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
,
223 static int slab_early_init
= 1;
225 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
227 static void kmem_cache_node_init(struct kmem_cache_node
*parent
)
229 INIT_LIST_HEAD(&parent
->slabs_full
);
230 INIT_LIST_HEAD(&parent
->slabs_partial
);
231 INIT_LIST_HEAD(&parent
->slabs_free
);
232 parent
->total_slabs
= 0;
233 parent
->free_slabs
= 0;
234 parent
->shared
= NULL
;
235 parent
->alien
= NULL
;
236 parent
->colour_next
= 0;
237 spin_lock_init(&parent
->list_lock
);
238 parent
->free_objects
= 0;
239 parent
->free_touched
= 0;
242 #define MAKE_LIST(cachep, listp, slab, nodeid) \
244 INIT_LIST_HEAD(listp); \
245 list_splice(&get_node(cachep, nodeid)->slab, listp); \
248 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
250 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
251 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
252 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
255 #define CFLGS_OBJFREELIST_SLAB ((slab_flags_t __force)0x40000000U)
256 #define CFLGS_OFF_SLAB ((slab_flags_t __force)0x80000000U)
257 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
258 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
260 #define BATCHREFILL_LIMIT 16
262 * Optimization question: fewer reaps means less probability for unnecessary
263 * cpucache drain/refill cycles.
265 * OTOH the cpuarrays can contain lots of objects,
266 * which could lock up otherwise freeable slabs.
268 #define REAPTIMEOUT_AC (2*HZ)
269 #define REAPTIMEOUT_NODE (4*HZ)
272 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
273 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
274 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
275 #define STATS_INC_GROWN(x) ((x)->grown++)
276 #define STATS_ADD_REAPED(x, y) ((x)->reaped += (y))
277 #define STATS_SET_HIGH(x) \
279 if ((x)->num_active > (x)->high_mark) \
280 (x)->high_mark = (x)->num_active; \
282 #define STATS_INC_ERR(x) ((x)->errors++)
283 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
284 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
285 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
286 #define STATS_SET_FREEABLE(x, i) \
288 if ((x)->max_freeable < i) \
289 (x)->max_freeable = i; \
291 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
292 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
293 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
294 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
296 #define STATS_INC_ACTIVE(x) do { } while (0)
297 #define STATS_DEC_ACTIVE(x) do { } while (0)
298 #define STATS_INC_ALLOCED(x) do { } while (0)
299 #define STATS_INC_GROWN(x) do { } while (0)
300 #define STATS_ADD_REAPED(x, y) do { (void)(y); } while (0)
301 #define STATS_SET_HIGH(x) do { } while (0)
302 #define STATS_INC_ERR(x) do { } while (0)
303 #define STATS_INC_NODEALLOCS(x) do { } while (0)
304 #define STATS_INC_NODEFREES(x) do { } while (0)
305 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
306 #define STATS_SET_FREEABLE(x, i) do { } while (0)
307 #define STATS_INC_ALLOCHIT(x) do { } while (0)
308 #define STATS_INC_ALLOCMISS(x) do { } while (0)
309 #define STATS_INC_FREEHIT(x) do { } while (0)
310 #define STATS_INC_FREEMISS(x) do { } while (0)
316 * memory layout of objects:
318 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
319 * the end of an object is aligned with the end of the real
320 * allocation. Catches writes behind the end of the allocation.
321 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
323 * cachep->obj_offset: The real object.
324 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
325 * cachep->size - 1* BYTES_PER_WORD: last caller address
326 * [BYTES_PER_WORD long]
328 static int obj_offset(struct kmem_cache
*cachep
)
330 return cachep
->obj_offset
;
333 static unsigned long long *dbg_redzone1(struct kmem_cache
*cachep
, void *objp
)
335 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
336 return (unsigned long long *) (objp
+ obj_offset(cachep
) -
337 sizeof(unsigned long long));
340 static unsigned long long *dbg_redzone2(struct kmem_cache
*cachep
, void *objp
)
342 BUG_ON(!(cachep
->flags
& SLAB_RED_ZONE
));
343 if (cachep
->flags
& SLAB_STORE_USER
)
344 return (unsigned long long *)(objp
+ cachep
->size
-
345 sizeof(unsigned long long) -
347 return (unsigned long long *) (objp
+ cachep
->size
-
348 sizeof(unsigned long long));
351 static void **dbg_userword(struct kmem_cache
*cachep
, void *objp
)
353 BUG_ON(!(cachep
->flags
& SLAB_STORE_USER
));
354 return (void **)(objp
+ cachep
->size
- BYTES_PER_WORD
);
359 #define obj_offset(x) 0
360 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
361 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
362 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
367 * Do not go above this order unless 0 objects fit into the slab or
368 * overridden on the command line.
370 #define SLAB_MAX_ORDER_HI 1
371 #define SLAB_MAX_ORDER_LO 0
372 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
373 static bool slab_max_order_set __initdata
;
375 static inline void *index_to_obj(struct kmem_cache
*cache
,
376 const struct slab
*slab
, unsigned int idx
)
378 return slab
->s_mem
+ cache
->size
* idx
;
381 #define BOOT_CPUCACHE_ENTRIES 1
382 /* internal cache of cache description objs */
383 static struct kmem_cache kmem_cache_boot
= {
385 .limit
= BOOT_CPUCACHE_ENTRIES
,
387 .size
= sizeof(struct kmem_cache
),
388 .name
= "kmem_cache",
391 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
393 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
395 return this_cpu_ptr(cachep
->cpu_cache
);
399 * Calculate the number of objects and left-over bytes for a given buffer size.
401 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
402 slab_flags_t flags
, size_t *left_over
)
405 size_t slab_size
= PAGE_SIZE
<< gfporder
;
408 * The slab management structure can be either off the slab or
409 * on it. For the latter case, the memory allocated for a
412 * - @buffer_size bytes for each object
413 * - One freelist_idx_t for each object
415 * We don't need to consider alignment of freelist because
416 * freelist will be at the end of slab page. The objects will be
417 * at the correct alignment.
419 * If the slab management structure is off the slab, then the
420 * alignment will already be calculated into the size. Because
421 * the slabs are all pages aligned, the objects will be at the
422 * correct alignment when allocated.
424 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
425 num
= slab_size
/ buffer_size
;
426 *left_over
= slab_size
% buffer_size
;
428 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
429 *left_over
= slab_size
%
430 (buffer_size
+ sizeof(freelist_idx_t
));
437 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
439 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
442 pr_err("slab error in %s(): cache `%s': %s\n",
443 function
, cachep
->name
, msg
);
445 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
450 * By default on NUMA we use alien caches to stage the freeing of
451 * objects allocated from other nodes. This causes massive memory
452 * inefficiencies when using fake NUMA setup to split memory into a
453 * large number of small nodes, so it can be disabled on the command
457 static int use_alien_caches __read_mostly
= 1;
458 static int __init
noaliencache_setup(char *s
)
460 use_alien_caches
= 0;
463 __setup("noaliencache", noaliencache_setup
);
465 static int __init
slab_max_order_setup(char *str
)
467 get_option(&str
, &slab_max_order
);
468 slab_max_order
= slab_max_order
< 0 ? 0 :
469 min(slab_max_order
, MAX_ORDER
- 1);
470 slab_max_order_set
= true;
474 __setup("slab_max_order=", slab_max_order_setup
);
478 * Special reaping functions for NUMA systems called from cache_reap().
479 * These take care of doing round robin flushing of alien caches (containing
480 * objects freed on different nodes from which they were allocated) and the
481 * flushing of remote pcps by calling drain_node_pages.
483 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
485 static void init_reap_node(int cpu
)
487 per_cpu(slab_reap_node
, cpu
) = next_node_in(cpu_to_mem(cpu
),
491 static void next_reap_node(void)
493 int node
= __this_cpu_read(slab_reap_node
);
495 node
= next_node_in(node
, node_online_map
);
496 __this_cpu_write(slab_reap_node
, node
);
500 #define init_reap_node(cpu) do { } while (0)
501 #define next_reap_node(void) do { } while (0)
505 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
506 * via the workqueue/eventd.
507 * Add the CPU number into the expiration time to minimize the possibility of
508 * the CPUs getting into lockstep and contending for the global cache chain
511 static void start_cpu_timer(int cpu
)
513 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
515 if (reap_work
->work
.func
== NULL
) {
517 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
518 schedule_delayed_work_on(cpu
, reap_work
,
519 __round_jiffies_relative(HZ
, cpu
));
523 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
528 ac
->batchcount
= batch
;
533 static struct array_cache
*alloc_arraycache(int node
, int entries
,
534 int batchcount
, gfp_t gfp
)
536 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
537 struct array_cache
*ac
= NULL
;
539 ac
= kmalloc_node(memsize
, gfp
, node
);
541 * The array_cache structures contain pointers to free object.
542 * However, when such objects are allocated or transferred to another
543 * cache the pointers are not cleared and they could be counted as
544 * valid references during a kmemleak scan. Therefore, kmemleak must
545 * not scan such objects.
547 kmemleak_no_scan(ac
);
548 init_arraycache(ac
, entries
, batchcount
);
552 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
553 struct slab
*slab
, void *objp
)
555 struct kmem_cache_node
*n
;
559 slab_node
= slab_nid(slab
);
560 n
= get_node(cachep
, slab_node
);
562 spin_lock(&n
->list_lock
);
563 free_block(cachep
, &objp
, 1, slab_node
, &list
);
564 spin_unlock(&n
->list_lock
);
566 slabs_destroy(cachep
, &list
);
570 * Transfer objects in one arraycache to another.
571 * Locking must be handled by the caller.
573 * Return the number of entries transferred.
575 static int transfer_objects(struct array_cache
*to
,
576 struct array_cache
*from
, unsigned int max
)
578 /* Figure out how many entries to transfer */
579 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
584 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
- nr
,
592 /* &alien->lock must be held by alien callers. */
593 static __always_inline
void __free_one(struct array_cache
*ac
, void *objp
)
595 /* Avoid trivial double-free. */
596 if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED
) &&
597 WARN_ON_ONCE(ac
->avail
> 0 && ac
->entry
[ac
->avail
- 1] == objp
))
599 ac
->entry
[ac
->avail
++] = objp
;
604 #define drain_alien_cache(cachep, alien) do { } while (0)
605 #define reap_alien(cachep, n) do { } while (0)
607 static inline struct alien_cache
**alloc_alien_cache(int node
,
608 int limit
, gfp_t gfp
)
613 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
617 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
622 static inline gfp_t
gfp_exact_node(gfp_t flags
)
624 return flags
& ~__GFP_NOFAIL
;
627 #else /* CONFIG_NUMA */
629 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
630 int batch
, gfp_t gfp
)
632 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
633 struct alien_cache
*alc
= NULL
;
635 alc
= kmalloc_node(memsize
, gfp
, node
);
637 kmemleak_no_scan(alc
);
638 init_arraycache(&alc
->ac
, entries
, batch
);
639 spin_lock_init(&alc
->lock
);
644 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
646 struct alien_cache
**alc_ptr
;
651 alc_ptr
= kcalloc_node(nr_node_ids
, sizeof(void *), gfp
, node
);
656 if (i
== node
|| !node_online(i
))
658 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
660 for (i
--; i
>= 0; i
--)
669 static void free_alien_cache(struct alien_cache
**alc_ptr
)
680 static void __drain_alien_cache(struct kmem_cache
*cachep
,
681 struct array_cache
*ac
, int node
,
682 struct list_head
*list
)
684 struct kmem_cache_node
*n
= get_node(cachep
, node
);
687 spin_lock(&n
->list_lock
);
689 * Stuff objects into the remote nodes shared array first.
