Loading...
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#include <trace/events/kmem.h>
125
126/*
127 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * STATS - 1 to collect stats for /proc/slabinfo.
131 * 0 for faster, smaller code (especially in the critical paths).
132 *
133 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
134 */
135
136#ifdef CONFIG_DEBUG_SLAB
137#define DEBUG 1
138#define STATS 1
139#define FORCED_DEBUG 1
140#else
141#define DEBUG 0
142#define STATS 0
143#define FORCED_DEBUG 0
144#endif
145
146/* Shouldn't this be in a header file somewhere? */
147#define BYTES_PER_WORD sizeof(void *)
148#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
149
150#ifndef ARCH_KMALLOC_FLAGS
151#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
152#endif
153
154/* Legal flag mask for kmem_cache_create(). */
155#if DEBUG
156# define CREATE_MASK (SLAB_RED_ZONE | \
157 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
158 SLAB_CACHE_DMA | \
159 SLAB_STORE_USER | \
160 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
161 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
162 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
163#else
164# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
165 SLAB_CACHE_DMA | \
166 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
167 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
168 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
169#endif
170
171/*
172 * kmem_bufctl_t:
173 *
174 * Bufctl's are used for linking objs within a slab
175 * linked offsets.
176 *
177 * This implementation relies on "struct page" for locating the cache &
178 * slab an object belongs to.
179 * This allows the bufctl structure to be small (one int), but limits
180 * the number of objects a slab (not a cache) can contain when off-slab
181 * bufctls are used. The limit is the size of the largest general cache
182 * that does not use off-slab slabs.
183 * For 32bit archs with 4 kB pages, is this 56.
184 * This is not serious, as it is only for large objects, when it is unwise
185 * to have too many per slab.
186 * Note: This limit can be raised by introducing a general cache whose size
187 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
188 */
189
190typedef unsigned int kmem_bufctl_t;
191#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
192#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
193#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
194#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
195
196/*
197 * struct slab_rcu
198 *
199 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
200 * arrange for kmem_freepages to be called via RCU. This is useful if
201 * we need to approach a kernel structure obliquely, from its address
202 * obtained without the usual locking. We can lock the structure to
203 * stabilize it and check it's still at the given address, only if we
204 * can be sure that the memory has not been meanwhile reused for some
205 * other kind of object (which our subsystem's lock might corrupt).
206 *
207 * rcu_read_lock before reading the address, then rcu_read_unlock after
208 * taking the spinlock within the structure expected at that address.
209 */
210struct slab_rcu {
211 struct rcu_head head;
212 struct kmem_cache *cachep;
213 void *addr;
214};
215
216/*
217 * struct slab
218 *
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 */
223struct slab {
224 union {
225 struct {
226 struct list_head list;
227 unsigned long colouroff;
228 void *s_mem; /* including colour offset */
229 unsigned int inuse; /* num of objs active in slab */
230 kmem_bufctl_t free;
231 unsigned short nodeid;
232 };
233 struct slab_rcu __slab_cover_slab_rcu;
234 };
235};
236
237/*
238 * struct array_cache
239 *
240 * Purpose:
241 * - LIFO ordering, to hand out cache-warm objects from _alloc
242 * - reduce the number of linked list operations
243 * - reduce spinlock operations
244 *
245 * The limit is stored in the per-cpu structure to reduce the data cache
246 * footprint.
247 *
248 */
249struct array_cache {
250 unsigned int avail;
251 unsigned int limit;
252 unsigned int batchcount;
253 unsigned int touched;
254 spinlock_t lock;
255 void *entry[]; /*
256 * Must have this definition in here for the proper
257 * alignment of array_cache. Also simplifies accessing
258 * the entries.
259 */
260};
261
262/*
263 * bootstrap: The caches do not work without cpuarrays anymore, but the
264 * cpuarrays are allocated from the generic caches...
265 */
266#define BOOT_CPUCACHE_ENTRIES 1
267struct arraycache_init {
268 struct array_cache cache;
269 void *entries[BOOT_CPUCACHE_ENTRIES];
270};
271
272/*
273 * The slab lists for all objects.
274 */
275struct kmem_list3 {
276 struct list_head slabs_partial; /* partial list first, better asm code */
277 struct list_head slabs_full;
278 struct list_head slabs_free;
279 unsigned long free_objects;
280 unsigned int free_limit;
281 unsigned int colour_next; /* Per-node cache coloring */
282 spinlock_t list_lock;
283 struct array_cache *shared; /* shared per node */
284 struct array_cache **alien; /* on other nodes */
285 unsigned long next_reap; /* updated without locking */
286 int free_touched; /* updated without locking */
287};
288
289/*
290 * Need this for bootstrapping a per node allocator.
291 */
292#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
293static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
294#define CACHE_CACHE 0
295#define SIZE_AC MAX_NUMNODES
296#define SIZE_L3 (2 * MAX_NUMNODES)
297
298static int drain_freelist(struct kmem_cache *cache,
299 struct kmem_list3 *l3, int tofree);
300static void free_block(struct kmem_cache *cachep, void **objpp, int len,
301 int node);
302static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
303static void cache_reap(struct work_struct *unused);
304
305/*
306 * This function must be completely optimized away if a constant is passed to
307 * it. Mostly the same as what is in linux/slab.h except it returns an index.
308 */
309static __always_inline int index_of(const size_t size)
310{
311 extern void __bad_size(void);
312
313 if (__builtin_constant_p(size)) {
314 int i = 0;
315
316#define CACHE(x) \
317 if (size <=x) \
318 return i; \
319 else \
320 i++;
321#include <linux/kmalloc_sizes.h>
322#undef CACHE
323 __bad_size();
324 } else
325 __bad_size();
326 return 0;
327}
328
329static int slab_early_init = 1;
330
331#define INDEX_AC index_of(sizeof(struct arraycache_init))
332#define INDEX_L3 index_of(sizeof(struct kmem_list3))
333
334static void kmem_list3_init(struct kmem_list3 *parent)
335{
336 INIT_LIST_HEAD(&parent->slabs_full);
337 INIT_LIST_HEAD(&parent->slabs_partial);
338 INIT_LIST_HEAD(&parent->slabs_free);
339 parent->shared = NULL;
340 parent->alien = NULL;
341 parent->colour_next = 0;
342 spin_lock_init(&parent->list_lock);
343 parent->free_objects = 0;
344 parent->free_touched = 0;
345}
346
347#define MAKE_LIST(cachep, listp, slab, nodeid) \
348 do { \
349 INIT_LIST_HEAD(listp); \
350 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
351 } while (0)
352
353#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
354 do { \
355 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
356 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
358 } while (0)
359
360#define CFLGS_OFF_SLAB (0x80000000UL)
361#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
362
363#define BATCHREFILL_LIMIT 16
364/*
365 * Optimization question: fewer reaps means less probability for unnessary
366 * cpucache drain/refill cycles.
367 *
368 * OTOH the cpuarrays can contain lots of objects,
369 * which could lock up otherwise freeable slabs.
370 */
371#define REAPTIMEOUT_CPUC (2*HZ)
372#define REAPTIMEOUT_LIST3 (4*HZ)
373
374#if STATS
375#define STATS_INC_ACTIVE(x) ((x)->num_active++)
376#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378#define STATS_INC_GROWN(x) ((x)->grown++)
379#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
380#define STATS_SET_HIGH(x) \
381 do { \
382 if ((x)->num_active > (x)->high_mark) \
383 (x)->high_mark = (x)->num_active; \
384 } while (0)
385#define STATS_INC_ERR(x) ((x)->errors++)
386#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
387#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
388#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
389#define STATS_SET_FREEABLE(x, i) \
390 do { \
391 if ((x)->max_freeable < i) \
392 (x)->max_freeable = i; \
393 } while (0)
394#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
395#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
396#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
397#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
398#else
399#define STATS_INC_ACTIVE(x) do { } while (0)
400#define STATS_DEC_ACTIVE(x) do { } while (0)
401#define STATS_INC_ALLOCED(x) do { } while (0)
402#define STATS_INC_GROWN(x) do { } while (0)
403#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
404#define STATS_SET_HIGH(x) do { } while (0)
405#define STATS_INC_ERR(x) do { } while (0)
406#define STATS_INC_NODEALLOCS(x) do { } while (0)
407#define STATS_INC_NODEFREES(x) do { } while (0)
408#define STATS_INC_ACOVERFLOW(x) do { } while (0)
409#define STATS_SET_FREEABLE(x, i) do { } while (0)
410#define STATS_INC_ALLOCHIT(x) do { } while (0)
411#define STATS_INC_ALLOCMISS(x) do { } while (0)
412#define STATS_INC_FREEHIT(x) do { } while (0)
413#define STATS_INC_FREEMISS(x) do { } while (0)
414#endif
415
416#if DEBUG
417
418/*
419 * memory layout of objects:
420 * 0 : objp
421 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
422 * the end of an object is aligned with the end of the real
423 * allocation. Catches writes behind the end of the allocation.
424 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
425 * redzone word.
426 * cachep->obj_offset: The real object.
427 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
428 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
429 * [BYTES_PER_WORD long]
430 */
431static int obj_offset(struct kmem_cache *cachep)
432{
433 return cachep->obj_offset;
434}
435
436static int obj_size(struct kmem_cache *cachep)
437{
438 return cachep->obj_size;
439}
440
441static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
442{
443 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
444 return (unsigned long long*) (objp + obj_offset(cachep) -
445 sizeof(unsigned long long));
446}
447
448static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
449{
450 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
451 if (cachep->flags & SLAB_STORE_USER)
452 return (unsigned long long *)(objp + cachep->buffer_size -
453 sizeof(unsigned long long) -
454 REDZONE_ALIGN);
455 return (unsigned long long *) (objp + cachep->buffer_size -
456 sizeof(unsigned long long));
457}
458
459static void **dbg_userword(struct kmem_cache *cachep, void *objp)
460{
461 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
462 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
463}
464
465#else
466
467#define obj_offset(x) 0
468#define obj_size(cachep) (cachep->buffer_size)
469#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
470#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
471#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
472
473#endif
474
475#ifdef CONFIG_TRACING
476size_t slab_buffer_size(struct kmem_cache *cachep)
477{
478 return cachep->buffer_size;
479}
480EXPORT_SYMBOL(slab_buffer_size);
481#endif
482
483/*
484 * Do not go above this order unless 0 objects fit into the slab or
485 * overridden on the command line.
486 */
487#define SLAB_MAX_ORDER_HI 1
488#define SLAB_MAX_ORDER_LO 0
489static int slab_max_order = SLAB_MAX_ORDER_LO;
490static bool slab_max_order_set __initdata;
491
492/*
493 * Functions for storing/retrieving the cachep and or slab from the page
494 * allocator. These are used to find the slab an obj belongs to. With kfree(),
495 * these are used to find the cache which an obj belongs to.
496 */
497static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
498{
499 page->lru.next = (struct list_head *)cache;
500}
501
502static inline struct kmem_cache *page_get_cache(struct page *page)
503{
504 page = compound_head(page);
505 BUG_ON(!PageSlab(page));
506 return (struct kmem_cache *)page->lru.next;
507}
508
509static inline void page_set_slab(struct page *page, struct slab *slab)
510{
511 page->lru.prev = (struct list_head *)slab;
512}
513
514static inline struct slab *page_get_slab(struct page *page)
515{
516 BUG_ON(!PageSlab(page));
517 return (struct slab *)page->lru.prev;
518}
519
520static inline struct kmem_cache *virt_to_cache(const void *obj)
521{
522 struct page *page = virt_to_head_page(obj);
523 return page_get_cache(page);
524}
525
526static inline struct slab *virt_to_slab(const void *obj)
527{
528 struct page *page = virt_to_head_page(obj);
529 return page_get_slab(page);
530}
531
532static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
533 unsigned int idx)
534{
535 return slab->s_mem + cache->buffer_size * idx;
536}
537
538/*
539 * We want to avoid an expensive divide : (offset / cache->buffer_size)
540 * Using the fact that buffer_size is a constant for a particular cache,
541 * we can replace (offset / cache->buffer_size) by
542 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
543 */
544static inline unsigned int obj_to_index(const struct kmem_cache *cache,
545 const struct slab *slab, void *obj)
546{
547 u32 offset = (obj - slab->s_mem);
548 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
549}
550
551/*
552 * These are the default caches for kmalloc. Custom caches can have other sizes.
553 */
554struct cache_sizes malloc_sizes[] = {
555#define CACHE(x) { .cs_size = (x) },
556#include <linux/kmalloc_sizes.h>
557 CACHE(ULONG_MAX)
558#undef CACHE
559};
560EXPORT_SYMBOL(malloc_sizes);
561
562/* Must match cache_sizes above. Out of line to keep cache footprint low. */
563struct cache_names {
564 char *name;
565 char *name_dma;
566};
567
568static struct cache_names __initdata cache_names[] = {
569#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
570#include <linux/kmalloc_sizes.h>
571 {NULL,}
572#undef CACHE
573};
574
575static struct arraycache_init initarray_cache __initdata =
576 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
577static struct arraycache_init initarray_generic =
578 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
579
580/* internal cache of cache description objs */
581static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
582static struct kmem_cache cache_cache = {
583 .nodelists = cache_cache_nodelists,
584 .batchcount = 1,
585 .limit = BOOT_CPUCACHE_ENTRIES,
586 .shared = 1,
587 .buffer_size = sizeof(struct kmem_cache),
588 .name = "kmem_cache",
589};
590
591#define BAD_ALIEN_MAGIC 0x01020304ul
592
593/*
594 * chicken and egg problem: delay the per-cpu array allocation
595 * until the general caches are up.
596 */
597static enum {
598 NONE,
599 PARTIAL_AC,
600 PARTIAL_L3,
601 EARLY,
602 LATE,
603 FULL
604} g_cpucache_up;
605
606/*
607 * used by boot code to determine if it can use slab based allocator
608 */
609int slab_is_available(void)
610{
611 return g_cpucache_up >= EARLY;
612}
613
614#ifdef CONFIG_LOCKDEP
615
616/*
617 * Slab sometimes uses the kmalloc slabs to store the slab headers
618 * for other slabs "off slab".
619 * The locking for this is tricky in that it nests within the locks
620 * of all other slabs in a few places; to deal with this special
621 * locking we put on-slab caches into a separate lock-class.
622 *
623 * We set lock class for alien array caches which are up during init.
624 * The lock annotation will be lost if all cpus of a node goes down and
625 * then comes back up during hotplug
626 */
627static struct lock_class_key on_slab_l3_key;
628static struct lock_class_key on_slab_alc_key;
629
630static struct lock_class_key debugobj_l3_key;
631static struct lock_class_key debugobj_alc_key;
632
633static void slab_set_lock_classes(struct kmem_cache *cachep,
634 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
635 int q)
636{
637 struct array_cache **alc;
638 struct kmem_list3 *l3;
639 int r;
640
641 l3 = cachep->nodelists[q];
642 if (!l3)
643 return;
644
645 lockdep_set_class(&l3->list_lock, l3_key);
646 alc = l3->alien;
647 /*
648 * FIXME: This check for BAD_ALIEN_MAGIC
649 * should go away when common slab code is taught to
650 * work even without alien caches.
651 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
652 * for alloc_alien_cache,
653 */
654 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
655 return;
656 for_each_node(r) {
657 if (alc[r])
658 lockdep_set_class(&alc[r]->lock, alc_key);
659 }
660}
661
662static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
663{
664 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
665}
666
667static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
668{
669 int node;
670
671 for_each_online_node(node)
672 slab_set_debugobj_lock_classes_node(cachep, node);
673}
674
675static void init_node_lock_keys(int q)
676{
677 struct cache_sizes *s = malloc_sizes;
678
679 if (g_cpucache_up < LATE)
680 return;
681
682 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
683 struct kmem_list3 *l3;
684
685 l3 = s->cs_cachep->nodelists[q];
686 if (!l3 || OFF_SLAB(s->cs_cachep))
687 continue;
688
689 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
690 &on_slab_alc_key, q);
691 }
692}
693
694static inline void init_lock_keys(void)
695{
696 int node;
697
698 for_each_node(node)
699 init_node_lock_keys(node);
700}
701#else
702static void init_node_lock_keys(int q)
703{
704}
705
706static inline void init_lock_keys(void)
707{
708}
709
710static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
711{
712}
713
714static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
715{
716}
717#endif
718
719/*
720 * Guard access to the cache-chain.
721 */
722static DEFINE_MUTEX(cache_chain_mutex);
723static struct list_head cache_chain;
724
725static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
726
727static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
728{
729 return cachep->array[smp_processor_id()];
730}
731
732static inline struct kmem_cache *__find_general_cachep(size_t size,
733 gfp_t gfpflags)
734{
735 struct cache_sizes *csizep = malloc_sizes;
736
737#if DEBUG
738 /* This happens if someone tries to call
739 * kmem_cache_create(), or __kmalloc(), before
740 * the generic caches are initialized.
