]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/slab.c
a684778b2b41af38e06aa47e5fe66a7a988ad4c7
[mirror_ubuntu-bionic-kernel.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 intializations 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/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
113
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
117
118 /*
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128 #ifdef CONFIG_DEBUG_SLAB
129 #define DEBUG 1
130 #define STATS 1
131 #define FORCED_DEBUG 1
132 #else
133 #define DEBUG 0
134 #define STATS 0
135 #define FORCED_DEBUG 0
136 #endif
137
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
140 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
141
142 #ifndef cache_line_size
143 #define cache_line_size() L1_CACHE_BYTES
144 #endif
145
146 #ifndef ARCH_KMALLOC_MINALIGN
147 /*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
155 */
156 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
157 #endif
158
159 #ifndef ARCH_SLAB_MINALIGN
160 /*
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
166 */
167 #define ARCH_SLAB_MINALIGN 0
168 #endif
169
170 #ifndef ARCH_KMALLOC_FLAGS
171 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 #endif
173
174 /* Legal flag mask for kmem_cache_create(). */
175 #if DEBUG
176 # define CREATE_MASK (SLAB_RED_ZONE | \
177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
178 SLAB_CACHE_DMA | \
179 SLAB_STORE_USER | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 #else
183 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
184 SLAB_CACHE_DMA | \
185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
187 #endif
188
189 /*
190 * kmem_bufctl_t:
191 *
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
194 *
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 */
207
208 typedef unsigned int kmem_bufctl_t;
209 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
211 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
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 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
228 };
229
230 /*
231 * struct slab_rcu
232 *
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
240 *
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
243 *
244 * We assume struct slab_rcu can overlay struct slab when destroying.
245 */
246 struct slab_rcu {
247 struct rcu_head head;
248 struct kmem_cache *cachep;
249 void *addr;
250 };
251
252 /*
253 * struct array_cache
254 *
255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
259 *
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
262 *
263 */
264 struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
269 spinlock_t lock;
270 void *entry[0]; /*
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
274 * [0] is for gcc 2.95. It should really be [].
275 */
276 };
277
278 /*
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
281 */
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
286 };
287
288 /*
289 * The slab lists for all objects.
290 */
291 struct kmem_list3 {
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
303 };
304
305 /*
306 * Need this for bootstrapping a per node allocator.
307 */
308 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
309 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
311 #define SIZE_AC 1
312 #define SIZE_L3 (1 + MAX_NUMNODES)
313
314 static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 int node);
318 static int enable_cpucache(struct kmem_cache *cachep);
319 static void cache_reap(struct work_struct *unused);
320
321 /*
322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 */
325 static __always_inline int index_of(const size_t size)
326 {
327 extern void __bad_size(void);
328
329 if (__builtin_constant_p(size)) {
330 int i = 0;
331
332 #define CACHE(x) \
333 if (size <=x) \
334 return i; \
335 else \
336 i++;
337 #include "linux/kmalloc_sizes.h"
338 #undef CACHE
339 __bad_size();
340 } else
341 __bad_size();
342 return 0;
343 }
344
345 static int slab_early_init = 1;
346
347 #define INDEX_AC index_of(sizeof(struct arraycache_init))
348 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349
350 static void kmem_list3_init(struct kmem_list3 *parent)
351 {
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
357 parent->colour_next = 0;
358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
361 }
362
363 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 do { \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
367 } while (0)
368
369 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 do { \
371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 } while (0)
375
376 /*
377 * struct kmem_cache
378 *
379 * manages a cache.
380 */
381
382 struct kmem_cache {
383 /* 1) per-cpu data, touched during every alloc/free */
384 struct array_cache *array[NR_CPUS];
385 /* 2) Cache tunables. Protected by cache_chain_mutex */
386 unsigned int batchcount;
387 unsigned int limit;
388 unsigned int shared;
389
390 unsigned int buffer_size;
391 u32 reciprocal_buffer_size;
392 /* 3) touched by every alloc & free from the backend */
393
394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
396
397 /* 4) cache_grow/shrink */
398 /* order of pgs per slab (2^n) */
399 unsigned int gfporder;
400
401 /* force GFP flags, e.g. GFP_DMA */
402 gfp_t gfpflags;
403
404 size_t colour; /* cache colouring range */
405 unsigned int colour_off; /* colour offset */
406 struct kmem_cache *slabp_cache;
407 unsigned int slab_size;
408 unsigned int dflags; /* dynamic flags */
409
410 /* constructor func */
411 void (*ctor) (void *, struct kmem_cache *, unsigned long);
412
413 /* 5) cache creation/removal */
414 const char *name;
415 struct list_head next;
416
417 /* 6) statistics */
418 #if STATS
419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
422 unsigned long grown;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
428 unsigned long node_overflow;
429 atomic_t allochit;
430 atomic_t allocmiss;
431 atomic_t freehit;
432 atomic_t freemiss;
433 #endif
434 #if DEBUG
435 /*
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
440 */
441 int obj_offset;
442 int obj_size;
443 #endif
444 /*
445 * We put nodelists[] at the end of kmem_cache, because we want to size
446 * this array to nr_node_ids slots instead of MAX_NUMNODES
447 * (see kmem_cache_init())
448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
449 * is statically defined, so we reserve the max number of nodes.
450 */
451 struct kmem_list3 *nodelists[MAX_NUMNODES];
452 /*
453 * Do not add fields after nodelists[]
454 */
455 };
456
457 #define CFLGS_OFF_SLAB (0x80000000UL)
458 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459
460 #define BATCHREFILL_LIMIT 16
461 /*
462 * Optimization question: fewer reaps means less probability for unnessary
463 * cpucache drain/refill cycles.
464 *
465 * OTOH the cpuarrays can contain lots of objects,
466 * which could lock up otherwise freeable slabs.
467 */
468 #define REAPTIMEOUT_CPUC (2*HZ)
469 #define REAPTIMEOUT_LIST3 (4*HZ)
470
471 #if STATS
472 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
473 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
474 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
475 #define STATS_INC_GROWN(x) ((x)->grown++)
476 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
477 #define STATS_SET_HIGH(x) \
478 do { \
479 if ((x)->num_active > (x)->high_mark) \
480 (x)->high_mark = (x)->num_active; \
481 } while (0)
482 #define STATS_INC_ERR(x) ((x)->errors++)
483 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
484 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
485 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
486 #define STATS_SET_FREEABLE(x, i) \
487 do { \
488 if ((x)->max_freeable < i) \
489 (x)->max_freeable = i; \
490 } while (0)
491 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
492 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
493 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
494 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
495 #else
496 #define STATS_INC_ACTIVE(x) do { } while (0)
497 #define STATS_DEC_ACTIVE(x) do { } while (0)
498 #define STATS_INC_ALLOCED(x) do { } while (0)
499 #define STATS_INC_GROWN(x) do { } while (0)
500 #define STATS_ADD_REAPED(x,y) do { } while (0)
501 #define STATS_SET_HIGH(x) do { } while (0)
502 #define STATS_INC_ERR(x) do { } while (0)
503 #define STATS_INC_NODEALLOCS(x) do { } while (0)
504 #define STATS_INC_NODEFREES(x) do { } while (0)
505 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
506 #define STATS_SET_FREEABLE(x, i) do { } while (0)
507 #define STATS_INC_ALLOCHIT(x) do { } while (0)
508 #define STATS_INC_ALLOCMISS(x) do { } while (0)
509 #define STATS_INC_FREEHIT(x) do { } while (0)
510 #define STATS_INC_FREEMISS(x) do { } while (0)
511 #endif
512
513 #if DEBUG
514
515 /*
516 * memory layout of objects:
517 * 0 : objp
518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
519 * the end of an object is aligned with the end of the real
520 * allocation. Catches writes behind the end of the allocation.
521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
522 * redzone word.
523 * cachep->obj_offset: The real object.
524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
526 * [BYTES_PER_WORD long]
527 */
528 static int obj_offset(struct kmem_cache *cachep)
529 {
530 return cachep->obj_offset;
531 }
532
533 static int obj_size(struct kmem_cache *cachep)
534 {
535 return cachep->obj_size;
536 }
537
538 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
539 {
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
541 return (unsigned long long*) (objp + obj_offset(cachep) -
542 sizeof(unsigned long long));
543 }
544
545 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
546 {
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 if (cachep->flags & SLAB_STORE_USER)
549 return (unsigned long long *)(objp + cachep->buffer_size -
550 sizeof(unsigned long long) -
551 REDZONE_ALIGN);
552 return (unsigned long long *) (objp + cachep->buffer_size -
553 sizeof(unsigned long long));
554 }
555
556 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
557 {
558 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
560 }
561
562 #else
563
564 #define obj_offset(x) 0
565 #define obj_size(cachep) (cachep->buffer_size)
566 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
568 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
569
570 #endif
571
572 /*
573 * Do not go above this order unless 0 objects fit into the slab.
574 */
575 #define BREAK_GFP_ORDER_HI 1
576 #define BREAK_GFP_ORDER_LO 0
577 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
578
579 /*
580 * Functions for storing/retrieving the cachep and or slab from the page
581 * allocator. These are used to find the slab an obj belongs to. With kfree(),
582 * these are used to find the cache which an obj belongs to.
583 */
584 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
585 {
586 page->lru.next = (struct list_head *)cache;
587 }
588
589 static inline struct kmem_cache *page_get_cache(struct page *page)
590 {
591 page = compound_head(page);
592 BUG_ON(!PageSlab(page));
593 return (struct kmem_cache *)page->lru.next;
594 }
595
596 static inline void page_set_slab(struct page *page, struct slab *slab)
597 {
598 page->lru.prev = (struct list_head *)slab;
599 }
600
601 static inline struct slab *page_get_slab(struct page *page)
602 {
603 BUG_ON(!PageSlab(page));
604 return (struct slab *)page->lru.prev;
605 }
606
607 static inline struct kmem_cache *virt_to_cache(const void *obj)
608 {
609 struct page *page = virt_to_head_page(obj);
610 return page_get_cache(page);
611 }
612
613 static inline struct slab *virt_to_slab(const void *obj)
614 {
615 struct page *page = virt_to_head_page(obj);
616 return page_get_slab(page);
617 }
618
619 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
620 unsigned int idx)
621 {
622 return slab->s_mem + cache->buffer_size * idx;
623 }
624
625 /*
626 * We want to avoid an expensive divide : (offset / cache->buffer_size)
627 * Using the fact that buffer_size is a constant for a particular cache,
628 * we can replace (offset / cache->buffer_size) by
629 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
630 */
631 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
632 const struct slab *slab, void *obj)
633 {
634 u32 offset = (obj - slab->s_mem);
635 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
636 }
637
638 /*
639 * These are the default caches for kmalloc. Custom caches can have other sizes.
640 */
641 struct cache_sizes malloc_sizes[] = {
642 #define CACHE(x) { .cs_size = (x) },
643 #include <linux/kmalloc_sizes.h>
644 CACHE(ULONG_MAX)
645 #undef CACHE
646 };
647 EXPORT_SYMBOL(malloc_sizes);
648
649 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
650 struct cache_names {
651 char *name;
652 char *name_dma;
653 };
654
655 static struct cache_names __initdata cache_names[] = {
656 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
657 #include <linux/kmalloc_sizes.h>
658 {NULL,}
659 #undef CACHE
660 };
661
662 static struct arraycache_init initarray_cache __initdata =
663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
664 static struct arraycache_init initarray_generic =
665 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
666
667 /* internal cache of cache description objs */
668 static struct kmem_cache cache_cache = {
669 .batchcount = 1,
670 .limit = BOOT_CPUCACHE_ENTRIES,
671 .shared = 1,
672 .buffer_size = sizeof(struct kmem_cache),
673 .name = "kmem_cache",
674 };
675
676 #define BAD_ALIEN_MAGIC 0x01020304ul
677
678 #ifdef CONFIG_LOCKDEP
679
680 /*
681 * Slab sometimes uses the kmalloc slabs to store the slab headers
682 * for other slabs "off slab".
683 * The locking for this is tricky in that it nests within the locks
684 * of all other slabs in a few places; to deal with this special
685 * locking we put on-slab caches into a separate lock-class.
686 *
687 * We set lock class for alien array caches which are up during init.
688 * The lock annotation will be lost if all cpus of a node goes down and
689 * then comes back up during hotplug
690 */
691 static struct lock_class_key on_slab_l3_key;
692 static struct lock_class_key on_slab_alc_key;
693
694 static inline void init_lock_keys(void)
695
696 {
697 int q;
698 struct cache_sizes *s = malloc_sizes;
699
700 while (s->cs_size != ULONG_MAX) {
701 for_each_node(q) {
702 struct array_cache **alc;
703 int r;
704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
705 if (!l3 || OFF_SLAB(s->cs_cachep))
706 continue;
707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
708 alc = l3->alien;
709 /*
710 * FIXME: This check for BAD_ALIEN_MAGIC
711 * should go away when common slab code is taught to
712 * work even without alien caches.
713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
714 * for alloc_alien_cache,
715 */
716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
717 continue;
718 for_each_node(r) {
719 if (alc[r])
720 lockdep_set_class(&alc[r]->lock,
721 &on_slab_alc_key);
722 }
723 }
724 s++;
725 }
726 }
727 #else
728 static inline void init_lock_keys(void)
729 {
730 }
731 #endif
732
733 /*
734 * 1. Guard access to the cache-chain.
735 * 2. Protect sanity of cpu_online_map against cpu hotplug events
736 */
737 static DEFINE_MUTEX(cache_chain_mutex);
738 static struct list_head cache_chain;
739
740 /*
741 * chicken and egg problem: delay the per-cpu array allocation
742 * until the general caches are up.
743 */
744 static enum {
745 NONE,
746 PARTIAL_AC,
747 PARTIAL_L3,
748 FULL
749 } g_cpucache_up;
750
751 /*
752 * used by boot code to determine if it can use slab based allocator
753 */
754 int slab_is_available(void)
755 {
756 return g_cpucache_up == FULL;
757 }
758
759 static DEFINE_PER_CPU(struct delayed_work, reap_work);
760
761 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
762 {
763 return cachep->array[smp_processor_id()];
764 }
765
766 static inline struct kmem_cache *__find_general_cachep(size_t size,
767 gfp_t gfpflags)
768 {
769 struct cache_sizes *csizep = malloc_sizes;
770
771 #if DEBUG
772 /* This happens if someone tries to call
773 * kmem_cache_create(), or __kmalloc(), before
774 * the generic caches are initialized.
775 */
776 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
777 #endif
778 if (!size)
779 return ZERO_SIZE_PTR;
780
781 while (size > csizep->cs_size)
782 csizep++;
783
784 /*
785 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
786 * has cs_{dma,}cachep==NULL. Thus no special case
787 * for large kmalloc calls required.
788 */
789 #ifdef CONFIG_ZONE_DMA
790 if (unlikely(gfpflags & GFP_DMA))
791 return csizep->cs_dmacachep;
792 #endif
793 return csizep->cs_cachep;
794 }
795
796 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
797 {
798 return __find_general_cachep(size, gfpflags);
799 }
800
801 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
802 {
803 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
804 }
805
806 /*
807 * Calculate the number of objects and left-over bytes for a given buffer size.
808 */
809 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
810 size_t align, int flags, size_t *left_over,
811 unsigned int *num)
812 {
813 int nr_objs;
814 size_t mgmt_size;
815 size_t slab_size = PAGE_SIZE << gfporder;
816
817 /*
818 * The slab management structure can be either off the slab or
819 * on it. For the latter case, the memory allocated for a
820 * slab is used for:
821 *
822 * - The struct slab
823 * - One kmem_bufctl_t for each object
824 * - Padding to respect alignment of @align
825 * - @buffer_size bytes for each object
826 *
827 * If the slab management structure is off the slab, then the
828 * alignment will already be calculated into the size. Because
829 * the slabs are all pages aligned, the objects will be at the
830 * correct alignment when allocated.
