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 /* Legal flag mask for kmem_cache_create(). */
134 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
135 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
136 SLAB_NO_REAP | SLAB_CACHE_DMA | \
137 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
138 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC)
140 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
141 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
142 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC)
148 * Bufctl's are used for linking objs within a slab
151 * This implementation relies on "struct page" for locating the cache &
152 * slab an object belongs to.
153 * This allows the bufctl structure to be small (one int), but limits
154 * the number of objects a slab (not a cache) can contain when off-slab
155 * bufctls are used. The limit is the size of the largest general cache
156 * that does not use off-slab slabs.
157 * For 32bit archs with 4 kB pages, is this 56.
158 * This is not serious, as it is only for large objects, when it is unwise
159 * to have too many per slab.
160 * Note: This limit can be raised by introducing a general cache whose size
161 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
164 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
165 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
166 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
168 /* Max number of objs-per-slab for caches which use off-slab slabs.
169 * Needed to avoid a possible looping condition in cache_grow().
171 static unsigned long offslab_limit;
176 * Manages the objs in a slab. Placed either at the beginning of mem allocated
177 * for a slab, or allocated from an general cache.
178 * Slabs are chained into three list: fully used, partial, fully free slabs.
181 struct list_head list;
182 unsigned long colouroff;
183 void *s_mem; /* including colour offset */
184 unsigned int inuse; /* num of objs active in slab */
193 * - LIFO ordering, to hand out cache-warm objects from _alloc
194 * - reduce the number of linked list operations
195 * - reduce spinlock operations
197 * The limit is stored in the per-cpu structure to reduce the data cache
204 unsigned int batchcount;
205 unsigned int touched;
208 /* bootstrap: The caches do not work without cpuarrays anymore,
209 * but the cpuarrays are allocated from the generic caches...
211 #define BOOT_CPUCACHE_ENTRIES 1
212 struct arraycache_init {
213 struct array_cache cache;
214 void * entries[BOOT_CPUCACHE_ENTRIES];
218 * The slab lists of all objects.
219 * Hopefully reduce the internal fragmentation
220 * NUMA: The spinlock could be moved from the kmem_cache_t
221 * into this structure, too. Figure out what causes
222 * fewer cross-node spinlock operations.
225 struct list_head slabs_partial; /* partial list first, better asm code */
226 struct list_head slabs_full;
227 struct list_head slabs_free;
228 unsigned long free_objects;
230 unsigned long next_reap;
231 struct array_cache *shared;
234 #define LIST3_INIT(parent) \
236 .slabs_full = LIST_HEAD_INIT(parent.slabs_full), \
237 .slabs_partial = LIST_HEAD_INIT(parent.slabs_partial), \
238 .slabs_free = LIST_HEAD_INIT(parent.slabs_free) \
240 #define list3_data(cachep) \
244 #define list3_data_ptr(cachep, ptr) \
253 struct kmem_cache_s {
254 /* 1) per-cpu data, touched during every alloc/free */
255 struct array_cache *array[NR_CPUS];
256 unsigned int batchcount;
258 /* 2) touched by every alloc & free from the backend */
259 struct kmem_list3 lists;
260 /* NUMA: kmem_3list_t *nodelists[MAX_NUMNODES] */
261 unsigned int objsize;
262 unsigned int flags; /* constant flags */
263 unsigned int num; /* # of objs per slab */
264 unsigned int free_limit; /* upper limit of objects in the lists */
267 /* 3) cache_grow/shrink */
268 /* order of pgs per slab (2^n) */
269 unsigned int gfporder;
271 /* force GFP flags, e.g. GFP_DMA */
272 unsigned int gfpflags;
274 size_t colour; /* cache colouring range */
275 unsigned int colour_off; /* colour offset */
276 unsigned int colour_next; /* cache colouring */
277 kmem_cache_t *slabp_cache;
278 unsigned int slab_size;
279 unsigned int dflags; /* dynamic flags */
281 /* constructor func */
282 void (*ctor)(void *, kmem_cache_t *, unsigned long);
284 /* de-constructor func */
285 void (*dtor)(void *, kmem_cache_t *, unsigned long);
287 /* 4) cache creation/removal */
289 struct list_head next;
293 unsigned long num_active;
294 unsigned long num_allocations;
295 unsigned long high_mark;
297 unsigned long reaped;
298 unsigned long errors;
299 unsigned long max_freeable;
311 #define CFLGS_OFF_SLAB (0x80000000UL)
312 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
314 #define BATCHREFILL_LIMIT 16
315 /* Optimization question: fewer reaps means less
316 * probability for unnessary cpucache drain/refill cycles.
318 * OTHO the cpuarrays can contain lots of objects,
319 * which could lock up otherwise freeable slabs.
321 #define REAPTIMEOUT_CPUC (2*HZ)
322 #define REAPTIMEOUT_LIST3 (4*HZ)
325 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
326 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
327 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
328 #define STATS_INC_GROWN(x) ((x)->grown++)
329 #define STATS_INC_REAPED(x) ((x)->reaped++)
330 #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
331 (x)->high_mark = (x)->num_active; \
333 #define STATS_INC_ERR(x) ((x)->errors++)
334 #define STATS_SET_FREEABLE(x, i) \
335 do { if ((x)->max_freeable < i) \
336 (x)->max_freeable = i; \
339 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
340 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
341 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
342 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
344 #define STATS_INC_ACTIVE(x) do { } while (0)
345 #define STATS_DEC_ACTIVE(x) do { } while (0)
346 #define STATS_INC_ALLOCED(x) do { } while (0)
347 #define STATS_INC_GROWN(x) do { } while (0)
348 #define STATS_INC_REAPED(x) do { } while (0)
349 #define STATS_SET_HIGH(x) do { } while (0)
350 #define STATS_INC_ERR(x) do { } while (0)
351 #define STATS_SET_FREEABLE(x, i) \
354 #define STATS_INC_ALLOCHIT(x) do { } while (0)
355 #define STATS_INC_ALLOCMISS(x) do { } while (0)
356 #define STATS_INC_FREEHIT(x) do { } while (0)
357 #define STATS_INC_FREEMISS(x) do { } while (0)
361 /* Magic nums for obj red zoning.
362 * Placed in the first word before and the first word after an obj.
364 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
365 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
367 /* ...and for poisoning */
368 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
369 #define POISON_FREE 0x6b /* for use-after-free poisoning */
370 #define POISON_END 0xa5 /* end-byte of poisoning */
372 /* memory layout of objects:
374 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
375 * the end of an object is aligned with the end of the real
376 * allocation. Catches writes behind the end of the allocation.
377 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
379 * cachep->dbghead: The real object.
380 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
381 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
383 static inline int obj_dbghead(kmem_cache_t *cachep)
385 return cachep->dbghead;
388 static inline int obj_reallen(kmem_cache_t *cachep)
390 return cachep->reallen;
393 static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
395 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
396 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
399 static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
401 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
402 if (cachep->flags & SLAB_STORE_USER)
403 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
404 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
407 static void **dbg_userword(kmem_cache_t *cachep, void *objp)
409 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
410 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
413 static inline int obj_dbghead(kmem_cache_t *cachep)
417 static inline int obj_reallen(kmem_cache_t *cachep)
419 return cachep->objsize;
421 static inline unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
426 static inline unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
431 static inline void **dbg_userword(kmem_cache_t *cachep, void *objp)
439 * Maximum size of an obj (in 2^order pages)
440 * and absolute limit for the gfp order.
442 #if defined(CONFIG_LARGE_ALLOCS)
443 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
444 #define MAX_GFP_ORDER 13 /* up to 32Mb */
445 #elif defined(CONFIG_MMU)
446 #define MAX_OBJ_ORDER 5 /* 32 pages */
447 #define MAX_GFP_ORDER 5 /* 32 pages */
449 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
450 #define MAX_GFP_ORDER 8 /* up to 1Mb */
454 * Do not go above this order unless 0 objects fit into the slab.
456 #define BREAK_GFP_ORDER_HI 1
457 #define BREAK_GFP_ORDER_LO 0
458 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
460 /* Macros for storing/retrieving the cachep and or slab from the
461 * global 'mem_map'. These are used to find the slab an obj belongs to.
462 * With kfree(), these are used to find the cache which an obj belongs to.
