blob: 35cb0c861508c898264c56818ac431086c6f39e1 [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>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700118#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
Mel Gorman381760e2012-07-31 16:44:30 -0700120#include <net/sock.h>
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500126#include <trace/events/kmem.h>
127
Mel Gorman072bb0a2012-07-31 16:43:58 -0700128#include "internal.h"
129
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800130#include "slab.h"
131
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
Mel Gorman072bb0a2012-07-31 16:43:58 -0700160/*
161 * true if a page was allocated from pfmemalloc reserves for network-based
162 * swap
163 */
164static bool pfmemalloc_active __read_mostly;
165
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166/*
167 * kmem_bufctl_t:
168 *
169 * Bufctl's are used for linking objs within a slab
170 * linked offsets.
171 *
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
183 */
184
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700185typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800188#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 * struct slab_rcu
193 *
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
201 *
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204 */
205struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800206 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800207 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800208 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209};
210
211/*
Lai Jiangshan5bfe53a2011-03-10 15:22:24 +0800212 * struct slab
213 *
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
217 */
218struct slab {
219 union {
220 struct {
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
227 };
228 struct slab_rcu __slab_cover_slab_rcu;
229 };
230};
231
232/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 * struct array_cache
234 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235 * Purpose:
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
239 *
240 * The limit is stored in the per-cpu structure to reduce the data cache
241 * footprint.
242 *
243 */
244struct array_cache {
245 unsigned int avail;
246 unsigned int limit;
247 unsigned int batchcount;
248 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700249 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700250 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
253 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700254 *
255 * Entries should not be directly dereferenced as
256 * entries belonging to slabs marked pfmemalloc will
257 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800258 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259};
260
Mel Gorman072bb0a2012-07-31 16:43:58 -0700261#define SLAB_OBJ_PFMEMALLOC 1
262static inline bool is_obj_pfmemalloc(void *objp)
263{
264 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
265}
266
267static inline void set_obj_pfmemalloc(void **objp)
268{
269 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
270 return;
271}
272
273static inline void clear_obj_pfmemalloc(void **objp)
274{
275 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
276}
277
Andrew Mortona737b3e2006-03-22 00:08:11 -0800278/*
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
282#define BOOT_CPUCACHE_ENTRIES 1
283struct arraycache_init {
284 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800285 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286};
287
288/*
Christoph Lametere498be72005-09-09 13:03:32 -0700289 * Need this for bootstrapping a per node allocator.
290 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200291#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000292static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700293#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200294#define SIZE_AC MAX_NUMNODES
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000295#define SIZE_NODE (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
Christoph Lametered11d9e2006-06-30 01:55:45 -0700297static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000298 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700299static void free_block(struct kmem_cache *cachep, void **objpp, int len,
300 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300301static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000302static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700303
Ingo Molnare0a42722006-06-23 02:03:46 -0700304static int slab_early_init = 1;
305
Christoph Lametere3366012013-01-10 19:14:18 +0000306#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000307#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700308
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000309static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700310{
311 INIT_LIST_HEAD(&parent->slabs_full);
312 INIT_LIST_HEAD(&parent->slabs_partial);
313 INIT_LIST_HEAD(&parent->slabs_free);
314 parent->shared = NULL;
315 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800316 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700317 spin_lock_init(&parent->list_lock);
318 parent->free_objects = 0;
319 parent->free_touched = 0;
320}
321
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322#define MAKE_LIST(cachep, listp, slab, nodeid) \
323 do { \
324 INIT_LIST_HEAD(listp); \
Christoph Lameter6a673682013-01-10 19:14:19 +0000325 list_splice(&(cachep->node[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700326 } while (0)
327
Andrew Mortona737b3e2006-03-22 00:08:11 -0800328#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
329 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700330 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
331 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
332 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
333 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335#define CFLGS_OFF_SLAB (0x80000000UL)
336#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
337
338#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800339/*
340 * Optimization question: fewer reaps means less probability for unnessary
341 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100343 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 * which could lock up otherwise freeable slabs.
345 */
346#define REAPTIMEOUT_CPUC (2*HZ)
347#define REAPTIMEOUT_LIST3 (4*HZ)
348
349#if STATS
350#define STATS_INC_ACTIVE(x) ((x)->num_active++)
351#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
352#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
353#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700354#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800355#define STATS_SET_HIGH(x) \
356 do { \
357 if ((x)->num_active > (x)->high_mark) \
358 (x)->high_mark = (x)->num_active; \
359 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360#define STATS_INC_ERR(x) ((x)->errors++)
361#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700362#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700363#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800364#define STATS_SET_FREEABLE(x, i) \
365 do { \
366 if ((x)->max_freeable < i) \
367 (x)->max_freeable = i; \
368 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
370#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
371#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
372#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
373#else
374#define STATS_INC_ACTIVE(x) do { } while (0)
375#define STATS_DEC_ACTIVE(x) do { } while (0)
376#define STATS_INC_ALLOCED(x) do { } while (0)
377#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700378#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379#define STATS_SET_HIGH(x) do { } while (0)
380#define STATS_INC_ERR(x) do { } while (0)
381#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700382#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700383#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800384#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385#define STATS_INC_ALLOCHIT(x) do { } while (0)
386#define STATS_INC_ALLOCMISS(x) do { } while (0)
387#define STATS_INC_FREEHIT(x) do { } while (0)
388#define STATS_INC_FREEMISS(x) do { } while (0)
389#endif
390
391#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392
Andrew Mortona737b3e2006-03-22 00:08:11 -0800393/*
394 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800396 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 * the end of an object is aligned with the end of the real
398 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800399 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800401 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500402 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
403 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800404 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800408 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409}
410
David Woodhouseb46b8f12007-05-08 00:22:59 -0700411static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
413 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700414 return (unsigned long long*) (objp + obj_offset(cachep) -
415 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416}
417
David Woodhouseb46b8f12007-05-08 00:22:59 -0700418static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419{
420 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
421 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500422 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700423 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400424 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500425 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700426 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427}
428
Pekka Enberg343e0d72006-02-01 03:05:50 -0800429static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430{
431 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500432 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433}
434
435#else
436
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800437#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700438#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
439#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
441
442#endif
443
444/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700445 * Do not go above this order unless 0 objects fit into the slab or
446 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 */
David Rientjes543585c2011-10-18 22:09:24 -0700448#define SLAB_MAX_ORDER_HI 1
449#define SLAB_MAX_ORDER_LO 0
450static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700451static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800453static inline struct kmem_cache *virt_to_cache(const void *obj)
454{
Christoph Lameterb49af682007-05-06 14:49:41 -0700455 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500456 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800457}
458
459static inline struct slab *virt_to_slab(const void *obj)
460{
Christoph Lameterb49af682007-05-06 14:49:41 -0700461 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500462
463 VM_BUG_ON(!PageSlab(page));
464 return page->slab_page;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800465}
466
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800467static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
468 unsigned int idx)
469{
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500470 return slab->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800471}
472
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800473/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500474 * We want to avoid an expensive divide : (offset / cache->size)
475 * Using the fact that size is a constant for a particular cache,
476 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800477 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
478 */
479static inline unsigned int obj_to_index(const struct kmem_cache *cache,
480 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800481{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800482 u32 offset = (obj - slab->s_mem);
483 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800484}
485
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800487 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488
489/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000490static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800491 .batchcount = 1,
492 .limit = BOOT_CPUCACHE_ENTRIES,
493 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500494 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800495 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496};
497
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700498#define BAD_ALIEN_MAGIC 0x01020304ul
499
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200500#ifdef CONFIG_LOCKDEP
501
502/*
503 * Slab sometimes uses the kmalloc slabs to store the slab headers
504 * for other slabs "off slab".
505 * The locking for this is tricky in that it nests within the locks
506 * of all other slabs in a few places; to deal with this special
507 * locking we put on-slab caches into a separate lock-class.
508 *
509 * We set lock class for alien array caches which are up during init.
510 * The lock annotation will be lost if all cpus of a node goes down and
511 * then comes back up during hotplug
512 */
513static struct lock_class_key on_slab_l3_key;
514static struct lock_class_key on_slab_alc_key;
515
Peter Zijlstra83835b32011-07-22 15:26:05 +0200516static struct lock_class_key debugobj_l3_key;
517static struct lock_class_key debugobj_alc_key;
518
519static void slab_set_lock_classes(struct kmem_cache *cachep,
520 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
521 int q)
522{
523 struct array_cache **alc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000524 struct kmem_cache_node *n;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200525 int r;
526
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000527 n = cachep->node[q];
528 if (!n)
Peter Zijlstra83835b32011-07-22 15:26:05 +0200529 return;
530
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000531 lockdep_set_class(&n->list_lock, l3_key);
532 alc = n->alien;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200533 /*
534 * FIXME: This check for BAD_ALIEN_MAGIC
535 * should go away when common slab code is taught to
536 * work even without alien caches.
537 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
538 * for alloc_alien_cache,
539 */
540 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
541 return;
542 for_each_node(r) {
543 if (alc[r])
544 lockdep_set_class(&alc[r]->lock, alc_key);
545 }
546}
547
548static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
549{
550 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
551}
552
553static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
554{
555 int node;
556
557 for_each_online_node(node)
558 slab_set_debugobj_lock_classes_node(cachep, node);
559}
560
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200561static void init_node_lock_keys(int q)
562{
Christoph Lametere3366012013-01-10 19:14:18 +0000563 int i;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200564
Christoph Lameter97d06602012-07-06 15:25:11 -0500565 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200566 return;
567
Christoph Lameter0f8f8092013-07-02 12:12:10 -0700568 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000569 struct kmem_cache_node *n;
Christoph Lametere3366012013-01-10 19:14:18 +0000570 struct kmem_cache *cache = kmalloc_caches[i];
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200571
Christoph Lametere3366012013-01-10 19:14:18 +0000572 if (!cache)
Pekka Enberg00afa752009-12-27 14:33:14 +0200573 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200574
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000575 n = cache->node[q];
576 if (!n || OFF_SLAB(cache))
Christoph Lametere3366012013-01-10 19:14:18 +0000577 continue;
578
579 slab_set_lock_classes(cache, &on_slab_l3_key,
Peter Zijlstra83835b32011-07-22 15:26:05 +0200580 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200581 }
582}
583
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800584static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
585{
Christoph Lameter6a673682013-01-10 19:14:19 +0000586 if (!cachep->node[q])
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800587 return;
588
589 slab_set_lock_classes(cachep, &on_slab_l3_key,
590 &on_slab_alc_key, q);
591}
592
593static inline void on_slab_lock_classes(struct kmem_cache *cachep)
594{
595 int node;
596
597 VM_BUG_ON(OFF_SLAB(cachep));
598 for_each_node(node)
599 on_slab_lock_classes_node(cachep, node);
600}
601
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200602static inline void init_lock_keys(void)
603{
604 int node;
605
606 for_each_node(node)
607 init_node_lock_keys(node);
608}
609#else
610static void init_node_lock_keys(int q)
611{
612}
613
614static inline void init_lock_keys(void)
615{
616}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200617
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800618static inline void on_slab_lock_classes(struct kmem_cache *cachep)
619{
620}
621
622static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
623{
624}
625
Peter Zijlstra83835b32011-07-22 15:26:05 +0200626static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
627{
628}
629
630static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
631{
632}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200633#endif
634
Tejun Heo1871e522009-10-29 22:34:13 +0900635static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636
Pekka Enberg343e0d72006-02-01 03:05:50 -0800637static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638{
639 return cachep->array[smp_processor_id()];
640}
641
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800642static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800644 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
645}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
Andrew Mortona737b3e2006-03-22 00:08:11 -0800647/*
648 * Calculate the number of objects and left-over bytes for a given buffer size.
