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 * Need this for bootstrapping a per node allocator.
230 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
231 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
232 #define CACHE_CACHE 0
233 #define SIZE_NODE (MAX_NUMNODES)
235 static int drain_freelist(struct kmem_cache *cache,
236 struct kmem_cache_node *n, int tofree);
237 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
238 int node, struct list_head *list);
239 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
240 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
241 static void cache_reap(struct work_struct *unused);
243 static int slab_early_init = 1;
245 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
247 static void kmem_cache_node_init(struct kmem_cache_node *parent)
249 INIT_LIST_HEAD(&parent->slabs_full);
250 INIT_LIST_HEAD(&parent->slabs_partial);
251 INIT_LIST_HEAD(&parent->slabs_free);
252 parent->shared = NULL;
253 parent->alien = NULL;
254 parent->colour_next = 0;
255 spin_lock_init(&parent->list_lock);
256 parent->free_objects = 0;
257 parent->free_touched = 0;
260 #define MAKE_LIST(cachep, listp, slab, nodeid) \
262 INIT_LIST_HEAD(listp); \
263 list_splice(&get_node(cachep, nodeid)->slab, listp); \
266 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
268 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
269 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
270 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
273 #define CFLGS_OFF_SLAB (0x80000000UL)
274 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
275 #define OFF_SLAB_MIN_SIZE (max_t(size_t, PAGE_SIZE >> 5, KMALLOC_MIN_SIZE + 1))
277 #define BATCHREFILL_LIMIT 16
279 * Optimization question: fewer reaps means less probability for unnessary
280 * cpucache drain/refill cycles.
282 * OTOH the cpuarrays can contain lots of objects,
283 * which could lock up otherwise freeable slabs.
285 #define REAPTIMEOUT_AC (2*HZ)
286 #define REAPTIMEOUT_NODE (4*HZ)
289 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
290 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
291 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
292 #define STATS_INC_GROWN(x) ((x)->grown++)
293 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
294 #define STATS_SET_HIGH(x) \
296 if ((x)->num_active > (x)->high_mark) \
297 (x)->high_mark = (x)->num_active; \
299 #define STATS_INC_ERR(x) ((x)->errors++)
300 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
301 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
302 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
303 #define STATS_SET_FREEABLE(x, i) \
305 if ((x)->max_freeable < i) \
306 (x)->max_freeable = i; \
308 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
309 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
310 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
311 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
313 #define STATS_INC_ACTIVE(x) do { } while (0)
314 #define STATS_DEC_ACTIVE(x) do { } while (0)
315 #define STATS_INC_ALLOCED(x) do { } while (0)
316 #define STATS_INC_GROWN(x) do { } while (0)
317 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
318 #define STATS_SET_HIGH(x) do { } while (0)
319 #define STATS_INC_ERR(x) do { } while (0)
320 #define STATS_INC_NODEALLOCS(x) do { } while (0)
321 #define STATS_INC_NODEFREES(x) do { } while (0)
322 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
323 #define STATS_SET_FREEABLE(x, i) do { } while (0)
324 #define STATS_INC_ALLOCHIT(x) do { } while (0)
325 #define STATS_INC_ALLOCMISS(x) do { } while (0)
326 #define STATS_INC_FREEHIT(x) do { } while (0)
327 #define STATS_INC_FREEMISS(x) do { } while (0)
333 * memory layout of objects:
335 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
336 * the end of an object is aligned with the end of the real
337 * allocation. Catches writes behind the end of the allocation.
338 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
340 * cachep->obj_offset: The real object.
341 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
342 * cachep->size - 1* BYTES_PER_WORD: last caller address
343 * [BYTES_PER_WORD long]
345 static int obj_offset(struct kmem_cache *cachep)
347 return cachep->obj_offset;
350 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
352 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
353 return (unsigned long long*) (objp + obj_offset(cachep) -
354 sizeof(unsigned long long));
357 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
359 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
360 if (cachep->flags & SLAB_STORE_USER)
361 return (unsigned long long *)(objp + cachep->size -
362 sizeof(unsigned long long) -
364 return (unsigned long long *) (objp + cachep->size -
365 sizeof(unsigned long long));
368 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
370 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
371 return (void **)(objp + cachep->size - BYTES_PER_WORD);
376 #define obj_offset(x) 0
377 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
378 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
379 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
383 #define OBJECT_FREE (0)
384 #define OBJECT_ACTIVE (1)
386 #ifdef CONFIG_DEBUG_SLAB_LEAK
388 static void set_obj_status(struct page *page, int idx, int val)
392 struct kmem_cache *cachep = page->slab_cache;
394 freelist_size = cachep->num * sizeof(freelist_idx_t);
395 status = (char *)page->freelist + freelist_size;
399 static inline unsigned int get_obj_status(struct page *page, int idx)
403 struct kmem_cache *cachep = page->slab_cache;
405 freelist_size = cachep->num * sizeof(freelist_idx_t);
406 status = (char *)page->freelist + freelist_size;
412 static inline void set_obj_status(struct page *page, int idx, int val) {}
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 #define BOOT_CPUCACHE_ENTRIES 1
451 /* internal cache of cache description objs */
452 static struct kmem_cache kmem_cache_boot = {
454 .limit = BOOT_CPUCACHE_ENTRIES,
456 .size = sizeof(struct kmem_cache),
457 .name = "kmem_cache",
460 #define BAD_ALIEN_MAGIC 0x01020304ul
462 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
464 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
466 return this_cpu_ptr(cachep->cpu_cache);
469 static size_t calculate_freelist_size(int nr_objs, size_t align)
471 size_t freelist_size;
473 freelist_size = nr_objs * sizeof(freelist_idx_t);
474 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
475 freelist_size += nr_objs * sizeof(char);
478 freelist_size = ALIGN(freelist_size, align);
480 return freelist_size;
483 static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
484 size_t idx_size, size_t align)
487 size_t remained_size;
488 size_t freelist_size;
491 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
492 extra_space = sizeof(char);
494 * Ignore padding for the initial guess. The padding
495 * is at most @align-1 bytes, and @buffer_size is at
496 * least @align. In the worst case, this result will
497 * be one greater than the number of objects that fit
498 * into the memory allocation when taking the padding
501 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
504 * This calculated number will be either the right
505 * amount, or one greater than what we want.
507 remained_size = slab_size - nr_objs * buffer_size;
508 freelist_size = calculate_freelist_size(nr_objs, align);
509 if (remained_size < freelist_size)
516 * Calculate the number of objects and left-over bytes for a given buffer size.
518 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
519 size_t align, int flags, size_t *left_over,
524 size_t slab_size = PAGE_SIZE << gfporder;
527 * The slab management structure can be either off the slab or
528 * on it. For the latter case, the memory allocated for a
531 * - One freelist_idx_t for each object
532 * - Padding to respect alignment of @align
533 * - @buffer_size bytes for each object
535 * If the slab management structure is off the slab, then the
536 * alignment will already be calculated into the size. Because
537 * the slabs are all pages aligned, the objects will be at the
538 * correct alignment when allocated.
540 if (flags & CFLGS_OFF_SLAB) {
542 nr_objs = slab_size / buffer_size;
545 nr_objs = calculate_nr_objs(slab_size, buffer_size,
546 sizeof(freelist_idx_t), align);
547 mgmt_size = calculate_freelist_size(nr_objs, align);
550 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
554 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
556 static void __slab_error(const char *function, struct kmem_cache *cachep,
559 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
560 function, cachep->name, msg);
562 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
567 * By default on NUMA we use alien caches to stage the freeing of
568 * objects allocated from other nodes. This causes massive memory
569 * inefficiencies when using fake NUMA setup to split memory into a
570 * large number of small nodes, so it can be disabled on the command
574 static int use_alien_caches __read_mostly = 1;
575 static int __init noaliencache_setup(char *s)
577 use_alien_caches = 0;
580 __setup("noaliencache", noaliencache_setup);
582 static int __init slab_max_order_setup(char *str)
584 get_option(&str, &slab_max_order);
585 slab_max_order = slab_max_order < 0 ? 0 :
586 min(slab_max_order, MAX_ORDER - 1);
587 slab_max_order_set = true;
591 __setup("slab_max_order=", slab_max_order_setup);
595 * Special reaping functions for NUMA systems called from cache_reap().
596 * These take care of doing round robin flushing of alien caches (containing
597 * objects freed on different nodes from which they were allocated) and the
598 * flushing of remote pcps by calling drain_node_pages.
600 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
602 static void init_reap_node(int cpu)
606 node = next_node(cpu_to_mem(cpu), node_online_map);
607 if (node == MAX_NUMNODES)
608 node = first_node(node_online_map);
610 per_cpu(slab_reap_node, cpu) = node;
613 static void next_reap_node(void)
615 int node = __this_cpu_read(slab_reap_node);
617 node = next_node(node, node_online_map);
618 if (unlikely(node >= MAX_NUMNODES))
619 node = first_node(node_online_map);
620 __this_cpu_write(slab_reap_node, node);
624 #define init_reap_node(cpu) do { } while (0)
625 #define next_reap_node(void) do { } while (0)
629 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
630 * via the workqueue/eventd.
631 * Add the CPU number into the expiration time to minimize the possibility of
632 * the CPUs getting into lockstep and contending for the global cache chain
635 static void start_cpu_timer(int cpu)
637 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
640 * When this gets called from do_initcalls via cpucache_init(),
641 * init_workqueues() has already run, so keventd will be setup
644 if (keventd_up() && reap_work->work.func == NULL) {
646 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
647 schedule_delayed_work_on(cpu, reap_work,
648 __round_jiffies_relative(HZ, cpu));
652 static void init_arraycache(struct array_cache *ac, int limit, int batch)
655 * The array_cache structures contain pointers to free object.
656 * However, when such objects are allocated or transferred to another
657 * cache the pointers are not cleared and they could be counted as
658 * valid references during a kmemleak scan. Therefore, kmemleak must
659 * not scan such objects.
661 kmemleak_no_scan(ac);
665 ac->batchcount = batch;
670 static struct array_cache *alloc_arraycache(int node, int entries,
671 int batchcount, gfp_t gfp)
673 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
674 struct array_cache *ac = NULL;
676 ac = kmalloc_node(memsize, gfp, node);
677 init_arraycache(ac, entries, batchcount);
681 static inline bool is_slab_pfmemalloc(struct page *page)
683 return PageSlabPfmemalloc(page);
686 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
687 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
688 struct array_cache *ac)
690 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
694 if (!pfmemalloc_active)
697 spin_lock_irqsave(&n->list_lock, flags);
698 list_for_each_entry(page, &n->slabs_full, lru)
699 if (is_slab_pfmemalloc(page))
702 list_for_each_entry(page, &n->slabs_partial, lru)
703 if (is_slab_pfmemalloc(page))
706 list_for_each_entry(page, &n->slabs_free, lru)
707 if (is_slab_pfmemalloc(page))
710 pfmemalloc_active = false;
712 spin_unlock_irqrestore(&n->list_lock, flags);
715 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
716 gfp_t flags, bool force_refill)
719 void *objp = ac->entry[--ac->avail];
721 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
722 if (unlikely(is_obj_pfmemalloc(objp))) {
723 struct kmem_cache_node *n;
725 if (gfp_pfmemalloc_allowed(flags)) {
726 clear_obj_pfmemalloc(&objp);
730 /* The caller cannot use PFMEMALLOC objects, find another one */
731 for (i = 0; i < ac->avail; i++) {
732 /* If a !PFMEMALLOC object is found, swap them */
733 if (!is_obj_pfmemalloc(ac->entry[i])) {
735 ac->entry[i] = ac->entry[ac->avail];
736 ac->entry[ac->avail] = objp;
742 * If there are empty slabs on the slabs_free list and we are
743 * being forced to refill the cache, mark this one !pfmemalloc.
