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_OBJFREELIST_SLAB (0x40000000UL)
274 #define CFLGS_OFF_SLAB (0x80000000UL)
275 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
276 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
278 #define BATCHREFILL_LIMIT 16
280 * Optimization question: fewer reaps means less probability for unnessary
281 * cpucache drain/refill cycles.
283 * OTOH the cpuarrays can contain lots of objects,
284 * which could lock up otherwise freeable slabs.
286 #define REAPTIMEOUT_AC (2*HZ)
287 #define REAPTIMEOUT_NODE (4*HZ)
290 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
291 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
292 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
293 #define STATS_INC_GROWN(x) ((x)->grown++)
294 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
295 #define STATS_SET_HIGH(x) \
297 if ((x)->num_active > (x)->high_mark) \
298 (x)->high_mark = (x)->num_active; \
300 #define STATS_INC_ERR(x) ((x)->errors++)
301 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
302 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
303 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
304 #define STATS_SET_FREEABLE(x, i) \
306 if ((x)->max_freeable < i) \
307 (x)->max_freeable = i; \
309 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
310 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
311 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
312 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
314 #define STATS_INC_ACTIVE(x) do { } while (0)
315 #define STATS_DEC_ACTIVE(x) do { } while (0)
316 #define STATS_INC_ALLOCED(x) do { } while (0)
317 #define STATS_INC_GROWN(x) do { } while (0)
318 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
319 #define STATS_SET_HIGH(x) do { } while (0)
320 #define STATS_INC_ERR(x) do { } while (0)
321 #define STATS_INC_NODEALLOCS(x) do { } while (0)
322 #define STATS_INC_NODEFREES(x) do { } while (0)
323 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
324 #define STATS_SET_FREEABLE(x, i) do { } while (0)
325 #define STATS_INC_ALLOCHIT(x) do { } while (0)
326 #define STATS_INC_ALLOCMISS(x) do { } while (0)
327 #define STATS_INC_FREEHIT(x) do { } while (0)
328 #define STATS_INC_FREEMISS(x) do { } while (0)
334 * memory layout of objects:
336 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
337 * the end of an object is aligned with the end of the real
338 * allocation. Catches writes behind the end of the allocation.
339 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
341 * cachep->obj_offset: The real object.
342 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
343 * cachep->size - 1* BYTES_PER_WORD: last caller address
344 * [BYTES_PER_WORD long]
346 static int obj_offset(struct kmem_cache *cachep)
348 return cachep->obj_offset;
351 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
353 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
354 return (unsigned long long*) (objp + obj_offset(cachep) -
355 sizeof(unsigned long long));
358 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
360 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
361 if (cachep->flags & SLAB_STORE_USER)
362 return (unsigned long long *)(objp + cachep->size -
363 sizeof(unsigned long long) -
365 return (unsigned long long *) (objp + cachep->size -
366 sizeof(unsigned long long));
369 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
371 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
372 return (void **)(objp + cachep->size - BYTES_PER_WORD);
377 #define obj_offset(x) 0
378 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
379 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
380 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
384 #ifdef CONFIG_DEBUG_SLAB_LEAK
386 static inline bool is_store_user_clean(struct kmem_cache *cachep)
388 return atomic_read(&cachep->store_user_clean) == 1;
391 static inline void set_store_user_clean(struct kmem_cache *cachep)
393 atomic_set(&cachep->store_user_clean, 1);
396 static inline void set_store_user_dirty(struct kmem_cache *cachep)
398 if (is_store_user_clean(cachep))
399 atomic_set(&cachep->store_user_clean, 0);
403 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
408 * Do not go above this order unless 0 objects fit into the slab or
409 * overridden on the command line.
411 #define SLAB_MAX_ORDER_HI 1
412 #define SLAB_MAX_ORDER_LO 0
413 static int slab_max_order = SLAB_MAX_ORDER_LO;
414 static bool slab_max_order_set __initdata;
416 static inline struct kmem_cache *virt_to_cache(const void *obj)
418 struct page *page = virt_to_head_page(obj);
419 return page->slab_cache;
422 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
425 return page->s_mem + cache->size * idx;
429 * We want to avoid an expensive divide : (offset / cache->size)
430 * Using the fact that size is a constant for a particular cache,
431 * we can replace (offset / cache->size) by
432 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
434 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
435 const struct page *page, void *obj)
437 u32 offset = (obj - page->s_mem);
438 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
441 #define BOOT_CPUCACHE_ENTRIES 1
442 /* internal cache of cache description objs */
443 static struct kmem_cache kmem_cache_boot = {
445 .limit = BOOT_CPUCACHE_ENTRIES,
447 .size = sizeof(struct kmem_cache),
448 .name = "kmem_cache",
451 #define BAD_ALIEN_MAGIC 0x01020304ul
453 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
455 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
457 return this_cpu_ptr(cachep->cpu_cache);
461 * Calculate the number of objects and left-over bytes for a given buffer size.
463 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
464 unsigned long flags, size_t *left_over)
467 size_t slab_size = PAGE_SIZE << gfporder;
470 * The slab management structure can be either off the slab or
471 * on it. For the latter case, the memory allocated for a
474 * - @buffer_size bytes for each object
475 * - One freelist_idx_t for each object
477 * We don't need to consider alignment of freelist because
478 * freelist will be at the end of slab page. The objects will be
479 * at the correct alignment.
481 * If the slab management structure is off the slab, then the
482 * alignment will already be calculated into the size. Because
483 * the slabs are all pages aligned, the objects will be at the
484 * correct alignment when allocated.
486 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
487 num = slab_size / buffer_size;
488 *left_over = slab_size % buffer_size;
490 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
491 *left_over = slab_size %
492 (buffer_size + sizeof(freelist_idx_t));
499 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
501 static void __slab_error(const char *function, struct kmem_cache *cachep,
504 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
505 function, cachep->name, msg);
507 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
512 * By default on NUMA we use alien caches to stage the freeing of
513 * objects allocated from other nodes. This causes massive memory
514 * inefficiencies when using fake NUMA setup to split memory into a
515 * large number of small nodes, so it can be disabled on the command
519 static int use_alien_caches __read_mostly = 1;
520 static int __init noaliencache_setup(char *s)
522 use_alien_caches = 0;
525 __setup("noaliencache", noaliencache_setup);
527 static int __init slab_max_order_setup(char *str)
529 get_option(&str, &slab_max_order);
530 slab_max_order = slab_max_order < 0 ? 0 :
531 min(slab_max_order, MAX_ORDER - 1);
532 slab_max_order_set = true;
536 __setup("slab_max_order=", slab_max_order_setup);
540 * Special reaping functions for NUMA systems called from cache_reap().
541 * These take care of doing round robin flushing of alien caches (containing
542 * objects freed on different nodes from which they were allocated) and the
543 * flushing of remote pcps by calling drain_node_pages.
545 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
547 static void init_reap_node(int cpu)
551 node = next_node(cpu_to_mem(cpu), node_online_map);
552 if (node == MAX_NUMNODES)
553 node = first_node(node_online_map);
555 per_cpu(slab_reap_node, cpu) = node;
558 static void next_reap_node(void)
560 int node = __this_cpu_read(slab_reap_node);
562 node = next_node(node, node_online_map);
563 if (unlikely(node >= MAX_NUMNODES))
564 node = first_node(node_online_map);
565 __this_cpu_write(slab_reap_node, node);
569 #define init_reap_node(cpu) do { } while (0)
570 #define next_reap_node(void) do { } while (0)
574 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
575 * via the workqueue/eventd.
576 * Add the CPU number into the expiration time to minimize the possibility of
577 * the CPUs getting into lockstep and contending for the global cache chain
580 static void start_cpu_timer(int cpu)
582 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
585 * When this gets called from do_initcalls via cpucache_init(),
586 * init_workqueues() has already run, so keventd will be setup
589 if (keventd_up() && reap_work->work.func == NULL) {
591 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
592 schedule_delayed_work_on(cpu, reap_work,
593 __round_jiffies_relative(HZ, cpu));
597 static void init_arraycache(struct array_cache *ac, int limit, int batch)
600 * The array_cache structures contain pointers to free object.
601 * However, when such objects are allocated or transferred to another
602 * cache the pointers are not cleared and they could be counted as
603 * valid references during a kmemleak scan. Therefore, kmemleak must
604 * not scan such objects.
606 kmemleak_no_scan(ac);
610 ac->batchcount = batch;
615 static struct array_cache *alloc_arraycache(int node, int entries,
616 int batchcount, gfp_t gfp)
618 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
619 struct array_cache *ac = NULL;
621 ac = kmalloc_node(memsize, gfp, node);
622 init_arraycache(ac, entries, batchcount);
626 static inline bool is_slab_pfmemalloc(struct page *page)
628 return PageSlabPfmemalloc(page);
631 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
632 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
633 struct array_cache *ac)
635 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
639 if (!pfmemalloc_active)
642 spin_lock_irqsave(&n->list_lock, flags);
643 list_for_each_entry(page, &n->slabs_full, lru)
644 if (is_slab_pfmemalloc(page))
647 list_for_each_entry(page, &n->slabs_partial, lru)
648 if (is_slab_pfmemalloc(page))
651 list_for_each_entry(page, &n->slabs_free, lru)
652 if (is_slab_pfmemalloc(page))
655 pfmemalloc_active = false;
657 spin_unlock_irqrestore(&n->list_lock, flags);
660 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
661 gfp_t flags, bool force_refill)
664 void *objp = ac->entry[--ac->avail];
666 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
667 if (unlikely(is_obj_pfmemalloc(objp))) {
668 struct kmem_cache_node *n;
670 if (gfp_pfmemalloc_allowed(flags)) {
671 clear_obj_pfmemalloc(&objp);
675 /* The caller cannot use PFMEMALLOC objects, find another one */
676 for (i = 0; i < ac->avail; i++) {
677 /* If a !PFMEMALLOC object is found, swap them */
678 if (!is_obj_pfmemalloc(ac->entry[i])) {
680 ac->entry[i] = ac->entry[ac->avail];
681 ac->entry[ac->avail] = objp;
687 * If there are empty slabs on the slabs_free list and we are
688 * being forced to refill the cache, mark this one !pfmemalloc.
690 n = get_node(cachep, numa_mem_id());
691 if (!list_empty(&n->slabs_free) && force_refill) {
692 struct page *page = virt_to_head_page(objp);
693 ClearPageSlabPfmemalloc(page);
694 clear_obj_pfmemalloc(&objp);
695 recheck_pfmemalloc_active(cachep, ac);
699 /* No !PFMEMALLOC objects available */
707 static inline void *ac_get_obj(struct kmem_cache *cachep,
708 struct array_cache *ac, gfp_t flags, bool force_refill)
712 if (unlikely(sk_memalloc_socks()))
713 objp = __ac_get_obj(cachep, ac, flags, force_refill);
715 objp = ac->entry[--ac->avail];
720 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
721 struct array_cache *ac, void *objp)
723 if (unlikely(pfmemalloc_active)) {
724 /* Some pfmemalloc slabs exist, check if this is one */
725 struct page *page = virt_to_head_page(objp);
726 if (PageSlabPfmemalloc(page))
727 set_obj_pfmemalloc(&objp);
733 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
736 if (unlikely(sk_memalloc_socks()))
737 objp = __ac_put_obj(cachep, ac, objp);
739 ac->entry[ac->avail++] = objp;
743 * Transfer objects in one arraycache to another.
