Loading...
Note: File does not exist in v6.8.
1/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
87 */
88
89#include <linux/__KEEPIDENTS__B.h>
90#include <linux/__KEEPIDENTS__C.h>
91#include <linux/__KEEPIDENTS__D.h>
92#include <linux/__KEEPIDENTS__E.h>
93#include <linux/__KEEPIDENTS__F.h>
94#include <linux/__KEEPIDENTS__G.h>
95#include <linux/__KEEPIDENTS__H.h>
96#include <linux/__KEEPIDENTS__I.h>
97#include <linux/__KEEPIDENTS__J.h>
98#include <linux/proc_fs.h>
99#include <linux/__KEEPIDENTS__BA.h>
100#include <linux/__KEEPIDENTS__BB.h>
101#include <linux/__KEEPIDENTS__BC.h>
102#include <linux/cpu.h>
103#include <linux/__KEEPIDENTS__BD.h>
104#include <linux/__KEEPIDENTS__BE.h>
105#include <linux/rcupdate.h>
106#include <linux/__KEEPIDENTS__BF.h>
107#include <linux/__KEEPIDENTS__BG.h>
108#include <linux/__KEEPIDENTS__BH.h>
109#include <linux/kmemleak.h>
110#include <linux/__KEEPIDENTS__BI.h>
111#include <linux/__KEEPIDENTS__BJ.h>
112#include <linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
113#include <linux/__KEEPIDENTS__CC.h>
114#include <linux/reciprocal_div.h>
115#include <linux/debugobjects.h>
116#include <linux/kmemcheck.h>
117#include <linux/__KEEPIDENTS__CD.h>
118#include <linux/__KEEPIDENTS__CE.h>
119
120#include <asm/cacheflush.h>
121#include <asm/tlbflush.h>
122#include <asm/page.h>
123
124/*
125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
126 * 0 for faster, smaller code (especially in the critical paths).
127 *
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
130 *
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 */
133
134#ifdef CONFIG_DEBUG_SLAB
135#define DEBUG 1
136#define STATS 1
137#define FORCED_DEBUG 1
138#else
139#define DEBUG 0
140#define STATS 0
141#define FORCED_DEBUG 0
142#endif
143
144/* Shouldn't this be in a header file somewhere? */
145#define BYTES_PER_WORD sizeof(void *)
146#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
147
148#ifndef ARCH_KMALLOC_FLAGS
149#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150#endif
151
152/* Legal flag mask for kmem_cache_create(). */
153#if DEBUG
154# define CREATE_MASK (SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
156 SLAB_CACHE_DMA | \
157 SLAB_STORE_USER | \
158 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
159 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
160 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161#else
162# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163 SLAB_CACHE_DMA | \
164 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
166 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
167#endif
168
169/*
170 * kmem_bufctl_t:
171 *
172 * Bufctl's are used for linking objs within a slab
173 * linked offsets.
174 *
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
186 */
187
188typedef unsigned int kmem_bufctl_t;
189#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
191#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
193
194/*
195 * struct slab_rcu
196 *
197 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
198 * arrange for kmem_freepages to be called via RCU. This is useful if
199 * we need to approach a kernel structure obliquely, from its address
200 * obtained without the usual locking. We can lock the structure to
201 * stabilize it and check it's still at the given address, only if we
202 * can be sure that the memory has not been meanwhile reused for some
203 * other kind of object (which our subsystem's lock might corrupt).
204 *
205 * rcu_read_lock before reading the address, then rcu_read_unlock after
206 * taking the spinlock within the structure expected at that address.
207 */
208struct slab_rcu {
209 struct rcu_head head;
210 struct kmem_cache *cachep;
211 void *addr;
212};
213
214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
222 union {
223 struct {
224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
228 kmem_bufctl_t free;
229 unsigned short nodeid;
230 };
231 struct slab_rcu __slab_cover_slab_rcu;
232 };
233};
234
235/*
236 * struct array_cache
237 *
238 * Purpose:
239 * - LIFO ordering, to hand out cache-warm objects from _alloc
240 * - reduce the number of linked list operations
241 * - reduce spinlock operations
242 *
243 * The limit is stored in the per-cpu structure to reduce the data cache
244 * footprint.
245 *
246 */
247struct array_cache {
248 unsigned int avail;
249 unsigned int limit;
250 unsigned int batchcount;
251 unsigned int touched;
252 spinlock_t lock;
253 void *entry[]; /*
254 * Must have this definition in here for the proper
255 * alignment of array_cache. Also simplifies accessing
256 * the entries.
257 */
258};
259
260/*
261 * bootstrap: The caches do not work without cpuarrays anymore, but the
262 * cpuarrays are allocated from the generic caches...
263 */
264#define BOOT_CPUCACHE_ENTRIES 1
265struct arraycache_init {
266 struct array_cache cache;
267 void *entries[BOOT_CPUCACHE_ENTRIES];
268};
269
270/*
271 * The slab lists for all objects.
272 */
273struct kmem_list3 {
274 struct list_head slabs_partial; /* partial list first, better asm code */
275 struct list_head slabs_full;
276 struct list_head slabs_free;
277 unsigned long free_objects;
278 unsigned int free_limit;
279 unsigned int colour_next; /* Per-node cache coloring */
280 spinlock_t list_lock;
281 struct array_cache *shared; /* shared per node */
282 struct array_cache **alien; /* on other nodes */
283 unsigned long next_reap; /* updated without locking */
284 int free_touched; /* updated without locking */
285};
286
287/*
288 * Need this for bootstrapping a per node allocator.
289 */
290#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
291static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
292#define CACHE_CACHE 0
293#define SIZE_AC MAX_NUMNODES
294#define SIZE_L3 (2 * MAX_NUMNODES)
295
296static int drain_freelist(struct kmem_cache *cache,
297 struct kmem_list3 *l3, int tofree);
298static void free_block(struct kmem_cache *cachep, void **objpp, int len,
299 int node);
300static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
301static void cache_reap(struct work_struct *unused);
302
303/*
304 * This function must be completely optimized away if a constant is passed to
305 * it. Mostly the same as what is in linux/slab.h except it returns an index.
306 */
307static __always_inline int index_of(const size_t size)
308{
309 extern void __bad_size(void);
310
311 if (__builtin_constant_p(size)) {
312 int i = 0;
313
314#define CACHE(x) \
315 if (size <=x) \
316 return i; \
317 else \
318 i++;
319#include <linux/kmalloc_sizes.h>
320#undef CACHE
321 __bad_size();
322 } else
323 __bad_size();
324 return 0;
325}
326
327static int slab_early_init = 1;
328
329#define INDEX_AC index_of(sizeof(struct arraycache_init))
330#define INDEX_L3 index_of(sizeof(struct kmem_list3))
331
332static void kmem_list3_init(struct kmem_list3 *parent)
333{
334 INIT_LIST_HEAD(&parent->slabs_full);
335 INIT_LIST_HEAD(&parent->slabs_partial);
336 INIT_LIST_HEAD(&parent->slabs_free);
337 parent->shared = NULL;
338 parent->alien = NULL;
339 parent->colour_next = 0;
340 spin_lock_init(&parent->list_lock);
341 parent->free_objects = 0;
342 parent->free_touched = 0;
343}
344
345#define MAKE_LIST(cachep, listp, slab, nodeid) \
346 do { \
347 INIT_LIST_HEAD(listp); \
348 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
349 } while (0)
350
351#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
352 do { \
353 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
354 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
355 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
356 } while (0)
357
358#define CFLGS_OFF_SLAB (0x80000000UL)
359#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
360
361#define BATCHREFILL_LIMIT 16
362/*
363 * Optimization question: fewer reaps means less probability for unnessary
364 * cpucache drain/refill cycles.
365 *
366 * OTOH the cpuarrays can contain lots of objects,
367 * which could lock up otherwise freeable slabs.
368 */
369#define REAPTIMEOUT_CPUC (2*HZ)
370#define REAPTIMEOUT_LIST3 (4*HZ)
371
372#if STATS
373#define STATS_INC_ACTIVE(x) ((x)->num_active++)
374#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
375#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
376#define STATS_INC_GROWN(x) ((x)->grown++)
377#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
378#define STATS_SET_HIGH(x) \
379 do { \
380 if ((x)->num_active > (x)->high_mark) \
381 (x)->high_mark = (x)->num_active; \
382 } while (0)
383#define STATS_INC_ERR(x) ((x)->errors++)
384#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
385#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
386#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
387#define STATS_SET_FREEABLE(x, i) \
388 do { \
389 if ((x)->max_freeable < i) \
390 (x)->max_freeable = i; \
391 } while (0)
392#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
393#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
394#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
395#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
396#else
397#define STATS_INC_ACTIVE(x) do { } while (0)
398#define STATS_DEC_ACTIVE(x) do { } while (0)
399#define STATS_INC_ALLOCED(x) do { } while (0)
400#define STATS_INC_GROWN(x) do { } while (0)
401#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
402#define STATS_SET_HIGH(x) do { } while (0)
403#define STATS_INC_ERR(x) do { } while (0)
404#define STATS_INC_NODEALLOCS(x) do { } while (0)
405#define STATS_INC_NODEFREES(x) do { } while (0)
406#define STATS_INC_ACOVERFLOW(x) do { } while (0)
407#define STATS_SET_FREEABLE(x, i) do { } while (0)
408#define STATS_INC_ALLOCHIT(x) do { } while (0)
409#define STATS_INC_ALLOCMISS(x) do { } while (0)
410#define STATS_INC_FREEHIT(x) do { } while (0)
411#define STATS_INC_FREEMISS(x) do { } while (0)
412#endif
413
414#if DEBUG
415
416/*
417 * memory layout of objects:
418 * 0 : objp
419 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
420 * the end of an object is aligned with the end of the real
421 * allocation. Catches writes behind the end of the allocation.
422 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
423 * redzone word.
424 * cachep->obj_offset: The real object.
425 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
426 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
427 * [BYTES_PER_WORD long]
428 */
429static int obj_offset(struct kmem_cache *cachep)
430{
431 return cachep->obj_offset;
432}
433
434static int obj_size(struct kmem_cache *cachep)
435{
436 return cachep->obj_size;
437}
438
439static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
440{
441 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
442 return (unsigned long long*) (objp + obj_offset(cachep) -
443 sizeof(unsigned long long));
444}
445
446static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
447{
448 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449 if (cachep->flags & SLAB_STORE_USER)
450 return (unsigned long long *)(objp + cachep->buffer_size -
451 sizeof(unsigned long long) -
452 REDZONE_ALIGN);
453 return (unsigned long long *) (objp + cachep->buffer_size -
454 sizeof(unsigned long long));
455}
456
457static void **dbg_userword(struct kmem_cache *cachep, void *objp)
458{
459 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
460 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
461}
462
463#else
464
465#define obj_offset(x) 0
466#define obj_size(cachep) (cachep->buffer_size)
467#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
469#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470
471#endif
472
473#ifdef CONFIG_TRACING
474size_t slab_buffer_size(struct kmem_cache *cachep)
475{
476 return cachep->buffer_size;
477}
478EXPORT_SYMBOL(slab_buffer_size);
479#endif
480
481/*
482 * Do not go above this order unless 0 objects fit into the slab.
483 */
484#define BREAK_GFP_ORDER_HI 1
485#define BREAK_GFP_ORDER_LO 0
486static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
487
488/*
489 * Functions for storing/retrieving the cachep and or slab from the page
490 * allocator. These are used to find the slab an obj belongs to. With kfree(),
491 * these are used to find the cache which an obj belongs to.
492 */
493static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
494{
495 page->lru.next = (struct list_head *)cache;
496}
497
498static inline struct kmem_cache *page_get_cache(struct page *page)
499{
500 page = compound_head(page);
501 BUG_ON(!PageSlab(page));
502 return (struct kmem_cache *)page->lru.next;
503}
504
505static inline void page_set_slab(struct page *page, struct slab *slab)
506{
507 page->lru.prev = (struct list_head *)slab;
508}
509
510static inline struct slab *page_get_slab(struct page *page)
511{
512 BUG_ON(!PageSlab(page));
513 return (struct slab *)page->lru.prev;
514}
515
516static inline struct kmem_cache *virt_to_cache(const void *obj)
517{
518 struct page *page = virt_to_head_page(obj);
519 return page_get_cache(page);
520}
521
522static inline struct slab *virt_to_slab(const void *obj)
523{
524 struct page *page = virt_to_head_page(obj);
525 return page_get_slab(page);
526}
527
528static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
529 unsigned int idx)
530{
531 return slab->s_mem + cache->buffer_size * idx;
532}
533
534/*
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
539 */
540static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541 const struct slab *slab, void *obj)
542{
543 u32 offset = (obj - slab->s_mem);
544 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
545}
546
547/*
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
549 */
550struct cache_sizes malloc_sizes[] = {
551#define CACHE(x) { .cs_size = (x) },
552#include <linux/kmalloc_sizes.h>
553 CACHE(ULONG_MAX)
554#undef CACHE
555};
556EXPORT_SYMBOL(malloc_sizes);
557
558/* Must match cache_sizes above. Out of line to keep cache footprint low. */
559struct cache_names {
560 char *name;
561 char *name_dma;
562};
563
564static struct cache_names __initdata cache_names[] = {
565#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566#include <linux/kmalloc_sizes.h>
567 {NULL,}
568#undef CACHE
569};
570
571static struct arraycache_init initarray_cache __initdata =
572 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
573static struct arraycache_init initarray_generic =
574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
575
576/* internal cache of cache description objs */
577static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
578static struct kmem_cache cache_cache = {
579 .nodelists = cache_cache_nodelists,
580 .batchcount = 1,
581 .limit = BOOT_CPUCACHE_ENTRIES,
582 .shared = 1,
583 .buffer_size = sizeof(struct kmem_cache),
584 .name = "kmem_cache",
585};
586
587#define BAD_ALIEN_MAGIC 0x01020304ul
588
589/*
590 * chicken and egg problem: delay the per-cpu array allocation
591 * until the general caches are up.
592 */
593static enum {
594 NONE,
595 PARTIAL_AC,
596 PARTIAL_L3,
597 EARLY,
598 FULL
599} g_cpucache_up;
600
601/*
602 * used by boot code to determine if it can use slab based allocator
603 */
604int slab_is_available(void)
605{
606 return g_cpucache_up >= EARLY;
607}
608
609#ifdef CONFIG_LOCKDEP
610
611/*
612 * Slab sometimes uses the kmalloc slabs to store the slab headers
613 * for other slabs "off slab".
614 * The locking for this is tricky in that it nests within the locks
615 * of all other slabs in a few places; to deal with this special
616 * locking we put on-slab caches into a separate lock-class.
617 *
618 * We set lock class for alien array caches which are up during init.
619 * The lock annotation will be lost if all cpus of a node goes down and
620 * then comes back up during hotplug
621 */
622static struct lock_class_key on_slab_l3_key;
623static struct lock_class_key on_slab_alc_key;
624
625static struct lock_class_key debugobj_l3_key;
626static struct lock_class_key debugobj_alc_key;
627
628static void slab_set_lock_classes(struct kmem_cache *cachep,
629 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
630 int q)
631{
632 struct array_cache **alc;
633 struct kmem_list3 *l3;
634 int r;
635
636 l3 = cachep->nodelists[q];
637 if (!l3)
638 return;
639
640 lockdep_set_class(&l3->list_lock, l3_key);
641 alc = l3->alien;
642 /*
643 * FIXME: This check for BAD_ALIEN_MAGIC
644 * should go away when common slab code is taught to
645 * work even without alien caches.
646 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
647 * for alloc_alien_cache,
648 */
649 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
650 return;
651 for_each_node(r) {
652 if (alc[r])
653 lockdep_set_class(&alc[r]->lock, alc_key);
654 }
655}
656
657static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
658{
659 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
660}
661
662static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
663{
664 int node;
665
666 for_each_online_node(node)
667 slab_set_debugobj_lock_classes_node(cachep, node);
668}
669
670static void init_node_lock_keys(int q)
671{
672 struct cache_sizes *s = malloc_sizes;
673
674 if (g_cpucache_up != FULL)
675 return;
676
677 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
678 struct kmem_list3 *l3;
679
680 l3 = s->cs_cachep->nodelists[q];
681 if (!l3 || OFF_SLAB(s->cs_cachep))
682 continue;
683
684 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
685 &on_slab_alc_key, q);
686 }
687}
688
689static inline void init_lock_keys(void)
690{
691 int node;
692
693 for_each_node(node)
694 init_node_lock_keys(node);
695}
696#else
697static void init_node_lock_keys(int q)
698{
699}
700
701static inline void init_lock_keys(void)
702{
703}
704
705static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
706{
707}
708
709static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
710{
711}
712#endif
713
714/*
715 * Guard access to the cache-chain.
