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. */
480 static struct cache_names {
484 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
485 #include <linux/kmalloc_sizes.h>
490 struct arraycache_init initarray_cache __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
491 struct arraycache_init initarray_generic __initdata = { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
493 /* internal cache of cache description objs */
494 static kmem_cache_t cache_cache = {
495 .lists = LIST3_INIT(cache_cache.lists),
497 .limit = BOOT_CPUCACHE_ENTRIES,
498 .objsize = sizeof(kmem_cache_t),
499 .flags = SLAB_NO_REAP,
500 .spinlock = SPIN_LOCK_UNLOCKED,
501 .name = "kmem_cache",
503 .reallen = sizeof(kmem_cache_t),
507 /* Guard access to the cache-chain. */
508 static struct semaphore cache_chain_sem;
510 struct list_head cache_chain;
513 * vm_enough_memory() looks at this to determine how many
514 * slab-allocated pages are possibly freeable under pressure
516 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
518 atomic_t slab_reclaim_pages;
519 EXPORT_SYMBOL(slab_reclaim_pages);
522 * chicken and egg problem: delay the per-cpu array allocation
523 * until the general caches are up.
531 static DEFINE_PER_CPU(struct timer_list, reap_timers);
533 static void reap_timer_fnc(unsigned long data);
534 static void free_block(kmem_cache_t* cachep, void** objpp, int len);
535 static void enable_cpucache (kmem_cache_t *cachep);
537 static inline void ** ac_entry(struct array_cache *ac)
539 return (void**)(ac+1);
542 static inline struct array_cache *ac_data(kmem_cache_t *cachep)
544 return cachep->array[smp_processor_id()];
547 /* Cal the num objs, wastage, and bytes left over for a given slab size. */
548 static void cache_estimate (unsigned long gfporder, size_t size, size_t align,
549 int flags, size_t *left_over, unsigned int *num)
552 size_t wastage = PAGE_SIZE<<gfporder;
556 if (!(flags & CFLGS_OFF_SLAB)) {
557 base = sizeof(struct slab);
558 extra = sizeof(kmem_bufctl_t);
561 while (i*size + ALIGN(base+i*extra, align) <= wastage)
571 wastage -= ALIGN(base+i*extra, align);
572 *left_over = wastage;
575 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
577 static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
579 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
580 function, cachep->name, msg);
585 * Start the reap timer running on the target CPU. We run at around 1 to 2Hz.
586 * Add the CPU number into the expiry time to minimize the possibility of the
587 * CPUs getting into lockstep and contending for the global cache chain lock.
589 static void __devinit start_cpu_timer(int cpu)
591 struct timer_list *rt = &per_cpu(reap_timers, cpu);
593 if (rt->function == NULL) {
595 rt->expires = jiffies + HZ + 3*cpu;
597 rt->function = reap_timer_fnc;
598 add_timer_on(rt, cpu);
602 #ifdef CONFIG_HOTPLUG_CPU
603 static void stop_cpu_timer(int cpu)
605 struct timer_list *rt = &per_cpu(reap_timers, cpu);
609 WARN_ON(timer_pending(rt));
615 static struct array_cache *alloc_arraycache(int cpu, int entries, int batchcount)
617 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
618 struct array_cache *nc = NULL;
621 nc = kmem_cache_alloc_node(kmem_find_general_cachep(memsize,
622 GFP_KERNEL), cpu_to_node(cpu));
625 nc = kmalloc(memsize, GFP_KERNEL);
629 nc->batchcount = batchcount;
635 static int __devinit cpuup_callback(struct notifier_block *nfb,
636 unsigned long action,
639 long cpu = (long)hcpu;
640 kmem_cache_t* cachep;
644 down(&cache_chain_sem);
645 list_for_each_entry(cachep, &cache_chain, next) {
646 struct array_cache *nc;
648 nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount);
652 spin_lock_irq(&cachep->spinlock);
653 cachep->array[cpu] = nc;
654 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
656 spin_unlock_irq(&cachep->spinlock);
659 up(&cache_chain_sem);
662 start_cpu_timer(cpu);
664 #ifdef CONFIG_HOTPLUG_CPU
668 case CPU_UP_CANCELED:
669 down(&cache_chain_sem);
671 list_for_each_entry(cachep, &cache_chain, next) {
672 struct array_cache *nc;
674 spin_lock_irq(&cachep->spinlock);
675 /* cpu is dead; no one can alloc from it. */
676 nc = cachep->array[cpu];
677 cachep->array[cpu] = NULL;
678 cachep->free_limit -= cachep->batchcount;
679 free_block(cachep, ac_entry(nc), nc->avail);
680 spin_unlock_irq(&cachep->spinlock);
683 up(&cache_chain_sem);
689 up(&cache_chain_sem);
693 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
696 * Called after the gfp() functions have been enabled, and before smp_init().
698 void __init kmem_cache_init(void)
701 struct cache_sizes *sizes;
702 struct cache_names *names;
705 * Fragmentation resistance on low memory - only use bigger
706 * page orders on machines with more than 32MB of memory.
708 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
709 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
712 /* Bootstrap is tricky, because several objects are allocated
713 * from caches that do not exist yet:
714 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
715 * structures of all caches, except cache_cache itself: cache_cache
716 * is statically allocated.
717 * Initially an __init data area is used for the head array, it's
718 * replaced with a kmalloc allocated array at the end of the bootstrap.
719 * 2) Create the first kmalloc cache.
720 * The kmem_cache_t for the new cache is allocated normally. An __init
721 * data area is used for the head array.
722 * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
723 * 4) Replace the __init data head arrays for cache_cache and the first
724 * kmalloc cache with kmalloc allocated arrays.
725 * 5) Resize the head arrays of the kmalloc caches to their final sizes.
728 /* 1) create the cache_cache */
729 init_MUTEX(&cache_chain_sem);
730 INIT_LIST_HEAD(&cache_chain);
731 list_add(&cache_cache.next, &cache_chain);
732 cache_cache.colour_off = cache_line_size();
733 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
735 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
737 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
738 &left_over, &cache_cache.num);
739 if (!cache_cache.num)
742 cache_cache.colour = left_over/cache_cache.colour_off;
743 cache_cache.colour_next = 0;
744 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
745 sizeof(struct slab), cache_line_size());
747 /* 2+3) create the kmalloc caches */
748 sizes = malloc_sizes;
751 while (sizes->cs_size) {
752 /* For performance, all the general caches are L1 aligned.
753 * This should be particularly beneficial on SMP boxes, as it
754 * eliminates "false sharing".
755 * Note for systems short on memory removing the alignment will
756 * allow tighter packing of the smaller caches. */
757 sizes->cs_cachep = kmem_cache_create(names->name,
758 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
759 SLAB_PANIC, NULL, NULL);
761 /* Inc off-slab bufctl limit until the ceiling is hit. */
762 if (!(OFF_SLAB(sizes->cs_cachep))) {
763 offslab_limit = sizes->cs_size-sizeof(struct slab);
764 offslab_limit /= sizeof(kmem_bufctl_t);
767 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
768 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
769 (SLAB_CACHE_DMA | SLAB_PANIC), NULL, NULL);
774 /* 4) Replace the bootstrap head arrays */
778 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
780 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
781 memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
782 cache_cache.array[smp_processor_id()] = ptr;
785 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
787 BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
788 memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
789 sizeof(struct arraycache_init));
790 malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
794 /* 5) resize the head arrays to their final sizes */
796 kmem_cache_t *cachep;
797 down(&cache_chain_sem);
798 list_for_each_entry(cachep, &cache_chain, next)
799 enable_cpucache(cachep);
800 up(&cache_chain_sem);
804 g_cpucache_up = FULL;
806 /* Register a cpu startup notifier callback
807 * that initializes ac_data for all new cpus
809 register_cpu_notifier(&cpucache_notifier);
812 /* The reap timers are started later, with a module init call:
813 * That part of the kernel is not yet operational.
