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/kfence.h>
104 #include <linux/cpu.h>
105 #include <linux/sysctl.h>
106 #include <linux/module.h>
107 #include <linux/rcupdate.h>
108 #include <linux/string.h>
109 #include <linux/uaccess.h>
110 #include <linux/nodemask.h>
111 #include <linux/kmemleak.h>
112 #include <linux/mempolicy.h>
113 #include <linux/mutex.h>
114 #include <linux/fault-inject.h>
115 #include <linux/rtmutex.h>
116 #include <linux/reciprocal_div.h>
117 #include <linux/debugobjects.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
120 #include <linux/sched/task_stack.h>
122 #include <net/sock.h>
124 #include <asm/cacheflush.h>
125 #include <asm/tlbflush.h>
126 #include <asm/page.h>
128 #include <trace/events/kmem.h>
130 #include "internal.h"
135 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * STATS - 1 to collect stats for /proc/slabinfo.
139 * 0 for faster, smaller code (especially in the critical paths).
141 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
144 #ifdef CONFIG_DEBUG_SLAB
147 #define FORCED_DEBUG 1
151 #define FORCED_DEBUG 0
154 /* Shouldn't this be in a header file somewhere? */
155 #define BYTES_PER_WORD sizeof(void *)
156 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
158 #ifndef ARCH_KMALLOC_FLAGS
159 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
162 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
163 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
165 #if FREELIST_BYTE_INDEX
166 typedef unsigned char freelist_idx_t
;
168 typedef unsigned short freelist_idx_t
;
171 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
177 * - LIFO ordering, to hand out cache-warm objects from _alloc
178 * - reduce the number of linked list operations
179 * - reduce spinlock operations
181 * The limit is stored in the per-cpu structure to reduce the data cache
188 unsigned int batchcount
;
189 unsigned int touched
;
191 * Must have this definition in here for the proper
192 * alignment of array_cache. Also simplifies accessing
199 struct array_cache ac
;
203 * Need this for bootstrapping a per node allocator.
205 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
206 static struct kmem_cache_node __initdata init_kmem_cache_node
[NUM_INIT_LISTS
];
207 #define CACHE_CACHE 0
208 #define SIZE_NODE (MAX_NUMNODES)
210 static int drain_freelist(struct kmem_cache
*cache
,
211 struct kmem_cache_node
*n
, int tofree
);
212 static void free_block(struct kmem_cache
*cachep
, void **objpp
, int len
,
213 int node
, struct list_head
*list
);
214 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
);
215 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
);
216 static void cache_reap(struct work_struct
*unused
);
218 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
220 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
221 struct kmem_cache_node
*n
, struct slab
*slab
,
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 raw_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 unnecessary
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
,
375 const struct slab
*slab
, unsigned int idx
)
377 return slab
->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
);
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 slab
*slab
, void *objp
)
554 struct kmem_cache_node
*n
;
558 slab_node
= slab_nid(slab
);
559 n
= get_node(cachep
, slab_node
);
561 raw_spin_lock(&n
->list_lock
);
562 free_block(cachep
, &objp
, 1, slab_node
, &list
);
563 raw_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
,
591 /* &alien->lock must be held by alien callers. */
592 static __always_inline
void __free_one(struct array_cache
*ac
, void *objp
)
594 /* Avoid trivial double-free. */
595 if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED
) &&
596 WARN_ON_ONCE(ac
->avail
> 0 && ac
->entry
[ac
->avail
- 1] == objp
))
598 ac
->entry
[ac
->avail
++] = objp
;
603 #define drain_alien_cache(cachep, alien) do { } while (0)
604 #define reap_alien(cachep, n) do { } while (0)
606 static inline struct alien_cache
**alloc_alien_cache(int node
,
607 int limit
, gfp_t gfp
)
612 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
616 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
621 static inline gfp_t
gfp_exact_node(gfp_t flags
)
623 return flags
& ~__GFP_NOFAIL
;
626 #else /* CONFIG_NUMA */
628 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
629 int batch
, gfp_t gfp
)
631 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
632 struct alien_cache
*alc
= NULL
;
634 alc
= kmalloc_node(memsize
, gfp
, node
);
636 kmemleak_no_scan(alc
);
637 init_arraycache(&alc
->ac
, entries
, batch
);
638 spin_lock_init(&alc
->lock
);
643 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
645 struct alien_cache
**alc_ptr
;
650 alc_ptr
= kcalloc_node(nr_node_ids
, sizeof(void *), gfp
, node
);
655 if (i
== node
|| !node_online(i
))
657 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
659 for (i
--; i
>= 0; i
--)
668 static void free_alien_cache(struct alien_cache
**alc_ptr
)
679 static void __drain_alien_cache(struct kmem_cache
*cachep
,
680 struct array_cache
*ac
, int node
,
681 struct list_head
*list
)
683 struct kmem_cache_node
*n
= get_node(cachep
, node
);
686 raw_spin_lock(&n
->list_lock
);
688 * Stuff objects into the remote nodes shared array first.
689 * That way we could avoid the overhead of putting the objects
690 * into the free lists and getting them back later.
693 transfer_objects(n
->shared
, ac
, ac
->limit
);
695 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
697 raw_spin_unlock(&n
->list_lock
);
702 * Called from cache_reap() to regularly drain alien caches round robin.
704 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
706 int node
= __this_cpu_read(slab_reap_node
);
709 struct alien_cache
*alc
= n
->alien
[node
];
710 struct array_cache
*ac
;
714 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
717 __drain_alien_cache(cachep
, ac
, node
, &list
);
718 spin_unlock_irq(&alc
->lock
);
719 slabs_destroy(cachep
, &list
);
725 static void drain_alien_cache(struct kmem_cache
*cachep
,
726 struct alien_cache
**alien
)
729 struct alien_cache
*alc
;
730 struct array_cache
*ac
;
733 for_each_online_node(i
) {
739 spin_lock_irqsave(&alc
->lock
, flags
);
740 __drain_alien_cache(cachep
, ac
, i
, &list
);
741 spin_unlock_irqrestore(&alc
->lock
, flags
);
742 slabs_destroy(cachep
, &list
);
747 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
748 int node
, int slab_node
)
750 struct kmem_cache_node
*n
;
751 struct alien_cache
*alien
= NULL
;
752 struct array_cache
*ac
;
755 n
= get_node(cachep
, node
);
756 STATS_INC_NODEFREES(cachep
);
757 if (n
->alien
&& n
->alien
[slab_node
]) {
758 alien
= n
->alien
[slab_node
];
760 spin_lock(&alien
->lock
);
761 if (unlikely(ac
->avail
== ac
->limit
)) {
762 STATS_INC_ACOVERFLOW(cachep
);
763 __drain_alien_cache(cachep
, ac
, slab_node
, &list
);
765 __free_one(ac
, objp
);
766 spin_unlock(&alien
->lock
);
767 slabs_destroy(cachep
, &list
);
769 n
= get_node(cachep
, slab_node
);
770 raw_spin_lock(&n
->list_lock
);
771 free_block(cachep
, &objp
, 1, slab_node
, &list
);
772 raw_spin_unlock(&n
->list_lock
);
773 slabs_destroy(cachep
, &list
);
778 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
780 int slab_node
= slab_nid(virt_to_slab(objp
));
781 int node
= numa_mem_id();
783 * Make sure we are not freeing an object from another node to the array
786 if (likely(node
== slab_node
))
789 return __cache_free_alien(cachep
, objp
, node
, slab_node
);
793 * Construct gfp mask to allocate from a specific node but do not reclaim or
794 * warn about failures.
796 static inline gfp_t
gfp_exact_node(gfp_t flags
)
798 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
802 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
804 struct kmem_cache_node
*n
;
807 * Set up the kmem_cache_node for cpu before we can
808 * begin anything. Make sure some other cpu on this
809 * node has not already allocated this
811 n
= get_node(cachep
, node
);
813 raw_spin_lock_irq(&n
->list_lock
);
814 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
816 raw_spin_unlock_irq(&n
->list_lock
);
821 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
825 kmem_cache_node_init(n
);
826 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
827 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
830 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
833 * The kmem_cache_nodes don't come and go as CPUs
834 * come and go. slab_mutex provides sufficient
837 cachep
->node
[node
] = n
;
842 #if defined(CONFIG_NUMA) || defined(CONFIG_SMP)
844 * Allocates and initializes node for a node on each slab cache, used for
845 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
846 * will be allocated off-node since memory is not yet online for the new node.
847 * When hotplugging memory or a cpu, existing nodes are not replaced if
850 * Must hold slab_mutex.
852 static int init_cache_node_node(int node
)
855 struct kmem_cache
*cachep
;
857 list_for_each_entry(cachep
, &slab_caches
, list
) {
858 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
867 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
868 int node
, gfp_t gfp
, bool force_change
)
871 struct kmem_cache_node
*n
;
872 struct array_cache
*old_shared
= NULL
;
873 struct array_cache
*new_shared
= NULL
;
874 struct alien_cache
**new_alien
= NULL
;
877 if (use_alien_caches
) {
878 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
883 if (cachep
->shared
) {
884 new_shared
= alloc_arraycache(node
,
885 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
890 ret
= init_cache_node(cachep
, node
, gfp
);
894 n
= get_node(cachep
, node
);
895 raw_spin_lock_irq(&n
->list_lock
);
896 if (n
->shared
&& force_change
) {
897 free_block(cachep
, n
->shared
->entry
,
898 n
->shared
->avail
, node
, &list
);
899 n
->shared
->avail
= 0;
902 if (!n
->shared
|| force_change
) {
903 old_shared
= n
->shared
;
904 n
->shared
= new_shared
;
909 n
->alien
= new_alien
;
913 raw_spin_unlock_irq(&n
->list_lock
);
914 slabs_destroy(cachep
, &list
);
917 * To protect lockless access to n->shared during irq disabled context.
918 * If n->shared isn't NULL in irq disabled context, accessing to it is
919 * guaranteed to be valid until irq is re-enabled, because it will be
920 * freed after synchronize_rcu().
