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 #ifdef CONFIG_DEBUG_SLAB_LEAK
385 static inline bool is_store_user_clean(struct kmem_cache *cachep)
387 return atomic_read(&cachep->store_user_clean) == 1;
390 static inline void set_store_user_clean(struct kmem_cache *cachep)
392 atomic_set(&cachep->store_user_clean, 1);
395 static inline void set_store_user_dirty(struct kmem_cache *cachep)
397 if (is_store_user_clean(cachep))
398 atomic_set(&cachep->store_user_clean, 0);
402 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
407 * Do not go above this order unless 0 objects fit into the slab or
408 * overridden on the command line.
410 #define SLAB_MAX_ORDER_HI 1
411 #define SLAB_MAX_ORDER_LO 0
412 static int slab_max_order = SLAB_MAX_ORDER_LO;
413 static bool slab_max_order_set __initdata;
415 static inline struct kmem_cache *virt_to_cache(const void *obj)
417 struct page *page = virt_to_head_page(obj);
418 return page->slab_cache;
421 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
424 return page->s_mem + cache->size * idx;
428 * We want to avoid an expensive divide : (offset / cache->size)
429 * Using the fact that size is a constant for a particular cache,
430 * we can replace (offset / cache->size) by
431 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
433 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
434 const struct page *page, void *obj)
436 u32 offset = (obj - page->s_mem);
437 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
440 #define BOOT_CPUCACHE_ENTRIES 1
441 /* internal cache of cache description objs */
442 static struct kmem_cache kmem_cache_boot = {
444 .limit = BOOT_CPUCACHE_ENTRIES,
446 .size = sizeof(struct kmem_cache),
447 .name = "kmem_cache",
450 #define BAD_ALIEN_MAGIC 0x01020304ul
452 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
454 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
456 return this_cpu_ptr(cachep->cpu_cache);
460 * Calculate the number of objects and left-over bytes for a given buffer size.
462 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
463 unsigned long flags, size_t *left_over, unsigned int *num)
465 size_t slab_size = PAGE_SIZE << gfporder;
468 * The slab management structure can be either off the slab or
469 * on it. For the latter case, the memory allocated for a
472 * - @buffer_size bytes for each object
473 * - One freelist_idx_t for each object
475 * We don't need to consider alignment of freelist because
476 * freelist will be at the end of slab page. The objects will be
477 * at the correct alignment.
479 * If the slab management structure is off the slab, then the
480 * alignment will already be calculated into the size. Because
481 * the slabs are all pages aligned, the objects will be at the
482 * correct alignment when allocated.
484 if (flags & CFLGS_OFF_SLAB) {
485 *num = slab_size / buffer_size;
486 *left_over = slab_size % buffer_size;
488 *num = slab_size / (buffer_size + sizeof(freelist_idx_t));
489 *left_over = slab_size %
490 (buffer_size + sizeof(freelist_idx_t));
495 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
497 static void __slab_error(const char *function, struct kmem_cache *cachep,
500 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
501 function, cachep->name, msg);
503 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
508 * By default on NUMA we use alien caches to stage the freeing of
509 * objects allocated from other nodes. This causes massive memory
510 * inefficiencies when using fake NUMA setup to split memory into a
511 * large number of small nodes, so it can be disabled on the command
515 static int use_alien_caches __read_mostly = 1;
516 static int __init noaliencache_setup(char *s)
518 use_alien_caches = 0;
521 __setup("noaliencache", noaliencache_setup);
523 static int __init slab_max_order_setup(char *str)
525 get_option(&str, &slab_max_order);
526 slab_max_order = slab_max_order < 0 ? 0 :
527 min(slab_max_order, MAX_ORDER - 1);
528 slab_max_order_set = true;
532 __setup("slab_max_order=", slab_max_order_setup);
536 * Special reaping functions for NUMA systems called from cache_reap().
537 * These take care of doing round robin flushing of alien caches (containing
538 * objects freed on different nodes from which they were allocated) and the
539 * flushing of remote pcps by calling drain_node_pages.
541 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
543 static void init_reap_node(int cpu)
547 node = next_node(cpu_to_mem(cpu), node_online_map);
548 if (node == MAX_NUMNODES)
549 node = first_node(node_online_map);
551 per_cpu(slab_reap_node, cpu) = node;
554 static void next_reap_node(void)
556 int node = __this_cpu_read(slab_reap_node);
558 node = next_node(node, node_online_map);
559 if (unlikely(node >= MAX_NUMNODES))
560 node = first_node(node_online_map);
561 __this_cpu_write(slab_reap_node, node);
565 #define init_reap_node(cpu) do { } while (0)
566 #define next_reap_node(void) do { } while (0)
570 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
571 * via the workqueue/eventd.
572 * Add the CPU number into the expiration time to minimize the possibility of
573 * the CPUs getting into lockstep and contending for the global cache chain
576 static void start_cpu_timer(int cpu)
578 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
581 * When this gets called from do_initcalls via cpucache_init(),
582 * init_workqueues() has already run, so keventd will be setup
585 if (keventd_up() && reap_work->work.func == NULL) {
587 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
588 schedule_delayed_work_on(cpu, reap_work,
589 __round_jiffies_relative(HZ, cpu));
593 static void init_arraycache(struct array_cache *ac, int limit, int batch)
596 * The array_cache structures contain pointers to free object.
597 * However, when such objects are allocated or transferred to another
598 * cache the pointers are not cleared and they could be counted as
599 * valid references during a kmemleak scan. Therefore, kmemleak must
600 * not scan such objects.
602 kmemleak_no_scan(ac);
606 ac->batchcount = batch;
611 static struct array_cache *alloc_arraycache(int node, int entries,
612 int batchcount, gfp_t gfp)
614 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
615 struct array_cache *ac = NULL;
617 ac = kmalloc_node(memsize, gfp, node);
618 init_arraycache(ac, entries, batchcount);
622 static inline bool is_slab_pfmemalloc(struct page *page)
624 return PageSlabPfmemalloc(page);
627 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
628 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
629 struct array_cache *ac)
631 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
635 if (!pfmemalloc_active)
638 spin_lock_irqsave(&n->list_lock, flags);
639 list_for_each_entry(page, &n->slabs_full, lru)
640 if (is_slab_pfmemalloc(page))
643 list_for_each_entry(page, &n->slabs_partial, lru)
644 if (is_slab_pfmemalloc(page))
647 list_for_each_entry(page, &n->slabs_free, lru)
648 if (is_slab_pfmemalloc(page))
651 pfmemalloc_active = false;
653 spin_unlock_irqrestore(&n->list_lock, flags);
656 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
657 gfp_t flags, bool force_refill)
660 void *objp = ac->entry[--ac->avail];
662 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
663 if (unlikely(is_obj_pfmemalloc(objp))) {
664 struct kmem_cache_node *n;
666 if (gfp_pfmemalloc_allowed(flags)) {
667 clear_obj_pfmemalloc(&objp);
671 /* The caller cannot use PFMEMALLOC objects, find another one */
672 for (i = 0; i < ac->avail; i++) {
673 /* If a !PFMEMALLOC object is found, swap them */
674 if (!is_obj_pfmemalloc(ac->entry[i])) {
676 ac->entry[i] = ac->entry[ac->avail];
677 ac->entry[ac->avail] = objp;
683 * If there are empty slabs on the slabs_free list and we are
684 * being forced to refill the cache, mark this one !pfmemalloc.
686 n = get_node(cachep, numa_mem_id());
687 if (!list_empty(&n->slabs_free) && force_refill) {
688 struct page *page = virt_to_head_page(objp);
689 ClearPageSlabPfmemalloc(page);
690 clear_obj_pfmemalloc(&objp);
691 recheck_pfmemalloc_active(cachep, ac);
695 /* No !PFMEMALLOC objects available */
703 static inline void *ac_get_obj(struct kmem_cache *cachep,
704 struct array_cache *ac, gfp_t flags, bool force_refill)
708 if (unlikely(sk_memalloc_socks()))
709 objp = __ac_get_obj(cachep, ac, flags, force_refill);
711 objp = ac->entry[--ac->avail];
716 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
717 struct array_cache *ac, void *objp)
719 if (unlikely(pfmemalloc_active)) {
720 /* Some pfmemalloc slabs exist, check if this is one */
721 struct page *page = virt_to_head_page(objp);
722 if (PageSlabPfmemalloc(page))
723 set_obj_pfmemalloc(&objp);
729 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
732 if (unlikely(sk_memalloc_socks()))
733 objp = __ac_put_obj(cachep, ac, objp);
735 ac->entry[ac->avail++] = objp;
739 * Transfer objects in one arraycache to another.
740 * Locking must be handled by the caller.
742 * Return the number of entries transferred.
744 static int transfer_objects(struct array_cache *to,
745 struct array_cache *from, unsigned int max)
747 /* Figure out how many entries to transfer */
748 int nr = min3(from->avail, max, to->limit - to->avail);
753 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
763 #define drain_alien_cache(cachep, alien) do { } while (0)
764 #define reap_alien(cachep, n) do { } while (0)
766 static inline struct alien_cache **alloc_alien_cache(int node,
767 int limit, gfp_t gfp)
769 return (struct alien_cache **)BAD_ALIEN_MAGIC;
772 static inline void free_alien_cache(struct alien_cache **ac_ptr)
776 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
781 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
787 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
788 gfp_t flags, int nodeid)
793 static inline gfp_t gfp_exact_node(gfp_t flags)
798 #else /* CONFIG_NUMA */
800 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
801 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
803 static struct alien_cache *__alloc_alien_cache(int node, int entries,
804 int batch, gfp_t gfp)
806 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
807 struct alien_cache *alc = NULL;
809 alc = kmalloc_node(memsize, gfp, node);
810 init_arraycache(&alc->ac, entries, batch);
811 spin_lock_init(&alc->lock);
815 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
817 struct alien_cache **alc_ptr;
818 size_t memsize = sizeof(void *) * nr_node_ids;
823 alc_ptr = kzalloc_node(memsize, gfp, node);
828 if (i == node || !node_online(i))
830 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
832 for (i--; i >= 0; i--)
841 static void free_alien_cache(struct alien_cache **alc_ptr)
852 static void __drain_alien_cache(struct kmem_cache *cachep,
853 struct array_cache *ac, int node,
854 struct list_head *list)
856 struct kmem_cache_node *n = get_node(cachep, node);
859 spin_lock(&n->list_lock);
861 * Stuff objects into the remote nodes shared array first.
862 * That way we could avoid the overhead of putting the objects
863 * into the free lists and getting them back later.
