blob: d011030e9613b4ca16e064cf2ea9261a7faf3434 [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
501/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700502 * Do not go above this order unless 0 objects fit into the slab or
503 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 */
David Rientjes543585c2011-10-18 22:09:24 -0700505#define SLAB_MAX_ORDER_HI 1
506#define SLAB_MAX_ORDER_LO 0
507static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700508static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800510static inline struct kmem_cache *virt_to_cache(const void *obj)
511{
Christoph Lameterb49af682007-05-06 14:49:41 -0700512 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500513 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800514}
515
516static inline struct slab *virt_to_slab(const void *obj)
517{
Christoph Lameterb49af682007-05-06 14:49:41 -0700518 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500519
520 VM_BUG_ON(!PageSlab(page));
521 return page->slab_page;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800522}
523
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800524static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
525 unsigned int idx)
526{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500527 return slab->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800528}
529
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800530/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500531 * We want to avoid an expensive divide : (offset / cache->size)
532 * Using the fact that size is a constant for a particular cache,
533 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
535 */
536static inline unsigned int obj_to_index(const struct kmem_cache *cache,
537 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800538{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800539 u32 offset = (obj - slab->s_mem);
540 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800541}
542
Andrew Mortona737b3e2006-03-22 00:08:11 -0800543/*
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
545 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546struct cache_sizes malloc_sizes[] = {
547#define CACHE(x) { .cs_size = (x) },
548#include <linux/kmalloc_sizes.h>
549 CACHE(ULONG_MAX)
550#undef CACHE
551};
552EXPORT_SYMBOL(malloc_sizes);
553
554/* Must match cache_sizes above. Out of line to keep cache footprint low. */
555struct cache_names {
556 char *name;
557 char *name_dma;
558};
559
560static struct cache_names __initdata cache_names[] = {
561#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800563 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564#undef CACHE
565};
566
567static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800568 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800570 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
572/* internal cache of cache description objs */
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200573static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
Pekka Enberg343e0d72006-02-01 03:05:50 -0800574static struct kmem_cache cache_cache = {
Eric Dumazetb56efcf2011-07-20 19:04:23 +0200575 .nodelists = cache_cache_nodelists,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800576 .batchcount = 1,
577 .limit = BOOT_CPUCACHE_ENTRIES,
578 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500579 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800580 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581};
582
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700583#define BAD_ALIEN_MAGIC 0x01020304ul
584
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200585#ifdef CONFIG_LOCKDEP
586
587/*
588 * Slab sometimes uses the kmalloc slabs to store the slab headers
589 * for other slabs "off slab".
590 * The locking for this is tricky in that it nests within the locks
591 * of all other slabs in a few places; to deal with this special
592 * locking we put on-slab caches into a separate lock-class.
593 *
594 * We set lock class for alien array caches which are up during init.
595 * The lock annotation will be lost if all cpus of a node goes down and
596 * then comes back up during hotplug
597 */
598static struct lock_class_key on_slab_l3_key;
599static struct lock_class_key on_slab_alc_key;
600
Peter Zijlstra83835b32011-07-22 15:26:05 +0200601static struct lock_class_key debugobj_l3_key;
602static struct lock_class_key debugobj_alc_key;
603
604static void slab_set_lock_classes(struct kmem_cache *cachep,
605 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
606 int q)
607{
608 struct array_cache **alc;
609 struct kmem_list3 *l3;
610 int r;
611
612 l3 = cachep->nodelists[q];
613 if (!l3)
614 return;
615
616 lockdep_set_class(&l3->list_lock, l3_key);
617 alc = l3->alien;
618 /*
619 * FIXME: This check for BAD_ALIEN_MAGIC
620 * should go away when common slab code is taught to
621 * work even without alien caches.
622 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
623 * for alloc_alien_cache,
624 */
625 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
626 return;
627 for_each_node(r) {
628 if (alc[r])
629 lockdep_set_class(&alc[r]->lock, alc_key);
630 }
631}
632
633static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
634{
635 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
636}
637
638static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
639{
640 int node;
641
642 for_each_online_node(node)
643 slab_set_debugobj_lock_classes_node(cachep, node);
644}
645
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200646static void init_node_lock_keys(int q)
647{
648 struct cache_sizes *s = malloc_sizes;
649
Christoph Lameter97d06602012-07-06 15:25:11 -0500650 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200651 return;
652
653 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200654 struct kmem_list3 *l3;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200655
656 l3 = s->cs_cachep->nodelists[q];
657 if (!l3 || OFF_SLAB(s->cs_cachep))
Pekka Enberg00afa752009-12-27 14:33:14 +0200658 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200659
660 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
661 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200662 }
663}
664
665static inline void init_lock_keys(void)
666{
667 int node;
668
669 for_each_node(node)
670 init_node_lock_keys(node);
671}
672#else
673static void init_node_lock_keys(int q)
674{
675}
676
677static inline void init_lock_keys(void)
678{
679}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200680
681static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
682{
683}
684
685static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
686{
687}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200688#endif
689
Tejun Heo1871e522009-10-29 22:34:13 +0900690static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
Pekka Enberg343e0d72006-02-01 03:05:50 -0800692static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693{
694 return cachep->array[smp_processor_id()];
695}
696
Andrew Mortona737b3e2006-03-22 00:08:11 -0800697static inline struct kmem_cache *__find_general_cachep(size_t size,
698 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699{
700 struct cache_sizes *csizep = malloc_sizes;
701
702#if DEBUG
703 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800704 * kmem_cache_create(), or __kmalloc(), before
705 * the generic caches are initialized.
706 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700707 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700709 if (!size)
710 return ZERO_SIZE_PTR;
711
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 while (size > csizep->cs_size)
713 csizep++;
714
715 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700716 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 * has cs_{dma,}cachep==NULL. Thus no special case
718 * for large kmalloc calls required.
719 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800720#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 if (unlikely(gfpflags & GFP_DMA))
722 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800723#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724 return csizep->cs_cachep;
725}
726
Adrian Bunkb2213852006-09-25 23:31:02 -0700727static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700728{
729 return __find_general_cachep(size, gfpflags);
730}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700731
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800732static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800734 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
735}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736
Andrew Mortona737b3e2006-03-22 00:08:11 -0800737/*
738 * Calculate the number of objects and left-over bytes for a given buffer size.
739 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800740static void cache_estimate(unsigned long gfporder, size_t buffer_size,
741 size_t align, int flags, size_t *left_over,
742 unsigned int *num)
743{
744 int nr_objs;
745 size_t mgmt_size;
746 size_t slab_size = PAGE_SIZE << gfporder;
747
748 /*
749 * The slab management structure can be either off the slab or
750 * on it. For the latter case, the memory allocated for a
751 * slab is used for:
752 *
753 * - The struct slab
754 * - One kmem_bufctl_t for each object
755 * - Padding to respect alignment of @align
756 * - @buffer_size bytes for each object
757 *
758 * If the slab management structure is off the slab, then the
759 * alignment will already be calculated into the size. Because
760 * the slabs are all pages aligned, the objects will be at the
761 * correct alignment when allocated.
762 */
763 if (flags & CFLGS_OFF_SLAB) {
764 mgmt_size = 0;
765 nr_objs = slab_size / buffer_size;
766
767 if (nr_objs > SLAB_LIMIT)
768 nr_objs = SLAB_LIMIT;
769 } else {
770 /*
771 * Ignore padding for the initial guess. The padding
772 * is at most @align-1 bytes, and @buffer_size is at
773 * least @align. In the worst case, this result will
774 * be one greater than the number of objects that fit
775 * into the memory allocation when taking the padding
776 * into account.
777 */
778 nr_objs = (slab_size - sizeof(struct slab)) /
779 (buffer_size + sizeof(kmem_bufctl_t));
780
781 /*
782 * This calculated number will be either the right
783 * amount, or one greater than what we want.
784 */
785 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
786 > slab_size)
787 nr_objs--;
788
789 if (nr_objs > SLAB_LIMIT)
790 nr_objs = SLAB_LIMIT;
791
792 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800794 *num = nr_objs;
795 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796}
797
Harvey Harrisond40cee22008-04-30 00:55:07 -0700798#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
Andrew Mortona737b3e2006-03-22 00:08:11 -0800800static void __slab_error(const char *function, struct kmem_cache *cachep,
801 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802{
803 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800804 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 dump_stack();
Dave Jones645df232012-09-18 15:54:12 -0400806 add_taint(TAINT_BAD_PAGE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807}
808
Paul Menage3395ee02006-12-06 20:32:16 -0800809/*
810 * By default on NUMA we use alien caches to stage the freeing of
811 * objects allocated from other nodes. This causes massive memory
812 * inefficiencies when using fake NUMA setup to split memory into a
813 * large number of small nodes, so it can be disabled on the command
814 * line
815 */
816
817static int use_alien_caches __read_mostly = 1;
818static int __init noaliencache_setup(char *s)
819{
820 use_alien_caches = 0;
821 return 1;
822}
823__setup("noaliencache", noaliencache_setup);
824
David Rientjes3df1ccc2011-10-18 22:09:28 -0700825static int __init slab_max_order_setup(char *str)
826{
827 get_option(&str, &slab_max_order);
828 slab_max_order = slab_max_order < 0 ? 0 :
829 min(slab_max_order, MAX_ORDER - 1);
830 slab_max_order_set = true;
831
832 return 1;
833}
834__setup("slab_max_order=", slab_max_order_setup);
835
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800836#ifdef CONFIG_NUMA
837/*
838 * Special reaping functions for NUMA systems called from cache_reap().
839 * These take care of doing round robin flushing of alien caches (containing
840 * objects freed on different nodes from which they were allocated) and the
841 * flushing of remote pcps by calling drain_node_pages.
842 */
Tejun Heo1871e522009-10-29 22:34:13 +0900843static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800844
845static void init_reap_node(int cpu)
846{
847 int node;
848
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700849 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800850 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800851 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800852
Tejun Heo1871e522009-10-29 22:34:13 +0900853 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800854}
855
856static void next_reap_node(void)
857{
Christoph Lameter909ea962010-12-08 16:22:55 +0100858 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800859
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800860 node = next_node(node, node_online_map);
861 if (unlikely(node >= MAX_NUMNODES))
862 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100863 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800864}
865
866#else
867#define init_reap_node(cpu) do { } while (0)
868#define next_reap_node(void) do { } while (0)
869#endif
870
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871/*
872 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
873 * via the workqueue/eventd.
874 * Add the CPU number into the expiration time to minimize the possibility of
875 * the CPUs getting into lockstep and contending for the global cache chain
876 * lock.
877 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700878static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879{
Tejun Heo1871e522009-10-29 22:34:13 +0900880 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881
882 /*
883 * When this gets called from do_initcalls via cpucache_init(),
884 * init_workqueues() has already run, so keventd will be setup
885 * at that time.
886 */
David Howells52bad642006-11-22 14:54:01 +0000887 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800888 init_reap_node(cpu);
Arjan van de Ven78b43532010-07-19 10:59:42 -0700889 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800890 schedule_delayed_work_on(cpu, reap_work,
891 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 }
893}
894
Christoph Lametere498be72005-09-09 13:03:32 -0700895static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300896 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800898 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 struct array_cache *nc = NULL;
900
Pekka Enberg83b519e2009-06-10 19:40:04 +0300901 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100902 /*
903 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300904 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100905 * cache the pointers are not cleared and they could be counted as
906 * valid references during a kmemleak scan. Therefore, kmemleak must
907 * not scan such objects.
908 */
909 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 if (nc) {
911 nc->avail = 0;
912 nc->limit = entries;
913 nc->batchcount = batchcount;
914 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700915 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 }
917 return nc;
918}
919
Mel Gorman072bb0a2012-07-31 16:43:58 -0700920static inline bool is_slab_pfmemalloc(struct slab *slabp)
921{
922 struct page *page = virt_to_page(slabp->s_mem);
923
924 return PageSlabPfmemalloc(page);
925}
926
927/* Clears pfmemalloc_active if no slabs have pfmalloc set */
928static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
929 struct array_cache *ac)
930{
931 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
932 struct slab *slabp;
933 unsigned long flags;
934
935 if (!pfmemalloc_active)
936 return;
937
938 spin_lock_irqsave(&l3->list_lock, flags);
939 list_for_each_entry(slabp, &l3->slabs_full, list)
940 if (is_slab_pfmemalloc(slabp))
941 goto out;
942
943 list_for_each_entry(slabp, &l3->slabs_partial, list)
944 if (is_slab_pfmemalloc(slabp))
945 goto out;
946
947 list_for_each_entry(slabp, &l3->slabs_free, list)
948 if (is_slab_pfmemalloc(slabp))
949 goto out;
950
951 pfmemalloc_active = false;
952out:
953 spin_unlock_irqrestore(&l3->list_lock, flags);
954}
955
Mel Gorman381760e2012-07-31 16:44:30 -0700956static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700957 gfp_t flags, bool force_refill)
958{
959 int i;
960 void *objp = ac->entry[--ac->avail];
961
962 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
963 if (unlikely(is_obj_pfmemalloc(objp))) {
964 struct kmem_list3 *l3;
965
966 if (gfp_pfmemalloc_allowed(flags)) {
967 clear_obj_pfmemalloc(&objp);
968 return objp;
969 }
970
971 /* The caller cannot use PFMEMALLOC objects, find another one */
972 for (i = 1; i < ac->avail; i++) {
973 /* If a !PFMEMALLOC object is found, swap them */
974 if (!is_obj_pfmemalloc(ac->entry[i])) {
975 objp = ac->entry[i];
976 ac->entry[i] = ac->entry[ac->avail];
977 ac->entry[ac->avail] = objp;
978 return objp;
979 }
980 }
981
982 /*
983 * If there are empty slabs on the slabs_free list and we are
984 * being forced to refill the cache, mark this one !pfmemalloc.