649 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800650static void cache_estimate(unsigned long gfporder, size_t buffer_size,
651 size_t align, int flags, size_t *left_over,
652 unsigned int *num)
653{
654 int nr_objs;
655 size_t mgmt_size;
656 size_t slab_size = PAGE_SIZE << gfporder;
657
658 /*
659 * The slab management structure can be either off the slab or
660 * on it. For the latter case, the memory allocated for a
661 * slab is used for:
662 *
663 * - The struct slab
664 * - One kmem_bufctl_t for each object
665 * - Padding to respect alignment of @align
666 * - @buffer_size bytes for each object
667 *
668 * If the slab management structure is off the slab, then the
669 * alignment will already be calculated into the size. Because
670 * the slabs are all pages aligned, the objects will be at the
671 * correct alignment when allocated.
672 */
673 if (flags & CFLGS_OFF_SLAB) {
674 mgmt_size = 0;
675 nr_objs = slab_size / buffer_size;
676
677 if (nr_objs > SLAB_LIMIT)
678 nr_objs = SLAB_LIMIT;
679 } else {
680 /*
681 * Ignore padding for the initial guess. The padding
682 * is at most @align-1 bytes, and @buffer_size is at
683 * least @align. In the worst case, this result will
684 * be one greater than the number of objects that fit
685 * into the memory allocation when taking the padding
686 * into account.
687 */
688 nr_objs = (slab_size - sizeof(struct slab)) /
689 (buffer_size + sizeof(kmem_bufctl_t));
690
691 /*
692 * This calculated number will be either the right
693 * amount, or one greater than what we want.
694 */
695 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
696 > slab_size)
697 nr_objs--;
698
699 if (nr_objs > SLAB_LIMIT)
700 nr_objs = SLAB_LIMIT;
701
702 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800704 *num = nr_objs;
705 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706}
707
Christoph Lameterf28510d2012-09-11 19:49:38 +0000708#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700709#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
Andrew Mortona737b3e2006-03-22 00:08:11 -0800711static void __slab_error(const char *function, struct kmem_cache *cachep,
712 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713{
714 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800715 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030717 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000719#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
Paul Menage3395ee02006-12-06 20:32:16 -0800721/*
722 * By default on NUMA we use alien caches to stage the freeing of
723 * objects allocated from other nodes. This causes massive memory
724 * inefficiencies when using fake NUMA setup to split memory into a
725 * large number of small nodes, so it can be disabled on the command
726 * line
727 */
728
729static int use_alien_caches __read_mostly = 1;
730static int __init noaliencache_setup(char *s)
731{
732 use_alien_caches = 0;
733 return 1;
734}
735__setup("noaliencache", noaliencache_setup);
736
David Rientjes3df1ccc2011-10-18 22:09:28 -0700737static int __init slab_max_order_setup(char *str)
738{
739 get_option(&str, &slab_max_order);
740 slab_max_order = slab_max_order < 0 ? 0 :
741 min(slab_max_order, MAX_ORDER - 1);
742 slab_max_order_set = true;
743
744 return 1;
745}
746__setup("slab_max_order=", slab_max_order_setup);
747
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800748#ifdef CONFIG_NUMA
749/*
750 * Special reaping functions for NUMA systems called from cache_reap().
751 * These take care of doing round robin flushing of alien caches (containing
752 * objects freed on different nodes from which they were allocated) and the
753 * flushing of remote pcps by calling drain_node_pages.
754 */
Tejun Heo1871e522009-10-29 22:34:13 +0900755static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800756
757static void init_reap_node(int cpu)
758{
759 int node;
760
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700761 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800762 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800763 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800764
Tejun Heo1871e522009-10-29 22:34:13 +0900765 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800766}
767
768static void next_reap_node(void)
769{
Christoph Lameter909ea962010-12-08 16:22:55 +0100770 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800771
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800772 node = next_node(node, node_online_map);
773 if (unlikely(node >= MAX_NUMNODES))
774 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100775 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800776}
777
778#else
779#define init_reap_node(cpu) do { } while (0)
780#define next_reap_node(void) do { } while (0)
781#endif
782
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783/*
784 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
785 * via the workqueue/eventd.
786 * Add the CPU number into the expiration time to minimize the possibility of
787 * the CPUs getting into lockstep and contending for the global cache chain
788 * lock.
789 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700790static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791{
Tejun Heo1871e522009-10-29 22:34:13 +0900792 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793
794 /*
795 * When this gets called from do_initcalls via cpucache_init(),
796 * init_workqueues() has already run, so keventd will be setup
797 * at that time.
798 */
David Howells52bad642006-11-22 14:54:01 +0000799 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800800 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700801 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800802 schedule_delayed_work_on(cpu, reap_work,
803 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 }
805}
806
Christoph Lametere498be72005-09-09 13:03:32 -0700807static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300808 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800810 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 struct array_cache *nc = NULL;
812
Pekka Enberg83b519e2009-06-10 19:40:04 +0300813 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100814 /*
815 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300816 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100817 * cache the pointers are not cleared and they could be counted as
818 * valid references during a kmemleak scan. Therefore, kmemleak must
819 * not scan such objects.
820 */
821 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 if (nc) {
823 nc->avail = 0;
824 nc->limit = entries;
825 nc->batchcount = batchcount;
826 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700827 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 }
829 return nc;
830}
831
Mel Gorman072bb0a2012-07-31 16:43:58 -0700832static inline bool is_slab_pfmemalloc(struct slab *slabp)
833{
834 struct page *page = virt_to_page(slabp->s_mem);
835
836 return PageSlabPfmemalloc(page);
837}
838
839/* Clears pfmemalloc_active if no slabs have pfmalloc set */
840static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
841 struct array_cache *ac)
842{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000843 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
Mel Gorman072bb0a2012-07-31 16:43:58 -0700844 struct slab *slabp;
845 unsigned long flags;
846
847 if (!pfmemalloc_active)
848 return;
849
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000850 spin_lock_irqsave(&n->list_lock, flags);
851 list_for_each_entry(slabp, &n->slabs_full, list)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700852 if (is_slab_pfmemalloc(slabp))
853 goto out;
854
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000855 list_for_each_entry(slabp, &n->slabs_partial, list)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700856 if (is_slab_pfmemalloc(slabp))
857 goto out;
858
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000859 list_for_each_entry(slabp, &n->slabs_free, list)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700860 if (is_slab_pfmemalloc(slabp))
861 goto out;
862
863 pfmemalloc_active = false;
864out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000865 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700866}
867
Mel Gorman381760e2012-07-31 16:44:30 -0700868static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700869 gfp_t flags, bool force_refill)
870{
871 int i;
872 void *objp = ac->entry[--ac->avail];
873
874 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
875 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000876 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700877
878 if (gfp_pfmemalloc_allowed(flags)) {
879 clear_obj_pfmemalloc(&objp);
880 return objp;
881 }
882
883 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700884 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700885 /* If a !PFMEMALLOC object is found, swap them */
886 if (!is_obj_pfmemalloc(ac->entry[i])) {
887 objp = ac->entry[i];
888 ac->entry[i] = ac->entry[ac->avail];
889 ac->entry[ac->avail] = objp;
890 return objp;
891 }
892 }
893
894 /*
895 * If there are empty slabs on the slabs_free list and we are
896 * being forced to refill the cache, mark this one !pfmemalloc.
897 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000898 n = cachep->node[numa_mem_id()];
899 if (!list_empty(&n->slabs_free) && force_refill) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700900 struct slab *slabp = virt_to_slab(objp);
Mel Gorman30c29be2012-09-17 14:09:03 -0700901 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
Mel Gorman072bb0a2012-07-31 16:43:58 -0700902 clear_obj_pfmemalloc(&objp);
903 recheck_pfmemalloc_active(cachep, ac);
904 return objp;
905 }
906
907 /* No !PFMEMALLOC objects available */
908 ac->avail++;
909 objp = NULL;
910 }
911
912 return objp;
913}
914
Mel Gorman381760e2012-07-31 16:44:30 -0700915static inline void *ac_get_obj(struct kmem_cache *cachep,
916 struct array_cache *ac, gfp_t flags, bool force_refill)
917{
918 void *objp;
919
920 if (unlikely(sk_memalloc_socks()))
921 objp = __ac_get_obj(cachep, ac, flags, force_refill);
922 else
923 objp = ac->entry[--ac->avail];
924
925 return objp;
926}
927
928static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700929 void *objp)
930{
931 if (unlikely(pfmemalloc_active)) {
932 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700933 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700934 if (PageSlabPfmemalloc(page))
935 set_obj_pfmemalloc(&objp);
936 }
937
Mel Gorman381760e2012-07-31 16:44:30 -0700938 return objp;
939}
940
941static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
942 void *objp)
943{
944 if (unlikely(sk_memalloc_socks()))
945 objp = __ac_put_obj(cachep, ac, objp);
946
Mel Gorman072bb0a2012-07-31 16:43:58 -0700947 ac->entry[ac->avail++] = objp;
948}
949
Christoph Lameter3ded1752006-03-25 03:06:44 -0800950/*
951 * Transfer objects in one arraycache to another.
952 * Locking must be handled by the caller.
953 *
954 * Return the number of entries transferred.
955 */
956static int transfer_objects(struct array_cache *to,
957 struct array_cache *from, unsigned int max)
958{
959 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700960 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800961
962 if (!nr)
963 return 0;
964
965 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
966 sizeof(void *) *nr);
967
968 from->avail -= nr;
969 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800970 return nr;
971}
972
Christoph Lameter765c4502006-09-27 01:50:08 -0700973#ifndef CONFIG_NUMA
974
975#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000976#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700977
Pekka Enberg83b519e2009-06-10 19:40:04 +0300978static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700979{
980 return (struct array_cache **)BAD_ALIEN_MAGIC;
981}
982
983static inline void free_alien_cache(struct array_cache **ac_ptr)
984{
985}
986
987static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
988{
989 return 0;
990}
991
992static inline void *alternate_node_alloc(struct kmem_cache *cachep,
993 gfp_t flags)
994{
995 return NULL;
996}
997
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800998static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700999 gfp_t flags, int nodeid)
1000{
1001 return NULL;
1002}
1003
1004#else /* CONFIG_NUMA */
1005
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001006static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001007static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001008
Pekka Enberg83b519e2009-06-10 19:40:04 +03001009static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001010{
1011 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001012 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001013 int i;
1014
1015 if (limit > 1)
1016 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001017 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001018 if (ac_ptr) {
1019 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001020 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001021 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001022 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001023 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001024 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001025 kfree(ac_ptr[i]);
1026 kfree(ac_ptr);
1027 return NULL;
1028 }
1029 }
1030 }
1031 return ac_ptr;
1032}
1033
Pekka Enberg5295a742006-02-01 03:05:48 -08001034static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001035{
1036 int i;
1037
1038 if (!ac_ptr)
1039 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001040 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001041 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001042 kfree(ac_ptr);
1043}
1044
Pekka Enberg343e0d72006-02-01 03:05:50 -08001045static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001046 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001047{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001048 struct kmem_cache_node *n = cachep->node[node];
Christoph Lametere498be72005-09-09 13:03:32 -07001049
1050 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001051 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001052 /*
1053 * Stuff objects into the remote nodes shared array first.
