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 intializations 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 kmem_cache_t 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 semaphore 'cache_chain_sem'.
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.
80 #include <linux/config.h>
81 #include <linux/slab.h>
83 #include <linux/swap.h>
84 #include <linux/cache.h>
85 #include <linux/interrupt.h>
86 #include <linux/init.h>
87 #include <linux/compiler.h>
88 #include <linux/seq_file.h>
89 #include <linux/notifier.h>
90 #include <linux/kallsyms.h>
91 #include <linux/cpu.h>
92 #include <linux/sysctl.h>
93 #include <linux/module.h>
95 #include <asm/uaccess.h>
96 #include <asm/cacheflush.h>
97 #include <asm/tlbflush.h>
100 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
101 * SLAB_RED_ZONE & SLAB_POISON.
102 * 0 for faster, smaller code (especially in the critical paths).
104 * STATS - 1 to collect stats for /proc/slabinfo.
105 * 0 for faster, smaller code (especially in the critical paths).
107 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
110 #ifdef CONFIG_DEBUG_SLAB
113 #define FORCED_DEBUG 1
117 #define FORCED_DEBUG 0
121 /* Shouldn't this be in a header file somewhere? */
122 #define BYTES_PER_WORD sizeof(void *)
124 #ifndef cache_line_size
125 #define cache_line_size() L1_CACHE_BYTES
128 #ifndef ARCH_KMALLOC_MINALIGN
129 #define ARCH_KMALLOC_MINALIGN 0
132 #ifndef ARCH_KMALLOC_FLAGS
133 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
136 /* Legal flag mask for kmem_cache_create(). */
138 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
139 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
140 SLAB_NO_REAP | SLAB_CACHE_DMA | \
141 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
142 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC)
144 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
145 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
146 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC)
152 * Bufctl's are used for linking objs within a slab
155 * This implementation relies on "struct page" for locating the cache &
156 * slab an object belongs to.
157 * This allows the bufctl structure to be small (one int), but limits
158 * the number of objects a slab (not a cache) can contain when off-slab
159 * bufctls are used. The limit is the size of the largest general cache
160 * that does not use off-slab slabs.
161 * For 32bit archs with 4 kB pages, is this 56.
162 * This is not serious, as it is only for large objects, when it is unwise
163 * to have too many per slab.
164 * Note: This limit can be raised by introducing a general cache whose size
165 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
168 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
169 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
170 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
172 /* Max number of objs-per-slab for caches which use off-slab slabs.
173 * Needed to avoid a possible looping condition in cache_grow().
175 static unsigned long offslab_limit;
180 * Manages the objs in a slab. Placed either at the beginning of mem allocated
181 * for a slab, or allocated from an general cache.
182 * Slabs are chained into three list: fully used, partial, fully free slabs.
185 struct list_head list;
186 unsigned long colouroff;
187 void *s_mem; /* including colour offset */
188 unsigned int inuse; /* num of objs active in slab */
197 * - LIFO ordering, to hand out cache-warm objects from _alloc
198 * - reduce the number of linked list operations
199 * - reduce spinlock operations
201 * The limit is stored in the per-cpu structure to reduce the data cache
208 unsigned int batchcount;
209 unsigned int touched;
212 /* bootstrap: The caches do not work without cpuarrays anymore,
213 * but the cpuarrays are allocated from the generic caches...
215 #define BOOT_CPUCACHE_ENTRIES 1
216 struct arraycache_init {
217 struct array_cache cache;
218 void * entries[BOOT_CPUCACHE_ENTRIES];
222 * The slab lists of all objects.
223 * Hopefully reduce the internal fragmentation
224 * NUMA: The spinlock could be moved from the kmem_cache_t
225 * into this structure, too. Figure out what causes
226 * fewer cross-node spinlock operations.
229 struct list_head slabs_partial; /* partial list first, better asm code */
230 struct list_head slabs_full;
231 struct list_head slabs_free;
232 unsigned long free_objects;
234 unsigned long next_reap;
235 struct array_cache *shared;
238 #define LIST3_INIT(parent) \
240 .slabs_full = LIST_HEAD_INIT(parent.slabs_full), \
241 .slabs_partial = LIST_HEAD_INIT(parent.slabs_partial), \
242 .slabs_free = LIST_HEAD_INIT(parent.slabs_free) \
244 #define list3_data(cachep) \
248 #define list3_data_ptr(cachep, ptr) \
257 struct kmem_cache_s {
258 /* 1) per-cpu data, touched during every alloc/free */
259 struct array_cache *array[NR_CPUS];
260 unsigned int batchcount;
262 /* 2) touched by every alloc & free from the backend */
263 struct kmem_list3 lists;
264 /* NUMA: kmem_3list_t *nodelists[MAX_NUMNODES] */
265 unsigned int objsize;
266 unsigned int flags; /* constant flags */
267 unsigned int num; /* # of objs per slab */
268 unsigned int free_limit; /* upper limit of objects in the lists */
271 /* 3) cache_grow/shrink */
272 /* order of pgs per slab (2^n) */
273 unsigned int gfporder;
275 /* force GFP flags, e.g. GFP_DMA */
276 unsigned int gfpflags;
278 size_t colour; /* cache colouring range */
279 unsigned int colour_off; /* colour offset */
280 unsigned int colour_next; /* cache colouring */
281 kmem_cache_t *slabp_cache;
282 unsigned int slab_size;
283 unsigned int dflags; /* dynamic flags */
285 /* constructor func */
286 void (*ctor)(void *, kmem_cache_t *, unsigned long);
288 /* de-constructor func */
289 void (*dtor)(void *, kmem_cache_t *, unsigned long);
291 /* 4) cache creation/removal */
293 struct list_head next;
297 unsigned long num_active;
298 unsigned long num_allocations;
299 unsigned long high_mark;
301 unsigned long reaped;
302 unsigned long errors;
303 unsigned long max_freeable;
315 #define CFLGS_OFF_SLAB (0x80000000UL)
316 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
318 #define BATCHREFILL_LIMIT 16
319 /* Optimization question: fewer reaps means less
320 * probability for unnessary cpucache drain/refill cycles.
322 * OTHO the cpuarrays can contain lots of objects,
323 * which could lock up otherwise freeable slabs.
325 #define REAPTIMEOUT_CPUC (2*HZ)
326 #define REAPTIMEOUT_LIST3 (4*HZ)
329 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
330 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
331 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
332 #define STATS_INC_GROWN(x) ((x)->grown++)
333 #define STATS_INC_REAPED(x) ((x)->reaped++)
334 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
335 (x)->high_mark = (x)->num_active; \
337 #define STATS_INC_ERR(x) ((x)->errors++)
338 #define STATS_SET_FREEABLE(x, i) \
339 do { if ((x)->max_freeable < i) \
340 (x)->max_freeable = i; \
343 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
344 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
345 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
346 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
348 #define STATS_INC_ACTIVE(x) do { } while (0)
349 #define STATS_DEC_ACTIVE(x) do { } while (0)
350 #define STATS_INC_ALLOCED(x) do { } while (0)
351 #define STATS_INC_GROWN(x) do { } while (0)
352 #define STATS_INC_REAPED(x) do { } while (0)
353 #define STATS_SET_HIGH(x) do { } while (0)
354 #define STATS_INC_ERR(x) do { } while (0)
355 #define STATS_SET_FREEABLE(x, i) \
358 #define STATS_INC_ALLOCHIT(x) do { } while (0)
359 #define STATS_INC_ALLOCMISS(x) do { } while (0)
360 #define STATS_INC_FREEHIT(x) do { } while (0)
361 #define STATS_INC_FREEMISS(x) do { } while (0)
365 /* Magic nums for obj red zoning.
366 * Placed in the first word before and the first word after an obj.
368 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
369 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
371 /* ...and for poisoning */
372 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
373 #define POISON_FREE 0x6b /* for use-after-free poisoning */
374 #define POISON_END 0xa5 /* end-byte of poisoning */
376 /* memory layout of objects:
378 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
379 * the end of an object is aligned with the end of the real
380 * allocation. Catches writes behind the end of the allocation.
381 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
383 * cachep->dbghead: The real object.
384 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
385 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
387 static int obj_dbghead(kmem_cache_t *cachep)
389 return cachep->dbghead;
392 static int obj_reallen(kmem_cache_t *cachep)
394 return cachep->reallen;
397 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
399 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
400 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
403 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
405 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
406 if (cachep->flags & SLAB_STORE_USER)
407 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
408 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
411 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
413 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
414 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
419 #define obj_dbghead(x) 0
420 #define obj_reallen(cachep) (cachep->objsize)
421 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
422 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
423 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
428 * Maximum size of an obj (in 2^order pages)
429 * and absolute limit for the gfp order.
431 #if defined(CONFIG_LARGE_ALLOCS)
432 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
433 #define MAX_GFP_ORDER 13 /* up to 32Mb */
434 #elif defined(CONFIG_MMU)
435 #define MAX_OBJ_ORDER 5 /* 32 pages */
436 #define MAX_GFP_ORDER 5 /* 32 pages */
438 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
439 #define MAX_GFP_ORDER 8 /* up to 1Mb */
443 * Do not go above this order unless 0 objects fit into the slab.
