1210 slab_max_order
= SLAB_MAX_ORDER_HI
;
1212 /* Bootstrap is tricky, because several objects are allocated
1213 * from caches that do not exist yet:
1214 * 1) initialize the kmem_cache cache: it contains the struct
1215 * kmem_cache structures of all caches, except kmem_cache itself:
1216 * kmem_cache is statically allocated.
1217 * Initially an __init data area is used for the head array and the
1218 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1219 * array at the end of the bootstrap.
1220 * 2) Create the first kmalloc cache.
1221 * The struct kmem_cache for the new cache is allocated normally.
1222 * An __init data area is used for the head array.
1223 * 3) Create the remaining kmalloc caches, with minimally sized
1225 * 4) Replace the __init data head arrays for kmem_cache and the first
1226 * kmalloc cache with kmalloc allocated arrays.
1227 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1228 * the other cache's with kmalloc allocated memory.
1229 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1232 /* 1) create the kmem_cache */
1235 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1237 create_boot_cache(kmem_cache
, "kmem_cache",
1238 offsetof(struct kmem_cache
, node
) +
1239 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1240 SLAB_HWCACHE_ALIGN
, 0, 0);
1241 list_add(&kmem_cache
->list
, &slab_caches
);
1242 memcg_link_cache(kmem_cache
, NULL
);
1243 slab_state
= PARTIAL
;
1246 * Initialize the caches that provide memory for the kmem_cache_node
1247 * structures first. Without this, further allocations will bug.
1249 kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
] = create_kmalloc_cache(
1250 kmalloc_info
[INDEX_NODE
].name
,
1251 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
,
1252 0, kmalloc_size(INDEX_NODE
));
1253 slab_state
= PARTIAL_NODE
;
1254 setup_kmalloc_cache_index_table();
1256 slab_early_init
= 0;
1258 /* 5) Replace the bootstrap kmem_cache_node */
1262 for_each_online_node(nid
) {
1263 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1265 init_list(kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
],
1266 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1270 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1273 void __init
kmem_cache_init_late(void)
1275 struct kmem_cache
*cachep
;
1277 /* 6) resize the head arrays to their final sizes */
1278 mutex_lock(&slab_mutex
);
1279 list_for_each_entry(cachep
, &slab_caches
, list
)
1280 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1282 mutex_unlock(&slab_mutex
);
1289 * Register a memory hotplug callback that initializes and frees
1292 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1296 * The reap timers are started later, with a module init call: That part
1297 * of the kernel is not yet operational.
1301 static int __init
cpucache_init(void)
1306 * Register the timers that return unneeded pages to the page allocator
1308 ret
= cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "SLAB online",
1309 slab_online_cpu
, slab_offline_cpu
);
1314 __initcall(cpucache_init
);
1316 static noinline
void
1317 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1320 struct kmem_cache_node
*n
;
1321 unsigned long flags
;
1323 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1324 DEFAULT_RATELIMIT_BURST
);
1326 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1329 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1330 nodeid
, gfpflags
, &gfpflags
);
1331 pr_warn(" cache: %s, object size: %d, order: %d\n",
1332 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1334 for_each_kmem_cache_node(cachep
, node
, n
) {
1335 unsigned long total_slabs
, free_slabs
, free_objs
;
1337 spin_lock_irqsave(&n
->list_lock
, flags
);
1338 total_slabs
= n
->total_slabs
;
1339 free_slabs
= n
->free_slabs
;
1340 free_objs
= n
->free_objects
;
1341 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1343 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1344 node
, total_slabs
- free_slabs
, total_slabs
,
1345 (total_slabs
* cachep
->num
) - free_objs
,
1346 total_slabs
* cachep
->num
);
1352 * Interface to system's page allocator. No need to hold the
1353 * kmem_cache_node ->list_lock.
1355 * If we requested dmaable memory, we will get it. Even if we
1356 * did not request dmaable memory, we might get it, but that
1357 * would be relatively rare and ignorable.
1359 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1364 flags
|= cachep
->allocflags
;
1366 page
= __alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1368 slab_out_of_memory(cachep
, flags
, nodeid
);
1372 if (charge_slab_page(page
, flags
, cachep
->gfporder
, cachep
)) {
1373 __free_pages(page
, cachep
->gfporder
);
1377 __SetPageSlab(page
);
1378 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1379 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1380 SetPageSlabPfmemalloc(page
);
1386 * Interface to system's page release.
1388 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1390 int order
= cachep
->gfporder
;
1392 BUG_ON(!PageSlab(page
));
1393 __ClearPageSlabPfmemalloc(page
);
1394 __ClearPageSlab(page
);
1395 page_mapcount_reset(page
);
1396 page
->mapping
= NULL
;
1398 if (current
->reclaim_state
)
1399 current
->reclaim_state
->reclaimed_slab
+= 1 << order
;
1400 uncharge_slab_page(page
, order
, cachep
);
1401 __free_pages(page
, order
);
1404 static void kmem_rcu_free(struct rcu_head
*head
)
1406 struct kmem_cache
*cachep
;
1409 page
= container_of(head
, struct page
, rcu_head
);
1410 cachep
= page
->slab_cache
;
1412 kmem_freepages(cachep
, page
);
1416 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1418 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1419 (cachep
->size
% PAGE_SIZE
) == 0)
1425 #ifdef CONFIG_DEBUG_PAGEALLOC
1426 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
, int map
)
1428 if (!is_debug_pagealloc_cache(cachep
))
1431 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1435 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1440 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1442 int size
= cachep
->object_size
;
1443 addr
= &((char *)addr
)[obj_offset(cachep
)];
1445 memset(addr
, val
, size
);
1446 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1449 static void dump_line(char *data
, int offset
, int limit
)
1452 unsigned char error
= 0;
1455 pr_err("%03x: ", offset
);
1456 for (i
= 0; i
< limit
; i
++) {
1457 if (data
[offset
+ i
] != POISON_FREE
) {
1458 error
= data
[offset
+ i
];
1462 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1463 &data
[offset
], limit
, 1);
1465 if (bad_count
== 1) {
1466 error
^= POISON_FREE
;
1467 if (!(error
& (error
- 1))) {
1468 pr_err("Single bit error detected. Probably bad RAM.\n");
1470 pr_err("Run memtest86+ or a similar memory test tool.\n");
1472 pr_err("Run a memory test tool.\n");
1481 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1486 if (cachep
->flags
& SLAB_RED_ZONE
) {
1487 pr_err("Redzone: 0x%llx/0x%llx\n",
1488 *dbg_redzone1(cachep
, objp
),
1489 *dbg_redzone2(cachep
, objp
));
1492 if (cachep
->flags
& SLAB_STORE_USER
)
1493 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep
, objp
));
1494 realobj
= (char *)objp
+ obj_offset(cachep
);
1495 size
= cachep
->object_size
;
1496 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1499 if (i
+ limit
> size
)
1501 dump_line(realobj
, i
, limit
);
1505 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1511 if (is_debug_pagealloc_cache(cachep
))
1514 realobj
= (char *)objp
+ obj_offset(cachep
);
1515 size
= cachep
->object_size
;
1517 for (i
= 0; i
< size
; i
++) {
1518 char exp
= POISON_FREE
;
1521 if (realobj
[i
] != exp
) {
1526 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1527 print_tainted(), cachep
->name
,
1529 print_objinfo(cachep
, objp
, 0);
1531 /* Hexdump the affected line */
1534 if (i
+ limit
> size
)
1536 dump_line(realobj
, i
, limit
);
1539 /* Limit to 5 lines */
1545 /* Print some data about the neighboring objects, if they
1548 struct page
*page
= virt_to_head_page(objp
);
1551 objnr
= obj_to_index(cachep
, page
, objp
);
1553 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1554 realobj
= (char *)objp
+ obj_offset(cachep
);
1555 pr_err("Prev obj: start=%px, len=%d\n", realobj
, size
);
1556 print_objinfo(cachep
, objp
, 2);
1558 if (objnr
+ 1 < cachep
->num
) {
1559 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1560 realobj
= (char *)objp
+ obj_offset(cachep
);
1561 pr_err("Next obj: start=%px, len=%d\n", realobj
, size
);
1562 print_objinfo(cachep
, objp
, 2);
1569 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1574 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1575 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1579 for (i
= 0; i
< cachep
->num
; i
++) {
1580 void *objp
= index_to_obj(cachep
, page
, i
);
1582 if (cachep
->flags
& SLAB_POISON
) {
1583 check_poison_obj(cachep
, objp
);
1584 slab_kernel_map(cachep
, objp
, 1);
1586 if (cachep
->flags
& SLAB_RED_ZONE
) {
1587 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1588 slab_error(cachep
, "start of a freed object was overwritten");
1589 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1590 slab_error(cachep
, "end of a freed object was overwritten");
1595 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1602 * slab_destroy - destroy and release all objects in a slab
1603 * @cachep: cache pointer being destroyed
1604 * @page: page pointer being destroyed
1606 * Destroy all the objs in a slab page, and release the mem back to the system.