745 n = get_node(cachep, numa_mem_id());
746 if (!list_empty(&n->slabs_free) && force_refill) {
747 struct page *page = virt_to_head_page(objp);
748 ClearPageSlabPfmemalloc(page);
749 clear_obj_pfmemalloc(&objp);
750 recheck_pfmemalloc_active(cachep, ac);
754 /* No !PFMEMALLOC objects available */
762 static inline void *ac_get_obj(struct kmem_cache *cachep,
763 struct array_cache *ac, gfp_t flags, bool force_refill)
767 if (unlikely(sk_memalloc_socks()))
768 objp = __ac_get_obj(cachep, ac, flags, force_refill);
770 objp = ac->entry[--ac->avail];
775 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
776 struct array_cache *ac, void *objp)
778 if (unlikely(pfmemalloc_active)) {
779 /* Some pfmemalloc slabs exist, check if this is one */
780 struct page *page = virt_to_head_page(objp);
781 if (PageSlabPfmemalloc(page))
782 set_obj_pfmemalloc(&objp);
788 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
791 if (unlikely(sk_memalloc_socks()))
792 objp = __ac_put_obj(cachep, ac, objp);
794 ac->entry[ac->avail++] = objp;
798 * Transfer objects in one arraycache to another.
799 * Locking must be handled by the caller.
801 * Return the number of entries transferred.
803 static int transfer_objects(struct array_cache *to,
804 struct array_cache *from, unsigned int max)
806 /* Figure out how many entries to transfer */
807 int nr = min3(from->avail, max, to->limit - to->avail);
812 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
822 #define drain_alien_cache(cachep, alien) do { } while (0)
823 #define reap_alien(cachep, n) do { } while (0)
825 static inline struct alien_cache **alloc_alien_cache(int node,
826 int limit, gfp_t gfp)
828 return (struct alien_cache **)BAD_ALIEN_MAGIC;
831 static inline void free_alien_cache(struct alien_cache **ac_ptr)
835 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
840 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
846 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
847 gfp_t flags, int nodeid)
852 static inline gfp_t gfp_exact_node(gfp_t flags)
857 #else /* CONFIG_NUMA */
859 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
860 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
862 static struct alien_cache *__alloc_alien_cache(int node, int entries,
863 int batch, gfp_t gfp)
865 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
866 struct alien_cache *alc = NULL;
868 alc = kmalloc_node(memsize, gfp, node);
869 init_arraycache(&alc->ac, entries, batch);
870 spin_lock_init(&alc->lock);
874 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
876 struct alien_cache **alc_ptr;
877 size_t memsize = sizeof(void *) * nr_node_ids;
882 alc_ptr = kzalloc_node(memsize, gfp, node);
887 if (i == node || !node_online(i))
889 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
891 for (i--; i >= 0; i--)
900 static void free_alien_cache(struct alien_cache **alc_ptr)
911 static void __drain_alien_cache(struct kmem_cache *cachep,
912 struct array_cache *ac, int node,
913 struct list_head *list)
915 struct kmem_cache_node *n = get_node(cachep, node);
918 spin_lock(&n->list_lock);
920 * Stuff objects into the remote nodes shared array first.
921 * That way we could avoid the overhead of putting the objects
922 * into the free lists and getting them back later.
925 transfer_objects(n->shared, ac, ac->limit);
927 free_block(cachep, ac->entry, ac->avail, node, list);
929 spin_unlock(&n->list_lock);
934 * Called from cache_reap() to regularly drain alien caches round robin.
936 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
938 int node = __this_cpu_read(slab_reap_node);
941 struct alien_cache *alc = n->alien[node];
942 struct array_cache *ac;
946 if (ac->avail && spin_trylock_irq(&alc->lock)) {
949 __drain_alien_cache(cachep, ac, node, &list);
950 spin_unlock_irq(&alc->lock);
951 slabs_destroy(cachep, &list);
957 static void drain_alien_cache(struct kmem_cache *cachep,
958 struct alien_cache **alien)
961 struct alien_cache *alc;
962 struct array_cache *ac;
965 for_each_online_node(i) {
971 spin_lock_irqsave(&alc->lock, flags);
972 __drain_alien_cache(cachep, ac, i, &list);
973 spin_unlock_irqrestore(&alc->lock, flags);
974 slabs_destroy(cachep, &list);
979 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
980 int node, int page_node)
982 struct kmem_cache_node *n;
983 struct alien_cache *alien = NULL;
984 struct array_cache *ac;
987 n = get_node(cachep, node);
988 STATS_INC_NODEFREES(cachep);
989 if (n->alien && n->alien[page_node]) {
990 alien = n->alien[page_node];
992 spin_lock(&alien->lock);
993 if (unlikely(ac->avail == ac->limit)) {
994 STATS_INC_ACOVERFLOW(cachep);
995 __drain_alien_cache(cachep, ac, page_node, &list);
997 ac_put_obj(cachep, ac, objp);
998 spin_unlock(&alien->lock);
999 slabs_destroy(cachep, &list);
1001 n = get_node(cachep, page_node);
1002 spin_lock(&n->list_lock);
1003 free_block(cachep, &objp, 1, page_node, &list);
1004 spin_unlock(&n->list_lock);
1005 slabs_destroy(cachep, &list);
1010 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1012 int page_node = page_to_nid(virt_to_page(objp));
1013 int node = numa_mem_id();
1015 * Make sure we are not freeing a object from another node to the array
1016 * cache on this cpu.
1018 if (likely(node == page_node))
1021 return __cache_free_alien(cachep, objp, node, page_node);
1025 * Construct gfp mask to allocate from a specific node but do not direct reclaim
1026 * or warn about failures. kswapd may still wake to reclaim in the background.
1028 static inline gfp_t gfp_exact_node(gfp_t flags)
1030 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
1035 * Allocates and initializes node for a node on each slab cache, used for
1036 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1037 * will be allocated off-node since memory is not yet online for the new node.
1038 * When hotplugging memory or a cpu, existing node are not replaced if
1041 * Must hold slab_mutex.
1043 static int init_cache_node_node(int node)
1045 struct kmem_cache *cachep;
1046 struct kmem_cache_node *n;
1047 const size_t memsize = sizeof(struct kmem_cache_node);
1049 list_for_each_entry(cachep, &slab_caches, list) {
1051 * Set up the kmem_cache_node for cpu before we can
1052 * begin anything. Make sure some other cpu on this
1053 * node has not already allocated this
1055 n = get_node(cachep, node);
1057 n = kmalloc_node(memsize, GFP_KERNEL, node);
1060 kmem_cache_node_init(n);
1061 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1062 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1065 * The kmem_cache_nodes don't come and go as CPUs
1066 * come and go. slab_mutex is sufficient
1069 cachep->node[node] = n;
1072 spin_lock_irq(&n->list_lock);
1074 (1 + nr_cpus_node(node)) *
1075 cachep->batchcount + cachep->num;
1076 spin_unlock_irq(&n->list_lock);
1081 static inline int slabs_tofree(struct kmem_cache *cachep,
1082 struct kmem_cache_node *n)
1084 return (n->free_objects + cachep->num - 1) / cachep->num;
1087 static void cpuup_canceled(long cpu)
1089 struct kmem_cache *cachep;
1090 struct kmem_cache_node *n = NULL;
1091 int node = cpu_to_mem(cpu);
1092 const struct cpumask *mask = cpumask_of_node(node);
1094 list_for_each_entry(cachep, &slab_caches, list) {
1095 struct array_cache *nc;
1096 struct array_cache *shared;
1097 struct alien_cache **alien;
1100 n = get_node(cachep, node);
1104 spin_lock_irq(&n->list_lock);
1106 /* Free limit for this kmem_cache_node */
1107 n->free_limit -= cachep->batchcount;
1109 /* cpu is dead; no one can alloc from it. */
1110 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1112 free_block(cachep, nc->entry, nc->avail, node, &list);
1116 if (!cpumask_empty(mask)) {
1117 spin_unlock_irq(&n->list_lock);
1123 free_block(cachep, shared->entry,
1124 shared->avail, node, &list);
1131 spin_unlock_irq(&n->list_lock);
1135 drain_alien_cache(cachep, alien);
1136 free_alien_cache(alien);
1140 slabs_destroy(cachep, &list);
1143 * In the previous loop, all the objects were freed to
1144 * the respective cache's slabs, now we can go ahead and
1145 * shrink each nodelist to its limit.
1147 list_for_each_entry(cachep, &slab_caches, list) {
1148 n = get_node(cachep, node);
1151 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1155 static int cpuup_prepare(long cpu)
1157 struct kmem_cache *cachep;
1158 struct kmem_cache_node *n = NULL;
1159 int node = cpu_to_mem(cpu);
1163 * We need to do this right in the beginning since
1164 * alloc_arraycache's are going to use this list.
1165 * kmalloc_node allows us to add the slab to the right
1166 * kmem_cache_node and not this cpu's kmem_cache_node
1168 err = init_cache_node_node(node);
1173 * Now we can go ahead with allocating the shared arrays and
1176 list_for_each_entry(cachep, &slab_caches, list) {
1177 struct array_cache *shared = NULL;
1178 struct alien_cache **alien = NULL;
1180 if (cachep->shared) {
1181 shared = alloc_arraycache(node,
1182 cachep->shared * cachep->batchcount,
1183 0xbaadf00d, GFP_KERNEL);
1187 if (use_alien_caches) {
1188 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1194 n = get_node(cachep, node);
1197 spin_lock_irq(&n->list_lock);
1200 * We are serialised from CPU_DEAD or
1201 * CPU_UP_CANCELLED by the cpucontrol lock
1212 spin_unlock_irq(&n->list_lock);
1214 free_alien_cache(alien);
1219 cpuup_canceled(cpu);
1223 static int cpuup_callback(struct notifier_block *nfb,
1224 unsigned long action, void *hcpu)
1226 long cpu = (long)hcpu;
1230 case CPU_UP_PREPARE:
1231 case CPU_UP_PREPARE_FROZEN:
1232 mutex_lock(&slab_mutex);
1233 err = cpuup_prepare(cpu);
1234 mutex_unlock(&slab_mutex);
1237 case CPU_ONLINE_FROZEN:
1238 start_cpu_timer(cpu);
1240 #ifdef CONFIG_HOTPLUG_CPU
1241 case CPU_DOWN_PREPARE:
1242 case CPU_DOWN_PREPARE_FROZEN:
1244 * Shutdown cache reaper. Note that the slab_mutex is
1245 * held so that if cache_reap() is invoked it cannot do
1246 * anything expensive but will only modify reap_work
1247 * and reschedule the timer.