445 #define BREAK_GFP_ORDER_HI 1
446 #define BREAK_GFP_ORDER_LO 0
447 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
449 /* Macros for storing/retrieving the cachep and or slab from the
450 * global 'mem_map'. These are used to find the slab an obj belongs to.
451 * With kfree(), these are used to find the cache which an obj belongs to.
453 #define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
454 #define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
455 #define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
456 #define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
458 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
459 struct cache_sizes malloc_sizes[] = {
460 #define CACHE(x) { .cs_size = (x) },
461 #include <linux/kmalloc_sizes.h>
466 EXPORT_SYMBOL(malloc_sizes);
468 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
474 static struct cache_names __initdata cache_names[] = {
475 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
476 #include <linux/kmalloc_sizes.h>
481 struct arraycache_init initarray_cache __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
482 struct arraycache_init initarray_generic __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
484 /* internal cache of cache description objs */
485 static kmem_cache_t cache_cache = {
486 .lists = LIST3_INIT(cache_cache.lists),
488 .limit = BOOT_CPUCACHE_ENTRIES,
489 .objsize = sizeof(kmem_cache_t),
490 .flags = SLAB_NO_REAP,
491 .spinlock = SPIN_LOCK_UNLOCKED,
492 .name = "kmem_cache",
494 .reallen = sizeof(kmem_cache_t),
498 /* Guard access to the cache-chain. */
499 static struct semaphore cache_chain_sem;
501 struct list_head cache_chain;
504 * vm_enough_memory() looks at this to determine how many
505 * slab-allocated pages are possibly freeable under pressure
507 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
509 atomic_t slab_reclaim_pages;
510 EXPORT_SYMBOL(slab_reclaim_pages);
513 * chicken and egg problem: delay the per-cpu array allocation
514 * until the general caches are up.
522 static DEFINE_PER_CPU(struct timer_list, reap_timers);
524 static void reap_timer_fnc(unsigned long data);
525 static void free_block(kmem_cache_t* cachep, void** objpp, int len);
526 static void enable_cpucache (kmem_cache_t *cachep);
528 static inline void ** ac_entry(struct array_cache *ac)
530 return (void**)(ac+1);
533 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
535 return cachep->array[smp_processor_id()];
538 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
539 static void cache_estimate (unsigned long gfporder, size_t size, size_t align,
540 int flags, size_t *left_over, unsigned int *num)
543 size_t wastage = PAGE_SIZE<<gfporder;
547 if (!(flags & CFLGS_OFF_SLAB)) {
548 base = sizeof(struct slab);
549 extra = sizeof(kmem_bufctl_t);
552 while (i*size + ALIGN(base+i*extra, align) <= wastage)
562 wastage -= ALIGN(base+i*extra, align);
563 *left_over = wastage;
566 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
568 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
570 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
571 function, cachep->name, msg);
576 * Start the reap timer running on the target CPU. We run at around 1 to 2Hz.
577 * Add the CPU number into the expiry time to minimize the possibility of the
578 * CPUs getting into lockstep and contending for the global cache chain lock.
580 static void __devinit start_cpu_timer(int cpu)
582 struct timer_list *rt = &per_cpu(reap_timers, cpu);
584 if (rt->function == NULL) {
586 rt->expires = jiffies + HZ + 3*cpu;
588 rt->function = reap_timer_fnc;
589 add_timer_on(rt, cpu);
593 #ifdef CONFIG_HOTPLUG_CPU
594 static void stop_cpu_timer(int cpu)
596 struct timer_list *rt = &per_cpu(reap_timers, cpu);
600 WARN_ON(timer_pending(rt));
606 static struct array_cache *alloc_arraycache(int cpu, int entries, int batchcount)
608 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
609 struct array_cache *nc = NULL;
612 nc = kmem_cache_alloc_node(kmem_find_general_cachep(memsize,
613 GFP_KERNEL), cpu_to_node(cpu));
616 nc = kmalloc(memsize, GFP_KERNEL);
620 nc->batchcount = batchcount;
626 static int __devinit cpuup_callback(struct notifier_block *nfb,
627 unsigned long action,
630 long cpu = (long)hcpu;
631 kmem_cache_t* cachep;
635 down(&cache_chain_sem);
636 list_for_each_entry(cachep, &cache_chain, next) {
637 struct array_cache *nc;
639 nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount);
643 spin_lock_irq(&cachep->spinlock);
644 cachep->array[cpu] = nc;
645 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
647 spin_unlock_irq(&cachep->spinlock);
650 up(&cache_chain_sem);
653 start_cpu_timer(cpu);
655 #ifdef CONFIG_HOTPLUG_CPU
659 case CPU_UP_CANCELED:
660 down(&cache_chain_sem);
662 list_for_each_entry(cachep, &cache_chain, next) {
663 struct array_cache *nc;
665 spin_lock_irq(&cachep->spinlock);
666 /* cpu is dead; no one can alloc from it. */
667 nc = cachep->array[cpu];
668 cachep->array[cpu] = NULL;
669 cachep->free_limit -= cachep->batchcount;
670 free_block(cachep, ac_entry(nc), nc->avail);
671 spin_unlock_irq(&cachep->spinlock);
674 up(&cache_chain_sem);
680 up(&cache_chain_sem);
684 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
687 * Called after the gfp() functions have been enabled, and before smp_init().
689 void __init kmem_cache_init(void)
692 struct cache_sizes *sizes;
693 struct cache_names *names;
696 * Fragmentation resistance on low memory - only use bigger
697 * page orders on machines with more than 32MB of memory.
699 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
700 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
703 /* Bootstrap is tricky, because several objects are allocated
704 * from caches that do not exist yet:
705 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
706 * structures of all caches, except cache_cache itself: cache_cache
707 * is statically allocated.
708 * Initially an __init data area is used for the head array, it's
709 * replaced with a kmalloc allocated array at the end of the bootstrap.
710 * 2) Create the first kmalloc cache.
711 * The kmem_cache_t for the new cache is allocated normally. An __init
712 * data area is used for the head array.
713 * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
714 * 4) Replace the __init data head arrays for cache_cache and the first
715 * kmalloc cache with kmalloc allocated arrays.
716 * 5) Resize the head arrays of the kmalloc caches to their final sizes.
719 /* 1) create the cache_cache */
720 init_MUTEX(&cache_chain_sem);
721 INIT_LIST_HEAD(&cache_chain);
722 list_add(&cache_cache.next, &cache_chain);
723 cache_cache.colour_off = cache_line_size();
724 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
726 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
728 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
729 &left_over, &cache_cache.num);
730 if (!cache_cache.num)
733 cache_cache.colour = left_over/cache_cache.colour_off;
734 cache_cache.colour_next = 0;
735 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
736 sizeof(struct slab), cache_line_size());
738 /* 2+3) create the kmalloc caches */
739 sizes = malloc_sizes;
742 while (sizes->cs_size) {
743 /* For performance, all the general caches are L1 aligned.
744 * This should be particularly beneficial on SMP boxes, as it
745 * eliminates "false sharing".
746 * Note for systems short on memory removing the alignment will
747 * allow tighter packing of the smaller caches. */
748 sizes->cs_cachep = kmem_cache_create(names->name,
749 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
750 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
752 /* Inc off-slab bufctl limit until the ceiling is hit. */
753 if (!(OFF_SLAB(sizes->cs_cachep))) {
754 offslab_limit = sizes->cs_size-sizeof(struct slab);
755 offslab_limit /= sizeof(kmem_bufctl_t);
758 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
759 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
760 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
766 /* 4) Replace the bootstrap head arrays */
770 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
772 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
773 memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
774 cache_cache.array[smp_processor_id()] = ptr;
777 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
779 BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
780 memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
781 sizeof(struct arraycache_init));
782 malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
786 /* 5) resize the head arrays to their final sizes */
788 kmem_cache_t *cachep;
789 down(&cache_chain_sem);
790 list_for_each_entry(cachep, &cache_chain, next)
791 enable_cpucache(cachep);
792 up(&cache_chain_sem);
796 g_cpucache_up = FULL;
798 /* Register a cpu startup notifier callback
799 * that initializes ac_data for all new cpus
801 register_cpu_notifier(&cpucache_notifier);
804 /* The reap timers are started later, with a module init call:
805 * That part of the kernel is not yet operational.
809 int __init cpucache_init(void)
814 * Register the timers that return unneeded
817 for (cpu = 0; cpu < NR_CPUS; cpu++) {
819 start_cpu_timer(cpu);
825 __initcall(cpucache_init);
828 * Interface to system's page allocator. No need to hold the cache-lock.
830 * If we requested dmaable memory, we will get it. Even if we
831 * did not request dmaable memory, we might get it, but that
832 * would be relatively rare and ignorable.
