blob: f8b0d539b4822af7812c8f9edcb51224450f3f64 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * 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 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * 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
Christoph Lameter18004c52012-07-06 15:25:12 -050071 * The global cache-chain is protected by the mutex 'slab_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * 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 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
Christoph Lameter97d06602012-07-06 15:25:11 -050090#include "slab.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070091#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070092#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070093#include <linux/swap.h>
94#include <linux/cache.h>
95#include <linux/interrupt.h>
96#include <linux/init.h>
97#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080098#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040099#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100#include <linux/seq_file.h>
101#include <linux/notifier.h>
102#include <linux/kallsyms.h>
103#include <linux/cpu.h>
104#include <linux/sysctl.h>
105#include <linux/module.h>
106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100110#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800111#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800112#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800113#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700114#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800115#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700116#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200117#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700118#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700119#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
Mel Gorman381760e2012-07-31 16:44:30 -0700121#include <net/sock.h>
122
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#include <asm/cacheflush.h>
124#include <asm/tlbflush.h>
125#include <asm/page.h>
126
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500127#include <trace/events/kmem.h>
128
Mel Gorman072bb0a2012-07-31 16:43:58 -0700129#include "internal.h"
130
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700132 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 * 0 for faster, smaller code (especially in the critical paths).
134 *
135 * STATS - 1 to collect stats for /proc/slabinfo.
136 * 0 for faster, smaller code (especially in the critical paths).
137 *
138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
139 */
140
141#ifdef CONFIG_DEBUG_SLAB
142#define DEBUG 1
143#define STATS 1
144#define FORCED_DEBUG 1
145#else
146#define DEBUG 0
147#define STATS 0
148#define FORCED_DEBUG 0
149#endif
150
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151/* Shouldn't this be in a header file somewhere? */
152#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400153#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155#ifndef ARCH_KMALLOC_FLAGS
156#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
157#endif
158
Mel Gorman072bb0a2012-07-31 16:43:58 -0700159/*
160 * true if a page was allocated from pfmemalloc reserves for network-based
161 * swap
162 */
163static bool pfmemalloc_active __read_mostly;
164
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165/* Legal flag mask for kmem_cache_create(). */
166#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700167# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800169 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700170 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700172 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200173 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800175# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700176 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200179 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180#endif
181
182/*
183 * kmem_bufctl_t:
184 *
185 * Bufctl's are used for linking objs within a slab
186 * linked offsets.
187 *
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
199 */
200
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700201typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800204#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
205#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 * struct slab_rcu
209 *
210 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
211 * arrange for kmem_freepages to be called via RCU. This is useful if
212 * we need to approach a kernel structure obliquely, from its address
213 * obtained without the usual locking. We can lock the structure to
214 * stabilize it and check it's still at the given address, only if we
215 * can be sure that the memory has not been meanwhile reused for some
216 * other kind of object (which our subsystem's lock might corrupt).
217 *
218 * rcu_read_lock before reading the address, then rcu_read_unlock after
219 * taking the spinlock within the structure expected at that address.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 */
221struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800222 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800223 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800224 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225};
226
227/*
Lai Jiangshan5bfe53a2011-03-10 15:22:24 +0800228 * struct slab
229 *
230 * Manages the objs in a slab. Placed either at the beginning of mem allocated
231 * for a slab, or allocated from an general cache.
232 * Slabs are chained into three list: fully used, partial, fully free slabs.
233 */
234struct slab {
235 union {
236 struct {
237 struct list_head list;
238 unsigned long colouroff;
239 void *s_mem; /* including colour offset */
240 unsigned int inuse; /* num of objs active in slab */
241 kmem_bufctl_t free;
242 unsigned short nodeid;
243 };
244 struct slab_rcu __slab_cover_slab_rcu;
245 };
246};
247
248/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 * struct array_cache
250 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 * Purpose:
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
255 *
256 * The limit is stored in the per-cpu structure to reduce the data cache
257 * footprint.
258 *
259 */
260struct array_cache {
261 unsigned int avail;
262 unsigned int limit;
263 unsigned int batchcount;
264 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700265 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700266 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
269 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700270 *
271 * Entries should not be directly dereferenced as
272 * entries belonging to slabs marked pfmemalloc will
273 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
Mel Gorman072bb0a2012-07-31 16:43:58 -0700277#define SLAB_OBJ_PFMEMALLOC 1
278static inline bool is_obj_pfmemalloc(void *objp)
279{
280 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
281}
282
283static inline void set_obj_pfmemalloc(void **objp)
284{
285 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
286 return;
287}
288
289static inline void clear_obj_pfmemalloc(void **objp)
290{
291 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
292}
293
Andrew Mortona737b3e2006-03-22 00:08:11 -0800294/*
295 * bootstrap: The caches do not work without cpuarrays anymore, but the
296 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 */
298#define BOOT_CPUCACHE_ENTRIES 1
299struct arraycache_init {
300 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800301 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
304/*
Christoph Lametere498be72005-09-09 13:03:32 -0700305 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 */
307struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800308 struct list_head slabs_partial; /* partial list first, better asm code */
309 struct list_head slabs_full;
310 struct list_head slabs_free;
311 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800312 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800313 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800314 spinlock_t list_lock;
315 struct array_cache *shared; /* shared per node */
316 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800317 unsigned long next_reap; /* updated without locking */
318 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319};
320
Christoph Lametere498be72005-09-09 13:03:32 -0700321/*
322 * Need this for bootstrapping a per node allocator.
323 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200324#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
H Hartley Sweeten68a1b192011-01-11 17:49:32 -0600325static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700326#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200327#define SIZE_AC MAX_NUMNODES
328#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329
Christoph Lametered11d9e2006-06-30 01:55:45 -0700330static int drain_freelist(struct kmem_cache *cache,
331 struct kmem_list3 *l3, int tofree);
332static void free_block(struct kmem_cache *cachep, void **objpp, int len,
333 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300334static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000335static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700336
Christoph Lametere498be72005-09-09 13:03:32 -0700337/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800338 * This function must be completely optimized away if a constant is passed to
339 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700340 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700341static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700342{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800343 extern void __bad_size(void);
344
Christoph Lametere498be72005-09-09 13:03:32 -0700345 if (__builtin_constant_p(size)) {
346 int i = 0;
347
348#define CACHE(x) \
349 if (size <=x) \
350 return i; \
351 else \
352 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800353#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700354#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800355 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700356 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800357 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700358 return 0;
359}
360
Ingo Molnare0a42722006-06-23 02:03:46 -0700361static int slab_early_init = 1;
362
Christoph Lametere498be72005-09-09 13:03:32 -0700363#define INDEX_AC index_of(sizeof(struct arraycache_init))
364#define INDEX_L3 index_of(sizeof(struct kmem_list3))
365
Pekka Enberg5295a742006-02-01 03:05:48 -0800366static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700367{
368 INIT_LIST_HEAD(&parent->slabs_full);
369 INIT_LIST_HEAD(&parent->slabs_partial);
370 INIT_LIST_HEAD(&parent->slabs_free);
371 parent->shared = NULL;
372 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800373 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700374 spin_lock_init(&parent->list_lock);
375 parent->free_objects = 0;
376 parent->free_touched = 0;
377}
378
Andrew Mortona737b3e2006-03-22 00:08:11 -0800379#define MAKE_LIST(cachep, listp, slab, nodeid) \
380 do { \
381 INIT_LIST_HEAD(listp); \
382 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700383 } while (0)
384
Andrew Mortona737b3e2006-03-22 00:08:11 -0800385#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
386 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700387 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
388 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
389 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
390 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392#define CFLGS_OFF_SLAB (0x80000000UL)
393#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
394
395#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800396/*
397 * Optimization question: fewer reaps means less probability for unnessary
398 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100400 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 * which could lock up otherwise freeable slabs.
402 */
403#define REAPTIMEOUT_CPUC (2*HZ)
404#define REAPTIMEOUT_LIST3 (4*HZ)
405
406#if STATS
407#define STATS_INC_ACTIVE(x) ((x)->num_active++)
408#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
409#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
410#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700411#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800412#define STATS_SET_HIGH(x) \
413 do { \
414 if ((x)->num_active > (x)->high_mark) \
415 (x)->high_mark = (x)->num_active; \
416 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417#define STATS_INC_ERR(x) ((x)->errors++)
418#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700419#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700420#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800421#define STATS_SET_FREEABLE(x, i) \
422 do { \
423 if ((x)->max_freeable < i) \
424 (x)->max_freeable = i; \
425 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
427#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
428#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
429#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
430#else
431#define STATS_INC_ACTIVE(x) do { } while (0)
432#define STATS_DEC_ACTIVE(x) do { } while (0)
433#define STATS_INC_ALLOCED(x) do { } while (0)
434#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700435#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436#define STATS_SET_HIGH(x) do { } while (0)
437#define STATS_INC_ERR(x) do { } while (0)
438#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700439#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700440#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800441#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442#define STATS_INC_ALLOCHIT(x) do { } while (0)
443#define STATS_INC_ALLOCMISS(x) do { } while (0)
444#define STATS_INC_FREEHIT(x) do { } while (0)
445#define STATS_INC_FREEMISS(x) do { } while (0)
446#endif
447
448#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
Andrew Mortona737b3e2006-03-22 00:08:11 -0800450/*
451 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800453 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 * the end of an object is aligned with the end of the real
455 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800456 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800458 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500459 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
460 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800461 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800463static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800465 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466}
467
David Woodhouseb46b8f12007-05-08 00:22:59 -0700468static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469{
470 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700471 return (unsigned long long*) (objp + obj_offset(cachep) -
472 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473}
474
David Woodhouseb46b8f12007-05-08 00:22:59 -0700475static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476{
477 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
478 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500479 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700480 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400481 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500482 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700483 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484}
485
Pekka Enberg343e0d72006-02-01 03:05:50 -0800486static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487{
488 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500489 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490}
491
492#else
493
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800494#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700495#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
496#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
498
499#endif
500
Li Zefan0f24f122009-12-11 15:45:30 +0800501#ifdef CONFIG_TRACING
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300502size_t slab_buffer_size(struct kmem_cache *cachep)
503{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500504 return cachep->size;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300505}
506EXPORT_SYMBOL(slab_buffer_size);
507#endif
508
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700510 * Do not go above this order unless 0 objects fit into the slab or
511 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 */
David Rientjes543585c2011-10-18 22:09:24 -0700513#define SLAB_MAX_ORDER_HI 1
514#define SLAB_MAX_ORDER_LO 0
515static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700516static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517
Pekka Enberg065d41c2005-11-13 16:06:46 -0800518static inline struct kmem_cache *page_get_cache(struct page *page)
519{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700520 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700521 BUG_ON(!PageSlab(page));
Christoph Lametere571b0a2012-06-13 10:24:55 -0500522 return page->slab_cache;
Pekka Enberg065d41c2005-11-13 16:06:46 -0800523}
524
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800525static inline struct kmem_cache *virt_to_cache(const void *obj)
526{
Christoph Lameterb49af682007-05-06 14:49:41 -0700527 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500528 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800529}
530
531static inline struct slab *virt_to_slab(const void *obj)
532{
Christoph Lameterb49af682007-05-06 14:49:41 -0700533 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500534
535 VM_BUG_ON(!PageSlab(page));
536 return page->slab_page;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800537}
538
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800539static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
540 unsigned int idx)
541{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500542 return slab->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800543}
544
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800545/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500546 * We want to avoid an expensive divide : (offset / cache->size)
547 * Using the fact that size is a constant for a particular cache,
548 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800549 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
550 */
551static inline unsigned int obj_to_index(const struct kmem_cache *cache,
552 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800553{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800554 u32 offset = (obj - slab->s_mem);
555 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800556}
557
Andrew Mortona737b3e2006-03-22 00:08:11 -0800558/*
559 * These are the default caches for kmalloc. Custom caches can have other sizes.
560 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561struct cache_sizes malloc_sizes[] = {
562#define CACHE(x) { .cs_size = (x) },
563#include <linux/kmalloc_sizes.h>
564 CACHE(ULONG_MAX)
565#undef CACHE
566};
567EXPORT_SYMBOL(malloc_sizes);
568
569/* Must match cache_sizes above. Out of line to keep cache footprint low. */
570struct cache_names {
571 char *name;
572 char *name_dma;
573};
574
575static struct cache_names __initdata cache_names[] = {
576#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
577#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800578 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579#undef CACHE
580};
581
582static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800583 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800585 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
587/* internal cache of cache description objs */
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200588static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
Pekka Enberg343e0d72006-02-01 03:05:50 -0800589static struct kmem_cache cache_cache = {
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200590 .nodelists = cache_cache_nodelists,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800591 .batchcount = 1,
592 .limit = BOOT_CPUCACHE_ENTRIES,
593 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500594 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800595 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596};
597
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700598#define BAD_ALIEN_MAGIC 0x01020304ul
599
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200600#ifdef CONFIG_LOCKDEP
601
602/*
603 * Slab sometimes uses the kmalloc slabs to store the slab headers
604 * for other slabs "off slab".
605 * The locking for this is tricky in that it nests within the locks
606 * of all other slabs in a few places; to deal with this special
607 * locking we put on-slab caches into a separate lock-class.
608 *
609 * We set lock class for alien array caches which are up during init.
610 * The lock annotation will be lost if all cpus of a node goes down and
611 * then comes back up during hotplug
612 */
613static struct lock_class_key on_slab_l3_key;
614static struct lock_class_key on_slab_alc_key;
615
Peter Zijlstra83835b32011-07-22 15:26:05 +0200616static struct lock_class_key debugobj_l3_key;
617static struct lock_class_key debugobj_alc_key;
618
619static void slab_set_lock_classes(struct kmem_cache *cachep,
620 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
621 int q)
622{
623 struct array_cache **alc;
624 struct kmem_list3 *l3;
625 int r;
626
627 l3 = cachep->nodelists[q];
628 if (!l3)
629 return;
630
631 lockdep_set_class(&l3->list_lock, l3_key);
632 alc = l3->alien;
633 /*
634 * FIXME: This check for BAD_ALIEN_MAGIC
635 * should go away when common slab code is taught to
636 * work even without alien caches.
637 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
638 * for alloc_alien_cache,
639 */
640 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
641 return;
642 for_each_node(r) {
643 if (alc[r])
644 lockdep_set_class(&alc[r]->lock, alc_key);
645 }
646}
647
648static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
649{
650 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
651}
652
653static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
654{
655 int node;
656
657 for_each_online_node(node)
658 slab_set_debugobj_lock_classes_node(cachep, node);
659}
660
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200661static void init_node_lock_keys(int q)
662{
663 struct cache_sizes *s = malloc_sizes;
664
Christoph Lameter97d06602012-07-06 15:25:11 -0500665 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200666 return;
667
668 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200669 struct kmem_list3 *l3;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200670
671 l3 = s->cs_cachep->nodelists[q];
672 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200673 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200674
675 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
676 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200677 }
678}
679
680static inline void init_lock_keys(void)
681{
682 int node;
683
684 for_each_node(node)
685 init_node_lock_keys(node);
686}
687#else
688static void init_node_lock_keys(int q)
689{
690}
691
692static inline void init_lock_keys(void)
693{
694}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200695
696static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
697{
698}
699
700static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
701{
702}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200703#endif
704
Tejun Heo1871e522009-10-29 22:34:13 +0900705static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
Pekka Enberg343e0d72006-02-01 03:05:50 -0800707static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708{
709 return cachep->array[smp_processor_id()];
710}
711
Andrew Mortona737b3e2006-03-22 00:08:11 -0800712static inline struct kmem_cache *__find_general_cachep(size_t size,
713 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714{
715 struct cache_sizes *csizep = malloc_sizes;
716
717#if DEBUG
718 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800719 * kmem_cache_create(), or __kmalloc(), before
720 * the generic caches are initialized.
