3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t;
166 typedef unsigned short freelist_idx_t;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
172 * true if a page was allocated from pfmemalloc reserves for network-based
175 static bool pfmemalloc_active __read_mostly;
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
185 * The limit is stored in the per-cpu structure to reduce the data cache
192 unsigned int batchcount;
193 unsigned int touched;
195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
207 struct array_cache ac;
210 #define SLAB_OBJ_PFMEMALLOC 1
211 static inline bool is_obj_pfmemalloc(void *objp)
213 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
216 static inline void set_obj_pfmemalloc(void **objp)
218 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
222 static inline void clear_obj_pfmemalloc(void **objp)
224 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
228 * bootstrap: The caches do not work without cpuarrays anymore, but the
229 * cpuarrays are allocated from the generic caches...
231 #define BOOT_CPUCACHE_ENTRIES 1
232 struct arraycache_init {
233 struct array_cache cache;
234 void *entries[BOOT_CPUCACHE_ENTRIES];
238 * Need this for bootstrapping a per node allocator.
240 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
241 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
242 #define CACHE_CACHE 0
243 #define SIZE_NODE (MAX_NUMNODES)
245 static int drain_freelist(struct kmem_cache *cache,
246 struct kmem_cache_node *n, int tofree);
247 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
248 int node, struct list_head *list);
249 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
250 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
251 static void cache_reap(struct work_struct *unused);
253 static int slab_early_init = 1;
255 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
257 static void kmem_cache_node_init(struct kmem_cache_node *parent)
259 INIT_LIST_HEAD(&parent->slabs_full);
260 INIT_LIST_HEAD(&parent->slabs_partial);
261 INIT_LIST_HEAD(&parent->slabs_free);
262 parent->shared = NULL;
263 parent->alien = NULL;
264 parent->colour_next = 0;
265 spin_lock_init(&parent->list_lock);
266 parent->free_objects = 0;
267 parent->free_touched = 0;
270 #define MAKE_LIST(cachep, listp, slab, nodeid) \
272 INIT_LIST_HEAD(listp); \
273 list_splice(&get_node(cachep, nodeid)->slab, listp); \
276 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
278 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
279 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
280 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
283 #define CFLGS_OFF_SLAB (0x80000000UL)
284 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
285 #define OFF_SLAB_MIN_SIZE (max_t(size_t, PAGE_SIZE >> 5, KMALLOC_MIN_SIZE + 1))
287 #define BATCHREFILL_LIMIT 16
289 * Optimization question: fewer reaps means less probability for unnessary
290 * cpucache drain/refill cycles.
292 * OTOH the cpuarrays can contain lots of objects,
293 * which could lock up otherwise freeable slabs.
295 #define REAPTIMEOUT_AC (2*HZ)
296 #define REAPTIMEOUT_NODE (4*HZ)
299 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
300 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
301 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
302 #define STATS_INC_GROWN(x) ((x)->grown++)
303 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
304 #define STATS_SET_HIGH(x) \
306 if ((x)->num_active > (x)->high_mark) \
307 (x)->high_mark = (x)->num_active; \
309 #define STATS_INC_ERR(x) ((x)->errors++)
310 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
311 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
312 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
313 #define STATS_SET_FREEABLE(x, i) \
315 if ((x)->max_freeable < i) \
316 (x)->max_freeable = i; \
318 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
319 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
320 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
321 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
323 #define STATS_INC_ACTIVE(x) do { } while (0)
324 #define STATS_DEC_ACTIVE(x) do { } while (0)
325 #define STATS_INC_ALLOCED(x) do { } while (0)
326 #define STATS_INC_GROWN(x) do { } while (0)
327 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
328 #define STATS_SET_HIGH(x) do { } while (0)
329 #define STATS_INC_ERR(x) do { } while (0)
330 #define STATS_INC_NODEALLOCS(x) do { } while (0)
331 #define STATS_INC_NODEFREES(x) do { } while (0)
332 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
333 #define STATS_SET_FREEABLE(x, i) do { } while (0)
334 #define STATS_INC_ALLOCHIT(x) do { } while (0)
335 #define STATS_INC_ALLOCMISS(x) do { } while (0)
336 #define STATS_INC_FREEHIT(x) do { } while (0)
337 #define STATS_INC_FREEMISS(x) do { } while (0)
343 * memory layout of objects:
345 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
346 * the end of an object is aligned with the end of the real
347 * allocation. Catches writes behind the end of the allocation.
348 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
350 * cachep->obj_offset: The real object.
351 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
352 * cachep->size - 1* BYTES_PER_WORD: last caller address
353 * [BYTES_PER_WORD long]
355 static int obj_offset(struct kmem_cache *cachep)
357 return cachep->obj_offset;
360 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
362 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
363 return (unsigned long long*) (objp + obj_offset(cachep) -
364 sizeof(unsigned long long));
367 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
369 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
370 if (cachep->flags & SLAB_STORE_USER)
371 return (unsigned long long *)(objp + cachep->size -
372 sizeof(unsigned long long) -
374 return (unsigned long long *) (objp + cachep->size -
375 sizeof(unsigned long long));
378 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
380 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
381 return (void **)(objp + cachep->size - BYTES_PER_WORD);
386 #define obj_offset(x) 0
387 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
388 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
393 #ifdef CONFIG_DEBUG_SLAB_LEAK
395 static inline bool is_store_user_clean(struct kmem_cache *cachep)
397 return atomic_read(&cachep->store_user_clean) == 1;
400 static inline void set_store_user_clean(struct kmem_cache *cachep)
402 atomic_set(&cachep->store_user_clean, 1);
405 static inline void set_store_user_dirty(struct kmem_cache *cachep)
407 if (is_store_user_clean(cachep))
408 atomic_set(&cachep->store_user_clean, 0);
412 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
417 * Do not go above this order unless 0 objects fit into the slab or
418 * overridden on the command line.
420 #define SLAB_MAX_ORDER_HI 1
421 #define SLAB_MAX_ORDER_LO 0
422 static int slab_max_order = SLAB_MAX_ORDER_LO;
423 static bool slab_max_order_set __initdata;
425 static inline struct kmem_cache *virt_to_cache(const void *obj)
427 struct page *page = virt_to_head_page(obj);
428 return page->slab_cache;
431 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
434 return page->s_mem + cache->size * idx;
438 * We want to avoid an expensive divide : (offset / cache->size)
439 * Using the fact that size is a constant for a particular cache,
440 * we can replace (offset / cache->size) by
441 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
443 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
444 const struct page *page, void *obj)
446 u32 offset = (obj - page->s_mem);
447 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
450 /* internal cache of cache description objs */
451 static struct kmem_cache kmem_cache_boot = {
453 .limit = BOOT_CPUCACHE_ENTRIES,
455 .size = sizeof(struct kmem_cache),
456 .name = "kmem_cache",
459 #define BAD_ALIEN_MAGIC 0x01020304ul
461 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
463 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
465 return this_cpu_ptr(cachep->cpu_cache);
468 static size_t calculate_freelist_size(int nr_objs, size_t align)
470 size_t freelist_size;
472 freelist_size = nr_objs * sizeof(freelist_idx_t);
474 freelist_size = ALIGN(freelist_size, align);
476 return freelist_size;
479 static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
480 size_t idx_size, size_t align)
483 size_t remained_size;
484 size_t freelist_size;
487 * Ignore padding for the initial guess. The padding
488 * is at most @align-1 bytes, and @buffer_size is at
489 * least @align. In the worst case, this result will
490 * be one greater than the number of objects that fit
491 * into the memory allocation when taking the padding
494 nr_objs = slab_size / (buffer_size + idx_size);
497 * This calculated number will be either the right
498 * amount, or one greater than what we want.
500 remained_size = slab_size - nr_objs * buffer_size;
501 freelist_size = calculate_freelist_size(nr_objs, align);
502 if (remained_size < freelist_size)
509 * Calculate the number of objects and left-over bytes for a given buffer size.
511 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
512 size_t align, int flags, size_t *left_over,
517 size_t slab_size = PAGE_SIZE << gfporder;
520 * The slab management structure can be either off the slab or
521 * on it. For the latter case, the memory allocated for a
524 * - One unsigned int for each object
525 * - Padding to respect alignment of @align
526 * - @buffer_size bytes for each object
528 * If the slab management structure is off the slab, then the
529 * alignment will already be calculated into the size. Because
530 * the slabs are all pages aligned, the objects will be at the
531 * correct alignment when allocated.
533 if (flags & CFLGS_OFF_SLAB) {
535 nr_objs = slab_size / buffer_size;
538 nr_objs = calculate_nr_objs(slab_size, buffer_size,
539 sizeof(freelist_idx_t), align);
540 mgmt_size = calculate_freelist_size(nr_objs, align);
543 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
547 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
549 static void __slab_error(const char *function, struct kmem_cache *cachep,
552 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
553 function, cachep->name, msg);
555 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
560 * By default on NUMA we use alien caches to stage the freeing of
561 * objects allocated from other nodes. This causes massive memory
562 * inefficiencies when using fake NUMA setup to split memory into a
563 * large number of small nodes, so it can be disabled on the command
567 static int use_alien_caches __read_mostly = 1;
568 static int __init noaliencache_setup(char *s)
570 use_alien_caches = 0;
573 __setup("noaliencache", noaliencache_setup);
575 static int __init slab_max_order_setup(char *str)
577 get_option(&str, &slab_max_order);
578 slab_max_order = slab_max_order < 0 ? 0 :
579 min(slab_max_order, MAX_ORDER - 1);
580 slab_max_order_set = true;
584 __setup("slab_max_order=", slab_max_order_setup);
588 * Special reaping functions for NUMA systems called from cache_reap().
589 * These take care of doing round robin flushing of alien caches (containing
590 * objects freed on different nodes from which they were allocated) and the
591 * flushing of remote pcps by calling drain_node_pages.
593 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
595 static void init_reap_node(int cpu)
599 node = next_node(cpu_to_mem(cpu), node_online_map);
600 if (node == MAX_NUMNODES)
601 node = first_node(node_online_map);
603 per_cpu(slab_reap_node, cpu) = node;
606 static void next_reap_node(void)
608 int node = __this_cpu_read(slab_reap_node);
610 node = next_node(node, node_online_map);
611 if (unlikely(node >= MAX_NUMNODES))
612 node = first_node(node_online_map);
613 __this_cpu_write(slab_reap_node, node);
617 #define init_reap_node(cpu) do { } while (0)
618 #define next_reap_node(void) do { } while (0)
622 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
623 * via the workqueue/eventd.
624 * Add the CPU number into the expiration time to minimize the possibility of
625 * the CPUs getting into lockstep and contending for the global cache chain
628 static void start_cpu_timer(int cpu)
630 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
633 * When this gets called from do_initcalls via cpucache_init(),
634 * init_workqueues() has already run, so keventd will be setup
637 if (keventd_up() && reap_work->work.func == NULL) {
639 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
640 schedule_delayed_work_on(cpu, reap_work,
641 __round_jiffies_relative(HZ, cpu));
645 static void init_arraycache(struct array_cache *ac, int limit, int batch)
648 * The array_cache structures contain pointers to free object.
649 * However, when such objects are allocated or transferred to another
650 * cache the pointers are not cleared and they could be counted as
651 * valid references during a kmemleak scan. Therefore, kmemleak must
652 * not scan such objects.