741 */
742 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
743#endif
744 if (!size)
745 return ZERO_SIZE_PTR;
746
747 while (size > csizep->cs_size)
748 csizep++;
749
750 /*
751 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
752 * has cs_{dma,}cachep==NULL. Thus no special case
753 * for large kmalloc calls required.
754 */
755#ifdef CONFIG_ZONE_DMA
756 if (unlikely(gfpflags & GFP_DMA))
757 return csizep->cs_dmacachep;
758#endif
759 return csizep->cs_cachep;
760}
761
762static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
763{
764 return __find_general_cachep(size, gfpflags);
765}
766
767static size_t slab_mgmt_size(size_t nr_objs, size_t align)
768{
769 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
770}
771
772/*
773 * Calculate the number of objects and left-over bytes for a given buffer size.
774 */
775static void cache_estimate(unsigned long gfporder, size_t buffer_size,
776 size_t align, int flags, size_t *left_over,
777 unsigned int *num)
778{
779 int nr_objs;
780 size_t mgmt_size;
781 size_t slab_size = PAGE_SIZE << gfporder;
782
783 /*
784 * The slab management structure can be either off the slab or
785 * on it. For the latter case, the memory allocated for a
786 * slab is used for:
787 *
788 * - The struct slab
789 * - One kmem_bufctl_t for each object
790 * - Padding to respect alignment of @align
791 * - @buffer_size bytes for each object
792 *
793 * If the slab management structure is off the slab, then the
794 * alignment will already be calculated into the size. Because
795 * the slabs are all pages aligned, the objects will be at the
796 * correct alignment when allocated.
797 */
798 if (flags & CFLGS_OFF_SLAB) {
799 mgmt_size = 0;
800 nr_objs = slab_size / buffer_size;
801
802 if (nr_objs > SLAB_LIMIT)
803 nr_objs = SLAB_LIMIT;
804 } else {
805 /*
806 * Ignore padding for the initial guess. The padding
807 * is at most @align-1 bytes, and @buffer_size is at
808 * least @align. In the worst case, this result will
809 * be one greater than the number of objects that fit
810 * into the memory allocation when taking the padding
811 * into account.
812 */
813 nr_objs = (slab_size - sizeof(struct slab)) /
814 (buffer_size + sizeof(kmem_bufctl_t));
815
816 /*
817 * This calculated number will be either the right
818 * amount, or one greater than what we want.
819 */
820 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
821 > slab_size)
822 nr_objs--;
823
824 if (nr_objs > SLAB_LIMIT)
825 nr_objs = SLAB_LIMIT;
826
827 mgmt_size = slab_mgmt_size(nr_objs, align);
828 }
829 *num = nr_objs;
830 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
831}
832
833#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
834
835static void __slab_error(const char *function, struct kmem_cache *cachep,
836 char *msg)
837{
838 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
839 function, cachep->name, msg);
840 dump_stack();
841}
842
843/*
844 * By default on NUMA we use alien caches to stage the freeing of
845 * objects allocated from other nodes. This causes massive memory
846 * inefficiencies when using fake NUMA setup to split memory into a
847 * large number of small nodes, so it can be disabled on the command
848 * line
849 */
850
851static int use_alien_caches __read_mostly = 1;
852static int __init noaliencache_setup(char *s)
853{
854 use_alien_caches = 0;
855 return 1;
856}
857__setup("noaliencache", noaliencache_setup);
858
859static int __init slab_max_order_setup(char *str)
860{
861 get_option(&str, &slab_max_order);
862 slab_max_order = slab_max_order < 0 ? 0 :
863 min(slab_max_order, MAX_ORDER - 1);
864 slab_max_order_set = true;
865
866 return 1;
867}
868__setup("slab_max_order=", slab_max_order_setup);
869
870#ifdef CONFIG_NUMA
871/*
872 * Special reaping functions for NUMA systems called from cache_reap().
873 * These take care of doing round robin flushing of alien caches (containing
874 * objects freed on different nodes from which they were allocated) and the
875 * flushing of remote pcps by calling drain_node_pages.
876 */
877static DEFINE_PER_CPU(unsigned long, slab_reap_node);
878
879static void init_reap_node(int cpu)
880{
881 int node;
882
883 node = next_node(cpu_to_mem(cpu), node_online_map);
884 if (node == MAX_NUMNODES)
885 node = first_node(node_online_map);
886
887 per_cpu(slab_reap_node, cpu) = node;
888}
889
890static void next_reap_node(void)
891{
892 int node = __this_cpu_read(slab_reap_node);
893
894 node = next_node(node, node_online_map);
895 if (unlikely(node >= MAX_NUMNODES))
896 node = first_node(node_online_map);
897 __this_cpu_write(slab_reap_node, node);
898}
899
900#else
901#define init_reap_node(cpu) do { } while (0)
902#define next_reap_node(void) do { } while (0)
903#endif
904
905/*
906 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
907 * via the workqueue/eventd.
908 * Add the CPU number into the expiration time to minimize the possibility of
909 * the CPUs getting into lockstep and contending for the global cache chain
910 * lock.
911 */
912static void __cpuinit start_cpu_timer(int cpu)
913{
914 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
915
916 /*
917 * When this gets called from do_initcalls via cpucache_init(),
918 * init_workqueues() has already run, so keventd will be setup
919 * at that time.
920 */
921 if (keventd_up() && reap_work->work.func == NULL) {
922 init_reap_node(cpu);
923 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
924 schedule_delayed_work_on(cpu, reap_work,
925 __round_jiffies_relative(HZ, cpu));
926 }
927}
928
929static struct array_cache *alloc_arraycache(int node, int entries,
930 int batchcount, gfp_t gfp)
931{
932 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
933 struct array_cache *nc = NULL;
934
935 nc = kmalloc_node(memsize, gfp, node);
936 /*
937 * The array_cache structures contain pointers to free object.
938 * However, when such objects are allocated or transferred to another
939 * cache the pointers are not cleared and they could be counted as
940 * valid references during a kmemleak scan. Therefore, kmemleak must
941 * not scan such objects.
942 */
943 kmemleak_no_scan(nc);
944 if (nc) {
945 nc->avail = 0;
946 nc->limit = entries;
947 nc->batchcount = batchcount;
948 nc->touched = 0;
949 spin_lock_init(&nc->lock);
950 }
951 return nc;
952}
953
954/*
955 * Transfer objects in one arraycache to another.
956 * Locking must be handled by the caller.
957 *
958 * Return the number of entries transferred.
959 */
960static int transfer_objects(struct array_cache *to,
961 struct array_cache *from, unsigned int max)
962{
963 /* Figure out how many entries to transfer */
964 int nr = min3(from->avail, max, to->limit - to->avail);
965
966 if (!nr)
967 return 0;
968
969 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
970 sizeof(void *) *nr);
971
972 from->avail -= nr;
973 to->avail += nr;
974 return nr;
975}
976
977#ifndef CONFIG_NUMA
978
979#define drain_alien_cache(cachep, alien) do { } while (0)
980#define reap_alien(cachep, l3) do { } while (0)
981
982static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
983{
984 return (struct array_cache **)BAD_ALIEN_MAGIC;
985}
986
987static inline void free_alien_cache(struct array_cache **ac_ptr)
988{
989}
990
991static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
992{
993 return 0;
994}
995
996static inline void *alternate_node_alloc(struct kmem_cache *cachep,
997 gfp_t flags)
998{
999 return NULL;
1000}
1001
1002static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1003 gfp_t flags, int nodeid)
1004{
1005 return NULL;
1006}
1007
1008#else /* CONFIG_NUMA */
1009
1010static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1011static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1012
1013static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1014{
1015 struct array_cache **ac_ptr;
1016 int memsize = sizeof(void *) * nr_node_ids;
1017 int i;
1018
1019 if (limit > 1)
1020 limit = 12;
1021 ac_ptr = kzalloc_node(memsize, gfp, node);
1022 if (ac_ptr) {
1023 for_each_node(i) {
1024 if (i == node || !node_online(i))
1025 continue;
1026 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1027 if (!ac_ptr[i]) {
1028 for (i--; i >= 0; i--)
1029 kfree(ac_ptr[i]);
1030 kfree(ac_ptr);
1031 return NULL;
1032 }
1033 }
1034 }
1035 return ac_ptr;
1036}
1037
1038static void free_alien_cache(struct array_cache **ac_ptr)
1039{
1040 int i;
1041
1042 if (!ac_ptr)
1043 return;
1044 for_each_node(i)
1045 kfree(ac_ptr[i]);
1046 kfree(ac_ptr);
1047}
1048
1049static void __drain_alien_cache(struct kmem_cache *cachep,
1050 struct array_cache *ac, int node)
1051{
1052 struct kmem_list3 *rl3 = cachep->nodelists[node];
1053
1054 if (ac->avail) {
1055 spin_lock(&rl3->list_lock);
1056 /*
1057 * Stuff objects into the remote nodes shared array first.
1058 * That way we could avoid the overhead of putting the objects
1059 * into the free lists and getting them back later.
1060 */
1061 if (rl3->shared)
1062 transfer_objects(rl3->shared, ac, ac->limit);
1063
1064 free_block(cachep, ac->entry, ac->avail, node);
1065 ac->avail = 0;
1066 spin_unlock(&rl3->list_lock);
1067 }
1068}
1069
1070/*
1071 * Called from cache_reap() to regularly drain alien caches round robin.
1072 */
1073static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1074{
1075 int node = __this_cpu_read(slab_reap_node);
1076
1077 if (l3->alien) {
1078 struct array_cache *ac = l3->alien[node];
1079
1080 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1081 __drain_alien_cache(cachep, ac, node);
1082 spin_unlock_irq(&ac->lock);
1083 }
1084 }
1085}
1086
1087static void drain_alien_cache(struct kmem_cache *cachep,
1088 struct array_cache **alien)
1089{
1090 int i = 0;
1091 struct array_cache *ac;
1092 unsigned long flags;
1093
1094 for_each_online_node(i) {
1095 ac = alien[i];
1096 if (ac) {
1097 spin_lock_irqsave(&ac->lock, flags);
1098 __drain_alien_cache(cachep, ac, i);
1099 spin_unlock_irqrestore(&ac->lock, flags);
1100 }
1101 }
1102}
1103
1104static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1105{
1106 struct slab *slabp = virt_to_slab(objp);
1107 int nodeid = slabp->nodeid;
1108 struct kmem_list3 *l3;
1109 struct array_cache *alien = NULL;
1110 int node;
1111
1112 node = numa_mem_id();
1113
1114 /*
1115 * Make sure we are not freeing a object from another node to the array
1116 * cache on this cpu.
1117 */
1118 if (likely(slabp->nodeid == node))
1119 return 0;
1120
1121 l3 = cachep->nodelists[node];
1122 STATS_INC_NODEFREES(cachep);
1123 if (l3->alien && l3->alien[nodeid]) {
1124 alien = l3->alien[nodeid];
1125 spin_lock(&alien->lock);
1126 if (unlikely(alien->avail == alien->limit)) {
1127 STATS_INC_ACOVERFLOW(cachep);
1128 __drain_alien_cache(cachep, alien, nodeid);
1129 }
1130 alien->entry[alien->avail++] = objp;
1131 spin_unlock(&alien->lock);
1132 } else {
1133 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1134 free_block(cachep, &objp, 1, nodeid);
1135 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1136 }
1137 return 1;
1138}
1139#endif
1140
1141/*
1142 * Allocates and initializes nodelists for a node on each slab cache, used for
1143 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1144 * will be allocated off-node since memory is not yet online for the new node.
1145 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1146 * already in use.
1147 *
1148 * Must hold cache_chain_mutex.
1149 */
1150static int init_cache_nodelists_node(int node)
1151{
1152 struct kmem_cache *cachep;
1153 struct kmem_list3 *l3;
1154 const int memsize = sizeof(struct kmem_list3);
1155
1156 list_for_each_entry(cachep, &cache_chain, next) {
1157 /*
1158 * Set up the size64 kmemlist for cpu before we can
1159 * begin anything. Make sure some other cpu on this
1160 * node has not already allocated this
1161 */
1162 if (!cachep->nodelists[node]) {
1163 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1164 if (!l3)
1165 return -ENOMEM;
1166 kmem_list3_init(l3);
1167 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1168 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1169
1170 /*
1171 * The l3s don't come and go as CPUs come and
1172 * go. cache_chain_mutex is sufficient
1173 * protection here.
1174 */
1175 cachep->nodelists[node] = l3;
1176 }
1177
1178 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1179 cachep->nodelists[node]->free_limit =
1180 (1 + nr_cpus_node(node)) *
1181 cachep->batchcount + cachep->num;
1182 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1183 }
1184 return 0;
1185}
1186
1187static void __cpuinit cpuup_canceled(long cpu)
1188{
1189 struct kmem_cache *cachep;
1190 struct kmem_list3 *l3 = NULL;
1191 int node = cpu_to_mem(cpu);
1192 const struct cpumask *mask = cpumask_of_node(node);
1193
1194 list_for_each_entry(cachep, &cache_chain, next) {
1195 struct array_cache *nc;
1196 struct array_cache *shared;
1197 struct array_cache **alien;
1198
1199 /* cpu is dead; no one can alloc from it. */
1200 nc = cachep->array[cpu];
1201 cachep->array[cpu] = NULL;
1202 l3 = cachep->nodelists[node];
1203
1204 if (!l3)
1205 goto free_array_cache;
1206
1207 spin_lock_irq(&l3->list_lock);
1208
1209 /* Free limit for this kmem_list3 */
1210 l3->free_limit -= cachep->batchcount;
1211 if (nc)
1212 free_block(cachep, nc->entry, nc->avail, node);
1213
1214 if (!cpumask_empty(mask)) {
1215 spin_unlock_irq(&l3->list_lock);
1216 goto free_array_cache;
1217 }
1218
1219 shared = l3->shared;
1220 if (shared) {
1221 free_block(cachep, shared->entry,
1222 shared->avail, node);
1223 l3->shared = NULL;
1224 }
1225
1226 alien = l3->alien;
1227 l3->alien = NULL;
1228
1229 spin_unlock_irq(&l3->list_lock);
1230
1231 kfree(shared);
1232 if (alien) {
1233 drain_alien_cache(cachep, alien);
1234 free_alien_cache(alien);
1235 }
1236free_array_cache:
1237 kfree(nc);
1238 }
1239 /*
1240 * In the previous loop, all the objects were freed to
1241 * the respective cache's slabs, now we can go ahead and
1242 * shrink each nodelist to its limit.
1243 */
1244 list_for_each_entry(cachep, &cache_chain, next) {
1245 l3 = cachep->nodelists[node];
1246 if (!l3)
1247 continue;
1248 drain_freelist(cachep, l3, l3->free_objects);
1249 }
1250}
1251
1252static int __cpuinit cpuup_prepare(long cpu)
1253{
1254 struct kmem_cache *cachep;
1255 struct kmem_list3 *l3 = NULL;
1256 int node = cpu_to_mem(cpu);
1257 int err;
1258
1259 /*
1260 * We need to do this right in the beginning since
1261 * alloc_arraycache's are going to use this list.
1262 * kmalloc_node allows us to add the slab to the right
1263 * kmem_list3 and not this cpu's kmem_list3
1264 */
1265 err = init_cache_nodelists_node(node);
1266 if (err < 0)
1267 goto bad;
1268
1269 /*
1270 * Now we can go ahead with allocating the shared arrays and
1271 * array caches
1272 */
1273 list_for_each_entry(cachep, &cache_chain, next) {
1274 struct array_cache *nc;
1275 struct array_cache *shared = NULL;
1276 struct array_cache **alien = NULL;
1277
1278 nc = alloc_arraycache(node, cachep->limit,
1279 cachep->batchcount, GFP_KERNEL);
1280 if (!nc)
1281 goto bad;
1282 if (cachep->shared) {
1283 shared = alloc_arraycache(node,
1284 cachep->shared * cachep->batchcount,
1285 0xbaadf00d, GFP_KERNEL);
1286 if (!shared) {
1287 kfree(nc);
1288 goto bad;
1289 }
1290 }
1291 if (use_alien_caches) {
1292 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1293 if (!alien) {
1294 kfree(shared);
1295 kfree(nc);
1296 goto bad;
1297 }
1298 }
1299 cachep->array[cpu] = nc;
1300 l3 = cachep->nodelists[node];
1301 BUG_ON(!l3);
1302
1303 spin_lock_irq(&l3->list_lock);
1304 if (!l3->shared) {
1305 /*
1306 * We are serialised from CPU_DEAD or
1307 * CPU_UP_CANCELLED by the cpucontrol lock
1308 */
1309 l3->shared = shared;
1310 shared = NULL;
1311 }
1312#ifdef CONFIG_NUMA
1313 if (!l3->alien) {
1314 l3->alien = alien;
1315 alien = NULL;
1316 }
1317#endif
1318 spin_unlock_irq(&l3->list_lock);
1319 kfree(shared);
1320 free_alien_cache(alien);
1321 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1322 slab_set_debugobj_lock_classes_node(cachep, node);
1323 }
1324 init_node_lock_keys(node);
1325
1326 return 0;
1327bad:
1328 cpuup_canceled(cpu);
1329 return -ENOMEM;
1330}
1331
1332static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1333 unsigned long action, void *hcpu)
1334{
1335 long cpu = (long)hcpu;
1336 int err = 0;
1337
1338 switch (action) {
1339 case CPU_UP_PREPARE:
1340 case CPU_UP_PREPARE_FROZEN:
1341 mutex_lock(&cache_chain_mutex);
1342 err = cpuup_prepare(cpu);
1343 mutex_unlock(&cache_chain_mutex);
1344 break;
1345 case CPU_ONLINE:
1346 case CPU_ONLINE_FROZEN:
1347 start_cpu_timer(cpu);
1348 break;
1349#ifdef CONFIG_HOTPLUG_CPU
1350 case CPU_DOWN_PREPARE:
1351 case CPU_DOWN_PREPARE_FROZEN:
1352 /*
1353 * Shutdown cache reaper. Note that the cache_chain_mutex is
1354 * held so that if cache_reap() is invoked it cannot do
1355 * anything expensive but will only modify reap_work
1356 * and reschedule the timer.