1607 * Before calling the slab page must have been unlinked from the cache. The
1608 * kmem_cache_node ->list_lock is not held/needed.
1610 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1614 freelist
= page
->freelist
;
1615 slab_destroy_debugcheck(cachep
, page
);
1616 if (unlikely(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
1617 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1619 kmem_freepages(cachep
, page
);
1622 * From now on, we don't use freelist
1623 * although actual page can be freed in rcu context
1625 if (OFF_SLAB(cachep
))
1626 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1629 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1631 struct page
*page
, *n
;
1633 list_for_each_entry_safe(page
, n
, list
, slab_list
) {
1634 list_del(&page
->slab_list
);
1635 slab_destroy(cachep
, page
);
1640 * calculate_slab_order - calculate size (page order) of slabs
1641 * @cachep: pointer to the cache that is being created
1642 * @size: size of objects to be created in this cache.
1643 * @flags: slab allocation flags
1645 * Also calculates the number of objects per slab.
1647 * This could be made much more intelligent. For now, try to avoid using
1648 * high order pages for slabs. When the gfp() functions are more friendly
1649 * towards high-order requests, this should be changed.
1651 * Return: number of left-over bytes in a slab
1653 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1654 size_t size
, slab_flags_t flags
)
1656 size_t left_over
= 0;
1659 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1663 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1667 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1668 if (num
> SLAB_OBJ_MAX_NUM
)
1671 if (flags
& CFLGS_OFF_SLAB
) {
1672 struct kmem_cache
*freelist_cache
;
1673 size_t freelist_size
;
1675 freelist_size
= num
* sizeof(freelist_idx_t
);
1676 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1677 if (!freelist_cache
)
1681 * Needed to avoid possible looping condition
1682 * in cache_grow_begin()
1684 if (OFF_SLAB(freelist_cache
))
1687 /* check if off slab has enough benefit */
1688 if (freelist_cache
->size
> cachep
->size
/ 2)
1692 /* Found something acceptable - save it away */
1694 cachep
->gfporder
= gfporder
;
1695 left_over
= remainder
;
1698 * A VFS-reclaimable slab tends to have most allocations
1699 * as GFP_NOFS and we really don't want to have to be allocating
1700 * higher-order pages when we are unable to shrink dcache.
1702 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1706 * Large number of objects is good, but very large slabs are
1707 * currently bad for the gfp()s.
1709 if (gfporder
>= slab_max_order
)
1713 * Acceptable internal fragmentation?
1715 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1721 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1722 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1726 struct array_cache __percpu
*cpu_cache
;
1728 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1729 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1734 for_each_possible_cpu(cpu
) {
1735 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1736 entries
, batchcount
);
1742 static int __ref
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1744 if (slab_state
>= FULL
)
1745 return enable_cpucache(cachep
, gfp
);
1747 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1748 if (!cachep
->cpu_cache
)
1751 if (slab_state
== DOWN
) {
1752 /* Creation of first cache (kmem_cache). */
1753 set_up_node(kmem_cache
, CACHE_CACHE
);
1754 } else if (slab_state
== PARTIAL
) {
1755 /* For kmem_cache_node */
1756 set_up_node(cachep
, SIZE_NODE
);
1760 for_each_online_node(node
) {
1761 cachep
->node
[node
] = kmalloc_node(
1762 sizeof(struct kmem_cache_node
), gfp
, node
);
1763 BUG_ON(!cachep
->node
[node
]);
1764 kmem_cache_node_init(cachep
->node
[node
]);
1768 cachep
->node
[numa_mem_id()]->next_reap
=
1769 jiffies
+ REAPTIMEOUT_NODE
+
1770 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1772 cpu_cache_get(cachep
)->avail
= 0;
1773 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1774 cpu_cache_get(cachep
)->batchcount
= 1;
1775 cpu_cache_get(cachep
)->touched
= 0;
1776 cachep
->batchcount
= 1;
1777 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1781 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1782 slab_flags_t flags
, const char *name
,
1783 void (*ctor
)(void *))
1789 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
1790 slab_flags_t flags
, void (*ctor
)(void *))
1792 struct kmem_cache
*cachep
;
1794 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1799 * Adjust the object sizes so that we clear
1800 * the complete object on kzalloc.
1802 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1807 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1808 size_t size
, slab_flags_t flags
)
1814 if (cachep
->ctor
|| flags
& SLAB_TYPESAFE_BY_RCU
)
1817 left
= calculate_slab_order(cachep
, size
,
1818 flags
| CFLGS_OBJFREELIST_SLAB
);
1822 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1825 cachep
->colour
= left
/ cachep
->colour_off
;
1830 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1831 size_t size
, slab_flags_t flags
)
1838 * Always use on-slab management when SLAB_NOLEAKTRACE
1839 * to avoid recursive calls into kmemleak.
1841 if (flags
& SLAB_NOLEAKTRACE
)
1845 * Size is large, assume best to place the slab management obj
1846 * off-slab (should allow better packing of objs).
1848 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1853 * If the slab has been placed off-slab, and we have enough space then
1854 * move it on-slab. This is at the expense of any extra colouring.
1856 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1859 cachep
->colour
= left
/ cachep
->colour_off
;
1864 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1865 size_t size
, slab_flags_t flags
)
1871 left
= calculate_slab_order(cachep
, size
, flags
);
1875 cachep
->colour
= left
/ cachep
->colour_off
;
1881 * __kmem_cache_create - Create a cache.
1882 * @cachep: cache management descriptor
1883 * @flags: SLAB flags
1885 * Returns a ptr to the cache on success, NULL on failure.
1886 * Cannot be called within a int, but can be interrupted.
1887 * The @ctor is run when new pages are allocated by the cache.
1891 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1892 * to catch references to uninitialised memory.
1894 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1895 * for buffer overruns.
1897 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1898 * cacheline. This can be beneficial if you're counting cycles as closely
1901 * Return: a pointer to the created cache or %NULL in case of error
1903 int __kmem_cache_create(struct kmem_cache
*cachep
, slab_flags_t flags
)
1905 size_t ralign
= BYTES_PER_WORD
;
1908 unsigned int size
= cachep
->size
;
1913 * Enable redzoning and last user accounting, except for caches with
1914 * large objects, if the increased size would increase the object size
1915 * above the next power of two: caches with object sizes just above a
1916 * power of two have a significant amount of internal fragmentation.
1918 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
1919 2 * sizeof(unsigned long long)))
1920 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1921 if (!(flags
& SLAB_TYPESAFE_BY_RCU
))
1922 flags
|= SLAB_POISON
;
1927 * Check that size is in terms of words. This is needed to avoid
1928 * unaligned accesses for some archs when redzoning is used, and makes
1929 * sure any on-slab bufctl's are also correctly aligned.
1931 size
= ALIGN(size
, BYTES_PER_WORD
);
1933 if (flags
& SLAB_RED_ZONE
) {
1934 ralign
= REDZONE_ALIGN
;
1935 /* If redzoning, ensure that the second redzone is suitably
1936 * aligned, by adjusting the object size accordingly. */
1937 size
= ALIGN(size
, REDZONE_ALIGN
);
1940 /* 3) caller mandated alignment */
1941 if (ralign
< cachep
->align
) {
1942 ralign
= cachep
->align
;
1944 /* disable debug if necessary */
1945 if (ralign
> __alignof__(unsigned long long))
1946 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1950 cachep
->align
= ralign
;
1951 cachep
->colour_off
= cache_line_size();
1952 /* Offset must be a multiple of the alignment. */
1953 if (cachep
->colour_off
< cachep
->align
)
1954 cachep
->colour_off
= cachep
->align
;
1956 if (slab_is_available())
1964 * Both debugging options require word-alignment which is calculated
1967 if (flags
& SLAB_RED_ZONE
) {
1968 /* add space for red zone words */
1969 cachep
->obj_offset
+= sizeof(unsigned long long);
1970 size
+= 2 * sizeof(unsigned long long);
1972 if (flags
& SLAB_STORE_USER
) {
1973 /* user store requires one word storage behind the end of
1974 * the real object. But if the second red zone needs to be
1975 * aligned to 64 bits, we must allow that much space.