817 int __init cpucache_init(void)
822 * Register the timers that return unneeded
825 for (cpu = 0; cpu < NR_CPUS; cpu++) {
827 start_cpu_timer(cpu);
833 __initcall(cpucache_init);
836 * Interface to system's page allocator. No need to hold the cache-lock.
838 * If we requested dmaable memory, we will get it. Even if we
839 * did not request dmaable memory, we might get it, but that
840 * would be relatively rare and ignorable.
842 static void *kmem_getpages(kmem_cache_t *cachep, int flags, int nodeid)
848 flags |= cachep->gfpflags;
849 if (likely(nodeid == -1)) {
850 addr = (void*)__get_free_pages(flags, cachep->gfporder);
853 page = virt_to_page(addr);
855 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
858 addr = page_address(page);
861 i = (1 << cachep->gfporder);
862 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
863 atomic_add(i, &slab_reclaim_pages);
864 add_page_state(nr_slab, i);
873 * Interface to system's page release.
875 static inline void kmem_freepages(kmem_cache_t *cachep, void *addr)
877 unsigned long i = (1<<cachep->gfporder);
878 struct page *page = virt_to_page(addr);
879 const unsigned long nr_freed = i;
882 if (!TestClearPageSlab(page))
886 sub_page_state(nr_slab, nr_freed);
887 if (current->reclaim_state)
888 current->reclaim_state->reclaimed_slab += nr_freed;
889 free_pages((unsigned long)addr, cachep->gfporder);
890 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
891 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
896 #ifdef CONFIG_DEBUG_PAGEALLOC
897 static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, unsigned long caller)
899 int size = obj_reallen(cachep);
901 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
903 if (size < 5*sizeof(unsigned long))
908 *addr++=smp_processor_id();
909 size -= 3*sizeof(unsigned long);
911 unsigned long *sptr = &caller;
912 unsigned long svalue;
914 while (!kstack_end(sptr)) {
916 if (kernel_text_address(svalue)) {
918 size -= sizeof(unsigned long);
919 if (size <= sizeof(unsigned long))
929 static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
931 int size = obj_reallen(cachep);
932 addr = &((char*)addr)[obj_dbghead(cachep)];
934 memset(addr, val, size);
935 *(unsigned char *)(addr+size-1) = POISON_END;
938 static void dump_line(char *data, int offset, int limit)
941 printk(KERN_ERR "%03x:", offset);
942 for (i=0;i<limit;i++) {
943 printk(" %02x", (unsigned char)data[offset+i]);
949 static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
955 if (cachep->flags & SLAB_RED_ZONE) {
956 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
957 *dbg_redzone1(cachep, objp),
958 *dbg_redzone2(cachep, objp));
961 if (cachep->flags & SLAB_STORE_USER) {
962 printk(KERN_ERR "Last user: [<%p>]", *dbg_userword(cachep, objp));
963 print_symbol("(%s)", (unsigned long)*dbg_userword(cachep, objp));
966 realobj = (char*)objp+obj_dbghead(cachep);
967 size = obj_reallen(cachep);
968 for (i=0; i<size && lines;i+=16, lines--) {
973 dump_line(realobj, i, limit);
980 static void check_poison_obj(kmem_cache_t *cachep, void *objp)
986 realobj = (char*)objp+obj_dbghead(cachep);
987 size = obj_reallen(cachep);
989 for (i=0;i<size;i++) {
990 char exp = POISON_FREE;
993 if (realobj[i] != exp) {
998 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1000 print_objinfo(cachep, objp, 0);
1002 /* Hexdump the affected line */
1007 dump_line(realobj, i, limit);
1010 /* Limit to 5 lines */
1016 /* Print some data about the neighboring objects, if they
1019 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1022 objnr = (objp-slabp->s_mem)/cachep->objsize;
1024 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1025 realobj = (char*)objp+obj_dbghead(cachep);
1026 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1028 print_objinfo(cachep, objp, 2);
1030 if (objnr+1 < cachep->num) {
1031 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1032 realobj = (char*)objp+obj_dbghead(cachep);
1033 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1035 print_objinfo(cachep, objp, 2);
1041 /* Destroy all the objs in a slab, and release the mem back to the system.
1042 * Before calling the slab must have been unlinked from the cache.
1043 * The cache-lock is not held/needed.
1045 static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1049 for (i = 0; i < cachep->num; i++) {
1050 void *objp = slabp->s_mem + cachep->objsize * i;
1052 if (cachep->flags & SLAB_POISON) {
1053 #ifdef CONFIG_DEBUG_PAGEALLOC
1054 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1055 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1057 check_poison_obj(cachep, objp);
1059 check_poison_obj(cachep, objp);
1062 if (cachep->flags & SLAB_RED_ZONE) {
1063 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1064 slab_error(cachep, "start of a freed object "
1066 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1067 slab_error(cachep, "end of a freed object "
1070 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1071 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1076 for (i = 0; i < cachep->num; i++) {
1077 void* objp = slabp->s_mem+cachep->objsize*i;
1078 (cachep->dtor)(objp, cachep, 0);
1083 kmem_freepages(cachep, slabp->s_mem-slabp->colouroff);
1084 if (OFF_SLAB(cachep))
1085 kmem_cache_free(cachep->slabp_cache, slabp);
1089 * kmem_cache_create - Create a cache.
1090 * @name: A string which is used in /proc/slabinfo to identify this cache.
1091 * @size: The size of objects to be created in this cache.
1092 * @align: The required alignment for the objects.
1093 * @flags: SLAB flags
1094 * @ctor: A constructor for the objects.
1095 * @dtor: A destructor for the objects.
1097 * Returns a ptr to the cache on success, NULL on failure.
1098 * Cannot be called within a int, but can be interrupted.
1099 * The @ctor is run when new pages are allocated by the cache
1100 * and the @dtor is run before the pages are handed back.
1102 * @name must be valid until the cache is destroyed. This implies that
1103 * the module calling this has to destroy the cache before getting
1108 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1109 * to catch references to uninitialised memory.
1111 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1112 * for buffer overruns.
1114 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1117 * %SLAB_HWCACHE_ALIGN - This flag has no effect and will be removed soon.
1121 kmem_cache_create (const char *name, size_t size, size_t align,
1122 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1123 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1125 size_t left_over, slab_size;
1126 kmem_cache_t *cachep = NULL;
1129 * Sanity checks... these are all serious usage bugs.
1133 (size < BYTES_PER_WORD) ||
1134 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1137 printk(KERN_ERR "%s: Early error in slab %s\n",
1138 __FUNCTION__, name);
1143 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1144 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1145 /* No constructor, but inital state check requested */
1146 printk(KERN_ERR "%s: No con, but init state check "
1147 "requested - %s\n", __FUNCTION__, name);
1148 flags &= ~SLAB_DEBUG_INITIAL;
1153 * Enable redzoning and last user accounting, except for caches with
1154 * large objects, if the increased size would increase the object size
1155 * above the next power of two: caches with object sizes just above a
1156 * power of two have a significant amount of internal fragmentation.
1158 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1159 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1160 flags |= SLAB_POISON;
1164 * Always checks flags, a caller might be expecting debug
1165 * support which isn't available.
1167 if (flags & ~CREATE_MASK)
1171 /* combinations of forced alignment and advanced debugging is
1172 * not yet implemented.
1174 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1176 if (flags & SLAB_HWCACHE_ALIGN) {
1177 /* Default alignment: as specified by the arch code.
1178 * Except if an object is really small, then squeeze multiple
1179 * into one cacheline.
1181 align = cache_line_size();
1182 while (size <= align/2)
1185 align = BYTES_PER_WORD;
1189 /* Get cache's description obj. */
1190 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1193 memset(cachep, 0, sizeof(kmem_cache_t));
1195 /* Check that size is in terms of words. This is needed to avoid
1196 * unaligned accesses for some archs when redzoning is used, and makes
1197 * sure any on-slab bufctl's are also correctly aligned.