985 */
986 l3 = cachep->nodelists[numa_mem_id()];
987 if (!list_empty(&l3->slabs_free) && force_refill) {
988 struct slab *slabp = virt_to_slab(objp);
989 ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem));
990 clear_obj_pfmemalloc(&objp);
991 recheck_pfmemalloc_active(cachep, ac);
992 return objp;
993 }
994
995 /* No !PFMEMALLOC objects available */
996 ac->avail++;
997 objp = NULL;
998 }
999
1000 return objp;
1001}
1002
Mel Gorman381760e2012-07-31 16:44:30 -07001003static inline void *ac_get_obj(struct kmem_cache *cachep,
1004 struct array_cache *ac, gfp_t flags, bool force_refill)
1005{
1006 void *objp;
1007
1008 if (unlikely(sk_memalloc_socks()))
1009 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1010 else
1011 objp = ac->entry[--ac->avail];
1012
1013 return objp;
1014}
1015
1016static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -07001017 void *objp)
1018{
1019 if (unlikely(pfmemalloc_active)) {
1020 /* Some pfmemalloc slabs exist, check if this is one */
1021 struct page *page = virt_to_page(objp);
1022 if (PageSlabPfmemalloc(page))
1023 set_obj_pfmemalloc(&objp);
1024 }
1025
Mel Gorman381760e2012-07-31 16:44:30 -07001026 return objp;
1027}
1028
1029static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1030 void *objp)
1031{
1032 if (unlikely(sk_memalloc_socks()))
1033 objp = __ac_put_obj(cachep, ac, objp);
1034
Mel Gorman072bb0a2012-07-31 16:43:58 -07001035 ac->entry[ac->avail++] = objp;
1036}
1037
Christoph Lameter3ded1752006-03-25 03:06:44 -08001038/*
1039 * Transfer objects in one arraycache to another.
1040 * Locking must be handled by the caller.
1041 *
1042 * Return the number of entries transferred.
1043 */
1044static int transfer_objects(struct array_cache *to,
1045 struct array_cache *from, unsigned int max)
1046{
1047 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -07001048 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -08001049
1050 if (!nr)
1051 return 0;
1052
1053 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1054 sizeof(void *) *nr);
1055
1056 from->avail -= nr;
1057 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -08001058 return nr;
1059}
1060
Christoph Lameter765c4502006-09-27 01:50:08 -07001061#ifndef CONFIG_NUMA
1062
1063#define drain_alien_cache(cachep, alien) do { } while (0)
1064#define reap_alien(cachep, l3) do { } while (0)
1065
Pekka Enberg83b519e2009-06-10 19:40:04 +03001066static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -07001067{
1068 return (struct array_cache **)BAD_ALIEN_MAGIC;
1069}
1070
1071static inline void free_alien_cache(struct array_cache **ac_ptr)
1072{
1073}
1074
1075static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1076{
1077 return 0;
1078}
1079
1080static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1081 gfp_t flags)
1082{
1083 return NULL;
1084}
1085
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001086static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001087 gfp_t flags, int nodeid)
1088{
1089 return NULL;
1090}
1091
1092#else /* CONFIG_NUMA */
1093
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001094static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001095static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001096
Pekka Enberg83b519e2009-06-10 19:40:04 +03001097static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001098{
1099 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001100 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001101 int i;
1102
1103 if (limit > 1)
1104 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001105 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001106 if (ac_ptr) {
1107 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001108 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001109 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001110 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001111 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001112 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001113 kfree(ac_ptr[i]);
1114 kfree(ac_ptr);
1115 return NULL;
1116 }
1117 }
1118 }
1119 return ac_ptr;
1120}
1121
Pekka Enberg5295a742006-02-01 03:05:48 -08001122static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001123{
1124 int i;
1125
1126 if (!ac_ptr)
1127 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001128 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001129 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001130 kfree(ac_ptr);
1131}
1132
Pekka Enberg343e0d72006-02-01 03:05:50 -08001133static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001134 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001135{
1136 struct kmem_list3 *rl3 = cachep->nodelists[node];
1137
1138 if (ac->avail) {
1139 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001140 /*
1141 * Stuff objects into the remote nodes shared array first.
1142 * That way we could avoid the overhead of putting the objects
1143 * into the free lists and getting them back later.
1144 */
shin, jacob693f7d32006-04-28 10:54:37 -05001145 if (rl3->shared)
1146 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001147
Christoph Lameterff694162005-09-22 21:44:02 -07001148 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001149 ac->avail = 0;
1150 spin_unlock(&rl3->list_lock);
1151 }
1152}
1153
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001154/*
1155 * Called from cache_reap() to regularly drain alien caches round robin.
1156 */
1157static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1158{
Christoph Lameter909ea962010-12-08 16:22:55 +01001159 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001160
1161 if (l3->alien) {
1162 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001163
1164 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001165 __drain_alien_cache(cachep, ac, node);
1166 spin_unlock_irq(&ac->lock);
1167 }
1168 }
1169}
1170
Andrew Mortona737b3e2006-03-22 00:08:11 -08001171static void drain_alien_cache(struct kmem_cache *cachep,
1172 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001173{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001174 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001175 struct array_cache *ac;
1176 unsigned long flags;
1177
1178 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001179 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001180 if (ac) {
1181 spin_lock_irqsave(&ac->lock, flags);
1182 __drain_alien_cache(cachep, ac, i);
1183 spin_unlock_irqrestore(&ac->lock, flags);
1184 }
1185 }
1186}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001187
Ingo Molnar873623d2006-07-13 14:44:38 +02001188static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001189{
1190 struct slab *slabp = virt_to_slab(objp);
1191 int nodeid = slabp->nodeid;
1192 struct kmem_list3 *l3;
1193 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001194 int node;
1195
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001196 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001197
1198 /*
1199 * Make sure we are not freeing a object from another node to the array
1200 * cache on this cpu.
1201 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001202 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001203 return 0;
1204
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001205 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001206 STATS_INC_NODEFREES(cachep);
1207 if (l3->alien && l3->alien[nodeid]) {
1208 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001209 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001210 if (unlikely(alien->avail == alien->limit)) {
1211 STATS_INC_ACOVERFLOW(cachep);
1212 __drain_alien_cache(cachep, alien, nodeid);
1213 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001214 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001215 spin_unlock(&alien->lock);
1216 } else {
1217 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1218 free_block(cachep, &objp, 1, nodeid);
1219 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1220 }
1221 return 1;
1222}
Christoph Lametere498be72005-09-09 13:03:32 -07001223#endif
1224
David Rientjes8f9f8d92010-03-27 19:40:47 -07001225/*
1226 * Allocates and initializes nodelists for a node on each slab cache, used for
1227 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1228 * will be allocated off-node since memory is not yet online for the new node.
1229 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1230 * already in use.
1231 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001232 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001233 */
1234static int init_cache_nodelists_node(int node)
1235{
1236 struct kmem_cache *cachep;
1237 struct kmem_list3 *l3;
1238 const int memsize = sizeof(struct kmem_list3);
1239
Christoph Lameter18004c52012-07-06 15:25:12 -05001240 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001241 /*
1242 * Set up the size64 kmemlist for cpu before we can
1243 * begin anything. Make sure some other cpu on this
1244 * node has not already allocated this
1245 */
1246 if (!cachep->nodelists[node]) {
1247 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1248 if (!l3)
1249 return -ENOMEM;
1250 kmem_list3_init(l3);
1251 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1252 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1253
1254 /*
1255 * The l3s don't come and go as CPUs come and
Christoph Lameter18004c52012-07-06 15:25:12 -05001256 * go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001257 * protection here.
1258 */
1259 cachep->nodelists[node] = l3;
1260 }
1261
1262 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1263 cachep->nodelists[node]->free_limit =
1264 (1 + nr_cpus_node(node)) *
1265 cachep->batchcount + cachep->num;
1266 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1267 }
1268 return 0;
1269}
1270
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001271static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001273 struct kmem_cache *cachep;
1274 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001275 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301276 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001277
Christoph Lameter18004c52012-07-06 15:25:12 -05001278 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001279 struct array_cache *nc;
1280 struct array_cache *shared;
1281 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001282
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001283 /* cpu is dead; no one can alloc from it. */
1284 nc = cachep->array[cpu];
1285 cachep->array[cpu] = NULL;
1286 l3 = cachep->nodelists[node];
1287
1288 if (!l3)
1289 goto free_array_cache;
1290
1291 spin_lock_irq(&l3->list_lock);
1292
1293 /* Free limit for this kmem_list3 */
1294 l3->free_limit -= cachep->batchcount;
1295 if (nc)
1296 free_block(cachep, nc->entry, nc->avail, node);
1297
Rusty Russell58463c12009-12-17 11:43:12 -06001298 if (!cpumask_empty(mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001299 spin_unlock_irq(&l3->list_lock);
1300 goto free_array_cache;
1301 }
1302
1303 shared = l3->shared;
1304 if (shared) {
1305 free_block(cachep, shared->entry,
1306 shared->avail, node);
1307 l3->shared = NULL;
1308 }
1309
1310 alien = l3->alien;
1311 l3->alien = NULL;
1312
1313 spin_unlock_irq(&l3->list_lock);
1314
1315 kfree(shared);
1316 if (alien) {
1317 drain_alien_cache(cachep, alien);
1318 free_alien_cache(alien);
1319 }
1320free_array_cache:
1321 kfree(nc);
1322 }
1323 /*
1324 * In the previous loop, all the objects were freed to
1325 * the respective cache's slabs, now we can go ahead and
1326 * shrink each nodelist to its limit.
1327 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001328 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001329 l3 = cachep->nodelists[node];
1330 if (!l3)
1331 continue;
1332 drain_freelist(cachep, l3, l3->free_objects);
1333 }
1334}
1335
1336static int __cpuinit cpuup_prepare(long cpu)
1337{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001338 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001339 struct kmem_list3 *l3 = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001340 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001341 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001343 /*
1344 * We need to do this right in the beginning since
1345 * alloc_arraycache's are going to use this list.
1346 * kmalloc_node allows us to add the slab to the right
1347 * kmem_list3 and not this cpu's kmem_list3
1348 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001349 err = init_cache_nodelists_node(node);
1350 if (err < 0)
1351 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001352
1353 /*
1354 * Now we can go ahead with allocating the shared arrays and
1355 * array caches
1356 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001357 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001358 struct array_cache *nc;
1359 struct array_cache *shared = NULL;
1360 struct array_cache **alien = NULL;
1361
1362 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001363 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001364 if (!nc)
1365 goto bad;
1366 if (cachep->shared) {
1367 shared = alloc_arraycache(node,
1368 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001369 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001370 if (!shared) {
1371 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001372 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001373 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001374 }
1375 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001376 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001377 if (!alien) {
1378 kfree(shared);
1379 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001380 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001381 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001382 }
1383 cachep->array[cpu] = nc;
1384 l3 = cachep->nodelists[node];
1385 BUG_ON(!l3);
1386
1387 spin_lock_irq(&l3->list_lock);
1388 if (!l3->shared) {
1389 /*
1390 * We are serialised from CPU_DEAD or
1391 * CPU_UP_CANCELLED by the cpucontrol lock
1392 */
1393 l3->shared = shared;
1394 shared = NULL;
1395 }
1396#ifdef CONFIG_NUMA
1397 if (!l3->alien) {
1398 l3->alien = alien;
1399 alien = NULL;
1400 }
1401#endif
1402 spin_unlock_irq(&l3->list_lock);
1403 kfree(shared);
1404 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001405 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1406 slab_set_debugobj_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001407 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001408 init_node_lock_keys(node);
1409
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001410 return 0;
1411bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001412 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001413 return -ENOMEM;
1414}
1415
1416static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1417 unsigned long action, void *hcpu)
1418{
1419 long cpu = (long)hcpu;
1420 int err = 0;
1421
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001423 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001424 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001425 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001426 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001427 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 break;
1429 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001430 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 start_cpu_timer(cpu);
1432 break;
1433#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001434 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001435 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001436 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001437 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001438 * held so that if cache_reap() is invoked it cannot do
1439 * anything expensive but will only modify reap_work
1440 * and reschedule the timer.
1441 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001442 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001443 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001444 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001445 break;
1446 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001447 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001448 start_cpu_timer(cpu);
1449 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001451 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001452 /*
1453 * Even if all the cpus of a node are down, we don't free the
1454 * kmem_list3 of any cache. This to avoid a race between
1455 * cpu_down, and a kmalloc allocation from another cpu for
1456 * memory from the node of the cpu going down. The list3
1457 * structure is usually allocated from kmem_cache_create() and
1458 * gets destroyed at kmem_cache_destroy().
1459 */
Simon Arlott183ff222007-10-20 01:27:18 +02001460 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001461#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001463 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001464 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001465 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001466 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001469 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470}
1471
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001472static struct notifier_block __cpuinitdata cpucache_notifier = {
1473 &cpuup_callback, NULL, 0
1474};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
David Rientjes8f9f8d92010-03-27 19:40:47 -07001476#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1477/*
1478 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1479 * Returns -EBUSY if all objects cannot be drained so that the node is not
1480 * removed.
