slub: add taint flag outputting to debug paths
[deliverable/linux.git] / mm / slab.c
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 */
88
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>
119
120 #include <asm/cacheflush.h>
121 #include <asm/tlbflush.h>
122 #include <asm/page.h>
123
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 */
133
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
143
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))
147
148 #ifndef ARCH_KMALLOC_FLAGS
149 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150 #endif
151
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
168
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 */
187
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)
193
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 };
213
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 };
234
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 };
259
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 };
269
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 };
286
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)
295
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);
302
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);
310
311 if (__builtin_constant_p(size)) {
312 int i = 0;
313
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 }
326
327 static int slab_early_init = 1;
328
329 #define INDEX_AC index_of(sizeof(struct arraycache_init))
330 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
331
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 }
344
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)
350
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)
357
358 #define CFLGS_OFF_SLAB (0x80000000UL)
359 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
360
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)
371
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
413
414 #if DEBUG
415
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 }
433
434 static int obj_size(struct kmem_cache *cachep)
435 {
436 return cachep->obj_size;
437 }
438
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 }
445
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 }
456
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 }
462
463 #else
464
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;})
470
471 #endif
472
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
480
481 /*
482 * Do not go above this order unless 0 objects fit into the slab or
483 * overridden on the command line.
484 */
485 #define SLAB_MAX_ORDER_HI 1
486 #define SLAB_MAX_ORDER_LO 0
487 static int slab_max_order = SLAB_MAX_ORDER_LO;
488 static bool slab_max_order_set __initdata;
489
490 /*
491 * Functions for storing/retrieving the cachep and or slab from the page
492 * allocator. These are used to find the slab an obj belongs to. With kfree(),
493 * these are used to find the cache which an obj belongs to.
494 */
495 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
496 {
497 page->lru.next = (struct list_head *)cache;
498 }
499
500 static inline struct kmem_cache *page_get_cache(struct page *page)
501 {
502 page = compound_head(page);
503 BUG_ON(!PageSlab(page));
504 return (struct kmem_cache *)page->lru.next;
505 }
506
507 static inline void page_set_slab(struct page *page, struct slab *slab)
508 {
509 page->lru.prev = (struct list_head *)slab;
510 }
511
512 static inline struct slab *page_get_slab(struct page *page)
513 {
514 BUG_ON(!PageSlab(page));
515 return (struct slab *)page->lru.prev;
516 }
517
518 static inline struct kmem_cache *virt_to_cache(const void *obj)
519 {
520 struct page *page = virt_to_head_page(obj);
521 return page_get_cache(page);
522 }
523
524 static inline struct slab *virt_to_slab(const void *obj)
525 {
526 struct page *page = virt_to_head_page(obj);
527 return page_get_slab(page);
528 }
529
530 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
531 unsigned int idx)
532 {
533 return slab->s_mem + cache->buffer_size * idx;
534 }
535
536 /*
537 * We want to avoid an expensive divide : (offset / cache->buffer_size)
538 * Using the fact that buffer_size is a constant for a particular cache,
539 * we can replace (offset / cache->buffer_size) by
540 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
541 */
542 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
543 const struct slab *slab, void *obj)
544 {
545 u32 offset = (obj - slab->s_mem);
546 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
547 }
548
549 /*
550 * These are the default caches for kmalloc. Custom caches can have other sizes.
551 */
552 struct cache_sizes malloc_sizes[] = {
553 #define CACHE(x) { .cs_size = (x) },
554 #include <linux/kmalloc_sizes.h>
555 CACHE(ULONG_MAX)
556 #undef CACHE
557 };
558 EXPORT_SYMBOL(malloc_sizes);
559
560 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
561 struct cache_names {
562 char *name;
563 char *name_dma;
564 };
565
566 static struct cache_names __initdata cache_names[] = {
567 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
568 #include <linux/kmalloc_sizes.h>
569 {NULL,}
570 #undef CACHE
571 };
572
573 static struct arraycache_init initarray_cache __initdata =
574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
575 static struct arraycache_init initarray_generic =
576 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
577
578 /* internal cache of cache description objs */
579 static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
580 static struct kmem_cache cache_cache = {
581 .nodelists = cache_cache_nodelists,
582 .batchcount = 1,
583 .limit = BOOT_CPUCACHE_ENTRIES,
584 .shared = 1,
585 .buffer_size = sizeof(struct kmem_cache),
586 .name = "kmem_cache",
587 };
588
589 #define BAD_ALIEN_MAGIC 0x01020304ul
590
591 /*
592 * chicken and egg problem: delay the per-cpu array allocation
593 * until the general caches are up.
594 */
595 static enum {
596 NONE,
597 PARTIAL_AC,
598 PARTIAL_L3,
599 EARLY,
600 FULL
601 } g_cpucache_up;
602
603 /*
604 * used by boot code to determine if it can use slab based allocator
605 */
606 int slab_is_available(void)
607 {
608 return g_cpucache_up >= EARLY;
609 }
610
611 #ifdef CONFIG_LOCKDEP
612
613 /*
614 * Slab sometimes uses the kmalloc slabs to store the slab headers
615 * for other slabs "off slab".
616 * The locking for this is tricky in that it nests within the locks
617 * of all other slabs in a few places; to deal with this special
618 * locking we put on-slab caches into a separate lock-class.
619 *
620 * We set lock class for alien array caches which are up during init.
621 * The lock annotation will be lost if all cpus of a node goes down and
622 * then comes back up during hotplug
623 */
624 static struct lock_class_key on_slab_l3_key;
625 static struct lock_class_key on_slab_alc_key;
626
627 static struct lock_class_key debugobj_l3_key;
628 static struct lock_class_key debugobj_alc_key;
629
630 static void slab_set_lock_classes(struct kmem_cache *cachep,
631 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
632 int q)
633 {
634 struct array_cache **alc;
635 struct kmem_list3 *l3;
636 int r;
637
638 l3 = cachep->nodelists[q];
639 if (!l3)
640 return;
641
642 lockdep_set_class(&l3->list_lock, l3_key);
643 alc = l3->alien;
644 /*
645 * FIXME: This check for BAD_ALIEN_MAGIC
646 * should go away when common slab code is taught to
647 * work even without alien caches.
648 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
649 * for alloc_alien_cache,
650 */
651 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
652 return;
653 for_each_node(r) {
654 if (alc[r])
655 lockdep_set_class(&alc[r]->lock, alc_key);
656 }
657 }
658
659 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
660 {
661 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
662 }
663
664 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
665 {
666 int node;
667
668 for_each_online_node(node)
669 slab_set_debugobj_lock_classes_node(cachep, node);
670 }
671
672 static void init_node_lock_keys(int q)
673 {
674 struct cache_sizes *s = malloc_sizes;
675
676 if (g_cpucache_up != FULL)
677 return;
678
679 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
680 struct kmem_list3 *l3;
681
682 l3 = s->cs_cachep->nodelists[q];
683 if (!l3 || OFF_SLAB(s->cs_cachep))
684 continue;
685
686 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
687 &on_slab_alc_key, q);
688 }
689 }
690
691 static inline void init_lock_keys(void)
692 {
693 int node;
694
695 for_each_node(node)
696 init_node_lock_keys(node);
697 }
698 #else
699 static void init_node_lock_keys(int q)
700 {
701 }
702
703 static inline void init_lock_keys(void)
704 {
705 }
706
707 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
708 {
709 }
710
711 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
712 {
713 }
714 #endif
715
716 /*
717 * Guard access to the cache-chain.
718 */
719 static DEFINE_MUTEX(cache_chain_mutex);
720 static struct list_head cache_chain;
721
722 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
723
724 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
725 {
726 return cachep->array[smp_processor_id()];
727 }
728
729 static inline struct kmem_cache *__find_general_cachep(size_t size,
730 gfp_t gfpflags)
731 {
732 struct cache_sizes *csizep = malloc_sizes;
733
734 #if DEBUG
735 /* This happens if someone tries to call
736 * kmem_cache_create(), or __kmalloc(), before
737 * the generic caches are initialized.
738 */
739 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
740 #endif
741 if (!size)
742 return ZERO_SIZE_PTR;
743
744 while (size > csizep->cs_size)
745 csizep++;
746
747 /*
748 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
749 * has cs_{dma,}cachep==NULL. Thus no special case
750 * for large kmalloc calls required.
751 */
752 #ifdef CONFIG_ZONE_DMA
753 if (unlikely(gfpflags & GFP_DMA))
754 return csizep->cs_dmacachep;
755 #endif
756 return csizep->cs_cachep;
757 }
758
759 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
760 {
761 return __find_general_cachep(size, gfpflags);
762 }
763
764 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
765 {
766 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
767 }
768
769 /*
770 * Calculate the number of objects and left-over bytes for a given buffer size.
771 */
772 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
773 size_t align, int flags, size_t *left_over,
774 unsigned int *num)
775 {
776 int nr_objs;
777 size_t mgmt_size;
778 size_t slab_size = PAGE_SIZE << gfporder;
779
780 /*
781 * The slab management structure can be either off the slab or
782 * on it. For the latter case, the memory allocated for a
783 * slab is used for:
784 *
785 * - The struct slab
786 * - One kmem_bufctl_t for each object
787 * - Padding to respect alignment of @align
788 * - @buffer_size bytes for each object
789 *
790 * If the slab management structure is off the slab, then the
791 * alignment will already be calculated into the size. Because
792 * the slabs are all pages aligned, the objects will be at the
793 * correct alignment when allocated.
794 */
795 if (flags & CFLGS_OFF_SLAB) {
796 mgmt_size = 0;
797 nr_objs = slab_size / buffer_size;
798
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
801 } else {
802 /*
803 * Ignore padding for the initial guess. The padding
804 * is at most @align-1 bytes, and @buffer_size is at
805 * least @align. In the worst case, this result will
806 * be one greater than the number of objects that fit
807 * into the memory allocation when taking the padding
808 * into account.
809 */
810 nr_objs = (slab_size - sizeof(struct slab)) /
811 (buffer_size + sizeof(kmem_bufctl_t));
812
813 /*
814 * This calculated number will be either the right
815 * amount, or one greater than what we want.
816 */
817 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
818 > slab_size)
819 nr_objs--;
820
821 if (nr_objs > SLAB_LIMIT)
822 nr_objs = SLAB_LIMIT;
823
824 mgmt_size = slab_mgmt_size(nr_objs, align);
825 }
826 *num = nr_objs;
827 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
828 }
829
830 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
831
832 static void __slab_error(const char *function, struct kmem_cache *cachep,
833 char *msg)
834 {
835 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
836 function, cachep->name, msg);
837 dump_stack();
838 }
839
840 /*
841 * By default on NUMA we use alien caches to stage the freeing of
842 * objects allocated from other nodes. This causes massive memory
843 * inefficiencies when using fake NUMA setup to split memory into a
844 * large number of small nodes, so it can be disabled on the command
845 * line
846 */
847
848 static int use_alien_caches __read_mostly = 1;
849 static int __init noaliencache_setup(char *s)
850 {
851 use_alien_caches = 0;
852 return 1;
853 }
854 __setup("noaliencache", noaliencache_setup);
855
856 static int __init slab_max_order_setup(char *str)
857 {
858 get_option(&str, &slab_max_order);
859 slab_max_order = slab_max_order < 0 ? 0 :
860 min(slab_max_order, MAX_ORDER - 1);
861 slab_max_order_set = true;
862
863 return 1;
864 }
865 __setup("slab_max_order=", slab_max_order_setup);
866
867 #ifdef CONFIG_NUMA
868 /*
869 * Special reaping functions for NUMA systems called from cache_reap().
870 * These take care of doing round robin flushing of alien caches (containing
871 * objects freed on different nodes from which they were allocated) and the
872 * flushing of remote pcps by calling drain_node_pages.
873 */
874 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
875
876 static void init_reap_node(int cpu)
877 {
878 int node;
879
880 node = next_node(cpu_to_mem(cpu), node_online_map);
881 if (node == MAX_NUMNODES)
882 node = first_node(node_online_map);
883
884 per_cpu(slab_reap_node, cpu) = node;
885 }
886
887 static void next_reap_node(void)
888 {
889 int node = __this_cpu_read(slab_reap_node);
890
891 node = next_node(node, node_online_map);
892 if (unlikely(node >= MAX_NUMNODES))
893 node = first_node(node_online_map);
894 __this_cpu_write(slab_reap_node, node);
895 }
896
897 #else
898 #define init_reap_node(cpu) do { } while (0)
899 #define next_reap_node(void) do { } while (0)
900 #endif
901
902 /*
903 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
904 * via the workqueue/eventd.
905 * Add the CPU number into the expiration time to minimize the possibility of
906 * the CPUs getting into lockstep and contending for the global cache chain
907 * lock.
908 */
909 static void __cpuinit start_cpu_timer(int cpu)
910 {
911 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
912
913 /*
914 * When this gets called from do_initcalls via cpucache_init(),
915 * init_workqueues() has already run, so keventd will be setup
916 * at that time.
917 */
918 if (keventd_up() && reap_work->work.func == NULL) {
919 init_reap_node(cpu);
920 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
921 schedule_delayed_work_on(cpu, reap_work,
922 __round_jiffies_relative(HZ, cpu));
923 }
924 }
925
926 static struct array_cache *alloc_arraycache(int node, int entries,
927 int batchcount, gfp_t gfp)
928 {
929 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
930 struct array_cache *nc = NULL;
931
932 nc = kmalloc_node(memsize, gfp, node);
933 /*
934 * The array_cache structures contain pointers to free object.
935 * However, when such objects are allocated or transferred to another
936 * cache the pointers are not cleared and they could be counted as
937 * valid references during a kmemleak scan. Therefore, kmemleak must
938 * not scan such objects.
939 */
940 kmemleak_no_scan(nc);
941 if (nc) {
942 nc->avail = 0;
943 nc->limit = entries;
944 nc->batchcount = batchcount;
945 nc->touched = 0;
946 spin_lock_init(&nc->lock);
947 }
948 return nc;
949 }
950
951 /*
952 * Transfer objects in one arraycache to another.
953 * Locking must be handled by the caller.
954 *
955 * Return the number of entries transferred.
956 */
957 static int transfer_objects(struct array_cache *to,
958 struct array_cache *from, unsigned int max)
959 {
960 /* Figure out how many entries to transfer */
961 int nr = min3(from->avail, max, to->limit - to->avail);
962
963 if (!nr)
964 return 0;
965
966 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
967 sizeof(void *) *nr);
968
969 from->avail -= nr;
970 to->avail += nr;
971 return nr;
972 }
973
974 #ifndef CONFIG_NUMA
975
976 #define drain_alien_cache(cachep, alien) do { } while (0)
977 #define reap_alien(cachep, l3) do { } while (0)
978
979 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
980 {
981 return (struct array_cache **)BAD_ALIEN_MAGIC;
982 }
983
984 static inline void free_alien_cache(struct array_cache **ac_ptr)
985 {
986 }
987
988 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
989 {
990 return 0;
991 }
992
993 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
994 gfp_t flags)
995 {
996 return NULL;
997 }
998
999 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1000 gfp_t flags, int nodeid)
1001 {
1002 return NULL;
1003 }
1004
1005 #else /* CONFIG_NUMA */
1006
1007 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1008 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1009
1010 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1011 {
1012 struct array_cache **ac_ptr;
1013 int memsize = sizeof(void *) * nr_node_ids;
1014 int i;
1015
1016 if (limit > 1)
1017 limit = 12;
1018 ac_ptr = kzalloc_node(memsize, gfp, node);
1019 if (ac_ptr) {
1020 for_each_node(i) {
1021 if (i == node || !node_online(i))
1022 continue;
1023 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1024 if (!ac_ptr[i]) {
1025 for (i--; i >= 0; i--)
1026 kfree(ac_ptr[i]);
1027 kfree(ac_ptr);
1028 return NULL;
1029 }
1030 }
1031 }
1032 return ac_ptr;
1033 }
1034
1035 static void free_alien_cache(struct array_cache **ac_ptr)
1036 {
1037 int i;
1038
1039 if (!ac_ptr)
1040 return;
1041 for_each_node(i)
1042 kfree(ac_ptr[i]);
1043 kfree(ac_ptr);
1044 }
1045
1046 static void __drain_alien_cache(struct kmem_cache *cachep,
1047 struct array_cache *ac, int node)
1048 {
1049 struct kmem_list3 *rl3 = cachep->nodelists[node];
1050
1051 if (ac->avail) {
1052 spin_lock(&rl3->list_lock);
1053 /*
1054 * Stuff objects into the remote nodes shared array first.