1199 if (size & (BYTES_PER_WORD-1)) {
1200 size += (BYTES_PER_WORD-1);
1201 size &= ~(BYTES_PER_WORD-1);
1205 cachep->reallen = size;
1207 if (flags & SLAB_RED_ZONE) {
1208 /* redzoning only works with word aligned caches */
1209 align = BYTES_PER_WORD;
1211 /* add space for red zone words */
1212 cachep->dbghead += BYTES_PER_WORD;
1213 size += 2*BYTES_PER_WORD;
1215 if (flags & SLAB_STORE_USER) {
1216 /* user store requires word alignment and
1217 * one word storage behind the end of the real
1220 align = BYTES_PER_WORD;
1221 size += BYTES_PER_WORD;
1223 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1224 if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1225 cachep->dbghead += PAGE_SIZE - size;
1231 /* Determine if the slab management is 'on' or 'off' slab. */
1232 if (size >= (PAGE_SIZE>>3))
1234 * Size is large, assume best to place the slab management obj
1235 * off-slab (should allow better packing of objs).
1237 flags |= CFLGS_OFF_SLAB;
1239 size = ALIGN(size, align);
1241 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1243 * A VFS-reclaimable slab tends to have most allocations
1244 * as GFP_NOFS and we really don't want to have to be allocating
1245 * higher-order pages when we are unable to shrink dcache.
1247 cachep->gfporder = 0;
1248 cache_estimate(cachep->gfporder, size, align, flags,
1249 &left_over, &cachep->num);
1252 * Calculate size (in pages) of slabs, and the num of objs per
1253 * slab. This could be made much more intelligent. For now,
1254 * try to avoid using high page-orders for slabs. When the
1255 * gfp() funcs are more friendly towards high-order requests,
1256 * this should be changed.
1259 unsigned int break_flag = 0;
1261 cache_estimate(cachep->gfporder, size, align, flags,
1262 &left_over, &cachep->num);
1265 if (cachep->gfporder >= MAX_GFP_ORDER)
1269 if (flags & CFLGS_OFF_SLAB &&
1270 cachep->num > offslab_limit) {
1271 /* This num of objs will cause problems. */
1278 * Large num of objs is good, but v. large slabs are
1279 * currently bad for the gfp()s.
1281 if (cachep->gfporder >= slab_break_gfp_order)
1284 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1285 break; /* Acceptable internal fragmentation. */
1292 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1293 kmem_cache_free(&cache_cache, cachep);
1297 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1298 + sizeof(struct slab), align);
1301 * If the slab has been placed off-slab, and we have enough space then
1302 * move it on-slab. This is at the expense of any extra colouring.
1304 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1305 flags &= ~CFLGS_OFF_SLAB;
1306 left_over -= slab_size;
1309 if (flags & CFLGS_OFF_SLAB) {
1310 /* really off slab. No need for manual alignment */
1311 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1314 cachep->colour_off = cache_line_size();
1315 /* Offset must be a multiple of the alignment. */
1316 if (cachep->colour_off < align)
1317 cachep->colour_off = align;
1318 cachep->colour = left_over/cachep->colour_off;
1319 cachep->slab_size = slab_size;
1320 cachep->flags = flags;
1321 cachep->gfpflags = 0;
1322 if (flags & SLAB_CACHE_DMA)
1323 cachep->gfpflags |= GFP_DMA;
1324 spin_lock_init(&cachep->spinlock);
1325 cachep->objsize = size;
1327 INIT_LIST_HEAD(&cachep->lists.slabs_full);
1328 INIT_LIST_HEAD(&cachep->lists.slabs_partial);
1329 INIT_LIST_HEAD(&cachep->lists.slabs_free);
1331 if (flags & CFLGS_OFF_SLAB)
1332 cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
1333 cachep->ctor = ctor;
1334 cachep->dtor = dtor;
1335 cachep->name = name;
1337 /* Don't let CPUs to come and go */
1340 if (g_cpucache_up == FULL) {
1341 enable_cpucache(cachep);
1343 if (g_cpucache_up == NONE) {
1344 /* Note: the first kmem_cache_create must create
1345 * the cache that's used by kmalloc(24), otherwise
1346 * the creation of further caches will BUG().
1348 cachep->array[smp_processor_id()] =
1349 &initarray_generic.cache;
1350 g_cpucache_up = PARTIAL;
1352 cachep->array[smp_processor_id()] =
1353 kmalloc(sizeof(struct arraycache_init),
1356 BUG_ON(!ac_data(cachep));
1357 ac_data(cachep)->avail = 0;
1358 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1359 ac_data(cachep)->batchcount = 1;
1360 ac_data(cachep)->touched = 0;
1361 cachep->batchcount = 1;
1362 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1363 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
1367 cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
1368 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1370 /* Need the semaphore to access the chain. */
1371 down(&cache_chain_sem);
1373 struct list_head *p;
1374 mm_segment_t old_fs;
1378 list_for_each(p, &cache_chain) {
1379 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1383 * This happens when the module gets unloaded and
1384 * doesn't destroy its slab cache and noone else reuses
1385 * the vmalloc area of the module. Print a warning.
1387 #ifdef CONFIG_X86_UACCESS_INDIRECT
1388 if (__direct_get_user(tmp,pc->name)) {
1390 if (__get_user(tmp,pc->name)) {
1392 printk("SLAB: cache with size %d has lost its "
1393 "name\n", pc->objsize);
1396 if (!strcmp(pc->name,name)) {
1397 printk("kmem_cache_create: duplicate "
1399 up(&cache_chain_sem);
1400 unlock_cpu_hotplug();
1407 /* cache setup completed, link it into the list */
1408 list_add(&cachep->next, &cache_chain);
1409 up(&cache_chain_sem);
1410 unlock_cpu_hotplug();
1412 if (!cachep && (flags & SLAB_PANIC))
1413 panic("kmem_cache_create(): failed to create slab `%s'\n",
1417 EXPORT_SYMBOL(kmem_cache_create);
1419 static inline void check_irq_off(void)
1422 BUG_ON(!irqs_disabled());
1426 static inline void check_irq_on(void)
1429 BUG_ON(irqs_disabled());
1433 static inline void check_spinlock_acquired(kmem_cache_t *cachep)
1437 BUG_ON(spin_trylock(&cachep->spinlock));
1442 * Waits for all CPUs to execute func().
1444 static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1449 local_irq_disable();
1453 if (smp_call_function(func, arg, 1, 1))
1459 static void drain_array_locked(kmem_cache_t* cachep,
1460 struct array_cache *ac, int force);
1462 static void do_drain(void *arg)
1464 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1465 struct array_cache *ac;
1468 ac = ac_data(cachep);
1469 spin_lock(&cachep->spinlock);
1470 free_block(cachep, &ac_entry(ac)[0], ac->avail);
1471 spin_unlock(&cachep->spinlock);
1475 static void drain_cpu_caches(kmem_cache_t *cachep)
1477 smp_call_function_all_cpus(do_drain, cachep);
1479 spin_lock_irq(&cachep->spinlock);
1480 if (cachep->lists.shared)
1481 drain_array_locked(cachep, cachep->lists.shared, 1);
1482 spin_unlock_irq(&cachep->spinlock);
1486 /* NUMA shrink all list3s */
1487 static int __cache_shrink(kmem_cache_t *cachep)
1492 drain_cpu_caches(cachep);
1495 spin_lock_irq(&cachep->spinlock);
1498 struct list_head *p;
1500 p = cachep->lists.slabs_free.prev;
1501 if (p == &cachep->lists.slabs_free)
1504 slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
1509 list_del(&slabp->list);
1511 cachep->lists.free_objects -= cachep->num;
1512 spin_unlock_irq(&cachep->spinlock);
1513 slab_destroy(cachep, slabp);
1514 spin_lock_irq(&cachep->spinlock);
1516 ret = !list_empty(&cachep->lists.slabs_full) ||
1517 !list_empty(&cachep->lists.slabs_partial);
1518 spin_unlock_irq(&cachep->spinlock);
1523 * kmem_cache_shrink - Shrink a cache.