1481 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001482 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001483 */
1484static int __meminit drain_cache_nodelists_node(int node)
1485{
1486 struct kmem_cache *cachep;
1487 int ret = 0;
1488
Christoph Lameter18004c52012-07-06 15:25:12 -05001489 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001490 struct kmem_list3 *l3;
1491
1492 l3 = cachep->nodelists[node];
1493 if (!l3)
1494 continue;
1495
1496 drain_freelist(cachep, l3, l3->free_objects);
1497
1498 if (!list_empty(&l3->slabs_full) ||
1499 !list_empty(&l3->slabs_partial)) {
1500 ret = -EBUSY;
1501 break;
1502 }
1503 }
1504 return ret;
1505}
1506
1507static int __meminit slab_memory_callback(struct notifier_block *self,
1508 unsigned long action, void *arg)
1509{
1510 struct memory_notify *mnb = arg;
1511 int ret = 0;
1512 int nid;
1513
1514 nid = mnb->status_change_nid;
1515 if (nid < 0)
1516 goto out;
1517
1518 switch (action) {
1519 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001520 mutex_lock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001521 ret = init_cache_nodelists_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001522 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001523 break;
1524 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001525 mutex_lock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001526 ret = drain_cache_nodelists_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001527 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001528 break;
1529 case MEM_ONLINE:
1530 case MEM_OFFLINE:
1531 case MEM_CANCEL_ONLINE:
1532 case MEM_CANCEL_OFFLINE:
1533 break;
1534 }
1535out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001536 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001537}
1538#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1539
Christoph Lametere498be72005-09-09 13:03:32 -07001540/*
1541 * swap the static kmem_list3 with kmalloced memory
1542 */
David Rientjes8f9f8d92010-03-27 19:40:47 -07001543static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1544 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001545{
1546 struct kmem_list3 *ptr;
1547
Pekka Enberg83b519e2009-06-10 19:40:04 +03001548 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001549 BUG_ON(!ptr);
1550
Christoph Lametere498be72005-09-09 13:03:32 -07001551 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001552 /*
1553 * Do not assume that spinlocks can be initialized via memcpy:
1554 */
1555 spin_lock_init(&ptr->list_lock);
1556
Christoph Lametere498be72005-09-09 13:03:32 -07001557 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1558 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001559}
1560
Andrew Mortona737b3e2006-03-22 00:08:11 -08001561/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001562 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1563 * size of kmem_list3.
1564 */
1565static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1566{
1567 int node;
1568
1569 for_each_online_node(node) {
1570 cachep->nodelists[node] = &initkmem_list3[index + node];
1571 cachep->nodelists[node]->next_reap = jiffies +
1572 REAPTIMEOUT_LIST3 +
1573 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1574 }
1575}
1576
1577/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001578 * Initialisation. Called after the page allocator have been initialised and
1579 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 */
1581void __init kmem_cache_init(void)
1582{
1583 size_t left_over;
1584 struct cache_sizes *sizes;
1585 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001586 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001587 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001588 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001589
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001590 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001591 use_alien_caches = 0;
1592
Christoph Lametere498be72005-09-09 13:03:32 -07001593 for (i = 0; i < NUM_INIT_LISTS; i++) {
1594 kmem_list3_init(&initkmem_list3[i]);
1595 if (i < MAX_NUMNODES)
1596 cache_cache.nodelists[i] = NULL;
1597 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001598 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599
1600 /*
1601 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001602 * page orders on machines with more than 32MB of memory if
1603 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001605 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001606 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 /* Bootstrap is tricky, because several objects are allocated
1609 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001610 * 1) initialize the cache_cache cache: it contains the struct
1611 * kmem_cache structures of all caches, except cache_cache itself:
1612 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001613 * Initially an __init data area is used for the head array and the
1614 * kmem_list3 structures, it's replaced with a kmalloc allocated
1615 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001617 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001618 * An __init data area is used for the head array.
1619 * 3) Create the remaining kmalloc caches, with minimally sized
1620 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 * 4) Replace the __init data head arrays for cache_cache and the first
1622 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001623 * 5) Replace the __init data for kmem_list3 for cache_cache and
1624 * the other cache's with kmalloc allocated memory.
1625 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 */
1627
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001628 node = numa_mem_id();
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001629
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 /* 1) create the cache_cache */
Christoph Lameter18004c52012-07-06 15:25:12 -05001631 INIT_LIST_HEAD(&slab_caches);
1632 list_add(&cache_cache.list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 cache_cache.colour_off = cache_line_size();
1634 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001635 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636
Eric Dumazet8da34302007-05-06 14:49:29 -07001637 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001638 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001639 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001640 cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001641 nr_node_ids * sizeof(struct kmem_list3 *);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001642 cache_cache.object_size = cache_cache.size;
1643 cache_cache.size = ALIGN(cache_cache.size,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001644 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001645 cache_cache.reciprocal_buffer_size =
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001646 reciprocal_value(cache_cache.size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Jack Steiner07ed76b2006-03-07 21:55:46 -08001648 for (order = 0; order < MAX_ORDER; order++) {
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001649 cache_estimate(order, cache_cache.size,
Jack Steiner07ed76b2006-03-07 21:55:46 -08001650 cache_line_size(), 0, &left_over, &cache_cache.num);
1651 if (cache_cache.num)
1652 break;
1653 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001654 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001655 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001656 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001657 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1658 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659
1660 /* 2+3) create the kmalloc caches */
1661 sizes = malloc_sizes;
1662 names = cache_names;
1663
Andrew Mortona737b3e2006-03-22 00:08:11 -08001664 /*
1665 * Initialize the caches that provide memory for the array cache and the
1666 * kmem_list3 structures first. Without this, further allocations will
1667 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001668 */
1669
Christoph Lameter039363f2012-07-06 15:25:10 -05001670 sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001671 sizes[INDEX_AC].cs_size,
1672 ARCH_KMALLOC_MINALIGN,
1673 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001674 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001675
Andrew Mortona737b3e2006-03-22 00:08:11 -08001676 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001677 sizes[INDEX_L3].cs_cachep =
Christoph Lameter039363f2012-07-06 15:25:10 -05001678 __kmem_cache_create(names[INDEX_L3].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001679 sizes[INDEX_L3].cs_size,
1680 ARCH_KMALLOC_MINALIGN,
1681 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001682 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001683 }
Christoph Lametere498be72005-09-09 13:03:32 -07001684
Ingo Molnare0a42722006-06-23 02:03:46 -07001685 slab_early_init = 0;
1686
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001688 /*
1689 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 * This should be particularly beneficial on SMP boxes, as it
1691 * eliminates "false sharing".
1692 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001693 * allow tighter packing of the smaller caches.
1694 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001695 if (!sizes->cs_cachep) {
Christoph Lameter039363f2012-07-06 15:25:10 -05001696 sizes->cs_cachep = __kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001697 sizes->cs_size,
1698 ARCH_KMALLOC_MINALIGN,
1699 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001700 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001701 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001702#ifdef CONFIG_ZONE_DMA
Christoph Lameter039363f2012-07-06 15:25:10 -05001703 sizes->cs_dmacachep = __kmem_cache_create(
Christoph Lameter4b51d662007-02-10 01:43:10 -08001704 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001705 sizes->cs_size,
1706 ARCH_KMALLOC_MINALIGN,
1707 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1708 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001709 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001710#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 sizes++;
1712 names++;
1713 }
1714 /* 4) Replace the bootstrap head arrays */
1715 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001716 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001717
Pekka Enberg83b519e2009-06-10 19:40:04 +03001718 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001719
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001720 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1721 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001722 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001723 /*
1724 * Do not assume that spinlocks can be initialized via memcpy:
1725 */
1726 spin_lock_init(&ptr->lock);
1727
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001729
Pekka Enberg83b519e2009-06-10 19:40:04 +03001730 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001731
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001732 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001733 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001734 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001735 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001736 /*
1737 * Do not assume that spinlocks can be initialized via memcpy:
1738 */
1739 spin_lock_init(&ptr->lock);
1740
Christoph Lametere498be72005-09-09 13:03:32 -07001741 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001742 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 }
Christoph Lametere498be72005-09-09 13:03:32 -07001744 /* 5) Replace the bootstrap kmem_list3's */
1745 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001746 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747
Mel Gorman9c09a952008-01-24 05:49:54 -08001748 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001749 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001750
Christoph Lametere498be72005-09-09 13:03:32 -07001751 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001752 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001753
1754 if (INDEX_AC != INDEX_L3) {
1755 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001756 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001757 }
1758 }
1759 }
1760
Christoph Lameter97d06602012-07-06 15:25:11 -05001761 slab_state = UP;
Pekka Enberg8429db52009-06-12 15:58:59 +03001762}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001763
Pekka Enberg8429db52009-06-12 15:58:59 +03001764void __init kmem_cache_init_late(void)
1765{
1766 struct kmem_cache *cachep;
1767
Christoph Lameter97d06602012-07-06 15:25:11 -05001768 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001769
Pekka Enberg8429db52009-06-12 15:58:59 +03001770 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001771 mutex_lock(&slab_mutex);
1772 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001773 if (enable_cpucache(cachep, GFP_NOWAIT))
1774 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001775 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001776
Michael Wang947ca182012-09-05 10:33:18 +08001777 /* Annotate slab for lockdep -- annotate the malloc caches */
1778 init_lock_keys();
1779
Christoph Lameter97d06602012-07-06 15:25:11 -05001780 /* Done! */
1781 slab_state = FULL;
1782
Andrew Mortona737b3e2006-03-22 00:08:11 -08001783 /*
1784 * Register a cpu startup notifier callback that initializes
1785 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 */
1787 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788
David Rientjes8f9f8d92010-03-27 19:40:47 -07001789#ifdef CONFIG_NUMA
1790 /*
1791 * Register a memory hotplug callback that initializes and frees
1792 * nodelists.
1793 */
1794 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1795#endif
1796
Andrew Mortona737b3e2006-03-22 00:08:11 -08001797 /*
1798 * The reap timers are started later, with a module init call: That part
1799 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 */
1801}
1802
1803static int __init cpucache_init(void)
1804{
1805 int cpu;
1806
Andrew Mortona737b3e2006-03-22 00:08:11 -08001807 /*
1808 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809 */
Christoph Lametere498be72005-09-09 13:03:32 -07001810 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001811 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001812
1813 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001814 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 return 0;
1816}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817__initcall(cpucache_init);
1818
Rafael Aquini8bdec192012-03-09 17:27:27 -03001819static noinline void
1820slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1821{
1822 struct kmem_list3 *l3;
1823 struct slab *slabp;
1824 unsigned long flags;
1825 int node;
1826
1827 printk(KERN_WARNING
1828 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1829 nodeid, gfpflags);
1830 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001831 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001832
1833 for_each_online_node(node) {
1834 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1835 unsigned long active_slabs = 0, num_slabs = 0;
1836
1837 l3 = cachep->nodelists[node];
1838 if (!l3)
1839 continue;
1840
1841 spin_lock_irqsave(&l3->list_lock, flags);
1842 list_for_each_entry(slabp, &l3->slabs_full, list) {
1843 active_objs += cachep->num;
1844 active_slabs++;
1845 }
1846 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1847 active_objs += slabp->inuse;
1848 active_slabs++;
1849 }
1850 list_for_each_entry(slabp, &l3->slabs_free, list)
1851 num_slabs++;
1852
1853 free_objects += l3->free_objects;
1854 spin_unlock_irqrestore(&l3->list_lock, flags);
1855
1856 num_slabs += active_slabs;
1857 num_objs = num_slabs * cachep->num;
1858 printk(KERN_WARNING
1859 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1860 node, active_slabs, num_slabs, active_objs, num_objs,
1861 free_objects);
1862 }
1863}
1864
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865/*
1866 * Interface to system's page allocator. No need to hold the cache-lock.
1867 *
1868 * If we requested dmaable memory, we will get it. Even if we
1869 * did not request dmaable memory, we might get it, but that
1870 * would be relatively rare and ignorable.
1871 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001872static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873{
1874 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001875 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 int i;
1877
Luke Yangd6fef9d2006-04-10 22:52:56 -07001878#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001879 /*
1880 * Nommu uses slab's for process anonymous memory allocations, and thus
1881 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001882 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001883 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001884#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001885
Glauber Costaa618e892012-06-14 16:17:21 +04001886 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001887 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1888 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001889
Linus Torvalds517d0862009-06-16 19:50:13 -07001890 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001891 if (!page) {
1892 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1893 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001895 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001897 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001898 if (unlikely(page->pfmemalloc))
1899 pfmemalloc_active = true;
1900
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001901 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001903 add_zone_page_state(page_zone(page),
1904 NR_SLAB_RECLAIMABLE, nr_pages);
1905 else
1906 add_zone_page_state(page_zone(page),
1907 NR_SLAB_UNRECLAIMABLE, nr_pages);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001908 for (i = 0; i < nr_pages; i++) {
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001909 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001910
Mel Gorman072bb0a2012-07-31 16:43:58 -07001911 if (page->pfmemalloc)
1912 SetPageSlabPfmemalloc(page + i);
1913 }
1914
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001915 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1916 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1917
1918 if (cachep->ctor)
1919 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1920 else
1921 kmemcheck_mark_unallocated_pages(page, nr_pages);
1922 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001923
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001924 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925}
1926
1927/*
1928 * Interface to system's page release.