866 transfer_objects(n->shared, ac, ac->limit);
868 free_block(cachep, ac->entry, ac->avail, node, list);
870 spin_unlock(&n->list_lock);
875 * Called from cache_reap() to regularly drain alien caches round robin.
877 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
879 int node = __this_cpu_read(slab_reap_node);
882 struct alien_cache *alc = n->alien[node];
883 struct array_cache *ac;
887 if (ac->avail && spin_trylock_irq(&alc->lock)) {
890 __drain_alien_cache(cachep, ac, node, &list);
891 spin_unlock_irq(&alc->lock);
892 slabs_destroy(cachep, &list);
898 static void drain_alien_cache(struct kmem_cache *cachep,
899 struct alien_cache **alien)
902 struct alien_cache *alc;
903 struct array_cache *ac;
906 for_each_online_node(i) {
912 spin_lock_irqsave(&alc->lock, flags);
913 __drain_alien_cache(cachep, ac, i, &list);
914 spin_unlock_irqrestore(&alc->lock, flags);
915 slabs_destroy(cachep, &list);
920 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
921 int node, int page_node)
923 struct kmem_cache_node *n;
924 struct alien_cache *alien = NULL;
925 struct array_cache *ac;
928 n = get_node(cachep, node);
929 STATS_INC_NODEFREES(cachep);
930 if (n->alien && n->alien[page_node]) {
931 alien = n->alien[page_node];
933 spin_lock(&alien->lock);
934 if (unlikely(ac->avail == ac->limit)) {
935 STATS_INC_ACOVERFLOW(cachep);
936 __drain_alien_cache(cachep, ac, page_node, &list);
938 ac_put_obj(cachep, ac, objp);
939 spin_unlock(&alien->lock);
940 slabs_destroy(cachep, &list);
942 n = get_node(cachep, page_node);
943 spin_lock(&n->list_lock);
944 free_block(cachep, &objp, 1, page_node, &list);
945 spin_unlock(&n->list_lock);
946 slabs_destroy(cachep, &list);
951 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
953 int page_node = page_to_nid(virt_to_page(objp));
954 int node = numa_mem_id();
956 * Make sure we are not freeing a object from another node to the array
959 if (likely(node == page_node))
962 return __cache_free_alien(cachep, objp, node, page_node);
966 * Construct gfp mask to allocate from a specific node but do not direct reclaim
967 * or warn about failures. kswapd may still wake to reclaim in the background.
969 static inline gfp_t gfp_exact_node(gfp_t flags)
971 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
976 * Allocates and initializes node for a node on each slab cache, used for
977 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
978 * will be allocated off-node since memory is not yet online for the new node.
979 * When hotplugging memory or a cpu, existing node are not replaced if
982 * Must hold slab_mutex.
984 static int init_cache_node_node(int node)
986 struct kmem_cache *cachep;
987 struct kmem_cache_node *n;
988 const size_t memsize = sizeof(struct kmem_cache_node);
990 list_for_each_entry(cachep, &slab_caches, list) {
992 * Set up the kmem_cache_node for cpu before we can
993 * begin anything. Make sure some other cpu on this
994 * node has not already allocated this
996 n = get_node(cachep, node);
998 n = kmalloc_node(memsize, GFP_KERNEL, node);
1001 kmem_cache_node_init(n);
1002 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1003 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1006 * The kmem_cache_nodes don't come and go as CPUs
1007 * come and go. slab_mutex is sufficient
1010 cachep->node[node] = n;
1013 spin_lock_irq(&n->list_lock);
1015 (1 + nr_cpus_node(node)) *
1016 cachep->batchcount + cachep->num;
1017 spin_unlock_irq(&n->list_lock);
1022 static inline int slabs_tofree(struct kmem_cache *cachep,
1023 struct kmem_cache_node *n)
1025 return (n->free_objects + cachep->num - 1) / cachep->num;
1028 static void cpuup_canceled(long cpu)
1030 struct kmem_cache *cachep;
1031 struct kmem_cache_node *n = NULL;
1032 int node = cpu_to_mem(cpu);
1033 const struct cpumask *mask = cpumask_of_node(node);
1035 list_for_each_entry(cachep, &slab_caches, list) {
1036 struct array_cache *nc;
1037 struct array_cache *shared;
1038 struct alien_cache **alien;
1041 n = get_node(cachep, node);
1045 spin_lock_irq(&n->list_lock);
1047 /* Free limit for this kmem_cache_node */
1048 n->free_limit -= cachep->batchcount;
1050 /* cpu is dead; no one can alloc from it. */
1051 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1053 free_block(cachep, nc->entry, nc->avail, node, &list);
1057 if (!cpumask_empty(mask)) {
1058 spin_unlock_irq(&n->list_lock);
1064 free_block(cachep, shared->entry,
1065 shared->avail, node, &list);
1072 spin_unlock_irq(&n->list_lock);
1076 drain_alien_cache(cachep, alien);
1077 free_alien_cache(alien);
1081 slabs_destroy(cachep, &list);
1084 * In the previous loop, all the objects were freed to
1085 * the respective cache's slabs, now we can go ahead and
1086 * shrink each nodelist to its limit.
1088 list_for_each_entry(cachep, &slab_caches, list) {
1089 n = get_node(cachep, node);
1092 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1096 static int cpuup_prepare(long cpu)
1098 struct kmem_cache *cachep;
1099 struct kmem_cache_node *n = NULL;
1100 int node = cpu_to_mem(cpu);
1104 * We need to do this right in the beginning since
1105 * alloc_arraycache's are going to use this list.
1106 * kmalloc_node allows us to add the slab to the right
1107 * kmem_cache_node and not this cpu's kmem_cache_node
1109 err = init_cache_node_node(node);
1114 * Now we can go ahead with allocating the shared arrays and
1117 list_for_each_entry(cachep, &slab_caches, list) {
1118 struct array_cache *shared = NULL;
1119 struct alien_cache **alien = NULL;
1121 if (cachep->shared) {
1122 shared = alloc_arraycache(node,
1123 cachep->shared * cachep->batchcount,
1124 0xbaadf00d, GFP_KERNEL);
1128 if (use_alien_caches) {
1129 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1135 n = get_node(cachep, node);
1138 spin_lock_irq(&n->list_lock);
1141 * We are serialised from CPU_DEAD or
1142 * CPU_UP_CANCELLED by the cpucontrol lock
1153 spin_unlock_irq(&n->list_lock);
1155 free_alien_cache(alien);
1160 cpuup_canceled(cpu);
1164 static int cpuup_callback(struct notifier_block *nfb,
1165 unsigned long action, void *hcpu)
1167 long cpu = (long)hcpu;
1171 case CPU_UP_PREPARE:
1172 case CPU_UP_PREPARE_FROZEN:
1173 mutex_lock(&slab_mutex);
1174 err = cpuup_prepare(cpu);
1175 mutex_unlock(&slab_mutex);
1178 case CPU_ONLINE_FROZEN:
1179 start_cpu_timer(cpu);
1181 #ifdef CONFIG_HOTPLUG_CPU
1182 case CPU_DOWN_PREPARE:
1183 case CPU_DOWN_PREPARE_FROZEN:
1185 * Shutdown cache reaper. Note that the slab_mutex is
1186 * held so that if cache_reap() is invoked it cannot do
1187 * anything expensive but will only modify reap_work
1188 * and reschedule the timer.
1190 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1191 /* Now the cache_reaper is guaranteed to be not running. */
1192 per_cpu(slab_reap_work, cpu).work.func = NULL;
1194 case CPU_DOWN_FAILED:
1195 case CPU_DOWN_FAILED_FROZEN:
1196 start_cpu_timer(cpu);
1199 case CPU_DEAD_FROZEN:
1201 * Even if all the cpus of a node are down, we don't free the
1202 * kmem_cache_node of any cache. This to avoid a race between
1203 * cpu_down, and a kmalloc allocation from another cpu for
1204 * memory from the node of the cpu going down. The node
1205 * structure is usually allocated from kmem_cache_create() and
1206 * gets destroyed at kmem_cache_destroy().
1210 case CPU_UP_CANCELED:
1211 case CPU_UP_CANCELED_FROZEN:
1212 mutex_lock(&slab_mutex);
1213 cpuup_canceled(cpu);
1214 mutex_unlock(&slab_mutex);
1217 return notifier_from_errno(err);
1220 static struct notifier_block cpucache_notifier = {
1221 &cpuup_callback, NULL, 0
1224 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1226 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1227 * Returns -EBUSY if all objects cannot be drained so that the node is not
1230 * Must hold slab_mutex.
1232 static int __meminit drain_cache_node_node(int node)
1234 struct kmem_cache *cachep;
1237 list_for_each_entry(cachep, &slab_caches, list) {
1238 struct kmem_cache_node *n;
1240 n = get_node(cachep, node);
1244 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1246 if (!list_empty(&n->slabs_full) ||
1247 !list_empty(&n->slabs_partial)) {
1255 static int __meminit slab_memory_callback(struct notifier_block *self,
1256 unsigned long action, void *arg)
1258 struct memory_notify *mnb = arg;
1262 nid = mnb->status_change_nid;
1267 case MEM_GOING_ONLINE:
1268 mutex_lock(&slab_mutex);
1269 ret = init_cache_node_node(nid);
1270 mutex_unlock(&slab_mutex);
1272 case MEM_GOING_OFFLINE:
1273 mutex_lock(&slab_mutex);
1274 ret = drain_cache_node_node(nid);
1275 mutex_unlock(&slab_mutex);
1279 case MEM_CANCEL_ONLINE:
1280 case MEM_CANCEL_OFFLINE:
1284 return notifier_from_errno(ret);
1286 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1289 * swap the static kmem_cache_node with kmalloced memory
1291 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1294 struct kmem_cache_node *ptr;
1296 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1299 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1301 * Do not assume that spinlocks can be initialized via memcpy:
1303 spin_lock_init(&ptr->list_lock);
1305 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1306 cachep->node[nodeid] = ptr;
1310 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1311 * size of kmem_cache_node.
1313 static void __init set_up_node(struct kmem_cache *cachep, int index)
1317 for_each_online_node(node) {
1318 cachep->node[node] = &init_kmem_cache_node[index + node];
1319 cachep->node[node]->next_reap = jiffies +
1321 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1326 * Initialisation. Called after the page allocator have been initialised and
1327 * before smp_init().