716 */
717static DEFINE_MUTEX(cache_chain_mutex);
718static struct list_head cache_chain;
719
720static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
721
722static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
723{
724 return cachep->array[smp_processor_id()];
725}
726
727static inline struct kmem_cache *__find_general_cachep(size_t size,
728 gfp_t gfpflags)
729{
730 struct cache_sizes *csizep = malloc_sizes;
731
732#if DEBUG
733 /* This happens if someone tries to call
734 * kmem_cache_create(), or __kmalloc(), before
735 * the generic caches are initialized.
736 */
737 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
738#endif
739 if (!size)
740 return ZERO_SIZE_PTR;
741
742 while (size > csizep->cs_size)
743 csizep++;
744
745 /*
746 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
747 * has cs_{dma,}cachep==NULL. Thus no special case
748 * for large kmalloc calls required.
749 */
750#ifdef CONFIG_ZONE_DMA
751 if (unlikely(gfpflags & GFP_DMA))
752 return csizep->cs_dmacachep;
753#endif
754 return csizep->cs_cachep;
755}
756
757static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
758{
759 return __find_general_cachep(size, gfpflags);
760}
761
762static size_t slab_mgmt_size(size_t nr_objs, size_t align)
763{
764 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
765}
766
767/*
768 * Calculate the number of objects and left-over bytes for a given buffer size.
769 */
770static void cache_estimate(unsigned long gfporder, size_t buffer_size,
771 size_t align, int flags, size_t *left_over,
772 unsigned int *num)
773{
774 int nr_objs;
775 size_t mgmt_size;
776 size_t slab_size = PAGE_SIZE << gfporder;
777
778 /*
779 * The slab management structure can be either off the slab or
780 * on it. For the latter case, the memory allocated for a
781 * slab is used for:
782 *
783 * - The struct slab
784 * - One kmem_bufctl_t for each object
785 * - Padding to respect alignment of @align
786 * - @buffer_size bytes for each object
787 *
788 * If the slab management structure is off the slab, then the
789 * alignment will already be calculated into the size. Because
790 * the slabs are all pages aligned, the objects will be at the
791 * correct alignment when allocated.
792 */
793 if (flags & CFLGS_OFF_SLAB) {
794 mgmt_size = 0;
795 nr_objs = slab_size / buffer_size;
796
797 if (nr_objs > SLAB_LIMIT)
798 nr_objs = SLAB_LIMIT;
799 } else {
800 /*
801 * Ignore padding for the initial guess. The padding
802 * is at most @align-1 bytes, and @buffer_size is at
803 * least @align. In the worst case, this result will
804 * be one greater than the number of objects that fit
805 * into the memory allocation when taking the padding
806 * into account.
807 */
808 nr_objs = (slab_size - sizeof(struct slab)) /
809 (buffer_size + sizeof(kmem_bufctl_t));
810
811 /*
812 * This calculated number will be either the right
813 * amount, or one greater than what we want.
814 */
815 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
816 > slab_size)
817 nr_objs--;
818
819 if (nr_objs > SLAB_LIMIT)
820 nr_objs = SLAB_LIMIT;
821
822 mgmt_size = slab_mgmt_size(nr_objs, align);
823 }
824 *num = nr_objs;
825 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
826}
827
828#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
829
830static void __slab_error(const char *function, struct kmem_cache *cachep,
831 char *msg)
832{
833 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
834 function, cachep->name, msg);
835 dump_stack();
836}
837
838/*
839 * By default on NUMA we use alien caches to stage the freeing of
840 * objects allocated from other nodes. This causes massive memory
841 * inefficiencies when using fake NUMA setup to split memory into a
842 * large number of small nodes, so it can be disabled on the command
843 * line
844 */
845
846static int use_alien_caches __read_mostly = 1;
847static int __init noaliencache_setup(char *s)
848{
849 use_alien_caches = 0;
850 return 1;
851}
852__setup("noaliencache", noaliencache_setup);
853
854#ifdef CONFIG_NUMA
855/*
856 * Special reaping functions for NUMA systems called from cache_reap().
857 * These take care of doing round robin flushing of alien caches (containing
858 * objects freed on different nodes from which they were allocated) and the
859 * flushing of remote pcps by calling drain_node_pages.
860 */
861static DEFINE_PER_CPU(unsigned long, slab_reap_node);
862
863static void init_reap_node(int cpu)
864{
865 int node;
866
867 node = next_node(cpu_to_mem(cpu), node_online_map);
868 if (node == MAX_NUMNODES)
869 node = first_node(node_online_map);
870
871 per_cpu(slab_reap_node, cpu) = node;
872}
873
874static void next_reap_node(void)
875{
876 int node = __this_cpu_read(slab_reap_node);
877
878 node = next_node(node, node_online_map);
879 if (unlikely(node >= MAX_NUMNODES))
880 node = first_node(node_online_map);
881 __this_cpu_write(slab_reap_node, node);
882}
883
884#else
885#define init_reap_node(cpu) do { } while (0)
886#define next_reap_node(void) do { } while (0)
887#endif
888
889/*
890 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
891 * via the workqueue/eventd.
892 * Add the CPU number into the expiration time to minimize the possibility of
893 * the CPUs getting into lockstep and contending for the global cache chain
894 * lock.
895 */
896static void __cpuinit start_cpu_timer(int cpu)
897{
898 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
899
900 /*
901 * When this gets called from do_initcalls via cpucache_init(),
902 * init_workqueues() has already run, so keventd will be setup
903 * at that time.
904 */
905 if (keventd_up() && reap_work->work.func == NULL) {
906 init_reap_node(cpu);
907 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
908 schedule_delayed_work_on(cpu, reap_work,
909 __round_jiffies_relative(HZ, cpu));
910 }
911}
912
913static struct array_cache *alloc_arraycache(int node, int entries,
914 int batchcount, gfp_t gfp)
915{
916 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
917 struct array_cache *nc = NULL;
918
919 nc = kmalloc_node(memsize, gfp, node);
920 /*
921 * The array_cache structures contain pointers to free object.
922 * However, when such objects are allocated or transferred to another
923 * cache the pointers are not cleared and they could be counted as
924 * valid references during a kmemleak scan. Therefore, kmemleak must
925 * not scan such objects.
926 */
927 kmemleak_no_scan(nc);
928 if (nc) {
929 nc->avail = 0;
930 nc->limit = entries;
931 nc->batchcount = batchcount;
932 nc->touched = 0;
933 spin_lock_init(&nc->lock);
934 }
935 return nc;
936}
937
938/*
939 * Transfer objects in one arraycache to another.
940 * Locking must be handled by the caller.
941 *
942 * Return the number of entries transferred.
943 */
944static int transfer_objects(struct array_cache *to,
945 struct array_cache *from, unsigned int max)
946{
947 /* Figure out how many entries to transfer */
948 int nr = min3(from->avail, max, to->limit - to->avail);
949
950 if (!nr)
951 return 0;
952
953 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
954 sizeof(void *) *nr);
955
956 from->avail -= nr;
957 to->avail += nr;
958 return nr;
959}
960
961#ifndef CONFIG_NUMA
962
963#define drain_alien_cache(cachep, alien) do { } while (0)
964#define reap_alien(cachep, l3) do { } while (0)
965
966static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
967{
968 return (struct array_cache **)BAD_ALIEN_MAGIC;
969}
970
971static inline void free_alien_cache(struct array_cache **ac_ptr)
972{
973}
974
975static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
976{
977 return 0;
978}
979
980static inline void *alternate_node_alloc(struct kmem_cache *cachep,
981 gfp_t flags)
982{
983 return NULL;
984}
985
986static inline void *____cache_alloc_node(struct kmem_cache *cachep,
987 gfp_t flags, int nodeid)
988{
989 return NULL;
990}
991
992#else /* CONFIG_NUMA */
993
994static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
995static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
996
997static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
998{
999 struct array_cache **ac_ptr;
1000 int memsize = sizeof(void *) * nr_node_ids;
1001 int i;
1002
1003 if (limit > 1)
1004 limit = 12;
1005 ac_ptr = kzalloc_node(memsize, gfp, node);
1006 if (ac_ptr) {
1007 for_each_node(i) {
1008 if (i == node || !node_online(i))
1009 continue;
1010 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1011 if (!ac_ptr[i]) {
1012 for (i--; i >= 0; i--)
1013 kfree(ac_ptr[i]);
1014 kfree(ac_ptr);
1015 return NULL;
1016 }
1017 }
1018 }
1019 return ac_ptr;
1020}
1021
1022static void free_alien_cache(struct array_cache **ac_ptr)
1023{
1024 int i;
1025
1026 if (!ac_ptr)
1027 return;
1028 for_each_node(i)
1029 kfree(ac_ptr[i]);
1030 kfree(ac_ptr);
1031}
1032
1033static void __drain_alien_cache(struct kmem_cache *cachep,
1034 struct array_cache *ac, int node)
1035{
1036 struct kmem_list3 *rl3 = cachep->nodelists[node];
1037
1038 if (ac->avail) {
1039 spin_lock(&rl3->list_lock);
1040 /*
1041 * Stuff objects into the remote nodes shared array first.
1042 * That way we could avoid the overhead of putting the objects
1043 * into the free lists and getting them back later.
1044 */
1045 if (rl3->shared)
1046 transfer_objects(rl3->shared, ac, ac->limit);
1047
1048 free_block(cachep, ac->entry, ac->avail, node);
1049 ac->avail = 0;
1050 spin_unlock(&rl3->list_lock);
1051 }
1052}
1053
1054/*
1055 * Called from cache_reap() to regularly drain alien caches round robin.
1056 */
1057static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1058{
1059 int node = __this_cpu_read(slab_reap_node);
1060
1061 if (l3->alien) {
1062 struct array_cache *ac = l3->alien[node];
1063
1064 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1065 __drain_alien_cache(cachep, ac, node);
1066 spin_unlock_irq(&ac->lock);
1067 }
1068 }
1069}
1070
1071static void drain_alien_cache(struct kmem_cache *cachep,
1072 struct array_cache **alien)
1073{
1074 int i = 0;
1075 struct array_cache *ac;
1076 unsigned long flags;
1077
1078 for_each_online_node(i) {
1079 ac = alien[i];
1080 if (ac) {
1081 spin_lock_irqsave(&ac->lock, flags);
1082 __drain_alien_cache(cachep, ac, i);
1083 spin_unlock_irqrestore(&ac->lock, flags);
1084 }
1085 }
1086}
1087
1088static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1089{
1090 struct slab *slabp = virt_to_slab(objp);
1091 int nodeid = slabp->nodeid;
1092 struct kmem_list3 *l3;
1093 struct array_cache *alien = NULL;
1094 int node;
1095
1096 node = numa_mem_id();
1097
1098 /*
1099 * Make sure we are not freeing a object from another node to the array
1100 * cache on this cpu.
1101 */
1102 if (likely(slabp->nodeid == node))
1103 return 0;
1104
1105 l3 = cachep->nodelists[node];
1106 STATS_INC_NODEFREES(cachep);
1107 if (l3->alien && l3->alien[nodeid]) {
1108 alien = l3->alien[nodeid];
1109 spin_lock(&alien->lock);
1110 if (unlikely(alien->avail == alien->limit)) {
1111 STATS_INC_ACOVERFLOW(cachep);
1112 __drain_alien_cache(cachep, alien, nodeid);
1113 }
1114 alien->entry[alien->avail++] = objp;
1115 spin_unlock(&alien->lock);
1116 } else {
1117 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1118 free_block(cachep, &objp, 1, nodeid);
1119 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1120 }
1121 return 1;
1122}
1123#endif
1124
1125/*
1126 * Allocates and initializes nodelists for a node on each slab cache, used for
1127 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1128 * will be allocated off-node since memory is not yet online for the new node.
1129 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1130 * already in use.
1131 *
1132 * Must hold cache_chain_mutex.
1133 */
1134static int init_cache_nodelists_node(int node)
1135{
1136 struct kmem_cache *cachep;
1137 struct kmem_list3 *l3;
1138 const int memsize = sizeof(struct kmem_list3);
1139
1140 list_for_each_entry(cachep, &cache_chain, next) {
1141 /*
1142 * Set up the size64 kmemlist for cpu before we can
1143 * begin anything. Make sure some other cpu on this
1144 * node has not already allocated this
1145 */
1146 if (!cachep->nodelists[node]) {
1147 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1148 if (!l3)
1149 return -ENOMEM;
1150 kmem_list3_init(l3);
1151 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1152 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1153
1154 /*
1155 * The l3s don't come and go as CPUs come and
1156 * go. cache_chain_mutex is sufficient
1157 * protection here.
1158 */
1159 cachep->nodelists[node] = l3;
1160 }
1161
1162 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1163 cachep->nodelists[node]->free_limit =
1164 (1 + nr_cpus_node(node)) *
1165 cachep->batchcount + cachep->num;
1166 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1167 }
1168 return 0;
1169}
1170
1171static void __cpuinit cpuup_canceled(long cpu)
1172{
1173 struct kmem_cache *cachep;
1174 struct kmem_list3 *l3 = NULL;
1175 int node = cpu_to_mem(cpu);
1176 const struct cpumask *mask = cpumask_of_node(node);
1177
1178 list_for_each_entry(cachep, &cache_chain, next) {
1179 struct array_cache *nc;
1180 struct array_cache *shared;
1181 struct array_cache **alien;
1182
1183 /* cpu is dead; no one can alloc from it. */
1184 nc = cachep->array[cpu];
1185 cachep->array[cpu] = NULL;
1186 l3 = cachep->nodelists[node];
1187
1188 if (!l3)
1189 goto free_array_cache;
1190
1191 spin_lock_irq(&l3->list_lock);
1192
1193 /* Free limit for this kmem_list3 */
1194 l3->free_limit -= cachep->batchcount;
1195 if (nc)
1196 free_block(cachep, nc->entry, nc->avail, node);
1197
1198 if (!cpumask_empty(mask)) {
1199 spin_unlock_irq(&l3->list_lock);
1200 goto free_array_cache;
1201 }
1202
1203 shared = l3->shared;
1204 if (shared) {
1205 free_block(cachep, shared->entry,
1206 shared->avail, node);
1207 l3->shared = NULL;
1208 }
1209
1210 alien = l3->alien;
1211 l3->alien = NULL;
1212
1213 spin_unlock_irq(&l3->list_lock);
1214
1215 kfree(shared);
1216 if (alien) {
1217 drain_alien_cache(cachep, alien);
1218 free_alien_cache(alien);
1219 }
1220free_array_cache:
1221 kfree(nc);
1222 }
1223 /*
1224 * In the previous loop, all the objects were freed to
1225 * the respective cache's slabs, now we can go ahead and
1226 * shrink each nodelist to its limit.
1227 */
1228 list_for_each_entry(cachep, &cache_chain, next) {
1229 l3 = cachep->nodelists[node];
1230 if (!l3)
1231 continue;
1232 drain_freelist(cachep, l3, l3->free_objects);
1233 }
1234}
1235
1236static int __cpuinit cpuup_prepare(long cpu)
1237{
1238 struct kmem_cache *cachep;
1239 struct kmem_list3 *l3 = NULL;
1240 int node = cpu_to_mem(cpu);
1241 int err;
1242
1243 /*
1244 * We need to do this right in the beginning since
1245 * alloc_arraycache's are going to use this list.
