1 // SPDX-License-Identifier: GPL-2.0
4 * Written by Mark Hemment, 1996/97.
5 * (markhe@nextd.demon.co.uk)
7 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
9 * Major cleanup, different bufctl logic, per-cpu arrays
10 * (c) 2000 Manfred Spraul
12 * Cleanup, make the head arrays unconditional, preparation for NUMA
13 * (c) 2002 Manfred Spraul
15 * An implementation of the Slab Allocator as described in outline in;
16 * UNIX Internals: The New Frontiers by Uresh Vahalia
17 * Pub: Prentice Hall ISBN 0-13-101908-2
18 * or with a little more detail in;
19 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
20 * Jeff Bonwick (Sun Microsystems).
21 * Presented at: USENIX Summer 1994 Technical Conference
23 * The memory is organized in caches, one cache for each object type.
24 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
25 * Each cache consists out of many slabs (they are small (usually one
26 * page long) and always contiguous), and each slab contains multiple
27 * initialized objects.
29 * This means, that your constructor is used only for newly allocated
30 * slabs and you must pass objects with the same initializations to
33 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
34 * normal). If you need a special memory type, then must create a new
35 * cache for that memory type.
37 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
38 * full slabs with 0 free objects
40 * empty slabs with no allocated objects
42 * If partial slabs exist, then new allocations come from these slabs,
43 * otherwise from empty slabs or new slabs are allocated.
45 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
46 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
48 * Each cache has a short per-cpu head array, most allocs
49 * and frees go into that array, and if that array overflows, then 1/2
50 * of the entries in the array are given back into the global cache.
51 * The head array is strictly LIFO and should improve the cache hit rates.
52 * On SMP, it additionally reduces the spinlock operations.
54 * The c_cpuarray may not be read with enabled local interrupts -
55 * it's changed with a smp_call_function().
57 * SMP synchronization:
58 * constructors and destructors are called without any locking.
59 * Several members in struct kmem_cache and struct slab never change, they
60 * are accessed without any locking.
61 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
62 * and local interrupts are disabled so slab code is preempt-safe.
63 * The non-constant members are protected with a per-cache irq spinlock.
65 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
66 * in 2000 - many ideas in the current implementation are derived from
69 * Further notes from the original documentation:
71 * 11 April '97. Started multi-threading - markhe
72 * The global cache-chain is protected by the mutex 'slab_mutex'.
73 * The sem is only needed when accessing/extending the cache-chain, which
74 * can never happen inside an interrupt (kmem_cache_create(),
75 * kmem_cache_shrink() and kmem_cache_reap()).
77 * At present, each engine can be growing a cache. This should be blocked.
79 * 15 March 2005. NUMA slab allocator.
80 * Shai Fultheim <shai@scalex86.org>.
81 * Shobhit Dayal <shobhit@calsoftinc.com>
82 * Alok N Kataria <alokk@calsoftinc.com>
83 * Christoph Lameter <christoph@lameter.com>
85 * Modified the slab allocator to be node aware on NUMA systems.
86 * Each node has its own list of partial, free and full slabs.
87 * All object allocations for a node occur from node specific slab lists.
90 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.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 page
*page
,
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 unnessary
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;})
366 #ifdef CONFIG_DEBUG_SLAB_LEAK
368 static inline bool is_store_user_clean(struct kmem_cache
*cachep
)
370 return atomic_read(&cachep
->store_user_clean
) == 1;
373 static inline void set_store_user_clean(struct kmem_cache
*cachep
)
375 atomic_set(&cachep
->store_user_clean
, 1);
378 static inline void set_store_user_dirty(struct kmem_cache
*cachep
)
380 if (is_store_user_clean(cachep
))
381 atomic_set(&cachep
->store_user_clean
, 0);
385 static inline void set_store_user_dirty(struct kmem_cache
*cachep
) {}
390 * Do not go above this order unless 0 objects fit into the slab or
391 * overridden on the command line.
393 #define SLAB_MAX_ORDER_HI 1
394 #define SLAB_MAX_ORDER_LO 0
395 static int slab_max_order
= SLAB_MAX_ORDER_LO
;
396 static bool slab_max_order_set __initdata
;
398 static inline struct kmem_cache
*virt_to_cache(const void *obj
)
400 struct page
*page
= virt_to_head_page(obj
);
401 return page
->slab_cache
;
404 static inline void *index_to_obj(struct kmem_cache
*cache
, struct page
*page
,
407 return page
->s_mem
+ cache
->size
* idx
;
411 * We want to avoid an expensive divide : (offset / cache->size)
412 * Using the fact that size is a constant for a particular cache,
413 * we can replace (offset / cache->size) by
414 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
416 static inline unsigned int obj_to_index(const struct kmem_cache
*cache
,
417 const struct page
*page
, void *obj
)
419 u32 offset
= (obj
- page
->s_mem
);
420 return reciprocal_divide(offset
, cache
->reciprocal_buffer_size
);
423 #define BOOT_CPUCACHE_ENTRIES 1
424 /* internal cache of cache description objs */
425 static struct kmem_cache kmem_cache_boot
= {
427 .limit
= BOOT_CPUCACHE_ENTRIES
,
429 .size
= sizeof(struct kmem_cache
),
430 .name
= "kmem_cache",
433 static DEFINE_PER_CPU(struct delayed_work
, slab_reap_work
);
435 static inline struct array_cache
*cpu_cache_get(struct kmem_cache
*cachep
)
437 return this_cpu_ptr(cachep
->cpu_cache
);
441 * Calculate the number of objects and left-over bytes for a given buffer size.
443 static unsigned int cache_estimate(unsigned long gfporder
, size_t buffer_size
,
444 slab_flags_t flags
, size_t *left_over
)
447 size_t slab_size
= PAGE_SIZE
<< gfporder
;
450 * The slab management structure can be either off the slab or
451 * on it. For the latter case, the memory allocated for a
454 * - @buffer_size bytes for each object
455 * - One freelist_idx_t for each object
457 * We don't need to consider alignment of freelist because
458 * freelist will be at the end of slab page. The objects will be
459 * at the correct alignment.
461 * If the slab management structure is off the slab, then the
462 * alignment will already be calculated into the size. Because
463 * the slabs are all pages aligned, the objects will be at the
464 * correct alignment when allocated.
466 if (flags
& (CFLGS_OBJFREELIST_SLAB
| CFLGS_OFF_SLAB
)) {
467 num
= slab_size
/ buffer_size
;
468 *left_over
= slab_size
% buffer_size
;
470 num
= slab_size
/ (buffer_size
+ sizeof(freelist_idx_t
));
471 *left_over
= slab_size
%
472 (buffer_size
+ sizeof(freelist_idx_t
));
479 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
481 static void __slab_error(const char *function
, struct kmem_cache
*cachep
,
484 pr_err("slab error in %s(): cache `%s': %s\n",
485 function
, cachep
->name
, msg
);
487 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
492 * By default on NUMA we use alien caches to stage the freeing of
493 * objects allocated from other nodes. This causes massive memory
494 * inefficiencies when using fake NUMA setup to split memory into a
495 * large number of small nodes, so it can be disabled on the command
499 static int use_alien_caches __read_mostly
= 1;
500 static int __init
noaliencache_setup(char *s
)
502 use_alien_caches
= 0;
505 __setup("noaliencache", noaliencache_setup
);
507 static int __init
slab_max_order_setup(char *str
)
509 get_option(&str
, &slab_max_order
);
510 slab_max_order
= slab_max_order
< 0 ? 0 :
511 min(slab_max_order
, MAX_ORDER
- 1);
512 slab_max_order_set
= true;
516 __setup("slab_max_order=", slab_max_order_setup
);
520 * Special reaping functions for NUMA systems called from cache_reap().
521 * These take care of doing round robin flushing of alien caches (containing
522 * objects freed on different nodes from which they were allocated) and the
523 * flushing of remote pcps by calling drain_node_pages.
525 static DEFINE_PER_CPU(unsigned long, slab_reap_node
);
527 static void init_reap_node(int cpu
)
529 per_cpu(slab_reap_node
, cpu
) = next_node_in(cpu_to_mem(cpu
),
533 static void next_reap_node(void)
535 int node
= __this_cpu_read(slab_reap_node
);
537 node
= next_node_in(node
, node_online_map
);
538 __this_cpu_write(slab_reap_node
, node
);
542 #define init_reap_node(cpu) do { } while (0)
543 #define next_reap_node(void) do { } while (0)
547 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
548 * via the workqueue/eventd.
549 * Add the CPU number into the expiration time to minimize the possibility of
550 * the CPUs getting into lockstep and contending for the global cache chain
553 static void start_cpu_timer(int cpu
)
555 struct delayed_work
*reap_work
= &per_cpu(slab_reap_work
, cpu
);
557 if (reap_work
->work
.func
== NULL
) {
559 INIT_DEFERRABLE_WORK(reap_work
, cache_reap
);
560 schedule_delayed_work_on(cpu
, reap_work
,
561 __round_jiffies_relative(HZ
, cpu
));
565 static void init_arraycache(struct array_cache
*ac
, int limit
, int batch
)
568 * The array_cache structures contain pointers to free object.
569 * However, when such objects are allocated or transferred to another
570 * cache the pointers are not cleared and they could be counted as
571 * valid references during a kmemleak scan. Therefore, kmemleak must
572 * not scan such objects.
574 kmemleak_no_scan(ac
);
578 ac
->batchcount
= batch
;
583 static struct array_cache
*alloc_arraycache(int node
, int entries
,
584 int batchcount
, gfp_t gfp
)
586 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
587 struct array_cache
*ac
= NULL
;
589 ac
= kmalloc_node(memsize
, gfp
, node
);
590 init_arraycache(ac
, entries
, batchcount
);
594 static noinline
void cache_free_pfmemalloc(struct kmem_cache
*cachep
,
595 struct page
*page
, void *objp
)
597 struct kmem_cache_node
*n
;
601 page_node
= page_to_nid(page
);
602 n
= get_node(cachep
, page_node
);
604 spin_lock(&n
->list_lock
);
605 free_block(cachep
, &objp
, 1, page_node
, &list
);
606 spin_unlock(&n
->list_lock
);
608 slabs_destroy(cachep
, &list
);
612 * Transfer objects in one arraycache to another.
613 * Locking must be handled by the caller.
615 * Return the number of entries transferred.
617 static int transfer_objects(struct array_cache
*to
,
618 struct array_cache
*from
, unsigned int max
)
620 /* Figure out how many entries to transfer */
621 int nr
= min3(from
->avail
, max
, to
->limit
- to
->avail
);
626 memcpy(to
->entry
+ to
->avail
, from
->entry
+ from
->avail
-nr
,
636 #define drain_alien_cache(cachep, alien) do { } while (0)
637 #define reap_alien(cachep, n) do { } while (0)
639 static inline struct alien_cache
**alloc_alien_cache(int node
,
640 int limit
, gfp_t gfp
)
645 static inline void free_alien_cache(struct alien_cache
**ac_ptr
)
649 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
654 static inline void *alternate_node_alloc(struct kmem_cache
*cachep
,
660 static inline void *____cache_alloc_node(struct kmem_cache
*cachep
,
661 gfp_t flags
, int nodeid
)
666 static inline gfp_t
gfp_exact_node(gfp_t flags
)
668 return flags
& ~__GFP_NOFAIL
;
671 #else /* CONFIG_NUMA */
673 static void *____cache_alloc_node(struct kmem_cache
*, gfp_t
, int);
674 static void *alternate_node_alloc(struct kmem_cache
*, gfp_t
);
676 static struct alien_cache
*__alloc_alien_cache(int node
, int entries
,
677 int batch
, gfp_t gfp
)
679 size_t memsize
= sizeof(void *) * entries
+ sizeof(struct alien_cache
);
680 struct alien_cache
*alc
= NULL
;
682 alc
= kmalloc_node(memsize
, gfp
, node
);
683 init_arraycache(&alc
->ac
, entries
, batch
);
684 spin_lock_init(&alc
->lock
);
688 static struct alien_cache
**alloc_alien_cache(int node
, int limit
, gfp_t gfp
)
690 struct alien_cache
**alc_ptr
;
691 size_t memsize
= sizeof(void *) * nr_node_ids
;
696 alc_ptr
= kzalloc_node(memsize
, gfp
, node
);
701 if (i
== node
|| !node_online(i
))
703 alc_ptr
[i
] = __alloc_alien_cache(node
, limit
, 0xbaadf00d, gfp
);
705 for (i
--; i
>= 0; i
--)
714 static void free_alien_cache(struct alien_cache
**alc_ptr
)
725 static void __drain_alien_cache(struct kmem_cache
*cachep
,
726 struct array_cache
*ac
, int node
,
727 struct list_head
*list
)
729 struct kmem_cache_node
*n
= get_node(cachep
, node
);
732 spin_lock(&n
->list_lock
);
734 * Stuff objects into the remote nodes shared array first.
735 * That way we could avoid the overhead of putting the objects
736 * into the free lists and getting them back later.
739 transfer_objects(n
->shared
, ac
, ac
->limit
);
741 free_block(cachep
, ac
->entry
, ac
->avail
, node
, list
);
743 spin_unlock(&n
->list_lock
);
748 * Called from cache_reap() to regularly drain alien caches round robin.
750 static void reap_alien(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
)
752 int node
= __this_cpu_read(slab_reap_node
);
755 struct alien_cache
*alc
= n
->alien
[node
];
756 struct array_cache
*ac
;
760 if (ac
->avail
&& spin_trylock_irq(&alc
->lock
)) {
763 __drain_alien_cache(cachep
, ac
, node
, &list
);
764 spin_unlock_irq(&alc
->lock
);
765 slabs_destroy(cachep
, &list
);
771 static void drain_alien_cache(struct kmem_cache
*cachep
,
772 struct alien_cache
**alien
)
775 struct alien_cache
*alc
;
776 struct array_cache
*ac
;
779 for_each_online_node(i
) {
785 spin_lock_irqsave(&alc
->lock
, flags
);
786 __drain_alien_cache(cachep
, ac
, i
, &list
);
787 spin_unlock_irqrestore(&alc
->lock
, flags
);
788 slabs_destroy(cachep
, &list
);
793 static int __cache_free_alien(struct kmem_cache
*cachep
, void *objp
,
794 int node
, int page_node
)
796 struct kmem_cache_node
*n
;
797 struct alien_cache
*alien
= NULL
;
798 struct array_cache
*ac
;
801 n
= get_node(cachep
, node
);
802 STATS_INC_NODEFREES(cachep
);
803 if (n
->alien
&& n
->alien
[page_node
]) {
804 alien
= n
->alien
[page_node
];
806 spin_lock(&alien
->lock
);
807 if (unlikely(ac
->avail
== ac
->limit
)) {
808 STATS_INC_ACOVERFLOW(cachep
);
809 __drain_alien_cache(cachep
, ac
, page_node
, &list
);
811 ac
->entry
[ac
->avail
++] = objp
;
812 spin_unlock(&alien
->lock
);
813 slabs_destroy(cachep
, &list
);
815 n
= get_node(cachep
, page_node
);
816 spin_lock(&n
->list_lock
);
817 free_block(cachep
, &objp
, 1, page_node
, &list
);
818 spin_unlock(&n
->list_lock
);
819 slabs_destroy(cachep
, &list
);
824 static inline int cache_free_alien(struct kmem_cache
*cachep
, void *objp
)
826 int page_node
= page_to_nid(virt_to_page(objp
));
827 int node
= numa_mem_id();
829 * Make sure we are not freeing a object from another node to the array
832 if (likely(node
== page_node
))
835 return __cache_free_alien(cachep
, objp
, node
, page_node
);
839 * Construct gfp mask to allocate from a specific node but do not reclaim or
840 * warn about failures.