1357 */
1358 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1359 /* Now the cache_reaper is guaranteed to be not running. */
1360 per_cpu(slab_reap_work, cpu).work.func = NULL;
1361 break;
1362 case CPU_DOWN_FAILED:
1363 case CPU_DOWN_FAILED_FROZEN:
1364 start_cpu_timer(cpu);
1365 break;
1366 case CPU_DEAD:
1367 case CPU_DEAD_FROZEN:
1368 /*
1369 * Even if all the cpus of a node are down, we don't free the
1370 * kmem_list3 of any cache. This to avoid a race between
1371 * cpu_down, and a kmalloc allocation from another cpu for
1372 * memory from the node of the cpu going down. The list3
1373 * structure is usually allocated from kmem_cache_create() and
1374 * gets destroyed at kmem_cache_destroy().
1375 */
1376 /* fall through */
1377#endif
1378 case CPU_UP_CANCELED:
1379 case CPU_UP_CANCELED_FROZEN:
1380 mutex_lock(&cache_chain_mutex);
1381 cpuup_canceled(cpu);
1382 mutex_unlock(&cache_chain_mutex);
1383 break;
1384 }
1385 return notifier_from_errno(err);
1386}
1387
1388static struct notifier_block __cpuinitdata cpucache_notifier = {
1389 &cpuup_callback, NULL, 0
1390};
1391
1392#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1393/*
1394 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1395 * Returns -EBUSY if all objects cannot be drained so that the node is not
1396 * removed.
1397 *
1398 * Must hold cache_chain_mutex.
1399 */
1400static int __meminit drain_cache_nodelists_node(int node)
1401{
1402 struct kmem_cache *cachep;
1403 int ret = 0;
1404
1405 list_for_each_entry(cachep, &cache_chain, next) {
1406 struct kmem_list3 *l3;
1407
1408 l3 = cachep->nodelists[node];
1409 if (!l3)
1410 continue;
1411
1412 drain_freelist(cachep, l3, l3->free_objects);
1413
1414 if (!list_empty(&l3->slabs_full) ||
1415 !list_empty(&l3->slabs_partial)) {
1416 ret = -EBUSY;
1417 break;
1418 }
1419 }
1420 return ret;
1421}
1422
1423static int __meminit slab_memory_callback(struct notifier_block *self,
1424 unsigned long action, void *arg)
1425{
1426 struct memory_notify *mnb = arg;
1427 int ret = 0;
1428 int nid;
1429
1430 nid = mnb->status_change_nid;
1431 if (nid < 0)
1432 goto out;
1433
1434 switch (action) {
1435 case MEM_GOING_ONLINE:
1436 mutex_lock(&cache_chain_mutex);
1437 ret = init_cache_nodelists_node(nid);
1438 mutex_unlock(&cache_chain_mutex);
1439 break;
1440 case MEM_GOING_OFFLINE:
1441 mutex_lock(&cache_chain_mutex);
1442 ret = drain_cache_nodelists_node(nid);
1443 mutex_unlock(&cache_chain_mutex);
1444 break;
1445 case MEM_ONLINE:
1446 case MEM_OFFLINE:
1447 case MEM_CANCEL_ONLINE:
1448 case MEM_CANCEL_OFFLINE:
1449 break;
1450 }
1451out:
1452 return notifier_from_errno(ret);
1453}
1454#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1455
1456/*
1457 * swap the static kmem_list3 with kmalloced memory
1458 */
1459static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1460 int nodeid)
1461{
1462 struct kmem_list3 *ptr;
1463
1464 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1465 BUG_ON(!ptr);
1466
1467 memcpy(ptr, list, sizeof(struct kmem_list3));
1468 /*
1469 * Do not assume that spinlocks can be initialized via memcpy:
1470 */
1471 spin_lock_init(&ptr->list_lock);
1472
1473 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1474 cachep->nodelists[nodeid] = ptr;
1475}
1476
1477/*
1478 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1479 * size of kmem_list3.
1480 */
1481static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1482{
1483 int node;
1484
1485 for_each_online_node(node) {
1486 cachep->nodelists[node] = &initkmem_list3[index + node];
1487 cachep->nodelists[node]->next_reap = jiffies +
1488 REAPTIMEOUT_LIST3 +
1489 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1490 }
1491}
1492
1493/*
1494 * Initialisation. Called after the page allocator have been initialised and
1495 * before smp_init().
1496 */
1497void __init kmem_cache_init(void)
1498{
1499 size_t left_over;
1500 struct cache_sizes *sizes;
1501 struct cache_names *names;
1502 int i;
1503 int order;
1504 int node;
1505
1506 if (num_possible_nodes() == 1)
1507 use_alien_caches = 0;
1508
1509 for (i = 0; i < NUM_INIT_LISTS; i++) {
1510 kmem_list3_init(&initkmem_list3[i]);
1511 if (i < MAX_NUMNODES)
1512 cache_cache.nodelists[i] = NULL;
1513 }
1514 set_up_list3s(&cache_cache, CACHE_CACHE);
1515
1516 /*
1517 * Fragmentation resistance on low memory - only use bigger
1518 * page orders on machines with more than 32MB of memory if
1519 * not overridden on the command line.
1520 */
1521 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1522 slab_max_order = SLAB_MAX_ORDER_HI;
1523
1524 /* Bootstrap is tricky, because several objects are allocated
1525 * from caches that do not exist yet:
1526 * 1) initialize the cache_cache cache: it contains the struct
1527 * kmem_cache structures of all caches, except cache_cache itself:
1528 * cache_cache is statically allocated.
1529 * Initially an __init data area is used for the head array and the
1530 * kmem_list3 structures, it's replaced with a kmalloc allocated
1531 * array at the end of the bootstrap.
1532 * 2) Create the first kmalloc cache.
1533 * The struct kmem_cache for the new cache is allocated normally.
1534 * An __init data area is used for the head array.
1535 * 3) Create the remaining kmalloc caches, with minimally sized
1536 * head arrays.
1537 * 4) Replace the __init data head arrays for cache_cache and the first
1538 * kmalloc cache with kmalloc allocated arrays.
1539 * 5) Replace the __init data for kmem_list3 for cache_cache and
1540 * the other cache's with kmalloc allocated memory.
1541 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1542 */
1543
1544 node = numa_mem_id();
1545
1546 /* 1) create the cache_cache */
1547 INIT_LIST_HEAD(&cache_chain);
1548 list_add(&cache_cache.next, &cache_chain);
1549 cache_cache.colour_off = cache_line_size();
1550 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1551 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1552
1553 /*
1554 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1555 */
1556 cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1557 nr_node_ids * sizeof(struct kmem_list3 *);
1558#if DEBUG
1559 cache_cache.obj_size = cache_cache.buffer_size;
1560#endif
1561 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1562 cache_line_size());
1563 cache_cache.reciprocal_buffer_size =
1564 reciprocal_value(cache_cache.buffer_size);
1565
1566 for (order = 0; order < MAX_ORDER; order++) {
1567 cache_estimate(order, cache_cache.buffer_size,
1568 cache_line_size(), 0, &left_over, &cache_cache.num);
1569 if (cache_cache.num)
1570 break;
1571 }
1572 BUG_ON(!cache_cache.num);
1573 cache_cache.gfporder = order;
1574 cache_cache.colour = left_over / cache_cache.colour_off;
1575 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1576 sizeof(struct slab), cache_line_size());
1577
1578 /* 2+3) create the kmalloc caches */
1579 sizes = malloc_sizes;
1580 names = cache_names;
1581
1582 /*
1583 * Initialize the caches that provide memory for the array cache and the
1584 * kmem_list3 structures first. Without this, further allocations will
1585 * bug.
1586 */
1587
1588 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1589 sizes[INDEX_AC].cs_size,
1590 ARCH_KMALLOC_MINALIGN,
1591 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1592 NULL);
1593
1594 if (INDEX_AC != INDEX_L3) {
1595 sizes[INDEX_L3].cs_cachep =
1596 kmem_cache_create(names[INDEX_L3].name,
1597 sizes[INDEX_L3].cs_size,
1598 ARCH_KMALLOC_MINALIGN,
1599 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1600 NULL);
1601 }
1602
1603 slab_early_init = 0;
1604
1605 while (sizes->cs_size != ULONG_MAX) {
1606 /*
1607 * For performance, all the general caches are L1 aligned.
1608 * This should be particularly beneficial on SMP boxes, as it
1609 * eliminates "false sharing".
1610 * Note for systems short on memory removing the alignment will
1611 * allow tighter packing of the smaller caches.
1612 */
1613 if (!sizes->cs_cachep) {
1614 sizes->cs_cachep = kmem_cache_create(names->name,
1615 sizes->cs_size,
1616 ARCH_KMALLOC_MINALIGN,
1617 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1618 NULL);
1619 }
1620#ifdef CONFIG_ZONE_DMA
1621 sizes->cs_dmacachep = kmem_cache_create(
1622 names->name_dma,
1623 sizes->cs_size,
1624 ARCH_KMALLOC_MINALIGN,
1625 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1626 SLAB_PANIC,
1627 NULL);
1628#endif
1629 sizes++;
1630 names++;
1631 }
1632 /* 4) Replace the bootstrap head arrays */
1633 {
1634 struct array_cache *ptr;
1635
1636 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1637
1638 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1639 memcpy(ptr, cpu_cache_get(&cache_cache),
1640 sizeof(struct arraycache_init));
1641 /*
1642 * Do not assume that spinlocks can be initialized via memcpy:
1643 */
1644 spin_lock_init(&ptr->lock);
1645
1646 cache_cache.array[smp_processor_id()] = ptr;
1647
1648 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1649
1650 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1651 != &initarray_generic.cache);
1652 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1653 sizeof(struct arraycache_init));
1654 /*
1655 * Do not assume that spinlocks can be initialized via memcpy:
1656 */
1657 spin_lock_init(&ptr->lock);
1658
1659 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1660 ptr;
1661 }
1662 /* 5) Replace the bootstrap kmem_list3's */
1663 {
1664 int nid;
1665
1666 for_each_online_node(nid) {
1667 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1668
1669 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1670 &initkmem_list3[SIZE_AC + nid], nid);
1671
1672 if (INDEX_AC != INDEX_L3) {
1673 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1674 &initkmem_list3[SIZE_L3 + nid], nid);
1675 }
1676 }
1677 }
1678
1679 g_cpucache_up = EARLY;
1680}
1681
1682void __init kmem_cache_init_late(void)
1683{
1684 struct kmem_cache *cachep;
1685
1686 g_cpucache_up = LATE;
1687
1688 /* Annotate slab for lockdep -- annotate the malloc caches */
1689 init_lock_keys();
1690
1691 /* 6) resize the head arrays to their final sizes */
1692 mutex_lock(&cache_chain_mutex);
1693 list_for_each_entry(cachep, &cache_chain, next)
1694 if (enable_cpucache(cachep, GFP_NOWAIT))
1695 BUG();
1696 mutex_unlock(&cache_chain_mutex);
1697
1698 /* Done! */
1699 g_cpucache_up = FULL;
1700
1701 /*
1702 * Register a cpu startup notifier callback that initializes
1703 * cpu_cache_get for all new cpus
1704 */
1705 register_cpu_notifier(&cpucache_notifier);
1706
1707#ifdef CONFIG_NUMA
1708 /*
1709 * Register a memory hotplug callback that initializes and frees
1710 * nodelists.
1711 */
1712 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1713#endif
1714
1715 /*
1716 * The reap timers are started later, with a module init call: That part
1717 * of the kernel is not yet operational.
1718 */
1719}
1720
1721static int __init cpucache_init(void)
1722{
1723 int cpu;
1724
1725 /*
1726 * Register the timers that return unneeded pages to the page allocator
1727 */
1728 for_each_online_cpu(cpu)
1729 start_cpu_timer(cpu);
1730 return 0;
1731}
1732__initcall(cpucache_init);
1733
1734static noinline void
1735slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1736{
1737 struct kmem_list3 *l3;
1738 struct slab *slabp;
1739 unsigned long flags;
1740 int node;
1741
1742 printk(KERN_WARNING
1743 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1744 nodeid, gfpflags);
1745 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1746 cachep->name, cachep->buffer_size, cachep->gfporder);
1747
1748 for_each_online_node(node) {
1749 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1750 unsigned long active_slabs = 0, num_slabs = 0;
1751
1752 l3 = cachep->nodelists[node];
1753 if (!l3)
1754 continue;
1755
1756 spin_lock_irqsave(&l3->list_lock, flags);
1757 list_for_each_entry(slabp, &l3->slabs_full, list) {
1758 active_objs += cachep->num;
1759 active_slabs++;
1760 }
1761 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1762 active_objs += slabp->inuse;
1763 active_slabs++;
1764 }
1765 list_for_each_entry(slabp, &l3->slabs_free, list)
1766 num_slabs++;
1767
1768 free_objects += l3->free_objects;
1769 spin_unlock_irqrestore(&l3->list_lock, flags);
1770
1771 num_slabs += active_slabs;
1772 num_objs = num_slabs * cachep->num;
1773 printk(KERN_WARNING
1774 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1775 node, active_slabs, num_slabs, active_objs, num_objs,
1776 free_objects);
1777 }
1778}
1779
1780/*
1781 * Interface to system's page allocator. No need to hold the cache-lock.
1782 *
1783 * If we requested dmaable memory, we will get it. Even if we
1784 * did not request dmaable memory, we might get it, but that
1785 * would be relatively rare and ignorable.
1786 */
1787static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1788{
1789 struct page *page;
1790 int nr_pages;
1791 int i;
1792
1793#ifndef CONFIG_MMU
1794 /*
1795 * Nommu uses slab's for process anonymous memory allocations, and thus
1796 * requires __GFP_COMP to properly refcount higher order allocations
1797 */
1798 flags |= __GFP_COMP;
1799#endif
1800
1801 flags |= cachep->gfpflags;
1802 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1803 flags |= __GFP_RECLAIMABLE;
1804
1805 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1806 if (!page) {
1807 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1808 slab_out_of_memory(cachep, flags, nodeid);
1809 return NULL;
1810 }
1811
1812 nr_pages = (1 << cachep->gfporder);
1813 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1814 add_zone_page_state(page_zone(page),
1815 NR_SLAB_RECLAIMABLE, nr_pages);
1816 else
1817 add_zone_page_state(page_zone(page),
1818 NR_SLAB_UNRECLAIMABLE, nr_pages);
1819 for (i = 0; i < nr_pages; i++)
1820 __SetPageSlab(page + i);
1821
1822 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1823 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1824
1825 if (cachep->ctor)
1826 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1827 else
1828 kmemcheck_mark_unallocated_pages(page, nr_pages);
1829 }
1830
1831 return page_address(page);
1832}
1833
1834/*
1835 * Interface to system's page release.