721 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700722 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700724 if (!size)
725 return ZERO_SIZE_PTR;
726
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 while (size > csizep->cs_size)
728 csizep++;
729
730 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700731 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 * has cs_{dma,}cachep==NULL. Thus no special case
733 * for large kmalloc calls required.
734 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800735#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736 if (unlikely(gfpflags & GFP_DMA))
737 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800738#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 return csizep->cs_cachep;
740}
741
Adrian Bunkb2213852006-09-25 23:31:02 -0700742static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700743{
744 return __find_general_cachep(size, gfpflags);
745}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700746
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800747static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800749 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
750}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751
Andrew Mortona737b3e2006-03-22 00:08:11 -0800752/*
753 * Calculate the number of objects and left-over bytes for a given buffer size.
754 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800755static void cache_estimate(unsigned long gfporder, size_t buffer_size,
756 size_t align, int flags, size_t *left_over,
757 unsigned int *num)
758{
759 int nr_objs;
760 size_t mgmt_size;
761 size_t slab_size = PAGE_SIZE << gfporder;
762
763 /*
764 * The slab management structure can be either off the slab or
765 * on it. For the latter case, the memory allocated for a
766 * slab is used for:
767 *
768 * - The struct slab
769 * - One kmem_bufctl_t for each object
770 * - Padding to respect alignment of @align
771 * - @buffer_size bytes for each object
772 *
773 * If the slab management structure is off the slab, then the
774 * alignment will already be calculated into the size. Because
775 * the slabs are all pages aligned, the objects will be at the
776 * correct alignment when allocated.
777 */
778 if (flags & CFLGS_OFF_SLAB) {
779 mgmt_size = 0;
780 nr_objs = slab_size / buffer_size;
781
782 if (nr_objs > SLAB_LIMIT)
783 nr_objs = SLAB_LIMIT;
784 } else {
785 /*
786 * Ignore padding for the initial guess. The padding
787 * is at most @align-1 bytes, and @buffer_size is at
788 * least @align. In the worst case, this result will
789 * be one greater than the number of objects that fit
790 * into the memory allocation when taking the padding
791 * into account.
792 */
793 nr_objs = (slab_size - sizeof(struct slab)) /
794 (buffer_size + sizeof(kmem_bufctl_t));
795
796 /*
797 * This calculated number will be either the right
798 * amount, or one greater than what we want.
799 */
800 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
801 > slab_size)
802 nr_objs--;
803
804 if (nr_objs > SLAB_LIMIT)
805 nr_objs = SLAB_LIMIT;
806
807 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800809 *num = nr_objs;
810 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811}
812
Harvey Harrisond40cee22008-04-30 00:55:07 -0700813#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814
Andrew Mortona737b3e2006-03-22 00:08:11 -0800815static void __slab_error(const char *function, struct kmem_cache *cachep,
816 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817{
818 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800819 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 dump_stack();
821}
822
Paul Menage3395ee02006-12-06 20:32:16 -0800823/*
824 * By default on NUMA we use alien caches to stage the freeing of
825 * objects allocated from other nodes. This causes massive memory
826 * inefficiencies when using fake NUMA setup to split memory into a
827 * large number of small nodes, so it can be disabled on the command
828 * line
829 */
830
831static int use_alien_caches __read_mostly = 1;
832static int __init noaliencache_setup(char *s)
833{
834 use_alien_caches = 0;
835 return 1;
836}
837__setup("noaliencache", noaliencache_setup);
838
David Rientjes3df1ccc2011-10-18 22:09:28 -0700839static int __init slab_max_order_setup(char *str)
840{
841 get_option(&str, &slab_max_order);
842 slab_max_order = slab_max_order < 0 ? 0 :
843 min(slab_max_order, MAX_ORDER - 1);
844 slab_max_order_set = true;
845
846 return 1;
847}
848__setup("slab_max_order=", slab_max_order_setup);
849
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800850#ifdef CONFIG_NUMA
851/*
852 * Special reaping functions for NUMA systems called from cache_reap().
853 * These take care of doing round robin flushing of alien caches (containing
854 * objects freed on different nodes from which they were allocated) and the
855 * flushing of remote pcps by calling drain_node_pages.
856 */
Tejun Heo1871e522009-10-29 22:34:13 +0900857static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800858
859static void init_reap_node(int cpu)
860{
861 int node;
862
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700863 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800864 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800865 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800866
Tejun Heo1871e522009-10-29 22:34:13 +0900867 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800868}
869
870static void next_reap_node(void)
871{
Christoph Lameter909ea962010-12-08 16:22:55 +0100872 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800873
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800874 node = next_node(node, node_online_map);
875 if (unlikely(node >= MAX_NUMNODES))
876 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100877 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800878}
879
880#else
881#define init_reap_node(cpu) do { } while (0)
882#define next_reap_node(void) do { } while (0)
883#endif
884
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885/*
886 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
887 * via the workqueue/eventd.
888 * Add the CPU number into the expiration time to minimize the possibility of
889 * the CPUs getting into lockstep and contending for the global cache chain
890 * lock.
891 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700892static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893{
Tejun Heo1871e522009-10-29 22:34:13 +0900894 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895
896 /*
897 * When this gets called from do_initcalls via cpucache_init(),
898 * init_workqueues() has already run, so keventd will be setup
899 * at that time.
900 */
David Howells52bad642006-11-22 14:54:01 +0000901 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800902 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700903 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800904 schedule_delayed_work_on(cpu, reap_work,
905 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 }
907}
908
Christoph Lametere498be72005-09-09 13:03:32 -0700909static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300910 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800912 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 struct array_cache *nc = NULL;
914
Pekka Enberg83b519e2009-06-10 19:40:04 +0300915 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100916 /*
917 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300918 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100919 * cache the pointers are not cleared and they could be counted as
920 * valid references during a kmemleak scan. Therefore, kmemleak must
921 * not scan such objects.
922 */
923 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 if (nc) {
925 nc->avail = 0;
926 nc->limit = entries;
927 nc->batchcount = batchcount;
928 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700929 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 }
931 return nc;
932}
933
Mel Gorman072bb0a2012-07-31 16:43:58 -0700934static inline bool is_slab_pfmemalloc(struct slab *slabp)
935{
936 struct page *page = virt_to_page(slabp->s_mem);
937
938 return PageSlabPfmemalloc(page);
939}
940
941/* Clears pfmemalloc_active if no slabs have pfmalloc set */
942static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
943 struct array_cache *ac)
944{
945 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
946 struct slab *slabp;
947 unsigned long flags;
948
949 if (!pfmemalloc_active)
950 return;
951
952 spin_lock_irqsave(&l3->list_lock, flags);
953 list_for_each_entry(slabp, &l3->slabs_full, list)
954 if (is_slab_pfmemalloc(slabp))
955 goto out;
956
957 list_for_each_entry(slabp, &l3->slabs_partial, list)
958 if (is_slab_pfmemalloc(slabp))
959 goto out;
960
961 list_for_each_entry(slabp, &l3->slabs_free, list)
962 if (is_slab_pfmemalloc(slabp))
963 goto out;
964
965 pfmemalloc_active = false;
966out:
967 spin_unlock_irqrestore(&l3->list_lock, flags);
968}
969
Mel Gorman381760e2012-07-31 16:44:30 -0700970static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700971 gfp_t flags, bool force_refill)
972{
973 int i;
974 void *objp = ac->entry[--ac->avail];
975
976 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
977 if (unlikely(is_obj_pfmemalloc(objp))) {
978 struct kmem_list3 *l3;
979
980 if (gfp_pfmemalloc_allowed(flags)) {
981 clear_obj_pfmemalloc(&objp);
982 return objp;
983 }
984
985 /* The caller cannot use PFMEMALLOC objects, find another one */
986 for (i = 1; i < ac->avail; i++) {
987 /* If a !PFMEMALLOC object is found, swap them */
988 if (!is_obj_pfmemalloc(ac->entry[i])) {
989 objp = ac->entry[i];
990 ac->entry[i] = ac->entry[ac->avail];
991 ac->entry[ac->avail] = objp;
992 return objp;
993 }
994 }
995
996 /*
997 * If there are empty slabs on the slabs_free list and we are
998 * being forced to refill the cache, mark this one !pfmemalloc.
999 */
1000 l3 = cachep->nodelists[numa_mem_id()];
1001 if (!list_empty(&l3->slabs_free) && force_refill) {
1002 struct slab *slabp = virt_to_slab(objp);
1003 ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem));
1004 clear_obj_pfmemalloc(&objp);
1005 recheck_pfmemalloc_active(cachep, ac);
1006 return objp;
1007 }
1008
1009 /* No !PFMEMALLOC objects available */
1010 ac->avail++;
1011 objp = NULL;
1012 }
1013
1014 return objp;
1015}
1016
Mel Gorman381760e2012-07-31 16:44:30 -07001017static inline void *ac_get_obj(struct kmem_cache *cachep,
1018 struct array_cache *ac, gfp_t flags, bool force_refill)
1019{
1020 void *objp;
1021
1022 if (unlikely(sk_memalloc_socks()))
1023 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1024 else
1025 objp = ac->entry[--ac->avail];
1026
1027 return objp;
1028}
1029
1030static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -07001031 void *objp)
1032{
1033 if (unlikely(pfmemalloc_active)) {
1034 /* Some pfmemalloc slabs exist, check if this is one */
1035 struct page *page = virt_to_page(objp);
1036 if (PageSlabPfmemalloc(page))
1037 set_obj_pfmemalloc(&objp);
1038 }
1039
Mel Gorman381760e2012-07-31 16:44:30 -07001040 return objp;
1041}
1042
1043static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1044 void *objp)
1045{
1046 if (unlikely(sk_memalloc_socks()))
1047 objp = __ac_put_obj(cachep, ac, objp);
1048
Mel Gorman072bb0a2012-07-31 16:43:58 -07001049 ac->entry[ac->avail++] = objp;
1050}
1051
Christoph Lameter3ded1752006-03-25 03:06:44 -08001052/*
1053 * Transfer objects in one arraycache to another.
1054 * Locking must be handled by the caller.
1055 *
1056 * Return the number of entries transferred.
1057 */
1058static int transfer_objects(struct array_cache *to,
1059 struct array_cache *from, unsigned int max)
1060{
1061 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -07001062 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -08001063
1064 if (!nr)
1065 return 0;
1066
1067 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1068 sizeof(void *) *nr);
1069
1070 from->avail -= nr;
1071 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -08001072 return nr;
1073}
1074
Christoph Lameter765c4502006-09-27 01:50:08 -07001075#ifndef CONFIG_NUMA
1076
1077#define drain_alien_cache(cachep, alien) do { } while (0)
1078#define reap_alien(cachep, l3) do { } while (0)
1079
Pekka Enberg83b519e2009-06-10 19:40:04 +03001080static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -07001081{
1082 return (struct array_cache **)BAD_ALIEN_MAGIC;
1083}
1084
1085static inline void free_alien_cache(struct array_cache **ac_ptr)
1086{
1087}
1088
1089static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1090{
1091 return 0;
1092}
1093
1094static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1095 gfp_t flags)
1096{
1097 return NULL;
1098}
1099
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001100static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001101 gfp_t flags, int nodeid)
1102{
1103 return NULL;
1104}
1105
1106#else /* CONFIG_NUMA */
1107
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001108static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001109static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001110
Pekka Enberg83b519e2009-06-10 19:40:04 +03001111static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001112{
1113 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001114 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001115 int i;
1116
1117 if (limit > 1)
1118 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001119 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001120 if (ac_ptr) {
1121 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001122 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001123 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001124 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001125 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001126 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001127 kfree(ac_ptr[i]);
1128 kfree(ac_ptr);
1129 return NULL;
1130 }
1131 }
1132 }
1133 return ac_ptr;
1134}
1135
Pekka Enberg5295a742006-02-01 03:05:48 -08001136static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001137{
1138 int i;
1139
1140 if (!ac_ptr)
1141 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001142 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001143 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001144 kfree(ac_ptr);
1145}
1146
Pekka Enberg343e0d72006-02-01 03:05:50 -08001147static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001148 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001149{
1150 struct kmem_list3 *rl3 = cachep->nodelists[node];
1151
1152 if (ac->avail) {
1153 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001154 /*
1155 * Stuff objects into the remote nodes shared array first.
1156 * That way we could avoid the overhead of putting the objects
1157 * into the free lists and getting them back later.
1158 */
shin, jacob693f7d32006-04-28 10:54:37 -05001159 if (rl3->shared)
1160 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001161
Christoph Lameterff694162005-09-22 21:44:02 -07001162 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001163 ac->avail = 0;
1164 spin_unlock(&rl3->list_lock);
1165 }
1166}
1167
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001168/*
1169 * Called from cache_reap() to regularly drain alien caches round robin.
1170 */
1171static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1172{
Christoph Lameter909ea962010-12-08 16:22:55 +01001173 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001174
1175 if (l3->alien) {
1176 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001177
1178 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001179 __drain_alien_cache(cachep, ac, node);
1180 spin_unlock_irq(&ac->lock);
1181 }
1182 }
1183}
1184
Andrew Mortona737b3e2006-03-22 00:08:11 -08001185static void drain_alien_cache(struct kmem_cache *cachep,
1186 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001187{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001188 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001189 struct array_cache *ac;
1190 unsigned long flags;
1191
1192 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001193 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001194 if (ac) {
1195 spin_lock_irqsave(&ac->lock, flags);
1196 __drain_alien_cache(cachep, ac, i);
1197 spin_unlock_irqrestore(&ac->lock, flags);
1198 }
1199 }
1200}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001201
Ingo Molnar873623d2006-07-13 14:44:38 +02001202static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001203{
1204 struct slab *slabp = virt_to_slab(objp);
1205 int nodeid = slabp->nodeid;
1206 struct kmem_list3 *l3;
1207 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001208 int node;
1209
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001210 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001211
1212 /*
1213 * Make sure we are not freeing a object from another node to the array
1214 * cache on this cpu.
1215 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001216 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001217 return 0;
1218
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001219 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001220 STATS_INC_NODEFREES(cachep);
1221 if (l3->alien && l3->alien[nodeid]) {
1222 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001223 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001224 if (unlikely(alien->avail == alien->limit)) {
1225 STATS_INC_ACOVERFLOW(cachep);
1226 __drain_alien_cache(cachep, alien, nodeid);
1227 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001228 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001229 spin_unlock(&alien->lock);
1230 } else {
1231 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1232 free_block(cachep, &objp, 1, nodeid);
1233 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1234 }
1235 return 1;
1236}
Christoph Lametere498be72005-09-09 13:03:32 -07001237#endif
1238
David Rientjes8f9f8d92010-03-27 19:40:47 -07001239/*
1240 * Allocates and initializes nodelists for a node on each slab cache, used for
1241 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1242 * will be allocated off-node since memory is not yet online for the new node.