922 if (old_shared
&& force_change
)
928 free_alien_cache(new_alien
);
935 static void cpuup_canceled(long cpu
)
937 struct kmem_cache
*cachep
;
938 struct kmem_cache_node
*n
= NULL
;
939 int node
= cpu_to_mem(cpu
);
940 const struct cpumask
*mask
= cpumask_of_node(node
);
942 list_for_each_entry(cachep
, &slab_caches
, list
) {
943 struct array_cache
*nc
;
944 struct array_cache
*shared
;
945 struct alien_cache
**alien
;
948 n
= get_node(cachep
, node
);
952 raw_spin_lock_irq(&n
->list_lock
);
954 /* Free limit for this kmem_cache_node */
955 n
->free_limit
-= cachep
->batchcount
;
957 /* cpu is dead; no one can alloc from it. */
958 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
959 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
962 if (!cpumask_empty(mask
)) {
963 raw_spin_unlock_irq(&n
->list_lock
);
969 free_block(cachep
, shared
->entry
,
970 shared
->avail
, node
, &list
);
977 raw_spin_unlock_irq(&n
->list_lock
);
981 drain_alien_cache(cachep
, alien
);
982 free_alien_cache(alien
);
986 slabs_destroy(cachep
, &list
);
989 * In the previous loop, all the objects were freed to
990 * the respective cache's slabs, now we can go ahead and
991 * shrink each nodelist to its limit.
993 list_for_each_entry(cachep
, &slab_caches
, list
) {
994 n
= get_node(cachep
, node
);
997 drain_freelist(cachep
, n
, INT_MAX
);
1001 static int cpuup_prepare(long cpu
)
1003 struct kmem_cache
*cachep
;
1004 int node
= cpu_to_mem(cpu
);
1008 * We need to do this right in the beginning since
1009 * alloc_arraycache's are going to use this list.
1010 * kmalloc_node allows us to add the slab to the right
1011 * kmem_cache_node and not this cpu's kmem_cache_node
1013 err
= init_cache_node_node(node
);
1018 * Now we can go ahead with allocating the shared arrays and
1021 list_for_each_entry(cachep
, &slab_caches
, list
) {
1022 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1029 cpuup_canceled(cpu
);
1033 int slab_prepare_cpu(unsigned int cpu
)
1037 mutex_lock(&slab_mutex
);
1038 err
= cpuup_prepare(cpu
);
1039 mutex_unlock(&slab_mutex
);
1044 * This is called for a failed online attempt and for a successful
1047 * Even if all the cpus of a node are down, we don't free the
1048 * kmem_cache_node of any cache. This is to avoid a race between cpu_down, and
1049 * a kmalloc allocation from another cpu for memory from the node of
1050 * the cpu going down. The kmem_cache_node structure is usually allocated from
1051 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1053 int slab_dead_cpu(unsigned int cpu
)
1055 mutex_lock(&slab_mutex
);
1056 cpuup_canceled(cpu
);
1057 mutex_unlock(&slab_mutex
);
1062 static int slab_online_cpu(unsigned int cpu
)
1064 start_cpu_timer(cpu
);
1068 static int slab_offline_cpu(unsigned int cpu
)
1071 * Shutdown cache reaper. Note that the slab_mutex is held so
1072 * that if cache_reap() is invoked it cannot do anything
1073 * expensive but will only modify reap_work and reschedule the
1076 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1077 /* Now the cache_reaper is guaranteed to be not running. */
1078 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1082 #if defined(CONFIG_NUMA)
1084 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1085 * Returns -EBUSY if all objects cannot be drained so that the node is not
1088 * Must hold slab_mutex.
1090 static int __meminit
drain_cache_node_node(int node
)
1092 struct kmem_cache
*cachep
;
1095 list_for_each_entry(cachep
, &slab_caches
, list
) {
1096 struct kmem_cache_node
*n
;
1098 n
= get_node(cachep
, node
);
1102 drain_freelist(cachep
, n
, INT_MAX
);
1104 if (!list_empty(&n
->slabs_full
) ||
1105 !list_empty(&n
->slabs_partial
)) {
1113 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1114 unsigned long action
, void *arg
)
1116 struct memory_notify
*mnb
= arg
;
1120 nid
= mnb
->status_change_nid
;
1125 case MEM_GOING_ONLINE
:
1126 mutex_lock(&slab_mutex
);
1127 ret
= init_cache_node_node(nid
);
1128 mutex_unlock(&slab_mutex
);
1130 case MEM_GOING_OFFLINE
:
1131 mutex_lock(&slab_mutex
);
1132 ret
= drain_cache_node_node(nid
);
1133 mutex_unlock(&slab_mutex
);
1137 case MEM_CANCEL_ONLINE
:
1138 case MEM_CANCEL_OFFLINE
:
1142 return notifier_from_errno(ret
);
1144 #endif /* CONFIG_NUMA */
1147 * swap the static kmem_cache_node with kmalloced memory
1149 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1152 struct kmem_cache_node
*ptr
;
1154 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1157 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1159 * Do not assume that spinlocks can be initialized via memcpy:
1161 raw_spin_lock_init(&ptr
->list_lock
);
1163 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1164 cachep
->node
[nodeid
] = ptr
;
1168 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1169 * size of kmem_cache_node.
1171 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1175 for_each_online_node(node
) {
1176 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1177 cachep
->node
[node
]->next_reap
= jiffies
+
1179 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1184 * Initialisation. Called after the page allocator have been initialised and
1185 * before smp_init().
1187 void __init
kmem_cache_init(void)
1191 kmem_cache
= &kmem_cache_boot
;
1193 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1194 use_alien_caches
= 0;
1196 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1197 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1200 * Fragmentation resistance on low memory - only use bigger
1201 * page orders on machines with more than 32MB of memory if
1202 * not overridden on the command line.
1204 if (!slab_max_order_set
&& totalram_pages() > (32 << 20) >> PAGE_SHIFT
)
1205 slab_max_order
= SLAB_MAX_ORDER_HI
;
1207 /* Bootstrap is tricky, because several objects are allocated
1208 * from caches that do not exist yet:
1209 * 1) initialize the kmem_cache cache: it contains the struct
1210 * kmem_cache structures of all caches, except kmem_cache itself:
1211 * kmem_cache is statically allocated.
1212 * Initially an __init data area is used for the head array and the
1213 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1214 * array at the end of the bootstrap.
1215 * 2) Create the first kmalloc cache.
1216 * The struct kmem_cache for the new cache is allocated normally.
1217 * An __init data area is used for the head array.
1218 * 3) Create the remaining kmalloc caches, with minimally sized
1220 * 4) Replace the __init data head arrays for kmem_cache and the first
1221 * kmalloc cache with kmalloc allocated arrays.
1222 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1223 * the other cache's with kmalloc allocated memory.
1224 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1227 /* 1) create the kmem_cache */
1230 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1232 create_boot_cache(kmem_cache
, "kmem_cache",
1233 offsetof(struct kmem_cache
, node
) +
1234 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1235 SLAB_HWCACHE_ALIGN
, 0, 0);
1236 list_add(&kmem_cache
->list
, &slab_caches
);
1237 slab_state
= PARTIAL
;
1240 * Initialize the caches that provide memory for the kmem_cache_node
1241 * structures first. Without this, further allocations will bug.
1243 new_kmalloc_cache(INDEX_NODE
, KMALLOC_NORMAL
, ARCH_KMALLOC_FLAGS
);
1244 slab_state
= PARTIAL_NODE
;
1245 setup_kmalloc_cache_index_table();
1247 /* 5) Replace the bootstrap kmem_cache_node */
1251 for_each_online_node(nid
) {
1252 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1254 init_list(kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
],
1255 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1259 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1262 void __init
kmem_cache_init_late(void)
1264 struct kmem_cache
*cachep
;
1266 /* 6) resize the head arrays to their final sizes */
1267 mutex_lock(&slab_mutex
);
1268 list_for_each_entry(cachep
, &slab_caches
, list
)
1269 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1271 mutex_unlock(&slab_mutex
);
1278 * Register a memory hotplug callback that initializes and frees
1281 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1285 * The reap timers are started later, with a module init call: That part
1286 * of the kernel is not yet operational.
1290 static int __init
cpucache_init(void)
1295 * Register the timers that return unneeded pages to the page allocator
1297 ret
= cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "SLAB online",
1298 slab_online_cpu
, slab_offline_cpu
);
1303 __initcall(cpucache_init
);
1305 static noinline
void
1306 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1309 struct kmem_cache_node
*n
;
1310 unsigned long flags
;
1312 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1313 DEFAULT_RATELIMIT_BURST
);
1315 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1318 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1319 nodeid
, gfpflags
, &gfpflags
);
1320 pr_warn(" cache: %s, object size: %d, order: %d\n",
1321 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1323 for_each_kmem_cache_node(cachep
, node
, n
) {
1324 unsigned long total_slabs
, free_slabs
, free_objs
;
1326 raw_spin_lock_irqsave(&n
->list_lock
, flags
);
1327 total_slabs
= n
->total_slabs
;
1328 free_slabs
= n
->free_slabs
;
1329 free_objs
= n
->free_objects
;
1330 raw_spin_unlock_irqrestore(&n
->list_lock
, flags
);
1332 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1333 node
, total_slabs
- free_slabs
, total_slabs
,
1334 (total_slabs
* cachep
->num
) - free_objs
,
1335 total_slabs
* cachep
->num
);
1341 * Interface to system's page allocator. No need to hold the
1342 * kmem_cache_node ->list_lock.
1344 * If we requested dmaable memory, we will get it. Even if we
1345 * did not request dmaable memory, we might get it, but that
1346 * would be relatively rare and ignorable.
1348 static struct slab
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1351 struct folio
*folio
;
1354 flags
|= cachep
->allocflags
;
1356 folio
= (struct folio
*) __alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1358 slab_out_of_memory(cachep
, flags
, nodeid
);
1362 slab
= folio_slab(folio
);
1364 account_slab(slab
, cachep
->gfporder
, cachep
, flags
);
1365 __folio_set_slab(folio
);
1366 /* Make the flag visible before any changes to folio->mapping */
1368 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1369 if (sk_memalloc_socks() && folio_is_pfmemalloc(folio
))
1370 slab_set_pfmemalloc(slab
);
1376 * Interface to system's page release.