1246 * kmalloc_node allows us to add the slab to the right
1247 * kmem_list3 and not this cpu's kmem_list3
1248 */
1249 err = init_cache_nodelists_node(node);
1250 if (err < 0)
1251 goto bad;
1252
1253 /*
1254 * Now we can go ahead with allocating the shared arrays and
1255 * array caches
1256 */
1257 list_for_each_entry(cachep, &cache_chain, next) {
1258 struct array_cache *nc;
1259 struct array_cache *shared = NULL;
1260 struct array_cache **alien = NULL;
1261
1262 nc = alloc_arraycache(node, cachep->limit,
1263 cachep->batchcount, GFP_KERNEL);
1264 if (!nc)
1265 goto bad;
1266 if (cachep->shared) {
1267 shared = alloc_arraycache(node,
1268 cachep->shared * cachep->batchcount,
1269 0xbaadf00d, GFP_KERNEL);
1270 if (!shared) {
1271 kfree(nc);
1272 goto bad;
1273 }
1274 }
1275 if (use_alien_caches) {
1276 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1277 if (!alien) {
1278 kfree(shared);
1279 kfree(nc);
1280 goto bad;
1281 }
1282 }
1283 cachep->array[cpu] = nc;
1284 l3 = cachep->nodelists[node];
1285 BUG_ON(!l3);
1286
1287 spin_lock_irq(&l3->list_lock);
1288 if (!l3->shared) {
1289 /*
1290 * We are serialised from CPU_DEAD or
1291 * CPU_UP_CANCELLED by the cpucontrol lock
1292 */
1293 l3->shared = shared;
1294 shared = NULL;
1295 }
1296#ifdef CONFIG_NUMA
1297 if (!l3->alien) {
1298 l3->alien = alien;
1299 alien = NULL;
1300 }
1301#endif
1302 spin_unlock_irq(&l3->list_lock);
1303 kfree(shared);
1304 free_alien_cache(alien);
1305 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1306 slab_set_debugobj_lock_classes_node(cachep, node);
1307 }
1308 init_node_lock_keys(node);
1309
1310 return 0;
1311bad:
1312 cpuup_canceled(cpu);
1313 return -ENOMEM;
1314}
1315
1316static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1317 unsigned long action, void *hcpu)
1318{
1319 long cpu = (long)hcpu;
1320 int err = 0;
1321
1322 switch (action) {
1323 case CPU_UP_PREPARE:
1324 case CPU_UP_PREPARE_FROZEN:
1325 mutex_lock(&cache_chain_mutex);
1326 err = cpuup_prepare(cpu);
1327 mutex_unlock(&cache_chain_mutex);
1328 break;
1329 case CPU_ONLINE:
1330 case CPU_ONLINE_FROZEN:
1331 start_cpu_timer(cpu);
1332 break;
1333#ifdef CONFIG_HOTPLUG_CPU
1334 case CPU_DOWN_PREPARE:
1335 case CPU_DOWN_PREPARE_FROZEN:
1336 /*
1337 * Shutdown cache reaper. Note that the cache_chain_mutex is
1338 * held so that if cache_reap() is invoked it cannot do
1339 * anything expensive but will only modify reap_work
1340 * and reschedule the timer.
1341 */
1342 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1343 /* Now the cache_reaper is guaranteed to be not running. */
1344 per_cpu(slab_reap_work, cpu).work.func = NULL;
1345 break;
1346 case CPU_DOWN_FAILED:
1347 case CPU_DOWN_FAILED_FROZEN:
1348 start_cpu_timer(cpu);
1349 break;
1350 case CPU_DEAD:
1351 case CPU_DEAD_FROZEN:
1352 /*
1353 * Even if all the cpus of a node are down, we don't free the
1354 * kmem_list3 of any cache. This to avoid a race between
1355 * cpu_down, and a kmalloc allocation from another cpu for
1356 * memory from the node of the cpu going down. The list3
1357 * structure is usually allocated from kmem_cache_create() and
1358 * gets destroyed at kmem_cache_destroy().
1359 */
1360 /* fall through */
1361#endif
1362 case CPU_UP_CANCELED:
1363 case CPU_UP_CANCELED_FROZEN:
1364 mutex_lock(&cache_chain_mutex);
1365 cpuup_canceled(cpu);
1366 mutex_unlock(&cache_chain_mutex);
1367 break;
1368 }
1369 return notifier_from_errno(err);
1370}
1371
1372static struct notifier_block __cpuinitdata cpucache_notifier = {
1373 &cpuup_callback, NULL, 0
1374};
1375
1376#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1377/*
1378 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1379 * Returns -EBUSY if all objects cannot be drained so that the node is not
1380 * removed.
1381 *
1382 * Must hold cache_chain_mutex.
1383 */
1384static int __meminit drain_cache_nodelists_node(int node)
1385{
1386 struct kmem_cache *cachep;
1387 int ret = 0;
1388
1389 list_for_each_entry(cachep, &cache_chain, next) {
1390 struct kmem_list3 *l3;
1391
1392 l3 = cachep->nodelists[node];
1393 if (!l3)
1394 continue;
1395
1396 drain_freelist(cachep, l3, l3->free_objects);
1397
1398 if (!list_empty(&l3->slabs_full) ||
1399 !list_empty(&l3->slabs_partial)) {
1400 ret = -EBUSY;
1401 break;
1402 }
1403 }
1404 return ret;
1405}
1406
1407static int __meminit slab_memory_callback(struct notifier_block *self,
1408 unsigned long action, void *arg)
1409{
1410 struct memory_notify *mnb = arg;
1411 int ret = 0;
1412 int nid;
1413
1414 nid = mnb->status_change_nid;
1415 if (nid < 0)
1416 goto out;
1417
1418 switch (action) {
1419 case MEM_GOING_ONLINE:
1420 mutex_lock(&cache_chain_mutex);
1421 ret = init_cache_nodelists_node(nid);
1422 mutex_unlock(&cache_chain_mutex);
1423 break;
1424 case MEM_GOING_OFFLINE:
1425 mutex_lock(&cache_chain_mutex);
1426 ret = drain_cache_nodelists_node(nid);
1427 mutex_unlock(&cache_chain_mutex);
1428 break;
1429 case MEM_ONLINE:
1430 case MEM_OFFLINE:
1431 case MEM_CANCEL_ONLINE:
1432 case MEM_CANCEL_OFFLINE:
1433 break;
1434 }
1435out:
1436 return notifier_from_errno(ret);
1437}
1438#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1439
1440/*
1441 * swap the static kmem_list3 with kmalloced memory
1442 */
1443static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1444 int nodeid)
1445{
1446 struct kmem_list3 *ptr;
1447
1448 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1449 BUG_ON(!ptr);
1450
1451 memcpy(ptr, list, sizeof(struct kmem_list3));
1452 /*
1453 * Do not assume that spinlocks can be initialized via memcpy:
1454 */
1455 spin_lock_init(&ptr->list_lock);
1456
1457 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1458 cachep->nodelists[nodeid] = ptr;
1459}
1460
1461/*
1462 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1463 * size of kmem_list3.
1464 */
1465static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1466{
1467 int node;
1468
1469 for_each_online_node(node) {
1470 cachep->nodelists[node] = &initkmem_list3[index + node];
1471 cachep->nodelists[node]->next_reap = jiffies +
1472 REAPTIMEOUT_LIST3 +
1473 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1474 }
1475}
1476
1477/*
1478 * Initialisation. Called after the page allocator have been initialised and
1479 * before smp_init().
1480 */
1481void __init kmem_cache_init(void)
1482{
1483 size_t left_over;
1484 struct cache_sizes *sizes;
1485 struct cache_names *names;
1486 int i;
1487 int order;
1488 int node;
1489
1490 if (num_possible_nodes() == 1)
1491 use_alien_caches = 0;
1492
1493 for (i = 0; i < NUM_INIT_LISTS; i++) {
1494 kmem_list3_init(&initkmem_list3[i]);
1495 if (i < MAX_NUMNODES)
1496 cache_cache.nodelists[i] = NULL;
1497 }
1498 set_up_list3s(&cache_cache, CACHE_CACHE);
1499
1500 /*
1501 * Fragmentation resistance on low memory - only use bigger
1502 * page orders on machines with more than 32MB of memory.
1503 */
1504 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1505 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1506
1507 /* Bootstrap is tricky, because several objects are allocated
1508 * from caches that do not exist yet:
1509 * 1) initialize the cache_cache cache: it contains the struct
1510 * kmem_cache structures of all caches, except cache_cache itself:
1511 * cache_cache is statically allocated.
1512 * Initially an __init data area is used for the head array and the
1513 * kmem_list3 structures, it's replaced with a kmalloc allocated
1514 * array at the end of the bootstrap.
1515 * 2) Create the first kmalloc cache.
1516 * The struct kmem_cache for the new cache is allocated normally.
1517 * An __init data area is used for the head array.
1518 * 3) Create the remaining kmalloc caches, with minimally sized
1519 * head arrays.
1520 * 4) Replace the __init data head arrays for cache_cache and the first
1521 * kmalloc cache with kmalloc allocated arrays.
1522 * 5) Replace the __init data for kmem_list3 for cache_cache and
1523 * the other cache's with kmalloc allocated memory.
1524 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1525 */
1526
1527 node = numa_mem_id();
1528
1529 /* 1) create the cache_cache */
1530 INIT_LIST_HEAD(&cache_chain);
1531 list_add(&cache_cache.next, &cache_chain);
1532 cache_cache.colour_off = cache_line_size();
1533 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1534 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1535
1536 /*
1537 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1538 */
1539 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1540 nr_node_ids * sizeof(struct kmem_list3 *);
1541#if DEBUG
1542 cache_cache.obj_size = cache_cache.buffer_size;
1543#endif
1544 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1545 cache_line_size());
1546 cache_cache.reciprocal_buffer_size =
1547 reciprocal_value(cache_cache.buffer_size);
1548
1549 for (order = 0; order < MAX_ORDER; order++) {
1550 cache_estimate(order, cache_cache.buffer_size,
1551 cache_line_size(), 0, &left_over, &cache_cache.num);
1552 if (cache_cache.num)
1553 break;
1554 }
1555 BUG_ON(!cache_cache.num);
1556 cache_cache.gfporder = order;
1557 cache_cache.colour = left_over / cache_cache.colour_off;
1558 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1559 sizeof(struct slab), cache_line_size());
1560
1561 /* 2+3) create the kmalloc caches */
1562 sizes = malloc_sizes;
1563 names = cache_names;
1564
1565 /*
1566 * Initialize the caches that provide memory for the array cache and the
1567 * kmem_list3 structures first. Without this, further allocations will
1568 * bug.
1569 */
1570
1571 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1572 sizes[INDEX_AC].cs_size,
1573 ARCH_KMALLOC_MINALIGN,
1574 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1575 NULL);
1576
1577 if (INDEX_AC != INDEX_L3) {
1578 sizes[INDEX_L3].cs_cachep =
1579 kmem_cache_create(names[INDEX_L3].name,
1580 sizes[INDEX_L3].cs_size,
1581 ARCH_KMALLOC_MINALIGN,
1582 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1583 NULL);
1584 }
1585
1586 slab_early_init = 0;
1587
1588 while (sizes->cs_size != ULONG_MAX) {
1589 /*
1590 * For performance, all the general caches are L1 aligned.
1591 * This should be particularly beneficial on SMP boxes, as it
1592 * eliminates "false sharing".
1593 * Note for systems short on memory removing the alignment will
1594 * allow tighter packing of the smaller caches.
1595 */
1596 if (!sizes->cs_cachep) {
1597 sizes->cs_cachep = kmem_cache_create(names->name,
1598 sizes->cs_size,
1599 ARCH_KMALLOC_MINALIGN,
1600 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1601 NULL);
1602 }
1603#ifdef CONFIG_ZONE_DMA
1604 sizes->cs_dmacachep = kmem_cache_create(
1605 names->name_dma,
1606 sizes->cs_size,
1607 ARCH_KMALLOC_MINALIGN,
1608 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1609 SLAB_PANIC,
1610 NULL);
1611#endif
1612 sizes++;
1613 names++;
1614 }
1615 /* 4) Replace the bootstrap head arrays */
1616 {
1617 struct array_cache *ptr;
1618
1619 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1620
1621 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1622 memcpy(ptr, cpu_cache_get(&cache_cache),
1623 sizeof(struct arraycache_init));
1624 /*
1625 * Do not assume that spinlocks can be initialized via memcpy:
1626 */
1627 spin_lock_init(&ptr->lock);
1628
1629 cache_cache.array[smp_processor_id()] = ptr;
1630
1631 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1632
1633 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1634 != &initarray_generic.cache);
1635 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1636 sizeof(struct arraycache_init));
1637 /*
1638 * Do not assume that spinlocks can be initialized via memcpy:
1639 */
1640 spin_lock_init(&ptr->lock);
1641
1642 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1643 ptr;
1644 }
1645 /* 5) Replace the bootstrap kmem_list3's */
1646 {
1647 int nid;
1648
1649 for_each_online_node(nid) {
1650 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1651
1652 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1653 &initkmem_list3[SIZE_AC + nid], nid);
1654
1655 if (INDEX_AC != INDEX_L3) {
1656 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1657 &initkmem_list3[SIZE_L3 + nid], nid);
1658 }
1659 }
1660 }
1661
1662 g_cpucache_up = EARLY;
1663}
1664
1665void __init kmem_cache_init_late(void)
1666{
1667 struct kmem_cache *cachep;
1668
1669 /* Annotate slab for lockdep -- annotate the malloc caches */
1670 init_lock_keys();
1671
1672 /* 6) resize the head arrays to their final sizes */
1673 mutex_lock(&cache_chain_mutex);
1674 list_for_each_entry(cachep, &cache_chain, next)
1675 if (enable_cpucache(cachep, GFP_NOWAIT))
1676 BUG();
1677 mutex_unlock(&cache_chain_mutex);
1678
1679 /* Done! */
1680 g_cpucache_up = FULL;
1681
1682 /*
1683 * Register a cpu startup notifier callback that initializes
1684 * cpu_cache_get for all new cpus
1685 */
1686 register_cpu_notifier(&cpucache_notifier);
1687
1688#ifdef CONFIG_NUMA
1689 /*
1690 * Register a memory hotplug callback that initializes and frees
1691 * nodelists.
1692 */
1693 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1694#endif
1695
1696 /*
1697 * The reap timers are started later, with a module init call: That part
1698 * of the kernel is not yet operational.
1699 */
1700}
1701
1702static int __init cpucache_init(void)
1703{
1704 int cpu;
1705
1706 /*
1707 * Register the timers that return unneeded pages to the page allocator
1708 */
1709 for_each_online_cpu(cpu)
1710 start_cpu_timer(cpu);
1711 return 0;
1712}
1713__initcall(cpucache_init);
1714
1715/*
1716 * Interface to system's page allocator. No need to hold the cache-lock.
1717 *
1718 * If we requested dmaable memory, we will get it. Even if we
1719 * did not request dmaable memory, we might get it, but that
1720 * would be relatively rare and ignorable.
1721 */
1722static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1723{
1724 struct page *page;
1725 int nr_pages;
1726 int i;
1727
1728#ifndef CONFIG_MMU
1729 /*
1730 * Nommu uses slab's for process anonymous memory allocations, and thus
1731 * requires __GFP_COMP to properly refcount higher order allocations
1732 */
1733 flags |= __GFP_COMP;
1734#endif
1735
1736 flags |= cachep->gfpflags;
1737 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1738 flags |= __GFP_RECLAIMABLE;
1739
1740 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1741 if (!page)
1742 return NULL;
1743
1744 nr_pages = (1 << cachep->gfporder);
1745 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1746 add_zone_page_state(page_zone(page),
1747 NR_SLAB_RECLAIMABLE, nr_pages);
1748 else
1749 add_zone_page_state(page_zone(page),
1750 NR_SLAB_UNRECLAIMABLE, nr_pages);
1751 for (i = 0; i < nr_pages; i++)
1752 __SetPageSlab(page + i);
1753
1754 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1755 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1756
1757 if (cachep->ctor)
1758 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1759 else
1760 kmemcheck_mark_unallocated_pages(page, nr_pages);
1761 }
1762
1763 return page_address(page);
1764}
1765
1766/*
1767 * Interface to system's page release.