744 * Locking must be handled by the caller.
746 * Return the number of entries transferred.
748 static int transfer_objects(struct array_cache *to,
749 struct array_cache *from, unsigned int max)
751 /* Figure out how many entries to transfer */
752 int nr = min3(from->avail, max, to->limit - to->avail);
757 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
767 #define drain_alien_cache(cachep, alien) do { } while (0)
768 #define reap_alien(cachep, n) do { } while (0)
770 static inline struct alien_cache **alloc_alien_cache(int node,
771 int limit, gfp_t gfp)
773 return (struct alien_cache **)BAD_ALIEN_MAGIC;
776 static inline void free_alien_cache(struct alien_cache **ac_ptr)
780 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
785 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
791 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
792 gfp_t flags, int nodeid)
797 static inline gfp_t gfp_exact_node(gfp_t flags)
802 #else /* CONFIG_NUMA */
804 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
805 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
807 static struct alien_cache *__alloc_alien_cache(int node, int entries,
808 int batch, gfp_t gfp)
810 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
811 struct alien_cache *alc = NULL;
813 alc = kmalloc_node(memsize, gfp, node);
814 init_arraycache(&alc->ac, entries, batch);
815 spin_lock_init(&alc->lock);
819 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
821 struct alien_cache **alc_ptr;
822 size_t memsize = sizeof(void *) * nr_node_ids;
827 alc_ptr = kzalloc_node(memsize, gfp, node);
832 if (i == node || !node_online(i))
834 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
836 for (i--; i >= 0; i--)
845 static void free_alien_cache(struct alien_cache **alc_ptr)
856 static void __drain_alien_cache(struct kmem_cache *cachep,
857 struct array_cache *ac, int node,
858 struct list_head *list)
860 struct kmem_cache_node *n = get_node(cachep, node);
863 spin_lock(&n->list_lock);
865 * Stuff objects into the remote nodes shared array first.
866 * That way we could avoid the overhead of putting the objects
867 * into the free lists and getting them back later.
870 transfer_objects(n->shared, ac, ac->limit);
872 free_block(cachep, ac->entry, ac->avail, node, list);
874 spin_unlock(&n->list_lock);
879 * Called from cache_reap() to regularly drain alien caches round robin.
881 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
883 int node = __this_cpu_read(slab_reap_node);
886 struct alien_cache *alc = n->alien[node];
887 struct array_cache *ac;
891 if (ac->avail && spin_trylock_irq(&alc->lock)) {
894 __drain_alien_cache(cachep, ac, node, &list);
895 spin_unlock_irq(&alc->lock);
896 slabs_destroy(cachep, &list);
902 static void drain_alien_cache(struct kmem_cache *cachep,
903 struct alien_cache **alien)
906 struct alien_cache *alc;
907 struct array_cache *ac;
910 for_each_online_node(i) {
916 spin_lock_irqsave(&alc->lock, flags);
917 __drain_alien_cache(cachep, ac, i, &list);
918 spin_unlock_irqrestore(&alc->lock, flags);
919 slabs_destroy(cachep, &list);
924 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
925 int node, int page_node)
927 struct kmem_cache_node *n;
928 struct alien_cache *alien = NULL;
929 struct array_cache *ac;
932 n = get_node(cachep, node);
933 STATS_INC_NODEFREES(cachep);
934 if (n->alien && n->alien[page_node]) {
935 alien = n->alien[page_node];
937 spin_lock(&alien->lock);
938 if (unlikely(ac->avail == ac->limit)) {
939 STATS_INC_ACOVERFLOW(cachep);
940 __drain_alien_cache(cachep, ac, page_node, &list);
942 ac_put_obj(cachep, ac, objp);
943 spin_unlock(&alien->lock);
944 slabs_destroy(cachep, &list);
946 n = get_node(cachep, page_node);
947 spin_lock(&n->list_lock);
948 free_block(cachep, &objp, 1, page_node, &list);
949 spin_unlock(&n->list_lock);
950 slabs_destroy(cachep, &list);
955 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
957 int page_node = page_to_nid(virt_to_page(objp));
958 int node = numa_mem_id();
960 * Make sure we are not freeing a object from another node to the array
963 if (likely(node == page_node))
966 return __cache_free_alien(cachep, objp, node, page_node);
970 * Construct gfp mask to allocate from a specific node but do not direct reclaim
971 * or warn about failures. kswapd may still wake to reclaim in the background.
973 static inline gfp_t gfp_exact_node(gfp_t flags)
975 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
980 * Allocates and initializes node for a node on each slab cache, used for
981 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
982 * will be allocated off-node since memory is not yet online for the new node.
983 * When hotplugging memory or a cpu, existing node are not replaced if
986 * Must hold slab_mutex.
988 static int init_cache_node_node(int node)
990 struct kmem_cache *cachep;
991 struct kmem_cache_node *n;
992 const size_t memsize = sizeof(struct kmem_cache_node);
994 list_for_each_entry(cachep, &slab_caches, list) {
996 * Set up the kmem_cache_node for cpu before we can
997 * begin anything. Make sure some other cpu on this
998 * node has not already allocated this
1000 n = get_node(cachep, node);
1002 n = kmalloc_node(memsize, GFP_KERNEL, node);
1005 kmem_cache_node_init(n);
1006 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1007 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1010 * The kmem_cache_nodes don't come and go as CPUs
1011 * come and go. slab_mutex is sufficient
1014 cachep->node[node] = n;
1017 spin_lock_irq(&n->list_lock);
1019 (1 + nr_cpus_node(node)) *
1020 cachep->batchcount + cachep->num;
1021 spin_unlock_irq(&n->list_lock);
1026 static inline int slabs_tofree(struct kmem_cache *cachep,
1027 struct kmem_cache_node *n)
1029 return (n->free_objects + cachep->num - 1) / cachep->num;
1032 static void cpuup_canceled(long cpu)
1034 struct kmem_cache *cachep;
1035 struct kmem_cache_node *n = NULL;
1036 int node = cpu_to_mem(cpu);
1037 const struct cpumask *mask = cpumask_of_node(node);
1039 list_for_each_entry(cachep, &slab_caches, list) {
1040 struct array_cache *nc;
1041 struct array_cache *shared;
1042 struct alien_cache **alien;
1045 n = get_node(cachep, node);
1049 spin_lock_irq(&n->list_lock);
1051 /* Free limit for this kmem_cache_node */
1052 n->free_limit -= cachep->batchcount;
1054 /* cpu is dead; no one can alloc from it. */
1055 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1057 free_block(cachep, nc->entry, nc->avail, node, &list);
1061 if (!cpumask_empty(mask)) {
1062 spin_unlock_irq(&n->list_lock);
1068 free_block(cachep, shared->entry,
1069 shared->avail, node, &list);
1076 spin_unlock_irq(&n->list_lock);
1080 drain_alien_cache(cachep, alien);
1081 free_alien_cache(alien);
1085 slabs_destroy(cachep, &list);
1088 * In the previous loop, all the objects were freed to
1089 * the respective cache's slabs, now we can go ahead and
1090 * shrink each nodelist to its limit.
1092 list_for_each_entry(cachep, &slab_caches, list) {
1093 n = get_node(cachep, node);
1096 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1100 static int cpuup_prepare(long cpu)
1102 struct kmem_cache *cachep;
1103 struct kmem_cache_node *n = NULL;
1104 int node = cpu_to_mem(cpu);
1108 * We need to do this right in the beginning since
1109 * alloc_arraycache's are going to use this list.
1110 * kmalloc_node allows us to add the slab to the right
1111 * kmem_cache_node and not this cpu's kmem_cache_node
1113 err = init_cache_node_node(node);
1118 * Now we can go ahead with allocating the shared arrays and
1121 list_for_each_entry(cachep, &slab_caches, list) {
1122 struct array_cache *shared = NULL;
1123 struct alien_cache **alien = NULL;
1125 if (cachep->shared) {
1126 shared = alloc_arraycache(node,
1127 cachep->shared * cachep->batchcount,
1128 0xbaadf00d, GFP_KERNEL);
1132 if (use_alien_caches) {
1133 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1139 n = get_node(cachep, node);
1142 spin_lock_irq(&n->list_lock);
1145 * We are serialised from CPU_DEAD or
1146 * CPU_UP_CANCELLED by the cpucontrol lock
1157 spin_unlock_irq(&n->list_lock);
1159 free_alien_cache(alien);
1164 cpuup_canceled(cpu);
1168 static int cpuup_callback(struct notifier_block *nfb,
1169 unsigned long action, void *hcpu)
1171 long cpu = (long)hcpu;
1175 case CPU_UP_PREPARE:
1176 case CPU_UP_PREPARE_FROZEN:
1177 mutex_lock(&slab_mutex);
1178 err = cpuup_prepare(cpu);
1179 mutex_unlock(&slab_mutex);
1182 case CPU_ONLINE_FROZEN:
1183 start_cpu_timer(cpu);
1185 #ifdef CONFIG_HOTPLUG_CPU
1186 case CPU_DOWN_PREPARE:
1187 case CPU_DOWN_PREPARE_FROZEN:
1189 * Shutdown cache reaper. Note that the slab_mutex is
1190 * held so that if cache_reap() is invoked it cannot do
1191 * anything expensive but will only modify reap_work
1192 * and reschedule the timer.
1194 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1195 /* Now the cache_reaper is guaranteed to be not running. */
1196 per_cpu(slab_reap_work, cpu).work.func = NULL;
1198 case CPU_DOWN_FAILED:
1199 case CPU_DOWN_FAILED_FROZEN:
1200 start_cpu_timer(cpu);
1203 case CPU_DEAD_FROZEN:
1205 * Even if all the cpus of a node are down, we don't free the
1206 * kmem_cache_node of any cache. This to avoid a race between
1207 * cpu_down, and a kmalloc allocation from another cpu for
1208 * memory from the node of the cpu going down. The node
1209 * structure is usually allocated from kmem_cache_create() and
1210 * gets destroyed at kmem_cache_destroy().
1214 case CPU_UP_CANCELED:
1215 case CPU_UP_CANCELED_FROZEN:
1216 mutex_lock(&slab_mutex);
1217 cpuup_canceled(cpu);
1218 mutex_unlock(&slab_mutex);
1221 return notifier_from_errno(err);
1224 static struct notifier_block cpucache_notifier = {
1225 &cpuup_callback, NULL, 0
1228 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1230 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1231 * Returns -EBUSY if all objects cannot be drained so that the node is not
1234 * Must hold slab_mutex.