1836 */
1837static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1838{
1839 unsigned long i = (1 << cachep->gfporder);
1840 struct page *page = virt_to_page(addr);
1841 const unsigned long nr_freed = i;
1842
1843 kmemcheck_free_shadow(page, cachep->gfporder);
1844
1845 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1846 sub_zone_page_state(page_zone(page),
1847 NR_SLAB_RECLAIMABLE, nr_freed);
1848 else
1849 sub_zone_page_state(page_zone(page),
1850 NR_SLAB_UNRECLAIMABLE, nr_freed);
1851 while (i--) {
1852 BUG_ON(!PageSlab(page));
1853 __ClearPageSlab(page);
1854 page++;
1855 }
1856 if (current->reclaim_state)
1857 current->reclaim_state->reclaimed_slab += nr_freed;
1858 free_pages((unsigned long)addr, cachep->gfporder);
1859}
1860
1861static void kmem_rcu_free(struct rcu_head *head)
1862{
1863 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1864 struct kmem_cache *cachep = slab_rcu->cachep;
1865
1866 kmem_freepages(cachep, slab_rcu->addr);
1867 if (OFF_SLAB(cachep))
1868 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1869}
1870
1871#if DEBUG
1872
1873#ifdef CONFIG_DEBUG_PAGEALLOC
1874static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1875 unsigned long caller)
1876{
1877 int size = obj_size(cachep);
1878
1879 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1880
1881 if (size < 5 * sizeof(unsigned long))
1882 return;
1883
1884 *addr++ = 0x12345678;
1885 *addr++ = caller;
1886 *addr++ = smp_processor_id();
1887 size -= 3 * sizeof(unsigned long);
1888 {
1889 unsigned long *sptr = &caller;
1890 unsigned long svalue;
1891
1892 while (!kstack_end(sptr)) {
1893 svalue = *sptr++;
1894 if (kernel_text_address(svalue)) {
1895 *addr++ = svalue;
1896 size -= sizeof(unsigned long);
1897 if (size <= sizeof(unsigned long))
1898 break;
1899 }
1900 }
1901
1902 }
1903 *addr++ = 0x87654321;
1904}
1905#endif
1906
1907static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1908{
1909 int size = obj_size(cachep);
1910 addr = &((char *)addr)[obj_offset(cachep)];
1911
1912 memset(addr, val, size);
1913 *(unsigned char *)(addr + size - 1) = POISON_END;
1914}
1915
1916static void dump_line(char *data, int offset, int limit)
1917{
1918 int i;
1919 unsigned char error = 0;
1920 int bad_count = 0;
1921
1922 printk(KERN_ERR "%03x: ", offset);
1923 for (i = 0; i < limit; i++) {
1924 if (data[offset + i] != POISON_FREE) {
1925 error = data[offset + i];
1926 bad_count++;
1927 }
1928 }
1929 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1930 &data[offset], limit, 1);
1931
1932 if (bad_count == 1) {
1933 error ^= POISON_FREE;
1934 if (!(error & (error - 1))) {
1935 printk(KERN_ERR "Single bit error detected. Probably "
1936 "bad RAM.\n");
1937#ifdef CONFIG_X86
1938 printk(KERN_ERR "Run memtest86+ or a similar memory "
1939 "test tool.\n");
1940#else
1941 printk(KERN_ERR "Run a memory test tool.\n");
1942#endif
1943 }
1944 }
1945}
1946#endif
1947
1948#if DEBUG
1949
1950static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1951{
1952 int i, size;
1953 char *realobj;
1954
1955 if (cachep->flags & SLAB_RED_ZONE) {
1956 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1957 *dbg_redzone1(cachep, objp),
1958 *dbg_redzone2(cachep, objp));
1959 }
1960
1961 if (cachep->flags & SLAB_STORE_USER) {
1962 printk(KERN_ERR "Last user: [<%p>]",
1963 *dbg_userword(cachep, objp));
1964 print_symbol("(%s)",
1965 (unsigned long)*dbg_userword(cachep, objp));
1966 printk("\n");
1967 }
1968 realobj = (char *)objp + obj_offset(cachep);
1969 size = obj_size(cachep);
1970 for (i = 0; i < size && lines; i += 16, lines--) {
1971 int limit;
1972 limit = 16;
1973 if (i + limit > size)
1974 limit = size - i;
1975 dump_line(realobj, i, limit);
1976 }
1977}
1978
1979static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1980{
1981 char *realobj;
1982 int size, i;
1983 int lines = 0;
1984
1985 realobj = (char *)objp + obj_offset(cachep);
1986 size = obj_size(cachep);
1987
1988 for (i = 0; i < size; i++) {
1989 char exp = POISON_FREE;
1990 if (i == size - 1)
1991 exp = POISON_END;
1992 if (realobj[i] != exp) {
1993 int limit;
1994 /* Mismatch ! */
1995 /* Print header */
1996 if (lines == 0) {
1997 printk(KERN_ERR
1998 "Slab corruption (%s): %s start=%p, len=%d\n",
1999 print_tainted(), cachep->name, realobj, size);
2000 print_objinfo(cachep, objp, 0);
2001 }
2002 /* Hexdump the affected line */
2003 i = (i / 16) * 16;
2004 limit = 16;
2005 if (i + limit > size)
2006 limit = size - i;
2007 dump_line(realobj, i, limit);
2008 i += 16;
2009 lines++;
2010 /* Limit to 5 lines */
2011 if (lines > 5)
2012 break;
2013 }
2014 }
2015 if (lines != 0) {
2016 /* Print some data about the neighboring objects, if they
2017 * exist:
2018 */
2019 struct slab *slabp = virt_to_slab(objp);
2020 unsigned int objnr;
2021
2022 objnr = obj_to_index(cachep, slabp, objp);
2023 if (objnr) {
2024 objp = index_to_obj(cachep, slabp, objnr - 1);
2025 realobj = (char *)objp + obj_offset(cachep);
2026 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2027 realobj, size);
2028 print_objinfo(cachep, objp, 2);
2029 }
2030 if (objnr + 1 < cachep->num) {
2031 objp = index_to_obj(cachep, slabp, objnr + 1);
2032 realobj = (char *)objp + obj_offset(cachep);
2033 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2034 realobj, size);
2035 print_objinfo(cachep, objp, 2);
2036 }
2037 }
2038}
2039#endif
2040
2041#if DEBUG
2042static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2043{
2044 int i;
2045 for (i = 0; i < cachep->num; i++) {
2046 void *objp = index_to_obj(cachep, slabp, i);
2047
2048 if (cachep->flags & SLAB_POISON) {
2049#ifdef CONFIG_DEBUG_PAGEALLOC
2050 if (cachep->buffer_size % PAGE_SIZE == 0 &&
2051 OFF_SLAB(cachep))
2052 kernel_map_pages(virt_to_page(objp),
2053 cachep->buffer_size / PAGE_SIZE, 1);
2054 else
2055 check_poison_obj(cachep, objp);
2056#else
2057 check_poison_obj(cachep, objp);
2058#endif
2059 }
2060 if (cachep->flags & SLAB_RED_ZONE) {
2061 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2062 slab_error(cachep, "start of a freed object "
2063 "was overwritten");
2064 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2065 slab_error(cachep, "end of a freed object "
2066 "was overwritten");
2067 }
2068 }
2069}
2070#else
2071static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2072{
2073}
2074#endif
2075
2076/**
2077 * slab_destroy - destroy and release all objects in a slab
2078 * @cachep: cache pointer being destroyed
2079 * @slabp: slab pointer being destroyed
2080 *
2081 * Destroy all the objs in a slab, and release the mem back to the system.
2082 * Before calling the slab must have been unlinked from the cache. The
2083 * cache-lock is not held/needed.
2084 */
2085static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2086{
2087 void *addr = slabp->s_mem - slabp->colouroff;
2088
2089 slab_destroy_debugcheck(cachep, slabp);
2090 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2091 struct slab_rcu *slab_rcu;
2092
2093 slab_rcu = (struct slab_rcu *)slabp;
2094 slab_rcu->cachep = cachep;
2095 slab_rcu->addr = addr;
2096 call_rcu(&slab_rcu->head, kmem_rcu_free);
2097 } else {
2098 kmem_freepages(cachep, addr);
2099 if (OFF_SLAB(cachep))
2100 kmem_cache_free(cachep->slabp_cache, slabp);
2101 }
2102}
2103
2104static void __kmem_cache_destroy(struct kmem_cache *cachep)
2105{
2106 int i;
2107 struct kmem_list3 *l3;
2108
2109 for_each_online_cpu(i)
2110 kfree(cachep->array[i]);
2111
2112 /* NUMA: free the list3 structures */
2113 for_each_online_node(i) {
2114 l3 = cachep->nodelists[i];
2115 if (l3) {
2116 kfree(l3->shared);
2117 free_alien_cache(l3->alien);
2118 kfree(l3);
2119 }
2120 }
2121 kmem_cache_free(&cache_cache, cachep);
2122}
2123
2124
2125/**
2126 * calculate_slab_order - calculate size (page order) of slabs
2127 * @cachep: pointer to the cache that is being created
2128 * @size: size of objects to be created in this cache.
2129 * @align: required alignment for the objects.
2130 * @flags: slab allocation flags
2131 *
2132 * Also calculates the number of objects per slab.
2133 *
2134 * This could be made much more intelligent. For now, try to avoid using
2135 * high order pages for slabs. When the gfp() functions are more friendly
2136 * towards high-order requests, this should be changed.
2137 */
2138static size_t calculate_slab_order(struct kmem_cache *cachep,
2139 size_t size, size_t align, unsigned long flags)
2140{
2141 unsigned long offslab_limit;
2142 size_t left_over = 0;
2143 int gfporder;
2144
2145 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2146 unsigned int num;
2147 size_t remainder;
2148
2149 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2150 if (!num)
2151 continue;
2152
2153 if (flags & CFLGS_OFF_SLAB) {
2154 /*
2155 * Max number of objs-per-slab for caches which
2156 * use off-slab slabs. Needed to avoid a possible
2157 * looping condition in cache_grow().
2158 */
2159 offslab_limit = size - sizeof(struct slab);
2160 offslab_limit /= sizeof(kmem_bufctl_t);
2161
2162 if (num > offslab_limit)
2163 break;
2164 }
2165
2166 /* Found something acceptable - save it away */
2167 cachep->num = num;
2168 cachep->gfporder = gfporder;
2169 left_over = remainder;
2170
2171 /*
2172 * A VFS-reclaimable slab tends to have most allocations
2173 * as GFP_NOFS and we really don't want to have to be allocating
2174 * higher-order pages when we are unable to shrink dcache.
2175 */
2176 if (flags & SLAB_RECLAIM_ACCOUNT)
2177 break;
2178
2179 /*
2180 * Large number of objects is good, but very large slabs are
2181 * currently bad for the gfp()s.
2182 */
2183 if (gfporder >= slab_max_order)
2184 break;
2185
2186 /*
2187 * Acceptable internal fragmentation?
2188 */
2189 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2190 break;
2191 }
2192 return left_over;
2193}
2194
2195static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2196{
2197 if (g_cpucache_up == FULL)
2198 return enable_cpucache(cachep, gfp);
2199
2200 if (g_cpucache_up == NONE) {
2201 /*
2202 * Note: the first kmem_cache_create must create the cache
2203 * that's used by kmalloc(24), otherwise the creation of
2204 * further caches will BUG().
2205 */
2206 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2207
2208 /*
2209 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2210 * the first cache, then we need to set up all its list3s,
2211 * otherwise the creation of further caches will BUG().
2212 */
2213 set_up_list3s(cachep, SIZE_AC);
2214 if (INDEX_AC == INDEX_L3)
2215 g_cpucache_up = PARTIAL_L3;
2216 else
2217 g_cpucache_up = PARTIAL_AC;
2218 } else {
2219 cachep->array[smp_processor_id()] =
2220 kmalloc(sizeof(struct arraycache_init), gfp);
2221
2222 if (g_cpucache_up == PARTIAL_AC) {
2223 set_up_list3s(cachep, SIZE_L3);
2224 g_cpucache_up = PARTIAL_L3;
2225 } else {
2226 int node;
2227 for_each_online_node(node) {
2228 cachep->nodelists[node] =
2229 kmalloc_node(sizeof(struct kmem_list3),
2230 gfp, node);
2231 BUG_ON(!cachep->nodelists[node]);
2232 kmem_list3_init(cachep->nodelists[node]);
2233 }
2234 }
2235 }
2236 cachep->nodelists[numa_mem_id()]->next_reap =
2237 jiffies + REAPTIMEOUT_LIST3 +
2238 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2239
2240 cpu_cache_get(cachep)->avail = 0;
2241 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2242 cpu_cache_get(cachep)->batchcount = 1;
2243 cpu_cache_get(cachep)->touched = 0;
2244 cachep->batchcount = 1;
2245 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2246 return 0;
2247}
2248
2249/**
2250 * kmem_cache_create - Create a cache.
2251 * @name: A string which is used in /proc/slabinfo to identify this cache.
2252 * @size: The size of objects to be created in this cache.
2253 * @align: The required alignment for the objects.
2254 * @flags: SLAB flags
2255 * @ctor: A constructor for the objects.
2256 *
2257 * Returns a ptr to the cache on success, NULL on failure.
2258 * Cannot be called within a int, but can be interrupted.
2259 * The @ctor is run when new pages are allocated by the cache.
2260 *
2261 * @name must be valid until the cache is destroyed. This implies that
2262 * the module calling this has to destroy the cache before getting unloaded.
2263 *
2264 * The flags are
2265 *
2266 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2267 * to catch references to uninitialised memory.
2268 *
2269 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2270 * for buffer overruns.
2271 *
2272 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2273 * cacheline. This can be beneficial if you're counting cycles as closely
2274 * as davem.
2275 */
2276struct kmem_cache *
2277kmem_cache_create (const char *name, size_t size, size_t align,
2278 unsigned long flags, void (*ctor)(void *))
2279{
2280 size_t left_over, slab_size, ralign;
2281 struct kmem_cache *cachep = NULL, *pc;
2282 gfp_t gfp;
2283
2284 /*
2285 * Sanity checks... these are all serious usage bugs.
2286 */
2287 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2288 size > KMALLOC_MAX_SIZE) {
2289 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2290 name);
2291 BUG();
2292 }
2293
2294 /*
2295 * We use cache_chain_mutex to ensure a consistent view of
2296 * cpu_online_mask as well. Please see cpuup_callback
2297 */
2298 if (slab_is_available()) {
2299 get_online_cpus();
2300 mutex_lock(&cache_chain_mutex);
2301 }
2302
2303 list_for_each_entry(pc, &cache_chain, next) {
2304 char tmp;
2305 int res;
2306
2307 /*
2308 * This happens when the module gets unloaded and doesn't
2309 * destroy its slab cache and no-one else reuses the vmalloc
2310 * area of the module. Print a warning.
2311 */
2312 res = probe_kernel_address(pc->name, tmp);
2313 if (res) {
2314 printk(KERN_ERR
2315 "SLAB: cache with size %d has lost its name\n",
2316 pc->buffer_size);
2317 continue;
2318 }
2319
2320 if (!strcmp(pc->name, name)) {
2321 printk(KERN_ERR
2322 "kmem_cache_create: duplicate cache %s\n", name);
2323 dump_stack();
2324 goto oops;
2325 }
2326 }
2327
2328#if DEBUG
2329 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2330#if FORCED_DEBUG
2331 /*
2332 * Enable redzoning and last user accounting, except for caches with
2333 * large objects, if the increased size would increase the object size
2334 * above the next power of two: caches with object sizes just above a
2335 * power of two have a significant amount of internal fragmentation.
2336 */
2337 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2338 2 * sizeof(unsigned long long)))
2339 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2340 if (!(flags & SLAB_DESTROY_BY_RCU))
2341 flags |= SLAB_POISON;
2342#endif
2343 if (flags & SLAB_DESTROY_BY_RCU)
2344 BUG_ON(flags & SLAB_POISON);
2345#endif
2346 /*
2347 * Always checks flags, a caller might be expecting debug support which
2348 * isn't available.
2349 */
2350 BUG_ON(flags & ~CREATE_MASK);
2351
2352 /*
2353 * Check that size is in terms of words. This is needed to avoid
2354 * unaligned accesses for some archs when redzoning is used, and makes
2355 * sure any on-slab bufctl's are also correctly aligned.
2356 */
2357 if (size & (BYTES_PER_WORD - 1)) {
2358 size += (BYTES_PER_WORD - 1);
2359 size &= ~(BYTES_PER_WORD - 1);
2360 }
2361
2362 /* calculate the final buffer alignment: */
2363
2364 /* 1) arch recommendation: can be overridden for debug */
2365 if (flags & SLAB_HWCACHE_ALIGN) {
2366 /*
2367 * Default alignment: as specified by the arch code. Except if
2368 * an object is really small, then squeeze multiple objects into
2369 * one cacheline.
2370 */
2371 ralign = cache_line_size();
2372 while (size <= ralign / 2)
2373 ralign /= 2;
2374 } else {
2375 ralign = BYTES_PER_WORD;
2376 }
2377
2378 /*
2379 * Redzoning and user store require word alignment or possibly larger.
2380 * Note this will be overridden by architecture or caller mandated
2381 * alignment if either is greater than BYTES_PER_WORD.
2382 */
2383 if (flags & SLAB_STORE_USER)
2384 ralign = BYTES_PER_WORD;
2385
2386 if (flags & SLAB_RED_ZONE) {
2387 ralign = REDZONE_ALIGN;
2388 /* If redzoning, ensure that the second redzone is suitably
2389 * aligned, by adjusting the object size accordingly. */
2390 size += REDZONE_ALIGN - 1;
2391 size &= ~(REDZONE_ALIGN - 1);
2392 }
2393
2394 /* 2) arch mandated alignment */
2395 if (ralign < ARCH_SLAB_MINALIGN) {
2396 ralign = ARCH_SLAB_MINALIGN;
2397 }
2398 /* 3) caller mandated alignment */
2399 if (ralign < align) {
2400 ralign = align;
2401 }
2402 /* disable debug if necessary */
2403 if (ralign > __alignof__(unsigned long long))
2404 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2405 /*
2406 * 4) Store it.