1243 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1244 * already in use.
1245 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001246 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001247 */
1248static int init_cache_nodelists_node(int node)
1249{
1250 struct kmem_cache *cachep;
1251 struct kmem_list3 *l3;
1252 const int memsize = sizeof(struct kmem_list3);
1253
Christoph Lameter18004c52012-07-06 15:25:12 -05001254 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001255 /*
1256 * Set up the size64 kmemlist for cpu before we can
1257 * begin anything. Make sure some other cpu on this
1258 * node has not already allocated this
1259 */
1260 if (!cachep->nodelists[node]) {
1261 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1262 if (!l3)
1263 return -ENOMEM;
1264 kmem_list3_init(l3);
1265 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1266 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1267
1268 /*
1269 * The l3s don't come and go as CPUs come and
Christoph Lameter18004c52012-07-06 15:25:12 -05001270 * go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001271 * protection here.
1272 */
1273 cachep->nodelists[node] = l3;
1274 }
1275
1276 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1277 cachep->nodelists[node]->free_limit =
1278 (1 + nr_cpus_node(node)) *
1279 cachep->batchcount + cachep->num;
1280 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1281 }
1282 return 0;
1283}
1284
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001285static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001287 struct kmem_cache *cachep;
1288 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001289 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301290 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001291
Christoph Lameter18004c52012-07-06 15:25:12 -05001292 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001293 struct array_cache *nc;
1294 struct array_cache *shared;
1295 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001296
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001297 /* cpu is dead; no one can alloc from it. */
1298 nc = cachep->array[cpu];
1299 cachep->array[cpu] = NULL;
1300 l3 = cachep->nodelists[node];
1301
1302 if (!l3)
1303 goto free_array_cache;
1304
1305 spin_lock_irq(&l3->list_lock);
1306
1307 /* Free limit for this kmem_list3 */
1308 l3->free_limit -= cachep->batchcount;
1309 if (nc)
1310 free_block(cachep, nc->entry, nc->avail, node);
1311
Rusty Russell58463c12009-12-17 11:43:12 -06001312 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001313 spin_unlock_irq(&l3->list_lock);
1314 goto free_array_cache;
1315 }
1316
1317 shared = l3->shared;
1318 if (shared) {
1319 free_block(cachep, shared->entry,
1320 shared->avail, node);
1321 l3->shared = NULL;
1322 }
1323
1324 alien = l3->alien;
1325 l3->alien = NULL;
1326
1327 spin_unlock_irq(&l3->list_lock);
1328
1329 kfree(shared);
1330 if (alien) {
1331 drain_alien_cache(cachep, alien);
1332 free_alien_cache(alien);
1333 }
1334free_array_cache:
1335 kfree(nc);
1336 }
1337 /*
1338 * In the previous loop, all the objects were freed to
1339 * the respective cache's slabs, now we can go ahead and
1340 * shrink each nodelist to its limit.
1341 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001342 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001343 l3 = cachep->nodelists[node];
1344 if (!l3)
1345 continue;
1346 drain_freelist(cachep, l3, l3->free_objects);
1347 }
1348}
1349
1350static int __cpuinit cpuup_prepare(long cpu)
1351{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001352 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001353 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001354 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001355 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001357 /*
1358 * We need to do this right in the beginning since
1359 * alloc_arraycache's are going to use this list.
1360 * kmalloc_node allows us to add the slab to the right
1361 * kmem_list3 and not this cpu's kmem_list3
1362 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001363 err = init_cache_nodelists_node(node);
1364 if (err < 0)
1365 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001366
1367 /*
1368 * Now we can go ahead with allocating the shared arrays and
1369 * array caches
1370 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001371 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001372 struct array_cache *nc;
1373 struct array_cache *shared = NULL;
1374 struct array_cache **alien = NULL;
1375
1376 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001377 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001378 if (!nc)
1379 goto bad;
1380 if (cachep->shared) {
1381 shared = alloc_arraycache(node,
1382 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001383 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001384 if (!shared) {
1385 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001386 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001387 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001388 }
1389 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001390 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001391 if (!alien) {
1392 kfree(shared);
1393 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001394 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001395 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001396 }
1397 cachep->array[cpu] = nc;
1398 l3 = cachep->nodelists[node];
1399 BUG_ON(!l3);
1400
1401 spin_lock_irq(&l3->list_lock);
1402 if (!l3->shared) {
1403 /*
1404 * We are serialised from CPU_DEAD or
1405 * CPU_UP_CANCELLED by the cpucontrol lock
1406 */
1407 l3->shared = shared;
1408 shared = NULL;
1409 }
1410#ifdef CONFIG_NUMA
1411 if (!l3->alien) {
1412 l3->alien = alien;
1413 alien = NULL;
1414 }
1415#endif
1416 spin_unlock_irq(&l3->list_lock);
1417 kfree(shared);
1418 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001419 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1420 slab_set_debugobj_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001421 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001422 init_node_lock_keys(node);
1423
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001424 return 0;
1425bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001426 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001427 return -ENOMEM;
1428}
1429
1430static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1431 unsigned long action, void *hcpu)
1432{
1433 long cpu = (long)hcpu;
1434 int err = 0;
1435
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001437 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001438 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001439 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001440 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001441 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 break;
1443 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001444 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 start_cpu_timer(cpu);
1446 break;
1447#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001448 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001449 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001450 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001451 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001452 * held so that if cache_reap() is invoked it cannot do
1453 * anything expensive but will only modify reap_work
1454 * and reschedule the timer.
1455 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001456 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001457 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001458 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001459 break;
1460 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001461 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001462 start_cpu_timer(cpu);
1463 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001465 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001466 /*
1467 * Even if all the cpus of a node are down, we don't free the
1468 * kmem_list3 of any cache. This to avoid a race between
1469 * cpu_down, and a kmalloc allocation from another cpu for
1470 * memory from the node of the cpu going down. The list3
1471 * structure is usually allocated from kmem_cache_create() and
1472 * gets destroyed at kmem_cache_destroy().
1473 */
Simon Arlott183ff222007-10-20 01:27:18 +02001474 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001475#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001477 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001478 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001479 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001480 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001483 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484}
1485
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001486static struct notifier_block __cpuinitdata cpucache_notifier = {
1487 &cpuup_callback, NULL, 0
1488};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
David Rientjes8f9f8d92010-03-27 19:40:47 -07001490#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1491/*
1492 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1493 * Returns -EBUSY if all objects cannot be drained so that the node is not
1494 * removed.
1495 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001496 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001497 */
1498static int __meminit drain_cache_nodelists_node(int node)
1499{
1500 struct kmem_cache *cachep;
1501 int ret = 0;
1502
Christoph Lameter18004c52012-07-06 15:25:12 -05001503 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001504 struct kmem_list3 *l3;
1505
1506 l3 = cachep->nodelists[node];
1507 if (!l3)
1508 continue;
1509
1510 drain_freelist(cachep, l3, l3->free_objects);
1511
1512 if (!list_empty(&l3->slabs_full) ||
1513 !list_empty(&l3->slabs_partial)) {
1514 ret = -EBUSY;
1515 break;
1516 }
1517 }
1518 return ret;
1519}
1520
1521static int __meminit slab_memory_callback(struct notifier_block *self,
1522 unsigned long action, void *arg)
1523{
1524 struct memory_notify *mnb = arg;
1525 int ret = 0;
1526 int nid;
1527
1528 nid = mnb->status_change_nid;
1529 if (nid < 0)
1530 goto out;
1531
1532 switch (action) {
1533 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001534 mutex_lock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001535 ret = init_cache_nodelists_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001536 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001537 break;
1538 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001539 mutex_lock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001540 ret = drain_cache_nodelists_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001541 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001542 break;
1543 case MEM_ONLINE:
1544 case MEM_OFFLINE:
1545 case MEM_CANCEL_ONLINE:
1546 case MEM_CANCEL_OFFLINE:
1547 break;
1548 }
1549out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001550 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001551}
1552#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1553
Christoph Lametere498be72005-09-09 13:03:32 -07001554/*
1555 * swap the static kmem_list3 with kmalloced memory
1556 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001557static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1558 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001559{
1560 struct kmem_list3 *ptr;
1561
Pekka Enberg83b519e2009-06-10 19:40:04 +03001562 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001563 BUG_ON(!ptr);
1564
Christoph Lametere498be72005-09-09 13:03:32 -07001565 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001566 /*
1567 * Do not assume that spinlocks can be initialized via memcpy:
1568 */
1569 spin_lock_init(&ptr->list_lock);
1570
Christoph Lametere498be72005-09-09 13:03:32 -07001571 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1572 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001573}
1574
Andrew Mortona737b3e2006-03-22 00:08:11 -08001575/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001576 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1577 * size of kmem_list3.
1578 */
1579static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1580{
1581 int node;
1582
1583 for_each_online_node(node) {
1584 cachep->nodelists[node] = &initkmem_list3[index + node];
1585 cachep->nodelists[node]->next_reap = jiffies +
1586 REAPTIMEOUT_LIST3 +
1587 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1588 }
1589}
1590
1591/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001592 * Initialisation. Called after the page allocator have been initialised and
1593 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 */
1595void __init kmem_cache_init(void)
1596{
1597 size_t left_over;
1598 struct cache_sizes *sizes;
1599 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001600 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001601 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001602 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001603
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001604 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001605 use_alien_caches = 0;
1606
Christoph Lametere498be72005-09-09 13:03:32 -07001607 for (i = 0; i < NUM_INIT_LISTS; i++) {
1608 kmem_list3_init(&initkmem_list3[i]);
1609 if (i < MAX_NUMNODES)
1610 cache_cache.nodelists[i] = NULL;
1611 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001612 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613
1614 /*
1615 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001616 * page orders on machines with more than 32MB of memory if
1617 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001619 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001620 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 /* Bootstrap is tricky, because several objects are allocated
1623 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001624 * 1) initialize the cache_cache cache: it contains the struct
1625 * kmem_cache structures of all caches, except cache_cache itself:
1626 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001627 * Initially an __init data area is used for the head array and the
1628 * kmem_list3 structures, it's replaced with a kmalloc allocated
1629 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001631 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001632 * An __init data area is used for the head array.
1633 * 3) Create the remaining kmalloc caches, with minimally sized
1634 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 * 4) Replace the __init data head arrays for cache_cache and the first
1636 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001637 * 5) Replace the __init data for kmem_list3 for cache_cache and
1638 * the other cache's with kmalloc allocated memory.
1639 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 */
1641
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001642 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001643
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 /* 1) create the cache_cache */
Christoph Lameter18004c52012-07-06 15:25:12 -05001645 INIT_LIST_HEAD(&slab_caches);
1646 list_add(&cache_cache.list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 cache_cache.colour_off = cache_line_size();
1648 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001649 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
Eric Dumazet8da34302007-05-06 14:49:29 -07001651 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001652 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001653 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001654 cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001655 nr_node_ids * sizeof(struct kmem_list3 *);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001656 cache_cache.object_size = cache_cache.size;
1657 cache_cache.size = ALIGN(cache_cache.size,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001658 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001659 cache_cache.reciprocal_buffer_size =
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001660 reciprocal_value(cache_cache.size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
Jack Steiner07ed76b2006-03-07 21:55:46 -08001662 for (order = 0; order < MAX_ORDER; order++) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001663 cache_estimate(order, cache_cache.size,
Jack Steiner07ed76b2006-03-07 21:55:46 -08001664 cache_line_size(), 0, &left_over, &cache_cache.num);
1665 if (cache_cache.num)
1666 break;
1667 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001668 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001669 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001670 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001671 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1672 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673
1674 /* 2+3) create the kmalloc caches */
1675 sizes = malloc_sizes;
1676 names = cache_names;
1677
Andrew Mortona737b3e2006-03-22 00:08:11 -08001678 /*
1679 * Initialize the caches that provide memory for the array cache and the
1680 * kmem_list3 structures first. Without this, further allocations will
1681 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001682 */
1683
Christoph Lameter039363f2012-07-06 15:25:10 -05001684 sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001685 sizes[INDEX_AC].cs_size,
1686 ARCH_KMALLOC_MINALIGN,
1687 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001688 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001689
Andrew Mortona737b3e2006-03-22 00:08:11 -08001690 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001691 sizes[INDEX_L3].cs_cachep =
Christoph Lameter039363f2012-07-06 15:25:10 -05001692 __kmem_cache_create(names[INDEX_L3].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001693 sizes[INDEX_L3].cs_size,
1694 ARCH_KMALLOC_MINALIGN,
1695 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001696 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001697 }
Christoph Lametere498be72005-09-09 13:03:32 -07001698
Ingo Molnare0a42722006-06-23 02:03:46 -07001699 slab_early_init = 0;
1700
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001702 /*
1703 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 * This should be particularly beneficial on SMP boxes, as it
1705 * eliminates "false sharing".
1706 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001707 * allow tighter packing of the smaller caches.
1708 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001709 if (!sizes->cs_cachep) {
Christoph Lameter039363f2012-07-06 15:25:10 -05001710 sizes->cs_cachep = __kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001711 sizes->cs_size,
1712 ARCH_KMALLOC_MINALIGN,
1713 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001714 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001715 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001716#ifdef CONFIG_ZONE_DMA
Christoph Lameter039363f2012-07-06 15:25:10 -05001717 sizes->cs_dmacachep = __kmem_cache_create(
Christoph Lameter4b51d662007-02-10 01:43:10 -08001718 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001719 sizes->cs_size,
1720 ARCH_KMALLOC_MINALIGN,
1721 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1722 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001723 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001724#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 sizes++;
1726 names++;
1727 }
1728 /* 4) Replace the bootstrap head arrays */
1729 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001730 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001731
Pekka Enberg83b519e2009-06-10 19:40:04 +03001732 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001733
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001734 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1735 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001736 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001737 /*
1738 * Do not assume that spinlocks can be initialized via memcpy:
1739 */
1740 spin_lock_init(&ptr->lock);
1741
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001743
Pekka Enberg83b519e2009-06-10 19:40:04 +03001744 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001745
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001746 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001747 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001748 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001749 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001750 /*
1751 * Do not assume that spinlocks can be initialized via memcpy:
1752 */
1753 spin_lock_init(&ptr->lock);
1754
Christoph Lametere498be72005-09-09 13:03:32 -07001755 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001756 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 }
Christoph Lametere498be72005-09-09 13:03:32 -07001758 /* 5) Replace the bootstrap kmem_list3's */
1759 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001760 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761
Mel Gorman9c09a952008-01-24 05:49:54 -08001762 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001763 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001764
Christoph Lametere498be72005-09-09 13:03:32 -07001765 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001766 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001767
1768 if (INDEX_AC != INDEX_L3) {
1769 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001770 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001771 }
1772 }
1773 }
1774
Christoph Lameter97d06602012-07-06 15:25:11 -05001775 slab_state = UP;
Pekka Enberg8429db52009-06-12 15:58:59 +03001776}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001777
Pekka Enberg8429db52009-06-12 15:58:59 +03001778void __init kmem_cache_init_late(void)
1779{
1780 struct kmem_cache *cachep;
1781
Christoph Lameter97d06602012-07-06 15:25:11 -05001782 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001783
Peter Zijlstra30765b92011-07-28 23:22:56 +02001784 /* Annotate slab for lockdep -- annotate the malloc caches */
1785 init_lock_keys();
1786
Pekka Enberg8429db52009-06-12 15:58:59 +03001787 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001788 mutex_lock(&slab_mutex);
1789 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001790 if (enable_cpucache(cachep, GFP_NOWAIT))
1791 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001792 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001793
Christoph Lameter97d06602012-07-06 15:25:11 -05001794 /* Done! */
1795 slab_state = FULL;
1796
Andrew Mortona737b3e2006-03-22 00:08:11 -08001797 /*
1798 * Register a cpu startup notifier callback that initializes
1799 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 */
1801 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802
David Rientjes8f9f8d92010-03-27 19:40:47 -07001803#ifdef CONFIG_NUMA
1804 /*
1805 * Register a memory hotplug callback that initializes and frees
1806 * nodelists.