1236 static int __meminit drain_cache_node_node(int node)
1238 struct kmem_cache *cachep;
1241 list_for_each_entry(cachep, &slab_caches, list) {
1242 struct kmem_cache_node *n;
1244 n = get_node(cachep, node);
1248 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1250 if (!list_empty(&n->slabs_full) ||
1251 !list_empty(&n->slabs_partial)) {
1259 static int __meminit slab_memory_callback(struct notifier_block *self,
1260 unsigned long action, void *arg)
1262 struct memory_notify *mnb = arg;
1266 nid = mnb->status_change_nid;
1271 case MEM_GOING_ONLINE:
1272 mutex_lock(&slab_mutex);
1273 ret = init_cache_node_node(nid);
1274 mutex_unlock(&slab_mutex);
1276 case MEM_GOING_OFFLINE:
1277 mutex_lock(&slab_mutex);
1278 ret = drain_cache_node_node(nid);
1279 mutex_unlock(&slab_mutex);
1283 case MEM_CANCEL_ONLINE:
1284 case MEM_CANCEL_OFFLINE:
1288 return notifier_from_errno(ret);
1290 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1293 * swap the static kmem_cache_node with kmalloced memory
1295 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1298 struct kmem_cache_node *ptr;
1300 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1303 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1305 * Do not assume that spinlocks can be initialized via memcpy:
1307 spin_lock_init(&ptr->list_lock);
1309 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1310 cachep->node[nodeid] = ptr;
1314 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1315 * size of kmem_cache_node.
1317 static void __init set_up_node(struct kmem_cache *cachep, int index)
1321 for_each_online_node(node) {
1322 cachep->node[node] = &init_kmem_cache_node[index + node];
1323 cachep->node[node]->next_reap = jiffies +
1325 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1330 * Initialisation. Called after the page allocator have been initialised and
1331 * before smp_init().
1333 void __init kmem_cache_init(void)
1337 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1338 sizeof(struct rcu_head));
1339 kmem_cache = &kmem_cache_boot;
1341 if (num_possible_nodes() == 1)
1342 use_alien_caches = 0;
1344 for (i = 0; i < NUM_INIT_LISTS; i++)
1345 kmem_cache_node_init(&init_kmem_cache_node[i]);
1348 * Fragmentation resistance on low memory - only use bigger
1349 * page orders on machines with more than 32MB of memory if
1350 * not overridden on the command line.
1352 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1353 slab_max_order = SLAB_MAX_ORDER_HI;
1355 /* Bootstrap is tricky, because several objects are allocated
1356 * from caches that do not exist yet:
1357 * 1) initialize the kmem_cache cache: it contains the struct
1358 * kmem_cache structures of all caches, except kmem_cache itself:
1359 * kmem_cache is statically allocated.
1360 * Initially an __init data area is used for the head array and the
1361 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1362 * array at the end of the bootstrap.
1363 * 2) Create the first kmalloc cache.
1364 * The struct kmem_cache for the new cache is allocated normally.
1365 * An __init data area is used for the head array.
1366 * 3) Create the remaining kmalloc caches, with minimally sized
1368 * 4) Replace the __init data head arrays for kmem_cache and the first
1369 * kmalloc cache with kmalloc allocated arrays.
1370 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1371 * the other cache's with kmalloc allocated memory.
1372 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1375 /* 1) create the kmem_cache */
1378 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1380 create_boot_cache(kmem_cache, "kmem_cache",
1381 offsetof(struct kmem_cache, node) +
1382 nr_node_ids * sizeof(struct kmem_cache_node *),
1383 SLAB_HWCACHE_ALIGN);
1384 list_add(&kmem_cache->list, &slab_caches);
1385 slab_state = PARTIAL;
1388 * Initialize the caches that provide memory for the kmem_cache_node
1389 * structures first. Without this, further allocations will bug.
1391 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1392 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1393 slab_state = PARTIAL_NODE;
1394 setup_kmalloc_cache_index_table();
1396 slab_early_init = 0;
1398 /* 5) Replace the bootstrap kmem_cache_node */
1402 for_each_online_node(nid) {
1403 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1405 init_list(kmalloc_caches[INDEX_NODE],
1406 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1410 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1413 void __init kmem_cache_init_late(void)
1415 struct kmem_cache *cachep;
1419 /* 6) resize the head arrays to their final sizes */
1420 mutex_lock(&slab_mutex);
1421 list_for_each_entry(cachep, &slab_caches, list)
1422 if (enable_cpucache(cachep, GFP_NOWAIT))
1424 mutex_unlock(&slab_mutex);
1430 * Register a cpu startup notifier callback that initializes
1431 * cpu_cache_get for all new cpus
1433 register_cpu_notifier(&cpucache_notifier);
1437 * Register a memory hotplug callback that initializes and frees
1440 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1444 * The reap timers are started later, with a module init call: That part
1445 * of the kernel is not yet operational.
1449 static int __init cpucache_init(void)
1454 * Register the timers that return unneeded pages to the page allocator
1456 for_each_online_cpu(cpu)
1457 start_cpu_timer(cpu);
1463 __initcall(cpucache_init);
1465 static noinline void
1466 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1469 struct kmem_cache_node *n;
1471 unsigned long flags;
1473 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1474 DEFAULT_RATELIMIT_BURST);
1476 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1480 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1482 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1483 cachep->name, cachep->size, cachep->gfporder);
1485 for_each_kmem_cache_node(cachep, node, n) {
1486 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1487 unsigned long active_slabs = 0, num_slabs = 0;
1489 spin_lock_irqsave(&n->list_lock, flags);
1490 list_for_each_entry(page, &n->slabs_full, lru) {
1491 active_objs += cachep->num;
1494 list_for_each_entry(page, &n->slabs_partial, lru) {
1495 active_objs += page->active;
1498 list_for_each_entry(page, &n->slabs_free, lru)
1501 free_objects += n->free_objects;
1502 spin_unlock_irqrestore(&n->list_lock, flags);
1504 num_slabs += active_slabs;
1505 num_objs = num_slabs * cachep->num;
1507 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1508 node, active_slabs, num_slabs, active_objs, num_objs,
1515 * Interface to system's page allocator. No need to hold the
1516 * kmem_cache_node ->list_lock.
1518 * If we requested dmaable memory, we will get it. Even if we
1519 * did not request dmaable memory, we might get it, but that
1520 * would be relatively rare and ignorable.
1522 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1528 flags |= cachep->allocflags;
1529 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1530 flags |= __GFP_RECLAIMABLE;
1532 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1534 slab_out_of_memory(cachep, flags, nodeid);
1538 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1539 __free_pages(page, cachep->gfporder);
1543 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1544 if (page_is_pfmemalloc(page))
1545 pfmemalloc_active = true;
1547 nr_pages = (1 << cachep->gfporder);
1548 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1549 add_zone_page_state(page_zone(page),
1550 NR_SLAB_RECLAIMABLE, nr_pages);
1552 add_zone_page_state(page_zone(page),
1553 NR_SLAB_UNRECLAIMABLE, nr_pages);
1554 __SetPageSlab(page);
1555 if (page_is_pfmemalloc(page))
1556 SetPageSlabPfmemalloc(page);
1558 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1559 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1562 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1564 kmemcheck_mark_unallocated_pages(page, nr_pages);
1571 * Interface to system's page release.
1573 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1575 const unsigned long nr_freed = (1 << cachep->gfporder);
1577 kmemcheck_free_shadow(page, cachep->gfporder);
1579 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1580 sub_zone_page_state(page_zone(page),
1581 NR_SLAB_RECLAIMABLE, nr_freed);
1583 sub_zone_page_state(page_zone(page),
1584 NR_SLAB_UNRECLAIMABLE, nr_freed);
1586 BUG_ON(!PageSlab(page));
1587 __ClearPageSlabPfmemalloc(page);
1588 __ClearPageSlab(page);
1589 page_mapcount_reset(page);
1590 page->mapping = NULL;
1592 if (current->reclaim_state)
1593 current->reclaim_state->reclaimed_slab += nr_freed;
1594 __free_kmem_pages(page, cachep->gfporder);
1597 static void kmem_rcu_free(struct rcu_head *head)
1599 struct kmem_cache *cachep;
1602 page = container_of(head, struct page, rcu_head);
1603 cachep = page->slab_cache;
1605 kmem_freepages(cachep, page);
1609 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1611 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1612 (cachep->size % PAGE_SIZE) == 0)
1618 #ifdef CONFIG_DEBUG_PAGEALLOC
1619 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1620 unsigned long caller)
1622 int size = cachep->object_size;
1624 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1626 if (size < 5 * sizeof(unsigned long))
1629 *addr++ = 0x12345678;
1631 *addr++ = smp_processor_id();
1632 size -= 3 * sizeof(unsigned long);
1634 unsigned long *sptr = &caller;
1635 unsigned long svalue;
1637 while (!kstack_end(sptr)) {
1639 if (kernel_text_address(svalue)) {
1641 size -= sizeof(unsigned long);
1642 if (size <= sizeof(unsigned long))
1648 *addr++ = 0x87654321;
1651 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1652 int map, unsigned long caller)
1654 if (!is_debug_pagealloc_cache(cachep))
1658 store_stackinfo(cachep, objp, caller);
1660 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1664 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1665 int map, unsigned long caller) {}
1669 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1671 int size = cachep->object_size;
1672 addr = &((char *)addr)[obj_offset(cachep)];
1674 memset(addr, val, size);
1675 *(unsigned char *)(addr + size - 1) = POISON_END;
1678 static void dump_line(char *data, int offset, int limit)
1681 unsigned char error = 0;
1684 printk(KERN_ERR "%03x: ", offset);
1685 for (i = 0; i < limit; i++) {
1686 if (data[offset + i] != POISON_FREE) {
1687 error = data[offset + i];
1691 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1692 &data[offset], limit, 1);
1694 if (bad_count == 1) {
1695 error ^= POISON_FREE;
1696 if (!(error & (error - 1))) {
1697 printk(KERN_ERR "Single bit error detected. Probably "
1700 printk(KERN_ERR "Run memtest86+ or a similar memory "
1703 printk(KERN_ERR "Run a memory test tool.\n");
1712 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1717 if (cachep->flags & SLAB_RED_ZONE) {
1718 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1719 *dbg_redzone1(cachep, objp),
1720 *dbg_redzone2(cachep, objp));
1723 if (cachep->flags & SLAB_STORE_USER) {
1724 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1725 *dbg_userword(cachep, objp),
1726 *dbg_userword(cachep, objp));
1728 realobj = (char *)objp + obj_offset(cachep);
1729 size = cachep->object_size;
1730 for (i = 0; i < size && lines; i += 16, lines--) {
1733 if (i + limit > size)
1735 dump_line(realobj, i, limit);
1739 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1745 if (is_debug_pagealloc_cache(cachep))
1748 realobj = (char *)objp + obj_offset(cachep);
1749 size = cachep->object_size;
1751 for (i = 0; i < size; i++) {
1752 char exp = POISON_FREE;
1755 if (realobj[i] != exp) {
1761 "Slab corruption (%s): %s start=%p, len=%d\n",
1762 print_tainted(), cachep->name, realobj, size);
1763 print_objinfo(cachep, objp, 0);
1765 /* Hexdump the affected line */
1768 if (i + limit > size)
1770 dump_line(realobj, i, limit);
1773 /* Limit to 5 lines */
1779 /* Print some data about the neighboring objects, if they
1782 struct page *page = virt_to_head_page(objp);
1785 objnr = obj_to_index(cachep, page, objp);
1787 objp = index_to_obj(cachep, page, objnr - 1);
1788 realobj = (char *)objp + obj_offset(cachep);
1789 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1791 print_objinfo(cachep, objp, 2);
1793 if (objnr + 1 < cachep->num) {
1794 objp = index_to_obj(cachep, page, objnr + 1);
1795 realobj = (char *)objp + obj_offset(cachep);
1796 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1798 print_objinfo(cachep, objp, 2);
1805 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1810 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1811 poison_obj(cachep, page->freelist - obj_offset(cachep),
1815 for (i = 0; i < cachep->num; i++) {
1816 void *objp = index_to_obj(cachep, page, i);
1818 if (cachep->flags & SLAB_POISON) {
1819 check_poison_obj(cachep, objp);
1820 slab_kernel_map(cachep, objp, 1, 0);
1822 if (cachep->flags & SLAB_RED_ZONE) {
1823 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1824 slab_error(cachep, "start of a freed object "
1826 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1827 slab_error(cachep, "end of a freed object "
1833 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1840 * slab_destroy - destroy and release all objects in a slab
1841 * @cachep: cache pointer being destroyed
1842 * @page: page pointer being destroyed
1844 * Destroy all the objs in a slab page, and release the mem back to the system.