1055 * That way we could avoid the overhead of putting the objects
1056 * into the free lists and getting them back later.
1057 */
1058 if (rl3->shared)
1059 transfer_objects(rl3->shared, ac, ac->limit);
1060
1061 free_block(cachep, ac->entry, ac->avail, node);
1062 ac->avail = 0;
1063 spin_unlock(&rl3->list_lock);
1064 }
1065 }
1066
1067 /*
1068 * Called from cache_reap() to regularly drain alien caches round robin.
1069 */
1070 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1071 {
1072 int node = __this_cpu_read(slab_reap_node);
1073
1074 if (l3->alien) {
1075 struct array_cache *ac = l3->alien[node];
1076
1077 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1078 __drain_alien_cache(cachep, ac, node);
1079 spin_unlock_irq(&ac->lock);
1080 }
1081 }
1082 }
1083
1084 static void drain_alien_cache(struct kmem_cache *cachep,
1085 struct array_cache **alien)
1086 {
1087 int i = 0;
1088 struct array_cache *ac;
1089 unsigned long flags;
1090
1091 for_each_online_node(i) {
1092 ac = alien[i];
1093 if (ac) {
1094 spin_lock_irqsave(&ac->lock, flags);
1095 __drain_alien_cache(cachep, ac, i);
1096 spin_unlock_irqrestore(&ac->lock, flags);
1097 }
1098 }
1099 }
1100
1101 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1102 {
1103 struct slab *slabp = virt_to_slab(objp);
1104 int nodeid = slabp->nodeid;
1105 struct kmem_list3 *l3;
1106 struct array_cache *alien = NULL;
1107 int node;
1108
1109 node = numa_mem_id();
1110
1111 /*
1112 * Make sure we are not freeing a object from another node to the array
1113 * cache on this cpu.
1114 */
1115 if (likely(slabp->nodeid == node))
1116 return 0;
1117
1118 l3 = cachep->nodelists[node];
1119 STATS_INC_NODEFREES(cachep);
1120 if (l3->alien && l3->alien[nodeid]) {
1121 alien = l3->alien[nodeid];
1122 spin_lock(&alien->lock);
1123 if (unlikely(alien->avail == alien->limit)) {
1124 STATS_INC_ACOVERFLOW(cachep);
1125 __drain_alien_cache(cachep, alien, nodeid);
1126 }
1127 alien->entry[alien->avail++] = objp;
1128 spin_unlock(&alien->lock);
1129 } else {
1130 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1131 free_block(cachep, &objp, 1, nodeid);
1132 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1133 }
1134 return 1;
1135 }
1136 #endif
1137
1138 /*
1139 * Allocates and initializes nodelists for a node on each slab cache, used for
1140 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1141 * will be allocated off-node since memory is not yet online for the new node.
1142 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1143 * already in use.
1144 *
1145 * Must hold cache_chain_mutex.
1146 */
1147 static int init_cache_nodelists_node(int node)
1148 {
1149 struct kmem_cache *cachep;
1150 struct kmem_list3 *l3;
1151 const int memsize = sizeof(struct kmem_list3);
1152
1153 list_for_each_entry(cachep, &cache_chain, next) {
1154 /*
1155 * Set up the size64 kmemlist for cpu before we can
1156 * begin anything. Make sure some other cpu on this
1157 * node has not already allocated this
1158 */
1159 if (!cachep->nodelists[node]) {
1160 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1161 if (!l3)
1162 return -ENOMEM;
1163 kmem_list3_init(l3);
1164 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1165 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1166
1167 /*
1168 * The l3s don't come and go as CPUs come and
1169 * go. cache_chain_mutex is sufficient
1170 * protection here.
1171 */
1172 cachep->nodelists[node] = l3;
1173 }
1174
1175 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1176 cachep->nodelists[node]->free_limit =
1177 (1 + nr_cpus_node(node)) *
1178 cachep->batchcount + cachep->num;
1179 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1180 }
1181 return 0;
1182 }
1183
1184 static void __cpuinit cpuup_canceled(long cpu)
1185 {
1186 struct kmem_cache *cachep;
1187 struct kmem_list3 *l3 = NULL;
1188 int node = cpu_to_mem(cpu);
1189 const struct cpumask *mask = cpumask_of_node(node);
1190
1191 list_for_each_entry(cachep, &cache_chain, next) {
1192 struct array_cache *nc;
1193 struct array_cache *shared;
1194 struct array_cache **alien;
1195
1196 /* cpu is dead; no one can alloc from it. */
1197 nc = cachep->array[cpu];
1198 cachep->array[cpu] = NULL;
1199 l3 = cachep->nodelists[node];
1200
1201 if (!l3)
1202 goto free_array_cache;
1203
1204 spin_lock_irq(&l3->list_lock);
1205
1206 /* Free limit for this kmem_list3 */
1207 l3->free_limit -= cachep->batchcount;
1208 if (nc)
1209 free_block(cachep, nc->entry, nc->avail, node);
1210
1211 if (!cpumask_empty(mask)) {
1212 spin_unlock_irq(&l3->list_lock);
1213 goto free_array_cache;
1214 }
1215
1216 shared = l3->shared;
1217 if (shared) {
1218 free_block(cachep, shared->entry,
1219 shared->avail, node);
1220 l3->shared = NULL;
1221 }
1222
1223 alien = l3->alien;
1224 l3->alien = NULL;
1225
1226 spin_unlock_irq(&l3->list_lock);
1227
1228 kfree(shared);
1229 if (alien) {
1230 drain_alien_cache(cachep, alien);
1231 free_alien_cache(alien);
1232 }
1233 free_array_cache:
1234 kfree(nc);
1235 }
1236 /*
1237 * In the previous loop, all the objects were freed to
1238 * the respective cache's slabs, now we can go ahead and
1239 * shrink each nodelist to its limit.
1240 */
1241 list_for_each_entry(cachep, &cache_chain, next) {
1242 l3 = cachep->nodelists[node];
1243 if (!l3)
1244 continue;
1245 drain_freelist(cachep, l3, l3->free_objects);
1246 }
1247 }
1248
1249 static int __cpuinit cpuup_prepare(long cpu)
1250 {
1251 struct kmem_cache *cachep;
1252 struct kmem_list3 *l3 = NULL;
1253 int node = cpu_to_mem(cpu);
1254 int err;
1255
1256 /*
1257 * We need to do this right in the beginning since
1258 * alloc_arraycache's are going to use this list.
1259 * kmalloc_node allows us to add the slab to the right
1260 * kmem_list3 and not this cpu's kmem_list3
1261 */
1262 err = init_cache_nodelists_node(node);
1263 if (err < 0)
1264 goto bad;
1265
1266 /*
1267 * Now we can go ahead with allocating the shared arrays and
1268 * array caches
1269 */
1270 list_for_each_entry(cachep, &cache_chain, next) {
1271 struct array_cache *nc;
1272 struct array_cache *shared = NULL;
1273 struct array_cache **alien = NULL;
1274
1275 nc = alloc_arraycache(node, cachep->limit,
1276 cachep->batchcount, GFP_KERNEL);
1277 if (!nc)
1278 goto bad;
1279 if (cachep->shared) {
1280 shared = alloc_arraycache(node,
1281 cachep->shared * cachep->batchcount,
1282 0xbaadf00d, GFP_KERNEL);
1283 if (!shared) {
1284 kfree(nc);
1285 goto bad;
1286 }
1287 }
1288 if (use_alien_caches) {
1289 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1290 if (!alien) {
1291 kfree(shared);
1292 kfree(nc);
1293 goto bad;
1294 }
1295 }
1296 cachep->array[cpu] = nc;
1297 l3 = cachep->nodelists[node];
1298 BUG_ON(!l3);
1299
1300 spin_lock_irq(&l3->list_lock);
1301 if (!l3->shared) {
1302 /*
1303 * We are serialised from CPU_DEAD or
1304 * CPU_UP_CANCELLED by the cpucontrol lock
1305 */
1306 l3->shared = shared;
1307 shared = NULL;
1308 }
1309 #ifdef CONFIG_NUMA
1310 if (!l3->alien) {
1311 l3->alien = alien;
1312 alien = NULL;
1313 }
1314 #endif
1315 spin_unlock_irq(&l3->list_lock);
1316 kfree(shared);
1317 free_alien_cache(alien);
1318 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1319 slab_set_debugobj_lock_classes_node(cachep, node);
1320 }
1321 init_node_lock_keys(node);
1322
1323 return 0;
1324 bad:
1325 cpuup_canceled(cpu);
1326 return -ENOMEM;
1327 }
1328
1329 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1330 unsigned long action, void *hcpu)
1331 {
1332 long cpu = (long)hcpu;
1333 int err = 0;
1334
1335 switch (action) {
1336 case CPU_UP_PREPARE:
1337 case CPU_UP_PREPARE_FROZEN:
1338 mutex_lock(&cache_chain_mutex);
1339 err = cpuup_prepare(cpu);
1340 mutex_unlock(&cache_chain_mutex);
1341 break;
1342 case CPU_ONLINE:
1343 case CPU_ONLINE_FROZEN:
1344 start_cpu_timer(cpu);
1345 break;
1346 #ifdef CONFIG_HOTPLUG_CPU
1347 case CPU_DOWN_PREPARE:
1348 case CPU_DOWN_PREPARE_FROZEN:
1349 /*
1350 * Shutdown cache reaper. Note that the cache_chain_mutex is
1351 * held so that if cache_reap() is invoked it cannot do
1352 * anything expensive but will only modify reap_work
1353 * and reschedule the timer.
1354 */
1355 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1356 /* Now the cache_reaper is guaranteed to be not running. */
1357 per_cpu(slab_reap_work, cpu).work.func = NULL;
1358 break;
1359 case CPU_DOWN_FAILED:
1360 case CPU_DOWN_FAILED_FROZEN:
1361 start_cpu_timer(cpu);
1362 break;
1363 case CPU_DEAD:
1364 case CPU_DEAD_FROZEN:
1365 /*
1366 * Even if all the cpus of a node are down, we don't free the
1367 * kmem_list3 of any cache. This to avoid a race between
1368 * cpu_down, and a kmalloc allocation from another cpu for
1369 * memory from the node of the cpu going down. The list3
1370 * structure is usually allocated from kmem_cache_create() and
1371 * gets destroyed at kmem_cache_destroy().
1372 */
1373 /* fall through */
1374 #endif
1375 case CPU_UP_CANCELED:
1376 case CPU_UP_CANCELED_FROZEN:
1377 mutex_lock(&cache_chain_mutex);
1378 cpuup_canceled(cpu);
1379 mutex_unlock(&cache_chain_mutex);
1380 break;
1381 }
1382 return notifier_from_errno(err);
1383 }
1384
1385 static struct notifier_block __cpuinitdata cpucache_notifier = {
1386 &cpuup_callback, NULL, 0
1387 };
1388
1389 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1390 /*
1391 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1392 * Returns -EBUSY if all objects cannot be drained so that the node is not
1393 * removed.
1394 *
1395 * Must hold cache_chain_mutex.
1396 */
1397 static int __meminit drain_cache_nodelists_node(int node)
1398 {
1399 struct kmem_cache *cachep;
1400 int ret = 0;
1401
1402 list_for_each_entry(cachep, &cache_chain, next) {
1403 struct kmem_list3 *l3;
1404
1405 l3 = cachep->nodelists[node];
1406 if (!l3)
1407 continue;
1408
1409 drain_freelist(cachep, l3, l3->free_objects);
1410
1411 if (!list_empty(&l3->slabs_full) ||
1412 !list_empty(&l3->slabs_partial)) {
1413 ret = -EBUSY;
1414 break;
1415 }
1416 }
1417 return ret;
1418 }
1419
1420 static int __meminit slab_memory_callback(struct notifier_block *self,
1421 unsigned long action, void *arg)
1422 {
1423 struct memory_notify *mnb = arg;
1424 int ret = 0;
1425 int nid;
1426
1427 nid = mnb->status_change_nid;
1428 if (nid < 0)
1429 goto out;
1430
1431 switch (action) {
1432 case MEM_GOING_ONLINE:
1433 mutex_lock(&cache_chain_mutex);
1434 ret = init_cache_nodelists_node(nid);
1435 mutex_unlock(&cache_chain_mutex);
1436 break;
1437 case MEM_GOING_OFFLINE:
1438 mutex_lock(&cache_chain_mutex);
1439 ret = drain_cache_nodelists_node(nid);
1440 mutex_unlock(&cache_chain_mutex);
1441 break;
1442 case MEM_ONLINE:
1443 case MEM_OFFLINE:
1444 case MEM_CANCEL_ONLINE:
1445 case MEM_CANCEL_OFFLINE:
1446 break;
1447 }
1448 out:
1449 return notifier_from_errno(ret);
1450 }
1451 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1452
1453 /*
1454 * swap the static kmem_list3 with kmalloced memory
1455 */
1456 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1457 int nodeid)
1458 {
1459 struct kmem_list3 *ptr;
1460
1461 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1462 BUG_ON(!ptr);
1463
1464 memcpy(ptr, list, sizeof(struct kmem_list3));
1465 /*
1466 * Do not assume that spinlocks can be initialized via memcpy:
1467 */
1468 spin_lock_init(&ptr->list_lock);
1469
1470 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1471 cachep->nodelists[nodeid] = ptr;
1472 }
1473
1474 /*
1475 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1476 * size of kmem_list3.
1477 */
1478 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1479 {
1480 int node;
1481
1482 for_each_online_node(node) {
1483 cachep->nodelists[node] = &initkmem_list3[index + node];
1484 cachep->nodelists[node]->next_reap = jiffies +
1485 REAPTIMEOUT_LIST3 +
1486 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1487 }
1488 }
1489
1490 /*
1491 * Initialisation. Called after the page allocator have been initialised and
1492 * before smp_init().
1493 */
1494 void __init kmem_cache_init(void)
1495 {
1496 size_t left_over;
1497 struct cache_sizes *sizes;
1498 struct cache_names *names;
1499 int i;
1500 int order;
1501 int node;
1502
1503 if (num_possible_nodes() == 1)
1504 use_alien_caches = 0;
1505
1506 for (i = 0; i < NUM_INIT_LISTS; i++) {
1507 kmem_list3_init(&initkmem_list3[i]);
1508 if (i < MAX_NUMNODES)
1509 cache_cache.nodelists[i] = NULL;
1510 }
1511 set_up_list3s(&cache_cache, CACHE_CACHE);
1512
1513 /*
1514 * Fragmentation resistance on low memory - only use bigger
1515 * page orders on machines with more than 32MB of memory if
1516 * not overridden on the command line.
1517 */
1518 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1519 slab_max_order = SLAB_MAX_ORDER_HI;
1520
1521 /* Bootstrap is tricky, because several objects are allocated
1522 * from caches that do not exist yet:
1523 * 1) initialize the cache_cache cache: it contains the struct
1524 * kmem_cache structures of all caches, except cache_cache itself:
1525 * cache_cache is statically allocated.