842 static inline gfp_t
gfp_exact_node(gfp_t flags
)
844 return (flags
| __GFP_THISNODE
| __GFP_NOWARN
) & ~(__GFP_RECLAIM
|__GFP_NOFAIL
);
848 static int init_cache_node(struct kmem_cache
*cachep
, int node
, gfp_t gfp
)
850 struct kmem_cache_node
*n
;
853 * Set up the kmem_cache_node for cpu before we can
854 * begin anything. Make sure some other cpu on this
855 * node has not already allocated this
857 n
= get_node(cachep
, node
);
859 spin_lock_irq(&n
->list_lock
);
860 n
->free_limit
= (1 + nr_cpus_node(node
)) * cachep
->batchcount
+
862 spin_unlock_irq(&n
->list_lock
);
867 n
= kmalloc_node(sizeof(struct kmem_cache_node
), gfp
, node
);
871 kmem_cache_node_init(n
);
872 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
+
873 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
876 (1 + nr_cpus_node(node
)) * cachep
->batchcount
+ cachep
->num
;
879 * The kmem_cache_nodes don't come and go as CPUs
880 * come and go. slab_mutex is sufficient
883 cachep
->node
[node
] = n
;
888 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
890 * Allocates and initializes node for a node on each slab cache, used for
891 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
892 * will be allocated off-node since memory is not yet online for the new node.
893 * When hotplugging memory or a cpu, existing node are not replaced if
896 * Must hold slab_mutex.
898 static int init_cache_node_node(int node
)
901 struct kmem_cache
*cachep
;
903 list_for_each_entry(cachep
, &slab_caches
, list
) {
904 ret
= init_cache_node(cachep
, node
, GFP_KERNEL
);
913 static int setup_kmem_cache_node(struct kmem_cache
*cachep
,
914 int node
, gfp_t gfp
, bool force_change
)
917 struct kmem_cache_node
*n
;
918 struct array_cache
*old_shared
= NULL
;
919 struct array_cache
*new_shared
= NULL
;
920 struct alien_cache
**new_alien
= NULL
;
923 if (use_alien_caches
) {
924 new_alien
= alloc_alien_cache(node
, cachep
->limit
, gfp
);
929 if (cachep
->shared
) {
930 new_shared
= alloc_arraycache(node
,
931 cachep
->shared
* cachep
->batchcount
, 0xbaadf00d, gfp
);
936 ret
= init_cache_node(cachep
, node
, gfp
);
940 n
= get_node(cachep
, node
);
941 spin_lock_irq(&n
->list_lock
);
942 if (n
->shared
&& force_change
) {
943 free_block(cachep
, n
->shared
->entry
,
944 n
->shared
->avail
, node
, &list
);
945 n
->shared
->avail
= 0;
948 if (!n
->shared
|| force_change
) {
949 old_shared
= n
->shared
;
950 n
->shared
= new_shared
;
955 n
->alien
= new_alien
;
959 spin_unlock_irq(&n
->list_lock
);
960 slabs_destroy(cachep
, &list
);
963 * To protect lockless access to n->shared during irq disabled context.
964 * If n->shared isn't NULL in irq disabled context, accessing to it is
965 * guaranteed to be valid until irq is re-enabled, because it will be
966 * freed after synchronize_sched().
968 if (old_shared
&& force_change
)
974 free_alien_cache(new_alien
);
981 static void cpuup_canceled(long cpu
)
983 struct kmem_cache
*cachep
;
984 struct kmem_cache_node
*n
= NULL
;
985 int node
= cpu_to_mem(cpu
);
986 const struct cpumask
*mask
= cpumask_of_node(node
);
988 list_for_each_entry(cachep
, &slab_caches
, list
) {
989 struct array_cache
*nc
;
990 struct array_cache
*shared
;
991 struct alien_cache
**alien
;
994 n
= get_node(cachep
, node
);
998 spin_lock_irq(&n
->list_lock
);
1000 /* Free limit for this kmem_cache_node */
1001 n
->free_limit
-= cachep
->batchcount
;
1003 /* cpu is dead; no one can alloc from it. */
1004 nc
= per_cpu_ptr(cachep
->cpu_cache
, cpu
);
1006 free_block(cachep
, nc
->entry
, nc
->avail
, node
, &list
);
1010 if (!cpumask_empty(mask
)) {
1011 spin_unlock_irq(&n
->list_lock
);
1017 free_block(cachep
, shared
->entry
,
1018 shared
->avail
, node
, &list
);
1025 spin_unlock_irq(&n
->list_lock
);
1029 drain_alien_cache(cachep
, alien
);
1030 free_alien_cache(alien
);
1034 slabs_destroy(cachep
, &list
);
1037 * In the previous loop, all the objects were freed to
1038 * the respective cache's slabs, now we can go ahead and
1039 * shrink each nodelist to its limit.
1041 list_for_each_entry(cachep
, &slab_caches
, list
) {
1042 n
= get_node(cachep
, node
);
1045 drain_freelist(cachep
, n
, INT_MAX
);
1049 static int cpuup_prepare(long cpu
)
1051 struct kmem_cache
*cachep
;
1052 int node
= cpu_to_mem(cpu
);
1056 * We need to do this right in the beginning since
1057 * alloc_arraycache's are going to use this list.
1058 * kmalloc_node allows us to add the slab to the right
1059 * kmem_cache_node and not this cpu's kmem_cache_node
1061 err
= init_cache_node_node(node
);
1066 * Now we can go ahead with allocating the shared arrays and
1069 list_for_each_entry(cachep
, &slab_caches
, list
) {
1070 err
= setup_kmem_cache_node(cachep
, node
, GFP_KERNEL
, false);
1077 cpuup_canceled(cpu
);
1081 int slab_prepare_cpu(unsigned int cpu
)
1085 mutex_lock(&slab_mutex
);
1086 err
= cpuup_prepare(cpu
);
1087 mutex_unlock(&slab_mutex
);
1092 * This is called for a failed online attempt and for a successful
1095 * Even if all the cpus of a node are down, we don't free the
1096 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1097 * a kmalloc allocation from another cpu for memory from the node of
1098 * the cpu going down. The list3 structure is usually allocated from
1099 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1101 int slab_dead_cpu(unsigned int cpu
)
1103 mutex_lock(&slab_mutex
);
1104 cpuup_canceled(cpu
);
1105 mutex_unlock(&slab_mutex
);
1110 static int slab_online_cpu(unsigned int cpu
)
1112 start_cpu_timer(cpu
);
1116 static int slab_offline_cpu(unsigned int cpu
)
1119 * Shutdown cache reaper. Note that the slab_mutex is held so
1120 * that if cache_reap() is invoked it cannot do anything
1121 * expensive but will only modify reap_work and reschedule the
1124 cancel_delayed_work_sync(&per_cpu(slab_reap_work
, cpu
));
1125 /* Now the cache_reaper is guaranteed to be not running. */
1126 per_cpu(slab_reap_work
, cpu
).work
.func
= NULL
;
1130 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1132 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1133 * Returns -EBUSY if all objects cannot be drained so that the node is not
1136 * Must hold slab_mutex.
1138 static int __meminit
drain_cache_node_node(int node
)
1140 struct kmem_cache
*cachep
;
1143 list_for_each_entry(cachep
, &slab_caches
, list
) {
1144 struct kmem_cache_node
*n
;
1146 n
= get_node(cachep
, node
);
1150 drain_freelist(cachep
, n
, INT_MAX
);
1152 if (!list_empty(&n
->slabs_full
) ||
1153 !list_empty(&n
->slabs_partial
)) {
1161 static int __meminit
slab_memory_callback(struct notifier_block
*self
,
1162 unsigned long action
, void *arg
)
1164 struct memory_notify
*mnb
= arg
;
1168 nid
= mnb
->status_change_nid
;
1173 case MEM_GOING_ONLINE
:
1174 mutex_lock(&slab_mutex
);
1175 ret
= init_cache_node_node(nid
);
1176 mutex_unlock(&slab_mutex
);
1178 case MEM_GOING_OFFLINE
:
1179 mutex_lock(&slab_mutex
);
1180 ret
= drain_cache_node_node(nid
);
1181 mutex_unlock(&slab_mutex
);
1185 case MEM_CANCEL_ONLINE
:
1186 case MEM_CANCEL_OFFLINE
:
1190 return notifier_from_errno(ret
);
1192 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1195 * swap the static kmem_cache_node with kmalloced memory
1197 static void __init
init_list(struct kmem_cache
*cachep
, struct kmem_cache_node
*list
,
1200 struct kmem_cache_node
*ptr
;
1202 ptr
= kmalloc_node(sizeof(struct kmem_cache_node
), GFP_NOWAIT
, nodeid
);
1205 memcpy(ptr
, list
, sizeof(struct kmem_cache_node
));
1207 * Do not assume that spinlocks can be initialized via memcpy:
1209 spin_lock_init(&ptr
->list_lock
);
1211 MAKE_ALL_LISTS(cachep
, ptr
, nodeid
);
1212 cachep
->node
[nodeid
] = ptr
;
1216 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1217 * size of kmem_cache_node.
1219 static void __init
set_up_node(struct kmem_cache
*cachep
, int index
)
1223 for_each_online_node(node
) {
1224 cachep
->node
[node
] = &init_kmem_cache_node
[index
+ node
];
1225 cachep
->node
[node
]->next_reap
= jiffies
+
1227 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1232 * Initialisation. Called after the page allocator have been initialised and
1233 * before smp_init().
1235 void __init
kmem_cache_init(void)
1239 BUILD_BUG_ON(sizeof(((struct page
*)NULL
)->lru
) <
1240 sizeof(struct rcu_head
));
1241 kmem_cache
= &kmem_cache_boot
;
1243 if (!IS_ENABLED(CONFIG_NUMA
) || num_possible_nodes() == 1)
1244 use_alien_caches
= 0;
1246 for (i
= 0; i
< NUM_INIT_LISTS
; i
++)
1247 kmem_cache_node_init(&init_kmem_cache_node
[i
]);
1250 * Fragmentation resistance on low memory - only use bigger
1251 * page orders on machines with more than 32MB of memory if
1252 * not overridden on the command line.
1254 if (!slab_max_order_set
&& totalram_pages
> (32 << 20) >> PAGE_SHIFT
)
1255 slab_max_order
= SLAB_MAX_ORDER_HI
;
1257 /* Bootstrap is tricky, because several objects are allocated
1258 * from caches that do not exist yet:
1259 * 1) initialize the kmem_cache cache: it contains the struct
1260 * kmem_cache structures of all caches, except kmem_cache itself:
1261 * kmem_cache is statically allocated.
1262 * Initially an __init data area is used for the head array and the
1263 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1264 * array at the end of the bootstrap.
1265 * 2) Create the first kmalloc cache.
1266 * The struct kmem_cache for the new cache is allocated normally.
1267 * An __init data area is used for the head array.
1268 * 3) Create the remaining kmalloc caches, with minimally sized
1270 * 4) Replace the __init data head arrays for kmem_cache and the first
1271 * kmalloc cache with kmalloc allocated arrays.
1272 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1273 * the other cache's with kmalloc allocated memory.
1274 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1277 /* 1) create the kmem_cache */
1280 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1282 create_boot_cache(kmem_cache
, "kmem_cache",
1283 offsetof(struct kmem_cache
, node
) +
1284 nr_node_ids
* sizeof(struct kmem_cache_node
*),
1285 SLAB_HWCACHE_ALIGN
);
1286 list_add(&kmem_cache
->list
, &slab_caches
);
1287 slab_state
= PARTIAL
;
1290 * Initialize the caches that provide memory for the kmem_cache_node
1291 * structures first. Without this, further allocations will bug.
1293 kmalloc_caches
[INDEX_NODE
] = create_kmalloc_cache(
1294 kmalloc_info
[INDEX_NODE
].name
,
1295 kmalloc_size(INDEX_NODE
), ARCH_KMALLOC_FLAGS
);
1296 slab_state
= PARTIAL_NODE
;
1297 setup_kmalloc_cache_index_table();
1299 slab_early_init
= 0;
1301 /* 5) Replace the bootstrap kmem_cache_node */
1305 for_each_online_node(nid
) {
1306 init_list(kmem_cache
, &init_kmem_cache_node
[CACHE_CACHE
+ nid
], nid
);
1308 init_list(kmalloc_caches
[INDEX_NODE
],
1309 &init_kmem_cache_node
[SIZE_NODE
+ nid
], nid
);
1313 create_kmalloc_caches(ARCH_KMALLOC_FLAGS
);
1316 void __init
kmem_cache_init_late(void)
1318 struct kmem_cache
*cachep
;
1322 /* 6) resize the head arrays to their final sizes */
1323 mutex_lock(&slab_mutex
);
1324 list_for_each_entry(cachep
, &slab_caches
, list
)
1325 if (enable_cpucache(cachep
, GFP_NOWAIT
))
1327 mutex_unlock(&slab_mutex
);
1334 * Register a memory hotplug callback that initializes and frees
1337 hotplug_memory_notifier(slab_memory_callback
, SLAB_CALLBACK_PRI
);
1341 * The reap timers are started later, with a module init call: That part
1342 * of the kernel is not yet operational.