654 kmemleak_no_scan(ac);
658 ac->batchcount = batch;
663 static struct array_cache *alloc_arraycache(int node, int entries,
664 int batchcount, gfp_t gfp)
666 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
667 struct array_cache *ac = NULL;
669 ac = kmalloc_node(memsize, gfp, node);
670 init_arraycache(ac, entries, batchcount);
674 static inline bool is_slab_pfmemalloc(struct page *page)
676 return PageSlabPfmemalloc(page);
679 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
680 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
681 struct array_cache *ac)
683 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
687 if (!pfmemalloc_active)
690 spin_lock_irqsave(&n->list_lock, flags);
691 list_for_each_entry(page, &n->slabs_full, lru)
692 if (is_slab_pfmemalloc(page))
695 list_for_each_entry(page, &n->slabs_partial, lru)
696 if (is_slab_pfmemalloc(page))
699 list_for_each_entry(page, &n->slabs_free, lru)
700 if (is_slab_pfmemalloc(page))
703 pfmemalloc_active = false;
705 spin_unlock_irqrestore(&n->list_lock, flags);
708 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
709 gfp_t flags, bool force_refill)
712 void *objp = ac->entry[--ac->avail];
714 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
715 if (unlikely(is_obj_pfmemalloc(objp))) {
716 struct kmem_cache_node *n;
718 if (gfp_pfmemalloc_allowed(flags)) {
719 clear_obj_pfmemalloc(&objp);
723 /* The caller cannot use PFMEMALLOC objects, find another one */
724 for (i = 0; i < ac->avail; i++) {
725 /* If a !PFMEMALLOC object is found, swap them */
726 if (!is_obj_pfmemalloc(ac->entry[i])) {
728 ac->entry[i] = ac->entry[ac->avail];
729 ac->entry[ac->avail] = objp;
735 * If there are empty slabs on the slabs_free list and we are
736 * being forced to refill the cache, mark this one !pfmemalloc.
738 n = get_node(cachep, numa_mem_id());
739 if (!list_empty(&n->slabs_free) && force_refill) {
740 struct page *page = virt_to_head_page(objp);
741 ClearPageSlabPfmemalloc(page);
742 clear_obj_pfmemalloc(&objp);
743 recheck_pfmemalloc_active(cachep, ac);
747 /* No !PFMEMALLOC objects available */
755 static inline void *ac_get_obj(struct kmem_cache *cachep,
756 struct array_cache *ac, gfp_t flags, bool force_refill)
760 if (unlikely(sk_memalloc_socks()))
761 objp = __ac_get_obj(cachep, ac, flags, force_refill);
763 objp = ac->entry[--ac->avail];
768 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
769 struct array_cache *ac, void *objp)
771 if (unlikely(pfmemalloc_active)) {
772 /* Some pfmemalloc slabs exist, check if this is one */
773 struct page *page = virt_to_head_page(objp);
774 if (PageSlabPfmemalloc(page))
775 set_obj_pfmemalloc(&objp);
781 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
784 if (unlikely(sk_memalloc_socks()))
785 objp = __ac_put_obj(cachep, ac, objp);
787 ac->entry[ac->avail++] = objp;
791 * Transfer objects in one arraycache to another.
792 * Locking must be handled by the caller.
794 * Return the number of entries transferred.
796 static int transfer_objects(struct array_cache *to,
797 struct array_cache *from, unsigned int max)
799 /* Figure out how many entries to transfer */
800 int nr = min3(from->avail, max, to->limit - to->avail);
805 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
815 #define drain_alien_cache(cachep, alien) do { } while (0)
816 #define reap_alien(cachep, n) do { } while (0)
818 static inline struct alien_cache **alloc_alien_cache(int node,
819 int limit, gfp_t gfp)
821 return (struct alien_cache **)BAD_ALIEN_MAGIC;
824 static inline void free_alien_cache(struct alien_cache **ac_ptr)
828 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
833 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
839 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
840 gfp_t flags, int nodeid)
845 static inline gfp_t gfp_exact_node(gfp_t flags)
850 #else /* CONFIG_NUMA */
852 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
853 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
855 static struct alien_cache *__alloc_alien_cache(int node, int entries,
856 int batch, gfp_t gfp)
858 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
859 struct alien_cache *alc = NULL;
861 alc = kmalloc_node(memsize, gfp, node);
862 init_arraycache(&alc->ac, entries, batch);
863 spin_lock_init(&alc->lock);
867 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
869 struct alien_cache **alc_ptr;
870 size_t memsize = sizeof(void *) * nr_node_ids;
875 alc_ptr = kzalloc_node(memsize, gfp, node);
880 if (i == node || !node_online(i))
882 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
884 for (i--; i >= 0; i--)
893 static void free_alien_cache(struct alien_cache **alc_ptr)
904 static void __drain_alien_cache(struct kmem_cache *cachep,
905 struct array_cache *ac, int node,
906 struct list_head *list)
908 struct kmem_cache_node *n = get_node(cachep, node);
911 spin_lock(&n->list_lock);
913 * Stuff objects into the remote nodes shared array first.
914 * That way we could avoid the overhead of putting the objects
915 * into the free lists and getting them back later.
918 transfer_objects(n->shared, ac, ac->limit);
920 free_block(cachep, ac->entry, ac->avail, node, list);
922 spin_unlock(&n->list_lock);
927 * Called from cache_reap() to regularly drain alien caches round robin.
929 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
931 int node = __this_cpu_read(slab_reap_node);
934 struct alien_cache *alc = n->alien[node];
935 struct array_cache *ac;
939 if (ac->avail && spin_trylock_irq(&alc->lock)) {
942 __drain_alien_cache(cachep, ac, node, &list);
943 spin_unlock_irq(&alc->lock);
944 slabs_destroy(cachep, &list);
950 static void drain_alien_cache(struct kmem_cache *cachep,
951 struct alien_cache **alien)
954 struct alien_cache *alc;
955 struct array_cache *ac;
958 for_each_online_node(i) {
964 spin_lock_irqsave(&alc->lock, flags);
965 __drain_alien_cache(cachep, ac, i, &list);
966 spin_unlock_irqrestore(&alc->lock, flags);
967 slabs_destroy(cachep, &list);
972 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
973 int node, int page_node)
975 struct kmem_cache_node *n;
976 struct alien_cache *alien = NULL;
977 struct array_cache *ac;
980 n = get_node(cachep, node);
981 STATS_INC_NODEFREES(cachep);
982 if (n->alien && n->alien[page_node]) {
983 alien = n->alien[page_node];
985 spin_lock(&alien->lock);
986 if (unlikely(ac->avail == ac->limit)) {
987 STATS_INC_ACOVERFLOW(cachep);
988 __drain_alien_cache(cachep, ac, page_node, &list);
990 ac_put_obj(cachep, ac, objp);
991 spin_unlock(&alien->lock);
992 slabs_destroy(cachep, &list);
994 n = get_node(cachep, page_node);
995 spin_lock(&n->list_lock);
996 free_block(cachep, &objp, 1, page_node, &list);
997 spin_unlock(&n->list_lock);
998 slabs_destroy(cachep, &list);
1003 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1005 int page_node = page_to_nid(virt_to_page(objp));
1006 int node = numa_mem_id();
1008 * Make sure we are not freeing a object from another node to the array
1009 * cache on this cpu.
1011 if (likely(node == page_node))
1014 return __cache_free_alien(cachep, objp, node, page_node);
1018 * Construct gfp mask to allocate from a specific node but do not direct reclaim
1019 * or warn about failures. kswapd may still wake to reclaim in the background.
1021 static inline gfp_t gfp_exact_node(gfp_t flags)
1023 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
1028 * Allocates and initializes node for a node on each slab cache, used for
1029 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1030 * will be allocated off-node since memory is not yet online for the new node.
1031 * When hotplugging memory or a cpu, existing node are not replaced if
1034 * Must hold slab_mutex.
1036 static int init_cache_node_node(int node)
1038 struct kmem_cache *cachep;
1039 struct kmem_cache_node *n;
1040 const size_t memsize = sizeof(struct kmem_cache_node);
1042 list_for_each_entry(cachep, &slab_caches, list) {
1044 * Set up the kmem_cache_node for cpu before we can
1045 * begin anything. Make sure some other cpu on this
1046 * node has not already allocated this
1048 n = get_node(cachep, node);
1050 n = kmalloc_node(memsize, GFP_KERNEL, node);
1053 kmem_cache_node_init(n);
1054 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1055 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1058 * The kmem_cache_nodes don't come and go as CPUs
1059 * come and go. slab_mutex is sufficient
1062 cachep->node[node] = n;
1065 spin_lock_irq(&n->list_lock);
1067 (1 + nr_cpus_node(node)) *
1068 cachep->batchcount + cachep->num;
1069 spin_unlock_irq(&n->list_lock);
1074 static inline int slabs_tofree(struct kmem_cache *cachep,
1075 struct kmem_cache_node *n)
1077 return (n->free_objects + cachep->num - 1) / cachep->num;
1080 static void cpuup_canceled(long cpu)
1082 struct kmem_cache *cachep;
1083 struct kmem_cache_node *n = NULL;
1084 int node = cpu_to_mem(cpu);
1085 const struct cpumask *mask = cpumask_of_node(node);
1087 list_for_each_entry(cachep, &slab_caches, list) {
1088 struct array_cache *nc;
1089 struct array_cache *shared;
1090 struct alien_cache **alien;
1093 n = get_node(cachep, node);
1097 spin_lock_irq(&n->list_lock);
1099 /* Free limit for this kmem_cache_node */
1100 n->free_limit -= cachep->batchcount;
1102 /* cpu is dead; no one can alloc from it. */
1103 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1105 free_block(cachep, nc->entry, nc->avail, node, &list);
1109 if (!cpumask_empty(mask)) {
1110 spin_unlock_irq(&n->list_lock);
1116 free_block(cachep, shared->entry,
1117 shared->avail, node, &list);
1124 spin_unlock_irq(&n->list_lock);
1128 drain_alien_cache(cachep, alien);
1129 free_alien_cache(alien);
1133 slabs_destroy(cachep, &list);
1136 * In the previous loop, all the objects were freed to
1137 * the respective cache's slabs, now we can go ahead and
1138 * shrink each nodelist to its limit.
1140 list_for_each_entry(cachep, &slab_caches, list) {
1141 n = get_node(cachep, node);
1144 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1148 static int cpuup_prepare(long cpu)
1150 struct kmem_cache *cachep;
1151 struct kmem_cache_node *n = NULL;
1152 int node = cpu_to_mem(cpu);
1156 * We need to do this right in the beginning since
1157 * alloc_arraycache's are going to use this list.
1158 * kmalloc_node allows us to add the slab to the right
1159 * kmem_cache_node and not this cpu's kmem_cache_node
1161 err = init_cache_node_node(node);
1166 * Now we can go ahead with allocating the shared arrays and
1169 list_for_each_entry(cachep, &slab_caches, list) {
1170 struct array_cache *shared = NULL;
1171 struct alien_cache **alien = NULL;
1173 if (cachep->shared) {
1174 shared = alloc_arraycache(node,
1175 cachep->shared * cachep->batchcount,
1176 0xbaadf00d, GFP_KERNEL);
1180 if (use_alien_caches) {
1181 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1187 n = get_node(cachep, node);
1190 spin_lock_irq(&n->list_lock);
1193 * We are serialised from CPU_DEAD or
1194 * CPU_UP_CANCELLED by the cpucontrol lock
1205 spin_unlock_irq(&n->list_lock);
1207 free_alien_cache(alien);
1212 cpuup_canceled(cpu);
1216 static int cpuup_callback(struct notifier_block *nfb,
1217 unsigned long action, void *hcpu)
1219 long cpu = (long)hcpu;
1223 case CPU_UP_PREPARE:
1224 case CPU_UP_PREPARE_FROZEN:
1225 mutex_lock(&slab_mutex);
1226 err = cpuup_prepare(cpu);
1227 mutex_unlock(&slab_mutex);
1230 case CPU_ONLINE_FROZEN:
1231 start_cpu_timer(cpu);
1233 #ifdef CONFIG_HOTPLUG_CPU
1234 case CPU_DOWN_PREPARE:
1235 case CPU_DOWN_PREPARE_FROZEN:
1237 * Shutdown cache reaper. Note that the slab_mutex is
1238 * held so that if cache_reap() is invoked it cannot do
1239 * anything expensive but will only modify reap_work
1240 * and reschedule the timer.