1526 * Initially an __init data area is used for the head array and the
1527 * kmem_list3 structures, it's replaced with a kmalloc allocated
1528 * array at the end of the bootstrap.
1529 * 2) Create the first kmalloc cache.
1530 * The struct kmem_cache for the new cache is allocated normally.
1531 * An __init data area is used for the head array.
1532 * 3) Create the remaining kmalloc caches, with minimally sized
1533 * head arrays.
1534 * 4) Replace the __init data head arrays for cache_cache and the first
1535 * kmalloc cache with kmalloc allocated arrays.
1536 * 5) Replace the __init data for kmem_list3 for cache_cache and
1537 * the other cache's with kmalloc allocated memory.
1538 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1539 */
1540
1541 node = numa_mem_id();
1542
1543 /* 1) create the cache_cache */
1544 INIT_LIST_HEAD(&cache_chain);
1545 list_add(&cache_cache.next, &cache_chain);
1546 cache_cache.colour_off = cache_line_size();
1547 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1548 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1549
1550 /*
1551 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1552 */
1553 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1554 nr_node_ids * sizeof(struct kmem_list3 *);
1555 #if DEBUG
1556 cache_cache.obj_size = cache_cache.buffer_size;
1557 #endif
1558 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1559 cache_line_size());
1560 cache_cache.reciprocal_buffer_size =
1561 reciprocal_value(cache_cache.buffer_size);
1562
1563 for (order = 0; order < MAX_ORDER; order++) {
1564 cache_estimate(order, cache_cache.buffer_size,
1565 cache_line_size(), 0, &left_over, &cache_cache.num);
1566 if (cache_cache.num)
1567 break;
1568 }
1569 BUG_ON(!cache_cache.num);
1570 cache_cache.gfporder = order;
1571 cache_cache.colour = left_over / cache_cache.colour_off;
1572 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1573 sizeof(struct slab), cache_line_size());
1574
1575 /* 2+3) create the kmalloc caches */
1576 sizes = malloc_sizes;
1577 names = cache_names;
1578
1579 /*
1580 * Initialize the caches that provide memory for the array cache and the
1581 * kmem_list3 structures first. Without this, further allocations will
1582 * bug.
1583 */
1584
1585 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1586 sizes[INDEX_AC].cs_size,
1587 ARCH_KMALLOC_MINALIGN,
1588 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1589 NULL);
1590
1591 if (INDEX_AC != INDEX_L3) {
1592 sizes[INDEX_L3].cs_cachep =
1593 kmem_cache_create(names[INDEX_L3].name,
1594 sizes[INDEX_L3].cs_size,
1595 ARCH_KMALLOC_MINALIGN,
1596 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1597 NULL);
1598 }
1599
1600 slab_early_init = 0;
1601
1602 while (sizes->cs_size != ULONG_MAX) {
1603 /*
1604 * For performance, all the general caches are L1 aligned.
1605 * This should be particularly beneficial on SMP boxes, as it
1606 * eliminates "false sharing".
1607 * Note for systems short on memory removing the alignment will
1608 * allow tighter packing of the smaller caches.
1609 */
1610 if (!sizes->cs_cachep) {
1611 sizes->cs_cachep = kmem_cache_create(names->name,
1612 sizes->cs_size,
1613 ARCH_KMALLOC_MINALIGN,
1614 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1615 NULL);
1616 }
1617 #ifdef CONFIG_ZONE_DMA
1618 sizes->cs_dmacachep = kmem_cache_create(
1619 names->name_dma,
1620 sizes->cs_size,
1621 ARCH_KMALLOC_MINALIGN,
1622 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1623 SLAB_PANIC,
1624 NULL);
1625 #endif
1626 sizes++;
1627 names++;
1628 }
1629 /* 4) Replace the bootstrap head arrays */
1630 {
1631 struct array_cache *ptr;
1632
1633 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1634
1635 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1636 memcpy(ptr, cpu_cache_get(&cache_cache),
1637 sizeof(struct arraycache_init));
1638 /*
1639 * Do not assume that spinlocks can be initialized via memcpy:
1640 */
1641 spin_lock_init(&ptr->lock);
1642
1643 cache_cache.array[smp_processor_id()] = ptr;
1644
1645 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1646
1647 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1648 != &initarray_generic.cache);
1649 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1650 sizeof(struct arraycache_init));
1651 /*
1652 * Do not assume that spinlocks can be initialized via memcpy:
1653 */
1654 spin_lock_init(&ptr->lock);
1655
1656 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1657 ptr;
1658 }
1659 /* 5) Replace the bootstrap kmem_list3's */
1660 {
1661 int nid;
1662
1663 for_each_online_node(nid) {
1664 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1665
1666 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1667 &initkmem_list3[SIZE_AC + nid], nid);
1668
1669 if (INDEX_AC != INDEX_L3) {
1670 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1671 &initkmem_list3[SIZE_L3 + nid], nid);
1672 }
1673 }
1674 }
1675
1676 g_cpucache_up = EARLY;
1677 }
1678
1679 void __init kmem_cache_init_late(void)
1680 {
1681 struct kmem_cache *cachep;
1682
1683 /* Annotate slab for lockdep -- annotate the malloc caches */
1684 init_lock_keys();
1685
1686 /* 6) resize the head arrays to their final sizes */
1687 mutex_lock(&cache_chain_mutex);
1688 list_for_each_entry(cachep, &cache_chain, next)
1689 if (enable_cpucache(cachep, GFP_NOWAIT))
1690 BUG();
1691 mutex_unlock(&cache_chain_mutex);
1692
1693 /* Done! */
1694 g_cpucache_up = FULL;
1695
1696 /*
1697 * Register a cpu startup notifier callback that initializes
1698 * cpu_cache_get for all new cpus
1699 */
1700 register_cpu_notifier(&cpucache_notifier);
1701
1702 #ifdef CONFIG_NUMA
1703 /*
1704 * Register a memory hotplug callback that initializes and frees
1705 * nodelists.
1706 */
1707 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1708 #endif
1709
1710 /*
1711 * The reap timers are started later, with a module init call: That part
1712 * of the kernel is not yet operational.
1713 */
1714 }
1715
1716 static int __init cpucache_init(void)
1717 {
1718 int cpu;
1719
1720 /*
1721 * Register the timers that return unneeded pages to the page allocator
1722 */
1723 for_each_online_cpu(cpu)
1724 start_cpu_timer(cpu);
1725 return 0;
1726 }
1727 __initcall(cpucache_init);
1728
1729 /*
1730 * Interface to system's page allocator. No need to hold the cache-lock.
1731 *
1732 * If we requested dmaable memory, we will get it. Even if we
1733 * did not request dmaable memory, we might get it, but that
1734 * would be relatively rare and ignorable.
1735 */
1736 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1737 {
1738 struct page *page;
1739 int nr_pages;
1740 int i;
1741
1742 #ifndef CONFIG_MMU
1743 /*
1744 * Nommu uses slab's for process anonymous memory allocations, and thus
1745 * requires __GFP_COMP to properly refcount higher order allocations
1746 */
1747 flags |= __GFP_COMP;
1748 #endif
1749
1750 flags |= cachep->gfpflags;
1751 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1752 flags |= __GFP_RECLAIMABLE;
1753
1754 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1755 if (!page)
1756 return NULL;
1757
1758 nr_pages = (1 << cachep->gfporder);
1759 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1760 add_zone_page_state(page_zone(page),
1761 NR_SLAB_RECLAIMABLE, nr_pages);
1762 else
1763 add_zone_page_state(page_zone(page),
1764 NR_SLAB_UNRECLAIMABLE, nr_pages);
1765 for (i = 0; i < nr_pages; i++)
1766 __SetPageSlab(page + i);
1767
1768 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1769 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1770
1771 if (cachep->ctor)
1772 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1773 else
1774 kmemcheck_mark_unallocated_pages(page, nr_pages);
1775 }
1776
1777 return page_address(page);
1778 }
1779
1780 /*
1781 * Interface to system's page release.
1782 */
1783 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1784 {
1785 unsigned long i = (1 << cachep->gfporder);
1786 struct page *page = virt_to_page(addr);
1787 const unsigned long nr_freed = i;
1788
1789 kmemcheck_free_shadow(page, cachep->gfporder);
1790
1791 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1792 sub_zone_page_state(page_zone(page),
1793 NR_SLAB_RECLAIMABLE, nr_freed);
1794 else
1795 sub_zone_page_state(page_zone(page),
1796 NR_SLAB_UNRECLAIMABLE, nr_freed);
1797 while (i--) {
1798 BUG_ON(!PageSlab(page));
1799 __ClearPageSlab(page);
1800 page++;
1801 }
1802 if (current->reclaim_state)
1803 current->reclaim_state->reclaimed_slab += nr_freed;
1804 free_pages((unsigned long)addr, cachep->gfporder);
1805 }
1806
1807 static void kmem_rcu_free(struct rcu_head *head)
1808 {
1809 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1810 struct kmem_cache *cachep = slab_rcu->cachep;
1811
1812 kmem_freepages(cachep, slab_rcu->addr);
1813 if (OFF_SLAB(cachep))
1814 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1815 }
1816
1817 #if DEBUG
1818
1819 #ifdef CONFIG_DEBUG_PAGEALLOC
1820 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1821 unsigned long caller)
1822 {
1823 int size = obj_size(cachep);
1824
1825 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1826
1827 if (size < 5 * sizeof(unsigned long))
1828 return;
1829
1830 *addr++ = 0x12345678;
1831 *addr++ = caller;
1832 *addr++ = smp_processor_id();
1833 size -= 3 * sizeof(unsigned long);
1834 {
1835 unsigned long *sptr = &caller;
1836 unsigned long svalue;
1837
1838 while (!kstack_end(sptr)) {
1839 svalue = *sptr++;
1840 if (kernel_text_address(svalue)) {
1841 *addr++ = svalue;
1842 size -= sizeof(unsigned long);
1843 if (size <= sizeof(unsigned long))
1844 break;
1845 }
1846 }
1847
1848 }
1849 *addr++ = 0x87654321;
1850 }
1851 #endif
1852
1853 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1854 {
1855 int size = obj_size(cachep);
1856 addr = &((char *)addr)[obj_offset(cachep)];
1857
1858 memset(addr, val, size);
1859 *(unsigned char *)(addr + size - 1) = POISON_END;
1860 }
1861
1862 static void dump_line(char *data, int offset, int limit)
1863 {
1864 int i;
1865 unsigned char error = 0;
1866 int bad_count = 0;
1867
1868 printk(KERN_ERR "%03x: ", offset);
1869 for (i = 0; i < limit; i++) {
1870 if (data[offset + i] != POISON_FREE) {
1871 error = data[offset + i];
1872 bad_count++;
1873 }
1874 }
1875 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1876 &data[offset], limit, 1);
1877
1878 if (bad_count == 1) {
1879 error ^= POISON_FREE;
1880 if (!(error & (error - 1))) {
1881 printk(KERN_ERR "Single bit error detected. Probably "
1882 "bad RAM.\n");
1883 #ifdef CONFIG_X86
1884 printk(KERN_ERR "Run memtest86+ or a similar memory "
1885 "test tool.\n");
1886 #else
1887 printk(KERN_ERR "Run a memory test tool.\n");
1888 #endif
1889 }
1890 }
1891 }
1892 #endif
1893
1894 #if DEBUG
1895
1896 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1897 {
1898 int i, size;
1899 char *realobj;
1900
1901 if (cachep->flags & SLAB_RED_ZONE) {
1902 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1903 *dbg_redzone1(cachep, objp),
1904 *dbg_redzone2(cachep, objp));
1905 }
1906
1907 if (cachep->flags & SLAB_STORE_USER) {
1908 printk(KERN_ERR "Last user: [<%p>]",
1909 *dbg_userword(cachep, objp));
1910 print_symbol("(%s)",
1911 (unsigned long)*dbg_userword(cachep, objp));
1912 printk("\n");
1913 }
1914 realobj = (char *)objp + obj_offset(cachep);
1915 size = obj_size(cachep);
1916 for (i = 0; i < size && lines; i += 16, lines--) {
1917 int limit;
1918 limit = 16;
1919 if (i + limit > size)
1920 limit = size - i;
1921 dump_line(realobj, i, limit);
1922 }
1923 }
1924
1925 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1926 {
1927 char *realobj;
1928 int size, i;
1929 int lines = 0;
1930
1931 realobj = (char *)objp + obj_offset(cachep);
1932 size = obj_size(cachep);
1933
1934 for (i = 0; i < size; i++) {
1935 char exp = POISON_FREE;
1936 if (i == size - 1)
1937 exp = POISON_END;
1938 if (realobj[i] != exp) {
1939 int limit;
1940 /* Mismatch ! */
1941 /* Print header */
1942 if (lines == 0) {
1943 printk(KERN_ERR
1944 "Slab corruption: %s start=%p, len=%d\n",
1945 cachep->name, realobj, size);
1946 print_objinfo(cachep, objp, 0);
1947 }
1948 /* Hexdump the affected line */
1949 i = (i / 16) * 16;
1950 limit = 16;
1951 if (i + limit > size)
1952 limit = size - i;
1953 dump_line(realobj, i, limit);
1954 i += 16;
1955 lines++;
1956 /* Limit to 5 lines */
1957 if (lines > 5)
1958 break;
1959 }
1960 }
1961 if (lines != 0) {
1962 /* Print some data about the neighboring objects, if they
1963 * exist:
1964 */
1965 struct slab *slabp = virt_to_slab(objp);
1966 unsigned int objnr;
1967
1968 objnr = obj_to_index(cachep, slabp, objp);
1969 if (objnr) {
1970 objp = index_to_obj(cachep, slabp, objnr - 1);
1971 realobj = (char *)objp + obj_offset(cachep);
1972 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1973 realobj, size);
1974 print_objinfo(cachep, objp, 2);
1975 }
1976 if (objnr + 1 < cachep->num) {
1977 objp = index_to_obj(cachep, slabp, objnr + 1);
1978 realobj = (char *)objp + obj_offset(cachep);
1979 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1980 realobj, size);
1981 print_objinfo(cachep, objp, 2);
1982 }
1983 }
1984 }
1985 #endif
1986
1987 #if DEBUG
1988 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1989 {
1990 int i;
1991 for (i = 0; i < cachep->num; i++) {
1992 void *objp = index_to_obj(cachep, slabp, i);
1993
1994 if (cachep->flags & SLAB_POISON) {
1995 #ifdef CONFIG_DEBUG_PAGEALLOC
1996 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1997 OFF_SLAB(cachep))
1998 kernel_map_pages(virt_to_page(objp),
1999 cachep->buffer_size / PAGE_SIZE, 1);
2000 else
2001 check_poison_obj(cachep, objp);
2002 #else
2003 check_poison_obj(cachep, objp);
2004 #endif
2005 }
2006 if (cachep->flags & SLAB_RED_ZONE) {
2007 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2008 slab_error(cachep, "start of a freed object "
2009 "was overwritten");
2010 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2011 slab_error(cachep, "end of a freed object "
2012 "was overwritten");
2013 }
2014 }
2015 }
2016 #else
2017 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2018 {
2019 }
2020 #endif
2021
2022 /**
2023 * slab_destroy - destroy and release all objects in a slab
2024 * @cachep: cache pointer being destroyed
2025 * @slabp: slab pointer being destroyed
2026 *
2027 * Destroy all the objs in a slab, and release the mem back to the system.
2028 * Before calling the slab must have been unlinked from the cache. The
2029 * cache-lock is not held/needed.