1807 */
1808 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1809#endif
1810
Andrew Mortona737b3e2006-03-22 00:08:11 -08001811 /*
1812 * The reap timers are started later, with a module init call: That part
1813 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 */
1815}
1816
1817static int __init cpucache_init(void)
1818{
1819 int cpu;
1820
Andrew Mortona737b3e2006-03-22 00:08:11 -08001821 /*
1822 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 */
Christoph Lametere498be72005-09-09 13:03:32 -07001824 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001825 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001826
1827 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001828 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 return 0;
1830}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831__initcall(cpucache_init);
1832
Rafael Aquini8bdec192012-03-09 17:27:27 -03001833static noinline void
1834slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1835{
1836 struct kmem_list3 *l3;
1837 struct slab *slabp;
1838 unsigned long flags;
1839 int node;
1840
1841 printk(KERN_WARNING
1842 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1843 nodeid, gfpflags);
1844 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001845 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001846
1847 for_each_online_node(node) {
1848 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1849 unsigned long active_slabs = 0, num_slabs = 0;
1850
1851 l3 = cachep->nodelists[node];
1852 if (!l3)
1853 continue;
1854
1855 spin_lock_irqsave(&l3->list_lock, flags);
1856 list_for_each_entry(slabp, &l3->slabs_full, list) {
1857 active_objs += cachep->num;
1858 active_slabs++;
1859 }
1860 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1861 active_objs += slabp->inuse;
1862 active_slabs++;
1863 }
1864 list_for_each_entry(slabp, &l3->slabs_free, list)
1865 num_slabs++;
1866
1867 free_objects += l3->free_objects;
1868 spin_unlock_irqrestore(&l3->list_lock, flags);
1869
1870 num_slabs += active_slabs;
1871 num_objs = num_slabs * cachep->num;
1872 printk(KERN_WARNING
1873 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1874 node, active_slabs, num_slabs, active_objs, num_objs,
1875 free_objects);
1876 }
1877}
1878
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879/*
1880 * Interface to system's page allocator. No need to hold the cache-lock.
1881 *
1882 * If we requested dmaable memory, we will get it. Even if we
1883 * did not request dmaable memory, we might get it, but that
1884 * would be relatively rare and ignorable.
1885 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001886static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887{
1888 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001889 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 int i;
1891
Luke Yangd6fef9d2006-04-10 22:52:56 -07001892#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001893 /*
1894 * Nommu uses slab's for process anonymous memory allocations, and thus
1895 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001896 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001897 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001898#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001899
Glauber Costaa618e892012-06-14 16:17:21 +04001900 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001901 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1902 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001903
Linus Torvalds517d0862009-06-16 19:50:13 -07001904 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001905 if (!page) {
1906 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1907 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001909 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001911 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001912 if (unlikely(page->pfmemalloc))
1913 pfmemalloc_active = true;
1914
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001915 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001917 add_zone_page_state(page_zone(page),
1918 NR_SLAB_RECLAIMABLE, nr_pages);
1919 else
1920 add_zone_page_state(page_zone(page),
1921 NR_SLAB_UNRECLAIMABLE, nr_pages);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001922 for (i = 0; i < nr_pages; i++) {
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001923 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001924
Mel Gorman072bb0a2012-07-31 16:43:58 -07001925 if (page->pfmemalloc)
1926 SetPageSlabPfmemalloc(page + i);
1927 }
1928
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001929 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1930 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1931
1932 if (cachep->ctor)
1933 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1934 else
1935 kmemcheck_mark_unallocated_pages(page, nr_pages);
1936 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001937
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001938 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939}
1940
1941/*
1942 * Interface to system's page release.
1943 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001944static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001946 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 struct page *page = virt_to_page(addr);
1948 const unsigned long nr_freed = i;
1949
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001950 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001951
Christoph Lameter972d1a72006-09-25 23:31:51 -07001952 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1953 sub_zone_page_state(page_zone(page),
1954 NR_SLAB_RECLAIMABLE, nr_freed);
1955 else
1956 sub_zone_page_state(page_zone(page),
1957 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001959 BUG_ON(!PageSlab(page));
Mel Gorman072bb0a2012-07-31 16:43:58 -07001960 __ClearPageSlabPfmemalloc(page);
Nick Pigginf205b2f2006-03-22 00:08:02 -08001961 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 page++;
1963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 if (current->reclaim_state)
1965 current->reclaim_state->reclaimed_slab += nr_freed;
1966 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967}
1968
1969static void kmem_rcu_free(struct rcu_head *head)
1970{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001971 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001972 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
1974 kmem_freepages(cachep, slab_rcu->addr);
1975 if (OFF_SLAB(cachep))
1976 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1977}
1978
1979#if DEBUG
1980
1981#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001982static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001983 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001985 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001987 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001989 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 return;
1991
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001992 *addr++ = 0x12345678;
1993 *addr++ = caller;
1994 *addr++ = smp_processor_id();
1995 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 {
1997 unsigned long *sptr = &caller;
1998 unsigned long svalue;
1999
2000 while (!kstack_end(sptr)) {
2001 svalue = *sptr++;
2002 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002003 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 size -= sizeof(unsigned long);
2005 if (size <= sizeof(unsigned long))
2006 break;
2007 }
2008 }
2009
2010 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002011 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012}
2013#endif
2014
Pekka Enberg343e0d72006-02-01 03:05:50 -08002015static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002017 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002018 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019
2020 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002021 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022}
2023
2024static void dump_line(char *data, int offset, int limit)
2025{
2026 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07002027 unsigned char error = 0;
2028 int bad_count = 0;
2029
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002030 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07002031 for (i = 0; i < limit; i++) {
2032 if (data[offset + i] != POISON_FREE) {
2033 error = data[offset + i];
2034 bad_count++;
2035 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07002036 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002037 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2038 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07002039
2040 if (bad_count == 1) {
2041 error ^= POISON_FREE;
2042 if (!(error & (error - 1))) {
2043 printk(KERN_ERR "Single bit error detected. Probably "
2044 "bad RAM.\n");
2045#ifdef CONFIG_X86
2046 printk(KERN_ERR "Run memtest86+ or a similar memory "
2047 "test tool.\n");
2048#else
2049 printk(KERN_ERR "Run a memory test tool.\n");
2050#endif
2051 }
2052 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053}
2054#endif
2055
2056#if DEBUG
2057
Pekka Enberg343e0d72006-02-01 03:05:50 -08002058static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059{
2060 int i, size;
2061 char *realobj;
2062
2063 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07002064 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002065 *dbg_redzone1(cachep, objp),
2066 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 }
2068
2069 if (cachep->flags & SLAB_STORE_USER) {
2070 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002071 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002073 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 printk("\n");
2075 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002076 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002077 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002078 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 int limit;
2080 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002081 if (i + limit > size)
2082 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 dump_line(realobj, i, limit);
2084 }
2085}
2086
Pekka Enberg343e0d72006-02-01 03:05:50 -08002087static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088{
2089 char *realobj;
2090 int size, i;
2091 int lines = 0;
2092
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002093 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002094 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002096 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002098 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 exp = POISON_END;
2100 if (realobj[i] != exp) {
2101 int limit;
2102 /* Mismatch ! */
2103 /* Print header */
2104 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002105 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08002106 "Slab corruption (%s): %s start=%p, len=%d\n",
2107 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 print_objinfo(cachep, objp, 0);
2109 }
2110 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002111 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002113 if (i + limit > size)
2114 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 dump_line(realobj, i, limit);
2116 i += 16;
2117 lines++;
2118 /* Limit to 5 lines */
2119 if (lines > 5)
2120 break;
2121 }
2122 }
2123 if (lines != 0) {
2124 /* Print some data about the neighboring objects, if they
2125 * exist:
2126 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002127 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002128 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002130 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002132 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002133 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002135 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 print_objinfo(cachep, objp, 2);
2137 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002138 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002139 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002140 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002142 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 print_objinfo(cachep, objp, 2);
2144 }
2145 }
2146}
2147#endif
2148
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05302150static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002151{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 int i;
2153 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002154 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155
2156 if (cachep->flags & SLAB_POISON) {
2157#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002158 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002159 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002160 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002161 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 else
2163 check_poison_obj(cachep, objp);
2164#else
2165 check_poison_obj(cachep, objp);
2166#endif
2167 }
2168 if (cachep->flags & SLAB_RED_ZONE) {
2169 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2170 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002171 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2173 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002174 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002177}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302179static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002180{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002181}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182#endif
2183
Randy Dunlap911851e2006-03-22 00:08:14 -08002184/**
2185 * slab_destroy - destroy and release all objects in a slab
2186 * @cachep: cache pointer being destroyed
2187 * @slabp: slab pointer being destroyed
2188 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002189 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002190 * Before calling the slab must have been unlinked from the cache. The
2191 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002192 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002193static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002194{
2195 void *addr = slabp->s_mem - slabp->colouroff;
2196
Rabin Vincente79aec22008-07-04 00:40:32 +05302197 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2199 struct slab_rcu *slab_rcu;
2200
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 slab_rcu->cachep = cachep;
2203 slab_rcu->addr = addr;
2204 call_rcu(&slab_rcu->head, kmem_rcu_free);
2205 } else {
2206 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002207 if (OFF_SLAB(cachep))
2208 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 }
2210}
2211
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002212static void __kmem_cache_destroy(struct kmem_cache *cachep)
2213{
2214 int i;
2215 struct kmem_list3 *l3;
2216
2217 for_each_online_cpu(i)
2218 kfree(cachep->array[i]);
2219
2220 /* NUMA: free the list3 structures */
2221 for_each_online_node(i) {
2222 l3 = cachep->nodelists[i];
2223 if (l3) {
2224 kfree(l3->shared);
2225 free_alien_cache(l3->alien);
2226 kfree(l3);
2227 }
2228 }
2229 kmem_cache_free(&cache_cache, cachep);
2230}
2231
2232
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002234 * calculate_slab_order - calculate size (page order) of slabs
2235 * @cachep: pointer to the cache that is being created
2236 * @size: size of objects to be created in this cache.
2237 * @align: required alignment for the objects.
2238 * @flags: slab allocation flags
2239 *
2240 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002241 *
2242 * This could be made much more intelligent. For now, try to avoid using
2243 * high order pages for slabs. When the gfp() functions are more friendly
2244 * towards high-order requests, this should be changed.
2245 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002246static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002247 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002248{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002249 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002250 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002251 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002252
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002253 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002254 unsigned int num;
2255 size_t remainder;
2256
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002257 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002258 if (!num)
2259 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002260
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002261 if (flags & CFLGS_OFF_SLAB) {
2262 /*
2263 * Max number of objs-per-slab for caches which
2264 * use off-slab slabs. Needed to avoid a possible
2265 * looping condition in cache_grow().
2266 */
2267 offslab_limit = size - sizeof(struct slab);
2268 offslab_limit /= sizeof(kmem_bufctl_t);
2269
2270 if (num > offslab_limit)
2271 break;
2272 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002273
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002274 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002275 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002276 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002277 left_over = remainder;
2278
2279 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002280 * A VFS-reclaimable slab tends to have most allocations
2281 * as GFP_NOFS and we really don't want to have to be allocating
2282 * higher-order pages when we are unable to shrink dcache.
2283 */
2284 if (flags & SLAB_RECLAIM_ACCOUNT)
2285 break;
2286
2287 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002288 * Large number of objects is good, but very large slabs are
2289 * currently bad for the gfp()s.
2290 */
David Rientjes543585c2011-10-18 22:09:24 -07002291 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002292 break;
2293
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002294 /*
2295 * Acceptable internal fragmentation?
2296 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002297 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002298 break;
2299 }
2300 return left_over;
2301}
2302
Pekka Enberg83b519e2009-06-10 19:40:04 +03002303static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002304{
Christoph Lameter97d06602012-07-06 15:25:11 -05002305 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002306 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002307
Christoph Lameter97d06602012-07-06 15:25:11 -05002308 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002309 /*
2310 * Note: the first kmem_cache_create must create the cache
2311 * that's used by kmalloc(24), otherwise the creation of
2312 * further caches will BUG().
2313 */
2314 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2315
2316 /*
2317 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2318 * the first cache, then we need to set up all its list3s,
2319 * otherwise the creation of further caches will BUG().
2320 */
2321 set_up_list3s(cachep, SIZE_AC);
2322 if (INDEX_AC == INDEX_L3)
Christoph Lameter97d06602012-07-06 15:25:11 -05002323 slab_state = PARTIAL_L3;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002324 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002325 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002326 } else {
2327 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002328 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002329
Christoph Lameter97d06602012-07-06 15:25:11 -05002330 if (slab_state == PARTIAL_ARRAYCACHE) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002331 set_up_list3s(cachep, SIZE_L3);
Christoph Lameter97d06602012-07-06 15:25:11 -05002332 slab_state = PARTIAL_L3;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002333 } else {
2334 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002335 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002336 cachep->nodelists[node] =
2337 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002338 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002339 BUG_ON(!cachep->nodelists[node]);
2340 kmem_list3_init(cachep->nodelists[node]);
2341 }
2342 }
2343 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002344 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002345 jiffies + REAPTIMEOUT_LIST3 +
2346 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2347
2348 cpu_cache_get(cachep)->avail = 0;
2349 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2350 cpu_cache_get(cachep)->batchcount = 1;
2351 cpu_cache_get(cachep)->touched = 0;
2352 cachep->batchcount = 1;
2353 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002354 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002355}
2356
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002357/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002358 * __kmem_cache_create - Create a cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 * @name: A string which is used in /proc/slabinfo to identify this cache.
2360 * @size: The size of objects to be created in this cache.
2361 * @align: The required alignment for the objects.
2362 * @flags: SLAB flags
2363 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 *
2365 * Returns a ptr to the cache on success, NULL on failure.
2366 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002367 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 *
2369 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002370 * the module calling this has to destroy the cache before getting unloaded.