690 * That way we could avoid the overhead of putting the objects
691 * into the free lists and getting them back later.
694 transfer_objects(n
->shared
, ac
, ac
->limit
);
696 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
698 spin_unlock(&n
->list_lock
);
703 * Called from cache_reap() to regularly drain alien caches round robin.
705 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
707 int node
= __this_cpu_read(slab_reap_node
);
710 struct alien_cache
*alc
= n
->alien
[node
];
711 struct array_cache
*ac
;
715 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
718 __drain_alien_cache(cachep
, ac
, node
, &list
);
719 spin_unlock_irq(&alc
->lock
);
720 slabs_destroy(cachep
, &list
);
726 static void drain_alien_cache(struct kmem_cache
*cachep
,
727 struct alien_cache
**alien
)
730 struct alien_cache
*alc
;
731 struct array_cache
*ac
;
734 for_each_online_node(i
) {
740 spin_lock_irqsave(&alc
->lock
, flags
);
741 __drain_alien_cache(cachep
, ac
, i
, &list
);
742 spin_unlock_irqrestore(&alc
->lock
, flags
);
743 slabs_destroy(cachep
, &list
);
748 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
749 int node
, int slab_node
)
751 struct kmem_cache_node
*n
;
752 struct alien_cache
*alien
= NULL
;
753 struct array_cache
*ac
;
756 n
= get_node(cachep
, node
);
757 STATS_INC_NODEFREES(cachep
);
758 if (n
->alien
&& n
->alien
[slab_node
]) {
759 alien
= n
->alien
[slab_node
];
761 spin_lock(&alien
->lock
);
762 if (unlikely(ac
->avail
== ac
->limit
)) {
763 STATS_INC_ACOVERFLOW(cachep
);
764 __drain_alien_cache(cachep
, ac
, slab_node
, &list
);
766 __free_one(ac
, objp
);
767 spin_unlock(&alien
->lock
);
768 slabs_destroy(cachep
, &list
);
770 n
= get_node(cachep
, slab_node
);
771 spin_lock(&n
->list_lock
);
772 free_block(cachep
, &objp
, 1, slab_node
, &list
);
773 spin_unlock(&n
->list_lock
);
774 slabs_destroy(cachep
, &list
);
779 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
781 int slab_node
= slab_nid(virt_to_slab(objp
));
782 int node
= numa_mem_id();
784 * Make sure we are not freeing an object from another node to the array
787 if (likely(node
== slab_node
))
790 return __cache_free_alien(cachep
, objp
, node
, slab_node
);
794 * Construct gfp mask to allocate from a specific node but do not reclaim or
795 * warn about failures.
797 static inline gfp_t
gfp_exact_node(gfp_t flags
)
799 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
803 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
805 struct kmem_cache_node
*n
;
808 * Set up the kmem_cache_node for cpu before we can
809 * begin anything. Make sure some other cpu on this
810 * node has not already allocated this
812 n
= get_node(cachep
, node
);
814 spin_lock_irq(&n
->list_lock
);
815 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
817 spin_unlock_irq(&n
->list_lock
);
822 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
826 kmem_cache_node_init(n
);
827 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
828 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
831 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
834 * The kmem_cache_nodes don't come and go as CPUs
835 * come and go. slab_mutex provides sufficient
838 cachep
->node
[node
] = n
;
843 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
845 * Allocates and initializes node for a node on each slab cache, used for
846 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
847 * will be allocated off-node since memory is not yet online for the new node.
848 * When hotplugging memory or a cpu, existing nodes are not replaced if
851 * Must hold slab_mutex.
853 static int init_cache_node_node(int node
)
856 struct kmem_cache
*cachep
;
858 list_for_each_entry(cachep
, &slab_caches
, list
) {
859 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
868 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
869 int node
, gfp_t gfp
, bool force_change
)
872 struct kmem_cache_node
*n
;
873 struct array_cache
*old_shared
= NULL
;
874 struct array_cache
*new_shared
= NULL
;
875 struct alien_cache
**new_alien
= NULL
;
878 if (use_alien_caches
) {
879 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
884 if (cachep
->shared
) {
885 new_shared
= alloc_arraycache(node
,
886 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
891 ret
= init_cache_node(cachep
, node
, gfp
);
895 n
= get_node(cachep
, node
);
896 spin_lock_irq(&n
->list_lock
);
897 if (n
->shared
&& force_change
) {
898 free_block(cachep
, n
->shared
->entry
,
899 n
->shared
->avail
, node
, &list
);
900 n
->shared
->avail
= 0;
903 if (!n
->shared
|| force_change
) {
904 old_shared
= n
->shared
;
905 n
->shared
= new_shared
;
910 n
->alien
= new_alien
;
914 spin_unlock_irq(&n
->list_lock
);
915 slabs_destroy(cachep
, &list
);
918 * To protect lockless access to n->shared during irq disabled context.
919 * If n->shared isn't NULL in irq disabled context, accessing to it is
920 * guaranteed to be valid until irq is re-enabled, because it will be
921 * freed after synchronize_rcu().
923 if (old_shared
&& force_change
)
929 free_alien_cache(new_alien
);
936 static void cpuup_canceled(long cpu
)
938 struct kmem_cache
*cachep
;
939 struct kmem_cache_node
*n
= NULL
;
940 int node
= cpu_to_mem(cpu
);
941 const struct cpumask
*mask
= cpumask_of_node(node
);
943 list_for_each_entry(cachep
, &slab_caches
, list
) {
944 struct array_cache
*nc
;
945 struct array_cache
*shared
;
946 struct alien_cache
**alien
;
949 n
= get_node(cachep
, node
);
953 spin_lock_irq(&n
->list_lock
);
955 /* Free limit for this kmem_cache_node */
956 n
->free_limit
-= cachep
->batchcount
;
958 /* cpu is dead; no one can alloc from it. */
959 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
960 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
963 if (!cpumask_empty(mask
)) {
964 spin_unlock_irq(&n
->list_lock
);
970 free_block(cachep
, shared
->entry
,
971 shared
->avail
, node
, &list
);
978 spin_unlock_irq(&n
->list_lock
);
982 drain_alien_cache(cachep
, alien
);
983 free_alien_cache(alien
);
987 slabs_destroy(cachep
, &list
);
990 * In the previous loop, all the objects were freed to
991 * the respective cache's slabs, now we can go ahead and
992 * shrink each nodelist to its limit.
994 list_for_each_entry(cachep
, &slab_caches
, list
) {
995 n
= get_node(cachep
, node
);
998 drain_freelist(cachep
, n
, INT_MAX
);
1002 static int cpuup_prepare(long cpu
)
1004 struct kmem_cache
*cachep
;
1005 int node
= cpu_to_mem(cpu
);
1009 * We need to do this right in the beginning since
1010 * alloc_arraycache's are going to use this list.
1011 * kmalloc_node allows us to add the slab to the right
1012 * kmem_cache_node and not this cpu's kmem_cache_node
1014 err
= init_cache_node_node(node
);
1019 * Now we can go ahead with allocating the shared arrays and
1022 list_for_each_entry(cachep
, &slab_caches
, list
) {
1023 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1030 cpuup_canceled(cpu
);
1034 int slab_prepare_cpu(unsigned int cpu
)
1038 mutex_lock(&slab_mutex
);
1039 err
= cpuup_prepare(cpu
);
1040 mutex_unlock(&slab_mutex
);
1045 * This is called for a failed online attempt and for a successful
1048 * Even if all the cpus of a node are down, we don't free the
1049 * kmem_cache_node of any cache. This is to avoid a race between cpu_down, and
1050 * a kmalloc allocation from another cpu for memory from the node of
1051 * the cpu going down. The kmem_cache_node structure is usually allocated from
1052 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1054 int slab_dead_cpu(unsigned int cpu
)
1056 mutex_lock(&slab_mutex
);
1057 cpuup_canceled(cpu
);
1058 mutex_unlock(&slab_mutex
);
1063 static int slab_online_cpu(unsigned int cpu
)
1065 start_cpu_timer(cpu
);
1069 static int slab_offline_cpu(unsigned int cpu
)
1072 * Shutdown cache reaper. Note that the slab_mutex is held so
1073 * that if cache_reap() is invoked it cannot do anything
1074 * expensive but will only modify reap_work and reschedule the
1077 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1078 /* Now the cache_reaper is guaranteed to be not running. */
1079 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1083 #if defined(CONFIG_NUMA)
1085 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1086 * Returns -EBUSY if all objects cannot be drained so that the node is not
1089 * Must hold slab_mutex.
1091 static int __meminit
drain_cache_node_node(int node
)
1093 struct kmem_cache
*cachep
;
1096 list_for_each_entry(cachep
, &slab_caches
, list
) {
1097 struct kmem_cache_node
*n
;
1099 n
= get_node(cachep
, node
);
1103 drain_freelist(cachep
, n
, INT_MAX
);
1105 if (!list_empty(&n
->slabs_full
) ||
1106 !list_empty(&n
->slabs_partial
)) {
1114 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1115 unsigned long action
, void *arg
)
1117 struct memory_notify
*mnb
= arg
;
1121 nid
= mnb
->status_change_nid
;
1126 case MEM_GOING_ONLINE
:
1127 mutex_lock(&slab_mutex
);
1128 ret
= init_cache_node_node(nid
);
1129 mutex_unlock(&slab_mutex
);
1131 case MEM_GOING_OFFLINE
:
1132 mutex_lock(&slab_mutex
);
1133 ret
= drain_cache_node_node(nid
);
1134 mutex_unlock(&slab_mutex
);
1138 case MEM_CANCEL_ONLINE
:
1139 case MEM_CANCEL_OFFLINE
:
1143 return notifier_from_errno(ret
);
1145 #endif /* CONFIG_NUMA */
1148 * swap the static kmem_cache_node with kmalloced memory
1150 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1153 struct kmem_cache_node
*ptr
;
1155 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1158 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1160 * Do not assume that spinlocks can be initialized via memcpy:
1162 spin_lock_init(&ptr
->list_lock
);
1164 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1165 cachep
->node
[nodeid
] = ptr
;
1169 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1170 * size of kmem_cache_node.
1172 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1176 for_each_online_node(node
) {
1177 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1178 cachep
->node
[node
]->next_reap
= jiffies
+
1180 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1185 * Initialisation. Called after the page allocator have been initialised and
1186 * before smp_init().
1188 void __init
kmem_cache_init(void)
1192 kmem_cache
= &kmem_cache_boot
;
1194 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1195 use_alien_caches
= 0;
1197 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1198 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1201 * Fragmentation resistance on low memory - only use bigger
1202 * page orders on machines with more than 32MB of memory if
1203 * not overridden on the command line.
1205 if (!slab_max_order_set
&& totalram_pages() > (32 << 20) >> PAGE_SHIFT
)
1206 slab_max_order
= SLAB_MAX_ORDER_HI
;
1208 /* Bootstrap is tricky, because several objects are allocated
1209 * from caches that do not exist yet:
1210 * 1) initialize the kmem_cache cache: it contains the struct
1211 * kmem_cache structures of all caches, except kmem_cache itself:
1212 * kmem_cache is statically allocated.
1213 * Initially an __init data area is used for the head array and the
1214 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1215 * array at the end of the bootstrap.
1216 * 2) Create the first kmalloc cache.
1217 * The struct kmem_cache for the new cache is allocated normally.
1218 * An __init data area is used for the head array.
1219 * 3) Create the remaining kmalloc caches, with minimally sized
1221 * 4) Replace the __init data head arrays for kmem_cache and the first
1222 * kmalloc cache with kmalloc allocated arrays.