1524 * @cachep: The cache to shrink.
1526 * Releases as many slabs as possible for a cache.
1527 * To help debugging, a zero exit status indicates all slabs were released.
1529 int kmem_cache_shrink(kmem_cache_t *cachep)
1531 if (!cachep || in_interrupt())
1534 return __cache_shrink(cachep);
1537 EXPORT_SYMBOL(kmem_cache_shrink);
1540 * kmem_cache_destroy - delete a cache
1541 * @cachep: the cache to destroy
1543 * Remove a kmem_cache_t object from the slab cache.
1544 * Returns 0 on success.
1546 * It is expected this function will be called by a module when it is
1547 * unloaded. This will remove the cache completely, and avoid a duplicate
1548 * cache being allocated each time a module is loaded and unloaded, if the
1549 * module doesn't have persistent in-kernel storage across loads and unloads.
1551 * The cache must be empty before calling this function.
1553 * The caller must guarantee that noone will allocate memory from the cache
1554 * during the kmem_cache_destroy().
1556 int kmem_cache_destroy (kmem_cache_t * cachep)
1560 if (!cachep || in_interrupt())
1563 /* Don't let CPUs to come and go */
1566 /* Find the cache in the chain of caches. */
1567 down(&cache_chain_sem);
1569 * the chain is never empty, cache_cache is never destroyed
1571 list_del(&cachep->next);
1572 up(&cache_chain_sem);
1574 if (__cache_shrink(cachep)) {
1575 slab_error(cachep, "Can't free all objects");
1576 down(&cache_chain_sem);
1577 list_add(&cachep->next,&cache_chain);
1578 up(&cache_chain_sem);
1579 unlock_cpu_hotplug();
1583 /* no cpu_online check required here since we clear the percpu
1584 * array on cpu offline and set this to NULL.
1586 for (i = 0; i < NR_CPUS; i++)
1587 kfree(cachep->array[i]);
1589 /* NUMA: free the list3 structures */
1590 kfree(cachep->lists.shared);
1591 cachep->lists.shared = NULL;
1592 kmem_cache_free(&cache_cache, cachep);
1594 unlock_cpu_hotplug();
1599 EXPORT_SYMBOL(kmem_cache_destroy);
1601 /* Get the memory for a slab management obj. */
1602 static inline struct slab* alloc_slabmgmt (kmem_cache_t *cachep,
1603 void *objp, int colour_off, int local_flags)
1607 if (OFF_SLAB(cachep)) {
1608 /* Slab management obj is off-slab. */
1609 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
1613 slabp = objp+colour_off;
1614 colour_off += cachep->slab_size;
1617 slabp->colouroff = colour_off;
1618 slabp->s_mem = objp+colour_off;
1623 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
1625 return (kmem_bufctl_t *)(slabp+1);
1628 static void cache_init_objs (kmem_cache_t * cachep,
1629 struct slab * slabp, unsigned long ctor_flags)
1633 for (i = 0; i < cachep->num; i++) {
1634 void* objp = slabp->s_mem+cachep->objsize*i;
1636 /* need to poison the objs? */
1637 if (cachep->flags & SLAB_POISON)
1638 poison_obj(cachep, objp, POISON_FREE);
1639 if (cachep->flags & SLAB_STORE_USER)
1640 *dbg_userword(cachep, objp) = NULL;
1642 if (cachep->flags & SLAB_RED_ZONE) {
1643 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1644 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1647 * Constructors are not allowed to allocate memory from
1648 * the same cache which they are a constructor for.
1649 * Otherwise, deadlock. They must also be threaded.
1651 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
1652 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
1654 if (cachep->flags & SLAB_RED_ZONE) {
1655 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1656 slab_error(cachep, "constructor overwrote the"
1657 " end of an object");
1658 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1659 slab_error(cachep, "constructor overwrote the"
1660 " start of an object");
1662 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
1663 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1666 cachep->ctor(objp, cachep, ctor_flags);
1668 slab_bufctl(slabp)[i] = i+1;
1670 slab_bufctl(slabp)[i-1] = BUFCTL_END;
1674 static void kmem_flagcheck(kmem_cache_t *cachep, int flags)
1676 if (flags & SLAB_DMA) {
1677 if (!(cachep->gfpflags & GFP_DMA))
1680 if (cachep->gfpflags & GFP_DMA)
1685 static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
1690 /* Nasty!!!!!! I hope this is OK. */
1691 i = 1 << cachep->gfporder;
1692 page = virt_to_page(objp);
1694 SET_PAGE_CACHE(page, cachep);
1695 SET_PAGE_SLAB(page, slabp);
1701 * Grow (by 1) the number of slabs within a cache. This is called by
1702 * kmem_cache_alloc() when there are no active objs left in a cache.
1704 static int cache_grow (kmem_cache_t * cachep, int flags)
1710 unsigned long ctor_flags;
1712 /* Be lazy and only check for valid flags here,
1713 * keeping it out of the critical path in kmem_cache_alloc().
1715 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
1717 if (flags & SLAB_NO_GROW)
1720 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
1721 local_flags = (flags & SLAB_LEVEL_MASK);
1722 if (!(local_flags & __GFP_WAIT))
1724 * Not allowed to sleep. Need to tell a constructor about
1725 * this - it might need to know...
1727 ctor_flags |= SLAB_CTOR_ATOMIC;
1729 /* About to mess with non-constant members - lock. */
1731 spin_lock(&cachep->spinlock);
1733 /* Get colour for the slab, and cal the next value. */
1734 offset = cachep->colour_next;
1735 cachep->colour_next++;
1736 if (cachep->colour_next >= cachep->colour)
1737 cachep->colour_next = 0;
1738 offset *= cachep->colour_off;
1740 spin_unlock(&cachep->spinlock);
1742 if (local_flags & __GFP_WAIT)
1746 * The test for missing atomic flag is performed here, rather than
1747 * the more obvious place, simply to reduce the critical path length
1748 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
1749 * will eventually be caught here (where it matters).
1751 kmem_flagcheck(cachep, flags);
1754 /* Get mem for the objs. */
1755 if (!(objp = kmem_getpages(cachep, flags, -1)))
1758 /* Get slab management. */
1759 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
1762 set_slab_attr(cachep, slabp, objp);
1764 cache_init_objs(cachep, slabp, ctor_flags);
1766 if (local_flags & __GFP_WAIT)
1767 local_irq_disable();
1769 spin_lock(&cachep->spinlock);
1771 /* Make slab active. */
1772 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
1773 STATS_INC_GROWN(cachep);
1774 list3_data(cachep)->free_objects += cachep->num;
1775 spin_unlock(&cachep->spinlock);
1778 kmem_freepages(cachep, objp);
1780 if (local_flags & __GFP_WAIT)
1781 local_irq_disable();
1786 * Perform extra freeing checks:
1787 * - detect bad pointers.
