1 /*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
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
21 *
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.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
31 *
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.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
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.
46 *
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.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
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
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
87 */
89 #include <linux/slab.h>
90 #include <linux/mm.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <asm/cacheflush.h>
121 #include <asm/tlbflush.h>
122 #include <asm/page.h>
124 /*
125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
126 * 0 for faster, smaller code (especially in the critical paths).
127 *
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
130 *
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 */
134 #ifdef CONFIG_DEBUG_SLAB
135 #define DEBUG 1
136 #define STATS 1
137 #define FORCED_DEBUG 1
138 #else
139 #define DEBUG 0
140 #define STATS 0
141 #define FORCED_DEBUG 0
142 #endif
144 /* Shouldn't this be in a header file somewhere? */
145 #define BYTES_PER_WORD sizeof(void *)
146 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
148 #ifndef ARCH_KMALLOC_FLAGS
149 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150 #endif
152 /* Legal flag mask for kmem_cache_create(). */
153 #if DEBUG
154 # define CREATE_MASK (SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
156 SLAB_CACHE_DMA | \
157 SLAB_STORE_USER | \
158 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
159 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
160 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161 #else
162 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163 SLAB_CACHE_DMA | \
164 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
166 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
167 #endif
169 /*
170 * kmem_bufctl_t:
171 *
172 * Bufctl's are used for linking objs within a slab
173 * linked offsets.
174 *
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
186 */
188 typedef unsigned int kmem_bufctl_t;
189 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
191 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
194 /*
195 * struct slab_rcu
196 *
197 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
198 * arrange for kmem_freepages to be called via RCU. This is useful if
199 * we need to approach a kernel structure obliquely, from its address
200 * obtained without the usual locking. We can lock the structure to
201 * stabilize it and check it's still at the given address, only if we
202 * can be sure that the memory has not been meanwhile reused for some
203 * other kind of object (which our subsystem's lock might corrupt).
204 *
205 * rcu_read_lock before reading the address, then rcu_read_unlock after
206 * taking the spinlock within the structure expected at that address.
207 */
208 struct slab_rcu {
209 struct rcu_head head;
210 struct kmem_cache *cachep;
211 void *addr;
212 };
214 /*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221 struct slab {
222 union {
223 struct {
224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
228 kmem_bufctl_t free;
229 unsigned short nodeid;
230 };
231 struct slab_rcu __slab_cover_slab_rcu;
232 };
233 };
235 /*
236 * struct array_cache
237 *
238 * Purpose:
239 * - LIFO ordering, to hand out cache-warm objects from _alloc
240 * - reduce the number of linked list operations
241 * - reduce spinlock operations
242 *
243 * The limit is stored in the per-cpu structure to reduce the data cache
244 * footprint.
245 *
246 */
247 struct array_cache {
248 unsigned int avail;
249 unsigned int limit;
250 unsigned int batchcount;
251 unsigned int touched;
252 spinlock_t lock;
253 void *entry[]; /*
254 * Must have this definition in here for the proper
255 * alignment of array_cache. Also simplifies accessing
256 * the entries.
257 */
258 };
260 /*
261 * bootstrap: The caches do not work without cpuarrays anymore, but the
262 * cpuarrays are allocated from the generic caches...
263 */
264 #define BOOT_CPUCACHE_ENTRIES 1
265 struct arraycache_init {
266 struct array_cache cache;
267 void *entries[BOOT_CPUCACHE_ENTRIES];
268 };
270 /*
271 * The slab lists for all objects.
272 */
273 struct kmem_list3 {
274 struct list_head slabs_partial; /* partial list first, better asm code */
275 struct list_head slabs_full;
276 struct list_head slabs_free;
277 unsigned long free_objects;
278 unsigned int free_limit;
279 unsigned int colour_next; /* Per-node cache coloring */
280 spinlock_t list_lock;
281 struct array_cache *shared; /* shared per node */
282 struct array_cache **alien; /* on other nodes */
283 unsigned long next_reap; /* updated without locking */
284 int free_touched; /* updated without locking */
285 };
287 /*
288 * Need this for bootstrapping a per node allocator.
289 */
290 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
291 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
292 #define CACHE_CACHE 0
293 #define SIZE_AC MAX_NUMNODES
294 #define SIZE_L3 (2 * MAX_NUMNODES)
296 static int drain_freelist(struct kmem_cache *cache,
297 struct kmem_list3 *l3, int tofree);
298 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
299 int node);
300 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
301 static void cache_reap(struct work_struct *unused);
303 /*
304 * This function must be completely optimized away if a constant is passed to
305 * it. Mostly the same as what is in linux/slab.h except it returns an index.
306 */
307 static __always_inline int index_of(const size_t size)
308 {
309 extern void __bad_size(void);
311 if (__builtin_constant_p(size)) {
312 int i = 0;
314 #define CACHE(x) \
315 if (size <=x) \
316 return i; \
317 else \
318 i++;
319 #include <linux/kmalloc_sizes.h>
320 #undef CACHE
321 __bad_size();
322 } else
323 __bad_size();
324 return 0;
325 }
327 static int slab_early_init = 1;
329 #define INDEX_AC index_of(sizeof(struct arraycache_init))
330 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
332 static void kmem_list3_init(struct kmem_list3 *parent)
333 {
334 INIT_LIST_HEAD(&parent->slabs_full);
335 INIT_LIST_HEAD(&parent->slabs_partial);
336 INIT_LIST_HEAD(&parent->slabs_free);
337 parent->shared = NULL;
338 parent->alien = NULL;
339 parent->colour_next = 0;
340 spin_lock_init(&parent->list_lock);
341 parent->free_objects = 0;
342 parent->free_touched = 0;
343 }
345 #define MAKE_LIST(cachep, listp, slab, nodeid) \
346 do { \
347 INIT_LIST_HEAD(listp); \
348 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
349 } while (0)
351 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
352 do { \
353 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
354 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
355 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
356 } while (0)
358 #define CFLGS_OFF_SLAB (0x80000000UL)
359 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
361 #define BATCHREFILL_LIMIT 16
362 /*
363 * Optimization question: fewer reaps means less probability for unnessary
364 * cpucache drain/refill cycles.
365 *
366 * OTOH the cpuarrays can contain lots of objects,
367 * which could lock up otherwise freeable slabs.
368 */
369 #define REAPTIMEOUT_CPUC (2*HZ)
370 #define REAPTIMEOUT_LIST3 (4*HZ)
372 #if STATS
373 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
374 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
375 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
376 #define STATS_INC_GROWN(x) ((x)->grown++)
377 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
378 #define STATS_SET_HIGH(x) \
379 do { \
380 if ((x)->num_active > (x)->high_mark) \
381 (x)->high_mark = (x)->num_active; \
382 } while (0)
383 #define STATS_INC_ERR(x) ((x)->errors++)
384 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
385 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
386 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
387 #define STATS_SET_FREEABLE(x, i) \
388 do { \
389 if ((x)->max_freeable < i) \
390 (x)->max_freeable = i; \
391 } while (0)
392 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
393 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
394 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
395 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
396 #else
397 #define STATS_INC_ACTIVE(x) do { } while (0)
398 #define STATS_DEC_ACTIVE(x) do { } while (0)
399 #define STATS_INC_ALLOCED(x) do { } while (0)
400 #define STATS_INC_GROWN(x) do { } while (0)
401 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
402 #define STATS_SET_HIGH(x) do { } while (0)
403 #define STATS_INC_ERR(x) do { } while (0)
404 #define STATS_INC_NODEALLOCS(x) do { } while (0)
405 #define STATS_INC_NODEFREES(x) do { } while (0)
406 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
407 #define STATS_SET_FREEABLE(x, i) do { } while (0)
408 #define STATS_INC_ALLOCHIT(x) do { } while (0)
409 #define STATS_INC_ALLOCMISS(x) do { } while (0)
410 #define STATS_INC_FREEHIT(x) do { } while (0)
411 #define STATS_INC_FREEMISS(x) do { } while (0)
412 #endif
414 #if DEBUG
416 /*
417 * memory layout of objects:
418 * 0 : objp
419 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
420 * the end of an object is aligned with the end of the real
421 * allocation. Catches writes behind the end of the allocation.
422 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
423 * redzone word.
424 * cachep->obj_offset: The real object.
425 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
426 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
427 * [BYTES_PER_WORD long]
428 */
429 static int obj_offset(struct kmem_cache *cachep)
430 {
431 return cachep->obj_offset;
432 }
434 static int obj_size(struct kmem_cache *cachep)
435 {
436 return cachep->obj_size;
437 }
439 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
440 {
441 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
442 return (unsigned long long*) (objp + obj_offset(cachep) -
443 sizeof(unsigned long long));
444 }
446 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
447 {
448 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449 if (cachep->flags & SLAB_STORE_USER)
450 return (unsigned long long *)(objp + cachep->buffer_size -
451 sizeof(unsigned long long) -
452 REDZONE_ALIGN);
453 return (unsigned long long *) (objp + cachep->buffer_size -
454 sizeof(unsigned long long));
455 }
457 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
458 {
459 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
460 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
461 }
463 #else
465 #define obj_offset(x) 0
466 #define obj_size(cachep) (cachep->buffer_size)
467 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
469 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
471 #endif
473 #ifdef CONFIG_TRACING
474 size_t slab_buffer_size(struct kmem_cache *cachep)
475 {
476 return cachep->buffer_size;
477 }
478 EXPORT_SYMBOL(slab_buffer_size);
479 #endif
481 /*
482 * Do not go above this order unless 0 objects fit into the slab.
483 */
484 #define BREAK_GFP_ORDER_HI 1
485 #define BREAK_GFP_ORDER_LO 0
486 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
488 /*
489 * Functions for storing/retrieving the cachep and or slab from the page
490 * allocator. These are used to find the slab an obj belongs to. With kfree(),
491 * these are used to find the cache which an obj belongs to.
492 */
493 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
494 {
495 page->lru.next = (struct list_head *)cache;
496 }
498 static inline struct kmem_cache *page_get_cache(struct page *page)
499 {
500 page = compound_head(page);
501 BUG_ON(!PageSlab(page));
502 return (struct kmem_cache *)page->lru.next;
503 }
505 static inline void page_set_slab(struct page *page, struct slab *slab)
506 {
507 page->lru.prev = (struct list_head *)slab;
508 }
510 static inline struct slab *page_get_slab(struct page *page)
511 {
512 BUG_ON(!PageSlab(page));
513 return (struct slab *)page->lru.prev;
514 }
516 static inline struct kmem_cache *virt_to_cache(const void *obj)
517 {
518 struct page *page = virt_to_head_page(obj);
519 return page_get_cache(page);
520 }
522 static inline struct slab *virt_to_slab(const void *obj)
523 {
524 struct page *page = virt_to_head_page(obj);
525 return page_get_slab(page);
526 }
528 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
529 unsigned int idx)
530 {
531 return slab->s_mem + cache->buffer_size * idx;
532 }
534 /*
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
539 */
540 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541 const struct slab *slab, void *obj)
542 {
543 u32 offset = (obj - slab->s_mem);
544 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
545 }
547 /*
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
549 */
550 struct cache_sizes malloc_sizes[] = {
551 #define CACHE(x) { .cs_size = (x) },
552 #include <linux/kmalloc_sizes.h>
553 CACHE(ULONG_MAX)
554 #undef CACHE
555 };
556 EXPORT_SYMBOL(malloc_sizes);
558 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
559 struct cache_names {
560 char *name;
561 char *name_dma;
562 };
564 static struct cache_names __initdata cache_names[] = {
565 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566 #include <linux/kmalloc_sizes.h>
567 {NULL,}
568 #undef CACHE
569 };
571 static struct arraycache_init initarray_cache __initdata =
572 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
573 static struct arraycache_init initarray_generic =
574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
576 /* internal cache of cache description objs */
577 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
578 static struct kmem_cache cache_cache = {
579 .nodelists = cache_cache_nodelists,
580 .batchcount = 1,
581 .limit = BOOT_CPUCACHE_ENTRIES,
582 .shared = 1,
583 .buffer_size = sizeof(struct kmem_cache),
584 .name = "kmem_cache",
585 };
587 #define BAD_ALIEN_MAGIC 0x01020304ul
589 /*
590 * chicken and egg problem: delay the per-cpu array allocation
591 * until the general caches are up.
592 */
593 static enum {
594 NONE,
595 PARTIAL_AC,
596 PARTIAL_L3,
597 EARLY,
598 LATE,
599 FULL
600 } g_cpucache_up;
602 /*
603 * used by boot code to determine if it can use slab based allocator
604 */
605 int slab_is_available(void)
606 {
607 return g_cpucache_up >= EARLY;
608 }
610 #ifdef CONFIG_LOCKDEP
612 /*
613 * Slab sometimes uses the kmalloc slabs to store the slab headers
614 * for other slabs "off slab".
615 * The locking for this is tricky in that it nests within the locks
616 * of all other slabs in a few places; to deal with this special
617 * locking we put on-slab caches into a separate lock-class.
618 *
619 * We set lock class for alien array caches which are up during init.
620 * The lock annotation will be lost if all cpus of a node goes down and
621 * then comes back up during hotplug
622 */
623 static struct lock_class_key on_slab_l3_key;
624 static struct lock_class_key on_slab_alc_key;
626 static struct lock_class_key debugobj_l3_key;
627 static struct lock_class_key debugobj_alc_key;
629 static void slab_set_lock_classes(struct kmem_cache *cachep,
630 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
631 int q)
632 {
633 struct array_cache **alc;
634 struct kmem_list3 *l3;
635 int r;
637 l3 = cachep->nodelists[q];
638 if (!l3)
639 return;
641 lockdep_set_class(&l3->list_lock, l3_key);
642 alc = l3->alien;
643 /*
644 * FIXME: This check for BAD_ALIEN_MAGIC
645 * should go away when common slab code is taught to
646 * work even without alien caches.
647 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
648 * for alloc_alien_cache,
649 */
650 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
651 return;
652 for_each_node(r) {
653 if (alc[r])
654 lockdep_set_class(&alc[r]->lock, alc_key);
655 }
656 }
658 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
659 {
660 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
661 }
663 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
664 {
665 int node;
667 for_each_online_node(node)
668 slab_set_debugobj_lock_classes_node(cachep, node);
669 }
671 static void init_node_lock_keys(int q)
672 {
673 struct cache_sizes *s = malloc_sizes;
675 if (g_cpucache_up < LATE)
676 return;
678 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
679 struct kmem_list3 *l3;
681 l3 = s->cs_cachep->nodelists[q];
682 if (!l3 || OFF_SLAB(s->cs_cachep))
683 continue;
685 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
686 &on_slab_alc_key, q);
687 }
688 }
690 static inline void init_lock_keys(void)
691 {
692 int node;
694 for_each_node(node)
695 init_node_lock_keys(node);
696 }
697 #else
698 static void init_node_lock_keys(int q)
699 {
700 }
702 static inline void init_lock_keys(void)
703 {
704 }
706 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
707 {
708 }
710 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
711 {
712 }
713 #endif
715 /*
716 * Guard access to the cache-chain.
717 */
718 static DEFINE_MUTEX(cache_chain_mutex);
719 static struct list_head cache_chain;
721 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
723 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
724 {
725 return cachep->array[smp_processor_id()];
726 }
728 static inline struct kmem_cache *__find_general_cachep(size_t size,
729 gfp_t gfpflags)
730 {
731 struct cache_sizes *csizep = malloc_sizes;
733 #if DEBUG
734 /* This happens if someone tries to call
735 * kmem_cache_create(), or __kmalloc(), before
736 * the generic caches are initialized.
737 */
738 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
739 #endif
740 if (!size)
741 return ZERO_SIZE_PTR;
743 while (size > csizep->cs_size)
744 csizep++;
746 /*
747 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
748 * has cs_{dma,}cachep==NULL. Thus no special case
749 * for large kmalloc calls required.
750 */
751 #ifdef CONFIG_ZONE_DMA
752 if (unlikely(gfpflags & GFP_DMA))
753 return csizep->cs_dmacachep;
754 #endif
755 return csizep->cs_cachep;
756 }
758 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
759 {
760 return __find_general_cachep(size, gfpflags);
761 }
763 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
764 {
765 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
766 }
768 /*
769 * Calculate the number of objects and left-over bytes for a given buffer size.