1768 */
1769static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1770{
1771 unsigned long i = (1 << cachep->gfporder);
1772 struct page *page = virt_to_page(addr);
1773 const unsigned long nr_freed = i;
1774
1775 kmemcheck_free_shadow(page, cachep->gfporder);
1776
1777 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1778 sub_zone_page_state(page_zone(page),
1779 NR_SLAB_RECLAIMABLE, nr_freed);
1780 else
1781 sub_zone_page_state(page_zone(page),
1782 NR_SLAB_UNRECLAIMABLE, nr_freed);
1783 while (i--) {
1784 BUG_ON(!PageSlab(page));
1785 __ClearPageSlab(page);
1786 page++;
1787 }
1788 if (current->reclaim_state)
1789 current->reclaim_state->reclaimed_slab += nr_freed;
1790 free_pages((unsigned long)addr, cachep->gfporder);
1791}
1792
1793static void kmem_rcu_free(struct rcu_head *head)
1794{
1795 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1796 struct kmem_cache *cachep = slab_rcu->cachep;
1797
1798 kmem_freepages(cachep, slab_rcu->addr);
1799 if (OFF_SLAB(cachep))
1800 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1801}
1802
1803#if DEBUG
1804
1805#ifdef CONFIG_DEBUG_PAGEALLOC
1806static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1807 unsigned long caller)
1808{
1809 int size = obj_size(cachep);
1810
1811 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1812
1813 if (size < 5 * sizeof(unsigned long))
1814 return;
1815
1816 *addr++ = 0x12345678;
1817 *addr++ = caller;
1818 *addr++ = smp_processor_id();
1819 size -= 3 * sizeof(unsigned long);
1820 {
1821 unsigned long *sptr = &caller;
1822 unsigned long svalue;
1823
1824 while (!kstack_end(sptr)) {
1825 svalue = *sptr++;
1826 if (kernel_text_address(svalue)) {
1827 *addr++ = svalue;
1828 size -= sizeof(unsigned long);
1829 if (size <= sizeof(unsigned long))
1830 break;
1831 }
1832 }
1833
1834 }
1835 *addr++ = 0x87654321;
1836}
1837#endif
1838
1839static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1840{
1841 int size = obj_size(cachep);
1842 addr = &((char *)addr)[obj_offset(cachep)];
1843
1844 memset(addr, val, size);
1845 *(unsigned char *)(addr + size - 1) = POISON_END;
1846}
1847
1848static void dump_line(char *data, int offset, int limit)
1849{
1850 int i;
1851 unsigned char error = 0;
1852 int bad_count = 0;
1853
1854 printk(KERN_ERR "%03x:", offset);
1855 for (i = 0; i < limit; i++) {
1856 if (data[offset + i] != POISON_FREE) {
1857 error = data[offset + i];
1858 bad_count++;
1859 }
1860 printk(" %02x", (unsigned char)data[offset + i]);
1861 }
1862 printk("\n");
1863
1864 if (bad_count == 1) {
1865 error ^= POISON_FREE;
1866 if (!(error & (error - 1))) {
1867 printk(KERN_ERR "Single bit error detected. Probably "
1868 "bad RAM.\n");
1869#ifdef CONFIG_X86
1870 printk(KERN_ERR "Run memtest86+ or a similar memory "
1871 "test tool.\n");
1872#else
1873 printk(KERN_ERR "Run a memory test tool.\n");
1874#endif
1875 }
1876 }
1877}
1878#endif
1879
1880#if DEBUG
1881
1882static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1883{
1884 int i, size;
1885 char *realobj;
1886
1887 if (cachep->flags & SLAB_RED_ZONE) {
1888 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1889 *dbg_redzone1(cachep, objp),
1890 *dbg_redzone2(cachep, objp));
1891 }
1892
1893 if (cachep->flags & SLAB_STORE_USER) {
1894 printk(KERN_ERR "Last user: [<%p>]",
1895 *dbg_userword(cachep, objp));
1896 print_symbol("(%s)",
1897 (unsigned long)*dbg_userword(cachep, objp));
1898 printk("\n");
1899 }
1900 realobj = (char *)objp + obj_offset(cachep);
1901 size = obj_size(cachep);
1902 for (i = 0; i < size && lines; i += 16, lines--) {
1903 int limit;
1904 limit = 16;
1905 if (i + limit > size)
1906 limit = size - i;
1907 dump_line(realobj, i, limit);
1908 }
1909}
1910
1911static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1912{
1913 char *realobj;
1914 int size, i;
1915 int lines = 0;
1916
1917 realobj = (char *)objp + obj_offset(cachep);
1918 size = obj_size(cachep);
1919
1920 for (i = 0; i < size; i++) {
1921 char exp = POISON_FREE;
1922 if (i == size - 1)
1923 exp = POISON_END;
1924 if (realobj[i] != exp) {
1925 int limit;
1926 /* Mismatch ! */
1927 /* Print header */
1928 if (lines == 0) {
1929 printk(KERN_ERR
1930 "Slab corruption: %s start=%p, len=%d\n",
1931 cachep->name, realobj, size);
1932 print_objinfo(cachep, objp, 0);
1933 }
1934 /* Hexdump the affected line */
1935 i = (i / 16) * 16;
1936 limit = 16;
1937 if (i + limit > size)
1938 limit = size - i;
1939 dump_line(realobj, i, limit);
1940 i += 16;
1941 lines++;
1942 /* Limit to 5 lines */
1943 if (lines > 5)
1944 break;
1945 }
1946 }
1947 if (lines != 0) {
1948 /* Print some data about the neighboring objects, if they
1949 * exist:
1950 */
1951 struct slab *slabp = virt_to_slab(objp);
1952 unsigned int objnr;
1953
1954 objnr = obj_to_index(cachep, slabp, objp);
1955 if (objnr) {
1956 objp = index_to_obj(cachep, slabp, objnr - 1);
1957 realobj = (char *)objp + obj_offset(cachep);
1958 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1959 realobj, size);
1960 print_objinfo(cachep, objp, 2);
1961 }
1962 if (objnr + 1 < cachep->num) {
1963 objp = index_to_obj(cachep, slabp, objnr + 1);
1964 realobj = (char *)objp + obj_offset(cachep);
1965 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1966 realobj, size);
1967 print_objinfo(cachep, objp, 2);
1968 }
1969 }
1970}
1971#endif
1972
1973#if DEBUG
1974static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1975{
1976 int i;
1977 for (i = 0; i < cachep->num; i++) {
1978 void *objp = index_to_obj(cachep, slabp, i);
1979
1980 if (cachep->flags & SLAB_POISON) {
1981#ifdef CONFIG_DEBUG_PAGEALLOC
1982 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1983 OFF_SLAB(cachep))
1984 kernel_map_pages(virt_to_page(objp),
1985 cachep->buffer_size / PAGE_SIZE, 1);
1986 else
1987 check_poison_obj(cachep, objp);
1988#else
1989 check_poison_obj(cachep, objp);
1990#endif
1991 }
1992 if (cachep->flags & SLAB_RED_ZONE) {
1993 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1994 slab_error(cachep, "start of a freed object "
1995 "was overwritten");
1996 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1997 slab_error(cachep, "end of a freed object "
1998 "was overwritten");
1999 }
2000 }
2001}
2002#else
2003static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2004{
2005}
2006#endif
2007
2008/**
2009 * slab_destroy - destroy and release all objects in a slab
2010 * @cachep: cache pointer being destroyed
2011 * @slabp: slab pointer being destroyed
2012 *
2013 * Destroy all the objs in a slab, and release the mem back to the system.
2014 * Before calling the slab must have been unlinked from the cache. The
2015 * cache-lock is not held/needed.
2016 */
2017static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2018{
2019 void *addr = slabp->s_mem - slabp->colouroff;
2020
2021 slab_destroy_debugcheck(cachep, slabp);
2022 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2023 struct slab_rcu *slab_rcu;
2024
2025 slab_rcu = (struct slab_rcu *)slabp;
2026 slab_rcu->cachep = cachep;
2027 slab_rcu->addr = addr;
2028 call_rcu(&slab_rcu->head, kmem_rcu_free);
2029 } else {
2030 kmem_freepages(cachep, addr);
2031 if (OFF_SLAB(cachep))
2032 kmem_cache_free(cachep->slabp_cache, slabp);
2033 }
2034}
2035
2036static void __kmem_cache_destroy(struct kmem_cache *cachep)
2037{
2038 int i;
2039 struct kmem_list3 *l3;
2040
2041 for_each_online_cpu(i)
2042 kfree(cachep->array[i]);
2043
2044 /* NUMA: free the list3 structures */
2045 for_each_online_node(i) {
2046 l3 = cachep->nodelists[i];
2047 if (l3) {
2048 kfree(l3->shared);
2049 free_alien_cache(l3->alien);
2050 kfree(l3);
2051 }
2052 }
2053 kmem_cache_free(&cache_cache, cachep);
2054}
2055
2056
2057/**
2058 * calculate_slab_order - calculate size (page order) of slabs
2059 * @cachep: pointer to the cache that is being created
2060 * @size: size of objects to be created in this cache.
2061 * @align: required alignment for the objects.
2062 * @flags: slab allocation flags
2063 *
2064 * Also calculates the number of objects per slab.
2065 *
2066 * This could be made much more intelligent. For now, try to avoid using
2067 * high order pages for slabs. When the gfp() functions are more friendly
2068 * towards high-order requests, this should be changed.
2069 */
2070static size_t calculate_slab_order(struct kmem_cache *cachep,
2071 size_t size, size_t align, unsigned long flags)
2072{
2073 unsigned long offslab_limit;
2074 size_t left_over = 0;
2075 int gfporder;
2076
2077 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2078 unsigned int num;
2079 size_t remainder;
2080
2081 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2082 if (!num)
2083 continue;
2084
2085 if (flags & CFLGS_OFF_SLAB) {
2086 /*
2087 * Max number of objs-per-slab for caches which
2088 * use off-slab slabs. Needed to avoid a possible
2089 * looping condition in cache_grow().
2090 */
2091 offslab_limit = size - sizeof(struct slab);
2092 offslab_limit /= sizeof(kmem_bufctl_t);
2093
2094 if (num > offslab_limit)
2095 break;
2096 }
2097
2098 /* Found something acceptable - save it away */
2099 cachep->num = num;
2100 cachep->gfporder = gfporder;
2101 left_over = remainder;
2102
2103 /*
2104 * A VFS-reclaimable slab tends to have most allocations
2105 * as GFP_NOFS and we really don't want to have to be allocating
2106 * higher-order pages when we are unable to shrink dcache.
2107 */
2108 if (flags & SLAB_RECLAIM_ACCOUNT)
2109 break;
2110
2111 /*
2112 * Large number of objects is good, but very large slabs are
2113 * currently bad for the gfp()s.
2114 */
2115 if (gfporder >= slab_break_gfp_order)
2116 break;
2117
2118 /*
2119 * Acceptable internal fragmentation?
2120 */
2121 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2122 break;
2123 }
2124 return left_over;
2125}
2126
2127static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2128{
2129 if (g_cpucache_up == FULL)
2130 return enable_cpucache(cachep, gfp);
2131
2132 if (g_cpucache_up == NONE) {
2133 /*
2134 * Note: the first kmem_cache_create must create the cache
2135 * that's used by kmalloc(24), otherwise the creation of
2136 * further caches will BUG().
2137 */
2138 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2139
2140 /*
2141 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2142 * the first cache, then we need to set up all its list3s,
2143 * otherwise the creation of further caches will BUG().
2144 */
2145 set_up_list3s(cachep, SIZE_AC);
2146 if (INDEX_AC == INDEX_L3)
2147 g_cpucache_up = PARTIAL_L3;
2148 else
2149 g_cpucache_up = PARTIAL_AC;
2150 } else {
2151 cachep->array[smp_processor_id()] =
2152 kmalloc(sizeof(struct arraycache_init), gfp);
2153
2154 if (g_cpucache_up == PARTIAL_AC) {
2155 set_up_list3s(cachep, SIZE_L3);
2156 g_cpucache_up = PARTIAL_L3;
2157 } else {
2158 int node;
2159 for_each_online_node(node) {
2160 cachep->nodelists[node] =
2161 kmalloc_node(sizeof(struct kmem_list3),
2162 gfp, node);
2163 BUG_ON(!cachep->nodelists[node]);
2164 kmem_list3_init(cachep->nodelists[node]);
2165 }
2166 }
2167 }
2168 cachep->nodelists[numa_mem_id()]->next_reap =
2169 jiffies + REAPTIMEOUT_LIST3 +
2170 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2171
2172 cpu_cache_get(cachep)->avail = 0;
2173 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2174 cpu_cache_get(cachep)->batchcount = 1;
2175 cpu_cache_get(cachep)->touched = 0;
2176 cachep->batchcount = 1;
2177 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2178 return 0;
2179}
2180
2181/**
2182 * kmem_cache_create - Create a cache.
2183 * @name: A string which is used in /proc/slabinfo to identify this cache.
2184 * @size: The size of objects to be created in this cache.
2185 * @align: The required alignment for the objects.
2186 * @flags: SLAB flags
2187 * @ctor: A constructor for the objects.
2188 *
2189 * Returns a ptr to the cache on success, NULL on failure.
2190 * Cannot be called within a int, but can be interrupted.
2191 * The @ctor is run when new pages are allocated by the cache.
2192 *
2193 * @name must be valid until the cache is destroyed. This implies that
2194 * the module calling this has to destroy the cache before getting unloaded.
2195 *
2196 * The flags are
2197 *
2198 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2199 * to catch references to uninitialised memory.
2200 *
2201 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2202 * for buffer overruns.
2203 *
2204 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2205 * cacheline. This can be beneficial if you're counting cycles as closely
2206 * as davem.
2207 */
2208struct kmem_cache *
2209kmem_cache_create (const char *name, size_t size, size_t align,
2210 unsigned long flags, void (*ctor)(void *))
2211{
2212 size_t left_over, slab_size, ralign;
2213 struct kmem_cache *cachep = NULL, *pc;
2214 gfp_t gfp;
2215
2216 /*
2217 * Sanity checks... these are all serious usage bugs.
2218 */
2219 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2220 size > KMALLOC_MAX_SIZE) {
2221 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2222 name);
2223 BUG();
2224 }
2225
2226 /*
2227 * We use cache_chain_mutex to ensure a consistent view of
2228 * cpu_online_mask as well. Please see cpuup_callback
2229 */
2230 if (slab_is_available()) {
2231 get_online_cpus();
2232 mutex_lock(&cache_chain_mutex);
2233 }
2234
2235 list_for_each_entry(pc, &cache_chain, next) {
2236 char tmp;
2237 int res;
2238
2239 /*
2240 * This happens when the module gets unloaded and doesn't
2241 * destroy its slab cache and no-one else reuses the vmalloc
2242 * area of the module. Print a warning.
2243 */
2244 res = probe_kernel_address(pc->name, tmp);
2245 if (res) {
2246 printk(KERN_ERR
2247 "SLAB: cache with size %d has lost its name\n",
2248 pc->buffer_size);
2249 continue;
2250 }
2251
2252 if (!strcmp(pc->name, name)) {
2253 printk(KERN_ERR
2254 "kmem_cache_create: duplicate cache %s\n", name);
2255 dump_stack();
2256 goto oops;
2257 }
2258 }
2259
2260#if DEBUG
2261 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2262#if FORCED_DEBUG
2263 /*
2264 * Enable redzoning and last user accounting, except for caches with
2265 * large objects, if the increased size would increase the object size
2266 * above the next power of two: caches with object sizes just above a
2267 * power of two have a significant amount of internal fragmentation.
2268 */
2269 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2270 2 * sizeof(unsigned long long)))
2271 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2272 if (!(flags & SLAB_DESTROY_BY_RCU))
2273 flags |= SLAB_POISON;
2274#endif
2275 if (flags & SLAB_DESTROY_BY_RCU)
2276 BUG_ON(flags & SLAB_POISON);
2277#endif
2278 /*
2279 * Always checks flags, a caller might be expecting debug support which
2280 * isn't available.
2281 */
2282 BUG_ON(flags & ~CREATE_MASK);
2283
2284 /*
2285 * Check that size is in terms of words. This is needed to avoid
2286 * unaligned accesses for some archs when redzoning is used, and makes
2287 * sure any on-slab bufctl's are also correctly aligned.
2288 */
2289 if (size & (BYTES_PER_WORD - 1)) {
2290 size += (BYTES_PER_WORD - 1);
2291 size &= ~(BYTES_PER_WORD - 1);
2292 }
2293
2294 /* calculate the final buffer alignment: */
2295
2296 /* 1) arch recommendation: can be overridden for debug */
2297 if (flags & SLAB_HWCACHE_ALIGN) {
2298 /*
2299 * Default alignment: as specified by the arch code. Except if
2300 * an object is really small, then squeeze multiple objects into
2301 * one cacheline.
2302 */
2303 ralign = cache_line_size();
2304 while (size <= ralign / 2)
2305 ralign /= 2;
2306 } else {
2307 ralign = BYTES_PER_WORD;
2308 }
2309
2310 /*
2311 * Redzoning and user store require word alignment or possibly larger.
2312 * Note this will be overridden by architecture or caller mandated
2313 * alignment if either is greater than BYTES_PER_WORD.
2314 */
2315 if (flags & SLAB_STORE_USER)
2316 ralign = BYTES_PER_WORD;
2317
2318 if (flags & SLAB_RED_ZONE) {
2319 ralign = REDZONE_ALIGN;
2320 /* If redzoning, ensure that the second redzone is suitably
2321 * aligned, by adjusting the object size accordingly. */
2322 size += REDZONE_ALIGN - 1;
2323 size &= ~(REDZONE_ALIGN - 1);
2324 }
2325
2326 /* 2) arch mandated alignment */
2327 if (ralign < ARCH_SLAB_MINALIGN) {
2328 ralign = ARCH_SLAB_MINALIGN;
2329 }
2330 /* 3) caller mandated alignment */
2331 if (ralign < align) {
2332 ralign = align;
2333 }
2334 /* disable debug if necessary */
2335 if (ralign > __alignof__(unsigned long long))
2336 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2337 /*
2338 * 4) Store it.
2339 */
2340 align = ralign;
2341
2342 if (slab_is_available())
2343 gfp = GFP_KERNEL;
2344 else
2345 gfp = GFP_NOWAIT;
2346
2347 /* Get cache's description obj. */
2348 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2349 if (!cachep)
2350 goto oops;
2351
2352 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2353#if DEBUG
2354 cachep->obj_size = size;
2355
2356 /*
2357 * Both debugging options require word-alignment which is calculated
2358 * into align above.