1242 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1243 /* Now the cache_reaper is guaranteed to be not running. */
1244 per_cpu(slab_reap_work, cpu).work.func = NULL;
1246 case CPU_DOWN_FAILED:
1247 case CPU_DOWN_FAILED_FROZEN:
1248 start_cpu_timer(cpu);
1251 case CPU_DEAD_FROZEN:
1253 * Even if all the cpus of a node are down, we don't free the
1254 * kmem_cache_node of any cache. This to avoid a race between
1255 * cpu_down, and a kmalloc allocation from another cpu for
1256 * memory from the node of the cpu going down. The node
1257 * structure is usually allocated from kmem_cache_create() and
1258 * gets destroyed at kmem_cache_destroy().
1262 case CPU_UP_CANCELED:
1263 case CPU_UP_CANCELED_FROZEN:
1264 mutex_lock(&slab_mutex);
1265 cpuup_canceled(cpu);
1266 mutex_unlock(&slab_mutex);
1269 return notifier_from_errno(err);
1272 static struct notifier_block cpucache_notifier = {
1273 &cpuup_callback, NULL, 0
1276 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1278 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1279 * Returns -EBUSY if all objects cannot be drained so that the node is not
1282 * Must hold slab_mutex.
1284 static int __meminit drain_cache_node_node(int node)
1286 struct kmem_cache *cachep;
1289 list_for_each_entry(cachep, &slab_caches, list) {
1290 struct kmem_cache_node *n;
1292 n = get_node(cachep, node);
1296 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1298 if (!list_empty(&n->slabs_full) ||
1299 !list_empty(&n->slabs_partial)) {
1307 static int __meminit slab_memory_callback(struct notifier_block *self,
1308 unsigned long action, void *arg)
1310 struct memory_notify *mnb = arg;
1314 nid = mnb->status_change_nid;
1319 case MEM_GOING_ONLINE:
1320 mutex_lock(&slab_mutex);
1321 ret = init_cache_node_node(nid);
1322 mutex_unlock(&slab_mutex);
1324 case MEM_GOING_OFFLINE:
1325 mutex_lock(&slab_mutex);
1326 ret = drain_cache_node_node(nid);
1327 mutex_unlock(&slab_mutex);
1331 case MEM_CANCEL_ONLINE:
1332 case MEM_CANCEL_OFFLINE:
1336 return notifier_from_errno(ret);
1338 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1341 * swap the static kmem_cache_node with kmalloced memory
1343 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1346 struct kmem_cache_node *ptr;
1348 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1351 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1353 * Do not assume that spinlocks can be initialized via memcpy:
1355 spin_lock_init(&ptr->list_lock);
1357 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1358 cachep->node[nodeid] = ptr;
1362 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1363 * size of kmem_cache_node.
1365 static void __init set_up_node(struct kmem_cache *cachep, int index)
1369 for_each_online_node(node) {
1370 cachep->node[node] = &init_kmem_cache_node[index + node];
1371 cachep->node[node]->next_reap = jiffies +
1373 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1378 * Initialisation. Called after the page allocator have been initialised and
1379 * before smp_init().
1381 void __init kmem_cache_init(void)
1385 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1386 sizeof(struct rcu_head));
1387 kmem_cache = &kmem_cache_boot;
1389 if (num_possible_nodes() == 1)
1390 use_alien_caches = 0;
1392 for (i = 0; i < NUM_INIT_LISTS; i++)
1393 kmem_cache_node_init(&init_kmem_cache_node[i]);
1396 * Fragmentation resistance on low memory - only use bigger
1397 * page orders on machines with more than 32MB of memory if
1398 * not overridden on the command line.
1400 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1401 slab_max_order = SLAB_MAX_ORDER_HI;
1403 /* Bootstrap is tricky, because several objects are allocated
1404 * from caches that do not exist yet:
1405 * 1) initialize the kmem_cache cache: it contains the struct
1406 * kmem_cache structures of all caches, except kmem_cache itself:
1407 * kmem_cache is statically allocated.
1408 * Initially an __init data area is used for the head array and the
1409 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1410 * array at the end of the bootstrap.
1411 * 2) Create the first kmalloc cache.
1412 * The struct kmem_cache for the new cache is allocated normally.
1413 * An __init data area is used for the head array.
1414 * 3) Create the remaining kmalloc caches, with minimally sized
1416 * 4) Replace the __init data head arrays for kmem_cache and the first
1417 * kmalloc cache with kmalloc allocated arrays.
1418 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1419 * the other cache's with kmalloc allocated memory.
1420 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1423 /* 1) create the kmem_cache */
1426 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1428 create_boot_cache(kmem_cache, "kmem_cache",
1429 offsetof(struct kmem_cache, node) +
1430 nr_node_ids * sizeof(struct kmem_cache_node *),
1431 SLAB_HWCACHE_ALIGN);
1432 list_add(&kmem_cache->list, &slab_caches);
1433 slab_state = PARTIAL;
1436 * Initialize the caches that provide memory for the kmem_cache_node
1437 * structures first. Without this, further allocations will bug.
1439 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1440 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1441 slab_state = PARTIAL_NODE;
1442 setup_kmalloc_cache_index_table();
1444 slab_early_init = 0;
1446 /* 5) Replace the bootstrap kmem_cache_node */
1450 for_each_online_node(nid) {
1451 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1453 init_list(kmalloc_caches[INDEX_NODE],
1454 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1458 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1461 void __init kmem_cache_init_late(void)
1463 struct kmem_cache *cachep;
1467 /* 6) resize the head arrays to their final sizes */
1468 mutex_lock(&slab_mutex);
1469 list_for_each_entry(cachep, &slab_caches, list)
1470 if (enable_cpucache(cachep, GFP_NOWAIT))
1472 mutex_unlock(&slab_mutex);
1478 * Register a cpu startup notifier callback that initializes
1479 * cpu_cache_get for all new cpus
1481 register_cpu_notifier(&cpucache_notifier);
1485 * Register a memory hotplug callback that initializes and frees
1488 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1492 * The reap timers are started later, with a module init call: That part
1493 * of the kernel is not yet operational.
1497 static int __init cpucache_init(void)
1502 * Register the timers that return unneeded pages to the page allocator
1504 for_each_online_cpu(cpu)
1505 start_cpu_timer(cpu);
1511 __initcall(cpucache_init);
1513 static noinline void
1514 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1517 struct kmem_cache_node *n;
1519 unsigned long flags;
1521 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1522 DEFAULT_RATELIMIT_BURST);
1524 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1528 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1530 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1531 cachep->name, cachep->size, cachep->gfporder);
1533 for_each_kmem_cache_node(cachep, node, n) {
1534 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1535 unsigned long active_slabs = 0, num_slabs = 0;
1537 spin_lock_irqsave(&n->list_lock, flags);
1538 list_for_each_entry(page, &n->slabs_full, lru) {
1539 active_objs += cachep->num;
1542 list_for_each_entry(page, &n->slabs_partial, lru) {
1543 active_objs += page->active;
1546 list_for_each_entry(page, &n->slabs_free, lru)
1549 free_objects += n->free_objects;
1550 spin_unlock_irqrestore(&n->list_lock, flags);
1552 num_slabs += active_slabs;
1553 num_objs = num_slabs * cachep->num;
1555 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1556 node, active_slabs, num_slabs, active_objs, num_objs,
1563 * Interface to system's page allocator. No need to hold the
1564 * kmem_cache_node ->list_lock.
1566 * If we requested dmaable memory, we will get it. Even if we
1567 * did not request dmaable memory, we might get it, but that
1568 * would be relatively rare and ignorable.
1570 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1576 flags |= cachep->allocflags;
1577 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1578 flags |= __GFP_RECLAIMABLE;
1580 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1582 slab_out_of_memory(cachep, flags, nodeid);
1586 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1587 __free_pages(page, cachep->gfporder);
1591 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1592 if (page_is_pfmemalloc(page))
1593 pfmemalloc_active = true;
1595 nr_pages = (1 << cachep->gfporder);
1596 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1597 add_zone_page_state(page_zone(page),
1598 NR_SLAB_RECLAIMABLE, nr_pages);
1600 add_zone_page_state(page_zone(page),
1601 NR_SLAB_UNRECLAIMABLE, nr_pages);
1602 __SetPageSlab(page);
1603 if (page_is_pfmemalloc(page))
1604 SetPageSlabPfmemalloc(page);
1606 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1607 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1610 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1612 kmemcheck_mark_unallocated_pages(page, nr_pages);
1619 * Interface to system's page release.