770 */
771 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
772 size_t align, int flags, size_t *left_over,
773 unsigned int *num)
774 {
775 int nr_objs;
776 size_t mgmt_size;
777 size_t slab_size = PAGE_SIZE << gfporder;
779 /*
780 * The slab management structure can be either off the slab or
781 * on it. For the latter case, the memory allocated for a
782 * slab is used for:
783 *
784 * - The struct slab
785 * - One kmem_bufctl_t for each object
786 * - Padding to respect alignment of @align
787 * - @buffer_size bytes for each object
788 *
789 * If the slab management structure is off the slab, then the
790 * alignment will already be calculated into the size. Because
791 * the slabs are all pages aligned, the objects will be at the
792 * correct alignment when allocated.
793 */
794 if (flags & CFLGS_OFF_SLAB) {
795 mgmt_size = 0;
796 nr_objs = slab_size / buffer_size;
798 if (nr_objs > SLAB_LIMIT)
799 nr_objs = SLAB_LIMIT;
800 } else {
801 /*
802 * Ignore padding for the initial guess. The padding
803 * is at most @align-1 bytes, and @buffer_size is at
804 * least @align. In the worst case, this result will
805 * be one greater than the number of objects that fit
806 * into the memory allocation when taking the padding
807 * into account.
808 */
809 nr_objs = (slab_size - sizeof(struct slab)) /
810 (buffer_size + sizeof(kmem_bufctl_t));
812 /*
813 * This calculated number will be either the right
814 * amount, or one greater than what we want.
815 */
816 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
817 > slab_size)
818 nr_objs--;
820 if (nr_objs > SLAB_LIMIT)
821 nr_objs = SLAB_LIMIT;
823 mgmt_size = slab_mgmt_size(nr_objs, align);
824 }
825 *num = nr_objs;
826 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
827 }
829 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
831 static void __slab_error(const char *function, struct kmem_cache *cachep,
832 char *msg)
833 {
834 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
835 function, cachep->name, msg);
836 dump_stack();
837 }
839 /*
840 * By default on NUMA we use alien caches to stage the freeing of
841 * objects allocated from other nodes. This causes massive memory
842 * inefficiencies when using fake NUMA setup to split memory into a
843 * large number of small nodes, so it can be disabled on the command
844 * line
845 */
847 static int use_alien_caches __read_mostly = 1;
848 static int __init noaliencache_setup(char *s)
849 {
850 use_alien_caches = 0;
851 return 1;
852 }
853 __setup("noaliencache", noaliencache_setup);
855 #ifdef CONFIG_NUMA
856 /*
857 * Special reaping functions for NUMA systems called from cache_reap().
858 * These take care of doing round robin flushing of alien caches (containing
859 * objects freed on different nodes from which they were allocated) and the
860 * flushing of remote pcps by calling drain_node_pages.
861 */
862 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
864 static void init_reap_node(int cpu)
865 {
866 int node;
868 node = next_node(cpu_to_mem(cpu), node_online_map);
869 if (node == MAX_NUMNODES)
870 node = first_node(node_online_map);
872 per_cpu(slab_reap_node, cpu) = node;
873 }
875 static void next_reap_node(void)
876 {
877 int node = __this_cpu_read(slab_reap_node);
879 node = next_node(node, node_online_map);
880 if (unlikely(node >= MAX_NUMNODES))
881 node = first_node(node_online_map);
882 __this_cpu_write(slab_reap_node, node);
883 }
885 #else
886 #define init_reap_node(cpu) do { } while (0)
887 #define next_reap_node(void) do { } while (0)
888 #endif
890 /*
891 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
892 * via the workqueue/eventd.
893 * Add the CPU number into the expiration time to minimize the possibility of
894 * the CPUs getting into lockstep and contending for the global cache chain
895 * lock.
896 */
897 static void __cpuinit start_cpu_timer(int cpu)
898 {
899 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
901 /*
902 * When this gets called from do_initcalls via cpucache_init(),
903 * init_workqueues() has already run, so keventd will be setup
904 * at that time.
905 */
906 if (keventd_up() && reap_work->work.func == NULL) {
907 init_reap_node(cpu);
908 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
909 schedule_delayed_work_on(cpu, reap_work,
910 __round_jiffies_relative(HZ, cpu));
911 }
912 }
914 static struct array_cache *alloc_arraycache(int node, int entries,
915 int batchcount, gfp_t gfp)
916 {
917 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
918 struct array_cache *nc = NULL;
920 nc = kmalloc_node(memsize, gfp, node);
921 /*
922 * The array_cache structures contain pointers to free object.
923 * However, when such objects are allocated or transferred to another
924 * cache the pointers are not cleared and they could be counted as
925 * valid references during a kmemleak scan. Therefore, kmemleak must
926 * not scan such objects.
927 */
928 kmemleak_no_scan(nc);
929 if (nc) {
930 nc->avail = 0;
931 nc->limit = entries;
932 nc->batchcount = batchcount;
933 nc->touched = 0;
934 spin_lock_init(&nc->lock);
935 }
936 return nc;
937 }
939 /*
940 * Transfer objects in one arraycache to another.
941 * Locking must be handled by the caller.
942 *
943 * Return the number of entries transferred.
944 */
945 static int transfer_objects(struct array_cache *to,
946 struct array_cache *from, unsigned int max)
947 {
948 /* Figure out how many entries to transfer */
949 int nr = min3(from->avail, max, to->limit - to->avail);
951 if (!nr)
952 return 0;
954 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
955 sizeof(void *) *nr);
957 from->avail -= nr;
958 to->avail += nr;
959 return nr;
960 }
962 #ifndef CONFIG_NUMA
964 #define drain_alien_cache(cachep, alien) do { } while (0)
965 #define reap_alien(cachep, l3) do { } while (0)
967 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
968 {
969 return (struct array_cache **)BAD_ALIEN_MAGIC;
970 }
972 static inline void free_alien_cache(struct array_cache **ac_ptr)
973 {
974 }
976 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
977 {
978 return 0;
979 }
981 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
982 gfp_t flags)
983 {
984 return NULL;
985 }
987 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
988 gfp_t flags, int nodeid)
989 {
990 return NULL;
991 }
993 #else /* CONFIG_NUMA */
995 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
996 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
998 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
999 {
1000 struct array_cache **ac_ptr;
1001 int memsize = sizeof(void *) * nr_node_ids;
1002 int i;
1004 if (limit > 1)
1005 limit = 12;
1006 ac_ptr = kzalloc_node(memsize, gfp, node);
1007 if (ac_ptr) {
1008 for_each_node(i) {
1009 if (i == node || !node_online(i))
1010 continue;
1011 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1012 if (!ac_ptr[i]) {
1013 for (i--; i >= 0; i--)
1014 kfree(ac_ptr[i]);
1015 kfree(ac_ptr);
1016 return NULL;
1017 }
1018 }
1019 }
1020 return ac_ptr;
1021 }
1023 static void free_alien_cache(struct array_cache **ac_ptr)
1024 {
1025 int i;
1027 if (!ac_ptr)
1028 return;
1029 for_each_node(i)
1030 kfree(ac_ptr[i]);
1031 kfree(ac_ptr);
1032 }
1034 static void __drain_alien_cache(struct kmem_cache *cachep,
1035 struct array_cache *ac, int node)
1036 {
1037 struct kmem_list3 *rl3 = cachep->nodelists[node];
1039 if (ac->avail) {
1040 spin_lock(&rl3->list_lock);
1041 /*
1042 * Stuff objects into the remote nodes shared array first.
1043 * That way we could avoid the overhead of putting the objects
1044 * into the free lists and getting them back later.
1045 */
1046 if (rl3->shared)
1047 transfer_objects(rl3->shared, ac, ac->limit);
1049 free_block(cachep, ac->entry, ac->avail, node);
1050 ac->avail = 0;
1051 spin_unlock(&rl3->list_lock);
1052 }
1053 }
1055 /*
1056 * Called from cache_reap() to regularly drain alien caches round robin.
1057 */
1058 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1059 {
1060 int node = __this_cpu_read(slab_reap_node);
1062 if (l3->alien) {
1063 struct array_cache *ac = l3->alien[node];
1065 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1066 __drain_alien_cache(cachep, ac, node);
1067 spin_unlock_irq(&ac->lock);
1068 }
1069 }
1070 }
1072 static void drain_alien_cache(struct kmem_cache *cachep,
1073 struct array_cache **alien)
1074 {
1075 int i = 0;
1076 struct array_cache *ac;
1077 unsigned long flags;
1079 for_each_online_node(i) {
1080 ac = alien[i];
1081 if (ac) {
1082 spin_lock_irqsave(&ac->lock, flags);
1083 __drain_alien_cache(cachep, ac, i);
1084 spin_unlock_irqrestore(&ac->lock, flags);
1085 }
1086 }
1087 }
1089 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1090 {
1091 struct slab *slabp = virt_to_slab(objp);
1092 int nodeid = slabp->nodeid;
1093 struct kmem_list3 *l3;
1094 struct array_cache *alien = NULL;
1095 int node;
1097 node = numa_mem_id();
1099 /*
1100 * Make sure we are not freeing a object from another node to the array
1101 * cache on this cpu.
1102 */
1103 if (likely(slabp->nodeid == node))
1104 return 0;
1106 l3 = cachep->nodelists[node];
1107 STATS_INC_NODEFREES(cachep);
1108 if (l3->alien && l3->alien[nodeid]) {
1109 alien = l3->alien[nodeid];
1110 spin_lock(&alien->lock);
1111 if (unlikely(alien->avail == alien->limit)) {
1112 STATS_INC_ACOVERFLOW(cachep);
1113 __drain_alien_cache(cachep, alien, nodeid);
1114 }
1115 alien->entry[alien->avail++] = objp;
1116 spin_unlock(&alien->lock);
1117 } else {
1118 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1119 free_block(cachep, &objp, 1, nodeid);
1120 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1121 }
1122 return 1;
1123 }
1124 #endif
1126 /*
1127 * Allocates and initializes nodelists for a node on each slab cache, used for
1128 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1129 * will be allocated off-node since memory is not yet online for the new node.
1130 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1131 * already in use.
1132 *
1133 * Must hold cache_chain_mutex.
1134 */
1135 static int init_cache_nodelists_node(int node)
1136 {
1137 struct kmem_cache *cachep;
1138 struct kmem_list3 *l3;
1139 const int memsize = sizeof(struct kmem_list3);
1141 list_for_each_entry(cachep, &cache_chain, next) {
1142 /*
1143 * Set up the size64 kmemlist for cpu before we can
1144 * begin anything. Make sure some other cpu on this
1145 * node has not already allocated this
1146 */
1147 if (!cachep->nodelists[node]) {
1148 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1149 if (!l3)
1150 return -ENOMEM;
1151 kmem_list3_init(l3);
1152 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1153 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1155 /*
1156 * The l3s don't come and go as CPUs come and
1157 * go. cache_chain_mutex is sufficient
1158 * protection here.
1159 */
1160 cachep->nodelists[node] = l3;
1161 }
1163 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1164 cachep->nodelists[node]->free_limit =
1165 (1 + nr_cpus_node(node)) *
1166 cachep->batchcount + cachep->num;
1167 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1168 }
1169 return 0;
1170 }
1172 static void __cpuinit cpuup_canceled(long cpu)
1173 {
1174 struct kmem_cache *cachep;
1175 struct kmem_list3 *l3 = NULL;
1176 int node = cpu_to_mem(cpu);
1177 const struct cpumask *mask = cpumask_of_node(node);
1179 list_for_each_entry(cachep, &cache_chain, next) {
1180 struct array_cache *nc;
1181 struct array_cache *shared;
1182 struct array_cache **alien;
1184 /* cpu is dead; no one can alloc from it. */
1185 nc = cachep->array[cpu];
1186 cachep->array[cpu] = NULL;
1187 l3 = cachep->nodelists[node];
1189 if (!l3)
1190 goto free_array_cache;
1192 spin_lock_irq(&l3->list_lock);
1194 /* Free limit for this kmem_list3 */
1195 l3->free_limit -= cachep->batchcount;
1196 if (nc)
1197 free_block(cachep, nc->entry, nc->avail, node);
1199 if (!cpumask_empty(mask)) {
1200 spin_unlock_irq(&l3->list_lock);
1201 goto free_array_cache;
1202 }
1204 shared = l3->shared;
1205 if (shared) {
1206 free_block(cachep, shared->entry,
1207 shared->avail, node);
1208 l3->shared = NULL;
1209 }
1211 alien = l3->alien;
1212 l3->alien = NULL;
1214 spin_unlock_irq(&l3->list_lock);
1216 kfree(shared);
1217 if (alien) {
1218 drain_alien_cache(cachep, alien);
1219 free_alien_cache(alien);
1220 }
1221 free_array_cache:
1222 kfree(nc);
1223 }
1224 /*
1225 * In the previous loop, all the objects were freed to
1226 * the respective cache's slabs, now we can go ahead and
1227 * shrink each nodelist to its limit.
1228 */
1229 list_for_each_entry(cachep, &cache_chain, next) {
1230 l3 = cachep->nodelists[node];
1231 if (!l3)
1232 continue;
1233 drain_freelist(cachep, l3, l3->free_objects);
1234 }
1235 }
1237 static int __cpuinit cpuup_prepare(long cpu)
1238 {
1239 struct kmem_cache *cachep;
1240 struct kmem_list3 *l3 = NULL;
1241 int node = cpu_to_mem(cpu);
1242 int err;
1244 /*
1245 * We need to do this right in the beginning since
1246 * alloc_arraycache's are going to use this list.
1247 * kmalloc_node allows us to add the slab to the right
1248 * kmem_list3 and not this cpu's kmem_list3
1249 */
1250 err = init_cache_nodelists_node(node);
1251 if (err < 0)
1252 goto bad;
1254 /*
1255 * Now we can go ahead with allocating the shared arrays and
1256 * array caches
1257 */
1258 list_for_each_entry(cachep, &cache_chain, next) {
1259 struct array_cache *nc;
1260 struct array_cache *shared = NULL;
1261 struct array_cache **alien = NULL;
1263 nc = alloc_arraycache(node, cachep->limit,
1264 cachep->batchcount, GFP_KERNEL);
1265 if (!nc)
1266 goto bad;
1267 if (cachep->shared) {
1268 shared = alloc_arraycache(node,
1269 cachep->shared * cachep->batchcount,
1270 0xbaadf00d, GFP_KERNEL);
1271 if (!shared) {
1272 kfree(nc);
1273 goto bad;
1274 }
1275 }
1276 if (use_alien_caches) {
1277 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1278 if (!alien) {
1279 kfree(shared);
1280 kfree(nc);
1281 goto bad;
1282 }
1283 }
1284 cachep->array[cpu] = nc;
1285 l3 = cachep->nodelists[node];
1286 BUG_ON(!l3);
1288 spin_lock_irq(&l3->list_lock);
1289 if (!l3->shared) {
1290 /*
1291 * We are serialised from CPU_DEAD or
1292 * CPU_UP_CANCELLED by the cpucontrol lock
1293 */
1294 l3->shared = shared;
1295 shared = NULL;
1296 }
1297 #ifdef CONFIG_NUMA
1298 if (!l3->alien) {
1299 l3->alien = alien;
1300 alien = NULL;
1301 }
1302 #endif
1303 spin_unlock_irq(&l3->list_lock);
1304 kfree(shared);
1305 free_alien_cache(alien);
1306 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1307 slab_set_debugobj_lock_classes_node(cachep, node);
1308 }
1309 init_node_lock_keys(node);
1311 return 0;
1312 bad:
1313 cpuup_canceled(cpu);
1314 return -ENOMEM;
1315 }
1317 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1318 unsigned long action, void *hcpu)
1319 {
1320 long cpu = (long)hcpu;
1321 int err = 0;
1323 switch (action) {
1324 case CPU_UP_PREPARE:
1325 case CPU_UP_PREPARE_FROZEN:
1326 mutex_lock(&cache_chain_mutex);
1327 err = cpuup_prepare(cpu);
1328 mutex_unlock(&cache_chain_mutex);
1329 break;
1330 case CPU_ONLINE:
1331 case CPU_ONLINE_FROZEN:
1332 start_cpu_timer(cpu);
1333 break;
1334 #ifdef CONFIG_HOTPLUG_CPU
1335 case CPU_DOWN_PREPARE:
1336 case CPU_DOWN_PREPARE_FROZEN:
1337 /*
1338 * Shutdown cache reaper. Note that the cache_chain_mutex is
1339 * held so that if cache_reap() is invoked it cannot do
1340 * anything expensive but will only modify reap_work
1341 * and reschedule the timer.