1329 void __init kmem_cache_init(void)
1333 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1334 sizeof(struct rcu_head));
1335 kmem_cache = &kmem_cache_boot;
1337 if (num_possible_nodes() == 1)
1338 use_alien_caches = 0;
1340 for (i = 0; i < NUM_INIT_LISTS; i++)
1341 kmem_cache_node_init(&init_kmem_cache_node[i]);
1344 * Fragmentation resistance on low memory - only use bigger
1345 * page orders on machines with more than 32MB of memory if
1346 * not overridden on the command line.
1348 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1349 slab_max_order = SLAB_MAX_ORDER_HI;
1351 /* Bootstrap is tricky, because several objects are allocated
1352 * from caches that do not exist yet:
1353 * 1) initialize the kmem_cache cache: it contains the struct
1354 * kmem_cache structures of all caches, except kmem_cache itself:
1355 * kmem_cache is statically allocated.
1356 * Initially an __init data area is used for the head array and the
1357 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1358 * array at the end of the bootstrap.
1359 * 2) Create the first kmalloc cache.
1360 * The struct kmem_cache for the new cache is allocated normally.
1361 * An __init data area is used for the head array.
1362 * 3) Create the remaining kmalloc caches, with minimally sized
1364 * 4) Replace the __init data head arrays for kmem_cache and the first
1365 * kmalloc cache with kmalloc allocated arrays.
1366 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1367 * the other cache's with kmalloc allocated memory.
1368 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1371 /* 1) create the kmem_cache */
1374 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1376 create_boot_cache(kmem_cache, "kmem_cache",
1377 offsetof(struct kmem_cache, node) +
1378 nr_node_ids * sizeof(struct kmem_cache_node *),
1379 SLAB_HWCACHE_ALIGN);
1380 list_add(&kmem_cache->list, &slab_caches);
1381 slab_state = PARTIAL;
1384 * Initialize the caches that provide memory for the kmem_cache_node
1385 * structures first. Without this, further allocations will bug.
1387 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1388 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1389 slab_state = PARTIAL_NODE;
1390 setup_kmalloc_cache_index_table();
1392 slab_early_init = 0;
1394 /* 5) Replace the bootstrap kmem_cache_node */
1398 for_each_online_node(nid) {
1399 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1401 init_list(kmalloc_caches[INDEX_NODE],
1402 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1406 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1409 void __init kmem_cache_init_late(void)
1411 struct kmem_cache *cachep;
1415 /* 6) resize the head arrays to their final sizes */
1416 mutex_lock(&slab_mutex);
1417 list_for_each_entry(cachep, &slab_caches, list)
1418 if (enable_cpucache(cachep, GFP_NOWAIT))
1420 mutex_unlock(&slab_mutex);
1426 * Register a cpu startup notifier callback that initializes
1427 * cpu_cache_get for all new cpus
1429 register_cpu_notifier(&cpucache_notifier);
1433 * Register a memory hotplug callback that initializes and frees
1436 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1440 * The reap timers are started later, with a module init call: That part
1441 * of the kernel is not yet operational.
1445 static int __init cpucache_init(void)
1450 * Register the timers that return unneeded pages to the page allocator
1452 for_each_online_cpu(cpu)
1453 start_cpu_timer(cpu);
1459 __initcall(cpucache_init);
1461 static noinline void
1462 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1465 struct kmem_cache_node *n;
1467 unsigned long flags;
1469 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1470 DEFAULT_RATELIMIT_BURST);
1472 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1476 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1478 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1479 cachep->name, cachep->size, cachep->gfporder);
1481 for_each_kmem_cache_node(cachep, node, n) {
1482 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1483 unsigned long active_slabs = 0, num_slabs = 0;
1485 spin_lock_irqsave(&n->list_lock, flags);
1486 list_for_each_entry(page, &n->slabs_full, lru) {
1487 active_objs += cachep->num;
1490 list_for_each_entry(page, &n->slabs_partial, lru) {
1491 active_objs += page->active;
1494 list_for_each_entry(page, &n->slabs_free, lru)
1497 free_objects += n->free_objects;
1498 spin_unlock_irqrestore(&n->list_lock, flags);
1500 num_slabs += active_slabs;
1501 num_objs = num_slabs * cachep->num;
1503 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1504 node, active_slabs, num_slabs, active_objs, num_objs,
1511 * Interface to system's page allocator. No need to hold the
1512 * kmem_cache_node ->list_lock.
1514 * If we requested dmaable memory, we will get it. Even if we
1515 * did not request dmaable memory, we might get it, but that
1516 * would be relatively rare and ignorable.
1518 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1524 flags |= cachep->allocflags;
1525 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1526 flags |= __GFP_RECLAIMABLE;
1528 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1530 slab_out_of_memory(cachep, flags, nodeid);
1534 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1535 __free_pages(page, cachep->gfporder);
1539 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1540 if (page_is_pfmemalloc(page))
1541 pfmemalloc_active = true;
1543 nr_pages = (1 << cachep->gfporder);
1544 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1545 add_zone_page_state(page_zone(page),
1546 NR_SLAB_RECLAIMABLE, nr_pages);
1548 add_zone_page_state(page_zone(page),
1549 NR_SLAB_UNRECLAIMABLE, nr_pages);
1550 __SetPageSlab(page);
1551 if (page_is_pfmemalloc(page))
1552 SetPageSlabPfmemalloc(page);
1554 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1555 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1558 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1560 kmemcheck_mark_unallocated_pages(page, nr_pages);
1567 * Interface to system's page release.
1569 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1571 const unsigned long nr_freed = (1 << cachep->gfporder);
1573 kmemcheck_free_shadow(page, cachep->gfporder);
1575 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1576 sub_zone_page_state(page_zone(page),
1577 NR_SLAB_RECLAIMABLE, nr_freed);
1579 sub_zone_page_state(page_zone(page),
1580 NR_SLAB_UNRECLAIMABLE, nr_freed);
1582 BUG_ON(!PageSlab(page));
1583 __ClearPageSlabPfmemalloc(page);
1584 __ClearPageSlab(page);
1585 page_mapcount_reset(page);
1586 page->mapping = NULL;
1588 if (current->reclaim_state)
1589 current->reclaim_state->reclaimed_slab += nr_freed;
1590 __free_kmem_pages(page, cachep->gfporder);
1593 static void kmem_rcu_free(struct rcu_head *head)
1595 struct kmem_cache *cachep;
1598 page = container_of(head, struct page, rcu_head);
1599 cachep = page->slab_cache;
1601 kmem_freepages(cachep, page);
1605 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1607 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1608 (cachep->size % PAGE_SIZE) == 0)
1614 #ifdef CONFIG_DEBUG_PAGEALLOC
1615 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1616 unsigned long caller)
1618 int size = cachep->object_size;
1620 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1622 if (size < 5 * sizeof(unsigned long))
1625 *addr++ = 0x12345678;
1627 *addr++ = smp_processor_id();
1628 size -= 3 * sizeof(unsigned long);
1630 unsigned long *sptr = &caller;
1631 unsigned long svalue;
1633 while (!kstack_end(sptr)) {
1635 if (kernel_text_address(svalue)) {
1637 size -= sizeof(unsigned long);
1638 if (size <= sizeof(unsigned long))
1644 *addr++ = 0x87654321;
1647 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1648 int map, unsigned long caller)
1650 if (!is_debug_pagealloc_cache(cachep))
1654 store_stackinfo(cachep, objp, caller);
1656 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1660 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1661 int map, unsigned long caller) {}
1665 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1667 int size = cachep->object_size;
1668 addr = &((char *)addr)[obj_offset(cachep)];
1670 memset(addr, val, size);
1671 *(unsigned char *)(addr + size - 1) = POISON_END;
1674 static void dump_line(char *data, int offset, int limit)
1677 unsigned char error = 0;
1680 printk(KERN_ERR "%03x: ", offset);
1681 for (i = 0; i < limit; i++) {
1682 if (data[offset + i] != POISON_FREE) {
1683 error = data[offset + i];
1687 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1688 &data[offset], limit, 1);
1690 if (bad_count == 1) {
1691 error ^= POISON_FREE;
1692 if (!(error & (error - 1))) {
1693 printk(KERN_ERR "Single bit error detected. Probably "
1696 printk(KERN_ERR "Run memtest86+ or a similar memory "
1699 printk(KERN_ERR "Run a memory test tool.\n");
1708 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1713 if (cachep->flags & SLAB_RED_ZONE) {
1714 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1715 *dbg_redzone1(cachep, objp),
1716 *dbg_redzone2(cachep, objp));
1719 if (cachep->flags & SLAB_STORE_USER) {
1720 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1721 *dbg_userword(cachep, objp),
1722 *dbg_userword(cachep, objp));
1724 realobj = (char *)objp + obj_offset(cachep);
1725 size = cachep->object_size;
1726 for (i = 0; i < size && lines; i += 16, lines--) {
1729 if (i + limit > size)
1731 dump_line(realobj, i, limit);
1735 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1741 if (is_debug_pagealloc_cache(cachep))
1744 realobj = (char *)objp + obj_offset(cachep);
1745 size = cachep->object_size;
1747 for (i = 0; i < size; i++) {
1748 char exp = POISON_FREE;
1751 if (realobj[i] != exp) {
1757 "Slab corruption (%s): %s start=%p, len=%d\n",
1758 print_tainted(), cachep->name, realobj, size);
1759 print_objinfo(cachep, objp, 0);
1761 /* Hexdump the affected line */
1764 if (i + limit > size)
1766 dump_line(realobj, i, limit);
1769 /* Limit to 5 lines */
1775 /* Print some data about the neighboring objects, if they
1778 struct page *page = virt_to_head_page(objp);
1781 objnr = obj_to_index(cachep, page, objp);
1783 objp = index_to_obj(cachep, page, objnr - 1);
1784 realobj = (char *)objp + obj_offset(cachep);
1785 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1787 print_objinfo(cachep, objp, 2);
1789 if (objnr + 1 < cachep->num) {
1790 objp = index_to_obj(cachep, page, objnr + 1);
1791 realobj = (char *)objp + obj_offset(cachep);
1792 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1794 print_objinfo(cachep, objp, 2);
1801 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1805 for (i = 0; i < cachep->num; i++) {
1806 void *objp = index_to_obj(cachep, page, i);
1808 if (cachep->flags & SLAB_POISON) {
1809 check_poison_obj(cachep, objp);
1810 slab_kernel_map(cachep, objp, 1, 0);
1812 if (cachep->flags & SLAB_RED_ZONE) {
1813 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1814 slab_error(cachep, "start of a freed object "
1816 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1817 slab_error(cachep, "end of a freed object "
1823 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1830 * slab_destroy - destroy and release all objects in a slab
1831 * @cachep: cache pointer being destroyed
1832 * @page: page pointer being destroyed
1834 * Destroy all the objs in a slab page, and release the mem back to the system.