1054 * That way we could avoid the overhead of putting the objects
1055 * into the free lists and getting them back later.
1056 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001057 if (n->shared)
1058 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001059
Christoph Lameterff694162005-09-22 21:44:02 -07001060 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001061 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001062 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001063 }
1064}
1065
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001066/*
1067 * Called from cache_reap() to regularly drain alien caches round robin.
1068 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001069static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001070{
Christoph Lameter909ea962010-12-08 16:22:55 +01001071 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001072
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001073 if (n->alien) {
1074 struct array_cache *ac = n->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001075
1076 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001077 __drain_alien_cache(cachep, ac, node);
1078 spin_unlock_irq(&ac->lock);
1079 }
1080 }
1081}
1082
Andrew Mortona737b3e2006-03-22 00:08:11 -08001083static void drain_alien_cache(struct kmem_cache *cachep,
1084 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001085{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001086 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001087 struct array_cache *ac;
1088 unsigned long flags;
1089
1090 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001091 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001092 if (ac) {
1093 spin_lock_irqsave(&ac->lock, flags);
1094 __drain_alien_cache(cachep, ac, i);
1095 spin_unlock_irqrestore(&ac->lock, flags);
1096 }
1097 }
1098}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001099
Ingo Molnar873623d2006-07-13 14:44:38 +02001100static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001101{
1102 struct slab *slabp = virt_to_slab(objp);
1103 int nodeid = slabp->nodeid;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001104 struct kmem_cache_node *n;
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001105 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001106 int node;
1107
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001108 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001109
1110 /*
1111 * Make sure we are not freeing a object from another node to the array
1112 * cache on this cpu.
1113 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001114 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001115 return 0;
1116
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001117 n = cachep->node[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001118 STATS_INC_NODEFREES(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001119 if (n->alien && n->alien[nodeid]) {
1120 alien = n->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001121 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001122 if (unlikely(alien->avail == alien->limit)) {
1123 STATS_INC_ACOVERFLOW(cachep);
1124 __drain_alien_cache(cachep, alien, nodeid);
1125 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001126 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001127 spin_unlock(&alien->lock);
1128 } else {
Christoph Lameter6a673682013-01-10 19:14:19 +00001129 spin_lock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001130 free_block(cachep, &objp, 1, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001131 spin_unlock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001132 }
1133 return 1;
1134}
Christoph Lametere498be72005-09-09 13:03:32 -07001135#endif
1136
David Rientjes8f9f8d92010-03-27 19:40:47 -07001137/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001138 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001139 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001140 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001141 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001142 * already in use.
1143 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001144 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001145 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001146static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001147{
1148 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001149 struct kmem_cache_node *n;
Christoph Lameter6744f082013-01-10 19:12:17 +00001150 const int memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001151
Christoph Lameter18004c52012-07-06 15:25:12 -05001152 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001153 /*
1154 * Set up the size64 kmemlist for cpu before we can
1155 * begin anything. Make sure some other cpu on this
1156 * node has not already allocated this
1157 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001158 if (!cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001159 n = kmalloc_node(memsize, GFP_KERNEL, node);
1160 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001161 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001162 kmem_cache_node_init(n);
1163 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
David Rientjes8f9f8d92010-03-27 19:40:47 -07001164 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1165
1166 /*
1167 * The l3s don't come and go as CPUs come and
Christoph Lameter18004c52012-07-06 15:25:12 -05001168 * go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001169 * protection here.
1170 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001171 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001172 }
1173
Christoph Lameter6a673682013-01-10 19:14:19 +00001174 spin_lock_irq(&cachep->node[node]->list_lock);
1175 cachep->node[node]->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001176 (1 + nr_cpus_node(node)) *
1177 cachep->batchcount + cachep->num;
Christoph Lameter6a673682013-01-10 19:14:19 +00001178 spin_unlock_irq(&cachep->node[node]->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001179 }
1180 return 0;
1181}
1182
Wanpeng Li0fa81032013-07-04 08:33:22 +08001183static inline int slabs_tofree(struct kmem_cache *cachep,
1184 struct kmem_cache_node *n)
1185{
1186 return (n->free_objects + cachep->num - 1) / cachep->num;
1187}
1188
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001189static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001191 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001192 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001193 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301194 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001195
Christoph Lameter18004c52012-07-06 15:25:12 -05001196 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001197 struct array_cache *nc;
1198 struct array_cache *shared;
1199 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001200
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001201 /* cpu is dead; no one can alloc from it. */
1202 nc = cachep->array[cpu];
1203 cachep->array[cpu] = NULL;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001204 n = cachep->node[node];
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001205
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001206 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001207 goto free_array_cache;
1208
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001209 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001210
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001211 /* Free limit for this kmem_cache_node */
1212 n->free_limit -= cachep->batchcount;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001213 if (nc)
1214 free_block(cachep, nc->entry, nc->avail, node);
1215
Rusty Russell58463c12009-12-17 11:43:12 -06001216 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001217 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001218 goto free_array_cache;
1219 }
1220
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001221 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001222 if (shared) {
1223 free_block(cachep, shared->entry,
1224 shared->avail, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001225 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001226 }
1227
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001228 alien = n->alien;
1229 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001230
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001231 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001232
1233 kfree(shared);
1234 if (alien) {
1235 drain_alien_cache(cachep, alien);
1236 free_alien_cache(alien);
1237 }
1238free_array_cache:
1239 kfree(nc);
1240 }
1241 /*
1242 * In the previous loop, all the objects were freed to
1243 * the respective cache's slabs, now we can go ahead and
1244 * shrink each nodelist to its limit.
1245 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001246 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001247 n = cachep->node[node];
1248 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001249 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001250 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001251 }
1252}
1253
1254static int __cpuinit cpuup_prepare(long cpu)
1255{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001256 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001257 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001258 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001259 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001261 /*
1262 * We need to do this right in the beginning since
1263 * alloc_arraycache's are going to use this list.
1264 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001265 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001266 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001267 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001268 if (err < 0)
1269 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001270
1271 /*
1272 * Now we can go ahead with allocating the shared arrays and
1273 * array caches
1274 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001275 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001276 struct array_cache *nc;
1277 struct array_cache *shared = NULL;
1278 struct array_cache **alien = NULL;
1279
1280 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001281 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001282 if (!nc)
1283 goto bad;
1284 if (cachep->shared) {
1285 shared = alloc_arraycache(node,
1286 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001287 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001288 if (!shared) {
1289 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001290 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001291 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001292 }
1293 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001294 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001295 if (!alien) {
1296 kfree(shared);
1297 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001298 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001299 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001300 }
1301 cachep->array[cpu] = nc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001302 n = cachep->node[node];
1303 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001304
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001305 spin_lock_irq(&n->list_lock);
1306 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001307 /*
1308 * We are serialised from CPU_DEAD or
1309 * CPU_UP_CANCELLED by the cpucontrol lock
1310 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001311 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001312 shared = NULL;
1313 }
1314#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001315 if (!n->alien) {
1316 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001317 alien = NULL;
1318 }
1319#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001320 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001321 kfree(shared);
1322 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001323 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1324 slab_set_debugobj_lock_classes_node(cachep, node);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08001325 else if (!OFF_SLAB(cachep) &&
1326 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1327 on_slab_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001328 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001329 init_node_lock_keys(node);
1330
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001331 return 0;
1332bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001333 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001334 return -ENOMEM;
1335}
1336
1337static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1338 unsigned long action, void *hcpu)
1339{
1340 long cpu = (long)hcpu;
1341 int err = 0;
1342
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001344 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001345 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001346 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001347 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001348 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 break;
1350 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001351 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 start_cpu_timer(cpu);
1353 break;
1354#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001355 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001356 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001357 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001358 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001359 * held so that if cache_reap() is invoked it cannot do
1360 * anything expensive but will only modify reap_work
1361 * and reschedule the timer.
1362 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001363 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001364 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001365 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001366 break;
1367 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001368 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001369 start_cpu_timer(cpu);
1370 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001372 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001373 /*
1374 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001375 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001376 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001377 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001378 * structure is usually allocated from kmem_cache_create() and
1379 * gets destroyed at kmem_cache_destroy().
1380 */
Simon Arlott183ff222007-10-20 01:27:18 +02001381 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001382#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001384 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001385 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001386 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001387 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001390 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391}
1392
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001393static struct notifier_block __cpuinitdata cpucache_notifier = {
1394 &cpuup_callback, NULL, 0
1395};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
David Rientjes8f9f8d92010-03-27 19:40:47 -07001397#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1398/*
1399 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1400 * Returns -EBUSY if all objects cannot be drained so that the node is not
1401 * removed.
1402 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001403 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001404 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001405static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001406{
1407 struct kmem_cache *cachep;
1408 int ret = 0;
1409
Christoph Lameter18004c52012-07-06 15:25:12 -05001410 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001411 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001412
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001413 n = cachep->node[node];
1414 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001415 continue;
1416
Wanpeng Li0fa81032013-07-04 08:33:22 +08001417 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001418
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001419 if (!list_empty(&n->slabs_full) ||
1420 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001421 ret = -EBUSY;
1422 break;
1423 }
1424 }
1425 return ret;
1426}
1427
1428static int __meminit slab_memory_callback(struct notifier_block *self,
1429 unsigned long action, void *arg)
1430{
1431 struct memory_notify *mnb = arg;
1432 int ret = 0;
1433 int nid;
1434
1435 nid = mnb->status_change_nid;
1436 if (nid < 0)
1437 goto out;
1438
1439 switch (action) {
1440 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001441 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001442 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001443 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001444 break;
1445 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001446 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001447 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001448 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001449 break;
1450 case MEM_ONLINE:
1451 case MEM_OFFLINE:
1452 case MEM_CANCEL_ONLINE:
1453 case MEM_CANCEL_OFFLINE:
1454 break;
1455 }
1456out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001457 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001458}
1459#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1460
Christoph Lametere498be72005-09-09 13:03:32 -07001461/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001462 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001463 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001464static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001465 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001466{
Christoph Lameter6744f082013-01-10 19:12:17 +00001467 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001468
Christoph Lameter6744f082013-01-10 19:12:17 +00001469 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001470 BUG_ON(!ptr);
1471
Christoph Lameter6744f082013-01-10 19:12:17 +00001472 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001473 /*
1474 * Do not assume that spinlocks can be initialized via memcpy:
1475 */
1476 spin_lock_init(&ptr->list_lock);
1477
Christoph Lametere498be72005-09-09 13:03:32 -07001478 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001479 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001480}
1481
Andrew Mortona737b3e2006-03-22 00:08:11 -08001482/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001483 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1484 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001485 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001486static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001487{
1488 int node;
1489
1490 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001491 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001492 cachep->node[node]->next_reap = jiffies +
Pekka Enberg556a1692008-01-25 08:20:51 +02001493 REAPTIMEOUT_LIST3 +
1494 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1495 }
1496}
1497
1498/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001499 * The memory after the last cpu cache pointer is used for the
Christoph Lameter6a673682013-01-10 19:14:19 +00001500 * the node pointer.