1929 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001930static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001932 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 struct page *page = virt_to_page(addr);
1934 const unsigned long nr_freed = i;
1935
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001936 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001937
Christoph Lameter972d1a72006-09-25 23:31:51 -07001938 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1939 sub_zone_page_state(page_zone(page),
1940 NR_SLAB_RECLAIMABLE, nr_freed);
1941 else
1942 sub_zone_page_state(page_zone(page),
1943 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001945 BUG_ON(!PageSlab(page));
Mel Gorman072bb0a2012-07-31 16:43:58 -07001946 __ClearPageSlabPfmemalloc(page);
Nick Pigginf205b2f2006-03-22 00:08:02 -08001947 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948 page++;
1949 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 if (current->reclaim_state)
1951 current->reclaim_state->reclaimed_slab += nr_freed;
1952 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953}
1954
1955static void kmem_rcu_free(struct rcu_head *head)
1956{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001957 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001958 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959
1960 kmem_freepages(cachep, slab_rcu->addr);
1961 if (OFF_SLAB(cachep))
1962 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1963}
1964
1965#if DEBUG
1966
1967#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001968static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001969 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001971 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001973 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001975 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 return;
1977
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001978 *addr++ = 0x12345678;
1979 *addr++ = caller;
1980 *addr++ = smp_processor_id();
1981 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 {
1983 unsigned long *sptr = &caller;
1984 unsigned long svalue;
1985
1986 while (!kstack_end(sptr)) {
1987 svalue = *sptr++;
1988 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001989 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 size -= sizeof(unsigned long);
1991 if (size <= sizeof(unsigned long))
1992 break;
1993 }
1994 }
1995
1996 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001997 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998}
1999#endif
2000
Pekka Enberg343e0d72006-02-01 03:05:50 -08002001static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002003 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002004 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005
2006 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002007 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008}
2009
2010static void dump_line(char *data, int offset, int limit)
2011{
2012 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07002013 unsigned char error = 0;
2014 int bad_count = 0;
2015
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002016 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07002017 for (i = 0; i < limit; i++) {
2018 if (data[offset + i] != POISON_FREE) {
2019 error = data[offset + i];
2020 bad_count++;
2021 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07002022 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002023 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2024 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07002025
2026 if (bad_count == 1) {
2027 error ^= POISON_FREE;
2028 if (!(error & (error - 1))) {
2029 printk(KERN_ERR "Single bit error detected. Probably "
2030 "bad RAM.\n");
2031#ifdef CONFIG_X86
2032 printk(KERN_ERR "Run memtest86+ or a similar memory "
2033 "test tool.\n");
2034#else
2035 printk(KERN_ERR "Run a memory test tool.\n");
2036#endif
2037 }
2038 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039}
2040#endif
2041
2042#if DEBUG
2043
Pekka Enberg343e0d72006-02-01 03:05:50 -08002044static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045{
2046 int i, size;
2047 char *realobj;
2048
2049 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07002050 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002051 *dbg_redzone1(cachep, objp),
2052 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 }
2054
2055 if (cachep->flags & SLAB_STORE_USER) {
2056 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002057 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002059 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 printk("\n");
2061 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002062 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002063 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002064 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 int limit;
2066 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002067 if (i + limit > size)
2068 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 dump_line(realobj, i, limit);
2070 }
2071}
2072
Pekka Enberg343e0d72006-02-01 03:05:50 -08002073static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074{
2075 char *realobj;
2076 int size, i;
2077 int lines = 0;
2078
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002079 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002080 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002082 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002084 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 exp = POISON_END;
2086 if (realobj[i] != exp) {
2087 int limit;
2088 /* Mismatch ! */
2089 /* Print header */
2090 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002091 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08002092 "Slab corruption (%s): %s start=%p, len=%d\n",
2093 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 print_objinfo(cachep, objp, 0);
2095 }
2096 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002097 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002099 if (i + limit > size)
2100 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 dump_line(realobj, i, limit);
2102 i += 16;
2103 lines++;
2104 /* Limit to 5 lines */
2105 if (lines > 5)
2106 break;
2107 }
2108 }
2109 if (lines != 0) {
2110 /* Print some data about the neighboring objects, if they
2111 * exist:
2112 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002113 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002114 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002116 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002118 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002119 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002121 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 print_objinfo(cachep, objp, 2);
2123 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002124 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002125 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002126 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002128 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 print_objinfo(cachep, objp, 2);
2130 }
2131 }
2132}
2133#endif
2134
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05302136static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002137{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 int i;
2139 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002140 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
2142 if (cachep->flags & SLAB_POISON) {
2143#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002144 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002145 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002146 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002147 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 else
2149 check_poison_obj(cachep, objp);
2150#else
2151 check_poison_obj(cachep, objp);
2152#endif
2153 }
2154 if (cachep->flags & SLAB_RED_ZONE) {
2155 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2156 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002157 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2159 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002160 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002163}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302165static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002166{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002167}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168#endif
2169
Randy Dunlap911851e2006-03-22 00:08:14 -08002170/**
2171 * slab_destroy - destroy and release all objects in a slab
2172 * @cachep: cache pointer being destroyed
2173 * @slabp: slab pointer being destroyed
2174 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002175 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002176 * Before calling the slab must have been unlinked from the cache. The
2177 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002178 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002179static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002180{
2181 void *addr = slabp->s_mem - slabp->colouroff;
2182
Rabin Vincente79aec22008-07-04 00:40:32 +05302183 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2185 struct slab_rcu *slab_rcu;
2186
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002187 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 slab_rcu->cachep = cachep;
2189 slab_rcu->addr = addr;
2190 call_rcu(&slab_rcu->head, kmem_rcu_free);
2191 } else {
2192 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002193 if (OFF_SLAB(cachep))
2194 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 }
2196}
2197
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002198static void __kmem_cache_destroy(struct kmem_cache *cachep)
2199{
2200 int i;
2201 struct kmem_list3 *l3;
2202
2203 for_each_online_cpu(i)
2204 kfree(cachep->array[i]);
2205
2206 /* NUMA: free the list3 structures */
2207 for_each_online_node(i) {
2208 l3 = cachep->nodelists[i];
2209 if (l3) {
2210 kfree(l3->shared);
2211 free_alien_cache(l3->alien);
2212 kfree(l3);
2213 }
2214 }
2215 kmem_cache_free(&cache_cache, cachep);
2216}
2217
2218
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002220 * calculate_slab_order - calculate size (page order) of slabs
2221 * @cachep: pointer to the cache that is being created
2222 * @size: size of objects to be created in this cache.
2223 * @align: required alignment for the objects.
2224 * @flags: slab allocation flags
2225 *
2226 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002227 *
2228 * This could be made much more intelligent. For now, try to avoid using
2229 * high order pages for slabs. When the gfp() functions are more friendly
2230 * towards high-order requests, this should be changed.
2231 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002232static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002233 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002234{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002235 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002236 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002237 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002238
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002239 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002240 unsigned int num;
2241 size_t remainder;
2242
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002243 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002244 if (!num)
2245 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002246
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002247 if (flags & CFLGS_OFF_SLAB) {
2248 /*
2249 * Max number of objs-per-slab for caches which
2250 * use off-slab slabs. Needed to avoid a possible
2251 * looping condition in cache_grow().
2252 */
2253 offslab_limit = size - sizeof(struct slab);
2254 offslab_limit /= sizeof(kmem_bufctl_t);
2255
2256 if (num > offslab_limit)
2257 break;
2258 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002259
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002260 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002261 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002262 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002263 left_over = remainder;
2264
2265 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002266 * A VFS-reclaimable slab tends to have most allocations
2267 * as GFP_NOFS and we really don't want to have to be allocating
2268 * higher-order pages when we are unable to shrink dcache.
2269 */
2270 if (flags & SLAB_RECLAIM_ACCOUNT)
2271 break;
2272
2273 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002274 * Large number of objects is good, but very large slabs are
2275 * currently bad for the gfp()s.
2276 */
David Rientjes543585c2011-10-18 22:09:24 -07002277 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002278 break;
2279
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002280 /*
2281 * Acceptable internal fragmentation?
2282 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002283 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002284 break;
2285 }
2286 return left_over;
2287}
2288
Pekka Enberg83b519e2009-06-10 19:40:04 +03002289static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002290{
Christoph Lameter97d06602012-07-06 15:25:11 -05002291 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002292 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002293
Christoph Lameter97d06602012-07-06 15:25:11 -05002294 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002295 /*
2296 * Note: the first kmem_cache_create must create the cache
2297 * that's used by kmalloc(24), otherwise the creation of
2298 * further caches will BUG().
2299 */
2300 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2301
2302 /*
2303 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2304 * the first cache, then we need to set up all its list3s,
2305 * otherwise the creation of further caches will BUG().
2306 */
2307 set_up_list3s(cachep, SIZE_AC);
2308 if (INDEX_AC == INDEX_L3)
Christoph Lameter97d06602012-07-06 15:25:11 -05002309 slab_state = PARTIAL_L3;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002310 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002311 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002312 } else {
2313 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002314 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002315
Christoph Lameter97d06602012-07-06 15:25:11 -05002316 if (slab_state == PARTIAL_ARRAYCACHE) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002317 set_up_list3s(cachep, SIZE_L3);
Christoph Lameter97d06602012-07-06 15:25:11 -05002318 slab_state = PARTIAL_L3;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002319 } else {
2320 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002321 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002322 cachep->nodelists[node] =
2323 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002324 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002325 BUG_ON(!cachep->nodelists[node]);
2326 kmem_list3_init(cachep->nodelists[node]);
2327 }
2328 }
2329 }
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002330 cachep->nodelists[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002331 jiffies + REAPTIMEOUT_LIST3 +
2332 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2333
2334 cpu_cache_get(cachep)->avail = 0;
2335 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2336 cpu_cache_get(cachep)->batchcount = 1;
2337 cpu_cache_get(cachep)->touched = 0;
2338 cachep->batchcount = 1;
2339 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002340 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002341}
2342
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002343/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002344 * __kmem_cache_create - Create a cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 * @name: A string which is used in /proc/slabinfo to identify this cache.
2346 * @size: The size of objects to be created in this cache.
2347 * @align: The required alignment for the objects.
2348 * @flags: SLAB flags
2349 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 *
2351 * Returns a ptr to the cache on success, NULL on failure.
2352 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002353 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 *
2355 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002356 * the module calling this has to destroy the cache before getting unloaded.
2357 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 * The flags are
2359 *
2360 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2361 * to catch references to uninitialised memory.
2362 *
2363 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2364 * for buffer overruns.
2365 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2367 * cacheline. This can be beneficial if you're counting cycles as closely
2368 * as davem.
2369 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002370struct kmem_cache *
Christoph Lameter039363f2012-07-06 15:25:10 -05002371__kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002372 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373{
2374 size_t left_over, slab_size, ralign;
Christoph Lameter20cea962012-07-06 15:25:13 -05002375 struct kmem_cache *cachep = NULL;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002376 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379#if FORCED_DEBUG
2380 /*
2381 * Enable redzoning and last user accounting, except for caches with
2382 * large objects, if the increased size would increase the object size
2383 * above the next power of two: caches with object sizes just above a
2384 * power of two have a significant amount of internal fragmentation.
2385 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002386 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2387 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002388 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 if (!(flags & SLAB_DESTROY_BY_RCU))
2390 flags |= SLAB_POISON;
2391#endif
2392 if (flags & SLAB_DESTROY_BY_RCU)
2393 BUG_ON(flags & SLAB_POISON);
2394#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002396 * Always checks flags, a caller might be expecting debug support which
2397 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002399 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400
Andrew Mortona737b3e2006-03-22 00:08:11 -08002401 /*
2402 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 * unaligned accesses for some archs when redzoning is used, and makes
2404 * sure any on-slab bufctl's are also correctly aligned.
2405 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002406 if (size & (BYTES_PER_WORD - 1)) {
2407 size += (BYTES_PER_WORD - 1);
2408 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 }
2410
Andrew Mortona737b3e2006-03-22 00:08:11 -08002411 /* calculate the final buffer alignment: */
2412
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 /* 1) arch recommendation: can be overridden for debug */
2414 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002415 /*
2416 * Default alignment: as specified by the arch code. Except if
2417 * an object is really small, then squeeze multiple objects into
2418 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 */
2420 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002421 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 ralign /= 2;
2423 } else {
2424 ralign = BYTES_PER_WORD;
2425 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002426
2427 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002428 * Redzoning and user store require word alignment or possibly larger.
2429 * Note this will be overridden by architecture or caller mandated
2430 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002431 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002432 if (flags & SLAB_STORE_USER)
2433 ralign = BYTES_PER_WORD;
2434
2435 if (flags & SLAB_RED_ZONE) {
2436 ralign = REDZONE_ALIGN;
2437 /* If redzoning, ensure that the second redzone is suitably
2438 * aligned, by adjusting the object size accordingly. */
2439 size += REDZONE_ALIGN - 1;
2440 size &= ~(REDZONE_ALIGN - 1);
2441 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002442
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002443 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 if (ralign < ARCH_SLAB_MINALIGN) {
2445 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002447 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 if (ralign < align) {
2449 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002451 /* disable debug if necessary */
2452 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002453 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002454 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002455 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 */
2457 align = ralign;
2458
Pekka Enberg83b519e2009-06-10 19:40:04 +03002459 if (slab_is_available())
2460 gfp = GFP_KERNEL;
2461 else
2462 gfp = GFP_NOWAIT;
2463
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002465 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 if (!cachep)
Christoph Lameter039363f2012-07-06 15:25:10 -05002467 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
Eric Dumazetb56efcf2011-07-20 19:04:23 +02002469 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002470 cachep->object_size = size;
2471 cachep->align = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473
Pekka Enbergca5f9702006-09-25 23:31:25 -07002474 /*
2475 * Both debugging options require word-alignment which is calculated
2476 * into align above.
2477 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002480 cachep->obj_offset += sizeof(unsigned long long);
2481 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 }
2483 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002484 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002485 * the real object. But if the second red zone needs to be
2486 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002488 if (flags & SLAB_RED_ZONE)
2489 size += REDZONE_ALIGN;
2490 else
2491 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 }
2493#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002494 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002495 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
Carsten Otte1ab335d2010-08-06 18:19:22 +02002496 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 size = PAGE_SIZE;
2498 }
2499#endif
2500#endif
2501
Ingo Molnare0a42722006-06-23 02:03:46 -07002502 /*
2503 * Determine if the slab management is 'on' or 'off' slab.
2504 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002505 * it too early on. Always use on-slab management when
2506 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002507 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002508 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2509 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510 /*
2511 * Size is large, assume best to place the slab management obj
2512 * off-slab (should allow better packing of objs).
2513 */
2514 flags |= CFLGS_OFF_SLAB;
2515
2516 size = ALIGN(size, align);
2517
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002518 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519
2520 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002521 printk(KERN_ERR
2522 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 kmem_cache_free(&cache_cache, cachep);
Christoph Lameter039363f2012-07-06 15:25:10 -05002524 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002526 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2527 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528
2529 /*
2530 * If the slab has been placed off-slab, and we have enough space then
2531 * move it on-slab. This is at the expense of any extra colouring.