2030 */
2031 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2032 {
2033 void *addr = slabp->s_mem - slabp->colouroff;
2034
2035 slab_destroy_debugcheck(cachep, slabp);
2036 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2037 struct slab_rcu *slab_rcu;
2038
2039 slab_rcu = (struct slab_rcu *)slabp;
2040 slab_rcu->cachep = cachep;
2041 slab_rcu->addr = addr;
2042 call_rcu(&slab_rcu->head, kmem_rcu_free);
2043 } else {
2044 kmem_freepages(cachep, addr);
2045 if (OFF_SLAB(cachep))
2046 kmem_cache_free(cachep->slabp_cache, slabp);
2047 }
2048 }
2049
2050 static void __kmem_cache_destroy(struct kmem_cache *cachep)
2051 {
2052 int i;
2053 struct kmem_list3 *l3;
2054
2055 for_each_online_cpu(i)
2056 kfree(cachep->array[i]);
2057
2058 /* NUMA: free the list3 structures */
2059 for_each_online_node(i) {
2060 l3 = cachep->nodelists[i];
2061 if (l3) {
2062 kfree(l3->shared);
2063 free_alien_cache(l3->alien);
2064 kfree(l3);
2065 }
2066 }
2067 kmem_cache_free(&cache_cache, cachep);
2068 }
2069
2070
2071 /**
2072 * calculate_slab_order - calculate size (page order) of slabs
2073 * @cachep: pointer to the cache that is being created
2074 * @size: size of objects to be created in this cache.
2075 * @align: required alignment for the objects.
2076 * @flags: slab allocation flags
2077 *
2078 * Also calculates the number of objects per slab.
2079 *
2080 * This could be made much more intelligent. For now, try to avoid using
2081 * high order pages for slabs. When the gfp() functions are more friendly
2082 * towards high-order requests, this should be changed.
2083 */
2084 static size_t calculate_slab_order(struct kmem_cache *cachep,
2085 size_t size, size_t align, unsigned long flags)
2086 {
2087 unsigned long offslab_limit;
2088 size_t left_over = 0;
2089 int gfporder;
2090
2091 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2092 unsigned int num;
2093 size_t remainder;
2094
2095 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2096 if (!num)
2097 continue;
2098
2099 if (flags & CFLGS_OFF_SLAB) {
2100 /*
2101 * Max number of objs-per-slab for caches which
2102 * use off-slab slabs. Needed to avoid a possible
2103 * looping condition in cache_grow().
2104 */
2105 offslab_limit = size - sizeof(struct slab);
2106 offslab_limit /= sizeof(kmem_bufctl_t);
2107
2108 if (num > offslab_limit)
2109 break;
2110 }
2111
2112 /* Found something acceptable - save it away */
2113 cachep->num = num;
2114 cachep->gfporder = gfporder;
2115 left_over = remainder;
2116
2117 /*
2118 * A VFS-reclaimable slab tends to have most allocations
2119 * as GFP_NOFS and we really don't want to have to be allocating
2120 * higher-order pages when we are unable to shrink dcache.
2121 */
2122 if (flags & SLAB_RECLAIM_ACCOUNT)
2123 break;
2124
2125 /*
2126 * Large number of objects is good, but very large slabs are
2127 * currently bad for the gfp()s.
2128 */
2129 if (gfporder >= slab_max_order)
2130 break;
2131
2132 /*
2133 * Acceptable internal fragmentation?
2134 */
2135 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2136 break;
2137 }
2138 return left_over;
2139 }
2140
2141 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2142 {
2143 if (g_cpucache_up == FULL)
2144 return enable_cpucache(cachep, gfp);
2145
2146 if (g_cpucache_up == NONE) {
2147 /*
2148 * Note: the first kmem_cache_create must create the cache
2149 * that's used by kmalloc(24), otherwise the creation of
2150 * further caches will BUG().
2151 */
2152 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2153
2154 /*
2155 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2156 * the first cache, then we need to set up all its list3s,
2157 * otherwise the creation of further caches will BUG().
2158 */
2159 set_up_list3s(cachep, SIZE_AC);
2160 if (INDEX_AC == INDEX_L3)
2161 g_cpucache_up = PARTIAL_L3;
2162 else
2163 g_cpucache_up = PARTIAL_AC;
2164 } else {
2165 cachep->array[smp_processor_id()] =
2166 kmalloc(sizeof(struct arraycache_init), gfp);
2167
2168 if (g_cpucache_up == PARTIAL_AC) {
2169 set_up_list3s(cachep, SIZE_L3);
2170 g_cpucache_up = PARTIAL_L3;
2171 } else {
2172 int node;
2173 for_each_online_node(node) {
2174 cachep->nodelists[node] =
2175 kmalloc_node(sizeof(struct kmem_list3),
2176 gfp, node);
2177 BUG_ON(!cachep->nodelists[node]);
2178 kmem_list3_init(cachep->nodelists[node]);
2179 }
2180 }
2181 }
2182 cachep->nodelists[numa_mem_id()]->next_reap =
2183 jiffies + REAPTIMEOUT_LIST3 +
2184 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2185
2186 cpu_cache_get(cachep)->avail = 0;
2187 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2188 cpu_cache_get(cachep)->batchcount = 1;
2189 cpu_cache_get(cachep)->touched = 0;
2190 cachep->batchcount = 1;
2191 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2192 return 0;
2193 }
2194
2195 /**
2196 * kmem_cache_create - Create a cache.
2197 * @name: A string which is used in /proc/slabinfo to identify this cache.
2198 * @size: The size of objects to be created in this cache.
2199 * @align: The required alignment for the objects.
2200 * @flags: SLAB flags
2201 * @ctor: A constructor for the objects.
2202 *
2203 * Returns a ptr to the cache on success, NULL on failure.
2204 * Cannot be called within a int, but can be interrupted.
2205 * The @ctor is run when new pages are allocated by the cache.
2206 *
2207 * @name must be valid until the cache is destroyed. This implies that
2208 * the module calling this has to destroy the cache before getting unloaded.
2209 *
2210 * The flags are
2211 *
2212 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2213 * to catch references to uninitialised memory.
2214 *
2215 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2216 * for buffer overruns.
2217 *
2218 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2219 * cacheline. This can be beneficial if you're counting cycles as closely
2220 * as davem.
2221 */
2222 struct kmem_cache *
2223 kmem_cache_create (const char *name, size_t size, size_t align,
2224 unsigned long flags, void (*ctor)(void *))
2225 {
2226 size_t left_over, slab_size, ralign;
2227 struct kmem_cache *cachep = NULL, *pc;
2228 gfp_t gfp;
2229
2230 /*
2231 * Sanity checks... these are all serious usage bugs.
2232 */
2233 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2234 size > KMALLOC_MAX_SIZE) {
2235 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2236 name);
2237 BUG();
2238 }
2239
2240 /*
2241 * We use cache_chain_mutex to ensure a consistent view of
2242 * cpu_online_mask as well. Please see cpuup_callback
2243 */
2244 if (slab_is_available()) {
2245 get_online_cpus();
2246 mutex_lock(&cache_chain_mutex);
2247 }
2248
2249 list_for_each_entry(pc, &cache_chain, next) {
2250 char tmp;
2251 int res;
2252
2253 /*
2254 * This happens when the module gets unloaded and doesn't
2255 * destroy its slab cache and no-one else reuses the vmalloc
2256 * area of the module. Print a warning.
2257 */
2258 res = probe_kernel_address(pc->name, tmp);
2259 if (res) {
2260 printk(KERN_ERR
2261 "SLAB: cache with size %d has lost its name\n",
2262 pc->buffer_size);
2263 continue;
2264 }
2265
2266 if (!strcmp(pc->name, name)) {
2267 printk(KERN_ERR
2268 "kmem_cache_create: duplicate cache %s\n", name);
2269 dump_stack();
2270 goto oops;
2271 }
2272 }
2273
2274 #if DEBUG
2275 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2276 #if FORCED_DEBUG
2277 /*
2278 * Enable redzoning and last user accounting, except for caches with
2279 * large objects, if the increased size would increase the object size
2280 * above the next power of two: caches with object sizes just above a
2281 * power of two have a significant amount of internal fragmentation.
2282 */
2283 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2284 2 * sizeof(unsigned long long)))
2285 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2286 if (!(flags & SLAB_DESTROY_BY_RCU))
2287 flags |= SLAB_POISON;
2288 #endif
2289 if (flags & SLAB_DESTROY_BY_RCU)
2290 BUG_ON(flags & SLAB_POISON);
2291 #endif
2292 /*
2293 * Always checks flags, a caller might be expecting debug support which
2294 * isn't available.
2295 */
2296 BUG_ON(flags & ~CREATE_MASK);
2297
2298 /*
2299 * Check that size is in terms of words. This is needed to avoid
2300 * unaligned accesses for some archs when redzoning is used, and makes
2301 * sure any on-slab bufctl's are also correctly aligned.
2302 */
2303 if (size & (BYTES_PER_WORD - 1)) {
2304 size += (BYTES_PER_WORD - 1);
2305 size &= ~(BYTES_PER_WORD - 1);
2306 }
2307
2308 /* calculate the final buffer alignment: */
2309
2310 /* 1) arch recommendation: can be overridden for debug */
2311 if (flags & SLAB_HWCACHE_ALIGN) {
2312 /*
2313 * Default alignment: as specified by the arch code. Except if
2314 * an object is really small, then squeeze multiple objects into
2315 * one cacheline.
2316 */
2317 ralign = cache_line_size();
2318 while (size <= ralign / 2)
2319 ralign /= 2;
2320 } else {
2321 ralign = BYTES_PER_WORD;
2322 }
2323
2324 /*
2325 * Redzoning and user store require word alignment or possibly larger.
2326 * Note this will be overridden by architecture or caller mandated
2327 * alignment if either is greater than BYTES_PER_WORD.
2328 */
2329 if (flags & SLAB_STORE_USER)
2330 ralign = BYTES_PER_WORD;
2331
2332 if (flags & SLAB_RED_ZONE) {
2333 ralign = REDZONE_ALIGN;
2334 /* If redzoning, ensure that the second redzone is suitably
2335 * aligned, by adjusting the object size accordingly. */
2336 size += REDZONE_ALIGN - 1;
2337 size &= ~(REDZONE_ALIGN - 1);
2338 }
2339
2340 /* 2) arch mandated alignment */
2341 if (ralign < ARCH_SLAB_MINALIGN) {
2342 ralign = ARCH_SLAB_MINALIGN;
2343 }
2344 /* 3) caller mandated alignment */
2345 if (ralign < align) {
2346 ralign = align;
2347 }
2348 /* disable debug if necessary */
2349 if (ralign > __alignof__(unsigned long long))
2350 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2351 /*
2352 * 4) Store it.
2353 */
2354 align = ralign;
2355
2356 if (slab_is_available())
2357 gfp = GFP_KERNEL;
2358 else
2359 gfp = GFP_NOWAIT;
2360
2361 /* Get cache's description obj. */
2362 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2363 if (!cachep)
2364 goto oops;
2365
2366 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2367 #if DEBUG
2368 cachep->obj_size = size;
2369
2370 /*
2371 * Both debugging options require word-alignment which is calculated
2372 * into align above.
2373 */
2374 if (flags & SLAB_RED_ZONE) {
2375 /* add space for red zone words */
2376 cachep->obj_offset += sizeof(unsigned long long);
2377 size += 2 * sizeof(unsigned long long);
2378 }
2379 if (flags & SLAB_STORE_USER) {
2380 /* user store requires one word storage behind the end of
2381 * the real object. But if the second red zone needs to be
2382 * aligned to 64 bits, we must allow that much space.
2383 */
2384 if (flags & SLAB_RED_ZONE)
2385 size += REDZONE_ALIGN;
2386 else
2387 size += BYTES_PER_WORD;
2388 }
2389 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2390 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2391 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2392 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2393 size = PAGE_SIZE;
2394 }
2395 #endif
2396 #endif
2397
2398 /*
2399 * Determine if the slab management is 'on' or 'off' slab.
2400 * (bootstrapping cannot cope with offslab caches so don't do
2401 * it too early on. Always use on-slab management when
2402 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2403 */
2404 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2405 !(flags & SLAB_NOLEAKTRACE))
2406 /*
2407 * Size is large, assume best to place the slab management obj
2408 * off-slab (should allow better packing of objs).
2409 */
2410 flags |= CFLGS_OFF_SLAB;
2411
2412 size = ALIGN(size, align);
2413
2414 left_over = calculate_slab_order(cachep, size, align, flags);
2415
2416 if (!cachep->num) {
2417 printk(KERN_ERR
2418 "kmem_cache_create: couldn't create cache %s.\n", name);
2419 kmem_cache_free(&cache_cache, cachep);
2420 cachep = NULL;
2421 goto oops;
2422 }
2423 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2424 + sizeof(struct slab), align);
2425
2426 /*
2427 * If the slab has been placed off-slab, and we have enough space then
2428 * move it on-slab. This is at the expense of any extra colouring.
2429 */
2430 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2431 flags &= ~CFLGS_OFF_SLAB;
2432 left_over -= slab_size;
2433 }
2434
2435 if (flags & CFLGS_OFF_SLAB) {
2436 /* really off slab. No need for manual alignment */
2437 slab_size =
2438 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2439
2440 #ifdef CONFIG_PAGE_POISONING
2441 /* If we're going to use the generic kernel_map_pages()
2442 * poisoning, then it's going to smash the contents of
2443 * the redzone and userword anyhow, so switch them off.
2444 */
2445 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2446 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2447 #endif
2448 }
2449
2450 cachep->colour_off = cache_line_size();
2451 /* Offset must be a multiple of the alignment. */
2452 if (cachep->colour_off < align)
2453 cachep->colour_off = align;
2454 cachep->colour = left_over / cachep->colour_off;
2455 cachep->slab_size = slab_size;
2456 cachep->flags = flags;
2457 cachep->gfpflags = 0;
2458 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2459 cachep->gfpflags |= GFP_DMA;
2460 cachep->buffer_size = size;
2461 cachep->reciprocal_buffer_size = reciprocal_value(size);
2462
2463 if (flags & CFLGS_OFF_SLAB) {
2464 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2465 /*
2466 * This is a possibility for one of the malloc_sizes caches.
2467 * But since we go off slab only for object size greater than
2468 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2469 * this should not happen at all.
2470 * But leave a BUG_ON for some lucky dude.
2471 */
2472 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2473 }
2474 cachep->ctor = ctor;
2475 cachep->name = name;
2476
2477 if (setup_cpu_cache(cachep, gfp)) {
2478 __kmem_cache_destroy(cachep);
2479 cachep = NULL;
2480 goto oops;
2481 }
2482
2483 if (flags & SLAB_DEBUG_OBJECTS) {
2484 /*
2485 * Would deadlock through slab_destroy()->call_rcu()->
2486 * debug_object_activate()->kmem_cache_alloc().