1249 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1250 /* Now the cache_reaper is guaranteed to be not running. */
1251 per_cpu(slab_reap_work, cpu).work.func = NULL;
1253 case CPU_DOWN_FAILED:
1254 case CPU_DOWN_FAILED_FROZEN:
1255 start_cpu_timer(cpu);
1258 case CPU_DEAD_FROZEN:
1260 * Even if all the cpus of a node are down, we don't free the
1261 * kmem_cache_node of any cache. This to avoid a race between
1262 * cpu_down, and a kmalloc allocation from another cpu for
1263 * memory from the node of the cpu going down. The node
1264 * structure is usually allocated from kmem_cache_create() and
1265 * gets destroyed at kmem_cache_destroy().
1269 case CPU_UP_CANCELED:
1270 case CPU_UP_CANCELED_FROZEN:
1271 mutex_lock(&slab_mutex);
1272 cpuup_canceled(cpu);
1273 mutex_unlock(&slab_mutex);
1276 return notifier_from_errno(err);
1279 static struct notifier_block cpucache_notifier = {
1280 &cpuup_callback, NULL, 0
1283 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1285 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1286 * Returns -EBUSY if all objects cannot be drained so that the node is not
1289 * Must hold slab_mutex.
1291 static int __meminit drain_cache_node_node(int node)
1293 struct kmem_cache *cachep;
1296 list_for_each_entry(cachep, &slab_caches, list) {
1297 struct kmem_cache_node *n;
1299 n = get_node(cachep, node);
1303 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1305 if (!list_empty(&n->slabs_full) ||
1306 !list_empty(&n->slabs_partial)) {
1314 static int __meminit slab_memory_callback(struct notifier_block *self,
1315 unsigned long action, void *arg)
1317 struct memory_notify *mnb = arg;
1321 nid = mnb->status_change_nid;
1326 case MEM_GOING_ONLINE:
1327 mutex_lock(&slab_mutex);
1328 ret = init_cache_node_node(nid);
1329 mutex_unlock(&slab_mutex);
1331 case MEM_GOING_OFFLINE:
1332 mutex_lock(&slab_mutex);
1333 ret = drain_cache_node_node(nid);
1334 mutex_unlock(&slab_mutex);
1338 case MEM_CANCEL_ONLINE:
1339 case MEM_CANCEL_OFFLINE:
1343 return notifier_from_errno(ret);
1345 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1348 * swap the static kmem_cache_node with kmalloced memory
1350 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1353 struct kmem_cache_node *ptr;
1355 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1358 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1360 * Do not assume that spinlocks can be initialized via memcpy:
1362 spin_lock_init(&ptr->list_lock);
1364 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1365 cachep->node[nodeid] = ptr;
1369 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1370 * size of kmem_cache_node.
1372 static void __init set_up_node(struct kmem_cache *cachep, int index)
1376 for_each_online_node(node) {
1377 cachep->node[node] = &init_kmem_cache_node[index + node];
1378 cachep->node[node]->next_reap = jiffies +
1380 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1385 * Initialisation. Called after the page allocator have been initialised and
1386 * before smp_init().
1388 void __init kmem_cache_init(void)
1392 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1393 sizeof(struct rcu_head));
1394 kmem_cache = &kmem_cache_boot;
1396 if (num_possible_nodes() == 1)
1397 use_alien_caches = 0;
1399 for (i = 0; i < NUM_INIT_LISTS; i++)
1400 kmem_cache_node_init(&init_kmem_cache_node[i]);
1403 * Fragmentation resistance on low memory - only use bigger
1404 * page orders on machines with more than 32MB of memory if
1405 * not overridden on the command line.
1407 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1408 slab_max_order = SLAB_MAX_ORDER_HI;
1410 /* Bootstrap is tricky, because several objects are allocated
1411 * from caches that do not exist yet:
1412 * 1) initialize the kmem_cache cache: it contains the struct
1413 * kmem_cache structures of all caches, except kmem_cache itself:
1414 * kmem_cache is statically allocated.
1415 * Initially an __init data area is used for the head array and the
1416 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1417 * array at the end of the bootstrap.
1418 * 2) Create the first kmalloc cache.
1419 * The struct kmem_cache for the new cache is allocated normally.
1420 * An __init data area is used for the head array.
1421 * 3) Create the remaining kmalloc caches, with minimally sized
1423 * 4) Replace the __init data head arrays for kmem_cache and the first
1424 * kmalloc cache with kmalloc allocated arrays.
1425 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1426 * the other cache's with kmalloc allocated memory.
1427 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1430 /* 1) create the kmem_cache */
1433 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1435 create_boot_cache(kmem_cache, "kmem_cache",
1436 offsetof(struct kmem_cache, node) +
1437 nr_node_ids * sizeof(struct kmem_cache_node *),
1438 SLAB_HWCACHE_ALIGN);
1439 list_add(&kmem_cache->list, &slab_caches);
1440 slab_state = PARTIAL;
1443 * Initialize the caches that provide memory for the kmem_cache_node
1444 * structures first. Without this, further allocations will bug.
1446 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1447 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1448 slab_state = PARTIAL_NODE;
1449 setup_kmalloc_cache_index_table();
1451 slab_early_init = 0;
1453 /* 5) Replace the bootstrap kmem_cache_node */
1457 for_each_online_node(nid) {
1458 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1460 init_list(kmalloc_caches[INDEX_NODE],
1461 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1465 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1468 void __init kmem_cache_init_late(void)
1470 struct kmem_cache *cachep;
1474 /* 6) resize the head arrays to their final sizes */
1475 mutex_lock(&slab_mutex);
1476 list_for_each_entry(cachep, &slab_caches, list)
1477 if (enable_cpucache(cachep, GFP_NOWAIT))
1479 mutex_unlock(&slab_mutex);
1485 * Register a cpu startup notifier callback that initializes
1486 * cpu_cache_get for all new cpus
1488 register_cpu_notifier(&cpucache_notifier);
1492 * Register a memory hotplug callback that initializes and frees
1495 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1499 * The reap timers are started later, with a module init call: That part
1500 * of the kernel is not yet operational.
1504 static int __init cpucache_init(void)
1509 * Register the timers that return unneeded pages to the page allocator
1511 for_each_online_cpu(cpu)
1512 start_cpu_timer(cpu);
1518 __initcall(cpucache_init);
1520 static noinline void
1521 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1524 struct kmem_cache_node *n;
1526 unsigned long flags;
1528 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1529 DEFAULT_RATELIMIT_BURST);
1531 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1535 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1537 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1538 cachep->name, cachep->size, cachep->gfporder);
1540 for_each_kmem_cache_node(cachep, node, n) {
1541 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1542 unsigned long active_slabs = 0, num_slabs = 0;
1544 spin_lock_irqsave(&n->list_lock, flags);
1545 list_for_each_entry(page, &n->slabs_full, lru) {
1546 active_objs += cachep->num;
1549 list_for_each_entry(page, &n->slabs_partial, lru) {
1550 active_objs += page->active;
1553 list_for_each_entry(page, &n->slabs_free, lru)
1556 free_objects += n->free_objects;
1557 spin_unlock_irqrestore(&n->list_lock, flags);
1559 num_slabs += active_slabs;
1560 num_objs = num_slabs * cachep->num;
1562 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1563 node, active_slabs, num_slabs, active_objs, num_objs,
1570 * Interface to system's page allocator. No need to hold the
1571 * kmem_cache_node ->list_lock.
1573 * If we requested dmaable memory, we will get it. Even if we
1574 * did not request dmaable memory, we might get it, but that
1575 * would be relatively rare and ignorable.
1577 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1583 flags |= cachep->allocflags;
1584 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1585 flags |= __GFP_RECLAIMABLE;
1587 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1589 slab_out_of_memory(cachep, flags, nodeid);
1593 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1594 __free_pages(page, cachep->gfporder);
1598 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1599 if (page_is_pfmemalloc(page))
1600 pfmemalloc_active = true;
1602 nr_pages = (1 << cachep->gfporder);
1603 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1604 add_zone_page_state(page_zone(page),
1605 NR_SLAB_RECLAIMABLE, nr_pages);
1607 add_zone_page_state(page_zone(page),
1608 NR_SLAB_UNRECLAIMABLE, nr_pages);
1609 __SetPageSlab(page);
1610 if (page_is_pfmemalloc(page))
1611 SetPageSlabPfmemalloc(page);
1613 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1614 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1617 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1619 kmemcheck_mark_unallocated_pages(page, nr_pages);
1626 * Interface to system's page release.