464 #define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
465 #define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
466 #define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
467 #define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
469 /* These are the default caches for kmalloc. Custom caches can have other sizes. */
470 struct cache_sizes malloc_sizes[] = {
471 #define CACHE(x) { .cs_size = (x) },
472 #include <linux/kmalloc_sizes.h>
477 EXPORT_SYMBOL(malloc_sizes);
479 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
485 static struct cache_names __initdata cache_names[] = {
486 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
487 #include <linux/kmalloc_sizes.h>
492 struct arraycache_init initarray_cache __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
493 struct arraycache_init initarray_generic __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
495 /* internal cache of cache description objs */
496 static kmem_cache_t cache_cache = {
497 .lists = LIST3_INIT(cache_cache.lists),
499 .limit = BOOT_CPUCACHE_ENTRIES,
500 .objsize = sizeof(kmem_cache_t),
501 .flags = SLAB_NO_REAP,
502 .spinlock = SPIN_LOCK_UNLOCKED,
503 .name = "kmem_cache",
505 .reallen = sizeof(kmem_cache_t),
509 /* Guard access to the cache-chain. */
510 static struct semaphore cache_chain_sem;
512 struct list_head cache_chain;
515 * vm_enough_memory() looks at this to determine how many
516 * slab-allocated pages are possibly freeable under pressure
518 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
520 atomic_t slab_reclaim_pages;
521 EXPORT_SYMBOL(slab_reclaim_pages);
524 * chicken and egg problem: delay the per-cpu array allocation
525 * until the general caches are up.
533 static DEFINE_PER_CPU(struct timer_list, reap_timers);
535 static void reap_timer_fnc(unsigned long data);
536 static void free_block(kmem_cache_t* cachep, void** objpp, int len);
537 static void enable_cpucache (kmem_cache_t *cachep);
539 static inline void ** ac_entry(struct array_cache *ac)
541 return (void**)(ac+1);
544 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
546 return cachep->array[smp_processor_id()];
549 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
550 static void cache_estimate (unsigned long gfporder, size_t size, size_t align,
551 int flags, size_t *left_over, unsigned int *num)
554 size_t wastage = PAGE_SIZE<<gfporder;
558 if (!(flags & CFLGS_OFF_SLAB)) {
559 base = sizeof(struct slab);
560 extra = sizeof(kmem_bufctl_t);
563 while (i*size + ALIGN(base+i*extra, align) <= wastage)
573 wastage -= ALIGN(base+i*extra, align);
574 *left_over = wastage;
577 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
579 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
581 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
582 function, cachep->name, msg);
587 * Start the reap timer running on the target CPU. We run at around 1 to 2Hz.
588 * Add the CPU number into the expiry time to minimize the possibility of the
589 * CPUs getting into lockstep and contending for the global cache chain lock.
591 static void __devinit start_cpu_timer(int cpu)
593 struct timer_list *rt = &per_cpu(reap_timers, cpu);
595 if (rt->function == NULL) {
597 rt->expires = jiffies + HZ + 3*cpu;
599 rt->function = reap_timer_fnc;
600 add_timer_on(rt, cpu);
604 #ifdef CONFIG_HOTPLUG_CPU
605 static void stop_cpu_timer(int cpu)
607 struct timer_list *rt = &per_cpu(reap_timers, cpu);
611 WARN_ON(timer_pending(rt));
617 static struct array_cache *alloc_arraycache(int cpu, int entries, int batchcount)
619 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
620 struct array_cache *nc = NULL;
623 nc = kmem_cache_alloc_node(kmem_find_general_cachep(memsize,
624 GFP_KERNEL), cpu_to_node(cpu));
627 nc = kmalloc(memsize, GFP_KERNEL);
631 nc->batchcount = batchcount;
637 static int __devinit cpuup_callback(struct notifier_block *nfb,
638 unsigned long action,
641 long cpu = (long)hcpu;
642 kmem_cache_t* cachep;
646 down(&cache_chain_sem);
647 list_for_each_entry(cachep, &cache_chain, next) {
648 struct array_cache *nc;
650 nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount);
654 spin_lock_irq(&cachep->spinlock);
655 cachep->array[cpu] = nc;
656 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
658 spin_unlock_irq(&cachep->spinlock);
661 up(&cache_chain_sem);
664 start_cpu_timer(cpu);
666 #ifdef CONFIG_HOTPLUG_CPU
670 case CPU_UP_CANCELED:
671 down(&cache_chain_sem);
673 list_for_each_entry(cachep, &cache_chain, next) {
674 struct array_cache *nc;
676 spin_lock_irq(&cachep->spinlock);
677 /* cpu is dead; no one can alloc from it. */
678 nc = cachep->array[cpu];
679 cachep->array[cpu] = NULL;
680 cachep->free_limit -= cachep->batchcount;
681 free_block(cachep, ac_entry(nc), nc->avail);
682 spin_unlock_irq(&cachep->spinlock);
685 up(&cache_chain_sem);
691 up(&cache_chain_sem);
695 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
698 * Called after the gfp() functions have been enabled, and before smp_init().
700 void __init kmem_cache_init(void)
703 struct cache_sizes *sizes;
704 struct cache_names *names;
707 * Fragmentation resistance on low memory - only use bigger
708 * page orders on machines with more than 32MB of memory.
710 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
711 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
714 /* Bootstrap is tricky, because several objects are allocated
715 * from caches that do not exist yet:
716 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
717 * structures of all caches, except cache_cache itself: cache_cache
718 * is statically allocated.
719 * Initially an __init data area is used for the head array, it's
720 * replaced with a kmalloc allocated array at the end of the bootstrap.
721 * 2) Create the first kmalloc cache.
722 * The kmem_cache_t for the new cache is allocated normally. An __init
723 * data area is used for the head array.
724 * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
725 * 4) Replace the __init data head arrays for cache_cache and the first
726 * kmalloc cache with kmalloc allocated arrays.
727 * 5) Resize the head arrays of the kmalloc caches to their final sizes.
730 /* 1) create the cache_cache */
731 init_MUTEX(&cache_chain_sem);
732 INIT_LIST_HEAD(&cache_chain);
733 list_add(&cache_cache.next, &cache_chain);
734 cache_cache.colour_off = cache_line_size();
735 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
737 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
739 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
740 &left_over, &cache_cache.num);
741 if (!cache_cache.num)
744 cache_cache.colour = left_over/cache_cache.colour_off;
745 cache_cache.colour_next = 0;
746 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
747 sizeof(struct slab), cache_line_size());
749 /* 2+3) create the kmalloc caches */
750 sizes = malloc_sizes;
753 while (sizes->cs_size) {
754 /* For performance, all the general caches are L1 aligned.
755 * This should be particularly beneficial on SMP boxes, as it
756 * eliminates "false sharing".
757 * Note for systems short on memory removing the alignment will
758 * allow tighter packing of the smaller caches. */
759 sizes->cs_cachep = kmem_cache_create(names->name,
760 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
761 SLAB_PANIC, NULL, NULL);
763 /* Inc off-slab bufctl limit until the ceiling is hit. */
764 if (!(OFF_SLAB(sizes->cs_cachep))) {
765 offslab_limit = sizes->cs_size-sizeof(struct slab);
766 offslab_limit /= sizeof(kmem_bufctl_t);
769 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
770 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
771 (SLAB_CACHE_DMA | SLAB_PANIC), NULL, NULL);
776 /* 4) Replace the bootstrap head arrays */
780 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
782 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
783 memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
784 cache_cache.array[smp_processor_id()] = ptr;
787 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
789 BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
790 memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
791 sizeof(struct arraycache_init));
792 malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
796 /* 5) resize the head arrays to their final sizes */
798 kmem_cache_t *cachep;
799 down(&cache_chain_sem);
800 list_for_each_entry(cachep, &cache_chain, next)
801 enable_cpucache(cachep);
802 up(&cache_chain_sem);
806 g_cpucache_up = FULL;
808 /* Register a cpu startup notifier callback
809 * that initializes ac_data for all new cpus
811 register_cpu_notifier(&cpucache_notifier);
814 /* The reap timers are started later, with a module init call:
815 * That part of the kernel is not yet operational.
819 int __init cpucache_init(void)
824 * Register the timers that return unneeded
827 for (cpu = 0; cpu < NR_CPUS; cpu++) {
829 start_cpu_timer(cpu);
835 __initcall(cpucache_init);
838 * Interface to system's page allocator. No need to hold the cache-lock.
840 * If we requested dmaable memory, we will get it. Even if we
841 * did not request dmaable memory, we might get it, but that
842 * would be relatively rare and ignorable.
844 static void *kmem_getpages(kmem_cache_t *cachep, int flags, int nodeid)
850 flags |= cachep->gfpflags;
851 if (likely(nodeid == -1)) {
852 addr = (void*)__get_free_pages(flags, cachep->gfporder);
855 page = virt_to_page(addr);
857 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
860 addr = page_address(page);
863 i = (1 << cachep->gfporder);
864 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
865 atomic_add(i, &slab_reclaim_pages);
866 add_page_state(nr_slab, i);
875 * Interface to system's page release.