834 static void *kmem_getpages(kmem_cache_t *cachep, int flags, int nodeid)
840 flags |= cachep->gfpflags;
841 if (likely(nodeid == -1)) {
842 addr = (void*)__get_free_pages(flags, cachep->gfporder);
845 page = virt_to_page(addr);
847 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
850 addr = page_address(page);
853 i = (1 << cachep->gfporder);
854 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
855 atomic_add(i, &slab_reclaim_pages);
856 add_page_state(nr_slab, i);
865 * Interface to system's page release.
867 static void kmem_freepages(kmem_cache_t *cachep, void *addr)
869 unsigned long i = (1<<cachep->gfporder);
870 struct page *page = virt_to_page(addr);
871 const unsigned long nr_freed = i;
874 if (!TestClearPageSlab(page))
878 sub_page_state(nr_slab, nr_freed);
879 if (current->reclaim_state)
880 current->reclaim_state->reclaimed_slab += nr_freed;
881 free_pages((unsigned long)addr, cachep->gfporder);
882 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
883 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
888 #ifdef CONFIG_DEBUG_PAGEALLOC
889 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, unsigned long caller)
891 int size = obj_reallen(cachep);
893 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
895 if (size < 5*sizeof(unsigned long))
900 *addr++=smp_processor_id();
901 size -= 3*sizeof(unsigned long);
903 unsigned long *sptr = &caller;
904 unsigned long svalue;
906 while (!kstack_end(sptr)) {
908 if (kernel_text_address(svalue)) {
910 size -= sizeof(unsigned long);
911 if (size <= sizeof(unsigned long))
921 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
923 int size = obj_reallen(cachep);
924 addr = &((char*)addr)[obj_dbghead(cachep)];
926 memset(addr, val, size);
927 *(unsigned char *)(addr+size-1) = POISON_END;
930 static void dump_line(char *data, int offset, int limit)
933 printk(KERN_ERR "%03x:", offset);
934 for (i=0;i<limit;i++) {
935 printk(" %02x", (unsigned char)data[offset+i]);
941 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
947 if (cachep->flags & SLAB_RED_ZONE) {
948 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
949 *dbg_redzone1(cachep, objp),
950 *dbg_redzone2(cachep, objp));
953 if (cachep->flags & SLAB_STORE_USER) {
954 printk(KERN_ERR "Last user: [<%p>]", *dbg_userword(cachep, objp));
955 print_symbol("(%s)", (unsigned long)*dbg_userword(cachep, objp));
958 realobj = (char*)objp+obj_dbghead(cachep);
959 size = obj_reallen(cachep);
960 for (i=0; i<size && lines;i+=16, lines--) {
965 dump_line(realobj, i, limit);
972 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
978 realobj = (char*)objp+obj_dbghead(cachep);
979 size = obj_reallen(cachep);
981 for (i=0;i<size;i++) {
982 char exp = POISON_FREE;
985 if (realobj[i] != exp) {
990 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
992 print_objinfo(cachep, objp, 0);
994 /* Hexdump the affected line */
999 dump_line(realobj, i, limit);
1002 /* Limit to 5 lines */
1008 /* Print some data about the neighboring objects, if they
1011 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1014 objnr = (objp-slabp->s_mem)/cachep->objsize;
1016 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1017 realobj = (char*)objp+obj_dbghead(cachep);
1018 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1020 print_objinfo(cachep, objp, 2);
1022 if (objnr+1 < cachep->num) {
1023 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1024 realobj = (char*)objp+obj_dbghead(cachep);
1025 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1027 print_objinfo(cachep, objp, 2);
1033 /* Destroy all the objs in a slab, and release the mem back to the system.
1034 * Before calling the slab must have been unlinked from the cache.
1035 * The cache-lock is not held/needed.
1037 static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1041 for (i = 0; i < cachep->num; i++) {
1042 void *objp = slabp->s_mem + cachep->objsize * i;
1044 if (cachep->flags & SLAB_POISON) {
1045 #ifdef CONFIG_DEBUG_PAGEALLOC
1046 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1047 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1049 check_poison_obj(cachep, objp);
1051 check_poison_obj(cachep, objp);
1054 if (cachep->flags & SLAB_RED_ZONE) {
1055 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1056 slab_error(cachep, "start of a freed object "
1058 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1059 slab_error(cachep, "end of a freed object "
1062 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1063 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1068 for (i = 0; i < cachep->num; i++) {
1069 void* objp = slabp->s_mem+cachep->objsize*i;
1070 (cachep->dtor)(objp, cachep, 0);
1075 kmem_freepages(cachep, slabp->s_mem-slabp->colouroff);
1076 if (OFF_SLAB(cachep))
1077 kmem_cache_free(cachep->slabp_cache, slabp);
1081 * kmem_cache_create - Create a cache.
1082 * @name: A string which is used in /proc/slabinfo to identify this cache.
1083 * @size: The size of objects to be created in this cache.
1084 * @align: The required alignment for the objects.
1085 * @flags: SLAB flags
1086 * @ctor: A constructor for the objects.
1087 * @dtor: A destructor for the objects.
1089 * Returns a ptr to the cache on success, NULL on failure.
1090 * Cannot be called within a int, but can be interrupted.
1091 * The @ctor is run when new pages are allocated by the cache
1092 * and the @dtor is run before the pages are handed back.
1094 * @name must be valid until the cache is destroyed. This implies that
1095 * the module calling this has to destroy the cache before getting
1100 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1101 * to catch references to uninitialised memory.
1103 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1104 * for buffer overruns.
1106 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1109 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1110 * cacheline. This can be beneficial if you're counting cycles as closely
1114 kmem_cache_create (const char *name, size_t size, size_t align,
1115 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1116 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1118 size_t left_over, slab_size;
1119 kmem_cache_t *cachep = NULL;
1122 * Sanity checks... these are all serious usage bugs.
1126 (size < BYTES_PER_WORD) ||
1127 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1129 printk(KERN_ERR "%s: Early error in slab %s\n",
1130 __FUNCTION__, name);
1135 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1136 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1137 /* No constructor, but inital state check requested */
1138 printk(KERN_ERR "%s: No con, but init state check "
1139 "requested - %s\n", __FUNCTION__, name);
1140 flags &= ~SLAB_DEBUG_INITIAL;
1145 * Enable redzoning and last user accounting, except for caches with
1146 * large objects, if the increased size would increase the object size
1147 * above the next power of two: caches with object sizes just above a
1148 * power of two have a significant amount of internal fragmentation.
1150 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1151 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1152 flags |= SLAB_POISON;
1156 * Always checks flags, a caller might be expecting debug
1157 * support which isn't available.
1159 if (flags & ~CREATE_MASK)
1163 /* combinations of forced alignment and advanced debugging is
1164 * not yet implemented.
1166 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1168 if (flags & SLAB_HWCACHE_ALIGN) {
1169 /* Default alignment: as specified by the arch code.
1170 * Except if an object is really small, then squeeze multiple
1171 * into one cacheline.
1173 align = cache_line_size();
1174 while (size <= align/2)
1177 align = BYTES_PER_WORD;
1181 /* Get cache's description obj. */
1182 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1185 memset(cachep, 0, sizeof(kmem_cache_t));
1187 /* Check that size is in terms of words. This is needed to avoid
1188 * unaligned accesses for some archs when redzoning is used, and makes
1189 * sure any on-slab bufctl's are also correctly aligned.
1191 if (size & (BYTES_PER_WORD-1)) {
1192 size += (BYTES_PER_WORD-1);
1193 size &= ~(BYTES_PER_WORD-1);
1197 cachep->reallen = size;
1199 if (flags & SLAB_RED_ZONE) {
1200 /* redzoning only works with word aligned caches */
1201 align = BYTES_PER_WORD;
1203 /* add space for red zone words */
1204 cachep->dbghead += BYTES_PER_WORD;
1205 size += 2*BYTES_PER_WORD;
1207 if (flags & SLAB_STORE_USER) {
1208 /* user store requires word alignment and
1209 * one word storage behind the end of the real
1212 align = BYTES_PER_WORD;
1213 size += BYTES_PER_WORD;
1215 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1216 if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1217 cachep->dbghead += PAGE_SIZE - size;
1223 /* Determine if the slab management is 'on' or 'off' slab. */
1224 if (size >= (PAGE_SIZE>>3))
1226 * Size is large, assume best to place the slab management obj
1227 * off-slab (should allow better packing of objs).
1229 flags |= CFLGS_OFF_SLAB;
1231 size = ALIGN(size, align);
1233 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1235 * A VFS-reclaimable slab tends to have most allocations
1236 * as GFP_NOFS and we really don't want to have to be allocating
1237 * higher-order pages when we are unable to shrink dcache.
1239 cachep->gfporder = 0;
1240 cache_estimate(cachep->gfporder, size, align, flags,
1241 &left_over, &cachep->num);
1244 * Calculate size (in pages) of slabs, and the num of objs per
1245 * slab. This could be made much more intelligent. For now,
1246 * try to avoid using high page-orders for slabs. When the
1247 * gfp() funcs are more friendly towards high-order requests,
1248 * this should be changed.