Christoph Lameter3c583462012-11-28 16:23:01 +00001501 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001502static void setup_node_pointer(struct kmem_cache *cachep)
Christoph Lameter3c583462012-11-28 16:23:01 +00001503{
Christoph Lameter6a673682013-01-10 19:14:19 +00001504 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
Christoph Lameter3c583462012-11-28 16:23:01 +00001505}
1506
1507/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001508 * Initialisation. Called after the page allocator have been initialised and
1509 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 */
1511void __init kmem_cache_init(void)
1512{
Christoph Lametere498be72005-09-09 13:03:32 -07001513 int i;
1514
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001515 kmem_cache = &kmem_cache_boot;
Christoph Lameter6a673682013-01-10 19:14:19 +00001516 setup_node_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001517
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001518 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001519 use_alien_caches = 0;
1520
Christoph Lameter3c583462012-11-28 16:23:01 +00001521 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001522 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001523
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001524 set_up_node(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525
1526 /*
1527 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001528 * page orders on machines with more than 32MB of memory if
1529 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001531 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001532 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 /* Bootstrap is tricky, because several objects are allocated
1535 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001536 * 1) initialize the kmem_cache cache: it contains the struct
1537 * kmem_cache structures of all caches, except kmem_cache itself:
1538 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001539 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001540 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001541 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001543 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001544 * An __init data area is used for the head array.
1545 * 3) Create the remaining kmalloc caches, with minimally sized
1546 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001547 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001549 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001550 * the other cache's with kmalloc allocated memory.
1551 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 */
1553
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001554 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555
Eric Dumazet8da34302007-05-06 14:49:29 -07001556 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001557 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001558 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001559 create_boot_cache(kmem_cache, "kmem_cache",
1560 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001561 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001562 SLAB_HWCACHE_ALIGN);
1563 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564
1565 /* 2+3) create the kmalloc caches */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566
Andrew Mortona737b3e2006-03-22 00:08:11 -08001567 /*
1568 * Initialize the caches that provide memory for the array cache and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001569 * kmem_cache_node structures first. Without this, further allocations will
Andrew Mortona737b3e2006-03-22 00:08:11 -08001570 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001571 */
1572
Christoph Lametere3366012013-01-10 19:14:18 +00001573 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1574 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001575
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001576 if (INDEX_AC != INDEX_NODE)
1577 kmalloc_caches[INDEX_NODE] =
1578 create_kmalloc_cache("kmalloc-node",
1579 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001580
Ingo Molnare0a42722006-06-23 02:03:46 -07001581 slab_early_init = 0;
1582
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 /* 4) Replace the bootstrap head arrays */
1584 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001585 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001586
Pekka Enberg83b519e2009-06-10 19:40:04 +03001587 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001588
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001589 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001590 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001591 /*
1592 * Do not assume that spinlocks can be initialized via memcpy:
1593 */
1594 spin_lock_init(&ptr->lock);
1595
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001596 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001597
Pekka Enberg83b519e2009-06-10 19:40:04 +03001598 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001599
Christoph Lametere3366012013-01-10 19:14:18 +00001600 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001601 != &initarray_generic.cache);
Christoph Lametere3366012013-01-10 19:14:18 +00001602 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001603 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001604 /*
1605 * Do not assume that spinlocks can be initialized via memcpy:
1606 */
1607 spin_lock_init(&ptr->lock);
1608
Christoph Lametere3366012013-01-10 19:14:18 +00001609 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001611 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001612 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001613 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614
Mel Gorman9c09a952008-01-24 05:49:54 -08001615 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001616 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001617
Christoph Lametere3366012013-01-10 19:14:18 +00001618 init_list(kmalloc_caches[INDEX_AC],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001619 &init_kmem_cache_node[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001620
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001621 if (INDEX_AC != INDEX_NODE) {
1622 init_list(kmalloc_caches[INDEX_NODE],
1623 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001624 }
1625 }
1626 }
1627
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001628 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001629}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001630
Pekka Enberg8429db52009-06-12 15:58:59 +03001631void __init kmem_cache_init_late(void)
1632{
1633 struct kmem_cache *cachep;
1634
Christoph Lameter97d06602012-07-06 15:25:11 -05001635 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001636
Pekka Enberg8429db52009-06-12 15:58:59 +03001637 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001638 mutex_lock(&slab_mutex);
1639 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001640 if (enable_cpucache(cachep, GFP_NOWAIT))
1641 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001642 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001643
Michael Wang947ca182012-09-05 10:33:18 +08001644 /* Annotate slab for lockdep -- annotate the malloc caches */
1645 init_lock_keys();
1646
Christoph Lameter97d06602012-07-06 15:25:11 -05001647 /* Done! */
1648 slab_state = FULL;
1649
Andrew Mortona737b3e2006-03-22 00:08:11 -08001650 /*
1651 * Register a cpu startup notifier callback that initializes
1652 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 */
1654 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655
David Rientjes8f9f8d92010-03-27 19:40:47 -07001656#ifdef CONFIG_NUMA
1657 /*
1658 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001659 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001660 */
1661 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1662#endif
1663
Andrew Mortona737b3e2006-03-22 00:08:11 -08001664 /*
1665 * The reap timers are started later, with a module init call: That part
1666 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 */
1668}
1669
1670static int __init cpucache_init(void)
1671{
1672 int cpu;
1673
Andrew Mortona737b3e2006-03-22 00:08:11 -08001674 /*
1675 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 */
Christoph Lametere498be72005-09-09 13:03:32 -07001677 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001678 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001679
1680 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001681 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 return 0;
1683}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684__initcall(cpucache_init);
1685
Rafael Aquini8bdec192012-03-09 17:27:27 -03001686static noinline void
1687slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1688{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001689 struct kmem_cache_node *n;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001690 struct slab *slabp;
1691 unsigned long flags;
1692 int node;
1693
1694 printk(KERN_WARNING
1695 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1696 nodeid, gfpflags);
1697 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001698 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001699
1700 for_each_online_node(node) {
1701 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1702 unsigned long active_slabs = 0, num_slabs = 0;
1703
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001704 n = cachep->node[node];
1705 if (!n)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001706 continue;
1707
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001708 spin_lock_irqsave(&n->list_lock, flags);
1709 list_for_each_entry(slabp, &n->slabs_full, list) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001710 active_objs += cachep->num;
1711 active_slabs++;
1712 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001713 list_for_each_entry(slabp, &n->slabs_partial, list) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001714 active_objs += slabp->inuse;
1715 active_slabs++;
1716 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001717 list_for_each_entry(slabp, &n->slabs_free, list)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001718 num_slabs++;
1719
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001720 free_objects += n->free_objects;
1721 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001722
1723 num_slabs += active_slabs;
1724 num_objs = num_slabs * cachep->num;
1725 printk(KERN_WARNING
1726 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1727 node, active_slabs, num_slabs, active_objs, num_objs,
1728 free_objects);
1729 }
1730}
1731
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732/*
1733 * Interface to system's page allocator. No need to hold the cache-lock.
1734 *
1735 * If we requested dmaable memory, we will get it. Even if we
1736 * did not request dmaable memory, we might get it, but that
1737 * would be relatively rare and ignorable.
1738 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001739static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740{
1741 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001742 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 int i;
1744
Luke Yangd6fef9d2006-04-10 22:52:56 -07001745#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001746 /*
1747 * Nommu uses slab's for process anonymous memory allocations, and thus
1748 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001749 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001750 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001751#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001752
Glauber Costaa618e892012-06-14 16:17:21 +04001753 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001754 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1755 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001756
Linus Torvalds517d0862009-06-16 19:50:13 -07001757 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001758 if (!page) {
1759 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1760 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001762 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001764 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001765 if (unlikely(page->pfmemalloc))
1766 pfmemalloc_active = true;
1767
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001768 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001770 add_zone_page_state(page_zone(page),
1771 NR_SLAB_RECLAIMABLE, nr_pages);
1772 else
1773 add_zone_page_state(page_zone(page),
1774 NR_SLAB_UNRECLAIMABLE, nr_pages);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001775 for (i = 0; i < nr_pages; i++) {
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001776 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001777
Mel Gorman072bb0a2012-07-31 16:43:58 -07001778 if (page->pfmemalloc)
1779 SetPageSlabPfmemalloc(page + i);
1780 }
Glauber Costa1f458cb2012-12-18 14:22:50 -08001781 memcg_bind_pages(cachep, cachep->gfporder);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001782
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001783 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1784 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1785
1786 if (cachep->ctor)
1787 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1788 else
1789 kmemcheck_mark_unallocated_pages(page, nr_pages);
1790 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001791
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001792 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793}
1794
1795/*
1796 * Interface to system's page release.