2532 */
2533 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2534 flags &= ~CFLGS_OFF_SLAB;
2535 left_over -= slab_size;
2536 }
2537
2538 if (flags & CFLGS_OFF_SLAB) {
2539 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002540 slab_size =
2541 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302542
2543#ifdef CONFIG_PAGE_POISONING
2544 /* If we're going to use the generic kernel_map_pages()
2545 * poisoning, then it's going to smash the contents of
2546 * the redzone and userword anyhow, so switch them off.
2547 */
2548 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2549 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2550#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 }
2552
2553 cachep->colour_off = cache_line_size();
2554 /* Offset must be a multiple of the alignment. */
2555 if (cachep->colour_off < align)
2556 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002557 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558 cachep->slab_size = slab_size;
2559 cachep->flags = flags;
Glauber Costaa618e892012-06-14 16:17:21 +04002560 cachep->allocflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002561 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002562 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002563 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002564 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002566 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002567 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002568 /*
2569 * This is a possibility for one of the malloc_sizes caches.
2570 * But since we go off slab only for object size greater than
2571 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2572 * this should not happen at all.
2573 * But leave a BUG_ON for some lucky dude.
2574 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002575 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002576 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 cachep->name = name;
2579
Pekka Enberg83b519e2009-06-10 19:40:04 +03002580 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002581 __kmem_cache_destroy(cachep);
Christoph Lameter039363f2012-07-06 15:25:10 -05002582 return NULL;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002583 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584
Peter Zijlstra83835b32011-07-22 15:26:05 +02002585 if (flags & SLAB_DEBUG_OBJECTS) {
2586 /*
2587 * Would deadlock through slab_destroy()->call_rcu()->
2588 * debug_object_activate()->kmem_cache_alloc().
2589 */
2590 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2591
2592 slab_set_debugobj_lock_classes(cachep);
2593 }
2594
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 /* cache setup completed, link it into the list */
Christoph Lameter18004c52012-07-06 15:25:12 -05002596 list_add(&cachep->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 return cachep;
2598}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
2600#if DEBUG
2601static void check_irq_off(void)
2602{
2603 BUG_ON(!irqs_disabled());
2604}
2605
2606static void check_irq_on(void)
2607{
2608 BUG_ON(irqs_disabled());
2609}
2610
Pekka Enberg343e0d72006-02-01 03:05:50 -08002611static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612{
2613#ifdef CONFIG_SMP
2614 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002615 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616#endif
2617}
Christoph Lametere498be72005-09-09 13:03:32 -07002618
Pekka Enberg343e0d72006-02-01 03:05:50 -08002619static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002620{
2621#ifdef CONFIG_SMP
2622 check_irq_off();
2623 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2624#endif
2625}
2626
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627#else
2628#define check_irq_off() do { } while(0)
2629#define check_irq_on() do { } while(0)
2630#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002631#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632#endif
2633
Christoph Lameteraab22072006-03-22 00:09:06 -08002634static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2635 struct array_cache *ac,
2636 int force, int node);
2637
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638static void do_drain(void *arg)
2639{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002640 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002642 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643
2644 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002645 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002646 spin_lock(&cachep->nodelists[node]->list_lock);
2647 free_block(cachep, ac->entry, ac->avail, node);
2648 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 ac->avail = 0;
2650}
2651
Pekka Enberg343e0d72006-02-01 03:05:50 -08002652static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653{
Christoph Lametere498be72005-09-09 13:03:32 -07002654 struct kmem_list3 *l3;
2655 int node;
2656
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002657 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002659 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002660 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002661 if (l3 && l3->alien)
2662 drain_alien_cache(cachep, l3->alien);
2663 }
2664
2665 for_each_online_node(node) {
2666 l3 = cachep->nodelists[node];
2667 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002668 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002669 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670}
2671
Christoph Lametered11d9e2006-06-30 01:55:45 -07002672/*
2673 * Remove slabs from the list of free slabs.
2674 * Specify the number of slabs to drain in tofree.
2675 *
2676 * Returns the actual number of slabs released.
2677 */
2678static int drain_freelist(struct kmem_cache *cache,
2679 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002681 struct list_head *p;
2682 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684
Christoph Lametered11d9e2006-06-30 01:55:45 -07002685 nr_freed = 0;
2686 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687
Christoph Lametered11d9e2006-06-30 01:55:45 -07002688 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002689 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002690 if (p == &l3->slabs_free) {
2691 spin_unlock_irq(&l3->list_lock);
2692 goto out;
2693 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694
Christoph Lametered11d9e2006-06-30 01:55:45 -07002695 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002697 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698#endif
2699 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002700 /*
2701 * Safe to drop the lock. The slab is no longer linked
2702 * to the cache.
2703 */
2704 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002705 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002706 slab_destroy(cache, slabp);
2707 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002709out:
2710 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711}
2712
Christoph Lameter18004c52012-07-06 15:25:12 -05002713/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002714static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002715{
2716 int ret = 0, i = 0;
2717 struct kmem_list3 *l3;
2718
2719 drain_cpu_caches(cachep);
2720
2721 check_irq_on();
2722 for_each_online_node(i) {
2723 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002724 if (!l3)
2725 continue;
2726
2727 drain_freelist(cachep, l3, l3->free_objects);
2728
2729 ret += !list_empty(&l3->slabs_full) ||
2730 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002731 }
2732 return (ret ? 1 : 0);
2733}
2734
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735/**
2736 * kmem_cache_shrink - Shrink a cache.
2737 * @cachep: The cache to shrink.
2738 *
2739 * Releases as many slabs as possible for a cache.
2740 * To help debugging, a zero exit status indicates all slabs were released.
2741 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002742int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002744 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002745 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002747 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002748 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002749 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002750 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002751 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002752 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753}
2754EXPORT_SYMBOL(kmem_cache_shrink);
2755
2756/**
2757 * kmem_cache_destroy - delete a cache
2758 * @cachep: the cache to destroy
2759 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002760 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 *
2762 * It is expected this function will be called by a module when it is
2763 * unloaded. This will remove the cache completely, and avoid a duplicate
2764 * cache being allocated each time a module is loaded and unloaded, if the
2765 * module doesn't have persistent in-kernel storage across loads and unloads.
2766 *
2767 * The cache must be empty before calling this function.
2768 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002769 * The caller must guarantee that no one will allocate memory from the cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 * during the kmem_cache_destroy().
2771 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002772void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002774 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002777 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002778 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 /*
2780 * the chain is never empty, cache_cache is never destroyed
2781 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002782 list_del(&cachep->list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 if (__cache_shrink(cachep)) {
2784 slab_error(cachep, "Can't free all objects");
Christoph Lameter18004c52012-07-06 15:25:12 -05002785 list_add(&cachep->list, &slab_caches);
2786 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002787 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002788 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 }
2790
2791 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenney7ed9f7e2009-06-25 12:31:37 -07002792 rcu_barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002794 __kmem_cache_destroy(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002795 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002796 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797}
2798EXPORT_SYMBOL(kmem_cache_destroy);
2799
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002800/*
2801 * Get the memory for a slab management obj.
2802 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2803 * always come from malloc_sizes caches. The slab descriptor cannot
2804 * come from the same cache which is getting created because,
2805 * when we are searching for an appropriate cache for these
2806 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2807 * If we are creating a malloc_sizes cache here it would not be visible to
2808 * kmem_find_general_cachep till the initialization is complete.
2809 * Hence we cannot have slabp_cache same as the original cache.
2810 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002811static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002812 int colour_off, gfp_t local_flags,
2813 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814{
2815 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002816
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 if (OFF_SLAB(cachep)) {
2818 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002819 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002820 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002821 /*
2822 * If the first object in the slab is leaked (it's allocated
2823 * but no one has a reference to it), we want to make sure
2824 * kmemleak does not treat the ->s_mem pointer as a reference
2825 * to the object. Otherwise we will not report the leak.
2826 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002827 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2828 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 if (!slabp)
2830 return NULL;
2831 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002832 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 colour_off += cachep->slab_size;
2834 }
2835 slabp->inuse = 0;
2836 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002837 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002838 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002839 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 return slabp;
2841}
2842
2843static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2844{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002845 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846}
2847
Pekka Enberg343e0d72006-02-01 03:05:50 -08002848static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002849 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850{
2851 int i;
2852
2853 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002854 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855#if DEBUG
2856 /* need to poison the objs? */
2857 if (cachep->flags & SLAB_POISON)
2858 poison_obj(cachep, objp, POISON_FREE);
2859 if (cachep->flags & SLAB_STORE_USER)
2860 *dbg_userword(cachep, objp) = NULL;
2861
2862 if (cachep->flags & SLAB_RED_ZONE) {
2863 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2864 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2865 }
2866 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002867 * Constructors are not allowed to allocate memory from the same
2868 * cache which they are a constructor for. Otherwise, deadlock.
2869 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 */
2871 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002872 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
2874 if (cachep->flags & SLAB_RED_ZONE) {
2875 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2876 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002877 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2879 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002880 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002882 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002883 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002884 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002885 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886#else
2887 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002888 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002890 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002892 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893}
2894
Pekka Enberg343e0d72006-02-01 03:05:50 -08002895static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002897 if (CONFIG_ZONE_DMA_FLAG) {
2898 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002899 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002900 else
Glauber Costaa618e892012-06-14 16:17:21 +04002901 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002902 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903}
2904
Andrew Mortona737b3e2006-03-22 00:08:11 -08002905static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2906 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002907{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002908 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002909 kmem_bufctl_t next;
2910
2911 slabp->inuse++;
2912 next = slab_bufctl(slabp)[slabp->free];
2913#if DEBUG
2914 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2915 WARN_ON(slabp->nodeid != nodeid);
2916#endif
2917 slabp->free = next;
2918
2919 return objp;
2920}
2921
Andrew Mortona737b3e2006-03-22 00:08:11 -08002922static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2923 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002924{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002925 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002926
2927#if DEBUG
2928 /* Verify that the slab belongs to the intended node */
2929 WARN_ON(slabp->nodeid != nodeid);
2930
Al Viro871751e2006-03-25 03:06:39 -08002931 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002932 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002933 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002934 BUG();
2935 }
2936#endif
2937 slab_bufctl(slabp)[objnr] = slabp->free;
2938 slabp->free = objnr;
2939 slabp->inuse--;
2940}
2941
Pekka Enberg47768742006-06-23 02:03:07 -07002942/*
2943 * Map pages beginning at addr to the given cache and slab. This is required
2944 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002945 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002946 */
2947static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2948 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949{
Pekka Enberg47768742006-06-23 02:03:07 -07002950 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 struct page *page;
2952
Pekka Enberg47768742006-06-23 02:03:07 -07002953 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002954
Pekka Enberg47768742006-06-23 02:03:07 -07002955 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002956 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002957 nr_pages <<= cache->gfporder;
2958
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 do {
Christoph Lameter35026082012-06-13 10:24:56 -05002960 page->slab_cache = cache;
2961 page->slab_page = slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002963 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964}
2965
2966/*
2967 * Grow (by 1) the number of slabs within a cache. This is called by
2968 * kmem_cache_alloc() when there are no active objs left in a cache.
2969 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002970static int cache_grow(struct kmem_cache *cachep,
2971 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002973 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002974 size_t offset;
2975 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002976 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977
Andrew Mortona737b3e2006-03-22 00:08:11 -08002978 /*
2979 * Be lazy and only check for valid flags here, keeping it out of the
2980 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002982 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2983 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002985 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002987 l3 = cachep->nodelists[nodeid];
2988 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989
2990 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002991 offset = l3->colour_next;
2992 l3->colour_next++;
2993 if (l3->colour_next >= cachep->colour)
2994 l3->colour_next = 0;
2995 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002997 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998
2999 if (local_flags & __GFP_WAIT)
3000 local_irq_enable();
3001
3002 /*
3003 * The test for missing atomic flag is performed here, rather than
3004 * the more obvious place, simply to reduce the critical path length
3005 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
3006 * will eventually be caught here (where it matters).
3007 */
3008 kmem_flagcheck(cachep, flags);
3009
Andrew Mortona737b3e2006-03-22 00:08:11 -08003010 /*
3011 * Get mem for the objs. Attempt to allocate a physical page from
3012 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07003013 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003014 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07003015 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003016 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 goto failed;
3018
3019 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003020 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07003021 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003022 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 goto opps1;
3024
Pekka Enberg47768742006-06-23 02:03:07 -07003025 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
Christoph Lametera35afb82007-05-16 22:10:57 -07003027 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028
3029 if (local_flags & __GFP_WAIT)
3030 local_irq_disable();
3031 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07003032 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033
3034 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07003035 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003037 l3->free_objects += cachep->num;
3038 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003040opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003042failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 if (local_flags & __GFP_WAIT)
3044 local_irq_disable();
3045 return 0;
3046}
3047
3048#if DEBUG
3049
3050/*
3051 * Perform extra freeing checks:
3052 * - detect bad pointers.
3053 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054 */
3055static void kfree_debugcheck(const void *objp)
3056{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 if (!virt_addr_valid(objp)) {
3058 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003059 (unsigned long)objp);
3060 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062}
3063
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003064static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3065{
David Woodhouseb46b8f12007-05-08 00:22:59 -07003066 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003067
3068 redzone1 = *dbg_redzone1(cache, obj);
3069 redzone2 = *dbg_redzone2(cache, obj);
3070
3071 /*
3072 * Redzone is ok.