1621 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1623 const unsigned long nr_freed = (1 << cachep->gfporder);
1625 kmemcheck_free_shadow(page, cachep->gfporder);
1627 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1628 sub_zone_page_state(page_zone(page),
1629 NR_SLAB_RECLAIMABLE, nr_freed);
1631 sub_zone_page_state(page_zone(page),
1632 NR_SLAB_UNRECLAIMABLE, nr_freed);
1634 BUG_ON(!PageSlab(page));
1635 __ClearPageSlabPfmemalloc(page);
1636 __ClearPageSlab(page);
1637 page_mapcount_reset(page);
1638 page->mapping = NULL;
1640 if (current->reclaim_state)
1641 current->reclaim_state->reclaimed_slab += nr_freed;
1642 __free_kmem_pages(page, cachep->gfporder);
1645 static void kmem_rcu_free(struct rcu_head *head)
1647 struct kmem_cache *cachep;
1650 page = container_of(head, struct page, rcu_head);
1651 cachep = page->slab_cache;
1653 kmem_freepages(cachep, page);
1657 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1659 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1660 (cachep->size % PAGE_SIZE) == 0)
1666 #ifdef CONFIG_DEBUG_PAGEALLOC
1667 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1668 unsigned long caller)
1670 int size = cachep->object_size;
1672 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1674 if (size < 5 * sizeof(unsigned long))
1677 *addr++ = 0x12345678;
1679 *addr++ = smp_processor_id();
1680 size -= 3 * sizeof(unsigned long);
1682 unsigned long *sptr = &caller;
1683 unsigned long svalue;
1685 while (!kstack_end(sptr)) {
1687 if (kernel_text_address(svalue)) {
1689 size -= sizeof(unsigned long);
1690 if (size <= sizeof(unsigned long))
1696 *addr++ = 0x87654321;
1699 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1700 int map, unsigned long caller)
1702 if (!is_debug_pagealloc_cache(cachep))
1706 store_stackinfo(cachep, objp, caller);
1708 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1712 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1713 int map, unsigned long caller) {}
1717 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1719 int size = cachep->object_size;
1720 addr = &((char *)addr)[obj_offset(cachep)];
1722 memset(addr, val, size);
1723 *(unsigned char *)(addr + size - 1) = POISON_END;
1726 static void dump_line(char *data, int offset, int limit)
1729 unsigned char error = 0;
1732 printk(KERN_ERR "%03x: ", offset);
1733 for (i = 0; i < limit; i++) {
1734 if (data[offset + i] != POISON_FREE) {
1735 error = data[offset + i];
1739 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1740 &data[offset], limit, 1);
1742 if (bad_count == 1) {
1743 error ^= POISON_FREE;
1744 if (!(error & (error - 1))) {
1745 printk(KERN_ERR "Single bit error detected. Probably bad RAM.\n");
1747 printk(KERN_ERR "Run memtest86+ or a similar memory test tool.\n");
1749 printk(KERN_ERR "Run a memory test tool.\n");
1758 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1763 if (cachep->flags & SLAB_RED_ZONE) {
1764 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1765 *dbg_redzone1(cachep, objp),
1766 *dbg_redzone2(cachep, objp));
1769 if (cachep->flags & SLAB_STORE_USER) {
1770 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1771 *dbg_userword(cachep, objp),
1772 *dbg_userword(cachep, objp));
1774 realobj = (char *)objp + obj_offset(cachep);
1775 size = cachep->object_size;
1776 for (i = 0; i < size && lines; i += 16, lines--) {
1779 if (i + limit > size)
1781 dump_line(realobj, i, limit);
1785 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1791 if (is_debug_pagealloc_cache(cachep))
1794 realobj = (char *)objp + obj_offset(cachep);
1795 size = cachep->object_size;
1797 for (i = 0; i < size; i++) {
1798 char exp = POISON_FREE;
1801 if (realobj[i] != exp) {
1807 "Slab corruption (%s): %s start=%p, len=%d\n",
1808 print_tainted(), cachep->name, realobj, size);
1809 print_objinfo(cachep, objp, 0);
1811 /* Hexdump the affected line */
1814 if (i + limit > size)
1816 dump_line(realobj, i, limit);
1819 /* Limit to 5 lines */
1825 /* Print some data about the neighboring objects, if they
1828 struct page *page = virt_to_head_page(objp);
1831 objnr = obj_to_index(cachep, page, objp);
1833 objp = index_to_obj(cachep, page, objnr - 1);
1834 realobj = (char *)objp + obj_offset(cachep);
1835 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1837 print_objinfo(cachep, objp, 2);
1839 if (objnr + 1 < cachep->num) {
1840 objp = index_to_obj(cachep, page, objnr + 1);
1841 realobj = (char *)objp + obj_offset(cachep);
1842 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1844 print_objinfo(cachep, objp, 2);
1851 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1855 for (i = 0; i < cachep->num; i++) {
1856 void *objp = index_to_obj(cachep, page, i);
1858 if (cachep->flags & SLAB_POISON) {
1859 check_poison_obj(cachep, objp);
1860 slab_kernel_map(cachep, objp, 1, 0);
1862 if (cachep->flags & SLAB_RED_ZONE) {
1863 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1864 slab_error(cachep, "start of a freed object was overwritten");
1865 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1866 slab_error(cachep, "end of a freed object was overwritten");
1871 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1878 * slab_destroy - destroy and release all objects in a slab
1879 * @cachep: cache pointer being destroyed
1880 * @page: page pointer being destroyed
1882 * Destroy all the objs in a slab page, and release the mem back to the system.
1883 * Before calling the slab page must have been unlinked from the cache. The
1884 * kmem_cache_node ->list_lock is not held/needed.
1886 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1890 freelist = page->freelist;
1891 slab_destroy_debugcheck(cachep, page);
1892 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1893 call_rcu(&page->rcu_head, kmem_rcu_free);
1895 kmem_freepages(cachep, page);
1898 * From now on, we don't use freelist
1899 * although actual page can be freed in rcu context
1901 if (OFF_SLAB(cachep))
1902 kmem_cache_free(cachep->freelist_cache, freelist);
1905 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1907 struct page *page, *n;
1909 list_for_each_entry_safe(page, n, list, lru) {
1910 list_del(&page->lru);
1911 slab_destroy(cachep, page);
1916 * calculate_slab_order - calculate size (page order) of slabs
1917 * @cachep: pointer to the cache that is being created
1918 * @size: size of objects to be created in this cache.
1919 * @align: required alignment for the objects.
1920 * @flags: slab allocation flags
1922 * Also calculates the number of objects per slab.
1924 * This could be made much more intelligent. For now, try to avoid using
1925 * high order pages for slabs. When the gfp() functions are more friendly
1926 * towards high-order requests, this should be changed.
1928 static size_t calculate_slab_order(struct kmem_cache *cachep,
1929 size_t size, size_t align, unsigned long flags)
1931 unsigned long offslab_limit;
1932 size_t left_over = 0;
1935 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1939 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1943 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1944 if (num > SLAB_OBJ_MAX_NUM)
1947 if (flags & CFLGS_OFF_SLAB) {
1949 * Max number of objs-per-slab for caches which
1950 * use off-slab slabs. Needed to avoid a possible
1951 * looping condition in cache_grow().
1953 offslab_limit = size;
1954 offslab_limit /= sizeof(freelist_idx_t);
1956 if (num > offslab_limit)
1960 /* Found something acceptable - save it away */
1962 cachep->gfporder = gfporder;
1963 left_over = remainder;
1966 * A VFS-reclaimable slab tends to have most allocations
1967 * as GFP_NOFS and we really don't want to have to be allocating
1968 * higher-order pages when we are unable to shrink dcache.
1970 if (flags & SLAB_RECLAIM_ACCOUNT)
1974 * Large number of objects is good, but very large slabs are
1975 * currently bad for the gfp()s.
1977 if (gfporder >= slab_max_order)
1981 * Acceptable internal fragmentation?
1983 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1989 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1990 struct kmem_cache *cachep, int entries, int batchcount)
1994 struct array_cache __percpu *cpu_cache;
1996 size = sizeof(void *) * entries + sizeof(struct array_cache);
1997 cpu_cache = __alloc_percpu(size, sizeof(void *));
2002 for_each_possible_cpu(cpu) {
2003 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2004 entries, batchcount);
2010 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2012 if (slab_state >= FULL)
2013 return enable_cpucache(cachep, gfp);
2015 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2016 if (!cachep->cpu_cache)
2019 if (slab_state == DOWN) {
2020 /* Creation of first cache (kmem_cache). */
2021 set_up_node(kmem_cache, CACHE_CACHE);
2022 } else if (slab_state == PARTIAL) {
2023 /* For kmem_cache_node */
2024 set_up_node(cachep, SIZE_NODE);
2028 for_each_online_node(node) {
2029 cachep->node[node] = kmalloc_node(
2030 sizeof(struct kmem_cache_node), gfp, node);
2031 BUG_ON(!cachep->node[node]);
2032 kmem_cache_node_init(cachep->node[node]);
2036 cachep->node[numa_mem_id()]->next_reap =
2037 jiffies + REAPTIMEOUT_NODE +
2038 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2040 cpu_cache_get(cachep)->avail = 0;
2041 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2042 cpu_cache_get(cachep)->batchcount = 1;
2043 cpu_cache_get(cachep)->touched = 0;
2044 cachep->batchcount = 1;
2045 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2049 unsigned long kmem_cache_flags(unsigned long object_size,
2050 unsigned long flags, const char *name,
2051 void (*ctor)(void *))
2057 __kmem_cache_alias(const char *name, size_t size, size_t align,
2058 unsigned long flags, void (*ctor)(void *))
2060 struct kmem_cache *cachep;
2062 cachep = find_mergeable(size, align, flags, name, ctor);
2067 * Adjust the object sizes so that we clear
2068 * the complete object on kzalloc.
2070 cachep->object_size = max_t(int, cachep->object_size, size);
2076 * __kmem_cache_create - Create a cache.
2077 * @cachep: cache management descriptor
2078 * @flags: SLAB flags
2080 * Returns a ptr to the cache on success, NULL on failure.
2081 * Cannot be called within a int, but can be interrupted.
2082 * The @ctor is run when new pages are allocated by the cache.
2086 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2087 * to catch references to uninitialised memory.
2089 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2090 * for buffer overruns.
2092 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2093 * cacheline. This can be beneficial if you're counting cycles as closely
2097 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2099 size_t left_over, freelist_size;
2100 size_t ralign = BYTES_PER_WORD;
2103 size_t size = cachep->size;
2108 * Enable redzoning and last user accounting, except for caches with
2109 * large objects, if the increased size would increase the object size
2110 * above the next power of two: caches with object sizes just above a
2111 * power of two have a significant amount of internal fragmentation.
2113 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2114 2 * sizeof(unsigned long long)))
2115 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2116 if (!(flags & SLAB_DESTROY_BY_RCU))
2117 flags |= SLAB_POISON;
2119 if (flags & SLAB_DESTROY_BY_RCU)
2120 BUG_ON(flags & SLAB_POISON);
2124 * Check that size is in terms of words. This is needed to avoid
2125 * unaligned accesses for some archs when redzoning is used, and makes
2126 * sure any on-slab bufctl's are also correctly aligned.
2128 if (size & (BYTES_PER_WORD - 1)) {
2129 size += (BYTES_PER_WORD - 1);
2130 size &= ~(BYTES_PER_WORD - 1);
2133 if (flags & SLAB_RED_ZONE) {
2134 ralign = REDZONE_ALIGN;
2135 /* If redzoning, ensure that the second redzone is suitably
2136 * aligned, by adjusting the object size accordingly. */
2137 size += REDZONE_ALIGN - 1;
2138 size &= ~(REDZONE_ALIGN - 1);
2141 /* 3) caller mandated alignment */
2142 if (ralign < cachep->align) {
2143 ralign = cachep->align;
2145 /* disable debug if necessary */
2146 if (ralign > __alignof__(unsigned long long))
2147 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2151 cachep->align = ralign;
2153 if (slab_is_available())
2161 * Both debugging options require word-alignment which is calculated
2164 if (flags & SLAB_RED_ZONE) {
2165 /* add space for red zone words */
2166 cachep->obj_offset += sizeof(unsigned long long);
2167 size += 2 * sizeof(unsigned long long);
2169 if (flags & SLAB_STORE_USER) {
2170 /* user store requires one word storage behind the end of
2171 * the real object. But if the second red zone needs to be
2172 * aligned to 64 bits, we must allow that much space.
2174 if (flags & SLAB_RED_ZONE)
2175 size += REDZONE_ALIGN;
2177 size += BYTES_PER_WORD;
2181 kasan_cache_create(cachep, &size, &flags);
2183 size = ALIGN(size, cachep->align);
2185 * We should restrict the number of objects in a slab to implement
2186 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2188 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2189 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2193 * To activate debug pagealloc, off-slab management is necessary
2194 * requirement. In early phase of initialization, small sized slab
2195 * doesn't get initialized so it would not be possible. So, we need
2196 * to check size >= 256. It guarantees that all necessary small
2197 * sized slab is initialized in current slab initialization sequence.
2199 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2200 !slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2201 size >= 256 && cachep->object_size > cache_line_size() &&
2203 cachep->obj_offset += PAGE_SIZE - size;
2209 * Determine if the slab management is 'on' or 'off' slab.
2210 * (bootstrapping cannot cope with offslab caches so don't do
2211 * it too early on. Always use on-slab management when
2212 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2214 if (size >= OFF_SLAB_MIN_SIZE && !slab_early_init &&
2215 !(flags & SLAB_NOLEAKTRACE)) {
2217 * Size is large, assume best to place the slab management obj
2218 * off-slab (should allow better packing of objs).