1342 */
1343 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1344 /* Now the cache_reaper is guaranteed to be not running. */
1345 per_cpu(slab_reap_work, cpu).work.func = NULL;
1346 break;
1347 case CPU_DOWN_FAILED:
1348 case CPU_DOWN_FAILED_FROZEN:
1349 start_cpu_timer(cpu);
1350 break;
1351 case CPU_DEAD:
1352 case CPU_DEAD_FROZEN:
1353 /*
1354 * Even if all the cpus of a node are down, we don't free the
1355 * kmem_list3 of any cache. This to avoid a race between
1356 * cpu_down, and a kmalloc allocation from another cpu for
1357 * memory from the node of the cpu going down. The list3
1358 * structure is usually allocated from kmem_cache_create() and
1359 * gets destroyed at kmem_cache_destroy().
1360 */
1361 /* fall through */
1362 #endif
1363 case CPU_UP_CANCELED:
1364 case CPU_UP_CANCELED_FROZEN:
1365 mutex_lock(&cache_chain_mutex);
1366 cpuup_canceled(cpu);
1367 mutex_unlock(&cache_chain_mutex);
1368 break;
1369 }
1370 return notifier_from_errno(err);
1371 }
1373 static struct notifier_block __cpuinitdata cpucache_notifier = {
1374 &cpuup_callback, NULL, 0
1375 };
1377 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1378 /*
1379 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1380 * Returns -EBUSY if all objects cannot be drained so that the node is not
1381 * removed.
1382 *
1383 * Must hold cache_chain_mutex.
1384 */
1385 static int __meminit drain_cache_nodelists_node(int node)
1386 {
1387 struct kmem_cache *cachep;
1388 int ret = 0;
1390 list_for_each_entry(cachep, &cache_chain, next) {
1391 struct kmem_list3 *l3;
1393 l3 = cachep->nodelists[node];
1394 if (!l3)
1395 continue;
1397 drain_freelist(cachep, l3, l3->free_objects);
1399 if (!list_empty(&l3->slabs_full) ||
1400 !list_empty(&l3->slabs_partial)) {
1401 ret = -EBUSY;
1402 break;
1403 }
1404 }
1405 return ret;
1406 }
1408 static int __meminit slab_memory_callback(struct notifier_block *self,
1409 unsigned long action, void *arg)
1410 {
1411 struct memory_notify *mnb = arg;
1412 int ret = 0;
1413 int nid;
1415 nid = mnb->status_change_nid;
1416 if (nid < 0)
1417 goto out;
1419 switch (action) {
1420 case MEM_GOING_ONLINE:
1421 mutex_lock(&cache_chain_mutex);
1422 ret = init_cache_nodelists_node(nid);
1423 mutex_unlock(&cache_chain_mutex);
1424 break;
1425 case MEM_GOING_OFFLINE:
1426 mutex_lock(&cache_chain_mutex);
1427 ret = drain_cache_nodelists_node(nid);
1428 mutex_unlock(&cache_chain_mutex);
1429 break;
1430 case MEM_ONLINE:
1431 case MEM_OFFLINE:
1432 case MEM_CANCEL_ONLINE:
1433 case MEM_CANCEL_OFFLINE:
1434 break;
1435 }
1436 out:
1437 return notifier_from_errno(ret);
1438 }
1439 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1441 /*
1442 * swap the static kmem_list3 with kmalloced memory
1443 */
1444 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1445 int nodeid)
1446 {
1447 struct kmem_list3 *ptr;
1449 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1450 BUG_ON(!ptr);
1452 memcpy(ptr, list, sizeof(struct kmem_list3));
1453 /*
1454 * Do not assume that spinlocks can be initialized via memcpy:
1455 */
1456 spin_lock_init(&ptr->list_lock);
1458 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1459 cachep->nodelists[nodeid] = ptr;
1460 }
1462 /*
1463 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1464 * size of kmem_list3.
1465 */
1466 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1467 {
1468 int node;
1470 for_each_online_node(node) {
1471 cachep->nodelists[node] = &initkmem_list3[index + node];
1472 cachep->nodelists[node]->next_reap = jiffies +
1473 REAPTIMEOUT_LIST3 +
1474 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1475 }
1476 }
1478 /*
1479 * Initialisation. Called after the page allocator have been initialised and
1480 * before smp_init().
1481 */
1482 void __init kmem_cache_init(void)
1483 {
1484 size_t left_over;
1485 struct cache_sizes *sizes;
1486 struct cache_names *names;
1487 int i;
1488 int order;
1489 int node;
1491 if (num_possible_nodes() == 1)
1492 use_alien_caches = 0;
1494 for (i = 0; i < NUM_INIT_LISTS; i++) {
1495 kmem_list3_init(&initkmem_list3[i]);
1496 if (i < MAX_NUMNODES)
1497 cache_cache.nodelists[i] = NULL;
1498 }
1499 set_up_list3s(&cache_cache, CACHE_CACHE);
1501 /*
1502 * Fragmentation resistance on low memory - only use bigger
1503 * page orders on machines with more than 32MB of memory.
1504 */
1505 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1506 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1508 /* Bootstrap is tricky, because several objects are allocated
1509 * from caches that do not exist yet:
1510 * 1) initialize the cache_cache cache: it contains the struct
1511 * kmem_cache structures of all caches, except cache_cache itself:
1512 * cache_cache is statically allocated.
1513 * Initially an __init data area is used for the head array and the
1514 * kmem_list3 structures, it's replaced with a kmalloc allocated
1515 * array at the end of the bootstrap.
1516 * 2) Create the first kmalloc cache.
1517 * The struct kmem_cache for the new cache is allocated normally.
1518 * An __init data area is used for the head array.
1519 * 3) Create the remaining kmalloc caches, with minimally sized
1520 * head arrays.
1521 * 4) Replace the __init data head arrays for cache_cache and the first
1522 * kmalloc cache with kmalloc allocated arrays.
1523 * 5) Replace the __init data for kmem_list3 for cache_cache and
1524 * the other cache's with kmalloc allocated memory.
1525 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1526 */
1528 node = numa_mem_id();
1530 /* 1) create the cache_cache */
1531 INIT_LIST_HEAD(&cache_chain);
1532 list_add(&cache_cache.next, &cache_chain);
1533 cache_cache.colour_off = cache_line_size();
1534 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1535 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1537 /*
1538 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1539 */
1540 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1541 nr_node_ids * sizeof(struct kmem_list3 *);
1542 #if DEBUG
1543 cache_cache.obj_size = cache_cache.buffer_size;
1544 #endif
1545 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1546 cache_line_size());
1547 cache_cache.reciprocal_buffer_size =
1548 reciprocal_value(cache_cache.buffer_size);
1550 for (order = 0; order < MAX_ORDER; order++) {
1551 cache_estimate(order, cache_cache.buffer_size,
1552 cache_line_size(), 0, &left_over, &cache_cache.num);
1553 if (cache_cache.num)
1554 break;
1555 }
1556 BUG_ON(!cache_cache.num);
1557 cache_cache.gfporder = order;
1558 cache_cache.colour = left_over / cache_cache.colour_off;
1559 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1560 sizeof(struct slab), cache_line_size());
1562 /* 2+3) create the kmalloc caches */
1563 sizes = malloc_sizes;
1564 names = cache_names;
1566 /*
1567 * Initialize the caches that provide memory for the array cache and the
1568 * kmem_list3 structures first. Without this, further allocations will
1569 * bug.
1570 */
1572 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1573 sizes[INDEX_AC].cs_size,
1574 ARCH_KMALLOC_MINALIGN,
1575 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1576 NULL);
1578 if (INDEX_AC != INDEX_L3) {
1579 sizes[INDEX_L3].cs_cachep =
1580 kmem_cache_create(names[INDEX_L3].name,
1581 sizes[INDEX_L3].cs_size,
1582 ARCH_KMALLOC_MINALIGN,
1583 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1584 NULL);
1585 }
1587 slab_early_init = 0;
1589 while (sizes->cs_size != ULONG_MAX) {
1590 /*
1591 * For performance, all the general caches are L1 aligned.
1592 * This should be particularly beneficial on SMP boxes, as it
1593 * eliminates "false sharing".
1594 * Note for systems short on memory removing the alignment will
1595 * allow tighter packing of the smaller caches.
1596 */
1597 if (!sizes->cs_cachep) {
1598 sizes->cs_cachep = kmem_cache_create(names->name,
1599 sizes->cs_size,
1600 ARCH_KMALLOC_MINALIGN,
1601 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1602 NULL);
1603 }
1604 #ifdef CONFIG_ZONE_DMA
1605 sizes->cs_dmacachep = kmem_cache_create(
1606 names->name_dma,
1607 sizes->cs_size,
1608 ARCH_KMALLOC_MINALIGN,
1609 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1610 SLAB_PANIC,
1611 NULL);
1612 #endif
1613 sizes++;
1614 names++;
1615 }
1616 /* 4) Replace the bootstrap head arrays */
1617 {
1618 struct array_cache *ptr;
1620 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1622 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1623 memcpy(ptr, cpu_cache_get(&cache_cache),
1624 sizeof(struct arraycache_init));
1625 /*
1626 * Do not assume that spinlocks can be initialized via memcpy:
1627 */
1628 spin_lock_init(&ptr->lock);
1630 cache_cache.array[smp_processor_id()] = ptr;
1632 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1634 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1635 != &initarray_generic.cache);
1636 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1637 sizeof(struct arraycache_init));
1638 /*
1639 * Do not assume that spinlocks can be initialized via memcpy:
1640 */
1641 spin_lock_init(&ptr->lock);
1643 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1644 ptr;
1645 }
1646 /* 5) Replace the bootstrap kmem_list3's */
1647 {
1648 int nid;
1650 for_each_online_node(nid) {
1651 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1653 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1654 &initkmem_list3[SIZE_AC + nid], nid);
1656 if (INDEX_AC != INDEX_L3) {
1657 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1658 &initkmem_list3[SIZE_L3 + nid], nid);
1659 }
1660 }
1661 }
1663 g_cpucache_up = EARLY;
1664 }
1666 void __init kmem_cache_init_late(void)
1667 {
1668 struct kmem_cache *cachep;
1670 g_cpucache_up = LATE;
1672 /* Annotate slab for lockdep -- annotate the malloc caches */
1673 init_lock_keys();
1675 /* 6) resize the head arrays to their final sizes */
1676 mutex_lock(&cache_chain_mutex);
1677 list_for_each_entry(cachep, &cache_chain, next)
1678 if (enable_cpucache(cachep, GFP_NOWAIT))
1679 BUG();
1680 mutex_unlock(&cache_chain_mutex);
1682 /* Done! */
1683 g_cpucache_up = FULL;
1685 /*
1686 * Register a cpu startup notifier callback that initializes
1687 * cpu_cache_get for all new cpus
1688 */
1689 register_cpu_notifier(&cpucache_notifier);
1691 #ifdef CONFIG_NUMA
1692 /*
1693 * Register a memory hotplug callback that initializes and frees
1694 * nodelists.
1695 */
1696 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1697 #endif
1699 /*
1700 * The reap timers are started later, with a module init call: That part
1701 * of the kernel is not yet operational.
1702 */
1703 }
1705 static int __init cpucache_init(void)
1706 {
1707 int cpu;
1709 /*
1710 * Register the timers that return unneeded pages to the page allocator
1711 */
1712 for_each_online_cpu(cpu)
1713 start_cpu_timer(cpu);
1714 return 0;
1715 }
1716 __initcall(cpucache_init);
1718 /*
1719 * Interface to system's page allocator. No need to hold the cache-lock.
1720 *
1721 * If we requested dmaable memory, we will get it. Even if we
1722 * did not request dmaable memory, we might get it, but that
1723 * would be relatively rare and ignorable.
1724 */
1725 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1726 {
1727 struct page *page;
1728 int nr_pages;
1729 int i;
1731 #ifndef CONFIG_MMU
1732 /*
1733 * Nommu uses slab's for process anonymous memory allocations, and thus
1734 * requires __GFP_COMP to properly refcount higher order allocations
1735 */
1736 flags |= __GFP_COMP;
1737 #endif
1739 flags |= cachep->gfpflags;
1740 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1741 flags |= __GFP_RECLAIMABLE;
1743 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1744 if (!page)
1745 return NULL;
1747 nr_pages = (1 << cachep->gfporder);
1748 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1749 add_zone_page_state(page_zone(page),
1750 NR_SLAB_RECLAIMABLE, nr_pages);
1751 else
1752 add_zone_page_state(page_zone(page),
1753 NR_SLAB_UNRECLAIMABLE, nr_pages);
1754 for (i = 0; i < nr_pages; i++)
1755 __SetPageSlab(page + i);
1757 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1758 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1760 if (cachep->ctor)
1761 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1762 else
1763 kmemcheck_mark_unallocated_pages(page, nr_pages);
1764 }
1766 return page_address(page);
1767 }
1769 /*
1770 * Interface to system's page release.