1223 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1224 * the other cache's with kmalloc allocated memory.
1225 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1228 /* 1) create the kmem_cache */
1231 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1233 create_boot_cache(kmem_cache
, "kmem_cache",
1234 offsetof(struct kmem_cache
, node
) +
1235 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1236 SLAB_HWCACHE_ALIGN
, 0, 0);
1237 list_add(&kmem_cache
->list
, &slab_caches
);
1238 slab_state
= PARTIAL
;
1241 * Initialize the caches that provide memory for the kmem_cache_node
1242 * structures first. Without this, further allocations will bug.
1244 kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
] = create_kmalloc_cache(
1245 kmalloc_info
[INDEX_NODE
].name
[KMALLOC_NORMAL
],
1246 kmalloc_info
[INDEX_NODE
].size
,
1247 ARCH_KMALLOC_FLAGS
, 0,
1248 kmalloc_info
[INDEX_NODE
].size
);
1249 slab_state
= PARTIAL_NODE
;
1250 setup_kmalloc_cache_index_table();
1252 slab_early_init
= 0;
1254 /* 5) Replace the bootstrap kmem_cache_node */
1258 for_each_online_node(nid
) {
1259 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1261 init_list(kmalloc_caches
[KMALLOC_NORMAL
][INDEX_NODE
],
1262 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1266 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1269 void __init
kmem_cache_init_late(void)
1271 struct kmem_cache
*cachep
;
1273 /* 6) resize the head arrays to their final sizes */
1274 mutex_lock(&slab_mutex
);
1275 list_for_each_entry(cachep
, &slab_caches
, list
)
1276 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1278 mutex_unlock(&slab_mutex
);
1285 * Register a memory hotplug callback that initializes and frees
1288 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1292 * The reap timers are started later, with a module init call: That part
1293 * of the kernel is not yet operational.
1297 static int __init
cpucache_init(void)
1302 * Register the timers that return unneeded pages to the page allocator
1304 ret
= cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "SLAB online",
1305 slab_online_cpu
, slab_offline_cpu
);
1310 __initcall(cpucache_init
);
1312 static noinline
void
1313 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1316 struct kmem_cache_node
*n
;
1317 unsigned long flags
;
1319 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1320 DEFAULT_RATELIMIT_BURST
);
1322 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1325 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1326 nodeid
, gfpflags
, &gfpflags
);
1327 pr_warn(" cache: %s, object size: %d, order: %d\n",
1328 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1330 for_each_kmem_cache_node(cachep
, node
, n
) {
1331 unsigned long total_slabs
, free_slabs
, free_objs
;
1333 spin_lock_irqsave(&n
->list_lock
, flags
);
1334 total_slabs
= n
->total_slabs
;
1335 free_slabs
= n
->free_slabs
;
1336 free_objs
= n
->free_objects
;
1337 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1339 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1340 node
, total_slabs
- free_slabs
, total_slabs
,
1341 (total_slabs
* cachep
->num
) - free_objs
,
1342 total_slabs
* cachep
->num
);
1348 * Interface to system's page allocator. No need to hold the
1349 * kmem_cache_node ->list_lock.
1351 * If we requested dmaable memory, we will get it. Even if we
1352 * did not request dmaable memory, we might get it, but that
1353 * would be relatively rare and ignorable.
1355 static struct slab
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1358 struct folio
*folio
;
1361 flags
|= cachep
->allocflags
;
1363 folio
= (struct folio
*) __alloc_pages_node(nodeid
, flags
, cachep
->gfporder
);
1365 slab_out_of_memory(cachep
, flags
, nodeid
);
1369 slab
= folio_slab(folio
);
1371 account_slab(slab
, cachep
->gfporder
, cachep
, flags
);
1372 __folio_set_slab(folio
);
1373 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1374 if (sk_memalloc_socks() && page_is_pfmemalloc(folio_page(folio
, 0)))
1375 slab_set_pfmemalloc(slab
);
1381 * Interface to system's page release.
1383 static void kmem_freepages(struct kmem_cache
*cachep
, struct slab
*slab
)
1385 int order
= cachep
->gfporder
;
1386 struct folio
*folio
= slab_folio(slab
);
1388 BUG_ON(!folio_test_slab(folio
));
1389 __slab_clear_pfmemalloc(slab
);
1390 __folio_clear_slab(folio
);
1391 page_mapcount_reset(folio_page(folio
, 0));
1392 folio
->mapping
= NULL
;
1394 if (current
->reclaim_state
)
1395 current
->reclaim_state
->reclaimed_slab
+= 1 << order
;
1396 unaccount_slab(slab
, order
, cachep
);
1397 __free_pages(folio_page(folio
, 0), order
);
1400 static void kmem_rcu_free(struct rcu_head
*head
)
1402 struct kmem_cache
*cachep
;
1405 slab
= container_of(head
, struct slab
, rcu_head
);
1406 cachep
= slab
->slab_cache
;
1408 kmem_freepages(cachep
, slab
);
1412 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1414 if (debug_pagealloc_enabled_static() && OFF_SLAB(cachep
) &&
1415 (cachep
->size
% PAGE_SIZE
) == 0)
1421 #ifdef CONFIG_DEBUG_PAGEALLOC
1422 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
, int map
)
1424 if (!is_debug_pagealloc_cache(cachep
))
1427 __kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1431 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1436 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1438 int size
= cachep
->object_size
;
1439 addr
= &((char *)addr
)[obj_offset(cachep
)];
1441 memset(addr
, val
, size
);
1442 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1445 static void dump_line(char *data
, int offset
, int limit
)
1448 unsigned char error
= 0;
1451 pr_err("%03x: ", offset
);
1452 for (i
= 0; i
< limit
; i
++) {
1453 if (data
[offset
+ i
] != POISON_FREE
) {
1454 error
= data
[offset
+ i
];
1458 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1459 &data
[offset
], limit
, 1);
1461 if (bad_count
== 1) {
1462 error
^= POISON_FREE
;
1463 if (!(error
& (error
- 1))) {
1464 pr_err("Single bit error detected. Probably bad RAM.\n");
1466 pr_err("Run memtest86+ or a similar memory test tool.\n");
1468 pr_err("Run a memory test tool.\n");
1477 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1482 if (cachep
->flags
& SLAB_RED_ZONE
) {
1483 pr_err("Redzone: 0x%llx/0x%llx\n",
1484 *dbg_redzone1(cachep
, objp
),
1485 *dbg_redzone2(cachep
, objp
));
1488 if (cachep
->flags
& SLAB_STORE_USER
)
1489 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep
, objp
));
1490 realobj
= (char *)objp
+ obj_offset(cachep
);
1491 size
= cachep
->object_size
;
1492 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1495 if (i
+ limit
> size
)
1497 dump_line(realobj
, i
, limit
);
1501 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1507 if (is_debug_pagealloc_cache(cachep
))
1510 realobj
= (char *)objp
+ obj_offset(cachep
);
1511 size
= cachep
->object_size
;
1513 for (i
= 0; i
< size
; i
++) {
1514 char exp
= POISON_FREE
;
1517 if (realobj
[i
] != exp
) {
1522 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1523 print_tainted(), cachep
->name
,
1525 print_objinfo(cachep
, objp
, 0);
1527 /* Hexdump the affected line */
1530 if (i
+ limit
> size
)
1532 dump_line(realobj
, i
, limit
);
1535 /* Limit to 5 lines */
1541 /* Print some data about the neighboring objects, if they
1544 struct slab
*slab
= virt_to_slab(objp
);
1547 objnr
= obj_to_index(cachep
, slab
, objp
);
1549 objp
= index_to_obj(cachep
, slab
, objnr
- 1);
1550 realobj
= (char *)objp
+ obj_offset(cachep
);
1551 pr_err("Prev obj: start=%px, len=%d\n", realobj
, size
);
1552 print_objinfo(cachep
, objp
, 2);
1554 if (objnr
+ 1 < cachep
->num
) {
1555 objp
= index_to_obj(cachep
, slab
, objnr
+ 1);
1556 realobj
= (char *)objp
+ obj_offset(cachep
);
1557 pr_err("Next obj: start=%px, len=%d\n", realobj
, size
);
1558 print_objinfo(cachep
, objp
, 2);
1565 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1570 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1571 poison_obj(cachep
, slab
->freelist
- obj_offset(cachep
),
1575 for (i
= 0; i
< cachep
->num
; i
++) {
1576 void *objp
= index_to_obj(cachep
, slab
, i
);
1578 if (cachep
->flags
& SLAB_POISON
) {
1579 check_poison_obj(cachep
, objp
);
1580 slab_kernel_map(cachep
, objp
, 1);
1582 if (cachep
->flags
& SLAB_RED_ZONE
) {
1583 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1584 slab_error(cachep
, "start of a freed object was overwritten");
1585 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1586 slab_error(cachep
, "end of a freed object was overwritten");
1591 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1598 * slab_destroy - destroy and release all objects in a slab
1599 * @cachep: cache pointer being destroyed
1600 * @slab: slab being destroyed
1602 * Destroy all the objs in a slab, and release the mem back to the system.
1603 * Before calling the slab must have been unlinked from the cache. The
1604 * kmem_cache_node ->list_lock is not held/needed.
1606 static void slab_destroy(struct kmem_cache
*cachep
, struct slab
*slab
)
1610 freelist
= slab
->freelist
;
1611 slab_destroy_debugcheck(cachep
, slab
);
1612 if (unlikely(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
1613 call_rcu(&slab
->rcu_head
, kmem_rcu_free
);
1615 kmem_freepages(cachep
, slab
);
1618 * From now on, we don't use freelist
1619 * although actual page can be freed in rcu context
1621 if (OFF_SLAB(cachep
))
1626 * Update the size of the caches before calling slabs_destroy as it may
1627 * recursively call kfree.
1629 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1631 struct slab
*slab
, *n
;
1633 list_for_each_entry_safe(slab
, n
, list
, slab_list
) {
1634 list_del(&slab
->slab_list
);
1635 slab_destroy(cachep
, slab
);
1640 * calculate_slab_order - calculate size (page order) of slabs
1641 * @cachep: pointer to the cache that is being created
1642 * @size: size of objects to be created in this cache.
1643 * @flags: slab allocation flags
1645 * Also calculates the number of objects per slab.
1647 * This could be made much more intelligent. For now, try to avoid using
1648 * high order pages for slabs. When the gfp() functions are more friendly
1649 * towards high-order requests, this should be changed.
1651 * Return: number of left-over bytes in a slab
1653 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1654 size_t size
, slab_flags_t flags
)
1656 size_t left_over
= 0;
1659 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1663 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1667 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1668 if (num
> SLAB_OBJ_MAX_NUM
)
1671 if (flags
& CFLGS_OFF_SLAB
) {
1672 struct kmem_cache
*freelist_cache
;
1673 size_t freelist_size
;
1674 size_t freelist_cache_size
;
1676 freelist_size
= num
* sizeof(freelist_idx_t
);
1677 if (freelist_size
> KMALLOC_MAX_CACHE_SIZE
) {
1678 freelist_cache_size
= PAGE_SIZE
<< get_order(freelist_size
);
1680 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1681 if (!freelist_cache
)
1683 freelist_cache_size
= freelist_cache
->size
;
1686 * Needed to avoid possible looping condition
1687 * in cache_grow_begin()
1689 if (OFF_SLAB(freelist_cache
))
1693 /* check if off slab has enough benefit */
1694 if (freelist_cache_size
> cachep
->size
/ 2)
1698 /* Found something acceptable - save it away */
1700 cachep
->gfporder
= gfporder
;
1701 left_over
= remainder
;
1704 * A VFS-reclaimable slab tends to have most allocations
1705 * as GFP_NOFS and we really don't want to have to be allocating
1706 * higher-order pages when we are unable to shrink dcache.