1977 if (flags
& SLAB_RED_ZONE
)
1978 size
+= REDZONE_ALIGN
;
1980 size
+= BYTES_PER_WORD
;
1984 kasan_cache_create(cachep
, &size
, &flags
);
1986 size
= ALIGN(size
, cachep
->align
);
1988 * We should restrict the number of objects in a slab to implement
1989 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
1991 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
1992 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
1996 * To activate debug pagealloc, off-slab management is necessary
1997 * requirement. In early phase of initialization, small sized slab
1998 * doesn't get initialized so it would not be possible. So, we need
1999 * to check size >= 256. It guarantees that all necessary small
2000 * sized slab is initialized in current slab initialization sequence.
2002 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2003 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2004 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2005 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2007 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2008 flags
|= CFLGS_OFF_SLAB
;
2009 cachep
->obj_offset
+= tmp_size
- size
;
2017 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2018 flags
|= CFLGS_OBJFREELIST_SLAB
;
2022 if (set_off_slab_cache(cachep
, size
, flags
)) {
2023 flags
|= CFLGS_OFF_SLAB
;
2027 if (set_on_slab_cache(cachep
, size
, flags
))
2033 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2034 cachep
->flags
= flags
;
2035 cachep
->allocflags
= __GFP_COMP
;
2036 if (flags
& SLAB_CACHE_DMA
)
2037 cachep
->allocflags
|= GFP_DMA
;
2038 if (flags
& SLAB_CACHE_DMA32
)
2039 cachep
->allocflags
|= GFP_DMA32
;
2040 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2041 cachep
->allocflags
|= __GFP_RECLAIMABLE
;
2042 cachep
->size
= size
;
2043 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2047 * If we're going to use the generic kernel_map_pages()
2048 * poisoning, then it's going to smash the contents of
2049 * the redzone and userword anyhow, so switch them off.
2051 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2052 (cachep
->flags
& SLAB_POISON
) &&
2053 is_debug_pagealloc_cache(cachep
))
2054 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2057 if (OFF_SLAB(cachep
)) {
2058 cachep
->freelist_cache
=
2059 kmalloc_slab(cachep
->freelist_size
, 0u);
2062 err
= setup_cpu_cache(cachep
, gfp
);
2064 __kmem_cache_release(cachep
);
2072 static void check_irq_off(void)
2074 BUG_ON(!irqs_disabled());
2077 static void check_irq_on(void)
2079 BUG_ON(irqs_disabled());
2082 static void check_mutex_acquired(void)
2084 BUG_ON(!mutex_is_locked(&slab_mutex
));
2087 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2091 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2095 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2099 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2104 #define check_irq_off() do { } while(0)
2105 #define check_irq_on() do { } while(0)
2106 #define check_mutex_acquired() do { } while(0)
2107 #define check_spinlock_acquired(x) do { } while(0)
2108 #define check_spinlock_acquired_node(x, y) do { } while(0)
2111 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2112 int node
, bool free_all
, struct list_head
*list
)
2116 if (!ac
|| !ac
->avail
)
2119 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2120 if (tofree
> ac
->avail
)
2121 tofree
= (ac
->avail
+ 1) / 2;
2123 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2124 ac
->avail
-= tofree
;
2125 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2128 static void do_drain(void *arg
)
2130 struct kmem_cache
*cachep
= arg
;
2131 struct array_cache
*ac
;
2132 int node
= numa_mem_id();
2133 struct kmem_cache_node
*n
;
2137 ac
= cpu_cache_get(cachep
);
2138 n
= get_node(cachep
, node
);
2139 spin_lock(&n
->list_lock
);
2140 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2141 spin_unlock(&n
->list_lock
);
2142 slabs_destroy(cachep
, &list
);
2146 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2148 struct kmem_cache_node
*n
;
2152 on_each_cpu(do_drain
, cachep
, 1);
2154 for_each_kmem_cache_node(cachep
, node
, n
)
2156 drain_alien_cache(cachep
, n
->alien
);
2158 for_each_kmem_cache_node(cachep
, node
, n
) {
2159 spin_lock_irq(&n
->list_lock
);
2160 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2161 spin_unlock_irq(&n
->list_lock
);
2163 slabs_destroy(cachep
, &list
);
2168 * Remove slabs from the list of free slabs.
2169 * Specify the number of slabs to drain in tofree.
2171 * Returns the actual number of slabs released.
2173 static int drain_freelist(struct kmem_cache
*cache
,
2174 struct kmem_cache_node
*n
, int tofree
)
2176 struct list_head
*p
;
2181 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2183 spin_lock_irq(&n
->list_lock
);
2184 p
= n
->slabs_free
.prev
;
2185 if (p
== &n
->slabs_free
) {
2186 spin_unlock_irq(&n
->list_lock
);
2190 page
= list_entry(p
, struct page
, slab_list
);
2191 list_del(&page
->slab_list
);
2195 * Safe to drop the lock. The slab is no longer linked
2198 n
->free_objects
-= cache
->num
;
2199 spin_unlock_irq(&n
->list_lock
);
2200 slab_destroy(cache
, page
);
2207 bool __kmem_cache_empty(struct kmem_cache
*s
)
2210 struct kmem_cache_node
*n
;
2212 for_each_kmem_cache_node(s
, node
, n
)
2213 if (!list_empty(&n
->slabs_full
) ||
2214 !list_empty(&n
->slabs_partial
))
2219 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2223 struct kmem_cache_node
*n
;
2225 drain_cpu_caches(cachep
);
2228 for_each_kmem_cache_node(cachep
, node
, n
) {
2229 drain_freelist(cachep
, n
, INT_MAX
);
2231 ret
+= !list_empty(&n
->slabs_full
) ||
2232 !list_empty(&n
->slabs_partial
);
2234 return (ret
? 1 : 0);
2238 void __kmemcg_cache_deactivate(struct kmem_cache
*cachep
)
2240 __kmem_cache_shrink(cachep
);
2243 void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache
*s
)
2248 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2250 return __kmem_cache_shrink(cachep
);
2253 void __kmem_cache_release(struct kmem_cache
*cachep
)
2256 struct kmem_cache_node
*n
;
2258 cache_random_seq_destroy(cachep
);
2260 free_percpu(cachep
->cpu_cache
);
2262 /* NUMA: free the node structures */
2263 for_each_kmem_cache_node(cachep
, i
, n
) {
2265 free_alien_cache(n
->alien
);
2267 cachep
->node
[i
] = NULL
;
2272 * Get the memory for a slab management obj.
2274 * For a slab cache when the slab descriptor is off-slab, the
2275 * slab descriptor can't come from the same cache which is being created,
2276 * Because if it is the case, that means we defer the creation of
2277 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2278 * And we eventually call down to __kmem_cache_create(), which
2279 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2280 * This is a "chicken-and-egg" problem.
2282 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2283 * which are all initialized during kmem_cache_init().
2285 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2286 struct page
*page
, int colour_off
,
2287 gfp_t local_flags
, int nodeid
)
2290 void *addr
= page_address(page
);
2292 page
->s_mem
= addr
+ colour_off
;
2295 if (OBJFREELIST_SLAB(cachep
))
2297 else if (OFF_SLAB(cachep
)) {
2298 /* Slab management obj is off-slab. */
2299 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2300 local_flags
, nodeid
);
2304 /* We will use last bytes at the slab for freelist */
2305 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2306 cachep
->freelist_size
;
2312 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2314 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2317 static inline void set_free_obj(struct page
*page
,
2318 unsigned int idx
, freelist_idx_t val
)
2320 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2323 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2328 for (i
= 0; i
< cachep
->num
; i
++) {
2329 void *objp
= index_to_obj(cachep
, page
, i
);
2331 if (cachep
->flags
& SLAB_STORE_USER
)
2332 *dbg_userword(cachep
, objp
) = NULL
;
2334 if (cachep
->flags
& SLAB_RED_ZONE
) {
2335 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2336 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2339 * Constructors are not allowed to allocate memory from the same
2340 * cache which they are a constructor for. Otherwise, deadlock.
2341 * They must also be threaded.