1835 * Before calling the slab page must have been unlinked from the cache. The
1836 * kmem_cache_node ->list_lock is not held/needed.
1838 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1842 freelist = page->freelist;
1843 slab_destroy_debugcheck(cachep, page);
1844 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1845 call_rcu(&page->rcu_head, kmem_rcu_free);
1847 kmem_freepages(cachep, page);
1850 * From now on, we don't use freelist
1851 * although actual page can be freed in rcu context
1853 if (OFF_SLAB(cachep))
1854 kmem_cache_free(cachep->freelist_cache, freelist);
1857 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1859 struct page *page, *n;
1861 list_for_each_entry_safe(page, n, list, lru) {
1862 list_del(&page->lru);
1863 slab_destroy(cachep, page);
1868 * calculate_slab_order - calculate size (page order) of slabs
1869 * @cachep: pointer to the cache that is being created
1870 * @size: size of objects to be created in this cache.
1871 * @flags: slab allocation flags
1873 * Also calculates the number of objects per slab.
1875 * This could be made much more intelligent. For now, try to avoid using
1876 * high order pages for slabs. When the gfp() functions are more friendly
1877 * towards high-order requests, this should be changed.
1879 static size_t calculate_slab_order(struct kmem_cache *cachep,
1880 size_t size, unsigned long flags)
1882 unsigned long offslab_limit;
1883 size_t left_over = 0;
1886 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1890 cache_estimate(gfporder, size, flags, &remainder, &num);
1894 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1895 if (num > SLAB_OBJ_MAX_NUM)
1898 if (flags & CFLGS_OFF_SLAB) {
1900 * Max number of objs-per-slab for caches which
1901 * use off-slab slabs. Needed to avoid a possible
1902 * looping condition in cache_grow().
1904 offslab_limit = size;
1905 offslab_limit /= sizeof(freelist_idx_t);
1907 if (num > offslab_limit)
1911 /* Found something acceptable - save it away */
1913 cachep->gfporder = gfporder;
1914 left_over = remainder;
1917 * A VFS-reclaimable slab tends to have most allocations
1918 * as GFP_NOFS and we really don't want to have to be allocating
1919 * higher-order pages when we are unable to shrink dcache.
1921 if (flags & SLAB_RECLAIM_ACCOUNT)
1925 * Large number of objects is good, but very large slabs are
1926 * currently bad for the gfp()s.
1928 if (gfporder >= slab_max_order)
1932 * Acceptable internal fragmentation?
1934 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1940 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1941 struct kmem_cache *cachep, int entries, int batchcount)
1945 struct array_cache __percpu *cpu_cache;
1947 size = sizeof(void *) * entries + sizeof(struct array_cache);
1948 cpu_cache = __alloc_percpu(size, sizeof(void *));
1953 for_each_possible_cpu(cpu) {
1954 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1955 entries, batchcount);
1961 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1963 if (slab_state >= FULL)
1964 return enable_cpucache(cachep, gfp);
1966 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1967 if (!cachep->cpu_cache)
1970 if (slab_state == DOWN) {
1971 /* Creation of first cache (kmem_cache). */
1972 set_up_node(kmem_cache, CACHE_CACHE);
1973 } else if (slab_state == PARTIAL) {
1974 /* For kmem_cache_node */
1975 set_up_node(cachep, SIZE_NODE);
1979 for_each_online_node(node) {
1980 cachep->node[node] = kmalloc_node(
1981 sizeof(struct kmem_cache_node), gfp, node);
1982 BUG_ON(!cachep->node[node]);
1983 kmem_cache_node_init(cachep->node[node]);
1987 cachep->node[numa_mem_id()]->next_reap =
1988 jiffies + REAPTIMEOUT_NODE +
1989 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1991 cpu_cache_get(cachep)->avail = 0;
1992 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1993 cpu_cache_get(cachep)->batchcount = 1;
1994 cpu_cache_get(cachep)->touched = 0;
1995 cachep->batchcount = 1;
1996 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2000 unsigned long kmem_cache_flags(unsigned long object_size,
2001 unsigned long flags, const char *name,
2002 void (*ctor)(void *))
2008 __kmem_cache_alias(const char *name, size_t size, size_t align,
2009 unsigned long flags, void (*ctor)(void *))
2011 struct kmem_cache *cachep;
2013 cachep = find_mergeable(size, align, flags, name, ctor);
2018 * Adjust the object sizes so that we clear
2019 * the complete object on kzalloc.
2021 cachep->object_size = max_t(int, cachep->object_size, size);
2026 static bool set_off_slab_cache(struct kmem_cache *cachep,
2027 size_t size, unsigned long flags)
2034 * Determine if the slab management is 'on' or 'off' slab.
2035 * (bootstrapping cannot cope with offslab caches so don't do
2036 * it too early on. Always use on-slab management when
2037 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2039 if (size < OFF_SLAB_MIN_SIZE)
2042 if (slab_early_init)
2045 if (flags & SLAB_NOLEAKTRACE)
2049 * Size is large, assume best to place the slab management obj
2050 * off-slab (should allow better packing of objs).
2052 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
2057 * If the slab has been placed off-slab, and we have enough space then
2058 * move it on-slab. This is at the expense of any extra colouring.
2060 if (left >= cachep->num * sizeof(freelist_idx_t))
2063 cachep->colour = left / cachep->colour_off;
2068 static bool set_on_slab_cache(struct kmem_cache *cachep,
2069 size_t size, unsigned long flags)
2075 left = calculate_slab_order(cachep, size, flags);
2079 cachep->colour = left / cachep->colour_off;
2085 * __kmem_cache_create - Create a cache.
2086 * @cachep: cache management descriptor
2087 * @flags: SLAB flags
2089 * Returns a ptr to the cache on success, NULL on failure.
2090 * Cannot be called within a int, but can be interrupted.
2091 * The @ctor is run when new pages are allocated by the cache.
2095 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2096 * to catch references to uninitialised memory.
2098 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2099 * for buffer overruns.
2101 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2102 * cacheline. This can be beneficial if you're counting cycles as closely
2106 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2108 size_t ralign = BYTES_PER_WORD;
2111 size_t size = cachep->size;
2116 * Enable redzoning and last user accounting, except for caches with
2117 * large objects, if the increased size would increase the object size
2118 * above the next power of two: caches with object sizes just above a
2119 * power of two have a significant amount of internal fragmentation.
2121 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2122 2 * sizeof(unsigned long long)))
2123 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2124 if (!(flags & SLAB_DESTROY_BY_RCU))
2125 flags |= SLAB_POISON;
2130 * Check that size is in terms of words. This is needed to avoid
2131 * unaligned accesses for some archs when redzoning is used, and makes
2132 * sure any on-slab bufctl's are also correctly aligned.
2134 if (size & (BYTES_PER_WORD - 1)) {
2135 size += (BYTES_PER_WORD - 1);
2136 size &= ~(BYTES_PER_WORD - 1);
2139 if (flags & SLAB_RED_ZONE) {
2140 ralign = REDZONE_ALIGN;
2141 /* If redzoning, ensure that the second redzone is suitably
2142 * aligned, by adjusting the object size accordingly. */
2143 size += REDZONE_ALIGN - 1;
2144 size &= ~(REDZONE_ALIGN - 1);
2147 /* 3) caller mandated alignment */
2148 if (ralign < cachep->align) {
2149 ralign = cachep->align;
2151 /* disable debug if necessary */
2152 if (ralign > __alignof__(unsigned long long))
2153 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2157 cachep->align = ralign;
2158 cachep->colour_off = cache_line_size();
2159 /* Offset must be a multiple of the alignment. */
2160 if (cachep->colour_off < cachep->align)
2161 cachep->colour_off = cachep->align;
2163 if (slab_is_available())
2171 * Both debugging options require word-alignment which is calculated
2174 if (flags & SLAB_RED_ZONE) {
2175 /* add space for red zone words */
2176 cachep->obj_offset += sizeof(unsigned long long);
2177 size += 2 * sizeof(unsigned long long);
2179 if (flags & SLAB_STORE_USER) {
2180 /* user store requires one word storage behind the end of
2181 * the real object. But if the second red zone needs to be
2182 * aligned to 64 bits, we must allow that much space.
2184 if (flags & SLAB_RED_ZONE)
2185 size += REDZONE_ALIGN;
2187 size += BYTES_PER_WORD;
2191 size = ALIGN(size, cachep->align);
2193 * We should restrict the number of objects in a slab to implement
2194 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2196 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2197 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2201 * To activate debug pagealloc, off-slab management is necessary
2202 * requirement. In early phase of initialization, small sized slab
2203 * doesn't get initialized so it would not be possible. So, we need
2204 * to check size >= 256. It guarantees that all necessary small
2205 * sized slab is initialized in current slab initialization sequence.
2207 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2208 !slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2209 size >= 256 && cachep->object_size > cache_line_size() &&
2211 cachep->obj_offset += PAGE_SIZE - size;
2216 if (set_off_slab_cache(cachep, size, flags)) {
2217 flags |= CFLGS_OFF_SLAB;
2221 if (set_on_slab_cache(cachep, size, flags))
2227 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2228 cachep->flags = flags;
2229 cachep->allocflags = __GFP_COMP;
2230 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2231 cachep->allocflags |= GFP_DMA;
2232 cachep->size = size;
2233 cachep->reciprocal_buffer_size = reciprocal_value(size);
2237 * If we're going to use the generic kernel_map_pages()
2238 * poisoning, then it's going to smash the contents of
2239 * the redzone and userword anyhow, so switch them off.
2241 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2242 (cachep->flags & SLAB_POISON) &&
2243 is_debug_pagealloc_cache(cachep))
2244 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2247 if (OFF_SLAB(cachep)) {
2248 cachep->freelist_cache =
2249 kmalloc_slab(cachep->freelist_size, 0u);
2251 * This is a possibility for one of the kmalloc_{dma,}_caches.
2252 * But since we go off slab only for object size greater than
2253 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
2254 * in ascending order,this should not happen at all.
2255 * But leave a BUG_ON for some lucky dude.