1845 * Before calling the slab page must have been unlinked from the cache. The
1846 * kmem_cache_node ->list_lock is not held/needed.
1848 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1852 freelist = page->freelist;
1853 slab_destroy_debugcheck(cachep, page);
1854 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1855 call_rcu(&page->rcu_head, kmem_rcu_free);
1857 kmem_freepages(cachep, page);
1860 * From now on, we don't use freelist
1861 * although actual page can be freed in rcu context
1863 if (OFF_SLAB(cachep))
1864 kmem_cache_free(cachep->freelist_cache, freelist);
1867 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1869 struct page *page, *n;
1871 list_for_each_entry_safe(page, n, list, lru) {
1872 list_del(&page->lru);
1873 slab_destroy(cachep, page);
1878 * calculate_slab_order - calculate size (page order) of slabs
1879 * @cachep: pointer to the cache that is being created
1880 * @size: size of objects to be created in this cache.
1881 * @flags: slab allocation flags
1883 * Also calculates the number of objects per slab.
1885 * This could be made much more intelligent. For now, try to avoid using
1886 * high order pages for slabs. When the gfp() functions are more friendly
1887 * towards high-order requests, this should be changed.
1889 static size_t calculate_slab_order(struct kmem_cache *cachep,
1890 size_t size, unsigned long flags)
1892 size_t left_over = 0;
1895 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1899 num = cache_estimate(gfporder, size, flags, &remainder);
1903 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1904 if (num > SLAB_OBJ_MAX_NUM)
1907 if (flags & CFLGS_OFF_SLAB) {
1908 struct kmem_cache *freelist_cache;
1909 size_t freelist_size;
1911 freelist_size = num * sizeof(freelist_idx_t);
1912 freelist_cache = kmalloc_slab(freelist_size, 0u);
1913 if (!freelist_cache)
1917 * Needed to avoid possible looping condition
1920 if (OFF_SLAB(freelist_cache))
1923 /* check if off slab has enough benefit */
1924 if (freelist_cache->size > cachep->size / 2)
1928 /* Found something acceptable - save it away */
1930 cachep->gfporder = gfporder;
1931 left_over = remainder;
1934 * A VFS-reclaimable slab tends to have most allocations
1935 * as GFP_NOFS and we really don't want to have to be allocating
1936 * higher-order pages when we are unable to shrink dcache.
1938 if (flags & SLAB_RECLAIM_ACCOUNT)
1942 * Large number of objects is good, but very large slabs are
1943 * currently bad for the gfp()s.
1945 if (gfporder >= slab_max_order)
1949 * Acceptable internal fragmentation?
1951 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1957 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1958 struct kmem_cache *cachep, int entries, int batchcount)
1962 struct array_cache __percpu *cpu_cache;
1964 size = sizeof(void *) * entries + sizeof(struct array_cache);
1965 cpu_cache = __alloc_percpu(size, sizeof(void *));
1970 for_each_possible_cpu(cpu) {
1971 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1972 entries, batchcount);
1978 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1980 if (slab_state >= FULL)
1981 return enable_cpucache(cachep, gfp);
1983 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1984 if (!cachep->cpu_cache)
1987 if (slab_state == DOWN) {
1988 /* Creation of first cache (kmem_cache). */
1989 set_up_node(kmem_cache, CACHE_CACHE);
1990 } else if (slab_state == PARTIAL) {
1991 /* For kmem_cache_node */
1992 set_up_node(cachep, SIZE_NODE);
1996 for_each_online_node(node) {
1997 cachep->node[node] = kmalloc_node(
1998 sizeof(struct kmem_cache_node), gfp, node);
1999 BUG_ON(!cachep->node[node]);
2000 kmem_cache_node_init(cachep->node[node]);
2004 cachep->node[numa_mem_id()]->next_reap =
2005 jiffies + REAPTIMEOUT_NODE +
2006 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2008 cpu_cache_get(cachep)->avail = 0;
2009 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2010 cpu_cache_get(cachep)->batchcount = 1;
2011 cpu_cache_get(cachep)->touched = 0;
2012 cachep->batchcount = 1;
2013 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2017 unsigned long kmem_cache_flags(unsigned long object_size,
2018 unsigned long flags, const char *name,
2019 void (*ctor)(void *))
2025 __kmem_cache_alias(const char *name, size_t size, size_t align,
2026 unsigned long flags, void (*ctor)(void *))
2028 struct kmem_cache *cachep;
2030 cachep = find_mergeable(size, align, flags, name, ctor);
2035 * Adjust the object sizes so that we clear
2036 * the complete object on kzalloc.
2038 cachep->object_size = max_t(int, cachep->object_size, size);
2043 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
2044 size_t size, unsigned long flags)
2050 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
2053 left = calculate_slab_order(cachep, size,
2054 flags | CFLGS_OBJFREELIST_SLAB);
2058 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
2061 cachep->colour = left / cachep->colour_off;
2066 static bool set_off_slab_cache(struct kmem_cache *cachep,
2067 size_t size, unsigned long flags)
2074 * Always use on-slab management when SLAB_NOLEAKTRACE
2075 * to avoid recursive calls into kmemleak.
2077 if (flags & SLAB_NOLEAKTRACE)
2081 * Size is large, assume best to place the slab management obj
2082 * off-slab (should allow better packing of objs).
2084 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
2089 * If the slab has been placed off-slab, and we have enough space then
2090 * move it on-slab. This is at the expense of any extra colouring.
2092 if (left >= cachep->num * sizeof(freelist_idx_t))
2095 cachep->colour = left / cachep->colour_off;
2100 static bool set_on_slab_cache(struct kmem_cache *cachep,
2101 size_t size, unsigned long flags)
2107 left = calculate_slab_order(cachep, size, flags);
2111 cachep->colour = left / cachep->colour_off;
2117 * __kmem_cache_create - Create a cache.
2118 * @cachep: cache management descriptor
2119 * @flags: SLAB flags
2121 * Returns a ptr to the cache on success, NULL on failure.
2122 * Cannot be called within a int, but can be interrupted.
2123 * The @ctor is run when new pages are allocated by the cache.
2127 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2128 * to catch references to uninitialised memory.
2130 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2131 * for buffer overruns.
2133 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2134 * cacheline. This can be beneficial if you're counting cycles as closely
2138 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2140 size_t ralign = BYTES_PER_WORD;
2143 size_t size = cachep->size;
2148 * Enable redzoning and last user accounting, except for caches with
2149 * large objects, if the increased size would increase the object size
2150 * above the next power of two: caches with object sizes just above a
2151 * power of two have a significant amount of internal fragmentation.
2153 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2154 2 * sizeof(unsigned long long)))
2155 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2156 if (!(flags & SLAB_DESTROY_BY_RCU))
2157 flags |= SLAB_POISON;
2162 * Check that size is in terms of words. This is needed to avoid
2163 * unaligned accesses for some archs when redzoning is used, and makes
2164 * sure any on-slab bufctl's are also correctly aligned.
2166 if (size & (BYTES_PER_WORD - 1)) {
2167 size += (BYTES_PER_WORD - 1);
2168 size &= ~(BYTES_PER_WORD - 1);
2171 if (flags & SLAB_RED_ZONE) {
2172 ralign = REDZONE_ALIGN;
2173 /* If redzoning, ensure that the second redzone is suitably
2174 * aligned, by adjusting the object size accordingly. */
2175 size += REDZONE_ALIGN - 1;
2176 size &= ~(REDZONE_ALIGN - 1);
2179 /* 3) caller mandated alignment */
2180 if (ralign < cachep->align) {
2181 ralign = cachep->align;
2183 /* disable debug if necessary */
2184 if (ralign > __alignof__(unsigned long long))
2185 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2189 cachep->align = ralign;
2190 cachep->colour_off = cache_line_size();
2191 /* Offset must be a multiple of the alignment. */
2192 if (cachep->colour_off < cachep->align)
2193 cachep->colour_off = cachep->align;
2195 if (slab_is_available())
2203 * Both debugging options require word-alignment which is calculated
2206 if (flags & SLAB_RED_ZONE) {
2207 /* add space for red zone words */
2208 cachep->obj_offset += sizeof(unsigned long long);
2209 size += 2 * sizeof(unsigned long long);
2211 if (flags & SLAB_STORE_USER) {
2212 /* user store requires one word storage behind the end of
2213 * the real object. But if the second red zone needs to be
2214 * aligned to 64 bits, we must allow that much space.
2216 if (flags & SLAB_RED_ZONE)
2217 size += REDZONE_ALIGN;
2219 size += BYTES_PER_WORD;
2223 size = ALIGN(size, cachep->align);
2225 * We should restrict the number of objects in a slab to implement
2226 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2228 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2229 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2233 * To activate debug pagealloc, off-slab management is necessary
2234 * requirement. In early phase of initialization, small sized slab
2235 * doesn't get initialized so it would not be possible. So, we need
2236 * to check size >= 256. It guarantees that all necessary small
2237 * sized slab is initialized in current slab initialization sequence.
2239 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2240 size >= 256 && cachep->object_size > cache_line_size()) {
2241 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2242 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2244 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2245 flags |= CFLGS_OFF_SLAB;
2246 cachep->obj_offset += tmp_size - size;
2254 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2255 flags |= CFLGS_OBJFREELIST_SLAB;
2259 if (set_off_slab_cache(cachep, size, flags)) {
2260 flags |= CFLGS_OFF_SLAB;
2264 if (set_on_slab_cache(cachep, size, flags))
2270 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2271 cachep->flags = flags;
2272 cachep->allocflags = __GFP_COMP;
2273 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2274 cachep->allocflags |= GFP_DMA;
2275 cachep->size = size;
2276 cachep->reciprocal_buffer_size = reciprocal_value(size);
2280 * If we're going to use the generic kernel_map_pages()
2281 * poisoning, then it's going to smash the contents of
2282 * the redzone and userword anyhow, so switch them off.