1251 unsigned int break_flag = 0;
1253 cache_estimate(cachep->gfporder, size, align, flags,
1254 &left_over, &cachep->num);
1257 if (cachep->gfporder >= MAX_GFP_ORDER)
1261 if (flags & CFLGS_OFF_SLAB &&
1262 cachep->num > offslab_limit) {
1263 /* This num of objs will cause problems. */
1270 * Large num of objs is good, but v. large slabs are
1271 * currently bad for the gfp()s.
1273 if (cachep->gfporder >= slab_break_gfp_order)
1276 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1277 break; /* Acceptable internal fragmentation. */
1284 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1285 kmem_cache_free(&cache_cache, cachep);
1289 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1290 + sizeof(struct slab), align);
1293 * If the slab has been placed off-slab, and we have enough space then
1294 * move it on-slab. This is at the expense of any extra colouring.
1296 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1297 flags &= ~CFLGS_OFF_SLAB;
1298 left_over -= slab_size;
1301 if (flags & CFLGS_OFF_SLAB) {
1302 /* really off slab. No need for manual alignment */
1303 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1306 cachep->colour_off = cache_line_size();
1307 /* Offset must be a multiple of the alignment. */
1308 if (cachep->colour_off < align)
1309 cachep->colour_off = align;
1310 cachep->colour = left_over/cachep->colour_off;
1311 cachep->slab_size = slab_size;
1312 cachep->flags = flags;
1313 cachep->gfpflags = 0;
1314 if (flags & SLAB_CACHE_DMA)
1315 cachep->gfpflags |= GFP_DMA;
1316 spin_lock_init(&cachep->spinlock);
1317 cachep->objsize = size;
1319 INIT_LIST_HEAD(&cachep->lists.slabs_full);
1320 INIT_LIST_HEAD(&cachep->lists.slabs_partial);
1321 INIT_LIST_HEAD(&cachep->lists.slabs_free);
1323 if (flags & CFLGS_OFF_SLAB)
1324 cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
1325 cachep->ctor = ctor;
1326 cachep->dtor = dtor;
1327 cachep->name = name;
1329 /* Don't let CPUs to come and go */
1332 if (g_cpucache_up == FULL) {
1333 enable_cpucache(cachep);
1335 if (g_cpucache_up == NONE) {
1336 /* Note: the first kmem_cache_create must create
1337 * the cache that's used by kmalloc(24), otherwise
1338 * the creation of further caches will BUG().
1340 cachep->array[smp_processor_id()] =
1341 &initarray_generic.cache;
1342 g_cpucache_up = PARTIAL;
1344 cachep->array[smp_processor_id()] =
1345 kmalloc(sizeof(struct arraycache_init),
1348 BUG_ON(!ac_data(cachep));
1349 ac_data(cachep)->avail = 0;
1350 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1351 ac_data(cachep)->batchcount = 1;
1352 ac_data(cachep)->touched = 0;
1353 cachep->batchcount = 1;
1354 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1355 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
1359 cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
1360 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1362 /* Need the semaphore to access the chain. */
1363 down(&cache_chain_sem);
1365 struct list_head *p;
1366 mm_segment_t old_fs;
1370 list_for_each(p, &cache_chain) {
1371 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1375 * This happens when the module gets unloaded and
1376 * doesn't destroy its slab cache and noone else reuses
1377 * the vmalloc area of the module. Print a warning.
1379 #ifdef CONFIG_X86_UACCESS_INDIRECT
1380 if (__direct_get_user(tmp,pc->name)) {
1382 if (__get_user(tmp,pc->name)) {
1384 printk("SLAB: cache with size %d has lost its "
1385 "name\n", pc->objsize);
1388 if (!strcmp(pc->name,name)) {
1389 printk("kmem_cache_create: duplicate "
1391 up(&cache_chain_sem);
1392 unlock_cpu_hotplug();
1399 /* cache setup completed, link it into the list */
1400 list_add(&cachep->next, &cache_chain);
1401 up(&cache_chain_sem);
1402 unlock_cpu_hotplug();
1404 if (!cachep && (flags & SLAB_PANIC))
1405 panic("kmem_cache_create(): failed to create slab `%s'\n",
1409 EXPORT_SYMBOL(kmem_cache_create);
1412 static void check_irq_off(void)
1414 BUG_ON(!irqs_disabled());
1417 static void check_irq_on(void)
1419 BUG_ON(irqs_disabled());
1422 static void check_spinlock_acquired(kmem_cache_t *cachep)
1426 BUG_ON(spin_trylock(&cachep->spinlock));
1430 #define check_irq_off() do { } while(0)
1431 #define check_irq_on() do { } while(0)
1432 #define check_spinlock_acquired(x) do { } while(0)
1436 * Waits for all CPUs to execute func().
1438 static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1443 local_irq_disable();
1447 if (smp_call_function(func, arg, 1, 1))
1453 static void drain_array_locked(kmem_cache_t* cachep,
1454 struct array_cache *ac, int force);
1456 static void do_drain(void *arg)
1458 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1459 struct array_cache *ac;
1462 ac = ac_data(cachep);
1463 spin_lock(&cachep->spinlock);
1464 free_block(cachep, &ac_entry(ac)[0], ac->avail);
1465 spin_unlock(&cachep->spinlock);
1469 static void drain_cpu_caches(kmem_cache_t *cachep)
1471 smp_call_function_all_cpus(do_drain, cachep);
1473 spin_lock_irq(&cachep->spinlock);
1474 if (cachep->lists.shared)
1475 drain_array_locked(cachep, cachep->lists.shared, 1);
1476 spin_unlock_irq(&cachep->spinlock);
1480 /* NUMA shrink all list3s */
1481 static int __cache_shrink(kmem_cache_t *cachep)
1486 drain_cpu_caches(cachep);
1489 spin_lock_irq(&cachep->spinlock);
1492 struct list_head *p;
1494 p = cachep->lists.slabs_free.prev;
1495 if (p == &cachep->lists.slabs_free)
1498 slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
1503 list_del(&slabp->list);
1505 cachep->lists.free_objects -= cachep->num;
1506 spin_unlock_irq(&cachep->spinlock);
1507 slab_destroy(cachep, slabp);
1508 spin_lock_irq(&cachep->spinlock);
1510 ret = !list_empty(&cachep->lists.slabs_full) ||
1511 !list_empty(&cachep->lists.slabs_partial);
1512 spin_unlock_irq(&cachep->spinlock);
1517 * kmem_cache_shrink - Shrink a cache.
1518 * @cachep: The cache to shrink.
1520 * Releases as many slabs as possible for a cache.
1521 * To help debugging, a zero exit status indicates all slabs were released.
1523 int kmem_cache_shrink(kmem_cache_t *cachep)
1525 if (!cachep || in_interrupt())
1528 return __cache_shrink(cachep);
1531 EXPORT_SYMBOL(kmem_cache_shrink);
1534 * kmem_cache_destroy - delete a cache
1535 * @cachep: the cache to destroy
1537 * Remove a kmem_cache_t object from the slab cache.
1538 * Returns 0 on success.
1540 * It is expected this function will be called by a module when it is
1541 * unloaded. This will remove the cache completely, and avoid a duplicate
1542 * cache being allocated each time a module is loaded and unloaded, if the
1543 * module doesn't have persistent in-kernel storage across loads and unloads.
1545 * The cache must be empty before calling this function.
1547 * The caller must guarantee that noone will allocate memory from the cache
1548 * during the kmem_cache_destroy().
1550 int kmem_cache_destroy (kmem_cache_t * cachep)
1554 if (!cachep || in_interrupt())
1557 /* Don't let CPUs to come and go */
1560 /* Find the cache in the chain of caches. */
1561 down(&cache_chain_sem);
1563 * the chain is never empty, cache_cache is never destroyed
1565 list_del(&cachep->next);
1566 up(&cache_chain_sem);
1568 if (__cache_shrink(cachep)) {
1569 slab_error(cachep, "Can't free all objects");
1570 down(&cache_chain_sem);
1571 list_add(&cachep->next,&cache_chain);
1572 up(&cache_chain_sem);
1573 unlock_cpu_hotplug();
1577 /* no cpu_online check required here since we clear the percpu
1578 * array on cpu offline and set this to NULL.