2220 flags |= CFLGS_OFF_SLAB;
2223 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2228 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2231 * If the slab has been placed off-slab, and we have enough space then
2232 * move it on-slab. This is at the expense of any extra colouring.
2234 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2235 flags &= ~CFLGS_OFF_SLAB;
2236 left_over -= freelist_size;
2239 if (flags & CFLGS_OFF_SLAB) {
2240 /* really off slab. No need for manual alignment */
2241 freelist_size = calculate_freelist_size(cachep->num, 0);
2244 cachep->colour_off = cache_line_size();
2245 /* Offset must be a multiple of the alignment. */
2246 if (cachep->colour_off < cachep->align)
2247 cachep->colour_off = cachep->align;
2248 cachep->colour = left_over / cachep->colour_off;
2249 cachep->freelist_size = freelist_size;
2250 cachep->flags = flags;
2251 cachep->allocflags = __GFP_COMP;
2252 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2253 cachep->allocflags |= GFP_DMA;
2254 cachep->size = size;
2255 cachep->reciprocal_buffer_size = reciprocal_value(size);
2259 * If we're going to use the generic kernel_map_pages()
2260 * poisoning, then it's going to smash the contents of
2261 * the redzone and userword anyhow, so switch them off.
2263 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2264 (cachep->flags & SLAB_POISON) &&
2265 is_debug_pagealloc_cache(cachep))
2266 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2269 if (OFF_SLAB(cachep)) {
2270 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2272 * This is a possibility for one of the kmalloc_{dma,}_caches.
2273 * But since we go off slab only for object size greater than
2274 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
2275 * in ascending order,this should not happen at all.
2276 * But leave a BUG_ON for some lucky dude.
2278 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2281 err = setup_cpu_cache(cachep, gfp);
2283 __kmem_cache_shutdown(cachep);
2291 static void check_irq_off(void)
2293 BUG_ON(!irqs_disabled());
2296 static void check_irq_on(void)
2298 BUG_ON(irqs_disabled());
2301 static void check_spinlock_acquired(struct kmem_cache *cachep)
2305 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2309 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2313 assert_spin_locked(&get_node(cachep, node)->list_lock);
2318 #define check_irq_off() do { } while(0)
2319 #define check_irq_on() do { } while(0)
2320 #define check_spinlock_acquired(x) do { } while(0)
2321 #define check_spinlock_acquired_node(x, y) do { } while(0)
2324 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2325 struct array_cache *ac,
2326 int force, int node);
2328 static void do_drain(void *arg)
2330 struct kmem_cache *cachep = arg;
2331 struct array_cache *ac;
2332 int node = numa_mem_id();
2333 struct kmem_cache_node *n;
2337 ac = cpu_cache_get(cachep);
2338 n = get_node(cachep, node);
2339 spin_lock(&n->list_lock);
2340 free_block(cachep, ac->entry, ac->avail, node, &list);
2341 spin_unlock(&n->list_lock);
2342 slabs_destroy(cachep, &list);
2346 static void drain_cpu_caches(struct kmem_cache *cachep)
2348 struct kmem_cache_node *n;
2351 on_each_cpu(do_drain, cachep, 1);
2353 for_each_kmem_cache_node(cachep, node, n)
2355 drain_alien_cache(cachep, n->alien);
2357 for_each_kmem_cache_node(cachep, node, n)
2358 drain_array(cachep, n, n->shared, 1, node);
2362 * Remove slabs from the list of free slabs.
2363 * Specify the number of slabs to drain in tofree.
2365 * Returns the actual number of slabs released.
2367 static int drain_freelist(struct kmem_cache *cache,
2368 struct kmem_cache_node *n, int tofree)
2370 struct list_head *p;
2375 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2377 spin_lock_irq(&n->list_lock);
2378 p = n->slabs_free.prev;
2379 if (p == &n->slabs_free) {
2380 spin_unlock_irq(&n->list_lock);
2384 page = list_entry(p, struct page, lru);
2386 BUG_ON(page->active);
2388 list_del(&page->lru);
2390 * Safe to drop the lock. The slab is no longer linked
2393 n->free_objects -= cache->num;
2394 spin_unlock_irq(&n->list_lock);
2395 slab_destroy(cache, page);
2402 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2406 struct kmem_cache_node *n;
2408 drain_cpu_caches(cachep);
2411 for_each_kmem_cache_node(cachep, node, n) {
2412 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2414 ret += !list_empty(&n->slabs_full) ||
2415 !list_empty(&n->slabs_partial);
2417 return (ret ? 1 : 0);
2420 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2423 struct kmem_cache_node *n;
2424 int rc = __kmem_cache_shrink(cachep, false);
2429 free_percpu(cachep->cpu_cache);
2431 /* NUMA: free the node structures */
2432 for_each_kmem_cache_node(cachep, i, n) {
2434 free_alien_cache(n->alien);
2436 cachep->node[i] = NULL;
2442 * Get the memory for a slab management obj.
2444 * For a slab cache when the slab descriptor is off-slab, the
2445 * slab descriptor can't come from the same cache which is being created,
2446 * Because if it is the case, that means we defer the creation of
2447 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2448 * And we eventually call down to __kmem_cache_create(), which
2449 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2450 * This is a "chicken-and-egg" problem.
2452 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2453 * which are all initialized during kmem_cache_init().
2455 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2456 struct page *page, int colour_off,
2457 gfp_t local_flags, int nodeid)
2460 void *addr = page_address(page);
2462 if (OFF_SLAB(cachep)) {
2463 /* Slab management obj is off-slab. */
2464 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2465 local_flags, nodeid);
2469 freelist = addr + colour_off;
2470 colour_off += cachep->freelist_size;
2473 page->s_mem = addr + colour_off;
2477 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2479 return ((freelist_idx_t *)page->freelist)[idx];
2482 static inline void set_free_obj(struct page *page,
2483 unsigned int idx, freelist_idx_t val)
2485 ((freelist_idx_t *)(page->freelist))[idx] = val;
2488 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2493 for (i = 0; i < cachep->num; i++) {
2494 void *objp = index_to_obj(cachep, page, i);
2495 kasan_init_slab_obj(cachep, objp);
2496 if (cachep->flags & SLAB_STORE_USER)
2497 *dbg_userword(cachep, objp) = NULL;
2499 if (cachep->flags & SLAB_RED_ZONE) {
2500 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2501 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2504 * Constructors are not allowed to allocate memory from the same
2505 * cache which they are a constructor for. Otherwise, deadlock.
2506 * They must also be threaded.
2508 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2509 kasan_unpoison_object_data(cachep,
2510 objp + obj_offset(cachep));
2511 cachep->ctor(objp + obj_offset(cachep));
2512 kasan_poison_object_data(
2513 cachep, objp + obj_offset(cachep));
2516 if (cachep->flags & SLAB_RED_ZONE) {
2517 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2518 slab_error(cachep, "constructor overwrote the end of an object");
2519 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2520 slab_error(cachep, "constructor overwrote the start of an object");
2522 /* need to poison the objs? */
2523 if (cachep->flags & SLAB_POISON) {
2524 poison_obj(cachep, objp, POISON_FREE);
2525 slab_kernel_map(cachep, objp, 0, 0);
2531 static void cache_init_objs(struct kmem_cache *cachep,
2537 cache_init_objs_debug(cachep, page);
2539 for (i = 0; i < cachep->num; i++) {
2540 /* constructor could break poison info */
2541 if (DEBUG == 0 && cachep->ctor) {
2542 objp = index_to_obj(cachep, page, i);
2543 kasan_unpoison_object_data(cachep, objp);
2545 kasan_poison_object_data(cachep, objp);
2548 set_free_obj(page, i, i);
2552 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2554 if (CONFIG_ZONE_DMA_FLAG) {
2555 if (flags & GFP_DMA)
2556 BUG_ON(!(cachep->allocflags & GFP_DMA));
2558 BUG_ON(cachep->allocflags & GFP_DMA);
2562 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
2567 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2570 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2574 if (cachep->flags & SLAB_STORE_USER)
2575 set_store_user_dirty(cachep);
2581 static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
2582 void *objp, int nodeid)
2584 unsigned int objnr = obj_to_index(cachep, page, objp);
2588 /* Verify that the slab belongs to the intended node */
2589 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2591 /* Verify double free bug */
2592 for (i = page->active; i < cachep->num; i++) {
2593 if (get_free_obj(page, i) == objnr) {
2594 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p\n",
2595 cachep->name, objp);
2601 set_free_obj(page, page->active, objnr);
2605 * Map pages beginning at addr to the given cache and slab. This is required
2606 * for the slab allocator to be able to lookup the cache and slab of a
2607 * virtual address for kfree, ksize, and slab debugging.
2609 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2612 page->slab_cache = cache;
2613 page->freelist = freelist;
2617 * Grow (by 1) the number of slabs within a cache. This is called by
2618 * kmem_cache_alloc() when there are no active objs left in a cache.
2620 static int cache_grow(struct kmem_cache *cachep,
2621 gfp_t flags, int nodeid, struct page *page)
2626 struct kmem_cache_node *n;
2629 * Be lazy and only check for valid flags here, keeping it out of the
2630 * critical path in kmem_cache_alloc().
2632 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2633 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2636 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2638 /* Take the node list lock to change the colour_next on this node */
2640 n = get_node(cachep, nodeid);
2641 spin_lock(&n->list_lock);
2643 /* Get colour for the slab, and cal the next value. */
2644 offset = n->colour_next;
2646 if (n->colour_next >= cachep->colour)
2648 spin_unlock(&n->list_lock);
2650 offset *= cachep->colour_off;
2652 if (gfpflags_allow_blocking(local_flags))
2656 * The test for missing atomic flag is performed here, rather than
2657 * the more obvious place, simply to reduce the critical path length
2658 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2659 * will eventually be caught here (where it matters).
2661 kmem_flagcheck(cachep, flags);
2664 * Get mem for the objs. Attempt to allocate a physical page from
2668 page = kmem_getpages(cachep, local_flags, nodeid);
2672 /* Get slab management. */
2673 freelist = alloc_slabmgmt(cachep, page, offset,
2674 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2678 slab_map_pages(cachep, page, freelist);
2680 kasan_poison_slab(page);
2681 cache_init_objs(cachep, page);
2683 if (gfpflags_allow_blocking(local_flags))
2684 local_irq_disable();
2686 spin_lock(&n->list_lock);
2688 /* Make slab active. */
2689 list_add_tail(&page->lru, &(n->slabs_free));
2690 STATS_INC_GROWN(cachep);
2691 n->free_objects += cachep->num;
2692 spin_unlock(&n->list_lock);
2695 kmem_freepages(cachep, page);
2697 if (gfpflags_allow_blocking(local_flags))
2698 local_irq_disable();
2705 * Perform extra freeing checks:
2706 * - detect bad pointers.