831 */
832 if (flags & CFLGS_OFF_SLAB) {
833 mgmt_size = 0;
834 nr_objs = slab_size / buffer_size;
835
836 if (nr_objs > SLAB_LIMIT)
837 nr_objs = SLAB_LIMIT;
838 } else {
839 /*
840 * Ignore padding for the initial guess. The padding
841 * is at most @align-1 bytes, and @buffer_size is at
842 * least @align. In the worst case, this result will
843 * be one greater than the number of objects that fit
844 * into the memory allocation when taking the padding
845 * into account.
846 */
847 nr_objs = (slab_size - sizeof(struct slab)) /
848 (buffer_size + sizeof(kmem_bufctl_t));
849
850 /*
851 * This calculated number will be either the right
852 * amount, or one greater than what we want.
853 */
854 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
855 > slab_size)
856 nr_objs--;
857
858 if (nr_objs > SLAB_LIMIT)
859 nr_objs = SLAB_LIMIT;
860
861 mgmt_size = slab_mgmt_size(nr_objs, align);
862 }
863 *num = nr_objs;
864 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
865 }
866
867 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
868
869 static void __slab_error(const char *function, struct kmem_cache *cachep,
870 char *msg)
871 {
872 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
873 function, cachep->name, msg);
874 dump_stack();
875 }
876
877 /*
878 * By default on NUMA we use alien caches to stage the freeing of
879 * objects allocated from other nodes. This causes massive memory
880 * inefficiencies when using fake NUMA setup to split memory into a
881 * large number of small nodes, so it can be disabled on the command
882 * line
883 */
884
885 static int use_alien_caches __read_mostly = 1;
886 static int __init noaliencache_setup(char *s)
887 {
888 use_alien_caches = 0;
889 return 1;
890 }
891 __setup("noaliencache", noaliencache_setup);
892
893 #ifdef CONFIG_NUMA
894 /*
895 * Special reaping functions for NUMA systems called from cache_reap().
896 * These take care of doing round robin flushing of alien caches (containing
897 * objects freed on different nodes from which they were allocated) and the
898 * flushing of remote pcps by calling drain_node_pages.
899 */
900 static DEFINE_PER_CPU(unsigned long, reap_node);
901
902 static void init_reap_node(int cpu)
903 {
904 int node;
905
906 node = next_node(cpu_to_node(cpu), node_online_map);
907 if (node == MAX_NUMNODES)
908 node = first_node(node_online_map);
909
910 per_cpu(reap_node, cpu) = node;
911 }
912
913 static void next_reap_node(void)
914 {
915 int node = __get_cpu_var(reap_node);
916
917 node = next_node(node, node_online_map);
918 if (unlikely(node >= MAX_NUMNODES))
919 node = first_node(node_online_map);
920 __get_cpu_var(reap_node) = node;
921 }
922
923 #else
924 #define init_reap_node(cpu) do { } while (0)
925 #define next_reap_node(void) do { } while (0)
926 #endif
927
928 /*
929 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
930 * via the workqueue/eventd.
931 * Add the CPU number into the expiration time to minimize the possibility of
932 * the CPUs getting into lockstep and contending for the global cache chain
933 * lock.
934 */
935 static void __cpuinit start_cpu_timer(int cpu)
936 {
937 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
938
939 /*
940 * When this gets called from do_initcalls via cpucache_init(),
941 * init_workqueues() has already run, so keventd will be setup
942 * at that time.
943 */
944 if (keventd_up() && reap_work->work.func == NULL) {
945 init_reap_node(cpu);
946 INIT_DELAYED_WORK(reap_work, cache_reap);
947 schedule_delayed_work_on(cpu, reap_work,
948 __round_jiffies_relative(HZ, cpu));
949 }
950 }
951
952 static struct array_cache *alloc_arraycache(int node, int entries,
953 int batchcount)
954 {
955 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
956 struct array_cache *nc = NULL;
957
958 nc = kmalloc_node(memsize, GFP_KERNEL, node);
959 if (nc) {
960 nc->avail = 0;
961 nc->limit = entries;
962 nc->batchcount = batchcount;
963 nc->touched = 0;
964 spin_lock_init(&nc->lock);
965 }
966 return nc;
967 }
968
969 /*
970 * Transfer objects in one arraycache to another.
971 * Locking must be handled by the caller.
972 *
973 * Return the number of entries transferred.
974 */
975 static int transfer_objects(struct array_cache *to,
976 struct array_cache *from, unsigned int max)
977 {
978 /* Figure out how many entries to transfer */
979 int nr = min(min(from->avail, max), to->limit - to->avail);
980
981 if (!nr)
982 return 0;
983
984 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
985 sizeof(void *) *nr);
986
987 from->avail -= nr;
988 to->avail += nr;
989 to->touched = 1;
990 return nr;
991 }
992
993 #ifndef CONFIG_NUMA
994
995 #define drain_alien_cache(cachep, alien) do { } while (0)
996 #define reap_alien(cachep, l3) do { } while (0)
997
998 static inline struct array_cache **alloc_alien_cache(int node, int limit)
999 {
1000 return (struct array_cache **)BAD_ALIEN_MAGIC;
1001 }
1002
1003 static inline void free_alien_cache(struct array_cache **ac_ptr)
1004 {
1005 }
1006
1007 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1008 {
1009 return 0;
1010 }
1011
1012 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1013 gfp_t flags)
1014 {
1015 return NULL;
1016 }
1017
1018 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1019 gfp_t flags, int nodeid)
1020 {
1021 return NULL;
1022 }
1023
1024 #else /* CONFIG_NUMA */
1025
1026 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1027 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1028
1029 static struct array_cache **alloc_alien_cache(int node, int limit)
1030 {
1031 struct array_cache **ac_ptr;
1032 int memsize = sizeof(void *) * nr_node_ids;
1033 int i;
1034
1035 if (limit > 1)
1036 limit = 12;
1037 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1038 if (ac_ptr) {
1039 for_each_node(i) {
1040 if (i == node || !node_online(i)) {
1041 ac_ptr[i] = NULL;
1042 continue;
1043 }
1044 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1045 if (!ac_ptr[i]) {
1046 for (i--; i <= 0; i--)
1047 kfree(ac_ptr[i]);
1048 kfree(ac_ptr);
1049 return NULL;
1050 }
1051 }
1052 }
1053 return ac_ptr;
1054 }
1055
1056 static void free_alien_cache(struct array_cache **ac_ptr)
1057 {
1058 int i;
1059
1060 if (!ac_ptr)
1061 return;
1062 for_each_node(i)
1063 kfree(ac_ptr[i]);
1064 kfree(ac_ptr);
1065 }
1066
1067 static void __drain_alien_cache(struct kmem_cache *cachep,
1068 struct array_cache *ac, int node)
1069 {
1070 struct kmem_list3 *rl3 = cachep->nodelists[node];
1071
1072 if (ac->avail) {
1073 spin_lock(&rl3->list_lock);
1074 /*
1075 * Stuff objects into the remote nodes shared array first.
1076 * That way we could avoid the overhead of putting the objects
1077 * into the free lists and getting them back later.
1078 */
1079 if (rl3->shared)
1080 transfer_objects(rl3->shared, ac, ac->limit);
1081
1082 free_block(cachep, ac->entry, ac->avail, node);
1083 ac->avail = 0;
1084 spin_unlock(&rl3->list_lock);
1085 }
1086 }
1087
1088 /*
1089 * Called from cache_reap() to regularly drain alien caches round robin.
1090 */
1091 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1092 {
1093 int node = __get_cpu_var(reap_node);
1094
1095 if (l3->alien) {
1096 struct array_cache *ac = l3->alien[node];
1097
1098 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1099 __drain_alien_cache(cachep, ac, node);
1100 spin_unlock_irq(&ac->lock);
1101 }
1102 }
1103 }
1104
1105 static void drain_alien_cache(struct kmem_cache *cachep,
1106 struct array_cache **alien)
1107 {
1108 int i = 0;
1109 struct array_cache *ac;
1110 unsigned long flags;
1111
1112 for_each_online_node(i) {
1113 ac = alien[i];
1114 if (ac) {
1115 spin_lock_irqsave(&ac->lock, flags);
1116 __drain_alien_cache(cachep, ac, i);
1117 spin_unlock_irqrestore(&ac->lock, flags);
1118 }
1119 }
1120 }
1121
1122 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1123 {
1124 struct slab *slabp = virt_to_slab(objp);
1125 int nodeid = slabp->nodeid;
1126 struct kmem_list3 *l3;
1127 struct array_cache *alien = NULL;
1128 int node;
1129
1130 node = numa_node_id();
1131
1132 /*
1133 * Make sure we are not freeing a object from another node to the array
1134 * cache on this cpu.
1135 */
1136 if (likely(slabp->nodeid == node))
1137 return 0;
1138
1139 l3 = cachep->nodelists[node];
1140 STATS_INC_NODEFREES(cachep);
1141 if (l3->alien && l3->alien[nodeid]) {
1142 alien = l3->alien[nodeid];
1143 spin_lock(&alien->lock);
1144 if (unlikely(alien->avail == alien->limit)) {
1145 STATS_INC_ACOVERFLOW(cachep);
1146 __drain_alien_cache(cachep, alien, nodeid);
1147 }
1148 alien->entry[alien->avail++] = objp;
1149 spin_unlock(&alien->lock);
1150 } else {
1151 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1152 free_block(cachep, &objp, 1, nodeid);
1153 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1154 }
1155 return 1;
1156 }
1157 #endif
1158
1159 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1160 unsigned long action, void *hcpu)
1161 {
1162 long cpu = (long)hcpu;
1163 struct kmem_cache *cachep;
1164 struct kmem_list3 *l3 = NULL;
1165 int node = cpu_to_node(cpu);
1166 const int memsize = sizeof(struct kmem_list3);
1167
1168 switch (action) {
1169 case CPU_LOCK_ACQUIRE:
1170 mutex_lock(&cache_chain_mutex);
1171 break;
1172 case CPU_UP_PREPARE:
1173 case CPU_UP_PREPARE_FROZEN:
1174 /*
1175 * We need to do this right in the beginning since
1176 * alloc_arraycache's are going to use this list.
1177 * kmalloc_node allows us to add the slab to the right
1178 * kmem_list3 and not this cpu's kmem_list3
1179 */
1180
1181 list_for_each_entry(cachep, &cache_chain, next) {
1182 /*
1183 * Set up the size64 kmemlist for cpu before we can
1184 * begin anything. Make sure some other cpu on this
1185 * node has not already allocated this
1186 */
1187 if (!cachep->nodelists[node]) {
1188 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1189 if (!l3)
1190 goto bad;
1191 kmem_list3_init(l3);
1192 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1193 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1194
1195 /*
1196 * The l3s don't come and go as CPUs come and
1197 * go. cache_chain_mutex is sufficient
1198 * protection here.
1199 */
1200 cachep->nodelists[node] = l3;
1201 }
1202
1203 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1204 cachep->nodelists[node]->free_limit =
1205 (1 + nr_cpus_node(node)) *
1206 cachep->batchcount + cachep->num;
1207 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1208 }
1209
1210 /*
1211 * Now we can go ahead with allocating the shared arrays and
1212 * array caches
1213 */
1214 list_for_each_entry(cachep, &cache_chain, next) {
1215 struct array_cache *nc;
1216 struct array_cache *shared = NULL;
1217 struct array_cache **alien = NULL;
1218
1219 nc = alloc_arraycache(node, cachep->limit,
1220 cachep->batchcount);
1221 if (!nc)
1222 goto bad;
1223 if (cachep->shared) {
1224 shared = alloc_arraycache(node,
1225 cachep->shared * cachep->batchcount,
1226 0xbaadf00d);
1227 if (!shared)
1228 goto bad;
1229 }
1230 if (use_alien_caches) {
1231 alien = alloc_alien_cache(node, cachep->limit);
1232 if (!alien)
1233 goto bad;
1234 }
1235 cachep->array[cpu] = nc;
1236 l3 = cachep->nodelists[node];
1237 BUG_ON(!l3);
1238
1239 spin_lock_irq(&l3->list_lock);
1240 if (!l3->shared) {
1241 /*
1242 * We are serialised from CPU_DEAD or
1243 * CPU_UP_CANCELLED by the cpucontrol lock
1244 */
1245 l3->shared = shared;
1246 shared = NULL;
1247 }
1248 #ifdef CONFIG_NUMA
1249 if (!l3->alien) {
1250 l3->alien = alien;
1251 alien = NULL;
1252 }
1253 #endif
1254 spin_unlock_irq(&l3->list_lock);
1255 kfree(shared);
1256 free_alien_cache(alien);
1257 }
1258 break;
1259 case CPU_ONLINE:
1260 case CPU_ONLINE_FROZEN:
1261 start_cpu_timer(cpu);
1262 break;
1263 #ifdef CONFIG_HOTPLUG_CPU
1264 case CPU_DOWN_PREPARE:
1265 case CPU_DOWN_PREPARE_FROZEN:
1266 /*
1267 * Shutdown cache reaper. Note that the cache_chain_mutex is
1268 * held so that if cache_reap() is invoked it cannot do
1269 * anything expensive but will only modify reap_work
1270 * and reschedule the timer.
1271 */
1272 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1273 /* Now the cache_reaper is guaranteed to be not running. */
1274 per_cpu(reap_work, cpu).work.func = NULL;
1275 break;
1276 case CPU_DOWN_FAILED:
1277 case CPU_DOWN_FAILED_FROZEN:
1278 start_cpu_timer(cpu);
1279 break;
1280 case CPU_DEAD:
1281 case CPU_DEAD_FROZEN:
1282 /*
1283 * Even if all the cpus of a node are down, we don't free the
1284 * kmem_list3 of any cache. This to avoid a race between
1285 * cpu_down, and a kmalloc allocation from another cpu for
1286 * memory from the node of the cpu going down. The list3
1287 * structure is usually allocated from kmem_cache_create() and
1288 * gets destroyed at kmem_cache_destroy().
1289 */
1290 /* fall thru */
1291 #endif
1292 case CPU_UP_CANCELED:
1293 case CPU_UP_CANCELED_FROZEN:
1294 list_for_each_entry(cachep, &cache_chain, next) {
1295 struct array_cache *nc;
1296 struct array_cache *shared;
1297 struct array_cache **alien;
1298 cpumask_t mask;
1299
1300 mask = node_to_cpumask(node);
1301 /* cpu is dead; no one can alloc from it. */
1302 nc = cachep->array[cpu];
1303 cachep->array[cpu] = NULL;
1304 l3 = cachep->nodelists[node];
1305
1306 if (!l3)
1307 goto free_array_cache;
1308
1309 spin_lock_irq(&l3->list_lock);
1310
1311 /* Free limit for this kmem_list3 */
1312 l3->free_limit -= cachep->batchcount;
1313 if (nc)
1314 free_block(cachep, nc->entry, nc->avail, node);
1315
1316 if (!cpus_empty(mask)) {
1317 spin_unlock_irq(&l3->list_lock);
1318 goto free_array_cache;
1319 }
1320
1321 shared = l3->shared;
1322 if (shared) {
1323 free_block(cachep, shared->entry,
1324 shared->avail, node);
1325 l3->shared = NULL;
1326 }
1327
1328 alien = l3->alien;
1329 l3->alien = NULL;
1330
1331 spin_unlock_irq(&l3->list_lock);
1332
1333 kfree(shared);
1334 if (alien) {
1335 drain_alien_cache(cachep, alien);
1336 free_alien_cache(alien);
1337 }
1338 free_array_cache:
1339 kfree(nc);
1340 }
1341 /*
1342 * In the previous loop, all the objects were freed to
1343 * the respective cache's slabs, now we can go ahead and
1344 * shrink each nodelist to its limit.