1346 static int __init
cpucache_init(void)
1351 * Register the timers that return unneeded pages to the page allocator
1353 ret
= cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "SLAB online",
1354 slab_online_cpu
, slab_offline_cpu
);
1361 __initcall(cpucache_init
);
1363 static noinline
void
1364 slab_out_of_memory(struct kmem_cache
*cachep
, gfp_t gfpflags
, int nodeid
)
1367 struct kmem_cache_node
*n
;
1368 unsigned long flags
;
1370 static DEFINE_RATELIMIT_STATE(slab_oom_rs
, DEFAULT_RATELIMIT_INTERVAL
,
1371 DEFAULT_RATELIMIT_BURST
);
1373 if ((gfpflags
& __GFP_NOWARN
) || !__ratelimit(&slab_oom_rs
))
1376 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1377 nodeid
, gfpflags
, &gfpflags
);
1378 pr_warn(" cache: %s, object size: %d, order: %d\n",
1379 cachep
->name
, cachep
->size
, cachep
->gfporder
);
1381 for_each_kmem_cache_node(cachep
, node
, n
) {
1382 unsigned long total_slabs
, free_slabs
, free_objs
;
1384 spin_lock_irqsave(&n
->list_lock
, flags
);
1385 total_slabs
= n
->total_slabs
;
1386 free_slabs
= n
->free_slabs
;
1387 free_objs
= n
->free_objects
;
1388 spin_unlock_irqrestore(&n
->list_lock
, flags
);
1390 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1391 node
, total_slabs
- free_slabs
, total_slabs
,
1392 (total_slabs
* cachep
->num
) - free_objs
,
1393 total_slabs
* cachep
->num
);
1399 * Interface to system's page allocator. No need to hold the
1400 * kmem_cache_node ->list_lock.
1402 * If we requested dmaable memory, we will get it. Even if we
1403 * did not request dmaable memory, we might get it, but that
1404 * would be relatively rare and ignorable.
1406 static struct page
*kmem_getpages(struct kmem_cache
*cachep
, gfp_t flags
,
1412 flags
|= cachep
->allocflags
;
1414 page
= __alloc_pages_node(nodeid
, flags
| __GFP_NOTRACK
, cachep
->gfporder
);
1416 slab_out_of_memory(cachep
, flags
, nodeid
);
1420 if (memcg_charge_slab(page
, flags
, cachep
->gfporder
, cachep
)) {
1421 __free_pages(page
, cachep
->gfporder
);
1425 nr_pages
= (1 << cachep
->gfporder
);
1426 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1427 mod_lruvec_page_state(page
, NR_SLAB_RECLAIMABLE
, nr_pages
);
1429 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE
, nr_pages
);
1431 __SetPageSlab(page
);
1432 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1433 if (sk_memalloc_socks() && page_is_pfmemalloc(page
))
1434 SetPageSlabPfmemalloc(page
);
1436 if (kmemcheck_enabled
&& !(cachep
->flags
& SLAB_NOTRACK
)) {
1437 kmemcheck_alloc_shadow(page
, cachep
->gfporder
, flags
, nodeid
);
1440 kmemcheck_mark_uninitialized_pages(page
, nr_pages
);
1442 kmemcheck_mark_unallocated_pages(page
, nr_pages
);
1449 * Interface to system's page release.
1451 static void kmem_freepages(struct kmem_cache
*cachep
, struct page
*page
)
1453 int order
= cachep
->gfporder
;
1454 unsigned long nr_freed
= (1 << order
);
1456 kmemcheck_free_shadow(page
, order
);
1458 if (cachep
->flags
& SLAB_RECLAIM_ACCOUNT
)
1459 mod_lruvec_page_state(page
, NR_SLAB_RECLAIMABLE
, -nr_freed
);
1461 mod_lruvec_page_state(page
, NR_SLAB_UNRECLAIMABLE
, -nr_freed
);
1463 BUG_ON(!PageSlab(page
));
1464 __ClearPageSlabPfmemalloc(page
);
1465 __ClearPageSlab(page
);
1466 page_mapcount_reset(page
);
1467 page
->mapping
= NULL
;
1469 if (current
->reclaim_state
)
1470 current
->reclaim_state
->reclaimed_slab
+= nr_freed
;
1471 memcg_uncharge_slab(page
, order
, cachep
);
1472 __free_pages(page
, order
);
1475 static void kmem_rcu_free(struct rcu_head
*head
)
1477 struct kmem_cache
*cachep
;
1480 page
= container_of(head
, struct page
, rcu_head
);
1481 cachep
= page
->slab_cache
;
1483 kmem_freepages(cachep
, page
);
1487 static bool is_debug_pagealloc_cache(struct kmem_cache
*cachep
)
1489 if (debug_pagealloc_enabled() && OFF_SLAB(cachep
) &&
1490 (cachep
->size
% PAGE_SIZE
) == 0)
1496 #ifdef CONFIG_DEBUG_PAGEALLOC
1497 static void store_stackinfo(struct kmem_cache
*cachep
, unsigned long *addr
,
1498 unsigned long caller
)
1500 int size
= cachep
->object_size
;
1502 addr
= (unsigned long *)&((char *)addr
)[obj_offset(cachep
)];
1504 if (size
< 5 * sizeof(unsigned long))
1507 *addr
++ = 0x12345678;
1509 *addr
++ = smp_processor_id();
1510 size
-= 3 * sizeof(unsigned long);
1512 unsigned long *sptr
= &caller
;
1513 unsigned long svalue
;
1515 while (!kstack_end(sptr
)) {
1517 if (kernel_text_address(svalue
)) {
1519 size
-= sizeof(unsigned long);
1520 if (size
<= sizeof(unsigned long))
1526 *addr
++ = 0x87654321;
1529 static void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1530 int map
, unsigned long caller
)
1532 if (!is_debug_pagealloc_cache(cachep
))
1536 store_stackinfo(cachep
, objp
, caller
);
1538 kernel_map_pages(virt_to_page(objp
), cachep
->size
/ PAGE_SIZE
, map
);
1542 static inline void slab_kernel_map(struct kmem_cache
*cachep
, void *objp
,
1543 int map
, unsigned long caller
) {}
1547 static void poison_obj(struct kmem_cache
*cachep
, void *addr
, unsigned char val
)
1549 int size
= cachep
->object_size
;
1550 addr
= &((char *)addr
)[obj_offset(cachep
)];
1552 memset(addr
, val
, size
);
1553 *(unsigned char *)(addr
+ size
- 1) = POISON_END
;
1556 static void dump_line(char *data
, int offset
, int limit
)
1559 unsigned char error
= 0;
1562 pr_err("%03x: ", offset
);
1563 for (i
= 0; i
< limit
; i
++) {
1564 if (data
[offset
+ i
] != POISON_FREE
) {
1565 error
= data
[offset
+ i
];
1569 print_hex_dump(KERN_CONT
, "", 0, 16, 1,
1570 &data
[offset
], limit
, 1);
1572 if (bad_count
== 1) {
1573 error
^= POISON_FREE
;
1574 if (!(error
& (error
- 1))) {
1575 pr_err("Single bit error detected. Probably bad RAM.\n");
1577 pr_err("Run memtest86+ or a similar memory test tool.\n");
1579 pr_err("Run a memory test tool.\n");
1588 static void print_objinfo(struct kmem_cache
*cachep
, void *objp
, int lines
)
1593 if (cachep
->flags
& SLAB_RED_ZONE
) {
1594 pr_err("Redzone: 0x%llx/0x%llx\n",
1595 *dbg_redzone1(cachep
, objp
),
1596 *dbg_redzone2(cachep
, objp
));
1599 if (cachep
->flags
& SLAB_STORE_USER
) {
1600 pr_err("Last user: [<%p>](%pSR)\n",
1601 *dbg_userword(cachep
, objp
),
1602 *dbg_userword(cachep
, objp
));
1604 realobj
= (char *)objp
+ obj_offset(cachep
);
1605 size
= cachep
->object_size
;
1606 for (i
= 0; i
< size
&& lines
; i
+= 16, lines
--) {
1609 if (i
+ limit
> size
)
1611 dump_line(realobj
, i
, limit
);
1615 static void check_poison_obj(struct kmem_cache
*cachep
, void *objp
)
1621 if (is_debug_pagealloc_cache(cachep
))
1624 realobj
= (char *)objp
+ obj_offset(cachep
);
1625 size
= cachep
->object_size
;
1627 for (i
= 0; i
< size
; i
++) {
1628 char exp
= POISON_FREE
;
1631 if (realobj
[i
] != exp
) {
1636 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1637 print_tainted(), cachep
->name
,
1639 print_objinfo(cachep
, objp
, 0);
1641 /* Hexdump the affected line */
1644 if (i
+ limit
> size
)
1646 dump_line(realobj
, i
, limit
);
1649 /* Limit to 5 lines */
1655 /* Print some data about the neighboring objects, if they
1658 struct page
*page
= virt_to_head_page(objp
);
1661 objnr
= obj_to_index(cachep
, page
, objp
);
1663 objp
= index_to_obj(cachep
, page
, objnr
- 1);
1664 realobj
= (char *)objp
+ obj_offset(cachep
);
1665 pr_err("Prev obj: start=%p, len=%d\n", realobj
, size
);
1666 print_objinfo(cachep
, objp
, 2);
1668 if (objnr
+ 1 < cachep
->num
) {
1669 objp
= index_to_obj(cachep
, page
, objnr
+ 1);
1670 realobj
= (char *)objp
+ obj_offset(cachep
);
1671 pr_err("Next obj: start=%p, len=%d\n", realobj
, size
);
1672 print_objinfo(cachep
, objp
, 2);
1679 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1684 if (OBJFREELIST_SLAB(cachep
) && cachep
->flags
& SLAB_POISON
) {
1685 poison_obj(cachep
, page
->freelist
- obj_offset(cachep
),
1689 for (i
= 0; i
< cachep
->num
; i
++) {
1690 void *objp
= index_to_obj(cachep
, page
, i
);
1692 if (cachep
->flags
& SLAB_POISON
) {
1693 check_poison_obj(cachep
, objp
);
1694 slab_kernel_map(cachep
, objp
, 1, 0);
1696 if (cachep
->flags
& SLAB_RED_ZONE
) {
1697 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
1698 slab_error(cachep
, "start of a freed object was overwritten");
1699 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
1700 slab_error(cachep
, "end of a freed object was overwritten");
1705 static void slab_destroy_debugcheck(struct kmem_cache
*cachep
,
1712 * slab_destroy - destroy and release all objects in a slab
1713 * @cachep: cache pointer being destroyed
1714 * @page: page pointer being destroyed
1716 * Destroy all the objs in a slab page, and release the mem back to the system.
1717 * Before calling the slab page must have been unlinked from the cache. The
1718 * kmem_cache_node ->list_lock is not held/needed.
1720 static void slab_destroy(struct kmem_cache
*cachep
, struct page
*page
)
1724 freelist
= page
->freelist
;
1725 slab_destroy_debugcheck(cachep
, page
);
1726 if (unlikely(cachep
->flags
& SLAB_TYPESAFE_BY_RCU
))
1727 call_rcu(&page
->rcu_head
, kmem_rcu_free
);
1729 kmem_freepages(cachep
, page
);
1732 * From now on, we don't use freelist
1733 * although actual page can be freed in rcu context
1735 if (OFF_SLAB(cachep
))
1736 kmem_cache_free(cachep
->freelist_cache
, freelist
);
1739 static void slabs_destroy(struct kmem_cache
*cachep
, struct list_head
*list
)
1741 struct page
*page
, *n
;
1743 list_for_each_entry_safe(page
, n
, list
, lru
) {
1744 list_del(&page
->lru
);
1745 slab_destroy(cachep
, page
);
1750 * calculate_slab_order - calculate size (page order) of slabs
1751 * @cachep: pointer to the cache that is being created
1752 * @size: size of objects to be created in this cache.
1753 * @flags: slab allocation flags
1755 * Also calculates the number of objects per slab.
1757 * This could be made much more intelligent. For now, try to avoid using
1758 * high order pages for slabs. When the gfp() functions are more friendly
1759 * towards high-order requests, this should be changed.
1761 static size_t calculate_slab_order(struct kmem_cache
*cachep
,
1762 size_t size
, slab_flags_t flags
)
1764 size_t left_over
= 0;
1767 for (gfporder
= 0; gfporder
<= KMALLOC_MAX_ORDER
; gfporder
++) {
1771 num
= cache_estimate(gfporder
, size
, flags
, &remainder
);
1775 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1776 if (num
> SLAB_OBJ_MAX_NUM
)
1779 if (flags
& CFLGS_OFF_SLAB
) {
1780 struct kmem_cache
*freelist_cache
;
1781 size_t freelist_size
;
1783 freelist_size
= num
* sizeof(freelist_idx_t
);
1784 freelist_cache
= kmalloc_slab(freelist_size
, 0u);
1785 if (!freelist_cache
)
1789 * Needed to avoid possible looping condition
1790 * in cache_grow_begin()
1792 if (OFF_SLAB(freelist_cache
))
1795 /* check if off slab has enough benefit */
1796 if (freelist_cache
->size
> cachep
->size
/ 2)
1800 /* Found something acceptable - save it away */
1802 cachep
->gfporder
= gfporder
;
1803 left_over
= remainder
;
1806 * A VFS-reclaimable slab tends to have most allocations
1807 * as GFP_NOFS and we really don't want to have to be allocating
1808 * higher-order pages when we are unable to shrink dcache.
1810 if (flags
& SLAB_RECLAIM_ACCOUNT
)
1814 * Large number of objects is good, but very large slabs are
1815 * currently bad for the gfp()s.
1817 if (gfporder
>= slab_max_order
)
1821 * Acceptable internal fragmentation?