2487 */
2488 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2489
2490 slab_set_debugobj_lock_classes(cachep);
2491 }
2492
2493 /* cache setup completed, link it into the list */
2494 list_add(&cachep->next, &cache_chain);
2495 oops:
2496 if (!cachep && (flags & SLAB_PANIC))
2497 panic("kmem_cache_create(): failed to create slab `%s'\n",
2498 name);
2499 if (slab_is_available()) {
2500 mutex_unlock(&cache_chain_mutex);
2501 put_online_cpus();
2502 }
2503 return cachep;
2504 }
2505 EXPORT_SYMBOL(kmem_cache_create);
2506
2507 #if DEBUG
2508 static void check_irq_off(void)
2509 {
2510 BUG_ON(!irqs_disabled());
2511 }
2512
2513 static void check_irq_on(void)
2514 {
2515 BUG_ON(irqs_disabled());
2516 }
2517
2518 static void check_spinlock_acquired(struct kmem_cache *cachep)
2519 {
2520 #ifdef CONFIG_SMP
2521 check_irq_off();
2522 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2523 #endif
2524 }
2525
2526 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2527 {
2528 #ifdef CONFIG_SMP
2529 check_irq_off();
2530 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2531 #endif
2532 }
2533
2534 #else
2535 #define check_irq_off() do { } while(0)
2536 #define check_irq_on() do { } while(0)
2537 #define check_spinlock_acquired(x) do { } while(0)
2538 #define check_spinlock_acquired_node(x, y) do { } while(0)
2539 #endif
2540
2541 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2542 struct array_cache *ac,
2543 int force, int node);
2544
2545 static void do_drain(void *arg)
2546 {
2547 struct kmem_cache *cachep = arg;
2548 struct array_cache *ac;
2549 int node = numa_mem_id();
2550
2551 check_irq_off();
2552 ac = cpu_cache_get(cachep);
2553 spin_lock(&cachep->nodelists[node]->list_lock);
2554 free_block(cachep, ac->entry, ac->avail, node);
2555 spin_unlock(&cachep->nodelists[node]->list_lock);
2556 ac->avail = 0;
2557 }
2558
2559 static void drain_cpu_caches(struct kmem_cache *cachep)
2560 {
2561 struct kmem_list3 *l3;
2562 int node;
2563
2564 on_each_cpu(do_drain, cachep, 1);
2565 check_irq_on();
2566 for_each_online_node(node) {
2567 l3 = cachep->nodelists[node];
2568 if (l3 && l3->alien)
2569 drain_alien_cache(cachep, l3->alien);
2570 }
2571
2572 for_each_online_node(node) {
2573 l3 = cachep->nodelists[node];
2574 if (l3)
2575 drain_array(cachep, l3, l3->shared, 1, node);
2576 }
2577 }
2578
2579 /*
2580 * Remove slabs from the list of free slabs.
2581 * Specify the number of slabs to drain in tofree.
2582 *
2583 * Returns the actual number of slabs released.
2584 */
2585 static int drain_freelist(struct kmem_cache *cache,
2586 struct kmem_list3 *l3, int tofree)
2587 {
2588 struct list_head *p;
2589 int nr_freed;
2590 struct slab *slabp;
2591
2592 nr_freed = 0;
2593 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2594
2595 spin_lock_irq(&l3->list_lock);
2596 p = l3->slabs_free.prev;
2597 if (p == &l3->slabs_free) {
2598 spin_unlock_irq(&l3->list_lock);
2599 goto out;
2600 }
2601
2602 slabp = list_entry(p, struct slab, list);
2603 #if DEBUG
2604 BUG_ON(slabp->inuse);
2605 #endif
2606 list_del(&slabp->list);
2607 /*
2608 * Safe to drop the lock. The slab is no longer linked
2609 * to the cache.
2610 */
2611 l3->free_objects -= cache->num;
2612 spin_unlock_irq(&l3->list_lock);
2613 slab_destroy(cache, slabp);
2614 nr_freed++;
2615 }
2616 out:
2617 return nr_freed;
2618 }
2619
2620 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2621 static int __cache_shrink(struct kmem_cache *cachep)
2622 {
2623 int ret = 0, i = 0;
2624 struct kmem_list3 *l3;
2625
2626 drain_cpu_caches(cachep);
2627
2628 check_irq_on();
2629 for_each_online_node(i) {
2630 l3 = cachep->nodelists[i];
2631 if (!l3)
2632 continue;
2633
2634 drain_freelist(cachep, l3, l3->free_objects);
2635
2636 ret += !list_empty(&l3->slabs_full) ||
2637 !list_empty(&l3->slabs_partial);
2638 }
2639 return (ret ? 1 : 0);
2640 }
2641
2642 /**
2643 * kmem_cache_shrink - Shrink a cache.
2644 * @cachep: The cache to shrink.
2645 *
2646 * Releases as many slabs as possible for a cache.
2647 * To help debugging, a zero exit status indicates all slabs were released.
2648 */
2649 int kmem_cache_shrink(struct kmem_cache *cachep)
2650 {
2651 int ret;
2652 BUG_ON(!cachep || in_interrupt());
2653
2654 get_online_cpus();
2655 mutex_lock(&cache_chain_mutex);
2656 ret = __cache_shrink(cachep);
2657 mutex_unlock(&cache_chain_mutex);
2658 put_online_cpus();
2659 return ret;
2660 }
2661 EXPORT_SYMBOL(kmem_cache_shrink);
2662
2663 /**
2664 * kmem_cache_destroy - delete a cache
2665 * @cachep: the cache to destroy
2666 *
2667 * Remove a &struct kmem_cache object from the slab cache.
2668 *
2669 * It is expected this function will be called by a module when it is
2670 * unloaded. This will remove the cache completely, and avoid a duplicate
2671 * cache being allocated each time a module is loaded and unloaded, if the
2672 * module doesn't have persistent in-kernel storage across loads and unloads.
2673 *
2674 * The cache must be empty before calling this function.
2675 *
2676 * The caller must guarantee that no one will allocate memory from the cache
2677 * during the kmem_cache_destroy().
2678 */
2679 void kmem_cache_destroy(struct kmem_cache *cachep)
2680 {
2681 BUG_ON(!cachep || in_interrupt());
2682
2683 /* Find the cache in the chain of caches. */
2684 get_online_cpus();
2685 mutex_lock(&cache_chain_mutex);
2686 /*
2687 * the chain is never empty, cache_cache is never destroyed
2688 */
2689 list_del(&cachep->next);
2690 if (__cache_shrink(cachep)) {
2691 slab_error(cachep, "Can't free all objects");
2692 list_add(&cachep->next, &cache_chain);
2693 mutex_unlock(&cache_chain_mutex);
2694 put_online_cpus();
2695 return;
2696 }
2697
2698 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2699 rcu_barrier();
2700
2701 __kmem_cache_destroy(cachep);
2702 mutex_unlock(&cache_chain_mutex);
2703 put_online_cpus();
2704 }
2705 EXPORT_SYMBOL(kmem_cache_destroy);
2706
2707 /*
2708 * Get the memory for a slab management obj.
2709 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2710 * always come from malloc_sizes caches. The slab descriptor cannot
2711 * come from the same cache which is getting created because,
2712 * when we are searching for an appropriate cache for these
2713 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2714 * If we are creating a malloc_sizes cache here it would not be visible to
2715 * kmem_find_general_cachep till the initialization is complete.
2716 * Hence we cannot have slabp_cache same as the original cache.
2717 */
2718 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2719 int colour_off, gfp_t local_flags,
2720 int nodeid)
2721 {
2722 struct slab *slabp;
2723
2724 if (OFF_SLAB(cachep)) {
2725 /* Slab management obj is off-slab. */
2726 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2727 local_flags, nodeid);
2728 /*
2729 * If the first object in the slab is leaked (it's allocated
2730 * but no one has a reference to it), we want to make sure
2731 * kmemleak does not treat the ->s_mem pointer as a reference
2732 * to the object. Otherwise we will not report the leak.
2733 */
2734 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2735 local_flags);
2736 if (!slabp)
2737 return NULL;
2738 } else {
2739 slabp = objp + colour_off;
2740 colour_off += cachep->slab_size;
2741 }
2742 slabp->inuse = 0;
2743 slabp->colouroff = colour_off;
2744 slabp->s_mem = objp + colour_off;
2745 slabp->nodeid = nodeid;
2746 slabp->free = 0;
2747 return slabp;
2748 }
2749
2750 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2751 {
2752 return (kmem_bufctl_t *) (slabp + 1);
2753 }
2754
2755 static void cache_init_objs(struct kmem_cache *cachep,
2756 struct slab *slabp)
2757 {
2758 int i;
2759
2760 for (i = 0; i < cachep->num; i++) {
2761 void *objp = index_to_obj(cachep, slabp, i);
2762 #if DEBUG
2763 /* need to poison the objs? */
2764 if (cachep->flags & SLAB_POISON)
2765 poison_obj(cachep, objp, POISON_FREE);
2766 if (cachep->flags & SLAB_STORE_USER)
2767 *dbg_userword(cachep, objp) = NULL;
2768
2769 if (cachep->flags & SLAB_RED_ZONE) {
2770 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2771 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2772 }
2773 /*
2774 * Constructors are not allowed to allocate memory from the same
2775 * cache which they are a constructor for. Otherwise, deadlock.
2776 * They must also be threaded.
2777 */
2778 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2779 cachep->ctor(objp + obj_offset(cachep));
2780
2781 if (cachep->flags & SLAB_RED_ZONE) {
2782 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2783 slab_error(cachep, "constructor overwrote the"
2784 " end of an object");
2785 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2786 slab_error(cachep, "constructor overwrote the"
2787 " start of an object");
2788 }
2789 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2790 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2791 kernel_map_pages(virt_to_page(objp),
2792 cachep->buffer_size / PAGE_SIZE, 0);
2793 #else
2794 if (cachep->ctor)
2795 cachep->ctor(objp);
2796 #endif
2797 slab_bufctl(slabp)[i] = i + 1;
2798 }
2799 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2800 }
2801
2802 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2803 {
2804 if (CONFIG_ZONE_DMA_FLAG) {
2805 if (flags & GFP_DMA)
2806 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2807 else
2808 BUG_ON(cachep->gfpflags & GFP_DMA);
2809 }
2810 }
2811
2812 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2813 int nodeid)
2814 {
2815 void *objp = index_to_obj(cachep, slabp, slabp->free);
2816 kmem_bufctl_t next;
2817
2818 slabp->inuse++;
2819 next = slab_bufctl(slabp)[slabp->free];
2820 #if DEBUG
2821 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2822 WARN_ON(slabp->nodeid != nodeid);
2823 #endif
2824 slabp->free = next;
2825
2826 return objp;
2827 }
2828
2829 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2830 void *objp, int nodeid)
2831 {
2832 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2833
2834 #if DEBUG
2835 /* Verify that the slab belongs to the intended node */
2836 WARN_ON(slabp->nodeid != nodeid);
2837
2838 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2839 printk(KERN_ERR "slab: double free detected in cache "
2840 "'%s', objp %p\n", cachep->name, objp);
2841 BUG();
2842 }
2843 #endif
2844 slab_bufctl(slabp)[objnr] = slabp->free;
2845 slabp->free = objnr;
2846 slabp->inuse--;
2847 }
2848
2849 /*
2850 * Map pages beginning at addr to the given cache and slab. This is required
2851 * for the slab allocator to be able to lookup the cache and slab of a
2852 * virtual address for kfree, ksize, and slab debugging.
2853 */
2854 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2855 void *addr)
2856 {
2857 int nr_pages;
2858 struct page *page;
2859
2860 page = virt_to_page(addr);
2861
2862 nr_pages = 1;
2863 if (likely(!PageCompound(page)))
2864 nr_pages <<= cache->gfporder;
2865
2866 do {
2867 page_set_cache(page, cache);
2868 page_set_slab(page, slab);
2869 page++;
2870 } while (--nr_pages);
2871 }
2872
2873 /*
2874 * Grow (by 1) the number of slabs within a cache. This is called by
2875 * kmem_cache_alloc() when there are no active objs left in a cache.
2876 */
2877 static int cache_grow(struct kmem_cache *cachep,
2878 gfp_t flags, int nodeid, void *objp)
2879 {
2880 struct slab *slabp;
2881 size_t offset;
2882 gfp_t local_flags;
2883 struct kmem_list3 *l3;
2884
2885 /*
2886 * Be lazy and only check for valid flags here, keeping it out of the
2887 * critical path in kmem_cache_alloc().
2888 */
2889 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2890 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2891
2892 /* Take the l3 list lock to change the colour_next on this node */
2893 check_irq_off();
2894 l3 = cachep->nodelists[nodeid];
2895 spin_lock(&l3->list_lock);
2896
2897 /* Get colour for the slab, and cal the next value. */
2898 offset = l3->colour_next;
2899 l3->colour_next++;
2900 if (l3->colour_next >= cachep->colour)
2901 l3->colour_next = 0;
2902 spin_unlock(&l3->list_lock);
2903
2904 offset *= cachep->colour_off;
2905
2906 if (local_flags & __GFP_WAIT)
2907 local_irq_enable();
2908
2909 /*
2910 * The test for missing atomic flag is performed here, rather than
2911 * the more obvious place, simply to reduce the critical path length
2912 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2913 * will eventually be caught here (where it matters).
2914 */
2915 kmem_flagcheck(cachep, flags);
2916
2917 /*
2918 * Get mem for the objs. Attempt to allocate a physical page from
2919 * 'nodeid'.
2920 */
2921 if (!objp)
2922 objp = kmem_getpages(cachep, local_flags, nodeid);
2923 if (!objp)
2924 goto failed;
2925
2926 /* Get slab management. */
2927 slabp = alloc_slabmgmt(cachep, objp, offset,
2928 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2929 if (!slabp)
2930 goto opps1;
2931
2932 slab_map_pages(cachep, slabp, objp);
2933
2934 cache_init_objs(cachep, slabp);
2935
2936 if (local_flags & __GFP_WAIT)
2937 local_irq_disable();
2938 check_irq_off();
2939 spin_lock(&l3->list_lock);
2940
2941 /* Make slab active. */
2942 list_add_tail(&slabp->list, &(l3->slabs_free));
2943 STATS_INC_GROWN(cachep);
2944 l3->free_objects += cachep->num;
2945 spin_unlock(&l3->list_lock);
2946 return 1;
2947 opps1:
2948 kmem_freepages(cachep, objp);
2949 failed:
2950 if (local_flags & __GFP_WAIT)
2951 local_irq_disable();
2952 return 0;
2953 }
2954
2955 #if DEBUG
2956
2957 /*
2958 * Perform extra freeing checks:
2959 * - detect bad pointers.
2960 * - POISON/RED_ZONE checking
2961 */
2962 static void kfree_debugcheck(const void *objp)
2963 {
2964 if (!virt_addr_valid(objp)) {
2965 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2966 (unsigned long)objp);
2967 BUG();
2968 }
2969 }
2970
2971 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2972 {
2973 unsigned long long redzone1, redzone2;
2974
2975 redzone1 = *dbg_redzone1(cache, obj);
2976 redzone2 = *dbg_redzone2(cache, obj);
2977
2978 /*
2979 * Redzone is ok.