877 static inline void kmem_freepages(kmem_cache_t *cachep, void *addr)
879 unsigned long i = (1<<cachep->gfporder);
880 struct page *page = virt_to_page(addr);
881 const unsigned long nr_freed = i;
884 if (!TestClearPageSlab(page))
888 sub_page_state(nr_slab, nr_freed);
889 if (current->reclaim_state)
890 current->reclaim_state->reclaimed_slab += nr_freed;
891 free_pages((unsigned long)addr, cachep->gfporder);
892 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
893 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
898 #ifdef CONFIG_DEBUG_PAGEALLOC
899 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, unsigned long caller)
901 int size = obj_reallen(cachep);
903 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
905 if (size < 5*sizeof(unsigned long))
910 *addr++=smp_processor_id();
911 size -= 3*sizeof(unsigned long);
913 unsigned long *sptr = &caller;
914 unsigned long svalue;
916 while (!kstack_end(sptr)) {
918 if (kernel_text_address(svalue)) {
920 size -= sizeof(unsigned long);
921 if (size <= sizeof(unsigned long))
931 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
933 int size = obj_reallen(cachep);
934 addr = &((char*)addr)[obj_dbghead(cachep)];
936 memset(addr, val, size);
937 *(unsigned char *)(addr+size-1) = POISON_END;
940 static void dump_line(char *data, int offset, int limit)
943 printk(KERN_ERR "%03x:", offset);
944 for (i=0;i<limit;i++) {
945 printk(" %02x", (unsigned char)data[offset+i]);
951 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
957 if (cachep->flags & SLAB_RED_ZONE) {
958 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
959 *dbg_redzone1(cachep, objp),
960 *dbg_redzone2(cachep, objp));
963 if (cachep->flags & SLAB_STORE_USER) {
964 printk(KERN_ERR "Last user: [<%p>]", *dbg_userword(cachep, objp));
965 print_symbol("(%s)", (unsigned long)*dbg_userword(cachep, objp));
968 realobj = (char*)objp+obj_dbghead(cachep);
969 size = obj_reallen(cachep);
970 for (i=0; i<size && lines;i+=16, lines--) {
975 dump_line(realobj, i, limit);
982 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
988 realobj = (char*)objp+obj_dbghead(cachep);
989 size = obj_reallen(cachep);
991 for (i=0;i<size;i++) {
992 char exp = POISON_FREE;
995 if (realobj[i] != exp) {
1000 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1002 print_objinfo(cachep, objp, 0);
1004 /* Hexdump the affected line */
1009 dump_line(realobj, i, limit);
1012 /* Limit to 5 lines */
1018 /* Print some data about the neighboring objects, if they
1021 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1024 objnr = (objp-slabp->s_mem)/cachep->objsize;
1026 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1027 realobj = (char*)objp+obj_dbghead(cachep);
1028 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1030 print_objinfo(cachep, objp, 2);
1032 if (objnr+1 < cachep->num) {
1033 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1034 realobj = (char*)objp+obj_dbghead(cachep);
1035 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1037 print_objinfo(cachep, objp, 2);
1043 /* Destroy all the objs in a slab, and release the mem back to the system.
1044 * Before calling the slab must have been unlinked from the cache.
1045 * The cache-lock is not held/needed.
1047 static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1051 for (i = 0; i < cachep->num; i++) {
1052 void *objp = slabp->s_mem + cachep->objsize * i;
1054 if (cachep->flags & SLAB_POISON) {
1055 #ifdef CONFIG_DEBUG_PAGEALLOC
1056 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1057 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1059 check_poison_obj(cachep, objp);
1061 check_poison_obj(cachep, objp);
1064 if (cachep->flags & SLAB_RED_ZONE) {
1065 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1066 slab_error(cachep, "start of a freed object "
1068 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1069 slab_error(cachep, "end of a freed object "
1072 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1073 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1078 for (i = 0; i < cachep->num; i++) {
1079 void* objp = slabp->s_mem+cachep->objsize*i;
1080 (cachep->dtor)(objp, cachep, 0);
1085 kmem_freepages(cachep, slabp->s_mem-slabp->colouroff);
1086 if (OFF_SLAB(cachep))
1087 kmem_cache_free(cachep->slabp_cache, slabp);
1091 * kmem_cache_create - Create a cache.
1092 * @name: A string which is used in /proc/slabinfo to identify this cache.
1093 * @size: The size of objects to be created in this cache.
1094 * @align: The required alignment for the objects.
1095 * @flags: SLAB flags
1096 * @ctor: A constructor for the objects.
1097 * @dtor: A destructor for the objects.
1099 * Returns a ptr to the cache on success, NULL on failure.
1100 * Cannot be called within a int, but can be interrupted.
1101 * The @ctor is run when new pages are allocated by the cache
1102 * and the @dtor is run before the pages are handed back.
1104 * @name must be valid until the cache is destroyed. This implies that
1105 * the module calling this has to destroy the cache before getting
1110 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1111 * to catch references to uninitialised memory.
1113 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1114 * for buffer overruns.
1116 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1119 * %SLAB_HWCACHE_ALIGN - This flag has no effect and will be removed soon.
1123 kmem_cache_create (const char *name, size_t size, size_t align,
1124 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1125 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1127 size_t left_over, slab_size;
1128 kmem_cache_t *cachep = NULL;
1131 * Sanity checks... these are all serious usage bugs.
1135 (size < BYTES_PER_WORD) ||
1136 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1139 printk(KERN_ERR "%s: Early error in slab %s\n",
1140 __FUNCTION__, name);
1145 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1146 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1147 /* No constructor, but inital state check requested */
1148 printk(KERN_ERR "%s: No con, but init state check "
1149 "requested - %s\n", __FUNCTION__, name);
1150 flags &= ~SLAB_DEBUG_INITIAL;
1155 * Enable redzoning and last user accounting, except for caches with
1156 * large objects, if the increased size would increase the object size
1157 * above the next power of two: caches with object sizes just above a
1158 * power of two have a significant amount of internal fragmentation.
1160 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1161 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1162 flags |= SLAB_POISON;
1166 * Always checks flags, a caller might be expecting debug
1167 * support which isn't available.
1169 if (flags & ~CREATE_MASK)
1173 /* combinations of forced alignment and advanced debugging is
1174 * not yet implemented.
1176 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1178 if (flags & SLAB_HWCACHE_ALIGN) {
1179 /* Default alignment: as specified by the arch code.
1180 * Except if an object is really small, then squeeze multiple
1181 * into one cacheline.
1183 align = cache_line_size();
1184 while (size <= align/2)
1187 align = BYTES_PER_WORD;
1191 /* Get cache's description obj. */
1192 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1195 memset(cachep, 0, sizeof(kmem_cache_t));
1197 /* Check that size is in terms of words. This is needed to avoid
1198 * unaligned accesses for some archs when redzoning is used, and makes
1199 * sure any on-slab bufctl's are also correctly aligned.
1201 if (size & (BYTES_PER_WORD-1)) {
1202 size += (BYTES_PER_WORD-1);
1203 size &= ~(BYTES_PER_WORD-1);
1207 cachep->reallen = size;
1209 if (flags & SLAB_RED_ZONE) {
1210 /* redzoning only works with word aligned caches */
1211 align = BYTES_PER_WORD;
1213 /* add space for red zone words */
1214 cachep->dbghead += BYTES_PER_WORD;
1215 size += 2*BYTES_PER_WORD;
1217 if (flags & SLAB_STORE_USER) {
1218 /* user store requires word alignment and
1219 * one word storage behind the end of the real
1222 align = BYTES_PER_WORD;
1223 size += BYTES_PER_WORD;
1225 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1226 if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1227 cachep->dbghead += PAGE_SIZE - size;
1233 /* Determine if the slab management is 'on' or 'off' slab. */
1234 if (size >= (PAGE_SIZE>>3))
1236 * Size is large, assume best to place the slab management obj
1237 * off-slab (should allow better packing of objs).
1239 flags |= CFLGS_OFF_SLAB;
1241 size = ALIGN(size, align);
1243 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1245 * A VFS-reclaimable slab tends to have most allocations
1246 * as GFP_NOFS and we really don't want to have to be allocating
1247 * higher-order pages when we are unable to shrink dcache.
1249 cachep->gfporder = 0;
1250 cache_estimate(cachep->gfporder, size, align, flags,
1251 &left_over, &cachep->num);
1254 * Calculate size (in pages) of slabs, and the num of objs per
1255 * slab. This could be made much more intelligent. For now,
1256 * try to avoid using high page-orders for slabs. When the
1257 * gfp() funcs are more friendly towards high-order requests,
1258 * this should be changed.
1261 unsigned int break_flag = 0;
1263 cache_estimate(cachep->gfporder, size, align, flags,
1264 &left_over, &cachep->num);
1267 if (cachep->gfporder >= MAX_GFP_ORDER)
1271 if (flags & CFLGS_OFF_SLAB &&
1272 cachep->num > offslab_limit) {
1273 /* This num of objs will cause problems. */
1280 * Large num of objs is good, but v. large slabs are
1281 * currently bad for the gfp()s.