2343 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2344 kasan_unpoison_object_data(cachep
,
2345 objp
+ obj_offset(cachep
));
2346 cachep
->ctor(objp
+ obj_offset(cachep
));
2347 kasan_poison_object_data(
2348 cachep
, objp
+ obj_offset(cachep
));
2351 if (cachep
->flags
& SLAB_RED_ZONE
) {
2352 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2353 slab_error(cachep
, "constructor overwrote the end of an object");
2354 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2355 slab_error(cachep
, "constructor overwrote the start of an object");
2357 /* need to poison the objs? */
2358 if (cachep
->flags
& SLAB_POISON
) {
2359 poison_obj(cachep
, objp
, POISON_FREE
);
2360 slab_kernel_map(cachep
, objp
, 0);
2366 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2367 /* Hold information during a freelist initialization */
2368 union freelist_init_state
{
2374 struct rnd_state rnd_state
;
2378 * Initialize the state based on the randomization methode available.
2379 * return true if the pre-computed list is available, false otherwize.
2381 static bool freelist_state_initialize(union freelist_init_state
*state
,
2382 struct kmem_cache
*cachep
,
2388 /* Use best entropy available to define a random shift */
2389 rand
= get_random_int();
2391 /* Use a random state if the pre-computed list is not available */
2392 if (!cachep
->random_seq
) {
2393 prandom_seed_state(&state
->rnd_state
, rand
);
2396 state
->list
= cachep
->random_seq
;
2397 state
->count
= count
;
2398 state
->pos
= rand
% count
;
2404 /* Get the next entry on the list and randomize it using a random shift */
2405 static freelist_idx_t
next_random_slot(union freelist_init_state
*state
)
2407 if (state
->pos
>= state
->count
)
2409 return state
->list
[state
->pos
++];
2412 /* Swap two freelist entries */
2413 static void swap_free_obj(struct page
*page
, unsigned int a
, unsigned int b
)
2415 swap(((freelist_idx_t
*)page
->freelist
)[a
],
2416 ((freelist_idx_t
*)page
->freelist
)[b
]);
2420 * Shuffle the freelist initialization state based on pre-computed lists.
2421 * return true if the list was successfully shuffled, false otherwise.
2423 static bool shuffle_freelist(struct kmem_cache
*cachep
, struct page
*page
)
2425 unsigned int objfreelist
= 0, i
, rand
, count
= cachep
->num
;
2426 union freelist_init_state state
;
2432 precomputed
= freelist_state_initialize(&state
, cachep
, count
);
2434 /* Take a random entry as the objfreelist */
2435 if (OBJFREELIST_SLAB(cachep
)) {
2437 objfreelist
= count
- 1;
2439 objfreelist
= next_random_slot(&state
);
2440 page
->freelist
= index_to_obj(cachep
, page
, objfreelist
) +
2446 * On early boot, generate the list dynamically.
2447 * Later use a pre-computed list for speed.
2450 for (i
= 0; i
< count
; i
++)
2451 set_free_obj(page
, i
, i
);
2453 /* Fisher-Yates shuffle */
2454 for (i
= count
- 1; i
> 0; i
--) {
2455 rand
= prandom_u32_state(&state
.rnd_state
);
2457 swap_free_obj(page
, i
, rand
);
2460 for (i
= 0; i
< count
; i
++)
2461 set_free_obj(page
, i
, next_random_slot(&state
));
2464 if (OBJFREELIST_SLAB(cachep
))
2465 set_free_obj(page
, cachep
->num
- 1, objfreelist
);
2470 static inline bool shuffle_freelist(struct kmem_cache
*cachep
,
2475 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2477 static void cache_init_objs(struct kmem_cache
*cachep
,
2484 cache_init_objs_debug(cachep
, page
);
2486 /* Try to randomize the freelist if enabled */
2487 shuffled
= shuffle_freelist(cachep
, page
);
2489 if (!shuffled
&& OBJFREELIST_SLAB(cachep
)) {
2490 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2494 for (i
= 0; i
< cachep
->num
; i
++) {
2495 objp
= index_to_obj(cachep
, page
, i
);
2496 objp
= kasan_init_slab_obj(cachep
, objp
);
2498 /* constructor could break poison info */
2499 if (DEBUG
== 0 && cachep
->ctor
) {
2500 kasan_unpoison_object_data(cachep
, objp
);
2502 kasan_poison_object_data(cachep
, objp
);
2506 set_free_obj(page
, i
, i
);
2510 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2514 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2520 static void slab_put_obj(struct kmem_cache
*cachep
,
2521 struct page
*page
, void *objp
)
2523 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2527 /* Verify double free bug */
2528 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2529 if (get_free_obj(page
, i
) == objnr
) {
2530 pr_err("slab: double free detected in cache '%s', objp %px\n",
2531 cachep
->name
, objp
);
2537 if (!page
->freelist
)
2538 page
->freelist
= objp
+ obj_offset(cachep
);
2540 set_free_obj(page
, page
->active
, objnr
);
2544 * Map pages beginning at addr to the given cache and slab. This is required
2545 * for the slab allocator to be able to lookup the cache and slab of a
2546 * virtual address for kfree, ksize, and slab debugging.
2548 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2551 page
->slab_cache
= cache
;
2552 page
->freelist
= freelist
;
2556 * Grow (by 1) the number of slabs within a cache. This is called by
2557 * kmem_cache_alloc() when there are no active objs left in a cache.
2559 static struct page
*cache_grow_begin(struct kmem_cache
*cachep
,
2560 gfp_t flags
, int nodeid
)
2566 struct kmem_cache_node
*n
;
2570 * Be lazy and only check for valid flags here, keeping it out of the
2571 * critical path in kmem_cache_alloc().
2573 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2574 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
2575 flags
&= ~GFP_SLAB_BUG_MASK
;
2576 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2577 invalid_mask
, &invalid_mask
, flags
, &flags
);
2580 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2581 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2584 if (gfpflags_allow_blocking(local_flags
))
2588 * Get mem for the objs. Attempt to allocate a physical page from
2591 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2595 page_node
= page_to_nid(page
);
2596 n
= get_node(cachep
, page_node
);
2598 /* Get colour for the slab, and cal the next value. */
2600 if (n
->colour_next
>= cachep
->colour
)
2603 offset
= n
->colour_next
;
2604 if (offset
>= cachep
->colour
)
2607 offset
*= cachep
->colour_off
;
2610 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2611 * page_address() in the latter returns a non-tagged pointer,
2612 * as it should be for slab pages.
2614 kasan_poison_slab(page
);
2616 /* Get slab management. */
2617 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2618 local_flags
& ~GFP_CONSTRAINT_MASK
, page_node
);
2619 if (OFF_SLAB(cachep
) && !freelist
)
2622 slab_map_pages(cachep
, page
, freelist
);
2624 cache_init_objs(cachep
, page
);
2626 if (gfpflags_allow_blocking(local_flags
))
2627 local_irq_disable();
2632 kmem_freepages(cachep
, page
);
2634 if (gfpflags_allow_blocking(local_flags
))
2635 local_irq_disable();
2639 static void cache_grow_end(struct kmem_cache
*cachep
, struct page
*page
)
2641 struct kmem_cache_node
*n
;
2649 INIT_LIST_HEAD(&page
->slab_list
);
2650 n
= get_node(cachep
, page_to_nid(page
));
2652 spin_lock(&n
->list_lock
);
2654 if (!page
->active
) {
2655 list_add_tail(&page
->slab_list
, &n
->slabs_free
);
2658 fixup_slab_list(cachep
, n
, page
, &list
);
2660 STATS_INC_GROWN(cachep
);
2661 n
->free_objects
+= cachep
->num
- page
->active
;
2662 spin_unlock(&n
->list_lock
);
2664 fixup_objfreelist_debug(cachep
, &list
);
2670 * Perform extra freeing checks:
2671 * - detect bad pointers.