1378 static void kmem_freepages(struct kmem_cache
*cachep
, struct slab
*slab
)
1380 int order
= cachep
->gfporder
;
1381 struct folio
*folio
= slab_folio(slab
);
1383 BUG_ON(!folio_test_slab(folio
));
1384 __slab_clear_pfmemalloc(slab
);
1385 page_mapcount_reset(&folio
->page
);
1386 folio
->mapping
= NULL
;
1387 /* Make the mapping reset visible before clearing the flag */
1389 __folio_clear_slab(folio
);
1391 mm_account_reclaimed_pages(1 << order
);
1392 unaccount_slab(slab
, order
, cachep
);
1393 __free_pages(&folio
->page
, order
);
1396 static void kmem_rcu_free(struct rcu_head
*head
)
1398 struct kmem_cache
*cachep
;
1401 slab
= container_of(head
, struct slab
, rcu_head
);
1402 cachep
= slab
->slab_cache
;
1404 kmem_freepages(cachep
, slab
);
1408 static inline bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1410 return debug_pagealloc_enabled_static() && OFF_SLAB(cachep
) &&
1411 ((cachep
->size
% PAGE_SIZE
) == 0);
1414 #ifdef CONFIG_DEBUG_PAGEALLOC
1415 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
, int map
)
1417 if (!is_debug_pagealloc_cache(cachep
))
1420 __kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1424 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1429 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1431 int size
= cachep
->object_size
;
1432 addr
= &((char *)addr
)[obj_offset(cachep
)];
1434 memset(addr
, val
, size
);
1435 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1438 static void dump_line(char *data
, int offset
, int limit
)
1441 unsigned char error
= 0;
1444 pr_err("%03x: ", offset
);
1445 for (i
= 0; i
< limit
; i
++) {
1446 if (data
[offset
+ i
] != POISON_FREE
) {
1447 error
= data
[offset
+ i
];
1451 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1452 &data
[offset
], limit
, 1);
1454 if (bad_count
== 1) {
1455 error
^= POISON_FREE
;
1456 if (!(error
& (error
- 1))) {
1457 pr_err("Single bit error detected. Probably bad RAM.\n");
1459 pr_err("Run memtest86+ or a similar memory test tool.\n");
1461 pr_err("Run a memory test tool.\n");
1470 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1475 if (cachep
->flags
& SLAB_RED_ZONE
) {
1476 pr_err("Redzone: 0x%llx/0x%llx\n",
1477 *dbg_redzone1(cachep
, objp
),
1478 *dbg_redzone2(cachep
, objp
));
1481 if (cachep
->flags
& SLAB_STORE_USER
)
1482 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep
, objp
));
1483 realobj
= (char *)objp
+ obj_offset(cachep
);
1484 size
= cachep
->object_size
;
1485 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1488 if (i
+ limit
> size
)
1490 dump_line(realobj
, i
, limit
);
1494 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1500 if (is_debug_pagealloc_cache(cachep
))
1503 realobj
= (char *)objp
+ obj_offset(cachep
);
1504 size
= cachep
->object_size
;
1506 for (i
= 0; i
< size
; i
++) {
1507 char exp
= POISON_FREE
;
1510 if (realobj
[i
] != exp
) {
1515 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1516 print_tainted(), cachep
->name
,
1518 print_objinfo(cachep
, objp
, 0);
1520 /* Hexdump the affected line */
1523 if (i
+ limit
> size
)
1525 dump_line(realobj
, i
, limit
);
1528 /* Limit to 5 lines */
1534 /* Print some data about the neighboring objects, if they
1537 struct slab
*slab
= virt_to_slab(objp
);
1540 objnr
= obj_to_index(cachep
, slab
, objp
);
1542 objp
= index_to_obj(cachep
, slab
, objnr
- 1);
1543 realobj
= (char *)objp
+ obj_offset(cachep
);
1544 pr_err("Prev obj: start=%px, len=%d\n", realobj
, size
);
1545 print_objinfo(cachep
, objp
, 2);
1547 if (objnr
+ 1 < cachep
->num
) {
1548 objp
= index_to_obj(cachep
, slab
, objnr
+ 1);
1549 realobj
= (char *)objp
+ obj_offset(cachep
);
1550 pr_err("Next obj: start=%px, len=%d\n", realobj
, size
);
1551 print_objinfo(cachep
, objp
, 2);
1558 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1563 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1564 poison_obj(cachep
, slab
->freelist
- obj_offset(cachep
),
1568 for (i
= 0; i
< cachep
->num
; i
++) {
1569 void *objp
= index_to_obj(cachep
, slab
, i
);
1571 if (cachep
->flags
& SLAB_POISON
) {
1572 check_poison_obj(cachep
, objp
);
1573 slab_kernel_map(cachep
, objp
, 1);
1575 if (cachep
->flags
& SLAB_RED_ZONE
) {
1576 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1577 slab_error(cachep
, "start of a freed object was overwritten");
1578 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1579 slab_error(cachep
, "end of a freed object was overwritten");
1584 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1591 * slab_destroy - destroy and release all objects in a slab
1592 * @cachep: cache pointer being destroyed
1593 * @slab: slab being destroyed
1595 * Destroy all the objs in a slab, and release the mem back to the system.
1596 * Before calling the slab must have been unlinked from the cache. The
1597 * kmem_cache_node ->list_lock is not held/needed.
1599 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slab
)
1603 freelist
= slab
->freelist
;
1604 slab_destroy_debugcheck(cachep
, slab
);
1605 if (unlikely(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
1606 call_rcu(&slab
->rcu_head
, kmem_rcu_free
);
1608 kmem_freepages(cachep
, slab
);
1611 * From now on, we don't use freelist
1612 * although actual page can be freed in rcu context
1614 if (OFF_SLAB(cachep
))
1619 * Update the size of the caches before calling slabs_destroy as it may
1620 * recursively call kfree.
1622 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1624 struct slab
*slab
, *n
;
1626 list_for_each_entry_safe(slab
, n
, list
, slab_list
) {
1627 list_del(&slab
->slab_list
);
1628 slab_destroy(cachep
, slab
);
1633 * calculate_slab_order - calculate size (page order) of slabs
1634 * @cachep: pointer to the cache that is being created
1635 * @size: size of objects to be created in this cache.
1636 * @flags: slab allocation flags
1638 * Also calculates the number of objects per slab.
1640 * This could be made much more intelligent. For now, try to avoid using
1641 * high order pages for slabs. When the gfp() functions are more friendly
1642 * towards high-order requests, this should be changed.
1644 * Return: number of left-over bytes in a slab
1646 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1647 size_t size
, slab_flags_t flags
)
1649 size_t left_over
= 0;
1652 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1656 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1660 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1661 if (num
> SLAB_OBJ_MAX_NUM
)
1664 if (flags
& CFLGS_OFF_SLAB
) {
1665 struct kmem_cache
*freelist_cache
;
1666 size_t freelist_size
;
1667 size_t freelist_cache_size
;
1669 freelist_size
= num
* sizeof(freelist_idx_t
);
1670 if (freelist_size
> KMALLOC_MAX_CACHE_SIZE
) {
1671 freelist_cache_size
= PAGE_SIZE
<< get_order(freelist_size
);
1673 freelist_cache
= kmalloc_slab(freelist_size
, 0u, _RET_IP_
);
1674 if (!freelist_cache
)
1676 freelist_cache_size
= freelist_cache
->size
;
1679 * Needed to avoid possible looping condition
1680 * in cache_grow_begin()
1682 if (OFF_SLAB(freelist_cache
))
1686 /* check if off slab has enough benefit */
1687 if (freelist_cache_size
> cachep
->size
/ 2)
1691 /* Found something acceptable - save it away */
1693 cachep
->gfporder
= gfporder
;
1694 left_over
= remainder
;
1697 * A VFS-reclaimable slab tends to have most allocations
1698 * as GFP_NOFS and we really don't want to have to be allocating
1699 * higher-order pages when we are unable to shrink dcache.
1701 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1705 * Large number of objects is good, but very large slabs are
1706 * currently bad for the gfp()s.
1708 if (gfporder
>= slab_max_order
)
1712 * Acceptable internal fragmentation?
1714 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1720 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1721 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1725 struct array_cache __percpu
*cpu_cache
;
1727 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1728 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1733 for_each_possible_cpu(cpu
) {
1734 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1735 entries
, batchcount
);
1741 static int __ref
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1743 if (slab_state
>= FULL
)
1744 return enable_cpucache(cachep
, gfp
);
1746 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1747 if (!cachep
->cpu_cache
)
1750 if (slab_state
== DOWN
) {
1751 /* Creation of first cache (kmem_cache). */
1752 set_up_node(kmem_cache
, CACHE_CACHE
);
1753 } else if (slab_state
== PARTIAL
) {
1754 /* For kmem_cache_node */
1755 set_up_node(cachep
, SIZE_NODE
);
1759 for_each_online_node(node
) {
1760 cachep
->node
[node
] = kmalloc_node(
1761 sizeof(struct kmem_cache_node
), gfp
, node
);
1762 BUG_ON(!cachep
->node
[node
]);
1763 kmem_cache_node_init(cachep
->node
[node
]);
1767 cachep
->node
[numa_mem_id()]->next_reap
=
1768 jiffies
+ REAPTIMEOUT_NODE
+
1769 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1771 cpu_cache_get(cachep
)->avail
= 0;
1772 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1773 cpu_cache_get(cachep
)->batchcount
= 1;
1774 cpu_cache_get(cachep
)->touched
= 0;
1775 cachep
->batchcount
= 1;
1776 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1780 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1781 slab_flags_t flags
, const char *name
)
1787 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
1788 slab_flags_t flags
, void (*ctor
)(void *))
1790 struct kmem_cache
*cachep
;
1792 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1797 * Adjust the object sizes so that we clear
1798 * the complete object on kzalloc.
1800 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1805 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1806 size_t size
, slab_flags_t flags
)
1813 * If slab auto-initialization on free is enabled, store the freelist
1814 * off-slab, so that its contents don't end up in one of the allocated
1817 if (unlikely(slab_want_init_on_free(cachep
)))
1820 if (cachep
->ctor
|| flags
& SLAB_TYPESAFE_BY_RCU
)
1823 left
= calculate_slab_order(cachep
, size
,
1824 flags
| CFLGS_OBJFREELIST_SLAB
);
1828 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1831 cachep
->colour
= left
/ cachep
->colour_off
;
1836 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1837 size_t size
, slab_flags_t flags
)
1844 * Always use on-slab management when SLAB_NOLEAKTRACE
1845 * to avoid recursive calls into kmemleak.
1847 if (flags
& SLAB_NOLEAKTRACE
)
1851 * Size is large, assume best to place the slab management obj
1852 * off-slab (should allow better packing of objs).
1854 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1859 * If the slab has been placed off-slab, and we have enough space then
1860 * move it on-slab. This is at the expense of any extra colouring.
1862 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1865 cachep
->colour
= left
/ cachep
->colour_off
;
1870 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1871 size_t size
, slab_flags_t flags
)
1877 left
= calculate_slab_order(cachep
, size
, flags
);
1881 cachep
->colour
= left
/ cachep
->colour_off
;
1887 * __kmem_cache_create - Create a cache.
1888 * @cachep: cache management descriptor
1889 * @flags: SLAB flags
1891 * Returns zero on success, nonzero on failure.
1895 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1896 * to catch references to uninitialised memory.
1898 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1899 * for buffer overruns.