1345 */
1346 list_for_each_entry(cachep, &cache_chain, next) {
1347 l3 = cachep->nodelists[node];
1348 if (!l3)
1349 continue;
1350 drain_freelist(cachep, l3, l3->free_objects);
1351 }
1352 break;
1353 case CPU_LOCK_RELEASE:
1354 mutex_unlock(&cache_chain_mutex);
1355 break;
1356 }
1357 return NOTIFY_OK;
1358 bad:
1359 return NOTIFY_BAD;
1360 }
1361
1362 static struct notifier_block __cpuinitdata cpucache_notifier = {
1363 &cpuup_callback, NULL, 0
1364 };
1365
1366 /*
1367 * swap the static kmem_list3 with kmalloced memory
1368 */
1369 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1370 int nodeid)
1371 {
1372 struct kmem_list3 *ptr;
1373
1374 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1375 BUG_ON(!ptr);
1376
1377 local_irq_disable();
1378 memcpy(ptr, list, sizeof(struct kmem_list3));
1379 /*
1380 * Do not assume that spinlocks can be initialized via memcpy:
1381 */
1382 spin_lock_init(&ptr->list_lock);
1383
1384 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1385 cachep->nodelists[nodeid] = ptr;
1386 local_irq_enable();
1387 }
1388
1389 /*
1390 * Initialisation. Called after the page allocator have been initialised and
1391 * before smp_init().
1392 */
1393 void __init kmem_cache_init(void)
1394 {
1395 size_t left_over;
1396 struct cache_sizes *sizes;
1397 struct cache_names *names;
1398 int i;
1399 int order;
1400 int node;
1401
1402 if (num_possible_nodes() == 1)
1403 use_alien_caches = 0;
1404
1405 for (i = 0; i < NUM_INIT_LISTS; i++) {
1406 kmem_list3_init(&initkmem_list3[i]);
1407 if (i < MAX_NUMNODES)
1408 cache_cache.nodelists[i] = NULL;
1409 }
1410
1411 /*
1412 * Fragmentation resistance on low memory - only use bigger
1413 * page orders on machines with more than 32MB of memory.
1414 */
1415 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1416 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1417
1418 /* Bootstrap is tricky, because several objects are allocated
1419 * from caches that do not exist yet:
1420 * 1) initialize the cache_cache cache: it contains the struct
1421 * kmem_cache structures of all caches, except cache_cache itself:
1422 * cache_cache is statically allocated.
1423 * Initially an __init data area is used for the head array and the
1424 * kmem_list3 structures, it's replaced with a kmalloc allocated
1425 * array at the end of the bootstrap.
1426 * 2) Create the first kmalloc cache.
1427 * The struct kmem_cache for the new cache is allocated normally.
1428 * An __init data area is used for the head array.
1429 * 3) Create the remaining kmalloc caches, with minimally sized
1430 * head arrays.
1431 * 4) Replace the __init data head arrays for cache_cache and the first
1432 * kmalloc cache with kmalloc allocated arrays.
1433 * 5) Replace the __init data for kmem_list3 for cache_cache and
1434 * the other cache's with kmalloc allocated memory.
1435 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1436 */
1437
1438 node = numa_node_id();
1439
1440 /* 1) create the cache_cache */
1441 INIT_LIST_HEAD(&cache_chain);
1442 list_add(&cache_cache.next, &cache_chain);
1443 cache_cache.colour_off = cache_line_size();
1444 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1445 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1446
1447 /*
1448 * struct kmem_cache size depends on nr_node_ids, which
1449 * can be less than MAX_NUMNODES.
1450 */
1451 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1452 nr_node_ids * sizeof(struct kmem_list3 *);
1453 #if DEBUG
1454 cache_cache.obj_size = cache_cache.buffer_size;
1455 #endif
1456 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1457 cache_line_size());
1458 cache_cache.reciprocal_buffer_size =
1459 reciprocal_value(cache_cache.buffer_size);
1460
1461 for (order = 0; order < MAX_ORDER; order++) {
1462 cache_estimate(order, cache_cache.buffer_size,
1463 cache_line_size(), 0, &left_over, &cache_cache.num);
1464 if (cache_cache.num)
1465 break;
1466 }
1467 BUG_ON(!cache_cache.num);
1468 cache_cache.gfporder = order;
1469 cache_cache.colour = left_over / cache_cache.colour_off;
1470 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1471 sizeof(struct slab), cache_line_size());
1472
1473 /* 2+3) create the kmalloc caches */
1474 sizes = malloc_sizes;
1475 names = cache_names;
1476
1477 /*
1478 * Initialize the caches that provide memory for the array cache and the
1479 * kmem_list3 structures first. Without this, further allocations will
1480 * bug.
1481 */
1482
1483 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1484 sizes[INDEX_AC].cs_size,
1485 ARCH_KMALLOC_MINALIGN,
1486 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1487 NULL);
1488
1489 if (INDEX_AC != INDEX_L3) {
1490 sizes[INDEX_L3].cs_cachep =
1491 kmem_cache_create(names[INDEX_L3].name,
1492 sizes[INDEX_L3].cs_size,
1493 ARCH_KMALLOC_MINALIGN,
1494 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1495 NULL);
1496 }
1497
1498 slab_early_init = 0;
1499
1500 while (sizes->cs_size != ULONG_MAX) {
1501 /*
1502 * For performance, all the general caches are L1 aligned.
1503 * This should be particularly beneficial on SMP boxes, as it
1504 * eliminates "false sharing".
1505 * Note for systems short on memory removing the alignment will
1506 * allow tighter packing of the smaller caches.
1507 */
1508 if (!sizes->cs_cachep) {
1509 sizes->cs_cachep = kmem_cache_create(names->name,
1510 sizes->cs_size,
1511 ARCH_KMALLOC_MINALIGN,
1512 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1513 NULL);
1514 }
1515 #ifdef CONFIG_ZONE_DMA
1516 sizes->cs_dmacachep = kmem_cache_create(
1517 names->name_dma,
1518 sizes->cs_size,
1519 ARCH_KMALLOC_MINALIGN,
1520 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1521 SLAB_PANIC,
1522 NULL);
1523 #endif
1524 sizes++;
1525 names++;
1526 }
1527 /* 4) Replace the bootstrap head arrays */
1528 {
1529 struct array_cache *ptr;
1530
1531 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1532
1533 local_irq_disable();
1534 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1535 memcpy(ptr, cpu_cache_get(&cache_cache),
1536 sizeof(struct arraycache_init));
1537 /*
1538 * Do not assume that spinlocks can be initialized via memcpy:
1539 */
1540 spin_lock_init(&ptr->lock);
1541
1542 cache_cache.array[smp_processor_id()] = ptr;
1543 local_irq_enable();
1544
1545 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1546
1547 local_irq_disable();
1548 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1549 != &initarray_generic.cache);
1550 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1551 sizeof(struct arraycache_init));
1552 /*
1553 * Do not assume that spinlocks can be initialized via memcpy:
1554 */
1555 spin_lock_init(&ptr->lock);
1556
1557 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1558 ptr;
1559 local_irq_enable();
1560 }
1561 /* 5) Replace the bootstrap kmem_list3's */
1562 {
1563 int nid;
1564
1565 /* Replace the static kmem_list3 structures for the boot cpu */
1566 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1567
1568 for_each_online_node(nid) {
1569 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1570 &initkmem_list3[SIZE_AC + nid], nid);
1571
1572 if (INDEX_AC != INDEX_L3) {
1573 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1574 &initkmem_list3[SIZE_L3 + nid], nid);
1575 }
1576 }
1577 }
1578
1579 /* 6) resize the head arrays to their final sizes */
1580 {
1581 struct kmem_cache *cachep;
1582 mutex_lock(&cache_chain_mutex);
1583 list_for_each_entry(cachep, &cache_chain, next)
1584 if (enable_cpucache(cachep))
1585 BUG();
1586 mutex_unlock(&cache_chain_mutex);
1587 }
1588
1589 /* Annotate slab for lockdep -- annotate the malloc caches */
1590 init_lock_keys();
1591
1592
1593 /* Done! */
1594 g_cpucache_up = FULL;
1595
1596 /*
1597 * Register a cpu startup notifier callback that initializes
1598 * cpu_cache_get for all new cpus
1599 */
1600 register_cpu_notifier(&cpucache_notifier);
1601
1602 /*
1603 * The reap timers are started later, with a module init call: That part
1604 * of the kernel is not yet operational.
1605 */
1606 }
1607
1608 static int __init cpucache_init(void)
1609 {
1610 int cpu;
1611
1612 /*
1613 * Register the timers that return unneeded pages to the page allocator
1614 */
1615 for_each_online_cpu(cpu)
1616 start_cpu_timer(cpu);
1617 return 0;
1618 }
1619 __initcall(cpucache_init);
1620
1621 /*
1622 * Interface to system's page allocator. No need to hold the cache-lock.
1623 *
1624 * If we requested dmaable memory, we will get it. Even if we
1625 * did not request dmaable memory, we might get it, but that
1626 * would be relatively rare and ignorable.
1627 */
1628 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1629 {
1630 struct page *page;
1631 int nr_pages;
1632 int i;
1633
1634 #ifndef CONFIG_MMU
1635 /*
1636 * Nommu uses slab's for process anonymous memory allocations, and thus
1637 * requires __GFP_COMP to properly refcount higher order allocations
1638 */
1639 flags |= __GFP_COMP;
1640 #endif
1641
1642 flags |= cachep->gfpflags;
1643
1644 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1645 if (!page)
1646 return NULL;
1647
1648 nr_pages = (1 << cachep->gfporder);
1649 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1650 add_zone_page_state(page_zone(page),
1651 NR_SLAB_RECLAIMABLE, nr_pages);
1652 else
1653 add_zone_page_state(page_zone(page),
1654 NR_SLAB_UNRECLAIMABLE, nr_pages);
1655 for (i = 0; i < nr_pages; i++)
1656 __SetPageSlab(page + i);
1657 return page_address(page);
1658 }
1659
1660 /*
1661 * Interface to system's page release.
1662 */
1663 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1664 {
1665 unsigned long i = (1 << cachep->gfporder);
1666 struct page *page = virt_to_page(addr);
1667 const unsigned long nr_freed = i;
1668
1669 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1670 sub_zone_page_state(page_zone(page),
1671 NR_SLAB_RECLAIMABLE, nr_freed);
1672 else
1673 sub_zone_page_state(page_zone(page),
1674 NR_SLAB_UNRECLAIMABLE, nr_freed);
1675 while (i--) {
1676 BUG_ON(!PageSlab(page));
1677 __ClearPageSlab(page);
1678 page++;
1679 }
1680 if (current->reclaim_state)
1681 current->reclaim_state->reclaimed_slab += nr_freed;
1682 free_pages((unsigned long)addr, cachep->gfporder);
1683 }
1684
1685 static void kmem_rcu_free(struct rcu_head *head)
1686 {
1687 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1688 struct kmem_cache *cachep = slab_rcu->cachep;
1689
1690 kmem_freepages(cachep, slab_rcu->addr);
1691 if (OFF_SLAB(cachep))
1692 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1693 }
1694
1695 #if DEBUG
1696
1697 #ifdef CONFIG_DEBUG_PAGEALLOC
1698 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1699 unsigned long caller)
1700 {
1701 int size = obj_size(cachep);
1702
1703 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1704
1705 if (size < 5 * sizeof(unsigned long))
1706 return;
1707
1708 *addr++ = 0x12345678;
1709 *addr++ = caller;
1710 *addr++ = smp_processor_id();
1711 size -= 3 * sizeof(unsigned long);
1712 {
1713 unsigned long *sptr = &caller;
1714 unsigned long svalue;
1715
1716 while (!kstack_end(sptr)) {
1717 svalue = *sptr++;
1718 if (kernel_text_address(svalue)) {
1719 *addr++ = svalue;
1720 size -= sizeof(unsigned long);
1721 if (size <= sizeof(unsigned long))
1722 break;
1723 }
1724 }
1725
1726 }
1727 *addr++ = 0x87654321;
1728 }
1729 #endif
1730
1731 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1732 {
1733 int size = obj_size(cachep);
1734 addr = &((char *)addr)[obj_offset(cachep)];
1735
1736 memset(addr, val, size);
1737 *(unsigned char *)(addr + size - 1) = POISON_END;
1738 }
1739
1740 static void dump_line(char *data, int offset, int limit)
1741 {
1742 int i;
1743 unsigned char error = 0;
1744 int bad_count = 0;
1745
1746 printk(KERN_ERR "%03x:", offset);
1747 for (i = 0; i < limit; i++) {
1748 if (data[offset + i] != POISON_FREE) {
1749 error = data[offset + i];
1750 bad_count++;
1751 }
1752 printk(" %02x", (unsigned char)data[offset + i]);
1753 }
1754 printk("\n");
1755
1756 if (bad_count == 1) {
1757 error ^= POISON_FREE;
1758 if (!(error & (error - 1))) {
1759 printk(KERN_ERR "Single bit error detected. Probably "
1760 "bad RAM.\n");
1761 #ifdef CONFIG_X86
1762 printk(KERN_ERR "Run memtest86+ or a similar memory "
1763 "test tool.\n");
1764 #else
1765 printk(KERN_ERR "Run a memory test tool.\n");
1766 #endif
1767 }
1768 }
1769 }
1770 #endif
1771
1772 #if DEBUG
1773
1774 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1775 {
1776 int i, size;
1777 char *realobj;
1778
1779 if (cachep->flags & SLAB_RED_ZONE) {
1780 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1781 *dbg_redzone1(cachep, objp),
1782 *dbg_redzone2(cachep, objp));
1783 }
1784
1785 if (cachep->flags & SLAB_STORE_USER) {
1786 printk(KERN_ERR "Last user: [<%p>]",
1787 *dbg_userword(cachep, objp));
1788 print_symbol("(%s)",
1789 (unsigned long)*dbg_userword(cachep, objp));
1790 printk("\n");
1791 }
1792 realobj = (char *)objp + obj_offset(cachep);
1793 size = obj_size(cachep);
1794 for (i = 0; i < size && lines; i += 16, lines--) {
1795 int limit;
1796 limit = 16;
1797 if (i + limit > size)
1798 limit = size - i;
1799 dump_line(realobj, i, limit);
1800 }
1801 }
1802
1803 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1804 {
1805 char *realobj;
1806 int size, i;
1807 int lines = 0;
1808
1809 realobj = (char *)objp + obj_offset(cachep);
1810 size = obj_size(cachep);
1811
1812 for (i = 0; i < size; i++) {
1813 char exp = POISON_FREE;
1814 if (i == size - 1)
1815 exp = POISON_END;
1816 if (realobj[i] != exp) {
1817 int limit;
1818 /* Mismatch ! */
1819 /* Print header */
1820 if (lines == 0) {
1821 printk(KERN_ERR
1822 "Slab corruption: %s start=%p, len=%d\n",
1823 cachep->name, realobj, size);
1824 print_objinfo(cachep, objp, 0);
1825 }
1826 /* Hexdump the affected line */
1827 i = (i / 16) * 16;
1828 limit = 16;
1829 if (i + limit > size)
1830 limit = size - i;
1831 dump_line(realobj, i, limit);
1832 i += 16;
1833 lines++;
1834 /* Limit to 5 lines */
1835 if (lines > 5)
1836 break;
1837 }
1838 }
1839 if (lines != 0) {
1840 /* Print some data about the neighboring objects, if they
1841 * exist:
1842 */
1843 struct slab *slabp = virt_to_slab(objp);
1844 unsigned int objnr;
1845
1846 objnr = obj_to_index(cachep, slabp, objp);
1847 if (objnr) {
1848 objp = index_to_obj(cachep, slabp, objnr - 1);
1849 realobj = (char *)objp + obj_offset(cachep);
1850 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1851 realobj, size);
1852 print_objinfo(cachep, objp, 2);
1853 }
1854 if (objnr + 1 < cachep->num) {
1855 objp = index_to_obj(cachep, slabp, objnr + 1);
1856 realobj = (char *)objp + obj_offset(cachep);
1857 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1858 realobj, size);
1859 print_objinfo(cachep, objp, 2);
1860 }
1861 }
1862 }
1863 #endif
1864
1865 #if DEBUG
1866 /**
1867 * slab_destroy_objs - destroy a slab and its objects
1868 * @cachep: cache pointer being destroyed
1869 * @slabp: slab pointer being destroyed
1870 *
1871 * Call the registered destructor for each object in a slab that is being
1872 * destroyed.