2672 * - POISON/RED_ZONE checking
2674 static void kfree_debugcheck(const void *objp
)
2676 if (!virt_addr_valid(objp
)) {
2677 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2678 (unsigned long)objp
);
2683 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2685 unsigned long long redzone1
, redzone2
;
2687 redzone1
= *dbg_redzone1(cache
, obj
);
2688 redzone2
= *dbg_redzone2(cache
, obj
);
2693 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2696 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2697 slab_error(cache
, "double free detected");
2699 slab_error(cache
, "memory outside object was overwritten");
2701 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2702 obj
, redzone1
, redzone2
);
2705 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2706 unsigned long caller
)
2711 BUG_ON(virt_to_cache(objp
) != cachep
);
2713 objp
-= obj_offset(cachep
);
2714 kfree_debugcheck(objp
);
2715 page
= virt_to_head_page(objp
);
2717 if (cachep
->flags
& SLAB_RED_ZONE
) {
2718 verify_redzone_free(cachep
, objp
);
2719 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2720 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2722 if (cachep
->flags
& SLAB_STORE_USER
)
2723 *dbg_userword(cachep
, objp
) = (void *)caller
;
2725 objnr
= obj_to_index(cachep
, page
, objp
);
2727 BUG_ON(objnr
>= cachep
->num
);
2728 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2730 if (cachep
->flags
& SLAB_POISON
) {
2731 poison_obj(cachep
, objp
, POISON_FREE
);
2732 slab_kernel_map(cachep
, objp
, 0);
2738 #define kfree_debugcheck(x) do { } while(0)
2739 #define cache_free_debugcheck(x,objp,z) (objp)
2742 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2750 objp
= next
- obj_offset(cachep
);
2751 next
= *(void **)next
;
2752 poison_obj(cachep
, objp
, POISON_FREE
);
2757 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2758 struct kmem_cache_node
*n
, struct page
*page
,
2761 /* move slabp to correct slabp list: */
2762 list_del(&page
->slab_list
);
2763 if (page
->active
== cachep
->num
) {
2764 list_add(&page
->slab_list
, &n
->slabs_full
);
2765 if (OBJFREELIST_SLAB(cachep
)) {
2767 /* Poisoning will be done without holding the lock */
2768 if (cachep
->flags
& SLAB_POISON
) {
2769 void **objp
= page
->freelist
;
2775 page
->freelist
= NULL
;
2778 list_add(&page
->slab_list
, &n
->slabs_partial
);
2781 /* Try to find non-pfmemalloc slab if needed */
2782 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2783 struct page
*page
, bool pfmemalloc
)
2791 if (!PageSlabPfmemalloc(page
))
2794 /* No need to keep pfmemalloc slab if we have enough free objects */
2795 if (n
->free_objects
> n
->free_limit
) {
2796 ClearPageSlabPfmemalloc(page
);
2800 /* Move pfmemalloc slab to the end of list to speed up next search */
2801 list_del(&page
->slab_list
);
2802 if (!page
->active
) {
2803 list_add_tail(&page
->slab_list
, &n
->slabs_free
);
2806 list_add_tail(&page
->slab_list
, &n
->slabs_partial
);
2808 list_for_each_entry(page
, &n
->slabs_partial
, slab_list
) {
2809 if (!PageSlabPfmemalloc(page
))
2813 n
->free_touched
= 1;
2814 list_for_each_entry(page
, &n
->slabs_free
, slab_list
) {
2815 if (!PageSlabPfmemalloc(page
)) {
2824 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2828 assert_spin_locked(&n
->list_lock
);
2829 page
= list_first_entry_or_null(&n
->slabs_partial
, struct page
,
2832 n
->free_touched
= 1;
2833 page
= list_first_entry_or_null(&n
->slabs_free
, struct page
,
2839 if (sk_memalloc_socks())
2840 page
= get_valid_first_slab(n
, page
, pfmemalloc
);
2845 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2846 struct kmem_cache_node
*n
, gfp_t flags
)
2852 if (!gfp_pfmemalloc_allowed(flags
))
2855 spin_lock(&n
->list_lock
);
2856 page
= get_first_slab(n
, true);
2858 spin_unlock(&n
->list_lock
);
2862 obj
= slab_get_obj(cachep
, page
);
2865 fixup_slab_list(cachep
, n
, page
, &list
);
2867 spin_unlock(&n
->list_lock
);
2868 fixup_objfreelist_debug(cachep
, &list
);
2874 * Slab list should be fixed up by fixup_slab_list() for existing slab
2875 * or cache_grow_end() for new slab
2877 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2878 struct array_cache
*ac
, struct page
*page
, int batchcount
)
2881 * There must be at least one object available for
2884 BUG_ON(page
->active
>= cachep
->num
);
2886 while (page
->active
< cachep
->num
&& batchcount
--) {
2887 STATS_INC_ALLOCED(cachep
);
2888 STATS_INC_ACTIVE(cachep
);
2889 STATS_SET_HIGH(cachep
);
2891 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2897 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2900 struct kmem_cache_node
*n
;
2901 struct array_cache
*ac
, *shared
;
2907 node
= numa_mem_id();
2909 ac
= cpu_cache_get(cachep
);
2910 batchcount
= ac
->batchcount
;
2911 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2913 * If there was little recent activity on this cache, then
2914 * perform only a partial refill. Otherwise we could generate
2917 batchcount
= BATCHREFILL_LIMIT
;
2919 n
= get_node(cachep
, node
);
2921 BUG_ON(ac
->avail
> 0 || !n
);
2922 shared
= READ_ONCE(n
->shared
);
2923 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
2926 spin_lock(&n
->list_lock
);
2927 shared
= READ_ONCE(n
->shared
);
2929 /* See if we can refill from the shared array */
2930 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
2931 shared
->touched
= 1;
2935 while (batchcount
> 0) {
2936 /* Get slab alloc is to come from. */
2937 page
= get_first_slab(n
, false);
2941 check_spinlock_acquired(cachep
);
2943 batchcount
= alloc_block(cachep
, ac
, page
, batchcount
);
2944 fixup_slab_list(cachep
, n
, page
, &list
);
2948 n
->free_objects
-= ac
->avail
;
2950 spin_unlock(&n
->list_lock
);
2951 fixup_objfreelist_debug(cachep
, &list
);
2954 if (unlikely(!ac
->avail
)) {
2955 /* Check if we can use obj in pfmemalloc slab */
2956 if (sk_memalloc_socks()) {
2957 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2963 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
2966 * cache_grow_begin() can reenable interrupts,
2967 * then ac could change.
2969 ac
= cpu_cache_get(cachep
);
2970 if (!ac
->avail
&& page
)
2971 alloc_block(cachep
, ac
, page
, batchcount
);
2972 cache_grow_end(cachep
, page
);
2979 return ac
->entry
[--ac
->avail
];
2982 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
2985 might_sleep_if(gfpflags_allow_blocking(flags
));
2989 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2990 gfp_t flags
, void *objp
, unsigned long caller
)
2992 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2995 if (cachep
->flags
& SLAB_POISON
) {
2996 check_poison_obj(cachep
, objp
);
2997 slab_kernel_map(cachep
, objp
, 1);
2998 poison_obj(cachep
, objp
, POISON_INUSE
);
3000 if (cachep
->flags
& SLAB_STORE_USER
)
3001 *dbg_userword(cachep
, objp
) = (void *)caller
;
3003 if (cachep
->flags
& SLAB_RED_ZONE
) {
3004 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3005 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3006 slab_error(cachep
, "double free, or memory outside object was overwritten");
3007 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3008 objp
, *dbg_redzone1(cachep
, objp
),
3009 *dbg_redzone2(cachep
, objp
));
3011 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3012 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3015 objp
+= obj_offset(cachep
);
3016 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3018 if (ARCH_SLAB_MINALIGN
&&
3019 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3020 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3021 objp
, (int)ARCH_SLAB_MINALIGN
);
3026 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3029 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3032 struct array_cache
*ac
;
3036 ac
= cpu_cache_get(cachep
);
3037 if (likely(ac
->avail
)) {
3039 objp
= ac
->entry
[--ac
->avail
];
3041 STATS_INC_ALLOCHIT(cachep
);
3045 STATS_INC_ALLOCMISS(cachep
);
3046 objp
= cache_alloc_refill(cachep
, flags
);
3048 * the 'ac' may be updated by cache_alloc_refill(),
3049 * and kmemleak_erase() requires its correct value.
3051 ac
= cpu_cache_get(cachep
);
3055 * To avoid a false negative, if an object that is in one of the
3056 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3057 * treat the array pointers as a reference to the object.
3060 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3066 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3068 * If we are in_interrupt, then process context, including cpusets and
3069 * mempolicy, may not apply and should not be used for allocation policy.
3071 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3073 int nid_alloc
, nid_here
;
3075 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3077 nid_alloc
= nid_here
= numa_mem_id();
3078 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3079 nid_alloc
= cpuset_slab_spread_node();
3080 else if (current
->mempolicy
)
3081 nid_alloc
= mempolicy_slab_node();
3082 if (nid_alloc
!= nid_here
)
3083 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3088 * Fallback function if there was no memory available and no objects on a
3089 * certain node and fall back is permitted. First we scan all the
3090 * available node for available objects. If that fails then we
3091 * perform an allocation without specifying a node. This allows the page
3092 * allocator to do its reclaim / fallback magic. We then insert the
3093 * slab into the proper nodelist and then allocate from it.