1708 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1712 * Large number of objects is good, but very large slabs are
1713 * currently bad for the gfp()s.
1715 if (gfporder
>= slab_max_order
)
1719 * Acceptable internal fragmentation?
1721 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1727 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1728 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1732 struct array_cache __percpu
*cpu_cache
;
1734 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1735 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1740 for_each_possible_cpu(cpu
) {
1741 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1742 entries
, batchcount
);
1748 static int __ref
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1750 if (slab_state
>= FULL
)
1751 return enable_cpucache(cachep
, gfp
);
1753 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1754 if (!cachep
->cpu_cache
)
1757 if (slab_state
== DOWN
) {
1758 /* Creation of first cache (kmem_cache). */
1759 set_up_node(kmem_cache
, CACHE_CACHE
);
1760 } else if (slab_state
== PARTIAL
) {
1761 /* For kmem_cache_node */
1762 set_up_node(cachep
, SIZE_NODE
);
1766 for_each_online_node(node
) {
1767 cachep
->node
[node
] = kmalloc_node(
1768 sizeof(struct kmem_cache_node
), gfp
, node
);
1769 BUG_ON(!cachep
->node
[node
]);
1770 kmem_cache_node_init(cachep
->node
[node
]);
1774 cachep
->node
[numa_mem_id()]->next_reap
=
1775 jiffies
+ REAPTIMEOUT_NODE
+
1776 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1778 cpu_cache_get(cachep
)->avail
= 0;
1779 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1780 cpu_cache_get(cachep
)->batchcount
= 1;
1781 cpu_cache_get(cachep
)->touched
= 0;
1782 cachep
->batchcount
= 1;
1783 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1787 slab_flags_t
kmem_cache_flags(unsigned int object_size
,
1788 slab_flags_t flags
, const char *name
)
1794 __kmem_cache_alias(const char *name
, unsigned int size
, unsigned int align
,
1795 slab_flags_t flags
, void (*ctor
)(void *))
1797 struct kmem_cache
*cachep
;
1799 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1804 * Adjust the object sizes so that we clear
1805 * the complete object on kzalloc.
1807 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1812 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1813 size_t size
, slab_flags_t flags
)
1820 * If slab auto-initialization on free is enabled, store the freelist
1821 * off-slab, so that its contents don't end up in one of the allocated
1824 if (unlikely(slab_want_init_on_free(cachep
)))
1827 if (cachep
->ctor
|| flags
& SLAB_TYPESAFE_BY_RCU
)
1830 left
= calculate_slab_order(cachep
, size
,
1831 flags
| CFLGS_OBJFREELIST_SLAB
);
1835 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1838 cachep
->colour
= left
/ cachep
->colour_off
;
1843 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1844 size_t size
, slab_flags_t flags
)
1851 * Always use on-slab management when SLAB_NOLEAKTRACE
1852 * to avoid recursive calls into kmemleak.
1854 if (flags
& SLAB_NOLEAKTRACE
)
1858 * Size is large, assume best to place the slab management obj
1859 * off-slab (should allow better packing of objs).
1861 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1866 * If the slab has been placed off-slab, and we have enough space then
1867 * move it on-slab. This is at the expense of any extra colouring.
1869 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1872 cachep
->colour
= left
/ cachep
->colour_off
;
1877 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1878 size_t size
, slab_flags_t flags
)
1884 left
= calculate_slab_order(cachep
, size
, flags
);
1888 cachep
->colour
= left
/ cachep
->colour_off
;
1894 * __kmem_cache_create - Create a cache.
1895 * @cachep: cache management descriptor
1896 * @flags: SLAB flags
1898 * Returns a ptr to the cache on success, NULL on failure.
1899 * Cannot be called within an int, but can be interrupted.
1900 * The @ctor is run when new pages are allocated by the cache.
1904 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1905 * to catch references to uninitialised memory.
1907 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1908 * for buffer overruns.
1910 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1911 * cacheline. This can be beneficial if you're counting cycles as closely
1914 * Return: a pointer to the created cache or %NULL in case of error
1916 int __kmem_cache_create(struct kmem_cache
*cachep
, slab_flags_t flags
)
1918 size_t ralign
= BYTES_PER_WORD
;
1921 unsigned int size
= cachep
->size
;
1926 * Enable redzoning and last user accounting, except for caches with
1927 * large objects, if the increased size would increase the object size
1928 * above the next power of two: caches with object sizes just above a
1929 * power of two have a significant amount of internal fragmentation.
1931 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
1932 2 * sizeof(unsigned long long)))
1933 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
1934 if (!(flags
& SLAB_TYPESAFE_BY_RCU
))
1935 flags
|= SLAB_POISON
;
1940 * Check that size is in terms of words. This is needed to avoid
1941 * unaligned accesses for some archs when redzoning is used, and makes
1942 * sure any on-slab bufctl's are also correctly aligned.
1944 size
= ALIGN(size
, BYTES_PER_WORD
);
1946 if (flags
& SLAB_RED_ZONE
) {
1947 ralign
= REDZONE_ALIGN
;
1948 /* If redzoning, ensure that the second redzone is suitably
1949 * aligned, by adjusting the object size accordingly. */
1950 size
= ALIGN(size
, REDZONE_ALIGN
);
1953 /* 3) caller mandated alignment */
1954 if (ralign
< cachep
->align
) {
1955 ralign
= cachep
->align
;
1957 /* disable debug if necessary */
1958 if (ralign
> __alignof__(unsigned long long))
1959 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
1963 cachep
->align
= ralign
;
1964 cachep
->colour_off
= cache_line_size();
1965 /* Offset must be a multiple of the alignment. */
1966 if (cachep
->colour_off
< cachep
->align
)
1967 cachep
->colour_off
= cachep
->align
;
1969 if (slab_is_available())
1977 * Both debugging options require word-alignment which is calculated
1980 if (flags
& SLAB_RED_ZONE
) {
1981 /* add space for red zone words */
1982 cachep
->obj_offset
+= sizeof(unsigned long long);
1983 size
+= 2 * sizeof(unsigned long long);
1985 if (flags
& SLAB_STORE_USER
) {
1986 /* user store requires one word storage behind the end of
1987 * the real object. But if the second red zone needs to be
1988 * aligned to 64 bits, we must allow that much space.
1990 if (flags
& SLAB_RED_ZONE
)
1991 size
+= REDZONE_ALIGN
;
1993 size
+= BYTES_PER_WORD
;
1997 kasan_cache_create(cachep
, &size
, &flags
);
1999 size
= ALIGN(size
, cachep
->align
);
2001 * We should restrict the number of objects in a slab to implement
2002 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2004 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2005 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2009 * To activate debug pagealloc, off-slab management is necessary
2010 * requirement. In early phase of initialization, small sized slab
2011 * doesn't get initialized so it would not be possible. So, we need
2012 * to check size >= 256. It guarantees that all necessary small
2013 * sized slab is initialized in current slab initialization sequence.
2015 if (debug_pagealloc_enabled_static() && (flags
& SLAB_POISON
) &&
2016 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2017 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2018 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2020 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2021 flags
|= CFLGS_OFF_SLAB
;
2022 cachep
->obj_offset
+= tmp_size
- size
;
2030 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2031 flags
|= CFLGS_OBJFREELIST_SLAB
;
2035 if (set_off_slab_cache(cachep
, size
, flags
)) {
2036 flags
|= CFLGS_OFF_SLAB
;
2040 if (set_on_slab_cache(cachep
, size
, flags
))
2046 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2047 cachep
->flags
= flags
;
2048 cachep
->allocflags
= __GFP_COMP
;
2049 if (flags
& SLAB_CACHE_DMA
)
2050 cachep
->allocflags
|= GFP_DMA
;
2051 if (flags
& SLAB_CACHE_DMA32
)
2052 cachep
->allocflags
|= GFP_DMA32
;
2053 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2054 cachep
->allocflags
|= __GFP_RECLAIMABLE
;
2055 cachep
->size
= size
;
2056 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2060 * If we're going to use the generic kernel_map_pages()
2061 * poisoning, then it's going to smash the contents of
2062 * the redzone and userword anyhow, so switch them off.
2064 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2065 (cachep
->flags
& SLAB_POISON
) &&
2066 is_debug_pagealloc_cache(cachep
))
2067 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2070 err
= setup_cpu_cache(cachep
, gfp
);
2072 __kmem_cache_release(cachep
);
2080 static void check_irq_off(void)
2082 BUG_ON(!irqs_disabled());
2085 static void check_irq_on(void)
2087 BUG_ON(irqs_disabled());
2090 static void check_mutex_acquired(void)
2092 BUG_ON(!mutex_is_locked(&slab_mutex
));
2095 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2099 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2103 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2107 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2112 #define check_irq_off() do { } while(0)
2113 #define check_irq_on() do { } while(0)
2114 #define check_mutex_acquired() do { } while(0)
2115 #define check_spinlock_acquired(x) do { } while(0)
2116 #define check_spinlock_acquired_node(x, y) do { } while(0)
2119 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2120 int node
, bool free_all
, struct list_head
*list
)
2124 if (!ac
|| !ac
->avail
)
2127 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2128 if (tofree
> ac
->avail
)
2129 tofree
= (ac
->avail
+ 1) / 2;
2131 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2132 ac
->avail
-= tofree
;
2133 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2136 static void do_drain(void *arg
)
2138 struct kmem_cache
*cachep
= arg
;
2139 struct array_cache
*ac
;
2140 int node
= numa_mem_id();
2141 struct kmem_cache_node
*n
;
2145 ac
= cpu_cache_get(cachep
);
2146 n
= get_node(cachep
, node
);
2147 spin_lock(&n
->list_lock
);
2148 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2149 spin_unlock(&n
->list_lock
);
2151 slabs_destroy(cachep
, &list
);
2154 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2156 struct kmem_cache_node
*n
;
2160 on_each_cpu(do_drain
, cachep
, 1);
2162 for_each_kmem_cache_node(cachep
, node
, n
)
2164 drain_alien_cache(cachep
, n
->alien
);
2166 for_each_kmem_cache_node(cachep
, node
, n
) {
2167 spin_lock_irq(&n
->list_lock
);
2168 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2169 spin_unlock_irq(&n
->list_lock
);
2171 slabs_destroy(cachep
, &list
);
2176 * Remove slabs from the list of free slabs.
2177 * Specify the number of slabs to drain in tofree.
2179 * Returns the actual number of slabs released.