2371 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 * The flags are
2373 *
2374 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2375 * to catch references to uninitialised memory.
2376 *
2377 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2378 * for buffer overruns.
2379 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2381 * cacheline. This can be beneficial if you're counting cycles as closely
2382 * as davem.
2383 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002384struct kmem_cache *
Christoph Lameter039363f2012-07-06 15:25:10 -05002385__kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002386 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387{
2388 size_t left_over, slab_size, ralign;
Christoph Lameter20cea962012-07-06 15:25:13 -05002389 struct kmem_cache *cachep = NULL;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002390 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393#if FORCED_DEBUG
2394 /*
2395 * Enable redzoning and last user accounting, except for caches with
2396 * large objects, if the increased size would increase the object size
2397 * above the next power of two: caches with object sizes just above a
2398 * power of two have a significant amount of internal fragmentation.
2399 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002400 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2401 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002402 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 if (!(flags & SLAB_DESTROY_BY_RCU))
2404 flags |= SLAB_POISON;
2405#endif
2406 if (flags & SLAB_DESTROY_BY_RCU)
2407 BUG_ON(flags & SLAB_POISON);
2408#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002410 * Always checks flags, a caller might be expecting debug support which
2411 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002413 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414
Andrew Mortona737b3e2006-03-22 00:08:11 -08002415 /*
2416 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 * unaligned accesses for some archs when redzoning is used, and makes
2418 * sure any on-slab bufctl's are also correctly aligned.
2419 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002420 if (size & (BYTES_PER_WORD - 1)) {
2421 size += (BYTES_PER_WORD - 1);
2422 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 }
2424
Andrew Mortona737b3e2006-03-22 00:08:11 -08002425 /* calculate the final buffer alignment: */
2426
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 /* 1) arch recommendation: can be overridden for debug */
2428 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002429 /*
2430 * Default alignment: as specified by the arch code. Except if
2431 * an object is really small, then squeeze multiple objects into
2432 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433 */
2434 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002435 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 ralign /= 2;
2437 } else {
2438 ralign = BYTES_PER_WORD;
2439 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002440
2441 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002442 * Redzoning and user store require word alignment or possibly larger.
2443 * Note this will be overridden by architecture or caller mandated
2444 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002445 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002446 if (flags & SLAB_STORE_USER)
2447 ralign = BYTES_PER_WORD;
2448
2449 if (flags & SLAB_RED_ZONE) {
2450 ralign = REDZONE_ALIGN;
2451 /* If redzoning, ensure that the second redzone is suitably
2452 * aligned, by adjusting the object size accordingly. */
2453 size += REDZONE_ALIGN - 1;
2454 size &= ~(REDZONE_ALIGN - 1);
2455 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002456
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002457 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 if (ralign < ARCH_SLAB_MINALIGN) {
2459 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002461 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 if (ralign < align) {
2463 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002465 /* disable debug if necessary */
2466 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002467 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002468 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002469 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 */
2471 align = ralign;
2472
Pekka Enberg83b519e2009-06-10 19:40:04 +03002473 if (slab_is_available())
2474 gfp = GFP_KERNEL;
2475 else
2476 gfp = GFP_NOWAIT;
2477
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002479 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 if (!cachep)
Christoph Lameter039363f2012-07-06 15:25:10 -05002481 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482
Eric Dumazetb56efcf2011-07-20 19:04:23 +02002483 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002484 cachep->object_size = size;
2485 cachep->align = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487
Pekka Enbergca5f9702006-09-25 23:31:25 -07002488 /*
2489 * Both debugging options require word-alignment which is calculated
2490 * into align above.
2491 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002494 cachep->obj_offset += sizeof(unsigned long long);
2495 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496 }
2497 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002498 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002499 * the real object. But if the second red zone needs to be
2500 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002502 if (flags & SLAB_RED_ZONE)
2503 size += REDZONE_ALIGN;
2504 else
2505 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 }
2507#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002508 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002509 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
Carsten Otte1ab335d2010-08-06 18:19:22 +02002510 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 size = PAGE_SIZE;
2512 }
2513#endif
2514#endif
2515
Ingo Molnare0a42722006-06-23 02:03:46 -07002516 /*
2517 * Determine if the slab management is 'on' or 'off' slab.
2518 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002519 * it too early on. Always use on-slab management when
2520 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002521 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002522 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2523 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 /*
2525 * Size is large, assume best to place the slab management obj
2526 * off-slab (should allow better packing of objs).
2527 */
2528 flags |= CFLGS_OFF_SLAB;
2529
2530 size = ALIGN(size, align);
2531
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002532 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
2534 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002535 printk(KERN_ERR
2536 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 kmem_cache_free(&cache_cache, cachep);
Christoph Lameter039363f2012-07-06 15:25:10 -05002538 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002540 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2541 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542
2543 /*
2544 * If the slab has been placed off-slab, and we have enough space then
2545 * move it on-slab. This is at the expense of any extra colouring.
2546 */
2547 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2548 flags &= ~CFLGS_OFF_SLAB;
2549 left_over -= slab_size;
2550 }
2551
2552 if (flags & CFLGS_OFF_SLAB) {
2553 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002554 slab_size =
2555 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302556
2557#ifdef CONFIG_PAGE_POISONING
2558 /* If we're going to use the generic kernel_map_pages()
2559 * poisoning, then it's going to smash the contents of
2560 * the redzone and userword anyhow, so switch them off.
2561 */
2562 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2563 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2564#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 }
2566
2567 cachep->colour_off = cache_line_size();
2568 /* Offset must be a multiple of the alignment. */
2569 if (cachep->colour_off < align)
2570 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002571 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 cachep->slab_size = slab_size;
2573 cachep->flags = flags;
Glauber Costaa618e892012-06-14 16:17:21 +04002574 cachep->allocflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002575 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002576 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002577 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002578 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002580 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002581 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002582 /*
2583 * This is a possibility for one of the malloc_sizes caches.
2584 * But since we go off slab only for object size greater than
2585 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2586 * this should not happen at all.
2587 * But leave a BUG_ON for some lucky dude.
2588 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002589 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002590 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 cachep->name = name;
2593
Pekka Enberg83b519e2009-06-10 19:40:04 +03002594 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002595 __kmem_cache_destroy(cachep);
Christoph Lameter039363f2012-07-06 15:25:10 -05002596 return NULL;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002597 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598
Peter Zijlstra83835b32011-07-22 15:26:05 +02002599 if (flags & SLAB_DEBUG_OBJECTS) {
2600 /*
2601 * Would deadlock through slab_destroy()->call_rcu()->
2602 * debug_object_activate()->kmem_cache_alloc().
2603 */
2604 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2605
2606 slab_set_debugobj_lock_classes(cachep);
2607 }
2608
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 /* cache setup completed, link it into the list */
Christoph Lameter18004c52012-07-06 15:25:12 -05002610 list_add(&cachep->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 return cachep;
2612}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613
2614#if DEBUG
2615static void check_irq_off(void)
2616{
2617 BUG_ON(!irqs_disabled());
2618}
2619
2620static void check_irq_on(void)
2621{
2622 BUG_ON(irqs_disabled());
2623}
2624
Pekka Enberg343e0d72006-02-01 03:05:50 -08002625static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626{
2627#ifdef CONFIG_SMP
2628 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002629 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630#endif
2631}
Christoph Lametere498be72005-09-09 13:03:32 -07002632
Pekka Enberg343e0d72006-02-01 03:05:50 -08002633static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002634{
2635#ifdef CONFIG_SMP
2636 check_irq_off();
2637 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2638#endif
2639}
2640
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641#else
2642#define check_irq_off() do { } while(0)
2643#define check_irq_on() do { } while(0)
2644#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002645#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646#endif
2647
Christoph Lameteraab22072006-03-22 00:09:06 -08002648static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2649 struct array_cache *ac,
2650 int force, int node);
2651
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652static void do_drain(void *arg)
2653{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002654 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002656 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657
2658 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002659 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002660 spin_lock(&cachep->nodelists[node]->list_lock);
2661 free_block(cachep, ac->entry, ac->avail, node);
2662 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 ac->avail = 0;
2664}
2665
Pekka Enberg343e0d72006-02-01 03:05:50 -08002666static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667{
Christoph Lametere498be72005-09-09 13:03:32 -07002668 struct kmem_list3 *l3;
2669 int node;
2670
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002671 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002673 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002674 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002675 if (l3 && l3->alien)
2676 drain_alien_cache(cachep, l3->alien);
2677 }
2678
2679 for_each_online_node(node) {
2680 l3 = cachep->nodelists[node];
2681 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002682 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002683 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684}
2685
Christoph Lametered11d9e2006-06-30 01:55:45 -07002686/*
2687 * Remove slabs from the list of free slabs.
2688 * Specify the number of slabs to drain in tofree.
2689 *
2690 * Returns the actual number of slabs released.
2691 */
2692static int drain_freelist(struct kmem_cache *cache,
2693 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002695 struct list_head *p;
2696 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698
Christoph Lametered11d9e2006-06-30 01:55:45 -07002699 nr_freed = 0;
2700 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701
Christoph Lametered11d9e2006-06-30 01:55:45 -07002702 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002703 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002704 if (p == &l3->slabs_free) {
2705 spin_unlock_irq(&l3->list_lock);
2706 goto out;
2707 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708
Christoph Lametered11d9e2006-06-30 01:55:45 -07002709 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002711 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712#endif
2713 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002714 /*
2715 * Safe to drop the lock. The slab is no longer linked
2716 * to the cache.
2717 */
2718 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002719 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002720 slab_destroy(cache, slabp);
2721 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002723out:
2724 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725}
2726
Christoph Lameter18004c52012-07-06 15:25:12 -05002727/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002728static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002729{
2730 int ret = 0, i = 0;
2731 struct kmem_list3 *l3;
2732
2733 drain_cpu_caches(cachep);
2734
2735 check_irq_on();
2736 for_each_online_node(i) {
2737 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002738 if (!l3)
2739 continue;
2740
2741 drain_freelist(cachep, l3, l3->free_objects);
2742
2743 ret += !list_empty(&l3->slabs_full) ||
2744 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002745 }
2746 return (ret ? 1 : 0);
2747}
2748
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749/**
2750 * kmem_cache_shrink - Shrink a cache.
2751 * @cachep: The cache to shrink.
2752 *
2753 * Releases as many slabs as possible for a cache.
2754 * To help debugging, a zero exit status indicates all slabs were released.
2755 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002756int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002758 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002759 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002761 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002762 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002763 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002764 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002765 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002766 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767}
2768EXPORT_SYMBOL(kmem_cache_shrink);
2769
2770/**
2771 * kmem_cache_destroy - delete a cache
2772 * @cachep: the cache to destroy
2773 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002774 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 *
2776 * It is expected this function will be called by a module when it is
2777 * unloaded. This will remove the cache completely, and avoid a duplicate
2778 * cache being allocated each time a module is loaded and unloaded, if the
2779 * module doesn't have persistent in-kernel storage across loads and unloads.
2780 *
2781 * The cache must be empty before calling this function.
2782 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002783 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 * during the kmem_cache_destroy().
2785 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002786void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002788 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002791 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002792 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 /*
2794 * the chain is never empty, cache_cache is never destroyed
2795 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002796 list_del(&cachep->list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 if (__cache_shrink(cachep)) {
2798 slab_error(cachep, "Can't free all objects");
Christoph Lameter18004c52012-07-06 15:25:12 -05002799 list_add(&cachep->list, &slab_caches);
2800 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002801 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002802 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 }
2804
2805 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002806 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002808 __kmem_cache_destroy(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002809 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002810 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811}
2812EXPORT_SYMBOL(kmem_cache_destroy);
2813
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002814/*
2815 * Get the memory for a slab management obj.
2816 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2817 * always come from malloc_sizes caches. The slab descriptor cannot
2818 * come from the same cache which is getting created because,
2819 * when we are searching for an appropriate cache for these
2820 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2821 * If we are creating a malloc_sizes cache here it would not be visible to
2822 * kmem_find_general_cachep till the initialization is complete.
2823 * Hence we cannot have slabp_cache same as the original cache.
2824 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002825static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002826 int colour_off, gfp_t local_flags,
2827 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828{
2829 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002830
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 if (OFF_SLAB(cachep)) {
2832 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002833 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002834 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002835 /*
2836 * If the first object in the slab is leaked (it's allocated
2837 * but no one has a reference to it), we want to make sure
2838 * kmemleak does not treat the ->s_mem pointer as a reference
2839 * to the object. Otherwise we will not report the leak.
2840 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002841 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2842 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 if (!slabp)
2844 return NULL;
2845 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002846 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 colour_off += cachep->slab_size;
2848 }
2849 slabp->inuse = 0;
2850 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002851 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002852 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002853 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 return slabp;
2855}
2856
2857static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2858{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002859 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860}
2861
Pekka Enberg343e0d72006-02-01 03:05:50 -08002862static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002863 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864{
2865 int i;
2866
2867 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002868 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869#if DEBUG
2870 /* need to poison the objs? */
2871 if (cachep->flags & SLAB_POISON)
2872 poison_obj(cachep, objp, POISON_FREE);
2873 if (cachep->flags & SLAB_STORE_USER)
2874 *dbg_userword(cachep, objp) = NULL;
2875
2876 if (cachep->flags & SLAB_RED_ZONE) {
2877 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2878 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2879 }
2880 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002881 * Constructors are not allowed to allocate memory from the same
2882 * cache which they are a constructor for. Otherwise, deadlock.
2883 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 */
2885 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002886 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887
2888 if (cachep->flags & SLAB_RED_ZONE) {
2889 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2890 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002891 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2893 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002894 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002896 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002897 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002898 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002899 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900#else
2901 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002902 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002904 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002906 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907}
2908
Pekka Enberg343e0d72006-02-01 03:05:50 -08002909static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002911 if (CONFIG_ZONE_DMA_FLAG) {
2912 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002913 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002914 else
Glauber Costaa618e892012-06-14 16:17:21 +04002915 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002916 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917}
2918
Andrew Mortona737b3e2006-03-22 00:08:11 -08002919static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2920 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002921{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002922 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002923 kmem_bufctl_t next;
2924
2925 slabp->inuse++;
2926 next = slab_bufctl(slabp)[slabp->free];
2927#if DEBUG
2928 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2929 WARN_ON(slabp->nodeid != nodeid);
2930#endif
2931 slabp->free = next;
2932
2933 return objp;
2934}
2935
Andrew Mortona737b3e2006-03-22 00:08:11 -08002936static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2937 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002938{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002939 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002940
2941#if DEBUG
2942 /* Verify that the slab belongs to the intended node */
2943 WARN_ON(slabp->nodeid != nodeid);
2944
Al Viro871751e2006-03-25 03:06:39 -08002945 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002946 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002947 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002948 BUG();
2949 }
2950#endif
2951 slab_bufctl(slabp)[objnr] = slabp->free;
2952 slabp->free = objnr;
2953 slabp->inuse--;
2954}
2955
Pekka Enberg47768742006-06-23 02:03:07 -07002956/*
2957 * Map pages beginning at addr to the given cache and slab. This is required
2958 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002959 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002960 */
2961static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2962 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963{
Pekka Enberg47768742006-06-23 02:03:07 -07002964 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 struct page *page;
2966
Pekka Enberg47768742006-06-23 02:03:07 -07002967 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002968
Pekka Enberg47768742006-06-23 02:03:07 -07002969 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002970 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002971 nr_pages <<= cache->gfporder;
2972
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973 do {
Christoph Lameter35026082012-06-13 10:24:56 -05002974 page->slab_cache = cache;
2975 page->slab_page = slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002977 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978}
2979
2980/*
2981 * Grow (by 1) the number of slabs within a cache. This is called by
2982 * kmem_cache_alloc() when there are no active objs left in a cache.