2284 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2285 (cachep->flags & SLAB_POISON) &&
2286 is_debug_pagealloc_cache(cachep))
2287 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2290 if (OFF_SLAB(cachep)) {
2291 cachep->freelist_cache =
2292 kmalloc_slab(cachep->freelist_size, 0u);
2295 err = setup_cpu_cache(cachep, gfp);
2297 __kmem_cache_release(cachep);
2305 static void check_irq_off(void)
2307 BUG_ON(!irqs_disabled());
2310 static void check_irq_on(void)
2312 BUG_ON(irqs_disabled());
2315 static void check_spinlock_acquired(struct kmem_cache *cachep)
2319 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2323 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2327 assert_spin_locked(&get_node(cachep, node)->list_lock);
2332 #define check_irq_off() do { } while(0)
2333 #define check_irq_on() do { } while(0)
2334 #define check_spinlock_acquired(x) do { } while(0)
2335 #define check_spinlock_acquired_node(x, y) do { } while(0)
2338 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2339 struct array_cache *ac,
2340 int force, int node);
2342 static void do_drain(void *arg)
2344 struct kmem_cache *cachep = arg;
2345 struct array_cache *ac;
2346 int node = numa_mem_id();
2347 struct kmem_cache_node *n;
2351 ac = cpu_cache_get(cachep);
2352 n = get_node(cachep, node);
2353 spin_lock(&n->list_lock);
2354 free_block(cachep, ac->entry, ac->avail, node, &list);
2355 spin_unlock(&n->list_lock);
2356 slabs_destroy(cachep, &list);
2360 static void drain_cpu_caches(struct kmem_cache *cachep)
2362 struct kmem_cache_node *n;
2365 on_each_cpu(do_drain, cachep, 1);
2367 for_each_kmem_cache_node(cachep, node, n)
2369 drain_alien_cache(cachep, n->alien);
2371 for_each_kmem_cache_node(cachep, node, n)
2372 drain_array(cachep, n, n->shared, 1, node);
2376 * Remove slabs from the list of free slabs.
2377 * Specify the number of slabs to drain in tofree.
2379 * Returns the actual number of slabs released.
2381 static int drain_freelist(struct kmem_cache *cache,
2382 struct kmem_cache_node *n, int tofree)
2384 struct list_head *p;
2389 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2391 spin_lock_irq(&n->list_lock);
2392 p = n->slabs_free.prev;
2393 if (p == &n->slabs_free) {
2394 spin_unlock_irq(&n->list_lock);
2398 page = list_entry(p, struct page, lru);
2399 list_del(&page->lru);
2401 * Safe to drop the lock. The slab is no longer linked
2404 n->free_objects -= cache->num;
2405 spin_unlock_irq(&n->list_lock);
2406 slab_destroy(cache, page);
2413 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2417 struct kmem_cache_node *n;
2419 drain_cpu_caches(cachep);
2422 for_each_kmem_cache_node(cachep, node, n) {
2423 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2425 ret += !list_empty(&n->slabs_full) ||
2426 !list_empty(&n->slabs_partial);
2428 return (ret ? 1 : 0);
2431 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2433 return __kmem_cache_shrink(cachep, false);
2436 void __kmem_cache_release(struct kmem_cache *cachep)
2439 struct kmem_cache_node *n;
2441 free_percpu(cachep->cpu_cache);
2443 /* NUMA: free the node structures */
2444 for_each_kmem_cache_node(cachep, i, n) {
2446 free_alien_cache(n->alien);
2448 cachep->node[i] = NULL;
2453 * Get the memory for a slab management obj.
2455 * For a slab cache when the slab descriptor is off-slab, the
2456 * slab descriptor can't come from the same cache which is being created,
2457 * Because if it is the case, that means we defer the creation of
2458 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2459 * And we eventually call down to __kmem_cache_create(), which
2460 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2461 * This is a "chicken-and-egg" problem.
2463 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2464 * which are all initialized during kmem_cache_init().
2466 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2467 struct page *page, int colour_off,
2468 gfp_t local_flags, int nodeid)
2471 void *addr = page_address(page);
2473 page->s_mem = addr + colour_off;
2476 if (OBJFREELIST_SLAB(cachep))
2478 else if (OFF_SLAB(cachep)) {
2479 /* Slab management obj is off-slab. */
2480 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2481 local_flags, nodeid);
2485 /* We will use last bytes at the slab for freelist */
2486 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2487 cachep->freelist_size;
2493 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2495 return ((freelist_idx_t *)page->freelist)[idx];
2498 static inline void set_free_obj(struct page *page,
2499 unsigned int idx, freelist_idx_t val)
2501 ((freelist_idx_t *)(page->freelist))[idx] = val;
2504 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2509 for (i = 0; i < cachep->num; i++) {
2510 void *objp = index_to_obj(cachep, page, i);
2512 if (cachep->flags & SLAB_STORE_USER)
2513 *dbg_userword(cachep, objp) = NULL;
2515 if (cachep->flags & SLAB_RED_ZONE) {
2516 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2517 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2520 * Constructors are not allowed to allocate memory from the same
2521 * cache which they are a constructor for. Otherwise, deadlock.
2522 * They must also be threaded.
2524 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2525 cachep->ctor(objp + obj_offset(cachep));
2527 if (cachep->flags & SLAB_RED_ZONE) {
2528 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2529 slab_error(cachep, "constructor overwrote the"
2530 " end of an object");
2531 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2532 slab_error(cachep, "constructor overwrote the"
2533 " start of an object");
2535 /* need to poison the objs? */
2536 if (cachep->flags & SLAB_POISON) {
2537 poison_obj(cachep, objp, POISON_FREE);
2538 slab_kernel_map(cachep, objp, 0, 0);
2544 static void cache_init_objs(struct kmem_cache *cachep,
2549 cache_init_objs_debug(cachep, page);
2551 if (OBJFREELIST_SLAB(cachep)) {
2552 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2556 for (i = 0; i < cachep->num; i++) {
2557 /* constructor could break poison info */
2558 if (DEBUG == 0 && cachep->ctor)
2559 cachep->ctor(index_to_obj(cachep, page, i));
2561 set_free_obj(page, i, i);
2565 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2567 if (CONFIG_ZONE_DMA_FLAG) {
2568 if (flags & GFP_DMA)
2569 BUG_ON(!(cachep->allocflags & GFP_DMA));
2571 BUG_ON(cachep->allocflags & GFP_DMA);
2575 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2579 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2583 if (cachep->flags & SLAB_STORE_USER)
2584 set_store_user_dirty(cachep);
2590 static void slab_put_obj(struct kmem_cache *cachep,
2591 struct page *page, void *objp)
2593 unsigned int objnr = obj_to_index(cachep, page, objp);
2597 /* Verify double free bug */
2598 for (i = page->active; i < cachep->num; i++) {
2599 if (get_free_obj(page, i) == objnr) {
2600 printk(KERN_ERR "slab: double free detected in cache "
2601 "'%s', objp %p\n", cachep->name, objp);
2607 if (!page->freelist)
2608 page->freelist = objp + obj_offset(cachep);
2610 set_free_obj(page, page->active, objnr);
2614 * Map pages beginning at addr to the given cache and slab. This is required
2615 * for the slab allocator to be able to lookup the cache and slab of a
2616 * virtual address for kfree, ksize, and slab debugging.
2618 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2621 page->slab_cache = cache;
2622 page->freelist = freelist;
2626 * Grow (by 1) the number of slabs within a cache. This is called by
2627 * kmem_cache_alloc() when there are no active objs left in a cache.
2629 static int cache_grow(struct kmem_cache *cachep,
2630 gfp_t flags, int nodeid, struct page *page)
2635 struct kmem_cache_node *n;
2638 * Be lazy and only check for valid flags here, keeping it out of the
2639 * critical path in kmem_cache_alloc().
2641 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2642 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2645 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2647 /* Take the node list lock to change the colour_next on this node */
2649 n = get_node(cachep, nodeid);
2650 spin_lock(&n->list_lock);
2652 /* Get colour for the slab, and cal the next value. */
2653 offset = n->colour_next;
2655 if (n->colour_next >= cachep->colour)
2657 spin_unlock(&n->list_lock);
2659 offset *= cachep->colour_off;
2661 if (gfpflags_allow_blocking(local_flags))
2665 * The test for missing atomic flag is performed here, rather than
2666 * the more obvious place, simply to reduce the critical path length
2667 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2668 * will eventually be caught here (where it matters).
2670 kmem_flagcheck(cachep, flags);
2673 * Get mem for the objs. Attempt to allocate a physical page from
2677 page = kmem_getpages(cachep, local_flags, nodeid);
2681 /* Get slab management. */
2682 freelist = alloc_slabmgmt(cachep, page, offset,
2683 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2684 if (OFF_SLAB(cachep) && !freelist)
2687 slab_map_pages(cachep, page, freelist);
2689 cache_init_objs(cachep, page);
2691 if (gfpflags_allow_blocking(local_flags))
2692 local_irq_disable();
2694 spin_lock(&n->list_lock);
2696 /* Make slab active. */
2697 list_add_tail(&page->lru, &(n->slabs_free));
2698 STATS_INC_GROWN(cachep);
2699 n->free_objects += cachep->num;
2700 spin_unlock(&n->list_lock);
2703 kmem_freepages(cachep, page);
2705 if (gfpflags_allow_blocking(local_flags))
2706 local_irq_disable();
2713 * Perform extra freeing checks:
2714 * - detect bad pointers.