2181 static int drain_freelist(struct kmem_cache
*cache
,
2182 struct kmem_cache_node
*n
, int tofree
)
2184 struct list_head
*p
;
2189 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2191 spin_lock_irq(&n
->list_lock
);
2192 p
= n
->slabs_free
.prev
;
2193 if (p
== &n
->slabs_free
) {
2194 spin_unlock_irq(&n
->list_lock
);
2198 slab
= list_entry(p
, struct slab
, slab_list
);
2199 list_del(&slab
->slab_list
);
2203 * Safe to drop the lock. The slab is no longer linked
2206 n
->free_objects
-= cache
->num
;
2207 spin_unlock_irq(&n
->list_lock
);
2208 slab_destroy(cache
, slab
);
2215 bool __kmem_cache_empty(struct kmem_cache
*s
)
2218 struct kmem_cache_node
*n
;
2220 for_each_kmem_cache_node(s
, node
, n
)
2221 if (!list_empty(&n
->slabs_full
) ||
2222 !list_empty(&n
->slabs_partial
))
2227 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2231 struct kmem_cache_node
*n
;
2233 drain_cpu_caches(cachep
);
2236 for_each_kmem_cache_node(cachep
, node
, n
) {
2237 drain_freelist(cachep
, n
, INT_MAX
);
2239 ret
+= !list_empty(&n
->slabs_full
) ||
2240 !list_empty(&n
->slabs_partial
);
2242 return (ret
? 1 : 0);
2245 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2247 return __kmem_cache_shrink(cachep
);
2250 void __kmem_cache_release(struct kmem_cache
*cachep
)
2253 struct kmem_cache_node
*n
;
2255 cache_random_seq_destroy(cachep
);
2257 free_percpu(cachep
->cpu_cache
);
2259 /* NUMA: free the node structures */
2260 for_each_kmem_cache_node(cachep
, i
, n
) {
2262 free_alien_cache(n
->alien
);
2264 cachep
->node
[i
] = NULL
;
2269 * Get the memory for a slab management obj.
2271 * For a slab cache when the slab descriptor is off-slab, the
2272 * slab descriptor can't come from the same cache which is being created,
2273 * Because if it is the case, that means we defer the creation of
2274 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2275 * And we eventually call down to __kmem_cache_create(), which
2276 * in turn looks up in the kmalloc_{dma,}_caches for the desired-size one.
2277 * This is a "chicken-and-egg" problem.
2279 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2280 * which are all initialized during kmem_cache_init().
2282 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2283 struct slab
*slab
, int colour_off
,
2284 gfp_t local_flags
, int nodeid
)
2287 void *addr
= slab_address(slab
);
2289 slab
->s_mem
= addr
+ colour_off
;
2292 if (OBJFREELIST_SLAB(cachep
))
2294 else if (OFF_SLAB(cachep
)) {
2295 /* Slab management obj is off-slab. */
2296 freelist
= kmalloc_node(cachep
->freelist_size
,
2297 local_flags
, nodeid
);
2299 /* We will use last bytes at the slab for freelist */
2300 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2301 cachep
->freelist_size
;
2307 static inline freelist_idx_t
get_free_obj(struct slab
*slab
, unsigned int idx
)
2309 return ((freelist_idx_t
*) slab
->freelist
)[idx
];
2312 static inline void set_free_obj(struct slab
*slab
,
2313 unsigned int idx
, freelist_idx_t val
)
2315 ((freelist_idx_t
*)(slab
->freelist
))[idx
] = val
;
2318 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct slab
*slab
)
2323 for (i
= 0; i
< cachep
->num
; i
++) {
2324 void *objp
= index_to_obj(cachep
, slab
, i
);
2326 if (cachep
->flags
& SLAB_STORE_USER
)
2327 *dbg_userword(cachep
, objp
) = NULL
;
2329 if (cachep
->flags
& SLAB_RED_ZONE
) {
2330 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2331 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2334 * Constructors are not allowed to allocate memory from the same
2335 * cache which they are a constructor for. Otherwise, deadlock.
2336 * They must also be threaded.
2338 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2339 kasan_unpoison_object_data(cachep
,
2340 objp
+ obj_offset(cachep
));
2341 cachep
->ctor(objp
+ obj_offset(cachep
));
2342 kasan_poison_object_data(
2343 cachep
, objp
+ obj_offset(cachep
));
2346 if (cachep
->flags
& SLAB_RED_ZONE
) {
2347 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2348 slab_error(cachep
, "constructor overwrote the end of an object");
2349 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2350 slab_error(cachep
, "constructor overwrote the start of an object");
2352 /* need to poison the objs? */
2353 if (cachep
->flags
& SLAB_POISON
) {
2354 poison_obj(cachep
, objp
, POISON_FREE
);
2355 slab_kernel_map(cachep
, objp
, 0);
2361 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2362 /* Hold information during a freelist initialization */
2363 union freelist_init_state
{
2369 struct rnd_state rnd_state
;
2373 * Initialize the state based on the randomization method available.
2374 * return true if the pre-computed list is available, false otherwise.
2376 static bool freelist_state_initialize(union freelist_init_state
*state
,
2377 struct kmem_cache
*cachep
,
2383 /* Use best entropy available to define a random shift */
2384 rand
= get_random_u32();
2386 /* Use a random state if the pre-computed list is not available */
2387 if (!cachep
->random_seq
) {
2388 prandom_seed_state(&state
->rnd_state
, rand
);
2391 state
->list
= cachep
->random_seq
;
2392 state
->count
= count
;
2393 state
->pos
= rand
% count
;
2399 /* Get the next entry on the list and randomize it using a random shift */
2400 static freelist_idx_t
next_random_slot(union freelist_init_state
*state
)
2402 if (state
->pos
>= state
->count
)
2404 return state
->list
[state
->pos
++];
2407 /* Swap two freelist entries */
2408 static void swap_free_obj(struct slab
*slab
, unsigned int a
, unsigned int b
)
2410 swap(((freelist_idx_t
*) slab
->freelist
)[a
],
2411 ((freelist_idx_t
*) slab
->freelist
)[b
]);
2415 * Shuffle the freelist initialization state based on pre-computed lists.
2416 * return true if the list was successfully shuffled, false otherwise.
2418 static bool shuffle_freelist(struct kmem_cache
*cachep
, struct slab
*slab
)
2420 unsigned int objfreelist
= 0, i
, rand
, count
= cachep
->num
;
2421 union freelist_init_state state
;
2427 precomputed
= freelist_state_initialize(&state
, cachep
, count
);
2429 /* Take a random entry as the objfreelist */
2430 if (OBJFREELIST_SLAB(cachep
)) {
2432 objfreelist
= count
- 1;
2434 objfreelist
= next_random_slot(&state
);
2435 slab
->freelist
= index_to_obj(cachep
, slab
, objfreelist
) +
2441 * On early boot, generate the list dynamically.
2442 * Later use a pre-computed list for speed.
2445 for (i
= 0; i
< count
; i
++)
2446 set_free_obj(slab
, i
, i
);
2448 /* Fisher-Yates shuffle */
2449 for (i
= count
- 1; i
> 0; i
--) {
2450 rand
= prandom_u32_state(&state
.rnd_state
);
2452 swap_free_obj(slab
, i
, rand
);
2455 for (i
= 0; i
< count
; i
++)
2456 set_free_obj(slab
, i
, next_random_slot(&state
));
2459 if (OBJFREELIST_SLAB(cachep
))
2460 set_free_obj(slab
, cachep
->num
- 1, objfreelist
);
2465 static inline bool shuffle_freelist(struct kmem_cache
*cachep
,
2470 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2472 static void cache_init_objs(struct kmem_cache
*cachep
,
2479 cache_init_objs_debug(cachep
, slab
);
2481 /* Try to randomize the freelist if enabled */
2482 shuffled
= shuffle_freelist(cachep
, slab
);
2484 if (!shuffled
&& OBJFREELIST_SLAB(cachep
)) {
2485 slab
->freelist
= index_to_obj(cachep
, slab
, cachep
->num
- 1) +
2489 for (i
= 0; i
< cachep
->num
; i
++) {
2490 objp
= index_to_obj(cachep
, slab
, i
);
2491 objp
= kasan_init_slab_obj(cachep
, objp
);
2493 /* constructor could break poison info */
2494 if (DEBUG
== 0 && cachep
->ctor
) {
2495 kasan_unpoison_object_data(cachep
, objp
);
2497 kasan_poison_object_data(cachep
, objp
);
2501 set_free_obj(slab
, i
, i
);
2505 static void *slab_get_obj(struct kmem_cache
*cachep
, struct slab
*slab
)
2509 objp
= index_to_obj(cachep
, slab
, get_free_obj(slab
, slab
->active
));
2515 static void slab_put_obj(struct kmem_cache
*cachep
,
2516 struct slab
*slab
, void *objp
)
2518 unsigned int objnr
= obj_to_index(cachep
, slab
, objp
);
2522 /* Verify double free bug */
2523 for (i
= slab
->active
; i
< cachep
->num
; i
++) {
2524 if (get_free_obj(slab
, i
) == objnr
) {
2525 pr_err("slab: double free detected in cache '%s', objp %px\n",
2526 cachep
->name
, objp
);
2532 if (!slab
->freelist
)
2533 slab
->freelist
= objp
+ obj_offset(cachep
);
2535 set_free_obj(slab
, slab
->active
, objnr
);
2539 * Grow (by 1) the number of slabs within a cache. This is called by
2540 * kmem_cache_alloc() when there are no active objs left in a cache.
2542 static struct slab
*cache_grow_begin(struct kmem_cache
*cachep
,
2543 gfp_t flags
, int nodeid
)
2549 struct kmem_cache_node
*n
;
2553 * Be lazy and only check for valid flags here, keeping it out of the
2554 * critical path in kmem_cache_alloc().
2556 if (unlikely(flags
& GFP_SLAB_BUG_MASK
))
2557 flags
= kmalloc_fix_flags(flags
);
2559 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2560 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2563 if (gfpflags_allow_blocking(local_flags
))
2567 * Get mem for the objs. Attempt to allocate a physical page from
2570 slab
= kmem_getpages(cachep
, local_flags
, nodeid
);
2574 slab_node
= slab_nid(slab
);
2575 n
= get_node(cachep
, slab_node
);
2577 /* Get colour for the slab, and cal the next value. */
2579 if (n
->colour_next
>= cachep
->colour
)
2582 offset
= n
->colour_next
;
2583 if (offset
>= cachep
->colour
)
2586 offset
*= cachep
->colour_off
;
2589 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2590 * page_address() in the latter returns a non-tagged pointer,
2591 * as it should be for slab pages.
2593 kasan_poison_slab(slab
);
2595 /* Get slab management. */
2596 freelist
= alloc_slabmgmt(cachep
, slab
, offset
,
2597 local_flags
& ~GFP_CONSTRAINT_MASK
, slab_node
);
2598 if (OFF_SLAB(cachep
) && !freelist
)
2601 slab
->slab_cache
= cachep
;
2602 slab
->freelist
= freelist
;
2604 cache_init_objs(cachep
, slab
);
2606 if (gfpflags_allow_blocking(local_flags
))
2607 local_irq_disable();
2612 kmem_freepages(cachep
, slab
);
2614 if (gfpflags_allow_blocking(local_flags
))
2615 local_irq_disable();
2619 static void cache_grow_end(struct kmem_cache
*cachep
, struct slab
*slab
)
2621 struct kmem_cache_node
*n
;
2629 INIT_LIST_HEAD(&slab
->slab_list
);
2630 n
= get_node(cachep
, slab_nid(slab
));
2632 spin_lock(&n
->list_lock
);
2634 if (!slab
->active
) {
2635 list_add_tail(&slab
->slab_list
, &n
->slabs_free
);
2638 fixup_slab_list(cachep
, n
, slab
, &list
);
2640 STATS_INC_GROWN(cachep
);
2641 n
->free_objects
+= cachep
->num
- slab
->active
;
2642 spin_unlock(&n
->list_lock
);
2644 fixup_objfreelist_debug(cachep
, &list
);
2650 * Perform extra freeing checks:
2651 * - detect bad pointers.