1797 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001798static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 struct page *page = virt_to_page(addr);
1802 const unsigned long nr_freed = i;
1803
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001804 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001805
Christoph Lameter972d1a72006-09-25 23:31:51 -07001806 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1807 sub_zone_page_state(page_zone(page),
1808 NR_SLAB_RECLAIMABLE, nr_freed);
1809 else
1810 sub_zone_page_state(page_zone(page),
1811 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001813 BUG_ON(!PageSlab(page));
Mel Gorman072bb0a2012-07-31 16:43:58 -07001814 __ClearPageSlabPfmemalloc(page);
Nick Pigginf205b2f2006-03-22 00:08:02 -08001815 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 page++;
1817 }
Glauber Costa1f458cb2012-12-18 14:22:50 -08001818
1819 memcg_release_pages(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 if (current->reclaim_state)
1821 current->reclaim_state->reclaimed_slab += nr_freed;
Glauber Costad79923f2012-12-18 14:22:48 -08001822 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823}
1824
1825static void kmem_rcu_free(struct rcu_head *head)
1826{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001827 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001828 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829
1830 kmem_freepages(cachep, slab_rcu->addr);
1831 if (OFF_SLAB(cachep))
1832 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1833}
1834
1835#if DEBUG
1836
1837#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001838static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001839 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001841 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001843 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001845 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 return;
1847
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001848 *addr++ = 0x12345678;
1849 *addr++ = caller;
1850 *addr++ = smp_processor_id();
1851 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 {
1853 unsigned long *sptr = &caller;
1854 unsigned long svalue;
1855
1856 while (!kstack_end(sptr)) {
1857 svalue = *sptr++;
1858 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001859 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 size -= sizeof(unsigned long);
1861 if (size <= sizeof(unsigned long))
1862 break;
1863 }
1864 }
1865
1866 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001867 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868}
1869#endif
1870
Pekka Enberg343e0d72006-02-01 03:05:50 -08001871static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001873 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001874 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875
1876 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878}
1879
1880static void dump_line(char *data, int offset, int limit)
1881{
1882 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001883 unsigned char error = 0;
1884 int bad_count = 0;
1885
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001886 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001887 for (i = 0; i < limit; i++) {
1888 if (data[offset + i] != POISON_FREE) {
1889 error = data[offset + i];
1890 bad_count++;
1891 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001892 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001893 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1894 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001895
1896 if (bad_count == 1) {
1897 error ^= POISON_FREE;
1898 if (!(error & (error - 1))) {
1899 printk(KERN_ERR "Single bit error detected. Probably "
1900 "bad RAM.\n");
1901#ifdef CONFIG_X86
1902 printk(KERN_ERR "Run memtest86+ or a similar memory "
1903 "test tool.\n");
1904#else
1905 printk(KERN_ERR "Run a memory test tool.\n");
1906#endif
1907 }
1908 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909}
1910#endif
1911
1912#if DEBUG
1913
Pekka Enberg343e0d72006-02-01 03:05:50 -08001914static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915{
1916 int i, size;
1917 char *realobj;
1918
1919 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001920 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001921 *dbg_redzone1(cachep, objp),
1922 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 }
1924
1925 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001926 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1927 *dbg_userword(cachep, objp),
1928 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001930 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001931 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001932 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 int limit;
1934 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001935 if (i + limit > size)
1936 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 dump_line(realobj, i, limit);
1938 }
1939}
1940
Pekka Enberg343e0d72006-02-01 03:05:50 -08001941static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942{
1943 char *realobj;
1944 int size, i;
1945 int lines = 0;
1946
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001947 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001948 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001950 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001952 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 exp = POISON_END;
1954 if (realobj[i] != exp) {
1955 int limit;
1956 /* Mismatch ! */
1957 /* Print header */
1958 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001959 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001960 "Slab corruption (%s): %s start=%p, len=%d\n",
1961 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 print_objinfo(cachep, objp, 0);
1963 }
1964 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001965 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001967 if (i + limit > size)
1968 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 dump_line(realobj, i, limit);
1970 i += 16;
1971 lines++;
1972 /* Limit to 5 lines */
1973 if (lines > 5)
1974 break;
1975 }
1976 }
1977 if (lines != 0) {
1978 /* Print some data about the neighboring objects, if they
1979 * exist:
1980 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001981 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001982 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001984 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001986 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001987 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001989 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 print_objinfo(cachep, objp, 2);
1991 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001992 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001993 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001994 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001996 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 print_objinfo(cachep, objp, 2);
1998 }
1999 }
2000}
2001#endif
2002
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05302004static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002005{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 int i;
2007 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002008 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009
2010 if (cachep->flags & SLAB_POISON) {
2011#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002012 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002013 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002014 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002015 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016 else
2017 check_poison_obj(cachep, objp);
2018#else
2019 check_poison_obj(cachep, objp);
2020#endif
2021 }
2022 if (cachep->flags & SLAB_RED_ZONE) {
2023 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2024 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002025 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2027 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002028 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002031}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032#else
Rabin Vincente79aec22008-07-04 00:40:32 +05302033static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002034{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002035}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036#endif
2037
Randy Dunlap911851e2006-03-22 00:08:14 -08002038/**
2039 * slab_destroy - destroy and release all objects in a slab
2040 * @cachep: cache pointer being destroyed
2041 * @slabp: slab pointer being destroyed
2042 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002043 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002044 * Before calling the slab must have been unlinked from the cache. The
2045 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002046 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002047static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002048{
2049 void *addr = slabp->s_mem - slabp->colouroff;
2050
Rabin Vincente79aec22008-07-04 00:40:32 +05302051 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2053 struct slab_rcu *slab_rcu;
2054
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002055 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 slab_rcu->cachep = cachep;
2057 slab_rcu->addr = addr;
2058 call_rcu(&slab_rcu->head, kmem_rcu_free);
2059 } else {
2060 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02002061 if (OFF_SLAB(cachep))
2062 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 }
2064}
2065
2066/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002067 * calculate_slab_order - calculate size (page order) of slabs
2068 * @cachep: pointer to the cache that is being created
2069 * @size: size of objects to be created in this cache.
2070 * @align: required alignment for the objects.
2071 * @flags: slab allocation flags
2072 *
2073 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002074 *
2075 * This could be made much more intelligent. For now, try to avoid using
2076 * high order pages for slabs. When the gfp() functions are more friendly
2077 * towards high-order requests, this should be changed.
2078 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002079static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002080 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002081{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002082 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002083 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002084 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002085
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002086 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002087 unsigned int num;
2088 size_t remainder;
2089
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002090 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002091 if (!num)
2092 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002093
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002094 if (flags & CFLGS_OFF_SLAB) {
2095 /*
2096 * Max number of objs-per-slab for caches which
2097 * use off-slab slabs. Needed to avoid a possible
2098 * looping condition in cache_grow().
2099 */
2100 offslab_limit = size - sizeof(struct slab);
2101 offslab_limit /= sizeof(kmem_bufctl_t);
2102
2103 if (num > offslab_limit)
2104 break;
2105 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002106
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002107 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002108 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002109 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002110 left_over = remainder;
2111
2112 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002113 * A VFS-reclaimable slab tends to have most allocations
2114 * as GFP_NOFS and we really don't want to have to be allocating
2115 * higher-order pages when we are unable to shrink dcache.
2116 */
2117 if (flags & SLAB_RECLAIM_ACCOUNT)
2118 break;
2119
2120 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002121 * Large number of objects is good, but very large slabs are
2122 * currently bad for the gfp()s.
2123 */
David Rientjes543585c2011-10-18 22:09:24 -07002124 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002125 break;
2126
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002127 /*
2128 * Acceptable internal fragmentation?
2129 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002130 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002131 break;
2132 }
2133 return left_over;
2134}
2135
Pekka Enberg83b519e2009-06-10 19:40:04 +03002136static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002137{
Christoph Lameter97d06602012-07-06 15:25:11 -05002138 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002139 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002140
Christoph Lameter97d06602012-07-06 15:25:11 -05002141 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002142 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002143 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002144 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002145 * of by the caller of __kmem_cache_create
2146 */
2147 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2148 slab_state = PARTIAL;
2149 } else if (slab_state == PARTIAL) {
2150 /*
2151 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002152 * that's used by kmalloc(24), otherwise the creation of
2153 * further caches will BUG().
2154 */
2155 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2156
2157 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002158 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2159 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002160 * otherwise the creation of further caches will BUG().
2161 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002162 set_up_node(cachep, SIZE_AC);
2163 if (INDEX_AC == INDEX_NODE)
2164 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002165 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002166 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002167 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002168 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002169 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002170 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002171
Christoph Lameter97d06602012-07-06 15:25:11 -05002172 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002173 set_up_node(cachep, SIZE_NODE);
2174 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002175 } else {
2176 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002177 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002178 cachep->node[node] =
Christoph Lameter6744f082013-01-10 19:12:17 +00002179 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002180 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002181 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002182 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002183 }
2184 }
2185 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002186 cachep->node[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002187 jiffies + REAPTIMEOUT_LIST3 +
2188 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2189
2190 cpu_cache_get(cachep)->avail = 0;
2191 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2192 cpu_cache_get(cachep)->batchcount = 1;
2193 cpu_cache_get(cachep)->touched = 0;
2194 cachep->batchcount = 1;
2195 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002196 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002197}
2198
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002199/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002200 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002201 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 *
2204 * Returns a ptr to the cache on success, NULL on failure.
2205 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002206 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 * The flags are
2209 *
2210 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2211 * to catch references to uninitialised memory.
2212 *
2213 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2214 * for buffer overruns.
2215 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2217 * cacheline. This can be beneficial if you're counting cycles as closely
2218 * as davem.
2219 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002220int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002221__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222{
2223 size_t left_over, slab_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002224 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002225 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002226 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229#if FORCED_DEBUG
2230 /*
2231 * Enable redzoning and last user accounting, except for caches with
2232 * large objects, if the increased size would increase the object size
2233 * above the next power of two: caches with object sizes just above a
2234 * power of two have a significant amount of internal fragmentation.
2235 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002236 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2237 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002238 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 if (!(flags & SLAB_DESTROY_BY_RCU))
2240 flags |= SLAB_POISON;
2241#endif
2242 if (flags & SLAB_DESTROY_BY_RCU)
2243 BUG_ON(flags & SLAB_POISON);
2244#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245
Andrew Mortona737b3e2006-03-22 00:08:11 -08002246 /*
2247 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 * unaligned accesses for some archs when redzoning is used, and makes
2249 * sure any on-slab bufctl's are also correctly aligned.
2250 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002251 if (size & (BYTES_PER_WORD - 1)) {
2252 size += (BYTES_PER_WORD - 1);
2253 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 }
2255
Pekka Enbergca5f9702006-09-25 23:31:25 -07002256 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002257 * Redzoning and user store require word alignment or possibly larger.
2258 * Note this will be overridden by architecture or caller mandated
2259 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002260 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002261 if (flags & SLAB_STORE_USER)
2262 ralign = BYTES_PER_WORD;
2263
2264 if (flags & SLAB_RED_ZONE) {
2265 ralign = REDZONE_ALIGN;
2266 /* If redzoning, ensure that the second redzone is suitably
2267 * aligned, by adjusting the object size accordingly. */
2268 size += REDZONE_ALIGN - 1;
2269 size &= ~(REDZONE_ALIGN - 1);
2270 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002271
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002272 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002273 if (ralign < cachep->align) {
2274 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002276 /* disable debug if necessary */
2277 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002278 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002279 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002280 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002282 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
Pekka Enberg83b519e2009-06-10 19:40:04 +03002284 if (slab_is_available())
2285 gfp = GFP_KERNEL;
2286 else
2287 gfp = GFP_NOWAIT;
2288
Christoph Lameter6a673682013-01-10 19:14:19 +00002289 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
Pekka Enbergca5f9702006-09-25 23:31:25 -07002292 /*
2293 * Both debugging options require word-alignment which is calculated
2294 * into align above.