1823 if (left_over
* 8 <= (PAGE_SIZE
<< gfporder
))
1829 static struct array_cache __percpu
*alloc_kmem_cache_cpus(
1830 struct kmem_cache
*cachep
, int entries
, int batchcount
)
1834 struct array_cache __percpu
*cpu_cache
;
1836 size
= sizeof(void *) * entries
+ sizeof(struct array_cache
);
1837 cpu_cache
= __alloc_percpu(size
, sizeof(void *));
1842 for_each_possible_cpu(cpu
) {
1843 init_arraycache(per_cpu_ptr(cpu_cache
, cpu
),
1844 entries
, batchcount
);
1850 static int __ref
setup_cpu_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
1852 if (slab_state
>= FULL
)
1853 return enable_cpucache(cachep
, gfp
);
1855 cachep
->cpu_cache
= alloc_kmem_cache_cpus(cachep
, 1, 1);
1856 if (!cachep
->cpu_cache
)
1859 if (slab_state
== DOWN
) {
1860 /* Creation of first cache (kmem_cache). */
1861 set_up_node(kmem_cache
, CACHE_CACHE
);
1862 } else if (slab_state
== PARTIAL
) {
1863 /* For kmem_cache_node */
1864 set_up_node(cachep
, SIZE_NODE
);
1868 for_each_online_node(node
) {
1869 cachep
->node
[node
] = kmalloc_node(
1870 sizeof(struct kmem_cache_node
), gfp
, node
);
1871 BUG_ON(!cachep
->node
[node
]);
1872 kmem_cache_node_init(cachep
->node
[node
]);
1876 cachep
->node
[numa_mem_id()]->next_reap
=
1877 jiffies
+ REAPTIMEOUT_NODE
+
1878 ((unsigned long)cachep
) % REAPTIMEOUT_NODE
;
1880 cpu_cache_get(cachep
)->avail
= 0;
1881 cpu_cache_get(cachep
)->limit
= BOOT_CPUCACHE_ENTRIES
;
1882 cpu_cache_get(cachep
)->batchcount
= 1;
1883 cpu_cache_get(cachep
)->touched
= 0;
1884 cachep
->batchcount
= 1;
1885 cachep
->limit
= BOOT_CPUCACHE_ENTRIES
;
1889 slab_flags_t
kmem_cache_flags(unsigned long object_size
,
1890 slab_flags_t flags
, const char *name
,
1891 void (*ctor
)(void *))
1897 __kmem_cache_alias(const char *name
, size_t size
, size_t align
,
1898 slab_flags_t flags
, void (*ctor
)(void *))
1900 struct kmem_cache
*cachep
;
1902 cachep
= find_mergeable(size
, align
, flags
, name
, ctor
);
1907 * Adjust the object sizes so that we clear
1908 * the complete object on kzalloc.
1910 cachep
->object_size
= max_t(int, cachep
->object_size
, size
);
1915 static bool set_objfreelist_slab_cache(struct kmem_cache
*cachep
,
1916 size_t size
, slab_flags_t flags
)
1922 if (cachep
->ctor
|| flags
& SLAB_TYPESAFE_BY_RCU
)
1925 left
= calculate_slab_order(cachep
, size
,
1926 flags
| CFLGS_OBJFREELIST_SLAB
);
1930 if (cachep
->num
* sizeof(freelist_idx_t
) > cachep
->object_size
)
1933 cachep
->colour
= left
/ cachep
->colour_off
;
1938 static bool set_off_slab_cache(struct kmem_cache
*cachep
,
1939 size_t size
, slab_flags_t flags
)
1946 * Always use on-slab management when SLAB_NOLEAKTRACE
1947 * to avoid recursive calls into kmemleak.
1949 if (flags
& SLAB_NOLEAKTRACE
)
1953 * Size is large, assume best to place the slab management obj
1954 * off-slab (should allow better packing of objs).
1956 left
= calculate_slab_order(cachep
, size
, flags
| CFLGS_OFF_SLAB
);
1961 * If the slab has been placed off-slab, and we have enough space then
1962 * move it on-slab. This is at the expense of any extra colouring.
1964 if (left
>= cachep
->num
* sizeof(freelist_idx_t
))
1967 cachep
->colour
= left
/ cachep
->colour_off
;
1972 static bool set_on_slab_cache(struct kmem_cache
*cachep
,
1973 size_t size
, slab_flags_t flags
)
1979 left
= calculate_slab_order(cachep
, size
, flags
);
1983 cachep
->colour
= left
/ cachep
->colour_off
;
1989 * __kmem_cache_create - Create a cache.
1990 * @cachep: cache management descriptor
1991 * @flags: SLAB flags
1993 * Returns a ptr to the cache on success, NULL on failure.
1994 * Cannot be called within a int, but can be interrupted.
1995 * The @ctor is run when new pages are allocated by the cache.
1999 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2000 * to catch references to uninitialised memory.
2002 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2003 * for buffer overruns.
2005 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2006 * cacheline. This can be beneficial if you're counting cycles as closely
2009 int __kmem_cache_create(struct kmem_cache
*cachep
, slab_flags_t flags
)
2011 size_t ralign
= BYTES_PER_WORD
;
2014 size_t size
= cachep
->size
;
2019 * Enable redzoning and last user accounting, except for caches with
2020 * large objects, if the increased size would increase the object size
2021 * above the next power of two: caches with object sizes just above a
2022 * power of two have a significant amount of internal fragmentation.
2024 if (size
< 4096 || fls(size
- 1) == fls(size
-1 + REDZONE_ALIGN
+
2025 2 * sizeof(unsigned long long)))
2026 flags
|= SLAB_RED_ZONE
| SLAB_STORE_USER
;
2027 if (!(flags
& SLAB_TYPESAFE_BY_RCU
))
2028 flags
|= SLAB_POISON
;
2033 * Check that size is in terms of words. This is needed to avoid
2034 * unaligned accesses for some archs when redzoning is used, and makes
2035 * sure any on-slab bufctl's are also correctly aligned.
2037 size
= ALIGN(size
, BYTES_PER_WORD
);
2039 if (flags
& SLAB_RED_ZONE
) {
2040 ralign
= REDZONE_ALIGN
;
2041 /* If redzoning, ensure that the second redzone is suitably
2042 * aligned, by adjusting the object size accordingly. */
2043 size
= ALIGN(size
, REDZONE_ALIGN
);
2046 /* 3) caller mandated alignment */
2047 if (ralign
< cachep
->align
) {
2048 ralign
= cachep
->align
;
2050 /* disable debug if necessary */
2051 if (ralign
> __alignof__(unsigned long long))
2052 flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2056 cachep
->align
= ralign
;
2057 cachep
->colour_off
= cache_line_size();
2058 /* Offset must be a multiple of the alignment. */
2059 if (cachep
->colour_off
< cachep
->align
)
2060 cachep
->colour_off
= cachep
->align
;
2062 if (slab_is_available())
2070 * Both debugging options require word-alignment which is calculated
2073 if (flags
& SLAB_RED_ZONE
) {
2074 /* add space for red zone words */
2075 cachep
->obj_offset
+= sizeof(unsigned long long);
2076 size
+= 2 * sizeof(unsigned long long);
2078 if (flags
& SLAB_STORE_USER
) {
2079 /* user store requires one word storage behind the end of
2080 * the real object. But if the second red zone needs to be
2081 * aligned to 64 bits, we must allow that much space.
2083 if (flags
& SLAB_RED_ZONE
)
2084 size
+= REDZONE_ALIGN
;
2086 size
+= BYTES_PER_WORD
;
2090 kasan_cache_create(cachep
, &size
, &flags
);
2092 size
= ALIGN(size
, cachep
->align
);
2094 * We should restrict the number of objects in a slab to implement
2095 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2097 if (FREELIST_BYTE_INDEX
&& size
< SLAB_OBJ_MIN_SIZE
)
2098 size
= ALIGN(SLAB_OBJ_MIN_SIZE
, cachep
->align
);
2102 * To activate debug pagealloc, off-slab management is necessary
2103 * requirement. In early phase of initialization, small sized slab
2104 * doesn't get initialized so it would not be possible. So, we need
2105 * to check size >= 256. It guarantees that all necessary small
2106 * sized slab is initialized in current slab initialization sequence.
2108 if (debug_pagealloc_enabled() && (flags
& SLAB_POISON
) &&
2109 size
>= 256 && cachep
->object_size
> cache_line_size()) {
2110 if (size
< PAGE_SIZE
|| size
% PAGE_SIZE
== 0) {
2111 size_t tmp_size
= ALIGN(size
, PAGE_SIZE
);
2113 if (set_off_slab_cache(cachep
, tmp_size
, flags
)) {
2114 flags
|= CFLGS_OFF_SLAB
;
2115 cachep
->obj_offset
+= tmp_size
- size
;
2123 if (set_objfreelist_slab_cache(cachep
, size
, flags
)) {
2124 flags
|= CFLGS_OBJFREELIST_SLAB
;
2128 if (set_off_slab_cache(cachep
, size
, flags
)) {
2129 flags
|= CFLGS_OFF_SLAB
;
2133 if (set_on_slab_cache(cachep
, size
, flags
))
2139 cachep
->freelist_size
= cachep
->num
* sizeof(freelist_idx_t
);
2140 cachep
->flags
= flags
;
2141 cachep
->allocflags
= __GFP_COMP
;
2142 if (flags
& SLAB_CACHE_DMA
)
2143 cachep
->allocflags
|= GFP_DMA
;
2144 if (flags
& SLAB_RECLAIM_ACCOUNT
)
2145 cachep
->allocflags
|= __GFP_RECLAIMABLE
;
2146 cachep
->size
= size
;
2147 cachep
->reciprocal_buffer_size
= reciprocal_value(size
);
2151 * If we're going to use the generic kernel_map_pages()
2152 * poisoning, then it's going to smash the contents of
2153 * the redzone and userword anyhow, so switch them off.
2155 if (IS_ENABLED(CONFIG_PAGE_POISONING
) &&
2156 (cachep
->flags
& SLAB_POISON
) &&
2157 is_debug_pagealloc_cache(cachep
))
2158 cachep
->flags
&= ~(SLAB_RED_ZONE
| SLAB_STORE_USER
);
2161 if (OFF_SLAB(cachep
)) {
2162 cachep
->freelist_cache
=
2163 kmalloc_slab(cachep
->freelist_size
, 0u);
2166 err
= setup_cpu_cache(cachep
, gfp
);
2168 __kmem_cache_release(cachep
);
2176 static void check_irq_off(void)
2178 BUG_ON(!irqs_disabled());
2181 static void check_irq_on(void)
2183 BUG_ON(irqs_disabled());
2186 static void check_mutex_acquired(void)
2188 BUG_ON(!mutex_is_locked(&slab_mutex
));
2191 static void check_spinlock_acquired(struct kmem_cache
*cachep
)
2195 assert_spin_locked(&get_node(cachep
, numa_mem_id())->list_lock
);
2199 static void check_spinlock_acquired_node(struct kmem_cache
*cachep
, int node
)
2203 assert_spin_locked(&get_node(cachep
, node
)->list_lock
);
2208 #define check_irq_off() do { } while(0)
2209 #define check_irq_on() do { } while(0)
2210 #define check_mutex_acquired() do { } while(0)
2211 #define check_spinlock_acquired(x) do { } while(0)
2212 #define check_spinlock_acquired_node(x, y) do { } while(0)
2215 static void drain_array_locked(struct kmem_cache
*cachep
, struct array_cache
*ac
,
2216 int node
, bool free_all
, struct list_head
*list
)
2220 if (!ac
|| !ac
->avail
)
2223 tofree
= free_all
? ac
->avail
: (ac
->limit
+ 4) / 5;
2224 if (tofree
> ac
->avail
)
2225 tofree
= (ac
->avail
+ 1) / 2;
2227 free_block(cachep
, ac
->entry
, tofree
, node
, list
);
2228 ac
->avail
-= tofree
;
2229 memmove(ac
->entry
, &(ac
->entry
[tofree
]), sizeof(void *) * ac
->avail
);
2232 static void do_drain(void *arg
)
2234 struct kmem_cache
*cachep
= arg
;
2235 struct array_cache
*ac
;
2236 int node
= numa_mem_id();
2237 struct kmem_cache_node
*n
;
2241 ac
= cpu_cache_get(cachep
);
2242 n
= get_node(cachep
, node
);
2243 spin_lock(&n
->list_lock
);
2244 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
2245 spin_unlock(&n
->list_lock
);
2246 slabs_destroy(cachep
, &list
);
2250 static void drain_cpu_caches(struct kmem_cache
*cachep
)
2252 struct kmem_cache_node
*n
;
2256 on_each_cpu(do_drain
, cachep
, 1);
2258 for_each_kmem_cache_node(cachep
, node
, n
)
2260 drain_alien_cache(cachep
, n
->alien
);
2262 for_each_kmem_cache_node(cachep
, node
, n
) {
2263 spin_lock_irq(&n
->list_lock
);
2264 drain_array_locked(cachep
, n
->shared
, node
, true, &list
);
2265 spin_unlock_irq(&n
->list_lock
);
2267 slabs_destroy(cachep
, &list
);
2272 * Remove slabs from the list of free slabs.
2273 * Specify the number of slabs to drain in tofree.
2275 * Returns the actual number of slabs released.
2277 static int drain_freelist(struct kmem_cache
*cache
,
2278 struct kmem_cache_node
*n
, int tofree
)
2280 struct list_head
*p
;
2285 while (nr_freed
< tofree
&& !list_empty(&n
->slabs_free
)) {
2287 spin_lock_irq(&n
->list_lock
);
2288 p
= n
->slabs_free
.prev
;
2289 if (p
== &n
->slabs_free
) {
2290 spin_unlock_irq(&n
->list_lock
);
2294 page
= list_entry(p
, struct page
, lru
);
2295 list_del(&page
->lru
);
2299 * Safe to drop the lock. The slab is no longer linked
2302 n
->free_objects
-= cache
->num
;
2303 spin_unlock_irq(&n
->list_lock
);
2304 slab_destroy(cache
, page
);
2311 int __kmem_cache_shrink(struct kmem_cache
*cachep
)
2315 struct kmem_cache_node
*n
;
2317 drain_cpu_caches(cachep
);
2320 for_each_kmem_cache_node(cachep
, node
, n
) {
2321 drain_freelist(cachep
, n
, INT_MAX
);
2323 ret
+= !list_empty(&n
->slabs_full
) ||
2324 !list_empty(&n
->slabs_partial
);
2326 return (ret
? 1 : 0);
2330 void __kmemcg_cache_deactivate(struct kmem_cache
*cachep
)
2332 __kmem_cache_shrink(cachep
);
2336 int __kmem_cache_shutdown(struct kmem_cache
*cachep
)
2338 return __kmem_cache_shrink(cachep
);
2341 void __kmem_cache_release(struct kmem_cache
*cachep
)
2344 struct kmem_cache_node
*n
;
2346 cache_random_seq_destroy(cachep
);
2348 free_percpu(cachep
->cpu_cache
);
2350 /* NUMA: free the node structures */
2351 for_each_kmem_cache_node(cachep
, i
, n
) {
2353 free_alien_cache(n
->alien
);
2355 cachep
->node
[i
] = NULL
;
2360 * Get the memory for a slab management obj.
2362 * For a slab cache when the slab descriptor is off-slab, the
2363 * slab descriptor can't come from the same cache which is being created,
2364 * Because if it is the case, that means we defer the creation of
2365 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2366 * And we eventually call down to __kmem_cache_create(), which
2367 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2368 * This is a "chicken-and-egg" problem.
2370 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2371 * which are all initialized during kmem_cache_init().