3095 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3097 struct zonelist
*zonelist
;
3100 enum zone_type high_zoneidx
= gfp_zone(flags
);
3104 unsigned int cpuset_mems_cookie
;
3106 if (flags
& __GFP_THISNODE
)
3110 cpuset_mems_cookie
= read_mems_allowed_begin();
3111 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3115 * Look through allowed nodes for objects available
3116 * from existing per node queues.
3118 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3119 nid
= zone_to_nid(zone
);
3121 if (cpuset_zone_allowed(zone
, flags
) &&
3122 get_node(cache
, nid
) &&
3123 get_node(cache
, nid
)->free_objects
) {
3124 obj
= ____cache_alloc_node(cache
,
3125 gfp_exact_node(flags
), nid
);
3133 * This allocation will be performed within the constraints
3134 * of the current cpuset / memory policy requirements.
3135 * We may trigger various forms of reclaim on the allowed
3136 * set and go into memory reserves if necessary.
3138 page
= cache_grow_begin(cache
, flags
, numa_mem_id());
3139 cache_grow_end(cache
, page
);
3141 nid
= page_to_nid(page
);
3142 obj
= ____cache_alloc_node(cache
,
3143 gfp_exact_node(flags
), nid
);
3146 * Another processor may allocate the objects in
3147 * the slab since we are not holding any locks.
3154 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3160 * A interface to enable slab creation on nodeid
3162 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3166 struct kmem_cache_node
*n
;
3170 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3171 n
= get_node(cachep
, nodeid
);
3175 spin_lock(&n
->list_lock
);
3176 page
= get_first_slab(n
, false);
3180 check_spinlock_acquired_node(cachep
, nodeid
);
3182 STATS_INC_NODEALLOCS(cachep
);
3183 STATS_INC_ACTIVE(cachep
);
3184 STATS_SET_HIGH(cachep
);
3186 BUG_ON(page
->active
== cachep
->num
);
3188 obj
= slab_get_obj(cachep
, page
);
3191 fixup_slab_list(cachep
, n
, page
, &list
);
3193 spin_unlock(&n
->list_lock
);
3194 fixup_objfreelist_debug(cachep
, &list
);
3198 spin_unlock(&n
->list_lock
);
3199 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3201 /* This slab isn't counted yet so don't update free_objects */
3202 obj
= slab_get_obj(cachep
, page
);
3204 cache_grow_end(cachep
, page
);
3206 return obj
? obj
: fallback_alloc(cachep
, flags
);
3209 static __always_inline
void *
3210 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3211 unsigned long caller
)
3213 unsigned long save_flags
;
3215 int slab_node
= numa_mem_id();
3217 flags
&= gfp_allowed_mask
;
3218 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3219 if (unlikely(!cachep
))
3222 cache_alloc_debugcheck_before(cachep
, flags
);
3223 local_irq_save(save_flags
);
3225 if (nodeid
== NUMA_NO_NODE
)
3228 if (unlikely(!get_node(cachep
, nodeid
))) {
3229 /* Node not bootstrapped yet */
3230 ptr
= fallback_alloc(cachep
, flags
);
3234 if (nodeid
== slab_node
) {
3236 * Use the locally cached objects if possible.
3237 * However ____cache_alloc does not allow fallback
3238 * to other nodes. It may fail while we still have
3239 * objects on other nodes available.
3241 ptr
= ____cache_alloc(cachep
, flags
);
3245 /* ___cache_alloc_node can fall back to other nodes */
3246 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3248 local_irq_restore(save_flags
);
3249 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3251 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3252 memset(ptr
, 0, cachep
->object_size
);
3254 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3258 static __always_inline
void *
3259 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3263 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3264 objp
= alternate_node_alloc(cache
, flags
);
3268 objp
= ____cache_alloc(cache
, flags
);
3271 * We may just have run out of memory on the local node.
3272 * ____cache_alloc_node() knows how to locate memory on other nodes
3275 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3282 static __always_inline
void *
3283 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3285 return ____cache_alloc(cachep
, flags
);
3288 #endif /* CONFIG_NUMA */
3290 static __always_inline
void *
3291 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3293 unsigned long save_flags
;
3296 flags
&= gfp_allowed_mask
;
3297 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3298 if (unlikely(!cachep
))
3301 cache_alloc_debugcheck_before(cachep
, flags
);
3302 local_irq_save(save_flags
);
3303 objp
= __do_cache_alloc(cachep
, flags
);
3304 local_irq_restore(save_flags
);
3305 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3308 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3309 memset(objp
, 0, cachep
->object_size
);
3311 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3316 * Caller needs to acquire correct kmem_cache_node's list_lock
3317 * @list: List of detached free slabs should be freed by caller
3319 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3320 int nr_objects
, int node
, struct list_head
*list
)
3323 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3326 n
->free_objects
+= nr_objects
;
3328 for (i
= 0; i
< nr_objects
; i
++) {
3334 page
= virt_to_head_page(objp
);
3335 list_del(&page
->slab_list
);
3336 check_spinlock_acquired_node(cachep
, node
);
3337 slab_put_obj(cachep
, page
, objp
);
3338 STATS_DEC_ACTIVE(cachep
);
3340 /* fixup slab chains */
3341 if (page
->active
== 0) {
3342 list_add(&page
->slab_list
, &n
->slabs_free
);
3345 /* Unconditionally move a slab to the end of the
3346 * partial list on free - maximum time for the
3347 * other objects to be freed, too.
3349 list_add_tail(&page
->slab_list
, &n
->slabs_partial
);
3353 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3354 n
->free_objects
-= cachep
->num
;
3356 page
= list_last_entry(&n
->slabs_free
, struct page
, slab_list
);
3357 list_move(&page
->slab_list
, list
);
3363 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3366 struct kmem_cache_node
*n
;
3367 int node
= numa_mem_id();
3370 batchcount
= ac
->batchcount
;
3373 n
= get_node(cachep
, node
);
3374 spin_lock(&n
->list_lock
);
3376 struct array_cache
*shared_array
= n
->shared
;
3377 int max
= shared_array
->limit
- shared_array
->avail
;
3379 if (batchcount
> max
)
3381 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3382 ac
->entry
, sizeof(void *) * batchcount
);
3383 shared_array
->avail
+= batchcount
;
3388 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3395 list_for_each_entry(page
, &n
->slabs_free
, slab_list
) {
3396 BUG_ON(page
->active
);
3400 STATS_SET_FREEABLE(cachep
, i
);
3403 spin_unlock(&n
->list_lock
);
3404 slabs_destroy(cachep
, &list
);
3405 ac
->avail
-= batchcount
;
3406 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3410 * Release an obj back to its cache. If the obj has a constructed state, it must
3411 * be in this state _before_ it is released. Called with disabled ints.
3413 static __always_inline
void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3414 unsigned long caller
)
3416 /* Put the object into the quarantine, don't touch it for now. */
3417 if (kasan_slab_free(cachep
, objp
, _RET_IP_
))
3420 ___cache_free(cachep
, objp
, caller
);
3423 void ___cache_free(struct kmem_cache
*cachep
, void *objp
,
3424 unsigned long caller
)
3426 struct array_cache
*ac
= cpu_cache_get(cachep
);
3429 kmemleak_free_recursive(objp
, cachep
->flags
);
3430 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3433 * Skip calling cache_free_alien() when the platform is not numa.
3434 * This will avoid cache misses that happen while accessing slabp (which
3435 * is per page memory reference) to get nodeid. Instead use a global
3436 * variable to skip the call, which is mostly likely to be present in
3439 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3442 if (ac
->avail
< ac
->limit
) {
3443 STATS_INC_FREEHIT(cachep
);
3445 STATS_INC_FREEMISS(cachep
);
3446 cache_flusharray(cachep
, ac
);
3449 if (sk_memalloc_socks()) {
3450 struct page
*page
= virt_to_head_page(objp
);
3452 if (unlikely(PageSlabPfmemalloc(page
))) {
3453 cache_free_pfmemalloc(cachep
, page
, objp
);
3458 ac
->entry
[ac
->avail
++] = objp
;
3462 * kmem_cache_alloc - Allocate an object
3463 * @cachep: The cache to allocate from.
3464 * @flags: See kmalloc().
3466 * Allocate an object from this cache. The flags are only relevant
3467 * if the cache has no available objects.