1873 */
1874 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1875 {
1876 int i;
1877 for (i = 0; i < cachep->num; i++) {
1878 void *objp = index_to_obj(cachep, slabp, i);
1879
1880 if (cachep->flags & SLAB_POISON) {
1881 #ifdef CONFIG_DEBUG_PAGEALLOC
1882 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1883 OFF_SLAB(cachep))
1884 kernel_map_pages(virt_to_page(objp),
1885 cachep->buffer_size / PAGE_SIZE, 1);
1886 else
1887 check_poison_obj(cachep, objp);
1888 #else
1889 check_poison_obj(cachep, objp);
1890 #endif
1891 }
1892 if (cachep->flags & SLAB_RED_ZONE) {
1893 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1894 slab_error(cachep, "start of a freed object "
1895 "was overwritten");
1896 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1897 slab_error(cachep, "end of a freed object "
1898 "was overwritten");
1899 }
1900 }
1901 }
1902 #else
1903 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1904 {
1905 }
1906 #endif
1907
1908 /**
1909 * slab_destroy - destroy and release all objects in a slab
1910 * @cachep: cache pointer being destroyed
1911 * @slabp: slab pointer being destroyed
1912 *
1913 * Destroy all the objs in a slab, and release the mem back to the system.
1914 * Before calling the slab must have been unlinked from the cache. The
1915 * cache-lock is not held/needed.
1916 */
1917 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1918 {
1919 void *addr = slabp->s_mem - slabp->colouroff;
1920
1921 slab_destroy_objs(cachep, slabp);
1922 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1923 struct slab_rcu *slab_rcu;
1924
1925 slab_rcu = (struct slab_rcu *)slabp;
1926 slab_rcu->cachep = cachep;
1927 slab_rcu->addr = addr;
1928 call_rcu(&slab_rcu->head, kmem_rcu_free);
1929 } else {
1930 kmem_freepages(cachep, addr);
1931 if (OFF_SLAB(cachep))
1932 kmem_cache_free(cachep->slabp_cache, slabp);
1933 }
1934 }
1935
1936 /*
1937 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1938 * size of kmem_list3.
1939 */
1940 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1941 {
1942 int node;
1943
1944 for_each_online_node(node) {
1945 cachep->nodelists[node] = &initkmem_list3[index + node];
1946 cachep->nodelists[node]->next_reap = jiffies +
1947 REAPTIMEOUT_LIST3 +
1948 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1949 }
1950 }
1951
1952 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1953 {
1954 int i;
1955 struct kmem_list3 *l3;
1956
1957 for_each_online_cpu(i)
1958 kfree(cachep->array[i]);
1959
1960 /* NUMA: free the list3 structures */
1961 for_each_online_node(i) {
1962 l3 = cachep->nodelists[i];
1963 if (l3) {
1964 kfree(l3->shared);
1965 free_alien_cache(l3->alien);
1966 kfree(l3);
1967 }
1968 }
1969 kmem_cache_free(&cache_cache, cachep);
1970 }
1971
1972
1973 /**
1974 * calculate_slab_order - calculate size (page order) of slabs
1975 * @cachep: pointer to the cache that is being created
1976 * @size: size of objects to be created in this cache.
1977 * @align: required alignment for the objects.
1978 * @flags: slab allocation flags
1979 *
1980 * Also calculates the number of objects per slab.
1981 *
1982 * This could be made much more intelligent. For now, try to avoid using
1983 * high order pages for slabs. When the gfp() functions are more friendly
1984 * towards high-order requests, this should be changed.
1985 */
1986 static size_t calculate_slab_order(struct kmem_cache *cachep,
1987 size_t size, size_t align, unsigned long flags)
1988 {
1989 unsigned long offslab_limit;
1990 size_t left_over = 0;
1991 int gfporder;
1992
1993 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1994 unsigned int num;
1995 size_t remainder;
1996
1997 cache_estimate(gfporder, size, align, flags, &remainder, &num);
1998 if (!num)
1999 continue;
2000
2001 if (flags & CFLGS_OFF_SLAB) {
2002 /*
2003 * Max number of objs-per-slab for caches which
2004 * use off-slab slabs. Needed to avoid a possible
2005 * looping condition in cache_grow().
2006 */
2007 offslab_limit = size - sizeof(struct slab);
2008 offslab_limit /= sizeof(kmem_bufctl_t);
2009
2010 if (num > offslab_limit)
2011 break;
2012 }
2013
2014 /* Found something acceptable - save it away */
2015 cachep->num = num;
2016 cachep->gfporder = gfporder;
2017 left_over = remainder;
2018
2019 /*
2020 * A VFS-reclaimable slab tends to have most allocations
2021 * as GFP_NOFS and we really don't want to have to be allocating
2022 * higher-order pages when we are unable to shrink dcache.
2023 */
2024 if (flags & SLAB_RECLAIM_ACCOUNT)
2025 break;
2026
2027 /*
2028 * Large number of objects is good, but very large slabs are
2029 * currently bad for the gfp()s.
2030 */
2031 if (gfporder >= slab_break_gfp_order)
2032 break;
2033
2034 /*
2035 * Acceptable internal fragmentation?
2036 */
2037 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2038 break;
2039 }
2040 return left_over;
2041 }
2042
2043 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
2044 {
2045 if (g_cpucache_up == FULL)
2046 return enable_cpucache(cachep);
2047
2048 if (g_cpucache_up == NONE) {
2049 /*
2050 * Note: the first kmem_cache_create must create the cache
2051 * that's used by kmalloc(24), otherwise the creation of
2052 * further caches will BUG().
2053 */
2054 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2055
2056 /*
2057 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2058 * the first cache, then we need to set up all its list3s,
2059 * otherwise the creation of further caches will BUG().
2060 */
2061 set_up_list3s(cachep, SIZE_AC);
2062 if (INDEX_AC == INDEX_L3)
2063 g_cpucache_up = PARTIAL_L3;
2064 else
2065 g_cpucache_up = PARTIAL_AC;
2066 } else {
2067 cachep->array[smp_processor_id()] =
2068 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2069
2070 if (g_cpucache_up == PARTIAL_AC) {
2071 set_up_list3s(cachep, SIZE_L3);
2072 g_cpucache_up = PARTIAL_L3;
2073 } else {
2074 int node;
2075 for_each_online_node(node) {
2076 cachep->nodelists[node] =
2077 kmalloc_node(sizeof(struct kmem_list3),
2078 GFP_KERNEL, node);
2079 BUG_ON(!cachep->nodelists[node]);
2080 kmem_list3_init(cachep->nodelists[node]);
2081 }
2082 }
2083 }
2084 cachep->nodelists[numa_node_id()]->next_reap =
2085 jiffies + REAPTIMEOUT_LIST3 +
2086 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2087
2088 cpu_cache_get(cachep)->avail = 0;
2089 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2090 cpu_cache_get(cachep)->batchcount = 1;
2091 cpu_cache_get(cachep)->touched = 0;
2092 cachep->batchcount = 1;
2093 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2094 return 0;
2095 }
2096
2097 /**
2098 * kmem_cache_create - Create a cache.
2099 * @name: A string which is used in /proc/slabinfo to identify this cache.
2100 * @size: The size of objects to be created in this cache.
2101 * @align: The required alignment for the objects.
2102 * @flags: SLAB flags
2103 * @ctor: A constructor for the objects.
2104 *
2105 * Returns a ptr to the cache on success, NULL on failure.
2106 * Cannot be called within a int, but can be interrupted.
2107 * The @ctor is run when new pages are allocated by the cache.
2108 *
2109 * @name must be valid until the cache is destroyed. This implies that
2110 * the module calling this has to destroy the cache before getting unloaded.
2111 *
2112 * The flags are
2113 *
2114 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2115 * to catch references to uninitialised memory.
2116 *
2117 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2118 * for buffer overruns.
2119 *
2120 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2121 * cacheline. This can be beneficial if you're counting cycles as closely
2122 * as davem.
2123 */
2124 struct kmem_cache *
2125 kmem_cache_create (const char *name, size_t size, size_t align,
2126 unsigned long flags,
2127 void (*ctor)(void*, struct kmem_cache *, unsigned long))
2128 {
2129 size_t left_over, slab_size, ralign;
2130 struct kmem_cache *cachep = NULL, *pc;
2131
2132 /*
2133 * Sanity checks... these are all serious usage bugs.
2134 */
2135 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2136 size > KMALLOC_MAX_SIZE) {
2137 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2138 name);
2139 BUG();
2140 }
2141
2142 /*
2143 * We use cache_chain_mutex to ensure a consistent view of
2144 * cpu_online_map as well. Please see cpuup_callback
2145 */
2146 mutex_lock(&cache_chain_mutex);
2147
2148 list_for_each_entry(pc, &cache_chain, next) {
2149 char tmp;
2150 int res;
2151
2152 /*
2153 * This happens when the module gets unloaded and doesn't
2154 * destroy its slab cache and no-one else reuses the vmalloc
2155 * area of the module. Print a warning.
2156 */
2157 res = probe_kernel_address(pc->name, tmp);
2158 if (res) {
2159 printk(KERN_ERR
2160 "SLAB: cache with size %d has lost its name\n",
2161 pc->buffer_size);
2162 continue;
2163 }
2164
2165 if (!strcmp(pc->name, name)) {
2166 printk(KERN_ERR
2167 "kmem_cache_create: duplicate cache %s\n", name);
2168 dump_stack();
2169 goto oops;
2170 }
2171 }
2172
2173 #if DEBUG
2174 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2175 #if FORCED_DEBUG
2176 /*
2177 * Enable redzoning and last user accounting, except for caches with
2178 * large objects, if the increased size would increase the object size
2179 * above the next power of two: caches with object sizes just above a
2180 * power of two have a significant amount of internal fragmentation.
2181 */
2182 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2183 2 * sizeof(unsigned long long)))
2184 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2185 if (!(flags & SLAB_DESTROY_BY_RCU))
2186 flags |= SLAB_POISON;
2187 #endif
2188 if (flags & SLAB_DESTROY_BY_RCU)
2189 BUG_ON(flags & SLAB_POISON);
2190 #endif
2191 /*
2192 * Always checks flags, a caller might be expecting debug support which
2193 * isn't available.
2194 */
2195 BUG_ON(flags & ~CREATE_MASK);
2196
2197 /*
2198 * Check that size is in terms of words. This is needed to avoid
2199 * unaligned accesses for some archs when redzoning is used, and makes
2200 * sure any on-slab bufctl's are also correctly aligned.
2201 */
2202 if (size & (BYTES_PER_WORD - 1)) {
2203 size += (BYTES_PER_WORD - 1);
2204 size &= ~(BYTES_PER_WORD - 1);
2205 }
2206
2207 /* calculate the final buffer alignment: */
2208
2209 /* 1) arch recommendation: can be overridden for debug */
2210 if (flags & SLAB_HWCACHE_ALIGN) {
2211 /*
2212 * Default alignment: as specified by the arch code. Except if
2213 * an object is really small, then squeeze multiple objects into
2214 * one cacheline.
2215 */
2216 ralign = cache_line_size();
2217 while (size <= ralign / 2)
2218 ralign /= 2;
2219 } else {
2220 ralign = BYTES_PER_WORD;
2221 }
2222
2223 /*
2224 * Redzoning and user store require word alignment or possibly larger.
2225 * Note this will be overridden by architecture or caller mandated
2226 * alignment if either is greater than BYTES_PER_WORD.
2227 */
2228 if (flags & SLAB_STORE_USER)
2229 ralign = BYTES_PER_WORD;
2230
2231 if (flags & SLAB_RED_ZONE) {
2232 ralign = REDZONE_ALIGN;
2233 /* If redzoning, ensure that the second redzone is suitably
2234 * aligned, by adjusting the object size accordingly. */
2235 size += REDZONE_ALIGN - 1;
2236 size &= ~(REDZONE_ALIGN - 1);
2237 }
2238
2239 /* 2) arch mandated alignment */
2240 if (ralign < ARCH_SLAB_MINALIGN) {
2241 ralign = ARCH_SLAB_MINALIGN;
2242 }
2243 /* 3) caller mandated alignment */
2244 if (ralign < align) {
2245 ralign = align;
2246 }
2247 /* disable debug if necessary */
2248 if (ralign > __alignof__(unsigned long long))
2249 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2250 /*
2251 * 4) Store it.
2252 */
2253 align = ralign;
2254
2255 /* Get cache's description obj. */
2256 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2257 if (!cachep)
2258 goto oops;
2259
2260 #if DEBUG
2261 cachep->obj_size = size;
2262
2263 /*
2264 * Both debugging options require word-alignment which is calculated
2265 * into align above.
2266 */
2267 if (flags & SLAB_RED_ZONE) {
2268 /* add space for red zone words */
2269 cachep->obj_offset += sizeof(unsigned long long);
2270 size += 2 * sizeof(unsigned long long);
2271 }
2272 if (flags & SLAB_STORE_USER) {
2273 /* user store requires one word storage behind the end of
2274 * the real object. But if the second red zone needs to be
2275 * aligned to 64 bits, we must allow that much space.
2276 */
2277 if (flags & SLAB_RED_ZONE)
2278 size += REDZONE_ALIGN;
2279 else
2280 size += BYTES_PER_WORD;
2281 }
2282 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2283 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2284 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2285 cachep->obj_offset += PAGE_SIZE - size;
2286 size = PAGE_SIZE;
2287 }
2288 #endif
2289 #endif
2290
2291 /*
2292 * Determine if the slab management is 'on' or 'off' slab.
2293 * (bootstrapping cannot cope with offslab caches so don't do
2294 * it too early on.)
2295 */
2296 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2297 /*
2298 * Size is large, assume best to place the slab management obj
2299 * off-slab (should allow better packing of objs).
2300 */
2301 flags |= CFLGS_OFF_SLAB;
2302
2303 size = ALIGN(size, align);
2304
2305 left_over = calculate_slab_order(cachep, size, align, flags);
2306
2307 if (!cachep->num) {
2308 printk(KERN_ERR
2309 "kmem_cache_create: couldn't create cache %s.\n", name);
2310 kmem_cache_free(&cache_cache, cachep);
2311 cachep = NULL;
2312 goto oops;
2313 }
2314 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2315 + sizeof(struct slab), align);
2316
2317 /*
2318 * If the slab has been placed off-slab, and we have enough space then
2319 * move it on-slab. This is at the expense of any extra colouring.
2320 */
2321 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2322 flags &= ~CFLGS_OFF_SLAB;
2323 left_over -= slab_size;
2324 }
2325
2326 if (flags & CFLGS_OFF_SLAB) {
2327 /* really off slab. No need for manual alignment */
2328 slab_size =
2329 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2330 }
2331
2332 cachep->colour_off = cache_line_size();
2333 /* Offset must be a multiple of the alignment. */
2334 if (cachep->colour_off < align)
2335 cachep->colour_off = align;
2336 cachep->colour = left_over / cachep->colour_off;
2337 cachep->slab_size = slab_size;
2338 cachep->flags = flags;
2339 cachep->gfpflags = 0;
2340 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2341 cachep->gfpflags |= GFP_DMA;
2342 cachep->buffer_size = size;
2343 cachep->reciprocal_buffer_size = reciprocal_value(size);
2344
2345 if (flags & CFLGS_OFF_SLAB) {
2346 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2347 /*
2348 * This is a possibility for one of the malloc_sizes caches.