2359 */
2360 if (flags & SLAB_RED_ZONE) {
2361 /* add space for red zone words */
2362 cachep->obj_offset += sizeof(unsigned long long);
2363 size += 2 * sizeof(unsigned long long);
2364 }
2365 if (flags & SLAB_STORE_USER) {
2366 /* user store requires one word storage behind the end of
2367 * the real object. But if the second red zone needs to be
2368 * aligned to 64 bits, we must allow that much space.
2369 */
2370 if (flags & SLAB_RED_ZONE)
2371 size += REDZONE_ALIGN;
2372 else
2373 size += BYTES_PER_WORD;
2374 }
2375#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2376 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2377 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2378 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2379 size = PAGE_SIZE;
2380 }
2381#endif
2382#endif
2383
2384 /*
2385 * Determine if the slab management is 'on' or 'off' slab.
2386 * (bootstrapping cannot cope with offslab caches so don't do
2387 * it too early on. Always use on-slab management when
2388 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2389 */
2390 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2391 !(flags & SLAB_NOLEAKTRACE))
2392 /*
2393 * Size is large, assume best to place the slab management obj
2394 * off-slab (should allow better packing of objs).
2395 */
2396 flags |= CFLGS_OFF_SLAB;
2397
2398 size = ALIGN(size, align);
2399
2400 left_over = calculate_slab_order(cachep, size, align, flags);
2401
2402 if (!cachep->num) {
2403 printk(KERN_ERR
2404 "kmem_cache_create: couldn't create cache %s.\n", name);
2405 kmem_cache_free(&cache_cache, cachep);
2406 cachep = NULL;
2407 goto oops;
2408 }
2409 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2410 + sizeof(struct slab), align);
2411
2412 /*
2413 * If the slab has been placed off-slab, and we have enough space then
2414 * move it on-slab. This is at the expense of any extra colouring.
2415 */
2416 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2417 flags &= ~CFLGS_OFF_SLAB;
2418 left_over -= slab_size;
2419 }
2420
2421 if (flags & CFLGS_OFF_SLAB) {
2422 /* really off slab. No need for manual alignment */
2423 slab_size =
2424 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2425
2426#ifdef CONFIG_PAGE_POISONING
2427 /* If we're going to use the generic kernel_map_pages()
2428 * poisoning, then it's going to smash the contents of
2429 * the redzone and userword anyhow, so switch them off.
2430 */
2431 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2432 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2433#endif
2434 }
2435
2436 cachep->colour_off = cache_line_size();
2437 /* Offset must be a multiple of the alignment. */
2438 if (cachep->colour_off < align)
2439 cachep->colour_off = align;
2440 cachep->colour = left_over / cachep->colour_off;
2441 cachep->slab_size = slab_size;
2442 cachep->flags = flags;
2443 cachep->gfpflags = 0;
2444 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2445 cachep->gfpflags |= GFP_DMA;
2446 cachep->buffer_size = size;
2447 cachep->reciprocal_buffer_size = reciprocal_value(size);
2448
2449 if (flags & CFLGS_OFF_SLAB) {
2450 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2451 /*
2452 * This is a possibility for one of the malloc_sizes caches.
2453 * But since we go off slab only for object size greater than
2454 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2455 * this should not happen at all.
2456 * But leave a BUG_ON for some lucky dude.
2457 */
2458 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2459 }
2460 cachep->ctor = ctor;
2461 cachep->name = name;
2462
2463 if (setup_cpu_cache(cachep, gfp)) {
2464 __kmem_cache_destroy(cachep);
2465 cachep = NULL;
2466 goto oops;
2467 }
2468
2469 if (flags & SLAB_DEBUG_OBJECTS) {
2470 /*
2471 * Would deadlock through slab_destroy()->call_rcu()->
2472 * debug_object_activate()->kmem_cache_alloc().
2473 */
2474 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2475
2476 slab_set_debugobj_lock_classes(cachep);
2477 }
2478
2479 /* cache setup completed, link it into the list */
2480 list_add(&cachep->next, &cache_chain);
2481oops:
2482 if (!cachep && (flags & SLAB_PANIC))
2483 panic("kmem_cache_create(): failed to create slab `%s'\n",
2484 name);
2485 if (slab_is_available()) {
2486 mutex_unlock(&cache_chain_mutex);
2487 put_online_cpus();
2488 }
2489 return cachep;
2490}
2491EXPORT_SYMBOL(kmem_cache_create);
2492
2493#if DEBUG
2494static void check_irq_off(void)
2495{
2496 BUG_ON(!irqs_disabled());
2497}
2498
2499static void check_irq_on(void)
2500{
2501 BUG_ON(irqs_disabled());
2502}
2503
2504static void check_spinlock_acquired(struct kmem_cache *cachep)
2505{
2506#ifdef CONFIG_SMP
2507 check_irq_off();
2508 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2509#endif
2510}
2511
2512static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2513{
2514#ifdef CONFIG_SMP
2515 check_irq_off();
2516 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2517#endif
2518}
2519
2520#else
2521#define check_irq_off() do { } while(0)
2522#define check_irq_on() do { } while(0)
2523#define check_spinlock_acquired(x) do { } while(0)
2524#define check_spinlock_acquired_node(x, y) do { } while(0)
2525#endif
2526
2527static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2528 struct array_cache *ac,
2529 int force, int node);
2530
2531static void do_drain(void *arg)
2532{
2533 struct kmem_cache *cachep = arg;
2534 struct array_cache *ac;
2535 int node = numa_mem_id();
2536
2537 check_irq_off();
2538 ac = cpu_cache_get(cachep);
2539 spin_lock(&cachep->nodelists[node]->list_lock);
2540 free_block(cachep, ac->entry, ac->avail, node);
2541 spin_unlock(&cachep->nodelists[node]->list_lock);
2542 ac->avail = 0;
2543}
2544
2545static void drain_cpu_caches(struct kmem_cache *cachep)
2546{
2547 struct kmem_list3 *l3;
2548 int node;
2549
2550 on_each_cpu(do_drain, cachep, 1);
2551 check_irq_on();
2552 for_each_online_node(node) {
2553 l3 = cachep->nodelists[node];
2554 if (l3 && l3->alien)
2555 drain_alien_cache(cachep, l3->alien);
2556 }
2557
2558 for_each_online_node(node) {
2559 l3 = cachep->nodelists[node];
2560 if (l3)
2561 drain_array(cachep, l3, l3->shared, 1, node);
2562 }
2563}
2564
2565/*
2566 * Remove slabs from the list of free slabs.
2567 * Specify the number of slabs to drain in tofree.
2568 *
2569 * Returns the actual number of slabs released.
2570 */
2571static int drain_freelist(struct kmem_cache *cache,
2572 struct kmem_list3 *l3, int tofree)
2573{
2574 struct list_head *p;
2575 int nr_freed;
2576 struct slab *slabp;
2577
2578 nr_freed = 0;
2579 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2580
2581 spin_lock_irq(&l3->list_lock);
2582 p = l3->slabs_free.prev;
2583 if (p == &l3->slabs_free) {
2584 spin_unlock_irq(&l3->list_lock);
2585 goto out;
2586 }
2587
2588 slabp = list_entry(p, struct slab, list);
2589#if DEBUG
2590 BUG_ON(slabp->inuse);
2591#endif
2592 list_del(&slabp->list);
2593 /*
2594 * Safe to drop the lock. The slab is no longer linked
2595 * to the cache.
2596 */
2597 l3->free_objects -= cache->num;
2598 spin_unlock_irq(&l3->list_lock);
2599 slab_destroy(cache, slabp);
2600 nr_freed++;
2601 }
2602out:
2603 return nr_freed;
2604}
2605
2606/* Called with cache_chain_mutex held to protect against cpu hotplug */
2607static int __cache_shrink(struct kmem_cache *cachep)
2608{
2609 int ret = 0, i = 0;
2610 struct kmem_list3 *l3;
2611
2612 drain_cpu_caches(cachep);
2613
2614 check_irq_on();
2615 for_each_online_node(i) {
2616 l3 = cachep->nodelists[i];
2617 if (!l3)
2618 continue;
2619
2620 drain_freelist(cachep, l3, l3->free_objects);
2621
2622 ret += !list_empty(&l3->slabs_full) ||
2623 !list_empty(&l3->slabs_partial);
2624 }
2625 return (ret ? 1 : 0);
2626}
2627
2628/**
2629 * kmem_cache_shrink - Shrink a cache.
2630 * @cachep: The cache to shrink.
2631 *
2632 * Releases as many slabs as possible for a cache.
2633 * To help debugging, a zero exit status indicates all slabs were released.
2634 */
2635int kmem_cache_shrink(struct kmem_cache *cachep)
2636{
2637 int ret;
2638 BUG_ON(!cachep || in_interrupt());
2639
2640 get_online_cpus();
2641 mutex_lock(&cache_chain_mutex);
2642 ret = __cache_shrink(cachep);
2643 mutex_unlock(&cache_chain_mutex);
2644 put_online_cpus();
2645 return ret;
2646}
2647EXPORT_SYMBOL(kmem_cache_shrink);
2648
2649/**
2650 * kmem_cache_destroy - delete a cache
2651 * @cachep: the cache to destroy
2652 *
2653 * Remove a &struct kmem_cache object from the slab cache.
2654 *
2655 * It is expected this function will be called by a module when it is
2656 * unloaded. This will remove the cache completely, and avoid a duplicate
2657 * cache being allocated each time a module is loaded and unloaded, if the
2658 * module doesn't have persistent in-kernel storage across loads and unloads.
2659 *
2660 * The cache must be empty before calling this function.
2661 *
2662 * The caller must guarantee that no one will allocate memory from the cache
2663 * during the kmem_cache_destroy().
2664 */
2665void kmem_cache_destroy(struct kmem_cache *cachep)
2666{
2667 BUG_ON(!cachep || in_interrupt());
2668
2669 /* Find the cache in the chain of caches. */
2670 get_online_cpus();
2671 mutex_lock(&cache_chain_mutex);
2672 /*
2673 * the chain is never empty, cache_cache is never destroyed
2674 */
2675 list_del(&cachep->next);
2676 if (__cache_shrink(cachep)) {
2677 slab_error(cachep, "Can't free all objects");
2678 list_add(&cachep->next, &cache_chain);
2679 mutex_unlock(&cache_chain_mutex);
2680 put_online_cpus();
2681 return;
2682 }
2683
2684 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2685 rcu_barrier();
2686
2687 __kmem_cache_destroy(cachep);
2688 mutex_unlock(&cache_chain_mutex);
2689 put_online_cpus();
2690}
2691EXPORT_SYMBOL(kmem_cache_destroy);
2692
2693/*
2694 * Get the memory for a slab management obj.
2695 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2696 * always come from malloc_sizes caches. The slab descriptor cannot
2697 * come from the same cache which is getting created because,
2698 * when we are searching for an appropriate cache for these
2699 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2700 * If we are creating a malloc_sizes cache here it would not be visible to
2701 * kmem_find_general_cachep till the initialization is complete.
2702 * Hence we cannot have slabp_cache same as the original cache.
2703 */
2704static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2705 int colour_off, gfp_t local_flags,
2706 int nodeid)
2707{
2708 struct slab *slabp;
2709
2710 if (OFF_SLAB(cachep)) {
2711 /* Slab management obj is off-slab. */
2712 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2713 local_flags, nodeid);
2714 /*
2715 * If the first object in the slab is leaked (it's allocated
2716 * but no one has a reference to it), we want to make sure
2717 * kmemleak does not treat the ->s_mem pointer as a reference
2718 * to the object. Otherwise we will not report the leak.
2719 */
2720 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2721 local_flags);
2722 if (!slabp)
2723 return NULL;
2724 } else {
2725 slabp = objp + colour_off;
2726 colour_off += cachep->slab_size;
2727 }
2728 slabp->inuse = 0;
2729 slabp->colouroff = colour_off;
2730 slabp->s_mem = objp + colour_off;
2731 slabp->nodeid = nodeid;
2732 slabp->free = 0;
2733 return slabp;
2734}
2735
2736static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2737{
2738 return (kmem_bufctl_t *) (slabp + 1);
2739}
2740
2741static void cache_init_objs(struct kmem_cache *cachep,
2742 struct slab *slabp)
2743{
2744 int i;
2745
2746 for (i = 0; i < cachep->num; i++) {
2747 void *objp = index_to_obj(cachep, slabp, i);
2748#if DEBUG
2749 /* need to poison the objs? */
2750 if (cachep->flags & SLAB_POISON)
2751 poison_obj(cachep, objp, POISON_FREE);
2752 if (cachep->flags & SLAB_STORE_USER)
2753 *dbg_userword(cachep, objp) = NULL;
2754
2755 if (cachep->flags & SLAB_RED_ZONE) {
2756 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2757 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2758 }
2759 /*
2760 * Constructors are not allowed to allocate memory from the same
2761 * cache which they are a constructor for. Otherwise, deadlock.
2762 * They must also be threaded.
2763 */
2764 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2765 cachep->ctor(objp + obj_offset(cachep));
2766
2767 if (cachep->flags & SLAB_RED_ZONE) {
2768 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2769 slab_error(cachep, "constructor overwrote the"
2770 " end of an object");
2771 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2772 slab_error(cachep, "constructor overwrote the"
2773 " start of an object");
2774 }
2775 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2776 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2777 kernel_map_pages(virt_to_page(objp),
2778 cachep->buffer_size / PAGE_SIZE, 0);
2779#else
2780 if (cachep->ctor)
2781 cachep->ctor(objp);
2782#endif
2783 slab_bufctl(slabp)[i] = i + 1;
2784 }
2785 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2786}
2787
2788static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2789{
2790 if (CONFIG_ZONE_DMA_FLAG) {
2791 if (flags & GFP_DMA)
2792 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2793 else
2794 BUG_ON(cachep->gfpflags & GFP_DMA);
2795 }
2796}
2797
2798static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2799 int nodeid)
2800{
2801 void *objp = index_to_obj(cachep, slabp, slabp->free);
2802 kmem_bufctl_t next;
2803
2804 slabp->inuse++;
2805 next = slab_bufctl(slabp)[slabp->free];
2806#if DEBUG
2807 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2808 WARN_ON(slabp->nodeid != nodeid);
2809#endif
2810 slabp->free = next;
2811
2812 return objp;
2813}
2814
2815static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2816 void *objp, int nodeid)
2817{
2818 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2819
2820#if DEBUG
2821 /* Verify that the slab belongs to the intended node */
2822 WARN_ON(slabp->nodeid != nodeid);
2823
2824 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2825 printk(KERN_ERR "slab: double free detected in cache "
2826 "'%s', objp %p\n", cachep->name, objp);
2827 BUG();
2828 }
2829#endif
2830 slab_bufctl(slabp)[objnr] = slabp->free;
2831 slabp->free = objnr;
2832 slabp->inuse--;
2833}
2834
2835/*
2836 * Map pages beginning at addr to the given cache and slab. This is required
2837 * for the slab allocator to be able to lookup the cache and slab of a
2838 * virtual address for kfree, ksize, and slab debugging.
2839 */
2840static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2841 void *addr)
2842{
2843 int nr_pages;
2844 struct page *page;
2845
2846 page = virt_to_page(addr);
2847
2848 nr_pages = 1;
2849 if (likely(!PageCompound(page)))
2850 nr_pages <<= cache->gfporder;
2851
2852 do {
2853 page_set_cache(page, cache);
2854 page_set_slab(page, slab);
2855 page++;
2856 } while (--nr_pages);
2857}
2858
2859/*
2860 * Grow (by 1) the number of slabs within a cache. This is called by
2861 * kmem_cache_alloc() when there are no active objs left in a cache.
2862 */
2863static int cache_grow(struct kmem_cache *cachep,
2864 gfp_t flags, int nodeid, void *objp)
2865{
2866 struct slab *slabp;
2867 size_t offset;
2868 gfp_t local_flags;
2869 struct kmem_list3 *l3;
2870
2871 /*
2872 * Be lazy and only check for valid flags here, keeping it out of the
2873 * critical path in kmem_cache_alloc().
2874 */
2875 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2876 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2877
2878 /* Take the l3 list lock to change the colour_next on this node */
2879 check_irq_off();
2880 l3 = cachep->nodelists[nodeid];
2881 spin_lock(&l3->list_lock);
2882
2883 /* Get colour for the slab, and cal the next value. */
2884 offset = l3->colour_next;
2885 l3->colour_next++;
2886 if (l3->colour_next >= cachep->colour)
2887 l3->colour_next = 0;
2888 spin_unlock(&l3->list_lock);
2889
2890 offset *= cachep->colour_off;
2891
2892 if (local_flags & __GFP_WAIT)
2893 local_irq_enable();
2894
2895 /*
2896 * The test for missing atomic flag is performed here, rather than
2897 * the more obvious place, simply to reduce the critical path length
2898 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2899 * will eventually be caught here (where it matters).