2983 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002984static int cache_grow(struct kmem_cache *cachep,
2985 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002987 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002988 size_t offset;
2989 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002990 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991
Andrew Mortona737b3e2006-03-22 00:08:11 -08002992 /*
2993 * Be lazy and only check for valid flags here, keeping it out of the
2994 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002996 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2997 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002999 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08003001 l3 = cachep->nodelists[nodeid];
3002 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003
3004 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08003005 offset = l3->colour_next;
3006 l3->colour_next++;
3007 if (l3->colour_next >= cachep->colour)
3008 l3->colour_next = 0;
3009 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08003011 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012
3013 if (local_flags & __GFP_WAIT)
3014 local_irq_enable();
3015
3016 /*
3017 * The test for missing atomic flag is performed here, rather than
3018 * the more obvious place, simply to reduce the critical path length
3019 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
3020 * will eventually be caught here (where it matters).
3021 */
3022 kmem_flagcheck(cachep, flags);
3023
Andrew Mortona737b3e2006-03-22 00:08:11 -08003024 /*
3025 * Get mem for the objs. Attempt to allocate a physical page from
3026 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07003027 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003028 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07003029 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003030 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 goto failed;
3032
3033 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003034 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07003035 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003036 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 goto opps1;
3038
Pekka Enberg47768742006-06-23 02:03:07 -07003039 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040
Christoph Lametera35afb82007-05-16 22:10:57 -07003041 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042
3043 if (local_flags & __GFP_WAIT)
3044 local_irq_disable();
3045 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07003046 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047
3048 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07003049 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003051 l3->free_objects += cachep->num;
3052 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003054opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003056failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 if (local_flags & __GFP_WAIT)
3058 local_irq_disable();
3059 return 0;
3060}
3061
3062#if DEBUG
3063
3064/*
3065 * Perform extra freeing checks:
3066 * - detect bad pointers.
3067 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 */
3069static void kfree_debugcheck(const void *objp)
3070{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 if (!virt_addr_valid(objp)) {
3072 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003073 (unsigned long)objp);
3074 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076}
3077
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003078static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3079{
David Woodhouseb46b8f12007-05-08 00:22:59 -07003080 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003081
3082 redzone1 = *dbg_redzone1(cache, obj);
3083 redzone2 = *dbg_redzone2(cache, obj);
3084
3085 /*
3086 * Redzone is ok.
3087 */
3088 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3089 return;
3090
3091 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3092 slab_error(cache, "double free detected");
3093 else
3094 slab_error(cache, "memory outside object was overwritten");
3095
David Woodhouseb46b8f12007-05-08 00:22:59 -07003096 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003097 obj, redzone1, redzone2);
3098}
3099
Pekka Enberg343e0d72006-02-01 03:05:50 -08003100static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003101 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102{
3103 struct page *page;
3104 unsigned int objnr;
3105 struct slab *slabp;
3106
Matthew Wilcox80cbd912007-11-29 12:05:13 -07003107 BUG_ON(virt_to_cache(objp) != cachep);
3108
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003109 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07003111 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112
Christoph Lameter35026082012-06-13 10:24:56 -05003113 slabp = page->slab_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114
3115 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003116 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3118 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3119 }
3120 if (cachep->flags & SLAB_STORE_USER)
3121 *dbg_userword(cachep, objp) = caller;
3122
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003123 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124
3125 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003126 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127
Al Viro871751e2006-03-25 03:06:39 -08003128#ifdef CONFIG_DEBUG_SLAB_LEAK
3129 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3130#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 if (cachep->flags & SLAB_POISON) {
3132#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003133 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003135 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003136 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 } else {
3138 poison_obj(cachep, objp, POISON_FREE);
3139 }
3140#else
3141 poison_obj(cachep, objp, POISON_FREE);
3142#endif
3143 }
3144 return objp;
3145}
3146
Pekka Enberg343e0d72006-02-01 03:05:50 -08003147static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148{
3149 kmem_bufctl_t i;
3150 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003151
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 /* Check slab's freelist to see if this obj is there. */
3153 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3154 entries++;
3155 if (entries > cachep->num || i >= cachep->num)
3156 goto bad;
3157 }
3158 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003159bad:
3160 printk(KERN_ERR "slab: Internal list corruption detected in "
Dave Jonesface37f2011-11-15 15:03:52 -08003161 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3162 cachep->name, cachep->num, slabp, slabp->inuse,
3163 print_tainted());
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02003164 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3165 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3166 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 BUG();
3168 }
3169}
3170#else
3171#define kfree_debugcheck(x) do { } while(0)
3172#define cache_free_debugcheck(x,objp,z) (objp)
3173#define check_slabp(x,y) do { } while(0)
3174#endif
3175
Mel Gorman072bb0a2012-07-31 16:43:58 -07003176static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3177 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003178{
3179 int batchcount;
3180 struct kmem_list3 *l3;
3181 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003182 int node;
3183
Joe Korty6d2144d2008-03-05 15:04:59 -08003184 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003185 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07003186 if (unlikely(force_refill))
3187 goto force_grow;
3188retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003189 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190 batchcount = ac->batchcount;
3191 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003192 /*
3193 * If there was little recent activity on this cache, then
3194 * perform only a partial refill. Otherwise we could generate
3195 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 */
3197 batchcount = BATCHREFILL_LIMIT;
3198 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003199 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003200
Christoph Lametere498be72005-09-09 13:03:32 -07003201 BUG_ON(ac->avail > 0 || !l3);
3202 spin_lock(&l3->list_lock);
3203
Christoph Lameter3ded1752006-03-25 03:06:44 -08003204 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003205 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3206 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003207 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003208 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003209
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 while (batchcount > 0) {
3211 struct list_head *entry;
3212 struct slab *slabp;
3213 /* Get slab alloc is to come from. */
3214 entry = l3->slabs_partial.next;
3215 if (entry == &l3->slabs_partial) {
3216 l3->free_touched = 1;
3217 entry = l3->slabs_free.next;
3218 if (entry == &l3->slabs_free)
3219 goto must_grow;
3220 }
3221
3222 slabp = list_entry(entry, struct slab, list);
3223 check_slabp(cachep, slabp);
3224 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003225
3226 /*
3227 * The slab was either on partial or free list so
3228 * there must be at least one object available for
3229 * allocation.
3230 */
roel kluin249b9f32008-10-29 17:18:07 -04003231 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003232
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234 STATS_INC_ALLOCED(cachep);
3235 STATS_INC_ACTIVE(cachep);
3236 STATS_SET_HIGH(cachep);
3237
Mel Gorman072bb0a2012-07-31 16:43:58 -07003238 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3239 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240 }
3241 check_slabp(cachep, slabp);
3242
3243 /* move slabp to correct slabp list: */
3244 list_del(&slabp->list);
3245 if (slabp->free == BUFCTL_END)
3246 list_add(&slabp->list, &l3->slabs_full);
3247 else
3248 list_add(&slabp->list, &l3->slabs_partial);
3249 }
3250
Andrew Mortona737b3e2006-03-22 00:08:11 -08003251must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003253alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003254 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255
3256 if (unlikely(!ac->avail)) {
3257 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003258force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08003259 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003260
Andrew Mortona737b3e2006-03-22 00:08:11 -08003261 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003262 ac = cpu_cache_get(cachep);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003263
3264 /* no objects in sight? abort */
3265 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266 return NULL;
3267
Andrew Mortona737b3e2006-03-22 00:08:11 -08003268 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269 goto retry;
3270 }
3271 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003272
3273 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274}
3275
Andrew Mortona737b3e2006-03-22 00:08:11 -08003276static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3277 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278{
3279 might_sleep_if(flags & __GFP_WAIT);
3280#if DEBUG
3281 kmem_flagcheck(cachep, flags);
3282#endif
3283}
3284
3285#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003286static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3287 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003289 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003293 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003294 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003295 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 else
3297 check_poison_obj(cachep, objp);
3298#else
3299 check_poison_obj(cachep, objp);
3300#endif
3301 poison_obj(cachep, objp, POISON_INUSE);
3302 }
3303 if (cachep->flags & SLAB_STORE_USER)
3304 *dbg_userword(cachep, objp) = caller;
3305
3306 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003307 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3308 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3309 slab_error(cachep, "double free, or memory outside"
3310 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003311 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003312 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003313 objp, *dbg_redzone1(cachep, objp),
3314 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 }
3316 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3317 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3318 }
Al Viro871751e2006-03-25 03:06:39 -08003319#ifdef CONFIG_DEBUG_SLAB_LEAK
3320 {
3321 struct slab *slabp;
3322 unsigned objnr;
3323
Christoph Lameter35026082012-06-13 10:24:56 -05003324 slabp = virt_to_head_page(objp)->slab_page;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003325 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
Al Viro871751e2006-03-25 03:06:39 -08003326 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3327 }
3328#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003329 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003330 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003331 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003332 if (ARCH_SLAB_MINALIGN &&
3333 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003334 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003335 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003336 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 return objp;
3338}
3339#else
3340#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3341#endif
3342
Akinobu Mita773ff602008-12-23 19:37:01 +09003343static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003344{
3345 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003346 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003347
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003348 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003349}
3350
Pekka Enberg343e0d72006-02-01 03:05:50 -08003351static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003353 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003355 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356
Alok N Kataria5c382302005-09-27 21:45:46 -07003357 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003358
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003359 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003362 objp = ac_get_obj(cachep, ac, flags, false);
3363
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003364 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003365 * Allow for the possibility all avail objects are not allowed
3366 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003367 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003368 if (objp) {
3369 STATS_INC_ALLOCHIT(cachep);
3370 goto out;
3371 }
3372 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003374
3375 STATS_INC_ALLOCMISS(cachep);
3376 objp = cache_alloc_refill(cachep, flags, force_refill);
3377 /*
3378 * the 'ac' may be updated by cache_alloc_refill(),
3379 * and kmemleak_erase() requires its correct value.
3380 */
3381 ac = cpu_cache_get(cachep);
3382
3383out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003384 /*
3385 * To avoid a false negative, if an object that is in one of the
3386 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3387 * treat the array pointers as a reference to the object.
3388 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003389 if (objp)
3390 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003391 return objp;
3392}
3393
Christoph Lametere498be72005-09-09 13:03:32 -07003394#ifdef CONFIG_NUMA
3395/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003396 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003397 *
3398 * If we are in_interrupt, then process context, including cpusets and
3399 * mempolicy, may not apply and should not be used for allocation policy.
3400 */
3401static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3402{
3403 int nid_alloc, nid_here;
3404
Christoph Lameter765c4502006-09-27 01:50:08 -07003405 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003406 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003407 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003408 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003409 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003410 else if (current->mempolicy)
Andi Kleene7b691b2012-06-09 02:40:03 -07003411 nid_alloc = slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003412 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003413 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003414 return NULL;
3415}
3416
3417/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003418 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003419 * certain node and fall back is permitted. First we scan all the
3420 * available nodelists for available objects. If that fails then we
3421 * perform an allocation without specifying a node. This allows the page
3422 * allocator to do its reclaim / fallback magic. We then insert the
3423 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003424 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003425static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003426{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003427 struct zonelist *zonelist;
3428 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003429 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003430 struct zone *zone;
3431 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003432 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003433 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003434 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003435
3436 if (flags & __GFP_THISNODE)
3437 return NULL;
3438
Christoph Lameter6cb06222007-10-16 01:25:41 -07003439 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003440
Mel Gormancc9a6c82012-03-21 16:34:11 -07003441retry_cpuset:
3442 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07003443 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003444
Christoph Lameter3c517a62006-12-06 20:33:29 -08003445retry:
3446 /*
3447 * Look through allowed nodes for objects available
3448 * from existing per node queues.
3449 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003450 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3451 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003452
Mel Gorman54a6eb52008-04-28 02:12:16 -07003453 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003454 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003455 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003456 obj = ____cache_alloc_node(cache,
3457 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003458 if (obj)
3459 break;
3460 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003461 }
3462
Christoph Lametercfce6602007-05-06 14:50:17 -07003463 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003464 /*
3465 * This allocation will be performed within the constraints
3466 * of the current cpuset / memory policy requirements.
3467 * We may trigger various forms of reclaim on the allowed
3468 * set and go into memory reserves if necessary.
3469 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003470 if (local_flags & __GFP_WAIT)
3471 local_irq_enable();
3472 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003473 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003474 if (local_flags & __GFP_WAIT)
3475 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003476 if (obj) {
3477 /*
3478 * Insert into the appropriate per node queues
3479 */
3480 nid = page_to_nid(virt_to_page(obj));
3481 if (cache_grow(cache, flags, nid, obj)) {
3482 obj = ____cache_alloc_node(cache,
3483 flags | GFP_THISNODE, nid);
3484 if (!obj)
3485 /*
3486 * Another processor may allocate the
3487 * objects in the slab since we are
3488 * not holding any locks.
3489 */
3490 goto retry;
3491 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003492 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003493 obj = NULL;
3494 }
3495 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003496 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003497
3498 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3499 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003500 return obj;
3501}
3502
3503/*
Christoph Lametere498be72005-09-09 13:03:32 -07003504 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003506static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003507 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003508{
3509 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003510 struct slab *slabp;
3511 struct kmem_list3 *l3;
3512 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003513 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003515 l3 = cachep->nodelists[nodeid];
3516 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003517
Andrew Mortona737b3e2006-03-22 00:08:11 -08003518retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003519 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003520 spin_lock(&l3->list_lock);
3521 entry = l3->slabs_partial.next;
3522 if (entry == &l3->slabs_partial) {
3523 l3->free_touched = 1;
3524 entry = l3->slabs_free.next;
3525 if (entry == &l3->slabs_free)
3526 goto must_grow;
3527 }
Christoph Lametere498be72005-09-09 13:03:32 -07003528
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003529 slabp = list_entry(entry, struct slab, list);
3530 check_spinlock_acquired_node(cachep, nodeid);
3531 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003532
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003533 STATS_INC_NODEALLOCS(cachep);
3534 STATS_INC_ACTIVE(cachep);
3535 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003536
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003537 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003538
Matthew Dobson78d382d2006-02-01 03:05:47 -08003539 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003540 check_slabp(cachep, slabp);
3541 l3->free_objects--;
3542 /* move slabp to correct slabp list: */
3543 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003544
Andrew Mortona737b3e2006-03-22 00:08:11 -08003545 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003546 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003547 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003548 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003549
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003550 spin_unlock(&l3->list_lock);
3551 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003552
Andrew Mortona737b3e2006-03-22 00:08:11 -08003553must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003554 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003555 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003556 if (x)
3557 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003558
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003559 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003560
Andrew Mortona737b3e2006-03-22 00:08:11 -08003561done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003562 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003563}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003564
3565/**
3566 * kmem_cache_alloc_node - Allocate an object on the specified node
3567 * @cachep: The cache to allocate from.