1283 if (cachep->gfporder >= slab_break_gfp_order)
1286 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1287 break; /* Acceptable internal fragmentation. */
1294 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1295 kmem_cache_free(&cache_cache, cachep);
1299 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1300 + sizeof(struct slab), align);
1303 * If the slab has been placed off-slab, and we have enough space then
1304 * move it on-slab. This is at the expense of any extra colouring.
1306 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1307 flags &= ~CFLGS_OFF_SLAB;
1308 left_over -= slab_size;
1311 if (flags & CFLGS_OFF_SLAB) {
1312 /* really off slab. No need for manual alignment */
1313 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1316 cachep->colour_off = cache_line_size();
1317 /* Offset must be a multiple of the alignment. */
1318 if (cachep->colour_off < align)
1319 cachep->colour_off = align;
1320 cachep->colour = left_over/cachep->colour_off;
1321 cachep->slab_size = slab_size;
1322 cachep->flags = flags;
1323 cachep->gfpflags = 0;
1324 if (flags & SLAB_CACHE_DMA)
1325 cachep->gfpflags |= GFP_DMA;
1326 spin_lock_init(&cachep->spinlock);
1327 cachep->objsize = size;
1329 INIT_LIST_HEAD(&cachep->lists.slabs_full);
1330 INIT_LIST_HEAD(&cachep->lists.slabs_partial);
1331 INIT_LIST_HEAD(&cachep->lists.slabs_free);
1333 if (flags & CFLGS_OFF_SLAB)
1334 cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
1335 cachep->ctor = ctor;
1336 cachep->dtor = dtor;
1337 cachep->name = name;
1339 /* Don't let CPUs to come and go */
1342 if (g_cpucache_up == FULL) {
1343 enable_cpucache(cachep);
1345 if (g_cpucache_up == NONE) {
1346 /* Note: the first kmem_cache_create must create
1347 * the cache that's used by kmalloc(24), otherwise
1348 * the creation of further caches will BUG().
1350 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1351 g_cpucache_up = PARTIAL;
1353 cachep->array[smp_processor_id()] = kmalloc(sizeof(struct arraycache_init),GFP_KERNEL);
1355 BUG_ON(!ac_data(cachep));
1356 ac_data(cachep)->avail = 0;
1357 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1358 ac_data(cachep)->batchcount = 1;
1359 ac_data(cachep)->touched = 0;
1360 cachep->batchcount = 1;
1361 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1362 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
1366 cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
1367 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1369 /* Need the semaphore to access the chain. */
1370 down(&cache_chain_sem);
1372 struct list_head *p;
1373 mm_segment_t old_fs;
1377 list_for_each(p, &cache_chain) {
1378 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1380 /* This happens when the module gets unloaded and doesn't
1381 destroy its slab cache and noone else reuses the vmalloc
1382 area of the module. Print a warning. */
1383 if (__get_user(tmp,pc->name)) {
1384 printk("SLAB: cache with size %d has lost its name\n",
1388 if (!strcmp(pc->name,name)) {
1389 printk("kmem_cache_create: duplicate cache %s\n",name);
1390 up(&cache_chain_sem);
1391 unlock_cpu_hotplug();
1398 /* cache setup completed, link it into the list */
1399 list_add(&cachep->next, &cache_chain);
1400 up(&cache_chain_sem);
1401 unlock_cpu_hotplug();
1403 if (!cachep && (flags & SLAB_PANIC))
1404 panic("kmem_cache_create(): failed to create slab `%s'\n",
1408 EXPORT_SYMBOL(kmem_cache_create);
1410 static inline void check_irq_off(void)
1413 BUG_ON(!irqs_disabled());
1417 static inline void check_irq_on(void)
1420 BUG_ON(irqs_disabled());
1424 static inline void check_spinlock_acquired(kmem_cache_t *cachep)
1428 BUG_ON(spin_trylock(&cachep->spinlock));
1433 * Waits for all CPUs to execute func().
1435 static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1440 local_irq_disable();
1444 if (smp_call_function(func, arg, 1, 1))
1450 static void drain_array_locked(kmem_cache_t* cachep,
1451 struct array_cache *ac, int force);
1453 static void do_drain(void *arg)
1455 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1456 struct array_cache *ac;
1459 ac = ac_data(cachep);
1460 spin_lock(&cachep->spinlock);
1461 free_block(cachep, &ac_entry(ac)[0], ac->avail);
1462 spin_unlock(&cachep->spinlock);
1466 static void drain_cpu_caches(kmem_cache_t *cachep)
1468 smp_call_function_all_cpus(do_drain, cachep);
1470 spin_lock_irq(&cachep->spinlock);
1471 if (cachep->lists.shared)
1472 drain_array_locked(cachep, cachep->lists.shared, 1);
1473 spin_unlock_irq(&cachep->spinlock);
1477 /* NUMA shrink all list3s */
1478 static int __cache_shrink(kmem_cache_t *cachep)
1483 drain_cpu_caches(cachep);
1486 spin_lock_irq(&cachep->spinlock);
1489 struct list_head *p;
1491 p = cachep->lists.slabs_free.prev;
1492 if (p == &cachep->lists.slabs_free)
1495 slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
1500 list_del(&slabp->list);
1502 cachep->lists.free_objects -= cachep->num;
1503 spin_unlock_irq(&cachep->spinlock);
1504 slab_destroy(cachep, slabp);
1505 spin_lock_irq(&cachep->spinlock);
1507 ret = !list_empty(&cachep->lists.slabs_full) ||
1508 !list_empty(&cachep->lists.slabs_partial);
1509 spin_unlock_irq(&cachep->spinlock);
1514 * kmem_cache_shrink - Shrink a cache.
1515 * @cachep: The cache to shrink.
1517 * Releases as many slabs as possible for a cache.
1518 * To help debugging, a zero exit status indicates all slabs were released.
1520 int kmem_cache_shrink(kmem_cache_t *cachep)
1522 if (!cachep || in_interrupt())
1525 return __cache_shrink(cachep);
1528 EXPORT_SYMBOL(kmem_cache_shrink);
1531 * kmem_cache_destroy - delete a cache
1532 * @cachep: the cache to destroy
1534 * Remove a kmem_cache_t object from the slab cache.
1535 * Returns 0 on success.
1537 * It is expected this function will be called by a module when it is
1538 * unloaded. This will remove the cache completely, and avoid a duplicate
1539 * cache being allocated each time a module is loaded and unloaded, if the
1540 * module doesn't have persistent in-kernel storage across loads and unloads.
1542 * The cache must be empty before calling this function.
1544 * The caller must guarantee that noone will allocate memory from the cache
1545 * during the kmem_cache_destroy().
1547 int kmem_cache_destroy (kmem_cache_t * cachep)
1551 if (!cachep || in_interrupt())
1554 /* Don't let CPUs to come and go */
1557 /* Find the cache in the chain of caches. */
1558 down(&cache_chain_sem);
1560 * the chain is never empty, cache_cache is never destroyed
1562 list_del(&cachep->next);
1563 up(&cache_chain_sem);
1565 if (__cache_shrink(cachep)) {
1566 slab_error(cachep, "Can't free all objects");
1567 down(&cache_chain_sem);
1568 list_add(&cachep->next,&cache_chain);
1569 up(&cache_chain_sem);
1570 unlock_cpu_hotplug();
1574 /* no cpu_online check required here since we clear the percpu
1575 * array on cpu offline and set this to NULL.
1577 for (i = 0; i < NR_CPUS; i++)
1578 kfree(cachep->array[i]);
1580 /* NUMA: free the list3 structures */
1581 kfree(cachep->lists.shared);
1582 cachep->lists.shared = NULL;
1583 kmem_cache_free(&cache_cache, cachep);
1585 unlock_cpu_hotplug();
1590 EXPORT_SYMBOL(kmem_cache_destroy);
1592 /* Get the memory for a slab management obj. */
1593 static inline struct slab* alloc_slabmgmt (kmem_cache_t *cachep,
1594 void *objp, int colour_off, int local_flags)
1598 if (OFF_SLAB(cachep)) {
1599 /* Slab management obj is off-slab. */
1600 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
1604 slabp = objp+colour_off;
1605 colour_off += cachep->slab_size;
1608 slabp->colouroff = colour_off;
1609 slabp->s_mem = objp+colour_off;
1614 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
1616 return (kmem_bufctl_t *)(slabp+1);
1619 static void cache_init_objs (kmem_cache_t * cachep,
1620 struct slab * slabp, unsigned long ctor_flags)
1624 for (i = 0; i < cachep->num; i++) {
1625 void* objp = slabp->s_mem+cachep->objsize*i;
1627 /* need to poison the objs? */
1628 if (cachep->flags & SLAB_POISON)
1629 poison_obj(cachep, objp, POISON_FREE);
1630 if (cachep->flags & SLAB_STORE_USER)
1631 *dbg_userword(cachep, objp) = NULL;
1633 if (cachep->flags & SLAB_RED_ZONE) {
1634 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1635 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1638 * Constructors are not allowed to allocate memory from
1639 * the same cache which they are a constructor for.