2707 * - POISON/RED_ZONE checking
2709 static void kfree_debugcheck(const void *objp)
2711 if (!virt_addr_valid(objp)) {
2712 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2713 (unsigned long)objp);
2718 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2720 unsigned long long redzone1, redzone2;
2722 redzone1 = *dbg_redzone1(cache, obj);
2723 redzone2 = *dbg_redzone2(cache, obj);
2728 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2731 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2732 slab_error(cache, "double free detected");
2734 slab_error(cache, "memory outside object was overwritten");
2736 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2737 obj, redzone1, redzone2);
2740 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2741 unsigned long caller)
2746 BUG_ON(virt_to_cache(objp) != cachep);
2748 objp -= obj_offset(cachep);
2749 kfree_debugcheck(objp);
2750 page = virt_to_head_page(objp);
2752 if (cachep->flags & SLAB_RED_ZONE) {
2753 verify_redzone_free(cachep, objp);
2754 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2755 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2757 if (cachep->flags & SLAB_STORE_USER) {
2758 set_store_user_dirty(cachep);
2759 *dbg_userword(cachep, objp) = (void *)caller;
2762 objnr = obj_to_index(cachep, page, objp);
2764 BUG_ON(objnr >= cachep->num);
2765 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2767 if (cachep->flags & SLAB_POISON) {
2768 poison_obj(cachep, objp, POISON_FREE);
2769 slab_kernel_map(cachep, objp, 0, caller);
2775 #define kfree_debugcheck(x) do { } while(0)
2776 #define cache_free_debugcheck(x,objp,z) (objp)
2779 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2783 struct kmem_cache_node *n;
2784 struct array_cache *ac;
2788 node = numa_mem_id();
2789 if (unlikely(force_refill))
2792 ac = cpu_cache_get(cachep);
2793 batchcount = ac->batchcount;
2794 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2796 * If there was little recent activity on this cache, then
2797 * perform only a partial refill. Otherwise we could generate
2800 batchcount = BATCHREFILL_LIMIT;
2802 n = get_node(cachep, node);
2804 BUG_ON(ac->avail > 0 || !n);
2805 spin_lock(&n->list_lock);
2807 /* See if we can refill from the shared array */
2808 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2809 n->shared->touched = 1;
2813 while (batchcount > 0) {
2814 struct list_head *entry;
2816 /* Get slab alloc is to come from. */
2817 entry = n->slabs_partial.next;
2818 if (entry == &n->slabs_partial) {
2819 n->free_touched = 1;
2820 entry = n->slabs_free.next;
2821 if (entry == &n->slabs_free)
2825 page = list_entry(entry, struct page, lru);
2826 check_spinlock_acquired(cachep);
2829 * The slab was either on partial or free list so
2830 * there must be at least one object available for
2833 BUG_ON(page->active >= cachep->num);
2835 while (page->active < cachep->num && batchcount--) {
2836 STATS_INC_ALLOCED(cachep);
2837 STATS_INC_ACTIVE(cachep);
2838 STATS_SET_HIGH(cachep);
2840 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
2844 /* move slabp to correct slabp list: */
2845 list_del(&page->lru);
2846 if (page->active == cachep->num)
2847 list_add(&page->lru, &n->slabs_full);
2849 list_add(&page->lru, &n->slabs_partial);
2853 n->free_objects -= ac->avail;
2855 spin_unlock(&n->list_lock);
2857 if (unlikely(!ac->avail)) {
2860 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2862 /* cache_grow can reenable interrupts, then ac could change. */
2863 ac = cpu_cache_get(cachep);
2864 node = numa_mem_id();
2866 /* no objects in sight? abort */
2867 if (!x && (ac->avail == 0 || force_refill))
2870 if (!ac->avail) /* objects refilled by interrupt? */
2875 return ac_get_obj(cachep, ac, flags, force_refill);
2878 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2881 might_sleep_if(gfpflags_allow_blocking(flags));
2883 kmem_flagcheck(cachep, flags);
2888 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2889 gfp_t flags, void *objp, unsigned long caller)
2893 if (cachep->flags & SLAB_POISON) {
2894 check_poison_obj(cachep, objp);
2895 slab_kernel_map(cachep, objp, 1, 0);
2896 poison_obj(cachep, objp, POISON_INUSE);
2898 if (cachep->flags & SLAB_STORE_USER)
2899 *dbg_userword(cachep, objp) = (void *)caller;
2901 if (cachep->flags & SLAB_RED_ZONE) {
2902 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2903 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2904 slab_error(cachep, "double free, or memory outside object was overwritten");
2906 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2907 objp, *dbg_redzone1(cachep, objp),
2908 *dbg_redzone2(cachep, objp));
2910 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2911 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2914 objp += obj_offset(cachep);
2915 if (cachep->ctor && cachep->flags & SLAB_POISON)
2917 if (ARCH_SLAB_MINALIGN &&
2918 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2919 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2920 objp, (int)ARCH_SLAB_MINALIGN);
2925 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2928 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
2930 if (unlikely(cachep == kmem_cache))
2933 return should_failslab(cachep->object_size, flags, cachep->flags);
2936 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2939 struct array_cache *ac;
2940 bool force_refill = false;
2944 ac = cpu_cache_get(cachep);
2945 if (likely(ac->avail)) {
2947 objp = ac_get_obj(cachep, ac, flags, false);
2950 * Allow for the possibility all avail objects are not allowed
2951 * by the current flags
2954 STATS_INC_ALLOCHIT(cachep);
2957 force_refill = true;
2960 STATS_INC_ALLOCMISS(cachep);
2961 objp = cache_alloc_refill(cachep, flags, force_refill);
2963 * the 'ac' may be updated by cache_alloc_refill(),
2964 * and kmemleak_erase() requires its correct value.
2966 ac = cpu_cache_get(cachep);
2970 * To avoid a false negative, if an object that is in one of the
2971 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2972 * treat the array pointers as a reference to the object.
2975 kmemleak_erase(&ac->entry[ac->avail]);
2981 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2983 * If we are in_interrupt, then process context, including cpusets and
2984 * mempolicy, may not apply and should not be used for allocation policy.
2986 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2988 int nid_alloc, nid_here;
2990 if (in_interrupt() || (flags & __GFP_THISNODE))
2992 nid_alloc = nid_here = numa_mem_id();
2993 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2994 nid_alloc = cpuset_slab_spread_node();
2995 else if (current->mempolicy)
2996 nid_alloc = mempolicy_slab_node();
2997 if (nid_alloc != nid_here)
2998 return ____cache_alloc_node(cachep, flags, nid_alloc);
3003 * Fallback function if there was no memory available and no objects on a
3004 * certain node and fall back is permitted. First we scan all the
3005 * available node for available objects. If that fails then we
3006 * perform an allocation without specifying a node. This allows the page
3007 * allocator to do its reclaim / fallback magic. We then insert the
3008 * slab into the proper nodelist and then allocate from it.
3010 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3012 struct zonelist *zonelist;
3016 enum zone_type high_zoneidx = gfp_zone(flags);
3019 unsigned int cpuset_mems_cookie;
3021 if (flags & __GFP_THISNODE)
3024 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3027 cpuset_mems_cookie = read_mems_allowed_begin();
3028 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3032 * Look through allowed nodes for objects available
3033 * from existing per node queues.
3035 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3036 nid = zone_to_nid(zone);
3038 if (cpuset_zone_allowed(zone, flags) &&
3039 get_node(cache, nid) &&
3040 get_node(cache, nid)->free_objects) {
3041 obj = ____cache_alloc_node(cache,
3042 gfp_exact_node(flags), nid);
3050 * This allocation will be performed within the constraints
3051 * of the current cpuset / memory policy requirements.
3052 * We may trigger various forms of reclaim on the allowed
3053 * set and go into memory reserves if necessary.
3057 if (gfpflags_allow_blocking(local_flags))
3059 kmem_flagcheck(cache, flags);
3060 page = kmem_getpages(cache, local_flags, numa_mem_id());
3061 if (gfpflags_allow_blocking(local_flags))
3062 local_irq_disable();
3065 * Insert into the appropriate per node queues
3067 nid = page_to_nid(page);
3068 if (cache_grow(cache, flags, nid, page)) {
3069 obj = ____cache_alloc_node(cache,
3070 gfp_exact_node(flags), nid);
3073 * Another processor may allocate the
3074 * objects in the slab since we are
3075 * not holding any locks.
3079 /* cache_grow already freed obj */
3085 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3091 * A interface to enable slab creation on nodeid
3093 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3096 struct list_head *entry;
3098 struct kmem_cache_node *n;
3102 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3103 n = get_node(cachep, nodeid);
3108 spin_lock(&n->list_lock);
3109 entry = n->slabs_partial.next;
3110 if (entry == &n->slabs_partial) {
3111 n->free_touched = 1;
3112 entry = n->slabs_free.next;
3113 if (entry == &n->slabs_free)
3117 page = list_entry(entry, struct page, lru);
3118 check_spinlock_acquired_node(cachep, nodeid);
3120 STATS_INC_NODEALLOCS(cachep);
3121 STATS_INC_ACTIVE(cachep);
3122 STATS_SET_HIGH(cachep);
3124 BUG_ON(page->active == cachep->num);
3126 obj = slab_get_obj(cachep, page, nodeid);
3128 /* move slabp to correct slabp list: */
3129 list_del(&page->lru);
3131 if (page->active == cachep->num)
3132 list_add(&page->lru, &n->slabs_full);
3134 list_add(&page->lru, &n->slabs_partial);
3136 spin_unlock(&n->list_lock);
3140 spin_unlock(&n->list_lock);
3141 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3145 return fallback_alloc(cachep, flags);
3151 static __always_inline void *
3152 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3153 unsigned long caller)
3155 unsigned long save_flags;
3157 int slab_node = numa_mem_id();
3159 flags &= gfp_allowed_mask;
3161 lockdep_trace_alloc(flags);
3163 if (slab_should_failslab(cachep, flags))
3166 cachep = memcg_kmem_get_cache(cachep, flags);
3168 cache_alloc_debugcheck_before(cachep, flags);
3169 local_irq_save(save_flags);
3171 if (nodeid == NUMA_NO_NODE)
3174 if (unlikely(!get_node(cachep, nodeid))) {
3175 /* Node not bootstrapped yet */
3176 ptr = fallback_alloc(cachep, flags);
3180 if (nodeid == slab_node) {
3182 * Use the locally cached objects if possible.
3183 * However ____cache_alloc does not allow fallback
3184 * to other nodes. It may fail while we still have
3185 * objects on other nodes available.
3187 ptr = ____cache_alloc(cachep, flags);
3191 /* ___cache_alloc_node can fall back to other nodes */
3192 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3194 local_irq_restore(save_flags);
3195 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3196 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3200 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3201 if (unlikely(flags & __GFP_ZERO))
3202 memset(ptr, 0, cachep->object_size);
3205 memcg_kmem_put_cache(cachep);
3209 static __always_inline void *
3210 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3214 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3215 objp = alternate_node_alloc(cache, flags);
3219 objp = ____cache_alloc(cache, flags);
3222 * We may just have run out of memory on the local node.