1901 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1902 * cacheline. This can be beneficial if you're counting cycles as closely
1905 int __kmem_cache_create(struct kmem_cache
*cachep
, slab_flags_t flags
)
1907 size_t ralign
= BYTES_PER_WORD
;
1910 unsigned int size
= cachep
->size
;
1915 * Enable redzoning and last user accounting, except for caches with
1916 * large objects, if the increased size would increase the object size
1917 * above the next power of two: caches with object sizes just above a
1918 * power of two have a significant amount of internal fragmentation.
1920 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
1921 2 * sizeof(unsigned long long)))
1922 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1923 if (!(flags
& SLAB_TYPESAFE_BY_RCU
))
1924 flags
|= SLAB_POISON
;
1929 * Check that size is in terms of words. This is needed to avoid
1930 * unaligned accesses for some archs when redzoning is used, and makes
1931 * sure any on-slab bufctl's are also correctly aligned.
1933 size
= ALIGN(size
, BYTES_PER_WORD
);
1935 if (flags
& SLAB_RED_ZONE
) {
1936 ralign
= REDZONE_ALIGN
;
1937 /* If redzoning, ensure that the second redzone is suitably
1938 * aligned, by adjusting the object size accordingly. */
1939 size
= ALIGN(size
, REDZONE_ALIGN
);
1942 /* 3) caller mandated alignment */
1943 if (ralign
< cachep
->align
) {
1944 ralign
= cachep
->align
;
1946 /* disable debug if necessary */
1947 if (ralign
> __alignof__(unsigned long long))
1948 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1952 cachep
->align
= ralign
;
1953 cachep
->colour_off
= cache_line_size();
1954 /* Offset must be a multiple of the alignment. */
1955 if (cachep
->colour_off
< cachep
->align
)
1956 cachep
->colour_off
= cachep
->align
;
1958 if (slab_is_available())
1966 * Both debugging options require word-alignment which is calculated
1969 if (flags
& SLAB_RED_ZONE
) {
1970 /* add space for red zone words */
1971 cachep
->obj_offset
+= sizeof(unsigned long long);
1972 size
+= 2 * sizeof(unsigned long long);
1974 if (flags
& SLAB_STORE_USER
) {
1975 /* user store requires one word storage behind the end of
1976 * the real object. But if the second red zone needs to be
1977 * aligned to 64 bits, we must allow that much space.
1979 if (flags
& SLAB_RED_ZONE
)
1980 size
+= REDZONE_ALIGN
;
1982 size
+= BYTES_PER_WORD
;
1986 kasan_cache_create(cachep
, &size
, &flags
);
1988 size
= ALIGN(size
, cachep
->align
);
1990 * We should restrict the number of objects in a slab to implement
1991 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
1993 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
1994 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
1998 * To activate debug pagealloc, off-slab management is necessary
1999 * requirement. In early phase of initialization, small sized slab
2000 * doesn't get initialized so it would not be possible. So, we need
2001 * to check size >= 256. It guarantees that all necessary small
2002 * sized slab is initialized in current slab initialization sequence.
2004 if (debug_pagealloc_enabled_static() && (flags
& SLAB_POISON
) &&
2005 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2006 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2007 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2009 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2010 flags
|= CFLGS_OFF_SLAB
;
2011 cachep
->obj_offset
+= tmp_size
- size
;
2019 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2020 flags
|= CFLGS_OBJFREELIST_SLAB
;
2024 if (set_off_slab_cache(cachep
, size
, flags
)) {
2025 flags
|= CFLGS_OFF_SLAB
;
2029 if (set_on_slab_cache(cachep
, size
, flags
))
2035 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2036 cachep
->flags
= flags
;
2037 cachep
->allocflags
= __GFP_COMP
;
2038 if (flags
& SLAB_CACHE_DMA
)
2039 cachep
->allocflags
|= GFP_DMA
;
2040 if (flags
& SLAB_CACHE_DMA32
)
2041 cachep
->allocflags
|= GFP_DMA32
;
2042 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2043 cachep
->allocflags
|= __GFP_RECLAIMABLE
;
2044 cachep
->size
= size
;
2045 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2049 * If we're going to use the generic kernel_map_pages()
2050 * poisoning, then it's going to smash the contents of
2051 * the redzone and userword anyhow, so switch them off.
2053 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2054 (cachep
->flags
& SLAB_POISON
) &&
2055 is_debug_pagealloc_cache(cachep
))
2056 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2059 err
= setup_cpu_cache(cachep
, gfp
);
2061 __kmem_cache_release(cachep
);
2069 static void check_irq_off(void)
2071 BUG_ON(!irqs_disabled());
2074 static void check_irq_on(void)
2076 BUG_ON(irqs_disabled());
2079 static void check_mutex_acquired(void)
2081 BUG_ON(!mutex_is_locked(&slab_mutex
));
2084 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2088 assert_raw_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2092 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2096 assert_raw_spin_locked(&get_node(cachep
, node
)->list_lock
);
2101 #define check_irq_off() do { } while(0)
2102 #define check_irq_on() do { } while(0)
2103 #define check_mutex_acquired() do { } while(0)
2104 #define check_spinlock_acquired(x) do { } while(0)
2105 #define check_spinlock_acquired_node(x, y) do { } while(0)
2108 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2109 int node
, bool free_all
, struct list_head
*list
)
2113 if (!ac
|| !ac
->avail
)
2116 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2117 if (tofree
> ac
->avail
)
2118 tofree
= (ac
->avail
+ 1) / 2;
2120 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2121 ac
->avail
-= tofree
;
2122 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2125 static void do_drain(void *arg
)
2127 struct kmem_cache
*cachep
= arg
;
2128 struct array_cache
*ac
;
2129 int node
= numa_mem_id();
2130 struct kmem_cache_node
*n
;
2134 ac
= cpu_cache_get(cachep
);
2135 n
= get_node(cachep
, node
);
2136 raw_spin_lock(&n
->list_lock
);
2137 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2138 raw_spin_unlock(&n
->list_lock
);
2140 slabs_destroy(cachep
, &list
);
2143 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2145 struct kmem_cache_node
*n
;
2149 on_each_cpu(do_drain
, cachep
, 1);
2151 for_each_kmem_cache_node(cachep
, node
, n
)
2153 drain_alien_cache(cachep
, n
->alien
);
2155 for_each_kmem_cache_node(cachep
, node
, n
) {
2156 raw_spin_lock_irq(&n
->list_lock
);
2157 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2158 raw_spin_unlock_irq(&n
->list_lock
);
2160 slabs_destroy(cachep
, &list
);
2165 * Remove slabs from the list of free slabs.
2166 * Specify the number of slabs to drain in tofree.
2168 * Returns the actual number of slabs released.
2170 static int drain_freelist(struct kmem_cache
*cache
,
2171 struct kmem_cache_node
*n
, int tofree
)
2173 struct list_head
*p
;
2178 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2180 raw_spin_lock_irq(&n
->list_lock
);
2181 p
= n
->slabs_free
.prev
;
2182 if (p
== &n
->slabs_free
) {
2183 raw_spin_unlock_irq(&n
->list_lock
);
2187 slab
= list_entry(p
, struct slab
, slab_list
);
2188 list_del(&slab
->slab_list
);
2192 * Safe to drop the lock. The slab is no longer linked
2195 n
->free_objects
-= cache
->num
;
2196 raw_spin_unlock_irq(&n
->list_lock
);
2197 slab_destroy(cache
, slab
);
2206 bool __kmem_cache_empty(struct kmem_cache
*s
)
2209 struct kmem_cache_node
*n
;
2211 for_each_kmem_cache_node(s
, node
, n
)
2212 if (!list_empty(&n
->slabs_full
) ||
2213 !list_empty(&n
->slabs_partial
))
2218 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2222 struct kmem_cache_node
*n
;
2224 drain_cpu_caches(cachep
);
2227 for_each_kmem_cache_node(cachep
, node
, n
) {
2228 drain_freelist(cachep
, n
, INT_MAX
);
2230 ret
+= !list_empty(&n
->slabs_full
) ||
2231 !list_empty(&n
->slabs_partial
);
2233 return (ret
? 1 : 0);
2236 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2238 return __kmem_cache_shrink(cachep
);
2241 void __kmem_cache_release(struct kmem_cache
*cachep
)
2244 struct kmem_cache_node
*n
;
2246 cache_random_seq_destroy(cachep
);
2248 free_percpu(cachep
->cpu_cache
);
2250 /* NUMA: free the node structures */
2251 for_each_kmem_cache_node(cachep
, i
, n
) {
2253 free_alien_cache(n
->alien
);
2255 cachep
->node
[i
] = NULL
;
2260 * Get the memory for a slab management obj.
2262 * For a slab cache when the slab descriptor is off-slab, the
2263 * slab descriptor can't come from the same cache which is being created,
2264 * Because if it is the case, that means we defer the creation of
2265 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2266 * And we eventually call down to __kmem_cache_create(), which
2267 * in turn looks up in the kmalloc_{dma,}_caches for the desired-size one.
2268 * This is a "chicken-and-egg" problem.
2270 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2271 * which are all initialized during kmem_cache_init().
2273 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2274 struct slab
*slab
, int colour_off
,
2275 gfp_t local_flags
, int nodeid
)
2278 void *addr
= slab_address(slab
);
2280 slab
->s_mem
= addr
+ colour_off
;
2283 if (OBJFREELIST_SLAB(cachep
))
2285 else if (OFF_SLAB(cachep
)) {
2286 /* Slab management obj is off-slab. */
2287 freelist
= kmalloc_node(cachep
->freelist_size
,
2288 local_flags
, nodeid
);
2290 /* We will use last bytes at the slab for freelist */
2291 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2292 cachep
->freelist_size
;
2298 static inline freelist_idx_t
get_free_obj(struct slab
*slab
, unsigned int idx
)
2300 return ((freelist_idx_t
*) slab
->freelist
)[idx
];
2303 static inline void set_free_obj(struct slab
*slab
,
2304 unsigned int idx
, freelist_idx_t val
)
2306 ((freelist_idx_t
*)(slab
->freelist
))[idx
] = val
;
2309 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct slab
*slab
)
2314 for (i
= 0; i
< cachep
->num
; i
++) {
2315 void *objp
= index_to_obj(cachep
, slab
, i
);
2317 if (cachep
->flags
& SLAB_STORE_USER
)
2318 *dbg_userword(cachep
, objp
) = NULL
;
2320 if (cachep
->flags
& SLAB_RED_ZONE
) {
2321 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2322 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2325 * Constructors are not allowed to allocate memory from the same
2326 * cache which they are a constructor for. Otherwise, deadlock.
2327 * They must also be threaded.