1788 * - POISON/RED_ZONE checking
1789 * - destructor calls, for caches with POISON+dtor
1791 static inline void kfree_debugcheck(const void *objp)
1796 if (!virt_addr_valid(objp)) {
1797 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
1798 (unsigned long)objp);
1801 page = virt_to_page(objp);
1802 if (!PageSlab(page)) {
1803 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
1809 static inline void *cache_free_debugcheck (kmem_cache_t * cachep, void * objp, void *caller)
1816 objp -= obj_dbghead(cachep);
1817 kfree_debugcheck(objp);
1818 page = virt_to_page(objp);
1820 if (GET_PAGE_CACHE(page) != cachep) {
1821 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
1822 GET_PAGE_CACHE(page),cachep);
1823 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
1824 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
1827 slabp = GET_PAGE_SLAB(page);
1829 if (cachep->flags & SLAB_RED_ZONE) {
1830 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
1831 slab_error(cachep, "double free, or memory outside"
1832 " object was overwritten");
1833 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1834 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
1836 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1837 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1839 if (cachep->flags & SLAB_STORE_USER)
1840 *dbg_userword(cachep, objp) = caller;
1842 objnr = (objp-slabp->s_mem)/cachep->objsize;
1844 BUG_ON(objnr >= cachep->num);
1845 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
1847 if (cachep->flags & SLAB_DEBUG_INITIAL) {
1848 /* Need to call the slab's constructor so the
1849 * caller can perform a verify of its state (debugging).
1850 * Called without the cache-lock held.
1852 cachep->ctor(objp+obj_dbghead(cachep),
1853 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
1855 if (cachep->flags & SLAB_POISON && cachep->dtor) {
1856 /* we want to cache poison the object,
1857 * call the destruction callback
1859 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
1861 if (cachep->flags & SLAB_POISON) {
1862 #ifdef CONFIG_DEBUG_PAGEALLOC
1863 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
1864 store_stackinfo(cachep, objp, (unsigned long)caller);
1865 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1867 poison_obj(cachep, objp, POISON_FREE);
1870 poison_obj(cachep, objp, POISON_FREE);
1877 static inline void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
1883 check_spinlock_acquired(cachep);
1884 /* Check slab's freelist to see if this obj is there. */
1885 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
1887 if (entries > cachep->num || i < 0 || i >= cachep->num)
1890 if (entries != cachep->num - slabp->inuse) {
1893 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
1894 cachep->name, cachep->num, slabp, slabp->inuse);
1895 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
1897 printk("\n%03x:", i);
1898 printk(" %02x", ((unsigned char*)slabp)[i]);
1906 static void* cache_alloc_refill(kmem_cache_t* cachep, int flags)
1909 struct kmem_list3 *l3;
1910 struct array_cache *ac;
1913 ac = ac_data(cachep);
1915 batchcount = ac->batchcount;
1916 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
1917 /* if there was little recent activity on this
1918 * cache, then perform only a partial refill.
1919 * Otherwise we could generate refill bouncing.
1921 batchcount = BATCHREFILL_LIMIT;
1923 l3 = list3_data(cachep);
1925 BUG_ON(ac->avail > 0);
1926 spin_lock(&cachep->spinlock);
1928 struct array_cache *shared_array = l3->shared;
1929 if (shared_array->avail) {
1930 if (batchcount > shared_array->avail)
1931 batchcount = shared_array->avail;
1932 shared_array->avail -= batchcount;
1933 ac->avail = batchcount;
1934 memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
1935 sizeof(void*)*batchcount);
1936 shared_array->touched = 1;
1940 while (batchcount > 0) {
1941 struct list_head *entry;
1943 /* Get slab alloc is to come from. */
1944 entry = l3->slabs_partial.next;
1945 if (entry == &l3->slabs_partial) {
1946 l3->free_touched = 1;
1947 entry = l3->slabs_free.next;
1948 if (entry == &l3->slabs_free)
1952 slabp = list_entry(entry, struct slab, list);
1953 check_slabp(cachep, slabp);
1954 check_spinlock_acquired(cachep);
1955 while (slabp->inuse < cachep->num && batchcount--) {
1957 STATS_INC_ALLOCED(cachep);
1958 STATS_INC_ACTIVE(cachep);
1959 STATS_SET_HIGH(cachep);
1961 /* get obj pointer */
1962 ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
1965 next = slab_bufctl(slabp)[slabp->free];
1967 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
1971 check_slabp(cachep, slabp);
1973 /* move slabp to correct slabp list: */
1974 list_del(&slabp->list);
1975 if (slabp->free == BUFCTL_END)
1976 list_add(&slabp->list, &l3->slabs_full);
1978 list_add(&slabp->list, &l3->slabs_partial);
1982 l3->free_objects -= ac->avail;
1984 spin_unlock(&cachep->spinlock);
1986 if (unlikely(!ac->avail)) {
1988 x = cache_grow(cachep, flags);
1990 // cache_grow can reenable interrupts, then ac could change.
1991 ac = ac_data(cachep);
1992 if (!x && ac->avail == 0) // no objects in sight? abort
1995 if (!ac->avail) // objects refilled by interrupt?
1999 return ac_entry(ac)[--ac->avail];
2003 cache_alloc_debugcheck_before(kmem_cache_t *cachep, int flags)
2005 might_sleep_if(flags & __GFP_WAIT);
2007 kmem_flagcheck(cachep, flags);
2011 static inline void *
2012 cache_alloc_debugcheck_after(kmem_cache_t *cachep,
2013 unsigned long flags, void *objp, void *caller)
2018 if (cachep->flags & SLAB_POISON) {
2019 #ifdef CONFIG_DEBUG_PAGEALLOC
2020 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2021 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2023 check_poison_obj(cachep, objp);
2025 check_poison_obj(cachep, objp);
2027 poison_obj(cachep, objp, POISON_INUSE);
2029 if (cachep->flags & SLAB_STORE_USER)
2030 *dbg_userword(cachep, objp) = caller;
2032 if (cachep->flags & SLAB_RED_ZONE) {
2033 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2034 slab_error(cachep, "double free, or memory outside"
2035 " object was overwritten");
2036 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2037 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2039 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2040 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2042 objp += obj_dbghead(cachep);
2043 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2044 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2046 if (!(flags & __GFP_WAIT))
2047 ctor_flags |= SLAB_CTOR_ATOMIC;
2049 cachep->ctor(objp, cachep, ctor_flags);
2056 static inline void * __cache_alloc (kmem_cache_t *cachep, int flags)
2058 unsigned long save_flags;
2060 struct array_cache *ac;
2062 cache_alloc_debugcheck_before(cachep, flags);
2064 local_irq_save(save_flags);
2065 ac = ac_data(cachep);
2066 if (likely(ac->avail)) {
2067 STATS_INC_ALLOCHIT(cachep);
2069 objp = ac_entry(ac)[--ac->avail];
2071 STATS_INC_ALLOCMISS(cachep);
2072 objp = cache_alloc_refill(cachep, flags);
2074 local_irq_restore(save_flags);
2075 objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0));
2080 * NUMA: different approach needed if the spinlock is moved into
2084 static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2088 check_spinlock_acquired(cachep);
2090 /* NUMA: move add into loop */
2091 cachep->lists.free_objects += nr_objects;
2093 for (i = 0; i < nr_objects; i++) {
2094 void *objp = objpp[i];
2098 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2099 list_del(&slabp->list);
2100 objnr = (objp - slabp->s_mem) / cachep->objsize;
2101 check_slabp(cachep, slabp);
2103 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2104 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
2105 cachep->name, objp);
2109 slab_bufctl(slabp)[objnr] = slabp->free;
2110 slabp->free = objnr;
2111 STATS_DEC_ACTIVE(cachep);
2113 check_slabp(cachep, slabp);
2115 /* fixup slab chains */
2116 if (slabp->inuse == 0) {
2117 if (cachep->lists.free_objects > cachep->free_limit) {
2118 cachep->lists.free_objects -= cachep->num;
2119 slab_destroy(cachep, slabp);
2121 list_add(&slabp->list,
2122 &list3_data_ptr(cachep, objp)->slabs_free);
2125 /* Unconditionally move a slab to the end of the
2126 * partial list on free - maximum time for the
2127 * other objects to be freed, too.