2349 * But since we go off slab only for object size greater than
2350 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2351 * this should not happen at all.
2352 * But leave a BUG_ON for some lucky dude.
2353 */
2354 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2355 }
2356 cachep->ctor = ctor;
2357 cachep->name = name;
2358
2359 if (setup_cpu_cache(cachep)) {
2360 __kmem_cache_destroy(cachep);
2361 cachep = NULL;
2362 goto oops;
2363 }
2364
2365 /* cache setup completed, link it into the list */
2366 list_add(&cachep->next, &cache_chain);
2367 oops:
2368 if (!cachep && (flags & SLAB_PANIC))
2369 panic("kmem_cache_create(): failed to create slab `%s'\n",
2370 name);
2371 mutex_unlock(&cache_chain_mutex);
2372 return cachep;
2373 }
2374 EXPORT_SYMBOL(kmem_cache_create);
2375
2376 #if DEBUG
2377 static void check_irq_off(void)
2378 {
2379 BUG_ON(!irqs_disabled());
2380 }
2381
2382 static void check_irq_on(void)
2383 {
2384 BUG_ON(irqs_disabled());
2385 }
2386
2387 static void check_spinlock_acquired(struct kmem_cache *cachep)
2388 {
2389 #ifdef CONFIG_SMP
2390 check_irq_off();
2391 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2392 #endif
2393 }
2394
2395 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2396 {
2397 #ifdef CONFIG_SMP
2398 check_irq_off();
2399 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2400 #endif
2401 }
2402
2403 #else
2404 #define check_irq_off() do { } while(0)
2405 #define check_irq_on() do { } while(0)
2406 #define check_spinlock_acquired(x) do { } while(0)
2407 #define check_spinlock_acquired_node(x, y) do { } while(0)
2408 #endif
2409
2410 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2411 struct array_cache *ac,
2412 int force, int node);
2413
2414 static void do_drain(void *arg)
2415 {
2416 struct kmem_cache *cachep = arg;
2417 struct array_cache *ac;
2418 int node = numa_node_id();
2419
2420 check_irq_off();
2421 ac = cpu_cache_get(cachep);
2422 spin_lock(&cachep->nodelists[node]->list_lock);
2423 free_block(cachep, ac->entry, ac->avail, node);
2424 spin_unlock(&cachep->nodelists[node]->list_lock);
2425 ac->avail = 0;
2426 }
2427
2428 static void drain_cpu_caches(struct kmem_cache *cachep)
2429 {
2430 struct kmem_list3 *l3;
2431 int node;
2432
2433 on_each_cpu(do_drain, cachep, 1, 1);
2434 check_irq_on();
2435 for_each_online_node(node) {
2436 l3 = cachep->nodelists[node];
2437 if (l3 && l3->alien)
2438 drain_alien_cache(cachep, l3->alien);
2439 }
2440
2441 for_each_online_node(node) {
2442 l3 = cachep->nodelists[node];
2443 if (l3)
2444 drain_array(cachep, l3, l3->shared, 1, node);
2445 }
2446 }
2447
2448 /*
2449 * Remove slabs from the list of free slabs.
2450 * Specify the number of slabs to drain in tofree.
2451 *
2452 * Returns the actual number of slabs released.
2453 */
2454 static int drain_freelist(struct kmem_cache *cache,
2455 struct kmem_list3 *l3, int tofree)
2456 {
2457 struct list_head *p;
2458 int nr_freed;
2459 struct slab *slabp;
2460
2461 nr_freed = 0;
2462 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2463
2464 spin_lock_irq(&l3->list_lock);
2465 p = l3->slabs_free.prev;
2466 if (p == &l3->slabs_free) {
2467 spin_unlock_irq(&l3->list_lock);
2468 goto out;
2469 }
2470
2471 slabp = list_entry(p, struct slab, list);
2472 #if DEBUG
2473 BUG_ON(slabp->inuse);
2474 #endif
2475 list_del(&slabp->list);
2476 /*
2477 * Safe to drop the lock. The slab is no longer linked
2478 * to the cache.
2479 */
2480 l3->free_objects -= cache->num;
2481 spin_unlock_irq(&l3->list_lock);
2482 slab_destroy(cache, slabp);
2483 nr_freed++;
2484 }
2485 out:
2486 return nr_freed;
2487 }
2488
2489 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2490 static int __cache_shrink(struct kmem_cache *cachep)
2491 {
2492 int ret = 0, i = 0;
2493 struct kmem_list3 *l3;
2494
2495 drain_cpu_caches(cachep);
2496
2497 check_irq_on();
2498 for_each_online_node(i) {
2499 l3 = cachep->nodelists[i];
2500 if (!l3)
2501 continue;
2502
2503 drain_freelist(cachep, l3, l3->free_objects);
2504
2505 ret += !list_empty(&l3->slabs_full) ||
2506 !list_empty(&l3->slabs_partial);
2507 }
2508 return (ret ? 1 : 0);
2509 }
2510
2511 /**
2512 * kmem_cache_shrink - Shrink a cache.
2513 * @cachep: The cache to shrink.
2514 *
2515 * Releases as many slabs as possible for a cache.
2516 * To help debugging, a zero exit status indicates all slabs were released.
2517 */
2518 int kmem_cache_shrink(struct kmem_cache *cachep)
2519 {
2520 int ret;
2521 BUG_ON(!cachep || in_interrupt());
2522
2523 mutex_lock(&cache_chain_mutex);
2524 ret = __cache_shrink(cachep);
2525 mutex_unlock(&cache_chain_mutex);
2526 return ret;
2527 }
2528 EXPORT_SYMBOL(kmem_cache_shrink);
2529
2530 /**
2531 * kmem_cache_destroy - delete a cache
2532 * @cachep: the cache to destroy
2533 *
2534 * Remove a &struct kmem_cache object from the slab cache.
2535 *
2536 * It is expected this function will be called by a module when it is
2537 * unloaded. This will remove the cache completely, and avoid a duplicate
2538 * cache being allocated each time a module is loaded and unloaded, if the
2539 * module doesn't have persistent in-kernel storage across loads and unloads.
2540 *
2541 * The cache must be empty before calling this function.
2542 *
2543 * The caller must guarantee that noone will allocate memory from the cache
2544 * during the kmem_cache_destroy().
2545 */
2546 void kmem_cache_destroy(struct kmem_cache *cachep)
2547 {
2548 BUG_ON(!cachep || in_interrupt());
2549
2550 /* Find the cache in the chain of caches. */
2551 mutex_lock(&cache_chain_mutex);
2552 /*
2553 * the chain is never empty, cache_cache is never destroyed
2554 */
2555 list_del(&cachep->next);
2556 if (__cache_shrink(cachep)) {
2557 slab_error(cachep, "Can't free all objects");
2558 list_add(&cachep->next, &cache_chain);
2559 mutex_unlock(&cache_chain_mutex);
2560 return;
2561 }
2562
2563 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2564 synchronize_rcu();
2565
2566 __kmem_cache_destroy(cachep);
2567 mutex_unlock(&cache_chain_mutex);
2568 }
2569 EXPORT_SYMBOL(kmem_cache_destroy);
2570
2571 /*
2572 * Get the memory for a slab management obj.
2573 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2574 * always come from malloc_sizes caches. The slab descriptor cannot
2575 * come from the same cache which is getting created because,
2576 * when we are searching for an appropriate cache for these
2577 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2578 * If we are creating a malloc_sizes cache here it would not be visible to
2579 * kmem_find_general_cachep till the initialization is complete.
2580 * Hence we cannot have slabp_cache same as the original cache.
2581 */
2582 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2583 int colour_off, gfp_t local_flags,
2584 int nodeid)
2585 {
2586 struct slab *slabp;
2587
2588 if (OFF_SLAB(cachep)) {
2589 /* Slab management obj is off-slab. */
2590 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2591 local_flags & ~GFP_THISNODE, nodeid);
2592 if (!slabp)
2593 return NULL;
2594 } else {
2595 slabp = objp + colour_off;
2596 colour_off += cachep->slab_size;
2597 }
2598 slabp->inuse = 0;
2599 slabp->colouroff = colour_off;
2600 slabp->s_mem = objp + colour_off;
2601 slabp->nodeid = nodeid;
2602 return slabp;
2603 }
2604
2605 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2606 {
2607 return (kmem_bufctl_t *) (slabp + 1);
2608 }
2609
2610 static void cache_init_objs(struct kmem_cache *cachep,
2611 struct slab *slabp)
2612 {
2613 int i;
2614
2615 for (i = 0; i < cachep->num; i++) {
2616 void *objp = index_to_obj(cachep, slabp, i);
2617 #if DEBUG
2618 /* need to poison the objs? */
2619 if (cachep->flags & SLAB_POISON)
2620 poison_obj(cachep, objp, POISON_FREE);
2621 if (cachep->flags & SLAB_STORE_USER)
2622 *dbg_userword(cachep, objp) = NULL;
2623
2624 if (cachep->flags & SLAB_RED_ZONE) {
2625 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2626 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2627 }
2628 /*
2629 * Constructors are not allowed to allocate memory from the same
2630 * cache which they are a constructor for. Otherwise, deadlock.
2631 * They must also be threaded.
2632 */
2633 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2634 cachep->ctor(objp + obj_offset(cachep), cachep,
2635 0);
2636
2637 if (cachep->flags & SLAB_RED_ZONE) {
2638 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2639 slab_error(cachep, "constructor overwrote the"
2640 " end of an object");
2641 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2642 slab_error(cachep, "constructor overwrote the"
2643 " start of an object");
2644 }
2645 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2646 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2647 kernel_map_pages(virt_to_page(objp),
2648 cachep->buffer_size / PAGE_SIZE, 0);
2649 #else
2650 if (cachep->ctor)
2651 cachep->ctor(objp, cachep, 0);
2652 #endif
2653 slab_bufctl(slabp)[i] = i + 1;
2654 }
2655 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2656 slabp->free = 0;
2657 }
2658
2659 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2660 {
2661 if (CONFIG_ZONE_DMA_FLAG) {
2662 if (flags & GFP_DMA)
2663 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2664 else
2665 BUG_ON(cachep->gfpflags & GFP_DMA);
2666 }
2667 }
2668
2669 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2670 int nodeid)
2671 {
2672 void *objp = index_to_obj(cachep, slabp, slabp->free);
2673 kmem_bufctl_t next;
2674
2675 slabp->inuse++;
2676 next = slab_bufctl(slabp)[slabp->free];
2677 #if DEBUG
2678 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2679 WARN_ON(slabp->nodeid != nodeid);
2680 #endif
2681 slabp->free = next;
2682
2683 return objp;
2684 }
2685
2686 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2687 void *objp, int nodeid)
2688 {
2689 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2690
2691 #if DEBUG
2692 /* Verify that the slab belongs to the intended node */
2693 WARN_ON(slabp->nodeid != nodeid);
2694
2695 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2696 printk(KERN_ERR "slab: double free detected in cache "
2697 "'%s', objp %p\n", cachep->name, objp);
2698 BUG();
2699 }
2700 #endif
2701 slab_bufctl(slabp)[objnr] = slabp->free;
2702 slabp->free = objnr;
2703 slabp->inuse--;
2704 }
2705
2706 /*
2707 * Map pages beginning at addr to the given cache and slab. This is required
2708 * for the slab allocator to be able to lookup the cache and slab of a
2709 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2710 */
2711 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2712 void *addr)
2713 {
2714 int nr_pages;
2715 struct page *page;
2716
2717 page = virt_to_page(addr);
2718
2719 nr_pages = 1;
2720 if (likely(!PageCompound(page)))
2721 nr_pages <<= cache->gfporder;
2722
2723 do {
2724 page_set_cache(page, cache);
2725 page_set_slab(page, slab);
2726 page++;
2727 } while (--nr_pages);
2728 }
2729
2730 /*
2731 * Grow (by 1) the number of slabs within a cache. This is called by
2732 * kmem_cache_alloc() when there are no active objs left in a cache.
2733 */
2734 static int cache_grow(struct kmem_cache *cachep,
2735 gfp_t flags, int nodeid, void *objp)
2736 {
2737 struct slab *slabp;
2738 size_t offset;
2739 gfp_t local_flags;
2740 struct kmem_list3 *l3;
2741
2742 /*
2743 * Be lazy and only check for valid flags here, keeping it out of the
2744 * critical path in kmem_cache_alloc().
2745 */
2746 BUG_ON(flags & ~(GFP_DMA | __GFP_ZERO | GFP_LEVEL_MASK));
2747
2748 local_flags = (flags & GFP_LEVEL_MASK);
2749 /* Take the l3 list lock to change the colour_next on this node */
2750 check_irq_off();
2751 l3 = cachep->nodelists[nodeid];
2752 spin_lock(&l3->list_lock);
2753
2754 /* Get colour for the slab, and cal the next value. */
2755 offset = l3->colour_next;
2756 l3->colour_next++;
2757 if (l3->colour_next >= cachep->colour)
2758 l3->colour_next = 0;
2759 spin_unlock(&l3->list_lock);
2760
2761 offset *= cachep->colour_off;
2762
2763 if (local_flags & __GFP_WAIT)
2764 local_irq_enable();
2765
2766 /*
2767 * The test for missing atomic flag is performed here, rather than
2768 * the more obvious place, simply to reduce the critical path length
2769 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2770 * will eventually be caught here (where it matters).
2771 */
2772 kmem_flagcheck(cachep, flags);
2773
2774 /*
2775 * Get mem for the objs. Attempt to allocate a physical page from
2776 * 'nodeid'.
2777 */
2778 if (!objp)
2779 objp = kmem_getpages(cachep, local_flags, nodeid);
2780 if (!objp)
2781 goto failed;
2782
2783 /* Get slab management. */
2784 slabp = alloc_slabmgmt(cachep, objp, offset,
2785 local_flags & ~GFP_THISNODE, nodeid);
2786 if (!slabp)
2787 goto opps1;
2788
2789 slabp->nodeid = nodeid;
2790 slab_map_pages(cachep, slabp, objp);
2791
2792 cache_init_objs(cachep, slabp);
2793
2794 if (local_flags & __GFP_WAIT)
2795 local_irq_disable();
2796 check_irq_off();
2797 spin_lock(&l3->list_lock);
2798
2799 /* Make slab active. */
2800 list_add_tail(&slabp->list, &(l3->slabs_free));
2801 STATS_INC_GROWN(cachep);
2802 l3->free_objects += cachep->num;
2803 spin_unlock(&l3->list_lock);
2804 return 1;
2805 opps1:
2806 kmem_freepages(cachep, objp);
2807 failed:
2808 if (local_flags & __GFP_WAIT)
2809 local_irq_disable();
2810 return 0;
2811 }
2812
2813 #if DEBUG
2814
2815 /*
2816 * Perform extra freeing checks:
2817 * - detect bad pointers.
2818 * - POISON/RED_ZONE checking
2819 */
2820 static void kfree_debugcheck(const void *objp)
2821 {
2822 if (!virt_addr_valid(objp)) {
2823 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2824 (unsigned long)objp);
2825 BUG();
2826 }
2827 }
2828
2829 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2830 {
2831 unsigned long long redzone1, redzone2;
2832
2833 redzone1 = *dbg_redzone1(cache, obj);
2834 redzone2 = *dbg_redzone2(cache, obj);
2835
2836 /*
2837 * Redzone is ok.