2329 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2330 kasan_unpoison_object_data(cachep
,
2331 objp
+ obj_offset(cachep
));
2332 cachep
->ctor(objp
+ obj_offset(cachep
));
2333 kasan_poison_object_data(
2334 cachep
, objp
+ obj_offset(cachep
));
2337 if (cachep
->flags
& SLAB_RED_ZONE
) {
2338 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2339 slab_error(cachep
, "constructor overwrote the end of an object");
2340 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2341 slab_error(cachep
, "constructor overwrote the start of an object");
2343 /* need to poison the objs? */
2344 if (cachep
->flags
& SLAB_POISON
) {
2345 poison_obj(cachep
, objp
, POISON_FREE
);
2346 slab_kernel_map(cachep
, objp
, 0);
2352 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2353 /* Hold information during a freelist initialization */
2354 struct freelist_init_state
{
2361 * Initialize the state based on the randomization method available.
2362 * return true if the pre-computed list is available, false otherwise.
2364 static bool freelist_state_initialize(struct freelist_init_state
*state
,
2365 struct kmem_cache
*cachep
,
2369 if (!cachep
->random_seq
) {
2372 state
->list
= cachep
->random_seq
;
2373 state
->count
= count
;
2374 state
->pos
= get_random_u32_below(count
);
2380 /* Get the next entry on the list and randomize it using a random shift */
2381 static freelist_idx_t
next_random_slot(struct freelist_init_state
*state
)
2383 if (state
->pos
>= state
->count
)
2385 return state
->list
[state
->pos
++];
2388 /* Swap two freelist entries */
2389 static void swap_free_obj(struct slab
*slab
, unsigned int a
, unsigned int b
)
2391 swap(((freelist_idx_t
*) slab
->freelist
)[a
],
2392 ((freelist_idx_t
*) slab
->freelist
)[b
]);
2396 * Shuffle the freelist initialization state based on pre-computed lists.
2397 * return true if the list was successfully shuffled, false otherwise.
2399 static bool shuffle_freelist(struct kmem_cache
*cachep
, struct slab
*slab
)
2401 unsigned int objfreelist
= 0, i
, rand
, count
= cachep
->num
;
2402 struct freelist_init_state state
;
2408 precomputed
= freelist_state_initialize(&state
, cachep
, count
);
2410 /* Take a random entry as the objfreelist */
2411 if (OBJFREELIST_SLAB(cachep
)) {
2413 objfreelist
= count
- 1;
2415 objfreelist
= next_random_slot(&state
);
2416 slab
->freelist
= index_to_obj(cachep
, slab
, objfreelist
) +
2422 * On early boot, generate the list dynamically.
2423 * Later use a pre-computed list for speed.
2426 for (i
= 0; i
< count
; i
++)
2427 set_free_obj(slab
, i
, i
);
2429 /* Fisher-Yates shuffle */
2430 for (i
= count
- 1; i
> 0; i
--) {
2431 rand
= get_random_u32_below(i
+ 1);
2432 swap_free_obj(slab
, i
, rand
);
2435 for (i
= 0; i
< count
; i
++)
2436 set_free_obj(slab
, i
, next_random_slot(&state
));
2439 if (OBJFREELIST_SLAB(cachep
))
2440 set_free_obj(slab
, cachep
->num
- 1, objfreelist
);
2445 static inline bool shuffle_freelist(struct kmem_cache
*cachep
,
2450 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2452 static void cache_init_objs(struct kmem_cache
*cachep
,
2459 cache_init_objs_debug(cachep
, slab
);
2461 /* Try to randomize the freelist if enabled */
2462 shuffled
= shuffle_freelist(cachep
, slab
);
2464 if (!shuffled
&& OBJFREELIST_SLAB(cachep
)) {
2465 slab
->freelist
= index_to_obj(cachep
, slab
, cachep
->num
- 1) +
2469 for (i
= 0; i
< cachep
->num
; i
++) {
2470 objp
= index_to_obj(cachep
, slab
, i
);
2471 objp
= kasan_init_slab_obj(cachep
, objp
);
2473 /* constructor could break poison info */
2474 if (DEBUG
== 0 && cachep
->ctor
) {
2475 kasan_unpoison_object_data(cachep
, objp
);
2477 kasan_poison_object_data(cachep
, objp
);
2481 set_free_obj(slab
, i
, i
);
2485 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slab
)
2489 objp
= index_to_obj(cachep
, slab
, get_free_obj(slab
, slab
->active
));
2495 static void slab_put_obj(struct kmem_cache
*cachep
,
2496 struct slab
*slab
, void *objp
)
2498 unsigned int objnr
= obj_to_index(cachep
, slab
, objp
);
2502 /* Verify double free bug */
2503 for (i
= slab
->active
; i
< cachep
->num
; i
++) {
2504 if (get_free_obj(slab
, i
) == objnr
) {
2505 pr_err("slab: double free detected in cache '%s', objp %px\n",
2506 cachep
->name
, objp
);
2512 if (!slab
->freelist
)
2513 slab
->freelist
= objp
+ obj_offset(cachep
);
2515 set_free_obj(slab
, slab
->active
, objnr
);
2519 * Grow (by 1) the number of slabs within a cache. This is called by
2520 * kmem_cache_alloc() when there are no active objs left in a cache.
2522 static struct slab
*cache_grow_begin(struct kmem_cache
*cachep
,
2523 gfp_t flags
, int nodeid
)
2529 struct kmem_cache_node
*n
;
2533 * Be lazy and only check for valid flags here, keeping it out of the
2534 * critical path in kmem_cache_alloc().
2536 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
2537 flags
= kmalloc_fix_flags(flags
);
2539 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2540 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2543 if (gfpflags_allow_blocking(local_flags
))
2547 * Get mem for the objs. Attempt to allocate a physical page from
2550 slab
= kmem_getpages(cachep
, local_flags
, nodeid
);
2554 slab_node
= slab_nid(slab
);
2555 n
= get_node(cachep
, slab_node
);
2557 /* Get colour for the slab, and cal the next value. */
2559 if (n
->colour_next
>= cachep
->colour
)
2562 offset
= n
->colour_next
;
2563 if (offset
>= cachep
->colour
)
2566 offset
*= cachep
->colour_off
;
2569 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2570 * page_address() in the latter returns a non-tagged pointer,
2571 * as it should be for slab pages.
2573 kasan_poison_slab(slab
);
2575 /* Get slab management. */
2576 freelist
= alloc_slabmgmt(cachep
, slab
, offset
,
2577 local_flags
& ~GFP_CONSTRAINT_MASK
, slab_node
);
2578 if (OFF_SLAB(cachep
) && !freelist
)
2581 slab
->slab_cache
= cachep
;
2582 slab
->freelist
= freelist
;
2584 cache_init_objs(cachep
, slab
);
2586 if (gfpflags_allow_blocking(local_flags
))
2587 local_irq_disable();
2592 kmem_freepages(cachep
, slab
);
2594 if (gfpflags_allow_blocking(local_flags
))
2595 local_irq_disable();
2599 static void cache_grow_end(struct kmem_cache
*cachep
, struct slab
*slab
)
2601 struct kmem_cache_node
*n
;
2609 INIT_LIST_HEAD(&slab
->slab_list
);
2610 n
= get_node(cachep
, slab_nid(slab
));
2612 raw_spin_lock(&n
->list_lock
);
2614 if (!slab
->active
) {
2615 list_add_tail(&slab
->slab_list
, &n
->slabs_free
);
2618 fixup_slab_list(cachep
, n
, slab
, &list
);
2620 STATS_INC_GROWN(cachep
);
2621 n
->free_objects
+= cachep
->num
- slab
->active
;
2622 raw_spin_unlock(&n
->list_lock
);
2624 fixup_objfreelist_debug(cachep
, &list
);
2630 * Perform extra freeing checks:
2631 * - detect bad pointers.