2715 * - POISON/RED_ZONE checking
2717 static void kfree_debugcheck(const void *objp)
2719 if (!virt_addr_valid(objp)) {
2720 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2721 (unsigned long)objp);
2726 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2728 unsigned long long redzone1, redzone2;
2730 redzone1 = *dbg_redzone1(cache, obj);
2731 redzone2 = *dbg_redzone2(cache, obj);
2736 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2739 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2740 slab_error(cache, "double free detected");
2742 slab_error(cache, "memory outside object was overwritten");
2744 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2745 obj, redzone1, redzone2);
2748 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2749 unsigned long caller)
2754 BUG_ON(virt_to_cache(objp) != cachep);
2756 objp -= obj_offset(cachep);
2757 kfree_debugcheck(objp);
2758 page = virt_to_head_page(objp);
2760 if (cachep->flags & SLAB_RED_ZONE) {
2761 verify_redzone_free(cachep, objp);
2762 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2763 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2765 if (cachep->flags & SLAB_STORE_USER) {
2766 set_store_user_dirty(cachep);
2767 *dbg_userword(cachep, objp) = (void *)caller;
2770 objnr = obj_to_index(cachep, page, objp);
2772 BUG_ON(objnr >= cachep->num);
2773 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2775 if (cachep->flags & SLAB_POISON) {
2776 poison_obj(cachep, objp, POISON_FREE);
2777 slab_kernel_map(cachep, objp, 0, caller);
2783 #define kfree_debugcheck(x) do { } while(0)
2784 #define cache_free_debugcheck(x,objp,z) (objp)
2787 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2795 objp = next - obj_offset(cachep);
2796 next = *(void **)next;
2797 poison_obj(cachep, objp, POISON_FREE);
2802 static inline void fixup_slab_list(struct kmem_cache *cachep,
2803 struct kmem_cache_node *n, struct page *page,
2806 /* move slabp to correct slabp list: */
2807 list_del(&page->lru);
2808 if (page->active == cachep->num) {
2809 list_add(&page->lru, &n->slabs_full);
2810 if (OBJFREELIST_SLAB(cachep)) {
2812 /* Poisoning will be done without holding the lock */
2813 if (cachep->flags & SLAB_POISON) {
2814 void **objp = page->freelist;
2820 page->freelist = NULL;
2823 list_add(&page->lru, &n->slabs_partial);
2826 static struct page *get_first_slab(struct kmem_cache_node *n)
2830 page = list_first_entry_or_null(&n->slabs_partial,
2833 n->free_touched = 1;
2834 page = list_first_entry_or_null(&n->slabs_free,
2841 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2845 struct kmem_cache_node *n;
2846 struct array_cache *ac;
2851 node = numa_mem_id();
2852 if (unlikely(force_refill))
2855 ac = cpu_cache_get(cachep);
2856 batchcount = ac->batchcount;
2857 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2859 * If there was little recent activity on this cache, then
2860 * perform only a partial refill. Otherwise we could generate
2863 batchcount = BATCHREFILL_LIMIT;
2865 n = get_node(cachep, node);
2867 BUG_ON(ac->avail > 0 || !n);
2868 spin_lock(&n->list_lock);
2870 /* See if we can refill from the shared array */
2871 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2872 n->shared->touched = 1;
2876 while (batchcount > 0) {
2878 /* Get slab alloc is to come from. */
2879 page = get_first_slab(n);
2883 check_spinlock_acquired(cachep);
2886 * The slab was either on partial or free list so
2887 * there must be at least one object available for
2890 BUG_ON(page->active >= cachep->num);
2892 while (page->active < cachep->num && batchcount--) {
2893 STATS_INC_ALLOCED(cachep);
2894 STATS_INC_ACTIVE(cachep);
2895 STATS_SET_HIGH(cachep);
2897 ac_put_obj(cachep, ac, slab_get_obj(cachep, page));
2900 fixup_slab_list(cachep, n, page, &list);
2904 n->free_objects -= ac->avail;
2906 spin_unlock(&n->list_lock);
2907 fixup_objfreelist_debug(cachep, &list);
2909 if (unlikely(!ac->avail)) {
2912 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2914 /* cache_grow can reenable interrupts, then ac could change. */
2915 ac = cpu_cache_get(cachep);
2916 node = numa_mem_id();
2918 /* no objects in sight? abort */
2919 if (!x && (ac->avail == 0 || force_refill))
2922 if (!ac->avail) /* objects refilled by interrupt? */
2927 return ac_get_obj(cachep, ac, flags, force_refill);
2930 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2933 might_sleep_if(gfpflags_allow_blocking(flags));
2935 kmem_flagcheck(cachep, flags);
2940 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2941 gfp_t flags, void *objp, unsigned long caller)
2945 if (cachep->flags & SLAB_POISON) {
2946 check_poison_obj(cachep, objp);
2947 slab_kernel_map(cachep, objp, 1, 0);
2948 poison_obj(cachep, objp, POISON_INUSE);
2950 if (cachep->flags & SLAB_STORE_USER)
2951 *dbg_userword(cachep, objp) = (void *)caller;
2953 if (cachep->flags & SLAB_RED_ZONE) {
2954 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2955 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2956 slab_error(cachep, "double free, or memory outside"
2957 " object was overwritten");
2959 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2960 objp, *dbg_redzone1(cachep, objp),
2961 *dbg_redzone2(cachep, objp));
2963 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2964 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2967 objp += obj_offset(cachep);
2968 if (cachep->ctor && cachep->flags & SLAB_POISON)
2970 if (ARCH_SLAB_MINALIGN &&
2971 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2972 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2973 objp, (int)ARCH_SLAB_MINALIGN);
2978 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2981 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2984 struct array_cache *ac;
2985 bool force_refill = false;
2989 ac = cpu_cache_get(cachep);
2990 if (likely(ac->avail)) {
2992 objp = ac_get_obj(cachep, ac, flags, false);
2995 * Allow for the possibility all avail objects are not allowed
2996 * by the current flags
2999 STATS_INC_ALLOCHIT(cachep);
3002 force_refill = true;
3005 STATS_INC_ALLOCMISS(cachep);
3006 objp = cache_alloc_refill(cachep, flags, force_refill);
3008 * the 'ac' may be updated by cache_alloc_refill(),
3009 * and kmemleak_erase() requires its correct value.
3011 ac = cpu_cache_get(cachep);
3015 * To avoid a false negative, if an object that is in one of the
3016 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3017 * treat the array pointers as a reference to the object.
3020 kmemleak_erase(&ac->entry[ac->avail]);
3026 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3028 * If we are in_interrupt, then process context, including cpusets and
3029 * mempolicy, may not apply and should not be used for allocation policy.
3031 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3033 int nid_alloc, nid_here;
3035 if (in_interrupt() || (flags & __GFP_THISNODE))
3037 nid_alloc = nid_here = numa_mem_id();
3038 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3039 nid_alloc = cpuset_slab_spread_node();
3040 else if (current->mempolicy)
3041 nid_alloc = mempolicy_slab_node();
3042 if (nid_alloc != nid_here)
3043 return ____cache_alloc_node(cachep, flags, nid_alloc);
3048 * Fallback function if there was no memory available and no objects on a
3049 * certain node and fall back is permitted. First we scan all the
3050 * available node for available objects. If that fails then we
3051 * perform an allocation without specifying a node. This allows the page
3052 * allocator to do its reclaim / fallback magic. We then insert the
3053 * slab into the proper nodelist and then allocate from it.
3055 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3057 struct zonelist *zonelist;
3061 enum zone_type high_zoneidx = gfp_zone(flags);
3064 unsigned int cpuset_mems_cookie;
3066 if (flags & __GFP_THISNODE)
3069 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3072 cpuset_mems_cookie = read_mems_allowed_begin();
3073 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3077 * Look through allowed nodes for objects available
3078 * from existing per node queues.
3080 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3081 nid = zone_to_nid(zone);
3083 if (cpuset_zone_allowed(zone, flags) &&
3084 get_node(cache, nid) &&
3085 get_node(cache, nid)->free_objects) {
3086 obj = ____cache_alloc_node(cache,
3087 gfp_exact_node(flags), nid);
3095 * This allocation will be performed within the constraints
3096 * of the current cpuset / memory policy requirements.
3097 * We may trigger various forms of reclaim on the allowed
3098 * set and go into memory reserves if necessary.
3102 if (gfpflags_allow_blocking(local_flags))
3104 kmem_flagcheck(cache, flags);
3105 page = kmem_getpages(cache, local_flags, numa_mem_id());
3106 if (gfpflags_allow_blocking(local_flags))
3107 local_irq_disable();
3110 * Insert into the appropriate per node queues
3112 nid = page_to_nid(page);
3113 if (cache_grow(cache, flags, nid, page)) {
3114 obj = ____cache_alloc_node(cache,
3115 gfp_exact_node(flags), nid);
3118 * Another processor may allocate the
3119 * objects in the slab since we are
3120 * not holding any locks.
3124 /* cache_grow already freed obj */
3130 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3136 * A interface to enable slab creation on nodeid
3138 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3142 struct kmem_cache_node *n;
3147 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3148 n = get_node(cachep, nodeid);
3153 spin_lock(&n->list_lock);
3154 page = get_first_slab(n);
3158 check_spinlock_acquired_node(cachep, nodeid);
3160 STATS_INC_NODEALLOCS(cachep);
3161 STATS_INC_ACTIVE(cachep);
3162 STATS_SET_HIGH(cachep);
3164 BUG_ON(page->active == cachep->num);
3166 obj = slab_get_obj(cachep, page);
3169 fixup_slab_list(cachep, n, page, &list);
3171 spin_unlock(&n->list_lock);
3172 fixup_objfreelist_debug(cachep, &list);
3176 spin_unlock(&n->list_lock);
3177 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3181 return fallback_alloc(cachep, flags);
3187 static __always_inline void *
3188 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3189 unsigned long caller)
3191 unsigned long save_flags;
3193 int slab_node = numa_mem_id();
3195 flags &= gfp_allowed_mask;
3196 cachep = slab_pre_alloc_hook(cachep, flags);
3197 if (unlikely(!cachep))
3200 cache_alloc_debugcheck_before(cachep, flags);
3201 local_irq_save(save_flags);
3203 if (nodeid == NUMA_NO_NODE)
3206 if (unlikely(!get_node(cachep, nodeid))) {
3207 /* Node not bootstrapped yet */
3208 ptr = fallback_alloc(cachep, flags);
3212 if (nodeid == slab_node) {
3214 * Use the locally cached objects if possible.
3215 * However ____cache_alloc does not allow fallback
3216 * to other nodes. It may fail while we still have
3217 * objects on other nodes available.
3219 ptr = ____cache_alloc(cachep, flags);
3223 /* ___cache_alloc_node can fall back to other nodes */
3224 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3226 local_irq_restore(save_flags);
3227 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3229 if (unlikely(flags & __GFP_ZERO) && ptr)
3230 memset(ptr, 0, cachep->object_size);
3232 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3236 static __always_inline void *
3237 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3241 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3242 objp = alternate_node_alloc(cache, flags);
3246 objp = ____cache_alloc(cache, flags);
3249 * We may just have run out of memory on the local node.
3250 * ____cache_alloc_node() knows how to locate memory on other nodes
3253 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3260 static __always_inline void *
3261 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3263 return ____cache_alloc(cachep, flags);
3266 #endif /* CONFIG_NUMA */
3268 static __always_inline void *
3269 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3271 unsigned long save_flags;
3274 flags &= gfp_allowed_mask;
3275 cachep = slab_pre_alloc_hook(cachep, flags);
3276 if (unlikely(!cachep))
3279 cache_alloc_debugcheck_before(cachep, flags);
3280 local_irq_save(save_flags);
3281 objp = __do_cache_alloc(cachep, flags);
3282 local_irq_restore(save_flags);
3283 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3286 if (unlikely(flags & __GFP_ZERO) && objp)
3287 memset(objp, 0, cachep->object_size);
3289 slab_post_alloc_hook(cachep, flags, 1, &objp);
3294 * Caller needs to acquire correct kmem_cache_node's list_lock
3295 * @list: List of detached free slabs should be freed by caller
3297 static void free_block(struct kmem_cache *cachep, void **objpp,
3298 int nr_objects, int node, struct list_head *list)
3301 struct kmem_cache_node *n = get_node(cachep, node);
3303 for (i = 0; i < nr_objects; i++) {
3307 clear_obj_pfmemalloc(&objpp[i]);
3310 page = virt_to_head_page(objp);
3311 list_del(&page->lru);
3312 check_spinlock_acquired_node(cachep, node);
3313 slab_put_obj(cachep, page, objp);
3314 STATS_DEC_ACTIVE(cachep);
3317 /* fixup slab chains */
3318 if (page->active == 0) {
3319 if (n->free_objects > n->free_limit) {
3320 n->free_objects -= cachep->num;
3321 list_add_tail(&page->lru, list);
3323 list_add(&page->lru, &n->slabs_free);
3326 /* Unconditionally move a slab to the end of the
3327 * partial list on free - maximum time for the
3328 * other objects to be freed, too.