3073 */
3074 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3075 return;
3076
3077 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3078 slab_error(cache, "double free detected");
3079 else
3080 slab_error(cache, "memory outside object was overwritten");
3081
David Woodhouseb46b8f12007-05-08 00:22:59 -07003082 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003083 obj, redzone1, redzone2);
3084}
3085
Pekka Enberg343e0d72006-02-01 03:05:50 -08003086static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003087 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088{
3089 struct page *page;
3090 unsigned int objnr;
3091 struct slab *slabp;
3092
Matthew Wilcox80cbd912007-11-29 12:05:13 -07003093 BUG_ON(virt_to_cache(objp) != cachep);
3094
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003095 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07003097 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098
Christoph Lameter35026082012-06-13 10:24:56 -05003099 slabp = page->slab_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100
3101 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07003102 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3104 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3105 }
3106 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003107 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003109 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110
3111 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08003112 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113
Al Viro871751e2006-03-25 03:06:39 -08003114#ifdef CONFIG_DEBUG_SLAB_LEAK
3115 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3116#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 if (cachep->flags & SLAB_POISON) {
3118#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003119 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003120 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003121 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003122 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123 } else {
3124 poison_obj(cachep, objp, POISON_FREE);
3125 }
3126#else
3127 poison_obj(cachep, objp, POISON_FREE);
3128#endif
3129 }
3130 return objp;
3131}
3132
Pekka Enberg343e0d72006-02-01 03:05:50 -08003133static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003134{
3135 kmem_bufctl_t i;
3136 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003137
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138 /* Check slab's freelist to see if this obj is there. */
3139 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3140 entries++;
3141 if (entries > cachep->num || i >= cachep->num)
3142 goto bad;
3143 }
3144 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003145bad:
3146 printk(KERN_ERR "slab: Internal list corruption detected in "
Dave Jonesface37f2011-11-15 15:03:52 -08003147 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3148 cachep->name, cachep->num, slabp, slabp->inuse,
3149 print_tainted());
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02003150 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3151 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3152 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003153 BUG();
3154 }
3155}
3156#else
3157#define kfree_debugcheck(x) do { } while(0)
3158#define cache_free_debugcheck(x,objp,z) (objp)
3159#define check_slabp(x,y) do { } while(0)
3160#endif
3161
Mel Gorman072bb0a2012-07-31 16:43:58 -07003162static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3163 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164{
3165 int batchcount;
3166 struct kmem_list3 *l3;
3167 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003168 int node;
3169
Joe Korty6d2144d2008-03-05 15:04:59 -08003170 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003171 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07003172 if (unlikely(force_refill))
3173 goto force_grow;
3174retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003175 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 batchcount = ac->batchcount;
3177 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003178 /*
3179 * If there was little recent activity on this cache, then
3180 * perform only a partial refill. Otherwise we could generate
3181 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 */
3183 batchcount = BATCHREFILL_LIMIT;
3184 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003185 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186
Christoph Lametere498be72005-09-09 13:03:32 -07003187 BUG_ON(ac->avail > 0 || !l3);
3188 spin_lock(&l3->list_lock);
3189
Christoph Lameter3ded1752006-03-25 03:06:44 -08003190 /* See if we can refill from the shared array */
Nick Piggin44b57f12010-01-27 22:27:40 +11003191 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3192 l3->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003193 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003194 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003195
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 while (batchcount > 0) {
3197 struct list_head *entry;
3198 struct slab *slabp;
3199 /* Get slab alloc is to come from. */
3200 entry = l3->slabs_partial.next;
3201 if (entry == &l3->slabs_partial) {
3202 l3->free_touched = 1;
3203 entry = l3->slabs_free.next;
3204 if (entry == &l3->slabs_free)
3205 goto must_grow;
3206 }
3207
3208 slabp = list_entry(entry, struct slab, list);
3209 check_slabp(cachep, slabp);
3210 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003211
3212 /*
3213 * The slab was either on partial or free list so
3214 * there must be at least one object available for
3215 * allocation.
3216 */
roel kluin249b9f32008-10-29 17:18:07 -04003217 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003218
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 STATS_INC_ALLOCED(cachep);
3221 STATS_INC_ACTIVE(cachep);
3222 STATS_SET_HIGH(cachep);
3223
Mel Gorman072bb0a2012-07-31 16:43:58 -07003224 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3225 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226 }
3227 check_slabp(cachep, slabp);
3228
3229 /* move slabp to correct slabp list: */
3230 list_del(&slabp->list);
3231 if (slabp->free == BUFCTL_END)
3232 list_add(&slabp->list, &l3->slabs_full);
3233 else
3234 list_add(&slabp->list, &l3->slabs_partial);
3235 }
3236
Andrew Mortona737b3e2006-03-22 00:08:11 -08003237must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003239alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003240 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241
3242 if (unlikely(!ac->avail)) {
3243 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003244force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08003245 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003246
Andrew Mortona737b3e2006-03-22 00:08:11 -08003247 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003248 ac = cpu_cache_get(cachep);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003249
3250 /* no objects in sight? abort */
3251 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 return NULL;
3253
Andrew Mortona737b3e2006-03-22 00:08:11 -08003254 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 goto retry;
3256 }
3257 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003258
3259 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260}
3261
Andrew Mortona737b3e2006-03-22 00:08:11 -08003262static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3263 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264{
3265 might_sleep_if(flags & __GFP_WAIT);
3266#if DEBUG
3267 kmem_flagcheck(cachep, flags);
3268#endif
3269}
3270
3271#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003272static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003273 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003275 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003277 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003279 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003280 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003281 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 else
3283 check_poison_obj(cachep, objp);
3284#else
3285 check_poison_obj(cachep, objp);
3286#endif
3287 poison_obj(cachep, objp, POISON_INUSE);
3288 }
3289 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003290 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291
3292 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003293 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3294 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3295 slab_error(cachep, "double free, or memory outside"
3296 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003297 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003298 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003299 objp, *dbg_redzone1(cachep, objp),
3300 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 }
3302 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3303 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3304 }
Al Viro871751e2006-03-25 03:06:39 -08003305#ifdef CONFIG_DEBUG_SLAB_LEAK
3306 {
3307 struct slab *slabp;
3308 unsigned objnr;
3309
Christoph Lameter35026082012-06-13 10:24:56 -05003310 slabp = virt_to_head_page(objp)->slab_page;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003311 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
Al Viro871751e2006-03-25 03:06:39 -08003312 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3313 }
3314#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003315 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003316 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003317 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003318 if (ARCH_SLAB_MINALIGN &&
3319 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003320 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003321 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003322 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 return objp;
3324}
3325#else
3326#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3327#endif
3328
Akinobu Mita773ff602008-12-23 19:37:01 +09003329static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003330{
3331 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003332 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003333
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003334 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003335}
3336
Pekka Enberg343e0d72006-02-01 03:05:50 -08003337static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003339 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003341 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003342
Alok N Kataria5c382302005-09-27 21:45:46 -07003343 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003344
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003345 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003348 objp = ac_get_obj(cachep, ac, flags, false);
3349
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003350 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003351 * Allow for the possibility all avail objects are not allowed
3352 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003353 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003354 if (objp) {
3355 STATS_INC_ALLOCHIT(cachep);
3356 goto out;
3357 }
3358 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003360
3361 STATS_INC_ALLOCMISS(cachep);
3362 objp = cache_alloc_refill(cachep, flags, force_refill);
3363 /*
3364 * the 'ac' may be updated by cache_alloc_refill(),
3365 * and kmemleak_erase() requires its correct value.
3366 */
3367 ac = cpu_cache_get(cachep);
3368
3369out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003370 /*
3371 * To avoid a false negative, if an object that is in one of the
3372 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3373 * treat the array pointers as a reference to the object.
3374 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003375 if (objp)
3376 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003377 return objp;
3378}
3379
Christoph Lametere498be72005-09-09 13:03:32 -07003380#ifdef CONFIG_NUMA
3381/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003382 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003383 *
3384 * If we are in_interrupt, then process context, including cpusets and
3385 * mempolicy, may not apply and should not be used for allocation policy.
3386 */
3387static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3388{
3389 int nid_alloc, nid_here;
3390
Christoph Lameter765c4502006-09-27 01:50:08 -07003391 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003392 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003393 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003394 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003395 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003396 else if (current->mempolicy)
Andi Kleene7b691b2012-06-09 02:40:03 -07003397 nid_alloc = slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003398 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003399 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003400 return NULL;
3401}
3402
3403/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003404 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003405 * certain node and fall back is permitted. First we scan all the
3406 * available nodelists for available objects. If that fails then we
3407 * perform an allocation without specifying a node. This allows the page
3408 * allocator to do its reclaim / fallback magic. We then insert the
3409 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003410 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003411static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003412{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003413 struct zonelist *zonelist;
3414 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003415 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003416 struct zone *zone;
3417 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003418 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003419 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003420 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003421
3422 if (flags & __GFP_THISNODE)
3423 return NULL;
3424
Christoph Lameter6cb06222007-10-16 01:25:41 -07003425 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003426
Mel Gormancc9a6c82012-03-21 16:34:11 -07003427retry_cpuset:
3428 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07003429 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003430
Christoph Lameter3c517a62006-12-06 20:33:29 -08003431retry:
3432 /*
3433 * Look through allowed nodes for objects available
3434 * from existing per node queues.
3435 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003436 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3437 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003438
Mel Gorman54a6eb52008-04-28 02:12:16 -07003439 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003440 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003441 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003442 obj = ____cache_alloc_node(cache,
3443 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003444 if (obj)
3445 break;
3446 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003447 }
3448
Christoph Lametercfce6602007-05-06 14:50:17 -07003449 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003450 /*
3451 * This allocation will be performed within the constraints
3452 * of the current cpuset / memory policy requirements.
3453 * We may trigger various forms of reclaim on the allowed
3454 * set and go into memory reserves if necessary.
3455 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003456 if (local_flags & __GFP_WAIT)
3457 local_irq_enable();
3458 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003459 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003460 if (local_flags & __GFP_WAIT)
3461 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003462 if (obj) {
3463 /*
3464 * Insert into the appropriate per node queues
3465 */
3466 nid = page_to_nid(virt_to_page(obj));
3467 if (cache_grow(cache, flags, nid, obj)) {
3468 obj = ____cache_alloc_node(cache,
3469 flags | GFP_THISNODE, nid);
3470 if (!obj)
3471 /*
3472 * Another processor may allocate the
3473 * objects in the slab since we are
3474 * not holding any locks.
3475 */
3476 goto retry;
3477 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003478 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003479 obj = NULL;
3480 }
3481 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003482 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003483
3484 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3485 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003486 return obj;
3487}
3488
3489/*
Christoph Lametere498be72005-09-09 13:03:32 -07003490 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003492static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003493 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003494{
3495 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003496 struct slab *slabp;
3497 struct kmem_list3 *l3;
3498 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003499 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003501 l3 = cachep->nodelists[nodeid];
3502 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003503
Andrew Mortona737b3e2006-03-22 00:08:11 -08003504retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003505 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003506 spin_lock(&l3->list_lock);
3507 entry = l3->slabs_partial.next;
3508 if (entry == &l3->slabs_partial) {
3509 l3->free_touched = 1;
3510 entry = l3->slabs_free.next;
3511 if (entry == &l3->slabs_free)
3512 goto must_grow;
3513 }
Christoph Lametere498be72005-09-09 13:03:32 -07003514
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003515 slabp = list_entry(entry, struct slab, list);
3516 check_spinlock_acquired_node(cachep, nodeid);
3517 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003518
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003519 STATS_INC_NODEALLOCS(cachep);
3520 STATS_INC_ACTIVE(cachep);
3521 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003522
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003523 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003524
Matthew Dobson78d382d2006-02-01 03:05:47 -08003525 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003526 check_slabp(cachep, slabp);
3527 l3->free_objects--;
3528 /* move slabp to correct slabp list: */
3529 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003530
Andrew Mortona737b3e2006-03-22 00:08:11 -08003531 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003532 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003533 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003534 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003535
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003536 spin_unlock(&l3->list_lock);
3537 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003538
Andrew Mortona737b3e2006-03-22 00:08:11 -08003539must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003540 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003541 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003542 if (x)
3543 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003544
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003545 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003546
Andrew Mortona737b3e2006-03-22 00:08:11 -08003547done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003548 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003549}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003550
3551/**
3552 * kmem_cache_alloc_node - Allocate an object on the specified node
3553 * @cachep: The cache to allocate from.
3554 * @flags: See kmalloc().
3555 * @nodeid: node number of the target node.
3556 * @caller: return address of caller, used for debug information
3557 *
3558 * Identical to kmem_cache_alloc but it will allocate memory on the given
3559 * node, which can improve the performance for cpu bound structures.
3560 *
3561 * Fallback to other node is possible if __GFP_THISNODE is not set.
3562 */
3563static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003564slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003565 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003566{
3567 unsigned long save_flags;
3568 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003569 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003570
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003571 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003572
Nick Piggincf40bd12009-01-21 08:12:39 +01003573 lockdep_trace_alloc(flags);
3574
Akinobu Mita773ff602008-12-23 19:37:01 +09003575 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003576 return NULL;
3577
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003578 cache_alloc_debugcheck_before(cachep, flags);
3579 local_irq_save(save_flags);
3580
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003581 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003582 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003583
3584 if (unlikely(!cachep->nodelists[nodeid])) {
3585 /* Node not bootstrapped yet */
3586 ptr = fallback_alloc(cachep, flags);
3587 goto out;
3588 }
3589
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003590 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003591 /*
3592 * Use the locally cached objects if possible.
3593 * However ____cache_alloc does not allow fallback
3594 * to other nodes. It may fail while we still have
3595 * objects on other nodes available.
3596 */
3597 ptr = ____cache_alloc(cachep, flags);
3598 if (ptr)
3599 goto out;
3600 }
3601 /* ___cache_alloc_node can fall back to other nodes */
3602 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3603 out:
3604 local_irq_restore(save_flags);
3605 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003606 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003607 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003608
Pekka Enbergc175eea2008-05-09 20:35:53 +02003609 if (likely(ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003610 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003611
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003612 if (unlikely((flags & __GFP_ZERO) && ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003613 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003614
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003615 return ptr;
3616}
3617
3618static __always_inline void *
3619__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3620{
3621 void *objp;
3622
3623 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3624 objp = alternate_node_alloc(cache, flags);
3625 if (objp)
3626 goto out;
3627 }
3628 objp = ____cache_alloc(cache, flags);
3629
3630 /*
3631 * We may just have run out of memory on the local node.