1640 * Otherwise, deadlock. They must also be threaded.
1642 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
1643 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
1645 if (cachep->flags & SLAB_RED_ZONE) {
1646 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1647 slab_error(cachep, "constructor overwrote the"
1648 " end of an object");
1649 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1650 slab_error(cachep, "constructor overwrote the"
1651 " start of an object");
1653 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
1654 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1657 cachep->ctor(objp, cachep, ctor_flags);
1659 slab_bufctl(slabp)[i] = i+1;
1661 slab_bufctl(slabp)[i-1] = BUFCTL_END;
1665 static void kmem_flagcheck(kmem_cache_t *cachep, int flags)
1667 if (flags & SLAB_DMA) {
1668 if (!(cachep->gfpflags & GFP_DMA))
1671 if (cachep->gfpflags & GFP_DMA)
1676 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
1681 /* Nasty!!!!!! I hope this is OK. */
1682 i = 1 << cachep->gfporder;
1683 page = virt_to_page(objp);
1685 SET_PAGE_CACHE(page, cachep);
1686 SET_PAGE_SLAB(page, slabp);
1692 * Grow (by 1) the number of slabs within a cache. This is called by
1693 * kmem_cache_alloc() when there are no active objs left in a cache.
1695 static int cache_grow (kmem_cache_t * cachep, int flags)
1701 unsigned long ctor_flags;
1703 /* Be lazy and only check for valid flags here,
1704 * keeping it out of the critical path in kmem_cache_alloc().
1706 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
1708 if (flags & SLAB_NO_GROW)
1711 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
1712 local_flags = (flags & SLAB_LEVEL_MASK);
1713 if (!(local_flags & __GFP_WAIT))
1715 * Not allowed to sleep. Need to tell a constructor about
1716 * this - it might need to know...
1718 ctor_flags |= SLAB_CTOR_ATOMIC;
1720 /* About to mess with non-constant members - lock. */
1722 spin_lock(&cachep->spinlock);
1724 /* Get colour for the slab, and cal the next value. */
1725 offset = cachep->colour_next;
1726 cachep->colour_next++;
1727 if (cachep->colour_next >= cachep->colour)
1728 cachep->colour_next = 0;
1729 offset *= cachep->colour_off;
1731 spin_unlock(&cachep->spinlock);
1733 if (local_flags & __GFP_WAIT)
1737 * The test for missing atomic flag is performed here, rather than
1738 * the more obvious place, simply to reduce the critical path length
1739 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
1740 * will eventually be caught here (where it matters).
1742 kmem_flagcheck(cachep, flags);
1745 /* Get mem for the objs. */
1746 if (!(objp = kmem_getpages(cachep, flags, -1)))
1749 /* Get slab management. */
1750 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
1753 set_slab_attr(cachep, slabp, objp);
1755 cache_init_objs(cachep, slabp, ctor_flags);
1757 if (local_flags & __GFP_WAIT)
1758 local_irq_disable();
1760 spin_lock(&cachep->spinlock);
1762 /* Make slab active. */
1763 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
1764 STATS_INC_GROWN(cachep);
1765 list3_data(cachep)->free_objects += cachep->num;
1766 spin_unlock(&cachep->spinlock);
1769 kmem_freepages(cachep, objp);
1771 if (local_flags & __GFP_WAIT)
1772 local_irq_disable();
1777 * Perform extra freeing checks:
1778 * - detect bad pointers.
1779 * - POISON/RED_ZONE checking
1780 * - destructor calls, for caches with POISON+dtor
1782 static inline void kfree_debugcheck(const void *objp)
1787 if (!virt_addr_valid(objp)) {
1788 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
1789 (unsigned long)objp);
1792 page = virt_to_page(objp);
1793 if (!PageSlab(page)) {
1794 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
1800 static inline void *cache_free_debugcheck (kmem_cache_t * cachep, void * objp, void *caller)
1807 objp -= obj_dbghead(cachep);
1808 kfree_debugcheck(objp);
1809 page = virt_to_page(objp);
1811 if (GET_PAGE_CACHE(page) != cachep) {
1812 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
1813 GET_PAGE_CACHE(page),cachep);
1814 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
1815 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
1818 slabp = GET_PAGE_SLAB(page);
1820 if (cachep->flags & SLAB_RED_ZONE) {
1821 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
1822 slab_error(cachep, "double free, or memory outside"
1823 " object was overwritten");
1824 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1825 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
1827 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1828 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1830 if (cachep->flags & SLAB_STORE_USER)
1831 *dbg_userword(cachep, objp) = caller;
1833 objnr = (objp-slabp->s_mem)/cachep->objsize;
1835 BUG_ON(objnr >= cachep->num);
1836 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
1838 if (cachep->flags & SLAB_DEBUG_INITIAL) {
1839 /* Need to call the slab's constructor so the
1840 * caller can perform a verify of its state (debugging).
1841 * Called without the cache-lock held.
1843 cachep->ctor(objp+obj_dbghead(cachep),
1844 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
1846 if (cachep->flags & SLAB_POISON && cachep->dtor) {
1847 /* we want to cache poison the object,
1848 * call the destruction callback
1850 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
1852 if (cachep->flags & SLAB_POISON) {
1853 #ifdef CONFIG_DEBUG_PAGEALLOC
1854 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
1855 store_stackinfo(cachep, objp, (unsigned long)caller);
1856 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1858 poison_obj(cachep, objp, POISON_FREE);
1861 poison_obj(cachep, objp, POISON_FREE);
1868 static inline 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]);
1897 static void* cache_alloc_refill(kmem_cache_t* cachep, int flags)
1900 struct kmem_list3 *l3;
1901 struct array_cache *ac;
1904 ac = ac_data(cachep);
1906 batchcount = ac->batchcount;
1907 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
1908 /* if there was little recent activity on this
1909 * cache, then perform only a partial refill.
1910 * Otherwise we could generate refill bouncing.
1912 batchcount = BATCHREFILL_LIMIT;
1914 l3 = list3_data(cachep);
1916 BUG_ON(ac->avail > 0);
1917 spin_lock(&cachep->spinlock);
1919 struct array_cache *shared_array = l3->shared;
1920 if (shared_array->avail) {
1921 if (batchcount > shared_array->avail)
1922 batchcount = shared_array->avail;
1923 shared_array->avail -= batchcount;
1924 ac->avail = batchcount;
1925 memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
1926 sizeof(void*)*batchcount);
1927 shared_array->touched = 1;
1931 while (batchcount > 0) {
1932 struct list_head *entry;
1934 /* Get slab alloc is to come from. */
1935 entry = l3->slabs_partial.next;
1936 if (entry == &l3->slabs_partial) {
1937 l3->free_touched = 1;
1938 entry = l3->slabs_free.next;
1939 if (entry == &l3->slabs_free)
1943 slabp = list_entry(entry, struct slab, list);
1944 check_slabp(cachep, slabp);
1945 check_spinlock_acquired(cachep);
1946 while (slabp->inuse < cachep->num && batchcount--) {
1948 STATS_INC_ALLOCED(cachep);
1949 STATS_INC_ACTIVE(cachep);
1950 STATS_SET_HIGH(cachep);
1952 /* get obj pointer */
1953 ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
1956 next = slab_bufctl(slabp)[slabp->free];
1958 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
1962 check_slabp(cachep, slabp);
1964 /* move slabp to correct slabp list: */
1965 list_del(&slabp->list);
1966 if (slabp->free == BUFCTL_END)
1967 list_add(&slabp->list, &l3->slabs_full);
1969 list_add(&slabp->list, &l3->slabs_partial);
1973 l3->free_objects -= ac->avail;
1975 spin_unlock(&cachep->spinlock);
1977 if (unlikely(!ac->avail)) {
1979 x = cache_grow(cachep, flags);
1981 // cache_grow can reenable interrupts, then ac could change.
1982 ac = ac_data(cachep);
1983 if (!x && ac->avail == 0) // no objects in sight? abort
1986 if (!ac->avail) // objects refilled by interrupt?