2257 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2260 err = setup_cpu_cache(cachep, gfp);
2262 __kmem_cache_release(cachep);
2270 static void check_irq_off(void)
2272 BUG_ON(!irqs_disabled());
2275 static void check_irq_on(void)
2277 BUG_ON(irqs_disabled());
2280 static void check_spinlock_acquired(struct kmem_cache *cachep)
2284 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2288 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2292 assert_spin_locked(&get_node(cachep, node)->list_lock);
2297 #define check_irq_off() do { } while(0)
2298 #define check_irq_on() do { } while(0)
2299 #define check_spinlock_acquired(x) do { } while(0)
2300 #define check_spinlock_acquired_node(x, y) do { } while(0)
2303 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2304 struct array_cache *ac,
2305 int force, int node);
2307 static void do_drain(void *arg)
2309 struct kmem_cache *cachep = arg;
2310 struct array_cache *ac;
2311 int node = numa_mem_id();
2312 struct kmem_cache_node *n;
2316 ac = cpu_cache_get(cachep);
2317 n = get_node(cachep, node);
2318 spin_lock(&n->list_lock);
2319 free_block(cachep, ac->entry, ac->avail, node, &list);
2320 spin_unlock(&n->list_lock);
2321 slabs_destroy(cachep, &list);
2325 static void drain_cpu_caches(struct kmem_cache *cachep)
2327 struct kmem_cache_node *n;
2330 on_each_cpu(do_drain, cachep, 1);
2332 for_each_kmem_cache_node(cachep, node, n)
2334 drain_alien_cache(cachep, n->alien);
2336 for_each_kmem_cache_node(cachep, node, n)
2337 drain_array(cachep, n, n->shared, 1, node);
2341 * Remove slabs from the list of free slabs.
2342 * Specify the number of slabs to drain in tofree.
2344 * Returns the actual number of slabs released.
2346 static int drain_freelist(struct kmem_cache *cache,
2347 struct kmem_cache_node *n, int tofree)
2349 struct list_head *p;
2354 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2356 spin_lock_irq(&n->list_lock);
2357 p = n->slabs_free.prev;
2358 if (p == &n->slabs_free) {
2359 spin_unlock_irq(&n->list_lock);
2363 page = list_entry(p, struct page, lru);
2364 list_del(&page->lru);
2366 * Safe to drop the lock. The slab is no longer linked
2369 n->free_objects -= cache->num;
2370 spin_unlock_irq(&n->list_lock);
2371 slab_destroy(cache, page);
2378 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2382 struct kmem_cache_node *n;
2384 drain_cpu_caches(cachep);
2387 for_each_kmem_cache_node(cachep, node, n) {
2388 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2390 ret += !list_empty(&n->slabs_full) ||
2391 !list_empty(&n->slabs_partial);
2393 return (ret ? 1 : 0);
2396 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2398 return __kmem_cache_shrink(cachep, false);
2401 void __kmem_cache_release(struct kmem_cache *cachep)
2404 struct kmem_cache_node *n;
2406 free_percpu(cachep->cpu_cache);
2408 /* NUMA: free the node structures */
2409 for_each_kmem_cache_node(cachep, i, n) {
2411 free_alien_cache(n->alien);
2413 cachep->node[i] = NULL;
2418 * Get the memory for a slab management obj.
2420 * For a slab cache when the slab descriptor is off-slab, the
2421 * slab descriptor can't come from the same cache which is being created,
2422 * Because if it is the case, that means we defer the creation of
2423 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2424 * And we eventually call down to __kmem_cache_create(), which
2425 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2426 * This is a "chicken-and-egg" problem.
2428 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2429 * which are all initialized during kmem_cache_init().
2431 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2432 struct page *page, int colour_off,
2433 gfp_t local_flags, int nodeid)
2436 void *addr = page_address(page);
2438 page->s_mem = addr + colour_off;
2441 if (OFF_SLAB(cachep)) {
2442 /* Slab management obj is off-slab. */
2443 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2444 local_flags, nodeid);
2448 /* We will use last bytes at the slab for freelist */
2449 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2450 cachep->freelist_size;
2456 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2458 return ((freelist_idx_t *)page->freelist)[idx];
2461 static inline void set_free_obj(struct page *page,
2462 unsigned int idx, freelist_idx_t val)
2464 ((freelist_idx_t *)(page->freelist))[idx] = val;
2467 static void cache_init_objs(struct kmem_cache *cachep,
2472 for (i = 0; i < cachep->num; i++) {
2473 void *objp = index_to_obj(cachep, page, i);
2475 if (cachep->flags & SLAB_STORE_USER)
2476 *dbg_userword(cachep, objp) = NULL;
2478 if (cachep->flags & SLAB_RED_ZONE) {
2479 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2480 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2483 * Constructors are not allowed to allocate memory from the same
2484 * cache which they are a constructor for. Otherwise, deadlock.
2485 * They must also be threaded.
2487 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2488 cachep->ctor(objp + obj_offset(cachep));
2490 if (cachep->flags & SLAB_RED_ZONE) {
2491 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2492 slab_error(cachep, "constructor overwrote the"
2493 " end of an object");
2494 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2495 slab_error(cachep, "constructor overwrote the"
2496 " start of an object");
2498 /* need to poison the objs? */
2499 if (cachep->flags & SLAB_POISON) {
2500 poison_obj(cachep, objp, POISON_FREE);
2501 slab_kernel_map(cachep, objp, 0, 0);
2507 set_free_obj(page, i, i);
2511 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2513 if (CONFIG_ZONE_DMA_FLAG) {
2514 if (flags & GFP_DMA)
2515 BUG_ON(!(cachep->allocflags & GFP_DMA));
2517 BUG_ON(cachep->allocflags & GFP_DMA);
2521 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2525 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2529 if (cachep->flags & SLAB_STORE_USER)
2530 set_store_user_dirty(cachep);
2536 static void slab_put_obj(struct kmem_cache *cachep,
2537 struct page *page, void *objp)
2539 unsigned int objnr = obj_to_index(cachep, page, objp);
2543 /* Verify double free bug */
2544 for (i = page->active; i < cachep->num; i++) {
2545 if (get_free_obj(page, i) == objnr) {
2546 printk(KERN_ERR "slab: double free detected in cache "
2547 "'%s', objp %p\n", cachep->name, objp);
2553 set_free_obj(page, page->active, objnr);
2557 * Map pages beginning at addr to the given cache and slab. This is required
2558 * for the slab allocator to be able to lookup the cache and slab of a
2559 * virtual address for kfree, ksize, and slab debugging.
2561 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2564 page->slab_cache = cache;
2565 page->freelist = freelist;
2569 * Grow (by 1) the number of slabs within a cache. This is called by
2570 * kmem_cache_alloc() when there are no active objs left in a cache.
2572 static int cache_grow(struct kmem_cache *cachep,
2573 gfp_t flags, int nodeid, struct page *page)
2578 struct kmem_cache_node *n;
2581 * Be lazy and only check for valid flags here, keeping it out of the
2582 * critical path in kmem_cache_alloc().
2584 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2585 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2588 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2590 /* Take the node list lock to change the colour_next on this node */
2592 n = get_node(cachep, nodeid);
2593 spin_lock(&n->list_lock);
2595 /* Get colour for the slab, and cal the next value. */
2596 offset = n->colour_next;
2598 if (n->colour_next >= cachep->colour)
2600 spin_unlock(&n->list_lock);
2602 offset *= cachep->colour_off;
2604 if (gfpflags_allow_blocking(local_flags))
2608 * The test for missing atomic flag is performed here, rather than
2609 * the more obvious place, simply to reduce the critical path length
2610 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2611 * will eventually be caught here (where it matters).
2613 kmem_flagcheck(cachep, flags);
2616 * Get mem for the objs. Attempt to allocate a physical page from
2620 page = kmem_getpages(cachep, local_flags, nodeid);
2624 /* Get slab management. */
2625 freelist = alloc_slabmgmt(cachep, page, offset,
2626 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2630 slab_map_pages(cachep, page, freelist);
2632 cache_init_objs(cachep, page);
2634 if (gfpflags_allow_blocking(local_flags))
2635 local_irq_disable();
2637 spin_lock(&n->list_lock);
2639 /* Make slab active. */
2640 list_add_tail(&page->lru, &(n->slabs_free));
2641 STATS_INC_GROWN(cachep);
2642 n->free_objects += cachep->num;
2643 spin_unlock(&n->list_lock);
2646 kmem_freepages(cachep, page);
2648 if (gfpflags_allow_blocking(local_flags))
2649 local_irq_disable();
2656 * Perform extra freeing checks:
2657 * - detect bad pointers.