3330 list_add_tail(&page->lru, &n->slabs_partial);
3335 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3338 struct kmem_cache_node *n;
3339 int node = numa_mem_id();
3342 batchcount = ac->batchcount;
3345 n = get_node(cachep, node);
3346 spin_lock(&n->list_lock);
3348 struct array_cache *shared_array = n->shared;
3349 int max = shared_array->limit - shared_array->avail;
3351 if (batchcount > max)
3353 memcpy(&(shared_array->entry[shared_array->avail]),
3354 ac->entry, sizeof(void *) * batchcount);
3355 shared_array->avail += batchcount;
3360 free_block(cachep, ac->entry, batchcount, node, &list);
3367 list_for_each_entry(page, &n->slabs_free, lru) {
3368 BUG_ON(page->active);
3372 STATS_SET_FREEABLE(cachep, i);
3375 spin_unlock(&n->list_lock);
3376 slabs_destroy(cachep, &list);
3377 ac->avail -= batchcount;
3378 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3382 * Release an obj back to its cache. If the obj has a constructed state, it must
3383 * be in this state _before_ it is released. Called with disabled ints.
3385 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3386 unsigned long caller)
3388 struct array_cache *ac = cpu_cache_get(cachep);
3391 kmemleak_free_recursive(objp, cachep->flags);
3392 objp = cache_free_debugcheck(cachep, objp, caller);
3394 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3397 * Skip calling cache_free_alien() when the platform is not numa.
3398 * This will avoid cache misses that happen while accessing slabp (which
3399 * is per page memory reference) to get nodeid. Instead use a global
3400 * variable to skip the call, which is mostly likely to be present in
3403 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3406 if (ac->avail < ac->limit) {
3407 STATS_INC_FREEHIT(cachep);
3409 STATS_INC_FREEMISS(cachep);
3410 cache_flusharray(cachep, ac);
3413 ac_put_obj(cachep, ac, objp);
3417 * kmem_cache_alloc - Allocate an object
3418 * @cachep: The cache to allocate from.
3419 * @flags: See kmalloc().
3421 * Allocate an object from this cache. The flags are only relevant
3422 * if the cache has no available objects.
3424 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3426 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3428 trace_kmem_cache_alloc(_RET_IP_, ret,
3429 cachep->object_size, cachep->size, flags);
3433 EXPORT_SYMBOL(kmem_cache_alloc);
3435 static __always_inline void
3436 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3437 size_t size, void **p, unsigned long caller)
3441 for (i = 0; i < size; i++)
3442 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3445 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3450 s = slab_pre_alloc_hook(s, flags);
3454 cache_alloc_debugcheck_before(s, flags);
3456 local_irq_disable();
3457 for (i = 0; i < size; i++) {
3458 void *objp = __do_cache_alloc(s, flags);
3460 if (unlikely(!objp))
3466 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3468 /* Clear memory outside IRQ disabled section */
3469 if (unlikely(flags & __GFP_ZERO))
3470 for (i = 0; i < size; i++)
3471 memset(p[i], 0, s->object_size);
3473 slab_post_alloc_hook(s, flags, size, p);
3474 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3478 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3479 slab_post_alloc_hook(s, flags, i, p);
3480 __kmem_cache_free_bulk(s, i, p);
3483 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3485 #ifdef CONFIG_TRACING
3487 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3491 ret = slab_alloc(cachep, flags, _RET_IP_);
3493 trace_kmalloc(_RET_IP_, ret,
3494 size, cachep->size, flags);
3497 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3502 * kmem_cache_alloc_node - Allocate an object on the specified node
3503 * @cachep: The cache to allocate from.
3504 * @flags: See kmalloc().
3505 * @nodeid: node number of the target node.
3507 * Identical to kmem_cache_alloc but it will allocate memory on the given
3508 * node, which can improve the performance for cpu bound structures.
3510 * Fallback to other node is possible if __GFP_THISNODE is not set.
3512 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3514 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3516 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3517 cachep->object_size, cachep->size,
3522 EXPORT_SYMBOL(kmem_cache_alloc_node);
3524 #ifdef CONFIG_TRACING
3525 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3532 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3534 trace_kmalloc_node(_RET_IP_, ret,
3539 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3542 static __always_inline void *
3543 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3545 struct kmem_cache *cachep;
3547 cachep = kmalloc_slab(size, flags);
3548 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3550 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3553 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3555 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3557 EXPORT_SYMBOL(__kmalloc_node);
3559 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3560 int node, unsigned long caller)
3562 return __do_kmalloc_node(size, flags, node, caller);
3564 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3565 #endif /* CONFIG_NUMA */
3568 * __do_kmalloc - allocate memory
3569 * @size: how many bytes of memory are required.
3570 * @flags: the type of memory to allocate (see kmalloc).
3571 * @caller: function caller for debug tracking of the caller
3573 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3574 unsigned long caller)
3576 struct kmem_cache *cachep;
3579 cachep = kmalloc_slab(size, flags);
3580 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3582 ret = slab_alloc(cachep, flags, caller);
3584 trace_kmalloc(caller, ret,
3585 size, cachep->size, flags);
3590 void *__kmalloc(size_t size, gfp_t flags)
3592 return __do_kmalloc(size, flags, _RET_IP_);
3594 EXPORT_SYMBOL(__kmalloc);
3596 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3598 return __do_kmalloc(size, flags, caller);
3600 EXPORT_SYMBOL(__kmalloc_track_caller);
3603 * kmem_cache_free - Deallocate an object
3604 * @cachep: The cache the allocation was from.
3605 * @objp: The previously allocated object.
3607 * Free an object which was previously allocated from this
3610 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3612 unsigned long flags;
3613 cachep = cache_from_obj(cachep, objp);
3617 local_irq_save(flags);
3618 debug_check_no_locks_freed(objp, cachep->object_size);
3619 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3620 debug_check_no_obj_freed(objp, cachep->object_size);
3621 __cache_free(cachep, objp, _RET_IP_);
3622 local_irq_restore(flags);
3624 trace_kmem_cache_free(_RET_IP_, objp);
3626 EXPORT_SYMBOL(kmem_cache_free);
3628 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3630 struct kmem_cache *s;
3633 local_irq_disable();
3634 for (i = 0; i < size; i++) {
3637 if (!orig_s) /* called via kfree_bulk */
3638 s = virt_to_cache(objp);
3640 s = cache_from_obj(orig_s, objp);
3642 debug_check_no_locks_freed(objp, s->object_size);
3643 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3644 debug_check_no_obj_freed(objp, s->object_size);
3646 __cache_free(s, objp, _RET_IP_);
3650 /* FIXME: add tracing */
3652 EXPORT_SYMBOL(kmem_cache_free_bulk);
3655 * kfree - free previously allocated memory
3656 * @objp: pointer returned by kmalloc.
3658 * If @objp is NULL, no operation is performed.
3660 * Don't free memory not originally allocated by kmalloc()
3661 * or you will run into trouble.
3663 void kfree(const void *objp)
3665 struct kmem_cache *c;
3666 unsigned long flags;
3668 trace_kfree(_RET_IP_, objp);
3670 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3672 local_irq_save(flags);
3673 kfree_debugcheck(objp);
3674 c = virt_to_cache(objp);
3675 debug_check_no_locks_freed(objp, c->object_size);
3677 debug_check_no_obj_freed(objp, c->object_size);
3678 __cache_free(c, (void *)objp, _RET_IP_);
3679 local_irq_restore(flags);
3681 EXPORT_SYMBOL(kfree);
3684 * This initializes kmem_cache_node or resizes various caches for all nodes.
3686 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3689 struct kmem_cache_node *n;
3690 struct array_cache *new_shared;
3691 struct alien_cache **new_alien = NULL;
3693 for_each_online_node(node) {
3695 if (use_alien_caches) {
3696 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3702 if (cachep->shared) {
3703 new_shared = alloc_arraycache(node,
3704 cachep->shared*cachep->batchcount,
3707 free_alien_cache(new_alien);
3712 n = get_node(cachep, node);
3714 struct array_cache *shared = n->shared;
3717 spin_lock_irq(&n->list_lock);
3720 free_block(cachep, shared->entry,
3721 shared->avail, node, &list);
3723 n->shared = new_shared;
3725 n->alien = new_alien;
3728 n->free_limit = (1 + nr_cpus_node(node)) *
3729 cachep->batchcount + cachep->num;
3730 spin_unlock_irq(&n->list_lock);
3731 slabs_destroy(cachep, &list);
3733 free_alien_cache(new_alien);
3736 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3738 free_alien_cache(new_alien);
3743 kmem_cache_node_init(n);
3744 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3745 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3746 n->shared = new_shared;
3747 n->alien = new_alien;
3748 n->free_limit = (1 + nr_cpus_node(node)) *
3749 cachep->batchcount + cachep->num;
3750 cachep->node[node] = n;
3755 if (!cachep->list.next) {
3756 /* Cache is not active yet. Roll back what we did */
3759 n = get_node(cachep, node);
3762 free_alien_cache(n->alien);
3764 cachep->node[node] = NULL;
3772 /* Always called with the slab_mutex held */
3773 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3774 int batchcount, int shared, gfp_t gfp)
3776 struct array_cache __percpu *cpu_cache, *prev;
3779 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3783 prev = cachep->cpu_cache;
3784 cachep->cpu_cache = cpu_cache;
3785 kick_all_cpus_sync();
3788 cachep->batchcount = batchcount;
3789 cachep->limit = limit;
3790 cachep->shared = shared;
3795 for_each_online_cpu(cpu) {
3798 struct kmem_cache_node *n;
3799 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3801 node = cpu_to_mem(cpu);
3802 n = get_node(cachep, node);
3803 spin_lock_irq(&n->list_lock);
3804 free_block(cachep, ac->entry, ac->avail, node, &list);
3805 spin_unlock_irq(&n->list_lock);
3806 slabs_destroy(cachep, &list);
3811 return alloc_kmem_cache_node(cachep, gfp);
3814 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3815 int batchcount, int shared, gfp_t gfp)
3818 struct kmem_cache *c;
3820 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3822 if (slab_state < FULL)
3825 if ((ret < 0) || !is_root_cache(cachep))
3828 lockdep_assert_held(&slab_mutex);
3829 for_each_memcg_cache(c, cachep) {
3830 /* return value determined by the root cache only */
3831 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3837 /* Called with slab_mutex held always */
3838 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3845 if (!is_root_cache(cachep)) {
3846 struct kmem_cache *root = memcg_root_cache(cachep);
3847 limit = root->limit;
3848 shared = root->shared;
3849 batchcount = root->batchcount;
3852 if (limit && shared && batchcount)
3855 * The head array serves three purposes:
3856 * - create a LIFO ordering, i.e. return objects that are cache-warm
3857 * - reduce the number of spinlock operations.