2129 list_add_tail(&slabp->list,
2130 &list3_data_ptr(cachep, objp)->slabs_partial);
2135 static void cache_flusharray (kmem_cache_t* cachep, struct array_cache *ac)
2139 batchcount = ac->batchcount;
2141 BUG_ON(!batchcount || batchcount > ac->avail);
2144 spin_lock(&cachep->spinlock);
2145 if (cachep->lists.shared) {
2146 struct array_cache *shared_array = cachep->lists.shared;
2147 int max = shared_array->limit-shared_array->avail;
2149 if (batchcount > max)
2151 memcpy(&ac_entry(shared_array)[shared_array->avail],
2153 sizeof(void*)*batchcount);
2154 shared_array->avail += batchcount;
2159 free_block(cachep, &ac_entry(ac)[0], batchcount);
2164 struct list_head *p;
2166 p = list3_data(cachep)->slabs_free.next;
2167 while (p != &(list3_data(cachep)->slabs_free)) {
2170 slabp = list_entry(p, struct slab, list);
2171 BUG_ON(slabp->inuse);
2176 STATS_SET_FREEABLE(cachep, i);
2179 spin_unlock(&cachep->spinlock);
2180 ac->avail -= batchcount;
2181 memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
2182 sizeof(void*)*ac->avail);
2187 * Release an obj back to its cache. If the obj has a constructed
2188 * state, it must be in this state _before_ it is released.
2190 * Called with disabled ints.
2192 static inline void __cache_free (kmem_cache_t *cachep, void* objp)
2194 struct array_cache *ac = ac_data(cachep);
2197 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2199 if (likely(ac->avail < ac->limit)) {
2200 STATS_INC_FREEHIT(cachep);
2201 ac_entry(ac)[ac->avail++] = objp;
2204 STATS_INC_FREEMISS(cachep);
2205 cache_flusharray(cachep, ac);
2206 ac_entry(ac)[ac->avail++] = objp;
2211 * kmem_cache_alloc - Allocate an object
2212 * @cachep: The cache to allocate from.
2213 * @flags: See kmalloc().
2215 * Allocate an object from this cache. The flags are only relevant
2216 * if the cache has no available objects.
2218 void * kmem_cache_alloc (kmem_cache_t *cachep, int flags)
2220 return __cache_alloc(cachep, flags);
2223 EXPORT_SYMBOL(kmem_cache_alloc);
2226 * kmem_ptr_validate - check if an untrusted pointer might
2228 * @cachep: the cache we're checking against
2229 * @ptr: pointer to validate
2231 * This verifies that the untrusted pointer looks sane:
2232 * it is _not_ a guarantee that the pointer is actually
2233 * part of the slab cache in question, but it at least
2234 * validates that the pointer can be dereferenced and
2235 * looks half-way sane.
2237 * Currently only used for dentry validation.
2239 int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2241 unsigned long addr = (unsigned long) ptr;
2242 unsigned long min_addr = PAGE_OFFSET;
2243 unsigned long align_mask = BYTES_PER_WORD-1;
2244 unsigned long size = cachep->objsize;
2247 if (unlikely(addr < min_addr))
2249 if (unlikely(addr > (unsigned long)high_memory - size))
2251 if (unlikely(addr & align_mask))
2253 if (unlikely(!kern_addr_valid(addr)))
2255 if (unlikely(!kern_addr_valid(addr + size - 1)))
2257 page = virt_to_page(ptr);
2258 if (unlikely(!PageSlab(page)))
2260 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2268 * kmem_cache_alloc_node - Allocate an object on the specified node
2269 * @cachep: The cache to allocate from.
2270 * @flags: See kmalloc().
2271 * @nodeid: node number of the target node.
2273 * Identical to kmem_cache_alloc, except that this function is slow
2274 * and can sleep. And it will allocate memory on the given node, which
2275 * can improve the performance for cpu bound structures.
2277 void *kmem_cache_alloc_node(kmem_cache_t *cachep, int nodeid)
2284 /* The main algorithms are not node aware, thus we have to cheat:
2285 * We bypass all caches and allocate a new slab.
2286 * The following code is a streamlined copy of cache_grow().
2289 /* Get colour for the slab, and update the next value. */
2290 spin_lock_irq(&cachep->spinlock);
2291 offset = cachep->colour_next;
2292 cachep->colour_next++;
2293 if (cachep->colour_next >= cachep->colour)
2294 cachep->colour_next = 0;
2295 offset *= cachep->colour_off;
2296 spin_unlock_irq(&cachep->spinlock);
2298 /* Get mem for the objs. */
2299 if (!(objp = kmem_getpages(cachep, GFP_KERNEL, nodeid)))
2302 /* Get slab management. */
2303 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, GFP_KERNEL)))
2306 set_slab_attr(cachep, slabp, objp);
2307 cache_init_objs(cachep, slabp, SLAB_CTOR_CONSTRUCTOR);
2309 /* The first object is ours: */
2310 objp = slabp->s_mem + slabp->free*cachep->objsize;
2312 next = slab_bufctl(slabp)[slabp->free];
2314 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2318 /* add the remaining objects into the cache */
2319 spin_lock_irq(&cachep->spinlock);
2320 check_slabp(cachep, slabp);
2321 STATS_INC_GROWN(cachep);
2322 /* Make slab active. */
2323 if (slabp->free == BUFCTL_END) {
2324 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_full));
2326 list_add_tail(&slabp->list,
2327 &(list3_data(cachep)->slabs_partial));
2328 list3_data(cachep)->free_objects += cachep->num-1;
2330 spin_unlock_irq(&cachep->spinlock);
2331 objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp,
2332 __builtin_return_address(0));
2335 kmem_freepages(cachep, objp);
2340 EXPORT_SYMBOL(kmem_cache_alloc_node);
2343 * kmalloc - allocate memory
2344 * @size: how many bytes of memory are required.
2345 * @flags: the type of memory to allocate.
2347 * kmalloc is the normal method of allocating memory
2350 * The @flags argument may be one of:
2352 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2354 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2356 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2358 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2359 * must be suitable for DMA. This can mean different things on different
2360 * platforms. For example, on i386, it means that the memory must come
2361 * from the first 16MB.
2363 void * __kmalloc (size_t size, int flags)
2365 struct cache_sizes *csizep = malloc_sizes;
2367 for (; csizep->cs_size; csizep++) {
2368 if (size > csizep->cs_size)
2371 /* This happens if someone tries to call
2372 * kmem_cache_create(), or kmalloc(), before
2373 * the generic caches are initialized.
2375 BUG_ON(csizep->cs_cachep == NULL);
2377 return __cache_alloc(flags & GFP_DMA ?
2378 csizep->cs_dmacachep : csizep->cs_cachep, flags);
2383 EXPORT_SYMBOL(__kmalloc);
2387 * __alloc_percpu - allocate one copy of the object for every present
2388 * cpu in the system, zeroing them.
2389 * Objects should be dereferenced using per_cpu_ptr/get_cpu_ptr
2392 * @size: how many bytes of memory are required.
2393 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2395 void *__alloc_percpu(size_t size, size_t align)
2398 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2403 for (i = 0; i < NR_CPUS; i++) {
2404 if (!cpu_possible(i))
2406 pdata->ptrs[i] = kmem_cache_alloc_node(
2407 kmem_find_general_cachep(size, GFP_KERNEL),
2410 if (!pdata->ptrs[i])
2412 memset(pdata->ptrs[i], 0, size);
2415 /* Catch derefs w/o wrappers */
2416 return (void *) (~(unsigned long) pdata);
2420 if (!cpu_possible(i))
2422 kfree(pdata->ptrs[i]);
2428 EXPORT_SYMBOL(__alloc_percpu);
2432 * kmem_cache_free - Deallocate an object
2433 * @cachep: The cache the allocation was from.
2434 * @objp: The previously allocated object.