1580 for (i = 0; i < NR_CPUS; i++)
1581 kfree(cachep->array[i]);
1583 /* NUMA: free the list3 structures */
1584 kfree(cachep->lists.shared);
1585 cachep->lists.shared = NULL;
1586 kmem_cache_free(&cache_cache, cachep);
1588 unlock_cpu_hotplug();
1593 EXPORT_SYMBOL(kmem_cache_destroy);
1595 /* Get the memory for a slab management obj. */
1596 static struct slab* alloc_slabmgmt (kmem_cache_t *cachep,
1597 void *objp, int colour_off, int local_flags)
1601 if (OFF_SLAB(cachep)) {
1602 /* Slab management obj is off-slab. */
1603 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
1607 slabp = objp+colour_off;
1608 colour_off += cachep->slab_size;
1611 slabp->colouroff = colour_off;
1612 slabp->s_mem = objp+colour_off;
1617 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
1619 return (kmem_bufctl_t *)(slabp+1);
1622 static void cache_init_objs (kmem_cache_t * cachep,
1623 struct slab * slabp, unsigned long ctor_flags)
1627 for (i = 0; i < cachep->num; i++) {
1628 void* objp = slabp->s_mem+cachep->objsize*i;
1630 /* need to poison the objs? */
1631 if (cachep->flags & SLAB_POISON)
1632 poison_obj(cachep, objp, POISON_FREE);
1633 if (cachep->flags & SLAB_STORE_USER)
1634 *dbg_userword(cachep, objp) = NULL;
1636 if (cachep->flags & SLAB_RED_ZONE) {
1637 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1638 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1641 * Constructors are not allowed to allocate memory from
1642 * the same cache which they are a constructor for.
1643 * Otherwise, deadlock. They must also be threaded.
1645 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
1646 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
1648 if (cachep->flags & SLAB_RED_ZONE) {
1649 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1650 slab_error(cachep, "constructor overwrote the"
1651 " end of an object");
1652 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1653 slab_error(cachep, "constructor overwrote the"
1654 " start of an object");
1656 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
1657 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1660 cachep->ctor(objp, cachep, ctor_flags);
1662 slab_bufctl(slabp)[i] = i+1;
1664 slab_bufctl(slabp)[i-1] = BUFCTL_END;
1668 static void kmem_flagcheck(kmem_cache_t *cachep, int flags)
1670 if (flags & SLAB_DMA) {
1671 if (!(cachep->gfpflags & GFP_DMA))
1674 if (cachep->gfpflags & GFP_DMA)
1679 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
1684 /* Nasty!!!!!! I hope this is OK. */
1685 i = 1 << cachep->gfporder;
1686 page = virt_to_page(objp);
1688 SET_PAGE_CACHE(page, cachep);
1689 SET_PAGE_SLAB(page, slabp);
1695 * Grow (by 1) the number of slabs within a cache. This is called by
1696 * kmem_cache_alloc() when there are no active objs left in a cache.
1698 static int cache_grow (kmem_cache_t * cachep, int flags)
1704 unsigned long ctor_flags;
1706 /* Be lazy and only check for valid flags here,
1707 * keeping it out of the critical path in kmem_cache_alloc().
1709 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
1711 if (flags & SLAB_NO_GROW)
1714 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
1715 local_flags = (flags & SLAB_LEVEL_MASK);
1716 if (!(local_flags & __GFP_WAIT))
1718 * Not allowed to sleep. Need to tell a constructor about
1719 * this - it might need to know...
1721 ctor_flags |= SLAB_CTOR_ATOMIC;
1723 /* About to mess with non-constant members - lock. */
1725 spin_lock(&cachep->spinlock);
1727 /* Get colour for the slab, and cal the next value. */
1728 offset = cachep->colour_next;
1729 cachep->colour_next++;
1730 if (cachep->colour_next >= cachep->colour)
1731 cachep->colour_next = 0;
1732 offset *= cachep->colour_off;
1734 spin_unlock(&cachep->spinlock);
1736 if (local_flags & __GFP_WAIT)
1740 * The test for missing atomic flag is performed here, rather than
1741 * the more obvious place, simply to reduce the critical path length
1742 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
1743 * will eventually be caught here (where it matters).
1745 kmem_flagcheck(cachep, flags);
1748 /* Get mem for the objs. */
1749 if (!(objp = kmem_getpages(cachep, flags, -1)))
1752 /* Get slab management. */
1753 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
1756 set_slab_attr(cachep, slabp, objp);
1758 cache_init_objs(cachep, slabp, ctor_flags);
1760 if (local_flags & __GFP_WAIT)
1761 local_irq_disable();
1763 spin_lock(&cachep->spinlock);
1765 /* Make slab active. */
1766 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
1767 STATS_INC_GROWN(cachep);
1768 list3_data(cachep)->free_objects += cachep->num;
1769 spin_unlock(&cachep->spinlock);
1772 kmem_freepages(cachep, objp);
1774 if (local_flags & __GFP_WAIT)
1775 local_irq_disable();
1782 * Perform extra freeing checks:
1783 * - detect bad pointers.
1784 * - POISON/RED_ZONE checking
1785 * - destructor calls, for caches with POISON+dtor
1787 static void kfree_debugcheck(const void *objp)
1791 if (!virt_addr_valid(objp)) {
1792 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
1793 (unsigned long)objp);
1796 page = virt_to_page(objp);
1797 if (!PageSlab(page)) {
1798 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
1803 static void *cache_free_debugcheck (kmem_cache_t * cachep, void * objp, void *caller)
1809 objp -= obj_dbghead(cachep);
1810 kfree_debugcheck(objp);
1811 page = virt_to_page(objp);
1813 if (GET_PAGE_CACHE(page) != cachep) {
1814 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
1815 GET_PAGE_CACHE(page),cachep);
1816 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
1817 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
1820 slabp = GET_PAGE_SLAB(page);
1822 if (cachep->flags & SLAB_RED_ZONE) {
1823 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
1824 slab_error(cachep, "double free, or memory outside"
1825 " object was overwritten");
1826 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1827 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
1829 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1830 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1832 if (cachep->flags & SLAB_STORE_USER)
1833 *dbg_userword(cachep, objp) = caller;
1835 objnr = (objp-slabp->s_mem)/cachep->objsize;
1837 BUG_ON(objnr >= cachep->num);
1838 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
1840 if (cachep->flags & SLAB_DEBUG_INITIAL) {
1841 /* Need to call the slab's constructor so the
1842 * caller can perform a verify of its state (debugging).
1843 * Called without the cache-lock held.
1845 cachep->ctor(objp+obj_dbghead(cachep),
1846 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
1848 if (cachep->flags & SLAB_POISON && cachep->dtor) {
1849 /* we want to cache poison the object,
1850 * call the destruction callback
1852 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
1854 if (cachep->flags & SLAB_POISON) {
1855 #ifdef CONFIG_DEBUG_PAGEALLOC
1856 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
1857 store_stackinfo(cachep, objp, (unsigned long)caller);
1858 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1860 poison_obj(cachep, objp, POISON_FREE);
1863 poison_obj(cachep, objp, POISON_FREE);
1869 static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
1874 check_spinlock_acquired(cachep);
1875 /* Check slab's freelist to see if this obj is there. */
1876 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
1878 if (entries > cachep->num || i < 0 || i >= cachep->num)
1881 if (entries != cachep->num - slabp->inuse) {
1884 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
1885 cachep->name, cachep->num, slabp, slabp->inuse);
1886 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
1888 printk("\n%03x:", i);
1889 printk(" %02x", ((unsigned char*)slabp)[i]);
1896 #define kfree_debugcheck(x) do { } while(0)
1897 #define cache_free_debugcheck(x,objp,z) (objp)
1898 #define check_slabp(x,y) do { } while(0)
1901 static void* cache_alloc_refill(kmem_cache_t* cachep, int flags)
1904 struct kmem_list3 *l3;
1905 struct array_cache *ac;
1908 ac = ac_data(cachep);
1910 batchcount = ac->batchcount;
1911 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
1912 /* if there was little recent activity on this
1913 * cache, then perform only a partial refill.
1914 * Otherwise we could generate refill bouncing.
1916 batchcount = BATCHREFILL_LIMIT;
1918 l3 = list3_data(cachep);
1920 BUG_ON(ac->avail > 0);
1921 spin_lock(&cachep->spinlock);
1923 struct array_cache *shared_array = l3->shared;
1924 if (shared_array->avail) {
1925 if (batchcount > shared_array->avail)
1926 batchcount = shared_array->avail;
1927 shared_array->avail -= batchcount;
1928 ac->avail = batchcount;
1929 memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
1930 sizeof(void*)*batchcount);
1931 shared_array->touched = 1;
1935 while (batchcount > 0) {
1936 struct list_head *entry;
1938 /* Get slab alloc is to come from. */
1939 entry = l3->slabs_partial.next;
1940 if (entry == &l3->slabs_partial) {
1941 l3->free_touched = 1;
1942 entry = l3->slabs_free.next;
1943 if (entry == &l3->slabs_free)
1947 slabp = list_entry(entry, struct slab, list);
1948 check_slabp(cachep, slabp);
1949 check_spinlock_acquired(cachep);
1950 while (slabp->inuse < cachep->num && batchcount--) {
1952 STATS_INC_ALLOCED(cachep);
1953 STATS_INC_ACTIVE(cachep);
1954 STATS_SET_HIGH(cachep);
1956 /* get obj pointer */
1957 ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
1960 next = slab_bufctl(slabp)[slabp->free];
1962 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
1966 check_slabp(cachep, slabp);
1968 /* move slabp to correct slabp list: */
1969 list_del(&slabp->list);
1970 if (slabp->free == BUFCTL_END)
1971 list_add(&slabp->list, &l3->slabs_full);
1973 list_add(&slabp->list, &l3->slabs_partial);
1977 l3->free_objects -= ac->avail;
1979 spin_unlock(&cachep->spinlock);
1981 if (unlikely(!ac->avail)) {
1983 x = cache_grow(cachep, flags);
1985 // cache_grow can reenable interrupts, then ac could change.