1771 */
1772 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1773 {
1774 unsigned long i = (1 << cachep->gfporder);
1775 struct page *page = virt_to_page(addr);
1776 const unsigned long nr_freed = i;
1778 kmemcheck_free_shadow(page, cachep->gfporder);
1780 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1781 sub_zone_page_state(page_zone(page),
1782 NR_SLAB_RECLAIMABLE, nr_freed);
1783 else
1784 sub_zone_page_state(page_zone(page),
1785 NR_SLAB_UNRECLAIMABLE, nr_freed);
1786 while (i--) {
1787 BUG_ON(!PageSlab(page));
1788 __ClearPageSlab(page);
1789 page++;
1790 }
1791 if (current->reclaim_state)
1792 current->reclaim_state->reclaimed_slab += nr_freed;
1793 free_pages((unsigned long)addr, cachep->gfporder);
1794 }
1796 static void kmem_rcu_free(struct rcu_head *head)
1797 {
1798 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1799 struct kmem_cache *cachep = slab_rcu->cachep;
1801 kmem_freepages(cachep, slab_rcu->addr);
1802 if (OFF_SLAB(cachep))
1803 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1804 }
1806 #if DEBUG
1808 #ifdef CONFIG_DEBUG_PAGEALLOC
1809 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1810 unsigned long caller)
1811 {
1812 int size = obj_size(cachep);
1814 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1816 if (size < 5 * sizeof(unsigned long))
1817 return;
1819 *addr++ = 0x12345678;
1820 *addr++ = caller;
1821 *addr++ = smp_processor_id();
1822 size -= 3 * sizeof(unsigned long);
1823 {
1824 unsigned long *sptr = &caller;
1825 unsigned long svalue;
1827 while (!kstack_end(sptr)) {
1828 svalue = *sptr++;
1829 if (kernel_text_address(svalue)) {
1830 *addr++ = svalue;
1831 size -= sizeof(unsigned long);
1832 if (size <= sizeof(unsigned long))
1833 break;
1834 }
1835 }
1837 }
1838 *addr++ = 0x87654321;
1839 }
1840 #endif
1842 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1843 {
1844 int size = obj_size(cachep);
1845 addr = &((char *)addr)[obj_offset(cachep)];
1847 memset(addr, val, size);
1848 *(unsigned char *)(addr + size - 1) = POISON_END;
1849 }
1851 static void dump_line(char *data, int offset, int limit)
1852 {
1853 int i;
1854 unsigned char error = 0;
1855 int bad_count = 0;
1857 printk(KERN_ERR "%03x: ", offset);
1858 for (i = 0; i < limit; i++) {
1859 if (data[offset + i] != POISON_FREE) {
1860 error = data[offset + i];
1861 bad_count++;
1862 }
1863 }
1864 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1865 &data[offset], limit, 1);
1867 if (bad_count == 1) {
1868 error ^= POISON_FREE;
1869 if (!(error & (error - 1))) {
1870 printk(KERN_ERR "Single bit error detected. Probably "
1871 "bad RAM.\n");
1872 #ifdef CONFIG_X86
1873 printk(KERN_ERR "Run memtest86+ or a similar memory "
1874 "test tool.\n");
1875 #else
1876 printk(KERN_ERR "Run a memory test tool.\n");
1877 #endif
1878 }
1879 }
1880 }
1881 #endif
1883 #if DEBUG
1885 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1886 {
1887 int i, size;
1888 char *realobj;
1890 if (cachep->flags & SLAB_RED_ZONE) {
1891 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1892 *dbg_redzone1(cachep, objp),
1893 *dbg_redzone2(cachep, objp));
1894 }
1896 if (cachep->flags & SLAB_STORE_USER) {
1897 printk(KERN_ERR "Last user: [<%p>]",
1898 *dbg_userword(cachep, objp));
1899 print_symbol("(%s)",
1900 (unsigned long)*dbg_userword(cachep, objp));
1901 printk("\n");
1902 }
1903 realobj = (char *)objp + obj_offset(cachep);
1904 size = obj_size(cachep);
1905 for (i = 0; i < size && lines; i += 16, lines--) {
1906 int limit;
1907 limit = 16;
1908 if (i + limit > size)
1909 limit = size - i;
1910 dump_line(realobj, i, limit);
1911 }
1912 }
1914 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1915 {
1916 char *realobj;
1917 int size, i;
1918 int lines = 0;
1920 realobj = (char *)objp + obj_offset(cachep);
1921 size = obj_size(cachep);
1923 for (i = 0; i < size; i++) {
1924 char exp = POISON_FREE;
1925 if (i == size - 1)
1926 exp = POISON_END;
1927 if (realobj[i] != exp) {
1928 int limit;
1929 /* Mismatch ! */
1930 /* Print header */
1931 if (lines == 0) {
1932 printk(KERN_ERR
1933 "Slab corruption: %s start=%p, len=%d\n",
1934 cachep->name, realobj, size);
1935 print_objinfo(cachep, objp, 0);
1936 }
1937 /* Hexdump the affected line */
1938 i = (i / 16) * 16;
1939 limit = 16;
1940 if (i + limit > size)
1941 limit = size - i;
1942 dump_line(realobj, i, limit);
1943 i += 16;
1944 lines++;
1945 /* Limit to 5 lines */
1946 if (lines > 5)
1947 break;
1948 }
1949 }
1950 if (lines != 0) {
1951 /* Print some data about the neighboring objects, if they
1952 * exist:
1953 */
1954 struct slab *slabp = virt_to_slab(objp);
1955 unsigned int objnr;
1957 objnr = obj_to_index(cachep, slabp, objp);
1958 if (objnr) {
1959 objp = index_to_obj(cachep, slabp, objnr - 1);
1960 realobj = (char *)objp + obj_offset(cachep);
1961 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1962 realobj, size);
1963 print_objinfo(cachep, objp, 2);
1964 }
1965 if (objnr + 1 < cachep->num) {
1966 objp = index_to_obj(cachep, slabp, objnr + 1);
1967 realobj = (char *)objp + obj_offset(cachep);
1968 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1969 realobj, size);
1970 print_objinfo(cachep, objp, 2);
1971 }
1972 }
1973 }
1974 #endif
1976 #if DEBUG
1977 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1978 {
1979 int i;
1980 for (i = 0; i < cachep->num; i++) {
1981 void *objp = index_to_obj(cachep, slabp, i);
1983 if (cachep->flags & SLAB_POISON) {
1984 #ifdef CONFIG_DEBUG_PAGEALLOC
1985 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1986 OFF_SLAB(cachep))
1987 kernel_map_pages(virt_to_page(objp),
1988 cachep->buffer_size / PAGE_SIZE, 1);
1989 else
1990 check_poison_obj(cachep, objp);
1991 #else
1992 check_poison_obj(cachep, objp);
1993 #endif
1994 }
1995 if (cachep->flags & SLAB_RED_ZONE) {
1996 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1997 slab_error(cachep, "start of a freed object "
1998 "was overwritten");
1999 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2000 slab_error(cachep, "end of a freed object "
2001 "was overwritten");
2002 }
2003 }
2004 }
2005 #else
2006 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2007 {
2008 }
2009 #endif
2011 /**
2012 * slab_destroy - destroy and release all objects in a slab
2013 * @cachep: cache pointer being destroyed
2014 * @slabp: slab pointer being destroyed
2015 *
2016 * Destroy all the objs in a slab, and release the mem back to the system.
2017 * Before calling the slab must have been unlinked from the cache. The
2018 * cache-lock is not held/needed.
2019 */
2020 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2021 {
2022 void *addr = slabp->s_mem - slabp->colouroff;
2024 slab_destroy_debugcheck(cachep, slabp);
2025 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2026 struct slab_rcu *slab_rcu;
2028 slab_rcu = (struct slab_rcu *)slabp;
2029 slab_rcu->cachep = cachep;
2030 slab_rcu->addr = addr;
2031 call_rcu(&slab_rcu->head, kmem_rcu_free);
2032 } else {
2033 kmem_freepages(cachep, addr);
2034 if (OFF_SLAB(cachep))
2035 kmem_cache_free(cachep->slabp_cache, slabp);
2036 }
2037 }
2039 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2040 {
2041 int i;
2042 struct kmem_list3 *l3;
2044 for_each_online_cpu(i)
2045 kfree(cachep->array[i]);
2047 /* NUMA: free the list3 structures */
2048 for_each_online_node(i) {
2049 l3 = cachep->nodelists[i];
2050 if (l3) {
2051 kfree(l3->shared);
2052 free_alien_cache(l3->alien);
2053 kfree(l3);
2054 }
2055 }
2056 kmem_cache_free(&cache_cache, cachep);
2057 }
2060 /**
2061 * calculate_slab_order - calculate size (page order) of slabs
2062 * @cachep: pointer to the cache that is being created
2063 * @size: size of objects to be created in this cache.
2064 * @align: required alignment for the objects.
2065 * @flags: slab allocation flags
2066 *
2067 * Also calculates the number of objects per slab.
2068 *
2069 * This could be made much more intelligent. For now, try to avoid using
2070 * high order pages for slabs. When the gfp() functions are more friendly
2071 * towards high-order requests, this should be changed.
2072 */
2073 static size_t calculate_slab_order(struct kmem_cache *cachep,
2074 size_t size, size_t align, unsigned long flags)
2075 {
2076 unsigned long offslab_limit;
2077 size_t left_over = 0;
2078 int gfporder;
2080 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2081 unsigned int num;
2082 size_t remainder;
2084 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2085 if (!num)
2086 continue;
2088 if (flags & CFLGS_OFF_SLAB) {
2089 /*
2090 * Max number of objs-per-slab for caches which
2091 * use off-slab slabs. Needed to avoid a possible
2092 * looping condition in cache_grow().
2093 */
2094 offslab_limit = size - sizeof(struct slab);
2095 offslab_limit /= sizeof(kmem_bufctl_t);
2097 if (num > offslab_limit)
2098 break;
2099 }
2101 /* Found something acceptable - save it away */
2102 cachep->num = num;
2103 cachep->gfporder = gfporder;
2104 left_over = remainder;
2106 /*
2107 * A VFS-reclaimable slab tends to have most allocations
2108 * as GFP_NOFS and we really don't want to have to be allocating
2109 * higher-order pages when we are unable to shrink dcache.
2110 */
2111 if (flags & SLAB_RECLAIM_ACCOUNT)
2112 break;
2114 /*
2115 * Large number of objects is good, but very large slabs are
2116 * currently bad for the gfp()s.
2117 */
2118 if (gfporder >= slab_break_gfp_order)
2119 break;
2121 /*
2122 * Acceptable internal fragmentation?
2123 */
2124 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2125 break;
2126 }
2127 return left_over;
2128 }
2130 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2131 {
2132 if (g_cpucache_up == FULL)
2133 return enable_cpucache(cachep, gfp);
2135 if (g_cpucache_up == NONE) {
2136 /*
2137 * Note: the first kmem_cache_create must create the cache
2138 * that's used by kmalloc(24), otherwise the creation of
2139 * further caches will BUG().
2140 */
2141 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2143 /*
2144 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2145 * the first cache, then we need to set up all its list3s,
2146 * otherwise the creation of further caches will BUG().
2147 */
2148 set_up_list3s(cachep, SIZE_AC);
2149 if (INDEX_AC == INDEX_L3)
2150 g_cpucache_up = PARTIAL_L3;
2151 else
2152 g_cpucache_up = PARTIAL_AC;
2153 } else {
2154 cachep->array[smp_processor_id()] =
2155 kmalloc(sizeof(struct arraycache_init), gfp);
2157 if (g_cpucache_up == PARTIAL_AC) {
2158 set_up_list3s(cachep, SIZE_L3);
2159 g_cpucache_up = PARTIAL_L3;
2160 } else {
2161 int node;
2162 for_each_online_node(node) {
2163 cachep->nodelists[node] =
2164 kmalloc_node(sizeof(struct kmem_list3),
2165 gfp, node);
2166 BUG_ON(!cachep->nodelists[node]);
2167 kmem_list3_init(cachep->nodelists[node]);
2168 }
2169 }
2170 }
2171 cachep->nodelists[numa_mem_id()]->next_reap =
2172 jiffies + REAPTIMEOUT_LIST3 +
2173 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2175 cpu_cache_get(cachep)->avail = 0;
2176 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2177 cpu_cache_get(cachep)->batchcount = 1;
2178 cpu_cache_get(cachep)->touched = 0;
2179 cachep->batchcount = 1;
2180 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2181 return 0;
2182 }
2184 /**
2185 * kmem_cache_create - Create a cache.
2186 * @name: A string which is used in /proc/slabinfo to identify this cache.
2187 * @size: The size of objects to be created in this cache.
2188 * @align: The required alignment for the objects.
2189 * @flags: SLAB flags
2190 * @ctor: A constructor for the objects.
2191 *
2192 * Returns a ptr to the cache on success, NULL on failure.
2193 * Cannot be called within a int, but can be interrupted.
2194 * The @ctor is run when new pages are allocated by the cache.
2195 *
2196 * @name must be valid until the cache is destroyed. This implies that
2197 * the module calling this has to destroy the cache before getting unloaded.
2198 *
2199 * The flags are
2200 *
2201 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2202 * to catch references to uninitialised memory.
2203 *
2204 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2205 * for buffer overruns.
2206 *
2207 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2208 * cacheline. This can be beneficial if you're counting cycles as closely
2209 * as davem.
2210 */
2211 struct kmem_cache *
2212 kmem_cache_create (const char *name, size_t size, size_t align,
2213 unsigned long flags, void (*ctor)(void *))
2214 {
2215 size_t left_over, slab_size, ralign;
2216 struct kmem_cache *cachep = NULL, *pc;
2217 gfp_t gfp;
2219 /*
2220 * Sanity checks... these are all serious usage bugs.
2221 */
2222 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2223 size > KMALLOC_MAX_SIZE) {
2224 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2225 name);
2226 BUG();
2227 }
2229 /*
2230 * We use cache_chain_mutex to ensure a consistent view of
2231 * cpu_online_mask as well. Please see cpuup_callback
2232 */
2233 if (slab_is_available()) {
2234 get_online_cpus();
2235 mutex_lock(&cache_chain_mutex);
2236 }
2238 list_for_each_entry(pc, &cache_chain, next) {
2239 char tmp;
2240 int res;
2242 /*
2243 * This happens when the module gets unloaded and doesn't
2244 * destroy its slab cache and no-one else reuses the vmalloc
2245 * area of the module. Print a warning.
2246 */
2247 res = probe_kernel_address(pc->name, tmp);
2248 if (res) {
2249 printk(KERN_ERR
2250 "SLAB: cache with size %d has lost its name\n",
2251 pc->buffer_size);
2252 continue;
2253 }
2255 if (!strcmp(pc->name, name)) {
2256 printk(KERN_ERR
2257 "kmem_cache_create: duplicate cache %s\n", name);
2258 dump_stack();
2259 goto oops;
2260 }
2261 }
2263 #if DEBUG
2264 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2265 #if FORCED_DEBUG
2266 /*
2267 * Enable redzoning and last user accounting, except for caches with
2268 * large objects, if the increased size would increase the object size
2269 * above the next power of two: caches with object sizes just above a
2270 * power of two have a significant amount of internal fragmentation.
2271 */
2272 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2273 2 * sizeof(unsigned long long)))
2274 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2275 if (!(flags & SLAB_DESTROY_BY_RCU))
2276 flags |= SLAB_POISON;
2277 #endif
2278 if (flags & SLAB_DESTROY_BY_RCU)
2279 BUG_ON(flags & SLAB_POISON);
2280 #endif
2281 /*
2282 * Always checks flags, a caller might be expecting debug support which
2283 * isn't available.
2284 */
2285 BUG_ON(flags & ~CREATE_MASK);
2287 /*
2288 * Check that size is in terms of words. This is needed to avoid
2289 * unaligned accesses for some archs when redzoning is used, and makes
2290 * sure any on-slab bufctl's are also correctly aligned.
2291 */
2292 if (size & (BYTES_PER_WORD - 1)) {
2293 size += (BYTES_PER_WORD - 1);
2294 size &= ~(BYTES_PER_WORD - 1);
2295 }
2297 /* calculate the final buffer alignment: */
2299 /* 1) arch recommendation: can be overridden for debug */
2300 if (flags & SLAB_HWCACHE_ALIGN) {
2301 /*
2302 * Default alignment: as specified by the arch code. Except if
2303 * an object is really small, then squeeze multiple objects into
2304 * one cacheline.
2305 */
2306 ralign = cache_line_size();
2307 while (size <= ralign / 2)
2308 ralign /= 2;
2309 } else {
2310 ralign = BYTES_PER_WORD;
2311 }
2313 /*
2314 * Redzoning and user store require word alignment or possibly larger.
2315 * Note this will be overridden by architecture or caller mandated
2316 * alignment if either is greater than BYTES_PER_WORD.
2317 */
2318 if (flags & SLAB_STORE_USER)
2319 ralign = BYTES_PER_WORD;
2321 if (flags & SLAB_RED_ZONE) {
2322 ralign = REDZONE_ALIGN;
2323 /* If redzoning, ensure that the second redzone is suitably
2324 * aligned, by adjusting the object size accordingly. */
2325 size += REDZONE_ALIGN - 1;
2326 size &= ~(REDZONE_ALIGN - 1);
2327 }
2329 /* 2) arch mandated alignment */
2330 if (ralign < ARCH_SLAB_MINALIGN) {
2331 ralign = ARCH_SLAB_MINALIGN;
2332 }
2333 /* 3) caller mandated alignment */
2334 if (ralign < align) {
2335 ralign = align;
2336 }
2337 /* disable debug if necessary */
2338 if (ralign > __alignof__(unsigned long long))
2339 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2340 /*
2341 * 4) Store it.
2342 */
2343 align = ralign;
2345 if (slab_is_available())
2346 gfp = GFP_KERNEL;
2347 else
2348 gfp = GFP_NOWAIT;
2350 /* Get cache's description obj. */
2351 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2352 if (!cachep)
2353 goto oops;
2355 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2356 #if DEBUG
2357 cachep->obj_size = size;
2359 /*
2360 * Both debugging options require word-alignment which is calculated
2361 * into align above.
2362 */
2363 if (flags & SLAB_RED_ZONE) {
2364 /* add space for red zone words */
2365 cachep->obj_offset += sizeof(unsigned long long);
2366 size += 2 * sizeof(unsigned long long);
2367 }
2368 if (flags & SLAB_STORE_USER) {
2369 /* user store requires one word storage behind the end of
2370 * the real object. But if the second red zone needs to be
2371 * aligned to 64 bits, we must allow that much space.
2372 */
2373 if (flags & SLAB_RED_ZONE)
2374 size += REDZONE_ALIGN;
2375 else
2376 size += BYTES_PER_WORD;
2377 }
2378 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2379 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2380 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2381 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2382 size = PAGE_SIZE;
2383 }
2384 #endif
2385 #endif
2387 /*
2388 * Determine if the slab management is 'on' or 'off' slab.