2373 static void *alloc_slabmgmt(struct kmem_cache
*cachep
,
2374 struct page
*page
, int colour_off
,
2375 gfp_t local_flags
, int nodeid
)
2378 void *addr
= page_address(page
);
2380 page
->s_mem
= addr
+ colour_off
;
2383 if (OBJFREELIST_SLAB(cachep
))
2385 else if (OFF_SLAB(cachep
)) {
2386 /* Slab management obj is off-slab. */
2387 freelist
= kmem_cache_alloc_node(cachep
->freelist_cache
,
2388 local_flags
, nodeid
);
2392 /* We will use last bytes at the slab for freelist */
2393 freelist
= addr
+ (PAGE_SIZE
<< cachep
->gfporder
) -
2394 cachep
->freelist_size
;
2400 static inline freelist_idx_t
get_free_obj(struct page
*page
, unsigned int idx
)
2402 return ((freelist_idx_t
*)page
->freelist
)[idx
];
2405 static inline void set_free_obj(struct page
*page
,
2406 unsigned int idx
, freelist_idx_t val
)
2408 ((freelist_idx_t
*)(page
->freelist
))[idx
] = val
;
2411 static void cache_init_objs_debug(struct kmem_cache
*cachep
, struct page
*page
)
2416 for (i
= 0; i
< cachep
->num
; i
++) {
2417 void *objp
= index_to_obj(cachep
, page
, i
);
2419 if (cachep
->flags
& SLAB_STORE_USER
)
2420 *dbg_userword(cachep
, objp
) = NULL
;
2422 if (cachep
->flags
& SLAB_RED_ZONE
) {
2423 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2424 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2427 * Constructors are not allowed to allocate memory from the same
2428 * cache which they are a constructor for. Otherwise, deadlock.
2429 * They must also be threaded.
2431 if (cachep
->ctor
&& !(cachep
->flags
& SLAB_POISON
)) {
2432 kasan_unpoison_object_data(cachep
,
2433 objp
+ obj_offset(cachep
));
2434 cachep
->ctor(objp
+ obj_offset(cachep
));
2435 kasan_poison_object_data(
2436 cachep
, objp
+ obj_offset(cachep
));
2439 if (cachep
->flags
& SLAB_RED_ZONE
) {
2440 if (*dbg_redzone2(cachep
, objp
) != RED_INACTIVE
)
2441 slab_error(cachep
, "constructor overwrote the end of an object");
2442 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
)
2443 slab_error(cachep
, "constructor overwrote the start of an object");
2445 /* need to poison the objs? */
2446 if (cachep
->flags
& SLAB_POISON
) {
2447 poison_obj(cachep
, objp
, POISON_FREE
);
2448 slab_kernel_map(cachep
, objp
, 0, 0);
2454 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2455 /* Hold information during a freelist initialization */
2456 union freelist_init_state
{
2462 struct rnd_state rnd_state
;
2466 * Initialize the state based on the randomization methode available.
2467 * return true if the pre-computed list is available, false otherwize.
2469 static bool freelist_state_initialize(union freelist_init_state
*state
,
2470 struct kmem_cache
*cachep
,
2476 /* Use best entropy available to define a random shift */
2477 rand
= get_random_int();
2479 /* Use a random state if the pre-computed list is not available */
2480 if (!cachep
->random_seq
) {
2481 prandom_seed_state(&state
->rnd_state
, rand
);
2484 state
->list
= cachep
->random_seq
;
2485 state
->count
= count
;
2486 state
->pos
= rand
% count
;
2492 /* Get the next entry on the list and randomize it using a random shift */
2493 static freelist_idx_t
next_random_slot(union freelist_init_state
*state
)
2495 if (state
->pos
>= state
->count
)
2497 return state
->list
[state
->pos
++];
2500 /* Swap two freelist entries */
2501 static void swap_free_obj(struct page
*page
, unsigned int a
, unsigned int b
)
2503 swap(((freelist_idx_t
*)page
->freelist
)[a
],
2504 ((freelist_idx_t
*)page
->freelist
)[b
]);
2508 * Shuffle the freelist initialization state based on pre-computed lists.
2509 * return true if the list was successfully shuffled, false otherwise.
2511 static bool shuffle_freelist(struct kmem_cache
*cachep
, struct page
*page
)
2513 unsigned int objfreelist
= 0, i
, rand
, count
= cachep
->num
;
2514 union freelist_init_state state
;
2520 precomputed
= freelist_state_initialize(&state
, cachep
, count
);
2522 /* Take a random entry as the objfreelist */
2523 if (OBJFREELIST_SLAB(cachep
)) {
2525 objfreelist
= count
- 1;
2527 objfreelist
= next_random_slot(&state
);
2528 page
->freelist
= index_to_obj(cachep
, page
, objfreelist
) +
2534 * On early boot, generate the list dynamically.
2535 * Later use a pre-computed list for speed.
2538 for (i
= 0; i
< count
; i
++)
2539 set_free_obj(page
, i
, i
);
2541 /* Fisher-Yates shuffle */
2542 for (i
= count
- 1; i
> 0; i
--) {
2543 rand
= prandom_u32_state(&state
.rnd_state
);
2545 swap_free_obj(page
, i
, rand
);
2548 for (i
= 0; i
< count
; i
++)
2549 set_free_obj(page
, i
, next_random_slot(&state
));
2552 if (OBJFREELIST_SLAB(cachep
))
2553 set_free_obj(page
, cachep
->num
- 1, objfreelist
);
2558 static inline bool shuffle_freelist(struct kmem_cache
*cachep
,
2563 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2565 static void cache_init_objs(struct kmem_cache
*cachep
,
2572 cache_init_objs_debug(cachep
, page
);
2574 /* Try to randomize the freelist if enabled */
2575 shuffled
= shuffle_freelist(cachep
, page
);
2577 if (!shuffled
&& OBJFREELIST_SLAB(cachep
)) {
2578 page
->freelist
= index_to_obj(cachep
, page
, cachep
->num
- 1) +
2582 for (i
= 0; i
< cachep
->num
; i
++) {
2583 objp
= index_to_obj(cachep
, page
, i
);
2584 kasan_init_slab_obj(cachep
, objp
);
2586 /* constructor could break poison info */
2587 if (DEBUG
== 0 && cachep
->ctor
) {
2588 kasan_unpoison_object_data(cachep
, objp
);
2590 kasan_poison_object_data(cachep
, objp
);
2594 set_free_obj(page
, i
, i
);
2598 static void *slab_get_obj(struct kmem_cache
*cachep
, struct page
*page
)
2602 objp
= index_to_obj(cachep
, page
, get_free_obj(page
, page
->active
));
2606 if (cachep
->flags
& SLAB_STORE_USER
)
2607 set_store_user_dirty(cachep
);
2613 static void slab_put_obj(struct kmem_cache
*cachep
,
2614 struct page
*page
, void *objp
)
2616 unsigned int objnr
= obj_to_index(cachep
, page
, objp
);
2620 /* Verify double free bug */
2621 for (i
= page
->active
; i
< cachep
->num
; i
++) {
2622 if (get_free_obj(page
, i
) == objnr
) {
2623 pr_err("slab: double free detected in cache '%s', objp %p\n",
2624 cachep
->name
, objp
);
2630 if (!page
->freelist
)
2631 page
->freelist
= objp
+ obj_offset(cachep
);
2633 set_free_obj(page
, page
->active
, objnr
);
2637 * Map pages beginning at addr to the given cache and slab. This is required
2638 * for the slab allocator to be able to lookup the cache and slab of a
2639 * virtual address for kfree, ksize, and slab debugging.
2641 static void slab_map_pages(struct kmem_cache
*cache
, struct page
*page
,
2644 page
->slab_cache
= cache
;
2645 page
->freelist
= freelist
;
2649 * Grow (by 1) the number of slabs within a cache. This is called by
2650 * kmem_cache_alloc() when there are no active objs left in a cache.
2652 static struct page
*cache_grow_begin(struct kmem_cache
*cachep
,
2653 gfp_t flags
, int nodeid
)
2659 struct kmem_cache_node
*n
;
2663 * Be lazy and only check for valid flags here, keeping it out of the
2664 * critical path in kmem_cache_alloc().
2666 if (unlikely(flags
& GFP_SLAB_BUG_MASK
)) {
2667 gfp_t invalid_mask
= flags
& GFP_SLAB_BUG_MASK
;
2668 flags
&= ~GFP_SLAB_BUG_MASK
;
2669 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2670 invalid_mask
, &invalid_mask
, flags
, &flags
);
2673 local_flags
= flags
& (GFP_CONSTRAINT_MASK
|GFP_RECLAIM_MASK
);
2676 if (gfpflags_allow_blocking(local_flags
))
2680 * Get mem for the objs. Attempt to allocate a physical page from
2683 page
= kmem_getpages(cachep
, local_flags
, nodeid
);
2687 page_node
= page_to_nid(page
);
2688 n
= get_node(cachep
, page_node
);
2690 /* Get colour for the slab, and cal the next value. */
2692 if (n
->colour_next
>= cachep
->colour
)
2695 offset
= n
->colour_next
;
2696 if (offset
>= cachep
->colour
)
2699 offset
*= cachep
->colour_off
;
2701 /* Get slab management. */
2702 freelist
= alloc_slabmgmt(cachep
, page
, offset
,
2703 local_flags
& ~GFP_CONSTRAINT_MASK
, page_node
);
2704 if (OFF_SLAB(cachep
) && !freelist
)
2707 slab_map_pages(cachep
, page
, freelist
);
2709 kasan_poison_slab(page
);
2710 cache_init_objs(cachep
, page
);
2712 if (gfpflags_allow_blocking(local_flags
))
2713 local_irq_disable();
2718 kmem_freepages(cachep
, page
);
2720 if (gfpflags_allow_blocking(local_flags
))
2721 local_irq_disable();
2725 static void cache_grow_end(struct kmem_cache
*cachep
, struct page
*page
)
2727 struct kmem_cache_node
*n
;
2735 INIT_LIST_HEAD(&page
->lru
);
2736 n
= get_node(cachep
, page_to_nid(page
));
2738 spin_lock(&n
->list_lock
);
2740 if (!page
->active
) {
2741 list_add_tail(&page
->lru
, &(n
->slabs_free
));
2744 fixup_slab_list(cachep
, n
, page
, &list
);
2746 STATS_INC_GROWN(cachep
);
2747 n
->free_objects
+= cachep
->num
- page
->active
;
2748 spin_unlock(&n
->list_lock
);
2750 fixup_objfreelist_debug(cachep
, &list
);
2756 * Perform extra freeing checks:
2757 * - detect bad pointers.
2758 * - POISON/RED_ZONE checking
2760 static void kfree_debugcheck(const void *objp
)
2762 if (!virt_addr_valid(objp
)) {
2763 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2764 (unsigned long)objp
);
2769 static inline void verify_redzone_free(struct kmem_cache
*cache
, void *obj
)
2771 unsigned long long redzone1
, redzone2
;
2773 redzone1
= *dbg_redzone1(cache
, obj
);
2774 redzone2
= *dbg_redzone2(cache
, obj
);
2779 if (redzone1
== RED_ACTIVE
&& redzone2
== RED_ACTIVE
)
2782 if (redzone1
== RED_INACTIVE
&& redzone2
== RED_INACTIVE
)
2783 slab_error(cache
, "double free detected");
2785 slab_error(cache
, "memory outside object was overwritten");
2787 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2788 obj
, redzone1
, redzone2
);
2791 static void *cache_free_debugcheck(struct kmem_cache
*cachep
, void *objp
,
2792 unsigned long caller
)
2797 BUG_ON(virt_to_cache(objp
) != cachep
);
2799 objp
-= obj_offset(cachep
);
2800 kfree_debugcheck(objp
);
2801 page
= virt_to_head_page(objp
);
2803 if (cachep
->flags
& SLAB_RED_ZONE
) {
2804 verify_redzone_free(cachep
, objp
);
2805 *dbg_redzone1(cachep
, objp
) = RED_INACTIVE
;
2806 *dbg_redzone2(cachep
, objp
) = RED_INACTIVE
;
2808 if (cachep
->flags
& SLAB_STORE_USER
) {
2809 set_store_user_dirty(cachep
);
2810 *dbg_userword(cachep
, objp
) = (void *)caller
;
2813 objnr
= obj_to_index(cachep
, page
, objp
);
2815 BUG_ON(objnr
>= cachep
->num
);
2816 BUG_ON(objp
!= index_to_obj(cachep
, page
, objnr
));
2818 if (cachep
->flags
& SLAB_POISON
) {
2819 poison_obj(cachep
, objp
, POISON_FREE
);
2820 slab_kernel_map(cachep
, objp
, 0, caller
);
2826 #define kfree_debugcheck(x) do { } while(0)
2827 #define cache_free_debugcheck(x,objp,z) (objp)
2830 static inline void fixup_objfreelist_debug(struct kmem_cache
*cachep
,
2838 objp
= next
- obj_offset(cachep
);
2839 next
= *(void **)next
;
2840 poison_obj(cachep
, objp
, POISON_FREE
);
2845 static inline void fixup_slab_list(struct kmem_cache
*cachep
,
2846 struct kmem_cache_node
*n
, struct page
*page
,
2849 /* move slabp to correct slabp list: */
2850 list_del(&page
->lru
);
2851 if (page
->active
== cachep
->num
) {
2852 list_add(&page
->lru
, &n
->slabs_full
);
2853 if (OBJFREELIST_SLAB(cachep
)) {
2855 /* Poisoning will be done without holding the lock */
2856 if (cachep
->flags
& SLAB_POISON
) {
2857 void **objp
= page
->freelist
;
2863 page
->freelist
= NULL
;
2866 list_add(&page
->lru
, &n
->slabs_partial
);
2869 /* Try to find non-pfmemalloc slab if needed */
2870 static noinline
struct page
*get_valid_first_slab(struct kmem_cache_node
*n
,
2871 struct page
*page
, bool pfmemalloc
)
2879 if (!PageSlabPfmemalloc(page
))
2882 /* No need to keep pfmemalloc slab if we have enough free objects */
2883 if (n
->free_objects
> n
->free_limit
) {
2884 ClearPageSlabPfmemalloc(page
);
2888 /* Move pfmemalloc slab to the end of list to speed up next search */
2889 list_del(&page
->lru
);
2890 if (!page
->active
) {
2891 list_add_tail(&page
->lru
, &n
->slabs_free
);
2894 list_add_tail(&page
->lru
, &n
->slabs_partial
);
2896 list_for_each_entry(page
, &n
->slabs_partial
, lru
) {
2897 if (!PageSlabPfmemalloc(page
))
2901 n
->free_touched
= 1;
2902 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
2903 if (!PageSlabPfmemalloc(page
)) {
2912 static struct page
*get_first_slab(struct kmem_cache_node
*n
, bool pfmemalloc
)
2916 assert_spin_locked(&n
->list_lock
);
2917 page
= list_first_entry_or_null(&n
->slabs_partial
, struct page
, lru
);
2919 n
->free_touched
= 1;
2920 page
= list_first_entry_or_null(&n
->slabs_free
, struct page
,
2926 if (sk_memalloc_socks())
2927 page
= get_valid_first_slab(n
, page
, pfmemalloc
);
2932 static noinline
void *cache_alloc_pfmemalloc(struct kmem_cache
*cachep
,
2933 struct kmem_cache_node
*n
, gfp_t flags
)
2939 if (!gfp_pfmemalloc_allowed(flags
))
2942 spin_lock(&n
->list_lock
);
2943 page
= get_first_slab(n
, true);
2945 spin_unlock(&n
->list_lock
);
2949 obj
= slab_get_obj(cachep
, page
);
2952 fixup_slab_list(cachep
, n
, page
, &list
);
2954 spin_unlock(&n
->list_lock
);
2955 fixup_objfreelist_debug(cachep
, &list
);
2961 * Slab list should be fixed up by fixup_slab_list() for existing slab
2962 * or cache_grow_end() for new slab
2964 static __always_inline
int alloc_block(struct kmem_cache
*cachep
,
2965 struct array_cache
*ac
, struct page
*page
, int batchcount
)
2968 * There must be at least one object available for
2971 BUG_ON(page
->active
>= cachep
->num
);
2973 while (page
->active
< cachep
->num
&& batchcount
--) {
2974 STATS_INC_ALLOCED(cachep
);
2975 STATS_INC_ACTIVE(cachep
);
2976 STATS_SET_HIGH(cachep
);
2978 ac
->entry
[ac
->avail
++] = slab_get_obj(cachep
, page
);
2984 static void *cache_alloc_refill(struct kmem_cache
*cachep
, gfp_t flags
)
2987 struct kmem_cache_node
*n
;
2988 struct array_cache
*ac
, *shared
;
2994 node
= numa_mem_id();
2996 ac
= cpu_cache_get(cachep
);
2997 batchcount
= ac
->batchcount
;
2998 if (!ac
->touched
&& batchcount
> BATCHREFILL_LIMIT
) {
3000 * If there was little recent activity on this cache, then
3001 * perform only a partial refill. Otherwise we could generate
3004 batchcount
= BATCHREFILL_LIMIT
;
3006 n
= get_node(cachep
, node
);
3008 BUG_ON(ac
->avail
> 0 || !n
);
3009 shared
= READ_ONCE(n
->shared
);
3010 if (!n
->free_objects
&& (!shared
|| !shared
->avail
))
3013 spin_lock(&n
->list_lock
);
3014 shared
= READ_ONCE(n
->shared
);
3016 /* See if we can refill from the shared array */
3017 if (shared
&& transfer_objects(ac
, shared
, batchcount
)) {
3018 shared
->touched
= 1;
3022 while (batchcount
> 0) {
3023 /* Get slab alloc is to come from. */
3024 page
= get_first_slab(n
, false);
3028 check_spinlock_acquired(cachep
);
3030 batchcount
= alloc_block(cachep
, ac
, page
, batchcount
);
3031 fixup_slab_list(cachep
, n
, page
, &list
);
3035 n
->free_objects
-= ac
->avail
;
3037 spin_unlock(&n
->list_lock
);
3038 fixup_objfreelist_debug(cachep
, &list
);
3041 if (unlikely(!ac
->avail
)) {
3042 /* Check if we can use obj in pfmemalloc slab */
3043 if (sk_memalloc_socks()) {
3044 void *obj
= cache_alloc_pfmemalloc(cachep
, n
, flags
);
3050 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), node
);
3053 * cache_grow_begin() can reenable interrupts,
3054 * then ac could change.