2295 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002298 cachep->obj_offset += sizeof(unsigned long long);
2299 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 }
2301 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002302 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002303 * the real object. But if the second red zone needs to be
2304 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002306 if (flags & SLAB_RED_ZONE)
2307 size += REDZONE_ALIGN;
2308 else
2309 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 }
2311#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002312 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002313 && cachep->object_size > cache_line_size()
2314 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2315 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 size = PAGE_SIZE;
2317 }
2318#endif
2319#endif
2320
Ingo Molnare0a42722006-06-23 02:03:46 -07002321 /*
2322 * Determine if the slab management is 'on' or 'off' slab.
2323 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002324 * it too early on. Always use on-slab management when
2325 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002326 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002327 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2328 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 /*
2330 * Size is large, assume best to place the slab management obj
2331 * off-slab (should allow better packing of objs).
2332 */
2333 flags |= CFLGS_OFF_SLAB;
2334
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002335 size = ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002337 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002339 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002340 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002341
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002342 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002343 + sizeof(struct slab), cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344
2345 /*
2346 * If the slab has been placed off-slab, and we have enough space then
2347 * move it on-slab. This is at the expense of any extra colouring.
2348 */
2349 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2350 flags &= ~CFLGS_OFF_SLAB;
2351 left_over -= slab_size;
2352 }
2353
2354 if (flags & CFLGS_OFF_SLAB) {
2355 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002356 slab_size =
2357 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302358
2359#ifdef CONFIG_PAGE_POISONING
2360 /* If we're going to use the generic kernel_map_pages()
2361 * poisoning, then it's going to smash the contents of
2362 * the redzone and userword anyhow, so switch them off.
2363 */
2364 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2365 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2366#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 }
2368
2369 cachep->colour_off = cache_line_size();
2370 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002371 if (cachep->colour_off < cachep->align)
2372 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002373 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 cachep->slab_size = slab_size;
2375 cachep->flags = flags;
Glauber Costaa618e892012-06-14 16:17:21 +04002376 cachep->allocflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002377 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002378 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002379 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002380 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002382 if (flags & CFLGS_OFF_SLAB) {
Christoph Lameter2c59dd62013-01-10 19:14:19 +00002383 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002384 /*
2385 * This is a possibility for one of the malloc_sizes caches.
2386 * But since we go off slab only for object size greater than
2387 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2388 * this should not happen at all.
2389 * But leave a BUG_ON for some lucky dude.
2390 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002391 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002392 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002394 err = setup_cpu_cache(cachep, gfp);
2395 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002396 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002397 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002398 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399
Peter Zijlstra83835b32011-07-22 15:26:05 +02002400 if (flags & SLAB_DEBUG_OBJECTS) {
2401 /*
2402 * Would deadlock through slab_destroy()->call_rcu()->
2403 * debug_object_activate()->kmem_cache_alloc().
2404 */
2405 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2406
2407 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002408 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2409 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002410
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002411 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413
2414#if DEBUG
2415static void check_irq_off(void)
2416{
2417 BUG_ON(!irqs_disabled());
2418}
2419
2420static void check_irq_on(void)
2421{
2422 BUG_ON(irqs_disabled());
2423}
2424
Pekka Enberg343e0d72006-02-01 03:05:50 -08002425static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426{
2427#ifdef CONFIG_SMP
2428 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002429 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430#endif
2431}
Christoph Lametere498be72005-09-09 13:03:32 -07002432
Pekka Enberg343e0d72006-02-01 03:05:50 -08002433static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002434{
2435#ifdef CONFIG_SMP
2436 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002437 assert_spin_locked(&cachep->node[node]->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002438#endif
2439}
2440
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441#else
2442#define check_irq_off() do { } while(0)
2443#define check_irq_on() do { } while(0)
2444#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002445#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446#endif
2447
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002448static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002449 struct array_cache *ac,
2450 int force, int node);
2451
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452static void do_drain(void *arg)
2453{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002454 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002456 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
2458 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002459 ac = cpu_cache_get(cachep);
Christoph Lameter6a673682013-01-10 19:14:19 +00002460 spin_lock(&cachep->node[node]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002461 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002462 spin_unlock(&cachep->node[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 ac->avail = 0;
2464}
2465
Pekka Enberg343e0d72006-02-01 03:05:50 -08002466static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002468 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002469 int node;
2470
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002471 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002473 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002474 n = cachep->node[node];
2475 if (n && n->alien)
2476 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002477 }
2478
2479 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002480 n = cachep->node[node];
2481 if (n)
2482 drain_array(cachep, n, n->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002483 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484}
2485
Christoph Lametered11d9e2006-06-30 01:55:45 -07002486/*
2487 * Remove slabs from the list of free slabs.
2488 * Specify the number of slabs to drain in tofree.
2489 *
2490 * Returns the actual number of slabs released.
2491 */
2492static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002493 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002495 struct list_head *p;
2496 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498
Christoph Lametered11d9e2006-06-30 01:55:45 -07002499 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002500 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002502 spin_lock_irq(&n->list_lock);
2503 p = n->slabs_free.prev;
2504 if (p == &n->slabs_free) {
2505 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002506 goto out;
2507 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
Christoph Lametered11d9e2006-06-30 01:55:45 -07002509 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002511 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512#endif
2513 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002514 /*
2515 * Safe to drop the lock. The slab is no longer linked
2516 * to the cache.
2517 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002518 n->free_objects -= cache->num;
2519 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002520 slab_destroy(cache, slabp);
2521 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002523out:
2524 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525}
2526
Christoph Lameter18004c52012-07-06 15:25:12 -05002527/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002528static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002529{
2530 int ret = 0, i = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002531 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002532
2533 drain_cpu_caches(cachep);
2534
2535 check_irq_on();
2536 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002537 n = cachep->node[i];
2538 if (!n)
Christoph Lametered11d9e2006-06-30 01:55:45 -07002539 continue;
2540
Wanpeng Li0fa81032013-07-04 08:33:22 +08002541 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002542
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002543 ret += !list_empty(&n->slabs_full) ||
2544 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002545 }
2546 return (ret ? 1 : 0);
2547}
2548
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549/**
2550 * kmem_cache_shrink - Shrink a cache.
2551 * @cachep: The cache to shrink.
2552 *
2553 * Releases as many slabs as possible for a cache.
2554 * To help debugging, a zero exit status indicates all slabs were released.
2555 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002556int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002558 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002559 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002561 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002562 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002563 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002564 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002565 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002566 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567}
2568EXPORT_SYMBOL(kmem_cache_shrink);
2569
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002570int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571{
Christoph Lameter12c36672012-09-04 23:38:33 +00002572 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002573 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002574 int rc = __cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575
Christoph Lameter12c36672012-09-04 23:38:33 +00002576 if (rc)
2577 return rc;
2578
2579 for_each_online_cpu(i)
2580 kfree(cachep->array[i]);
2581
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002582 /* NUMA: free the node structures */
Christoph Lameter12c36672012-09-04 23:38:33 +00002583 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002584 n = cachep->node[i];
2585 if (n) {
2586 kfree(n->shared);
2587 free_alien_cache(n->alien);
2588 kfree(n);
Christoph Lameter12c36672012-09-04 23:38:33 +00002589 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002591 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002594/*
2595 * Get the memory for a slab management obj.
2596 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2597 * always come from malloc_sizes caches. The slab descriptor cannot
2598 * come from the same cache which is getting created because,
2599 * when we are searching for an appropriate cache for these
2600 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2601 * If we are creating a malloc_sizes cache here it would not be visible to
2602 * kmem_find_general_cachep till the initialization is complete.
2603 * Hence we cannot have slabp_cache same as the original cache.
2604 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002605static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002606 int colour_off, gfp_t local_flags,
2607 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608{
2609 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002610
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 if (OFF_SLAB(cachep)) {
2612 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002613 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002614 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002615 /*
2616 * If the first object in the slab is leaked (it's allocated
2617 * but no one has a reference to it), we want to make sure
2618 * kmemleak does not treat the ->s_mem pointer as a reference
2619 * to the object. Otherwise we will not report the leak.
2620 */
Catalin Marinasc017b4b2009-10-28 13:33:09 +00002621 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2622 local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 if (!slabp)
2624 return NULL;
2625 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002626 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 colour_off += cachep->slab_size;
2628 }
2629 slabp->inuse = 0;
2630 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002631 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002632 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002633 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634 return slabp;
2635}
2636
2637static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2638{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002639 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640}
2641
Pekka Enberg343e0d72006-02-01 03:05:50 -08002642static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002643 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644{
2645 int i;
2646
2647 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002648 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649#if DEBUG
2650 /* need to poison the objs? */
2651 if (cachep->flags & SLAB_POISON)
2652 poison_obj(cachep, objp, POISON_FREE);
2653 if (cachep->flags & SLAB_STORE_USER)
2654 *dbg_userword(cachep, objp) = NULL;
2655
2656 if (cachep->flags & SLAB_RED_ZONE) {
2657 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2658 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2659 }
2660 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002661 * Constructors are not allowed to allocate memory from the same
2662 * cache which they are a constructor for. Otherwise, deadlock.
2663 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 */
2665 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002666 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667
2668 if (cachep->flags & SLAB_RED_ZONE) {
2669 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2670 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002671 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2673 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002674 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002676 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002677 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002678 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002679 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680#else
2681 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002682 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002684 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002686 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687}
2688
Pekka Enberg343e0d72006-02-01 03:05:50 -08002689static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002691 if (CONFIG_ZONE_DMA_FLAG) {
2692 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002693 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002694 else
Glauber Costaa618e892012-06-14 16:17:21 +04002695 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002696 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697}
2698
Andrew Mortona737b3e2006-03-22 00:08:11 -08002699static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2700 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002701{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002702 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002703 kmem_bufctl_t next;
2704
2705 slabp->inuse++;
2706 next = slab_bufctl(slabp)[slabp->free];
2707#if DEBUG
2708 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2709 WARN_ON(slabp->nodeid != nodeid);
2710#endif
2711 slabp->free = next;
2712
2713 return objp;
2714}
2715
Andrew Mortona737b3e2006-03-22 00:08:11 -08002716static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2717 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002718{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002719 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002720
2721#if DEBUG
2722 /* Verify that the slab belongs to the intended node */
2723 WARN_ON(slabp->nodeid != nodeid);
2724
Al Viro871751e2006-03-25 03:06:39 -08002725 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002726 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002727 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002728 BUG();
2729 }
2730#endif
2731 slab_bufctl(slabp)[objnr] = slabp->free;
2732 slabp->free = objnr;
2733 slabp->inuse--;
2734}
2735
Pekka Enberg47768742006-06-23 02:03:07 -07002736/*
2737 * Map pages beginning at addr to the given cache and slab. This is required
2738 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002739 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002740 */
2741static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2742 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743{
Pekka Enberg47768742006-06-23 02:03:07 -07002744 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 struct page *page;
2746
Pekka Enberg47768742006-06-23 02:03:07 -07002747 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002748
Pekka Enberg47768742006-06-23 02:03:07 -07002749 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002750 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002751 nr_pages <<= cache->gfporder;
2752
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 do {
Christoph Lameter35026082012-06-13 10:24:56 -05002754 page->slab_cache = cache;
2755 page->slab_page = slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002757 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758}
2759
2760/*
2761 * Grow (by 1) the number of slabs within a cache. This is called by
2762 * kmem_cache_alloc() when there are no active objs left in a cache.