2632 * - POISON/RED_ZONE checking
2634 static void kfree_debugcheck(const void *objp
)
2636 if (!virt_addr_valid(objp
)) {
2637 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2638 (unsigned long)objp
);
2643 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2645 unsigned long long redzone1
, redzone2
;
2647 redzone1
= *dbg_redzone1(cache
, obj
);
2648 redzone2
= *dbg_redzone2(cache
, obj
);
2653 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2656 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2657 slab_error(cache
, "double free detected");
2659 slab_error(cache
, "memory outside object was overwritten");
2661 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2662 obj
, redzone1
, redzone2
);
2665 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2666 unsigned long caller
)
2671 BUG_ON(virt_to_cache(objp
) != cachep
);
2673 objp
-= obj_offset(cachep
);
2674 kfree_debugcheck(objp
);
2675 slab
= virt_to_slab(objp
);
2677 if (cachep
->flags
& SLAB_RED_ZONE
) {
2678 verify_redzone_free(cachep
, objp
);
2679 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2680 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2682 if (cachep
->flags
& SLAB_STORE_USER
)
2683 *dbg_userword(cachep
, objp
) = (void *)caller
;
2685 objnr
= obj_to_index(cachep
, slab
, objp
);
2687 BUG_ON(objnr
>= cachep
->num
);
2688 BUG_ON(objp
!= index_to_obj(cachep
, slab
, objnr
));
2690 if (cachep
->flags
& SLAB_POISON
) {
2691 poison_obj(cachep
, objp
, POISON_FREE
);
2692 slab_kernel_map(cachep
, objp
, 0);
2698 #define kfree_debugcheck(x) do { } while(0)
2699 #define cache_free_debugcheck(x, objp, z) (objp)
2702 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2710 objp
= next
- obj_offset(cachep
);
2711 next
= *(void **)next
;
2712 poison_obj(cachep
, objp
, POISON_FREE
);
2717 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2718 struct kmem_cache_node
*n
, struct slab
*slab
,
2721 /* move slabp to correct slabp list: */
2722 list_del(&slab
->slab_list
);
2723 if (slab
->active
== cachep
->num
) {
2724 list_add(&slab
->slab_list
, &n
->slabs_full
);
2725 if (OBJFREELIST_SLAB(cachep
)) {
2727 /* Poisoning will be done without holding the lock */
2728 if (cachep
->flags
& SLAB_POISON
) {
2729 void **objp
= slab
->freelist
;
2735 slab
->freelist
= NULL
;
2738 list_add(&slab
->slab_list
, &n
->slabs_partial
);
2741 /* Try to find non-pfmemalloc slab if needed */
2742 static noinline
struct slab
*get_valid_first_slab(struct kmem_cache_node
*n
,
2743 struct slab
*slab
, bool pfmemalloc
)
2751 if (!slab_test_pfmemalloc(slab
))
2754 /* No need to keep pfmemalloc slab if we have enough free objects */
2755 if (n
->free_objects
> n
->free_limit
) {
2756 slab_clear_pfmemalloc(slab
);
2760 /* Move pfmemalloc slab to the end of list to speed up next search */
2761 list_del(&slab
->slab_list
);
2762 if (!slab
->active
) {
2763 list_add_tail(&slab
->slab_list
, &n
->slabs_free
);
2766 list_add_tail(&slab
->slab_list
, &n
->slabs_partial
);
2768 list_for_each_entry(slab
, &n
->slabs_partial
, slab_list
) {
2769 if (!slab_test_pfmemalloc(slab
))
2773 n
->free_touched
= 1;
2774 list_for_each_entry(slab
, &n
->slabs_free
, slab_list
) {
2775 if (!slab_test_pfmemalloc(slab
)) {
2784 static struct slab
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2788 assert_raw_spin_locked(&n
->list_lock
);
2789 slab
= list_first_entry_or_null(&n
->slabs_partial
, struct slab
,
2792 n
->free_touched
= 1;
2793 slab
= list_first_entry_or_null(&n
->slabs_free
, struct slab
,
2799 if (sk_memalloc_socks())
2800 slab
= get_valid_first_slab(n
, slab
, pfmemalloc
);
2805 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2806 struct kmem_cache_node
*n
, gfp_t flags
)
2812 if (!gfp_pfmemalloc_allowed(flags
))
2815 raw_spin_lock(&n
->list_lock
);
2816 slab
= get_first_slab(n
, true);
2818 raw_spin_unlock(&n
->list_lock
);
2822 obj
= slab_get_obj(cachep
, slab
);
2825 fixup_slab_list(cachep
, n
, slab
, &list
);
2827 raw_spin_unlock(&n
->list_lock
);
2828 fixup_objfreelist_debug(cachep
, &list
);
2834 * Slab list should be fixed up by fixup_slab_list() for existing slab
2835 * or cache_grow_end() for new slab
2837 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2838 struct array_cache
*ac
, struct slab
*slab
, int batchcount
)
2841 * There must be at least one object available for
2844 BUG_ON(slab
->active
>= cachep
->num
);
2846 while (slab
->active
< cachep
->num
&& batchcount
--) {
2847 STATS_INC_ALLOCED(cachep
);
2848 STATS_INC_ACTIVE(cachep
);
2849 STATS_SET_HIGH(cachep
);
2851 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slab
);
2857 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2860 struct kmem_cache_node
*n
;
2861 struct array_cache
*ac
, *shared
;
2867 node
= numa_mem_id();
2869 ac
= cpu_cache_get(cachep
);
2870 batchcount
= ac
->batchcount
;
2871 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2873 * If there was little recent activity on this cache, then
2874 * perform only a partial refill. Otherwise we could generate
2877 batchcount
= BATCHREFILL_LIMIT
;
2879 n
= get_node(cachep
, node
);
2881 BUG_ON(ac
->avail
> 0 || !n
);
2882 shared
= READ_ONCE(n
->shared
);
2883 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
2886 raw_spin_lock(&n
->list_lock
);
2887 shared
= READ_ONCE(n
->shared
);
2889 /* See if we can refill from the shared array */
2890 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
2891 shared
->touched
= 1;
2895 while (batchcount
> 0) {
2896 /* Get slab alloc is to come from. */
2897 slab
= get_first_slab(n
, false);
2901 check_spinlock_acquired(cachep
);
2903 batchcount
= alloc_block(cachep
, ac
, slab
, batchcount
);
2904 fixup_slab_list(cachep
, n
, slab
, &list
);
2908 n
->free_objects
-= ac
->avail
;
2910 raw_spin_unlock(&n
->list_lock
);
2911 fixup_objfreelist_debug(cachep
, &list
);
2914 if (unlikely(!ac
->avail
)) {
2915 /* Check if we can use obj in pfmemalloc slab */
2916 if (sk_memalloc_socks()) {
2917 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2923 slab
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
2926 * cache_grow_begin() can reenable interrupts,
2927 * then ac could change.
2929 ac
= cpu_cache_get(cachep
);
2930 if (!ac
->avail
&& slab
)
2931 alloc_block(cachep
, ac
, slab
, batchcount
);
2932 cache_grow_end(cachep
, slab
);
2939 return ac
->entry
[--ac
->avail
];
2943 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2944 gfp_t flags
, void *objp
, unsigned long caller
)
2946 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2947 if (!objp
|| is_kfence_address(objp
))
2949 if (cachep
->flags
& SLAB_POISON
) {
2950 check_poison_obj(cachep
, objp
);
2951 slab_kernel_map(cachep
, objp
, 1);
2952 poison_obj(cachep
, objp
, POISON_INUSE
);
2954 if (cachep
->flags
& SLAB_STORE_USER
)
2955 *dbg_userword(cachep
, objp
) = (void *)caller
;
2957 if (cachep
->flags
& SLAB_RED_ZONE
) {
2958 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2959 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2960 slab_error(cachep
, "double free, or memory outside object was overwritten");
2961 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2962 objp
, *dbg_redzone1(cachep
, objp
),
2963 *dbg_redzone2(cachep
, objp
));
2965 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2966 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2969 objp
+= obj_offset(cachep
);
2970 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2972 if ((unsigned long)objp
& (arch_slab_minalign() - 1)) {
2973 pr_err("0x%px: not aligned to arch_slab_minalign()=%u\n", objp
,
2974 arch_slab_minalign());
2979 #define cache_alloc_debugcheck_after(a, b, objp, d) (objp)
2982 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
2985 struct array_cache
*ac
;
2989 ac
= cpu_cache_get(cachep
);
2990 if (likely(ac
->avail
)) {
2992 objp
= ac
->entry
[--ac
->avail
];
2994 STATS_INC_ALLOCHIT(cachep
);
2998 STATS_INC_ALLOCMISS(cachep
);
2999 objp
= cache_alloc_refill(cachep
, flags
);
3001 * the 'ac' may be updated by cache_alloc_refill(),
3002 * and kmemleak_erase() requires its correct value.
3004 ac
= cpu_cache_get(cachep
);
3008 * To avoid a false negative, if an object that is in one of the
3009 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3010 * treat the array pointers as a reference to the object.
3013 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3018 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
3021 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3023 * If we are in_interrupt, then process context, including cpusets and
3024 * mempolicy, may not apply and should not be used for allocation policy.
3026 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3028 int nid_alloc
, nid_here
;
3030 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3032 nid_alloc
= nid_here
= numa_mem_id();
3033 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3034 nid_alloc
= cpuset_slab_spread_node();
3035 else if (current
->mempolicy
)
3036 nid_alloc
= mempolicy_slab_node();
3037 if (nid_alloc
!= nid_here
)
3038 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3043 * Fallback function if there was no memory available and no objects on a
3044 * certain node and fall back is permitted. First we scan all the
3045 * available node for available objects. If that fails then we
3046 * perform an allocation without specifying a node. This allows the page
3047 * allocator to do its reclaim / fallback magic. We then insert the
3048 * slab into the proper nodelist and then allocate from it.
3050 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3052 struct zonelist
*zonelist
;
3055 enum zone_type highest_zoneidx
= gfp_zone(flags
);
3059 unsigned int cpuset_mems_cookie
;
3061 if (flags
& __GFP_THISNODE
)
3065 cpuset_mems_cookie
= read_mems_allowed_begin();
3066 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3070 * Look through allowed nodes for objects available
3071 * from existing per node queues.
3073 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
3074 nid
= zone_to_nid(zone
);
3076 if (cpuset_zone_allowed(zone
, flags
) &&
3077 get_node(cache
, nid
) &&
3078 get_node(cache
, nid
)->free_objects
) {
3079 obj
= ____cache_alloc_node(cache
,
3080 gfp_exact_node(flags
), nid
);
3088 * This allocation will be performed within the constraints
3089 * of the current cpuset / memory policy requirements.
3090 * We may trigger various forms of reclaim on the allowed
3091 * set and go into memory reserves if necessary.
3093 slab
= cache_grow_begin(cache
, flags
, numa_mem_id());
3094 cache_grow_end(cache
, slab
);
3096 nid
= slab_nid(slab
);
3097 obj
= ____cache_alloc_node(cache
,
3098 gfp_exact_node(flags
), nid
);
3101 * Another processor may allocate the objects in
3102 * the slab since we are not holding any locks.
3109 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3115 * An interface to enable slab creation on nodeid
3117 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3121 struct kmem_cache_node
*n
;
3125 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3126 n
= get_node(cachep
, nodeid
);
3130 raw_spin_lock(&n
->list_lock
);
3131 slab
= get_first_slab(n
, false);
3135 check_spinlock_acquired_node(cachep
, nodeid
);
3137 STATS_INC_NODEALLOCS(cachep
);
3138 STATS_INC_ACTIVE(cachep
);
3139 STATS_SET_HIGH(cachep
);
3141 BUG_ON(slab
->active
== cachep
->num
);
3143 obj
= slab_get_obj(cachep
, slab
);
3146 fixup_slab_list(cachep
, n
, slab
, &list
);
3148 raw_spin_unlock(&n
->list_lock
);
3149 fixup_objfreelist_debug(cachep
, &list
);
3153 raw_spin_unlock(&n
->list_lock
);
3154 slab
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3156 /* This slab isn't counted yet so don't update free_objects */
3157 obj
= slab_get_obj(cachep
, slab
);
3159 cache_grow_end(cachep
, slab
);
3161 return obj
? obj
: fallback_alloc(cachep
, flags
);
3164 static __always_inline
void *
3165 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3168 int slab_node
= numa_mem_id();
3170 if (nodeid
== NUMA_NO_NODE
) {
3171 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3172 objp
= alternate_node_alloc(cachep
, flags
);
3177 * Use the locally cached objects if possible.
3178 * However ____cache_alloc does not allow fallback
3179 * to other nodes. It may fail while we still have
3180 * objects on other nodes available.
3182 objp
= ____cache_alloc(cachep
, flags
);
3184 } else if (nodeid
== slab_node
) {
3185 objp
= ____cache_alloc(cachep
, flags
);
3186 } else if (!get_node(cachep
, nodeid
)) {
3187 /* Node not bootstrapped yet */
3188 objp
= fallback_alloc(cachep
, flags
);
3193 * We may just have run out of memory on the local node.