3056 ac
= cpu_cache_get(cachep
);
3057 if (!ac
->avail
&& page
)
3058 alloc_block(cachep
, ac
, page
, batchcount
);
3059 cache_grow_end(cachep
, page
);
3066 return ac
->entry
[--ac
->avail
];
3069 static inline void cache_alloc_debugcheck_before(struct kmem_cache
*cachep
,
3072 might_sleep_if(gfpflags_allow_blocking(flags
));
3076 static void *cache_alloc_debugcheck_after(struct kmem_cache
*cachep
,
3077 gfp_t flags
, void *objp
, unsigned long caller
)
3081 if (cachep
->flags
& SLAB_POISON
) {
3082 check_poison_obj(cachep
, objp
);
3083 slab_kernel_map(cachep
, objp
, 1, 0);
3084 poison_obj(cachep
, objp
, POISON_INUSE
);
3086 if (cachep
->flags
& SLAB_STORE_USER
)
3087 *dbg_userword(cachep
, objp
) = (void *)caller
;
3089 if (cachep
->flags
& SLAB_RED_ZONE
) {
3090 if (*dbg_redzone1(cachep
, objp
) != RED_INACTIVE
||
3091 *dbg_redzone2(cachep
, objp
) != RED_INACTIVE
) {
3092 slab_error(cachep
, "double free, or memory outside object was overwritten");
3093 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3094 objp
, *dbg_redzone1(cachep
, objp
),
3095 *dbg_redzone2(cachep
, objp
));
3097 *dbg_redzone1(cachep
, objp
) = RED_ACTIVE
;
3098 *dbg_redzone2(cachep
, objp
) = RED_ACTIVE
;
3101 objp
+= obj_offset(cachep
);
3102 if (cachep
->ctor
&& cachep
->flags
& SLAB_POISON
)
3104 if (ARCH_SLAB_MINALIGN
&&
3105 ((unsigned long)objp
& (ARCH_SLAB_MINALIGN
-1))) {
3106 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3107 objp
, (int)ARCH_SLAB_MINALIGN
);
3112 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3115 static inline void *____cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3118 struct array_cache
*ac
;
3122 ac
= cpu_cache_get(cachep
);
3123 if (likely(ac
->avail
)) {
3125 objp
= ac
->entry
[--ac
->avail
];
3127 STATS_INC_ALLOCHIT(cachep
);
3131 STATS_INC_ALLOCMISS(cachep
);
3132 objp
= cache_alloc_refill(cachep
, flags
);
3134 * the 'ac' may be updated by cache_alloc_refill(),
3135 * and kmemleak_erase() requires its correct value.
3137 ac
= cpu_cache_get(cachep
);
3141 * To avoid a false negative, if an object that is in one of the
3142 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3143 * treat the array pointers as a reference to the object.
3146 kmemleak_erase(&ac
->entry
[ac
->avail
]);
3152 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3154 * If we are in_interrupt, then process context, including cpusets and
3155 * mempolicy, may not apply and should not be used for allocation policy.
3157 static void *alternate_node_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3159 int nid_alloc
, nid_here
;
3161 if (in_interrupt() || (flags
& __GFP_THISNODE
))
3163 nid_alloc
= nid_here
= numa_mem_id();
3164 if (cpuset_do_slab_mem_spread() && (cachep
->flags
& SLAB_MEM_SPREAD
))
3165 nid_alloc
= cpuset_slab_spread_node();
3166 else if (current
->mempolicy
)
3167 nid_alloc
= mempolicy_slab_node();
3168 if (nid_alloc
!= nid_here
)
3169 return ____cache_alloc_node(cachep
, flags
, nid_alloc
);
3174 * Fallback function if there was no memory available and no objects on a
3175 * certain node and fall back is permitted. First we scan all the
3176 * available node for available objects. If that fails then we
3177 * perform an allocation without specifying a node. This allows the page
3178 * allocator to do its reclaim / fallback magic. We then insert the
3179 * slab into the proper nodelist and then allocate from it.
3181 static void *fallback_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3183 struct zonelist
*zonelist
;
3186 enum zone_type high_zoneidx
= gfp_zone(flags
);
3190 unsigned int cpuset_mems_cookie
;
3192 if (flags
& __GFP_THISNODE
)
3196 cpuset_mems_cookie
= read_mems_allowed_begin();
3197 zonelist
= node_zonelist(mempolicy_slab_node(), flags
);
3201 * Look through allowed nodes for objects available
3202 * from existing per node queues.
3204 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
3205 nid
= zone_to_nid(zone
);
3207 if (cpuset_zone_allowed(zone
, flags
) &&
3208 get_node(cache
, nid
) &&
3209 get_node(cache
, nid
)->free_objects
) {
3210 obj
= ____cache_alloc_node(cache
,
3211 gfp_exact_node(flags
), nid
);
3219 * This allocation will be performed within the constraints
3220 * of the current cpuset / memory policy requirements.
3221 * We may trigger various forms of reclaim on the allowed
3222 * set and go into memory reserves if necessary.
3224 page
= cache_grow_begin(cache
, flags
, numa_mem_id());
3225 cache_grow_end(cache
, page
);
3227 nid
= page_to_nid(page
);
3228 obj
= ____cache_alloc_node(cache
,
3229 gfp_exact_node(flags
), nid
);
3232 * Another processor may allocate the objects in
3233 * the slab since we are not holding any locks.
3240 if (unlikely(!obj
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3246 * A interface to enable slab creation on nodeid
3248 static void *____cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
,
3252 struct kmem_cache_node
*n
;
3256 VM_BUG_ON(nodeid
< 0 || nodeid
>= MAX_NUMNODES
);
3257 n
= get_node(cachep
, nodeid
);
3261 spin_lock(&n
->list_lock
);
3262 page
= get_first_slab(n
, false);
3266 check_spinlock_acquired_node(cachep
, nodeid
);
3268 STATS_INC_NODEALLOCS(cachep
);
3269 STATS_INC_ACTIVE(cachep
);
3270 STATS_SET_HIGH(cachep
);
3272 BUG_ON(page
->active
== cachep
->num
);
3274 obj
= slab_get_obj(cachep
, page
);
3277 fixup_slab_list(cachep
, n
, page
, &list
);
3279 spin_unlock(&n
->list_lock
);
3280 fixup_objfreelist_debug(cachep
, &list
);
3284 spin_unlock(&n
->list_lock
);
3285 page
= cache_grow_begin(cachep
, gfp_exact_node(flags
), nodeid
);
3287 /* This slab isn't counted yet so don't update free_objects */
3288 obj
= slab_get_obj(cachep
, page
);
3290 cache_grow_end(cachep
, page
);
3292 return obj
? obj
: fallback_alloc(cachep
, flags
);
3295 static __always_inline
void *
3296 slab_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
,
3297 unsigned long caller
)
3299 unsigned long save_flags
;
3301 int slab_node
= numa_mem_id();
3303 flags
&= gfp_allowed_mask
;
3304 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3305 if (unlikely(!cachep
))
3308 cache_alloc_debugcheck_before(cachep
, flags
);
3309 local_irq_save(save_flags
);
3311 if (nodeid
== NUMA_NO_NODE
)
3314 if (unlikely(!get_node(cachep
, nodeid
))) {
3315 /* Node not bootstrapped yet */
3316 ptr
= fallback_alloc(cachep
, flags
);
3320 if (nodeid
== slab_node
) {
3322 * Use the locally cached objects if possible.
3323 * However ____cache_alloc does not allow fallback
3324 * to other nodes. It may fail while we still have
3325 * objects on other nodes available.
3327 ptr
= ____cache_alloc(cachep
, flags
);
3331 /* ___cache_alloc_node can fall back to other nodes */
3332 ptr
= ____cache_alloc_node(cachep
, flags
, nodeid
);
3334 local_irq_restore(save_flags
);
3335 ptr
= cache_alloc_debugcheck_after(cachep
, flags
, ptr
, caller
);
3337 if (unlikely(flags
& __GFP_ZERO
) && ptr
)
3338 memset(ptr
, 0, cachep
->object_size
);
3340 slab_post_alloc_hook(cachep
, flags
, 1, &ptr
);
3344 static __always_inline
void *
3345 __do_cache_alloc(struct kmem_cache
*cache
, gfp_t flags
)
3349 if (current
->mempolicy
|| cpuset_do_slab_mem_spread()) {
3350 objp
= alternate_node_alloc(cache
, flags
);
3354 objp
= ____cache_alloc(cache
, flags
);
3357 * We may just have run out of memory on the local node.
3358 * ____cache_alloc_node() knows how to locate memory on other nodes
3361 objp
= ____cache_alloc_node(cache
, flags
, numa_mem_id());
3368 static __always_inline
void *
3369 __do_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3371 return ____cache_alloc(cachep
, flags
);
3374 #endif /* CONFIG_NUMA */
3376 static __always_inline
void *
3377 slab_alloc(struct kmem_cache
*cachep
, gfp_t flags
, unsigned long caller
)
3379 unsigned long save_flags
;
3382 flags
&= gfp_allowed_mask
;
3383 cachep
= slab_pre_alloc_hook(cachep
, flags
);
3384 if (unlikely(!cachep
))
3387 cache_alloc_debugcheck_before(cachep
, flags
);
3388 local_irq_save(save_flags
);
3389 objp
= __do_cache_alloc(cachep
, flags
);
3390 local_irq_restore(save_flags
);
3391 objp
= cache_alloc_debugcheck_after(cachep
, flags
, objp
, caller
);
3394 if (unlikely(flags
& __GFP_ZERO
) && objp
)
3395 memset(objp
, 0, cachep
->object_size
);
3397 slab_post_alloc_hook(cachep
, flags
, 1, &objp
);
3402 * Caller needs to acquire correct kmem_cache_node's list_lock
3403 * @list: List of detached free slabs should be freed by caller
3405 static void free_block(struct kmem_cache
*cachep
, void **objpp
,
3406 int nr_objects
, int node
, struct list_head
*list
)
3409 struct kmem_cache_node
*n
= get_node(cachep
, node
);
3412 n
->free_objects
+= nr_objects
;
3414 for (i
= 0; i
< nr_objects
; i
++) {
3420 page
= virt_to_head_page(objp
);
3421 list_del(&page
->lru
);
3422 check_spinlock_acquired_node(cachep
, node
);
3423 slab_put_obj(cachep
, page
, objp
);
3424 STATS_DEC_ACTIVE(cachep
);
3426 /* fixup slab chains */
3427 if (page
->active
== 0) {
3428 list_add(&page
->lru
, &n
->slabs_free
);
3431 /* Unconditionally move a slab to the end of the
3432 * partial list on free - maximum time for the
3433 * other objects to be freed, too.