2658 * - POISON/RED_ZONE checking
2660 static void kfree_debugcheck(const void *objp)
2662 if (!virt_addr_valid(objp)) {
2663 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2664 (unsigned long)objp);
2669 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2671 unsigned long long redzone1, redzone2;
2673 redzone1 = *dbg_redzone1(cache, obj);
2674 redzone2 = *dbg_redzone2(cache, obj);
2679 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2682 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2683 slab_error(cache, "double free detected");
2685 slab_error(cache, "memory outside object was overwritten");
2687 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2688 obj, redzone1, redzone2);
2691 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2692 unsigned long caller)
2697 BUG_ON(virt_to_cache(objp) != cachep);
2699 objp -= obj_offset(cachep);
2700 kfree_debugcheck(objp);
2701 page = virt_to_head_page(objp);
2703 if (cachep->flags & SLAB_RED_ZONE) {
2704 verify_redzone_free(cachep, objp);
2705 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2706 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2708 if (cachep->flags & SLAB_STORE_USER) {
2709 set_store_user_dirty(cachep);
2710 *dbg_userword(cachep, objp) = (void *)caller;
2713 objnr = obj_to_index(cachep, page, objp);
2715 BUG_ON(objnr >= cachep->num);
2716 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2718 if (cachep->flags & SLAB_POISON) {
2719 poison_obj(cachep, objp, POISON_FREE);
2720 slab_kernel_map(cachep, objp, 0, caller);
2726 #define kfree_debugcheck(x) do { } while(0)
2727 #define cache_free_debugcheck(x,objp,z) (objp)
2730 static struct page *get_first_slab(struct kmem_cache_node *n)
2734 page = list_first_entry_or_null(&n->slabs_partial,
2737 n->free_touched = 1;
2738 page = list_first_entry_or_null(&n->slabs_free,
2745 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2749 struct kmem_cache_node *n;
2750 struct array_cache *ac;
2754 node = numa_mem_id();
2755 if (unlikely(force_refill))
2758 ac = cpu_cache_get(cachep);
2759 batchcount = ac->batchcount;
2760 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2762 * If there was little recent activity on this cache, then
2763 * perform only a partial refill. Otherwise we could generate
2766 batchcount = BATCHREFILL_LIMIT;
2768 n = get_node(cachep, node);
2770 BUG_ON(ac->avail > 0 || !n);
2771 spin_lock(&n->list_lock);
2773 /* See if we can refill from the shared array */
2774 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2775 n->shared->touched = 1;
2779 while (batchcount > 0) {
2781 /* Get slab alloc is to come from. */
2782 page = get_first_slab(n);
2786 check_spinlock_acquired(cachep);
2789 * The slab was either on partial or free list so
2790 * there must be at least one object available for
2793 BUG_ON(page->active >= cachep->num);
2795 while (page->active < cachep->num && batchcount--) {
2796 STATS_INC_ALLOCED(cachep);
2797 STATS_INC_ACTIVE(cachep);
2798 STATS_SET_HIGH(cachep);
2800 ac_put_obj(cachep, ac, slab_get_obj(cachep, page));
2803 /* move slabp to correct slabp list: */
2804 list_del(&page->lru);
2805 if (page->active == cachep->num)
2806 list_add(&page->lru, &n->slabs_full);
2808 list_add(&page->lru, &n->slabs_partial);
2812 n->free_objects -= ac->avail;
2814 spin_unlock(&n->list_lock);
2816 if (unlikely(!ac->avail)) {
2819 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2821 /* cache_grow can reenable interrupts, then ac could change. */
2822 ac = cpu_cache_get(cachep);
2823 node = numa_mem_id();
2825 /* no objects in sight? abort */
2826 if (!x && (ac->avail == 0 || force_refill))
2829 if (!ac->avail) /* objects refilled by interrupt? */
2834 return ac_get_obj(cachep, ac, flags, force_refill);
2837 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2840 might_sleep_if(gfpflags_allow_blocking(flags));
2842 kmem_flagcheck(cachep, flags);
2847 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2848 gfp_t flags, void *objp, unsigned long caller)
2852 if (cachep->flags & SLAB_POISON) {
2853 check_poison_obj(cachep, objp);
2854 slab_kernel_map(cachep, objp, 1, 0);
2855 poison_obj(cachep, objp, POISON_INUSE);
2857 if (cachep->flags & SLAB_STORE_USER)
2858 *dbg_userword(cachep, objp) = (void *)caller;
2860 if (cachep->flags & SLAB_RED_ZONE) {
2861 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2862 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2863 slab_error(cachep, "double free, or memory outside"
2864 " object was overwritten");
2866 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2867 objp, *dbg_redzone1(cachep, objp),
2868 *dbg_redzone2(cachep, objp));
2870 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2871 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2874 objp += obj_offset(cachep);
2875 if (cachep->ctor && cachep->flags & SLAB_POISON)
2877 if (ARCH_SLAB_MINALIGN &&
2878 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2879 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2880 objp, (int)ARCH_SLAB_MINALIGN);
2885 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2888 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2891 struct array_cache *ac;
2892 bool force_refill = false;
2896 ac = cpu_cache_get(cachep);
2897 if (likely(ac->avail)) {
2899 objp = ac_get_obj(cachep, ac, flags, false);
2902 * Allow for the possibility all avail objects are not allowed
2903 * by the current flags
2906 STATS_INC_ALLOCHIT(cachep);
2909 force_refill = true;
2912 STATS_INC_ALLOCMISS(cachep);
2913 objp = cache_alloc_refill(cachep, flags, force_refill);
2915 * the 'ac' may be updated by cache_alloc_refill(),
2916 * and kmemleak_erase() requires its correct value.
2918 ac = cpu_cache_get(cachep);
2922 * To avoid a false negative, if an object that is in one of the
2923 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2924 * treat the array pointers as a reference to the object.
2927 kmemleak_erase(&ac->entry[ac->avail]);
2933 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2935 * If we are in_interrupt, then process context, including cpusets and
2936 * mempolicy, may not apply and should not be used for allocation policy.
2938 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2940 int nid_alloc, nid_here;
2942 if (in_interrupt() || (flags & __GFP_THISNODE))
2944 nid_alloc = nid_here = numa_mem_id();
2945 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2946 nid_alloc = cpuset_slab_spread_node();
2947 else if (current->mempolicy)
2948 nid_alloc = mempolicy_slab_node();
2949 if (nid_alloc != nid_here)
2950 return ____cache_alloc_node(cachep, flags, nid_alloc);
2955 * Fallback function if there was no memory available and no objects on a
2956 * certain node and fall back is permitted. First we scan all the
2957 * available node for available objects. If that fails then we
2958 * perform an allocation without specifying a node. This allows the page
2959 * allocator to do its reclaim / fallback magic. We then insert the
2960 * slab into the proper nodelist and then allocate from it.
2962 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2964 struct zonelist *zonelist;
2968 enum zone_type high_zoneidx = gfp_zone(flags);
2971 unsigned int cpuset_mems_cookie;
2973 if (flags & __GFP_THISNODE)
2976 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2979 cpuset_mems_cookie = read_mems_allowed_begin();
2980 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2984 * Look through allowed nodes for objects available
2985 * from existing per node queues.
2987 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2988 nid = zone_to_nid(zone);
2990 if (cpuset_zone_allowed(zone, flags) &&
2991 get_node(cache, nid) &&
2992 get_node(cache, nid)->free_objects) {
2993 obj = ____cache_alloc_node(cache,
2994 gfp_exact_node(flags), nid);
3002 * This allocation will be performed within the constraints
3003 * of the current cpuset / memory policy requirements.
3004 * We may trigger various forms of reclaim on the allowed
3005 * set and go into memory reserves if necessary.
3009 if (gfpflags_allow_blocking(local_flags))
3011 kmem_flagcheck(cache, flags);
3012 page = kmem_getpages(cache, local_flags, numa_mem_id());
3013 if (gfpflags_allow_blocking(local_flags))
3014 local_irq_disable();
3017 * Insert into the appropriate per node queues
3019 nid = page_to_nid(page);
3020 if (cache_grow(cache, flags, nid, page)) {
3021 obj = ____cache_alloc_node(cache,
3022 gfp_exact_node(flags), nid);
3025 * Another processor may allocate the
3026 * objects in the slab since we are
3027 * not holding any locks.
3031 /* cache_grow already freed obj */
3037 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3043 * A interface to enable slab creation on nodeid
3045 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3049 struct kmem_cache_node *n;
3053 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3054 n = get_node(cachep, nodeid);
3059 spin_lock(&n->list_lock);
3060 page = get_first_slab(n);
3064 check_spinlock_acquired_node(cachep, nodeid);
3066 STATS_INC_NODEALLOCS(cachep);
3067 STATS_INC_ACTIVE(cachep);
3068 STATS_SET_HIGH(cachep);
3070 BUG_ON(page->active == cachep->num);
3072 obj = slab_get_obj(cachep, page);
3074 /* move slabp to correct slabp list: */
3075 list_del(&page->lru);
3077 if (page->active == cachep->num)
3078 list_add(&page->lru, &n->slabs_full);
3080 list_add(&page->lru, &n->slabs_partial);
3082 spin_unlock(&n->list_lock);
3086 spin_unlock(&n->list_lock);
3087 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3091 return fallback_alloc(cachep, flags);
3097 static __always_inline void *
3098 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3099 unsigned long caller)
3101 unsigned long save_flags;
3103 int slab_node = numa_mem_id();
3105 flags &= gfp_allowed_mask;
3106 cachep = slab_pre_alloc_hook(cachep, flags);
3107 if (unlikely(!cachep))
3110 cache_alloc_debugcheck_before(cachep, flags);
3111 local_irq_save(save_flags);
3113 if (nodeid == NUMA_NO_NODE)
3116 if (unlikely(!get_node(cachep, nodeid))) {
3117 /* Node not bootstrapped yet */
3118 ptr = fallback_alloc(cachep, flags);
3122 if (nodeid == slab_node) {
3124 * Use the locally cached objects if possible.
3125 * However ____cache_alloc does not allow fallback
3126 * to other nodes. It may fail while we still have
3127 * objects on other nodes available.
3129 ptr = ____cache_alloc(cachep, flags);
3133 /* ___cache_alloc_node can fall back to other nodes */
3134 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3136 local_irq_restore(save_flags);
3137 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3139 if (unlikely(flags & __GFP_ZERO) && ptr)
3140 memset(ptr, 0, cachep->object_size);
3142 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3146 static __always_inline void *
3147 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3151 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3152 objp = alternate_node_alloc(cache, flags);
3156 objp = ____cache_alloc(cache, flags);
3159 * We may just have run out of memory on the local node.
3160 * ____cache_alloc_node() knows how to locate memory on other nodes
3163 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3170 static __always_inline void *
3171 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3173 return ____cache_alloc(cachep, flags);
3176 #endif /* CONFIG_NUMA */
3178 static __always_inline void *
3179 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3181 unsigned long save_flags;
3184 flags &= gfp_allowed_mask;
3185 cachep = slab_pre_alloc_hook(cachep, flags);
3186 if (unlikely(!cachep))
3189 cache_alloc_debugcheck_before(cachep, flags);
3190 local_irq_save(save_flags);
3191 objp = __do_cache_alloc(cachep, flags);
3192 local_irq_restore(save_flags);
3193 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3196 if (unlikely(flags & __GFP_ZERO) && objp)
3197 memset(objp, 0, cachep->object_size);
3199 slab_post_alloc_hook(cachep, flags, 1, &objp);
3204 * Caller needs to acquire correct kmem_cache_node's list_lock
3205 * @list: List of detached free slabs should be freed by caller
3207 static void free_block(struct kmem_cache *cachep, void **objpp,
3208 int nr_objects, int node, struct list_head *list)
3211 struct kmem_cache_node *n = get_node(cachep, node);
3213 for (i = 0; i < nr_objects; i++) {
3217 clear_obj_pfmemalloc(&objpp[i]);
3220 page = virt_to_head_page(objp);
3221 list_del(&page->lru);
3222 check_spinlock_acquired_node(cachep, node);
3223 slab_put_obj(cachep, page, objp);
3224 STATS_DEC_ACTIVE(cachep);
3227 /* fixup slab chains */
3228 if (page->active == 0) {
3229 if (n->free_objects > n->free_limit) {
3230 n->free_objects -= cachep->num;
3231 list_add_tail(&page->lru, list);
3233 list_add(&page->lru, &n->slabs_free);
3236 /* Unconditionally move a slab to the end of the
3237 * partial list on free - maximum time for the
3238 * other objects to be freed, too.