2389 * (bootstrapping cannot cope with offslab caches so don't do
2390 * it too early on. Always use on-slab management when
2391 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2392 */
2393 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2394 !(flags & SLAB_NOLEAKTRACE))
2395 /*
2396 * Size is large, assume best to place the slab management obj
2397 * off-slab (should allow better packing of objs).
2398 */
2399 flags |= CFLGS_OFF_SLAB;
2401 size = ALIGN(size, align);
2403 left_over = calculate_slab_order(cachep, size, align, flags);
2405 if (!cachep->num) {
2406 printk(KERN_ERR
2407 "kmem_cache_create: couldn't create cache %s.\n", name);
2408 kmem_cache_free(&cache_cache, cachep);
2409 cachep = NULL;
2410 goto oops;
2411 }
2412 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2413 + sizeof(struct slab), align);
2415 /*
2416 * If the slab has been placed off-slab, and we have enough space then
2417 * move it on-slab. This is at the expense of any extra colouring.
2418 */
2419 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2420 flags &= ~CFLGS_OFF_SLAB;
2421 left_over -= slab_size;
2422 }
2424 if (flags & CFLGS_OFF_SLAB) {
2425 /* really off slab. No need for manual alignment */
2426 slab_size =
2427 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2429 #ifdef CONFIG_PAGE_POISONING
2430 /* If we're going to use the generic kernel_map_pages()
2431 * poisoning, then it's going to smash the contents of
2432 * the redzone and userword anyhow, so switch them off.
2433 */
2434 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2435 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2436 #endif
2437 }
2439 cachep->colour_off = cache_line_size();
2440 /* Offset must be a multiple of the alignment. */
2441 if (cachep->colour_off < align)
2442 cachep->colour_off = align;
2443 cachep->colour = left_over / cachep->colour_off;
2444 cachep->slab_size = slab_size;
2445 cachep->flags = flags;
2446 cachep->gfpflags = 0;
2447 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2448 cachep->gfpflags |= GFP_DMA;
2449 cachep->buffer_size = size;
2450 cachep->reciprocal_buffer_size = reciprocal_value(size);
2452 if (flags & CFLGS_OFF_SLAB) {
2453 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2454 /*
2455 * This is a possibility for one of the malloc_sizes caches.
2456 * But since we go off slab only for object size greater than
2457 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2458 * this should not happen at all.
2459 * But leave a BUG_ON for some lucky dude.
2460 */
2461 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2462 }
2463 cachep->ctor = ctor;
2464 cachep->name = name;
2466 if (setup_cpu_cache(cachep, gfp)) {
2467 __kmem_cache_destroy(cachep);
2468 cachep = NULL;
2469 goto oops;
2470 }
2472 if (flags & SLAB_DEBUG_OBJECTS) {
2473 /*
2474 * Would deadlock through slab_destroy()->call_rcu()->
2475 * debug_object_activate()->kmem_cache_alloc().
2476 */
2477 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2479 slab_set_debugobj_lock_classes(cachep);
2480 }
2482 /* cache setup completed, link it into the list */
2483 list_add(&cachep->next, &cache_chain);
2484 oops:
2485 if (!cachep && (flags & SLAB_PANIC))
2486 panic("kmem_cache_create(): failed to create slab `%s'\n",
2487 name);
2488 if (slab_is_available()) {
2489 mutex_unlock(&cache_chain_mutex);
2490 put_online_cpus();
2491 }
2492 return cachep;
2493 }
2494 EXPORT_SYMBOL(kmem_cache_create);
2496 #if DEBUG
2497 static void check_irq_off(void)
2498 {
2499 BUG_ON(!irqs_disabled());
2500 }
2502 static void check_irq_on(void)
2503 {
2504 BUG_ON(irqs_disabled());
2505 }
2507 static void check_spinlock_acquired(struct kmem_cache *cachep)
2508 {
2509 #ifdef CONFIG_SMP
2510 check_irq_off();
2511 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2512 #endif
2513 }
2515 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2516 {
2517 #ifdef CONFIG_SMP
2518 check_irq_off();
2519 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2520 #endif
2521 }
2523 #else
2524 #define check_irq_off() do { } while(0)
2525 #define check_irq_on() do { } while(0)
2526 #define check_spinlock_acquired(x) do { } while(0)
2527 #define check_spinlock_acquired_node(x, y) do { } while(0)
2528 #endif
2530 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2531 struct array_cache *ac,
2532 int force, int node);
2534 static void do_drain(void *arg)
2535 {
2536 struct kmem_cache *cachep = arg;
2537 struct array_cache *ac;
2538 int node = numa_mem_id();
2540 check_irq_off();
2541 ac = cpu_cache_get(cachep);
2542 spin_lock(&cachep->nodelists[node]->list_lock);
2543 free_block(cachep, ac->entry, ac->avail, node);
2544 spin_unlock(&cachep->nodelists[node]->list_lock);
2545 ac->avail = 0;
2546 }
2548 static void drain_cpu_caches(struct kmem_cache *cachep)
2549 {
2550 struct kmem_list3 *l3;
2551 int node;
2553 on_each_cpu(do_drain, cachep, 1);
2554 check_irq_on();
2555 for_each_online_node(node) {
2556 l3 = cachep->nodelists[node];
2557 if (l3 && l3->alien)
2558 drain_alien_cache(cachep, l3->alien);
2559 }
2561 for_each_online_node(node) {
2562 l3 = cachep->nodelists[node];
2563 if (l3)
2564 drain_array(cachep, l3, l3->shared, 1, node);
2565 }
2566 }
2568 /*
2569 * Remove slabs from the list of free slabs.
2570 * Specify the number of slabs to drain in tofree.
2571 *
2572 * Returns the actual number of slabs released.
2573 */
2574 static int drain_freelist(struct kmem_cache *cache,
2575 struct kmem_list3 *l3, int tofree)
2576 {
2577 struct list_head *p;
2578 int nr_freed;
2579 struct slab *slabp;
2581 nr_freed = 0;
2582 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2584 spin_lock_irq(&l3->list_lock);
2585 p = l3->slabs_free.prev;
2586 if (p == &l3->slabs_free) {
2587 spin_unlock_irq(&l3->list_lock);
2588 goto out;
2589 }
2591 slabp = list_entry(p, struct slab, list);
2592 #if DEBUG
2593 BUG_ON(slabp->inuse);
2594 #endif
2595 list_del(&slabp->list);
2596 /*
2597 * Safe to drop the lock. The slab is no longer linked
2598 * to the cache.
2599 */
2600 l3->free_objects -= cache->num;
2601 spin_unlock_irq(&l3->list_lock);
2602 slab_destroy(cache, slabp);
2603 nr_freed++;
2604 }
2605 out:
2606 return nr_freed;
2607 }
2609 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2610 static int __cache_shrink(struct kmem_cache *cachep)
2611 {
2612 int ret = 0, i = 0;
2613 struct kmem_list3 *l3;
2615 drain_cpu_caches(cachep);
2617 check_irq_on();
2618 for_each_online_node(i) {
2619 l3 = cachep->nodelists[i];
2620 if (!l3)
2621 continue;
2623 drain_freelist(cachep, l3, l3->free_objects);
2625 ret += !list_empty(&l3->slabs_full) ||
2626 !list_empty(&l3->slabs_partial);
2627 }
2628 return (ret ? 1 : 0);
2629 }
2631 /**
2632 * kmem_cache_shrink - Shrink a cache.
2633 * @cachep: The cache to shrink.
2634 *
2635 * Releases as many slabs as possible for a cache.
2636 * To help debugging, a zero exit status indicates all slabs were released.
2637 */
2638 int kmem_cache_shrink(struct kmem_cache *cachep)
2639 {
2640 int ret;
2641 BUG_ON(!cachep || in_interrupt());
2643 get_online_cpus();
2644 mutex_lock(&cache_chain_mutex);
2645 ret = __cache_shrink(cachep);
2646 mutex_unlock(&cache_chain_mutex);
2647 put_online_cpus();
2648 return ret;
2649 }
2650 EXPORT_SYMBOL(kmem_cache_shrink);
2652 /**
2653 * kmem_cache_destroy - delete a cache
2654 * @cachep: the cache to destroy
2655 *
2656 * Remove a &struct kmem_cache object from the slab cache.
2657 *
2658 * It is expected this function will be called by a module when it is
2659 * unloaded. This will remove the cache completely, and avoid a duplicate
2660 * cache being allocated each time a module is loaded and unloaded, if the
2661 * module doesn't have persistent in-kernel storage across loads and unloads.
2662 *
2663 * The cache must be empty before calling this function.
2664 *
2665 * The caller must guarantee that no one will allocate memory from the cache
2666 * during the kmem_cache_destroy().
2667 */
2668 void kmem_cache_destroy(struct kmem_cache *cachep)
2669 {
2670 BUG_ON(!cachep || in_interrupt());
2672 /* Find the cache in the chain of caches. */
2673 get_online_cpus();
2674 mutex_lock(&cache_chain_mutex);
2675 /*
2676 * the chain is never empty, cache_cache is never destroyed
2677 */
2678 list_del(&cachep->next);
2679 if (__cache_shrink(cachep)) {
2680 slab_error(cachep, "Can't free all objects");
2681 list_add(&cachep->next, &cache_chain);
2682 mutex_unlock(&cache_chain_mutex);
2683 put_online_cpus();
2684 return;
2685 }
2687 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2688 rcu_barrier();
2690 __kmem_cache_destroy(cachep);
2691 mutex_unlock(&cache_chain_mutex);
2692 put_online_cpus();
2693 }
2694 EXPORT_SYMBOL(kmem_cache_destroy);
2696 /*
2697 * Get the memory for a slab management obj.
2698 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2699 * always come from malloc_sizes caches. The slab descriptor cannot
2700 * come from the same cache which is getting created because,
2701 * when we are searching for an appropriate cache for these
2702 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2703 * If we are creating a malloc_sizes cache here it would not be visible to
2704 * kmem_find_general_cachep till the initialization is complete.
2705 * Hence we cannot have slabp_cache same as the original cache.
2706 */
2707 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2708 int colour_off, gfp_t local_flags,
2709 int nodeid)
2710 {
2711 struct slab *slabp;
2713 if (OFF_SLAB(cachep)) {
2714 /* Slab management obj is off-slab. */
2715 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2716 local_flags, nodeid);
2717 /*
2718 * If the first object in the slab is leaked (it's allocated
2719 * but no one has a reference to it), we want to make sure
2720 * kmemleak does not treat the ->s_mem pointer as a reference
2721 * to the object. Otherwise we will not report the leak.
2722 */
2723 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2724 local_flags);
2725 if (!slabp)
2726 return NULL;
2727 } else {
2728 slabp = objp + colour_off;
2729 colour_off += cachep->slab_size;
2730 }
2731 slabp->inuse = 0;
2732 slabp->colouroff = colour_off;
2733 slabp->s_mem = objp + colour_off;
2734 slabp->nodeid = nodeid;
2735 slabp->free = 0;
2736 return slabp;
2737 }
2739 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2740 {
2741 return (kmem_bufctl_t *) (slabp + 1);
2742 }
2744 static void cache_init_objs(struct kmem_cache *cachep,
2745 struct slab *slabp)
2746 {
2747 int i;
2749 for (i = 0; i < cachep->num; i++) {
2750 void *objp = index_to_obj(cachep, slabp, i);
2751 #if DEBUG
2752 /* need to poison the objs? */
2753 if (cachep->flags & SLAB_POISON)
2754 poison_obj(cachep, objp, POISON_FREE);
2755 if (cachep->flags & SLAB_STORE_USER)
2756 *dbg_userword(cachep, objp) = NULL;
2758 if (cachep->flags & SLAB_RED_ZONE) {
2759 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2760 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2761 }
2762 /*
2763 * Constructors are not allowed to allocate memory from the same
2764 * cache which they are a constructor for. Otherwise, deadlock.
2765 * They must also be threaded.
2766 */
2767 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2768 cachep->ctor(objp + obj_offset(cachep));
2770 if (cachep->flags & SLAB_RED_ZONE) {
2771 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2772 slab_error(cachep, "constructor overwrote the"
2773 " end of an object");
2774 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2775 slab_error(cachep, "constructor overwrote the"
2776 " start of an object");
2777 }
2778 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2779 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2780 kernel_map_pages(virt_to_page(objp),
2781 cachep->buffer_size / PAGE_SIZE, 0);
2782 #else
2783 if (cachep->ctor)
2784 cachep->ctor(objp);
2785 #endif
2786 slab_bufctl(slabp)[i] = i + 1;
2787 }
2788 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2789 }
2791 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2792 {
2793 if (CONFIG_ZONE_DMA_FLAG) {
2794 if (flags & GFP_DMA)
2795 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2796 else
2797 BUG_ON(cachep->gfpflags & GFP_DMA);
2798 }
2799 }
2801 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2802 int nodeid)
2803 {
2804 void *objp = index_to_obj(cachep, slabp, slabp->free);
2805 kmem_bufctl_t next;
2807 slabp->inuse++;
2808 next = slab_bufctl(slabp)[slabp->free];
2809 #if DEBUG
2810 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2811 WARN_ON(slabp->nodeid != nodeid);
2812 #endif
2813 slabp->free = next;
2815 return objp;
2816 }
2818 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2819 void *objp, int nodeid)
2820 {
2821 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2823 #if DEBUG
2824 /* Verify that the slab belongs to the intended node */
2825 WARN_ON(slabp->nodeid != nodeid);
2827 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2828 printk(KERN_ERR "slab: double free detected in cache "
2829 "'%s', objp %p\n", cachep->name, objp);
2830 BUG();
2831 }
2832 #endif
2833 slab_bufctl(slabp)[objnr] = slabp->free;
2834 slabp->free = objnr;
2835 slabp->inuse--;
2836 }
2838 /*
2839 * Map pages beginning at addr to the given cache and slab. This is required
2840 * for the slab allocator to be able to lookup the cache and slab of a
2841 * virtual address for kfree, ksize, and slab debugging.
2842 */
2843 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2844 void *addr)
2845 {
2846 int nr_pages;
2847 struct page *page;
2849 page = virt_to_page(addr);
2851 nr_pages = 1;
2852 if (likely(!PageCompound(page)))
2853 nr_pages <<= cache->gfporder;
2855 do {
2856 page_set_cache(page, cache);
2857 page_set_slab(page, slab);
2858 page++;
2859 } while (--nr_pages);
2860 }
2862 /*
2863 * Grow (by 1) the number of slabs within a cache. This is called by
2864 * kmem_cache_alloc() when there are no active objs left in a cache.
2865 */
2866 static int cache_grow(struct kmem_cache *cachep,
2867 gfp_t flags, int nodeid, void *objp)
2868 {
2869 struct slab *slabp;
2870 size_t offset;
2871 gfp_t local_flags;
2872 struct kmem_list3 *l3;
2874 /*
2875 * Be lazy and only check for valid flags here, keeping it out of the
2876 * critical path in kmem_cache_alloc().
2877 */
2878 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2879 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2881 /* Take the l3 list lock to change the colour_next on this node */
2882 check_irq_off();
2883 l3 = cachep->nodelists[nodeid];
2884 spin_lock(&l3->list_lock);
2886 /* Get colour for the slab, and cal the next value. */
2887 offset = l3->colour_next;
2888 l3->colour_next++;
2889 if (l3->colour_next >= cachep->colour)
2890 l3->colour_next = 0;
2891 spin_unlock(&l3->list_lock);
2893 offset *= cachep->colour_off;
2895 if (local_flags & __GFP_WAIT)
2896 local_irq_enable();
2898 /*
2899 * The test for missing atomic flag is performed here, rather than
2900 * the more obvious place, simply to reduce the critical path length
2901 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2902 * will eventually be caught here (where it matters).
2903 */
2904 kmem_flagcheck(cachep, flags);
2906 /*
2907 * Get mem for the objs. Attempt to allocate a physical page from
2908 * 'nodeid'.