2652 * - POISON/RED_ZONE checking
2654 static void kfree_debugcheck(const void *objp
)
2656 if (!virt_addr_valid(objp
)) {
2657 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2658 (unsigned long)objp
);
2663 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2665 unsigned long long redzone1
, redzone2
;
2667 redzone1
= *dbg_redzone1(cache
, obj
);
2668 redzone2
= *dbg_redzone2(cache
, obj
);
2673 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2676 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2677 slab_error(cache
, "double free detected");
2679 slab_error(cache
, "memory outside object was overwritten");
2681 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2682 obj
, redzone1
, redzone2
);
2685 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2686 unsigned long caller
)
2691 BUG_ON(virt_to_cache(objp
) != cachep
);
2693 objp
-= obj_offset(cachep
);
2694 kfree_debugcheck(objp
);
2695 slab
= virt_to_slab(objp
);
2697 if (cachep
->flags
& SLAB_RED_ZONE
) {
2698 verify_redzone_free(cachep
, objp
);
2699 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2700 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2702 if (cachep
->flags
& SLAB_STORE_USER
)
2703 *dbg_userword(cachep
, objp
) = (void *)caller
;
2705 objnr
= obj_to_index(cachep
, slab
, objp
);
2707 BUG_ON(objnr
>= cachep
->num
);
2708 BUG_ON(objp
!= index_to_obj(cachep
, slab
, objnr
));
2710 if (cachep
->flags
& SLAB_POISON
) {
2711 poison_obj(cachep
, objp
, POISON_FREE
);
2712 slab_kernel_map(cachep
, objp
, 0);
2718 #define kfree_debugcheck(x) do { } while(0)
2719 #define cache_free_debugcheck(x, objp, z) (objp)
2722 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2730 objp
= next
- obj_offset(cachep
);
2731 next
= *(void **)next
;
2732 poison_obj(cachep
, objp
, POISON_FREE
);
2737 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2738 struct kmem_cache_node
*n
, struct slab
*slab
,
2741 /* move slabp to correct slabp list: */
2742 list_del(&slab
->slab_list
);
2743 if (slab
->active
== cachep
->num
) {
2744 list_add(&slab
->slab_list
, &n
->slabs_full
);
2745 if (OBJFREELIST_SLAB(cachep
)) {
2747 /* Poisoning will be done without holding the lock */
2748 if (cachep
->flags
& SLAB_POISON
) {
2749 void **objp
= slab
->freelist
;
2755 slab
->freelist
= NULL
;
2758 list_add(&slab
->slab_list
, &n
->slabs_partial
);
2761 /* Try to find non-pfmemalloc slab if needed */
2762 static noinline
struct slab
*get_valid_first_slab(struct kmem_cache_node
*n
,
2763 struct slab
*slab
, bool pfmemalloc
)
2771 if (!slab_test_pfmemalloc(slab
))
2774 /* No need to keep pfmemalloc slab if we have enough free objects */
2775 if (n
->free_objects
> n
->free_limit
) {
2776 slab_clear_pfmemalloc(slab
);
2780 /* Move pfmemalloc slab to the end of list to speed up next search */
2781 list_del(&slab
->slab_list
);
2782 if (!slab
->active
) {
2783 list_add_tail(&slab
->slab_list
, &n
->slabs_free
);
2786 list_add_tail(&slab
->slab_list
, &n
->slabs_partial
);
2788 list_for_each_entry(slab
, &n
->slabs_partial
, slab_list
) {
2789 if (!slab_test_pfmemalloc(slab
))
2793 n
->free_touched
= 1;
2794 list_for_each_entry(slab
, &n
->slabs_free
, slab_list
) {
2795 if (!slab_test_pfmemalloc(slab
)) {
2804 static struct slab
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2808 assert_spin_locked(&n
->list_lock
);
2809 slab
= list_first_entry_or_null(&n
->slabs_partial
, struct slab
,
2812 n
->free_touched
= 1;
2813 slab
= list_first_entry_or_null(&n
->slabs_free
, struct slab
,
2819 if (sk_memalloc_socks())
2820 slab
= get_valid_first_slab(n
, slab
, pfmemalloc
);
2825 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2826 struct kmem_cache_node
*n
, gfp_t flags
)
2832 if (!gfp_pfmemalloc_allowed(flags
))
2835 spin_lock(&n
->list_lock
);
2836 slab
= get_first_slab(n
, true);
2838 spin_unlock(&n
->list_lock
);
2842 obj
= slab_get_obj(cachep
, slab
);
2845 fixup_slab_list(cachep
, n
, slab
, &list
);
2847 spin_unlock(&n
->list_lock
);
2848 fixup_objfreelist_debug(cachep
, &list
);
2854 * Slab list should be fixed up by fixup_slab_list() for existing slab
2855 * or cache_grow_end() for new slab
2857 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2858 struct array_cache
*ac
, struct slab
*slab
, int batchcount
)
2861 * There must be at least one object available for
2864 BUG_ON(slab
->active
>= cachep
->num
);
2866 while (slab
->active
< cachep
->num
&& batchcount
--) {
2867 STATS_INC_ALLOCED(cachep
);
2868 STATS_INC_ACTIVE(cachep
);
2869 STATS_SET_HIGH(cachep
);
2871 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, slab
);
2877 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2880 struct kmem_cache_node
*n
;
2881 struct array_cache
*ac
, *shared
;
2887 node
= numa_mem_id();
2889 ac
= cpu_cache_get(cachep
);
2890 batchcount
= ac
->batchcount
;
2891 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
2893 * If there was little recent activity on this cache, then
2894 * perform only a partial refill. Otherwise we could generate
2897 batchcount
= BATCHREFILL_LIMIT
;
2899 n
= get_node(cachep
, node
);
2901 BUG_ON(ac
->avail
> 0 || !n
);
2902 shared
= READ_ONCE(n
->shared
);
2903 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
2906 spin_lock(&n
->list_lock
);
2907 shared
= READ_ONCE(n
->shared
);
2909 /* See if we can refill from the shared array */
2910 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
2911 shared
->touched
= 1;
2915 while (batchcount
> 0) {
2916 /* Get slab alloc is to come from. */
2917 slab
= get_first_slab(n
, false);
2921 check_spinlock_acquired(cachep
);
2923 batchcount
= alloc_block(cachep
, ac
, slab
, batchcount
);
2924 fixup_slab_list(cachep
, n
, slab
, &list
);
2928 n
->free_objects
-= ac
->avail
;
2930 spin_unlock(&n
->list_lock
);
2931 fixup_objfreelist_debug(cachep
, &list
);
2934 if (unlikely(!ac
->avail
)) {
2935 /* Check if we can use obj in pfmemalloc slab */
2936 if (sk_memalloc_socks()) {
2937 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
2943 slab
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
2946 * cache_grow_begin() can reenable interrupts,
2947 * then ac could change.
2949 ac
= cpu_cache_get(cachep
);
2950 if (!ac
->avail
&& slab
)
2951 alloc_block(cachep
, ac
, slab
, batchcount
);
2952 cache_grow_end(cachep
, slab
);
2959 return ac
->entry
[--ac
->avail
];
2963 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
2964 gfp_t flags
, void *objp
, unsigned long caller
)
2966 WARN_ON_ONCE(cachep
->ctor
&& (flags
& __GFP_ZERO
));
2967 if (!objp
|| is_kfence_address(objp
))
2969 if (cachep
->flags
& SLAB_POISON
) {
2970 check_poison_obj(cachep
, objp
);
2971 slab_kernel_map(cachep
, objp
, 1);
2972 poison_obj(cachep
, objp
, POISON_INUSE
);
2974 if (cachep
->flags
& SLAB_STORE_USER
)
2975 *dbg_userword(cachep
, objp
) = (void *)caller
;
2977 if (cachep
->flags
& SLAB_RED_ZONE
) {
2978 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
2979 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
2980 slab_error(cachep
, "double free, or memory outside object was overwritten");
2981 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2982 objp
, *dbg_redzone1(cachep
, objp
),
2983 *dbg_redzone2(cachep
, objp
));
2985 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
2986 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
2989 objp
+= obj_offset(cachep
);
2990 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
2992 if ((unsigned long)objp
& (arch_slab_minalign() - 1)) {
2993 pr_err("0x%px: not aligned to arch_slab_minalign()=%u\n", objp
,
2994 arch_slab_minalign());
2999 #define cache_alloc_debugcheck_after(a, b, objp, d) (objp)
3002 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3005 struct array_cache
*ac
;
3009 ac
= cpu_cache_get(cachep
);
3010 if (likely(ac
->avail
)) {
3012 objp
= ac
->entry
[--ac
->avail
];
3014 STATS_INC_ALLOCHIT(cachep
);
3018 STATS_INC_ALLOCMISS(cachep
);
3019 objp
= cache_alloc_refill(cachep
, flags
);
3021 * the 'ac' may be updated by cache_alloc_refill(),
3022 * and kmemleak_erase() requires its correct value.
3024 ac
= cpu_cache_get(cachep
);
3028 * To avoid a false negative, if an object that is in one of the
3029 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3030 * treat the array pointers as a reference to the object.
3033 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3038 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
3041 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3043 * If we are in_interrupt, then process context, including cpusets and
3044 * mempolicy, may not apply and should not be used for allocation policy.
3046 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3048 int nid_alloc
, nid_here
;
3050 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3052 nid_alloc
= nid_here
= numa_mem_id();
3053 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3054 nid_alloc
= cpuset_slab_spread_node();
3055 else if (current
->mempolicy
)
3056 nid_alloc
= mempolicy_slab_node();
3057 if (nid_alloc
!= nid_here
)
3058 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3063 * Fallback function if there was no memory available and no objects on a
3064 * certain node and fall back is permitted. First we scan all the
3065 * available node for available objects. If that fails then we
3066 * perform an allocation without specifying a node. This allows the page
3067 * allocator to do its reclaim / fallback magic. We then insert the
3068 * slab into the proper nodelist and then allocate from it.
3070 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3072 struct zonelist
*zonelist
;
3075 enum zone_type highest_zoneidx
= gfp_zone(flags
);
3079 unsigned int cpuset_mems_cookie
;
3081 if (flags
& __GFP_THISNODE
)
3085 cpuset_mems_cookie
= read_mems_allowed_begin();
3086 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3090 * Look through allowed nodes for objects available
3091 * from existing per node queues.
3093 for_each_zone_zonelist(zone
, z
, zonelist
, highest_zoneidx
) {
3094 nid
= zone_to_nid(zone
);
3096 if (cpuset_zone_allowed(zone
, flags
) &&
3097 get_node(cache
, nid
) &&
3098 get_node(cache
, nid
)->free_objects
) {
3099 obj
= ____cache_alloc_node(cache
,
3100 gfp_exact_node(flags
), nid
);
3108 * This allocation will be performed within the constraints
3109 * of the current cpuset / memory policy requirements.
3110 * We may trigger various forms of reclaim on the allowed
3111 * set and go into memory reserves if necessary.
3113 slab
= cache_grow_begin(cache
, flags
, numa_mem_id());
3114 cache_grow_end(cache
, slab
);
3116 nid
= slab_nid(slab
);
3117 obj
= ____cache_alloc_node(cache
,
3118 gfp_exact_node(flags
), nid
);
3121 * Another processor may allocate the objects in
3122 * the slab since we are not holding any locks.