3240 list_add_tail(&page->lru, &n->slabs_partial);
3245 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3248 struct kmem_cache_node *n;
3249 int node = numa_mem_id();
3252 batchcount = ac->batchcount;
3255 n = get_node(cachep, node);
3256 spin_lock(&n->list_lock);
3258 struct array_cache *shared_array = n->shared;
3259 int max = shared_array->limit - shared_array->avail;
3261 if (batchcount > max)
3263 memcpy(&(shared_array->entry[shared_array->avail]),
3264 ac->entry, sizeof(void *) * batchcount);
3265 shared_array->avail += batchcount;
3270 free_block(cachep, ac->entry, batchcount, node, &list);
3277 list_for_each_entry(page, &n->slabs_free, lru) {
3278 BUG_ON(page->active);
3282 STATS_SET_FREEABLE(cachep, i);
3285 spin_unlock(&n->list_lock);
3286 slabs_destroy(cachep, &list);
3287 ac->avail -= batchcount;
3288 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3292 * Release an obj back to its cache. If the obj has a constructed state, it must
3293 * be in this state _before_ it is released. Called with disabled ints.
3295 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3296 unsigned long caller)
3298 struct array_cache *ac = cpu_cache_get(cachep);
3301 kmemleak_free_recursive(objp, cachep->flags);
3302 objp = cache_free_debugcheck(cachep, objp, caller);
3304 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3307 * Skip calling cache_free_alien() when the platform is not numa.
3308 * This will avoid cache misses that happen while accessing slabp (which
3309 * is per page memory reference) to get nodeid. Instead use a global
3310 * variable to skip the call, which is mostly likely to be present in
3313 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3316 if (ac->avail < ac->limit) {
3317 STATS_INC_FREEHIT(cachep);
3319 STATS_INC_FREEMISS(cachep);
3320 cache_flusharray(cachep, ac);
3323 ac_put_obj(cachep, ac, objp);
3327 * kmem_cache_alloc - Allocate an object
3328 * @cachep: The cache to allocate from.
3329 * @flags: See kmalloc().
3331 * Allocate an object from this cache. The flags are only relevant
3332 * if the cache has no available objects.
3334 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3336 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3338 trace_kmem_cache_alloc(_RET_IP_, ret,
3339 cachep->object_size, cachep->size, flags);
3343 EXPORT_SYMBOL(kmem_cache_alloc);
3345 static __always_inline void
3346 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3347 size_t size, void **p, unsigned long caller)
3351 for (i = 0; i < size; i++)
3352 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3355 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3360 s = slab_pre_alloc_hook(s, flags);
3364 cache_alloc_debugcheck_before(s, flags);
3366 local_irq_disable();
3367 for (i = 0; i < size; i++) {
3368 void *objp = __do_cache_alloc(s, flags);
3370 if (unlikely(!objp))
3376 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3378 /* Clear memory outside IRQ disabled section */
3379 if (unlikely(flags & __GFP_ZERO))
3380 for (i = 0; i < size; i++)
3381 memset(p[i], 0, s->object_size);
3383 slab_post_alloc_hook(s, flags, size, p);
3384 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3388 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3389 slab_post_alloc_hook(s, flags, i, p);
3390 __kmem_cache_free_bulk(s, i, p);
3393 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3395 #ifdef CONFIG_TRACING
3397 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3401 ret = slab_alloc(cachep, flags, _RET_IP_);
3403 trace_kmalloc(_RET_IP_, ret,
3404 size, cachep->size, flags);
3407 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3412 * kmem_cache_alloc_node - Allocate an object on the specified node
3413 * @cachep: The cache to allocate from.
3414 * @flags: See kmalloc().
3415 * @nodeid: node number of the target node.
3417 * Identical to kmem_cache_alloc but it will allocate memory on the given
3418 * node, which can improve the performance for cpu bound structures.
3420 * Fallback to other node is possible if __GFP_THISNODE is not set.
3422 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3424 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3426 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3427 cachep->object_size, cachep->size,
3432 EXPORT_SYMBOL(kmem_cache_alloc_node);
3434 #ifdef CONFIG_TRACING
3435 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3442 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3444 trace_kmalloc_node(_RET_IP_, ret,
3449 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3452 static __always_inline void *
3453 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3455 struct kmem_cache *cachep;
3457 cachep = kmalloc_slab(size, flags);
3458 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3460 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3463 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3465 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3467 EXPORT_SYMBOL(__kmalloc_node);
3469 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3470 int node, unsigned long caller)
3472 return __do_kmalloc_node(size, flags, node, caller);
3474 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3475 #endif /* CONFIG_NUMA */
3478 * __do_kmalloc - allocate memory
3479 * @size: how many bytes of memory are required.
3480 * @flags: the type of memory to allocate (see kmalloc).
3481 * @caller: function caller for debug tracking of the caller
3483 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3484 unsigned long caller)
3486 struct kmem_cache *cachep;
3489 cachep = kmalloc_slab(size, flags);
3490 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3492 ret = slab_alloc(cachep, flags, caller);
3494 trace_kmalloc(caller, ret,
3495 size, cachep->size, flags);
3500 void *__kmalloc(size_t size, gfp_t flags)
3502 return __do_kmalloc(size, flags, _RET_IP_);
3504 EXPORT_SYMBOL(__kmalloc);
3506 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3508 return __do_kmalloc(size, flags, caller);
3510 EXPORT_SYMBOL(__kmalloc_track_caller);
3513 * kmem_cache_free - Deallocate an object
3514 * @cachep: The cache the allocation was from.
3515 * @objp: The previously allocated object.
3517 * Free an object which was previously allocated from this
3520 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3522 unsigned long flags;
3523 cachep = cache_from_obj(cachep, objp);
3527 local_irq_save(flags);
3528 debug_check_no_locks_freed(objp, cachep->object_size);
3529 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3530 debug_check_no_obj_freed(objp, cachep->object_size);
3531 __cache_free(cachep, objp, _RET_IP_);
3532 local_irq_restore(flags);
3534 trace_kmem_cache_free(_RET_IP_, objp);
3536 EXPORT_SYMBOL(kmem_cache_free);
3538 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3540 struct kmem_cache *s;
3543 local_irq_disable();
3544 for (i = 0; i < size; i++) {
3547 if (!orig_s) /* called via kfree_bulk */
3548 s = virt_to_cache(objp);
3550 s = cache_from_obj(orig_s, objp);
3552 debug_check_no_locks_freed(objp, s->object_size);
3553 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3554 debug_check_no_obj_freed(objp, s->object_size);
3556 __cache_free(s, objp, _RET_IP_);
3560 /* FIXME: add tracing */
3562 EXPORT_SYMBOL(kmem_cache_free_bulk);
3565 * kfree - free previously allocated memory
3566 * @objp: pointer returned by kmalloc.
3568 * If @objp is NULL, no operation is performed.
3570 * Don't free memory not originally allocated by kmalloc()
3571 * or you will run into trouble.
3573 void kfree(const void *objp)
3575 struct kmem_cache *c;
3576 unsigned long flags;
3578 trace_kfree(_RET_IP_, objp);
3580 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3582 local_irq_save(flags);
3583 kfree_debugcheck(objp);
3584 c = virt_to_cache(objp);
3585 debug_check_no_locks_freed(objp, c->object_size);
3587 debug_check_no_obj_freed(objp, c->object_size);
3588 __cache_free(c, (void *)objp, _RET_IP_);
3589 local_irq_restore(flags);
3591 EXPORT_SYMBOL(kfree);
3594 * This initializes kmem_cache_node or resizes various caches for all nodes.
3596 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3599 struct kmem_cache_node *n;
3600 struct array_cache *new_shared;
3601 struct alien_cache **new_alien = NULL;
3603 for_each_online_node(node) {
3605 if (use_alien_caches) {
3606 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3612 if (cachep->shared) {
3613 new_shared = alloc_arraycache(node,
3614 cachep->shared*cachep->batchcount,
3617 free_alien_cache(new_alien);
3622 n = get_node(cachep, node);
3624 struct array_cache *shared = n->shared;
3627 spin_lock_irq(&n->list_lock);
3630 free_block(cachep, shared->entry,
3631 shared->avail, node, &list);
3633 n->shared = new_shared;
3635 n->alien = new_alien;
3638 n->free_limit = (1 + nr_cpus_node(node)) *
3639 cachep->batchcount + cachep->num;
3640 spin_unlock_irq(&n->list_lock);
3641 slabs_destroy(cachep, &list);
3643 free_alien_cache(new_alien);
3646 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3648 free_alien_cache(new_alien);
3653 kmem_cache_node_init(n);
3654 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3655 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3656 n->shared = new_shared;
3657 n->alien = new_alien;
3658 n->free_limit = (1 + nr_cpus_node(node)) *
3659 cachep->batchcount + cachep->num;
3660 cachep->node[node] = n;
3665 if (!cachep->list.next) {
3666 /* Cache is not active yet. Roll back what we did */
3669 n = get_node(cachep, node);
3672 free_alien_cache(n->alien);
3674 cachep->node[node] = NULL;
3682 /* Always called with the slab_mutex held */
3683 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3684 int batchcount, int shared, gfp_t gfp)
3686 struct array_cache __percpu *cpu_cache, *prev;
3689 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3693 prev = cachep->cpu_cache;
3694 cachep->cpu_cache = cpu_cache;
3695 kick_all_cpus_sync();
3698 cachep->batchcount = batchcount;
3699 cachep->limit = limit;
3700 cachep->shared = shared;
3705 for_each_online_cpu(cpu) {
3708 struct kmem_cache_node *n;
3709 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3711 node = cpu_to_mem(cpu);
3712 n = get_node(cachep, node);
3713 spin_lock_irq(&n->list_lock);
3714 free_block(cachep, ac->entry, ac->avail, node, &list);
3715 spin_unlock_irq(&n->list_lock);
3716 slabs_destroy(cachep, &list);
3721 return alloc_kmem_cache_node(cachep, gfp);
3724 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3725 int batchcount, int shared, gfp_t gfp)
3728 struct kmem_cache *c;
3730 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3732 if (slab_state < FULL)
3735 if ((ret < 0) || !is_root_cache(cachep))
3738 lockdep_assert_held(&slab_mutex);
3739 for_each_memcg_cache(c, cachep) {
3740 /* return value determined by the root cache only */
3741 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3747 /* Called with slab_mutex held always */
3748 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3755 if (!is_root_cache(cachep)) {
3756 struct kmem_cache *root = memcg_root_cache(cachep);
3757 limit = root->limit;
3758 shared = root->shared;
3759 batchcount = root->batchcount;
3762 if (limit && shared && batchcount)
3765 * The head array serves three purposes:
3766 * - create a LIFO ordering, i.e. return objects that are cache-warm
3767 * - reduce the number of spinlock operations.