2980 */
2981 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2982 return;
2983
2984 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2985 slab_error(cache, "double free detected");
2986 else
2987 slab_error(cache, "memory outside object was overwritten");
2988
2989 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2990 obj, redzone1, redzone2);
2991 }
2992
2993 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2994 void *caller)
2995 {
2996 struct page *page;
2997 unsigned int objnr;
2998 struct slab *slabp;
2999
3000 BUG_ON(virt_to_cache(objp) != cachep);
3001
3002 objp -= obj_offset(cachep);
3003 kfree_debugcheck(objp);
3004 page = virt_to_head_page(objp);
3005
3006 slabp = page_get_slab(page);
3007
3008 if (cachep->flags & SLAB_RED_ZONE) {
3009 verify_redzone_free(cachep, objp);
3010 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3011 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3012 }
3013 if (cachep->flags & SLAB_STORE_USER)
3014 *dbg_userword(cachep, objp) = caller;
3015
3016 objnr = obj_to_index(cachep, slabp, objp);
3017
3018 BUG_ON(objnr >= cachep->num);
3019 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3020
3021 #ifdef CONFIG_DEBUG_SLAB_LEAK
3022 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3023 #endif
3024 if (cachep->flags & SLAB_POISON) {
3025 #ifdef CONFIG_DEBUG_PAGEALLOC
3026 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3027 store_stackinfo(cachep, objp, (unsigned long)caller);
3028 kernel_map_pages(virt_to_page(objp),
3029 cachep->buffer_size / PAGE_SIZE, 0);
3030 } else {
3031 poison_obj(cachep, objp, POISON_FREE);
3032 }
3033 #else
3034 poison_obj(cachep, objp, POISON_FREE);
3035 #endif
3036 }
3037 return objp;
3038 }
3039
3040 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3041 {
3042 kmem_bufctl_t i;
3043 int entries = 0;
3044
3045 /* Check slab's freelist to see if this obj is there. */
3046 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3047 entries++;
3048 if (entries > cachep->num || i >= cachep->num)
3049 goto bad;
3050 }
3051 if (entries != cachep->num - slabp->inuse) {
3052 bad:
3053 printk(KERN_ERR "slab: Internal list corruption detected in "
3054 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3055 cachep->name, cachep->num, slabp, slabp->inuse);
3056 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3057 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3058 1);
3059 BUG();
3060 }
3061 }
3062 #else
3063 #define kfree_debugcheck(x) do { } while(0)
3064 #define cache_free_debugcheck(x,objp,z) (objp)
3065 #define check_slabp(x,y) do { } while(0)
3066 #endif
3067
3068 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3069 {
3070 int batchcount;
3071 struct kmem_list3 *l3;
3072 struct array_cache *ac;
3073 int node;
3074
3075 retry:
3076 check_irq_off();
3077 node = numa_mem_id();
3078 ac = cpu_cache_get(cachep);
3079 batchcount = ac->batchcount;
3080 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3081 /*
3082 * If there was little recent activity on this cache, then
3083 * perform only a partial refill. Otherwise we could generate
3084 * refill bouncing.
3085 */
3086 batchcount = BATCHREFILL_LIMIT;
3087 }
3088 l3 = cachep->nodelists[node];
3089
3090 BUG_ON(ac->avail > 0 || !l3);
3091 spin_lock(&l3->list_lock);
3092
3093 /* See if we can refill from the shared array */
3094 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3095 l3->shared->touched = 1;
3096 goto alloc_done;
3097 }
3098
3099 while (batchcount > 0) {
3100 struct list_head *entry;
3101 struct slab *slabp;
3102 /* Get slab alloc is to come from. */
3103 entry = l3->slabs_partial.next;
3104 if (entry == &l3->slabs_partial) {
3105 l3->free_touched = 1;
3106 entry = l3->slabs_free.next;
3107 if (entry == &l3->slabs_free)
3108 goto must_grow;
3109 }
3110
3111 slabp = list_entry(entry, struct slab, list);
3112 check_slabp(cachep, slabp);
3113 check_spinlock_acquired(cachep);
3114
3115 /*
3116 * The slab was either on partial or free list so
3117 * there must be at least one object available for
3118 * allocation.
3119 */
3120 BUG_ON(slabp->inuse >= cachep->num);
3121
3122 while (slabp->inuse < cachep->num && batchcount--) {
3123 STATS_INC_ALLOCED(cachep);
3124 STATS_INC_ACTIVE(cachep);
3125 STATS_SET_HIGH(cachep);
3126
3127 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3128 node);
3129 }
3130 check_slabp(cachep, slabp);
3131
3132 /* move slabp to correct slabp list: */
3133 list_del(&slabp->list);
3134 if (slabp->free == BUFCTL_END)
3135 list_add(&slabp->list, &l3->slabs_full);
3136 else
3137 list_add(&slabp->list, &l3->slabs_partial);
3138 }
3139
3140 must_grow:
3141 l3->free_objects -= ac->avail;
3142 alloc_done:
3143 spin_unlock(&l3->list_lock);
3144
3145 if (unlikely(!ac->avail)) {
3146 int x;
3147 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3148
3149 /* cache_grow can reenable interrupts, then ac could change. */
3150 ac = cpu_cache_get(cachep);
3151 if (!x && ac->avail == 0) /* no objects in sight? abort */
3152 return NULL;
3153
3154 if (!ac->avail) /* objects refilled by interrupt? */
3155 goto retry;
3156 }
3157 ac->touched = 1;
3158 return ac->entry[--ac->avail];
3159 }
3160
3161 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3162 gfp_t flags)
3163 {
3164 might_sleep_if(flags & __GFP_WAIT);
3165 #if DEBUG
3166 kmem_flagcheck(cachep, flags);
3167 #endif
3168 }
3169
3170 #if DEBUG
3171 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3172 gfp_t flags, void *objp, void *caller)
3173 {
3174 if (!objp)
3175 return objp;
3176 if (cachep->flags & SLAB_POISON) {
3177 #ifdef CONFIG_DEBUG_PAGEALLOC
3178 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3179 kernel_map_pages(virt_to_page(objp),
3180 cachep->buffer_size / PAGE_SIZE, 1);
3181 else
3182 check_poison_obj(cachep, objp);
3183 #else
3184 check_poison_obj(cachep, objp);
3185 #endif
3186 poison_obj(cachep, objp, POISON_INUSE);
3187 }
3188 if (cachep->flags & SLAB_STORE_USER)
3189 *dbg_userword(cachep, objp) = caller;
3190
3191 if (cachep->flags & SLAB_RED_ZONE) {
3192 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3193 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3194 slab_error(cachep, "double free, or memory outside"
3195 " object was overwritten");
3196 printk(KERN_ERR
3197 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3198 objp, *dbg_redzone1(cachep, objp),
3199 *dbg_redzone2(cachep, objp));
3200 }
3201 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3202 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3203 }
3204 #ifdef CONFIG_DEBUG_SLAB_LEAK
3205 {
3206 struct slab *slabp;
3207 unsigned objnr;
3208
3209 slabp = page_get_slab(virt_to_head_page(objp));
3210 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3211 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3212 }
3213 #endif
3214 objp += obj_offset(cachep);
3215 if (cachep->ctor && cachep->flags & SLAB_POISON)
3216 cachep->ctor(objp);
3217 if (ARCH_SLAB_MINALIGN &&
3218 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3219 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3220 objp, (int)ARCH_SLAB_MINALIGN);
3221 }
3222 return objp;
3223 }
3224 #else
3225 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3226 #endif
3227
3228 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3229 {
3230 if (cachep == &cache_cache)
3231 return false;
3232
3233 return should_failslab(obj_size(cachep), flags, cachep->flags);
3234 }
3235
3236 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3237 {
3238 void *objp;
3239 struct array_cache *ac;
3240
3241 check_irq_off();
3242
3243 ac = cpu_cache_get(cachep);
3244 if (likely(ac->avail)) {
3245 STATS_INC_ALLOCHIT(cachep);
3246 ac->touched = 1;
3247 objp = ac->entry[--ac->avail];
3248 } else {
3249 STATS_INC_ALLOCMISS(cachep);
3250 objp = cache_alloc_refill(cachep, flags);
3251 /*
3252 * the 'ac' may be updated by cache_alloc_refill(),
3253 * and kmemleak_erase() requires its correct value.
3254 */
3255 ac = cpu_cache_get(cachep);
3256 }
3257 /*
3258 * To avoid a false negative, if an object that is in one of the
3259 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3260 * treat the array pointers as a reference to the object.
3261 */
3262 if (objp)
3263 kmemleak_erase(&ac->entry[ac->avail]);
3264 return objp;
3265 }
3266
3267 #ifdef CONFIG_NUMA
3268 /*
3269 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3270 *
3271 * If we are in_interrupt, then process context, including cpusets and
3272 * mempolicy, may not apply and should not be used for allocation policy.
3273 */
3274 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3275 {
3276 int nid_alloc, nid_here;
3277
3278 if (in_interrupt() || (flags & __GFP_THISNODE))
3279 return NULL;
3280 nid_alloc = nid_here = numa_mem_id();
3281 get_mems_allowed();
3282 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3283 nid_alloc = cpuset_slab_spread_node();
3284 else if (current->mempolicy)
3285 nid_alloc = slab_node(current->mempolicy);
3286 put_mems_allowed();
3287 if (nid_alloc != nid_here)
3288 return ____cache_alloc_node(cachep, flags, nid_alloc);
3289 return NULL;
3290 }
3291
3292 /*
3293 * Fallback function if there was no memory available and no objects on a
3294 * certain node and fall back is permitted. First we scan all the
3295 * available nodelists for available objects. If that fails then we
3296 * perform an allocation without specifying a node. This allows the page
3297 * allocator to do its reclaim / fallback magic. We then insert the
3298 * slab into the proper nodelist and then allocate from it.
3299 */
3300 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3301 {
3302 struct zonelist *zonelist;
3303 gfp_t local_flags;
3304 struct zoneref *z;
3305 struct zone *zone;
3306 enum zone_type high_zoneidx = gfp_zone(flags);
3307 void *obj = NULL;
3308 int nid;
3309
3310 if (flags & __GFP_THISNODE)
3311 return NULL;
3312
3313 get_mems_allowed();
3314 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3315 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3316
3317 retry:
3318 /*
3319 * Look through allowed nodes for objects available
3320 * from existing per node queues.
3321 */
3322 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3323 nid = zone_to_nid(zone);
3324
3325 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3326 cache->nodelists[nid] &&
3327 cache->nodelists[nid]->free_objects) {
3328 obj = ____cache_alloc_node(cache,
3329 flags | GFP_THISNODE, nid);
3330 if (obj)
3331 break;
3332 }
3333 }
3334
3335 if (!obj) {
3336 /*
3337 * This allocation will be performed within the constraints
3338 * of the current cpuset / memory policy requirements.
3339 * We may trigger various forms of reclaim on the allowed
3340 * set and go into memory reserves if necessary.
3341 */
3342 if (local_flags & __GFP_WAIT)
3343 local_irq_enable();
3344 kmem_flagcheck(cache, flags);
3345 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3346 if (local_flags & __GFP_WAIT)
3347 local_irq_disable();
3348 if (obj) {
3349 /*
3350 * Insert into the appropriate per node queues
3351 */
3352 nid = page_to_nid(virt_to_page(obj));
3353 if (cache_grow(cache, flags, nid, obj)) {
3354 obj = ____cache_alloc_node(cache,
3355 flags | GFP_THISNODE, nid);
3356 if (!obj)
3357 /*
3358 * Another processor may allocate the
3359 * objects in the slab since we are
3360 * not holding any locks.
3361 */
3362 goto retry;
3363 } else {
3364 /* cache_grow already freed obj */
3365 obj = NULL;
3366 }
3367 }
3368 }
3369 put_mems_allowed();
3370 return obj;
3371 }
3372
3373 /*
3374 * A interface to enable slab creation on nodeid
3375 */
3376 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3377 int nodeid)
3378 {
3379 struct list_head *entry;
3380 struct slab *slabp;
3381 struct kmem_list3 *l3;
3382 void *obj;
3383 int x;
3384
3385 l3 = cachep->nodelists[nodeid];
3386 BUG_ON(!l3);
3387
3388 retry:
3389 check_irq_off();
3390 spin_lock(&l3->list_lock);
3391 entry = l3->slabs_partial.next;
3392 if (entry == &l3->slabs_partial) {
3393 l3->free_touched = 1;
3394 entry = l3->slabs_free.next;
3395 if (entry == &l3->slabs_free)
3396 goto must_grow;
3397 }
3398
3399 slabp = list_entry(entry, struct slab, list);
3400 check_spinlock_acquired_node(cachep, nodeid);
3401 check_slabp(cachep, slabp);
3402
3403 STATS_INC_NODEALLOCS(cachep);
3404 STATS_INC_ACTIVE(cachep);
3405 STATS_SET_HIGH(cachep);
3406
3407 BUG_ON(slabp->inuse == cachep->num);
3408
3409 obj = slab_get_obj(cachep, slabp, nodeid);
3410 check_slabp(cachep, slabp);
3411 l3->free_objects--;
3412 /* move slabp to correct slabp list: */
3413 list_del(&slabp->list);
3414
3415 if (slabp->free == BUFCTL_END)
3416 list_add(&slabp->list, &l3->slabs_full);
3417 else
3418 list_add(&slabp->list, &l3->slabs_partial);
3419
3420 spin_unlock(&l3->list_lock);
3421 goto done;
3422
3423 must_grow:
3424 spin_unlock(&l3->list_lock);
3425 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3426 if (x)
3427 goto retry;
3428
3429 return fallback_alloc(cachep, flags);
3430
3431 done:
3432 return obj;
3433 }
3434
3435 /**
3436 * kmem_cache_alloc_node - Allocate an object on the specified node
3437 * @cachep: The cache to allocate from.
3438 * @flags: See kmalloc().
3439 * @nodeid: node number of the target node.
3440 * @caller: return address of caller, used for debug information
3441 *
3442 * Identical to kmem_cache_alloc but it will allocate memory on the given
3443 * node, which can improve the performance for cpu bound structures.
3444 *
3445 * Fallback to other node is possible if __GFP_THISNODE is not set.
3446 */
3447 static __always_inline void *
3448 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3449 void *caller)
3450 {
3451 unsigned long save_flags;
3452 void *ptr;
3453 int slab_node = numa_mem_id();
3454
3455 flags &= gfp_allowed_mask;
3456
3457 lockdep_trace_alloc(flags);
3458
3459 if (slab_should_failslab(cachep, flags))
3460 return NULL;
3461
3462 cache_alloc_debugcheck_before(cachep, flags);
3463 local_irq_save(save_flags);
3464
3465 if (nodeid == NUMA_NO_NODE)
3466 nodeid = slab_node;
3467
3468 if (unlikely(!cachep->nodelists[nodeid])) {
3469 /* Node not bootstrapped yet */
3470 ptr = fallback_alloc(cachep, flags);
3471 goto out;
3472 }
3473
3474 if (nodeid == slab_node) {
3475 /*
3476 * Use the locally cached objects if possible.
3477 * However ____cache_alloc does not allow fallback
3478 * to other nodes. It may fail while we still have
3479 * objects on other nodes available.
3480 */
3481 ptr = ____cache_alloc(cachep, flags);
3482 if (ptr)
3483 goto out;
3484 }
3485 /* ___cache_alloc_node can fall back to other nodes */
3486 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3487 out:
3488 local_irq_restore(save_flags);
3489 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3490 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3491 flags);
3492
3493 if (likely(ptr))
3494 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3495
3496 if (unlikely((flags & __GFP_ZERO) && ptr))
3497 memset(ptr, 0, obj_size(cachep));
3498
3499 return ptr;
3500 }
3501
3502 static __always_inline void *
3503 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3504 {
3505 void *objp;
3506
3507 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3508 objp = alternate_node_alloc(cache, flags);
3509 if (objp)
3510 goto out;
3511 }
3512 objp = ____cache_alloc(cache, flags);
3513
3514 /*
3515 * We may just have run out of memory on the local node.