2838 */
2839 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2840 return;
2841
2842 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2843 slab_error(cache, "double free detected");
2844 else
2845 slab_error(cache, "memory outside object was overwritten");
2846
2847 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2848 obj, redzone1, redzone2);
2849 }
2850
2851 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2852 void *caller)
2853 {
2854 struct page *page;
2855 unsigned int objnr;
2856 struct slab *slabp;
2857
2858 objp -= obj_offset(cachep);
2859 kfree_debugcheck(objp);
2860 page = virt_to_head_page(objp);
2861
2862 slabp = page_get_slab(page);
2863
2864 if (cachep->flags & SLAB_RED_ZONE) {
2865 verify_redzone_free(cachep, objp);
2866 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2867 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2868 }
2869 if (cachep->flags & SLAB_STORE_USER)
2870 *dbg_userword(cachep, objp) = caller;
2871
2872 objnr = obj_to_index(cachep, slabp, objp);
2873
2874 BUG_ON(objnr >= cachep->num);
2875 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2876
2877 #ifdef CONFIG_DEBUG_SLAB_LEAK
2878 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2879 #endif
2880 if (cachep->flags & SLAB_POISON) {
2881 #ifdef CONFIG_DEBUG_PAGEALLOC
2882 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2883 store_stackinfo(cachep, objp, (unsigned long)caller);
2884 kernel_map_pages(virt_to_page(objp),
2885 cachep->buffer_size / PAGE_SIZE, 0);
2886 } else {
2887 poison_obj(cachep, objp, POISON_FREE);
2888 }
2889 #else
2890 poison_obj(cachep, objp, POISON_FREE);
2891 #endif
2892 }
2893 return objp;
2894 }
2895
2896 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2897 {
2898 kmem_bufctl_t i;
2899 int entries = 0;
2900
2901 /* Check slab's freelist to see if this obj is there. */
2902 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2903 entries++;
2904 if (entries > cachep->num || i >= cachep->num)
2905 goto bad;
2906 }
2907 if (entries != cachep->num - slabp->inuse) {
2908 bad:
2909 printk(KERN_ERR "slab: Internal list corruption detected in "
2910 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2911 cachep->name, cachep->num, slabp, slabp->inuse);
2912 for (i = 0;
2913 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2914 i++) {
2915 if (i % 16 == 0)
2916 printk("\n%03x:", i);
2917 printk(" %02x", ((unsigned char *)slabp)[i]);
2918 }
2919 printk("\n");
2920 BUG();
2921 }
2922 }
2923 #else
2924 #define kfree_debugcheck(x) do { } while(0)
2925 #define cache_free_debugcheck(x,objp,z) (objp)
2926 #define check_slabp(x,y) do { } while(0)
2927 #endif
2928
2929 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2930 {
2931 int batchcount;
2932 struct kmem_list3 *l3;
2933 struct array_cache *ac;
2934 int node;
2935
2936 node = numa_node_id();
2937
2938 check_irq_off();
2939 ac = cpu_cache_get(cachep);
2940 retry:
2941 batchcount = ac->batchcount;
2942 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2943 /*
2944 * If there was little recent activity on this cache, then
2945 * perform only a partial refill. Otherwise we could generate
2946 * refill bouncing.
2947 */
2948 batchcount = BATCHREFILL_LIMIT;
2949 }
2950 l3 = cachep->nodelists[node];
2951
2952 BUG_ON(ac->avail > 0 || !l3);
2953 spin_lock(&l3->list_lock);
2954
2955 /* See if we can refill from the shared array */
2956 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2957 goto alloc_done;
2958
2959 while (batchcount > 0) {
2960 struct list_head *entry;
2961 struct slab *slabp;
2962 /* Get slab alloc is to come from. */
2963 entry = l3->slabs_partial.next;
2964 if (entry == &l3->slabs_partial) {
2965 l3->free_touched = 1;
2966 entry = l3->slabs_free.next;
2967 if (entry == &l3->slabs_free)
2968 goto must_grow;
2969 }
2970
2971 slabp = list_entry(entry, struct slab, list);
2972 check_slabp(cachep, slabp);
2973 check_spinlock_acquired(cachep);
2974
2975 /*
2976 * The slab was either on partial or free list so
2977 * there must be at least one object available for
2978 * allocation.
2979 */
2980 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2981
2982 while (slabp->inuse < cachep->num && batchcount--) {
2983 STATS_INC_ALLOCED(cachep);
2984 STATS_INC_ACTIVE(cachep);
2985 STATS_SET_HIGH(cachep);
2986
2987 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2988 node);
2989 }
2990 check_slabp(cachep, slabp);
2991
2992 /* move slabp to correct slabp list: */
2993 list_del(&slabp->list);
2994 if (slabp->free == BUFCTL_END)
2995 list_add(&slabp->list, &l3->slabs_full);
2996 else
2997 list_add(&slabp->list, &l3->slabs_partial);
2998 }
2999
3000 must_grow:
3001 l3->free_objects -= ac->avail;
3002 alloc_done:
3003 spin_unlock(&l3->list_lock);
3004
3005 if (unlikely(!ac->avail)) {
3006 int x;
3007 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3008
3009 /* cache_grow can reenable interrupts, then ac could change. */
3010 ac = cpu_cache_get(cachep);
3011 if (!x && ac->avail == 0) /* no objects in sight? abort */
3012 return NULL;
3013
3014 if (!ac->avail) /* objects refilled by interrupt? */
3015 goto retry;
3016 }
3017 ac->touched = 1;
3018 return ac->entry[--ac->avail];
3019 }
3020
3021 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3022 gfp_t flags)
3023 {
3024 might_sleep_if(flags & __GFP_WAIT);
3025 #if DEBUG
3026 kmem_flagcheck(cachep, flags);
3027 #endif
3028 }
3029
3030 #if DEBUG
3031 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3032 gfp_t flags, void *objp, void *caller)
3033 {
3034 if (!objp)
3035 return objp;
3036 if (cachep->flags & SLAB_POISON) {
3037 #ifdef CONFIG_DEBUG_PAGEALLOC
3038 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3039 kernel_map_pages(virt_to_page(objp),
3040 cachep->buffer_size / PAGE_SIZE, 1);
3041 else
3042 check_poison_obj(cachep, objp);
3043 #else
3044 check_poison_obj(cachep, objp);
3045 #endif
3046 poison_obj(cachep, objp, POISON_INUSE);
3047 }
3048 if (cachep->flags & SLAB_STORE_USER)
3049 *dbg_userword(cachep, objp) = caller;
3050
3051 if (cachep->flags & SLAB_RED_ZONE) {
3052 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3053 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3054 slab_error(cachep, "double free, or memory outside"
3055 " object was overwritten");
3056 printk(KERN_ERR
3057 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3058 objp, *dbg_redzone1(cachep, objp),
3059 *dbg_redzone2(cachep, objp));
3060 }
3061 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3062 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3063 }
3064 #ifdef CONFIG_DEBUG_SLAB_LEAK
3065 {
3066 struct slab *slabp;
3067 unsigned objnr;
3068
3069 slabp = page_get_slab(virt_to_head_page(objp));
3070 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3071 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3072 }
3073 #endif
3074 objp += obj_offset(cachep);
3075 if (cachep->ctor && cachep->flags & SLAB_POISON)
3076 cachep->ctor(objp, cachep, 0);
3077 #if ARCH_SLAB_MINALIGN
3078 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3079 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3080 objp, ARCH_SLAB_MINALIGN);
3081 }
3082 #endif
3083 return objp;
3084 }
3085 #else
3086 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3087 #endif
3088
3089 #ifdef CONFIG_FAILSLAB
3090
3091 static struct failslab_attr {
3092
3093 struct fault_attr attr;
3094
3095 u32 ignore_gfp_wait;
3096 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3097 struct dentry *ignore_gfp_wait_file;
3098 #endif
3099
3100 } failslab = {
3101 .attr = FAULT_ATTR_INITIALIZER,
3102 .ignore_gfp_wait = 1,
3103 };
3104
3105 static int __init setup_failslab(char *str)
3106 {
3107 return setup_fault_attr(&failslab.attr, str);
3108 }
3109 __setup("failslab=", setup_failslab);
3110
3111 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3112 {
3113 if (cachep == &cache_cache)
3114 return 0;
3115 if (flags & __GFP_NOFAIL)
3116 return 0;
3117 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3118 return 0;
3119
3120 return should_fail(&failslab.attr, obj_size(cachep));
3121 }
3122
3123 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3124
3125 static int __init failslab_debugfs(void)
3126 {
3127 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3128 struct dentry *dir;
3129 int err;
3130
3131 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3132 if (err)
3133 return err;
3134 dir = failslab.attr.dentries.dir;
3135
3136 failslab.ignore_gfp_wait_file =
3137 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3138 &failslab.ignore_gfp_wait);
3139
3140 if (!failslab.ignore_gfp_wait_file) {
3141 err = -ENOMEM;
3142 debugfs_remove(failslab.ignore_gfp_wait_file);
3143 cleanup_fault_attr_dentries(&failslab.attr);
3144 }
3145
3146 return err;
3147 }
3148
3149 late_initcall(failslab_debugfs);
3150
3151 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3152
3153 #else /* CONFIG_FAILSLAB */
3154
3155 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3156 {
3157 return 0;
3158 }
3159
3160 #endif /* CONFIG_FAILSLAB */
3161
3162 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3163 {
3164 void *objp;
3165 struct array_cache *ac;
3166
3167 check_irq_off();
3168
3169 ac = cpu_cache_get(cachep);
3170 if (likely(ac->avail)) {
3171 STATS_INC_ALLOCHIT(cachep);
3172 ac->touched = 1;
3173 objp = ac->entry[--ac->avail];
3174 } else {
3175 STATS_INC_ALLOCMISS(cachep);
3176 objp = cache_alloc_refill(cachep, flags);
3177 }
3178 return objp;
3179 }
3180
3181 #ifdef CONFIG_NUMA
3182 /*
3183 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3184 *
3185 * If we are in_interrupt, then process context, including cpusets and
3186 * mempolicy, may not apply and should not be used for allocation policy.
3187 */
3188 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3189 {
3190 int nid_alloc, nid_here;
3191
3192 if (in_interrupt() || (flags & __GFP_THISNODE))
3193 return NULL;
3194 nid_alloc = nid_here = numa_node_id();
3195 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3196 nid_alloc = cpuset_mem_spread_node();
3197 else if (current->mempolicy)
3198 nid_alloc = slab_node(current->mempolicy);
3199 if (nid_alloc != nid_here)
3200 return ____cache_alloc_node(cachep, flags, nid_alloc);
3201 return NULL;
3202 }
3203
3204 /*
3205 * Fallback function if there was no memory available and no objects on a
3206 * certain node and fall back is permitted. First we scan all the
3207 * available nodelists for available objects. If that fails then we
3208 * perform an allocation without specifying a node. This allows the page
3209 * allocator to do its reclaim / fallback magic. We then insert the
3210 * slab into the proper nodelist and then allocate from it.
3211 */
3212 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3213 {
3214 struct zonelist *zonelist;
3215 gfp_t local_flags;
3216 struct zone **z;
3217 void *obj = NULL;
3218 int nid;
3219
3220 if (flags & __GFP_THISNODE)
3221 return NULL;
3222
3223 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3224 ->node_zonelists[gfp_zone(flags)];
3225 local_flags = (flags & GFP_LEVEL_MASK);
3226
3227 retry:
3228 /*
3229 * Look through allowed nodes for objects available
3230 * from existing per node queues.
3231 */
3232 for (z = zonelist->zones; *z && !obj; z++) {
3233 nid = zone_to_nid(*z);
3234
3235 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3236 cache->nodelists[nid] &&
3237 cache->nodelists[nid]->free_objects)
3238 obj = ____cache_alloc_node(cache,
3239 flags | GFP_THISNODE, nid);
3240 }
3241
3242 if (!obj) {
3243 /*
3244 * This allocation will be performed within the constraints
3245 * of the current cpuset / memory policy requirements.
3246 * We may trigger various forms of reclaim on the allowed
3247 * set and go into memory reserves if necessary.
3248 */
3249 if (local_flags & __GFP_WAIT)
3250 local_irq_enable();
3251 kmem_flagcheck(cache, flags);
3252 obj = kmem_getpages(cache, flags, -1);
3253 if (local_flags & __GFP_WAIT)
3254 local_irq_disable();
3255 if (obj) {
3256 /*
3257 * Insert into the appropriate per node queues
3258 */
3259 nid = page_to_nid(virt_to_page(obj));
3260 if (cache_grow(cache, flags, nid, obj)) {
3261 obj = ____cache_alloc_node(cache,
3262 flags | GFP_THISNODE, nid);
3263 if (!obj)
3264 /*
3265 * Another processor may allocate the
3266 * objects in the slab since we are
3267 * not holding any locks.
3268 */
3269 goto retry;
3270 } else {
3271 /* cache_grow already freed obj */
3272 obj = NULL;
3273 }
3274 }
3275 }
3276 return obj;
3277 }
3278
3279 /*
3280 * A interface to enable slab creation on nodeid
3281 */
3282 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3283 int nodeid)
3284 {
3285 struct list_head *entry;
3286 struct slab *slabp;
3287 struct kmem_list3 *l3;
3288 void *obj;
3289 int x;
3290
3291 l3 = cachep->nodelists[nodeid];
3292 BUG_ON(!l3);
3293
3294 retry:
3295 check_irq_off();
3296 spin_lock(&l3->list_lock);
3297 entry = l3->slabs_partial.next;
3298 if (entry == &l3->slabs_partial) {
3299 l3->free_touched = 1;
3300 entry = l3->slabs_free.next;
3301 if (entry == &l3->slabs_free)
3302 goto must_grow;
3303 }
3304
3305 slabp = list_entry(entry, struct slab, list);
3306 check_spinlock_acquired_node(cachep, nodeid);
3307 check_slabp(cachep, slabp);
3308
3309 STATS_INC_NODEALLOCS(cachep);
3310 STATS_INC_ACTIVE(cachep);
3311 STATS_SET_HIGH(cachep);
3312
3313 BUG_ON(slabp->inuse == cachep->num);
3314
3315 obj = slab_get_obj(cachep, slabp, nodeid);
3316 check_slabp(cachep, slabp);
3317 l3->free_objects--;
3318 /* move slabp to correct slabp list: */
3319 list_del(&slabp->list);
3320
3321 if (slabp->free == BUFCTL_END)
3322 list_add(&slabp->list, &l3->slabs_full);
3323 else
3324 list_add(&slabp->list, &l3->slabs_partial);
3325
3326 spin_unlock(&l3->list_lock);
3327 goto done;
3328
3329 must_grow:
3330 spin_unlock(&l3->list_lock);
3331 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3332 if (x)
3333 goto retry;
3334
3335 return fallback_alloc(cachep, flags);
3336
3337 done:
3338 return obj;
3339 }
3340
3341 /**
3342 * kmem_cache_alloc_node - Allocate an object on the specified node
3343 * @cachep: The cache to allocate from.
3344 * @flags: See kmalloc().
3345 * @nodeid: node number of the target node.
3346 * @caller: return address of caller, used for debug information
3347 *
3348 * Identical to kmem_cache_alloc but it will allocate memory on the given
3349 * node, which can improve the performance for cpu bound structures.
3350 *
3351 * Fallback to other node is possible if __GFP_THISNODE is not set.
3352 */
3353 static __always_inline void *
3354 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3355 void *caller)
3356 {
3357 unsigned long save_flags;
3358 void *ptr;
3359
3360 if (should_failslab(cachep, flags))
3361 return NULL;
3362
3363 cache_alloc_debugcheck_before(cachep, flags);
3364 local_irq_save(save_flags);
3365
3366 if (unlikely(nodeid == -1))
3367 nodeid = numa_node_id();
3368
3369 if (unlikely(!cachep->nodelists[nodeid])) {
3370 /* Node not bootstrapped yet */
3371 ptr = fallback_alloc(cachep, flags);
3372 goto out;
3373 }
3374
3375 if (nodeid == numa_node_id()) {
3376 /*
3377 * Use the locally cached objects if possible.
3378 * However ____cache_alloc does not allow fallback
3379 * to other nodes. It may fail while we still have
3380 * objects on other nodes available.