2900 */
2901 kmem_flagcheck(cachep, flags);
2902
2903 /*
2904 * Get mem for the objs. Attempt to allocate a physical page from
2905 * 'nodeid'.
2906 */
2907 if (!objp)
2908 objp = kmem_getpages(cachep, local_flags, nodeid);
2909 if (!objp)
2910 goto failed;
2911
2912 /* Get slab management. */
2913 slabp = alloc_slabmgmt(cachep, objp, offset,
2914 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2915 if (!slabp)
2916 goto opps1;
2917
2918 slab_map_pages(cachep, slabp, objp);
2919
2920 cache_init_objs(cachep, slabp);
2921
2922 if (local_flags & __GFP_WAIT)
2923 local_irq_disable();
2924 check_irq_off();
2925 spin_lock(&l3->list_lock);
2926
2927 /* Make slab active. */
2928 list_add_tail(&slabp->list, &(l3->slabs_free));
2929 STATS_INC_GROWN(cachep);
2930 l3->free_objects += cachep->num;
2931 spin_unlock(&l3->list_lock);
2932 return 1;
2933opps1:
2934 kmem_freepages(cachep, objp);
2935failed:
2936 if (local_flags & __GFP_WAIT)
2937 local_irq_disable();
2938 return 0;
2939}
2940
2941#if DEBUG
2942
2943/*
2944 * Perform extra freeing checks:
2945 * - detect bad pointers.
2946 * - POISON/RED_ZONE checking
2947 */
2948static void kfree_debugcheck(const void *objp)
2949{
2950 if (!virt_addr_valid(objp)) {
2951 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2952 (unsigned long)objp);
2953 BUG();
2954 }
2955}
2956
2957static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2958{
2959 unsigned long long redzone1, redzone2;
2960
2961 redzone1 = *dbg_redzone1(cache, obj);
2962 redzone2 = *dbg_redzone2(cache, obj);
2963
2964 /*
2965 * Redzone is ok.
2966 */
2967 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2968 return;
2969
2970 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2971 slab_error(cache, "double free detected");
2972 else
2973 slab_error(cache, "memory outside object was overwritten");
2974
2975 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2976 obj, redzone1, redzone2);
2977}
2978
2979static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2980 void *caller)
2981{
2982 struct page *page;
2983 unsigned int objnr;
2984 struct slab *slabp;
2985
2986 BUG_ON(virt_to_cache(objp) != cachep);
2987
2988 objp -= obj_offset(cachep);
2989 kfree_debugcheck(objp);
2990 page = virt_to_head_page(objp);
2991
2992 slabp = page_get_slab(page);
2993
2994 if (cachep->flags & SLAB_RED_ZONE) {
2995 verify_redzone_free(cachep, objp);
2996 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2997 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2998 }
2999 if (cachep->flags & SLAB_STORE_USER)
3000 *dbg_userword(cachep, objp) = caller;
3001
3002 objnr = obj_to_index(cachep, slabp, objp);
3003
3004 BUG_ON(objnr >= cachep->num);
3005 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3006
3007#ifdef CONFIG_DEBUG_SLAB_LEAK
3008 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3009#endif
3010 if (cachep->flags & SLAB_POISON) {
3011#ifdef CONFIG_DEBUG_PAGEALLOC
3012 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3013 store_stackinfo(cachep, objp, (unsigned long)caller);
3014 kernel_map_pages(virt_to_page(objp),
3015 cachep->buffer_size / PAGE_SIZE, 0);
3016 } else {
3017 poison_obj(cachep, objp, POISON_FREE);
3018 }
3019#else
3020 poison_obj(cachep, objp, POISON_FREE);
3021#endif
3022 }
3023 return objp;
3024}
3025
3026static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3027{
3028 kmem_bufctl_t i;
3029 int entries = 0;
3030
3031 /* Check slab's freelist to see if this obj is there. */
3032 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3033 entries++;
3034 if (entries > cachep->num || i >= cachep->num)
3035 goto bad;
3036 }
3037 if (entries != cachep->num - slabp->inuse) {
3038bad:
3039 printk(KERN_ERR "slab: Internal list corruption detected in "
3040 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
3041 cachep->name, cachep->num, slabp, slabp->inuse);
3042 for (i = 0;
3043 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
3044 i++) {
3045 if (i % 16 == 0)
3046 printk("\n%03x:", i);
3047 printk(" %02x", ((unsigned char *)slabp)[i]);
3048 }
3049 printk("\n");
3050 BUG();
3051 }
3052}
3053#else
3054#define kfree_debugcheck(x) do { } while(0)
3055#define cache_free_debugcheck(x,objp,z) (objp)
3056#define check_slabp(x,y) do { } while(0)
3057#endif
3058
3059static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3060{
3061 int batchcount;
3062 struct kmem_list3 *l3;
3063 struct array_cache *ac;
3064 int node;
3065
3066retry:
3067 check_irq_off();
3068 node = numa_mem_id();
3069 ac = cpu_cache_get(cachep);
3070 batchcount = ac->batchcount;
3071 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3072 /*
3073 * If there was little recent activity on this cache, then
3074 * perform only a partial refill. Otherwise we could generate
3075 * refill bouncing.
3076 */
3077 batchcount = BATCHREFILL_LIMIT;
3078 }
3079 l3 = cachep->nodelists[node];
3080
3081 BUG_ON(ac->avail > 0 || !l3);
3082 spin_lock(&l3->list_lock);
3083
3084 /* See if we can refill from the shared array */
3085 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3086 l3->shared->touched = 1;
3087 goto alloc_done;
3088 }
3089
3090 while (batchcount > 0) {
3091 struct list_head *entry;
3092 struct slab *slabp;
3093 /* Get slab alloc is to come from. */
3094 entry = l3->slabs_partial.next;
3095 if (entry == &l3->slabs_partial) {
3096 l3->free_touched = 1;
3097 entry = l3->slabs_free.next;
3098 if (entry == &l3->slabs_free)
3099 goto must_grow;
3100 }
3101
3102 slabp = list_entry(entry, struct slab, list);
3103 check_slabp(cachep, slabp);
3104 check_spinlock_acquired(cachep);
3105
3106 /*
3107 * The slab was either on partial or free list so
3108 * there must be at least one object available for
3109 * allocation.
3110 */
3111 BUG_ON(slabp->inuse >= cachep->num);
3112
3113 while (slabp->inuse < cachep->num && batchcount--) {
3114 STATS_INC_ALLOCED(cachep);
3115 STATS_INC_ACTIVE(cachep);
3116 STATS_SET_HIGH(cachep);
3117
3118 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3119 node);
3120 }
3121 check_slabp(cachep, slabp);
3122
3123 /* move slabp to correct slabp list: */
3124 list_del(&slabp->list);
3125 if (slabp->free == BUFCTL_END)
3126 list_add(&slabp->list, &l3->slabs_full);
3127 else
3128 list_add(&slabp->list, &l3->slabs_partial);
3129 }
3130
3131must_grow:
3132 l3->free_objects -= ac->avail;
3133alloc_done:
3134 spin_unlock(&l3->list_lock);
3135
3136 if (unlikely(!ac->avail)) {
3137 int x;
3138 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3139
3140 /* cache_grow can reenable interrupts, then ac could change. */
3141 ac = cpu_cache_get(cachep);
3142 if (!x && ac->avail == 0) /* no objects in sight? abort */
3143 return NULL;
3144
3145 if (!ac->avail) /* objects refilled by interrupt? */
3146 goto retry;
3147 }
3148 ac->touched = 1;
3149 return ac->entry[--ac->avail];
3150}
3151
3152static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3153 gfp_t flags)
3154{
3155 might_sleep_if(flags & __GFP_WAIT);
3156#if DEBUG
3157 kmem_flagcheck(cachep, flags);
3158#endif
3159}
3160
3161#if DEBUG
3162static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3163 gfp_t flags, void *objp, void *caller)
3164{
3165 if (!objp)
3166 return objp;
3167 if (cachep->flags & SLAB_POISON) {
3168#ifdef CONFIG_DEBUG_PAGEALLOC
3169 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3170 kernel_map_pages(virt_to_page(objp),
3171 cachep->buffer_size / PAGE_SIZE, 1);
3172 else
3173 check_poison_obj(cachep, objp);
3174#else
3175 check_poison_obj(cachep, objp);
3176#endif
3177 poison_obj(cachep, objp, POISON_INUSE);
3178 }
3179 if (cachep->flags & SLAB_STORE_USER)
3180 *dbg_userword(cachep, objp) = caller;
3181
3182 if (cachep->flags & SLAB_RED_ZONE) {
3183 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3184 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3185 slab_error(cachep, "double free, or memory outside"
3186 " object was overwritten");
3187 printk(KERN_ERR
3188 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3189 objp, *dbg_redzone1(cachep, objp),
3190 *dbg_redzone2(cachep, objp));
3191 }
3192 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3193 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3194 }
3195#ifdef CONFIG_DEBUG_SLAB_LEAK
3196 {
3197 struct slab *slabp;
3198 unsigned objnr;
3199
3200 slabp = page_get_slab(virt_to_head_page(objp));
3201 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3202 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3203 }
3204#endif
3205 objp += obj_offset(cachep);
3206 if (cachep->ctor && cachep->flags & SLAB_POISON)
3207 cachep->ctor(objp);
3208 if (ARCH_SLAB_MINALIGN &&
3209 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3210 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3211 objp, (int)ARCH_SLAB_MINALIGN);
3212 }
3213 return objp;
3214}
3215#else
3216#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3217#endif
3218
3219static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3220{
3221 if (cachep == &cache_cache)
3222 return false;
3223
3224 return should_failslab(obj_size(cachep), flags, cachep->flags);
3225}
3226
3227static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3228{
3229 void *objp;
3230 struct array_cache *ac;
3231
3232 check_irq_off();
3233
3234 ac = cpu_cache_get(cachep);
3235 if (likely(ac->avail)) {
3236 STATS_INC_ALLOCHIT(cachep);
3237 ac->touched = 1;
3238 objp = ac->entry[--ac->avail];
3239 } else {
3240 STATS_INC_ALLOCMISS(cachep);
3241 objp = cache_alloc_refill(cachep, flags);
3242 /*
3243 * the 'ac' may be updated by cache_alloc_refill(),
3244 * and kmemleak_erase() requires its correct value.
3245 */
3246 ac = cpu_cache_get(cachep);
3247 }
3248 /*
3249 * To avoid a false negative, if an object that is in one of the
3250 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3251 * treat the array pointers as a reference to the object.
3252 */
3253 if (objp)
3254 kmemleak_erase(&ac->entry[ac->avail]);
3255 return objp;
3256}
3257
3258#ifdef CONFIG_NUMA
3259/*
3260 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3261 *
3262 * If we are in_interrupt, then process context, including cpusets and
3263 * mempolicy, may not apply and should not be used for allocation policy.
3264 */
3265static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3266{
3267 int nid_alloc, nid_here;
3268
3269 if (in_interrupt() || (flags & __GFP_THISNODE))
3270 return NULL;
3271 nid_alloc = nid_here = numa_mem_id();
3272 get_mems_allowed();
3273 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3274 nid_alloc = cpuset_slab_spread_node();
3275 else if (current->mempolicy)
3276 nid_alloc = slab_node(current->mempolicy);
3277 put_mems_allowed();
3278 if (nid_alloc != nid_here)
3279 return ____cache_alloc_node(cachep, flags, nid_alloc);
3280 return NULL;
3281}
3282
3283/*
3284 * Fallback function if there was no memory available and no objects on a
3285 * certain node and fall back is permitted. First we scan all the
3286 * available nodelists for available objects. If that fails then we
3287 * perform an allocation without specifying a node. This allows the page
3288 * allocator to do its reclaim / fallback magic. We then insert the
3289 * slab into the proper nodelist and then allocate from it.
3290 */
3291static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3292{
3293 struct zonelist *zonelist;
3294 gfp_t local_flags;
3295 struct zoneref *z;
3296 struct zone *zone;
3297 enum zone_type high_zoneidx = gfp_zone(flags);
3298 void *obj = NULL;
3299 int nid;
3300
3301 if (flags & __GFP_THISNODE)
3302 return NULL;
3303
3304 get_mems_allowed();
3305 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3306 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3307
3308retry:
3309 /*
3310 * Look through allowed nodes for objects available
3311 * from existing per node queues.
3312 */
3313 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3314 nid = zone_to_nid(zone);
3315
3316 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3317 cache->nodelists[nid] &&
3318 cache->nodelists[nid]->free_objects) {
3319 obj = ____cache_alloc_node(cache,
3320 flags | GFP_THISNODE, nid);
3321 if (obj)
3322 break;
3323 }
3324 }
3325
3326 if (!obj) {
3327 /*
3328 * This allocation will be performed within the constraints
3329 * of the current cpuset / memory policy requirements.
3330 * We may trigger various forms of reclaim on the allowed
3331 * set and go into memory reserves if necessary.
3332 */
3333 if (local_flags & __GFP_WAIT)
3334 local_irq_enable();
3335 kmem_flagcheck(cache, flags);
3336 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3337 if (local_flags & __GFP_WAIT)
3338 local_irq_disable();
3339 if (obj) {
3340 /*
3341 * Insert into the appropriate per node queues
3342 */
3343 nid = page_to_nid(virt_to_page(obj));
3344 if (cache_grow(cache, flags, nid, obj)) {
3345 obj = ____cache_alloc_node(cache,
3346 flags | GFP_THISNODE, nid);
3347 if (!obj)
3348 /*
3349 * Another processor may allocate the
3350 * objects in the slab since we are
3351 * not holding any locks.
3352 */
3353 goto retry;
3354 } else {
3355 /* cache_grow already freed obj */
3356 obj = NULL;
3357 }
3358 }
3359 }
3360 put_mems_allowed();
3361 return obj;
3362}
3363
3364/*
3365 * A interface to enable slab creation on nodeid
3366 */
3367static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3368 int nodeid)
3369{
3370 struct list_head *entry;
3371 struct slab *slabp;
3372 struct kmem_list3 *l3;
3373 void *obj;
3374 int x;
3375
3376 l3 = cachep->nodelists[nodeid];
3377 BUG_ON(!l3);
3378
3379retry:
3380 check_irq_off();
3381 spin_lock(&l3->list_lock);
3382 entry = l3->slabs_partial.next;
3383 if (entry == &l3->slabs_partial) {
3384 l3->free_touched = 1;
3385 entry = l3->slabs_free.next;
3386 if (entry == &l3->slabs_free)
3387 goto must_grow;
3388 }
3389
3390 slabp = list_entry(entry, struct slab, list);
3391 check_spinlock_acquired_node(cachep, nodeid);
3392 check_slabp(cachep, slabp);
3393
3394 STATS_INC_NODEALLOCS(cachep);
3395 STATS_INC_ACTIVE(cachep);
3396 STATS_SET_HIGH(cachep);
3397
3398 BUG_ON(slabp->inuse == cachep->num);
3399
3400 obj = slab_get_obj(cachep, slabp, nodeid);
3401 check_slabp(cachep, slabp);
3402 l3->free_objects--;
3403 /* move slabp to correct slabp list: */
3404 list_del(&slabp->list);
3405
3406 if (slabp->free == BUFCTL_END)
3407 list_add(&slabp->list, &l3->slabs_full);
3408 else
3409 list_add(&slabp->list, &l3->slabs_partial);
3410
3411 spin_unlock(&l3->list_lock);
3412 goto done;
3413
3414must_grow:
3415 spin_unlock(&l3->list_lock);
3416 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3417 if (x)
3418 goto retry;
3419
3420 return fallback_alloc(cachep, flags);
3421
3422done:
3423 return obj;
3424}
3425
3426/**
3427 * kmem_cache_alloc_node - Allocate an object on the specified node
3428 * @cachep: The cache to allocate from.
3429 * @flags: See kmalloc().
3430 * @nodeid: node number of the target node.
3431 * @caller: return address of caller, used for debug information
3432 *
3433 * Identical to kmem_cache_alloc but it will allocate memory on the given
3434 * node, which can improve the performance for cpu bound structures.
3435 *
3436 * Fallback to other node is possible if __GFP_THISNODE is not set.
3437 */
3438static __always_inline void *
3439__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3440 void *caller)
3441{
3442 unsigned long save_flags;
3443 void *ptr;
3444 int slab_node = numa_mem_id();
3445
3446 flags &= gfp_allowed_mask;
3447
3448 lockdep_trace_alloc(flags);
3449
3450 if (slab_should_failslab(cachep, flags))
3451 return NULL;
3452
3453 cache_alloc_debugcheck_before(cachep, flags);
3454 local_irq_save(save_flags);
3455
3456 if (nodeid == NUMA_NO_NODE)
3457 nodeid = slab_node;
3458
3459 if (unlikely(!cachep->nodelists[nodeid])) {
3460 /* Node not bootstrapped yet */
3461 ptr = fallback_alloc(cachep, flags);
3462 goto out;
3463 }
3464
3465 if (nodeid == slab_node) {
3466 /*
3467 * Use the locally cached objects if possible.