3469 * Return: pointer to the new object or %NULL in case of error
3471 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3473 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3475 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3476 cachep
->object_size
, cachep
->size
, flags
);
3480 EXPORT_SYMBOL(kmem_cache_alloc
);
3482 static __always_inline
void
3483 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3484 size_t size
, void **p
, unsigned long caller
)
3488 for (i
= 0; i
< size
; i
++)
3489 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3492 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3497 s
= slab_pre_alloc_hook(s
, flags
);
3501 cache_alloc_debugcheck_before(s
, flags
);
3503 local_irq_disable();
3504 for (i
= 0; i
< size
; i
++) {
3505 void *objp
= __do_cache_alloc(s
, flags
);
3507 if (unlikely(!objp
))
3513 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3515 /* Clear memory outside IRQ disabled section */
3516 if (unlikely(flags
& __GFP_ZERO
))
3517 for (i
= 0; i
< size
; i
++)
3518 memset(p
[i
], 0, s
->object_size
);
3520 slab_post_alloc_hook(s
, flags
, size
, p
);
3521 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3525 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3526 slab_post_alloc_hook(s
, flags
, i
, p
);
3527 __kmem_cache_free_bulk(s
, i
, p
);
3530 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3532 #ifdef CONFIG_TRACING
3534 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3538 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3540 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3541 trace_kmalloc(_RET_IP_
, ret
,
3542 size
, cachep
->size
, flags
);
3545 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3550 * kmem_cache_alloc_node - Allocate an object on the specified node
3551 * @cachep: The cache to allocate from.
3552 * @flags: See kmalloc().
3553 * @nodeid: node number of the target node.
3555 * Identical to kmem_cache_alloc but it will allocate memory on the given
3556 * node, which can improve the performance for cpu bound structures.
3558 * Fallback to other node is possible if __GFP_THISNODE is not set.
3560 * Return: pointer to the new object or %NULL in case of error
3562 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3564 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3566 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3567 cachep
->object_size
, cachep
->size
,
3572 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3574 #ifdef CONFIG_TRACING
3575 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3582 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3584 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3585 trace_kmalloc_node(_RET_IP_
, ret
,
3590 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3593 static __always_inline
void *
3594 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3596 struct kmem_cache
*cachep
;
3599 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3601 cachep
= kmalloc_slab(size
, flags
);
3602 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3604 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3605 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3610 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3612 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3614 EXPORT_SYMBOL(__kmalloc_node
);
3616 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3617 int node
, unsigned long caller
)
3619 return __do_kmalloc_node(size
, flags
, node
, caller
);
3621 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3622 #endif /* CONFIG_NUMA */
3625 * __do_kmalloc - allocate memory
3626 * @size: how many bytes of memory are required.
3627 * @flags: the type of memory to allocate (see kmalloc).
3628 * @caller: function caller for debug tracking of the caller
3630 * Return: pointer to the allocated memory or %NULL in case of error
3632 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3633 unsigned long caller
)
3635 struct kmem_cache
*cachep
;
3638 if (unlikely(size
> KMALLOC_MAX_CACHE_SIZE
))
3640 cachep
= kmalloc_slab(size
, flags
);
3641 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3643 ret
= slab_alloc(cachep
, flags
, caller
);
3645 ret
= kasan_kmalloc(cachep
, ret
, size
, flags
);
3646 trace_kmalloc(caller
, ret
,
3647 size
, cachep
->size
, flags
);
3652 void *__kmalloc(size_t size
, gfp_t flags
)
3654 return __do_kmalloc(size
, flags
, _RET_IP_
);
3656 EXPORT_SYMBOL(__kmalloc
);
3658 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3660 return __do_kmalloc(size
, flags
, caller
);
3662 EXPORT_SYMBOL(__kmalloc_track_caller
);
3665 * kmem_cache_free - Deallocate an object
3666 * @cachep: The cache the allocation was from.
3667 * @objp: The previously allocated object.
3669 * Free an object which was previously allocated from this
3672 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3674 unsigned long flags
;
3675 cachep
= cache_from_obj(cachep
, objp
);
3679 local_irq_save(flags
);
3680 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3681 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3682 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3683 __cache_free(cachep
, objp
, _RET_IP_
);
3684 local_irq_restore(flags
);
3686 trace_kmem_cache_free(_RET_IP_
, objp
);
3688 EXPORT_SYMBOL(kmem_cache_free
);
3690 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3692 struct kmem_cache
*s
;
3695 local_irq_disable();
3696 for (i
= 0; i
< size
; i
++) {
3699 if (!orig_s
) /* called via kfree_bulk */
3700 s
= virt_to_cache(objp
);
3702 s
= cache_from_obj(orig_s
, objp
);
3706 debug_check_no_locks_freed(objp
, s
->object_size
);
3707 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3708 debug_check_no_obj_freed(objp
, s
->object_size
);
3710 __cache_free(s
, objp
, _RET_IP_
);
3714 /* FIXME: add tracing */
3716 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3719 * kfree - free previously allocated memory
3720 * @objp: pointer returned by kmalloc.
3722 * If @objp is NULL, no operation is performed.
3724 * Don't free memory not originally allocated by kmalloc()
3725 * or you will run into trouble.
3727 void kfree(const void *objp
)
3729 struct kmem_cache
*c
;
3730 unsigned long flags
;
3732 trace_kfree(_RET_IP_
, objp
);
3734 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3736 local_irq_save(flags
);
3737 kfree_debugcheck(objp
);
3738 c
= virt_to_cache(objp
);
3740 local_irq_restore(flags
);
3743 debug_check_no_locks_freed(objp
, c
->object_size
);
3745 debug_check_no_obj_freed(objp
, c
->object_size
);
3746 __cache_free(c
, (void *)objp
, _RET_IP_
);
3747 local_irq_restore(flags
);
3749 EXPORT_SYMBOL(kfree
);
3752 * This initializes kmem_cache_node or resizes various caches for all nodes.
3754 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3758 struct kmem_cache_node
*n
;
3760 for_each_online_node(node
) {
3761 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3770 if (!cachep
->list
.next
) {
3771 /* Cache is not active yet. Roll back what we did */
3774 n
= get_node(cachep
, node
);
3777 free_alien_cache(n
->alien
);
3779 cachep
->node
[node
] = NULL
;
3787 /* Always called with the slab_mutex held */
3788 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3789 int batchcount
, int shared
, gfp_t gfp
)
3791 struct array_cache __percpu
*cpu_cache
, *prev
;
3794 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3798 prev
= cachep
->cpu_cache
;
3799 cachep
->cpu_cache
= cpu_cache
;
3801 * Without a previous cpu_cache there's no need to synchronize remote
3802 * cpus, so skip the IPIs.
3805 kick_all_cpus_sync();
3808 cachep
->batchcount
= batchcount
;
3809 cachep
->limit
= limit
;
3810 cachep
->shared
= shared
;
3815 for_each_online_cpu(cpu
) {
3818 struct kmem_cache_node
*n
;
3819 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3821 node
= cpu_to_mem(cpu
);
3822 n
= get_node(cachep
, node
);
3823 spin_lock_irq(&n
->list_lock
);
3824 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3825 spin_unlock_irq(&n
->list_lock
);
3826 slabs_destroy(cachep
, &list
);
3831 return setup_kmem_cache_nodes(cachep
, gfp
);
3834 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3835 int batchcount
, int shared
, gfp_t gfp
)
3838 struct kmem_cache
*c
;
3840 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3842 if (slab_state
< FULL
)
3845 if ((ret
< 0) || !is_root_cache(cachep
))
3848 lockdep_assert_held(&slab_mutex
);
3849 for_each_memcg_cache(c
, cachep
) {
3850 /* return value determined by the root cache only */
3851 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3857 /* Called with slab_mutex held always */
3858 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3865 err
= cache_random_seq_create(cachep
, cachep
->num
, gfp
);
3869 if (!is_root_cache(cachep
)) {
3870 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3871 limit
= root
->limit
;
3872 shared
= root
->shared
;
3873 batchcount
= root
->batchcount
;
3876 if (limit
&& shared
&& batchcount
)
3879 * The head array serves three purposes:
3880 * - create a LIFO ordering, i.e. return objects that are cache-warm
3881 * - reduce the number of spinlock operations.
3882 * - reduce the number of linked list operations on the slab and
3883 * bufctl chains: array operations are cheaper.
3884 * The numbers are guessed, we should auto-tune as described by
3887 if (cachep
->size
> 131072)
3889 else if (cachep
->size
> PAGE_SIZE
)
3891 else if (cachep
->size
> 1024)
3893 else if (cachep
->size
> 256)
3899 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3900 * allocation behaviour: Most allocs on one cpu, most free operations
3901 * on another cpu. For these cases, an efficient object passing between
3902 * cpus is necessary. This is provided by a shared array. The array
3903 * replaces Bonwick's magazine layer.