3381 */
3382 ptr = ____cache_alloc(cachep, flags);
3383 if (ptr)
3384 goto out;
3385 }
3386 /* ___cache_alloc_node can fall back to other nodes */
3387 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3388 out:
3389 local_irq_restore(save_flags);
3390 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3391
3392 if (unlikely((flags & __GFP_ZERO) && ptr))
3393 memset(ptr, 0, obj_size(cachep));
3394
3395 return ptr;
3396 }
3397
3398 static __always_inline void *
3399 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3400 {
3401 void *objp;
3402
3403 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3404 objp = alternate_node_alloc(cache, flags);
3405 if (objp)
3406 goto out;
3407 }
3408 objp = ____cache_alloc(cache, flags);
3409
3410 /*
3411 * We may just have run out of memory on the local node.
3412 * ____cache_alloc_node() knows how to locate memory on other nodes
3413 */
3414 if (!objp)
3415 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3416
3417 out:
3418 return objp;
3419 }
3420 #else
3421
3422 static __always_inline void *
3423 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3424 {
3425 return ____cache_alloc(cachep, flags);
3426 }
3427
3428 #endif /* CONFIG_NUMA */
3429
3430 static __always_inline void *
3431 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3432 {
3433 unsigned long save_flags;
3434 void *objp;
3435
3436 if (should_failslab(cachep, flags))
3437 return NULL;
3438
3439 cache_alloc_debugcheck_before(cachep, flags);
3440 local_irq_save(save_flags);
3441 objp = __do_cache_alloc(cachep, flags);
3442 local_irq_restore(save_flags);
3443 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3444 prefetchw(objp);
3445
3446 if (unlikely((flags & __GFP_ZERO) && objp))
3447 memset(objp, 0, obj_size(cachep));
3448
3449 return objp;
3450 }
3451
3452 /*
3453 * Caller needs to acquire correct kmem_list's list_lock
3454 */
3455 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3456 int node)
3457 {
3458 int i;
3459 struct kmem_list3 *l3;
3460
3461 for (i = 0; i < nr_objects; i++) {
3462 void *objp = objpp[i];
3463 struct slab *slabp;
3464
3465 slabp = virt_to_slab(objp);
3466 l3 = cachep->nodelists[node];
3467 list_del(&slabp->list);
3468 check_spinlock_acquired_node(cachep, node);
3469 check_slabp(cachep, slabp);
3470 slab_put_obj(cachep, slabp, objp, node);
3471 STATS_DEC_ACTIVE(cachep);
3472 l3->free_objects++;
3473 check_slabp(cachep, slabp);
3474
3475 /* fixup slab chains */
3476 if (slabp->inuse == 0) {
3477 if (l3->free_objects > l3->free_limit) {
3478 l3->free_objects -= cachep->num;
3479 /* No need to drop any previously held
3480 * lock here, even if we have a off-slab slab
3481 * descriptor it is guaranteed to come from
3482 * a different cache, refer to comments before
3483 * alloc_slabmgmt.
3484 */
3485 slab_destroy(cachep, slabp);
3486 } else {
3487 list_add(&slabp->list, &l3->slabs_free);
3488 }
3489 } else {
3490 /* Unconditionally move a slab to the end of the
3491 * partial list on free - maximum time for the
3492 * other objects to be freed, too.
3493 */
3494 list_add_tail(&slabp->list, &l3->slabs_partial);
3495 }
3496 }
3497 }
3498
3499 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3500 {
3501 int batchcount;
3502 struct kmem_list3 *l3;
3503 int node = numa_node_id();
3504
3505 batchcount = ac->batchcount;
3506 #if DEBUG
3507 BUG_ON(!batchcount || batchcount > ac->avail);
3508 #endif
3509 check_irq_off();
3510 l3 = cachep->nodelists[node];
3511 spin_lock(&l3->list_lock);
3512 if (l3->shared) {
3513 struct array_cache *shared_array = l3->shared;
3514 int max = shared_array->limit - shared_array->avail;
3515 if (max) {
3516 if (batchcount > max)
3517 batchcount = max;
3518 memcpy(&(shared_array->entry[shared_array->avail]),
3519 ac->entry, sizeof(void *) * batchcount);
3520 shared_array->avail += batchcount;
3521 goto free_done;
3522 }
3523 }
3524
3525 free_block(cachep, ac->entry, batchcount, node);
3526 free_done:
3527 #if STATS
3528 {
3529 int i = 0;
3530 struct list_head *p;
3531
3532 p = l3->slabs_free.next;
3533 while (p != &(l3->slabs_free)) {
3534 struct slab *slabp;
3535
3536 slabp = list_entry(p, struct slab, list);
3537 BUG_ON(slabp->inuse);
3538
3539 i++;
3540 p = p->next;
3541 }
3542 STATS_SET_FREEABLE(cachep, i);
3543 }
3544 #endif
3545 spin_unlock(&l3->list_lock);
3546 ac->avail -= batchcount;
3547 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3548 }
3549
3550 /*
3551 * Release an obj back to its cache. If the obj has a constructed state, it must
3552 * be in this state _before_ it is released. Called with disabled ints.
3553 */
3554 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3555 {
3556 struct array_cache *ac = cpu_cache_get(cachep);
3557
3558 check_irq_off();
3559 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3560
3561 if (cache_free_alien(cachep, objp))
3562 return;
3563
3564 if (likely(ac->avail < ac->limit)) {
3565 STATS_INC_FREEHIT(cachep);
3566 ac->entry[ac->avail++] = objp;
3567 return;
3568 } else {
3569 STATS_INC_FREEMISS(cachep);
3570 cache_flusharray(cachep, ac);
3571 ac->entry[ac->avail++] = objp;
3572 }
3573 }
3574
3575 /**
3576 * kmem_cache_alloc - Allocate an object
3577 * @cachep: The cache to allocate from.
3578 * @flags: See kmalloc().
3579 *
3580 * Allocate an object from this cache. The flags are only relevant
3581 * if the cache has no available objects.
3582 */
3583 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3584 {
3585 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3586 }
3587 EXPORT_SYMBOL(kmem_cache_alloc);
3588
3589 /**
3590 * kmem_ptr_validate - check if an untrusted pointer might
3591 * be a slab entry.
3592 * @cachep: the cache we're checking against
3593 * @ptr: pointer to validate
3594 *
3595 * This verifies that the untrusted pointer looks sane:
3596 * it is _not_ a guarantee that the pointer is actually
3597 * part of the slab cache in question, but it at least
3598 * validates that the pointer can be dereferenced and
3599 * looks half-way sane.
3600 *
3601 * Currently only used for dentry validation.
3602 */
3603 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3604 {
3605 unsigned long addr = (unsigned long)ptr;
3606 unsigned long min_addr = PAGE_OFFSET;
3607 unsigned long align_mask = BYTES_PER_WORD - 1;
3608 unsigned long size = cachep->buffer_size;
3609 struct page *page;
3610
3611 if (unlikely(addr < min_addr))
3612 goto out;
3613 if (unlikely(addr > (unsigned long)high_memory - size))
3614 goto out;
3615 if (unlikely(addr & align_mask))
3616 goto out;
3617 if (unlikely(!kern_addr_valid(addr)))
3618 goto out;
3619 if (unlikely(!kern_addr_valid(addr + size - 1)))
3620 goto out;
3621 page = virt_to_page(ptr);
3622 if (unlikely(!PageSlab(page)))
3623 goto out;
3624 if (unlikely(page_get_cache(page) != cachep))
3625 goto out;
3626 return 1;
3627 out:
3628 return 0;
3629 }
3630
3631 #ifdef CONFIG_NUMA
3632 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3633 {
3634 return __cache_alloc_node(cachep, flags, nodeid,
3635 __builtin_return_address(0));
3636 }
3637 EXPORT_SYMBOL(kmem_cache_alloc_node);
3638
3639 static __always_inline void *
3640 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3641 {
3642 struct kmem_cache *cachep;
3643
3644 cachep = kmem_find_general_cachep(size, flags);
3645 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3646 return cachep;
3647 return kmem_cache_alloc_node(cachep, flags, node);
3648 }
3649
3650 #ifdef CONFIG_DEBUG_SLAB
3651 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3652 {
3653 return __do_kmalloc_node(size, flags, node,
3654 __builtin_return_address(0));
3655 }
3656 EXPORT_SYMBOL(__kmalloc_node);
3657
3658 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3659 int node, void *caller)
3660 {
3661 return __do_kmalloc_node(size, flags, node, caller);
3662 }
3663 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3664 #else
3665 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3666 {
3667 return __do_kmalloc_node(size, flags, node, NULL);
3668 }
3669 EXPORT_SYMBOL(__kmalloc_node);
3670 #endif /* CONFIG_DEBUG_SLAB */
3671 #endif /* CONFIG_NUMA */
3672
3673 /**
3674 * __do_kmalloc - allocate memory
3675 * @size: how many bytes of memory are required.
3676 * @flags: the type of memory to allocate (see kmalloc).
3677 * @caller: function caller for debug tracking of the caller
3678 */
3679 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3680 void *caller)
3681 {
3682 struct kmem_cache *cachep;
3683
3684 /* If you want to save a few bytes .text space: replace
3685 * __ with kmem_.
3686 * Then kmalloc uses the uninlined functions instead of the inline
3687 * functions.
3688 */
3689 cachep = __find_general_cachep(size, flags);
3690 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3691 return cachep;
3692 return __cache_alloc(cachep, flags, caller);
3693 }
3694
3695
3696 #ifdef CONFIG_DEBUG_SLAB
3697 void *__kmalloc(size_t size, gfp_t flags)
3698 {
3699 return __do_kmalloc(size, flags, __builtin_return_address(0));
3700 }
3701 EXPORT_SYMBOL(__kmalloc);
3702
3703 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3704 {
3705 return __do_kmalloc(size, flags, caller);
3706 }
3707 EXPORT_SYMBOL(__kmalloc_track_caller);
3708
3709 #else
3710 void *__kmalloc(size_t size, gfp_t flags)
3711 {
3712 return __do_kmalloc(size, flags, NULL);
3713 }
3714 EXPORT_SYMBOL(__kmalloc);
3715 #endif
3716
3717 /**
3718 * kmem_cache_free - Deallocate an object
3719 * @cachep: The cache the allocation was from.
3720 * @objp: The previously allocated object.
3721 *
3722 * Free an object which was previously allocated from this
3723 * cache.
3724 */
3725 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3726 {
3727 unsigned long flags;
3728
3729 BUG_ON(virt_to_cache(objp) != cachep);
3730
3731 local_irq_save(flags);
3732 debug_check_no_locks_freed(objp, obj_size(cachep));
3733 __cache_free(cachep, objp);
3734 local_irq_restore(flags);
3735 }
3736 EXPORT_SYMBOL(kmem_cache_free);
3737
3738 /**
3739 * kfree - free previously allocated memory
3740 * @objp: pointer returned by kmalloc.
3741 *
3742 * If @objp is NULL, no operation is performed.
3743 *
3744 * Don't free memory not originally allocated by kmalloc()
3745 * or you will run into trouble.
3746 */
3747 void kfree(const void *objp)
3748 {
3749 struct kmem_cache *c;
3750 unsigned long flags;
3751
3752 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3753 return;
3754 local_irq_save(flags);
3755 kfree_debugcheck(objp);
3756 c = virt_to_cache(objp);
3757 debug_check_no_locks_freed(objp, obj_size(c));
3758 __cache_free(c, (void *)objp);
3759 local_irq_restore(flags);
3760 }
3761 EXPORT_SYMBOL(kfree);
3762
3763 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3764 {
3765 return obj_size(cachep);
3766 }
3767 EXPORT_SYMBOL(kmem_cache_size);
3768
3769 const char *kmem_cache_name(struct kmem_cache *cachep)
3770 {
3771 return cachep->name;
3772 }
3773 EXPORT_SYMBOL_GPL(kmem_cache_name);
3774
3775 /*
3776 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3777 */
3778 static int alloc_kmemlist(struct kmem_cache *cachep)
3779 {
3780 int node;
3781 struct kmem_list3 *l3;
3782 struct array_cache *new_shared;
3783 struct array_cache **new_alien = NULL;
3784
3785 for_each_online_node(node) {
3786
3787 if (use_alien_caches) {
3788 new_alien = alloc_alien_cache(node, cachep->limit);
3789 if (!new_alien)
3790 goto fail;
3791 }
3792
3793 new_shared = NULL;
3794 if (cachep->shared) {
3795 new_shared = alloc_arraycache(node,
3796 cachep->shared*cachep->batchcount,
3797 0xbaadf00d);
3798 if (!new_shared) {
3799 free_alien_cache(new_alien);
3800 goto fail;
3801 }
3802 }
3803
3804 l3 = cachep->nodelists[node];
3805 if (l3) {
3806 struct array_cache *shared = l3->shared;
3807
3808 spin_lock_irq(&l3->list_lock);
3809
3810 if (shared)
3811 free_block(cachep, shared->entry,
3812 shared->avail, node);
3813
3814 l3->shared = new_shared;
3815 if (!l3->alien) {
3816 l3->alien = new_alien;
3817 new_alien = NULL;
3818 }
3819 l3->free_limit = (1 + nr_cpus_node(node)) *
3820 cachep->batchcount + cachep->num;
3821 spin_unlock_irq(&l3->list_lock);
3822 kfree(shared);
3823 free_alien_cache(new_alien);
3824 continue;
3825 }
3826 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3827 if (!l3) {
3828 free_alien_cache(new_alien);
3829 kfree(new_shared);
3830 goto fail;
3831 }
3832
3833 kmem_list3_init(l3);
3834 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3835 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3836 l3->shared = new_shared;
3837 l3->alien = new_alien;
3838 l3->free_limit = (1 + nr_cpus_node(node)) *
3839 cachep->batchcount + cachep->num;
3840 cachep->nodelists[node] = l3;
3841 }
3842 return 0;
3843
3844 fail:
3845 if (!cachep->next.next) {
3846 /* Cache is not active yet. Roll back what we did */
3847 node--;
3848 while (node >= 0) {
3849 if (cachep->nodelists[node]) {
3850 l3 = cachep->nodelists[node];
3851
3852 kfree(l3->shared);
3853 free_alien_cache(l3->alien);
3854 kfree(l3);
3855 cachep->nodelists[node] = NULL;
3856 }
3857 node--;
3858 }
3859 }
3860 return -ENOMEM;
3861 }
3862
3863 struct ccupdate_struct {
3864 struct kmem_cache *cachep;
3865 struct array_cache *new[NR_CPUS];
3866 };
3867
3868 static void do_ccupdate_local(void *info)
3869 {
3870 struct ccupdate_struct *new = info;
3871 struct array_cache *old;
3872
3873 check_irq_off();
3874 old = cpu_cache_get(new->cachep);
3875
3876 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3877 new->new[smp_processor_id()] = old;
3878 }
3879
3880 /* Always called with the cache_chain_mutex held */
3881 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3882 int batchcount, int shared)
3883 {
3884 struct ccupdate_struct *new;
3885 int i;
3886
3887 new = kzalloc(sizeof(*new), GFP_KERNEL);
3888 if (!new)
3889 return -ENOMEM;
3890
3891 for_each_online_cpu(i) {
3892 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3893 batchcount);
3894 if (!new->new[i]) {
3895 for (i--; i >= 0; i--)
3896 kfree(new->new[i]);
3897 kfree(new);
3898 return -ENOMEM;
3899 }
3900 }
3901 new->cachep = cachep;
3902
3903 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3904
3905 check_irq_on();
3906 cachep->batchcount = batchcount;
3907 cachep->limit = limit;
3908 cachep->shared = shared;
3909
3910 for_each_online_cpu(i) {
3911 struct array_cache *ccold = new->new[i];
3912 if (!ccold)
3913 continue;
3914 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3915 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3916 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3917 kfree(ccold);
3918 }
3919 kfree(new);
3920 return alloc_kmemlist(cachep);
3921 }
3922
3923 /* Called with cache_chain_mutex held always */
3924 static int enable_cpucache(struct kmem_cache *cachep)
3925 {
3926 int err;
3927 int limit, shared;
3928
3929 /*
3930 * The head array serves three purposes:
3931 * - create a LIFO ordering, i.e. return objects that are cache-warm
3932 * - reduce the number of spinlock operations.