1986 ac = ac_data(cachep);
1987 if (!x && ac->avail == 0) // no objects in sight? abort
1990 if (!ac->avail) // objects refilled by interrupt?
1994 return ac_entry(ac)[--ac->avail];
1998 cache_alloc_debugcheck_before(kmem_cache_t *cachep, int flags)
2000 might_sleep_if(flags & __GFP_WAIT);
2002 kmem_flagcheck(cachep, flags);
2008 cache_alloc_debugcheck_after(kmem_cache_t *cachep,
2009 unsigned long flags, void *objp, void *caller)
2013 if (cachep->flags & SLAB_POISON) {
2014 #ifdef CONFIG_DEBUG_PAGEALLOC
2015 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2016 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2018 check_poison_obj(cachep, objp);
2020 check_poison_obj(cachep, objp);
2022 poison_obj(cachep, objp, POISON_INUSE);
2024 if (cachep->flags & SLAB_STORE_USER)
2025 *dbg_userword(cachep, objp) = caller;
2027 if (cachep->flags & SLAB_RED_ZONE) {
2028 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2029 slab_error(cachep, "double free, or memory outside"
2030 " object was overwritten");
2031 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2032 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2034 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2035 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2037 objp += obj_dbghead(cachep);
2038 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2039 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2041 if (!(flags & __GFP_WAIT))
2042 ctor_flags |= SLAB_CTOR_ATOMIC;
2044 cachep->ctor(objp, cachep, ctor_flags);
2049 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2053 static inline void * __cache_alloc (kmem_cache_t *cachep, int flags)
2055 unsigned long save_flags;
2057 struct array_cache *ac;
2059 cache_alloc_debugcheck_before(cachep, flags);
2061 local_irq_save(save_flags);
2062 ac = ac_data(cachep);
2063 if (likely(ac->avail)) {
2064 STATS_INC_ALLOCHIT(cachep);
2066 objp = ac_entry(ac)[--ac->avail];
2068 STATS_INC_ALLOCMISS(cachep);
2069 objp = cache_alloc_refill(cachep, flags);
2071 local_irq_restore(save_flags);
2072 objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0));
2077 * NUMA: different approach needed if the spinlock is moved into
2081 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2085 check_spinlock_acquired(cachep);
2087 /* NUMA: move add into loop */
2088 cachep->lists.free_objects += nr_objects;
2090 for (i = 0; i < nr_objects; i++) {
2091 void *objp = objpp[i];
2095 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2096 list_del(&slabp->list);
2097 objnr = (objp - slabp->s_mem) / cachep->objsize;
2098 check_slabp(cachep, slabp);
2100 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2101 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
2102 cachep->name, objp);
2106 slab_bufctl(slabp)[objnr] = slabp->free;
2107 slabp->free = objnr;
2108 STATS_DEC_ACTIVE(cachep);
2110 check_slabp(cachep, slabp);
2112 /* fixup slab chains */
2113 if (slabp->inuse == 0) {
2114 if (cachep->lists.free_objects > cachep->free_limit) {
2115 cachep->lists.free_objects -= cachep->num;
2116 slab_destroy(cachep, slabp);
2118 list_add(&slabp->list,
2119 &list3_data_ptr(cachep, objp)->slabs_free);
2122 /* Unconditionally move a slab to the end of the
2123 * partial list on free - maximum time for the
2124 * other objects to be freed, too.
2126 list_add_tail(&slabp->list,
2127 &list3_data_ptr(cachep, objp)->slabs_partial);
2132 static void cache_flusharray (kmem_cache_t* cachep, struct array_cache *ac)
2136 batchcount = ac->batchcount;
2138 BUG_ON(!batchcount || batchcount > ac->avail);
2141 spin_lock(&cachep->spinlock);
2142 if (cachep->lists.shared) {
2143 struct array_cache *shared_array = cachep->lists.shared;
2144 int max = shared_array->limit-shared_array->avail;
2146 if (batchcount > max)
2148 memcpy(&ac_entry(shared_array)[shared_array->avail],
2150 sizeof(void*)*batchcount);
2151 shared_array->avail += batchcount;
2156 free_block(cachep, &ac_entry(ac)[0], batchcount);
2161 struct list_head *p;
2163 p = list3_data(cachep)->slabs_free.next;
2164 while (p != &(list3_data(cachep)->slabs_free)) {
2167 slabp = list_entry(p, struct slab, list);
2168 BUG_ON(slabp->inuse);
2173 STATS_SET_FREEABLE(cachep, i);
2176 spin_unlock(&cachep->spinlock);
2177 ac->avail -= batchcount;
2178 memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
2179 sizeof(void*)*ac->avail);
2184 * Release an obj back to its cache. If the obj has a constructed
2185 * state, it must be in this state _before_ it is released.
2187 * Called with disabled ints.
2189 static inline void __cache_free (kmem_cache_t *cachep, void* objp)
2191 struct array_cache *ac = ac_data(cachep);
2194 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2196 if (likely(ac->avail < ac->limit)) {
2197 STATS_INC_FREEHIT(cachep);
2198 ac_entry(ac)[ac->avail++] = objp;
2201 STATS_INC_FREEMISS(cachep);
2202 cache_flusharray(cachep, ac);
2203 ac_entry(ac)[ac->avail++] = objp;
2208 * kmem_cache_alloc - Allocate an object
2209 * @cachep: The cache to allocate from.
2210 * @flags: See kmalloc().
2212 * Allocate an object from this cache. The flags are only relevant
2213 * if the cache has no available objects.
2215 void * kmem_cache_alloc (kmem_cache_t *cachep, int flags)
2217 return __cache_alloc(cachep, flags);
2220 EXPORT_SYMBOL(kmem_cache_alloc);
2223 * kmem_ptr_validate - check if an untrusted pointer might
2225 * @cachep: the cache we're checking against
2226 * @ptr: pointer to validate
2228 * This verifies that the untrusted pointer looks sane:
2229 * it is _not_ a guarantee that the pointer is actually
2230 * part of the slab cache in question, but it at least
2231 * validates that the pointer can be dereferenced and
2232 * looks half-way sane.
2234 * Currently only used for dentry validation.
2236 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2238 unsigned long addr = (unsigned long) ptr;
2239 unsigned long min_addr = PAGE_OFFSET;
2240 unsigned long align_mask = BYTES_PER_WORD-1;
2241 unsigned long size = cachep->objsize;
2244 if (unlikely(addr < min_addr))
2246 if (unlikely(addr > (unsigned long)high_memory - size))
2248 if (unlikely(addr & align_mask))
2250 if (unlikely(!kern_addr_valid(addr)))
2252 if (unlikely(!kern_addr_valid(addr + size - 1)))
2254 page = virt_to_page(ptr);
2255 if (unlikely(!PageSlab(page)))
2257 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2265 * kmem_cache_alloc_node - Allocate an object on the specified node
2266 * @cachep: The cache to allocate from.
2267 * @flags: See kmalloc().
2268 * @nodeid: node number of the target node.
2270 * Identical to kmem_cache_alloc, except that this function is slow
2271 * and can sleep. And it will allocate memory on the given node, which
2272 * can improve the performance for cpu bound structures.
2274 void *kmem_cache_alloc_node(kmem_cache_t *cachep, int nodeid)
2281 /* The main algorithms are not node aware, thus we have to cheat:
2282 * We bypass all caches and allocate a new slab.
2283 * The following code is a streamlined copy of cache_grow().
2286 /* Get colour for the slab, and update the next value. */
2287 spin_lock_irq(&cachep->spinlock);
2288 offset = cachep->colour_next;
2289 cachep->colour_next++;
2290 if (cachep->colour_next >= cachep->colour)
2291 cachep->colour_next = 0;
2292 offset *= cachep->colour_off;
2293 spin_unlock_irq(&cachep->spinlock);
2295 /* Get mem for the objs. */
2296 if (!(objp = kmem_getpages(cachep, GFP_KERNEL, nodeid)))
2299 /* Get slab management. */
2300 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, GFP_KERNEL)))
2303 set_slab_attr(cachep, slabp, objp);
2304 cache_init_objs(cachep, slabp, SLAB_CTOR_CONSTRUCTOR);
2306 /* The first object is ours: */
2307 objp = slabp->s_mem + slabp->free*cachep->objsize;
2309 next = slab_bufctl(slabp)[slabp->free];
2311 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2315 /* add the remaining objects into the cache */
2316 spin_lock_irq(&cachep->spinlock);
2317 check_slabp(cachep, slabp);
2318 STATS_INC_GROWN(cachep);
2319 /* Make slab active. */
2320 if (slabp->free == BUFCTL_END) {
2321 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_full));
2323 list_add_tail(&slabp->list,
2324 &(list3_data(cachep)->slabs_partial));
2325 list3_data(cachep)->free_objects += cachep->num-1;
2327 spin_unlock_irq(&cachep->spinlock);
2328 objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp,
2329 __builtin_return_address(0));
2332 kmem_freepages(cachep, objp);
2337 EXPORT_SYMBOL(kmem_cache_alloc_node);
2340 * kmalloc - allocate memory
2341 * @size: how many bytes of memory are required.