2436 * Free an object which was previously allocated from this
2439 void kmem_cache_free (kmem_cache_t *cachep, void *objp)
2441 unsigned long flags;
2443 local_irq_save(flags);
2444 __cache_free(cachep, objp);
2445 local_irq_restore(flags);
2448 EXPORT_SYMBOL(kmem_cache_free);
2451 * kfree - free previously allocated memory
2452 * @objp: pointer returned by kmalloc.
2454 * Don't free memory not originally allocated by kmalloc()
2455 * or you will run into trouble.
2457 void kfree (const void *objp)
2460 unsigned long flags;
2464 local_irq_save(flags);
2465 kfree_debugcheck(objp);
2466 c = GET_PAGE_CACHE(virt_to_page(objp));
2467 __cache_free(c, (void*)objp);
2468 local_irq_restore(flags);
2471 EXPORT_SYMBOL(kfree);
2475 * free_percpu - free previously allocated percpu memory
2476 * @objp: pointer returned by alloc_percpu.
2478 * Don't free memory not originally allocated by alloc_percpu()
2479 * The complemented objp is to check for that.
2482 free_percpu(const void *objp)
2485 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
2487 for (i = 0; i < NR_CPUS; i++) {
2488 if (!cpu_possible(i))
2494 EXPORT_SYMBOL(free_percpu);
2497 unsigned int kmem_cache_size(kmem_cache_t *cachep)
2499 return obj_reallen(cachep);
2502 EXPORT_SYMBOL(kmem_cache_size);
2504 kmem_cache_t * kmem_find_general_cachep (size_t size, int gfpflags)
2506 struct cache_sizes *csizep = malloc_sizes;
2508 /* This function could be moved to the header file, and
2509 * made inline so consumers can quickly determine what
2510 * cache pointer they require.
2512 for ( ; csizep->cs_size; csizep++) {
2513 if (size > csizep->cs_size)
2517 return (gfpflags & GFP_DMA) ? csizep->cs_dmacachep : csizep->cs_cachep;
2520 EXPORT_SYMBOL(kmem_find_general_cachep);
2522 struct ccupdate_struct {
2523 kmem_cache_t *cachep;
2524 struct array_cache *new[NR_CPUS];
2527 static void do_ccupdate_local(void *info)
2529 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
2530 struct array_cache *old;
2533 old = ac_data(new->cachep);
2535 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
2536 new->new[smp_processor_id()] = old;
2540 static int do_tune_cpucache (kmem_cache_t* cachep, int limit, int batchcount, int shared)
2542 struct ccupdate_struct new;
2543 struct array_cache *new_shared;
2546 memset(&new.new,0,sizeof(new.new));
2547 for (i = 0; i < NR_CPUS; i++) {
2548 if (cpu_online(i)) {
2549 new.new[i] = alloc_arraycache(i, limit, batchcount);
2551 for (i--; i >= 0; i--) kfree(new.new[i]);
2558 new.cachep = cachep;
2560 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
2563 spin_lock_irq(&cachep->spinlock);
2564 cachep->batchcount = batchcount;
2565 cachep->limit = limit;
2566 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
2567 spin_unlock_irq(&cachep->spinlock);
2569 for (i = 0; i < NR_CPUS; i++) {
2570 struct array_cache *ccold = new.new[i];
2573 spin_lock_irq(&cachep->spinlock);
2574 free_block(cachep, ac_entry(ccold), ccold->avail);
2575 spin_unlock_irq(&cachep->spinlock);
2578 new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d);
2580 struct array_cache *old;
2582 spin_lock_irq(&cachep->spinlock);
2583 old = cachep->lists.shared;
2584 cachep->lists.shared = new_shared;
2586 free_block(cachep, ac_entry(old), old->avail);
2587 spin_unlock_irq(&cachep->spinlock);
2595 static void enable_cpucache (kmem_cache_t *cachep)
2600 /* The head array serves three purposes:
2601 * - create a LIFO ordering, i.e. return objects that are cache-warm
2602 * - reduce the number of spinlock operations.
2603 * - reduce the number of linked list operations on the slab and
2604 * bufctl chains: array operations are cheaper.
2605 * The numbers are guessed, we should auto-tune as described by
2608 if (cachep->objsize > 131072)
2610 else if (cachep->objsize > PAGE_SIZE)
2612 else if (cachep->objsize > 1024)
2614 else if (cachep->objsize > 256)
2619 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
2620 * allocation behaviour: Most allocs on one cpu, most free operations
2621 * on another cpu. For these cases, an efficient object passing between
2622 * cpus is necessary. This is provided by a shared array. The array
2623 * replaces Bonwick's magazine layer.
2624 * On uniprocessor, it's functionally equivalent (but less efficient)
2625 * to a larger limit. Thus disabled by default.
2629 if (cachep->objsize <= PAGE_SIZE)
2634 /* With debugging enabled, large batchcount lead to excessively
2635 * long periods with disabled local interrupts. Limit the
2641 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
2643 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
2644 cachep->name, -err);
2647 static void drain_array(kmem_cache_t *cachep, struct array_cache *ac)
2654 } else if (ac->avail) {
2655 tofree = (ac->limit+4)/5;
2656 if (tofree > ac->avail) {
2657 tofree = (ac->avail+1)/2;
2659 spin_lock(&cachep->spinlock);
2660 free_block(cachep, ac_entry(ac), tofree);
2661 spin_unlock(&cachep->spinlock);
2662 ac->avail -= tofree;
2663 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2664 sizeof(void*)*ac->avail);
2668 static void drain_array_locked(kmem_cache_t *cachep,
2669 struct array_cache *ac, int force)
2673 check_spinlock_acquired(cachep);
2674 if (ac->touched && !force) {
2676 } else if (ac->avail) {
2677 tofree = force ? ac->avail : (ac->limit+4)/5;
2678 if (tofree > ac->avail) {
2679 tofree = (ac->avail+1)/2;
2681 free_block(cachep, ac_entry(ac), tofree);
2682 ac->avail -= tofree;
2683 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2684 sizeof(void*)*ac->avail);
2689 * cache_reap - Reclaim memory from caches.
2691 * Called from a timer, every few seconds
2693 * - clear the per-cpu caches for this CPU.
2694 * - return freeable pages to the main free memory pool.
2696 * If we cannot acquire the cache chain semaphore then just give up - we'll
2697 * try again next timer interrupt.
2699 static inline void cache_reap (void)
2701 struct list_head *walk;
2704 BUG_ON(!in_interrupt());
2707 if (down_trylock(&cache_chain_sem))
2710 list_for_each(walk, &cache_chain) {
2711 kmem_cache_t *searchp;
2712 struct list_head* p;
2716 searchp = list_entry(walk, kmem_cache_t, next);
2718 if (searchp->flags & SLAB_NO_REAP)
2722 local_irq_disable();
2723 drain_array(searchp, ac_data(searchp));
2725 if(time_after(searchp->lists.next_reap, jiffies))
2728 spin_lock(&searchp->spinlock);
2729 if(time_after(searchp->lists.next_reap, jiffies)) {
2732 searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
2734 if (searchp->lists.shared)
2735 drain_array_locked(searchp, searchp->lists.shared, 0);
2737 if (searchp->lists.free_touched) {
2738 searchp->lists.free_touched = 0;
2742 tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
2744 p = list3_data(searchp)->slabs_free.next;
2745 if (p == &(list3_data(searchp)->slabs_free))
2748 slabp = list_entry(p, struct slab, list);
2749 BUG_ON(slabp->inuse);
2750 list_del(&slabp->list);
2751 STATS_INC_REAPED(searchp);
2753 /* Safe to drop the lock. The slab is no longer
2754 * linked to the cache.
2755 * searchp cannot disappear, we hold
2758 searchp->lists.free_objects -= searchp->num;
2759 spin_unlock_irq(&searchp->spinlock);
2760 slab_destroy(searchp, slabp);
2761 spin_lock_irq(&searchp->spinlock);
2762 } while(--tofree > 0);
2764 spin_unlock(&searchp->spinlock);
2771 up(&cache_chain_sem);
2775 * This is a timer handler. There is one per CPU. It is called periodially
2776 * to shrink this CPU's caches. Otherwise there could be memory tied up
2777 * for long periods (or for ever) due to load changes.