3568 * @flags: See kmalloc().
3569 * @nodeid: node number of the target node.
3570 * @caller: return address of caller, used for debug information
3571 *
3572 * Identical to kmem_cache_alloc but it will allocate memory on the given
3573 * node, which can improve the performance for cpu bound structures.
3574 *
3575 * Fallback to other node is possible if __GFP_THISNODE is not set.
3576 */
3577static __always_inline void *
3578__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3579 void *caller)
3580{
3581 unsigned long save_flags;
3582 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003583 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003584
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003585 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003586
Nick Piggincf40bd12009-01-21 08:12:39 +01003587 lockdep_trace_alloc(flags);
3588
Akinobu Mita773ff602008-12-23 19:37:01 +09003589 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003590 return NULL;
3591
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003592 cache_alloc_debugcheck_before(cachep, flags);
3593 local_irq_save(save_flags);
3594
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003595 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003596 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003597
3598 if (unlikely(!cachep->nodelists[nodeid])) {
3599 /* Node not bootstrapped yet */
3600 ptr = fallback_alloc(cachep, flags);
3601 goto out;
3602 }
3603
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003604 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003605 /*
3606 * Use the locally cached objects if possible.
3607 * However ____cache_alloc does not allow fallback
3608 * to other nodes. It may fail while we still have
3609 * objects on other nodes available.
3610 */
3611 ptr = ____cache_alloc(cachep, flags);
3612 if (ptr)
3613 goto out;
3614 }
3615 /* ___cache_alloc_node can fall back to other nodes */
3616 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3617 out:
3618 local_irq_restore(save_flags);
3619 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003620 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003621 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003622
Pekka Enbergc175eea2008-05-09 20:35:53 +02003623 if (likely(ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003624 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003625
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003626 if (unlikely((flags & __GFP_ZERO) && ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003627 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003628
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003629 return ptr;
3630}
3631
3632static __always_inline void *
3633__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3634{
3635 void *objp;
3636
3637 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3638 objp = alternate_node_alloc(cache, flags);
3639 if (objp)
3640 goto out;
3641 }
3642 objp = ____cache_alloc(cache, flags);
3643
3644 /*
3645 * We may just have run out of memory on the local node.
3646 * ____cache_alloc_node() knows how to locate memory on other nodes
3647 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003648 if (!objp)
3649 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003650
3651 out:
3652 return objp;
3653}
3654#else
3655
3656static __always_inline void *
3657__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3658{
3659 return ____cache_alloc(cachep, flags);
3660}
3661
3662#endif /* CONFIG_NUMA */
3663
3664static __always_inline void *
3665__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3666{
3667 unsigned long save_flags;
3668 void *objp;
3669
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003670 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003671
Nick Piggincf40bd12009-01-21 08:12:39 +01003672 lockdep_trace_alloc(flags);
3673
Akinobu Mita773ff602008-12-23 19:37:01 +09003674 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003675 return NULL;
3676
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003677 cache_alloc_debugcheck_before(cachep, flags);
3678 local_irq_save(save_flags);
3679 objp = __do_cache_alloc(cachep, flags);
3680 local_irq_restore(save_flags);
3681 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003682 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003683 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003684 prefetchw(objp);
3685
Pekka Enbergc175eea2008-05-09 20:35:53 +02003686 if (likely(objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003687 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003688
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003689 if (unlikely((flags & __GFP_ZERO) && objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003690 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003691
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003692 return objp;
3693}
Christoph Lametere498be72005-09-09 13:03:32 -07003694
3695/*
3696 * Caller needs to acquire correct kmem_list's list_lock
3697 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003698static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003699 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700{
3701 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003702 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703
3704 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003705 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707
Mel Gorman072bb0a2012-07-31 16:43:58 -07003708 clear_obj_pfmemalloc(&objpp[i]);
3709 objp = objpp[i];
3710
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003711 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003712 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003714 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003716 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003718 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719 check_slabp(cachep, slabp);
3720
3721 /* fixup slab chains */
3722 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003723 if (l3->free_objects > l3->free_limit) {
3724 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003725 /* No need to drop any previously held
3726 * lock here, even if we have a off-slab slab
3727 * descriptor it is guaranteed to come from
3728 * a different cache, refer to comments before
3729 * alloc_slabmgmt.
3730 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 slab_destroy(cachep, slabp);
3732 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003733 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 }
3735 } else {
3736 /* Unconditionally move a slab to the end of the
3737 * partial list on free - maximum time for the
3738 * other objects to be freed, too.
3739 */
Christoph Lametere498be72005-09-09 13:03:32 -07003740 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 }
3742 }
3743}
3744
Pekka Enberg343e0d72006-02-01 03:05:50 -08003745static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746{
3747 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003748 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003749 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750
3751 batchcount = ac->batchcount;
3752#if DEBUG
3753 BUG_ON(!batchcount || batchcount > ac->avail);
3754#endif
3755 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003756 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003757 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003758 if (l3->shared) {
3759 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003760 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761 if (max) {
3762 if (batchcount > max)
3763 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003764 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003765 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 shared_array->avail += batchcount;
3767 goto free_done;
3768 }
3769 }
3770
Christoph Lameterff694162005-09-22 21:44:02 -07003771 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003772free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773#if STATS
3774 {
3775 int i = 0;
3776 struct list_head *p;
3777
Christoph Lametere498be72005-09-09 13:03:32 -07003778 p = l3->slabs_free.next;
3779 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780 struct slab *slabp;
3781
3782 slabp = list_entry(p, struct slab, list);
3783 BUG_ON(slabp->inuse);
3784
3785 i++;
3786 p = p->next;
3787 }
3788 STATS_SET_FREEABLE(cachep, i);
3789 }
3790#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003791 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003793 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794}
3795
3796/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003797 * Release an obj back to its cache. If the obj has a constructed state, it must
3798 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003800static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3801 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003803 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804
3805 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003806 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003807 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003809 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003810
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003811 /*
3812 * Skip calling cache_free_alien() when the platform is not numa.
3813 * This will avoid cache misses that happen while accessing slabp (which
3814 * is per page memory reference) to get nodeid. Instead use a global
3815 * variable to skip the call, which is mostly likely to be present in
3816 * the cache.
3817 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003818 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003819 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003820
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 if (likely(ac->avail < ac->limit)) {
3822 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823 } else {
3824 STATS_INC_FREEMISS(cachep);
3825 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003827
Mel Gorman072bb0a2012-07-31 16:43:58 -07003828 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829}
3830
3831/**
3832 * kmem_cache_alloc - Allocate an object
3833 * @cachep: The cache to allocate from.
3834 * @flags: See kmalloc().
3835 *
3836 * Allocate an object from this cache. The flags are only relevant
3837 * if the cache has no available objects.
3838 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003839void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003841 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3842
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003843 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003844 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003845
3846 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847}
3848EXPORT_SYMBOL(kmem_cache_alloc);
3849
Li Zefan0f24f122009-12-11 15:45:30 +08003850#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003851void *
3852kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003853{
Steven Rostedt85beb582010-11-24 16:23:34 -05003854 void *ret;
3855
3856 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3857
3858 trace_kmalloc(_RET_IP_, ret,
3859 size, slab_buffer_size(cachep), flags);
3860 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003861}
Steven Rostedt85beb582010-11-24 16:23:34 -05003862EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003863#endif
3864
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003866void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3867{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003868 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3869 __builtin_return_address(0));
3870
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003871 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003872 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003873 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003874
3875 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003876}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877EXPORT_SYMBOL(kmem_cache_alloc_node);
3878
Li Zefan0f24f122009-12-11 15:45:30 +08003879#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003880void *kmem_cache_alloc_node_trace(size_t size,
3881 struct kmem_cache *cachep,
3882 gfp_t flags,
3883 int nodeid)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003884{
Steven Rostedt85beb582010-11-24 16:23:34 -05003885 void *ret;
3886
3887 ret = __cache_alloc_node(cachep, flags, nodeid,
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003888 __builtin_return_address(0));
Steven Rostedt85beb582010-11-24 16:23:34 -05003889 trace_kmalloc_node(_RET_IP_, ret,
3890 size, slab_buffer_size(cachep),
3891 flags, nodeid);
3892 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003893}
Steven Rostedt85beb582010-11-24 16:23:34 -05003894EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003895#endif
3896
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003897static __always_inline void *
3898__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003899{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003900 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003901
3902 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003903 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3904 return cachep;
Steven Rostedt85beb582010-11-24 16:23:34 -05003905 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003906}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003907
Li Zefan0bb38a52009-12-11 15:45:50 +08003908#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003909void *__kmalloc_node(size_t size, gfp_t flags, int node)
3910{
3911 return __do_kmalloc_node(size, flags, node,
3912 __builtin_return_address(0));
3913}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003914EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003915
3916void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003917 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003918{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003919 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003920}
3921EXPORT_SYMBOL(__kmalloc_node_track_caller);
3922#else
3923void *__kmalloc_node(size_t size, gfp_t flags, int node)
3924{
3925 return __do_kmalloc_node(size, flags, node, NULL);
3926}
3927EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003928#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003929#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930
3931/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003932 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003934 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003935 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003937static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3938 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003940 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003941 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003943 /* If you want to save a few bytes .text space: replace
3944 * __ with kmem_.
3945 * Then kmalloc uses the uninlined functions instead of the inline
3946 * functions.
3947 */
3948 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003949 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3950 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003951 ret = __cache_alloc(cachep, flags, caller);
3952
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003953 trace_kmalloc((unsigned long) caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003954 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003955
3956 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003957}
3958
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003959
Li Zefan0bb38a52009-12-11 15:45:50 +08003960#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003961void *__kmalloc(size_t size, gfp_t flags)
3962{
Al Viro871751e2006-03-25 03:06:39 -08003963 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964}
3965EXPORT_SYMBOL(__kmalloc);
3966
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003967void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003968{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003969 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003970}
3971EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003972
3973#else
3974void *__kmalloc(size_t size, gfp_t flags)
3975{
3976 return __do_kmalloc(size, flags, NULL);
3977}
3978EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003979#endif
3980
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981/**
3982 * kmem_cache_free - Deallocate an object
3983 * @cachep: The cache the allocation was from.
3984 * @objp: The previously allocated object.
3985 *
3986 * Free an object which was previously allocated from this
3987 * cache.
3988 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003989void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990{
3991 unsigned long flags;
3992
3993 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003994 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003995 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003996 debug_check_no_obj_freed(objp, cachep->object_size);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003997 __cache_free(cachep, objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003999
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02004000 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001}
4002EXPORT_SYMBOL(kmem_cache_free);
4003
4004/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 * kfree - free previously allocated memory
4006 * @objp: pointer returned by kmalloc.
4007 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07004008 * If @objp is NULL, no operation is performed.
4009 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 * Don't free memory not originally allocated by kmalloc()
4011 * or you will run into trouble.
4012 */
4013void kfree(const void *objp)
4014{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004015 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 unsigned long flags;
4017
Pekka Enberg2121db72009-03-25 11:05:57 +02004018 trace_kfree(_RET_IP_, objp);
4019
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004020 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 return;
4022 local_irq_save(flags);
4023 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004024 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004025 debug_check_no_locks_freed(objp, c->object_size);
4026
4027 debug_check_no_obj_freed(objp, c->object_size);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07004028 __cache_free(c, (void *)objp, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 local_irq_restore(flags);
4030}
4031EXPORT_SYMBOL(kfree);
4032
Pekka Enberg343e0d72006-02-01 03:05:50 -08004033unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004035 return cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036}
4037EXPORT_SYMBOL(kmem_cache_size);
4038
Christoph Lametere498be72005-09-09 13:03:32 -07004039/*
Simon Arlott183ff222007-10-20 01:27:18 +02004040 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07004041 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004042static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07004043{
4044 int node;
4045 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08004046 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08004047 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004048
Mel Gorman9c09a952008-01-24 05:49:54 -08004049 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08004050
Paul Menage3395ee02006-12-06 20:32:16 -08004051 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03004052 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08004053 if (!new_alien)
4054 goto fail;
4055 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004056
Eric Dumazet63109842007-05-06 14:49:28 -07004057 new_shared = NULL;
4058 if (cachep->shared) {
4059 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08004060 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004061 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07004062 if (!new_shared) {
4063 free_alien_cache(new_alien);
4064 goto fail;
4065 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08004066 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004067
Andrew Mortona737b3e2006-03-22 00:08:11 -08004068 l3 = cachep->nodelists[node];
4069 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08004070 struct array_cache *shared = l3->shared;
4071
Christoph Lametere498be72005-09-09 13:03:32 -07004072 spin_lock_irq(&l3->list_lock);
4073
Christoph Lametercafeb022006-03-25 03:06:46 -08004074 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08004075 free_block(cachep, shared->entry,
4076 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07004077
Christoph Lametercafeb022006-03-25 03:06:46 -08004078 l3->shared = new_shared;
4079 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07004080 l3->alien = new_alien;
4081 new_alien = NULL;
4082 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004083 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004084 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004085 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08004086 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004087 free_alien_cache(new_alien);
4088 continue;
4089 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03004090 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08004091 if (!l3) {
4092 free_alien_cache(new_alien);
4093 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004094 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004095 }
Christoph Lametere498be72005-09-09 13:03:32 -07004096
4097 kmem_list3_init(l3);
4098 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08004099 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08004100 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07004101 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004102 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004103 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004104 cachep->nodelists[node] = l3;
4105 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004106 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004107
Andrew Mortona737b3e2006-03-22 00:08:11 -08004108fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004109 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08004110 /* Cache is not active yet. Roll back what we did */
4111 node--;
4112 while (node >= 0) {
4113 if (cachep->nodelists[node]) {
4114 l3 = cachep->nodelists[node];
4115
4116 kfree(l3->shared);
4117 free_alien_cache(l3->alien);
4118 kfree(l3);
4119 cachep->nodelists[node] = NULL;
4120 }
4121 node--;
4122 }
4123 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004124 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07004125}
4126
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08004128 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004129 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130};
4131
4132static void do_ccupdate_local(void *info)
4133{
Andrew Mortona737b3e2006-03-22 00:08:11 -08004134 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135 struct array_cache *old;
4136
4137 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08004138 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07004139
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4141 new->new[smp_processor_id()] = old;
4142}
4143
Christoph Lameter18004c52012-07-06 15:25:12 -05004144/* Always called with the slab_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08004145static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004146 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004148 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004149 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004151 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4152 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004153 if (!new)
4154 return -ENOMEM;
4155
Christoph Lametere498be72005-09-09 13:03:32 -07004156 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004157 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004158 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004159 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004160 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004161 kfree(new->new[i]);
4162 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07004163 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164 }
4165 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004166 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004168 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004169
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171 cachep->batchcount = batchcount;
4172 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004173 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174
Christoph Lametere498be72005-09-09 13:03:32 -07004175 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004176 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177 if (!ccold)
4178 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004179 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4180 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4181 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004182 kfree(ccold);
4183 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004184 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004185 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186}
4187
Christoph Lameter18004c52012-07-06 15:25:12 -05004188/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004189static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190{
4191 int err;
4192 int limit, shared;
4193
Andrew Mortona737b3e2006-03-22 00:08:11 -08004194 /*
4195 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196 * - create a LIFO ordering, i.e. return objects that are cache-warm
4197 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004198 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199 * bufctl chains: array operations are cheaper.