3858 * - reduce the number of linked list operations on the slab and
3859 * bufctl chains: array operations are cheaper.
3860 * The numbers are guessed, we should auto-tune as described by
3863 if (cachep->size > 131072)
3865 else if (cachep->size > PAGE_SIZE)
3867 else if (cachep->size > 1024)
3869 else if (cachep->size > 256)
3875 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3876 * allocation behaviour: Most allocs on one cpu, most free operations
3877 * on another cpu. For these cases, an efficient object passing between
3878 * cpus is necessary. This is provided by a shared array. The array
3879 * replaces Bonwick's magazine layer.
3880 * On uniprocessor, it's functionally equivalent (but less efficient)
3881 * to a larger limit. Thus disabled by default.
3884 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3889 * With debugging enabled, large batchcount lead to excessively long
3890 * periods with disabled local interrupts. Limit the batchcount
3895 batchcount = (limit + 1) / 2;
3897 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3899 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3900 cachep->name, -err);
3905 * Drain an array if it contains any elements taking the node lock only if
3906 * necessary. Note that the node listlock also protects the array_cache
3907 * if drain_array() is used on the shared array.
3909 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3910 struct array_cache *ac, int force, int node)
3915 if (!ac || !ac->avail)
3917 if (ac->touched && !force) {
3920 spin_lock_irq(&n->list_lock);
3922 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3923 if (tofree > ac->avail)
3924 tofree = (ac->avail + 1) / 2;
3925 free_block(cachep, ac->entry, tofree, node, &list);
3926 ac->avail -= tofree;
3927 memmove(ac->entry, &(ac->entry[tofree]),
3928 sizeof(void *) * ac->avail);
3930 spin_unlock_irq(&n->list_lock);
3931 slabs_destroy(cachep, &list);
3936 * cache_reap - Reclaim memory from caches.
3937 * @w: work descriptor
3939 * Called from workqueue/eventd every few seconds.
3941 * - clear the per-cpu caches for this CPU.
3942 * - return freeable pages to the main free memory pool.
3944 * If we cannot acquire the cache chain mutex then just give up - we'll try
3945 * again on the next iteration.
3947 static void cache_reap(struct work_struct *w)
3949 struct kmem_cache *searchp;
3950 struct kmem_cache_node *n;
3951 int node = numa_mem_id();
3952 struct delayed_work *work = to_delayed_work(w);
3954 if (!mutex_trylock(&slab_mutex))
3955 /* Give up. Setup the next iteration. */
3958 list_for_each_entry(searchp, &slab_caches, list) {
3962 * We only take the node lock if absolutely necessary and we
3963 * have established with reasonable certainty that
3964 * we can do some work if the lock was obtained.
3966 n = get_node(searchp, node);
3968 reap_alien(searchp, n);
3970 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3973 * These are racy checks but it does not matter
3974 * if we skip one check or scan twice.
3976 if (time_after(n->next_reap, jiffies))
3979 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3981 drain_array(searchp, n, n->shared, 0, node);
3983 if (n->free_touched)
3984 n->free_touched = 0;
3988 freed = drain_freelist(searchp, n, (n->free_limit +
3989 5 * searchp->num - 1) / (5 * searchp->num));
3990 STATS_ADD_REAPED(searchp, freed);
3996 mutex_unlock(&slab_mutex);
3999 /* Set up the next iteration */
4000 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4003 #ifdef CONFIG_SLABINFO
4004 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4007 unsigned long active_objs;
4008 unsigned long num_objs;
4009 unsigned long active_slabs = 0;
4010 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4014 struct kmem_cache_node *n;
4018 for_each_kmem_cache_node(cachep, node, n) {
4021 spin_lock_irq(&n->list_lock);
4023 list_for_each_entry(page, &n->slabs_full, lru) {
4024 if (page->active != cachep->num && !error)
4025 error = "slabs_full accounting error";
4026 active_objs += cachep->num;
4029 list_for_each_entry(page, &n->slabs_partial, lru) {
4030 if (page->active == cachep->num && !error)
4031 error = "slabs_partial accounting error";
4032 if (!page->active && !error)
4033 error = "slabs_partial accounting error";
4034 active_objs += page->active;
4037 list_for_each_entry(page, &n->slabs_free, lru) {
4038 if (page->active && !error)
4039 error = "slabs_free accounting error";
4042 free_objects += n->free_objects;
4044 shared_avail += n->shared->avail;
4046 spin_unlock_irq(&n->list_lock);
4048 num_slabs += active_slabs;
4049 num_objs = num_slabs * cachep->num;
4050 if (num_objs - active_objs != free_objects && !error)
4051 error = "free_objects accounting error";
4053 name = cachep->name;
4055 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4057 sinfo->active_objs = active_objs;
4058 sinfo->num_objs = num_objs;
4059 sinfo->active_slabs = active_slabs;
4060 sinfo->num_slabs = num_slabs;
4061 sinfo->shared_avail = shared_avail;
4062 sinfo->limit = cachep->limit;
4063 sinfo->batchcount = cachep->batchcount;
4064 sinfo->shared = cachep->shared;
4065 sinfo->objects_per_slab = cachep->num;
4066 sinfo->cache_order = cachep->gfporder;
4069 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4073 unsigned long high = cachep->high_mark;
4074 unsigned long allocs = cachep->num_allocations;
4075 unsigned long grown = cachep->grown;
4076 unsigned long reaped = cachep->reaped;
4077 unsigned long errors = cachep->errors;
4078 unsigned long max_freeable = cachep->max_freeable;
4079 unsigned long node_allocs = cachep->node_allocs;
4080 unsigned long node_frees = cachep->node_frees;
4081 unsigned long overflows = cachep->node_overflow;
4083 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4084 "%4lu %4lu %4lu %4lu %4lu",
4085 allocs, high, grown,
4086 reaped, errors, max_freeable, node_allocs,
4087 node_frees, overflows);
4091 unsigned long allochit = atomic_read(&cachep->allochit);
4092 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4093 unsigned long freehit = atomic_read(&cachep->freehit);
4094 unsigned long freemiss = atomic_read(&cachep->freemiss);
4096 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4097 allochit, allocmiss, freehit, freemiss);
4102 #define MAX_SLABINFO_WRITE 128
4104 * slabinfo_write - Tuning for the slab allocator
4106 * @buffer: user buffer
4107 * @count: data length
4110 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4111 size_t count, loff_t *ppos)
4113 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4114 int limit, batchcount, shared, res;
4115 struct kmem_cache *cachep;
4117 if (count > MAX_SLABINFO_WRITE)
4119 if (copy_from_user(&kbuf, buffer, count))
4121 kbuf[MAX_SLABINFO_WRITE] = '\0';
4123 tmp = strchr(kbuf, ' ');
4128 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4131 /* Find the cache in the chain of caches. */
4132 mutex_lock(&slab_mutex);
4134 list_for_each_entry(cachep, &slab_caches, list) {
4135 if (!strcmp(cachep->name, kbuf)) {
4136 if (limit < 1 || batchcount < 1 ||
4137 batchcount > limit || shared < 0) {
4140 res = do_tune_cpucache(cachep, limit,
4147 mutex_unlock(&slab_mutex);
4153 #ifdef CONFIG_DEBUG_SLAB_LEAK
4155 static inline int add_caller(unsigned long *n, unsigned long v)
4165 unsigned long *q = p + 2 * i;
4179 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4185 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4194 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4197 for (j = page->active; j < c->num; j++) {
4198 if (get_free_obj(page, j) == i) {
4208 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4209 * mapping is established when actual object allocation and
4210 * we could mistakenly access the unmapped object in the cpu
4213 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4216 if (!add_caller(n, v))
4221 static void show_symbol(struct seq_file *m, unsigned long address)
4223 #ifdef CONFIG_KALLSYMS
4224 unsigned long offset, size;
4225 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4227 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4228 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4230 seq_printf(m, " [%s]", modname);
4234 seq_printf(m, "%p", (void *)address);
4237 static int leaks_show(struct seq_file *m, void *p)
4239 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4241 struct kmem_cache_node *n;
4243 unsigned long *x = m->private;
4247 if (!(cachep->flags & SLAB_STORE_USER))
4249 if (!(cachep->flags & SLAB_RED_ZONE))
4253 * Set store_user_clean and start to grab stored user information
4254 * for all objects on this cache. If some alloc/free requests comes
4255 * during the processing, information would be wrong so restart
4259 set_store_user_clean(cachep);
4260 drain_cpu_caches(cachep);
4264 for_each_kmem_cache_node(cachep, node, n) {
4267 spin_lock_irq(&n->list_lock);
4269 list_for_each_entry(page, &n->slabs_full, lru)
4270 handle_slab(x, cachep, page);
4271 list_for_each_entry(page, &n->slabs_partial, lru)
4272 handle_slab(x, cachep, page);
4273 spin_unlock_irq(&n->list_lock);
4275 } while (!is_store_user_clean(cachep));
4277 name = cachep->name;
4279 /* Increase the buffer size */
4280 mutex_unlock(&slab_mutex);
4281 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4283 /* Too bad, we are really out */
4285 mutex_lock(&slab_mutex);
4288 *(unsigned long *)m->private = x[0] * 2;
4290 mutex_lock(&slab_mutex);
4291 /* Now make sure this entry will be retried */
4295 for (i = 0; i < x[1]; i++) {
4296 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4297 show_symbol(m, x[2*i+2]);
4304 static const struct seq_operations slabstats_op = {
4305 .start = slab_start,
4311 static int slabstats_open(struct inode *inode, struct file *file)
4315 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4319 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4324 static const struct file_operations proc_slabstats_operations = {
4325 .open = slabstats_open,
4327 .llseek = seq_lseek,
4328 .release = seq_release_private,
4332 static int __init slab_proc_init(void)
4334 #ifdef CONFIG_DEBUG_SLAB_LEAK
4335 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4339 module_init(slab_proc_init);
4343 * ksize - get the actual amount of memory allocated for a given object
4344 * @objp: Pointer to the object
4346 * kmalloc may internally round up allocations and return more memory
4347 * than requested. ksize() can be used to determine the actual amount of
4348 * memory allocated. The caller may use this additional memory, even though
4349 * a smaller amount of memory was initially specified with the kmalloc call.
4350 * The caller must guarantee that objp points to a valid object previously
4351 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4352 * must not be freed during the duration of the call.
4354 size_t ksize(const void *objp)
4357 if (unlikely(objp == ZERO_SIZE_PTR))
4360 return virt_to_cache(objp)->object_size;
4362 EXPORT_SYMBOL(ksize);