2779 static void reap_timer_fnc(unsigned long cpu)
2781 struct timer_list *rt = &__get_cpu_var(reap_timers);
2783 /* CPU hotplug can drag us off cpu: don't run on wrong CPU */
2784 if (!cpu_is_offline(cpu)) {
2786 mod_timer(rt, jiffies + REAPTIMEOUT_CPUC + cpu);
2790 #ifdef CONFIG_PROC_FS
2792 static void *s_start(struct seq_file *m, loff_t *pos)
2795 struct list_head *p;
2797 down(&cache_chain_sem);
2800 * Output format version, so at least we can change it
2801 * without _too_ many complaints.
2804 seq_puts(m, "slabinfo - version: 2.0 (statistics)\n");
2806 seq_puts(m, "slabinfo - version: 2.0\n");
2808 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
2809 seq_puts(m, " : tunables <batchcount> <limit> <sharedfactor>");
2810 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
2812 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <freelimit>");
2813 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
2817 p = cache_chain.next;
2820 if (p == &cache_chain)
2823 return list_entry(p, kmem_cache_t, next);
2826 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2828 kmem_cache_t *cachep = p;
2830 return cachep->next.next == &cache_chain ? NULL
2831 : list_entry(cachep->next.next, kmem_cache_t, next);
2834 static void s_stop(struct seq_file *m, void *p)
2836 up(&cache_chain_sem);
2839 static int s_show(struct seq_file *m, void *p)
2841 kmem_cache_t *cachep = p;
2842 struct list_head *q;
2844 unsigned long active_objs;
2845 unsigned long num_objs;
2846 unsigned long active_slabs = 0;
2847 unsigned long num_slabs;
2850 mm_segment_t old_fs;
2854 spin_lock_irq(&cachep->spinlock);
2857 list_for_each(q,&cachep->lists.slabs_full) {
2858 slabp = list_entry(q, struct slab, list);
2859 if (slabp->inuse != cachep->num && !error)
2860 error = "slabs_full accounting error";
2861 active_objs += cachep->num;
2864 list_for_each(q,&cachep->lists.slabs_partial) {
2865 slabp = list_entry(q, struct slab, list);
2866 if (slabp->inuse == cachep->num && !error)
2867 error = "slabs_partial inuse accounting error";
2868 if (!slabp->inuse && !error)
2869 error = "slabs_partial/inuse accounting error";
2870 active_objs += slabp->inuse;
2873 list_for_each(q,&cachep->lists.slabs_free) {
2874 slabp = list_entry(q, struct slab, list);
2875 if (slabp->inuse && !error)
2876 error = "slabs_free/inuse accounting error";
2879 num_slabs+=active_slabs;
2880 num_objs = num_slabs*cachep->num;
2881 if (num_objs - active_objs != cachep->lists.free_objects && !error)
2882 error = "free_objects accounting error";
2884 name = cachep->name;
2887 * Check to see if `name' resides inside a module which has been
2888 * unloaded (someone forgot to destroy their cache)
2892 if (__get_user(tmp, name))
2897 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
2899 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
2900 name, active_objs, num_objs, cachep->objsize,
2901 cachep->num, (1<<cachep->gfporder));
2902 seq_printf(m, " : tunables %4u %4u %4u",
2903 cachep->limit, cachep->batchcount,
2904 cachep->lists.shared->limit/cachep->batchcount);
2905 seq_printf(m, " : slabdata %6lu %6lu %6u",
2906 active_slabs, num_slabs, cachep->lists.shared->avail);
2909 unsigned long high = cachep->high_mark;
2910 unsigned long allocs = cachep->num_allocations;
2911 unsigned long grown = cachep->grown;
2912 unsigned long reaped = cachep->reaped;
2913 unsigned long errors = cachep->errors;
2914 unsigned long max_freeable = cachep->max_freeable;
2915 unsigned long free_limit = cachep->free_limit;
2917 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu",
2918 allocs, high, grown, reaped, errors,
2919 max_freeable, free_limit);
2923 unsigned long allochit = atomic_read(&cachep->allochit);
2924 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
2925 unsigned long freehit = atomic_read(&cachep->freehit);
2926 unsigned long freemiss = atomic_read(&cachep->freemiss);
2928 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
2929 allochit, allocmiss, freehit, freemiss);
2933 spin_unlock_irq(&cachep->spinlock);
2938 * slabinfo_op - iterator that generates /proc/slabinfo
2947 * num-pages-per-slab
2948 * + further values on SMP and with statistics enabled
2951 struct seq_operations slabinfo_op = {
2958 #define MAX_SLABINFO_WRITE 128
2960 * slabinfo_write - Tuning for the slab allocator
2962 * @buffer: user buffer
2963 * @count: data length
2966 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
2967 size_t count, loff_t *ppos)
2969 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
2970 int limit, batchcount, shared, res;
2971 struct list_head *p;
2973 if (count > MAX_SLABINFO_WRITE)
2975 if (copy_from_user(&kbuf, buffer, count))
2977 kbuf[MAX_SLABINFO_WRITE] = '\0';
2979 tmp = strchr(kbuf, ' ');
2984 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
2987 /* Find the cache in the chain of caches. */
2988 down(&cache_chain_sem);
2990 list_for_each(p,&cache_chain) {
2991 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
2993 if (!strcmp(cachep->name, kbuf)) {
2996 batchcount > limit ||
3000 res = do_tune_cpucache(cachep, limit, batchcount, shared);
3005 up(&cache_chain_sem);
3012 unsigned int ksize(const void *objp)
3015 unsigned long flags;
3016 unsigned int size = 0;
3018 if (likely(objp != NULL)) {
3019 local_irq_save(flags);
3020 c = GET_PAGE_CACHE(virt_to_page(objp));
3021 size = kmem_cache_size(c);
3022 local_irq_restore(flags);
3028 void ptrinfo(unsigned long addr)
3032 printk("Dumping data about address %p.\n", (void*)addr);
3033 if (!virt_addr_valid((void*)addr)) {
3034 printk("virt addr invalid.\n");
3039 pgd_t *pgd = pgd_offset_k(addr);
3041 if (pgd_none(*pgd)) {
3042 printk("No pgd.\n");
3045 pmd = pmd_offset(pgd, addr);
3046 if (pmd_none(*pmd)) {
3047 printk("No pmd.\n");
3051 if (pmd_large(*pmd)) {
3052 printk("Large page.\n");
3056 printk("normal page, pte_val 0x%llx\n",
3057 (unsigned long long)pte_val(*pte_offset_kernel(pmd, addr)));
3061 page = virt_to_page((void*)addr);
3062 printk("struct page at %p, flags %08lx\n",
3063 page, (unsigned long)page->flags);
3064 if (PageSlab(page)) {
3067 unsigned long flags;
3071 c = GET_PAGE_CACHE(page);
3072 printk("belongs to cache %s.\n",c->name);
3074 spin_lock_irqsave(&c->spinlock, flags);
3075 s = GET_PAGE_SLAB(page);
3076 printk("slabp %p with %d inuse objects (from %d).\n",
3077 s, s->inuse, c->num);
3080 objnr = (addr-(unsigned long)s->s_mem)/c->objsize;
3081 objp = s->s_mem+c->objsize*objnr;
3082 printk("points into object no %d, starting at %p, len %d.\n",
3083 objnr, objp, c->objsize);
3084 if (objnr >= c->num) {
3085 printk("Bad obj number.\n");
3087 kernel_map_pages(virt_to_page(objp),
3088 c->objsize/PAGE_SIZE, 1);
3090 print_objinfo(c, objp, 2);
3092 spin_unlock_irqrestore(&c->spinlock, flags);