2407 */
2408 align = ralign;
2409
2410 if (slab_is_available())
2411 gfp = GFP_KERNEL;
2412 else
2413 gfp = GFP_NOWAIT;
2414
2415 /* Get cache's description obj. */
2416 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2417 if (!cachep)
2418 goto oops;
2419
2420 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2421#if DEBUG
2422 cachep->obj_size = size;
2423
2424 /*
2425 * Both debugging options require word-alignment which is calculated
2426 * into align above.
2427 */
2428 if (flags & SLAB_RED_ZONE) {
2429 /* add space for red zone words */
2430 cachep->obj_offset += sizeof(unsigned long long);
2431 size += 2 * sizeof(unsigned long long);
2432 }
2433 if (flags & SLAB_STORE_USER) {
2434 /* user store requires one word storage behind the end of
2435 * the real object. But if the second red zone needs to be
2436 * aligned to 64 bits, we must allow that much space.
2437 */
2438 if (flags & SLAB_RED_ZONE)
2439 size += REDZONE_ALIGN;
2440 else
2441 size += BYTES_PER_WORD;
2442 }
2443#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2444 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2445 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2446 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2447 size = PAGE_SIZE;
2448 }
2449#endif
2450#endif
2451
2452 /*
2453 * Determine if the slab management is 'on' or 'off' slab.
2454 * (bootstrapping cannot cope with offslab caches so don't do
2455 * it too early on. Always use on-slab management when
2456 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2457 */
2458 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2459 !(flags & SLAB_NOLEAKTRACE))
2460 /*
2461 * Size is large, assume best to place the slab management obj
2462 * off-slab (should allow better packing of objs).
2463 */
2464 flags |= CFLGS_OFF_SLAB;
2465
2466 size = ALIGN(size, align);
2467
2468 left_over = calculate_slab_order(cachep, size, align, flags);
2469
2470 if (!cachep->num) {
2471 printk(KERN_ERR
2472 "kmem_cache_create: couldn't create cache %s.\n", name);
2473 kmem_cache_free(&cache_cache, cachep);
2474 cachep = NULL;
2475 goto oops;
2476 }
2477 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2478 + sizeof(struct slab), align);
2479
2480 /*
2481 * If the slab has been placed off-slab, and we have enough space then
2482 * move it on-slab. This is at the expense of any extra colouring.
2483 */
2484 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2485 flags &= ~CFLGS_OFF_SLAB;
2486 left_over -= slab_size;
2487 }
2488
2489 if (flags & CFLGS_OFF_SLAB) {
2490 /* really off slab. No need for manual alignment */
2491 slab_size =
2492 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2493
2494#ifdef CONFIG_PAGE_POISONING
2495 /* If we're going to use the generic kernel_map_pages()
2496 * poisoning, then it's going to smash the contents of
2497 * the redzone and userword anyhow, so switch them off.
2498 */
2499 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2500 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2501#endif
2502 }
2503
2504 cachep->colour_off = cache_line_size();
2505 /* Offset must be a multiple of the alignment. */
2506 if (cachep->colour_off < align)
2507 cachep->colour_off = align;
2508 cachep->colour = left_over / cachep->colour_off;
2509 cachep->slab_size = slab_size;
2510 cachep->flags = flags;
2511 cachep->gfpflags = 0;
2512 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2513 cachep->gfpflags |= GFP_DMA;
2514 cachep->buffer_size = size;
2515 cachep->reciprocal_buffer_size = reciprocal_value(size);
2516
2517 if (flags & CFLGS_OFF_SLAB) {
2518 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2519 /*
2520 * This is a possibility for one of the malloc_sizes caches.
2521 * But since we go off slab only for object size greater than
2522 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2523 * this should not happen at all.
2524 * But leave a BUG_ON for some lucky dude.
2525 */
2526 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2527 }
2528 cachep->ctor = ctor;
2529 cachep->name = name;
2530
2531 if (setup_cpu_cache(cachep, gfp)) {
2532 __kmem_cache_destroy(cachep);
2533 cachep = NULL;
2534 goto oops;
2535 }
2536
2537 if (flags & SLAB_DEBUG_OBJECTS) {
2538 /*
2539 * Would deadlock through slab_destroy()->call_rcu()->
2540 * debug_object_activate()->kmem_cache_alloc().
2541 */
2542 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2543
2544 slab_set_debugobj_lock_classes(cachep);
2545 }
2546
2547 /* cache setup completed, link it into the list */
2548 list_add(&cachep->next, &cache_chain);
2549oops:
2550 if (!cachep && (flags & SLAB_PANIC))
2551 panic("kmem_cache_create(): failed to create slab `%s'\n",
2552 name);
2553 if (slab_is_available()) {
2554 mutex_unlock(&cache_chain_mutex);
2555 put_online_cpus();
2556 }
2557 return cachep;
2558}
2559EXPORT_SYMBOL(kmem_cache_create);
2560
2561#if DEBUG
2562static void check_irq_off(void)
2563{
2564 BUG_ON(!irqs_disabled());
2565}
2566
2567static void check_irq_on(void)
2568{
2569 BUG_ON(irqs_disabled());
2570}
2571
2572static void check_spinlock_acquired(struct kmem_cache *cachep)
2573{
2574#ifdef CONFIG_SMP
2575 check_irq_off();
2576 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2577#endif
2578}
2579
2580static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2581{
2582#ifdef CONFIG_SMP
2583 check_irq_off();
2584 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2585#endif
2586}
2587
2588#else
2589#define check_irq_off() do { } while(0)
2590#define check_irq_on() do { } while(0)
2591#define check_spinlock_acquired(x) do { } while(0)
2592#define check_spinlock_acquired_node(x, y) do { } while(0)
2593#endif
2594
2595static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2596 struct array_cache *ac,
2597 int force, int node);
2598
2599static void do_drain(void *arg)
2600{
2601 struct kmem_cache *cachep = arg;
2602 struct array_cache *ac;
2603 int node = numa_mem_id();
2604
2605 check_irq_off();
2606 ac = cpu_cache_get(cachep);
2607 spin_lock(&cachep->nodelists[node]->list_lock);
2608 free_block(cachep, ac->entry, ac->avail, node);
2609 spin_unlock(&cachep->nodelists[node]->list_lock);
2610 ac->avail = 0;
2611}
2612
2613static void drain_cpu_caches(struct kmem_cache *cachep)
2614{
2615 struct kmem_list3 *l3;
2616 int node;
2617
2618 on_each_cpu(do_drain, cachep, 1);
2619 check_irq_on();
2620 for_each_online_node(node) {
2621 l3 = cachep->nodelists[node];
2622 if (l3 && l3->alien)
2623 drain_alien_cache(cachep, l3->alien);
2624 }
2625
2626 for_each_online_node(node) {
2627 l3 = cachep->nodelists[node];
2628 if (l3)
2629 drain_array(cachep, l3, l3->shared, 1, node);
2630 }
2631}
2632
2633/*
2634 * Remove slabs from the list of free slabs.
2635 * Specify the number of slabs to drain in tofree.
2636 *
2637 * Returns the actual number of slabs released.
2638 */
2639static int drain_freelist(struct kmem_cache *cache,
2640 struct kmem_list3 *l3, int tofree)
2641{
2642 struct list_head *p;
2643 int nr_freed;
2644 struct slab *slabp;
2645
2646 nr_freed = 0;
2647 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2648
2649 spin_lock_irq(&l3->list_lock);
2650 p = l3->slabs_free.prev;
2651 if (p == &l3->slabs_free) {
2652 spin_unlock_irq(&l3->list_lock);
2653 goto out;
2654 }
2655
2656 slabp = list_entry(p, struct slab, list);
2657#if DEBUG
2658 BUG_ON(slabp->inuse);
2659#endif
2660 list_del(&slabp->list);
2661 /*
2662 * Safe to drop the lock. The slab is no longer linked
2663 * to the cache.
2664 */
2665 l3->free_objects -= cache->num;
2666 spin_unlock_irq(&l3->list_lock);
2667 slab_destroy(cache, slabp);
2668 nr_freed++;
2669 }
2670out:
2671 return nr_freed;
2672}
2673
2674/* Called with cache_chain_mutex held to protect against cpu hotplug */
2675static int __cache_shrink(struct kmem_cache *cachep)
2676{
2677 int ret = 0, i = 0;
2678 struct kmem_list3 *l3;
2679
2680 drain_cpu_caches(cachep);
2681
2682 check_irq_on();
2683 for_each_online_node(i) {
2684 l3 = cachep->nodelists[i];
2685 if (!l3)
2686 continue;
2687
2688 drain_freelist(cachep, l3, l3->free_objects);
2689
2690 ret += !list_empty(&l3->slabs_full) ||
2691 !list_empty(&l3->slabs_partial);
2692 }
2693 return (ret ? 1 : 0);
2694}
2695
2696/**
2697 * kmem_cache_shrink - Shrink a cache.
2698 * @cachep: The cache to shrink.
2699 *
2700 * Releases as many slabs as possible for a cache.
2701 * To help debugging, a zero exit status indicates all slabs were released.
2702 */
2703int kmem_cache_shrink(struct kmem_cache *cachep)
2704{
2705 int ret;
2706 BUG_ON(!cachep || in_interrupt());
2707
2708 get_online_cpus();
2709 mutex_lock(&cache_chain_mutex);
2710 ret = __cache_shrink(cachep);
2711 mutex_unlock(&cache_chain_mutex);
2712 put_online_cpus();
2713 return ret;
2714}
2715EXPORT_SYMBOL(kmem_cache_shrink);
2716
2717/**
2718 * kmem_cache_destroy - delete a cache
2719 * @cachep: the cache to destroy
2720 *
2721 * Remove a &struct kmem_cache object from the slab cache.
2722 *
2723 * It is expected this function will be called by a module when it is
2724 * unloaded. This will remove the cache completely, and avoid a duplicate
2725 * cache being allocated each time a module is loaded and unloaded, if the
2726 * module doesn't have persistent in-kernel storage across loads and unloads.
2727 *
2728 * The cache must be empty before calling this function.
2729 *
2730 * The caller must guarantee that no one will allocate memory from the cache
2731 * during the kmem_cache_destroy().
2732 */
2733void kmem_cache_destroy(struct kmem_cache *cachep)
2734{
2735 BUG_ON(!cachep || in_interrupt());
2736
2737 /* Find the cache in the chain of caches. */
2738 get_online_cpus();
2739 mutex_lock(&cache_chain_mutex);
2740 /*
2741 * the chain is never empty, cache_cache is never destroyed
2742 */
2743 list_del(&cachep->next);
2744 if (__cache_shrink(cachep)) {
2745 slab_error(cachep, "Can't free all objects");
2746 list_add(&cachep->next, &cache_chain);
2747 mutex_unlock(&cache_chain_mutex);
2748 put_online_cpus();
2749 return;
2750 }
2751
2752 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2753 rcu_barrier();
2754
2755 __kmem_cache_destroy(cachep);
2756 mutex_unlock(&cache_chain_mutex);
2757 put_online_cpus();
2758}
2759EXPORT_SYMBOL(kmem_cache_destroy);
2760
2761/*
2762 * Get the memory for a slab management obj.
2763 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2764 * always come from malloc_sizes caches. The slab descriptor cannot
2765 * come from the same cache which is getting created because,
2766 * when we are searching for an appropriate cache for these
2767 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2768 * If we are creating a malloc_sizes cache here it would not be visible to
2769 * kmem_find_general_cachep till the initialization is complete.
2770 * Hence we cannot have slabp_cache same as the original cache.
2771 */
2772static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2773 int colour_off, gfp_t local_flags,
2774 int nodeid)
2775{
2776 struct slab *slabp;
2777
2778 if (OFF_SLAB(cachep)) {
2779 /* Slab management obj is off-slab. */
2780 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2781 local_flags, nodeid);
2782 /*
2783 * If the first object in the slab is leaked (it's allocated
2784 * but no one has a reference to it), we want to make sure
2785 * kmemleak does not treat the ->s_mem pointer as a reference
2786 * to the object. Otherwise we will not report the leak.
2787 */
2788 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2789 local_flags);
2790 if (!slabp)
2791 return NULL;
2792 } else {
2793 slabp = objp + colour_off;
2794 colour_off += cachep->slab_size;
2795 }
2796 slabp->inuse = 0;
2797 slabp->colouroff = colour_off;
2798 slabp->s_mem = objp + colour_off;
2799 slabp->nodeid = nodeid;
2800 slabp->free = 0;
2801 return slabp;
2802}
2803
2804static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2805{
2806 return (kmem_bufctl_t *) (slabp + 1);
2807}
2808
2809static void cache_init_objs(struct kmem_cache *cachep,
2810 struct slab *slabp)
2811{
2812 int i;
2813
2814 for (i = 0; i < cachep->num; i++) {
2815 void *objp = index_to_obj(cachep, slabp, i);
2816#if DEBUG
2817 /* need to poison the objs? */
2818 if (cachep->flags & SLAB_POISON)
2819 poison_obj(cachep, objp, POISON_FREE);
2820 if (cachep->flags & SLAB_STORE_USER)
2821 *dbg_userword(cachep, objp) = NULL;
2822
2823 if (cachep->flags & SLAB_RED_ZONE) {
2824 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2825 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2826 }
2827 /*
2828 * Constructors are not allowed to allocate memory from the same
2829 * cache which they are a constructor for. Otherwise, deadlock.
2830 * They must also be threaded.
2831 */
2832 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2833 cachep->ctor(objp + obj_offset(cachep));
2834
2835 if (cachep->flags & SLAB_RED_ZONE) {
2836 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2837 slab_error(cachep, "constructor overwrote the"
2838 " end of an object");
2839 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2840 slab_error(cachep, "constructor overwrote the"
2841 " start of an object");
2842 }
2843 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2844 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2845 kernel_map_pages(virt_to_page(objp),
2846 cachep->buffer_size / PAGE_SIZE, 0);
2847#else
2848 if (cachep->ctor)
2849 cachep->ctor(objp);
2850#endif
2851 slab_bufctl(slabp)[i] = i + 1;
2852 }
2853 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2854}
2855
2856static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2857{
2858 if (CONFIG_ZONE_DMA_FLAG) {
2859 if (flags & GFP_DMA)
2860 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2861 else
2862 BUG_ON(cachep->gfpflags & GFP_DMA);
2863 }
2864}
2865
2866static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2867 int nodeid)
2868{
2869 void *objp = index_to_obj(cachep, slabp, slabp->free);
2870 kmem_bufctl_t next;
2871
2872 slabp->inuse++;
2873 next = slab_bufctl(slabp)[slabp->free];
2874#if DEBUG
2875 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2876 WARN_ON(slabp->nodeid != nodeid);
2877#endif
2878 slabp->free = next;
2879
2880 return objp;
2881}
2882
2883static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2884 void *objp, int nodeid)
2885{
2886 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2887
2888#if DEBUG
2889 /* Verify that the slab belongs to the intended node */
2890 WARN_ON(slabp->nodeid != nodeid);
2891
2892 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2893 printk(KERN_ERR "slab: double free detected in cache "
2894 "'%s', objp %p\n", cachep->name, objp);
2895 BUG();
2896 }
2897#endif
2898 slab_bufctl(slabp)[objnr] = slabp->free;
2899 slabp->free = objnr;
2900 slabp->inuse--;
2901}
2902
2903/*
2904 * Map pages beginning at addr to the given cache and slab. This is required
2905 * for the slab allocator to be able to lookup the cache and slab of a
2906 * virtual address for kfree, ksize, and slab debugging.
2907 */
2908static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2909 void *addr)
2910{
2911 int nr_pages;
2912 struct page *page;
2913
2914 page = virt_to_page(addr);
2915
2916 nr_pages = 1;
2917 if (likely(!PageCompound(page)))
2918 nr_pages <<= cache->gfporder;
2919
2920 do {
2921 page_set_cache(page, cache);
2922 page_set_slab(page, slab);
2923 page++;
2924 } while (--nr_pages);
2925}
2926
2927/*
2928 * Grow (by 1) the number of slabs within a cache. This is called by
2929 * kmem_cache_alloc() when there are no active objs left in a cache.
2930 */
2931static int cache_grow(struct kmem_cache *cachep,
2932 gfp_t flags, int nodeid, void *objp)
2933{
2934 struct slab *slabp;
2935 size_t offset;
2936 gfp_t local_flags;
2937 struct kmem_list3 *l3;
2938
2939 /*
2940 * Be lazy and only check for valid flags here, keeping it out of the
2941 * critical path in kmem_cache_alloc().