4200 * The numbers are guessed, we should auto-tune as described by
4201 * Bonwick.
4202 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004203 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004205 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004206 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004207 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004209 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210 limit = 54;
4211 else
4212 limit = 120;
4213
Andrew Mortona737b3e2006-03-22 00:08:11 -08004214 /*
4215 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004216 * allocation behaviour: Most allocs on one cpu, most free operations
4217 * on another cpu. For these cases, an efficient object passing between
4218 * cpus is necessary. This is provided by a shared array. The array
4219 * replaces Bonwick's magazine layer.
4220 * On uniprocessor, it's functionally equivalent (but less efficient)
4221 * to a larger limit. Thus disabled by default.
4222 */
4223 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004224 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226
4227#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004228 /*
4229 * With debugging enabled, large batchcount lead to excessively long
4230 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231 */
4232 if (limit > 32)
4233 limit = 32;
4234#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004235 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 if (err)
4237 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004238 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004239 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240}
4241
Christoph Lameter1b552532006-03-22 00:09:07 -08004242/*
4243 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004244 * necessary. Note that the l3 listlock also protects the array_cache
4245 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004246 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004247static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004248 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249{
4250 int tofree;
4251
Christoph Lameter1b552532006-03-22 00:09:07 -08004252 if (!ac || !ac->avail)
4253 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254 if (ac->touched && !force) {
4255 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004256 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004257 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004258 if (ac->avail) {
4259 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4260 if (tofree > ac->avail)
4261 tofree = (ac->avail + 1) / 2;
4262 free_block(cachep, ac->entry, tofree, node);
4263 ac->avail -= tofree;
4264 memmove(ac->entry, &(ac->entry[tofree]),
4265 sizeof(void *) * ac->avail);
4266 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004267 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004268 }
4269}
4270
4271/**
4272 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004273 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004274 *
4275 * Called from workqueue/eventd every few seconds.
4276 * Purpose:
4277 * - clear the per-cpu caches for this CPU.
4278 * - return freeable pages to the main free memory pool.
4279 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004280 * If we cannot acquire the cache chain mutex then just give up - we'll try
4281 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004282 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004283static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004284{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004285 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004286 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004287 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004288 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289
Christoph Lameter18004c52012-07-06 15:25:12 -05004290 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004292 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004293
Christoph Lameter18004c52012-07-06 15:25:12 -05004294 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004295 check_irq_on();
4296
Christoph Lameter35386e32006-03-22 00:09:05 -08004297 /*
4298 * We only take the l3 lock if absolutely necessary and we
4299 * have established with reasonable certainty that
4300 * we can do some work if the lock was obtained.
4301 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004302 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004303
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004304 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004305
Christoph Lameteraab22072006-03-22 00:09:06 -08004306 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004307
Christoph Lameter35386e32006-03-22 00:09:05 -08004308 /*
4309 * These are racy checks but it does not matter
4310 * if we skip one check or scan twice.
4311 */
Christoph Lametere498be72005-09-09 13:03:32 -07004312 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004313 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314
Christoph Lametere498be72005-09-09 13:03:32 -07004315 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316
Christoph Lameteraab22072006-03-22 00:09:06 -08004317 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004318
Christoph Lametered11d9e2006-06-30 01:55:45 -07004319 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004320 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004321 else {
4322 int freed;
4323
4324 freed = drain_freelist(searchp, l3, (l3->free_limit +
4325 5 * searchp->num - 1) / (5 * searchp->num));
4326 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004328next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329 cond_resched();
4330 }
4331 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004332 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004333 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004334out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004335 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004336 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337}
4338
Linus Torvalds158a9622008-01-02 13:04:48 -08004339#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004340
Pekka Enberg85289f92006-01-08 01:00:36 -08004341static void print_slabinfo_header(struct seq_file *m)
4342{
4343 /*
4344 * Output format version, so at least we can change it
4345 * without _too_ many complaints.
4346 */
4347#if STATS
4348 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4349#else
4350 seq_puts(m, "slabinfo - version: 2.1\n");
4351#endif
4352 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4353 "<objperslab> <pagesperslab>");
4354 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4355 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4356#if STATS
4357 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004358 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004359 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4360#endif
4361 seq_putc(m, '\n');
4362}
4363
Linus Torvalds1da177e2005-04-16 15:20:36 -07004364static void *s_start(struct seq_file *m, loff_t *pos)
4365{
4366 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004367
Christoph Lameter18004c52012-07-06 15:25:12 -05004368 mutex_lock(&slab_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004369 if (!n)
4370 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004371
Christoph Lameter18004c52012-07-06 15:25:12 -05004372 return seq_list_start(&slab_caches, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373}
4374
4375static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4376{
Christoph Lameter18004c52012-07-06 15:25:12 -05004377 return seq_list_next(p, &slab_caches, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004378}
4379
4380static void s_stop(struct seq_file *m, void *p)
4381{
Christoph Lameter18004c52012-07-06 15:25:12 -05004382 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004383}
4384
4385static int s_show(struct seq_file *m, void *p)
4386{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004387 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004388 struct slab *slabp;
4389 unsigned long active_objs;
4390 unsigned long num_objs;
4391 unsigned long active_slabs = 0;
4392 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004393 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004394 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004395 int node;
4396 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004397
Linus Torvalds1da177e2005-04-16 15:20:36 -07004398 active_objs = 0;
4399 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004400 for_each_online_node(node) {
4401 l3 = cachep->nodelists[node];
4402 if (!l3)
4403 continue;
4404
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004405 check_irq_on();
4406 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004407
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004408 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004409 if (slabp->inuse != cachep->num && !error)
4410 error = "slabs_full accounting error";
4411 active_objs += cachep->num;
4412 active_slabs++;
4413 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004414 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004415 if (slabp->inuse == cachep->num && !error)
4416 error = "slabs_partial inuse accounting error";
4417 if (!slabp->inuse && !error)
4418 error = "slabs_partial/inuse accounting error";
4419 active_objs += slabp->inuse;
4420 active_slabs++;
4421 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004422 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004423 if (slabp->inuse && !error)
4424 error = "slabs_free/inuse accounting error";
4425 num_slabs++;
4426 }
4427 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004428 if (l3->shared)
4429 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004430
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004431 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004432 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004433 num_slabs += active_slabs;
4434 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004435 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004436 error = "free_objects accounting error";
4437
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004438 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004439 if (error)
4440 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4441
4442 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004443 name, active_objs, num_objs, cachep->size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004444 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004445 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004446 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004447 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004448 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004449#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004450 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004451 unsigned long high = cachep->high_mark;
4452 unsigned long allocs = cachep->num_allocations;
4453 unsigned long grown = cachep->grown;
4454 unsigned long reaped = cachep->reaped;
4455 unsigned long errors = cachep->errors;
4456 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004457 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004458 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004459 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460
Joe Perchese92dd4f2010-03-26 19:27:58 -07004461 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4462 "%4lu %4lu %4lu %4lu %4lu",
4463 allocs, high, grown,
4464 reaped, errors, max_freeable, node_allocs,
4465 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004466 }
4467 /* cpu stats */
4468 {
4469 unsigned long allochit = atomic_read(&cachep->allochit);
4470 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4471 unsigned long freehit = atomic_read(&cachep->freehit);
4472 unsigned long freemiss = atomic_read(&cachep->freemiss);
4473
4474 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004475 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004476 }
4477#endif
4478 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479 return 0;
4480}
4481
4482/*
4483 * slabinfo_op - iterator that generates /proc/slabinfo
4484 *
4485 * Output layout:
4486 * cache-name
4487 * num-active-objs
4488 * total-objs
4489 * object size
4490 * num-active-slabs
4491 * total-slabs
4492 * num-pages-per-slab
4493 * + further values on SMP and with statistics enabled
4494 */
4495
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004496static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004497 .start = s_start,
4498 .next = s_next,
4499 .stop = s_stop,
4500 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004501};
4502
4503#define MAX_SLABINFO_WRITE 128
4504/**
4505 * slabinfo_write - Tuning for the slab allocator
4506 * @file: unused
4507 * @buffer: user buffer
4508 * @count: data length
4509 * @ppos: unused
4510 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004511static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004512 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004513{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004514 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004515 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004516 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004517
Linus Torvalds1da177e2005-04-16 15:20:36 -07004518 if (count > MAX_SLABINFO_WRITE)
4519 return -EINVAL;
4520 if (copy_from_user(&kbuf, buffer, count))
4521 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004522 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004523
4524 tmp = strchr(kbuf, ' ');
4525 if (!tmp)
4526 return -EINVAL;
4527 *tmp = '\0';
4528 tmp++;
4529 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4530 return -EINVAL;
4531
4532 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004533 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004534 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004535 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004536 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004537 if (limit < 1 || batchcount < 1 ||
4538 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004539 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004540 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004541 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004542 batchcount, shared,
4543 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004544 }
4545 break;
4546 }
4547 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004548 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004549 if (res >= 0)
4550 res = count;
4551 return res;
4552}
Al Viro871751e2006-03-25 03:06:39 -08004553
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004554static int slabinfo_open(struct inode *inode, struct file *file)
4555{
4556 return seq_open(file, &slabinfo_op);
4557}
4558
4559static const struct file_operations proc_slabinfo_operations = {
4560 .open = slabinfo_open,
4561 .read = seq_read,
4562 .write = slabinfo_write,
4563 .llseek = seq_lseek,
4564 .release = seq_release,
4565};
4566
Al Viro871751e2006-03-25 03:06:39 -08004567#ifdef CONFIG_DEBUG_SLAB_LEAK
4568
4569static void *leaks_start(struct seq_file *m, loff_t *pos)
4570{
Christoph Lameter18004c52012-07-06 15:25:12 -05004571 mutex_lock(&slab_mutex);
4572 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004573}
4574
4575static inline int add_caller(unsigned long *n, unsigned long v)
4576{
4577 unsigned long *p;
4578 int l;
4579 if (!v)
4580 return 1;
4581 l = n[1];
4582 p = n + 2;
4583 while (l) {
4584 int i = l/2;
4585 unsigned long *q = p + 2 * i;
4586 if (*q == v) {
4587 q[1]++;
4588 return 1;
4589 }
4590 if (*q > v) {
4591 l = i;
4592 } else {
4593 p = q + 2;
4594 l -= i + 1;
4595 }
4596 }
4597 if (++n[1] == n[0])
4598 return 0;
4599 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4600 p[0] = v;
4601 p[1] = 1;
4602 return 1;
4603}
4604
4605static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4606{
4607 void *p;
4608 int i;
4609 if (n[0] == n[1])
4610 return;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004611 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
Al Viro871751e2006-03-25 03:06:39 -08004612 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4613 continue;
4614 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4615 return;
4616 }
4617}
4618
4619static void show_symbol(struct seq_file *m, unsigned long address)
4620{
4621#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004622 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004623 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004624
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004625 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004626 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004627 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004628 seq_printf(m, " [%s]", modname);
4629 return;
4630 }
4631#endif
4632 seq_printf(m, "%p", (void *)address);
4633}
4634
4635static int leaks_show(struct seq_file *m, void *p)
4636{
Thierry Reding0672aa72012-06-22 19:42:49 +02004637 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Al Viro871751e2006-03-25 03:06:39 -08004638 struct slab *slabp;
4639 struct kmem_list3 *l3;
4640 const char *name;
4641 unsigned long *n = m->private;
4642 int node;
4643 int i;
4644
4645 if (!(cachep->flags & SLAB_STORE_USER))
4646 return 0;
4647 if (!(cachep->flags & SLAB_RED_ZONE))
4648 return 0;
4649
4650 /* OK, we can do it */
4651
4652 n[1] = 0;
4653
4654 for_each_online_node(node) {
4655 l3 = cachep->nodelists[node];
4656 if (!l3)
4657 continue;
4658
4659 check_irq_on();
4660 spin_lock_irq(&l3->list_lock);
4661
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004662 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004663 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004664 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004665 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004666 spin_unlock_irq(&l3->list_lock);
4667 }
4668 name = cachep->name;
4669 if (n[0] == n[1]) {
4670 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004671 mutex_unlock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004672 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4673 if (!m->private) {
4674 /* Too bad, we are really out */
4675 m->private = n;
Christoph Lameter18004c52012-07-06 15:25:12 -05004676 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004677 return -ENOMEM;
4678 }
4679 *(unsigned long *)m->private = n[0] * 2;
4680 kfree(n);
Christoph Lameter18004c52012-07-06 15:25:12 -05004681 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004682 /* Now make sure this entry will be retried */
4683 m->count = m->size;
4684 return 0;
4685 }
4686 for (i = 0; i < n[1]; i++) {
4687 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4688 show_symbol(m, n[2*i+2]);
4689 seq_putc(m, '\n');
4690 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004691
Al Viro871751e2006-03-25 03:06:39 -08004692 return 0;
4693}
4694
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004695static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004696 .start = leaks_start,
4697 .next = s_next,
4698 .stop = s_stop,
4699 .show = leaks_show,
4700};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004701
4702static int slabstats_open(struct inode *inode, struct file *file)
4703{
4704 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4705 int ret = -ENOMEM;
4706 if (n) {
4707 ret = seq_open(file, &slabstats_op);
4708 if (!ret) {
4709 struct seq_file *m = file->private_data;
4710 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4711 m->private = n;
4712 n = NULL;
4713 }
4714 kfree(n);
4715 }
4716 return ret;
4717}
4718
4719static const struct file_operations proc_slabstats_operations = {
4720 .open = slabstats_open,
4721 .read = seq_read,
4722 .llseek = seq_lseek,
4723 .release = seq_release_private,
4724};
Al Viro871751e2006-03-25 03:06:39 -08004725#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004726
4727static int __init slab_proc_init(void)
4728{
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004729 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004730#ifdef CONFIG_DEBUG_SLAB_LEAK
4731 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4732#endif
4733 return 0;
4734}
4735module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004736#endif
4737
Manfred Spraul00e145b2005-09-03 15:55:07 -07004738/**
4739 * ksize - get the actual amount of memory allocated for a given object
4740 * @objp: Pointer to the object
4741 *
4742 * kmalloc may internally round up allocations and return more memory
4743 * than requested. ksize() can be used to determine the actual amount of
4744 * memory allocated. The caller may use this additional memory, even though
4745 * a smaller amount of memory was initially specified with the kmalloc call.
4746 * The caller must guarantee that objp points to a valid object previously
4747 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4748 * must not be freed during the duration of the call.
4749 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004750size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004751{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004752 BUG_ON(!objp);
4753 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004754 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004755
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004756 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004757}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004758EXPORT_SYMBOL(ksize);