2909 */
2910 if (!objp)
2911 objp = kmem_getpages(cachep, local_flags, nodeid);
2912 if (!objp)
2913 goto failed;
2915 /* Get slab management. */
2916 slabp = alloc_slabmgmt(cachep, objp, offset,
2917 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2918 if (!slabp)
2919 goto opps1;
2921 slab_map_pages(cachep, slabp, objp);
2923 cache_init_objs(cachep, slabp);
2925 if (local_flags & __GFP_WAIT)
2926 local_irq_disable();
2927 check_irq_off();
2928 spin_lock(&l3->list_lock);
2930 /* Make slab active. */
2931 list_add_tail(&slabp->list, &(l3->slabs_free));
2932 STATS_INC_GROWN(cachep);
2933 l3->free_objects += cachep->num;
2934 spin_unlock(&l3->list_lock);
2935 return 1;
2936 opps1:
2937 kmem_freepages(cachep, objp);
2938 failed:
2939 if (local_flags & __GFP_WAIT)
2940 local_irq_disable();
2941 return 0;
2942 }
2944 #if DEBUG
2946 /*
2947 * Perform extra freeing checks:
2948 * - detect bad pointers.
2949 * - POISON/RED_ZONE checking
2950 */
2951 static void kfree_debugcheck(const void *objp)
2952 {
2953 if (!virt_addr_valid(objp)) {
2954 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2955 (unsigned long)objp);
2956 BUG();
2957 }
2958 }
2960 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2961 {
2962 unsigned long long redzone1, redzone2;
2964 redzone1 = *dbg_redzone1(cache, obj);
2965 redzone2 = *dbg_redzone2(cache, obj);
2967 /*
2968 * Redzone is ok.
2969 */
2970 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2971 return;
2973 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2974 slab_error(cache, "double free detected");
2975 else
2976 slab_error(cache, "memory outside object was overwritten");
2978 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2979 obj, redzone1, redzone2);
2980 }
2982 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2983 void *caller)
2984 {
2985 struct page *page;
2986 unsigned int objnr;
2987 struct slab *slabp;
2989 BUG_ON(virt_to_cache(objp) != cachep);
2991 objp -= obj_offset(cachep);
2992 kfree_debugcheck(objp);
2993 page = virt_to_head_page(objp);
2995 slabp = page_get_slab(page);
2997 if (cachep->flags & SLAB_RED_ZONE) {
2998 verify_redzone_free(cachep, objp);
2999 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3000 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3001 }
3002 if (cachep->flags & SLAB_STORE_USER)
3003 *dbg_userword(cachep, objp) = caller;
3005 objnr = obj_to_index(cachep, slabp, objp);
3007 BUG_ON(objnr >= cachep->num);
3008 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3010 #ifdef CONFIG_DEBUG_SLAB_LEAK
3011 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3012 #endif
3013 if (cachep->flags & SLAB_POISON) {
3014 #ifdef CONFIG_DEBUG_PAGEALLOC
3015 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3016 store_stackinfo(cachep, objp, (unsigned long)caller);
3017 kernel_map_pages(virt_to_page(objp),
3018 cachep->buffer_size / PAGE_SIZE, 0);
3019 } else {
3020 poison_obj(cachep, objp, POISON_FREE);
3021 }
3022 #else
3023 poison_obj(cachep, objp, POISON_FREE);
3024 #endif
3025 }
3026 return objp;
3027 }
3029 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3030 {
3031 kmem_bufctl_t i;
3032 int entries = 0;
3034 /* Check slab's freelist to see if this obj is there. */
3035 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3036 entries++;
3037 if (entries > cachep->num || i >= cachep->num)
3038 goto bad;
3039 }
3040 if (entries != cachep->num - slabp->inuse) {
3041 bad:
3042 printk(KERN_ERR "slab: Internal list corruption detected in "
3043 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3044 cachep->name, cachep->num, slabp, slabp->inuse);
3045 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3046 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3047 1);
3048 BUG();
3049 }
3050 }
3051 #else
3052 #define kfree_debugcheck(x) do { } while(0)
3053 #define cache_free_debugcheck(x,objp,z) (objp)
3054 #define check_slabp(x,y) do { } while(0)
3055 #endif
3057 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3058 {
3059 int batchcount;
3060 struct kmem_list3 *l3;
3061 struct array_cache *ac;
3062 int node;
3064 retry:
3065 check_irq_off();
3066 node = numa_mem_id();
3067 ac = cpu_cache_get(cachep);
3068 batchcount = ac->batchcount;
3069 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3070 /*
3071 * If there was little recent activity on this cache, then
3072 * perform only a partial refill. Otherwise we could generate
3073 * refill bouncing.
3074 */
3075 batchcount = BATCHREFILL_LIMIT;
3076 }
3077 l3 = cachep->nodelists[node];
3079 BUG_ON(ac->avail > 0 || !l3);
3080 spin_lock(&l3->list_lock);
3082 /* See if we can refill from the shared array */
3083 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3084 l3->shared->touched = 1;
3085 goto alloc_done;
3086 }
3088 while (batchcount > 0) {
3089 struct list_head *entry;
3090 struct slab *slabp;
3091 /* Get slab alloc is to come from. */
3092 entry = l3->slabs_partial.next;
3093 if (entry == &l3->slabs_partial) {
3094 l3->free_touched = 1;
3095 entry = l3->slabs_free.next;
3096 if (entry == &l3->slabs_free)
3097 goto must_grow;
3098 }
3100 slabp = list_entry(entry, struct slab, list);
3101 check_slabp(cachep, slabp);
3102 check_spinlock_acquired(cachep);
3104 /*
3105 * The slab was either on partial or free list so
3106 * there must be at least one object available for
3107 * allocation.
3108 */
3109 BUG_ON(slabp->inuse >= cachep->num);
3111 while (slabp->inuse < cachep->num && batchcount--) {
3112 STATS_INC_ALLOCED(cachep);
3113 STATS_INC_ACTIVE(cachep);
3114 STATS_SET_HIGH(cachep);
3116 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3117 node);
3118 }
3119 check_slabp(cachep, slabp);
3121 /* move slabp to correct slabp list: */
3122 list_del(&slabp->list);
3123 if (slabp->free == BUFCTL_END)
3124 list_add(&slabp->list, &l3->slabs_full);
3125 else
3126 list_add(&slabp->list, &l3->slabs_partial);
3127 }
3129 must_grow:
3130 l3->free_objects -= ac->avail;
3131 alloc_done:
3132 spin_unlock(&l3->list_lock);
3134 if (unlikely(!ac->avail)) {
3135 int x;
3136 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3138 /* cache_grow can reenable interrupts, then ac could change. */
3139 ac = cpu_cache_get(cachep);
3140 if (!x && ac->avail == 0) /* no objects in sight? abort */
3141 return NULL;
3143 if (!ac->avail) /* objects refilled by interrupt? */
3144 goto retry;
3145 }
3146 ac->touched = 1;
3147 return ac->entry[--ac->avail];
3148 }
3150 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3151 gfp_t flags)
3152 {
3153 might_sleep_if(flags & __GFP_WAIT);
3154 #if DEBUG
3155 kmem_flagcheck(cachep, flags);
3156 #endif
3157 }
3159 #if DEBUG
3160 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3161 gfp_t flags, void *objp, void *caller)
3162 {
3163 if (!objp)
3164 return objp;
3165 if (cachep->flags & SLAB_POISON) {
3166 #ifdef CONFIG_DEBUG_PAGEALLOC
3167 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3168 kernel_map_pages(virt_to_page(objp),
3169 cachep->buffer_size / PAGE_SIZE, 1);
3170 else
3171 check_poison_obj(cachep, objp);
3172 #else
3173 check_poison_obj(cachep, objp);
3174 #endif
3175 poison_obj(cachep, objp, POISON_INUSE);
3176 }
3177 if (cachep->flags & SLAB_STORE_USER)
3178 *dbg_userword(cachep, objp) = caller;
3180 if (cachep->flags & SLAB_RED_ZONE) {
3181 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3182 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3183 slab_error(cachep, "double free, or memory outside"
3184 " object was overwritten");
3185 printk(KERN_ERR
3186 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3187 objp, *dbg_redzone1(cachep, objp),
3188 *dbg_redzone2(cachep, objp));
3189 }
3190 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3191 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3192 }
3193 #ifdef CONFIG_DEBUG_SLAB_LEAK
3194 {
3195 struct slab *slabp;
3196 unsigned objnr;
3198 slabp = page_get_slab(virt_to_head_page(objp));
3199 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3200 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3201 }
3202 #endif
3203 objp += obj_offset(cachep);
3204 if (cachep->ctor && cachep->flags & SLAB_POISON)
3205 cachep->ctor(objp);
3206 if (ARCH_SLAB_MINALIGN &&
3207 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3208 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3209 objp, (int)ARCH_SLAB_MINALIGN);
3210 }
3211 return objp;
3212 }
3213 #else
3214 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3215 #endif
3217 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3218 {
3219 if (cachep == &cache_cache)
3220 return false;
3222 return should_failslab(obj_size(cachep), flags, cachep->flags);
3223 }
3225 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3226 {
3227 void *objp;
3228 struct array_cache *ac;
3230 check_irq_off();
3232 ac = cpu_cache_get(cachep);
3233 if (likely(ac->avail)) {
3234 STATS_INC_ALLOCHIT(cachep);
3235 ac->touched = 1;
3236 objp = ac->entry[--ac->avail];
3237 } else {
3238 STATS_INC_ALLOCMISS(cachep);
3239 objp = cache_alloc_refill(cachep, flags);
3240 /*
3241 * the 'ac' may be updated by cache_alloc_refill(),
3242 * and kmemleak_erase() requires its correct value.
3243 */
3244 ac = cpu_cache_get(cachep);
3245 }
3246 /*
3247 * To avoid a false negative, if an object that is in one of the
3248 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3249 * treat the array pointers as a reference to the object.
3250 */
3251 if (objp)
3252 kmemleak_erase(&ac->entry[ac->avail]);
3253 return objp;
3254 }
3256 #ifdef CONFIG_NUMA
3257 /*
3258 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3259 *
3260 * If we are in_interrupt, then process context, including cpusets and
3261 * mempolicy, may not apply and should not be used for allocation policy.
3262 */
3263 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3264 {
3265 int nid_alloc, nid_here;
3267 if (in_interrupt() || (flags & __GFP_THISNODE))
3268 return NULL;
3269 nid_alloc = nid_here = numa_mem_id();
3270 get_mems_allowed();
3271 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3272 nid_alloc = cpuset_slab_spread_node();
3273 else if (current->mempolicy)
3274 nid_alloc = slab_node(current->mempolicy);
3275 put_mems_allowed();
3276 if (nid_alloc != nid_here)
3277 return ____cache_alloc_node(cachep, flags, nid_alloc);
3278 return NULL;
3279 }
3281 /*
3282 * Fallback function if there was no memory available and no objects on a
3283 * certain node and fall back is permitted. First we scan all the
3284 * available nodelists for available objects. If that fails then we
3285 * perform an allocation without specifying a node. This allows the page
3286 * allocator to do its reclaim / fallback magic. We then insert the
3287 * slab into the proper nodelist and then allocate from it.
3288 */
3289 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3290 {
3291 struct zonelist *zonelist;
3292 gfp_t local_flags;
3293 struct zoneref *z;
3294 struct zone *zone;
3295 enum zone_type high_zoneidx = gfp_zone(flags);
3296 void *obj = NULL;
3297 int nid;
3299 if (flags & __GFP_THISNODE)
3300 return NULL;
3302 get_mems_allowed();
3303 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3304 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3306 retry:
3307 /*
3308 * Look through allowed nodes for objects available
3309 * from existing per node queues.
3310 */
3311 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3312 nid = zone_to_nid(zone);
3314 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3315 cache->nodelists[nid] &&
3316 cache->nodelists[nid]->free_objects) {
3317 obj = ____cache_alloc_node(cache,
3318 flags | GFP_THISNODE, nid);
3319 if (obj)
3320 break;
3321 }
3322 }
3324 if (!obj) {
3325 /*
3326 * This allocation will be performed within the constraints
3327 * of the current cpuset / memory policy requirements.
3328 * We may trigger various forms of reclaim on the allowed
3329 * set and go into memory reserves if necessary.
3330 */
3331 if (local_flags & __GFP_WAIT)
3332 local_irq_enable();
3333 kmem_flagcheck(cache, flags);
3334 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3335 if (local_flags & __GFP_WAIT)
3336 local_irq_disable();
3337 if (obj) {
3338 /*
3339 * Insert into the appropriate per node queues
3340 */
3341 nid = page_to_nid(virt_to_page(obj));
3342 if (cache_grow(cache, flags, nid, obj)) {
3343 obj = ____cache_alloc_node(cache,
3344 flags | GFP_THISNODE, nid);
3345 if (!obj)
3346 /*
3347 * Another processor may allocate the
3348 * objects in the slab since we are
3349 * not holding any locks.
3350 */
3351 goto retry;
3352 } else {
3353 /* cache_grow already freed obj */
3354 obj = NULL;
3355 }
3356 }
3357 }
3358 put_mems_allowed();
3359 return obj;
3360 }
3362 /*
3363 * A interface to enable slab creation on nodeid
3364 */
3365 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3366 int nodeid)
3367 {
3368 struct list_head *entry;
3369 struct slab *slabp;
3370 struct kmem_list3 *l3;
3371 void *obj;
3372 int x;
3374 l3 = cachep->nodelists[nodeid];
3375 BUG_ON(!l3);
3377 retry:
3378 check_irq_off();
3379 spin_lock(&l3->list_lock);
3380 entry = l3->slabs_partial.next;
3381 if (entry == &l3->slabs_partial) {
3382 l3->free_touched = 1;
3383 entry = l3->slabs_free.next;
3384 if (entry == &l3->slabs_free)
3385 goto must_grow;
3386 }
3388 slabp = list_entry(entry, struct slab, list);
3389 check_spinlock_acquired_node(cachep, nodeid);
3390 check_slabp(cachep, slabp);
3392 STATS_INC_NODEALLOCS(cachep);
3393 STATS_INC_ACTIVE(cachep);
3394 STATS_SET_HIGH(cachep);
3396 BUG_ON(slabp->inuse == cachep->num);
3398 obj = slab_get_obj(cachep, slabp, nodeid);
3399 check_slabp(cachep, slabp);
3400 l3->free_objects--;
3401 /* move slabp to correct slabp list: */
3402 list_del(&slabp->list);
3404 if (slabp->free == BUFCTL_END)
3405 list_add(&slabp->list, &l3->slabs_full);
3406 else
3407 list_add(&slabp->list, &l3->slabs_partial);
3409 spin_unlock(&l3->list_lock);
3410 goto done;
3412 must_grow:
3413 spin_unlock(&l3->list_lock);
3414 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3415 if (x)
3416 goto retry;
3418 return fallback_alloc(cachep, flags);
3420 done:
3421 return obj;
3422 }
3424 /**
3425 * kmem_cache_alloc_node - Allocate an object on the specified node
3426 * @cachep: The cache to allocate from.
3427 * @flags: See kmalloc().
3428 * @nodeid: node number of the target node.
3429 * @caller: return address of caller, used for debug information
3430 *
3431 * Identical to kmem_cache_alloc but it will allocate memory on the given
3432 * node, which can improve the performance for cpu bound structures.
3433 *
3434 * Fallback to other node is possible if __GFP_THISNODE is not set.
3435 */
3436 static __always_inline void *
3437 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3438 void *caller)
3439 {
3440 unsigned long save_flags;
3441 void *ptr;
3442 int slab_node = numa_mem_id();
3444 flags &= gfp_allowed_mask;
3446 lockdep_trace_alloc(flags);
3448 if (slab_should_failslab(cachep, flags))
3449 return NULL;
3451 cache_alloc_debugcheck_before(cachep, flags);
3452 local_irq_save(save_flags);
3454 if (nodeid == NUMA_NO_NODE)
3455 nodeid = slab_node;
3457 if (unlikely(!cachep->nodelists[nodeid])) {
3458 /* Node not bootstrapped yet */
3459 ptr = fallback_alloc(cachep, flags);
3460 goto out;
3461 }
3463 if (nodeid == slab_node) {
3464 /*
3465 * Use the locally cached objects if possible.
3466 * However ____cache_alloc does not allow fallback
3467 * to other nodes. It may fail while we still have
3468 * objects on other nodes available.