3516 * ____cache_alloc_node() knows how to locate memory on other nodes
3517 */
3518 if (!objp)
3519 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3520
3521 out:
3522 return objp;
3523 }
3524 #else
3525
3526 static __always_inline void *
3527 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3528 {
3529 return ____cache_alloc(cachep, flags);
3530 }
3531
3532 #endif /* CONFIG_NUMA */
3533
3534 static __always_inline void *
3535 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3536 {
3537 unsigned long save_flags;
3538 void *objp;
3539
3540 flags &= gfp_allowed_mask;
3541
3542 lockdep_trace_alloc(flags);
3543
3544 if (slab_should_failslab(cachep, flags))
3545 return NULL;
3546
3547 cache_alloc_debugcheck_before(cachep, flags);
3548 local_irq_save(save_flags);
3549 objp = __do_cache_alloc(cachep, flags);
3550 local_irq_restore(save_flags);
3551 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3552 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3553 flags);
3554 prefetchw(objp);
3555
3556 if (likely(objp))
3557 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3558
3559 if (unlikely((flags & __GFP_ZERO) && objp))
3560 memset(objp, 0, obj_size(cachep));
3561
3562 return objp;
3563 }
3564
3565 /*
3566 * Caller needs to acquire correct kmem_list's list_lock
3567 */
3568 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3569 int node)
3570 {
3571 int i;
3572 struct kmem_list3 *l3;
3573
3574 for (i = 0; i < nr_objects; i++) {
3575 void *objp = objpp[i];
3576 struct slab *slabp;
3577
3578 slabp = virt_to_slab(objp);
3579 l3 = cachep->nodelists[node];
3580 list_del(&slabp->list);
3581 check_spinlock_acquired_node(cachep, node);
3582 check_slabp(cachep, slabp);
3583 slab_put_obj(cachep, slabp, objp, node);
3584 STATS_DEC_ACTIVE(cachep);
3585 l3->free_objects++;
3586 check_slabp(cachep, slabp);
3587
3588 /* fixup slab chains */
3589 if (slabp->inuse == 0) {
3590 if (l3->free_objects > l3->free_limit) {
3591 l3->free_objects -= cachep->num;
3592 /* No need to drop any previously held
3593 * lock here, even if we have a off-slab slab
3594 * descriptor it is guaranteed to come from
3595 * a different cache, refer to comments before
3596 * alloc_slabmgmt.
3597 */
3598 slab_destroy(cachep, slabp);
3599 } else {
3600 list_add(&slabp->list, &l3->slabs_free);
3601 }
3602 } else {
3603 /* Unconditionally move a slab to the end of the
3604 * partial list on free - maximum time for the
3605 * other objects to be freed, too.
3606 */
3607 list_add_tail(&slabp->list, &l3->slabs_partial);
3608 }
3609 }
3610 }
3611
3612 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3613 {
3614 int batchcount;
3615 struct kmem_list3 *l3;
3616 int node = numa_mem_id();
3617
3618 batchcount = ac->batchcount;
3619 #if DEBUG
3620 BUG_ON(!batchcount || batchcount > ac->avail);
3621 #endif
3622 check_irq_off();
3623 l3 = cachep->nodelists[node];
3624 spin_lock(&l3->list_lock);
3625 if (l3->shared) {
3626 struct array_cache *shared_array = l3->shared;
3627 int max = shared_array->limit - shared_array->avail;
3628 if (max) {
3629 if (batchcount > max)
3630 batchcount = max;
3631 memcpy(&(shared_array->entry[shared_array->avail]),
3632 ac->entry, sizeof(void *) * batchcount);
3633 shared_array->avail += batchcount;
3634 goto free_done;
3635 }
3636 }
3637
3638 free_block(cachep, ac->entry, batchcount, node);
3639 free_done:
3640 #if STATS
3641 {
3642 int i = 0;
3643 struct list_head *p;
3644
3645 p = l3->slabs_free.next;
3646 while (p != &(l3->slabs_free)) {
3647 struct slab *slabp;
3648
3649 slabp = list_entry(p, struct slab, list);
3650 BUG_ON(slabp->inuse);
3651
3652 i++;
3653 p = p->next;
3654 }
3655 STATS_SET_FREEABLE(cachep, i);
3656 }
3657 #endif
3658 spin_unlock(&l3->list_lock);
3659 ac->avail -= batchcount;
3660 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3661 }
3662
3663 /*
3664 * Release an obj back to its cache. If the obj has a constructed state, it must
3665 * be in this state _before_ it is released. Called with disabled ints.
3666 */
3667 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3668 void *caller)
3669 {
3670 struct array_cache *ac = cpu_cache_get(cachep);
3671
3672 check_irq_off();
3673 kmemleak_free_recursive(objp, cachep->flags);
3674 objp = cache_free_debugcheck(cachep, objp, caller);
3675
3676 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3677
3678 /*
3679 * Skip calling cache_free_alien() when the platform is not numa.
3680 * This will avoid cache misses that happen while accessing slabp (which
3681 * is per page memory reference) to get nodeid. Instead use a global
3682 * variable to skip the call, which is mostly likely to be present in
3683 * the cache.
3684 */
3685 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3686 return;
3687
3688 if (likely(ac->avail < ac->limit)) {
3689 STATS_INC_FREEHIT(cachep);
3690 ac->entry[ac->avail++] = objp;
3691 return;
3692 } else {
3693 STATS_INC_FREEMISS(cachep);
3694 cache_flusharray(cachep, ac);
3695 ac->entry[ac->avail++] = objp;
3696 }
3697 }
3698
3699 /**
3700 * kmem_cache_alloc - Allocate an object
3701 * @cachep: The cache to allocate from.
3702 * @flags: See kmalloc().
3703 *
3704 * Allocate an object from this cache. The flags are only relevant
3705 * if the cache has no available objects.
3706 */
3707 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3708 {
3709 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3710
3711 trace_kmem_cache_alloc(_RET_IP_, ret,
3712 obj_size(cachep), cachep->buffer_size, flags);
3713
3714 return ret;
3715 }
3716 EXPORT_SYMBOL(kmem_cache_alloc);
3717
3718 #ifdef CONFIG_TRACING
3719 void *
3720 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3721 {
3722 void *ret;
3723
3724 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3725
3726 trace_kmalloc(_RET_IP_, ret,
3727 size, slab_buffer_size(cachep), flags);
3728 return ret;
3729 }
3730 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3731 #endif
3732
3733 #ifdef CONFIG_NUMA
3734 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3735 {
3736 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3737 __builtin_return_address(0));
3738
3739 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3740 obj_size(cachep), cachep->buffer_size,
3741 flags, nodeid);
3742
3743 return ret;
3744 }
3745 EXPORT_SYMBOL(kmem_cache_alloc_node);
3746
3747 #ifdef CONFIG_TRACING
3748 void *kmem_cache_alloc_node_trace(size_t size,
3749 struct kmem_cache *cachep,
3750 gfp_t flags,
3751 int nodeid)
3752 {
3753 void *ret;
3754
3755 ret = __cache_alloc_node(cachep, flags, nodeid,
3756 __builtin_return_address(0));
3757 trace_kmalloc_node(_RET_IP_, ret,
3758 size, slab_buffer_size(cachep),
3759 flags, nodeid);
3760 return ret;
3761 }
3762 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3763 #endif
3764
3765 static __always_inline void *
3766 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3767 {
3768 struct kmem_cache *cachep;
3769
3770 cachep = kmem_find_general_cachep(size, flags);
3771 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3772 return cachep;
3773 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3774 }
3775
3776 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3777 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3778 {
3779 return __do_kmalloc_node(size, flags, node,
3780 __builtin_return_address(0));
3781 }
3782 EXPORT_SYMBOL(__kmalloc_node);
3783
3784 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3785 int node, unsigned long caller)
3786 {
3787 return __do_kmalloc_node(size, flags, node, (void *)caller);
3788 }
3789 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3790 #else
3791 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3792 {
3793 return __do_kmalloc_node(size, flags, node, NULL);
3794 }
3795 EXPORT_SYMBOL(__kmalloc_node);
3796 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3797 #endif /* CONFIG_NUMA */
3798
3799 /**
3800 * __do_kmalloc - allocate memory
3801 * @size: how many bytes of memory are required.
3802 * @flags: the type of memory to allocate (see kmalloc).
3803 * @caller: function caller for debug tracking of the caller
3804 */
3805 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3806 void *caller)
3807 {
3808 struct kmem_cache *cachep;
3809 void *ret;
3810
3811 /* If you want to save a few bytes .text space: replace
3812 * __ with kmem_.
3813 * Then kmalloc uses the uninlined functions instead of the inline
3814 * functions.
3815 */
3816 cachep = __find_general_cachep(size, flags);
3817 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3818 return cachep;
3819 ret = __cache_alloc(cachep, flags, caller);
3820
3821 trace_kmalloc((unsigned long) caller, ret,
3822 size, cachep->buffer_size, flags);
3823
3824 return ret;
3825 }
3826
3827
3828 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3829 void *__kmalloc(size_t size, gfp_t flags)
3830 {
3831 return __do_kmalloc(size, flags, __builtin_return_address(0));
3832 }
3833 EXPORT_SYMBOL(__kmalloc);
3834
3835 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3836 {
3837 return __do_kmalloc(size, flags, (void *)caller);
3838 }
3839 EXPORT_SYMBOL(__kmalloc_track_caller);
3840
3841 #else
3842 void *__kmalloc(size_t size, gfp_t flags)
3843 {
3844 return __do_kmalloc(size, flags, NULL);
3845 }
3846 EXPORT_SYMBOL(__kmalloc);
3847 #endif
3848
3849 /**
3850 * kmem_cache_free - Deallocate an object
3851 * @cachep: The cache the allocation was from.
3852 * @objp: The previously allocated object.
3853 *
3854 * Free an object which was previously allocated from this
3855 * cache.
3856 */
3857 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3858 {
3859 unsigned long flags;
3860
3861 local_irq_save(flags);
3862 debug_check_no_locks_freed(objp, obj_size(cachep));
3863 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3864 debug_check_no_obj_freed(objp, obj_size(cachep));
3865 __cache_free(cachep, objp, __builtin_return_address(0));
3866 local_irq_restore(flags);
3867
3868 trace_kmem_cache_free(_RET_IP_, objp);
3869 }
3870 EXPORT_SYMBOL(kmem_cache_free);
3871
3872 /**
3873 * kfree - free previously allocated memory
3874 * @objp: pointer returned by kmalloc.
3875 *
3876 * If @objp is NULL, no operation is performed.
3877 *
3878 * Don't free memory not originally allocated by kmalloc()
3879 * or you will run into trouble.
3880 */
3881 void kfree(const void *objp)
3882 {
3883 struct kmem_cache *c;
3884 unsigned long flags;
3885
3886 trace_kfree(_RET_IP_, objp);
3887
3888 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3889 return;
3890 local_irq_save(flags);
3891 kfree_debugcheck(objp);
3892 c = virt_to_cache(objp);
3893 debug_check_no_locks_freed(objp, obj_size(c));
3894 debug_check_no_obj_freed(objp, obj_size(c));
3895 __cache_free(c, (void *)objp, __builtin_return_address(0));
3896 local_irq_restore(flags);
3897 }
3898 EXPORT_SYMBOL(kfree);
3899
3900 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3901 {
3902 return obj_size(cachep);
3903 }
3904 EXPORT_SYMBOL(kmem_cache_size);
3905
3906 /*
3907 * This initializes kmem_list3 or resizes various caches for all nodes.
3908 */
3909 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3910 {
3911 int node;
3912 struct kmem_list3 *l3;
3913 struct array_cache *new_shared;
3914 struct array_cache **new_alien = NULL;
3915
3916 for_each_online_node(node) {
3917
3918 if (use_alien_caches) {
3919 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3920 if (!new_alien)
3921 goto fail;
3922 }
3923
3924 new_shared = NULL;
3925 if (cachep->shared) {
3926 new_shared = alloc_arraycache(node,
3927 cachep->shared*cachep->batchcount,
3928 0xbaadf00d, gfp);
3929 if (!new_shared) {
3930 free_alien_cache(new_alien);
3931 goto fail;
3932 }
3933 }
3934
3935 l3 = cachep->nodelists[node];
3936 if (l3) {
3937 struct array_cache *shared = l3->shared;
3938
3939 spin_lock_irq(&l3->list_lock);
3940
3941 if (shared)
3942 free_block(cachep, shared->entry,
3943 shared->avail, node);
3944
3945 l3->shared = new_shared;
3946 if (!l3->alien) {
3947 l3->alien = new_alien;
3948 new_alien = NULL;
3949 }
3950 l3->free_limit = (1 + nr_cpus_node(node)) *
3951 cachep->batchcount + cachep->num;
3952 spin_unlock_irq(&l3->list_lock);
3953 kfree(shared);
3954 free_alien_cache(new_alien);
3955 continue;
3956 }
3957 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3958 if (!l3) {
3959 free_alien_cache(new_alien);
3960 kfree(new_shared);
3961 goto fail;
3962 }
3963
3964 kmem_list3_init(l3);
3965 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3966 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3967 l3->shared = new_shared;
3968 l3->alien = new_alien;
3969 l3->free_limit = (1 + nr_cpus_node(node)) *
3970 cachep->batchcount + cachep->num;
3971 cachep->nodelists[node] = l3;
3972 }
3973 return 0;
3974
3975 fail:
3976 if (!cachep->next.next) {
3977 /* Cache is not active yet. Roll back what we did */
3978 node--;
3979 while (node >= 0) {
3980 if (cachep->nodelists[node]) {
3981 l3 = cachep->nodelists[node];
3982
3983 kfree(l3->shared);
3984 free_alien_cache(l3->alien);
3985 kfree(l3);
3986 cachep->nodelists[node] = NULL;
3987 }
3988 node--;
3989 }
3990 }
3991 return -ENOMEM;
3992 }
3993
3994 struct ccupdate_struct {
3995 struct kmem_cache *cachep;
3996 struct array_cache *new[0];
3997 };
3998
3999 static void do_ccupdate_local(void *info)
4000 {
4001 struct ccupdate_struct *new = info;
4002 struct array_cache *old;
4003
4004 check_irq_off();
4005 old = cpu_cache_get(new->cachep);
4006
4007 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4008 new->new[smp_processor_id()] = old;
4009 }
4010
4011 /* Always called with the cache_chain_mutex held */
4012 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4013 int batchcount, int shared, gfp_t gfp)
4014 {
4015 struct ccupdate_struct *new;
4016 int i;
4017
4018 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4019 gfp);
4020 if (!new)
4021 return -ENOMEM;
4022
4023 for_each_online_cpu(i) {
4024 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4025 batchcount, gfp);
4026 if (!new->new[i]) {
4027 for (i--; i >= 0; i--)
4028 kfree(new->new[i]);
4029 kfree(new);
4030 return -ENOMEM;
4031 }
4032 }
4033 new->cachep = cachep;
4034
4035 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4036
4037 check_irq_on();
4038 cachep->batchcount = batchcount;
4039 cachep->limit = limit;
4040 cachep->shared = shared;
4041
4042 for_each_online_cpu(i) {
4043 struct array_cache *ccold = new->new[i];
4044 if (!ccold)
4045 continue;
4046 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4047 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4048 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4049 kfree(ccold);
4050 }
4051 kfree(new);
4052 return alloc_kmemlist(cachep, gfp);
4053 }
4054
4055 /* Called with cache_chain_mutex held always */
4056 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4057 {
4058 int err;
4059 int limit, shared;
4060
4061 /*
4062 * The head array serves three purposes:
4063 * - create a LIFO ordering, i.e. return objects that are cache-warm
4064 * - reduce the number of spinlock operations.
4065 * - reduce the number of linked list operations on the slab and
4066 * bufctl chains: array operations are cheaper.
4067 * The numbers are guessed, we should auto-tune as described by
4068 * Bonwick.
4069 */
4070 if (cachep->buffer_size > 131072)
4071 limit = 1;
4072 else if (cachep->buffer_size > PAGE_SIZE)
4073 limit = 8;
4074 else if (cachep->buffer_size > 1024)
4075 limit = 24;
4076 else if (cachep->buffer_size > 256)
4077 limit = 54;
4078 else
4079 limit = 120;
4080
4081 /*
4082 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4083 * allocation behaviour: Most allocs on one cpu, most free operations
4084 * on another cpu. For these cases, an efficient object passing between
4085 * cpus is necessary. This is provided by a shared array. The array
4086 * replaces Bonwick's magazine layer.
4087 * On uniprocessor, it's functionally equivalent (but less efficient)
4088 * to a larger limit. Thus disabled by default.