2942 */
2943 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2944 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2945
2946 /* Take the l3 list lock to change the colour_next on this node */
2947 check_irq_off();
2948 l3 = cachep->nodelists[nodeid];
2949 spin_lock(&l3->list_lock);
2950
2951 /* Get colour for the slab, and cal the next value. */
2952 offset = l3->colour_next;
2953 l3->colour_next++;
2954 if (l3->colour_next >= cachep->colour)
2955 l3->colour_next = 0;
2956 spin_unlock(&l3->list_lock);
2957
2958 offset *= cachep->colour_off;
2959
2960 if (local_flags & __GFP_WAIT)
2961 local_irq_enable();
2962
2963 /*
2964 * The test for missing atomic flag is performed here, rather than
2965 * the more obvious place, simply to reduce the critical path length
2966 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2967 * will eventually be caught here (where it matters).
2968 */
2969 kmem_flagcheck(cachep, flags);
2970
2971 /*
2972 * Get mem for the objs. Attempt to allocate a physical page from
2973 * 'nodeid'.
2974 */
2975 if (!objp)
2976 objp = kmem_getpages(cachep, local_flags, nodeid);
2977 if (!objp)
2978 goto failed;
2979
2980 /* Get slab management. */
2981 slabp = alloc_slabmgmt(cachep, objp, offset,
2982 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2983 if (!slabp)
2984 goto opps1;
2985
2986 slab_map_pages(cachep, slabp, objp);
2987
2988 cache_init_objs(cachep, slabp);
2989
2990 if (local_flags & __GFP_WAIT)
2991 local_irq_disable();
2992 check_irq_off();
2993 spin_lock(&l3->list_lock);
2994
2995 /* Make slab active. */
2996 list_add_tail(&slabp->list, &(l3->slabs_free));
2997 STATS_INC_GROWN(cachep);
2998 l3->free_objects += cachep->num;
2999 spin_unlock(&l3->list_lock);
3000 return 1;
3001opps1:
3002 kmem_freepages(cachep, objp);
3003failed:
3004 if (local_flags & __GFP_WAIT)
3005 local_irq_disable();
3006 return 0;
3007}
3008
3009#if DEBUG
3010
3011/*
3012 * Perform extra freeing checks:
3013 * - detect bad pointers.
3014 * - POISON/RED_ZONE checking
3015 */
3016static void kfree_debugcheck(const void *objp)
3017{
3018 if (!virt_addr_valid(objp)) {
3019 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3020 (unsigned long)objp);
3021 BUG();
3022 }
3023}
3024
3025static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3026{
3027 unsigned long long redzone1, redzone2;
3028
3029 redzone1 = *dbg_redzone1(cache, obj);
3030 redzone2 = *dbg_redzone2(cache, obj);
3031
3032 /*
3033 * Redzone is ok.
3034 */
3035 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3036 return;
3037
3038 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3039 slab_error(cache, "double free detected");
3040 else
3041 slab_error(cache, "memory outside object was overwritten");
3042
3043 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3044 obj, redzone1, redzone2);
3045}
3046
3047static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3048 void *caller)
3049{
3050 struct page *page;
3051 unsigned int objnr;
3052 struct slab *slabp;
3053
3054 BUG_ON(virt_to_cache(objp) != cachep);
3055
3056 objp -= obj_offset(cachep);
3057 kfree_debugcheck(objp);
3058 page = virt_to_head_page(objp);
3059
3060 slabp = page_get_slab(page);
3061
3062 if (cachep->flags & SLAB_RED_ZONE) {
3063 verify_redzone_free(cachep, objp);
3064 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3065 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3066 }
3067 if (cachep->flags & SLAB_STORE_USER)
3068 *dbg_userword(cachep, objp) = caller;
3069
3070 objnr = obj_to_index(cachep, slabp, objp);
3071
3072 BUG_ON(objnr >= cachep->num);
3073 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3074
3075#ifdef CONFIG_DEBUG_SLAB_LEAK
3076 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3077#endif
3078 if (cachep->flags & SLAB_POISON) {
3079#ifdef CONFIG_DEBUG_PAGEALLOC
3080 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3081 store_stackinfo(cachep, objp, (unsigned long)caller);
3082 kernel_map_pages(virt_to_page(objp),
3083 cachep->buffer_size / PAGE_SIZE, 0);
3084 } else {
3085 poison_obj(cachep, objp, POISON_FREE);
3086 }
3087#else
3088 poison_obj(cachep, objp, POISON_FREE);
3089#endif
3090 }
3091 return objp;
3092}
3093
3094static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3095{
3096 kmem_bufctl_t i;
3097 int entries = 0;
3098
3099 /* Check slab's freelist to see if this obj is there. */
3100 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3101 entries++;
3102 if (entries > cachep->num || i >= cachep->num)
3103 goto bad;
3104 }
3105 if (entries != cachep->num - slabp->inuse) {
3106bad:
3107 printk(KERN_ERR "slab: Internal list corruption detected in "
3108 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3109 cachep->name, cachep->num, slabp, slabp->inuse,
3110 print_tainted());
3111 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3112 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3113 1);
3114 BUG();
3115 }
3116}
3117#else
3118#define kfree_debugcheck(x) do { } while(0)
3119#define cache_free_debugcheck(x,objp,z) (objp)
3120#define check_slabp(x,y) do { } while(0)
3121#endif
3122
3123static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3124{
3125 int batchcount;
3126 struct kmem_list3 *l3;
3127 struct array_cache *ac;
3128 int node;
3129
3130retry:
3131 check_irq_off();
3132 node = numa_mem_id();
3133 ac = cpu_cache_get(cachep);
3134 batchcount = ac->batchcount;
3135 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3136 /*
3137 * If there was little recent activity on this cache, then
3138 * perform only a partial refill. Otherwise we could generate
3139 * refill bouncing.
3140 */
3141 batchcount = BATCHREFILL_LIMIT;
3142 }
3143 l3 = cachep->nodelists[node];
3144
3145 BUG_ON(ac->avail > 0 || !l3);
3146 spin_lock(&l3->list_lock);
3147
3148 /* See if we can refill from the shared array */
3149 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3150 l3->shared->touched = 1;
3151 goto alloc_done;
3152 }
3153
3154 while (batchcount > 0) {
3155 struct list_head *entry;
3156 struct slab *slabp;
3157 /* Get slab alloc is to come from. */
3158 entry = l3->slabs_partial.next;
3159 if (entry == &l3->slabs_partial) {
3160 l3->free_touched = 1;
3161 entry = l3->slabs_free.next;
3162 if (entry == &l3->slabs_free)
3163 goto must_grow;
3164 }
3165
3166 slabp = list_entry(entry, struct slab, list);
3167 check_slabp(cachep, slabp);
3168 check_spinlock_acquired(cachep);
3169
3170 /*
3171 * The slab was either on partial or free list so
3172 * there must be at least one object available for
3173 * allocation.
3174 */
3175 BUG_ON(slabp->inuse >= cachep->num);
3176
3177 while (slabp->inuse < cachep->num && batchcount--) {
3178 STATS_INC_ALLOCED(cachep);
3179 STATS_INC_ACTIVE(cachep);
3180 STATS_SET_HIGH(cachep);
3181
3182 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3183 node);
3184 }
3185 check_slabp(cachep, slabp);
3186
3187 /* move slabp to correct slabp list: */
3188 list_del(&slabp->list);
3189 if (slabp->free == BUFCTL_END)
3190 list_add(&slabp->list, &l3->slabs_full);
3191 else
3192 list_add(&slabp->list, &l3->slabs_partial);
3193 }
3194
3195must_grow:
3196 l3->free_objects -= ac->avail;
3197alloc_done:
3198 spin_unlock(&l3->list_lock);
3199
3200 if (unlikely(!ac->avail)) {
3201 int x;
3202 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3203
3204 /* cache_grow can reenable interrupts, then ac could change. */
3205 ac = cpu_cache_get(cachep);
3206 if (!x && ac->avail == 0) /* no objects in sight? abort */
3207 return NULL;
3208
3209 if (!ac->avail) /* objects refilled by interrupt? */
3210 goto retry;
3211 }
3212 ac->touched = 1;
3213 return ac->entry[--ac->avail];
3214}
3215
3216static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3217 gfp_t flags)
3218{
3219 might_sleep_if(flags & __GFP_WAIT);
3220#if DEBUG
3221 kmem_flagcheck(cachep, flags);
3222#endif
3223}
3224
3225#if DEBUG
3226static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3227 gfp_t flags, void *objp, void *caller)
3228{
3229 if (!objp)
3230 return objp;
3231 if (cachep->flags & SLAB_POISON) {
3232#ifdef CONFIG_DEBUG_PAGEALLOC
3233 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3234 kernel_map_pages(virt_to_page(objp),
3235 cachep->buffer_size / PAGE_SIZE, 1);
3236 else
3237 check_poison_obj(cachep, objp);
3238#else
3239 check_poison_obj(cachep, objp);
3240#endif
3241 poison_obj(cachep, objp, POISON_INUSE);
3242 }
3243 if (cachep->flags & SLAB_STORE_USER)
3244 *dbg_userword(cachep, objp) = caller;
3245
3246 if (cachep->flags & SLAB_RED_ZONE) {
3247 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3248 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3249 slab_error(cachep, "double free, or memory outside"
3250 " object was overwritten");
3251 printk(KERN_ERR
3252 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3253 objp, *dbg_redzone1(cachep, objp),
3254 *dbg_redzone2(cachep, objp));
3255 }
3256 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3257 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3258 }
3259#ifdef CONFIG_DEBUG_SLAB_LEAK
3260 {
3261 struct slab *slabp;
3262 unsigned objnr;
3263
3264 slabp = page_get_slab(virt_to_head_page(objp));
3265 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3266 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3267 }
3268#endif
3269 objp += obj_offset(cachep);
3270 if (cachep->ctor && cachep->flags & SLAB_POISON)
3271 cachep->ctor(objp);
3272 if (ARCH_SLAB_MINALIGN &&
3273 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3274 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3275 objp, (int)ARCH_SLAB_MINALIGN);
3276 }
3277 return objp;
3278}
3279#else
3280#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3281#endif
3282
3283static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3284{
3285 if (cachep == &cache_cache)
3286 return false;
3287
3288 return should_failslab(obj_size(cachep), flags, cachep->flags);
3289}
3290
3291static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3292{
3293 void *objp;
3294 struct array_cache *ac;
3295
3296 check_irq_off();
3297
3298 ac = cpu_cache_get(cachep);
3299 if (likely(ac->avail)) {
3300 STATS_INC_ALLOCHIT(cachep);
3301 ac->touched = 1;
3302 objp = ac->entry[--ac->avail];
3303 } else {
3304 STATS_INC_ALLOCMISS(cachep);
3305 objp = cache_alloc_refill(cachep, flags);
3306 /*
3307 * the 'ac' may be updated by cache_alloc_refill(),
3308 * and kmemleak_erase() requires its correct value.
3309 */
3310 ac = cpu_cache_get(cachep);
3311 }
3312 /*
3313 * To avoid a false negative, if an object that is in one of the
3314 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3315 * treat the array pointers as a reference to the object.
3316 */
3317 if (objp)
3318 kmemleak_erase(&ac->entry[ac->avail]);
3319 return objp;
3320}
3321
3322#ifdef CONFIG_NUMA
3323/*
3324 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3325 *
3326 * If we are in_interrupt, then process context, including cpusets and
3327 * mempolicy, may not apply and should not be used for allocation policy.
3328 */
3329static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3330{
3331 int nid_alloc, nid_here;
3332
3333 if (in_interrupt() || (flags & __GFP_THISNODE))
3334 return NULL;
3335 nid_alloc = nid_here = numa_mem_id();
3336 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3337 nid_alloc = cpuset_slab_spread_node();
3338 else if (current->mempolicy)
3339 nid_alloc = slab_node(current->mempolicy);
3340 if (nid_alloc != nid_here)
3341 return ____cache_alloc_node(cachep, flags, nid_alloc);
3342 return NULL;
3343}
3344
3345/*
3346 * Fallback function if there was no memory available and no objects on a
3347 * certain node and fall back is permitted. First we scan all the
3348 * available nodelists for available objects. If that fails then we
3349 * perform an allocation without specifying a node. This allows the page
3350 * allocator to do its reclaim / fallback magic. We then insert the
3351 * slab into the proper nodelist and then allocate from it.
3352 */
3353static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3354{
3355 struct zonelist *zonelist;
3356 gfp_t local_flags;
3357 struct zoneref *z;
3358 struct zone *zone;
3359 enum zone_type high_zoneidx = gfp_zone(flags);
3360 void *obj = NULL;
3361 int nid;
3362 unsigned int cpuset_mems_cookie;
3363
3364 if (flags & __GFP_THISNODE)
3365 return NULL;
3366
3367 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3368
3369retry_cpuset:
3370 cpuset_mems_cookie = get_mems_allowed();
3371 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3372
3373retry:
3374 /*
3375 * Look through allowed nodes for objects available
3376 * from existing per node queues.
3377 */
3378 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3379 nid = zone_to_nid(zone);
3380
3381 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3382 cache->nodelists[nid] &&
3383 cache->nodelists[nid]->free_objects) {
3384 obj = ____cache_alloc_node(cache,
3385 flags | GFP_THISNODE, nid);
3386 if (obj)
3387 break;
3388 }
3389 }
3390
3391 if (!obj) {
3392 /*
3393 * This allocation will be performed within the constraints
3394 * of the current cpuset / memory policy requirements.
3395 * We may trigger various forms of reclaim on the allowed
3396 * set and go into memory reserves if necessary.
3397 */
3398 if (local_flags & __GFP_WAIT)
3399 local_irq_enable();
3400 kmem_flagcheck(cache, flags);
3401 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3402 if (local_flags & __GFP_WAIT)
3403 local_irq_disable();
3404 if (obj) {
3405 /*
3406 * Insert into the appropriate per node queues
3407 */
3408 nid = page_to_nid(virt_to_page(obj));
3409 if (cache_grow(cache, flags, nid, obj)) {
3410 obj = ____cache_alloc_node(cache,
3411 flags | GFP_THISNODE, nid);
3412 if (!obj)
3413 /*
3414 * Another processor may allocate the
3415 * objects in the slab since we are
3416 * not holding any locks.
3417 */
3418 goto retry;
3419 } else {
3420 /* cache_grow already freed obj */
3421 obj = NULL;
3422 }
3423 }
3424 }
3425
3426 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3427 goto retry_cpuset;
3428 return obj;
3429}
3430
3431/*
3432 * A interface to enable slab creation on nodeid
3433 */
3434static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3435 int nodeid)
3436{
3437 struct list_head *entry;
3438 struct slab *slabp;
3439 struct kmem_list3 *l3;
3440 void *obj;
3441 int x;
3442
3443 l3 = cachep->nodelists[nodeid];
3444 BUG_ON(!l3);
3445
3446retry:
3447 check_irq_off();
3448 spin_lock(&l3->list_lock);
3449 entry = l3->slabs_partial.next;
3450 if (entry == &l3->slabs_partial) {
3451 l3->free_touched = 1;
3452 entry = l3->slabs_free.next;
3453 if (entry == &l3->slabs_free)
3454 goto must_grow;
3455 }
3456
3457 slabp = list_entry(entry, struct slab, list);
3458 check_spinlock_acquired_node(cachep, nodeid);
3459 check_slabp(cachep, slabp);
3460
3461 STATS_INC_NODEALLOCS(cachep);
3462 STATS_INC_ACTIVE(cachep);
3463 STATS_SET_HIGH(cachep);
3464
3465 BUG_ON(slabp->inuse == cachep->num);
3466
3467 obj = slab_get_obj(cachep, slabp, nodeid);
3468 check_slabp(cachep, slabp);
3469 l3->free_objects--;
3470 /* move slabp to correct slabp list: */
3471 list_del(&slabp->list);
3472
3473 if (slabp->free == BUFCTL_END)
3474 list_add(&slabp->list, &l3->slabs_full);
3475 else
3476 list_add(&slabp->list, &l3->slabs_partial);
3477
3478 spin_unlock(&l3->list_lock);
3479 goto done;
3480
3481must_grow:
3482 spin_unlock(&l3->list_lock);
3483 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3484 if (x)
3485 goto retry;
3486
3487 return fallback_alloc(cachep, flags);
3488
3489done:
3490 return obj;
3491}
3492
3493/**
3494 * kmem_cache_alloc_node - Allocate an object on the specified node
3495 * @cachep: The cache to allocate from.
3496 * @flags: See kmalloc().
3497 * @nodeid: node number of the target node.
3498 * @caller: return address of caller, used for debug information
3499 *
3500 * Identical to kmem_cache_alloc but it will allocate memory on the given
3501 * node, which can improve the performance for cpu bound structures.
3502 *
3503 * Fallback to other node is possible if __GFP_THISNODE is not set.