3194 * ____cache_alloc_node() knows how to locate memory on other nodes
3197 objp
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3203 static __always_inline
void *
3204 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid __maybe_unused
)
3206 return ____cache_alloc(cachep
, flags
);
3209 #endif /* CONFIG_NUMA */
3211 static __always_inline
void *
3212 slab_alloc_node(struct kmem_cache
*cachep
, struct list_lru
*lru
, gfp_t flags
,
3213 int nodeid
, size_t orig_size
, unsigned long caller
)
3215 unsigned long save_flags
;
3217 struct obj_cgroup
*objcg
= NULL
;
3220 flags
&= gfp_allowed_mask
;
3221 cachep
= slab_pre_alloc_hook(cachep
, lru
, &objcg
, 1, flags
);
3222 if (unlikely(!cachep
))
3225 objp
= kfence_alloc(cachep
, orig_size
, flags
);
3229 local_irq_save(save_flags
);
3230 objp
= __do_cache_alloc(cachep
, flags
, nodeid
);
3231 local_irq_restore(save_flags
);
3232 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3234 init
= slab_want_init_on_alloc(flags
, cachep
);
3237 slab_post_alloc_hook(cachep
, objcg
, flags
, 1, &objp
, init
,
3238 cachep
->object_size
);
3242 static __always_inline
void *
3243 slab_alloc(struct kmem_cache
*cachep
, struct list_lru
*lru
, gfp_t flags
,
3244 size_t orig_size
, unsigned long caller
)
3246 return slab_alloc_node(cachep
, lru
, flags
, NUMA_NO_NODE
, orig_size
,
3251 * Caller needs to acquire correct kmem_cache_node's list_lock
3252 * @list: List of detached free slabs should be freed by caller
3254 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3255 int nr_objects
, int node
, struct list_head
*list
)
3258 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3261 n
->free_objects
+= nr_objects
;
3263 for (i
= 0; i
< nr_objects
; i
++) {
3269 slab
= virt_to_slab(objp
);
3270 list_del(&slab
->slab_list
);
3271 check_spinlock_acquired_node(cachep
, node
);
3272 slab_put_obj(cachep
, slab
, objp
);
3273 STATS_DEC_ACTIVE(cachep
);
3275 /* fixup slab chains */
3276 if (slab
->active
== 0) {
3277 list_add(&slab
->slab_list
, &n
->slabs_free
);
3280 /* Unconditionally move a slab to the end of the
3281 * partial list on free - maximum time for the
3282 * other objects to be freed, too.
3284 list_add_tail(&slab
->slab_list
, &n
->slabs_partial
);
3288 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3289 n
->free_objects
-= cachep
->num
;
3291 slab
= list_last_entry(&n
->slabs_free
, struct slab
, slab_list
);
3292 list_move(&slab
->slab_list
, list
);
3298 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3301 struct kmem_cache_node
*n
;
3302 int node
= numa_mem_id();
3305 batchcount
= ac
->batchcount
;
3308 n
= get_node(cachep
, node
);
3309 raw_spin_lock(&n
->list_lock
);
3311 struct array_cache
*shared_array
= n
->shared
;
3312 int max
= shared_array
->limit
- shared_array
->avail
;
3314 if (batchcount
> max
)
3316 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3317 ac
->entry
, sizeof(void *) * batchcount
);
3318 shared_array
->avail
+= batchcount
;
3323 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3330 list_for_each_entry(slab
, &n
->slabs_free
, slab_list
) {
3331 BUG_ON(slab
->active
);
3335 STATS_SET_FREEABLE(cachep
, i
);
3338 raw_spin_unlock(&n
->list_lock
);
3339 ac
->avail
-= batchcount
;
3340 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3341 slabs_destroy(cachep
, &list
);
3345 * Release an obj back to its cache. If the obj has a constructed state, it must
3346 * be in this state _before_ it is released. Called with disabled ints.
3348 static __always_inline
void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3349 unsigned long caller
)
3353 memcg_slab_free_hook(cachep
, virt_to_slab(objp
), &objp
, 1);
3355 if (is_kfence_address(objp
)) {
3356 kmemleak_free_recursive(objp
, cachep
->flags
);
3357 __kfence_free(objp
);
3362 * As memory initialization might be integrated into KASAN,
3363 * kasan_slab_free and initialization memset must be
3364 * kept together to avoid discrepancies in behavior.
3366 init
= slab_want_init_on_free(cachep
);
3367 if (init
&& !kasan_has_integrated_init())
3368 memset(objp
, 0, cachep
->object_size
);
3369 /* KASAN might put objp into memory quarantine, delaying its reuse. */
3370 if (kasan_slab_free(cachep
, objp
, init
))
3373 /* Use KCSAN to help debug racy use-after-free. */
3374 if (!(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
3375 __kcsan_check_access(objp
, cachep
->object_size
,
3376 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
3378 ___cache_free(cachep
, objp
, caller
);
3381 void ___cache_free(struct kmem_cache
*cachep
, void *objp
,
3382 unsigned long caller
)
3384 struct array_cache
*ac
= cpu_cache_get(cachep
);
3387 kmemleak_free_recursive(objp
, cachep
->flags
);
3388 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3391 * Skip calling cache_free_alien() when the platform is not numa.
3392 * This will avoid cache misses that happen while accessing slabp (which
3393 * is per page memory reference) to get nodeid. Instead use a global
3394 * variable to skip the call, which is mostly likely to be present in
3397 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3400 if (ac
->avail
< ac
->limit
) {
3401 STATS_INC_FREEHIT(cachep
);
3403 STATS_INC_FREEMISS(cachep
);
3404 cache_flusharray(cachep
, ac
);
3407 if (sk_memalloc_socks()) {
3408 struct slab
*slab
= virt_to_slab(objp
);
3410 if (unlikely(slab_test_pfmemalloc(slab
))) {
3411 cache_free_pfmemalloc(cachep
, slab
, objp
);
3416 __free_one(ac
, objp
);
3419 static __always_inline
3420 void *__kmem_cache_alloc_lru(struct kmem_cache
*cachep
, struct list_lru
*lru
,
3423 void *ret
= slab_alloc(cachep
, lru
, flags
, cachep
->object_size
, _RET_IP_
);
3425 trace_kmem_cache_alloc(_RET_IP_
, ret
, cachep
, flags
, NUMA_NO_NODE
);
3430 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3432 return __kmem_cache_alloc_lru(cachep
, NULL
, flags
);
3434 EXPORT_SYMBOL(kmem_cache_alloc
);
3436 void *kmem_cache_alloc_lru(struct kmem_cache
*cachep
, struct list_lru
*lru
,
3439 return __kmem_cache_alloc_lru(cachep
, lru
, flags
);
3441 EXPORT_SYMBOL(kmem_cache_alloc_lru
);
3443 static __always_inline
void
3444 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3445 size_t size
, void **p
, unsigned long caller
)
3449 for (i
= 0; i
< size
; i
++)
3450 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3453 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3456 struct obj_cgroup
*objcg
= NULL
;
3457 unsigned long irqflags
;
3460 s
= slab_pre_alloc_hook(s
, NULL
, &objcg
, size
, flags
);
3464 local_irq_save(irqflags
);
3465 for (i
= 0; i
< size
; i
++) {
3466 void *objp
= kfence_alloc(s
, s
->object_size
, flags
) ?:
3467 __do_cache_alloc(s
, flags
, NUMA_NO_NODE
);
3469 if (unlikely(!objp
))
3473 local_irq_restore(irqflags
);
3475 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3478 * memcg and kmem_cache debug support and memory initialization.
3479 * Done outside of the IRQ disabled section.
3481 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
,
3482 slab_want_init_on_alloc(flags
, s
), s
->object_size
);
3483 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3486 local_irq_restore(irqflags
);
3487 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3488 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
, false, s
->object_size
);
3489 kmem_cache_free_bulk(s
, i
, p
);
3492 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3495 * kmem_cache_alloc_node - Allocate an object on the specified node
3496 * @cachep: The cache to allocate from.
3497 * @flags: See kmalloc().
3498 * @nodeid: node number of the target node.
3500 * Identical to kmem_cache_alloc but it will allocate memory on the given
3501 * node, which can improve the performance for cpu bound structures.
3503 * Fallback to other node is possible if __GFP_THISNODE is not set.
3505 * Return: pointer to the new object or %NULL in case of error
3507 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3509 void *ret
= slab_alloc_node(cachep
, NULL
, flags
, nodeid
, cachep
->object_size
, _RET_IP_
);
3511 trace_kmem_cache_alloc(_RET_IP_
, ret
, cachep
, flags
, nodeid
);
3515 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3517 void *__kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3518 int nodeid
, size_t orig_size
,
3519 unsigned long caller
)
3521 return slab_alloc_node(cachep
, NULL
, flags
, nodeid
,
3525 #ifdef CONFIG_PRINTK
3526 void __kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct slab
*slab
)
3528 struct kmem_cache
*cachep
;
3532 kpp
->kp_ptr
= object
;
3533 kpp
->kp_slab
= slab
;
3534 cachep
= slab
->slab_cache
;
3535 kpp
->kp_slab_cache
= cachep
;
3536 objp
= object
- obj_offset(cachep
);
3537 kpp
->kp_data_offset
= obj_offset(cachep
);
3538 slab
= virt_to_slab(objp
);
3539 objnr
= obj_to_index(cachep
, slab
, objp
);
3540 objp
= index_to_obj(cachep
, slab
, objnr
);
3541 kpp
->kp_objp
= objp
;
3542 if (DEBUG
&& cachep
->flags
& SLAB_STORE_USER
)
3543 kpp
->kp_ret
= *dbg_userword(cachep
, objp
);
3547 static __always_inline
3548 void __do_kmem_cache_free(struct kmem_cache
*cachep
, void *objp
,
3549 unsigned long caller
)
3551 unsigned long flags
;
3553 local_irq_save(flags
);
3554 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3555 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3556 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3557 __cache_free(cachep
, objp
, caller
);
3558 local_irq_restore(flags
);
3561 void __kmem_cache_free(struct kmem_cache
*cachep
, void *objp
,
3562 unsigned long caller
)
3564 __do_kmem_cache_free(cachep
, objp
, caller
);
3568 * kmem_cache_free - Deallocate an object
3569 * @cachep: The cache the allocation was from.
3570 * @objp: The previously allocated object.