1628 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1630 const unsigned long nr_freed = (1 << cachep->gfporder);
1632 kmemcheck_free_shadow(page, cachep->gfporder);
1634 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1635 sub_zone_page_state(page_zone(page),
1636 NR_SLAB_RECLAIMABLE, nr_freed);
1638 sub_zone_page_state(page_zone(page),
1639 NR_SLAB_UNRECLAIMABLE, nr_freed);
1641 BUG_ON(!PageSlab(page));
1642 __ClearPageSlabPfmemalloc(page);
1643 __ClearPageSlab(page);
1644 page_mapcount_reset(page);
1645 page->mapping = NULL;
1647 if (current->reclaim_state)
1648 current->reclaim_state->reclaimed_slab += nr_freed;
1649 __free_kmem_pages(page, cachep->gfporder);
1652 static void kmem_rcu_free(struct rcu_head *head)
1654 struct kmem_cache *cachep;
1657 page = container_of(head, struct page, rcu_head);
1658 cachep = page->slab_cache;
1660 kmem_freepages(cachep, page);
1664 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1666 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1667 (cachep->size % PAGE_SIZE) == 0)
1673 #ifdef CONFIG_DEBUG_PAGEALLOC
1674 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1675 unsigned long caller)
1677 int size = cachep->object_size;
1679 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1681 if (size < 5 * sizeof(unsigned long))
1684 *addr++ = 0x12345678;
1686 *addr++ = smp_processor_id();
1687 size -= 3 * sizeof(unsigned long);
1689 unsigned long *sptr = &caller;
1690 unsigned long svalue;
1692 while (!kstack_end(sptr)) {
1694 if (kernel_text_address(svalue)) {
1696 size -= sizeof(unsigned long);
1697 if (size <= sizeof(unsigned long))
1703 *addr++ = 0x87654321;
1706 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1707 int map, unsigned long caller)
1709 if (!is_debug_pagealloc_cache(cachep))
1713 store_stackinfo(cachep, objp, caller);
1715 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1719 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1720 int map, unsigned long caller) {}
1724 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1726 int size = cachep->object_size;
1727 addr = &((char *)addr)[obj_offset(cachep)];
1729 memset(addr, val, size);
1730 *(unsigned char *)(addr + size - 1) = POISON_END;
1733 static void dump_line(char *data, int offset, int limit)
1736 unsigned char error = 0;
1739 printk(KERN_ERR "%03x: ", offset);
1740 for (i = 0; i < limit; i++) {
1741 if (data[offset + i] != POISON_FREE) {
1742 error = data[offset + i];
1746 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1747 &data[offset], limit, 1);
1749 if (bad_count == 1) {
1750 error ^= POISON_FREE;
1751 if (!(error & (error - 1))) {
1752 printk(KERN_ERR "Single bit error detected. Probably "
1755 printk(KERN_ERR "Run memtest86+ or a similar memory "
1758 printk(KERN_ERR "Run a memory test tool.\n");
1767 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1772 if (cachep->flags & SLAB_RED_ZONE) {
1773 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1774 *dbg_redzone1(cachep, objp),
1775 *dbg_redzone2(cachep, objp));
1778 if (cachep->flags & SLAB_STORE_USER) {
1779 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1780 *dbg_userword(cachep, objp),
1781 *dbg_userword(cachep, objp));
1783 realobj = (char *)objp + obj_offset(cachep);
1784 size = cachep->object_size;
1785 for (i = 0; i < size && lines; i += 16, lines--) {
1788 if (i + limit > size)
1790 dump_line(realobj, i, limit);
1794 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1800 if (is_debug_pagealloc_cache(cachep))
1803 realobj = (char *)objp + obj_offset(cachep);
1804 size = cachep->object_size;
1806 for (i = 0; i < size; i++) {
1807 char exp = POISON_FREE;
1810 if (realobj[i] != exp) {
1816 "Slab corruption (%s): %s start=%p, len=%d\n",
1817 print_tainted(), cachep->name, realobj, size);
1818 print_objinfo(cachep, objp, 0);
1820 /* Hexdump the affected line */
1823 if (i + limit > size)
1825 dump_line(realobj, i, limit);
1828 /* Limit to 5 lines */
1834 /* Print some data about the neighboring objects, if they
1837 struct page *page = virt_to_head_page(objp);
1840 objnr = obj_to_index(cachep, page, objp);
1842 objp = index_to_obj(cachep, page, objnr - 1);
1843 realobj = (char *)objp + obj_offset(cachep);
1844 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1846 print_objinfo(cachep, objp, 2);
1848 if (objnr + 1 < cachep->num) {
1849 objp = index_to_obj(cachep, page, objnr + 1);
1850 realobj = (char *)objp + obj_offset(cachep);
1851 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1853 print_objinfo(cachep, objp, 2);
1860 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1864 for (i = 0; i < cachep->num; i++) {
1865 void *objp = index_to_obj(cachep, page, i);
1867 if (cachep->flags & SLAB_POISON) {
1868 check_poison_obj(cachep, objp);
1869 slab_kernel_map(cachep, objp, 1, 0);
1871 if (cachep->flags & SLAB_RED_ZONE) {
1872 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1873 slab_error(cachep, "start of a freed object "
1875 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1876 slab_error(cachep, "end of a freed object "
1882 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1889 * slab_destroy - destroy and release all objects in a slab
1890 * @cachep: cache pointer being destroyed
1891 * @page: page pointer being destroyed
1893 * Destroy all the objs in a slab page, and release the mem back to the system.
1894 * Before calling the slab page must have been unlinked from the cache. The
1895 * kmem_cache_node ->list_lock is not held/needed.
1897 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1901 freelist = page->freelist;
1902 slab_destroy_debugcheck(cachep, page);
1903 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1904 call_rcu(&page->rcu_head, kmem_rcu_free);
1906 kmem_freepages(cachep, page);
1909 * From now on, we don't use freelist
1910 * although actual page can be freed in rcu context
1912 if (OFF_SLAB(cachep))
1913 kmem_cache_free(cachep->freelist_cache, freelist);
1916 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1918 struct page *page, *n;
1920 list_for_each_entry_safe(page, n, list, lru) {
1921 list_del(&page->lru);
1922 slab_destroy(cachep, page);
1927 * calculate_slab_order - calculate size (page order) of slabs
1928 * @cachep: pointer to the cache that is being created
1929 * @size: size of objects to be created in this cache.
1930 * @align: required alignment for the objects.
1931 * @flags: slab allocation flags
1933 * Also calculates the number of objects per slab.
1935 * This could be made much more intelligent. For now, try to avoid using
1936 * high order pages for slabs. When the gfp() functions are more friendly
1937 * towards high-order requests, this should be changed.
1939 static size_t calculate_slab_order(struct kmem_cache *cachep,
1940 size_t size, size_t align, unsigned long flags)
1942 unsigned long offslab_limit;
1943 size_t left_over = 0;
1946 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1950 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1954 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1955 if (num > SLAB_OBJ_MAX_NUM)
1958 if (flags & CFLGS_OFF_SLAB) {
1959 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
1961 * Max number of objs-per-slab for caches which
1962 * use off-slab slabs. Needed to avoid a possible
1963 * looping condition in cache_grow().
1965 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1966 freelist_size_per_obj += sizeof(char);
1967 offslab_limit = size;
1968 offslab_limit /= freelist_size_per_obj;
1970 if (num > offslab_limit)
1974 /* Found something acceptable - save it away */
1976 cachep->gfporder = gfporder;
1977 left_over = remainder;
1980 * A VFS-reclaimable slab tends to have most allocations
1981 * as GFP_NOFS and we really don't want to have to be allocating
1982 * higher-order pages when we are unable to shrink dcache.
1984 if (flags & SLAB_RECLAIM_ACCOUNT)
1988 * Large number of objects is good, but very large slabs are
1989 * currently bad for the gfp()s.
1991 if (gfporder >= slab_max_order)
1995 * Acceptable internal fragmentation?
1997 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2003 static struct array_cache __percpu *alloc_kmem_cache_cpus(
2004 struct kmem_cache *cachep, int entries, int batchcount)
2008 struct array_cache __percpu *cpu_cache;
2010 size = sizeof(void *) * entries + sizeof(struct array_cache);
2011 cpu_cache = __alloc_percpu(size, sizeof(void *));
2016 for_each_possible_cpu(cpu) {
2017 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2018 entries, batchcount);
2024 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2026 if (slab_state >= FULL)
2027 return enable_cpucache(cachep, gfp);
2029 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2030 if (!cachep->cpu_cache)
2033 if (slab_state == DOWN) {
2034 /* Creation of first cache (kmem_cache). */
2035 set_up_node(kmem_cache, CACHE_CACHE);
2036 } else if (slab_state == PARTIAL) {
2037 /* For kmem_cache_node */
2038 set_up_node(cachep, SIZE_NODE);
2042 for_each_online_node(node) {
2043 cachep->node[node] = kmalloc_node(
2044 sizeof(struct kmem_cache_node), gfp, node);
2045 BUG_ON(!cachep->node[node]);
2046 kmem_cache_node_init(cachep->node[node]);
2050 cachep->node[numa_mem_id()]->next_reap =
2051 jiffies + REAPTIMEOUT_NODE +
2052 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2054 cpu_cache_get(cachep)->avail = 0;
2055 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2056 cpu_cache_get(cachep)->batchcount = 1;
2057 cpu_cache_get(cachep)->touched = 0;
2058 cachep->batchcount = 1;
2059 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2063 unsigned long kmem_cache_flags(unsigned long object_size,
2064 unsigned long flags, const char *name,
2065 void (*ctor)(void *))
2071 __kmem_cache_alias(const char *name, size_t size, size_t align,
2072 unsigned long flags, void (*ctor)(void *))
2074 struct kmem_cache *cachep;
2076 cachep = find_mergeable(size, align, flags, name, ctor);
2081 * Adjust the object sizes so that we clear
2082 * the complete object on kzalloc.
2084 cachep->object_size = max_t(int, cachep->object_size, size);
2090 * __kmem_cache_create - Create a cache.
2091 * @cachep: cache management descriptor
2092 * @flags: SLAB flags
2094 * Returns a ptr to the cache on success, NULL on failure.
2095 * Cannot be called within a int, but can be interrupted.
2096 * The @ctor is run when new pages are allocated by the cache.
2100 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2101 * to catch references to uninitialised memory.
2103 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2104 * for buffer overruns.
2106 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2107 * cacheline. This can be beneficial if you're counting cycles as closely
2111 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2113 size_t left_over, freelist_size;
2114 size_t ralign = BYTES_PER_WORD;
2117 size_t size = cachep->size;
2122 * Enable redzoning and last user accounting, except for caches with
2123 * large objects, if the increased size would increase the object size
2124 * above the next power of two: caches with object sizes just above a
2125 * power of two have a significant amount of internal fragmentation.
2127 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2128 2 * sizeof(unsigned long long)))
2129 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2130 if (!(flags & SLAB_DESTROY_BY_RCU))
2131 flags |= SLAB_POISON;
2136 * Check that size is in terms of words. This is needed to avoid
2137 * unaligned accesses for some archs when redzoning is used, and makes
2138 * sure any on-slab bufctl's are also correctly aligned.
2140 if (size & (BYTES_PER_WORD - 1)) {
2141 size += (BYTES_PER_WORD - 1);
2142 size &= ~(BYTES_PER_WORD - 1);
2145 if (flags & SLAB_RED_ZONE) {
2146 ralign = REDZONE_ALIGN;
2147 /* If redzoning, ensure that the second redzone is suitably
2148 * aligned, by adjusting the object size accordingly. */
2149 size += REDZONE_ALIGN - 1;
2150 size &= ~(REDZONE_ALIGN - 1);
2153 /* 3) caller mandated alignment */
2154 if (ralign < cachep->align) {
2155 ralign = cachep->align;
2157 /* disable debug if necessary */
2158 if (ralign > __alignof__(unsigned long long))
2159 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2163 cachep->align = ralign;
2165 if (slab_is_available())
2173 * Both debugging options require word-alignment which is calculated
2176 if (flags & SLAB_RED_ZONE) {
2177 /* add space for red zone words */
2178 cachep->obj_offset += sizeof(unsigned long long);
2179 size += 2 * sizeof(unsigned long long);
2181 if (flags & SLAB_STORE_USER) {
2182 /* user store requires one word storage behind the end of
2183 * the real object. But if the second red zone needs to be
2184 * aligned to 64 bits, we must allow that much space.
2186 if (flags & SLAB_RED_ZONE)
2187 size += REDZONE_ALIGN;
2189 size += BYTES_PER_WORD;
2192 * To activate debug pagealloc, off-slab management is necessary
2193 * requirement. In early phase of initialization, small sized slab
2194 * doesn't get initialized so it would not be possible. So, we need
2195 * to check size >= 256. It guarantees that all necessary small
2196 * sized slab is initialized in current slab initialization sequence.
2198 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2199 !slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2200 size >= 256 && cachep->object_size > cache_line_size() &&
2201 ALIGN(size, cachep->align) < PAGE_SIZE) {
2202 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2208 * Determine if the slab management is 'on' or 'off' slab.