1990 return ac_entry(ac)[--ac->avail];
1994 cache_alloc_debugcheck_before(kmem_cache_t *cachep, int flags)
1996 might_sleep_if(flags & __GFP_WAIT);
1998 kmem_flagcheck(cachep, flags);
2002 static inline void *
2003 cache_alloc_debugcheck_after(kmem_cache_t *cachep,
2004 unsigned long flags, void *objp, void *caller)
2009 if (cachep->flags & SLAB_POISON) {
2010 #ifdef CONFIG_DEBUG_PAGEALLOC
2011 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2012 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2014 check_poison_obj(cachep, objp);
2016 check_poison_obj(cachep, objp);
2018 poison_obj(cachep, objp, POISON_INUSE);
2020 if (cachep->flags & SLAB_STORE_USER)
2021 *dbg_userword(cachep, objp) = caller;
2023 if (cachep->flags & SLAB_RED_ZONE) {
2024 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2025 slab_error(cachep, "double free, or memory outside"
2026 " object was overwritten");
2027 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2028 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2030 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2031 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2033 objp += obj_dbghead(cachep);
2034 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2035 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2037 if (!(flags & __GFP_WAIT))
2038 ctor_flags |= SLAB_CTOR_ATOMIC;
2040 cachep->ctor(objp, cachep, ctor_flags);
2047 static inline void * __cache_alloc (kmem_cache_t *cachep, int flags)
2049 unsigned long save_flags;
2051 struct array_cache *ac;
2053 cache_alloc_debugcheck_before(cachep, flags);
2055 local_irq_save(save_flags);
2056 ac = ac_data(cachep);
2057 if (likely(ac->avail)) {
2058 STATS_INC_ALLOCHIT(cachep);
2060 objp = ac_entry(ac)[--ac->avail];
2062 STATS_INC_ALLOCMISS(cachep);
2063 objp = cache_alloc_refill(cachep, flags);
2065 local_irq_restore(save_flags);
2066 objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0));
2071 * NUMA: different approach needed if the spinlock is moved into
2075 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2079 check_spinlock_acquired(cachep);
2081 /* NUMA: move add into loop */
2082 cachep->lists.free_objects += nr_objects;
2084 for (i = 0; i < nr_objects; i++) {
2085 void *objp = objpp[i];
2089 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2090 list_del(&slabp->list);
2091 objnr = (objp - slabp->s_mem) / cachep->objsize;
2092 check_slabp(cachep, slabp);
2094 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2095 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
2096 cachep->name, objp);
2100 slab_bufctl(slabp)[objnr] = slabp->free;
2101 slabp->free = objnr;
2102 STATS_DEC_ACTIVE(cachep);
2104 check_slabp(cachep, slabp);
2106 /* fixup slab chains */
2107 if (slabp->inuse == 0) {
2108 if (cachep->lists.free_objects > cachep->free_limit) {
2109 cachep->lists.free_objects -= cachep->num;
2110 slab_destroy(cachep, slabp);
2112 list_add(&slabp->list,
2113 &list3_data_ptr(cachep, objp)->slabs_free);
2116 /* Unconditionally move a slab to the end of the
2117 * partial list on free - maximum time for the
2118 * other objects to be freed, too.
2120 list_add_tail(&slabp->list,
2121 &list3_data_ptr(cachep, objp)->slabs_partial);
2126 static void cache_flusharray (kmem_cache_t* cachep, struct array_cache *ac)
2130 batchcount = ac->batchcount;
2132 BUG_ON(!batchcount || batchcount > ac->avail);
2135 spin_lock(&cachep->spinlock);
2136 if (cachep->lists.shared) {
2137 struct array_cache *shared_array = cachep->lists.shared;
2138 int max = shared_array->limit-shared_array->avail;
2140 if (batchcount > max)
2142 memcpy(&ac_entry(shared_array)[shared_array->avail],
2144 sizeof(void*)*batchcount);
2145 shared_array->avail += batchcount;
2150 free_block(cachep, &ac_entry(ac)[0], batchcount);
2155 struct list_head *p;
2157 p = list3_data(cachep)->slabs_free.next;
2158 while (p != &(list3_data(cachep)->slabs_free)) {
2161 slabp = list_entry(p, struct slab, list);
2162 BUG_ON(slabp->inuse);
2167 STATS_SET_FREEABLE(cachep, i);
2170 spin_unlock(&cachep->spinlock);
2171 ac->avail -= batchcount;
2172 memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
2173 sizeof(void*)*ac->avail);
2178 * Release an obj back to its cache. If the obj has a constructed
2179 * state, it must be in this state _before_ it is released.
2181 * Called with disabled ints.
2183 static inline void __cache_free (kmem_cache_t *cachep, void* objp)
2185 struct array_cache *ac = ac_data(cachep);
2188 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2190 if (likely(ac->avail < ac->limit)) {
2191 STATS_INC_FREEHIT(cachep);
2192 ac_entry(ac)[ac->avail++] = objp;
2195 STATS_INC_FREEMISS(cachep);
2196 cache_flusharray(cachep, ac);
2197 ac_entry(ac)[ac->avail++] = objp;
2202 * kmem_cache_alloc - Allocate an object
2203 * @cachep: The cache to allocate from.
2204 * @flags: See kmalloc().
2206 * Allocate an object from this cache. The flags are only relevant
2207 * if the cache has no available objects.
2209 void * kmem_cache_alloc (kmem_cache_t *cachep, int flags)
2211 return __cache_alloc(cachep, flags);
2214 EXPORT_SYMBOL(kmem_cache_alloc);
2217 * kmem_ptr_validate - check if an untrusted pointer might
2219 * @cachep: the cache we're checking against
2220 * @ptr: pointer to validate
2222 * This verifies that the untrusted pointer looks sane:
2223 * it is _not_ a guarantee that the pointer is actually
2224 * part of the slab cache in question, but it at least
2225 * validates that the pointer can be dereferenced and
2226 * looks half-way sane.
2228 * Currently only used for dentry validation.
2230 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2232 unsigned long addr = (unsigned long) ptr;
2233 unsigned long min_addr = PAGE_OFFSET;
2234 unsigned long align_mask = BYTES_PER_WORD-1;
2235 unsigned long size = cachep->objsize;
2238 if (unlikely(addr < min_addr))
2240 if (unlikely(addr > (unsigned long)high_memory - size))
2242 if (unlikely(addr & align_mask))
2244 if (unlikely(!kern_addr_valid(addr)))
2246 if (unlikely(!kern_addr_valid(addr + size - 1)))
2248 page = virt_to_page(ptr);
2249 if (unlikely(!PageSlab(page)))
2251 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2259 * kmem_cache_alloc_node - Allocate an object on the specified node
2260 * @cachep: The cache to allocate from.
2261 * @flags: See kmalloc().
2262 * @nodeid: node number of the target node.
2264 * Identical to kmem_cache_alloc, except that this function is slow
2265 * and can sleep. And it will allocate memory on the given node, which
2266 * can improve the performance for cpu bound structures.
2268 void *kmem_cache_alloc_node(kmem_cache_t *cachep, int nodeid)
2275 /* The main algorithms are not node aware, thus we have to cheat:
2276 * We bypass all caches and allocate a new slab.
2277 * The following code is a streamlined copy of cache_grow().
2280 /* Get colour for the slab, and update the next value. */
2281 spin_lock_irq(&cachep->spinlock);
2282 offset = cachep->colour_next;
2283 cachep->colour_next++;
2284 if (cachep->colour_next >= cachep->colour)
2285 cachep->colour_next = 0;
2286 offset *= cachep->colour_off;
2287 spin_unlock_irq(&cachep->spinlock);
2289 /* Get mem for the objs. */
2290 if (!(objp = kmem_getpages(cachep, GFP_KERNEL, nodeid)))
2293 /* Get slab management. */
2294 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, GFP_KERNEL)))
2297 set_slab_attr(cachep, slabp, objp);
2298 cache_init_objs(cachep, slabp, SLAB_CTOR_CONSTRUCTOR);
2300 /* The first object is ours: */
2301 objp = slabp->s_mem + slabp->free*cachep->objsize;
2303 next = slab_bufctl(slabp)[slabp->free];
2305 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2309 /* add the remaining objects into the cache */
2310 spin_lock_irq(&cachep->spinlock);
2311 check_slabp(cachep, slabp);
2312 STATS_INC_GROWN(cachep);
2313 /* Make slab active. */
2314 if (slabp->free == BUFCTL_END) {
2315 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_full));
2317 list_add_tail(&slabp->list,
2318 &(list3_data(cachep)->slabs_partial));
2319 list3_data(cachep)->free_objects += cachep->num-1;
2321 spin_unlock_irq(&cachep->spinlock);
2322 objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp,
2323 __builtin_return_address(0));
2326 kmem_freepages(cachep, objp);
2331 EXPORT_SYMBOL(kmem_cache_alloc_node);
2334 * kmalloc - allocate memory
2335 * @size: how many bytes of memory are required.
2336 * @flags: the type of memory to allocate.
2338 * kmalloc is the normal method of allocating memory
2341 * The @flags argument may be one of:
2343 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2345 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2347 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2349 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2350 * must be suitable for DMA. This can mean different things on different
2351 * platforms. For example, on i386, it means that the memory must come
2352 * from the first 16MB.