4089 */
4090 shared = 0;
4091 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4092 shared = 8;
4093
4094 #if DEBUG
4095 /*
4096 * With debugging enabled, large batchcount lead to excessively long
4097 * periods with disabled local interrupts. Limit the batchcount
4098 */
4099 if (limit > 32)
4100 limit = 32;
4101 #endif
4102 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4103 if (err)
4104 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4105 cachep->name, -err);
4106 return err;
4107 }
4108
4109 /*
4110 * Drain an array if it contains any elements taking the l3 lock only if
4111 * necessary. Note that the l3 listlock also protects the array_cache
4112 * if drain_array() is used on the shared array.
4113 */
4114 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4115 struct array_cache *ac, int force, int node)
4116 {
4117 int tofree;
4118
4119 if (!ac || !ac->avail)
4120 return;
4121 if (ac->touched && !force) {
4122 ac->touched = 0;
4123 } else {
4124 spin_lock_irq(&l3->list_lock);
4125 if (ac->avail) {
4126 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4127 if (tofree > ac->avail)
4128 tofree = (ac->avail + 1) / 2;
4129 free_block(cachep, ac->entry, tofree, node);
4130 ac->avail -= tofree;
4131 memmove(ac->entry, &(ac->entry[tofree]),
4132 sizeof(void *) * ac->avail);
4133 }
4134 spin_unlock_irq(&l3->list_lock);
4135 }
4136 }
4137
4138 /**
4139 * cache_reap - Reclaim memory from caches.
4140 * @w: work descriptor
4141 *
4142 * Called from workqueue/eventd every few seconds.
4143 * Purpose:
4144 * - clear the per-cpu caches for this CPU.
4145 * - return freeable pages to the main free memory pool.
4146 *
4147 * If we cannot acquire the cache chain mutex then just give up - we'll try
4148 * again on the next iteration.
4149 */
4150 static void cache_reap(struct work_struct *w)
4151 {
4152 struct kmem_cache *searchp;
4153 struct kmem_list3 *l3;
4154 int node = numa_mem_id();
4155 struct delayed_work *work = to_delayed_work(w);
4156
4157 if (!mutex_trylock(&cache_chain_mutex))
4158 /* Give up. Setup the next iteration. */
4159 goto out;
4160
4161 list_for_each_entry(searchp, &cache_chain, next) {
4162 check_irq_on();
4163
4164 /*
4165 * We only take the l3 lock if absolutely necessary and we
4166 * have established with reasonable certainty that
4167 * we can do some work if the lock was obtained.
4168 */
4169 l3 = searchp->nodelists[node];
4170
4171 reap_alien(searchp, l3);
4172
4173 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4174
4175 /*
4176 * These are racy checks but it does not matter
4177 * if we skip one check or scan twice.
4178 */
4179 if (time_after(l3->next_reap, jiffies))
4180 goto next;
4181
4182 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4183
4184 drain_array(searchp, l3, l3->shared, 0, node);
4185
4186 if (l3->free_touched)
4187 l3->free_touched = 0;
4188 else {
4189 int freed;
4190
4191 freed = drain_freelist(searchp, l3, (l3->free_limit +
4192 5 * searchp->num - 1) / (5 * searchp->num));
4193 STATS_ADD_REAPED(searchp, freed);
4194 }
4195 next:
4196 cond_resched();
4197 }
4198 check_irq_on();
4199 mutex_unlock(&cache_chain_mutex);
4200 next_reap_node();
4201 out:
4202 /* Set up the next iteration */
4203 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4204 }
4205
4206 #ifdef CONFIG_SLABINFO
4207
4208 static void print_slabinfo_header(struct seq_file *m)
4209 {
4210 /*
4211 * Output format version, so at least we can change it
4212 * without _too_ many complaints.
4213 */
4214 #if STATS
4215 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4216 #else
4217 seq_puts(m, "slabinfo - version: 2.1\n");
4218 #endif
4219 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4220 "<objperslab> <pagesperslab>");
4221 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4222 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4223 #if STATS
4224 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4225 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4226 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4227 #endif
4228 seq_putc(m, '\n');
4229 }
4230
4231 static void *s_start(struct seq_file *m, loff_t *pos)
4232 {
4233 loff_t n = *pos;
4234
4235 mutex_lock(&cache_chain_mutex);
4236 if (!n)
4237 print_slabinfo_header(m);
4238
4239 return seq_list_start(&cache_chain, *pos);
4240 }
4241
4242 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4243 {
4244 return seq_list_next(p, &cache_chain, pos);
4245 }
4246
4247 static void s_stop(struct seq_file *m, void *p)
4248 {
4249 mutex_unlock(&cache_chain_mutex);
4250 }
4251
4252 static int s_show(struct seq_file *m, void *p)
4253 {
4254 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4255 struct slab *slabp;
4256 unsigned long active_objs;
4257 unsigned long num_objs;
4258 unsigned long active_slabs = 0;
4259 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4260 const char *name;
4261 char *error = NULL;
4262 int node;
4263 struct kmem_list3 *l3;
4264
4265 active_objs = 0;
4266 num_slabs = 0;
4267 for_each_online_node(node) {
4268 l3 = cachep->nodelists[node];
4269 if (!l3)
4270 continue;
4271
4272 check_irq_on();
4273 spin_lock_irq(&l3->list_lock);
4274
4275 list_for_each_entry(slabp, &l3->slabs_full, list) {
4276 if (slabp->inuse != cachep->num && !error)
4277 error = "slabs_full accounting error";
4278 active_objs += cachep->num;
4279 active_slabs++;
4280 }
4281 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4282 if (slabp->inuse == cachep->num && !error)
4283 error = "slabs_partial inuse accounting error";
4284 if (!slabp->inuse && !error)
4285 error = "slabs_partial/inuse accounting error";
4286 active_objs += slabp->inuse;
4287 active_slabs++;
4288 }
4289 list_for_each_entry(slabp, &l3->slabs_free, list) {
4290 if (slabp->inuse && !error)
4291 error = "slabs_free/inuse accounting error";
4292 num_slabs++;
4293 }
4294 free_objects += l3->free_objects;
4295 if (l3->shared)
4296 shared_avail += l3->shared->avail;
4297
4298 spin_unlock_irq(&l3->list_lock);
4299 }
4300 num_slabs += active_slabs;
4301 num_objs = num_slabs * cachep->num;
4302 if (num_objs - active_objs != free_objects && !error)
4303 error = "free_objects accounting error";
4304
4305 name = cachep->name;
4306 if (error)
4307 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4308
4309 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4310 name, active_objs, num_objs, cachep->buffer_size,
4311 cachep->num, (1 << cachep->gfporder));
4312 seq_printf(m, " : tunables %4u %4u %4u",
4313 cachep->limit, cachep->batchcount, cachep->shared);
4314 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4315 active_slabs, num_slabs, shared_avail);
4316 #if STATS
4317 { /* list3 stats */
4318 unsigned long high = cachep->high_mark;
4319 unsigned long allocs = cachep->num_allocations;
4320 unsigned long grown = cachep->grown;
4321 unsigned long reaped = cachep->reaped;
4322 unsigned long errors = cachep->errors;
4323 unsigned long max_freeable = cachep->max_freeable;
4324 unsigned long node_allocs = cachep->node_allocs;
4325 unsigned long node_frees = cachep->node_frees;
4326 unsigned long overflows = cachep->node_overflow;
4327
4328 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4329 "%4lu %4lu %4lu %4lu %4lu",
4330 allocs, high, grown,
4331 reaped, errors, max_freeable, node_allocs,
4332 node_frees, overflows);
4333 }
4334 /* cpu stats */
4335 {
4336 unsigned long allochit = atomic_read(&cachep->allochit);
4337 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4338 unsigned long freehit = atomic_read(&cachep->freehit);
4339 unsigned long freemiss = atomic_read(&cachep->freemiss);
4340
4341 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4342 allochit, allocmiss, freehit, freemiss);
4343 }
4344 #endif
4345 seq_putc(m, '\n');
4346 return 0;
4347 }
4348
4349 /*
4350 * slabinfo_op - iterator that generates /proc/slabinfo
4351 *
4352 * Output layout:
4353 * cache-name
4354 * num-active-objs
4355 * total-objs
4356 * object size
4357 * num-active-slabs
4358 * total-slabs
4359 * num-pages-per-slab
4360 * + further values on SMP and with statistics enabled
4361 */
4362
4363 static const struct seq_operations slabinfo_op = {
4364 .start = s_start,
4365 .next = s_next,
4366 .stop = s_stop,
4367 .show = s_show,
4368 };
4369
4370 #define MAX_SLABINFO_WRITE 128
4371 /**
4372 * slabinfo_write - Tuning for the slab allocator
4373 * @file: unused
4374 * @buffer: user buffer
4375 * @count: data length
4376 * @ppos: unused
4377 */
4378 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4379 size_t count, loff_t *ppos)
4380 {
4381 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4382 int limit, batchcount, shared, res;
4383 struct kmem_cache *cachep;
4384
4385 if (count > MAX_SLABINFO_WRITE)
4386 return -EINVAL;
4387 if (copy_from_user(&kbuf, buffer, count))
4388 return -EFAULT;
4389 kbuf[MAX_SLABINFO_WRITE] = '\0';
4390
4391 tmp = strchr(kbuf, ' ');
4392 if (!tmp)
4393 return -EINVAL;
4394 *tmp = '\0';
4395 tmp++;
4396 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4397 return -EINVAL;
4398
4399 /* Find the cache in the chain of caches. */
4400 mutex_lock(&cache_chain_mutex);
4401 res = -EINVAL;
4402 list_for_each_entry(cachep, &cache_chain, next) {
4403 if (!strcmp(cachep->name, kbuf)) {
4404 if (limit < 1 || batchcount < 1 ||
4405 batchcount > limit || shared < 0) {
4406 res = 0;
4407 } else {
4408 res = do_tune_cpucache(cachep, limit,
4409 batchcount, shared,
4410 GFP_KERNEL);
4411 }
4412 break;
4413 }
4414 }
4415 mutex_unlock(&cache_chain_mutex);
4416 if (res >= 0)
4417 res = count;
4418 return res;
4419 }
4420
4421 static int slabinfo_open(struct inode *inode, struct file *file)
4422 {
4423 return seq_open(file, &slabinfo_op);
4424 }
4425
4426 static const struct file_operations proc_slabinfo_operations = {
4427 .open = slabinfo_open,
4428 .read = seq_read,
4429 .write = slabinfo_write,
4430 .llseek = seq_lseek,
4431 .release = seq_release,
4432 };
4433
4434 #ifdef CONFIG_DEBUG_SLAB_LEAK
4435
4436 static void *leaks_start(struct seq_file *m, loff_t *pos)
4437 {
4438 mutex_lock(&cache_chain_mutex);
4439 return seq_list_start(&cache_chain, *pos);
4440 }
4441
4442 static inline int add_caller(unsigned long *n, unsigned long v)
4443 {
4444 unsigned long *p;
4445 int l;
4446 if (!v)
4447 return 1;
4448 l = n[1];
4449 p = n + 2;
4450 while (l) {
4451 int i = l/2;
4452 unsigned long *q = p + 2 * i;
4453 if (*q == v) {
4454 q[1]++;
4455 return 1;
4456 }
4457 if (*q > v) {
4458 l = i;
4459 } else {
4460 p = q + 2;
4461 l -= i + 1;
4462 }
4463 }
4464 if (++n[1] == n[0])
4465 return 0;
4466 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4467 p[0] = v;
4468 p[1] = 1;
4469 return 1;
4470 }
4471
4472 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4473 {
4474 void *p;
4475 int i;
4476 if (n[0] == n[1])
4477 return;
4478 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4479 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4480 continue;
4481 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4482 return;
4483 }
4484 }
4485
4486 static void show_symbol(struct seq_file *m, unsigned long address)
4487 {
4488 #ifdef CONFIG_KALLSYMS
4489 unsigned long offset, size;
4490 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4491
4492 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4493 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4494 if (modname[0])
4495 seq_printf(m, " [%s]", modname);
4496 return;
4497 }
4498 #endif
4499 seq_printf(m, "%p", (void *)address);
4500 }
4501
4502 static int leaks_show(struct seq_file *m, void *p)
4503 {
4504 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4505 struct slab *slabp;
4506 struct kmem_list3 *l3;
4507 const char *name;
4508 unsigned long *n = m->private;
4509 int node;
4510 int i;
4511
4512 if (!(cachep->flags & SLAB_STORE_USER))
4513 return 0;
4514 if (!(cachep->flags & SLAB_RED_ZONE))
4515 return 0;
4516
4517 /* OK, we can do it */
4518
4519 n[1] = 0;
4520
4521 for_each_online_node(node) {
4522 l3 = cachep->nodelists[node];
4523 if (!l3)
4524 continue;
4525
4526 check_irq_on();
4527 spin_lock_irq(&l3->list_lock);
4528
4529 list_for_each_entry(slabp, &l3->slabs_full, list)
4530 handle_slab(n, cachep, slabp);
4531 list_for_each_entry(slabp, &l3->slabs_partial, list)
4532 handle_slab(n, cachep, slabp);
4533 spin_unlock_irq(&l3->list_lock);
4534 }
4535 name = cachep->name;
4536 if (n[0] == n[1]) {
4537 /* Increase the buffer size */
4538 mutex_unlock(&cache_chain_mutex);
4539 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4540 if (!m->private) {
4541 /* Too bad, we are really out */
4542 m->private = n;
4543 mutex_lock(&cache_chain_mutex);
4544 return -ENOMEM;
4545 }
4546 *(unsigned long *)m->private = n[0] * 2;
4547 kfree(n);
4548 mutex_lock(&cache_chain_mutex);
4549 /* Now make sure this entry will be retried */
4550 m->count = m->size;
4551 return 0;
4552 }
4553 for (i = 0; i < n[1]; i++) {
4554 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4555 show_symbol(m, n[2*i+2]);
4556 seq_putc(m, '\n');
4557 }
4558
4559 return 0;
4560 }
4561
4562 static const struct seq_operations slabstats_op = {
4563 .start = leaks_start,
4564 .next = s_next,
4565 .stop = s_stop,
4566 .show = leaks_show,
4567 };
4568
4569 static int slabstats_open(struct inode *inode, struct file *file)
4570 {
4571 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4572 int ret = -ENOMEM;
4573 if (n) {
4574 ret = seq_open(file, &slabstats_op);
4575 if (!ret) {
4576 struct seq_file *m = file->private_data;
4577 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4578 m->private = n;
4579 n = NULL;
4580 }
4581 kfree(n);
4582 }
4583 return ret;
4584 }
4585
4586 static const struct file_operations proc_slabstats_operations = {
4587 .open = slabstats_open,
4588 .read = seq_read,
4589 .llseek = seq_lseek,
4590 .release = seq_release_private,
4591 };
4592 #endif
4593
4594 static int __init slab_proc_init(void)
4595 {
4596 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4597 #ifdef CONFIG_DEBUG_SLAB_LEAK
4598 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4599 #endif
4600 return 0;
4601 }
4602 module_init(slab_proc_init);
4603 #endif
4604
4605 /**
4606 * ksize - get the actual amount of memory allocated for a given object
4607 * @objp: Pointer to the object
4608 *
4609 * kmalloc may internally round up allocations and return more memory
4610 * than requested. ksize() can be used to determine the actual amount of
4611 * memory allocated. The caller may use this additional memory, even though
4612 * a smaller amount of memory was initially specified with the kmalloc call.
4613 * The caller must guarantee that objp points to a valid object previously
4614 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4615 * must not be freed during the duration of the call.
4616 */
4617 size_t ksize(const void *objp)
4618 {
4619 BUG_ON(!objp);
4620 if (unlikely(objp == ZERO_SIZE_PTR))
4621 return 0;
4622
4623 return obj_size(virt_to_cache(objp));
4624 }
4625 EXPORT_SYMBOL(ksize);
This page took 0.133786 seconds and 6 git commands to generate.