2209 * (bootstrapping cannot cope with offslab caches so don't do
2210 * it too early on. Always use on-slab management when
2211 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2213 if (size >= OFF_SLAB_MIN_SIZE && !slab_early_init &&
2214 !(flags & SLAB_NOLEAKTRACE))
2216 * Size is large, assume best to place the slab management obj
2217 * off-slab (should allow better packing of objs).
2219 flags |= CFLGS_OFF_SLAB;
2221 size = ALIGN(size, cachep->align);
2223 * We should restrict the number of objects in a slab to implement
2224 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2226 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2227 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2229 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2234 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2237 * If the slab has been placed off-slab, and we have enough space then
2238 * move it on-slab. This is at the expense of any extra colouring.
2240 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2241 flags &= ~CFLGS_OFF_SLAB;
2242 left_over -= freelist_size;
2245 if (flags & CFLGS_OFF_SLAB) {
2246 /* really off slab. No need for manual alignment */
2247 freelist_size = calculate_freelist_size(cachep->num, 0);
2250 cachep->colour_off = cache_line_size();
2251 /* Offset must be a multiple of the alignment. */
2252 if (cachep->colour_off < cachep->align)
2253 cachep->colour_off = cachep->align;
2254 cachep->colour = left_over / cachep->colour_off;
2255 cachep->freelist_size = freelist_size;
2256 cachep->flags = flags;
2257 cachep->allocflags = __GFP_COMP;
2258 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2259 cachep->allocflags |= GFP_DMA;
2260 cachep->size = size;
2261 cachep->reciprocal_buffer_size = reciprocal_value(size);
2265 * If we're going to use the generic kernel_map_pages()
2266 * poisoning, then it's going to smash the contents of
2267 * the redzone and userword anyhow, so switch them off.
2269 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2270 (cachep->flags & SLAB_POISON) &&
2271 is_debug_pagealloc_cache(cachep))
2272 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2275 if (OFF_SLAB(cachep)) {
2276 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2278 * This is a possibility for one of the kmalloc_{dma,}_caches.
2279 * But since we go off slab only for object size greater than
2280 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
2281 * in ascending order,this should not happen at all.
2282 * But leave a BUG_ON for some lucky dude.
2284 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2287 err = setup_cpu_cache(cachep, gfp);
2289 __kmem_cache_release(cachep);
2297 static void check_irq_off(void)
2299 BUG_ON(!irqs_disabled());
2302 static void check_irq_on(void)
2304 BUG_ON(irqs_disabled());
2307 static void check_spinlock_acquired(struct kmem_cache *cachep)
2311 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2315 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2319 assert_spin_locked(&get_node(cachep, node)->list_lock);
2324 #define check_irq_off() do { } while(0)
2325 #define check_irq_on() do { } while(0)
2326 #define check_spinlock_acquired(x) do { } while(0)
2327 #define check_spinlock_acquired_node(x, y) do { } while(0)
2330 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2331 struct array_cache *ac,
2332 int force, int node);
2334 static void do_drain(void *arg)
2336 struct kmem_cache *cachep = arg;
2337 struct array_cache *ac;
2338 int node = numa_mem_id();
2339 struct kmem_cache_node *n;
2343 ac = cpu_cache_get(cachep);
2344 n = get_node(cachep, node);
2345 spin_lock(&n->list_lock);
2346 free_block(cachep, ac->entry, ac->avail, node, &list);
2347 spin_unlock(&n->list_lock);
2348 slabs_destroy(cachep, &list);
2352 static void drain_cpu_caches(struct kmem_cache *cachep)
2354 struct kmem_cache_node *n;
2357 on_each_cpu(do_drain, cachep, 1);
2359 for_each_kmem_cache_node(cachep, node, n)
2361 drain_alien_cache(cachep, n->alien);
2363 for_each_kmem_cache_node(cachep, node, n)
2364 drain_array(cachep, n, n->shared, 1, node);
2368 * Remove slabs from the list of free slabs.
2369 * Specify the number of slabs to drain in tofree.
2371 * Returns the actual number of slabs released.
2373 static int drain_freelist(struct kmem_cache *cache,
2374 struct kmem_cache_node *n, int tofree)
2376 struct list_head *p;
2381 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2383 spin_lock_irq(&n->list_lock);
2384 p = n->slabs_free.prev;
2385 if (p == &n->slabs_free) {
2386 spin_unlock_irq(&n->list_lock);
2390 page = list_entry(p, struct page, lru);
2391 list_del(&page->lru);
2393 * Safe to drop the lock. The slab is no longer linked
2396 n->free_objects -= cache->num;
2397 spin_unlock_irq(&n->list_lock);
2398 slab_destroy(cache, page);
2405 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2409 struct kmem_cache_node *n;
2411 drain_cpu_caches(cachep);
2414 for_each_kmem_cache_node(cachep, node, n) {
2415 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2417 ret += !list_empty(&n->slabs_full) ||
2418 !list_empty(&n->slabs_partial);
2420 return (ret ? 1 : 0);
2423 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2425 return __kmem_cache_shrink(cachep, false);
2428 void __kmem_cache_release(struct kmem_cache *cachep)
2431 struct kmem_cache_node *n;
2433 free_percpu(cachep->cpu_cache);
2435 /* NUMA: free the node structures */
2436 for_each_kmem_cache_node(cachep, i, n) {
2438 free_alien_cache(n->alien);
2440 cachep->node[i] = NULL;
2445 * Get the memory for a slab management obj.
2447 * For a slab cache when the slab descriptor is off-slab, the
2448 * slab descriptor can't come from the same cache which is being created,
2449 * Because if it is the case, that means we defer the creation of
2450 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2451 * And we eventually call down to __kmem_cache_create(), which
2452 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2453 * This is a "chicken-and-egg" problem.
2455 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2456 * which are all initialized during kmem_cache_init().
2458 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2459 struct page *page, int colour_off,
2460 gfp_t local_flags, int nodeid)
2463 void *addr = page_address(page);
2465 if (OFF_SLAB(cachep)) {
2466 /* Slab management obj is off-slab. */
2467 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2468 local_flags, nodeid);
2472 freelist = addr + colour_off;
2473 colour_off += cachep->freelist_size;
2476 page->s_mem = addr + colour_off;
2480 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2482 return ((freelist_idx_t *)page->freelist)[idx];
2485 static inline void set_free_obj(struct page *page,
2486 unsigned int idx, freelist_idx_t val)
2488 ((freelist_idx_t *)(page->freelist))[idx] = val;
2491 static void cache_init_objs(struct kmem_cache *cachep,
2496 for (i = 0; i < cachep->num; i++) {
2497 void *objp = index_to_obj(cachep, page, i);
2499 if (cachep->flags & SLAB_STORE_USER)
2500 *dbg_userword(cachep, objp) = NULL;
2502 if (cachep->flags & SLAB_RED_ZONE) {
2503 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2504 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2507 * Constructors are not allowed to allocate memory from the same
2508 * cache which they are a constructor for. Otherwise, deadlock.
2509 * They must also be threaded.
2511 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2512 cachep->ctor(objp + obj_offset(cachep));
2514 if (cachep->flags & SLAB_RED_ZONE) {
2515 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2516 slab_error(cachep, "constructor overwrote the"
2517 " end of an object");
2518 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2519 slab_error(cachep, "constructor overwrote the"
2520 " 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 set_obj_status(page, i, OBJECT_FREE);
2532 set_free_obj(page, i, i);
2536 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2538 if (CONFIG_ZONE_DMA_FLAG) {
2539 if (flags & GFP_DMA)
2540 BUG_ON(!(cachep->allocflags & GFP_DMA));
2542 BUG_ON(cachep->allocflags & GFP_DMA);
2546 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2550 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2556 static void slab_put_obj(struct kmem_cache *cachep,
2557 struct page *page, void *objp)
2559 unsigned int objnr = obj_to_index(cachep, page, objp);
2563 /* Verify double free bug */
2564 for (i = page->active; i < cachep->num; i++) {
2565 if (get_free_obj(page, i) == objnr) {
2566 printk(KERN_ERR "slab: double free detected in cache "
2567 "'%s', objp %p\n", cachep->name, objp);
2573 set_free_obj(page, page->active, objnr);
2577 * Map pages beginning at addr to the given cache and slab. This is required
2578 * for the slab allocator to be able to lookup the cache and slab of a
2579 * virtual address for kfree, ksize, and slab debugging.
2581 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2584 page->slab_cache = cache;
2585 page->freelist = freelist;
2589 * Grow (by 1) the number of slabs within a cache. This is called by
2590 * kmem_cache_alloc() when there are no active objs left in a cache.
2592 static int cache_grow(struct kmem_cache *cachep,
2593 gfp_t flags, int nodeid, struct page *page)
2598 struct kmem_cache_node *n;
2601 * Be lazy and only check for valid flags here, keeping it out of the
2602 * critical path in kmem_cache_alloc().
2604 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2605 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2608 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2610 /* Take the node list lock to change the colour_next on this node */
2612 n = get_node(cachep, nodeid);
2613 spin_lock(&n->list_lock);
2615 /* Get colour for the slab, and cal the next value. */
2616 offset = n->colour_next;
2618 if (n->colour_next >= cachep->colour)
2620 spin_unlock(&n->list_lock);
2622 offset *= cachep->colour_off;
2624 if (gfpflags_allow_blocking(local_flags))
2628 * The test for missing atomic flag is performed here, rather than
2629 * the more obvious place, simply to reduce the critical path length
2630 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2631 * will eventually be caught here (where it matters).
2633 kmem_flagcheck(cachep, flags);
2636 * Get mem for the objs. Attempt to allocate a physical page from
2640 page = kmem_getpages(cachep, local_flags, nodeid);
2644 /* Get slab management. */
2645 freelist = alloc_slabmgmt(cachep, page, offset,
2646 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2650 slab_map_pages(cachep, page, freelist);
2652 cache_init_objs(cachep, page);
2654 if (gfpflags_allow_blocking(local_flags))
2655 local_irq_disable();
2657 spin_lock(&n->list_lock);
2659 /* Make slab active. */
2660 list_add_tail(&page->lru, &(n->slabs_free));
2661 STATS_INC_GROWN(cachep);
2662 n->free_objects += cachep->num;
2663 spin_unlock(&n->list_lock);
2666 kmem_freepages(cachep, page);
2668 if (gfpflags_allow_blocking(local_flags))
2669 local_irq_disable();
2676 * Perform extra freeing checks:
2677 * - detect bad pointers.