2354 void * __kmalloc (size_t size, int flags)
2356 struct cache_sizes *csizep = malloc_sizes;
2358 for (; csizep->cs_size; csizep++) {
2359 if (size > csizep->cs_size)
2362 /* This happens if someone tries to call
2363 * kmem_cache_create(), or kmalloc(), before
2364 * the generic caches are initialized.
2366 BUG_ON(csizep->cs_cachep == NULL);
2368 return __cache_alloc(flags & GFP_DMA ?
2369 csizep->cs_dmacachep : csizep->cs_cachep, flags);
2374 EXPORT_SYMBOL(__kmalloc);
2378 * __alloc_percpu - allocate one copy of the object for every present
2379 * cpu in the system, zeroing them.
2380 * Objects should be dereferenced using per_cpu_ptr/get_cpu_ptr
2383 * @size: how many bytes of memory are required.
2384 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2386 void *__alloc_percpu(size_t size, size_t align)
2389 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2394 for (i = 0; i < NR_CPUS; i++) {
2395 if (!cpu_possible(i))
2397 pdata->ptrs[i] = kmem_cache_alloc_node(
2398 kmem_find_general_cachep(size, GFP_KERNEL),
2401 if (!pdata->ptrs[i])
2403 memset(pdata->ptrs[i], 0, size);
2406 /* Catch derefs w/o wrappers */
2407 return (void *) (~(unsigned long) pdata);
2411 if (!cpu_possible(i))
2413 kfree(pdata->ptrs[i]);
2419 EXPORT_SYMBOL(__alloc_percpu);
2423 * kmem_cache_free - Deallocate an object
2424 * @cachep: The cache the allocation was from.
2425 * @objp: The previously allocated object.
2427 * Free an object which was previously allocated from this
2430 void kmem_cache_free (kmem_cache_t *cachep, void *objp)
2432 unsigned long flags;
2434 local_irq_save(flags);
2435 __cache_free(cachep, objp);
2436 local_irq_restore(flags);
2439 EXPORT_SYMBOL(kmem_cache_free);
2442 * kfree - free previously allocated memory
2443 * @objp: pointer returned by kmalloc.
2445 * Don't free memory not originally allocated by kmalloc()
2446 * or you will run into trouble.
2448 void kfree (const void *objp)
2451 unsigned long flags;
2455 local_irq_save(flags);
2456 kfree_debugcheck(objp);
2457 c = GET_PAGE_CACHE(virt_to_page(objp));
2458 __cache_free(c, (void*)objp);
2459 local_irq_restore(flags);
2462 EXPORT_SYMBOL(kfree);
2466 * free_percpu - free previously allocated percpu memory
2467 * @objp: pointer returned by alloc_percpu.
2469 * Don't free memory not originally allocated by alloc_percpu()
2470 * The complemented objp is to check for that.
2473 free_percpu(const void *objp)
2476 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
2478 for (i = 0; i < NR_CPUS; i++) {
2479 if (!cpu_possible(i))
2485 EXPORT_SYMBOL(free_percpu);
2488 unsigned int kmem_cache_size(kmem_cache_t *cachep)
2490 return obj_reallen(cachep);
2493 EXPORT_SYMBOL(kmem_cache_size);
2495 kmem_cache_t * kmem_find_general_cachep (size_t size, int gfpflags)
2497 struct cache_sizes *csizep = malloc_sizes;
2499 /* This function could be moved to the header file, and
2500 * made inline so consumers can quickly determine what
2501 * cache pointer they require.
2503 for ( ; csizep->cs_size; csizep++) {
2504 if (size > csizep->cs_size)
2508 return (gfpflags & GFP_DMA) ? csizep->cs_dmacachep : csizep->cs_cachep;
2511 EXPORT_SYMBOL(kmem_find_general_cachep);
2513 struct ccupdate_struct {
2514 kmem_cache_t *cachep;
2515 struct array_cache *new[NR_CPUS];
2518 static void do_ccupdate_local(void *info)
2520 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
2521 struct array_cache *old;
2524 old = ac_data(new->cachep);
2526 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
2527 new->new[smp_processor_id()] = old;
2531 static int do_tune_cpucache (kmem_cache_t* cachep, int limit, int batchcount, int shared)
2533 struct ccupdate_struct new;
2534 struct array_cache *new_shared;
2537 memset(&new.new,0,sizeof(new.new));
2538 for (i = 0; i < NR_CPUS; i++) {
2539 if (cpu_online(i)) {
2540 new.new[i] = alloc_arraycache(i, limit, batchcount);
2542 for (i--; i >= 0; i--) kfree(new.new[i]);
2549 new.cachep = cachep;
2551 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
2554 spin_lock_irq(&cachep->spinlock);
2555 cachep->batchcount = batchcount;
2556 cachep->limit = limit;
2557 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
2558 spin_unlock_irq(&cachep->spinlock);
2560 for (i = 0; i < NR_CPUS; i++) {
2561 struct array_cache *ccold = new.new[i];
2564 spin_lock_irq(&cachep->spinlock);
2565 free_block(cachep, ac_entry(ccold), ccold->avail);
2566 spin_unlock_irq(&cachep->spinlock);
2569 new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d);
2571 struct array_cache *old;
2573 spin_lock_irq(&cachep->spinlock);
2574 old = cachep->lists.shared;
2575 cachep->lists.shared = new_shared;
2577 free_block(cachep, ac_entry(old), old->avail);
2578 spin_unlock_irq(&cachep->spinlock);
2586 static void enable_cpucache (kmem_cache_t *cachep)
2591 /* The head array serves three purposes:
2592 * - create a LIFO ordering, i.e. return objects that are cache-warm
2593 * - reduce the number of spinlock operations.
2594 * - reduce the number of linked list operations on the slab and
2595 * bufctl chains: array operations are cheaper.
2596 * The numbers are guessed, we should auto-tune as described by
2599 if (cachep->objsize > 131072)
2601 else if (cachep->objsize > PAGE_SIZE)
2603 else if (cachep->objsize > 1024)
2605 else if (cachep->objsize > 256)
2610 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
2611 * allocation behaviour: Most allocs on one cpu, most free operations
2612 * on another cpu. For these cases, an efficient object passing between
2613 * cpus is necessary. This is provided by a shared array. The array
2614 * replaces Bonwick's magazine layer.
2615 * On uniprocessor, it's functionally equivalent (but less efficient)
2616 * to a larger limit. Thus disabled by default.
2620 if (cachep->objsize <= PAGE_SIZE)
2625 /* With debugging enabled, large batchcount lead to excessively
2626 * long periods with disabled local interrupts. Limit the
2632 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
2634 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
2635 cachep->name, -err);
2638 static void drain_array(kmem_cache_t *cachep, struct array_cache *ac)
2645 } else if (ac->avail) {
2646 tofree = (ac->limit+4)/5;
2647 if (tofree > ac->avail) {
2648 tofree = (ac->avail+1)/2;
2650 spin_lock(&cachep->spinlock);
2651 free_block(cachep, ac_entry(ac), tofree);
2652 spin_unlock(&cachep->spinlock);
2653 ac->avail -= tofree;
2654 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2655 sizeof(void*)*ac->avail);
2659 static void drain_array_locked(kmem_cache_t *cachep,
2660 struct array_cache *ac, int force)
2664 check_spinlock_acquired(cachep);
2665 if (ac->touched && !force) {
2667 } else if (ac->avail) {
2668 tofree = force ? ac->avail : (ac->limit+4)/5;
2669 if (tofree > ac->avail) {
2670 tofree = (ac->avail+1)/2;
2672 free_block(cachep, ac_entry(ac), tofree);
2673 ac->avail -= tofree;
2674 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2675 sizeof(void*)*ac->avail);
2680 * cache_reap - Reclaim memory from caches.
2682 * Called from a timer, every few seconds
2684 * - clear the per-cpu caches for this CPU.
2685 * - return freeable pages to the main free memory pool.
2687 * If we cannot acquire the cache chain semaphore then just give up - we'll
2688 * try again next timer interrupt.
2690 static inline void cache_reap (void)
2692 struct list_head *walk;
2695 BUG_ON(!in_interrupt());
2698 if (down_trylock(&cache_chain_sem))
2701 list_for_each(walk, &cache_chain) {
2702 kmem_cache_t *searchp;
2703 struct list_head* p;
2707 searchp = list_entry(walk, kmem_cache_t, next);
2709 if (searchp->flags & SLAB_NO_REAP)
2713 local_irq_disable();
2714 drain_array(searchp, ac_data(searchp));
2716 if(time_after(searchp->lists.next_reap, jiffies))
2719 spin_lock(&searchp->spinlock);
2720 if(time_after(searchp->lists.next_reap, jiffies)) {
2723 searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
2725 if (searchp->lists.shared)
2726 drain_array_locked(searchp, searchp->lists.shared, 0);
2728 if (searchp->lists.free_touched) {
2729 searchp->lists.free_touched = 0;
2733 tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
2735 p = list3_data(searchp)->slabs_free.next;
2736 if (p == &(list3_data(searchp)->slabs_free))
2739 slabp = list_entry(p, struct slab, list);
2740 BUG_ON(slabp->inuse);
2741 list_del(&slabp->list);
2742 STATS_INC_REAPED(searchp);
2744 /* Safe to drop the lock. The slab is no longer
2745 * linked to the cache.