2342 * @flags: the type of memory to allocate.
2344 * kmalloc is the normal method of allocating memory
2347 * The @flags argument may be one of:
2349 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2351 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2353 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2355 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2356 * must be suitable for DMA. This can mean different things on different
2357 * platforms. For example, on i386, it means that the memory must come
2358 * from the first 16MB.
2360 void * __kmalloc (size_t size, int flags)
2362 struct cache_sizes *csizep = malloc_sizes;
2364 for (; csizep->cs_size; csizep++) {
2365 if (size > csizep->cs_size)
2368 /* This happens if someone tries to call
2369 * kmem_cache_create(), or kmalloc(), before
2370 * the generic caches are initialized.
2372 BUG_ON(csizep->cs_cachep == NULL);
2374 return __cache_alloc(flags & GFP_DMA ?
2375 csizep->cs_dmacachep : csizep->cs_cachep, flags);
2380 EXPORT_SYMBOL(__kmalloc);
2384 * __alloc_percpu - allocate one copy of the object for every present
2385 * cpu in the system, zeroing them.
2386 * Objects should be dereferenced using per_cpu_ptr/get_cpu_ptr
2389 * @size: how many bytes of memory are required.
2390 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2392 void *__alloc_percpu(size_t size, size_t align)
2395 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2400 for (i = 0; i < NR_CPUS; i++) {
2401 if (!cpu_possible(i))
2403 pdata->ptrs[i] = kmem_cache_alloc_node(
2404 kmem_find_general_cachep(size, GFP_KERNEL),
2407 if (!pdata->ptrs[i])
2409 memset(pdata->ptrs[i], 0, size);
2412 /* Catch derefs w/o wrappers */
2413 return (void *) (~(unsigned long) pdata);
2417 if (!cpu_possible(i))
2419 kfree(pdata->ptrs[i]);
2425 EXPORT_SYMBOL(__alloc_percpu);
2429 * kmem_cache_free - Deallocate an object
2430 * @cachep: The cache the allocation was from.
2431 * @objp: The previously allocated object.
2433 * Free an object which was previously allocated from this
2436 void kmem_cache_free (kmem_cache_t *cachep, void *objp)
2438 unsigned long flags;
2440 local_irq_save(flags);
2441 __cache_free(cachep, objp);
2442 local_irq_restore(flags);
2445 EXPORT_SYMBOL(kmem_cache_free);
2448 * kfree - free previously allocated memory
2449 * @objp: pointer returned by kmalloc.
2451 * Don't free memory not originally allocated by kmalloc()
2452 * or you will run into trouble.
2454 void kfree (const void *objp)
2457 unsigned long flags;
2461 local_irq_save(flags);
2462 kfree_debugcheck(objp);
2463 c = GET_PAGE_CACHE(virt_to_page(objp));
2464 __cache_free(c, (void*)objp);
2465 local_irq_restore(flags);
2468 EXPORT_SYMBOL(kfree);
2472 * free_percpu - free previously allocated percpu memory
2473 * @objp: pointer returned by alloc_percpu.
2475 * Don't free memory not originally allocated by alloc_percpu()
2476 * The complemented objp is to check for that.
2479 free_percpu(const void *objp)
2482 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
2484 for (i = 0; i < NR_CPUS; i++) {
2485 if (!cpu_possible(i))
2491 EXPORT_SYMBOL(free_percpu);
2494 unsigned int kmem_cache_size(kmem_cache_t *cachep)
2496 return obj_reallen(cachep);
2499 EXPORT_SYMBOL(kmem_cache_size);
2501 kmem_cache_t * kmem_find_general_cachep (size_t size, int gfpflags)
2503 struct cache_sizes *csizep = malloc_sizes;
2505 /* This function could be moved to the header file, and
2506 * made inline so consumers can quickly determine what
2507 * cache pointer they require.
2509 for ( ; csizep->cs_size; csizep++) {
2510 if (size > csizep->cs_size)
2514 return (gfpflags & GFP_DMA) ? csizep->cs_dmacachep : csizep->cs_cachep;
2517 EXPORT_SYMBOL(kmem_find_general_cachep);
2519 struct ccupdate_struct {
2520 kmem_cache_t *cachep;
2521 struct array_cache *new[NR_CPUS];
2524 static void do_ccupdate_local(void *info)
2526 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
2527 struct array_cache *old;
2530 old = ac_data(new->cachep);
2532 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
2533 new->new[smp_processor_id()] = old;
2537 static int do_tune_cpucache (kmem_cache_t* cachep, int limit, int batchcount, int shared)
2539 struct ccupdate_struct new;
2540 struct array_cache *new_shared;
2543 memset(&new.new,0,sizeof(new.new));
2544 for (i = 0; i < NR_CPUS; i++) {
2545 if (cpu_online(i)) {
2546 new.new[i] = alloc_arraycache(i, limit, batchcount);
2548 for (i--; i >= 0; i--) kfree(new.new[i]);
2555 new.cachep = cachep;
2557 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
2560 spin_lock_irq(&cachep->spinlock);
2561 cachep->batchcount = batchcount;
2562 cachep->limit = limit;
2563 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
2564 spin_unlock_irq(&cachep->spinlock);
2566 for (i = 0; i < NR_CPUS; i++) {
2567 struct array_cache *ccold = new.new[i];
2570 spin_lock_irq(&cachep->spinlock);
2571 free_block(cachep, ac_entry(ccold), ccold->avail);
2572 spin_unlock_irq(&cachep->spinlock);
2575 new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d);
2577 struct array_cache *old;
2579 spin_lock_irq(&cachep->spinlock);
2580 old = cachep->lists.shared;
2581 cachep->lists.shared = new_shared;
2583 free_block(cachep, ac_entry(old), old->avail);
2584 spin_unlock_irq(&cachep->spinlock);
2592 static void enable_cpucache (kmem_cache_t *cachep)
2597 /* The head array serves three purposes:
2598 * - create a LIFO ordering, i.e. return objects that are cache-warm
2599 * - reduce the number of spinlock operations.
2600 * - reduce the number of linked list operations on the slab and
2601 * bufctl chains: array operations are cheaper.
2602 * The numbers are guessed, we should auto-tune as described by
2605 if (cachep->objsize > 131072)
2607 else if (cachep->objsize > PAGE_SIZE)
2609 else if (cachep->objsize > 1024)
2611 else if (cachep->objsize > 256)
2616 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
2617 * allocation behaviour: Most allocs on one cpu, most free operations
2618 * on another cpu. For these cases, an efficient object passing between
2619 * cpus is necessary. This is provided by a shared array. The array
2620 * replaces Bonwick's magazine layer.
2621 * On uniprocessor, it's functionally equivalent (but less efficient)
2622 * to a larger limit. Thus disabled by default.
2626 if (cachep->objsize <= PAGE_SIZE)
2631 /* With debugging enabled, large batchcount lead to excessively
2632 * long periods with disabled local interrupts. Limit the
2638 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
2640 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
2641 cachep->name, -err);
2644 static void drain_array(kmem_cache_t *cachep, struct array_cache *ac)
2651 } else if (ac->avail) {
2652 tofree = (ac->limit+4)/5;
2653 if (tofree > ac->avail) {
2654 tofree = (ac->avail+1)/2;
2656 spin_lock(&cachep->spinlock);
2657 free_block(cachep, ac_entry(ac), tofree);
2658 spin_unlock(&cachep->spinlock);
2659 ac->avail -= tofree;
2660 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2661 sizeof(void*)*ac->avail);
2665 static void drain_array_locked(kmem_cache_t *cachep,
2666 struct array_cache *ac, int force)
2670 check_spinlock_acquired(cachep);
2671 if (ac->touched && !force) {
2673 } else if (ac->avail) {
2674 tofree = force ? ac->avail : (ac->limit+4)/5;
2675 if (tofree > ac->avail) {
2676 tofree = (ac->avail+1)/2;
2678 free_block(cachep, ac_entry(ac), tofree);
2679 ac->avail -= tofree;
2680 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2681 sizeof(void*)*ac->avail);
2686 * cache_reap - Reclaim memory from caches.
2688 * Called from a timer, every few seconds
2690 * - clear the per-cpu caches for this CPU.
2691 * - return freeable pages to the main free memory pool.
2693 * If we cannot acquire the cache chain semaphore then just give up - we'll
2694 * try again next timer interrupt.