3632 * ____cache_alloc_node() knows how to locate memory on other nodes
3633 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003634 if (!objp)
3635 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003636
3637 out:
3638 return objp;
3639}
3640#else
3641
3642static __always_inline void *
3643__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3644{
3645 return ____cache_alloc(cachep, flags);
3646}
3647
3648#endif /* CONFIG_NUMA */
3649
3650static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003651slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003652{
3653 unsigned long save_flags;
3654 void *objp;
3655
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003656 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003657
Nick Piggincf40bd12009-01-21 08:12:39 +01003658 lockdep_trace_alloc(flags);
3659
Akinobu Mita773ff602008-12-23 19:37:01 +09003660 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003661 return NULL;
3662
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003663 cache_alloc_debugcheck_before(cachep, flags);
3664 local_irq_save(save_flags);
3665 objp = __do_cache_alloc(cachep, flags);
3666 local_irq_restore(save_flags);
3667 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003668 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003669 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003670 prefetchw(objp);
3671
Pekka Enbergc175eea2008-05-09 20:35:53 +02003672 if (likely(objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003673 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003674
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003675 if (unlikely((flags & __GFP_ZERO) && objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003676 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003677
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003678 return objp;
3679}
Christoph Lametere498be72005-09-09 13:03:32 -07003680
3681/*
3682 * Caller needs to acquire correct kmem_list's list_lock
3683 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003684static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003685 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686{
3687 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003688 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689
3690 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003691 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003692 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693
Mel Gorman072bb0a2012-07-31 16:43:58 -07003694 clear_obj_pfmemalloc(&objpp[i]);
3695 objp = objpp[i];
3696
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003697 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003698 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003700 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003702 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003704 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 check_slabp(cachep, slabp);
3706
3707 /* fixup slab chains */
3708 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003709 if (l3->free_objects > l3->free_limit) {
3710 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003711 /* No need to drop any previously held
3712 * lock here, even if we have a off-slab slab
3713 * descriptor it is guaranteed to come from
3714 * a different cache, refer to comments before
3715 * alloc_slabmgmt.
3716 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 slab_destroy(cachep, slabp);
3718 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003719 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720 }
3721 } else {
3722 /* Unconditionally move a slab to the end of the
3723 * partial list on free - maximum time for the
3724 * other objects to be freed, too.
3725 */
Christoph Lametere498be72005-09-09 13:03:32 -07003726 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 }
3728 }
3729}
3730
Pekka Enberg343e0d72006-02-01 03:05:50 -08003731static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732{
3733 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003734 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003735 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736
3737 batchcount = ac->batchcount;
3738#if DEBUG
3739 BUG_ON(!batchcount || batchcount > ac->avail);
3740#endif
3741 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003742 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003743 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003744 if (l3->shared) {
3745 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003746 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747 if (max) {
3748 if (batchcount > max)
3749 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003750 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003751 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 shared_array->avail += batchcount;
3753 goto free_done;
3754 }
3755 }
3756
Christoph Lameterff694162005-09-22 21:44:02 -07003757 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003758free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759#if STATS
3760 {
3761 int i = 0;
3762 struct list_head *p;
3763
Christoph Lametere498be72005-09-09 13:03:32 -07003764 p = l3->slabs_free.next;
3765 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 struct slab *slabp;
3767
3768 slabp = list_entry(p, struct slab, list);
3769 BUG_ON(slabp->inuse);
3770
3771 i++;
3772 p = p->next;
3773 }
3774 STATS_SET_FREEABLE(cachep, i);
3775 }
3776#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003777 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003779 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780}
3781
3782/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003783 * Release an obj back to its cache. If the obj has a constructed state, it must
3784 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003786static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003787 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003789 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790
3791 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003792 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003793 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003795 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003796
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003797 /*
3798 * Skip calling cache_free_alien() when the platform is not numa.
3799 * This will avoid cache misses that happen while accessing slabp (which
3800 * is per page memory reference) to get nodeid. Instead use a global
3801 * variable to skip the call, which is mostly likely to be present in
3802 * the cache.
3803 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003804 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003805 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003806
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 if (likely(ac->avail < ac->limit)) {
3808 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809 } else {
3810 STATS_INC_FREEMISS(cachep);
3811 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003812 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003813
Mel Gorman072bb0a2012-07-31 16:43:58 -07003814 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003815}
3816
3817/**
3818 * kmem_cache_alloc - Allocate an object
3819 * @cachep: The cache to allocate from.
3820 * @flags: See kmalloc().
3821 *
3822 * Allocate an object from this cache. The flags are only relevant
3823 * if the cache has no available objects.
3824 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003825void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003827 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003828
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003829 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003830 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003831
3832 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833}
3834EXPORT_SYMBOL(kmem_cache_alloc);
3835
Li Zefan0f24f122009-12-11 15:45:30 +08003836#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003837void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003838kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003839{
Steven Rostedt85beb582010-11-24 16:23:34 -05003840 void *ret;
3841
Ezequiel Garcia48356302012-09-08 17:47:57 -03003842 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003843
3844 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003845 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003846 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003847}
Steven Rostedt85beb582010-11-24 16:23:34 -05003848EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003849#endif
3850
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003852void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3853{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003854 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003855
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003856 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003857 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003858 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003859
3860 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003861}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862EXPORT_SYMBOL(kmem_cache_alloc_node);
3863
Li Zefan0f24f122009-12-11 15:45:30 +08003864#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003865void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003866 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003867 int nodeid,
3868 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003869{
Steven Rostedt85beb582010-11-24 16:23:34 -05003870 void *ret;
3871
Ezequiel Garcia48356302012-09-08 17:47:57 -03003872 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003873
Steven Rostedt85beb582010-11-24 16:23:34 -05003874 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003875 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003876 flags, nodeid);
3877 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003878}
Steven Rostedt85beb582010-11-24 16:23:34 -05003879EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003880#endif
3881
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003882static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003883__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003884{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003885 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003886
3887 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003888 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3889 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003890 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003891}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003892
Li Zefan0bb38a52009-12-11 15:45:50 +08003893#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003894void *__kmalloc_node(size_t size, gfp_t flags, int node)
3895{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003896 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003897}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003898EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003899
3900void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003901 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003902{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003903 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003904}
3905EXPORT_SYMBOL(__kmalloc_node_track_caller);
3906#else
3907void *__kmalloc_node(size_t size, gfp_t flags, int node)
3908{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003909 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003910}
3911EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003912#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003913#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914
3915/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003916 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003918 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003919 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003921static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003922 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003924 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003925 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003927 /* If you want to save a few bytes .text space: replace
3928 * __ with kmem_.
3929 * Then kmalloc uses the uninlined functions instead of the inline
3930 * functions.
3931 */
3932 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003933 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3934 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003935 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003936
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003937 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003938 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003939
3940 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003941}
3942
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003943
Li Zefan0bb38a52009-12-11 15:45:50 +08003944#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003945void *__kmalloc(size_t size, gfp_t flags)
3946{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003947 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948}
3949EXPORT_SYMBOL(__kmalloc);
3950
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003951void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003952{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003953 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003954}
3955EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003956
3957#else
3958void *__kmalloc(size_t size, gfp_t flags)
3959{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003960 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003961}
3962EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003963#endif
3964
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965/**
3966 * kmem_cache_free - Deallocate an object
3967 * @cachep: The cache the allocation was from.
3968 * @objp: The previously allocated object.
3969 *
3970 * Free an object which was previously allocated from this
3971 * cache.
3972 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003973void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974{
3975 unsigned long flags;
3976
3977 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003978 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003979 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003980 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003981 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003983
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003984 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985}
3986EXPORT_SYMBOL(kmem_cache_free);
3987
3988/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989 * kfree - free previously allocated memory
3990 * @objp: pointer returned by kmalloc.
3991 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003992 * If @objp is NULL, no operation is performed.
3993 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994 * Don't free memory not originally allocated by kmalloc()
3995 * or you will run into trouble.
3996 */
3997void kfree(const void *objp)
3998{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003999 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 unsigned long flags;
4001
Pekka Enberg2121db72009-03-25 11:05:57 +02004002 trace_kfree(_RET_IP_, objp);
4003
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004004 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 return;
4006 local_irq_save(flags);
4007 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004008 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004009 debug_check_no_locks_freed(objp, c->object_size);
4010
4011 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03004012 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013 local_irq_restore(flags);
4014}
4015EXPORT_SYMBOL(kfree);
4016
Pekka Enberg343e0d72006-02-01 03:05:50 -08004017unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004019 return cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020}
4021EXPORT_SYMBOL(kmem_cache_size);
4022
Christoph Lametere498be72005-09-09 13:03:32 -07004023/*
Simon Arlott183ff222007-10-20 01:27:18 +02004024 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07004025 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004026static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07004027{
4028 int node;
4029 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08004030 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08004031 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004032
Mel Gorman9c09a952008-01-24 05:49:54 -08004033 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08004034
Paul Menage3395ee02006-12-06 20:32:16 -08004035 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03004036 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08004037 if (!new_alien)
4038 goto fail;
4039 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004040
Eric Dumazet63109842007-05-06 14:49:28 -07004041 new_shared = NULL;
4042 if (cachep->shared) {
4043 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08004044 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004045 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07004046 if (!new_shared) {
4047 free_alien_cache(new_alien);
4048 goto fail;
4049 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08004050 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004051
Andrew Mortona737b3e2006-03-22 00:08:11 -08004052 l3 = cachep->nodelists[node];
4053 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08004054 struct array_cache *shared = l3->shared;
4055
Christoph Lametere498be72005-09-09 13:03:32 -07004056 spin_lock_irq(&l3->list_lock);
4057
Christoph Lametercafeb022006-03-25 03:06:46 -08004058 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08004059 free_block(cachep, shared->entry,
4060 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07004061
Christoph Lametercafeb022006-03-25 03:06:46 -08004062 l3->shared = new_shared;
4063 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07004064 l3->alien = new_alien;
4065 new_alien = NULL;
4066 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004067 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004068 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004069 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08004070 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004071 free_alien_cache(new_alien);
4072 continue;
4073 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03004074 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08004075 if (!l3) {
4076 free_alien_cache(new_alien);
4077 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004078 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004079 }
Christoph Lametere498be72005-09-09 13:03:32 -07004080
4081 kmem_list3_init(l3);
4082 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08004083 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08004084 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07004085 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004086 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004087 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004088 cachep->nodelists[node] = l3;
4089 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004090 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08004091
Andrew Mortona737b3e2006-03-22 00:08:11 -08004092fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004093 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08004094 /* Cache is not active yet. Roll back what we did */
4095 node--;
4096 while (node >= 0) {
4097 if (cachep->nodelists[node]) {
4098 l3 = cachep->nodelists[node];
4099
4100 kfree(l3->shared);
4101 free_alien_cache(l3->alien);
4102 kfree(l3);
4103 cachep->nodelists[node] = NULL;
4104 }
4105 node--;
4106 }
4107 }
Christoph Lametercafeb022006-03-25 03:06:46 -08004108 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07004109}
4110
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08004112 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004113 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114};
4115
4116static void do_ccupdate_local(void *info)
4117{
Andrew Mortona737b3e2006-03-22 00:08:11 -08004118 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119 struct array_cache *old;
4120
4121 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08004122 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07004123
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4125 new->new[smp_processor_id()] = old;
4126}
4127
Christoph Lameter18004c52012-07-06 15:25:12 -05004128/* Always called with the slab_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08004129static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004130 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004132 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004133 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134
Eric Dumazetacfe7d72011-07-25 08:55:42 +02004135 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4136 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004137 if (!new)
4138 return -ENOMEM;
4139
Christoph Lametere498be72005-09-09 13:03:32 -07004140 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004141 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004142 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004143 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004144 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004145 kfree(new->new[i]);
4146 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07004147 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 }
4149 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004150 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151
Jens Axboe15c8b6c2008-05-09 09:39:44 +02004152 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07004153
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 cachep->batchcount = batchcount;
4156 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07004157 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158
Christoph Lametere498be72005-09-09 13:03:32 -07004159 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004160 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 if (!ccold)
4162 continue;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004163 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4164 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4165 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166 kfree(ccold);
4167 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004168 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03004169 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170}
4171
Christoph Lameter18004c52012-07-06 15:25:12 -05004172/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004173static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174{
4175 int err;
4176 int limit, shared;
4177
Andrew Mortona737b3e2006-03-22 00:08:11 -08004178 /*
4179 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 * - create a LIFO ordering, i.e. return objects that are cache-warm
4181 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004182 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 * bufctl chains: array operations are cheaper.
4184 * The numbers are guessed, we should auto-tune as described by
4185 * Bonwick.
4186 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004187 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004189 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004191 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004193 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194 limit = 54;
4195 else
4196 limit = 120;
4197
Andrew Mortona737b3e2006-03-22 00:08:11 -08004198 /*
4199 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200 * allocation behaviour: Most allocs on one cpu, most free operations
4201 * on another cpu. For these cases, an efficient object passing between
4202 * cpus is necessary. This is provided by a shared array. The array
4203 * replaces Bonwick's magazine layer.
4204 * On uniprocessor, it's functionally equivalent (but less efficient)
4205 * to a larger limit. Thus disabled by default.