2746 * searchp cannot disappear, we hold
2749 searchp->lists.free_objects -= searchp->num;
2750 spin_unlock_irq(&searchp->spinlock);
2751 slab_destroy(searchp, slabp);
2752 spin_lock_irq(&searchp->spinlock);
2753 } while(--tofree > 0);
2755 spin_unlock(&searchp->spinlock);
2762 up(&cache_chain_sem);
2766 * This is a timer handler. There is one per CPU. It is called periodially
2767 * to shrink this CPU's caches. Otherwise there could be memory tied up
2768 * for long periods (or for ever) due to load changes.
2770 static void reap_timer_fnc(unsigned long cpu)
2772 struct timer_list *rt = &__get_cpu_var(reap_timers);
2774 /* CPU hotplug can drag us off cpu: don't run on wrong CPU */
2775 if (!cpu_is_offline(cpu)) {
2777 mod_timer(rt, jiffies + REAPTIMEOUT_CPUC + cpu);
2781 #ifdef CONFIG_PROC_FS
2783 static void *s_start(struct seq_file *m, loff_t *pos)
2786 struct list_head *p;
2788 down(&cache_chain_sem);
2791 * Output format version, so at least we can change it
2792 * without _too_ many complaints.
2795 seq_puts(m, "slabinfo - version: 2.0 (statistics)\n");
2797 seq_puts(m, "slabinfo - version: 2.0\n");
2799 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
2800 seq_puts(m, " : tunables <batchcount> <limit> <sharedfactor>");
2801 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
2803 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <freelimit>");
2804 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
2808 p = cache_chain.next;
2811 if (p == &cache_chain)
2814 return list_entry(p, kmem_cache_t, next);
2817 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2819 kmem_cache_t *cachep = p;
2821 return cachep->next.next == &cache_chain ? NULL
2822 : list_entry(cachep->next.next, kmem_cache_t, next);
2825 static void s_stop(struct seq_file *m, void *p)
2827 up(&cache_chain_sem);
2830 static int s_show(struct seq_file *m, void *p)
2832 kmem_cache_t *cachep = p;
2833 struct list_head *q;
2835 unsigned long active_objs;
2836 unsigned long num_objs;
2837 unsigned long active_slabs = 0;
2838 unsigned long num_slabs;
2841 mm_segment_t old_fs;
2845 spin_lock_irq(&cachep->spinlock);
2848 list_for_each(q,&cachep->lists.slabs_full) {
2849 slabp = list_entry(q, struct slab, list);
2850 if (slabp->inuse != cachep->num && !error)
2851 error = "slabs_full accounting error";
2852 active_objs += cachep->num;
2855 list_for_each(q,&cachep->lists.slabs_partial) {
2856 slabp = list_entry(q, struct slab, list);
2857 if (slabp->inuse == cachep->num && !error)
2858 error = "slabs_partial inuse accounting error";
2859 if (!slabp->inuse && !error)
2860 error = "slabs_partial/inuse accounting error";
2861 active_objs += slabp->inuse;
2864 list_for_each(q,&cachep->lists.slabs_free) {
2865 slabp = list_entry(q, struct slab, list);
2866 if (slabp->inuse && !error)
2867 error = "slabs_free/inuse accounting error";
2870 num_slabs+=active_slabs;
2871 num_objs = num_slabs*cachep->num;
2872 if (num_objs - active_objs != cachep->lists.free_objects && !error)
2873 error = "free_objects accounting error";
2875 name = cachep->name;
2878 * Check to see if `name' resides inside a module which has been
2879 * unloaded (someone forgot to destroy their cache)
2883 if (__get_user(tmp, name))
2888 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
2890 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
2891 name, active_objs, num_objs, cachep->objsize,
2892 cachep->num, (1<<cachep->gfporder));
2893 seq_printf(m, " : tunables %4u %4u %4u",
2894 cachep->limit, cachep->batchcount,
2895 cachep->lists.shared->limit/cachep->batchcount);
2896 seq_printf(m, " : slabdata %6lu %6lu %6u",
2897 active_slabs, num_slabs, cachep->lists.shared->avail);
2900 unsigned long high = cachep->high_mark;
2901 unsigned long allocs = cachep->num_allocations;
2902 unsigned long grown = cachep->grown;
2903 unsigned long reaped = cachep->reaped;
2904 unsigned long errors = cachep->errors;
2905 unsigned long max_freeable = cachep->max_freeable;
2906 unsigned long free_limit = cachep->free_limit;
2908 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu",
2909 allocs, high, grown, reaped, errors,
2910 max_freeable, free_limit);
2914 unsigned long allochit = atomic_read(&cachep->allochit);
2915 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
2916 unsigned long freehit = atomic_read(&cachep->freehit);
2917 unsigned long freemiss = atomic_read(&cachep->freemiss);
2919 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
2920 allochit, allocmiss, freehit, freemiss);
2924 spin_unlock_irq(&cachep->spinlock);
2929 * slabinfo_op - iterator that generates /proc/slabinfo
2938 * num-pages-per-slab
2939 * + further values on SMP and with statistics enabled
2942 struct seq_operations slabinfo_op = {
2949 #define MAX_SLABINFO_WRITE 128
2951 * slabinfo_write - Tuning for the slab allocator
2953 * @buffer: user buffer
2954 * @count: data length
2957 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
2958 size_t count, loff_t *ppos)
2960 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
2961 int limit, batchcount, shared, res;
2962 struct list_head *p;
2964 if (count > MAX_SLABINFO_WRITE)
2966 if (copy_from_user(&kbuf, buffer, count))
2968 kbuf[MAX_SLABINFO_WRITE] = '\0';
2970 tmp = strchr(kbuf, ' ');
2975 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
2978 /* Find the cache in the chain of caches. */
2979 down(&cache_chain_sem);
2981 list_for_each(p,&cache_chain) {
2982 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
2984 if (!strcmp(cachep->name, kbuf)) {
2987 batchcount > limit ||
2991 res = do_tune_cpucache(cachep, limit, batchcount, shared);
2996 up(&cache_chain_sem);
3003 unsigned int ksize(const void *objp)
3006 unsigned long flags;
3007 unsigned int size = 0;
3009 if (likely(objp != NULL)) {
3010 local_irq_save(flags);
3011 c = GET_PAGE_CACHE(virt_to_page(objp));
3012 size = kmem_cache_size(c);
3013 local_irq_restore(flags);
3019 void ptrinfo(unsigned long addr)
3023 printk("Dumping data about address %p.\n", (void*)addr);
3024 if (!virt_addr_valid((void*)addr)) {
3025 printk("virt addr invalid.\n");
3030 pgd_t *pgd = pgd_offset_k(addr);
3032 if (pgd_none(*pgd)) {
3033 printk("No pgd.\n");
3036 pmd = pmd_offset(pgd, addr);
3037 if (pmd_none(*pmd)) {
3038 printk("No pmd.\n");
3042 if (pmd_large(*pmd)) {
3043 printk("Large page.\n");
3047 printk("normal page, pte_val 0x%llx\n",
3048 (unsigned long long)pte_val(*pte_offset_kernel(pmd, addr)));
3052 page = virt_to_page((void*)addr);
3053 printk("struct page at %p, flags %08lx\n",
3054 page, (unsigned long)page->flags);
3055 if (PageSlab(page)) {
3058 unsigned long flags;
3062 c = GET_PAGE_CACHE(page);
3063 printk("belongs to cache %s.\n",c->name);
3065 spin_lock_irqsave(&c->spinlock, flags);
3066 s = GET_PAGE_SLAB(page);
3067 printk("slabp %p with %d inuse objects (from %d).\n",
3068 s, s->inuse, c->num);
3071 objnr = (addr-(unsigned long)s->s_mem)/c->objsize;
3072 objp = s->s_mem+c->objsize*objnr;
3073 printk("points into object no %d, starting at %p, len %d.\n",
3074 objnr, objp, c->objsize);
3075 if (objnr >= c->num) {
3076 printk("Bad obj number.\n");
3078 kernel_map_pages(virt_to_page(objp),
3079 c->objsize/PAGE_SIZE, 1);
3081 print_objinfo(c, objp, 2);
3083 spin_unlock_irqrestore(&c->spinlock, flags);