3468 * However ____cache_alloc does not allow fallback
3469 * to other nodes. It may fail while we still have
3470 * objects on other nodes available.
3471 */
3472 ptr = ____cache_alloc(cachep, flags);
3473 if (ptr)
3474 goto out;
3475 }
3476 /* ___cache_alloc_node can fall back to other nodes */
3477 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3478 out:
3479 local_irq_restore(save_flags);
3480 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3481 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3482 flags);
3483
3484 if (likely(ptr))
3485 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3486
3487 if (unlikely((flags & __GFP_ZERO) && ptr))
3488 memset(ptr, 0, obj_size(cachep));
3489
3490 return ptr;
3491}
3492
3493static __always_inline void *
3494__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3495{
3496 void *objp;
3497
3498 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3499 objp = alternate_node_alloc(cache, flags);
3500 if (objp)
3501 goto out;
3502 }
3503 objp = ____cache_alloc(cache, flags);
3504
3505 /*
3506 * We may just have run out of memory on the local node.
3507 * ____cache_alloc_node() knows how to locate memory on other nodes
3508 */
3509 if (!objp)
3510 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3511
3512 out:
3513 return objp;
3514}
3515#else
3516
3517static __always_inline void *
3518__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3519{
3520 return ____cache_alloc(cachep, flags);
3521}
3522
3523#endif /* CONFIG_NUMA */
3524
3525static __always_inline void *
3526__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3527{
3528 unsigned long save_flags;
3529 void *objp;
3530
3531 flags &= gfp_allowed_mask;
3532
3533 lockdep_trace_alloc(flags);
3534
3535 if (slab_should_failslab(cachep, flags))
3536 return NULL;
3537
3538 cache_alloc_debugcheck_before(cachep, flags);
3539 local_irq_save(save_flags);
3540 objp = __do_cache_alloc(cachep, flags);
3541 local_irq_restore(save_flags);
3542 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3543 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3544 flags);
3545 prefetchw(objp);
3546
3547 if (likely(objp))
3548 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3549
3550 if (unlikely((flags & __GFP_ZERO) && objp))
3551 memset(objp, 0, obj_size(cachep));
3552
3553 return objp;
3554}
3555
3556/*
3557 * Caller needs to acquire correct kmem_list's list_lock
3558 */
3559static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3560 int node)
3561{
3562 int i;
3563 struct kmem_list3 *l3;
3564
3565 for (i = 0; i < nr_objects; i++) {
3566 void *objp = objpp[i];
3567 struct slab *slabp;
3568
3569 slabp = virt_to_slab(objp);
3570 l3 = cachep->nodelists[node];
3571 list_del(&slabp->list);
3572 check_spinlock_acquired_node(cachep, node);
3573 check_slabp(cachep, slabp);
3574 slab_put_obj(cachep, slabp, objp, node);
3575 STATS_DEC_ACTIVE(cachep);
3576 l3->free_objects++;
3577 check_slabp(cachep, slabp);
3578
3579 /* fixup slab chains */
3580 if (slabp->inuse == 0) {
3581 if (l3->free_objects > l3->free_limit) {
3582 l3->free_objects -= cachep->num;
3583 /* No need to drop any previously held
3584 * lock here, even if we have a off-slab slab
3585 * descriptor it is guaranteed to come from
3586 * a different cache, refer to comments before
3587 * alloc_slabmgmt.
3588 */
3589 slab_destroy(cachep, slabp);
3590 } else {
3591 list_add(&slabp->list, &l3->slabs_free);
3592 }
3593 } else {
3594 /* Unconditionally move a slab to the end of the
3595 * partial list on free - maximum time for the
3596 * other objects to be freed, too.
3597 */
3598 list_add_tail(&slabp->list, &l3->slabs_partial);
3599 }
3600 }
3601}
3602
3603static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3604{
3605 int batchcount;
3606 struct kmem_list3 *l3;
3607 int node = numa_mem_id();
3608
3609 batchcount = ac->batchcount;
3610#if DEBUG
3611 BUG_ON(!batchcount || batchcount > ac->avail);
3612#endif
3613 check_irq_off();
3614 l3 = cachep->nodelists[node];
3615 spin_lock(&l3->list_lock);
3616 if (l3->shared) {
3617 struct array_cache *shared_array = l3->shared;
3618 int max = shared_array->limit - shared_array->avail;
3619 if (max) {
3620 if (batchcount > max)
3621 batchcount = max;
3622 memcpy(&(shared_array->entry[shared_array->avail]),
3623 ac->entry, sizeof(void *) * batchcount);
3624 shared_array->avail += batchcount;
3625 goto free_done;
3626 }
3627 }
3628
3629 free_block(cachep, ac->entry, batchcount, node);
3630free_done:
3631#if STATS
3632 {
3633 int i = 0;
3634 struct list_head *p;
3635
3636 p = l3->slabs_free.next;
3637 while (p != &(l3->slabs_free)) {
3638 struct slab *slabp;
3639
3640 slabp = list_entry(p, struct slab, list);
3641 BUG_ON(slabp->inuse);
3642
3643 i++;
3644 p = p->next;
3645 }
3646 STATS_SET_FREEABLE(cachep, i);
3647 }
3648#endif
3649 spin_unlock(&l3->list_lock);
3650 ac->avail -= batchcount;
3651 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3652}
3653
3654/*
3655 * Release an obj back to its cache. If the obj has a constructed state, it must
3656 * be in this state _before_ it is released. Called with disabled ints.
3657 */
3658static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3659 void *caller)
3660{
3661 struct array_cache *ac = cpu_cache_get(cachep);
3662
3663 check_irq_off();
3664 kmemleak_free_recursive(objp, cachep->flags);
3665 objp = cache_free_debugcheck(cachep, objp, caller);
3666
3667 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3668
3669 /*
3670 * Skip calling cache_free_alien() when the platform is not numa.
3671 * This will avoid cache misses that happen while accessing slabp (which
3672 * is per page memory reference) to get nodeid. Instead use a global
3673 * variable to skip the call, which is mostly likely to be present in
3674 * the cache.
3675 */
3676 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3677 return;
3678
3679 if (likely(ac->avail < ac->limit)) {
3680 STATS_INC_FREEHIT(cachep);
3681 ac->entry[ac->avail++] = objp;
3682 return;
3683 } else {
3684 STATS_INC_FREEMISS(cachep);
3685 cache_flusharray(cachep, ac);
3686 ac->entry[ac->avail++] = objp;
3687 }
3688}
3689
3690/**
3691 * kmem_cache_alloc - Allocate an object
3692 * @cachep: The cache to allocate from.
3693 * @flags: See kmalloc().
3694 *
3695 * Allocate an object from this cache. The flags are only relevant
3696 * if the cache has no available objects.
3697 */
3698void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3699{
3700 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3701
3702 trace_kmem_cache_alloc(_RET_IP_, ret,
3703 obj_size(cachep), cachep->buffer_size, flags);
3704
3705 return ret;
3706}
3707EXPORT_SYMBOL(kmem_cache_alloc);
3708
3709#ifdef CONFIG_TRACING
3710void *
3711kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3712{
3713 void *ret;
3714
3715 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3716
3717 trace_kmalloc(_RET_IP_, ret,
3718 size, slab_buffer_size(cachep), flags);
3719 return ret;
3720}
3721EXPORT_SYMBOL(kmem_cache_alloc_trace);
3722#endif
3723
3724#ifdef CONFIG_NUMA
3725void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3726{
3727 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3728 __builtin_return_address(0));
3729
3730 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3731 obj_size(cachep), cachep->buffer_size,
3732 flags, nodeid);
3733
3734 return ret;
3735}
3736EXPORT_SYMBOL(kmem_cache_alloc_node);
3737
3738#ifdef CONFIG_TRACING
3739void *kmem_cache_alloc_node_trace(size_t size,
3740 struct kmem_cache *cachep,
3741 gfp_t flags,
3742 int nodeid)
3743{
3744 void *ret;
3745
3746 ret = __cache_alloc_node(cachep, flags, nodeid,
3747 __builtin_return_address(0));
3748 trace_kmalloc_node(_RET_IP_, ret,
3749 size, slab_buffer_size(cachep),
3750 flags, nodeid);
3751 return ret;
3752}
3753EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3754#endif
3755
3756static __always_inline void *
3757__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3758{
3759 struct kmem_cache *cachep;
3760
3761 cachep = kmem_find_general_cachep(size, flags);
3762 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3763 return cachep;
3764 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3765}
3766
3767#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3768void *__kmalloc_node(size_t size, gfp_t flags, int node)
3769{
3770 return __do_kmalloc_node(size, flags, node,
3771 __builtin_return_address(0));
3772}
3773EXPORT_SYMBOL(__kmalloc_node);
3774
3775void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3776 int node, unsigned long caller)
3777{
3778 return __do_kmalloc_node(size, flags, node, (void *)caller);
3779}
3780EXPORT_SYMBOL(__kmalloc_node_track_caller);
3781#else
3782void *__kmalloc_node(size_t size, gfp_t flags, int node)
3783{
3784 return __do_kmalloc_node(size, flags, node, NULL);
3785}
3786EXPORT_SYMBOL(__kmalloc_node);
3787#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3788#endif /* CONFIG_NUMA */
3789
3790/**
3791 * __do_kmalloc - allocate memory
3792 * @size: how many bytes of memory are required.
3793 * @flags: the type of memory to allocate (see kmalloc).
3794 * @caller: function caller for debug tracking of the caller
3795 */
3796static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3797 void *caller)
3798{
3799 struct kmem_cache *cachep;
3800 void *ret;
3801
3802 /* If you want to save a few bytes .text space: replace
3803 * __ with kmem_.
3804 * Then kmalloc uses the uninlined functions instead of the inline
3805 * functions.
3806 */
3807 cachep = __find_general_cachep(size, flags);
3808 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3809 return cachep;
3810 ret = __cache_alloc(cachep, flags, caller);
3811
3812 trace_kmalloc((unsigned long) caller, ret,
3813 size, cachep->buffer_size, flags);
3814
3815 return ret;
3816}
3817
3818
3819#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3820void *__kmalloc(size_t size, gfp_t flags)
3821{
3822 return __do_kmalloc(size, flags, __builtin_return_address(0));
3823}
3824EXPORT_SYMBOL(__kmalloc);
3825
3826void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3827{
3828 return __do_kmalloc(size, flags, (void *)caller);
3829}
3830EXPORT_SYMBOL(__kmalloc_track_caller);
3831
3832#else
3833void *__kmalloc(size_t size, gfp_t flags)
3834{
3835 return __do_kmalloc(size, flags, NULL);
3836}
3837EXPORT_SYMBOL(__kmalloc);
3838#endif
3839
3840/**
3841 * kmem_cache_free - Deallocate an object
3842 * @cachep: The cache the allocation was from.
3843 * @objp: The previously allocated object.
3844 *
3845 * Free an object which was previously allocated from this
3846 * cache.
3847 */
3848void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3849{
3850 unsigned long flags;
3851
3852 local_irq_save(flags);
3853 debug_check_no_locks_freed(objp, obj_size(cachep));
3854 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3855 debug_check_no_obj_freed(objp, obj_size(cachep));
3856 __cache_free(cachep, objp, __builtin_return_address(0));
3857 local_irq_restore(flags);
3858
3859 trace_kmem_cache_free(_RET_IP_, objp);
3860}
3861EXPORT_SYMBOL(kmem_cache_free);
3862
3863/**
3864 * kfree - free previously allocated memory
3865 * @objp: pointer returned by kmalloc.
3866 *
3867 * If @objp is NULL, no operation is performed.
3868 *
3869 * Don't free memory not originally allocated by kmalloc()
3870 * or you will run into trouble.
3871 */
3872void kfree(const void *objp)
3873{
3874 struct kmem_cache *c;
3875 unsigned long flags;
3876
3877 trace_kfree(_RET_IP_, objp);
3878
3879 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3880 return;
3881 local_irq_save(flags);
3882 kfree_debugcheck(objp);
3883 c = virt_to_cache(objp);
3884 debug_check_no_locks_freed(objp, obj_size(c));
3885 debug_check_no_obj_freed(objp, obj_size(c));
3886 __cache_free(c, (void *)objp, __builtin_return_address(0));
3887 local_irq_restore(flags);
3888}
3889EXPORT_SYMBOL(kfree);
3890
3891unsigned int kmem_cache_size(struct kmem_cache *cachep)
3892{
3893 return obj_size(cachep);
3894}
3895EXPORT_SYMBOL(kmem_cache_size);
3896
3897/*
3898 * This initializes kmem_list3 or resizes various caches for all nodes.
3899 */
3900static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3901{
3902 int node;
3903 struct kmem_list3 *l3;
3904 struct array_cache *new_shared;
3905 struct array_cache **new_alien = NULL;
3906
3907 for_each_online_node(node) {
3908
3909 if (use_alien_caches) {
3910 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3911 if (!new_alien)
3912 goto fail;
3913 }
3914
3915 new_shared = NULL;
3916 if (cachep->shared) {
3917 new_shared = alloc_arraycache(node,
3918 cachep->shared*cachep->batchcount,
3919 0xbaadf00d, gfp);
3920 if (!new_shared) {
3921 free_alien_cache(new_alien);
3922 goto fail;
3923 }
3924 }
3925
3926 l3 = cachep->nodelists[node];
3927 if (l3) {
3928 struct array_cache *shared = l3->shared;
3929
3930 spin_lock_irq(&l3->list_lock);
3931
3932 if (shared)
3933 free_block(cachep, shared->entry,
3934 shared->avail, node);
3935
3936 l3->shared = new_shared;
3937 if (!l3->alien) {
3938 l3->alien = new_alien;
3939 new_alien = NULL;
3940 }
3941 l3->free_limit = (1 + nr_cpus_node(node)) *
3942 cachep->batchcount + cachep->num;
3943 spin_unlock_irq(&l3->list_lock);
3944 kfree(shared);
3945 free_alien_cache(new_alien);
3946 continue;
3947 }
3948 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3949 if (!l3) {
3950 free_alien_cache(new_alien);
3951 kfree(new_shared);
3952 goto fail;
3953 }
3954
3955 kmem_list3_init(l3);
3956 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3957 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3958 l3->shared = new_shared;
3959 l3->alien = new_alien;
3960 l3->free_limit = (1 + nr_cpus_node(node)) *
3961 cachep->batchcount + cachep->num;
3962 cachep->nodelists[node] = l3;
3963 }
3964 return 0;
3965
3966fail:
3967 if (!cachep->next.next) {
3968 /* Cache is not active yet. Roll back what we did */
3969 node--;
3970 while (node >= 0) {
3971 if (cachep->nodelists[node]) {
3972 l3 = cachep->nodelists[node];
3973
3974 kfree(l3->shared);
3975 free_alien_cache(l3->alien);
3976 kfree(l3);
3977 cachep->nodelists[node] = NULL;
3978 }
3979 node--;
3980 }
3981 }
3982 return -ENOMEM;
3983}
3984
3985struct ccupdate_struct {
3986 struct kmem_cache *cachep;
3987 struct array_cache *new[0];
3988};
3989
3990static void do_ccupdate_local(void *info)
3991{
3992 struct ccupdate_struct *new = info;
3993 struct array_cache *old;
3994
3995 check_irq_off();
3996 old = cpu_cache_get(new->cachep);
3997
3998 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3999 new->new[smp_processor_id()] = old;
4000}
4001
4002/* Always called with the cache_chain_mutex held */
4003static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4004 int batchcount, int shared, gfp_t gfp)
4005{
4006 struct ccupdate_struct *new;
4007 int i;
4008
4009 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4010 gfp);
4011 if (!new)
4012 return -ENOMEM;
4013
4014 for_each_online_cpu(i) {
4015 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4016 batchcount, gfp);
4017 if (!new->new[i]) {
4018 for (i--; i >= 0; i--)
4019 kfree(new->new[i]);
4020 kfree(new);
4021 return -ENOMEM;
4022 }
4023 }
4024 new->cachep = cachep;
4025
4026 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4027
4028 check_irq_on();
4029 cachep->batchcount = batchcount;
4030 cachep->limit = limit;
4031 cachep->shared = shared;
4032
4033 for_each_online_cpu(i) {
4034 struct array_cache *ccold = new->new[i];
4035 if (!ccold)
4036 continue;
4037 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4038 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4039 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4040 kfree(ccold);
4041 }
4042 kfree(new);
4043 return alloc_kmemlist(cachep, gfp);
4044}
4045
4046/* Called with cache_chain_mutex held always */
4047static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4048{
4049 int err;
4050 int limit, shared;
4051
4052 /*
4053 * The head array serves three purposes:
4054 * - create a LIFO ordering, i.e. return objects that are cache-warm
4055 * - reduce the number of spinlock operations.