2763 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002764static int cache_grow(struct kmem_cache *cachep,
2765 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002767 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002768 size_t offset;
2769 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002770 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
Andrew Mortona737b3e2006-03-22 00:08:11 -08002772 /*
2773 * Be lazy and only check for valid flags here, keeping it out of the
2774 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002776 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2777 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002779 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002781 n = cachep->node[nodeid];
2782 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783
2784 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002785 offset = n->colour_next;
2786 n->colour_next++;
2787 if (n->colour_next >= cachep->colour)
2788 n->colour_next = 0;
2789 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002791 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792
2793 if (local_flags & __GFP_WAIT)
2794 local_irq_enable();
2795
2796 /*
2797 * The test for missing atomic flag is performed here, rather than
2798 * the more obvious place, simply to reduce the critical path length
2799 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2800 * will eventually be caught here (where it matters).
2801 */
2802 kmem_flagcheck(cachep, flags);
2803
Andrew Mortona737b3e2006-03-22 00:08:11 -08002804 /*
2805 * Get mem for the objs. Attempt to allocate a physical page from
2806 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002807 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002808 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002809 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002810 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 goto failed;
2812
2813 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002814 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002815 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002816 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 goto opps1;
2818
Pekka Enberg47768742006-06-23 02:03:07 -07002819 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
Christoph Lametera35afb82007-05-16 22:10:57 -07002821 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822
2823 if (local_flags & __GFP_WAIT)
2824 local_irq_disable();
2825 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002826 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827
2828 /* Make slab active. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002829 list_add_tail(&slabp->list, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002831 n->free_objects += cachep->num;
2832 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002834opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002836failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837 if (local_flags & __GFP_WAIT)
2838 local_irq_disable();
2839 return 0;
2840}
2841
2842#if DEBUG
2843
2844/*
2845 * Perform extra freeing checks:
2846 * - detect bad pointers.
2847 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 */
2849static void kfree_debugcheck(const void *objp)
2850{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 if (!virt_addr_valid(objp)) {
2852 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002853 (unsigned long)objp);
2854 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856}
2857
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002858static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2859{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002860 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002861
2862 redzone1 = *dbg_redzone1(cache, obj);
2863 redzone2 = *dbg_redzone2(cache, obj);
2864
2865 /*
2866 * Redzone is ok.
2867 */
2868 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2869 return;
2870
2871 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2872 slab_error(cache, "double free detected");
2873 else
2874 slab_error(cache, "memory outside object was overwritten");
2875
David Woodhouseb46b8f12007-05-08 00:22:59 -07002876 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002877 obj, redzone1, redzone2);
2878}
2879
Pekka Enberg343e0d72006-02-01 03:05:50 -08002880static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002881 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882{
2883 struct page *page;
2884 unsigned int objnr;
2885 struct slab *slabp;
2886
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002887 BUG_ON(virt_to_cache(objp) != cachep);
2888
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002889 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002891 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892
Christoph Lameter35026082012-06-13 10:24:56 -05002893 slabp = page->slab_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894
2895 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002896 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2898 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2899 }
2900 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002901 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002903 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904
2905 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002906 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907
Al Viro871751e2006-03-25 03:06:39 -08002908#ifdef CONFIG_DEBUG_SLAB_LEAK
2909 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2910#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911 if (cachep->flags & SLAB_POISON) {
2912#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002913 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002914 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002915 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002916 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917 } else {
2918 poison_obj(cachep, objp, POISON_FREE);
2919 }
2920#else
2921 poison_obj(cachep, objp, POISON_FREE);
2922#endif
2923 }
2924 return objp;
2925}
2926
Pekka Enberg343e0d72006-02-01 03:05:50 -08002927static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928{
2929 kmem_bufctl_t i;
2930 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002931
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932 /* Check slab's freelist to see if this obj is there. */
2933 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2934 entries++;
2935 if (entries > cachep->num || i >= cachep->num)
2936 goto bad;
2937 }
2938 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002939bad:
2940 printk(KERN_ERR "slab: Internal list corruption detected in "
Dave Jonesface37f2011-11-15 15:03:52 -08002941 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2942 cachep->name, cachep->num, slabp, slabp->inuse,
2943 print_tainted());
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02002944 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2945 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2946 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 BUG();
2948 }
2949}
2950#else
2951#define kfree_debugcheck(x) do { } while(0)
2952#define cache_free_debugcheck(x,objp,z) (objp)
2953#define check_slabp(x,y) do { } while(0)
2954#endif
2955
Mel Gorman072bb0a2012-07-31 16:43:58 -07002956static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2957 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958{
2959 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002960 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002962 int node;
2963
Joe Korty6d2144d2008-03-05 15:04:59 -08002964 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002965 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002966 if (unlikely(force_refill))
2967 goto force_grow;
2968retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002969 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 batchcount = ac->batchcount;
2971 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002972 /*
2973 * If there was little recent activity on this cache, then
2974 * perform only a partial refill. Otherwise we could generate
2975 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 */
2977 batchcount = BATCHREFILL_LIMIT;
2978 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002979 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002981 BUG_ON(ac->avail > 0 || !n);
2982 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002983
Christoph Lameter3ded1752006-03-25 03:06:44 -08002984 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002985 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2986 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002987 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002988 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002989
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 while (batchcount > 0) {
2991 struct list_head *entry;
2992 struct slab *slabp;
2993 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002994 entry = n->slabs_partial.next;
2995 if (entry == &n->slabs_partial) {
2996 n->free_touched = 1;
2997 entry = n->slabs_free.next;
2998 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 goto must_grow;
3000 }
3001
3002 slabp = list_entry(entry, struct slab, list);
3003 check_slabp(cachep, slabp);
3004 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003005
3006 /*
3007 * The slab was either on partial or free list so
3008 * there must be at least one object available for
3009 * allocation.
3010 */
roel kluin249b9f32008-10-29 17:18:07 -04003011 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003012
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 STATS_INC_ALLOCED(cachep);
3015 STATS_INC_ACTIVE(cachep);
3016 STATS_SET_HIGH(cachep);
3017
Mel Gorman072bb0a2012-07-31 16:43:58 -07003018 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3019 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 }
3021 check_slabp(cachep, slabp);
3022
3023 /* move slabp to correct slabp list: */
3024 list_del(&slabp->list);
3025 if (slabp->free == BUFCTL_END)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003026 list_add(&slabp->list, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 else
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003028 list_add(&slabp->list, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 }
3030
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003032 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003033alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003034 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035
3036 if (unlikely(!ac->avail)) {
3037 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003038force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08003039 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003040
Andrew Mortona737b3e2006-03-22 00:08:11 -08003041 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003042 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07003043 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07003044
3045 /* no objects in sight? abort */
3046 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 return NULL;
3048
Andrew Mortona737b3e2006-03-22 00:08:11 -08003049 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 goto retry;
3051 }
3052 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003053
3054 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055}
3056
Andrew Mortona737b3e2006-03-22 00:08:11 -08003057static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3058 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059{
3060 might_sleep_if(flags & __GFP_WAIT);
3061#if DEBUG
3062 kmem_flagcheck(cachep, flags);
3063#endif
3064}
3065
3066#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003067static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003068 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003070 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003072 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003074 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003075 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003076 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 else
3078 check_poison_obj(cachep, objp);
3079#else
3080 check_poison_obj(cachep, objp);
3081#endif
3082 poison_obj(cachep, objp, POISON_INUSE);
3083 }
3084 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003085 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086
3087 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003088 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3089 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3090 slab_error(cachep, "double free, or memory outside"
3091 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003092 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003093 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003094 objp, *dbg_redzone1(cachep, objp),
3095 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 }
3097 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3098 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3099 }
Al Viro871751e2006-03-25 03:06:39 -08003100#ifdef CONFIG_DEBUG_SLAB_LEAK
3101 {
3102 struct slab *slabp;
3103 unsigned objnr;
3104
Christoph Lameter35026082012-06-13 10:24:56 -05003105 slabp = virt_to_head_page(objp)->slab_page;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003106 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
Al Viro871751e2006-03-25 03:06:39 -08003107 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3108 }
3109#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003110 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003111 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003112 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003113 if (ARCH_SLAB_MINALIGN &&
3114 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003115 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003116 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003117 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 return objp;
3119}
3120#else
3121#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3122#endif
3123
Akinobu Mita773ff602008-12-23 19:37:01 +09003124static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003125{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00003126 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003127 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003128
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003129 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003130}
3131
Pekka Enberg343e0d72006-02-01 03:05:50 -08003132static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003134 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003136 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137
Alok N Kataria5c382302005-09-27 21:45:46 -07003138 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003139
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003140 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003143 objp = ac_get_obj(cachep, ac, flags, false);
3144
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003145 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003146 * Allow for the possibility all avail objects are not allowed
3147 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003148 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003149 if (objp) {
3150 STATS_INC_ALLOCHIT(cachep);
3151 goto out;
3152 }
3153 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003154 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003155
3156 STATS_INC_ALLOCMISS(cachep);
3157 objp = cache_alloc_refill(cachep, flags, force_refill);
3158 /*
3159 * the 'ac' may be updated by cache_alloc_refill(),
3160 * and kmemleak_erase() requires its correct value.
3161 */
3162 ac = cpu_cache_get(cachep);
3163
3164out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003165 /*
3166 * To avoid a false negative, if an object that is in one of the
3167 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3168 * treat the array pointers as a reference to the object.
3169 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003170 if (objp)
3171 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003172 return objp;
3173}
3174
Christoph Lametere498be72005-09-09 13:03:32 -07003175#ifdef CONFIG_NUMA
3176/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003177 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003178 *
3179 * If we are in_interrupt, then process context, including cpusets and
3180 * mempolicy, may not apply and should not be used for allocation policy.
3181 */
3182static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3183{
3184 int nid_alloc, nid_here;
3185
Christoph Lameter765c4502006-09-27 01:50:08 -07003186 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003187 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003188 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003189 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003190 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003191 else if (current->mempolicy)
Andi Kleene7b691b2012-06-09 02:40:03 -07003192 nid_alloc = slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003193 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003194 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003195 return NULL;
3196}
3197
3198/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003199 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003200 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003201 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003202 * perform an allocation without specifying a node. This allows the page
3203 * allocator to do its reclaim / fallback magic. We then insert the
3204 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003205 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003206static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003207{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003208 struct zonelist *zonelist;
3209 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003210 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003211 struct zone *zone;
3212 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003213 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003214 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003215 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003216
3217 if (flags & __GFP_THISNODE)
3218 return NULL;
3219
Christoph Lameter6cb06222007-10-16 01:25:41 -07003220 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003221
Mel Gormancc9a6c82012-03-21 16:34:11 -07003222retry_cpuset:
3223 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07003224 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003225
Christoph Lameter3c517a62006-12-06 20:33:29 -08003226retry:
3227 /*
3228 * Look through allowed nodes for objects available
3229 * from existing per node queues.