3904 * On uniprocessor, it's functionally equivalent (but less efficient)
3905 * to a larger limit. Thus disabled by default.
3908 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3913 * With debugging enabled, large batchcount lead to excessively long
3914 * periods with disabled local interrupts. Limit the batchcount
3919 batchcount
= (limit
+ 1) / 2;
3921 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3924 pr_err("enable_cpucache failed for %s, error %d\n",
3925 cachep
->name
, -err
);
3930 * Drain an array if it contains any elements taking the node lock only if
3931 * necessary. Note that the node listlock also protects the array_cache
3932 * if drain_array() is used on the shared array.
3934 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3935 struct array_cache
*ac
, int node
)
3939 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3940 check_mutex_acquired();
3942 if (!ac
|| !ac
->avail
)
3950 spin_lock_irq(&n
->list_lock
);
3951 drain_array_locked(cachep
, ac
, node
, false, &list
);
3952 spin_unlock_irq(&n
->list_lock
);
3954 slabs_destroy(cachep
, &list
);
3958 * cache_reap - Reclaim memory from caches.
3959 * @w: work descriptor
3961 * Called from workqueue/eventd every few seconds.
3963 * - clear the per-cpu caches for this CPU.
3964 * - return freeable pages to the main free memory pool.
3966 * If we cannot acquire the cache chain mutex then just give up - we'll try
3967 * again on the next iteration.
3969 static void cache_reap(struct work_struct
*w
)
3971 struct kmem_cache
*searchp
;
3972 struct kmem_cache_node
*n
;
3973 int node
= numa_mem_id();
3974 struct delayed_work
*work
= to_delayed_work(w
);
3976 if (!mutex_trylock(&slab_mutex
))
3977 /* Give up. Setup the next iteration. */
3980 list_for_each_entry(searchp
, &slab_caches
, list
) {
3984 * We only take the node lock if absolutely necessary and we
3985 * have established with reasonable certainty that
3986 * we can do some work if the lock was obtained.
3988 n
= get_node(searchp
, node
);
3990 reap_alien(searchp
, n
);
3992 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3995 * These are racy checks but it does not matter
3996 * if we skip one check or scan twice.
3998 if (time_after(n
->next_reap
, jiffies
))
4001 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
4003 drain_array(searchp
, n
, n
->shared
, node
);
4005 if (n
->free_touched
)
4006 n
->free_touched
= 0;
4010 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4011 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4012 STATS_ADD_REAPED(searchp
, freed
);
4018 mutex_unlock(&slab_mutex
);
4021 /* Set up the next iteration */
4022 schedule_delayed_work_on(smp_processor_id(), work
,
4023 round_jiffies_relative(REAPTIMEOUT_AC
));
4026 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4028 unsigned long active_objs
, num_objs
, active_slabs
;
4029 unsigned long total_slabs
= 0, free_objs
= 0, shared_avail
= 0;
4030 unsigned long free_slabs
= 0;
4032 struct kmem_cache_node
*n
;
4034 for_each_kmem_cache_node(cachep
, node
, n
) {
4036 spin_lock_irq(&n
->list_lock
);
4038 total_slabs
+= n
->total_slabs
;
4039 free_slabs
+= n
->free_slabs
;
4040 free_objs
+= n
->free_objects
;
4043 shared_avail
+= n
->shared
->avail
;
4045 spin_unlock_irq(&n
->list_lock
);
4047 num_objs
= total_slabs
* cachep
->num
;
4048 active_slabs
= total_slabs
- free_slabs
;
4049 active_objs
= num_objs
- free_objs
;
4051 sinfo
->active_objs
= active_objs
;
4052 sinfo
->num_objs
= num_objs
;
4053 sinfo
->active_slabs
= active_slabs
;
4054 sinfo
->num_slabs
= total_slabs
;
4055 sinfo
->shared_avail
= shared_avail
;
4056 sinfo
->limit
= cachep
->limit
;
4057 sinfo
->batchcount
= cachep
->batchcount
;
4058 sinfo
->shared
= cachep
->shared
;
4059 sinfo
->objects_per_slab
= cachep
->num
;
4060 sinfo
->cache_order
= cachep
->gfporder
;
4063 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4067 unsigned long high
= cachep
->high_mark
;
4068 unsigned long allocs
= cachep
->num_allocations
;
4069 unsigned long grown
= cachep
->grown
;
4070 unsigned long reaped
= cachep
->reaped
;
4071 unsigned long errors
= cachep
->errors
;
4072 unsigned long max_freeable
= cachep
->max_freeable
;
4073 unsigned long node_allocs
= cachep
->node_allocs
;
4074 unsigned long node_frees
= cachep
->node_frees
;
4075 unsigned long overflows
= cachep
->node_overflow
;
4077 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4078 allocs
, high
, grown
,
4079 reaped
, errors
, max_freeable
, node_allocs
,
4080 node_frees
, overflows
);
4084 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4085 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4086 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4087 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4089 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4090 allochit
, allocmiss
, freehit
, freemiss
);
4095 #define MAX_SLABINFO_WRITE 128
4097 * slabinfo_write - Tuning for the slab allocator
4099 * @buffer: user buffer
4100 * @count: data length
4103 * Return: %0 on success, negative error code otherwise.
4105 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4106 size_t count
, loff_t
*ppos
)
4108 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4109 int limit
, batchcount
, shared
, res
;
4110 struct kmem_cache
*cachep
;
4112 if (count
> MAX_SLABINFO_WRITE
)
4114 if (copy_from_user(&kbuf
, buffer
, count
))
4116 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4118 tmp
= strchr(kbuf
, ' ');
4123 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4126 /* Find the cache in the chain of caches. */
4127 mutex_lock(&slab_mutex
);
4129 list_for_each_entry(cachep
, &slab_caches
, list
) {
4130 if (!strcmp(cachep
->name
, kbuf
)) {
4131 if (limit
< 1 || batchcount
< 1 ||
4132 batchcount
> limit
|| shared
< 0) {
4135 res
= do_tune_cpucache(cachep
, limit
,
4142 mutex_unlock(&slab_mutex
);
4148 #ifdef CONFIG_HARDENED_USERCOPY
4150 * Rejects incorrectly sized objects and objects that are to be copied
4151 * to/from userspace but do not fall entirely within the containing slab
4152 * cache's usercopy region.
4154 * Returns NULL if check passes, otherwise const char * to name of cache
4155 * to indicate an error.
4157 void __check_heap_object(const void *ptr
, unsigned long n
, struct page
*page
,
4160 struct kmem_cache
*cachep
;
4162 unsigned long offset
;
4164 ptr
= kasan_reset_tag(ptr
);
4166 /* Find and validate object. */
4167 cachep
= page
->slab_cache
;
4168 objnr
= obj_to_index(cachep
, page
, (void *)ptr
);
4169 BUG_ON(objnr
>= cachep
->num
);
4171 /* Find offset within object. */
4172 offset
= ptr
- index_to_obj(cachep
, page
, objnr
) - obj_offset(cachep
);
4174 /* Allow address range falling entirely within usercopy region. */
4175 if (offset
>= cachep
->useroffset
&&
4176 offset
- cachep
->useroffset
<= cachep
->usersize
&&
4177 n
<= cachep
->useroffset
- offset
+ cachep
->usersize
)
4181 * If the copy is still within the allocated object, produce
4182 * a warning instead of rejecting the copy. This is intended
4183 * to be a temporary method to find any missing usercopy
4186 if (usercopy_fallback
&&
4187 offset
<= cachep
->object_size
&&
4188 n
<= cachep
->object_size
- offset
) {
4189 usercopy_warn("SLAB object", cachep
->name
, to_user
, offset
, n
);
4193 usercopy_abort("SLAB object", cachep
->name
, to_user
, offset
, n
);
4195 #endif /* CONFIG_HARDENED_USERCOPY */
4198 * __ksize -- Uninstrumented ksize.
4200 * Unlike ksize(), __ksize() is uninstrumented, and does not provide the same
4201 * safety checks as ksize() with KASAN instrumentation enabled.
4203 size_t __ksize(const void *objp
)
4205 struct kmem_cache
*c
;
4209 if (unlikely(objp
== ZERO_SIZE_PTR
))
4212 c
= virt_to_cache(objp
);
4213 size
= c
? c
->object_size
: 0;
4217 EXPORT_SYMBOL(__ksize
);