4206 */
4207 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004208 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210
4211#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004212 /*
4213 * With debugging enabled, large batchcount lead to excessively long
4214 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004215 */
4216 if (limit > 32)
4217 limit = 32;
4218#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004219 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220 if (err)
4221 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004222 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004223 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004224}
4225
Christoph Lameter1b552532006-03-22 00:09:07 -08004226/*
4227 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004228 * necessary. Note that the l3 listlock also protects the array_cache
4229 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004230 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004231static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
Christoph Lameter1b552532006-03-22 00:09:07 -08004232 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233{
4234 int tofree;
4235
Christoph Lameter1b552532006-03-22 00:09:07 -08004236 if (!ac || !ac->avail)
4237 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 if (ac->touched && !force) {
4239 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004240 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004241 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004242 if (ac->avail) {
4243 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4244 if (tofree > ac->avail)
4245 tofree = (ac->avail + 1) / 2;
4246 free_block(cachep, ac->entry, tofree, node);
4247 ac->avail -= tofree;
4248 memmove(ac->entry, &(ac->entry[tofree]),
4249 sizeof(void *) * ac->avail);
4250 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004251 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252 }
4253}
4254
4255/**
4256 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004257 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258 *
4259 * Called from workqueue/eventd every few seconds.
4260 * Purpose:
4261 * - clear the per-cpu caches for this CPU.
4262 * - return freeable pages to the main free memory pool.
4263 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004264 * If we cannot acquire the cache chain mutex then just give up - we'll try
4265 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004266 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004267static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004268{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004269 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004270 struct kmem_list3 *l3;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004271 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004272 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273
Christoph Lameter18004c52012-07-06 15:25:12 -05004274 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004276 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004277
Christoph Lameter18004c52012-07-06 15:25:12 -05004278 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004279 check_irq_on();
4280
Christoph Lameter35386e32006-03-22 00:09:05 -08004281 /*
4282 * We only take the l3 lock if absolutely necessary and we
4283 * have established with reasonable certainty that
4284 * we can do some work if the lock was obtained.
4285 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004286 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004287
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004288 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004289
Christoph Lameteraab22072006-03-22 00:09:06 -08004290 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291
Christoph Lameter35386e32006-03-22 00:09:05 -08004292 /*
4293 * These are racy checks but it does not matter
4294 * if we skip one check or scan twice.
4295 */
Christoph Lametere498be72005-09-09 13:03:32 -07004296 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004297 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004298
Christoph Lametere498be72005-09-09 13:03:32 -07004299 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004300
Christoph Lameteraab22072006-03-22 00:09:06 -08004301 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302
Christoph Lametered11d9e2006-06-30 01:55:45 -07004303 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004304 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004305 else {
4306 int freed;
4307
4308 freed = drain_freelist(searchp, l3, (l3->free_limit +
4309 5 * searchp->num - 1) / (5 * searchp->num));
4310 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004312next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313 cond_resched();
4314 }
4315 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004316 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004317 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004318out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004319 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004320 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004321}
4322
Linus Torvalds158a9622008-01-02 13:04:48 -08004323#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324
Pekka Enberg85289f92006-01-08 01:00:36 -08004325static void print_slabinfo_header(struct seq_file *m)
4326{
4327 /*
4328 * Output format version, so at least we can change it
4329 * without _too_ many complaints.
4330 */
4331#if STATS
4332 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4333#else
4334 seq_puts(m, "slabinfo - version: 2.1\n");
4335#endif
4336 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4337 "<objperslab> <pagesperslab>");
4338 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4339 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4340#if STATS
4341 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004342 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004343 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4344#endif
4345 seq_putc(m, '\n');
4346}
4347
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348static void *s_start(struct seq_file *m, loff_t *pos)
4349{
4350 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004351
Christoph Lameter18004c52012-07-06 15:25:12 -05004352 mutex_lock(&slab_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004353 if (!n)
4354 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004355
Christoph Lameter18004c52012-07-06 15:25:12 -05004356 return seq_list_start(&slab_caches, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004357}
4358
4359static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4360{
Christoph Lameter18004c52012-07-06 15:25:12 -05004361 return seq_list_next(p, &slab_caches, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004362}
4363
4364static void s_stop(struct seq_file *m, void *p)
4365{
Christoph Lameter18004c52012-07-06 15:25:12 -05004366 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004367}
4368
4369static int s_show(struct seq_file *m, void *p)
4370{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004371 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004372 struct slab *slabp;
4373 unsigned long active_objs;
4374 unsigned long num_objs;
4375 unsigned long active_slabs = 0;
4376 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004377 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004378 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004379 int node;
4380 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381
Linus Torvalds1da177e2005-04-16 15:20:36 -07004382 active_objs = 0;
4383 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004384 for_each_online_node(node) {
4385 l3 = cachep->nodelists[node];
4386 if (!l3)
4387 continue;
4388
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004389 check_irq_on();
4390 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004391
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004392 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004393 if (slabp->inuse != cachep->num && !error)
4394 error = "slabs_full accounting error";
4395 active_objs += cachep->num;
4396 active_slabs++;
4397 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004398 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004399 if (slabp->inuse == cachep->num && !error)
4400 error = "slabs_partial inuse accounting error";
4401 if (!slabp->inuse && !error)
4402 error = "slabs_partial/inuse accounting error";
4403 active_objs += slabp->inuse;
4404 active_slabs++;
4405 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004406 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004407 if (slabp->inuse && !error)
4408 error = "slabs_free/inuse accounting error";
4409 num_slabs++;
4410 }
4411 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004412 if (l3->shared)
4413 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004414
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004415 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004417 num_slabs += active_slabs;
4418 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004419 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004420 error = "free_objects accounting error";
4421
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004422 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004423 if (error)
4424 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4425
4426 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004427 name, active_objs, num_objs, cachep->size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004428 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004429 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004430 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004431 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004432 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004433#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004434 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435 unsigned long high = cachep->high_mark;
4436 unsigned long allocs = cachep->num_allocations;
4437 unsigned long grown = cachep->grown;
4438 unsigned long reaped = cachep->reaped;
4439 unsigned long errors = cachep->errors;
4440 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004442 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004443 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004444
Joe Perchese92dd4f2010-03-26 19:27:58 -07004445 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4446 "%4lu %4lu %4lu %4lu %4lu",
4447 allocs, high, grown,
4448 reaped, errors, max_freeable, node_allocs,
4449 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004450 }
4451 /* cpu stats */
4452 {
4453 unsigned long allochit = atomic_read(&cachep->allochit);
4454 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4455 unsigned long freehit = atomic_read(&cachep->freehit);
4456 unsigned long freemiss = atomic_read(&cachep->freemiss);
4457
4458 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004459 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460 }
4461#endif
4462 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004463 return 0;
4464}
4465
4466/*
4467 * slabinfo_op - iterator that generates /proc/slabinfo
4468 *
4469 * Output layout:
4470 * cache-name
4471 * num-active-objs
4472 * total-objs
4473 * object size
4474 * num-active-slabs
4475 * total-slabs
4476 * num-pages-per-slab
4477 * + further values on SMP and with statistics enabled
4478 */
4479
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004480static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004481 .start = s_start,
4482 .next = s_next,
4483 .stop = s_stop,
4484 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004485};
4486
4487#define MAX_SLABINFO_WRITE 128
4488/**
4489 * slabinfo_write - Tuning for the slab allocator
4490 * @file: unused
4491 * @buffer: user buffer
4492 * @count: data length
4493 * @ppos: unused
4494 */
H Hartley Sweeten68a1b192011-01-11 17:49:32 -06004495static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004496 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004497{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004498 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004499 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004500 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004501
Linus Torvalds1da177e2005-04-16 15:20:36 -07004502 if (count > MAX_SLABINFO_WRITE)
4503 return -EINVAL;
4504 if (copy_from_user(&kbuf, buffer, count))
4505 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004506 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004507
4508 tmp = strchr(kbuf, ' ');
4509 if (!tmp)
4510 return -EINVAL;
4511 *tmp = '\0';
4512 tmp++;
4513 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4514 return -EINVAL;
4515
4516 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004517 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004518 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004519 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004520 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004521 if (limit < 1 || batchcount < 1 ||
4522 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004523 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004524 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004525 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004526 batchcount, shared,
4527 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004528 }
4529 break;
4530 }
4531 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004532 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004533 if (res >= 0)
4534 res = count;
4535 return res;
4536}
Al Viro871751e2006-03-25 03:06:39 -08004537
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004538static int slabinfo_open(struct inode *inode, struct file *file)
4539{
4540 return seq_open(file, &slabinfo_op);
4541}
4542
4543static const struct file_operations proc_slabinfo_operations = {
4544 .open = slabinfo_open,
4545 .read = seq_read,
4546 .write = slabinfo_write,
4547 .llseek = seq_lseek,
4548 .release = seq_release,
4549};
4550
Al Viro871751e2006-03-25 03:06:39 -08004551#ifdef CONFIG_DEBUG_SLAB_LEAK
4552
4553static void *leaks_start(struct seq_file *m, loff_t *pos)
4554{
Christoph Lameter18004c52012-07-06 15:25:12 -05004555 mutex_lock(&slab_mutex);
4556 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004557}
4558
4559static inline int add_caller(unsigned long *n, unsigned long v)
4560{
4561 unsigned long *p;
4562 int l;
4563 if (!v)
4564 return 1;
4565 l = n[1];
4566 p = n + 2;
4567 while (l) {
4568 int i = l/2;
4569 unsigned long *q = p + 2 * i;
4570 if (*q == v) {
4571 q[1]++;
4572 return 1;
4573 }
4574 if (*q > v) {
4575 l = i;
4576 } else {
4577 p = q + 2;
4578 l -= i + 1;
4579 }
4580 }
4581 if (++n[1] == n[0])
4582 return 0;
4583 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4584 p[0] = v;
4585 p[1] = 1;
4586 return 1;
4587}
4588
4589static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4590{
4591 void *p;
4592 int i;
4593 if (n[0] == n[1])
4594 return;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004595 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
Al Viro871751e2006-03-25 03:06:39 -08004596 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4597 continue;
4598 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4599 return;
4600 }
4601}
4602
4603static void show_symbol(struct seq_file *m, unsigned long address)
4604{
4605#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004606 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004607 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004608
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004609 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004610 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004611 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004612 seq_printf(m, " [%s]", modname);
4613 return;
4614 }
4615#endif
4616 seq_printf(m, "%p", (void *)address);
4617}
4618
4619static int leaks_show(struct seq_file *m, void *p)
4620{
Thierry Reding0672aa72012-06-22 19:42:49 +02004621 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Al Viro871751e2006-03-25 03:06:39 -08004622 struct slab *slabp;
4623 struct kmem_list3 *l3;
4624 const char *name;
4625 unsigned long *n = m->private;
4626 int node;
4627 int i;
4628
4629 if (!(cachep->flags & SLAB_STORE_USER))
4630 return 0;
4631 if (!(cachep->flags & SLAB_RED_ZONE))
4632 return 0;
4633
4634 /* OK, we can do it */
4635
4636 n[1] = 0;
4637
4638 for_each_online_node(node) {
4639 l3 = cachep->nodelists[node];
4640 if (!l3)
4641 continue;
4642
4643 check_irq_on();
4644 spin_lock_irq(&l3->list_lock);
4645
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004646 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004647 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004648 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004649 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004650 spin_unlock_irq(&l3->list_lock);
4651 }
4652 name = cachep->name;
4653 if (n[0] == n[1]) {
4654 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004655 mutex_unlock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004656 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4657 if (!m->private) {
4658 /* Too bad, we are really out */
4659 m->private = n;
Christoph Lameter18004c52012-07-06 15:25:12 -05004660 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004661 return -ENOMEM;
4662 }
4663 *(unsigned long *)m->private = n[0] * 2;
4664 kfree(n);
Christoph Lameter18004c52012-07-06 15:25:12 -05004665 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004666 /* Now make sure this entry will be retried */
4667 m->count = m->size;
4668 return 0;
4669 }
4670 for (i = 0; i < n[1]; i++) {
4671 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4672 show_symbol(m, n[2*i+2]);
4673 seq_putc(m, '\n');
4674 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004675
Al Viro871751e2006-03-25 03:06:39 -08004676 return 0;
4677}
4678
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004679static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004680 .start = leaks_start,
4681 .next = s_next,
4682 .stop = s_stop,
4683 .show = leaks_show,
4684};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004685
4686static int slabstats_open(struct inode *inode, struct file *file)
4687{
4688 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4689 int ret = -ENOMEM;
4690 if (n) {
4691 ret = seq_open(file, &slabstats_op);
4692 if (!ret) {
4693 struct seq_file *m = file->private_data;
4694 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4695 m->private = n;
4696 n = NULL;
4697 }
4698 kfree(n);
4699 }
4700 return ret;
4701}
4702
4703static const struct file_operations proc_slabstats_operations = {
4704 .open = slabstats_open,
4705 .read = seq_read,
4706 .llseek = seq_lseek,
4707 .release = seq_release_private,
4708};
Al Viro871751e2006-03-25 03:06:39 -08004709#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004710
4711static int __init slab_proc_init(void)
4712{
Vasiliy Kulikovab067e92011-09-27 21:54:53 +04004713 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004714#ifdef CONFIG_DEBUG_SLAB_LEAK
4715 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4716#endif
4717 return 0;
4718}
4719module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004720#endif
4721
Manfred Spraul00e145b2005-09-03 15:55:07 -07004722/**
4723 * ksize - get the actual amount of memory allocated for a given object
4724 * @objp: Pointer to the object
4725 *
4726 * kmalloc may internally round up allocations and return more memory
4727 * than requested. ksize() can be used to determine the actual amount of
4728 * memory allocated. The caller may use this additional memory, even though
4729 * a smaller amount of memory was initially specified with the kmalloc call.
4730 * The caller must guarantee that objp points to a valid object previously
4731 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4732 * must not be freed during the duration of the call.
4733 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004734size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004736 BUG_ON(!objp);
4737 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004738 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004739
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004740 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004742EXPORT_SYMBOL(ksize);