3933 * - reduce the number of linked list operations on the slab and
3934 * bufctl chains: array operations are cheaper.
3935 * The numbers are guessed, we should auto-tune as described by
3936 * Bonwick.
3937 */
3938 if (cachep->buffer_size > 131072)
3939 limit = 1;
3940 else if (cachep->buffer_size > PAGE_SIZE)
3941 limit = 8;
3942 else if (cachep->buffer_size > 1024)
3943 limit = 24;
3944 else if (cachep->buffer_size > 256)
3945 limit = 54;
3946 else
3947 limit = 120;
3948
3949 /*
3950 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3951 * allocation behaviour: Most allocs on one cpu, most free operations
3952 * on another cpu. For these cases, an efficient object passing between
3953 * cpus is necessary. This is provided by a shared array. The array
3954 * replaces Bonwick's magazine layer.
3955 * On uniprocessor, it's functionally equivalent (but less efficient)
3956 * to a larger limit. Thus disabled by default.
3957 */
3958 shared = 0;
3959 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3960 shared = 8;
3961
3962 #if DEBUG
3963 /*
3964 * With debugging enabled, large batchcount lead to excessively long
3965 * periods with disabled local interrupts. Limit the batchcount
3966 */
3967 if (limit > 32)
3968 limit = 32;
3969 #endif
3970 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3971 if (err)
3972 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3973 cachep->name, -err);
3974 return err;
3975 }
3976
3977 /*
3978 * Drain an array if it contains any elements taking the l3 lock only if
3979 * necessary. Note that the l3 listlock also protects the array_cache
3980 * if drain_array() is used on the shared array.
3981 */
3982 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3983 struct array_cache *ac, int force, int node)
3984 {
3985 int tofree;
3986
3987 if (!ac || !ac->avail)
3988 return;
3989 if (ac->touched && !force) {
3990 ac->touched = 0;
3991 } else {
3992 spin_lock_irq(&l3->list_lock);
3993 if (ac->avail) {
3994 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3995 if (tofree > ac->avail)
3996 tofree = (ac->avail + 1) / 2;
3997 free_block(cachep, ac->entry, tofree, node);
3998 ac->avail -= tofree;
3999 memmove(ac->entry, &(ac->entry[tofree]),
4000 sizeof(void *) * ac->avail);
4001 }
4002 spin_unlock_irq(&l3->list_lock);
4003 }
4004 }
4005
4006 /**
4007 * cache_reap - Reclaim memory from caches.
4008 * @w: work descriptor
4009 *
4010 * Called from workqueue/eventd every few seconds.
4011 * Purpose:
4012 * - clear the per-cpu caches for this CPU.
4013 * - return freeable pages to the main free memory pool.
4014 *
4015 * If we cannot acquire the cache chain mutex then just give up - we'll try
4016 * again on the next iteration.
4017 */
4018 static void cache_reap(struct work_struct *w)
4019 {
4020 struct kmem_cache *searchp;
4021 struct kmem_list3 *l3;
4022 int node = numa_node_id();
4023 struct delayed_work *work =
4024 container_of(w, struct delayed_work, work);
4025
4026 if (!mutex_trylock(&cache_chain_mutex))
4027 /* Give up. Setup the next iteration. */
4028 goto out;
4029
4030 list_for_each_entry(searchp, &cache_chain, next) {
4031 check_irq_on();
4032
4033 /*
4034 * We only take the l3 lock if absolutely necessary and we
4035 * have established with reasonable certainty that
4036 * we can do some work if the lock was obtained.
4037 */
4038 l3 = searchp->nodelists[node];
4039
4040 reap_alien(searchp, l3);
4041
4042 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4043
4044 /*
4045 * These are racy checks but it does not matter
4046 * if we skip one check or scan twice.
4047 */
4048 if (time_after(l3->next_reap, jiffies))
4049 goto next;
4050
4051 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4052
4053 drain_array(searchp, l3, l3->shared, 0, node);
4054
4055 if (l3->free_touched)
4056 l3->free_touched = 0;
4057 else {
4058 int freed;
4059
4060 freed = drain_freelist(searchp, l3, (l3->free_limit +
4061 5 * searchp->num - 1) / (5 * searchp->num));
4062 STATS_ADD_REAPED(searchp, freed);
4063 }
4064 next:
4065 cond_resched();
4066 }
4067 check_irq_on();
4068 mutex_unlock(&cache_chain_mutex);
4069 next_reap_node();
4070 out:
4071 /* Set up the next iteration */
4072 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4073 }
4074
4075 #ifdef CONFIG_PROC_FS
4076
4077 static void print_slabinfo_header(struct seq_file *m)
4078 {
4079 /*
4080 * Output format version, so at least we can change it
4081 * without _too_ many complaints.
4082 */
4083 #if STATS
4084 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4085 #else
4086 seq_puts(m, "slabinfo - version: 2.1\n");
4087 #endif
4088 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4089 "<objperslab> <pagesperslab>");
4090 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4091 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4092 #if STATS
4093 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4094 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4095 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4096 #endif
4097 seq_putc(m, '\n');
4098 }
4099
4100 static void *s_start(struct seq_file *m, loff_t *pos)
4101 {
4102 loff_t n = *pos;
4103
4104 mutex_lock(&cache_chain_mutex);
4105 if (!n)
4106 print_slabinfo_header(m);
4107
4108 return seq_list_start(&cache_chain, *pos);
4109 }
4110
4111 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4112 {
4113 return seq_list_next(p, &cache_chain, pos);
4114 }
4115
4116 static void s_stop(struct seq_file *m, void *p)
4117 {
4118 mutex_unlock(&cache_chain_mutex);
4119 }
4120
4121 static int s_show(struct seq_file *m, void *p)
4122 {
4123 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4124 struct slab *slabp;
4125 unsigned long active_objs;
4126 unsigned long num_objs;
4127 unsigned long active_slabs = 0;
4128 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4129 const char *name;
4130 char *error = NULL;
4131 int node;
4132 struct kmem_list3 *l3;
4133
4134 active_objs = 0;
4135 num_slabs = 0;
4136 for_each_online_node(node) {
4137 l3 = cachep->nodelists[node];
4138 if (!l3)
4139 continue;
4140
4141 check_irq_on();
4142 spin_lock_irq(&l3->list_lock);
4143
4144 list_for_each_entry(slabp, &l3->slabs_full, list) {
4145 if (slabp->inuse != cachep->num && !error)
4146 error = "slabs_full accounting error";
4147 active_objs += cachep->num;
4148 active_slabs++;
4149 }
4150 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4151 if (slabp->inuse == cachep->num && !error)
4152 error = "slabs_partial inuse accounting error";
4153 if (!slabp->inuse && !error)
4154 error = "slabs_partial/inuse accounting error";
4155 active_objs += slabp->inuse;
4156 active_slabs++;
4157 }
4158 list_for_each_entry(slabp, &l3->slabs_free, list) {
4159 if (slabp->inuse && !error)
4160 error = "slabs_free/inuse accounting error";
4161 num_slabs++;
4162 }
4163 free_objects += l3->free_objects;
4164 if (l3->shared)
4165 shared_avail += l3->shared->avail;
4166
4167 spin_unlock_irq(&l3->list_lock);
4168 }
4169 num_slabs += active_slabs;
4170 num_objs = num_slabs * cachep->num;
4171 if (num_objs - active_objs != free_objects && !error)
4172 error = "free_objects accounting error";
4173
4174 name = cachep->name;
4175 if (error)
4176 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4177
4178 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4179 name, active_objs, num_objs, cachep->buffer_size,
4180 cachep->num, (1 << cachep->gfporder));
4181 seq_printf(m, " : tunables %4u %4u %4u",
4182 cachep->limit, cachep->batchcount, cachep->shared);
4183 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4184 active_slabs, num_slabs, shared_avail);
4185 #if STATS
4186 { /* list3 stats */
4187 unsigned long high = cachep->high_mark;
4188 unsigned long allocs = cachep->num_allocations;
4189 unsigned long grown = cachep->grown;
4190 unsigned long reaped = cachep->reaped;
4191 unsigned long errors = cachep->errors;
4192 unsigned long max_freeable = cachep->max_freeable;
4193 unsigned long node_allocs = cachep->node_allocs;
4194 unsigned long node_frees = cachep->node_frees;
4195 unsigned long overflows = cachep->node_overflow;
4196
4197 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4198 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4199 reaped, errors, max_freeable, node_allocs,
4200 node_frees, overflows);
4201 }
4202 /* cpu stats */
4203 {
4204 unsigned long allochit = atomic_read(&cachep->allochit);
4205 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4206 unsigned long freehit = atomic_read(&cachep->freehit);
4207 unsigned long freemiss = atomic_read(&cachep->freemiss);
4208
4209 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4210 allochit, allocmiss, freehit, freemiss);
4211 }
4212 #endif
4213 seq_putc(m, '\n');
4214 return 0;
4215 }
4216
4217 /*
4218 * slabinfo_op - iterator that generates /proc/slabinfo
4219 *
4220 * Output layout:
4221 * cache-name
4222 * num-active-objs
4223 * total-objs
4224 * object size
4225 * num-active-slabs
4226 * total-slabs
4227 * num-pages-per-slab
4228 * + further values on SMP and with statistics enabled
4229 */
4230
4231 const struct seq_operations slabinfo_op = {
4232 .start = s_start,
4233 .next = s_next,
4234 .stop = s_stop,
4235 .show = s_show,
4236 };
4237
4238 #define MAX_SLABINFO_WRITE 128
4239 /**
4240 * slabinfo_write - Tuning for the slab allocator
4241 * @file: unused
4242 * @buffer: user buffer
4243 * @count: data length
4244 * @ppos: unused
4245 */
4246 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4247 size_t count, loff_t *ppos)
4248 {
4249 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4250 int limit, batchcount, shared, res;
4251 struct kmem_cache *cachep;
4252
4253 if (count > MAX_SLABINFO_WRITE)
4254 return -EINVAL;
4255 if (copy_from_user(&kbuf, buffer, count))
4256 return -EFAULT;
4257 kbuf[MAX_SLABINFO_WRITE] = '\0';
4258
4259 tmp = strchr(kbuf, ' ');
4260 if (!tmp)
4261 return -EINVAL;
4262 *tmp = '\0';
4263 tmp++;
4264 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4265 return -EINVAL;
4266
4267 /* Find the cache in the chain of caches. */
4268 mutex_lock(&cache_chain_mutex);
4269 res = -EINVAL;
4270 list_for_each_entry(cachep, &cache_chain, next) {
4271 if (!strcmp(cachep->name, kbuf)) {
4272 if (limit < 1 || batchcount < 1 ||
4273 batchcount > limit || shared < 0) {
4274 res = 0;
4275 } else {
4276 res = do_tune_cpucache(cachep, limit,
4277 batchcount, shared);
4278 }
4279 break;
4280 }
4281 }
4282 mutex_unlock(&cache_chain_mutex);
4283 if (res >= 0)
4284 res = count;
4285 return res;
4286 }
4287
4288 #ifdef CONFIG_DEBUG_SLAB_LEAK
4289
4290 static void *leaks_start(struct seq_file *m, loff_t *pos)
4291 {
4292 mutex_lock(&cache_chain_mutex);
4293 return seq_list_start(&cache_chain, *pos);
4294 }
4295
4296 static inline int add_caller(unsigned long *n, unsigned long v)
4297 {
4298 unsigned long *p;
4299 int l;
4300 if (!v)
4301 return 1;
4302 l = n[1];
4303 p = n + 2;
4304 while (l) {
4305 int i = l/2;
4306 unsigned long *q = p + 2 * i;
4307 if (*q == v) {
4308 q[1]++;
4309 return 1;
4310 }
4311 if (*q > v) {
4312 l = i;
4313 } else {
4314 p = q + 2;
4315 l -= i + 1;
4316 }
4317 }
4318 if (++n[1] == n[0])
4319 return 0;
4320 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4321 p[0] = v;
4322 p[1] = 1;
4323 return 1;
4324 }
4325
4326 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4327 {
4328 void *p;
4329 int i;
4330 if (n[0] == n[1])
4331 return;
4332 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4333 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4334 continue;
4335 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4336 return;
4337 }
4338 }
4339
4340 static void show_symbol(struct seq_file *m, unsigned long address)
4341 {
4342 #ifdef CONFIG_KALLSYMS
4343 unsigned long offset, size;
4344 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4345
4346 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4347 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4348 if (modname[0])
4349 seq_printf(m, " [%s]", modname);
4350 return;
4351 }
4352 #endif
4353 seq_printf(m, "%p", (void *)address);
4354 }
4355
4356 static int leaks_show(struct seq_file *m, void *p)
4357 {
4358 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4359 struct slab *slabp;
4360 struct kmem_list3 *l3;
4361 const char *name;
4362 unsigned long *n = m->private;
4363 int node;
4364 int i;
4365
4366 if (!(cachep->flags & SLAB_STORE_USER))
4367 return 0;
4368 if (!(cachep->flags & SLAB_RED_ZONE))
4369 return 0;
4370
4371 /* OK, we can do it */
4372
4373 n[1] = 0;
4374
4375 for_each_online_node(node) {
4376 l3 = cachep->nodelists[node];
4377 if (!l3)
4378 continue;
4379
4380 check_irq_on();
4381 spin_lock_irq(&l3->list_lock);
4382
4383 list_for_each_entry(slabp, &l3->slabs_full, list)
4384 handle_slab(n, cachep, slabp);
4385 list_for_each_entry(slabp, &l3->slabs_partial, list)
4386 handle_slab(n, cachep, slabp);
4387 spin_unlock_irq(&l3->list_lock);
4388 }
4389 name = cachep->name;
4390 if (n[0] == n[1]) {
4391 /* Increase the buffer size */
4392 mutex_unlock(&cache_chain_mutex);
4393 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4394 if (!m->private) {
4395 /* Too bad, we are really out */
4396 m->private = n;
4397 mutex_lock(&cache_chain_mutex);
4398 return -ENOMEM;
4399 }
4400 *(unsigned long *)m->private = n[0] * 2;
4401 kfree(n);
4402 mutex_lock(&cache_chain_mutex);
4403 /* Now make sure this entry will be retried */
4404 m->count = m->size;
4405 return 0;
4406 }
4407 for (i = 0; i < n[1]; i++) {
4408 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4409 show_symbol(m, n[2*i+2]);
4410 seq_putc(m, '\n');
4411 }
4412
4413 return 0;
4414 }
4415
4416 const struct seq_operations slabstats_op = {
4417 .start = leaks_start,
4418 .next = s_next,
4419 .stop = s_stop,
4420 .show = leaks_show,
4421 };
4422 #endif
4423 #endif
4424
4425 /**
4426 * ksize - get the actual amount of memory allocated for a given object
4427 * @objp: Pointer to the object
4428 *
4429 * kmalloc may internally round up allocations and return more memory
4430 * than requested. ksize() can be used to determine the actual amount of
4431 * memory allocated. The caller may use this additional memory, even though
4432 * a smaller amount of memory was initially specified with the kmalloc call.
4433 * The caller must guarantee that objp points to a valid object previously
4434 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4435 * must not be freed during the duration of the call.
4436 */
4437 size_t ksize(const void *objp)
4438 {
4439 if (unlikely(ZERO_OR_NULL_PTR(objp)))
4440 return 0;
4441
4442 return obj_size(virt_to_cache(objp));
4443 }