3435 list_add_tail(&page
->lru
, &n
->slabs_partial
);
3439 while (n
->free_objects
> n
->free_limit
&& !list_empty(&n
->slabs_free
)) {
3440 n
->free_objects
-= cachep
->num
;
3442 page
= list_last_entry(&n
->slabs_free
, struct page
, lru
);
3443 list_move(&page
->lru
, list
);
3449 static void cache_flusharray(struct kmem_cache
*cachep
, struct array_cache
*ac
)
3452 struct kmem_cache_node
*n
;
3453 int node
= numa_mem_id();
3456 batchcount
= ac
->batchcount
;
3459 n
= get_node(cachep
, node
);
3460 spin_lock(&n
->list_lock
);
3462 struct array_cache
*shared_array
= n
->shared
;
3463 int max
= shared_array
->limit
- shared_array
->avail
;
3465 if (batchcount
> max
)
3467 memcpy(&(shared_array
->entry
[shared_array
->avail
]),
3468 ac
->entry
, sizeof(void *) * batchcount
);
3469 shared_array
->avail
+= batchcount
;
3474 free_block(cachep
, ac
->entry
, batchcount
, node
, &list
);
3481 list_for_each_entry(page
, &n
->slabs_free
, lru
) {
3482 BUG_ON(page
->active
);
3486 STATS_SET_FREEABLE(cachep
, i
);
3489 spin_unlock(&n
->list_lock
);
3490 slabs_destroy(cachep
, &list
);
3491 ac
->avail
-= batchcount
;
3492 memmove(ac
->entry
, &(ac
->entry
[batchcount
]), sizeof(void *)*ac
->avail
);
3496 * Release an obj back to its cache. If the obj has a constructed state, it must
3497 * be in this state _before_ it is released. Called with disabled ints.
3499 static inline void __cache_free(struct kmem_cache
*cachep
, void *objp
,
3500 unsigned long caller
)
3502 /* Put the object into the quarantine, don't touch it for now. */
3503 if (kasan_slab_free(cachep
, objp
))
3506 ___cache_free(cachep
, objp
, caller
);
3509 void ___cache_free(struct kmem_cache
*cachep
, void *objp
,
3510 unsigned long caller
)
3512 struct array_cache
*ac
= cpu_cache_get(cachep
);
3515 kmemleak_free_recursive(objp
, cachep
->flags
);
3516 objp
= cache_free_debugcheck(cachep
, objp
, caller
);
3518 kmemcheck_slab_free(cachep
, objp
, cachep
->object_size
);
3521 * Skip calling cache_free_alien() when the platform is not numa.
3522 * This will avoid cache misses that happen while accessing slabp (which
3523 * is per page memory reference) to get nodeid. Instead use a global
3524 * variable to skip the call, which is mostly likely to be present in
3527 if (nr_online_nodes
> 1 && cache_free_alien(cachep
, objp
))
3530 if (ac
->avail
< ac
->limit
) {
3531 STATS_INC_FREEHIT(cachep
);
3533 STATS_INC_FREEMISS(cachep
);
3534 cache_flusharray(cachep
, ac
);
3537 if (sk_memalloc_socks()) {
3538 struct page
*page
= virt_to_head_page(objp
);
3540 if (unlikely(PageSlabPfmemalloc(page
))) {
3541 cache_free_pfmemalloc(cachep
, page
, objp
);
3546 ac
->entry
[ac
->avail
++] = objp
;
3550 * kmem_cache_alloc - Allocate an object
3551 * @cachep: The cache to allocate from.
3552 * @flags: See kmalloc().
3554 * Allocate an object from this cache. The flags are only relevant
3555 * if the cache has no available objects.
3557 void *kmem_cache_alloc(struct kmem_cache
*cachep
, gfp_t flags
)
3559 void *ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3561 kasan_slab_alloc(cachep
, ret
, flags
);
3562 trace_kmem_cache_alloc(_RET_IP_
, ret
,
3563 cachep
->object_size
, cachep
->size
, flags
);
3567 EXPORT_SYMBOL(kmem_cache_alloc
);
3569 static __always_inline
void
3570 cache_alloc_debugcheck_after_bulk(struct kmem_cache
*s
, gfp_t flags
,
3571 size_t size
, void **p
, unsigned long caller
)
3575 for (i
= 0; i
< size
; i
++)
3576 p
[i
] = cache_alloc_debugcheck_after(s
, flags
, p
[i
], caller
);
3579 int kmem_cache_alloc_bulk(struct kmem_cache
*s
, gfp_t flags
, size_t size
,
3584 s
= slab_pre_alloc_hook(s
, flags
);
3588 cache_alloc_debugcheck_before(s
, flags
);
3590 local_irq_disable();
3591 for (i
= 0; i
< size
; i
++) {
3592 void *objp
= __do_cache_alloc(s
, flags
);
3594 if (unlikely(!objp
))
3600 cache_alloc_debugcheck_after_bulk(s
, flags
, size
, p
, _RET_IP_
);
3602 /* Clear memory outside IRQ disabled section */
3603 if (unlikely(flags
& __GFP_ZERO
))
3604 for (i
= 0; i
< size
; i
++)
3605 memset(p
[i
], 0, s
->object_size
);
3607 slab_post_alloc_hook(s
, flags
, size
, p
);
3608 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3612 cache_alloc_debugcheck_after_bulk(s
, flags
, i
, p
, _RET_IP_
);
3613 slab_post_alloc_hook(s
, flags
, i
, p
);
3614 __kmem_cache_free_bulk(s
, i
, p
);
3617 EXPORT_SYMBOL(kmem_cache_alloc_bulk
);
3619 #ifdef CONFIG_TRACING
3621 kmem_cache_alloc_trace(struct kmem_cache
*cachep
, gfp_t flags
, size_t size
)
3625 ret
= slab_alloc(cachep
, flags
, _RET_IP_
);
3627 kasan_kmalloc(cachep
, ret
, size
, flags
);
3628 trace_kmalloc(_RET_IP_
, ret
,
3629 size
, cachep
->size
, flags
);
3632 EXPORT_SYMBOL(kmem_cache_alloc_trace
);
3637 * kmem_cache_alloc_node - Allocate an object on the specified node
3638 * @cachep: The cache to allocate from.
3639 * @flags: See kmalloc().
3640 * @nodeid: node number of the target node.
3642 * Identical to kmem_cache_alloc but it will allocate memory on the given
3643 * node, which can improve the performance for cpu bound structures.
3645 * Fallback to other node is possible if __GFP_THISNODE is not set.
3647 void *kmem_cache_alloc_node(struct kmem_cache
*cachep
, gfp_t flags
, int nodeid
)
3649 void *ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3651 kasan_slab_alloc(cachep
, ret
, flags
);
3652 trace_kmem_cache_alloc_node(_RET_IP_
, ret
,
3653 cachep
->object_size
, cachep
->size
,
3658 EXPORT_SYMBOL(kmem_cache_alloc_node
);
3660 #ifdef CONFIG_TRACING
3661 void *kmem_cache_alloc_node_trace(struct kmem_cache
*cachep
,
3668 ret
= slab_alloc_node(cachep
, flags
, nodeid
, _RET_IP_
);
3670 kasan_kmalloc(cachep
, ret
, size
, flags
);
3671 trace_kmalloc_node(_RET_IP_
, ret
,
3676 EXPORT_SYMBOL(kmem_cache_alloc_node_trace
);
3679 static __always_inline
void *
3680 __do_kmalloc_node(size_t size
, gfp_t flags
, int node
, unsigned long caller
)
3682 struct kmem_cache
*cachep
;
3685 cachep
= kmalloc_slab(size
, flags
);
3686 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3688 ret
= kmem_cache_alloc_node_trace(cachep
, flags
, node
, size
);
3689 kasan_kmalloc(cachep
, ret
, size
, flags
);
3694 void *__kmalloc_node(size_t size
, gfp_t flags
, int node
)
3696 return __do_kmalloc_node(size
, flags
, node
, _RET_IP_
);
3698 EXPORT_SYMBOL(__kmalloc_node
);
3700 void *__kmalloc_node_track_caller(size_t size
, gfp_t flags
,
3701 int node
, unsigned long caller
)
3703 return __do_kmalloc_node(size
, flags
, node
, caller
);
3705 EXPORT_SYMBOL(__kmalloc_node_track_caller
);
3706 #endif /* CONFIG_NUMA */
3709 * __do_kmalloc - allocate memory
3710 * @size: how many bytes of memory are required.
3711 * @flags: the type of memory to allocate (see kmalloc).
3712 * @caller: function caller for debug tracking of the caller
3714 static __always_inline
void *__do_kmalloc(size_t size
, gfp_t flags
,
3715 unsigned long caller
)
3717 struct kmem_cache
*cachep
;
3720 cachep
= kmalloc_slab(size
, flags
);
3721 if (unlikely(ZERO_OR_NULL_PTR(cachep
)))
3723 ret
= slab_alloc(cachep
, flags
, caller
);
3725 kasan_kmalloc(cachep
, ret
, size
, flags
);
3726 trace_kmalloc(caller
, ret
,
3727 size
, cachep
->size
, flags
);
3732 void *__kmalloc(size_t size
, gfp_t flags
)
3734 return __do_kmalloc(size
, flags
, _RET_IP_
);
3736 EXPORT_SYMBOL(__kmalloc
);
3738 void *__kmalloc_track_caller(size_t size
, gfp_t flags
, unsigned long caller
)
3740 return __do_kmalloc(size
, flags
, caller
);
3742 EXPORT_SYMBOL(__kmalloc_track_caller
);
3745 * kmem_cache_free - Deallocate an object
3746 * @cachep: The cache the allocation was from.
3747 * @objp: The previously allocated object.
3749 * Free an object which was previously allocated from this
3752 void kmem_cache_free(struct kmem_cache
*cachep
, void *objp
)
3754 unsigned long flags
;
3755 cachep
= cache_from_obj(cachep
, objp
);
3759 local_irq_save(flags
);
3760 debug_check_no_locks_freed(objp
, cachep
->object_size
);
3761 if (!(cachep
->flags
& SLAB_DEBUG_OBJECTS
))
3762 debug_check_no_obj_freed(objp
, cachep
->object_size
);
3763 __cache_free(cachep
, objp
, _RET_IP_
);
3764 local_irq_restore(flags
);
3766 trace_kmem_cache_free(_RET_IP_
, objp
);
3768 EXPORT_SYMBOL(kmem_cache_free
);
3770 void kmem_cache_free_bulk(struct kmem_cache
*orig_s
, size_t size
, void **p
)
3772 struct kmem_cache
*s
;
3775 local_irq_disable();
3776 for (i
= 0; i
< size
; i
++) {
3779 if (!orig_s
) /* called via kfree_bulk */
3780 s
= virt_to_cache(objp
);
3782 s
= cache_from_obj(orig_s
, objp
);
3784 debug_check_no_locks_freed(objp
, s
->object_size
);
3785 if (!(s
->flags
& SLAB_DEBUG_OBJECTS
))
3786 debug_check_no_obj_freed(objp
, s
->object_size
);
3788 __cache_free(s
, objp
, _RET_IP_
);
3792 /* FIXME: add tracing */
3794 EXPORT_SYMBOL(kmem_cache_free_bulk
);
3797 * kfree - free previously allocated memory
3798 * @objp: pointer returned by kmalloc.
3800 * If @objp is NULL, no operation is performed.
3802 * Don't free memory not originally allocated by kmalloc()
3803 * or you will run into trouble.
3805 void kfree(const void *objp
)
3807 struct kmem_cache
*c
;
3808 unsigned long flags
;
3810 trace_kfree(_RET_IP_
, objp
);
3812 if (unlikely(ZERO_OR_NULL_PTR(objp
)))
3814 local_irq_save(flags
);
3815 kfree_debugcheck(objp
);
3816 c
= virt_to_cache(objp
);
3817 debug_check_no_locks_freed(objp
, c
->object_size
);
3819 debug_check_no_obj_freed(objp
, c
->object_size
);
3820 __cache_free(c
, (void *)objp
, _RET_IP_
);
3821 local_irq_restore(flags
);
3823 EXPORT_SYMBOL(kfree
);
3826 * This initializes kmem_cache_node or resizes various caches for all nodes.
3828 static int setup_kmem_cache_nodes(struct kmem_cache
*cachep
, gfp_t gfp
)
3832 struct kmem_cache_node
*n
;
3834 for_each_online_node(node
) {
3835 ret
= setup_kmem_cache_node(cachep
, node
, gfp
, true);
3844 if (!cachep
->list
.next
) {
3845 /* Cache is not active yet. Roll back what we did */
3848 n
= get_node(cachep
, node
);
3851 free_alien_cache(n
->alien
);
3853 cachep
->node
[node
] = NULL
;
3861 /* Always called with the slab_mutex held */
3862 static int __do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3863 int batchcount
, int shared
, gfp_t gfp
)
3865 struct array_cache __percpu
*cpu_cache
, *prev
;
3868 cpu_cache
= alloc_kmem_cache_cpus(cachep
, limit
, batchcount
);
3872 prev
= cachep
->cpu_cache
;
3873 cachep
->cpu_cache
= cpu_cache
;
3875 * Without a previous cpu_cache there's no need to synchronize remote
3876 * cpus, so skip the IPIs.
3879 kick_all_cpus_sync();
3882 cachep
->batchcount
= batchcount
;
3883 cachep
->limit
= limit
;
3884 cachep
->shared
= shared
;
3889 for_each_online_cpu(cpu
) {
3892 struct kmem_cache_node
*n
;
3893 struct array_cache
*ac
= per_cpu_ptr(prev
, cpu
);
3895 node
= cpu_to_mem(cpu
);
3896 n
= get_node(cachep
, node
);
3897 spin_lock_irq(&n
->list_lock
);
3898 free_block(cachep
, ac
->entry
, ac
->avail
, node
, &list
);
3899 spin_unlock_irq(&n
->list_lock
);
3900 slabs_destroy(cachep
, &list
);
3905 return setup_kmem_cache_nodes(cachep
, gfp
);
3908 static int do_tune_cpucache(struct kmem_cache
*cachep
, int limit
,
3909 int batchcount
, int shared
, gfp_t gfp
)
3912 struct kmem_cache
*c
;
3914 ret
= __do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3916 if (slab_state
< FULL
)
3919 if ((ret
< 0) || !is_root_cache(cachep
))
3922 lockdep_assert_held(&slab_mutex
);
3923 for_each_memcg_cache(c
, cachep
) {
3924 /* return value determined by the root cache only */
3925 __do_tune_cpucache(c
, limit
, batchcount
, shared
, gfp
);
3931 /* Called with slab_mutex held always */
3932 static int enable_cpucache(struct kmem_cache
*cachep
, gfp_t gfp
)
3939 err
= cache_random_seq_create(cachep
, cachep
->num
, gfp
);
3943 if (!is_root_cache(cachep
)) {
3944 struct kmem_cache
*root
= memcg_root_cache(cachep
);
3945 limit
= root
->limit
;
3946 shared
= root
->shared
;
3947 batchcount
= root
->batchcount
;
3950 if (limit
&& shared
&& batchcount
)
3953 * The head array serves three purposes:
3954 * - create a LIFO ordering, i.e. return objects that are cache-warm
3955 * - reduce the number of spinlock operations.