3223 * ____cache_alloc_node() knows how to locate memory on other nodes
3226 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3233 static __always_inline void *
3234 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3236 return ____cache_alloc(cachep, flags);
3239 #endif /* CONFIG_NUMA */
3241 static __always_inline void *
3242 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3244 unsigned long save_flags;
3247 flags &= gfp_allowed_mask;
3249 lockdep_trace_alloc(flags);
3251 if (slab_should_failslab(cachep, flags))
3254 cachep = memcg_kmem_get_cache(cachep, flags);
3256 cache_alloc_debugcheck_before(cachep, flags);
3257 local_irq_save(save_flags);
3258 objp = __do_cache_alloc(cachep, flags);
3259 local_irq_restore(save_flags);
3260 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3261 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3266 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3267 if (unlikely(flags & __GFP_ZERO))
3268 memset(objp, 0, cachep->object_size);
3271 memcg_kmem_put_cache(cachep);
3276 * Caller needs to acquire correct kmem_cache_node's list_lock
3277 * @list: List of detached free slabs should be freed by caller
3279 static void free_block(struct kmem_cache *cachep, void **objpp,
3280 int nr_objects, int node, struct list_head *list)
3283 struct kmem_cache_node *n = get_node(cachep, node);
3285 for (i = 0; i < nr_objects; i++) {
3289 clear_obj_pfmemalloc(&objpp[i]);
3292 page = virt_to_head_page(objp);
3293 list_del(&page->lru);
3294 check_spinlock_acquired_node(cachep, node);
3295 slab_put_obj(cachep, page, objp, node);
3296 STATS_DEC_ACTIVE(cachep);
3299 /* fixup slab chains */
3300 if (page->active == 0) {
3301 if (n->free_objects > n->free_limit) {
3302 n->free_objects -= cachep->num;
3303 list_add_tail(&page->lru, list);
3305 list_add(&page->lru, &n->slabs_free);
3308 /* Unconditionally move a slab to the end of the
3309 * partial list on free - maximum time for the
3310 * other objects to be freed, too.
3312 list_add_tail(&page->lru, &n->slabs_partial);
3317 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3320 struct kmem_cache_node *n;
3321 int node = numa_mem_id();
3324 batchcount = ac->batchcount;
3326 BUG_ON(!batchcount || batchcount > ac->avail);
3329 n = get_node(cachep, node);
3330 spin_lock(&n->list_lock);
3332 struct array_cache *shared_array = n->shared;
3333 int max = shared_array->limit - shared_array->avail;
3335 if (batchcount > max)
3337 memcpy(&(shared_array->entry[shared_array->avail]),
3338 ac->entry, sizeof(void *) * batchcount);
3339 shared_array->avail += batchcount;
3344 free_block(cachep, ac->entry, batchcount, node, &list);
3349 struct list_head *p;
3351 p = n->slabs_free.next;
3352 while (p != &(n->slabs_free)) {
3355 page = list_entry(p, struct page, lru);
3356 BUG_ON(page->active);
3361 STATS_SET_FREEABLE(cachep, i);
3364 spin_unlock(&n->list_lock);
3365 slabs_destroy(cachep, &list);
3366 ac->avail -= batchcount;
3367 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3371 * Release an obj back to its cache. If the obj has a constructed state, it must
3372 * be in this state _before_ it is released. Called with disabled ints.
3374 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3375 unsigned long caller)
3377 /* Put the object into the quarantine, don't touch it for now. */
3378 if (kasan_slab_free(cachep, objp))
3381 ___cache_free(cachep, objp, caller);
3384 void ___cache_free(struct kmem_cache *cachep, void *objp,
3385 unsigned long caller)
3387 struct array_cache *ac = cpu_cache_get(cachep);
3390 kmemleak_free_recursive(objp, cachep->flags);
3391 objp = cache_free_debugcheck(cachep, objp, caller);
3393 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3396 * Skip calling cache_free_alien() when the platform is not numa.
3397 * This will avoid cache misses that happen while accessing slabp (which
3398 * is per page memory reference) to get nodeid. Instead use a global
3399 * variable to skip the call, which is mostly likely to be present in
3402 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3405 if (ac->avail < ac->limit) {
3406 STATS_INC_FREEHIT(cachep);
3408 STATS_INC_FREEMISS(cachep);
3409 cache_flusharray(cachep, ac);
3412 ac_put_obj(cachep, ac, objp);
3416 * kmem_cache_alloc - Allocate an object
3417 * @cachep: The cache to allocate from.
3418 * @flags: See kmalloc().
3420 * Allocate an object from this cache. The flags are only relevant
3421 * if the cache has no available objects.
3423 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3425 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3427 kasan_slab_alloc(cachep, ret, flags);
3428 trace_kmem_cache_alloc(_RET_IP_, ret,
3429 cachep->object_size, cachep->size, flags);
3433 EXPORT_SYMBOL(kmem_cache_alloc);
3435 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3437 __kmem_cache_free_bulk(s, size, p);
3439 EXPORT_SYMBOL(kmem_cache_free_bulk);
3441 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3444 return __kmem_cache_alloc_bulk(s, flags, size, p);
3446 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3448 #ifdef CONFIG_TRACING
3450 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3454 ret = slab_alloc(cachep, flags, _RET_IP_);
3456 kasan_kmalloc(cachep, ret, size, flags);
3457 trace_kmalloc(_RET_IP_, ret,
3458 size, cachep->size, flags);
3461 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3466 * kmem_cache_alloc_node - Allocate an object on the specified node
3467 * @cachep: The cache to allocate from.
3468 * @flags: See kmalloc().
3469 * @nodeid: node number of the target node.
3471 * Identical to kmem_cache_alloc but it will allocate memory on the given
3472 * node, which can improve the performance for cpu bound structures.
3474 * Fallback to other node is possible if __GFP_THISNODE is not set.
3476 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3478 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3480 kasan_slab_alloc(cachep, ret, flags);
3481 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3482 cachep->object_size, cachep->size,
3487 EXPORT_SYMBOL(kmem_cache_alloc_node);
3489 #ifdef CONFIG_TRACING
3490 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3497 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3499 kasan_kmalloc(cachep, ret, size, flags);
3500 trace_kmalloc_node(_RET_IP_, ret,
3505 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3508 static __always_inline void *
3509 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3511 struct kmem_cache *cachep;
3514 cachep = kmalloc_slab(size, flags);
3515 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3517 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3518 kasan_kmalloc(cachep, ret, size, flags);
3523 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3525 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3527 EXPORT_SYMBOL(__kmalloc_node);
3529 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3530 int node, unsigned long caller)
3532 return __do_kmalloc_node(size, flags, node, caller);
3534 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3535 #endif /* CONFIG_NUMA */
3538 * __do_kmalloc - allocate memory
3539 * @size: how many bytes of memory are required.
3540 * @flags: the type of memory to allocate (see kmalloc).
3541 * @caller: function caller for debug tracking of the caller
3543 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3544 unsigned long caller)
3546 struct kmem_cache *cachep;
3549 cachep = kmalloc_slab(size, flags);
3550 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3552 ret = slab_alloc(cachep, flags, caller);
3554 kasan_kmalloc(cachep, ret, size, flags);
3555 trace_kmalloc(caller, ret,
3556 size, cachep->size, flags);
3561 void *__kmalloc(size_t size, gfp_t flags)
3563 return __do_kmalloc(size, flags, _RET_IP_);
3565 EXPORT_SYMBOL(__kmalloc);
3567 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3569 return __do_kmalloc(size, flags, caller);
3571 EXPORT_SYMBOL(__kmalloc_track_caller);
3574 * kmem_cache_free - Deallocate an object
3575 * @cachep: The cache the allocation was from.
3576 * @objp: The previously allocated object.
3578 * Free an object which was previously allocated from this
3581 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3583 unsigned long flags;
3584 cachep = cache_from_obj(cachep, objp);
3588 local_irq_save(flags);
3589 debug_check_no_locks_freed(objp, cachep->object_size);
3590 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3591 debug_check_no_obj_freed(objp, cachep->object_size);
3592 __cache_free(cachep, objp, _RET_IP_);
3593 local_irq_restore(flags);
3595 trace_kmem_cache_free(_RET_IP_, objp);
3597 EXPORT_SYMBOL(kmem_cache_free);
3600 * kfree - free previously allocated memory
3601 * @objp: pointer returned by kmalloc.
3603 * If @objp is NULL, no operation is performed.
3605 * Don't free memory not originally allocated by kmalloc()
3606 * or you will run into trouble.
3608 void kfree(const void *objp)
3610 struct kmem_cache *c;
3611 unsigned long flags;
3613 trace_kfree(_RET_IP_, objp);
3615 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3617 local_irq_save(flags);
3618 kfree_debugcheck(objp);
3619 c = virt_to_cache(objp);
3620 debug_check_no_locks_freed(objp, c->object_size);
3622 debug_check_no_obj_freed(objp, c->object_size);
3623 __cache_free(c, (void *)objp, _RET_IP_);
3624 local_irq_restore(flags);
3626 EXPORT_SYMBOL(kfree);
3629 * This initializes kmem_cache_node or resizes various caches for all nodes.
3631 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3634 struct kmem_cache_node *n;
3635 struct array_cache *new_shared;
3636 struct alien_cache **new_alien = NULL;
3638 for_each_online_node(node) {
3640 if (use_alien_caches) {
3641 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3647 if (cachep->shared) {
3648 new_shared = alloc_arraycache(node,
3649 cachep->shared*cachep->batchcount,
3652 free_alien_cache(new_alien);
3657 n = get_node(cachep, node);
3659 struct array_cache *shared = n->shared;
3662 spin_lock_irq(&n->list_lock);
3665 free_block(cachep, shared->entry,
3666 shared->avail, node, &list);
3668 n->shared = new_shared;
3670 n->alien = new_alien;
3673 n->free_limit = (1 + nr_cpus_node(node)) *
3674 cachep->batchcount + cachep->num;
3675 spin_unlock_irq(&n->list_lock);
3676 slabs_destroy(cachep, &list);
3678 free_alien_cache(new_alien);
3681 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3683 free_alien_cache(new_alien);
3688 kmem_cache_node_init(n);
3689 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3690 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3691 n->shared = new_shared;
3692 n->alien = new_alien;
3693 n->free_limit = (1 + nr_cpus_node(node)) *
3694 cachep->batchcount + cachep->num;
3695 cachep->node[node] = n;
3700 if (!cachep->list.next) {
3701 /* Cache is not active yet. Roll back what we did */
3704 n = get_node(cachep, node);
3707 free_alien_cache(n->alien);
3709 cachep->node[node] = NULL;
3717 /* Always called with the slab_mutex held */
3718 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3719 int batchcount, int shared, gfp_t gfp)
3721 struct array_cache __percpu *cpu_cache, *prev;
3724 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3728 prev = cachep->cpu_cache;
3729 cachep->cpu_cache = cpu_cache;
3730 kick_all_cpus_sync();
3733 cachep->batchcount = batchcount;
3734 cachep->limit = limit;
3735 cachep->shared = shared;
3740 for_each_online_cpu(cpu) {
3743 struct kmem_cache_node *n;
3744 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3746 node = cpu_to_mem(cpu);
3747 n = get_node(cachep, node);
3748 spin_lock_irq(&n->list_lock);
3749 free_block(cachep, ac->entry, ac->avail, node, &list);
3750 spin_unlock_irq(&n->list_lock);
3751 slabs_destroy(cachep, &list);
3756 return alloc_kmem_cache_node(cachep, gfp);
3759 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3760 int batchcount, int shared, gfp_t gfp)
3763 struct kmem_cache *c;
3765 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3767 if (slab_state < FULL)
3770 if ((ret < 0) || !is_root_cache(cachep))
3773 lockdep_assert_held(&slab_mutex);
3774 for_each_memcg_cache(c, cachep) {
3775 /* return value determined by the root cache only */
3776 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3782 /* Called with slab_mutex held always */
3783 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3790 if (!is_root_cache(cachep)) {
3791 struct kmem_cache *root = memcg_root_cache(cachep);
3792 limit = root->limit;
3793 shared = root->shared;
3794 batchcount = root->batchcount;
3797 if (limit && shared && batchcount)
3800 * The head array serves three purposes:
3801 * - create a LIFO ordering, i.e. return objects that are cache-warm
3802 * - reduce the number of spinlock operations.