2678 * - POISON/RED_ZONE checking
2680 static void kfree_debugcheck(const void *objp)
2682 if (!virt_addr_valid(objp)) {
2683 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2684 (unsigned long)objp);
2689 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2691 unsigned long long redzone1, redzone2;
2693 redzone1 = *dbg_redzone1(cache, obj);
2694 redzone2 = *dbg_redzone2(cache, obj);
2699 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2702 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2703 slab_error(cache, "double free detected");
2705 slab_error(cache, "memory outside object was overwritten");
2707 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2708 obj, redzone1, redzone2);
2711 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2712 unsigned long caller)
2717 BUG_ON(virt_to_cache(objp) != cachep);
2719 objp -= obj_offset(cachep);
2720 kfree_debugcheck(objp);
2721 page = virt_to_head_page(objp);
2723 if (cachep->flags & SLAB_RED_ZONE) {
2724 verify_redzone_free(cachep, objp);
2725 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2726 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2728 if (cachep->flags & SLAB_STORE_USER)
2729 *dbg_userword(cachep, objp) = (void *)caller;
2731 objnr = obj_to_index(cachep, page, objp);
2733 BUG_ON(objnr >= cachep->num);
2734 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2736 set_obj_status(page, objnr, OBJECT_FREE);
2737 if (cachep->flags & SLAB_POISON) {
2738 poison_obj(cachep, objp, POISON_FREE);
2739 slab_kernel_map(cachep, objp, 0, caller);
2745 #define kfree_debugcheck(x) do { } while(0)
2746 #define cache_free_debugcheck(x,objp,z) (objp)
2749 static struct page *get_first_slab(struct kmem_cache_node *n)
2753 page = list_first_entry_or_null(&n->slabs_partial,
2756 n->free_touched = 1;
2757 page = list_first_entry_or_null(&n->slabs_free,
2764 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2768 struct kmem_cache_node *n;
2769 struct array_cache *ac;
2773 node = numa_mem_id();
2774 if (unlikely(force_refill))
2777 ac = cpu_cache_get(cachep);
2778 batchcount = ac->batchcount;
2779 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2781 * If there was little recent activity on this cache, then
2782 * perform only a partial refill. Otherwise we could generate
2785 batchcount = BATCHREFILL_LIMIT;
2787 n = get_node(cachep, node);
2789 BUG_ON(ac->avail > 0 || !n);
2790 spin_lock(&n->list_lock);
2792 /* See if we can refill from the shared array */
2793 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2794 n->shared->touched = 1;
2798 while (batchcount > 0) {
2800 /* Get slab alloc is to come from. */
2801 page = get_first_slab(n);
2805 check_spinlock_acquired(cachep);
2808 * The slab was either on partial or free list so
2809 * there must be at least one object available for
2812 BUG_ON(page->active >= cachep->num);
2814 while (page->active < cachep->num && batchcount--) {
2815 STATS_INC_ALLOCED(cachep);
2816 STATS_INC_ACTIVE(cachep);
2817 STATS_SET_HIGH(cachep);
2819 ac_put_obj(cachep, ac, slab_get_obj(cachep, page));
2822 /* move slabp to correct slabp list: */
2823 list_del(&page->lru);
2824 if (page->active == cachep->num)
2825 list_add(&page->lru, &n->slabs_full);
2827 list_add(&page->lru, &n->slabs_partial);
2831 n->free_objects -= ac->avail;
2833 spin_unlock(&n->list_lock);
2835 if (unlikely(!ac->avail)) {
2838 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2840 /* cache_grow can reenable interrupts, then ac could change. */
2841 ac = cpu_cache_get(cachep);
2842 node = numa_mem_id();
2844 /* no objects in sight? abort */
2845 if (!x && (ac->avail == 0 || force_refill))
2848 if (!ac->avail) /* objects refilled by interrupt? */
2853 return ac_get_obj(cachep, ac, flags, force_refill);
2856 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2859 might_sleep_if(gfpflags_allow_blocking(flags));
2861 kmem_flagcheck(cachep, flags);
2866 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2867 gfp_t flags, void *objp, unsigned long caller)
2873 if (cachep->flags & SLAB_POISON) {
2874 check_poison_obj(cachep, objp);
2875 slab_kernel_map(cachep, objp, 1, 0);
2876 poison_obj(cachep, objp, POISON_INUSE);
2878 if (cachep->flags & SLAB_STORE_USER)
2879 *dbg_userword(cachep, objp) = (void *)caller;
2881 if (cachep->flags & SLAB_RED_ZONE) {
2882 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2883 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2884 slab_error(cachep, "double free, or memory outside"
2885 " object was overwritten");
2887 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2888 objp, *dbg_redzone1(cachep, objp),
2889 *dbg_redzone2(cachep, objp));
2891 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2892 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2895 page = virt_to_head_page(objp);
2896 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
2897 objp += obj_offset(cachep);
2898 if (cachep->ctor && cachep->flags & SLAB_POISON)
2900 if (ARCH_SLAB_MINALIGN &&
2901 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2902 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2903 objp, (int)ARCH_SLAB_MINALIGN);
2908 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2911 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2914 struct array_cache *ac;
2915 bool force_refill = false;
2919 ac = cpu_cache_get(cachep);
2920 if (likely(ac->avail)) {
2922 objp = ac_get_obj(cachep, ac, flags, false);
2925 * Allow for the possibility all avail objects are not allowed
2926 * by the current flags
2929 STATS_INC_ALLOCHIT(cachep);
2932 force_refill = true;
2935 STATS_INC_ALLOCMISS(cachep);
2936 objp = cache_alloc_refill(cachep, flags, force_refill);
2938 * the 'ac' may be updated by cache_alloc_refill(),
2939 * and kmemleak_erase() requires its correct value.
2941 ac = cpu_cache_get(cachep);
2945 * To avoid a false negative, if an object that is in one of the
2946 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2947 * treat the array pointers as a reference to the object.
2950 kmemleak_erase(&ac->entry[ac->avail]);
2956 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2958 * If we are in_interrupt, then process context, including cpusets and
2959 * mempolicy, may not apply and should not be used for allocation policy.
2961 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2963 int nid_alloc, nid_here;
2965 if (in_interrupt() || (flags & __GFP_THISNODE))
2967 nid_alloc = nid_here = numa_mem_id();
2968 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2969 nid_alloc = cpuset_slab_spread_node();
2970 else if (current->mempolicy)
2971 nid_alloc = mempolicy_slab_node();
2972 if (nid_alloc != nid_here)
2973 return ____cache_alloc_node(cachep, flags, nid_alloc);
2978 * Fallback function if there was no memory available and no objects on a
2979 * certain node and fall back is permitted. First we scan all the
2980 * available node for available objects. If that fails then we
2981 * perform an allocation without specifying a node. This allows the page
2982 * allocator to do its reclaim / fallback magic. We then insert the
2983 * slab into the proper nodelist and then allocate from it.
2985 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2987 struct zonelist *zonelist;
2991 enum zone_type high_zoneidx = gfp_zone(flags);
2994 unsigned int cpuset_mems_cookie;
2996 if (flags & __GFP_THISNODE)
2999 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3002 cpuset_mems_cookie = read_mems_allowed_begin();
3003 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3007 * Look through allowed nodes for objects available
3008 * from existing per node queues.
3010 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3011 nid = zone_to_nid(zone);
3013 if (cpuset_zone_allowed(zone, flags) &&
3014 get_node(cache, nid) &&
3015 get_node(cache, nid)->free_objects) {
3016 obj = ____cache_alloc_node(cache,
3017 gfp_exact_node(flags), nid);
3025 * This allocation will be performed within the constraints
3026 * of the current cpuset / memory policy requirements.
3027 * We may trigger various forms of reclaim on the allowed
3028 * set and go into memory reserves if necessary.
3032 if (gfpflags_allow_blocking(local_flags))
3034 kmem_flagcheck(cache, flags);
3035 page = kmem_getpages(cache, local_flags, numa_mem_id());
3036 if (gfpflags_allow_blocking(local_flags))
3037 local_irq_disable();
3040 * Insert into the appropriate per node queues
3042 nid = page_to_nid(page);
3043 if (cache_grow(cache, flags, nid, page)) {
3044 obj = ____cache_alloc_node(cache,
3045 gfp_exact_node(flags), nid);
3048 * Another processor may allocate the
3049 * objects in the slab since we are
3050 * not holding any locks.
3054 /* cache_grow already freed obj */
3060 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3066 * A interface to enable slab creation on nodeid
3068 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3072 struct kmem_cache_node *n;
3076 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3077 n = get_node(cachep, nodeid);
3082 spin_lock(&n->list_lock);
3083 page = get_first_slab(n);
3087 check_spinlock_acquired_node(cachep, nodeid);
3089 STATS_INC_NODEALLOCS(cachep);
3090 STATS_INC_ACTIVE(cachep);
3091 STATS_SET_HIGH(cachep);
3093 BUG_ON(page->active == cachep->num);
3095 obj = slab_get_obj(cachep, page);
3097 /* move slabp to correct slabp list: */
3098 list_del(&page->lru);
3100 if (page->active == cachep->num)
3101 list_add(&page->lru, &n->slabs_full);
3103 list_add(&page->lru, &n->slabs_partial);
3105 spin_unlock(&n->list_lock);
3109 spin_unlock(&n->list_lock);
3110 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3114 return fallback_alloc(cachep, flags);
3120 static __always_inline void *
3121 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3122 unsigned long caller)
3124 unsigned long save_flags;
3126 int slab_node = numa_mem_id();
3128 flags &= gfp_allowed_mask;
3129 cachep = slab_pre_alloc_hook(cachep, flags);
3130 if (unlikely(!cachep))
3133 cache_alloc_debugcheck_before(cachep, flags);
3134 local_irq_save(save_flags);
3136 if (nodeid == NUMA_NO_NODE)
3139 if (unlikely(!get_node(cachep, nodeid))) {
3140 /* Node not bootstrapped yet */
3141 ptr = fallback_alloc(cachep, flags);
3145 if (nodeid == slab_node) {
3147 * Use the locally cached objects if possible.
3148 * However ____cache_alloc does not allow fallback
3149 * to other nodes. It may fail while we still have
3150 * objects on other nodes available.
3152 ptr = ____cache_alloc(cachep, flags);
3156 /* ___cache_alloc_node can fall back to other nodes */
3157 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3159 local_irq_restore(save_flags);
3160 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3162 if (unlikely(flags & __GFP_ZERO) && ptr)
3163 memset(ptr, 0, cachep->object_size);
3165 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3169 static __always_inline void *
3170 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3174 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3175 objp = alternate_node_alloc(cache, flags);
3179 objp = ____cache_alloc(cache, flags);
3182 * We may just have run out of memory on the local node.