3129 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3135 * An interface to enable slab creation on nodeid
3137 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3141 struct kmem_cache_node
*n
;
3145 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3146 n
= get_node(cachep
, nodeid
);
3150 spin_lock(&n
->list_lock
);
3151 slab
= get_first_slab(n
, false);
3155 check_spinlock_acquired_node(cachep
, nodeid
);
3157 STATS_INC_NODEALLOCS(cachep
);
3158 STATS_INC_ACTIVE(cachep
);
3159 STATS_SET_HIGH(cachep
);
3161 BUG_ON(slab
->active
== cachep
->num
);
3163 obj
= slab_get_obj(cachep
, slab
);
3166 fixup_slab_list(cachep
, n
, slab
, &list
);
3168 spin_unlock(&n
->list_lock
);
3169 fixup_objfreelist_debug(cachep
, &list
);
3173 spin_unlock(&n
->list_lock
);
3174 slab
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3176 /* This slab isn't counted yet so don't update free_objects */
3177 obj
= slab_get_obj(cachep
, slab
);
3179 cache_grow_end(cachep
, slab
);
3181 return obj
? obj
: fallback_alloc(cachep
, flags
);
3184 static __always_inline
void *
3185 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3188 int slab_node
= numa_mem_id();
3190 if (nodeid
== NUMA_NO_NODE
) {
3191 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3192 objp
= alternate_node_alloc(cachep
, flags
);
3197 * Use the locally cached objects if possible.
3198 * However ____cache_alloc does not allow fallback
3199 * to other nodes. It may fail while we still have
3200 * objects on other nodes available.
3202 objp
= ____cache_alloc(cachep
, flags
);
3204 } else if (nodeid
== slab_node
) {
3205 objp
= ____cache_alloc(cachep
, flags
);
3206 } else if (!get_node(cachep
, nodeid
)) {
3207 /* Node not bootstrapped yet */
3208 objp
= fallback_alloc(cachep
, flags
);
3213 * We may just have run out of memory on the local node.
3214 * ____cache_alloc_node() knows how to locate memory on other nodes
3217 objp
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3223 static __always_inline
void *
3224 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid __maybe_unused
)
3226 return ____cache_alloc(cachep
, flags
);
3229 #endif /* CONFIG_NUMA */
3231 static __always_inline
void *
3232 slab_alloc_node(struct kmem_cache
*cachep
, struct list_lru
*lru
, gfp_t flags
,
3233 int nodeid
, size_t orig_size
, unsigned long caller
)
3235 unsigned long save_flags
;
3237 struct obj_cgroup
*objcg
= NULL
;
3240 flags
&= gfp_allowed_mask
;
3241 cachep
= slab_pre_alloc_hook(cachep
, lru
, &objcg
, 1, flags
);
3242 if (unlikely(!cachep
))
3245 objp
= kfence_alloc(cachep
, orig_size
, flags
);
3249 local_irq_save(save_flags
);
3250 objp
= __do_cache_alloc(cachep
, flags
, nodeid
);
3251 local_irq_restore(save_flags
);
3252 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3254 init
= slab_want_init_on_alloc(flags
, cachep
);
3257 slab_post_alloc_hook(cachep
, objcg
, flags
, 1, &objp
, init
);
3261 static __always_inline
void *
3262 slab_alloc(struct kmem_cache
*cachep
, struct list_lru
*lru
, gfp_t flags
,
3263 size_t orig_size
, unsigned long caller
)
3265 return slab_alloc_node(cachep
, lru
, flags
, NUMA_NO_NODE
, orig_size
,
3270 * Caller needs to acquire correct kmem_cache_node's list_lock
3271 * @list: List of detached free slabs should be freed by caller
3273 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3274 int nr_objects
, int node
, struct list_head
*list
)
3277 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3280 n
->free_objects
+= nr_objects
;
3282 for (i
= 0; i
< nr_objects
; i
++) {
3288 slab
= virt_to_slab(objp
);
3289 list_del(&slab
->slab_list
);
3290 check_spinlock_acquired_node(cachep
, node
);
3291 slab_put_obj(cachep
, slab
, objp
);
3292 STATS_DEC_ACTIVE(cachep
);
3294 /* fixup slab chains */
3295 if (slab
->active
== 0) {
3296 list_add(&slab
->slab_list
, &n
->slabs_free
);
3299 /* Unconditionally move a slab to the end of the
3300 * partial list on free - maximum time for the
3301 * other objects to be freed, too.
3303 list_add_tail(&slab
->slab_list
, &n
->slabs_partial
);
3307 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3308 n
->free_objects
-= cachep
->num
;
3310 slab
= list_last_entry(&n
->slabs_free
, struct slab
, slab_list
);
3311 list_move(&slab
->slab_list
, list
);
3317 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3320 struct kmem_cache_node
*n
;
3321 int node
= numa_mem_id();
3324 batchcount
= ac
->batchcount
;
3327 n
= get_node(cachep
, node
);
3328 spin_lock(&n
->list_lock
);
3330 struct array_cache
*shared_array
= n
->shared
;
3331 int max
= shared_array
->limit
- shared_array
->avail
;
3333 if (batchcount
> max
)
3335 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3336 ac
->entry
, sizeof(void *) * batchcount
);
3337 shared_array
->avail
+= batchcount
;
3342 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3349 list_for_each_entry(slab
, &n
->slabs_free
, slab_list
) {
3350 BUG_ON(slab
->active
);
3354 STATS_SET_FREEABLE(cachep
, i
);
3357 spin_unlock(&n
->list_lock
);
3358 ac
->avail
-= batchcount
;
3359 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3360 slabs_destroy(cachep
, &list
);
3364 * Release an obj back to its cache. If the obj has a constructed state, it must
3365 * be in this state _before_ it is released. Called with disabled ints.
3367 static __always_inline
void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3368 unsigned long caller
)
3372 memcg_slab_free_hook(cachep
, virt_to_slab(objp
), &objp
, 1);
3374 if (is_kfence_address(objp
)) {
3375 kmemleak_free_recursive(objp
, cachep
->flags
);
3376 __kfence_free(objp
);
3381 * As memory initialization might be integrated into KASAN,
3382 * kasan_slab_free and initialization memset must be
3383 * kept together to avoid discrepancies in behavior.
3385 init
= slab_want_init_on_free(cachep
);
3386 if (init
&& !kasan_has_integrated_init())
3387 memset(objp
, 0, cachep
->object_size
);
3388 /* KASAN might put objp into memory quarantine, delaying its reuse. */
3389 if (kasan_slab_free(cachep
, objp
, init
))
3392 /* Use KCSAN to help debug racy use-after-free. */
3393 if (!(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
3394 __kcsan_check_access(objp
, cachep
->object_size
,
3395 KCSAN_ACCESS_WRITE
| KCSAN_ACCESS_ASSERT
);
3397 ___cache_free(cachep
, objp
, caller
);
3400 void ___cache_free(struct kmem_cache
*cachep
, void *objp
,
3401 unsigned long caller
)
3403 struct array_cache
*ac
= cpu_cache_get(cachep
);
3406 kmemleak_free_recursive(objp
, cachep
->flags
);
3407 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3410 * Skip calling cache_free_alien() when the platform is not numa.
3411 * This will avoid cache misses that happen while accessing slabp (which
3412 * is per page memory reference) to get nodeid. Instead use a global
3413 * variable to skip the call, which is mostly likely to be present in
3416 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3419 if (ac
->avail
< ac
->limit
) {
3420 STATS_INC_FREEHIT(cachep
);
3422 STATS_INC_FREEMISS(cachep
);
3423 cache_flusharray(cachep
, ac
);
3426 if (sk_memalloc_socks()) {
3427 struct slab
*slab
= virt_to_slab(objp
);
3429 if (unlikely(slab_test_pfmemalloc(slab
))) {
3430 cache_free_pfmemalloc(cachep
, slab
, objp
);
3435 __free_one(ac
, objp
);
3438 static __always_inline
3439 void *__kmem_cache_alloc_lru(struct kmem_cache
*cachep
, struct list_lru
*lru
,
3442 void *ret
= slab_alloc(cachep
, lru
, flags
, cachep
->object_size
, _RET_IP_
);
3444 trace_kmem_cache_alloc(_RET_IP_
, ret
, cachep
, flags
, NUMA_NO_NODE
);
3450 * kmem_cache_alloc - Allocate an object
3451 * @cachep: The cache to allocate from.
3452 * @flags: See kmalloc().
3454 * Allocate an object from this cache. The flags are only relevant
3455 * if the cache has no available objects.
3457 * Return: pointer to the new object or %NULL in case of error
3459 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3461 return __kmem_cache_alloc_lru(cachep
, NULL
, flags
);
3463 EXPORT_SYMBOL(kmem_cache_alloc
);
3465 void *kmem_cache_alloc_lru(struct kmem_cache
*cachep
, struct list_lru
*lru
,
3468 return __kmem_cache_alloc_lru(cachep
, lru
, flags
);
3470 EXPORT_SYMBOL(kmem_cache_alloc_lru
);
3472 static __always_inline
void
3473 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3474 size_t size
, void **p
, unsigned long caller
)
3478 for (i
= 0; i
< size
; i
++)
3479 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3482 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3486 struct obj_cgroup
*objcg
= NULL
;
3488 s
= slab_pre_alloc_hook(s
, NULL
, &objcg
, size
, flags
);
3492 local_irq_disable();
3493 for (i
= 0; i
< size
; i
++) {
3494 void *objp
= kfence_alloc(s
, s
->object_size
, flags
) ?:
3495 __do_cache_alloc(s
, flags
, NUMA_NO_NODE
);
3497 if (unlikely(!objp
))
3503 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3506 * memcg and kmem_cache debug support and memory initialization.
3507 * Done outside of the IRQ disabled section.
3509 slab_post_alloc_hook(s
, objcg
, flags
, size
, p
,
3510 slab_want_init_on_alloc(flags
, s
));
3511 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3515 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3516 slab_post_alloc_hook(s
, objcg
, flags
, i
, p
, false);
3517 kmem_cache_free_bulk(s
, i
, p
);
3520 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3523 * kmem_cache_alloc_node - Allocate an object on the specified node
3524 * @cachep: The cache to allocate from.
3525 * @flags: See kmalloc().
3526 * @nodeid: node number of the target node.
3528 * Identical to kmem_cache_alloc but it will allocate memory on the given
3529 * node, which can improve the performance for cpu bound structures.
3531 * Fallback to other node is possible if __GFP_THISNODE is not set.
3533 * Return: pointer to the new object or %NULL in case of error
3535 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3537 void *ret
= slab_alloc_node(cachep
, NULL
, flags
, nodeid
, cachep
->object_size
, _RET_IP_
);
3539 trace_kmem_cache_alloc(_RET_IP_
, ret
, cachep
, flags
, nodeid
);
3543 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3545 void *__kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3546 int nodeid
, size_t orig_size
,
3547 unsigned long caller
)
3549 return slab_alloc_node(cachep
, NULL
, flags
, nodeid
,
3553 #ifdef CONFIG_PRINTK
3554 void __kmem_obj_info(struct kmem_obj_info
*kpp
, void *object
, struct slab
*slab
)
3556 struct kmem_cache
*cachep
;
3560 kpp
->kp_ptr
= object
;
3561 kpp
->kp_slab
= slab
;
3562 cachep
= slab
->slab_cache
;
3563 kpp
->kp_slab_cache
= cachep
;
3564 objp
= object
- obj_offset(cachep
);
3565 kpp
->kp_data_offset
= obj_offset(cachep
);
3566 slab
= virt_to_slab(objp
);
3567 objnr
= obj_to_index(cachep
, slab
, objp
);
3568 objp
= index_to_obj(cachep
, slab
, objnr
);
3569 kpp
->kp_objp
= objp
;
3570 if (DEBUG
&& cachep
->flags
& SLAB_STORE_USER
)
3571 kpp
->kp_ret
= *dbg_userword(cachep
, objp
);
3575 static __always_inline
3576 void __do_kmem_cache_free(struct kmem_cache
*cachep
, void *objp
,
3577 unsigned long caller
)
3579 unsigned long flags
;
3581 local_irq_save(flags
);
3582 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3583 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3584 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3585 __cache_free(cachep
, objp
, caller
);
3586 local_irq_restore(flags
);
3589 void __kmem_cache_free(struct kmem_cache
*cachep
, void *objp
,
3590 unsigned long caller
)
3592 __do_kmem_cache_free(cachep
, objp
, caller
);
3596 * kmem_cache_free - Deallocate an object
3597 * @cachep: The cache the allocation was from.