3504 */
3505static __always_inline void *
3506__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3507 void *caller)
3508{
3509 unsigned long save_flags;
3510 void *ptr;
3511 int slab_node = numa_mem_id();
3512
3513 flags &= gfp_allowed_mask;
3514
3515 lockdep_trace_alloc(flags);
3516
3517 if (slab_should_failslab(cachep, flags))
3518 return NULL;
3519
3520 cache_alloc_debugcheck_before(cachep, flags);
3521 local_irq_save(save_flags);
3522
3523 if (nodeid == NUMA_NO_NODE)
3524 nodeid = slab_node;
3525
3526 if (unlikely(!cachep->nodelists[nodeid])) {
3527 /* Node not bootstrapped yet */
3528 ptr = fallback_alloc(cachep, flags);
3529 goto out;
3530 }
3531
3532 if (nodeid == slab_node) {
3533 /*
3534 * Use the locally cached objects if possible.
3535 * However ____cache_alloc does not allow fallback
3536 * to other nodes. It may fail while we still have
3537 * objects on other nodes available.
3538 */
3539 ptr = ____cache_alloc(cachep, flags);
3540 if (ptr)
3541 goto out;
3542 }
3543 /* ___cache_alloc_node can fall back to other nodes */
3544 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3545 out:
3546 local_irq_restore(save_flags);
3547 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3548 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3549 flags);
3550
3551 if (likely(ptr))
3552 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3553
3554 if (unlikely((flags & __GFP_ZERO) && ptr))
3555 memset(ptr, 0, obj_size(cachep));
3556
3557 return ptr;
3558}
3559
3560static __always_inline void *
3561__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3562{
3563 void *objp;
3564
3565 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3566 objp = alternate_node_alloc(cache, flags);
3567 if (objp)
3568 goto out;
3569 }
3570 objp = ____cache_alloc(cache, flags);
3571
3572 /*
3573 * We may just have run out of memory on the local node.
3574 * ____cache_alloc_node() knows how to locate memory on other nodes
3575 */
3576 if (!objp)
3577 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3578
3579 out:
3580 return objp;
3581}
3582#else
3583
3584static __always_inline void *
3585__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3586{
3587 return ____cache_alloc(cachep, flags);
3588}
3589
3590#endif /* CONFIG_NUMA */
3591
3592static __always_inline void *
3593__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3594{
3595 unsigned long save_flags;
3596 void *objp;
3597
3598 flags &= gfp_allowed_mask;
3599
3600 lockdep_trace_alloc(flags);
3601
3602 if (slab_should_failslab(cachep, flags))
3603 return NULL;
3604
3605 cache_alloc_debugcheck_before(cachep, flags);
3606 local_irq_save(save_flags);
3607 objp = __do_cache_alloc(cachep, flags);
3608 local_irq_restore(save_flags);
3609 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3610 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3611 flags);
3612 prefetchw(objp);
3613
3614 if (likely(objp))
3615 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3616
3617 if (unlikely((flags & __GFP_ZERO) && objp))
3618 memset(objp, 0, obj_size(cachep));
3619
3620 return objp;
3621}
3622
3623/*
3624 * Caller needs to acquire correct kmem_list's list_lock
3625 */
3626static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3627 int node)
3628{
3629 int i;
3630 struct kmem_list3 *l3;
3631
3632 for (i = 0; i < nr_objects; i++) {
3633 void *objp = objpp[i];
3634 struct slab *slabp;
3635
3636 slabp = virt_to_slab(objp);
3637 l3 = cachep->nodelists[node];
3638 list_del(&slabp->list);
3639 check_spinlock_acquired_node(cachep, node);
3640 check_slabp(cachep, slabp);
3641 slab_put_obj(cachep, slabp, objp, node);
3642 STATS_DEC_ACTIVE(cachep);
3643 l3->free_objects++;
3644 check_slabp(cachep, slabp);
3645
3646 /* fixup slab chains */
3647 if (slabp->inuse == 0) {
3648 if (l3->free_objects > l3->free_limit) {
3649 l3->free_objects -= cachep->num;
3650 /* No need to drop any previously held
3651 * lock here, even if we have a off-slab slab
3652 * descriptor it is guaranteed to come from
3653 * a different cache, refer to comments before
3654 * alloc_slabmgmt.
3655 */
3656 slab_destroy(cachep, slabp);
3657 } else {
3658 list_add(&slabp->list, &l3->slabs_free);
3659 }
3660 } else {
3661 /* Unconditionally move a slab to the end of the
3662 * partial list on free - maximum time for the
3663 * other objects to be freed, too.
3664 */
3665 list_add_tail(&slabp->list, &l3->slabs_partial);
3666 }
3667 }
3668}
3669
3670static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3671{
3672 int batchcount;
3673 struct kmem_list3 *l3;
3674 int node = numa_mem_id();
3675
3676 batchcount = ac->batchcount;
3677#if DEBUG
3678 BUG_ON(!batchcount || batchcount > ac->avail);
3679#endif
3680 check_irq_off();
3681 l3 = cachep->nodelists[node];
3682 spin_lock(&l3->list_lock);
3683 if (l3->shared) {
3684 struct array_cache *shared_array = l3->shared;
3685 int max = shared_array->limit - shared_array->avail;
3686 if (max) {
3687 if (batchcount > max)
3688 batchcount = max;
3689 memcpy(&(shared_array->entry[shared_array->avail]),
3690 ac->entry, sizeof(void *) * batchcount);
3691 shared_array->avail += batchcount;
3692 goto free_done;
3693 }
3694 }
3695
3696 free_block(cachep, ac->entry, batchcount, node);
3697free_done:
3698#if STATS
3699 {
3700 int i = 0;
3701 struct list_head *p;
3702
3703 p = l3->slabs_free.next;
3704 while (p != &(l3->slabs_free)) {
3705 struct slab *slabp;
3706
3707 slabp = list_entry(p, struct slab, list);
3708 BUG_ON(slabp->inuse);
3709
3710 i++;
3711 p = p->next;
3712 }
3713 STATS_SET_FREEABLE(cachep, i);
3714 }
3715#endif
3716 spin_unlock(&l3->list_lock);
3717 ac->avail -= batchcount;
3718 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3719}
3720
3721/*
3722 * Release an obj back to its cache. If the obj has a constructed state, it must
3723 * be in this state _before_ it is released. Called with disabled ints.
3724 */
3725static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3726 void *caller)
3727{
3728 struct array_cache *ac = cpu_cache_get(cachep);
3729
3730 check_irq_off();
3731 kmemleak_free_recursive(objp, cachep->flags);
3732 objp = cache_free_debugcheck(cachep, objp, caller);
3733
3734 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3735
3736 /*
3737 * Skip calling cache_free_alien() when the platform is not numa.
3738 * This will avoid cache misses that happen while accessing slabp (which
3739 * is per page memory reference) to get nodeid. Instead use a global
3740 * variable to skip the call, which is mostly likely to be present in
3741 * the cache.
3742 */
3743 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3744 return;
3745
3746 if (likely(ac->avail < ac->limit)) {
3747 STATS_INC_FREEHIT(cachep);
3748 } else {
3749 STATS_INC_FREEMISS(cachep);
3750 cache_flusharray(cachep, ac);
3751 }
3752
3753 ac->entry[ac->avail++] = objp;
3754}
3755
3756/**
3757 * kmem_cache_alloc - Allocate an object
3758 * @cachep: The cache to allocate from.
3759 * @flags: See kmalloc().
3760 *
3761 * Allocate an object from this cache. The flags are only relevant
3762 * if the cache has no available objects.
3763 */
3764void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3765{
3766 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3767
3768 trace_kmem_cache_alloc(_RET_IP_, ret,
3769 obj_size(cachep), cachep->buffer_size, flags);
3770
3771 return ret;
3772}
3773EXPORT_SYMBOL(kmem_cache_alloc);
3774
3775#ifdef CONFIG_TRACING
3776void *
3777kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3778{
3779 void *ret;
3780
3781 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3782
3783 trace_kmalloc(_RET_IP_, ret,
3784 size, slab_buffer_size(cachep), flags);
3785 return ret;
3786}
3787EXPORT_SYMBOL(kmem_cache_alloc_trace);
3788#endif
3789
3790#ifdef CONFIG_NUMA
3791void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3792{
3793 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3794 __builtin_return_address(0));
3795
3796 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3797 obj_size(cachep), cachep->buffer_size,
3798 flags, nodeid);
3799
3800 return ret;
3801}
3802EXPORT_SYMBOL(kmem_cache_alloc_node);
3803
3804#ifdef CONFIG_TRACING
3805void *kmem_cache_alloc_node_trace(size_t size,
3806 struct kmem_cache *cachep,
3807 gfp_t flags,
3808 int nodeid)
3809{
3810 void *ret;
3811
3812 ret = __cache_alloc_node(cachep, flags, nodeid,
3813 __builtin_return_address(0));
3814 trace_kmalloc_node(_RET_IP_, ret,
3815 size, slab_buffer_size(cachep),
3816 flags, nodeid);
3817 return ret;
3818}
3819EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3820#endif
3821
3822static __always_inline void *
3823__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3824{
3825 struct kmem_cache *cachep;
3826
3827 cachep = kmem_find_general_cachep(size, flags);
3828 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3829 return cachep;
3830 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3831}
3832
3833#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3834void *__kmalloc_node(size_t size, gfp_t flags, int node)
3835{
3836 return __do_kmalloc_node(size, flags, node,
3837 __builtin_return_address(0));
3838}
3839EXPORT_SYMBOL(__kmalloc_node);
3840
3841void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3842 int node, unsigned long caller)
3843{
3844 return __do_kmalloc_node(size, flags, node, (void *)caller);
3845}
3846EXPORT_SYMBOL(__kmalloc_node_track_caller);
3847#else
3848void *__kmalloc_node(size_t size, gfp_t flags, int node)
3849{
3850 return __do_kmalloc_node(size, flags, node, NULL);
3851}
3852EXPORT_SYMBOL(__kmalloc_node);
3853#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3854#endif /* CONFIG_NUMA */
3855
3856/**
3857 * __do_kmalloc - allocate memory
3858 * @size: how many bytes of memory are required.
3859 * @flags: the type of memory to allocate (see kmalloc).
3860 * @caller: function caller for debug tracking of the caller
3861 */
3862static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3863 void *caller)
3864{
3865 struct kmem_cache *cachep;
3866 void *ret;
3867
3868 /* If you want to save a few bytes .text space: replace
3869 * __ with kmem_.
3870 * Then kmalloc uses the uninlined functions instead of the inline
3871 * functions.
3872 */
3873 cachep = __find_general_cachep(size, flags);
3874 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3875 return cachep;
3876 ret = __cache_alloc(cachep, flags, caller);
3877
3878 trace_kmalloc((unsigned long) caller, ret,
3879 size, cachep->buffer_size, flags);
3880
3881 return ret;
3882}
3883
3884
3885#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3886void *__kmalloc(size_t size, gfp_t flags)
3887{
3888 return __do_kmalloc(size, flags, __builtin_return_address(0));
3889}
3890EXPORT_SYMBOL(__kmalloc);
3891
3892void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3893{
3894 return __do_kmalloc(size, flags, (void *)caller);
3895}
3896EXPORT_SYMBOL(__kmalloc_track_caller);
3897
3898#else
3899void *__kmalloc(size_t size, gfp_t flags)
3900{
3901 return __do_kmalloc(size, flags, NULL);
3902}
3903EXPORT_SYMBOL(__kmalloc);
3904#endif
3905
3906/**
3907 * kmem_cache_free - Deallocate an object
3908 * @cachep: The cache the allocation was from.
3909 * @objp: The previously allocated object.
3910 *
3911 * Free an object which was previously allocated from this
3912 * cache.
3913 */
3914void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3915{
3916 unsigned long flags;
3917
3918 local_irq_save(flags);
3919 debug_check_no_locks_freed(objp, obj_size(cachep));
3920 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3921 debug_check_no_obj_freed(objp, obj_size(cachep));
3922 __cache_free(cachep, objp, __builtin_return_address(0));
3923 local_irq_restore(flags);
3924
3925 trace_kmem_cache_free(_RET_IP_, objp);
3926}
3927EXPORT_SYMBOL(kmem_cache_free);
3928
3929/**
3930 * kfree - free previously allocated memory
3931 * @objp: pointer returned by kmalloc.
3932 *
3933 * If @objp is NULL, no operation is performed.
3934 *
3935 * Don't free memory not originally allocated by kmalloc()
3936 * or you will run into trouble.
3937 */
3938void kfree(const void *objp)
3939{
3940 struct kmem_cache *c;
3941 unsigned long flags;
3942
3943 trace_kfree(_RET_IP_, objp);
3944
3945 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3946 return;
3947 local_irq_save(flags);
3948 kfree_debugcheck(objp);
3949 c = virt_to_cache(objp);
3950 debug_check_no_locks_freed(objp, obj_size(c));
3951 debug_check_no_obj_freed(objp, obj_size(c));
3952 __cache_free(c, (void *)objp, __builtin_return_address(0));
3953 local_irq_restore(flags);
3954}
3955EXPORT_SYMBOL(kfree);
3956
3957unsigned int kmem_cache_size(struct kmem_cache *cachep)
3958{
3959 return obj_size(cachep);
3960}
3961EXPORT_SYMBOL(kmem_cache_size);
3962
3963/*
3964 * This initializes kmem_list3 or resizes various caches for all nodes.
3965 */
3966static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3967{
3968 int node;
3969 struct kmem_list3 *l3;
3970 struct array_cache *new_shared;
3971 struct array_cache **new_alien = NULL;
3972
3973 for_each_online_node(node) {
3974
3975 if (use_alien_caches) {
3976 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3977 if (!new_alien)
3978 goto fail;
3979 }
3980
3981 new_shared = NULL;
3982 if (cachep->shared) {
3983 new_shared = alloc_arraycache(node,
3984 cachep->shared*cachep->batchcount,
3985 0xbaadf00d, gfp);
3986 if (!new_shared) {
3987 free_alien_cache(new_alien);
3988 goto fail;
3989 }
3990 }
3991
3992 l3 = cachep->nodelists[node];
3993 if (l3) {
3994 struct array_cache *shared = l3->shared;
3995
3996 spin_lock_irq(&l3->list_lock);
3997
3998 if (shared)
3999 free_block(cachep, shared->entry,
4000 shared->avail, node);
4001
4002 l3->shared = new_shared;
4003 if (!l3->alien) {
4004 l3->alien = new_alien;
4005 new_alien = NULL;
4006 }
4007 l3->free_limit = (1 + nr_cpus_node(node)) *
4008 cachep->batchcount + cachep->num;
4009 spin_unlock_irq(&l3->list_lock);
4010 kfree(shared);
4011 free_alien_cache(new_alien);
4012 continue;
4013 }
4014 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4015 if (!l3) {
4016 free_alien_cache(new_alien);
4017 kfree(new_shared);
4018 goto fail;
4019 }
4020
4021 kmem_list3_init(l3);
4022 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4023 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4024 l3->shared = new_shared;
4025 l3->alien = new_alien;
4026 l3->free_limit = (1 + nr_cpus_node(node)) *
4027 cachep->batchcount + cachep->num;
4028 cachep->nodelists[node] = l3;
4029 }
4030 return 0;
4031
4032fail:
4033 if (!cachep->next.next) {
4034 /* Cache is not active yet. Roll back what we did */
4035 node--;
4036 while (node >= 0) {
4037 if (cachep->nodelists[node]) {
4038 l3 = cachep->nodelists[node];
4039
4040 kfree(l3->shared);
4041 free_alien_cache(l3->alien);
4042 kfree(l3);
4043 cachep->nodelists[node] = NULL;
4044 }
4045 node--;
4046 }
4047 }
4048 return -ENOMEM;
4049}
4050
4051struct ccupdate_struct {
4052 struct kmem_cache *cachep;
4053 struct array_cache *new[0];
4054};
4055
4056static void do_ccupdate_local(void *info)
4057{
4058 struct ccupdate_struct *new = info;
4059 struct array_cache *old;
4060
4061 check_irq_off();
4062 old = cpu_cache_get(new->cachep);
4063
4064 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4065 new->new[smp_processor_id()] = old;
4066}
4067
4068/* Always called with the cache_chain_mutex held */
4069static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4070 int batchcount, int shared, gfp_t gfp)
4071{
4072 struct ccupdate_struct *new;
4073 int i;
4074
4075 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4076 gfp);
4077 if (!new)
4078 return -ENOMEM;
4079
4080 for_each_online_cpu(i) {
4081 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4082 batchcount, gfp);
4083 if (!new->new[i]) {
4084 for (i--; i >= 0; i--)
4085 kfree(new->new[i]);
4086 kfree(new);
4087 return -ENOMEM;
4088 }
4089 }
4090 new->cachep = cachep;
4091
4092 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4093
4094 check_irq_on();
4095 cachep->batchcount = batchcount;
4096 cachep->limit = limit;
4097 cachep->shared = shared;
4098
4099 for_each_online_cpu(i) {
4100 struct array_cache *ccold = new->new[i];
4101 if (!ccold)
4102 continue;
4103 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4104 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4105 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4106 kfree(ccold);
4107 }
4108 kfree(new);
4109 return alloc_kmemlist(cachep, gfp);
4110}
4111
4112/* Called with cache_chain_mutex held always */
4113static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4114{
4115 int err;
4116 int limit, shared;
4117
4118 /*
4119 * The head array serves three purposes:
4120 * - create a LIFO ordering, i.e. return objects that are cache-warm
4121 * - reduce the number of spinlock operations.