4056 * - reduce the number of linked list operations on the slab and
4057 * bufctl chains: array operations are cheaper.
4058 * The numbers are guessed, we should auto-tune as described by
4059 * Bonwick.
4060 */
4061 if (cachep->buffer_size > 131072)
4062 limit = 1;
4063 else if (cachep->buffer_size > PAGE_SIZE)
4064 limit = 8;
4065 else if (cachep->buffer_size > 1024)
4066 limit = 24;
4067 else if (cachep->buffer_size > 256)
4068 limit = 54;
4069 else
4070 limit = 120;
4071
4072 /*
4073 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4074 * allocation behaviour: Most allocs on one cpu, most free operations
4075 * on another cpu. For these cases, an efficient object passing between
4076 * cpus is necessary. This is provided by a shared array. The array
4077 * replaces Bonwick's magazine layer.
4078 * On uniprocessor, it's functionally equivalent (but less efficient)
4079 * to a larger limit. Thus disabled by default.
4080 */
4081 shared = 0;
4082 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4083 shared = 8;
4084
4085#if DEBUG
4086 /*
4087 * With debugging enabled, large batchcount lead to excessively long
4088 * periods with disabled local interrupts. Limit the batchcount
4089 */
4090 if (limit > 32)
4091 limit = 32;
4092#endif
4093 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4094 if (err)
4095 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4096 cachep->name, -err);
4097 return err;
4098}
4099
4100/*
4101 * Drain an array if it contains any elements taking the l3 lock only if
4102 * necessary. Note that the l3 listlock also protects the array_cache
4103 * if drain_array() is used on the shared array.
4104 */
4105static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4106 struct array_cache *ac, int force, int node)
4107{
4108 int tofree;
4109
4110 if (!ac || !ac->avail)
4111 return;
4112 if (ac->touched && !force) {
4113 ac->touched = 0;
4114 } else {
4115 spin_lock_irq(&l3->list_lock);
4116 if (ac->avail) {
4117 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4118 if (tofree > ac->avail)
4119 tofree = (ac->avail + 1) / 2;
4120 free_block(cachep, ac->entry, tofree, node);
4121 ac->avail -= tofree;
4122 memmove(ac->entry, &(ac->entry[tofree]),
4123 sizeof(void *) * ac->avail);
4124 }
4125 spin_unlock_irq(&l3->list_lock);
4126 }
4127}
4128
4129/**
4130 * cache_reap - Reclaim memory from caches.
4131 * @w: work descriptor
4132 *
4133 * Called from workqueue/eventd every few seconds.
4134 * Purpose:
4135 * - clear the per-cpu caches for this CPU.
4136 * - return freeable pages to the main free memory pool.
4137 *
4138 * If we cannot acquire the cache chain mutex then just give up - we'll try
4139 * again on the next iteration.
4140 */
4141static void cache_reap(struct work_struct *w)
4142{
4143 struct kmem_cache *searchp;
4144 struct kmem_list3 *l3;
4145 int node = numa_mem_id();
4146 struct delayed_work *work = to_delayed_work(w);
4147
4148 if (!mutex_trylock(&cache_chain_mutex))
4149 /* Give up. Setup the next iteration. */
4150 goto out;
4151
4152 list_for_each_entry(searchp, &cache_chain, next) {
4153 check_irq_on();
4154
4155 /*
4156 * We only take the l3 lock if absolutely necessary and we
4157 * have established with reasonable certainty that
4158 * we can do some work if the lock was obtained.
4159 */
4160 l3 = searchp->nodelists[node];
4161
4162 reap_alien(searchp, l3);
4163
4164 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4165
4166 /*
4167 * These are racy checks but it does not matter
4168 * if we skip one check or scan twice.
4169 */
4170 if (time_after(l3->next_reap, jiffies))
4171 goto next;
4172
4173 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4174
4175 drain_array(searchp, l3, l3->shared, 0, node);
4176
4177 if (l3->free_touched)
4178 l3->free_touched = 0;
4179 else {
4180 int freed;
4181
4182 freed = drain_freelist(searchp, l3, (l3->free_limit +
4183 5 * searchp->num - 1) / (5 * searchp->num));
4184 STATS_ADD_REAPED(searchp, freed);
4185 }
4186next:
4187 cond_resched();
4188 }
4189 check_irq_on();
4190 mutex_unlock(&cache_chain_mutex);
4191 next_reap_node();
4192out:
4193 /* Set up the next iteration */
4194 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4195}
4196
4197#ifdef CONFIG_SLABINFO
4198
4199static void print_slabinfo_header(struct seq_file *m)
4200{
4201 /*
4202 * Output format version, so at least we can change it
4203 * without _too_ many complaints.
4204 */
4205#if STATS
4206 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4207#else
4208 seq_puts(m, "slabinfo - version: 2.1\n");
4209#endif
4210 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4211 "<objperslab> <pagesperslab>");
4212 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4213 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4214#if STATS
4215 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4216 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4217 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4218#endif
4219 seq_putc(m, '\n');
4220}
4221
4222static void *s_start(struct seq_file *m, loff_t *pos)
4223{
4224 loff_t n = *pos;
4225
4226 mutex_lock(&cache_chain_mutex);
4227 if (!n)
4228 print_slabinfo_header(m);
4229
4230 return seq_list_start(&cache_chain, *pos);
4231}
4232
4233static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4234{
4235 return seq_list_next(p, &cache_chain, pos);
4236}
4237
4238static void s_stop(struct seq_file *m, void *p)
4239{
4240 mutex_unlock(&cache_chain_mutex);
4241}
4242
4243static int s_show(struct seq_file *m, void *p)
4244{
4245 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4246 struct slab *slabp;
4247 unsigned long active_objs;
4248 unsigned long num_objs;
4249 unsigned long active_slabs = 0;
4250 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4251 const char *name;
4252 char *error = NULL;
4253 int node;
4254 struct kmem_list3 *l3;
4255
4256 active_objs = 0;
4257 num_slabs = 0;
4258 for_each_online_node(node) {
4259 l3 = cachep->nodelists[node];
4260 if (!l3)
4261 continue;
4262
4263 check_irq_on();
4264 spin_lock_irq(&l3->list_lock);
4265
4266 list_for_each_entry(slabp, &l3->slabs_full, list) {
4267 if (slabp->inuse != cachep->num && !error)
4268 error = "slabs_full accounting error";
4269 active_objs += cachep->num;
4270 active_slabs++;
4271 }
4272 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4273 if (slabp->inuse == cachep->num && !error)
4274 error = "slabs_partial inuse accounting error";
4275 if (!slabp->inuse && !error)
4276 error = "slabs_partial/inuse accounting error";
4277 active_objs += slabp->inuse;
4278 active_slabs++;
4279 }
4280 list_for_each_entry(slabp, &l3->slabs_free, list) {
4281 if (slabp->inuse && !error)
4282 error = "slabs_free/inuse accounting error";
4283 num_slabs++;
4284 }
4285 free_objects += l3->free_objects;
4286 if (l3->shared)
4287 shared_avail += l3->shared->avail;
4288
4289 spin_unlock_irq(&l3->list_lock);
4290 }
4291 num_slabs += active_slabs;
4292 num_objs = num_slabs * cachep->num;
4293 if (num_objs - active_objs != free_objects && !error)
4294 error = "free_objects accounting error";
4295
4296 name = cachep->name;
4297 if (error)
4298 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4299
4300 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4301 name, active_objs, num_objs, cachep->buffer_size,
4302 cachep->num, (1 << cachep->gfporder));
4303 seq_printf(m, " : tunables %4u %4u %4u",
4304 cachep->limit, cachep->batchcount, cachep->shared);
4305 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4306 active_slabs, num_slabs, shared_avail);
4307#if STATS
4308 { /* list3 stats */
4309 unsigned long high = cachep->high_mark;
4310 unsigned long allocs = cachep->num_allocations;
4311 unsigned long grown = cachep->grown;
4312 unsigned long reaped = cachep->reaped;
4313 unsigned long errors = cachep->errors;
4314 unsigned long max_freeable = cachep->max_freeable;
4315 unsigned long node_allocs = cachep->node_allocs;
4316 unsigned long node_frees = cachep->node_frees;
4317 unsigned long overflows = cachep->node_overflow;
4318
4319 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4320 "%4lu %4lu %4lu %4lu %4lu",
4321 allocs, high, grown,
4322 reaped, errors, max_freeable, node_allocs,
4323 node_frees, overflows);
4324 }
4325 /* cpu stats */
4326 {
4327 unsigned long allochit = atomic_read(&cachep->allochit);
4328 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4329 unsigned long freehit = atomic_read(&cachep->freehit);
4330 unsigned long freemiss = atomic_read(&cachep->freemiss);
4331
4332 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4333 allochit, allocmiss, freehit, freemiss);
4334 }
4335#endif
4336 seq_putc(m, '\n');
4337 return 0;
4338}
4339
4340/*
4341 * slabinfo_op - iterator that generates /proc/slabinfo
4342 *
4343 * Output layout:
4344 * cache-name
4345 * num-active-objs
4346 * total-objs
4347 * object size
4348 * num-active-slabs
4349 * total-slabs
4350 * num-pages-per-slab
4351 * + further values on SMP and with statistics enabled
4352 */
4353
4354static const struct seq_operations slabinfo_op = {
4355 .start = s_start,
4356 .next = s_next,
4357 .stop = s_stop,
4358 .show = s_show,
4359};
4360
4361#define MAX_SLABINFO_WRITE 128
4362/**
4363 * slabinfo_write - Tuning for the slab allocator
4364 * @file: unused
4365 * @buffer: user buffer
4366 * @count: data length
4367 * @ppos: unused
4368 */
4369static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4370 size_t count, loff_t *ppos)
4371{
4372 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4373 int limit, batchcount, shared, res;
4374 struct kmem_cache *cachep;
4375
4376 if (count > MAX_SLABINFO_WRITE)
4377 return -EINVAL;
4378 if (copy_from_user(&kbuf, buffer, count))
4379 return -EFAULT;
4380 kbuf[MAX_SLABINFO_WRITE] = '\0';
4381
4382 tmp = strchr(kbuf, ' ');
4383 if (!tmp)
4384 return -EINVAL;
4385 *tmp = '\0';
4386 tmp++;
4387 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4388 return -EINVAL;
4389
4390 /* Find the cache in the chain of caches. */
4391 mutex_lock(&cache_chain_mutex);
4392 res = -EINVAL;
4393 list_for_each_entry(cachep, &cache_chain, next) {
4394 if (!strcmp(cachep->name, kbuf)) {
4395 if (limit < 1 || batchcount < 1 ||
4396 batchcount > limit || shared < 0) {
4397 res = 0;
4398 } else {
4399 res = do_tune_cpucache(cachep, limit,
4400 batchcount, shared,
4401 GFP_KERNEL);
4402 }
4403 break;
4404 }
4405 }
4406 mutex_unlock(&cache_chain_mutex);
4407 if (res >= 0)
4408 res = count;
4409 return res;
4410}
4411
4412static int slabinfo_open(struct inode *inode, struct file *file)
4413{
4414 return seq_open(file, &slabinfo_op);
4415}
4416
4417static const struct file_operations proc_slabinfo_operations = {
4418 .open = slabinfo_open,
4419 .read = seq_read,
4420 .write = slabinfo_write,
4421 .llseek = seq_lseek,
4422 .release = seq_release,
4423};
4424
4425#ifdef CONFIG_DEBUG_SLAB_LEAK
4426
4427static void *leaks_start(struct seq_file *m, loff_t *pos)
4428{
4429 mutex_lock(&cache_chain_mutex);
4430 return seq_list_start(&cache_chain, *pos);
4431}
4432
4433static inline int add_caller(unsigned long *n, unsigned long v)
4434{
4435 unsigned long *p;
4436 int l;
4437 if (!v)
4438 return 1;
4439 l = n[1];
4440 p = n + 2;
4441 while (l) {
4442 int i = l/2;
4443 unsigned long *q = p + 2 * i;
4444 if (*q == v) {
4445 q[1]++;
4446 return 1;
4447 }
4448 if (*q > v) {
4449 l = i;
4450 } else {
4451 p = q + 2;
4452 l -= i + 1;
4453 }
4454 }
4455 if (++n[1] == n[0])
4456 return 0;
4457 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4458 p[0] = v;
4459 p[1] = 1;
4460 return 1;
4461}
4462
4463static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4464{
4465 void *p;
4466 int i;
4467 if (n[0] == n[1])
4468 return;
4469 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4470 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4471 continue;
4472 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4473 return;
4474 }
4475}
4476
4477static void show_symbol(struct seq_file *m, unsigned long address)
4478{
4479#ifdef CONFIG_KALLSYMS
4480 unsigned long offset, size;
4481 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4482
4483 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4484 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4485 if (modname[0])
4486 seq_printf(m, " [%s]", modname);
4487 return;
4488 }
4489#endif
4490 seq_printf(m, "%p", (void *)address);
4491}
4492
4493static int leaks_show(struct seq_file *m, void *p)
4494{
4495 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4496 struct slab *slabp;
4497 struct kmem_list3 *l3;
4498 const char *name;
4499 unsigned long *n = m->private;
4500 int node;
4501 int i;
4502
4503 if (!(cachep->flags & SLAB_STORE_USER))
4504 return 0;
4505 if (!(cachep->flags & SLAB_RED_ZONE))
4506 return 0;
4507
4508 /* OK, we can do it */
4509
4510 n[1] = 0;
4511
4512 for_each_online_node(node) {
4513 l3 = cachep->nodelists[node];
4514 if (!l3)
4515 continue;
4516
4517 check_irq_on();
4518 spin_lock_irq(&l3->list_lock);
4519
4520 list_for_each_entry(slabp, &l3->slabs_full, list)
4521 handle_slab(n, cachep, slabp);
4522 list_for_each_entry(slabp, &l3->slabs_partial, list)
4523 handle_slab(n, cachep, slabp);
4524 spin_unlock_irq(&l3->list_lock);
4525 }
4526 name = cachep->name;
4527 if (n[0] == n[1]) {
4528 /* Increase the buffer size */
4529 mutex_unlock(&cache_chain_mutex);
4530 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4531 if (!m->private) {
4532 /* Too bad, we are really out */
4533 m->private = n;
4534 mutex_lock(&cache_chain_mutex);
4535 return -ENOMEM;
4536 }
4537 *(unsigned long *)m->private = n[0] * 2;
4538 kfree(n);
4539 mutex_lock(&cache_chain_mutex);
4540 /* Now make sure this entry will be retried */
4541 m->count = m->size;
4542 return 0;
4543 }
4544 for (i = 0; i < n[1]; i++) {
4545 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4546 show_symbol(m, n[2*i+2]);
4547 seq_putc(m, '\n');
4548 }
4549
4550 return 0;
4551}
4552
4553static const struct seq_operations slabstats_op = {
4554 .start = leaks_start,
4555 .next = s_next,
4556 .stop = s_stop,
4557 .show = leaks_show,
4558};
4559
4560static int slabstats_open(struct inode *inode, struct file *file)
4561{
4562 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4563 int ret = -ENOMEM;
4564 if (n) {
4565 ret = seq_open(file, &slabstats_op);
4566 if (!ret) {
4567 struct seq_file *m = file->private_data;
4568 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4569 m->private = n;
4570 n = NULL;
4571 }
4572 kfree(n);
4573 }
4574 return ret;
4575}
4576
4577static const struct file_operations proc_slabstats_operations = {
4578 .open = slabstats_open,
4579 .read = seq_read,
4580 .llseek = seq_lseek,
4581 .release = seq_release_private,
4582};
4583#endif
4584
4585static int __init slab_proc_init(void)
4586{
4587 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4588#ifdef CONFIG_DEBUG_SLAB_LEAK
4589 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4590#endif
4591 return 0;
4592}
4593module_init(slab_proc_init);
4594#endif
4595
4596/**
4597 * ksize - get the actual amount of memory allocated for a given object
4598 * @objp: Pointer to the object
4599 *
4600 * kmalloc may internally round up allocations and return more memory
4601 * than requested. ksize() can be used to determine the actual amount of
4602 * memory allocated. The caller may use this additional memory, even though
4603 * a smaller amount of memory was initially specified with the kmalloc call.
4604 * The caller must guarantee that objp points to a valid object previously
4605 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4606 * must not be freed during the duration of the call.
4607 */
4608size_t ksize(const void *objp)
4609{
4610 BUG_ON(!objp);
4611 if (unlikely(objp == ZERO_SIZE_PTR))
4612 return 0;
4613
4614 return obj_size(virt_to_cache(objp));
4615}
4616EXPORT_SYMBOL(ksize);