3572 * Free an object which was previously allocated from this
3575 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3577 cachep
= cache_from_obj(cachep
, objp
);
3581 trace_kmem_cache_free(_RET_IP_
, objp
, cachep
);
3582 __do_kmem_cache_free(cachep
, objp
, _RET_IP_
);
3584 EXPORT_SYMBOL(kmem_cache_free
);
3586 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3588 unsigned long flags
;
3590 local_irq_save(flags
);
3591 for (int i
= 0; i
< size
; i
++) {
3593 struct kmem_cache
*s
;
3596 struct folio
*folio
= virt_to_folio(objp
);
3598 /* called via kfree_bulk */
3599 if (!folio_test_slab(folio
)) {
3600 local_irq_restore(flags
);
3601 free_large_kmalloc(folio
, objp
);
3602 local_irq_save(flags
);
3605 s
= folio_slab(folio
)->slab_cache
;
3607 s
= cache_from_obj(orig_s
, objp
);
3613 debug_check_no_locks_freed(objp
, s
->object_size
);
3614 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3615 debug_check_no_obj_freed(objp
, s
->object_size
);
3617 __cache_free(s
, objp
, _RET_IP_
);
3619 local_irq_restore(flags
);
3621 /* FIXME: add tracing */
3623 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3626 * This initializes kmem_cache_node or resizes various caches for all nodes.
3628 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3632 struct kmem_cache_node
*n
;
3634 for_each_online_node(node
) {
3635 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3644 if (!cachep
->list
.next
) {
3645 /* Cache is not active yet. Roll back what we did */
3648 n
= get_node(cachep
, node
);
3651 free_alien_cache(n
->alien
);
3653 cachep
->node
[node
] = NULL
;
3661 /* Always called with the slab_mutex held */
3662 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3663 int batchcount
, int shared
, gfp_t gfp
)
3665 struct array_cache __percpu
*cpu_cache
, *prev
;
3668 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3672 prev
= cachep
->cpu_cache
;
3673 cachep
->cpu_cache
= cpu_cache
;
3675 * Without a previous cpu_cache there's no need to synchronize remote
3676 * cpus, so skip the IPIs.
3679 kick_all_cpus_sync();
3682 cachep
->batchcount
= batchcount
;
3683 cachep
->limit
= limit
;
3684 cachep
->shared
= shared
;
3689 for_each_online_cpu(cpu
) {
3692 struct kmem_cache_node
*n
;
3693 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3695 node
= cpu_to_mem(cpu
);
3696 n
= get_node(cachep
, node
);
3697 raw_spin_lock_irq(&n
->list_lock
);
3698 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3699 raw_spin_unlock_irq(&n
->list_lock
);
3700 slabs_destroy(cachep
, &list
);
3705 return setup_kmem_cache_nodes(cachep
, gfp
);
3708 /* Called with slab_mutex held always */
3709 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3716 err
= cache_random_seq_create(cachep
, cachep
->num
, gfp
);
3721 * The head array serves three purposes:
3722 * - create a LIFO ordering, i.e. return objects that are cache-warm
3723 * - reduce the number of spinlock operations.
3724 * - reduce the number of linked list operations on the slab and
3725 * bufctl chains: array operations are cheaper.
3726 * The numbers are guessed, we should auto-tune as described by
3729 if (cachep
->size
> 131072)
3731 else if (cachep
->size
> PAGE_SIZE
)
3733 else if (cachep
->size
> 1024)
3735 else if (cachep
->size
> 256)
3741 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3742 * allocation behaviour: Most allocs on one cpu, most free operations
3743 * on another cpu. For these cases, an efficient object passing between
3744 * cpus is necessary. This is provided by a shared array. The array
3745 * replaces Bonwick's magazine layer.
3746 * On uniprocessor, it's functionally equivalent (but less efficient)
3747 * to a larger limit. Thus disabled by default.
3750 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3755 * With debugging enabled, large batchcount lead to excessively long
3756 * periods with disabled local interrupts. Limit the batchcount
3761 batchcount
= (limit
+ 1) / 2;
3762 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3765 pr_err("enable_cpucache failed for %s, error %d\n",
3766 cachep
->name
, -err
);
3771 * Drain an array if it contains any elements taking the node lock only if
3772 * necessary. Note that the node listlock also protects the array_cache
3773 * if drain_array() is used on the shared array.
3775 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3776 struct array_cache
*ac
, int node
)
3780 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3781 check_mutex_acquired();
3783 if (!ac
|| !ac
->avail
)
3791 raw_spin_lock_irq(&n
->list_lock
);
3792 drain_array_locked(cachep
, ac
, node
, false, &list
);
3793 raw_spin_unlock_irq(&n
->list_lock
);
3795 slabs_destroy(cachep
, &list
);
3799 * cache_reap - Reclaim memory from caches.
3800 * @w: work descriptor
3802 * Called from workqueue/eventd every few seconds.
3804 * - clear the per-cpu caches for this CPU.
3805 * - return freeable pages to the main free memory pool.
3807 * If we cannot acquire the cache chain mutex then just give up - we'll try
3808 * again on the next iteration.
3810 static void cache_reap(struct work_struct
*w
)
3812 struct kmem_cache
*searchp
;
3813 struct kmem_cache_node
*n
;
3814 int node
= numa_mem_id();
3815 struct delayed_work
*work
= to_delayed_work(w
);
3817 if (!mutex_trylock(&slab_mutex
))
3818 /* Give up. Setup the next iteration. */
3821 list_for_each_entry(searchp
, &slab_caches
, list
) {
3825 * We only take the node lock if absolutely necessary and we
3826 * have established with reasonable certainty that
3827 * we can do some work if the lock was obtained.
3829 n
= get_node(searchp
, node
);
3831 reap_alien(searchp
, n
);
3833 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3836 * These are racy checks but it does not matter
3837 * if we skip one check or scan twice.
3839 if (time_after(n
->next_reap
, jiffies
))
3842 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3844 drain_array(searchp
, n
, n
->shared
, node
);
3846 if (n
->free_touched
)
3847 n
->free_touched
= 0;
3851 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3852 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3853 STATS_ADD_REAPED(searchp
, freed
);
3859 mutex_unlock(&slab_mutex
);
3862 /* Set up the next iteration */
3863 schedule_delayed_work_on(smp_processor_id(), work
,
3864 round_jiffies_relative(REAPTIMEOUT_AC
));
3867 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3869 unsigned long active_objs
, num_objs
, active_slabs
;
3870 unsigned long total_slabs
= 0, free_objs
= 0, shared_avail
= 0;
3871 unsigned long free_slabs
= 0;
3873 struct kmem_cache_node
*n
;
3875 for_each_kmem_cache_node(cachep
, node
, n
) {
3877 raw_spin_lock_irq(&n
->list_lock
);
3879 total_slabs
+= n
->total_slabs
;
3880 free_slabs
+= n
->free_slabs
;
3881 free_objs
+= n
->free_objects
;
3884 shared_avail
+= n
->shared
->avail
;
3886 raw_spin_unlock_irq(&n
->list_lock
);
3888 num_objs
= total_slabs
* cachep
->num
;
3889 active_slabs
= total_slabs
- free_slabs
;
3890 active_objs
= num_objs
- free_objs
;
3892 sinfo
->active_objs
= active_objs
;
3893 sinfo
->num_objs
= num_objs
;
3894 sinfo
->active_slabs
= active_slabs
;
3895 sinfo
->num_slabs
= total_slabs
;
3896 sinfo
->shared_avail
= shared_avail
;
3897 sinfo
->limit
= cachep
->limit
;
3898 sinfo
->batchcount
= cachep
->batchcount
;
3899 sinfo
->shared
= cachep
->shared
;
3900 sinfo
->objects_per_slab
= cachep
->num
;
3901 sinfo
->cache_order
= cachep
->gfporder
;
3904 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
3908 unsigned long high
= cachep
->high_mark
;
3909 unsigned long allocs
= cachep
->num_allocations
;
3910 unsigned long grown
= cachep
->grown
;
3911 unsigned long reaped
= cachep
->reaped
;
3912 unsigned long errors
= cachep
->errors
;
3913 unsigned long max_freeable
= cachep
->max_freeable
;
3914 unsigned long node_allocs
= cachep
->node_allocs
;
3915 unsigned long node_frees
= cachep
->node_frees
;
3916 unsigned long overflows
= cachep
->node_overflow
;
3918 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
3919 allocs
, high
, grown
,
3920 reaped
, errors
, max_freeable
, node_allocs
,
3921 node_frees
, overflows
);
3925 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3926 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3927 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3928 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3930 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3931 allochit
, allocmiss
, freehit
, freemiss
);
3936 #define MAX_SLABINFO_WRITE 128
3938 * slabinfo_write - Tuning for the slab allocator
3940 * @buffer: user buffer
3941 * @count: data length
3944 * Return: %0 on success, negative error code otherwise.
3946 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
3947 size_t count
, loff_t
*ppos
)
3949 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3950 int limit
, batchcount
, shared
, res
;
3951 struct kmem_cache
*cachep
;
3953 if (count
> MAX_SLABINFO_WRITE
)
3955 if (copy_from_user(&kbuf
, buffer
, count
))
3957 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3959 tmp
= strchr(kbuf
, ' ');
3964 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3967 /* Find the cache in the chain of caches. */
3968 mutex_lock(&slab_mutex
);
3970 list_for_each_entry(cachep
, &slab_caches
, list
) {
3971 if (!strcmp(cachep
->name
, kbuf
)) {
3972 if (limit
< 1 || batchcount
< 1 ||
3973 batchcount
> limit
|| shared
< 0) {
3976 res
= do_tune_cpucache(cachep
, limit
,
3983 mutex_unlock(&slab_mutex
);
3989 #ifdef CONFIG_HARDENED_USERCOPY
3991 * Rejects incorrectly sized objects and objects that are to be copied
3992 * to/from userspace but do not fall entirely within the containing slab
3993 * cache's usercopy region.
3995 * Returns NULL if check passes, otherwise const char * to name of cache
3996 * to indicate an error.
3998 void __check_heap_object(const void *ptr
, unsigned long n
,
3999 const struct slab
*slab
, bool to_user
)
4001 struct kmem_cache
*cachep
;
4003 unsigned long offset
;
4005 ptr
= kasan_reset_tag(ptr
);
4007 /* Find and validate object. */
4008 cachep
= slab
->slab_cache
;
4009 objnr
= obj_to_index(cachep
, slab
, (void *)ptr
);
4010 BUG_ON(objnr
>= cachep
->num
);
4012 /* Find offset within object. */
4013 if (is_kfence_address(ptr
))
4014 offset
= ptr
- kfence_object_start(ptr
);
4016 offset
= ptr
- index_to_obj(cachep
, slab
, objnr
) - obj_offset(cachep
);
4018 /* Allow address range falling entirely within usercopy region. */
4019 if (offset
>= cachep
->useroffset
&&
4020 offset
- cachep
->useroffset
<= cachep
->usersize
&&
4021 n
<= cachep
->useroffset
- offset
+ cachep
->usersize
)
4024 usercopy_abort("SLAB object", cachep
->name
, to_user
, offset
, n
);
4026 #endif /* CONFIG_HARDENED_USERCOPY */