3183 * ____cache_alloc_node() knows how to locate memory on other nodes
3186 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3193 static __always_inline void *
3194 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3196 return ____cache_alloc(cachep, flags);
3199 #endif /* CONFIG_NUMA */
3201 static __always_inline void *
3202 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3204 unsigned long save_flags;
3207 flags &= gfp_allowed_mask;
3208 cachep = slab_pre_alloc_hook(cachep, flags);
3209 if (unlikely(!cachep))
3212 cache_alloc_debugcheck_before(cachep, flags);
3213 local_irq_save(save_flags);
3214 objp = __do_cache_alloc(cachep, flags);
3215 local_irq_restore(save_flags);
3216 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3219 if (unlikely(flags & __GFP_ZERO) && objp)
3220 memset(objp, 0, cachep->object_size);
3222 slab_post_alloc_hook(cachep, flags, 1, &objp);
3227 * Caller needs to acquire correct kmem_cache_node's list_lock
3228 * @list: List of detached free slabs should be freed by caller
3230 static void free_block(struct kmem_cache *cachep, void **objpp,
3231 int nr_objects, int node, struct list_head *list)
3234 struct kmem_cache_node *n = get_node(cachep, node);
3236 for (i = 0; i < nr_objects; i++) {
3240 clear_obj_pfmemalloc(&objpp[i]);
3243 page = virt_to_head_page(objp);
3244 list_del(&page->lru);
3245 check_spinlock_acquired_node(cachep, node);
3246 slab_put_obj(cachep, page, objp);
3247 STATS_DEC_ACTIVE(cachep);
3250 /* fixup slab chains */
3251 if (page->active == 0) {
3252 if (n->free_objects > n->free_limit) {
3253 n->free_objects -= cachep->num;
3254 list_add_tail(&page->lru, list);
3256 list_add(&page->lru, &n->slabs_free);
3259 /* Unconditionally move a slab to the end of the
3260 * partial list on free - maximum time for the
3261 * other objects to be freed, too.
3263 list_add_tail(&page->lru, &n->slabs_partial);
3268 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3271 struct kmem_cache_node *n;
3272 int node = numa_mem_id();
3275 batchcount = ac->batchcount;
3278 n = get_node(cachep, node);
3279 spin_lock(&n->list_lock);
3281 struct array_cache *shared_array = n->shared;
3282 int max = shared_array->limit - shared_array->avail;
3284 if (batchcount > max)
3286 memcpy(&(shared_array->entry[shared_array->avail]),
3287 ac->entry, sizeof(void *) * batchcount);
3288 shared_array->avail += batchcount;
3293 free_block(cachep, ac->entry, batchcount, node, &list);
3300 list_for_each_entry(page, &n->slabs_free, lru) {
3301 BUG_ON(page->active);
3305 STATS_SET_FREEABLE(cachep, i);
3308 spin_unlock(&n->list_lock);
3309 slabs_destroy(cachep, &list);
3310 ac->avail -= batchcount;
3311 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3315 * Release an obj back to its cache. If the obj has a constructed state, it must
3316 * be in this state _before_ it is released. Called with disabled ints.
3318 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3319 unsigned long caller)
3321 struct array_cache *ac = cpu_cache_get(cachep);
3324 kmemleak_free_recursive(objp, cachep->flags);
3325 objp = cache_free_debugcheck(cachep, objp, caller);
3327 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3330 * Skip calling cache_free_alien() when the platform is not numa.
3331 * This will avoid cache misses that happen while accessing slabp (which
3332 * is per page memory reference) to get nodeid. Instead use a global
3333 * variable to skip the call, which is mostly likely to be present in
3336 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3339 if (ac->avail < ac->limit) {
3340 STATS_INC_FREEHIT(cachep);
3342 STATS_INC_FREEMISS(cachep);
3343 cache_flusharray(cachep, ac);
3346 ac_put_obj(cachep, ac, objp);
3350 * kmem_cache_alloc - Allocate an object
3351 * @cachep: The cache to allocate from.
3352 * @flags: See kmalloc().
3354 * Allocate an object from this cache. The flags are only relevant
3355 * if the cache has no available objects.
3357 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3359 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3361 trace_kmem_cache_alloc(_RET_IP_, ret,
3362 cachep->object_size, cachep->size, flags);
3366 EXPORT_SYMBOL(kmem_cache_alloc);
3368 static __always_inline void
3369 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3370 size_t size, void **p, unsigned long caller)
3374 for (i = 0; i < size; i++)
3375 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3378 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3383 s = slab_pre_alloc_hook(s, flags);
3387 cache_alloc_debugcheck_before(s, flags);
3389 local_irq_disable();
3390 for (i = 0; i < size; i++) {
3391 void *objp = __do_cache_alloc(s, flags);
3393 if (unlikely(!objp))
3399 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3401 /* Clear memory outside IRQ disabled section */
3402 if (unlikely(flags & __GFP_ZERO))
3403 for (i = 0; i < size; i++)
3404 memset(p[i], 0, s->object_size);
3406 slab_post_alloc_hook(s, flags, size, p);
3407 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3411 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3412 slab_post_alloc_hook(s, flags, i, p);
3413 __kmem_cache_free_bulk(s, i, p);
3416 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3418 #ifdef CONFIG_TRACING
3420 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3424 ret = slab_alloc(cachep, flags, _RET_IP_);
3426 trace_kmalloc(_RET_IP_, ret,
3427 size, cachep->size, flags);
3430 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3435 * kmem_cache_alloc_node - Allocate an object on the specified node
3436 * @cachep: The cache to allocate from.
3437 * @flags: See kmalloc().
3438 * @nodeid: node number of the target node.
3440 * Identical to kmem_cache_alloc but it will allocate memory on the given
3441 * node, which can improve the performance for cpu bound structures.
3443 * Fallback to other node is possible if __GFP_THISNODE is not set.
3445 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3447 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3449 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3450 cachep->object_size, cachep->size,
3455 EXPORT_SYMBOL(kmem_cache_alloc_node);
3457 #ifdef CONFIG_TRACING
3458 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3465 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3467 trace_kmalloc_node(_RET_IP_, ret,
3472 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3475 static __always_inline void *
3476 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3478 struct kmem_cache *cachep;
3480 cachep = kmalloc_slab(size, flags);
3481 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3483 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3486 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3488 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3490 EXPORT_SYMBOL(__kmalloc_node);
3492 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3493 int node, unsigned long caller)
3495 return __do_kmalloc_node(size, flags, node, caller);
3497 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3498 #endif /* CONFIG_NUMA */
3501 * __do_kmalloc - allocate memory
3502 * @size: how many bytes of memory are required.
3503 * @flags: the type of memory to allocate (see kmalloc).
3504 * @caller: function caller for debug tracking of the caller
3506 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3507 unsigned long caller)
3509 struct kmem_cache *cachep;
3512 cachep = kmalloc_slab(size, flags);
3513 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3515 ret = slab_alloc(cachep, flags, caller);
3517 trace_kmalloc(caller, ret,
3518 size, cachep->size, flags);
3523 void *__kmalloc(size_t size, gfp_t flags)
3525 return __do_kmalloc(size, flags, _RET_IP_);
3527 EXPORT_SYMBOL(__kmalloc);
3529 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3531 return __do_kmalloc(size, flags, caller);
3533 EXPORT_SYMBOL(__kmalloc_track_caller);
3536 * kmem_cache_free - Deallocate an object
3537 * @cachep: The cache the allocation was from.
3538 * @objp: The previously allocated object.
3540 * Free an object which was previously allocated from this
3543 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3545 unsigned long flags;
3546 cachep = cache_from_obj(cachep, objp);
3550 local_irq_save(flags);
3551 debug_check_no_locks_freed(objp, cachep->object_size);
3552 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3553 debug_check_no_obj_freed(objp, cachep->object_size);
3554 __cache_free(cachep, objp, _RET_IP_);
3555 local_irq_restore(flags);
3557 trace_kmem_cache_free(_RET_IP_, objp);
3559 EXPORT_SYMBOL(kmem_cache_free);
3561 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3563 struct kmem_cache *s;
3566 local_irq_disable();
3567 for (i = 0; i < size; i++) {
3570 if (!orig_s) /* called via kfree_bulk */
3571 s = virt_to_cache(objp);
3573 s = cache_from_obj(orig_s, objp);
3575 debug_check_no_locks_freed(objp, s->object_size);
3576 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3577 debug_check_no_obj_freed(objp, s->object_size);
3579 __cache_free(s, objp, _RET_IP_);
3583 /* FIXME: add tracing */
3585 EXPORT_SYMBOL(kmem_cache_free_bulk);
3588 * kfree - free previously allocated memory
3589 * @objp: pointer returned by kmalloc.
3591 * If @objp is NULL, no operation is performed.
3593 * Don't free memory not originally allocated by kmalloc()
3594 * or you will run into trouble.
3596 void kfree(const void *objp)
3598 struct kmem_cache *c;
3599 unsigned long flags;
3601 trace_kfree(_RET_IP_, objp);
3603 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3605 local_irq_save(flags);
3606 kfree_debugcheck(objp);
3607 c = virt_to_cache(objp);
3608 debug_check_no_locks_freed(objp, c->object_size);
3610 debug_check_no_obj_freed(objp, c->object_size);
3611 __cache_free(c, (void *)objp, _RET_IP_);
3612 local_irq_restore(flags);
3614 EXPORT_SYMBOL(kfree);
3617 * This initializes kmem_cache_node or resizes various caches for all nodes.
3619 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3622 struct kmem_cache_node *n;
3623 struct array_cache *new_shared;
3624 struct alien_cache **new_alien = NULL;
3626 for_each_online_node(node) {
3628 if (use_alien_caches) {
3629 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3635 if (cachep->shared) {
3636 new_shared = alloc_arraycache(node,
3637 cachep->shared*cachep->batchcount,
3640 free_alien_cache(new_alien);
3645 n = get_node(cachep, node);
3647 struct array_cache *shared = n->shared;
3650 spin_lock_irq(&n->list_lock);
3653 free_block(cachep, shared->entry,
3654 shared->avail, node, &list);
3656 n->shared = new_shared;
3658 n->alien = new_alien;
3661 n->free_limit = (1 + nr_cpus_node(node)) *
3662 cachep->batchcount + cachep->num;
3663 spin_unlock_irq(&n->list_lock);
3664 slabs_destroy(cachep, &list);
3666 free_alien_cache(new_alien);
3669 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3671 free_alien_cache(new_alien);
3676 kmem_cache_node_init(n);
3677 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3678 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3679 n->shared = new_shared;
3680 n->alien = new_alien;
3681 n->free_limit = (1 + nr_cpus_node(node)) *
3682 cachep->batchcount + cachep->num;
3683 cachep->node[node] = n;
3688 if (!cachep->list.next) {
3689 /* Cache is not active yet. Roll back what we did */
3692 n = get_node(cachep, node);
3695 free_alien_cache(n->alien);
3697 cachep->node[node] = NULL;
3705 /* Always called with the slab_mutex held */
3706 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3707 int batchcount, int shared, gfp_t gfp)
3709 struct array_cache __percpu *cpu_cache, *prev;
3712 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3716 prev = cachep->cpu_cache;
3717 cachep->cpu_cache = cpu_cache;
3718 kick_all_cpus_sync();
3721 cachep->batchcount = batchcount;
3722 cachep->limit = limit;
3723 cachep->shared = shared;
3728 for_each_online_cpu(cpu) {
3731 struct kmem_cache_node *n;
3732 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3734 node = cpu_to_mem(cpu);
3735 n = get_node(cachep, node);
3736 spin_lock_irq(&n->list_lock);
3737 free_block(cachep, ac->entry, ac->avail, node, &list);
3738 spin_unlock_irq(&n->list_lock);
3739 slabs_destroy(cachep, &list);
3744 return alloc_kmem_cache_node(cachep, gfp);
3747 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3748 int batchcount, int shared, gfp_t gfp)
3751 struct kmem_cache *c;
3753 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3755 if (slab_state < FULL)
3758 if ((ret < 0) || !is_root_cache(cachep))
3761 lockdep_assert_held(&slab_mutex);
3762 for_each_memcg_cache(c, cachep) {
3763 /* return value determined by the root cache only */
3764 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3770 /* Called with slab_mutex held always */
3771 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3778 if (!is_root_cache(cachep)) {
3779 struct kmem_cache *root = memcg_root_cache(cachep);
3780 limit = root->limit;
3781 shared = root->shared;
3782 batchcount = root->batchcount;
3785 if (limit && shared && batchcount)
3788 * The head array serves three purposes:
3789 * - create a LIFO ordering, i.e. return objects that are cache-warm
3790 * - reduce the number of spinlock operations.