3469 */
3470 ptr = ____cache_alloc(cachep, flags);
3471 if (ptr)
3472 goto out;
3473 }
3474 /* ___cache_alloc_node can fall back to other nodes */
3475 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3476 out:
3477 local_irq_restore(save_flags);
3478 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3479 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3480 flags);
3482 if (likely(ptr))
3483 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3485 if (unlikely((flags & __GFP_ZERO) && ptr))
3486 memset(ptr, 0, obj_size(cachep));
3488 return ptr;
3489 }
3491 static __always_inline void *
3492 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3493 {
3494 void *objp;
3496 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3497 objp = alternate_node_alloc(cache, flags);
3498 if (objp)
3499 goto out;
3500 }
3501 objp = ____cache_alloc(cache, flags);
3503 /*
3504 * We may just have run out of memory on the local node.
3505 * ____cache_alloc_node() knows how to locate memory on other nodes
3506 */
3507 if (!objp)
3508 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3510 out:
3511 return objp;
3512 }
3513 #else
3515 static __always_inline void *
3516 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3517 {
3518 return ____cache_alloc(cachep, flags);
3519 }
3521 #endif /* CONFIG_NUMA */
3523 static __always_inline void *
3524 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3525 {
3526 unsigned long save_flags;
3527 void *objp;
3529 flags &= gfp_allowed_mask;
3531 lockdep_trace_alloc(flags);
3533 if (slab_should_failslab(cachep, flags))
3534 return NULL;
3536 cache_alloc_debugcheck_before(cachep, flags);
3537 local_irq_save(save_flags);
3538 objp = __do_cache_alloc(cachep, flags);
3539 local_irq_restore(save_flags);
3540 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3541 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3542 flags);
3543 prefetchw(objp);
3545 if (likely(objp))
3546 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3548 if (unlikely((flags & __GFP_ZERO) && objp))
3549 memset(objp, 0, obj_size(cachep));
3551 return objp;
3552 }
3554 /*
3555 * Caller needs to acquire correct kmem_list's list_lock
3556 */
3557 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3558 int node)
3559 {
3560 int i;
3561 struct kmem_list3 *l3;
3563 for (i = 0; i < nr_objects; i++) {
3564 void *objp = objpp[i];
3565 struct slab *slabp;
3567 slabp = virt_to_slab(objp);
3568 l3 = cachep->nodelists[node];
3569 list_del(&slabp->list);
3570 check_spinlock_acquired_node(cachep, node);
3571 check_slabp(cachep, slabp);
3572 slab_put_obj(cachep, slabp, objp, node);
3573 STATS_DEC_ACTIVE(cachep);
3574 l3->free_objects++;
3575 check_slabp(cachep, slabp);
3577 /* fixup slab chains */
3578 if (slabp->inuse == 0) {
3579 if (l3->free_objects > l3->free_limit) {
3580 l3->free_objects -= cachep->num;
3581 /* No need to drop any previously held
3582 * lock here, even if we have a off-slab slab
3583 * descriptor it is guaranteed to come from
3584 * a different cache, refer to comments before
3585 * alloc_slabmgmt.
3586 */
3587 slab_destroy(cachep, slabp);
3588 } else {
3589 list_add(&slabp->list, &l3->slabs_free);
3590 }
3591 } else {
3592 /* Unconditionally move a slab to the end of the
3593 * partial list on free - maximum time for the
3594 * other objects to be freed, too.
3595 */
3596 list_add_tail(&slabp->list, &l3->slabs_partial);
3597 }
3598 }
3599 }
3601 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3602 {
3603 int batchcount;
3604 struct kmem_list3 *l3;
3605 int node = numa_mem_id();
3607 batchcount = ac->batchcount;
3608 #if DEBUG
3609 BUG_ON(!batchcount || batchcount > ac->avail);
3610 #endif
3611 check_irq_off();
3612 l3 = cachep->nodelists[node];
3613 spin_lock(&l3->list_lock);
3614 if (l3->shared) {
3615 struct array_cache *shared_array = l3->shared;
3616 int max = shared_array->limit - shared_array->avail;
3617 if (max) {
3618 if (batchcount > max)
3619 batchcount = max;
3620 memcpy(&(shared_array->entry[shared_array->avail]),
3621 ac->entry, sizeof(void *) * batchcount);
3622 shared_array->avail += batchcount;
3623 goto free_done;
3624 }
3625 }
3627 free_block(cachep, ac->entry, batchcount, node);
3628 free_done:
3629 #if STATS
3630 {
3631 int i = 0;
3632 struct list_head *p;
3634 p = l3->slabs_free.next;
3635 while (p != &(l3->slabs_free)) {
3636 struct slab *slabp;
3638 slabp = list_entry(p, struct slab, list);
3639 BUG_ON(slabp->inuse);
3641 i++;
3642 p = p->next;
3643 }
3644 STATS_SET_FREEABLE(cachep, i);
3645 }
3646 #endif
3647 spin_unlock(&l3->list_lock);
3648 ac->avail -= batchcount;
3649 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3650 }
3652 /*
3653 * Release an obj back to its cache. If the obj has a constructed state, it must
3654 * be in this state _before_ it is released. Called with disabled ints.
3655 */
3656 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3657 void *caller)
3658 {
3659 struct array_cache *ac = cpu_cache_get(cachep);
3661 check_irq_off();
3662 kmemleak_free_recursive(objp, cachep->flags);
3663 objp = cache_free_debugcheck(cachep, objp, caller);
3665 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3667 /*
3668 * Skip calling cache_free_alien() when the platform is not numa.
3669 * This will avoid cache misses that happen while accessing slabp (which
3670 * is per page memory reference) to get nodeid. Instead use a global
3671 * variable to skip the call, which is mostly likely to be present in
3672 * the cache.
3673 */
3674 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3675 return;
3677 if (likely(ac->avail < ac->limit)) {
3678 STATS_INC_FREEHIT(cachep);
3679 ac->entry[ac->avail++] = objp;
3680 return;
3681 } else {
3682 STATS_INC_FREEMISS(cachep);
3683 cache_flusharray(cachep, ac);
3684 ac->entry[ac->avail++] = objp;
3685 }
3686 }
3688 /**
3689 * kmem_cache_alloc - Allocate an object
3690 * @cachep: The cache to allocate from.
3691 * @flags: See kmalloc().
3692 *
3693 * Allocate an object from this cache. The flags are only relevant
3694 * if the cache has no available objects.
3695 */
3696 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3697 {
3698 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3700 trace_kmem_cache_alloc(_RET_IP_, ret,
3701 obj_size(cachep), cachep->buffer_size, flags);
3703 return ret;
3704 }
3705 EXPORT_SYMBOL(kmem_cache_alloc);
3707 #ifdef CONFIG_TRACING
3708 void *
3709 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3710 {
3711 void *ret;
3713 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3715 trace_kmalloc(_RET_IP_, ret,
3716 size, slab_buffer_size(cachep), flags);
3717 return ret;
3718 }
3719 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3720 #endif
3722 #ifdef CONFIG_NUMA
3723 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3724 {
3725 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3726 __builtin_return_address(0));
3728 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3729 obj_size(cachep), cachep->buffer_size,
3730 flags, nodeid);
3732 return ret;
3733 }
3734 EXPORT_SYMBOL(kmem_cache_alloc_node);
3736 #ifdef CONFIG_TRACING
3737 void *kmem_cache_alloc_node_trace(size_t size,
3738 struct kmem_cache *cachep,
3739 gfp_t flags,
3740 int nodeid)
3741 {
3742 void *ret;
3744 ret = __cache_alloc_node(cachep, flags, nodeid,
3745 __builtin_return_address(0));
3746 trace_kmalloc_node(_RET_IP_, ret,
3747 size, slab_buffer_size(cachep),
3748 flags, nodeid);
3749 return ret;
3750 }
3751 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3752 #endif
3754 static __always_inline void *
3755 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3756 {
3757 struct kmem_cache *cachep;
3759 cachep = kmem_find_general_cachep(size, flags);
3760 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3761 return cachep;
3762 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3763 }
3765 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3766 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3767 {
3768 return __do_kmalloc_node(size, flags, node,
3769 __builtin_return_address(0));
3770 }
3771 EXPORT_SYMBOL(__kmalloc_node);
3773 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3774 int node, unsigned long caller)
3775 {
3776 return __do_kmalloc_node(size, flags, node, (void *)caller);
3777 }
3778 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3779 #else
3780 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3781 {
3782 return __do_kmalloc_node(size, flags, node, NULL);
3783 }
3784 EXPORT_SYMBOL(__kmalloc_node);
3785 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3786 #endif /* CONFIG_NUMA */
3788 /**
3789 * __do_kmalloc - allocate memory
3790 * @size: how many bytes of memory are required.
3791 * @flags: the type of memory to allocate (see kmalloc).
3792 * @caller: function caller for debug tracking of the caller
3793 */
3794 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3795 void *caller)
3796 {
3797 struct kmem_cache *cachep;
3798 void *ret;
3800 /* If you want to save a few bytes .text space: replace
3801 * __ with kmem_.
3802 * Then kmalloc uses the uninlined functions instead of the inline
3803 * functions.
3804 */
3805 cachep = __find_general_cachep(size, flags);
3806 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3807 return cachep;
3808 ret = __cache_alloc(cachep, flags, caller);
3810 trace_kmalloc((unsigned long) caller, ret,
3811 size, cachep->buffer_size, flags);
3813 return ret;
3814 }
3817 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3818 void *__kmalloc(size_t size, gfp_t flags)
3819 {
3820 return __do_kmalloc(size, flags, __builtin_return_address(0));
3821 }
3822 EXPORT_SYMBOL(__kmalloc);
3824 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3825 {
3826 return __do_kmalloc(size, flags, (void *)caller);
3827 }
3828 EXPORT_SYMBOL(__kmalloc_track_caller);
3830 #else
3831 void *__kmalloc(size_t size, gfp_t flags)
3832 {
3833 return __do_kmalloc(size, flags, NULL);
3834 }
3835 EXPORT_SYMBOL(__kmalloc);
3836 #endif
3838 /**
3839 * kmem_cache_free - Deallocate an object
3840 * @cachep: The cache the allocation was from.
3841 * @objp: The previously allocated object.
3842 *
3843 * Free an object which was previously allocated from this
3844 * cache.
3845 */
3846 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3847 {
3848 unsigned long flags;
3850 local_irq_save(flags);
3851 debug_check_no_locks_freed(objp, obj_size(cachep));
3852 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3853 debug_check_no_obj_freed(objp, obj_size(cachep));
3854 __cache_free(cachep, objp, __builtin_return_address(0));
3855 local_irq_restore(flags);
3857 trace_kmem_cache_free(_RET_IP_, objp);
3858 }
3859 EXPORT_SYMBOL(kmem_cache_free);
3861 /**
3862 * kfree - free previously allocated memory
3863 * @objp: pointer returned by kmalloc.
3864 *
3865 * If @objp is NULL, no operation is performed.
3866 *
3867 * Don't free memory not originally allocated by kmalloc()
3868 * or you will run into trouble.
3869 */
3870 void kfree(const void *objp)
3871 {
3872 struct kmem_cache *c;
3873 unsigned long flags;
3875 trace_kfree(_RET_IP_, objp);
3877 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3878 return;
3879 local_irq_save(flags);
3880 kfree_debugcheck(objp);
3881 c = virt_to_cache(objp);
3882 debug_check_no_locks_freed(objp, obj_size(c));
3883 debug_check_no_obj_freed(objp, obj_size(c));
3884 __cache_free(c, (void *)objp, __builtin_return_address(0));
3885 local_irq_restore(flags);
3886 }
3887 EXPORT_SYMBOL(kfree);
3889 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3890 {
3891 return obj_size(cachep);
3892 }
3893 EXPORT_SYMBOL(kmem_cache_size);
3895 /*
3896 * This initializes kmem_list3 or resizes various caches for all nodes.
3897 */
3898 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3899 {
3900 int node;
3901 struct kmem_list3 *l3;
3902 struct array_cache *new_shared;
3903 struct array_cache **new_alien = NULL;
3905 for_each_online_node(node) {
3907 if (use_alien_caches) {
3908 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3909 if (!new_alien)
3910 goto fail;
3911 }
3913 new_shared = NULL;
3914 if (cachep->shared) {
3915 new_shared = alloc_arraycache(node,
3916 cachep->shared*cachep->batchcount,
3917 0xbaadf00d, gfp);
3918 if (!new_shared) {
3919 free_alien_cache(new_alien);
3920 goto fail;
3921 }
3922 }
3924 l3 = cachep->nodelists[node];
3925 if (l3) {
3926 struct array_cache *shared = l3->shared;
3928 spin_lock_irq(&l3->list_lock);
3930 if (shared)
3931 free_block(cachep, shared->entry,
3932 shared->avail, node);
3934 l3->shared = new_shared;
3935 if (!l3->alien) {
3936 l3->alien = new_alien;
3937 new_alien = NULL;
3938 }
3939 l3->free_limit = (1 + nr_cpus_node(node)) *
3940 cachep->batchcount + cachep->num;
3941 spin_unlock_irq(&l3->list_lock);
3942 kfree(shared);
3943 free_alien_cache(new_alien);
3944 continue;
3945 }
3946 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3947 if (!l3) {
3948 free_alien_cache(new_alien);
3949 kfree(new_shared);
3950 goto fail;
3951 }
3953 kmem_list3_init(l3);
3954 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3955 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3956 l3->shared = new_shared;
3957 l3->alien = new_alien;
3958 l3->free_limit = (1 + nr_cpus_node(node)) *
3959 cachep->batchcount + cachep->num;
3960 cachep->nodelists[node] = l3;
3961 }
3962 return 0;
3964 fail:
3965 if (!cachep->next.next) {
3966 /* Cache is not active yet. Roll back what we did */
3967 node--;
3968 while (node >= 0) {
3969 if (cachep->nodelists[node]) {
3970 l3 = cachep->nodelists[node];
3972 kfree(l3->shared);
3973 free_alien_cache(l3->alien);
3974 kfree(l3);
3975 cachep->nodelists[node] = NULL;
3976 }
3977 node--;
3978 }
3979 }
3980 return -ENOMEM;
3981 }
3983 struct ccupdate_struct {
3984 struct kmem_cache *cachep;
3985 struct array_cache *new[0];
3986 };
3988 static void do_ccupdate_local(void *info)
3989 {
3990 struct ccupdate_struct *new = info;
3991 struct array_cache *old;
3993 check_irq_off();
3994 old = cpu_cache_get(new->cachep);
3996 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3997 new->new[smp_processor_id()] = old;
3998 }
4000 /* Always called with the cache_chain_mutex held */
4001 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4002 int batchcount, int shared, gfp_t gfp)
4003 {
4004 struct ccupdate_struct *new;
4005 int i;
4007 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4008 gfp);
4009 if (!new)
4010 return -ENOMEM;
4012 for_each_online_cpu(i) {
4013 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4014 batchcount, gfp);
4015 if (!new->new[i]) {
4016 for (i--; i >= 0; i--)
4017 kfree(new->new[i]);
4018 kfree(new);
4019 return -ENOMEM;
4020 }
4021 }
4022 new->cachep = cachep;
4024 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4026 check_irq_on();
4027 cachep->batchcount = batchcount;
4028 cachep->limit = limit;
4029 cachep->shared = shared;
4031 for_each_online_cpu(i) {
4032 struct array_cache *ccold = new->new[i];
4033 if (!ccold)
4034 continue;
4035 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4036 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4037 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4038 kfree(ccold);
4039 }
4040 kfree(new);
4041 return alloc_kmemlist(cachep, gfp);
4042 }
4044 /* Called with cache_chain_mutex held always */
4045 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4046 {
4047 int err;
4048 int limit, shared;
4050 /*
4051 * The head array serves three purposes:
4052 * - create a LIFO ordering, i.e. return objects that are cache-warm
4053 * - reduce the number of spinlock operations.
4054 * - reduce the number of linked list operations on the slab and
4055 * bufctl chains: array operations are cheaper.
4056 * The numbers are guessed, we should auto-tune as described by
4057 * Bonwick.