3768 * - reduce the number of linked list operations on the slab and
3769 * bufctl chains: array operations are cheaper.
3770 * The numbers are guessed, we should auto-tune as described by
3773 if (cachep->size > 131072)
3775 else if (cachep->size > PAGE_SIZE)
3777 else if (cachep->size > 1024)
3779 else if (cachep->size > 256)
3785 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3786 * allocation behaviour: Most allocs on one cpu, most free operations
3787 * on another cpu. For these cases, an efficient object passing between
3788 * cpus is necessary. This is provided by a shared array. The array
3789 * replaces Bonwick's magazine layer.
3790 * On uniprocessor, it's functionally equivalent (but less efficient)
3791 * to a larger limit. Thus disabled by default.
3794 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3799 * With debugging enabled, large batchcount lead to excessively long
3800 * periods with disabled local interrupts. Limit the batchcount
3805 batchcount = (limit + 1) / 2;
3807 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3809 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3810 cachep->name, -err);
3815 * Drain an array if it contains any elements taking the node lock only if
3816 * necessary. Note that the node listlock also protects the array_cache
3817 * if drain_array() is used on the shared array.
3819 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3820 struct array_cache *ac, int force, int node)
3825 if (!ac || !ac->avail)
3827 if (ac->touched && !force) {
3830 spin_lock_irq(&n->list_lock);
3832 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3833 if (tofree > ac->avail)
3834 tofree = (ac->avail + 1) / 2;
3835 free_block(cachep, ac->entry, tofree, node, &list);
3836 ac->avail -= tofree;
3837 memmove(ac->entry, &(ac->entry[tofree]),
3838 sizeof(void *) * ac->avail);
3840 spin_unlock_irq(&n->list_lock);
3841 slabs_destroy(cachep, &list);
3846 * cache_reap - Reclaim memory from caches.
3847 * @w: work descriptor
3849 * Called from workqueue/eventd every few seconds.
3851 * - clear the per-cpu caches for this CPU.
3852 * - return freeable pages to the main free memory pool.
3854 * If we cannot acquire the cache chain mutex then just give up - we'll try
3855 * again on the next iteration.
3857 static void cache_reap(struct work_struct *w)
3859 struct kmem_cache *searchp;
3860 struct kmem_cache_node *n;
3861 int node = numa_mem_id();
3862 struct delayed_work *work = to_delayed_work(w);
3864 if (!mutex_trylock(&slab_mutex))
3865 /* Give up. Setup the next iteration. */
3868 list_for_each_entry(searchp, &slab_caches, list) {
3872 * We only take the node lock if absolutely necessary and we
3873 * have established with reasonable certainty that
3874 * we can do some work if the lock was obtained.
3876 n = get_node(searchp, node);
3878 reap_alien(searchp, n);
3880 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3883 * These are racy checks but it does not matter
3884 * if we skip one check or scan twice.
3886 if (time_after(n->next_reap, jiffies))
3889 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3891 drain_array(searchp, n, n->shared, 0, node);
3893 if (n->free_touched)
3894 n->free_touched = 0;
3898 freed = drain_freelist(searchp, n, (n->free_limit +
3899 5 * searchp->num - 1) / (5 * searchp->num));
3900 STATS_ADD_REAPED(searchp, freed);
3906 mutex_unlock(&slab_mutex);
3909 /* Set up the next iteration */
3910 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3913 #ifdef CONFIG_SLABINFO
3914 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3917 unsigned long active_objs;
3918 unsigned long num_objs;
3919 unsigned long active_slabs = 0;
3920 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3924 struct kmem_cache_node *n;
3928 for_each_kmem_cache_node(cachep, node, n) {
3931 spin_lock_irq(&n->list_lock);
3933 list_for_each_entry(page, &n->slabs_full, lru) {
3934 if (page->active != cachep->num && !error)
3935 error = "slabs_full accounting error";
3936 active_objs += cachep->num;
3939 list_for_each_entry(page, &n->slabs_partial, lru) {
3940 if (page->active == cachep->num && !error)
3941 error = "slabs_partial accounting error";
3942 if (!page->active && !error)
3943 error = "slabs_partial accounting error";
3944 active_objs += page->active;
3947 list_for_each_entry(page, &n->slabs_free, lru) {
3948 if (page->active && !error)
3949 error = "slabs_free accounting error";
3952 free_objects += n->free_objects;
3954 shared_avail += n->shared->avail;
3956 spin_unlock_irq(&n->list_lock);
3958 num_slabs += active_slabs;
3959 num_objs = num_slabs * cachep->num;
3960 if (num_objs - active_objs != free_objects && !error)
3961 error = "free_objects accounting error";
3963 name = cachep->name;
3965 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3967 sinfo->active_objs = active_objs;
3968 sinfo->num_objs = num_objs;
3969 sinfo->active_slabs = active_slabs;
3970 sinfo->num_slabs = num_slabs;
3971 sinfo->shared_avail = shared_avail;
3972 sinfo->limit = cachep->limit;
3973 sinfo->batchcount = cachep->batchcount;
3974 sinfo->shared = cachep->shared;
3975 sinfo->objects_per_slab = cachep->num;
3976 sinfo->cache_order = cachep->gfporder;
3979 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3983 unsigned long high = cachep->high_mark;
3984 unsigned long allocs = cachep->num_allocations;
3985 unsigned long grown = cachep->grown;
3986 unsigned long reaped = cachep->reaped;
3987 unsigned long errors = cachep->errors;
3988 unsigned long max_freeable = cachep->max_freeable;
3989 unsigned long node_allocs = cachep->node_allocs;
3990 unsigned long node_frees = cachep->node_frees;
3991 unsigned long overflows = cachep->node_overflow;
3993 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
3994 "%4lu %4lu %4lu %4lu %4lu",
3995 allocs, high, grown,
3996 reaped, errors, max_freeable, node_allocs,
3997 node_frees, overflows);
4001 unsigned long allochit = atomic_read(&cachep->allochit);
4002 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4003 unsigned long freehit = atomic_read(&cachep->freehit);
4004 unsigned long freemiss = atomic_read(&cachep->freemiss);
4006 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4007 allochit, allocmiss, freehit, freemiss);
4012 #define MAX_SLABINFO_WRITE 128
4014 * slabinfo_write - Tuning for the slab allocator
4016 * @buffer: user buffer
4017 * @count: data length
4020 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4021 size_t count, loff_t *ppos)
4023 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4024 int limit, batchcount, shared, res;
4025 struct kmem_cache *cachep;
4027 if (count > MAX_SLABINFO_WRITE)
4029 if (copy_from_user(&kbuf, buffer, count))
4031 kbuf[MAX_SLABINFO_WRITE] = '\0';
4033 tmp = strchr(kbuf, ' ');
4038 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4041 /* Find the cache in the chain of caches. */
4042 mutex_lock(&slab_mutex);
4044 list_for_each_entry(cachep, &slab_caches, list) {
4045 if (!strcmp(cachep->name, kbuf)) {
4046 if (limit < 1 || batchcount < 1 ||
4047 batchcount > limit || shared < 0) {
4050 res = do_tune_cpucache(cachep, limit,
4057 mutex_unlock(&slab_mutex);
4063 #ifdef CONFIG_DEBUG_SLAB_LEAK
4065 static inline int add_caller(unsigned long *n, unsigned long v)
4075 unsigned long *q = p + 2 * i;
4089 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4095 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4104 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4107 for (j = page->active; j < c->num; j++) {
4108 if (get_free_obj(page, j) == i) {
4118 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4119 * mapping is established when actual object allocation and
4120 * we could mistakenly access the unmapped object in the cpu
4123 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4126 if (!add_caller(n, v))
4131 static void show_symbol(struct seq_file *m, unsigned long address)
4133 #ifdef CONFIG_KALLSYMS
4134 unsigned long offset, size;
4135 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4137 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4138 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4140 seq_printf(m, " [%s]", modname);
4144 seq_printf(m, "%p", (void *)address);
4147 static int leaks_show(struct seq_file *m, void *p)
4149 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4151 struct kmem_cache_node *n;
4153 unsigned long *x = m->private;
4157 if (!(cachep->flags & SLAB_STORE_USER))
4159 if (!(cachep->flags & SLAB_RED_ZONE))
4163 * Set store_user_clean and start to grab stored user information
4164 * for all objects on this cache. If some alloc/free requests comes
4165 * during the processing, information would be wrong so restart
4169 set_store_user_clean(cachep);
4170 drain_cpu_caches(cachep);
4174 for_each_kmem_cache_node(cachep, node, n) {
4177 spin_lock_irq(&n->list_lock);
4179 list_for_each_entry(page, &n->slabs_full, lru)
4180 handle_slab(x, cachep, page);
4181 list_for_each_entry(page, &n->slabs_partial, lru)
4182 handle_slab(x, cachep, page);
4183 spin_unlock_irq(&n->list_lock);
4185 } while (!is_store_user_clean(cachep));
4187 name = cachep->name;
4189 /* Increase the buffer size */
4190 mutex_unlock(&slab_mutex);
4191 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4193 /* Too bad, we are really out */
4195 mutex_lock(&slab_mutex);
4198 *(unsigned long *)m->private = x[0] * 2;
4200 mutex_lock(&slab_mutex);
4201 /* Now make sure this entry will be retried */
4205 for (i = 0; i < x[1]; i++) {
4206 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4207 show_symbol(m, x[2*i+2]);
4214 static const struct seq_operations slabstats_op = {
4215 .start = slab_start,
4221 static int slabstats_open(struct inode *inode, struct file *file)
4225 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4229 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4234 static const struct file_operations proc_slabstats_operations = {
4235 .open = slabstats_open,
4237 .llseek = seq_lseek,
4238 .release = seq_release_private,
4242 static int __init slab_proc_init(void)
4244 #ifdef CONFIG_DEBUG_SLAB_LEAK
4245 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4249 module_init(slab_proc_init);
4253 * ksize - get the actual amount of memory allocated for a given object
4254 * @objp: Pointer to the object
4256 * kmalloc may internally round up allocations and return more memory
4257 * than requested. ksize() can be used to determine the actual amount of
4258 * memory allocated. The caller may use this additional memory, even though
4259 * a smaller amount of memory was initially specified with the kmalloc call.
4260 * The caller must guarantee that objp points to a valid object previously
4261 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4262 * must not be freed during the duration of the call.
4264 size_t ksize(const void *objp)
4267 if (unlikely(objp == ZERO_SIZE_PTR))
4270 return virt_to_cache(objp)->object_size;
4272 EXPORT_SYMBOL(ksize);