3791 * - reduce the number of linked list operations on the slab and
3792 * bufctl chains: array operations are cheaper.
3793 * The numbers are guessed, we should auto-tune as described by
3796 if (cachep->size > 131072)
3798 else if (cachep->size > PAGE_SIZE)
3800 else if (cachep->size > 1024)
3802 else if (cachep->size > 256)
3808 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3809 * allocation behaviour: Most allocs on one cpu, most free operations
3810 * on another cpu. For these cases, an efficient object passing between
3811 * cpus is necessary. This is provided by a shared array. The array
3812 * replaces Bonwick's magazine layer.
3813 * On uniprocessor, it's functionally equivalent (but less efficient)
3814 * to a larger limit. Thus disabled by default.
3817 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3822 * With debugging enabled, large batchcount lead to excessively long
3823 * periods with disabled local interrupts. Limit the batchcount
3828 batchcount = (limit + 1) / 2;
3830 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3832 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3833 cachep->name, -err);
3838 * Drain an array if it contains any elements taking the node lock only if
3839 * necessary. Note that the node listlock also protects the array_cache
3840 * if drain_array() is used on the shared array.
3842 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3843 struct array_cache *ac, int force, int node)
3848 if (!ac || !ac->avail)
3850 if (ac->touched && !force) {
3853 spin_lock_irq(&n->list_lock);
3855 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3856 if (tofree > ac->avail)
3857 tofree = (ac->avail + 1) / 2;
3858 free_block(cachep, ac->entry, tofree, node, &list);
3859 ac->avail -= tofree;
3860 memmove(ac->entry, &(ac->entry[tofree]),
3861 sizeof(void *) * ac->avail);
3863 spin_unlock_irq(&n->list_lock);
3864 slabs_destroy(cachep, &list);
3869 * cache_reap - Reclaim memory from caches.
3870 * @w: work descriptor
3872 * Called from workqueue/eventd every few seconds.
3874 * - clear the per-cpu caches for this CPU.
3875 * - return freeable pages to the main free memory pool.
3877 * If we cannot acquire the cache chain mutex then just give up - we'll try
3878 * again on the next iteration.
3880 static void cache_reap(struct work_struct *w)
3882 struct kmem_cache *searchp;
3883 struct kmem_cache_node *n;
3884 int node = numa_mem_id();
3885 struct delayed_work *work = to_delayed_work(w);
3887 if (!mutex_trylock(&slab_mutex))
3888 /* Give up. Setup the next iteration. */
3891 list_for_each_entry(searchp, &slab_caches, list) {
3895 * We only take the node lock if absolutely necessary and we
3896 * have established with reasonable certainty that
3897 * we can do some work if the lock was obtained.
3899 n = get_node(searchp, node);
3901 reap_alien(searchp, n);
3903 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3906 * These are racy checks but it does not matter
3907 * if we skip one check or scan twice.
3909 if (time_after(n->next_reap, jiffies))
3912 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3914 drain_array(searchp, n, n->shared, 0, node);
3916 if (n->free_touched)
3917 n->free_touched = 0;
3921 freed = drain_freelist(searchp, n, (n->free_limit +
3922 5 * searchp->num - 1) / (5 * searchp->num));
3923 STATS_ADD_REAPED(searchp, freed);
3929 mutex_unlock(&slab_mutex);
3932 /* Set up the next iteration */
3933 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3936 #ifdef CONFIG_SLABINFO
3937 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3940 unsigned long active_objs;
3941 unsigned long num_objs;
3942 unsigned long active_slabs = 0;
3943 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3947 struct kmem_cache_node *n;
3951 for_each_kmem_cache_node(cachep, node, n) {
3954 spin_lock_irq(&n->list_lock);
3956 list_for_each_entry(page, &n->slabs_full, lru) {
3957 if (page->active != cachep->num && !error)
3958 error = "slabs_full accounting error";
3959 active_objs += cachep->num;
3962 list_for_each_entry(page, &n->slabs_partial, lru) {
3963 if (page->active == cachep->num && !error)
3964 error = "slabs_partial accounting error";
3965 if (!page->active && !error)
3966 error = "slabs_partial accounting error";
3967 active_objs += page->active;
3970 list_for_each_entry(page, &n->slabs_free, lru) {
3971 if (page->active && !error)
3972 error = "slabs_free accounting error";
3975 free_objects += n->free_objects;
3977 shared_avail += n->shared->avail;
3979 spin_unlock_irq(&n->list_lock);
3981 num_slabs += active_slabs;
3982 num_objs = num_slabs * cachep->num;
3983 if (num_objs - active_objs != free_objects && !error)
3984 error = "free_objects accounting error";
3986 name = cachep->name;
3988 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3990 sinfo->active_objs = active_objs;
3991 sinfo->num_objs = num_objs;
3992 sinfo->active_slabs = active_slabs;
3993 sinfo->num_slabs = num_slabs;
3994 sinfo->shared_avail = shared_avail;
3995 sinfo->limit = cachep->limit;
3996 sinfo->batchcount = cachep->batchcount;
3997 sinfo->shared = cachep->shared;
3998 sinfo->objects_per_slab = cachep->num;
3999 sinfo->cache_order = cachep->gfporder;
4002 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4006 unsigned long high = cachep->high_mark;
4007 unsigned long allocs = cachep->num_allocations;
4008 unsigned long grown = cachep->grown;
4009 unsigned long reaped = cachep->reaped;
4010 unsigned long errors = cachep->errors;
4011 unsigned long max_freeable = cachep->max_freeable;
4012 unsigned long node_allocs = cachep->node_allocs;
4013 unsigned long node_frees = cachep->node_frees;
4014 unsigned long overflows = cachep->node_overflow;
4016 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4017 "%4lu %4lu %4lu %4lu %4lu",
4018 allocs, high, grown,
4019 reaped, errors, max_freeable, node_allocs,
4020 node_frees, overflows);
4024 unsigned long allochit = atomic_read(&cachep->allochit);
4025 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4026 unsigned long freehit = atomic_read(&cachep->freehit);
4027 unsigned long freemiss = atomic_read(&cachep->freemiss);
4029 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4030 allochit, allocmiss, freehit, freemiss);
4035 #define MAX_SLABINFO_WRITE 128
4037 * slabinfo_write - Tuning for the slab allocator
4039 * @buffer: user buffer
4040 * @count: data length
4043 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4044 size_t count, loff_t *ppos)
4046 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4047 int limit, batchcount, shared, res;
4048 struct kmem_cache *cachep;
4050 if (count > MAX_SLABINFO_WRITE)
4052 if (copy_from_user(&kbuf, buffer, count))
4054 kbuf[MAX_SLABINFO_WRITE] = '\0';
4056 tmp = strchr(kbuf, ' ');
4061 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4064 /* Find the cache in the chain of caches. */
4065 mutex_lock(&slab_mutex);
4067 list_for_each_entry(cachep, &slab_caches, list) {
4068 if (!strcmp(cachep->name, kbuf)) {
4069 if (limit < 1 || batchcount < 1 ||
4070 batchcount > limit || shared < 0) {
4073 res = do_tune_cpucache(cachep, limit,
4080 mutex_unlock(&slab_mutex);
4086 #ifdef CONFIG_DEBUG_SLAB_LEAK
4088 static inline int add_caller(unsigned long *n, unsigned long v)
4098 unsigned long *q = p + 2 * i;
4112 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4118 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4126 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4127 if (get_obj_status(page, i) != OBJECT_ACTIVE)
4130 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4135 static void show_symbol(struct seq_file *m, unsigned long address)
4137 #ifdef CONFIG_KALLSYMS
4138 unsigned long offset, size;
4139 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4141 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4142 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4144 seq_printf(m, " [%s]", modname);
4148 seq_printf(m, "%p", (void *)address);
4151 static int leaks_show(struct seq_file *m, void *p)
4153 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4155 struct kmem_cache_node *n;
4157 unsigned long *x = m->private;
4161 if (!(cachep->flags & SLAB_STORE_USER))
4163 if (!(cachep->flags & SLAB_RED_ZONE))
4166 /* OK, we can do it */
4170 for_each_kmem_cache_node(cachep, node, n) {
4173 spin_lock_irq(&n->list_lock);
4175 list_for_each_entry(page, &n->slabs_full, lru)
4176 handle_slab(x, cachep, page);
4177 list_for_each_entry(page, &n->slabs_partial, lru)
4178 handle_slab(x, cachep, page);
4179 spin_unlock_irq(&n->list_lock);
4181 name = cachep->name;
4183 /* Increase the buffer size */
4184 mutex_unlock(&slab_mutex);
4185 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4187 /* Too bad, we are really out */
4189 mutex_lock(&slab_mutex);
4192 *(unsigned long *)m->private = x[0] * 2;
4194 mutex_lock(&slab_mutex);
4195 /* Now make sure this entry will be retried */
4199 for (i = 0; i < x[1]; i++) {
4200 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4201 show_symbol(m, x[2*i+2]);
4208 static const struct seq_operations slabstats_op = {
4209 .start = slab_start,
4215 static int slabstats_open(struct inode *inode, struct file *file)
4219 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4223 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4228 static const struct file_operations proc_slabstats_operations = {
4229 .open = slabstats_open,
4231 .llseek = seq_lseek,
4232 .release = seq_release_private,
4236 static int __init slab_proc_init(void)
4238 #ifdef CONFIG_DEBUG_SLAB_LEAK
4239 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4243 module_init(slab_proc_init);
4247 * ksize - get the actual amount of memory allocated for a given object
4248 * @objp: Pointer to the object
4250 * kmalloc may internally round up allocations and return more memory
4251 * than requested. ksize() can be used to determine the actual amount of
4252 * memory allocated. The caller may use this additional memory, even though
4253 * a smaller amount of memory was initially specified with the kmalloc call.
4254 * The caller must guarantee that objp points to a valid object previously
4255 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4256 * must not be freed during the duration of the call.
4258 size_t ksize(const void *objp)
4261 if (unlikely(objp == ZERO_SIZE_PTR))
4264 return virt_to_cache(objp)->object_size;
4266 EXPORT_SYMBOL(ksize);