3598 * @objp: The previously allocated object.
3600 * Free an object which was previously allocated from this
3603 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3605 cachep
= cache_from_obj(cachep
, objp
);
3609 trace_kmem_cache_free(_RET_IP_
, objp
, cachep
);
3610 __do_kmem_cache_free(cachep
, objp
, _RET_IP_
);
3612 EXPORT_SYMBOL(kmem_cache_free
);
3614 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3617 local_irq_disable();
3618 for (int i
= 0; i
< size
; i
++) {
3620 struct kmem_cache
*s
;
3623 struct folio
*folio
= virt_to_folio(objp
);
3625 /* called via kfree_bulk */
3626 if (!folio_test_slab(folio
)) {
3628 free_large_kmalloc(folio
, objp
);
3629 local_irq_disable();
3632 s
= folio_slab(folio
)->slab_cache
;
3634 s
= cache_from_obj(orig_s
, objp
);
3640 debug_check_no_locks_freed(objp
, s
->object_size
);
3641 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3642 debug_check_no_obj_freed(objp
, s
->object_size
);
3644 __cache_free(s
, objp
, _RET_IP_
);
3648 /* FIXME: add tracing */
3650 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3653 * This initializes kmem_cache_node or resizes various caches for all nodes.
3655 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3659 struct kmem_cache_node
*n
;
3661 for_each_online_node(node
) {
3662 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3671 if (!cachep
->list
.next
) {
3672 /* Cache is not active yet. Roll back what we did */
3675 n
= get_node(cachep
, node
);
3678 free_alien_cache(n
->alien
);
3680 cachep
->node
[node
] = NULL
;
3688 /* Always called with the slab_mutex held */
3689 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3690 int batchcount
, int shared
, gfp_t gfp
)
3692 struct array_cache __percpu
*cpu_cache
, *prev
;
3695 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3699 prev
= cachep
->cpu_cache
;
3700 cachep
->cpu_cache
= cpu_cache
;
3702 * Without a previous cpu_cache there's no need to synchronize remote
3703 * cpus, so skip the IPIs.
3706 kick_all_cpus_sync();
3709 cachep
->batchcount
= batchcount
;
3710 cachep
->limit
= limit
;
3711 cachep
->shared
= shared
;
3716 for_each_online_cpu(cpu
) {
3719 struct kmem_cache_node
*n
;
3720 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3722 node
= cpu_to_mem(cpu
);
3723 n
= get_node(cachep
, node
);
3724 spin_lock_irq(&n
->list_lock
);
3725 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3726 spin_unlock_irq(&n
->list_lock
);
3727 slabs_destroy(cachep
, &list
);
3732 return setup_kmem_cache_nodes(cachep
, gfp
);
3735 /* Called with slab_mutex held always */
3736 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3743 err
= cache_random_seq_create(cachep
, cachep
->num
, gfp
);
3748 * The head array serves three purposes:
3749 * - create a LIFO ordering, i.e. return objects that are cache-warm
3750 * - reduce the number of spinlock operations.
3751 * - reduce the number of linked list operations on the slab and
3752 * bufctl chains: array operations are cheaper.
3753 * The numbers are guessed, we should auto-tune as described by
3756 if (cachep
->size
> 131072)
3758 else if (cachep
->size
> PAGE_SIZE
)
3760 else if (cachep
->size
> 1024)
3762 else if (cachep
->size
> 256)
3768 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3769 * allocation behaviour: Most allocs on one cpu, most free operations
3770 * on another cpu. For these cases, an efficient object passing between
3771 * cpus is necessary. This is provided by a shared array. The array
3772 * replaces Bonwick's magazine layer.
3773 * On uniprocessor, it's functionally equivalent (but less efficient)
3774 * to a larger limit. Thus disabled by default.
3777 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3782 * With debugging enabled, large batchcount lead to excessively long
3783 * periods with disabled local interrupts. Limit the batchcount
3788 batchcount
= (limit
+ 1) / 2;
3789 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3792 pr_err("enable_cpucache failed for %s, error %d\n",
3793 cachep
->name
, -err
);
3798 * Drain an array if it contains any elements taking the node lock only if
3799 * necessary. Note that the node listlock also protects the array_cache
3800 * if drain_array() is used on the shared array.
3802 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
3803 struct array_cache
*ac
, int node
)
3807 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3808 check_mutex_acquired();
3810 if (!ac
|| !ac
->avail
)
3818 spin_lock_irq(&n
->list_lock
);
3819 drain_array_locked(cachep
, ac
, node
, false, &list
);
3820 spin_unlock_irq(&n
->list_lock
);
3822 slabs_destroy(cachep
, &list
);
3826 * cache_reap - Reclaim memory from caches.
3827 * @w: work descriptor
3829 * Called from workqueue/eventd every few seconds.
3831 * - clear the per-cpu caches for this CPU.
3832 * - return freeable pages to the main free memory pool.
3834 * If we cannot acquire the cache chain mutex then just give up - we'll try
3835 * again on the next iteration.
3837 static void cache_reap(struct work_struct
*w
)
3839 struct kmem_cache
*searchp
;
3840 struct kmem_cache_node
*n
;
3841 int node
= numa_mem_id();
3842 struct delayed_work
*work
= to_delayed_work(w
);
3844 if (!mutex_trylock(&slab_mutex
))
3845 /* Give up. Setup the next iteration. */
3848 list_for_each_entry(searchp
, &slab_caches
, list
) {
3852 * We only take the node lock if absolutely necessary and we
3853 * have established with reasonable certainty that
3854 * we can do some work if the lock was obtained.
3856 n
= get_node(searchp
, node
);
3858 reap_alien(searchp
, n
);
3860 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
3863 * These are racy checks but it does not matter
3864 * if we skip one check or scan twice.
3866 if (time_after(n
->next_reap
, jiffies
))
3869 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
3871 drain_array(searchp
, n
, n
->shared
, node
);
3873 if (n
->free_touched
)
3874 n
->free_touched
= 0;
3878 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
3879 5 * searchp
->num
- 1) / (5 * searchp
->num
));
3880 STATS_ADD_REAPED(searchp
, freed
);
3886 mutex_unlock(&slab_mutex
);
3889 /* Set up the next iteration */
3890 schedule_delayed_work_on(smp_processor_id(), work
,
3891 round_jiffies_relative(REAPTIMEOUT_AC
));
3894 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
3896 unsigned long active_objs
, num_objs
, active_slabs
;
3897 unsigned long total_slabs
= 0, free_objs
= 0, shared_avail
= 0;
3898 unsigned long free_slabs
= 0;
3900 struct kmem_cache_node
*n
;
3902 for_each_kmem_cache_node(cachep
, node
, n
) {
3904 spin_lock_irq(&n
->list_lock
);
3906 total_slabs
+= n
->total_slabs
;
3907 free_slabs
+= n
->free_slabs
;
3908 free_objs
+= n
->free_objects
;
3911 shared_avail
+= n
->shared
->avail
;
3913 spin_unlock_irq(&n
->list_lock
);
3915 num_objs
= total_slabs
* cachep
->num
;
3916 active_slabs
= total_slabs
- free_slabs
;
3917 active_objs
= num_objs
- free_objs
;
3919 sinfo
->active_objs
= active_objs
;
3920 sinfo
->num_objs
= num_objs
;
3921 sinfo
->active_slabs
= active_slabs
;
3922 sinfo
->num_slabs
= total_slabs
;
3923 sinfo
->shared_avail
= shared_avail
;
3924 sinfo
->limit
= cachep
->limit
;
3925 sinfo
->batchcount
= cachep
->batchcount
;
3926 sinfo
->shared
= cachep
->shared
;
3927 sinfo
->objects_per_slab
= cachep
->num
;
3928 sinfo
->cache_order
= cachep
->gfporder
;
3931 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
3935 unsigned long high
= cachep
->high_mark
;
3936 unsigned long allocs
= cachep
->num_allocations
;
3937 unsigned long grown
= cachep
->grown
;
3938 unsigned long reaped
= cachep
->reaped
;
3939 unsigned long errors
= cachep
->errors
;
3940 unsigned long max_freeable
= cachep
->max_freeable
;
3941 unsigned long node_allocs
= cachep
->node_allocs
;
3942 unsigned long node_frees
= cachep
->node_frees
;
3943 unsigned long overflows
= cachep
->node_overflow
;
3945 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
3946 allocs
, high
, grown
,
3947 reaped
, errors
, max_freeable
, node_allocs
,
3948 node_frees
, overflows
);
3952 unsigned long allochit
= atomic_read(&cachep
->allochit
);
3953 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
3954 unsigned long freehit
= atomic_read(&cachep
->freehit
);
3955 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
3957 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
3958 allochit
, allocmiss
, freehit
, freemiss
);
3963 #define MAX_SLABINFO_WRITE 128
3965 * slabinfo_write - Tuning for the slab allocator
3967 * @buffer: user buffer
3968 * @count: data length
3971 * Return: %0 on success, negative error code otherwise.
3973 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
3974 size_t count
, loff_t
*ppos
)
3976 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
3977 int limit
, batchcount
, shared
, res
;
3978 struct kmem_cache
*cachep
;
3980 if (count
> MAX_SLABINFO_WRITE
)
3982 if (copy_from_user(&kbuf
, buffer
, count
))
3984 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
3986 tmp
= strchr(kbuf
, ' ');
3991 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
3994 /* Find the cache in the chain of caches. */
3995 mutex_lock(&slab_mutex
);
3997 list_for_each_entry(cachep
, &slab_caches
, list
) {
3998 if (!strcmp(cachep
->name
, kbuf
)) {
3999 if (limit
< 1 || batchcount
< 1 ||
4000 batchcount
> limit
|| shared
< 0) {
4003 res
= do_tune_cpucache(cachep
, limit
,
4010 mutex_unlock(&slab_mutex
);
4016 #ifdef CONFIG_HARDENED_USERCOPY
4018 * Rejects incorrectly sized objects and objects that are to be copied
4019 * to/from userspace but do not fall entirely within the containing slab
4020 * cache's usercopy region.
4022 * Returns NULL if check passes, otherwise const char * to name of cache
4023 * to indicate an error.
4025 void __check_heap_object(const void *ptr
, unsigned long n
,
4026 const struct slab
*slab
, bool to_user
)
4028 struct kmem_cache
*cachep
;
4030 unsigned long offset
;
4032 ptr
= kasan_reset_tag(ptr
);
4034 /* Find and validate object. */
4035 cachep
= slab
->slab_cache
;
4036 objnr
= obj_to_index(cachep
, slab
, (void *)ptr
);
4037 BUG_ON(objnr
>= cachep
->num
);
4039 /* Find offset within object. */
4040 if (is_kfence_address(ptr
))
4041 offset
= ptr
- kfence_object_start(ptr
);
4043 offset
= ptr
- index_to_obj(cachep
, slab
, objnr
) - obj_offset(cachep
);
4045 /* Allow address range falling entirely within usercopy region. */
4046 if (offset
>= cachep
->useroffset
&&
4047 offset
- cachep
->useroffset
<= cachep
->usersize
&&
4048 n
<= cachep
->useroffset
- offset
+ cachep
->usersize
)
4051 usercopy_abort("SLAB object", cachep
->name
, to_user
, offset
, n
);
4053 #endif /* CONFIG_HARDENED_USERCOPY */