3956 * - reduce the number of linked list operations on the slab and
3957 * bufctl chains: array operations are cheaper.
3958 * The numbers are guessed, we should auto-tune as described by
3961 if (cachep
->size
> 131072)
3963 else if (cachep
->size
> PAGE_SIZE
)
3965 else if (cachep
->size
> 1024)
3967 else if (cachep
->size
> 256)
3973 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3974 * allocation behaviour: Most allocs on one cpu, most free operations
3975 * on another cpu. For these cases, an efficient object passing between
3976 * cpus is necessary. This is provided by a shared array. The array
3977 * replaces Bonwick's magazine layer.
3978 * On uniprocessor, it's functionally equivalent (but less efficient)
3979 * to a larger limit. Thus disabled by default.
3982 if (cachep
->size
<= PAGE_SIZE
&& num_possible_cpus() > 1)
3987 * With debugging enabled, large batchcount lead to excessively long
3988 * periods with disabled local interrupts. Limit the batchcount
3993 batchcount
= (limit
+ 1) / 2;
3995 err
= do_tune_cpucache(cachep
, limit
, batchcount
, shared
, gfp
);
3998 pr_err("enable_cpucache failed for %s, error %d\n",
3999 cachep
->name
, -err
);
4004 * Drain an array if it contains any elements taking the node lock only if
4005 * necessary. Note that the node listlock also protects the array_cache
4006 * if drain_array() is used on the shared array.
4008 static void drain_array(struct kmem_cache
*cachep
, struct kmem_cache_node
*n
,
4009 struct array_cache
*ac
, int node
)
4013 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4014 check_mutex_acquired();
4016 if (!ac
|| !ac
->avail
)
4024 spin_lock_irq(&n
->list_lock
);
4025 drain_array_locked(cachep
, ac
, node
, false, &list
);
4026 spin_unlock_irq(&n
->list_lock
);
4028 slabs_destroy(cachep
, &list
);
4032 * cache_reap - Reclaim memory from caches.
4033 * @w: work descriptor
4035 * Called from workqueue/eventd every few seconds.
4037 * - clear the per-cpu caches for this CPU.
4038 * - return freeable pages to the main free memory pool.
4040 * If we cannot acquire the cache chain mutex then just give up - we'll try
4041 * again on the next iteration.
4043 static void cache_reap(struct work_struct
*w
)
4045 struct kmem_cache
*searchp
;
4046 struct kmem_cache_node
*n
;
4047 int node
= numa_mem_id();
4048 struct delayed_work
*work
= to_delayed_work(w
);
4050 if (!mutex_trylock(&slab_mutex
))
4051 /* Give up. Setup the next iteration. */
4054 list_for_each_entry(searchp
, &slab_caches
, list
) {
4058 * We only take the node lock if absolutely necessary and we
4059 * have established with reasonable certainty that
4060 * we can do some work if the lock was obtained.
4062 n
= get_node(searchp
, node
);
4064 reap_alien(searchp
, n
);
4066 drain_array(searchp
, n
, cpu_cache_get(searchp
), node
);
4069 * These are racy checks but it does not matter
4070 * if we skip one check or scan twice.
4072 if (time_after(n
->next_reap
, jiffies
))
4075 n
->next_reap
= jiffies
+ REAPTIMEOUT_NODE
;
4077 drain_array(searchp
, n
, n
->shared
, node
);
4079 if (n
->free_touched
)
4080 n
->free_touched
= 0;
4084 freed
= drain_freelist(searchp
, n
, (n
->free_limit
+
4085 5 * searchp
->num
- 1) / (5 * searchp
->num
));
4086 STATS_ADD_REAPED(searchp
, freed
);
4092 mutex_unlock(&slab_mutex
);
4095 /* Set up the next iteration */
4096 schedule_delayed_work(work
, round_jiffies_relative(REAPTIMEOUT_AC
));
4099 void get_slabinfo(struct kmem_cache
*cachep
, struct slabinfo
*sinfo
)
4101 unsigned long active_objs
, num_objs
, active_slabs
;
4102 unsigned long total_slabs
= 0, free_objs
= 0, shared_avail
= 0;
4103 unsigned long free_slabs
= 0;
4105 struct kmem_cache_node
*n
;
4107 for_each_kmem_cache_node(cachep
, node
, n
) {
4109 spin_lock_irq(&n
->list_lock
);
4111 total_slabs
+= n
->total_slabs
;
4112 free_slabs
+= n
->free_slabs
;
4113 free_objs
+= n
->free_objects
;
4116 shared_avail
+= n
->shared
->avail
;
4118 spin_unlock_irq(&n
->list_lock
);
4120 num_objs
= total_slabs
* cachep
->num
;
4121 active_slabs
= total_slabs
- free_slabs
;
4122 active_objs
= num_objs
- free_objs
;
4124 sinfo
->active_objs
= active_objs
;
4125 sinfo
->num_objs
= num_objs
;
4126 sinfo
->active_slabs
= active_slabs
;
4127 sinfo
->num_slabs
= total_slabs
;
4128 sinfo
->shared_avail
= shared_avail
;
4129 sinfo
->limit
= cachep
->limit
;
4130 sinfo
->batchcount
= cachep
->batchcount
;
4131 sinfo
->shared
= cachep
->shared
;
4132 sinfo
->objects_per_slab
= cachep
->num
;
4133 sinfo
->cache_order
= cachep
->gfporder
;
4136 void slabinfo_show_stats(struct seq_file
*m
, struct kmem_cache
*cachep
)
4140 unsigned long high
= cachep
->high_mark
;
4141 unsigned long allocs
= cachep
->num_allocations
;
4142 unsigned long grown
= cachep
->grown
;
4143 unsigned long reaped
= cachep
->reaped
;
4144 unsigned long errors
= cachep
->errors
;
4145 unsigned long max_freeable
= cachep
->max_freeable
;
4146 unsigned long node_allocs
= cachep
->node_allocs
;
4147 unsigned long node_frees
= cachep
->node_frees
;
4148 unsigned long overflows
= cachep
->node_overflow
;
4150 seq_printf(m
, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4151 allocs
, high
, grown
,
4152 reaped
, errors
, max_freeable
, node_allocs
,
4153 node_frees
, overflows
);
4157 unsigned long allochit
= atomic_read(&cachep
->allochit
);
4158 unsigned long allocmiss
= atomic_read(&cachep
->allocmiss
);
4159 unsigned long freehit
= atomic_read(&cachep
->freehit
);
4160 unsigned long freemiss
= atomic_read(&cachep
->freemiss
);
4162 seq_printf(m
, " : cpustat %6lu %6lu %6lu %6lu",
4163 allochit
, allocmiss
, freehit
, freemiss
);
4168 #define MAX_SLABINFO_WRITE 128
4170 * slabinfo_write - Tuning for the slab allocator
4172 * @buffer: user buffer
4173 * @count: data length
4176 ssize_t
slabinfo_write(struct file
*file
, const char __user
*buffer
,
4177 size_t count
, loff_t
*ppos
)
4179 char kbuf
[MAX_SLABINFO_WRITE
+ 1], *tmp
;
4180 int limit
, batchcount
, shared
, res
;
4181 struct kmem_cache
*cachep
;
4183 if (count
> MAX_SLABINFO_WRITE
)
4185 if (copy_from_user(&kbuf
, buffer
, count
))
4187 kbuf
[MAX_SLABINFO_WRITE
] = '\0';
4189 tmp
= strchr(kbuf
, ' ');
4194 if (sscanf(tmp
, " %d %d %d", &limit
, &batchcount
, &shared
) != 3)
4197 /* Find the cache in the chain of caches. */
4198 mutex_lock(&slab_mutex
);
4200 list_for_each_entry(cachep
, &slab_caches
, list
) {
4201 if (!strcmp(cachep
->name
, kbuf
)) {
4202 if (limit
< 1 || batchcount
< 1 ||
4203 batchcount
> limit
|| shared
< 0) {
4206 res
= do_tune_cpucache(cachep
, limit
,
4213 mutex_unlock(&slab_mutex
);
4219 #ifdef CONFIG_DEBUG_SLAB_LEAK
4221 static inline int add_caller(unsigned long *n
, unsigned long v
)
4231 unsigned long *q
= p
+ 2 * i
;
4245 memmove(p
+ 2, p
, n
[1] * 2 * sizeof(unsigned long) - ((void *)p
- (void *)n
));
4251 static void handle_slab(unsigned long *n
, struct kmem_cache
*c
,
4260 for (i
= 0, p
= page
->s_mem
; i
< c
->num
; i
++, p
+= c
->size
) {
4263 for (j
= page
->active
; j
< c
->num
; j
++) {
4264 if (get_free_obj(page
, j
) == i
) {
4274 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4275 * mapping is established when actual object allocation and
4276 * we could mistakenly access the unmapped object in the cpu
4279 if (probe_kernel_read(&v
, dbg_userword(c
, p
), sizeof(v
)))
4282 if (!add_caller(n
, v
))
4287 static void show_symbol(struct seq_file
*m
, unsigned long address
)
4289 #ifdef CONFIG_KALLSYMS
4290 unsigned long offset
, size
;
4291 char modname
[MODULE_NAME_LEN
], name
[KSYM_NAME_LEN
];
4293 if (lookup_symbol_attrs(address
, &size
, &offset
, modname
, name
) == 0) {
4294 seq_printf(m
, "%s+%#lx/%#lx", name
, offset
, size
);
4296 seq_printf(m
, " [%s]", modname
);
4300 seq_printf(m
, "%p", (void *)address
);
4303 static int leaks_show(struct seq_file
*m
, void *p
)
4305 struct kmem_cache
*cachep
= list_entry(p
, struct kmem_cache
, list
);
4307 struct kmem_cache_node
*n
;
4309 unsigned long *x
= m
->private;
4313 if (!(cachep
->flags
& SLAB_STORE_USER
))
4315 if (!(cachep
->flags
& SLAB_RED_ZONE
))
4319 * Set store_user_clean and start to grab stored user information
4320 * for all objects on this cache. If some alloc/free requests comes
4321 * during the processing, information would be wrong so restart
4325 set_store_user_clean(cachep
);
4326 drain_cpu_caches(cachep
);
4330 for_each_kmem_cache_node(cachep
, node
, n
) {
4333 spin_lock_irq(&n
->list_lock
);
4335 list_for_each_entry(page
, &n
->slabs_full
, lru
)
4336 handle_slab(x
, cachep
, page
);
4337 list_for_each_entry(page
, &n
->slabs_partial
, lru
)
4338 handle_slab(x
, cachep
, page
);
4339 spin_unlock_irq(&n
->list_lock
);
4341 } while (!is_store_user_clean(cachep
));
4343 name
= cachep
->name
;
4345 /* Increase the buffer size */
4346 mutex_unlock(&slab_mutex
);
4347 m
->private = kzalloc(x
[0] * 4 * sizeof(unsigned long), GFP_KERNEL
);
4349 /* Too bad, we are really out */
4351 mutex_lock(&slab_mutex
);
4354 *(unsigned long *)m
->private = x
[0] * 2;
4356 mutex_lock(&slab_mutex
);
4357 /* Now make sure this entry will be retried */
4361 for (i
= 0; i
< x
[1]; i
++) {
4362 seq_printf(m
, "%s: %lu ", name
, x
[2*i
+3]);
4363 show_symbol(m
, x
[2*i
+2]);
4370 static const struct seq_operations slabstats_op
= {
4371 .start
= slab_start
,
4377 static int slabstats_open(struct inode
*inode
, struct file
*file
)
4381 n
= __seq_open_private(file
, &slabstats_op
, PAGE_SIZE
);
4385 *n
= PAGE_SIZE
/ (2 * sizeof(unsigned long));
4390 static const struct file_operations proc_slabstats_operations
= {
4391 .open
= slabstats_open
,
4393 .llseek
= seq_lseek
,
4394 .release
= seq_release_private
,
4398 static int __init
slab_proc_init(void)
4400 #ifdef CONFIG_DEBUG_SLAB_LEAK
4401 proc_create("slab_allocators", 0, NULL
, &proc_slabstats_operations
);
4405 module_init(slab_proc_init
);
4407 #ifdef CONFIG_HARDENED_USERCOPY
4409 * Rejects objects that are incorrectly sized.
4411 * Returns NULL if check passes, otherwise const char * to name of cache
4412 * to indicate an error.
4414 const char *__check_heap_object(const void *ptr
, unsigned long n
,
4417 struct kmem_cache
*cachep
;
4419 unsigned long offset
;
4421 /* Find and validate object. */
4422 cachep
= page
->slab_cache
;
4423 objnr
= obj_to_index(cachep
, page
, (void *)ptr
);
4424 BUG_ON(objnr
>= cachep
->num
);
4426 /* Find offset within object. */
4427 offset
= ptr
- index_to_obj(cachep
, page
, objnr
) - obj_offset(cachep
);
4429 /* Allow address range falling entirely within object size. */
4430 if (offset
<= cachep
->object_size
&& n
<= cachep
->object_size
- offset
)
4433 return cachep
->name
;
4435 #endif /* CONFIG_HARDENED_USERCOPY */
4438 * ksize - get the actual amount of memory allocated for a given object
4439 * @objp: Pointer to the object
4441 * kmalloc may internally round up allocations and return more memory
4442 * than requested. ksize() can be used to determine the actual amount of
4443 * memory allocated. The caller may use this additional memory, even though
4444 * a smaller amount of memory was initially specified with the kmalloc call.
4445 * The caller must guarantee that objp points to a valid object previously
4446 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4447 * must not be freed during the duration of the call.
4449 size_t ksize(const void *objp
)
4454 if (unlikely(objp
== ZERO_SIZE_PTR
))
4457 size
= virt_to_cache(objp
)->object_size
;
4458 /* We assume that ksize callers could use the whole allocated area,
4459 * so we need to unpoison this area.
4461 kasan_unpoison_shadow(objp
, size
);
4465 EXPORT_SYMBOL(ksize
);