2696 static void cache_reap (void)
2698 struct list_head *walk;
2701 BUG_ON(!in_interrupt());
2704 if (down_trylock(&cache_chain_sem))
2707 list_for_each(walk, &cache_chain) {
2708 kmem_cache_t *searchp;
2709 struct list_head* p;
2713 searchp = list_entry(walk, kmem_cache_t, next);
2715 if (searchp->flags & SLAB_NO_REAP)
2719 local_irq_disable();
2720 drain_array(searchp, ac_data(searchp));
2722 if(time_after(searchp->lists.next_reap, jiffies))
2725 spin_lock(&searchp->spinlock);
2726 if(time_after(searchp->lists.next_reap, jiffies)) {
2729 searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
2731 if (searchp->lists.shared)
2732 drain_array_locked(searchp, searchp->lists.shared, 0);
2734 if (searchp->lists.free_touched) {
2735 searchp->lists.free_touched = 0;
2739 tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
2741 p = list3_data(searchp)->slabs_free.next;
2742 if (p == &(list3_data(searchp)->slabs_free))
2745 slabp = list_entry(p, struct slab, list);
2746 BUG_ON(slabp->inuse);
2747 list_del(&slabp->list);
2748 STATS_INC_REAPED(searchp);
2750 /* Safe to drop the lock. The slab is no longer
2751 * linked to the cache.
2752 * searchp cannot disappear, we hold
2755 searchp->lists.free_objects -= searchp->num;
2756 spin_unlock_irq(&searchp->spinlock);
2757 slab_destroy(searchp, slabp);
2758 spin_lock_irq(&searchp->spinlock);
2759 } while(--tofree > 0);
2761 spin_unlock(&searchp->spinlock);
2768 up(&cache_chain_sem);
2772 * This is a timer handler. There is one per CPU. It is called periodially
2773 * to shrink this CPU's caches. Otherwise there could be memory tied up
2774 * for long periods (or for ever) due to load changes.
2776 static void reap_timer_fnc(unsigned long cpu)
2778 struct timer_list *rt = &__get_cpu_var(reap_timers);
2780 /* CPU hotplug can drag us off cpu: don't run on wrong CPU */
2781 if (!cpu_is_offline(cpu)) {
2783 mod_timer(rt, jiffies + REAPTIMEOUT_CPUC + cpu);
2787 #ifdef CONFIG_PROC_FS
2789 static void *s_start(struct seq_file *m, loff_t *pos)
2792 struct list_head *p;
2794 down(&cache_chain_sem);
2797 * Output format version, so at least we can change it
2798 * without _too_ many complaints.
2801 seq_puts(m, "slabinfo - version: 2.0 (statistics)\n");
2803 seq_puts(m, "slabinfo - version: 2.0\n");
2805 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
2806 seq_puts(m, " : tunables <batchcount> <limit> <sharedfactor>");
2807 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
2809 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <freelimit>");
2810 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
2814 p = cache_chain.next;
2817 if (p == &cache_chain)
2820 return list_entry(p, kmem_cache_t, next);
2823 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2825 kmem_cache_t *cachep = p;
2827 return cachep->next.next == &cache_chain ? NULL
2828 : list_entry(cachep->next.next, kmem_cache_t, next);
2831 static void s_stop(struct seq_file *m, void *p)
2833 up(&cache_chain_sem);
2836 static int s_show(struct seq_file *m, void *p)
2838 kmem_cache_t *cachep = p;
2839 struct list_head *q;
2841 unsigned long active_objs;
2842 unsigned long num_objs;
2843 unsigned long active_slabs = 0;
2844 unsigned long num_slabs;
2849 spin_lock_irq(&cachep->spinlock);
2852 list_for_each(q,&cachep->lists.slabs_full) {
2853 slabp = list_entry(q, struct slab, list);
2854 if (slabp->inuse != cachep->num && !error)
2855 error = "slabs_full accounting error";
2856 active_objs += cachep->num;
2859 list_for_each(q,&cachep->lists.slabs_partial) {
2860 slabp = list_entry(q, struct slab, list);
2861 if (slabp->inuse == cachep->num && !error)
2862 error = "slabs_partial inuse accounting error";
2863 if (!slabp->inuse && !error)
2864 error = "slabs_partial/inuse accounting error";
2865 active_objs += slabp->inuse;
2868 list_for_each(q,&cachep->lists.slabs_free) {
2869 slabp = list_entry(q, struct slab, list);
2870 if (slabp->inuse && !error)
2871 error = "slabs_free/inuse accounting error";
2874 num_slabs+=active_slabs;
2875 num_objs = num_slabs*cachep->num;
2876 if (num_objs - active_objs != cachep->lists.free_objects && !error)
2877 error = "free_objects accounting error";
2879 name = cachep->name;
2881 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
2883 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
2884 name, active_objs, num_objs, cachep->objsize,
2885 cachep->num, (1<<cachep->gfporder));
2886 seq_printf(m, " : tunables %4u %4u %4u",
2887 cachep->limit, cachep->batchcount,
2888 cachep->lists.shared->limit/cachep->batchcount);
2889 seq_printf(m, " : slabdata %6lu %6lu %6u",
2890 active_slabs, num_slabs, cachep->lists.shared->avail);
2893 unsigned long high = cachep->high_mark;
2894 unsigned long allocs = cachep->num_allocations;
2895 unsigned long grown = cachep->grown;
2896 unsigned long reaped = cachep->reaped;
2897 unsigned long errors = cachep->errors;
2898 unsigned long max_freeable = cachep->max_freeable;
2899 unsigned long free_limit = cachep->free_limit;
2901 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu",
2902 allocs, high, grown, reaped, errors,
2903 max_freeable, free_limit);
2907 unsigned long allochit = atomic_read(&cachep->allochit);
2908 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
2909 unsigned long freehit = atomic_read(&cachep->freehit);
2910 unsigned long freemiss = atomic_read(&cachep->freemiss);
2912 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
2913 allochit, allocmiss, freehit, freemiss);
2917 spin_unlock_irq(&cachep->spinlock);
2922 * slabinfo_op - iterator that generates /proc/slabinfo
2931 * num-pages-per-slab
2932 * + further values on SMP and with statistics enabled
2935 struct seq_operations slabinfo_op = {
2942 #define MAX_SLABINFO_WRITE 128
2944 * slabinfo_write - Tuning for the slab allocator
2946 * @buffer: user buffer
2947 * @count: data length
2950 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
2951 size_t count, loff_t *ppos)
2953 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
2954 int limit, batchcount, shared, res;
2955 struct list_head *p;
2957 if (count > MAX_SLABINFO_WRITE)
2959 if (copy_from_user(&kbuf, buffer, count))
2961 kbuf[MAX_SLABINFO_WRITE] = '\0';
2963 tmp = strchr(kbuf, ' ');
2968 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
2971 /* Find the cache in the chain of caches. */
2972 down(&cache_chain_sem);
2974 list_for_each(p,&cache_chain) {
2975 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
2977 if (!strcmp(cachep->name, kbuf)) {
2980 batchcount > limit ||
2984 res = do_tune_cpucache(cachep, limit, batchcount, shared);
2989 up(&cache_chain_sem);
2996 unsigned int ksize(const void *objp)
2999 unsigned long flags;
3000 unsigned int size = 0;
3002 if (likely(objp != NULL)) {
3003 local_irq_save(flags);
3004 c = GET_PAGE_CACHE(virt_to_page(objp));
3005 size = kmem_cache_size(c);
3006 local_irq_restore(flags);
3012 void ptrinfo(unsigned long addr)
3016 printk("Dumping data about address %p.\n", (void*)addr);
3017 if (!virt_addr_valid((void*)addr)) {
3018 printk("virt addr invalid.\n");
3023 pgd_t *pgd = pgd_offset_k(addr);
3025 if (pgd_none(*pgd)) {
3026 printk("No pgd.\n");
3029 pmd = pmd_offset(pgd, addr);
3030 if (pmd_none(*pmd)) {
3031 printk("No pmd.\n");
3035 if (pmd_large(*pmd)) {
3036 printk("Large page.\n");
3040 printk("normal page, pte_val 0x%llx\n",
3041 (unsigned long long)pte_val(*pte_offset_kernel(pmd, addr)));
3045 page = virt_to_page((void*)addr);
3046 printk("struct page at %p, flags %08lx\n",
3047 page, (unsigned long)page->flags);
3048 if (PageSlab(page)) {
3051 unsigned long flags;
3055 c = GET_PAGE_CACHE(page);
3056 printk("belongs to cache %s.\n",c->name);
3058 spin_lock_irqsave(&c->spinlock, flags);
3059 s = GET_PAGE_SLAB(page);
3060 printk("slabp %p with %d inuse objects (from %d).\n",
3061 s, s->inuse, c->num);
3064 objnr = (addr-(unsigned long)s->s_mem)/c->objsize;
3065 objp = s->s_mem+c->objsize*objnr;
3066 printk("points into object no %d, starting at %p, len %d.\n",
3067 objnr, objp, c->objsize);
3068 if (objnr >= c->num) {
3069 printk("Bad obj number.\n");
3071 kernel_map_pages(virt_to_page(objp),
3072 c->objsize/PAGE_SIZE, 1);
3074 print_objinfo(c, objp, 2);
3076 spin_unlock_irqrestore(&c->spinlock, flags);