3803 * - reduce the number of linked list operations on the slab and
3804 * bufctl chains: array operations are cheaper.
3805 * The numbers are guessed, we should auto-tune as described by
3808 if (cachep->size > 131072)
3810 else if (cachep->size > PAGE_SIZE)
3812 else if (cachep->size > 1024)
3814 else if (cachep->size > 256)
3820 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3821 * allocation behaviour: Most allocs on one cpu, most free operations
3822 * on another cpu. For these cases, an efficient object passing between
3823 * cpus is necessary. This is provided by a shared array. The array
3824 * replaces Bonwick's magazine layer.
3825 * On uniprocessor, it's functionally equivalent (but less efficient)
3826 * to a larger limit. Thus disabled by default.
3829 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3834 * With debugging enabled, large batchcount lead to excessively long
3835 * periods with disabled local interrupts. Limit the batchcount
3840 batchcount = (limit + 1) / 2;
3842 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3844 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3845 cachep->name, -err);
3850 * Drain an array if it contains any elements taking the node lock only if
3851 * necessary. Note that the node listlock also protects the array_cache
3852 * if drain_array() is used on the shared array.
3854 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3855 struct array_cache *ac, int force, int node)
3860 if (!ac || !ac->avail)
3862 if (ac->touched && !force) {
3865 spin_lock_irq(&n->list_lock);
3867 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3868 if (tofree > ac->avail)
3869 tofree = (ac->avail + 1) / 2;
3870 free_block(cachep, ac->entry, tofree, node, &list);
3871 ac->avail -= tofree;
3872 memmove(ac->entry, &(ac->entry[tofree]),
3873 sizeof(void *) * ac->avail);
3875 spin_unlock_irq(&n->list_lock);
3876 slabs_destroy(cachep, &list);
3881 * cache_reap - Reclaim memory from caches.
3882 * @w: work descriptor
3884 * Called from workqueue/eventd every few seconds.
3886 * - clear the per-cpu caches for this CPU.
3887 * - return freeable pages to the main free memory pool.
3889 * If we cannot acquire the cache chain mutex then just give up - we'll try
3890 * again on the next iteration.
3892 static void cache_reap(struct work_struct *w)
3894 struct kmem_cache *searchp;
3895 struct kmem_cache_node *n;
3896 int node = numa_mem_id();
3897 struct delayed_work *work = to_delayed_work(w);
3899 if (!mutex_trylock(&slab_mutex))
3900 /* Give up. Setup the next iteration. */
3903 list_for_each_entry(searchp, &slab_caches, list) {
3907 * We only take the node lock if absolutely necessary and we
3908 * have established with reasonable certainty that
3909 * we can do some work if the lock was obtained.
3911 n = get_node(searchp, node);
3913 reap_alien(searchp, n);
3915 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3918 * These are racy checks but it does not matter
3919 * if we skip one check or scan twice.
3921 if (time_after(n->next_reap, jiffies))
3924 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3926 drain_array(searchp, n, n->shared, 0, node);
3928 if (n->free_touched)
3929 n->free_touched = 0;
3933 freed = drain_freelist(searchp, n, (n->free_limit +
3934 5 * searchp->num - 1) / (5 * searchp->num));
3935 STATS_ADD_REAPED(searchp, freed);
3941 mutex_unlock(&slab_mutex);
3944 /* Set up the next iteration */
3945 schedule_delayed_work_on(smp_processor_id(), work,
3946 round_jiffies_relative(REAPTIMEOUT_AC));
3949 #ifdef CONFIG_SLABINFO
3950 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3953 unsigned long active_objs;
3954 unsigned long num_objs;
3955 unsigned long active_slabs = 0;
3956 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3960 struct kmem_cache_node *n;
3964 for_each_kmem_cache_node(cachep, node, n) {
3967 spin_lock_irq(&n->list_lock);
3969 list_for_each_entry(page, &n->slabs_full, lru) {
3970 if (page->active != cachep->num && !error)
3971 error = "slabs_full accounting error";
3972 active_objs += cachep->num;
3975 list_for_each_entry(page, &n->slabs_partial, lru) {
3976 if (page->active == cachep->num && !error)
3977 error = "slabs_partial accounting error";
3978 if (!page->active && !error)
3979 error = "slabs_partial accounting error";
3980 active_objs += page->active;
3983 list_for_each_entry(page, &n->slabs_free, lru) {
3984 if (page->active && !error)
3985 error = "slabs_free accounting error";
3988 free_objects += n->free_objects;
3990 shared_avail += n->shared->avail;
3992 spin_unlock_irq(&n->list_lock);
3994 num_slabs += active_slabs;
3995 num_objs = num_slabs * cachep->num;
3996 if (num_objs - active_objs != free_objects && !error)
3997 error = "free_objects accounting error";
3999 name = cachep->name;
4001 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4003 sinfo->active_objs = active_objs;
4004 sinfo->num_objs = num_objs;
4005 sinfo->active_slabs = active_slabs;
4006 sinfo->num_slabs = num_slabs;
4007 sinfo->shared_avail = shared_avail;
4008 sinfo->limit = cachep->limit;
4009 sinfo->batchcount = cachep->batchcount;
4010 sinfo->shared = cachep->shared;
4011 sinfo->objects_per_slab = cachep->num;
4012 sinfo->cache_order = cachep->gfporder;
4015 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4019 unsigned long high = cachep->high_mark;
4020 unsigned long allocs = cachep->num_allocations;
4021 unsigned long grown = cachep->grown;
4022 unsigned long reaped = cachep->reaped;
4023 unsigned long errors = cachep->errors;
4024 unsigned long max_freeable = cachep->max_freeable;
4025 unsigned long node_allocs = cachep->node_allocs;
4026 unsigned long node_frees = cachep->node_frees;
4027 unsigned long overflows = cachep->node_overflow;
4029 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4030 allocs, high, grown,
4031 reaped, errors, max_freeable, node_allocs,
4032 node_frees, overflows);
4036 unsigned long allochit = atomic_read(&cachep->allochit);
4037 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4038 unsigned long freehit = atomic_read(&cachep->freehit);
4039 unsigned long freemiss = atomic_read(&cachep->freemiss);
4041 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4042 allochit, allocmiss, freehit, freemiss);
4047 #define MAX_SLABINFO_WRITE 128
4049 * slabinfo_write - Tuning for the slab allocator
4051 * @buffer: user buffer
4052 * @count: data length
4055 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4056 size_t count, loff_t *ppos)
4058 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4059 int limit, batchcount, shared, res;
4060 struct kmem_cache *cachep;
4062 if (count > MAX_SLABINFO_WRITE)
4064 if (copy_from_user(&kbuf, buffer, count))
4066 kbuf[MAX_SLABINFO_WRITE] = '\0';
4068 tmp = strchr(kbuf, ' ');
4073 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4076 /* Find the cache in the chain of caches. */
4077 mutex_lock(&slab_mutex);
4079 list_for_each_entry(cachep, &slab_caches, list) {
4080 if (!strcmp(cachep->name, kbuf)) {
4081 if (limit < 1 || batchcount < 1 ||
4082 batchcount > limit || shared < 0) {
4085 res = do_tune_cpucache(cachep, limit,
4092 mutex_unlock(&slab_mutex);
4098 #ifdef CONFIG_DEBUG_SLAB_LEAK
4100 static inline int add_caller(unsigned long *n, unsigned long v)
4110 unsigned long *q = p + 2 * i;
4124 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4130 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4139 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4142 for (j = page->active; j < c->num; j++) {
4143 if (get_free_obj(page, j) == i) {
4153 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4154 * mapping is established when actual object allocation and
4155 * we could mistakenly access the unmapped object in the cpu
4158 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4161 if (!add_caller(n, v))
4166 static void show_symbol(struct seq_file *m, unsigned long address)
4168 #ifdef CONFIG_KALLSYMS
4169 unsigned long offset, size;
4170 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4172 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4173 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4175 seq_printf(m, " [%s]", modname);
4179 seq_printf(m, "%p", (void *)address);
4182 static int leaks_show(struct seq_file *m, void *p)
4184 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4186 struct kmem_cache_node *n;
4188 unsigned long *x = m->private;
4192 if (!(cachep->flags & SLAB_STORE_USER))
4194 if (!(cachep->flags & SLAB_RED_ZONE))
4198 * Set store_user_clean and start to grab stored user information
4199 * for all objects on this cache. If some alloc/free requests comes
4200 * during the processing, information would be wrong so restart
4204 set_store_user_clean(cachep);
4205 drain_cpu_caches(cachep);
4209 for_each_kmem_cache_node(cachep, node, n) {
4212 spin_lock_irq(&n->list_lock);
4214 list_for_each_entry(page, &n->slabs_full, lru)
4215 handle_slab(x, cachep, page);
4216 list_for_each_entry(page, &n->slabs_partial, lru)
4217 handle_slab(x, cachep, page);
4218 spin_unlock_irq(&n->list_lock);
4220 } while (!is_store_user_clean(cachep));
4222 name = cachep->name;
4224 /* Increase the buffer size */
4225 mutex_unlock(&slab_mutex);
4226 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4228 /* Too bad, we are really out */
4230 mutex_lock(&slab_mutex);
4233 *(unsigned long *)m->private = x[0] * 2;
4235 mutex_lock(&slab_mutex);
4236 /* Now make sure this entry will be retried */
4240 for (i = 0; i < x[1]; i++) {
4241 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4242 show_symbol(m, x[2*i+2]);
4249 static const struct seq_operations slabstats_op = {
4250 .start = slab_start,
4256 static int slabstats_open(struct inode *inode, struct file *file)
4260 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4264 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4269 static const struct file_operations proc_slabstats_operations = {
4270 .open = slabstats_open,
4272 .llseek = seq_lseek,
4273 .release = seq_release_private,
4277 static int __init slab_proc_init(void)
4279 #ifdef CONFIG_DEBUG_SLAB_LEAK
4280 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4284 module_init(slab_proc_init);
4287 #ifdef CONFIG_HARDENED_USERCOPY
4289 * Rejects objects that are incorrectly sized.
4291 * Returns NULL if check passes, otherwise const char * to name of cache
4292 * to indicate an error.
4294 const char *__check_heap_object(const void *ptr, unsigned long n,
4297 struct kmem_cache *cachep;
4299 unsigned long offset;
4301 /* Find and validate object. */
4302 cachep = page->slab_cache;
4303 objnr = obj_to_index(cachep, page, (void *)ptr);
4304 BUG_ON(objnr >= cachep->num);
4306 /* Find offset within object. */
4307 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4309 /* Allow address range falling entirely within object size. */
4310 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4313 return cachep->name;
4315 #endif /* CONFIG_HARDENED_USERCOPY */
4318 * ksize - get the actual amount of memory allocated for a given object
4319 * @objp: Pointer to the object
4321 * kmalloc may internally round up allocations and return more memory
4322 * than requested. ksize() can be used to determine the actual amount of
4323 * memory allocated. The caller may use this additional memory, even though
4324 * a smaller amount of memory was initially specified with the kmalloc call.
4325 * The caller must guarantee that objp points to a valid object previously
4326 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4327 * must not be freed during the duration of the call.
4329 size_t ksize(const void *objp)
4334 if (unlikely(objp == ZERO_SIZE_PTR))
4337 size = virt_to_cache(objp)->object_size;
4338 /* We assume that ksize callers could use the whole allocated area,
4339 * so we need to unpoison this area.
4341 kasan_krealloc(objp, size, GFP_NOWAIT);
4345 EXPORT_SYMBOL(ksize);