4122 * - reduce the number of linked list operations on the slab and
4123 * bufctl chains: array operations are cheaper.
4124 * The numbers are guessed, we should auto-tune as described by
4125 * Bonwick.
4126 */
4127 if (cachep->buffer_size > 131072)
4128 limit = 1;
4129 else if (cachep->buffer_size > PAGE_SIZE)
4130 limit = 8;
4131 else if (cachep->buffer_size > 1024)
4132 limit = 24;
4133 else if (cachep->buffer_size > 256)
4134 limit = 54;
4135 else
4136 limit = 120;
4137
4138 /*
4139 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4140 * allocation behaviour: Most allocs on one cpu, most free operations
4141 * on another cpu. For these cases, an efficient object passing between
4142 * cpus is necessary. This is provided by a shared array. The array
4143 * replaces Bonwick's magazine layer.
4144 * On uniprocessor, it's functionally equivalent (but less efficient)
4145 * to a larger limit. Thus disabled by default.
4146 */
4147 shared = 0;
4148 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4149 shared = 8;
4150
4151#if DEBUG
4152 /*
4153 * With debugging enabled, large batchcount lead to excessively long
4154 * periods with disabled local interrupts. Limit the batchcount
4155 */
4156 if (limit > 32)
4157 limit = 32;
4158#endif
4159 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4160 if (err)
4161 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4162 cachep->name, -err);
4163 return err;
4164}
4165
4166/*
4167 * Drain an array if it contains any elements taking the l3 lock only if
4168 * necessary. Note that the l3 listlock also protects the array_cache
4169 * if drain_array() is used on the shared array.
4170 */
4171static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4172 struct array_cache *ac, int force, int node)
4173{
4174 int tofree;
4175
4176 if (!ac || !ac->avail)
4177 return;
4178 if (ac->touched && !force) {
4179 ac->touched = 0;
4180 } else {
4181 spin_lock_irq(&l3->list_lock);
4182 if (ac->avail) {
4183 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4184 if (tofree > ac->avail)
4185 tofree = (ac->avail + 1) / 2;
4186 free_block(cachep, ac->entry, tofree, node);
4187 ac->avail -= tofree;
4188 memmove(ac->entry, &(ac->entry[tofree]),
4189 sizeof(void *) * ac->avail);
4190 }
4191 spin_unlock_irq(&l3->list_lock);
4192 }
4193}
4194
4195/**
4196 * cache_reap - Reclaim memory from caches.
4197 * @w: work descriptor
4198 *
4199 * Called from workqueue/eventd every few seconds.
4200 * Purpose:
4201 * - clear the per-cpu caches for this CPU.
4202 * - return freeable pages to the main free memory pool.
4203 *
4204 * If we cannot acquire the cache chain mutex then just give up - we'll try
4205 * again on the next iteration.
4206 */
4207static void cache_reap(struct work_struct *w)
4208{
4209 struct kmem_cache *searchp;
4210 struct kmem_list3 *l3;
4211 int node = numa_mem_id();
4212 struct delayed_work *work = to_delayed_work(w);
4213
4214 if (!mutex_trylock(&cache_chain_mutex))
4215 /* Give up. Setup the next iteration. */
4216 goto out;
4217
4218 list_for_each_entry(searchp, &cache_chain, next) {
4219 check_irq_on();
4220
4221 /*
4222 * We only take the l3 lock if absolutely necessary and we
4223 * have established with reasonable certainty that
4224 * we can do some work if the lock was obtained.
4225 */
4226 l3 = searchp->nodelists[node];
4227
4228 reap_alien(searchp, l3);
4229
4230 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4231
4232 /*
4233 * These are racy checks but it does not matter
4234 * if we skip one check or scan twice.
4235 */
4236 if (time_after(l3->next_reap, jiffies))
4237 goto next;
4238
4239 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4240
4241 drain_array(searchp, l3, l3->shared, 0, node);
4242
4243 if (l3->free_touched)
4244 l3->free_touched = 0;
4245 else {
4246 int freed;
4247
4248 freed = drain_freelist(searchp, l3, (l3->free_limit +
4249 5 * searchp->num - 1) / (5 * searchp->num));
4250 STATS_ADD_REAPED(searchp, freed);
4251 }
4252next:
4253 cond_resched();
4254 }
4255 check_irq_on();
4256 mutex_unlock(&cache_chain_mutex);
4257 next_reap_node();
4258out:
4259 /* Set up the next iteration */
4260 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4261}
4262
4263#ifdef CONFIG_SLABINFO
4264
4265static void print_slabinfo_header(struct seq_file *m)
4266{
4267 /*
4268 * Output format version, so at least we can change it
4269 * without _too_ many complaints.
4270 */
4271#if STATS
4272 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4273#else
4274 seq_puts(m, "slabinfo - version: 2.1\n");
4275#endif
4276 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4277 "<objperslab> <pagesperslab>");
4278 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4279 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4280#if STATS
4281 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4282 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4283 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4284#endif
4285 seq_putc(m, '\n');
4286}
4287
4288static void *s_start(struct seq_file *m, loff_t *pos)
4289{
4290 loff_t n = *pos;
4291
4292 mutex_lock(&cache_chain_mutex);
4293 if (!n)
4294 print_slabinfo_header(m);
4295
4296 return seq_list_start(&cache_chain, *pos);
4297}
4298
4299static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4300{
4301 return seq_list_next(p, &cache_chain, pos);
4302}
4303
4304static void s_stop(struct seq_file *m, void *p)
4305{
4306 mutex_unlock(&cache_chain_mutex);
4307}
4308
4309static int s_show(struct seq_file *m, void *p)
4310{
4311 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4312 struct slab *slabp;
4313 unsigned long active_objs;
4314 unsigned long num_objs;
4315 unsigned long active_slabs = 0;
4316 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4317 const char *name;
4318 char *error = NULL;
4319 int node;
4320 struct kmem_list3 *l3;
4321
4322 active_objs = 0;
4323 num_slabs = 0;
4324 for_each_online_node(node) {
4325 l3 = cachep->nodelists[node];
4326 if (!l3)
4327 continue;
4328
4329 check_irq_on();
4330 spin_lock_irq(&l3->list_lock);
4331
4332 list_for_each_entry(slabp, &l3->slabs_full, list) {
4333 if (slabp->inuse != cachep->num && !error)
4334 error = "slabs_full accounting error";
4335 active_objs += cachep->num;
4336 active_slabs++;
4337 }
4338 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4339 if (slabp->inuse == cachep->num && !error)
4340 error = "slabs_partial inuse accounting error";
4341 if (!slabp->inuse && !error)
4342 error = "slabs_partial/inuse accounting error";
4343 active_objs += slabp->inuse;
4344 active_slabs++;
4345 }
4346 list_for_each_entry(slabp, &l3->slabs_free, list) {
4347 if (slabp->inuse && !error)
4348 error = "slabs_free/inuse accounting error";
4349 num_slabs++;
4350 }
4351 free_objects += l3->free_objects;
4352 if (l3->shared)
4353 shared_avail += l3->shared->avail;
4354
4355 spin_unlock_irq(&l3->list_lock);
4356 }
4357 num_slabs += active_slabs;
4358 num_objs = num_slabs * cachep->num;
4359 if (num_objs - active_objs != free_objects && !error)
4360 error = "free_objects accounting error";
4361
4362 name = cachep->name;
4363 if (error)
4364 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4365
4366 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4367 name, active_objs, num_objs, cachep->buffer_size,
4368 cachep->num, (1 << cachep->gfporder));
4369 seq_printf(m, " : tunables %4u %4u %4u",
4370 cachep->limit, cachep->batchcount, cachep->shared);
4371 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4372 active_slabs, num_slabs, shared_avail);
4373#if STATS
4374 { /* list3 stats */
4375 unsigned long high = cachep->high_mark;
4376 unsigned long allocs = cachep->num_allocations;
4377 unsigned long grown = cachep->grown;
4378 unsigned long reaped = cachep->reaped;
4379 unsigned long errors = cachep->errors;
4380 unsigned long max_freeable = cachep->max_freeable;
4381 unsigned long node_allocs = cachep->node_allocs;
4382 unsigned long node_frees = cachep->node_frees;
4383 unsigned long overflows = cachep->node_overflow;
4384
4385 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4386 "%4lu %4lu %4lu %4lu %4lu",
4387 allocs, high, grown,
4388 reaped, errors, max_freeable, node_allocs,
4389 node_frees, overflows);
4390 }
4391 /* cpu stats */
4392 {
4393 unsigned long allochit = atomic_read(&cachep->allochit);
4394 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4395 unsigned long freehit = atomic_read(&cachep->freehit);
4396 unsigned long freemiss = atomic_read(&cachep->freemiss);
4397
4398 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4399 allochit, allocmiss, freehit, freemiss);
4400 }
4401#endif
4402 seq_putc(m, '\n');
4403 return 0;
4404}
4405
4406/*
4407 * slabinfo_op - iterator that generates /proc/slabinfo
4408 *
4409 * Output layout:
4410 * cache-name
4411 * num-active-objs
4412 * total-objs
4413 * object size
4414 * num-active-slabs
4415 * total-slabs
4416 * num-pages-per-slab
4417 * + further values on SMP and with statistics enabled
4418 */
4419
4420static const struct seq_operations slabinfo_op = {
4421 .start = s_start,
4422 .next = s_next,
4423 .stop = s_stop,
4424 .show = s_show,
4425};
4426
4427#define MAX_SLABINFO_WRITE 128
4428/**
4429 * slabinfo_write - Tuning for the slab allocator
4430 * @file: unused
4431 * @buffer: user buffer
4432 * @count: data length
4433 * @ppos: unused
4434 */
4435static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4436 size_t count, loff_t *ppos)
4437{
4438 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4439 int limit, batchcount, shared, res;
4440 struct kmem_cache *cachep;
4441
4442 if (count > MAX_SLABINFO_WRITE)
4443 return -EINVAL;
4444 if (copy_from_user(&kbuf, buffer, count))
4445 return -EFAULT;
4446 kbuf[MAX_SLABINFO_WRITE] = '\0';
4447
4448 tmp = strchr(kbuf, ' ');
4449 if (!tmp)
4450 return -EINVAL;
4451 *tmp = '\0';
4452 tmp++;
4453 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4454 return -EINVAL;
4455
4456 /* Find the cache in the chain of caches. */
4457 mutex_lock(&cache_chain_mutex);
4458 res = -EINVAL;
4459 list_for_each_entry(cachep, &cache_chain, next) {
4460 if (!strcmp(cachep->name, kbuf)) {
4461 if (limit < 1 || batchcount < 1 ||
4462 batchcount > limit || shared < 0) {
4463 res = 0;
4464 } else {
4465 res = do_tune_cpucache(cachep, limit,
4466 batchcount, shared,
4467 GFP_KERNEL);
4468 }
4469 break;
4470 }
4471 }
4472 mutex_unlock(&cache_chain_mutex);
4473 if (res >= 0)
4474 res = count;
4475 return res;
4476}
4477
4478static int slabinfo_open(struct inode *inode, struct file *file)
4479{
4480 return seq_open(file, &slabinfo_op);
4481}
4482
4483static const struct file_operations proc_slabinfo_operations = {
4484 .open = slabinfo_open,
4485 .read = seq_read,
4486 .write = slabinfo_write,
4487 .llseek = seq_lseek,
4488 .release = seq_release,
4489};
4490
4491#ifdef CONFIG_DEBUG_SLAB_LEAK
4492
4493static void *leaks_start(struct seq_file *m, loff_t *pos)
4494{
4495 mutex_lock(&cache_chain_mutex);
4496 return seq_list_start(&cache_chain, *pos);
4497}
4498
4499static inline int add_caller(unsigned long *n, unsigned long v)
4500{
4501 unsigned long *p;
4502 int l;
4503 if (!v)
4504 return 1;
4505 l = n[1];
4506 p = n + 2;
4507 while (l) {
4508 int i = l/2;
4509 unsigned long *q = p + 2 * i;
4510 if (*q == v) {
4511 q[1]++;
4512 return 1;
4513 }
4514 if (*q > v) {
4515 l = i;
4516 } else {
4517 p = q + 2;
4518 l -= i + 1;
4519 }
4520 }
4521 if (++n[1] == n[0])
4522 return 0;
4523 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4524 p[0] = v;
4525 p[1] = 1;
4526 return 1;
4527}
4528
4529static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4530{
4531 void *p;
4532 int i;
4533 if (n[0] == n[1])
4534 return;
4535 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4536 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4537 continue;
4538 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4539 return;
4540 }
4541}
4542
4543static void show_symbol(struct seq_file *m, unsigned long address)
4544{
4545#ifdef CONFIG_KALLSYMS
4546 unsigned long offset, size;
4547 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4548
4549 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4550 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4551 if (modname[0])
4552 seq_printf(m, " [%s]", modname);
4553 return;
4554 }
4555#endif
4556 seq_printf(m, "%p", (void *)address);
4557}
4558
4559static int leaks_show(struct seq_file *m, void *p)
4560{
4561 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4562 struct slab *slabp;
4563 struct kmem_list3 *l3;
4564 const char *name;
4565 unsigned long *n = m->private;
4566 int node;
4567 int i;
4568
4569 if (!(cachep->flags & SLAB_STORE_USER))
4570 return 0;
4571 if (!(cachep->flags & SLAB_RED_ZONE))
4572 return 0;
4573
4574 /* OK, we can do it */
4575
4576 n[1] = 0;
4577
4578 for_each_online_node(node) {
4579 l3 = cachep->nodelists[node];
4580 if (!l3)
4581 continue;
4582
4583 check_irq_on();
4584 spin_lock_irq(&l3->list_lock);
4585
4586 list_for_each_entry(slabp, &l3->slabs_full, list)
4587 handle_slab(n, cachep, slabp);
4588 list_for_each_entry(slabp, &l3->slabs_partial, list)
4589 handle_slab(n, cachep, slabp);
4590 spin_unlock_irq(&l3->list_lock);
4591 }
4592 name = cachep->name;
4593 if (n[0] == n[1]) {
4594 /* Increase the buffer size */
4595 mutex_unlock(&cache_chain_mutex);
4596 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4597 if (!m->private) {
4598 /* Too bad, we are really out */
4599 m->private = n;
4600 mutex_lock(&cache_chain_mutex);
4601 return -ENOMEM;
4602 }
4603 *(unsigned long *)m->private = n[0] * 2;
4604 kfree(n);
4605 mutex_lock(&cache_chain_mutex);
4606 /* Now make sure this entry will be retried */
4607 m->count = m->size;
4608 return 0;
4609 }
4610 for (i = 0; i < n[1]; i++) {
4611 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4612 show_symbol(m, n[2*i+2]);
4613 seq_putc(m, '\n');
4614 }
4615
4616 return 0;
4617}
4618
4619static const struct seq_operations slabstats_op = {
4620 .start = leaks_start,
4621 .next = s_next,
4622 .stop = s_stop,
4623 .show = leaks_show,
4624};
4625
4626static int slabstats_open(struct inode *inode, struct file *file)
4627{
4628 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4629 int ret = -ENOMEM;
4630 if (n) {
4631 ret = seq_open(file, &slabstats_op);
4632 if (!ret) {
4633 struct seq_file *m = file->private_data;
4634 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4635 m->private = n;
4636 n = NULL;
4637 }
4638 kfree(n);
4639 }
4640 return ret;
4641}
4642
4643static const struct file_operations proc_slabstats_operations = {
4644 .open = slabstats_open,
4645 .read = seq_read,
4646 .llseek = seq_lseek,
4647 .release = seq_release_private,
4648};
4649#endif
4650
4651static int __init slab_proc_init(void)
4652{
4653 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4654#ifdef CONFIG_DEBUG_SLAB_LEAK
4655 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4656#endif
4657 return 0;
4658}
4659module_init(slab_proc_init);
4660#endif
4661
4662/**
4663 * ksize - get the actual amount of memory allocated for a given object
4664 * @objp: Pointer to the object
4665 *
4666 * kmalloc may internally round up allocations and return more memory
4667 * than requested. ksize() can be used to determine the actual amount of
4668 * memory allocated. The caller may use this additional memory, even though
4669 * a smaller amount of memory was initially specified with the kmalloc call.
4670 * The caller must guarantee that objp points to a valid object previously
4671 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4672 * must not be freed during the duration of the call.
4673 */
4674size_t ksize(const void *objp)
4675{
4676 BUG_ON(!objp);
4677 if (unlikely(objp == ZERO_SIZE_PTR))
4678 return 0;
4679
4680 return obj_size(virt_to_cache(objp));
4681}
4682EXPORT_SYMBOL(ksize);