4058 */
4059 if (cachep->buffer_size > 131072)
4060 limit = 1;
4061 else if (cachep->buffer_size > PAGE_SIZE)
4062 limit = 8;
4063 else if (cachep->buffer_size > 1024)
4064 limit = 24;
4065 else if (cachep->buffer_size > 256)
4066 limit = 54;
4067 else
4068 limit = 120;
4070 /*
4071 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4072 * allocation behaviour: Most allocs on one cpu, most free operations
4073 * on another cpu. For these cases, an efficient object passing between
4074 * cpus is necessary. This is provided by a shared array. The array
4075 * replaces Bonwick's magazine layer.
4076 * On uniprocessor, it's functionally equivalent (but less efficient)
4077 * to a larger limit. Thus disabled by default.
4078 */
4079 shared = 0;
4080 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4081 shared = 8;
4083 #if DEBUG
4084 /*
4085 * With debugging enabled, large batchcount lead to excessively long
4086 * periods with disabled local interrupts. Limit the batchcount
4087 */
4088 if (limit > 32)
4089 limit = 32;
4090 #endif
4091 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4092 if (err)
4093 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4094 cachep->name, -err);
4095 return err;
4096 }
4098 /*
4099 * Drain an array if it contains any elements taking the l3 lock only if
4100 * necessary. Note that the l3 listlock also protects the array_cache
4101 * if drain_array() is used on the shared array.
4102 */
4103 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4104 struct array_cache *ac, int force, int node)
4105 {
4106 int tofree;
4108 if (!ac || !ac->avail)
4109 return;
4110 if (ac->touched && !force) {
4111 ac->touched = 0;
4112 } else {
4113 spin_lock_irq(&l3->list_lock);
4114 if (ac->avail) {
4115 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4116 if (tofree > ac->avail)
4117 tofree = (ac->avail + 1) / 2;
4118 free_block(cachep, ac->entry, tofree, node);
4119 ac->avail -= tofree;
4120 memmove(ac->entry, &(ac->entry[tofree]),
4121 sizeof(void *) * ac->avail);
4122 }
4123 spin_unlock_irq(&l3->list_lock);
4124 }
4125 }
4127 /**
4128 * cache_reap - Reclaim memory from caches.
4129 * @w: work descriptor
4130 *
4131 * Called from workqueue/eventd every few seconds.
4132 * Purpose:
4133 * - clear the per-cpu caches for this CPU.
4134 * - return freeable pages to the main free memory pool.
4135 *
4136 * If we cannot acquire the cache chain mutex then just give up - we'll try
4137 * again on the next iteration.
4138 */
4139 static void cache_reap(struct work_struct *w)
4140 {
4141 struct kmem_cache *searchp;
4142 struct kmem_list3 *l3;
4143 int node = numa_mem_id();
4144 struct delayed_work *work = to_delayed_work(w);
4146 if (!mutex_trylock(&cache_chain_mutex))
4147 /* Give up. Setup the next iteration. */
4148 goto out;
4150 list_for_each_entry(searchp, &cache_chain, next) {
4151 check_irq_on();
4153 /*
4154 * We only take the l3 lock if absolutely necessary and we
4155 * have established with reasonable certainty that
4156 * we can do some work if the lock was obtained.
4157 */
4158 l3 = searchp->nodelists[node];
4160 reap_alien(searchp, l3);
4162 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4164 /*
4165 * These are racy checks but it does not matter
4166 * if we skip one check or scan twice.
4167 */
4168 if (time_after(l3->next_reap, jiffies))
4169 goto next;
4171 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4173 drain_array(searchp, l3, l3->shared, 0, node);
4175 if (l3->free_touched)
4176 l3->free_touched = 0;
4177 else {
4178 int freed;
4180 freed = drain_freelist(searchp, l3, (l3->free_limit +
4181 5 * searchp->num - 1) / (5 * searchp->num));
4182 STATS_ADD_REAPED(searchp, freed);
4183 }
4184 next:
4185 cond_resched();
4186 }
4187 check_irq_on();
4188 mutex_unlock(&cache_chain_mutex);
4189 next_reap_node();
4190 out:
4191 /* Set up the next iteration */
4192 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4193 }
4195 #ifdef CONFIG_SLABINFO
4197 static void print_slabinfo_header(struct seq_file *m)
4198 {
4199 /*
4200 * Output format version, so at least we can change it
4201 * without _too_ many complaints.
4202 */
4203 #if STATS
4204 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4205 #else
4206 seq_puts(m, "slabinfo - version: 2.1\n");
4207 #endif
4208 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4209 "<objperslab> <pagesperslab>");
4210 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4211 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4212 #if STATS
4213 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4214 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4215 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4216 #endif
4217 seq_putc(m, '\n');
4218 }
4220 static void *s_start(struct seq_file *m, loff_t *pos)
4221 {
4222 loff_t n = *pos;
4224 mutex_lock(&cache_chain_mutex);
4225 if (!n)
4226 print_slabinfo_header(m);
4228 return seq_list_start(&cache_chain, *pos);
4229 }
4231 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4232 {
4233 return seq_list_next(p, &cache_chain, pos);
4234 }
4236 static void s_stop(struct seq_file *m, void *p)
4237 {
4238 mutex_unlock(&cache_chain_mutex);
4239 }
4241 static int s_show(struct seq_file *m, void *p)
4242 {
4243 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4244 struct slab *slabp;
4245 unsigned long active_objs;
4246 unsigned long num_objs;
4247 unsigned long active_slabs = 0;
4248 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4249 const char *name;
4250 char *error = NULL;
4251 int node;
4252 struct kmem_list3 *l3;
4254 active_objs = 0;
4255 num_slabs = 0;
4256 for_each_online_node(node) {
4257 l3 = cachep->nodelists[node];
4258 if (!l3)
4259 continue;
4261 check_irq_on();
4262 spin_lock_irq(&l3->list_lock);
4264 list_for_each_entry(slabp, &l3->slabs_full, list) {
4265 if (slabp->inuse != cachep->num && !error)
4266 error = "slabs_full accounting error";
4267 active_objs += cachep->num;
4268 active_slabs++;
4269 }
4270 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4271 if (slabp->inuse == cachep->num && !error)
4272 error = "slabs_partial inuse accounting error";
4273 if (!slabp->inuse && !error)
4274 error = "slabs_partial/inuse accounting error";
4275 active_objs += slabp->inuse;
4276 active_slabs++;
4277 }
4278 list_for_each_entry(slabp, &l3->slabs_free, list) {
4279 if (slabp->inuse && !error)
4280 error = "slabs_free/inuse accounting error";
4281 num_slabs++;
4282 }
4283 free_objects += l3->free_objects;
4284 if (l3->shared)
4285 shared_avail += l3->shared->avail;
4287 spin_unlock_irq(&l3->list_lock);
4288 }
4289 num_slabs += active_slabs;
4290 num_objs = num_slabs * cachep->num;
4291 if (num_objs - active_objs != free_objects && !error)
4292 error = "free_objects accounting error";
4294 name = cachep->name;
4295 if (error)
4296 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4298 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4299 name, active_objs, num_objs, cachep->buffer_size,
4300 cachep->num, (1 << cachep->gfporder));
4301 seq_printf(m, " : tunables %4u %4u %4u",
4302 cachep->limit, cachep->batchcount, cachep->shared);
4303 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4304 active_slabs, num_slabs, shared_avail);
4305 #if STATS
4306 { /* list3 stats */
4307 unsigned long high = cachep->high_mark;
4308 unsigned long allocs = cachep->num_allocations;
4309 unsigned long grown = cachep->grown;
4310 unsigned long reaped = cachep->reaped;
4311 unsigned long errors = cachep->errors;
4312 unsigned long max_freeable = cachep->max_freeable;
4313 unsigned long node_allocs = cachep->node_allocs;
4314 unsigned long node_frees = cachep->node_frees;
4315 unsigned long overflows = cachep->node_overflow;
4317 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4318 "%4lu %4lu %4lu %4lu %4lu",
4319 allocs, high, grown,
4320 reaped, errors, max_freeable, node_allocs,
4321 node_frees, overflows);
4322 }
4323 /* cpu stats */
4324 {
4325 unsigned long allochit = atomic_read(&cachep->allochit);
4326 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4327 unsigned long freehit = atomic_read(&cachep->freehit);
4328 unsigned long freemiss = atomic_read(&cachep->freemiss);
4330 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4331 allochit, allocmiss, freehit, freemiss);
4332 }
4333 #endif
4334 seq_putc(m, '\n');
4335 return 0;
4336 }
4338 /*
4339 * slabinfo_op - iterator that generates /proc/slabinfo
4340 *
4341 * Output layout:
4342 * cache-name
4343 * num-active-objs
4344 * total-objs
4345 * object size
4346 * num-active-slabs
4347 * total-slabs
4348 * num-pages-per-slab
4349 * + further values on SMP and with statistics enabled
4350 */
4352 static const struct seq_operations slabinfo_op = {
4353 .start = s_start,
4354 .next = s_next,
4355 .stop = s_stop,
4356 .show = s_show,
4357 };
4359 #define MAX_SLABINFO_WRITE 128
4360 /**
4361 * slabinfo_write - Tuning for the slab allocator
4362 * @file: unused
4363 * @buffer: user buffer
4364 * @count: data length
4365 * @ppos: unused
4366 */
4367 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4368 size_t count, loff_t *ppos)
4369 {
4370 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4371 int limit, batchcount, shared, res;
4372 struct kmem_cache *cachep;
4374 if (count > MAX_SLABINFO_WRITE)
4375 return -EINVAL;
4376 if (copy_from_user(&kbuf, buffer, count))
4377 return -EFAULT;
4378 kbuf[MAX_SLABINFO_WRITE] = '\0';
4380 tmp = strchr(kbuf, ' ');
4381 if (!tmp)
4382 return -EINVAL;
4383 *tmp = '\0';
4384 tmp++;
4385 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4386 return -EINVAL;
4388 /* Find the cache in the chain of caches. */
4389 mutex_lock(&cache_chain_mutex);
4390 res = -EINVAL;
4391 list_for_each_entry(cachep, &cache_chain, next) {
4392 if (!strcmp(cachep->name, kbuf)) {
4393 if (limit < 1 || batchcount < 1 ||
4394 batchcount > limit || shared < 0) {
4395 res = 0;
4396 } else {
4397 res = do_tune_cpucache(cachep, limit,
4398 batchcount, shared,
4399 GFP_KERNEL);
4400 }
4401 break;
4402 }
4403 }
4404 mutex_unlock(&cache_chain_mutex);
4405 if (res >= 0)
4406 res = count;
4407 return res;
4408 }
4410 static int slabinfo_open(struct inode *inode, struct file *file)
4411 {
4412 return seq_open(file, &slabinfo_op);
4413 }
4415 static const struct file_operations proc_slabinfo_operations = {
4416 .open = slabinfo_open,
4417 .read = seq_read,
4418 .write = slabinfo_write,
4419 .llseek = seq_lseek,
4420 .release = seq_release,
4421 };
4423 #ifdef CONFIG_DEBUG_SLAB_LEAK
4425 static void *leaks_start(struct seq_file *m, loff_t *pos)
4426 {
4427 mutex_lock(&cache_chain_mutex);
4428 return seq_list_start(&cache_chain, *pos);
4429 }
4431 static inline int add_caller(unsigned long *n, unsigned long v)
4432 {
4433 unsigned long *p;
4434 int l;
4435 if (!v)
4436 return 1;
4437 l = n[1];
4438 p = n + 2;
4439 while (l) {
4440 int i = l/2;
4441 unsigned long *q = p + 2 * i;
4442 if (*q == v) {
4443 q[1]++;
4444 return 1;
4445 }
4446 if (*q > v) {
4447 l = i;
4448 } else {
4449 p = q + 2;
4450 l -= i + 1;
4451 }
4452 }
4453 if (++n[1] == n[0])
4454 return 0;
4455 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4456 p[0] = v;
4457 p[1] = 1;
4458 return 1;
4459 }
4461 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4462 {
4463 void *p;
4464 int i;
4465 if (n[0] == n[1])
4466 return;
4467 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4468 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4469 continue;
4470 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4471 return;
4472 }
4473 }
4475 static void show_symbol(struct seq_file *m, unsigned long address)
4476 {
4477 #ifdef CONFIG_KALLSYMS
4478 unsigned long offset, size;
4479 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4481 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4482 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4483 if (modname[0])
4484 seq_printf(m, " [%s]", modname);
4485 return;
4486 }
4487 #endif
4488 seq_printf(m, "%p", (void *)address);
4489 }
4491 static int leaks_show(struct seq_file *m, void *p)
4492 {
4493 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4494 struct slab *slabp;
4495 struct kmem_list3 *l3;
4496 const char *name;
4497 unsigned long *n = m->private;
4498 int node;
4499 int i;
4501 if (!(cachep->flags & SLAB_STORE_USER))
4502 return 0;
4503 if (!(cachep->flags & SLAB_RED_ZONE))
4504 return 0;
4506 /* OK, we can do it */
4508 n[1] = 0;
4510 for_each_online_node(node) {
4511 l3 = cachep->nodelists[node];
4512 if (!l3)
4513 continue;
4515 check_irq_on();
4516 spin_lock_irq(&l3->list_lock);
4518 list_for_each_entry(slabp, &l3->slabs_full, list)
4519 handle_slab(n, cachep, slabp);
4520 list_for_each_entry(slabp, &l3->slabs_partial, list)
4521 handle_slab(n, cachep, slabp);
4522 spin_unlock_irq(&l3->list_lock);
4523 }
4524 name = cachep->name;
4525 if (n[0] == n[1]) {
4526 /* Increase the buffer size */
4527 mutex_unlock(&cache_chain_mutex);
4528 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4529 if (!m->private) {
4530 /* Too bad, we are really out */
4531 m->private = n;
4532 mutex_lock(&cache_chain_mutex);
4533 return -ENOMEM;
4534 }
4535 *(unsigned long *)m->private = n[0] * 2;
4536 kfree(n);
4537 mutex_lock(&cache_chain_mutex);
4538 /* Now make sure this entry will be retried */
4539 m->count = m->size;
4540 return 0;
4541 }
4542 for (i = 0; i < n[1]; i++) {
4543 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4544 show_symbol(m, n[2*i+2]);
4545 seq_putc(m, '\n');
4546 }
4548 return 0;
4549 }
4551 static const struct seq_operations slabstats_op = {
4552 .start = leaks_start,
4553 .next = s_next,
4554 .stop = s_stop,
4555 .show = leaks_show,
4556 };
4558 static int slabstats_open(struct inode *inode, struct file *file)
4559 {
4560 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4561 int ret = -ENOMEM;
4562 if (n) {
4563 ret = seq_open(file, &slabstats_op);
4564 if (!ret) {
4565 struct seq_file *m = file->private_data;
4566 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4567 m->private = n;
4568 n = NULL;
4569 }
4570 kfree(n);
4571 }
4572 return ret;
4573 }
4575 static const struct file_operations proc_slabstats_operations = {
4576 .open = slabstats_open,
4577 .read = seq_read,
4578 .llseek = seq_lseek,
4579 .release = seq_release_private,
4580 };
4581 #endif
4583 static int __init slab_proc_init(void)
4584 {
4585 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4586 #ifdef CONFIG_DEBUG_SLAB_LEAK
4587 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4588 #endif
4589 return 0;
4590 }
4591 module_init(slab_proc_init);
4592 #endif
4594 /**
4595 * ksize - get the actual amount of memory allocated for a given object
4596 * @objp: Pointer to the object
4597 *
4598 * kmalloc may internally round up allocations and return more memory
4599 * than requested. ksize() can be used to determine the actual amount of
4600 * memory allocated. The caller may use this additional memory, even though
4601 * a smaller amount of memory was initially specified with the kmalloc call.
4602 * The caller must guarantee that objp points to a valid object previously
4603 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4604 * must not be freed during the duration of the call.
4605 */
4606 size_t ksize(const void *objp)
4607 {
4608 BUG_ON(!objp);
4609 if (unlikely(objp == ZERO_SIZE_PTR))
4610 return 0;
4612 return obj_size(virt_to_cache(objp));
4613 }
4614 EXPORT_SYMBOL(ksize);