3230 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003231 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3232 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003233
Mel Gorman54a6eb52008-04-28 02:12:16 -07003234 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter6a673682013-01-10 19:14:19 +00003235 cache->node[nid] &&
3236 cache->node[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003237 obj = ____cache_alloc_node(cache,
3238 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003239 if (obj)
3240 break;
3241 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003242 }
3243
Christoph Lametercfce6602007-05-06 14:50:17 -07003244 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003245 /*
3246 * This allocation will be performed within the constraints
3247 * of the current cpuset / memory policy requirements.
3248 * We may trigger various forms of reclaim on the allowed
3249 * set and go into memory reserves if necessary.
3250 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003251 if (local_flags & __GFP_WAIT)
3252 local_irq_enable();
3253 kmem_flagcheck(cache, flags);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003254 obj = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003255 if (local_flags & __GFP_WAIT)
3256 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003257 if (obj) {
3258 /*
3259 * Insert into the appropriate per node queues
3260 */
3261 nid = page_to_nid(virt_to_page(obj));
3262 if (cache_grow(cache, flags, nid, obj)) {
3263 obj = ____cache_alloc_node(cache,
3264 flags | GFP_THISNODE, nid);
3265 if (!obj)
3266 /*
3267 * Another processor may allocate the
3268 * objects in the slab since we are
3269 * not holding any locks.
3270 */
3271 goto retry;
3272 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003273 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003274 obj = NULL;
3275 }
3276 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003277 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003278
3279 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3280 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003281 return obj;
3282}
3283
3284/*
Christoph Lametere498be72005-09-09 13:03:32 -07003285 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003287static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003288 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003289{
3290 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 struct slab *slabp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003292 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003293 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003294 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003296 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003297 n = cachep->node[nodeid];
3298 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003299
Andrew Mortona737b3e2006-03-22 00:08:11 -08003300retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003301 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003302 spin_lock(&n->list_lock);
3303 entry = n->slabs_partial.next;
3304 if (entry == &n->slabs_partial) {
3305 n->free_touched = 1;
3306 entry = n->slabs_free.next;
3307 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003308 goto must_grow;
3309 }
Christoph Lametere498be72005-09-09 13:03:32 -07003310
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003311 slabp = list_entry(entry, struct slab, list);
3312 check_spinlock_acquired_node(cachep, nodeid);
3313 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003314
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003315 STATS_INC_NODEALLOCS(cachep);
3316 STATS_INC_ACTIVE(cachep);
3317 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003318
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003319 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003320
Matthew Dobson78d382d2006-02-01 03:05:47 -08003321 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003322 check_slabp(cachep, slabp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003323 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003324 /* move slabp to correct slabp list: */
3325 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003326
Andrew Mortona737b3e2006-03-22 00:08:11 -08003327 if (slabp->free == BUFCTL_END)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003328 list_add(&slabp->list, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003329 else
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003330 list_add(&slabp->list, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003331
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003332 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003333 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003334
Andrew Mortona737b3e2006-03-22 00:08:11 -08003335must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003336 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003337 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003338 if (x)
3339 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003340
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003341 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003342
Andrew Mortona737b3e2006-03-22 00:08:11 -08003343done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003344 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003345}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003346
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003347static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003348slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003349 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003350{
3351 unsigned long save_flags;
3352 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003353 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003354
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003355 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003356
Nick Piggincf40bd12009-01-21 08:12:39 +01003357 lockdep_trace_alloc(flags);
3358
Akinobu Mita773ff602008-12-23 19:37:01 +09003359 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003360 return NULL;
3361
Glauber Costad79923f2012-12-18 14:22:48 -08003362 cachep = memcg_kmem_get_cache(cachep, flags);
3363
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003364 cache_alloc_debugcheck_before(cachep, flags);
3365 local_irq_save(save_flags);
3366
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003367 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003368 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003369
Christoph Lameter6a673682013-01-10 19:14:19 +00003370 if (unlikely(!cachep->node[nodeid])) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003371 /* Node not bootstrapped yet */
3372 ptr = fallback_alloc(cachep, flags);
3373 goto out;
3374 }
3375
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003376 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003377 /*
3378 * Use the locally cached objects if possible.
3379 * However ____cache_alloc does not allow fallback
3380 * to other nodes. It may fail while we still have
3381 * objects on other nodes available.
3382 */
3383 ptr = ____cache_alloc(cachep, flags);
3384 if (ptr)
3385 goto out;
3386 }
3387 /* ___cache_alloc_node can fall back to other nodes */
3388 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3389 out:
3390 local_irq_restore(save_flags);
3391 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003392 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003393 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003394
Pekka Enbergc175eea2008-05-09 20:35:53 +02003395 if (likely(ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003396 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003397
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003398 if (unlikely((flags & __GFP_ZERO) && ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003399 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003400
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003401 return ptr;
3402}
3403
3404static __always_inline void *
3405__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3406{
3407 void *objp;
3408
3409 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3410 objp = alternate_node_alloc(cache, flags);
3411 if (objp)
3412 goto out;
3413 }
3414 objp = ____cache_alloc(cache, flags);
3415
3416 /*
3417 * We may just have run out of memory on the local node.
3418 * ____cache_alloc_node() knows how to locate memory on other nodes
3419 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003420 if (!objp)
3421 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003422
3423 out:
3424 return objp;
3425}
3426#else
3427
3428static __always_inline void *
3429__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3430{
3431 return ____cache_alloc(cachep, flags);
3432}
3433
3434#endif /* CONFIG_NUMA */
3435
3436static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003437slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003438{
3439 unsigned long save_flags;
3440 void *objp;
3441
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003442 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003443
Nick Piggincf40bd12009-01-21 08:12:39 +01003444 lockdep_trace_alloc(flags);
3445
Akinobu Mita773ff602008-12-23 19:37:01 +09003446 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003447 return NULL;
3448
Glauber Costad79923f2012-12-18 14:22:48 -08003449 cachep = memcg_kmem_get_cache(cachep, flags);
3450
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003451 cache_alloc_debugcheck_before(cachep, flags);
3452 local_irq_save(save_flags);
3453 objp = __do_cache_alloc(cachep, flags);
3454 local_irq_restore(save_flags);
3455 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003456 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003457 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003458 prefetchw(objp);
3459
Pekka Enbergc175eea2008-05-09 20:35:53 +02003460 if (likely(objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003461 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003462
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003463 if (unlikely((flags & __GFP_ZERO) && objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003464 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003465
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003466 return objp;
3467}
Christoph Lametere498be72005-09-09 13:03:32 -07003468
3469/*
3470 * Caller needs to acquire correct kmem_list's list_lock
3471 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003472static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003473 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474{
3475 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003476 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477
3478 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003479 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481
Mel Gorman072bb0a2012-07-31 16:43:58 -07003482 clear_obj_pfmemalloc(&objpp[i]);
3483 objp = objpp[i];
3484
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003485 slabp = virt_to_slab(objp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003486 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003488 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003490 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003492 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 check_slabp(cachep, slabp);
3494
3495 /* fixup slab chains */
3496 if (slabp->inuse == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003497 if (n->free_objects > n->free_limit) {
3498 n->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003499 /* No need to drop any previously held
3500 * lock here, even if we have a off-slab slab
3501 * descriptor it is guaranteed to come from
3502 * a different cache, refer to comments before
3503 * alloc_slabmgmt.
3504 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 slab_destroy(cachep, slabp);
3506 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003507 list_add(&slabp->list, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 }
3509 } else {
3510 /* Unconditionally move a slab to the end of the
3511 * partial list on free - maximum time for the
3512 * other objects to be freed, too.
3513 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003514 list_add_tail(&slabp->list, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 }
3516 }
3517}
3518
Pekka Enberg343e0d72006-02-01 03:05:50 -08003519static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520{
3521 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003522 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003523 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524
3525 batchcount = ac->batchcount;
3526#if DEBUG
3527 BUG_ON(!batchcount || batchcount > ac->avail);
3528#endif
3529 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003530 n = cachep->node[node];
3531 spin_lock(&n->list_lock);
3532 if (n->shared) {
3533 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003534 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 if (max) {
3536 if (batchcount > max)
3537 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003538 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003539 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 shared_array->avail += batchcount;
3541 goto free_done;
3542 }
3543 }
3544
Christoph Lameterff694162005-09-22 21:44:02 -07003545 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003546free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547#if STATS
3548 {
3549 int i = 0;
3550 struct list_head *p;
3551
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003552 p = n->slabs_free.next;
3553 while (p != &(n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554 struct slab *slabp;
3555
3556 slabp = list_entry(p, struct slab, list);
3557 BUG_ON(slabp->inuse);
3558
3559 i++;
3560 p = p->next;
3561 }
3562 STATS_SET_FREEABLE(cachep, i);
3563 }
3564#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003565 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003567 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568}
3569
3570/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003571 * Release an obj back to its cache. If the obj has a constructed state, it must
3572 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003574static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003575 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003577 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578
3579 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003580 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003581 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003583 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003584
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003585 /*
3586 * Skip calling cache_free_alien() when the platform is not numa.
3587 * This will avoid cache misses that happen while accessing slabp (which
3588 * is per page memory reference) to get nodeid. Instead use a global
3589 * variable to skip the call, which is mostly likely to be present in
3590 * the cache.
3591 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003592 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003593 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003594
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595 if (likely(ac->avail < ac->limit)) {
3596 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 } else {
3598 STATS_INC_FREEMISS(cachep);
3599 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003601
Mel Gorman072bb0a2012-07-31 16:43:58 -07003602 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603}
3604
3605/**
3606 * kmem_cache_alloc - Allocate an object
3607 * @cachep: The cache to allocate from.
3608 * @flags: See kmalloc().
3609 *
3610 * Allocate an object from this cache. The flags are only relevant
3611 * if the cache has no available objects.
3612 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003613void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003615 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003616
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003617 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003618 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003619
3620 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621}
3622EXPORT_SYMBOL(kmem_cache_alloc);
3623
Li Zefan0f24f122009-12-11 15:45:30 +08003624#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003625void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003626kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003627{
Steven Rostedt85beb582010-11-24 16:23:34 -05003628 void *ret;
3629
Ezequiel Garcia48356302012-09-08 17:47:57 -03003630 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003631
3632 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003633 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003634 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003635}
Steven Rostedt85beb582010-11-24 16:23:34 -05003636EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003637#endif
3638
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003640/**
3641 * kmem_cache_alloc_node - Allocate an object on the specified node
3642 * @cachep: The cache to allocate from.
3643 * @flags: See kmalloc().
3644 * @nodeid: node number of the target node.
3645 *
3646 * Identical to kmem_cache_alloc but it will allocate memory on the given
3647 * node, which can improve the performance for cpu bound structures.
3648 *
3649 * Fallback to other node is possible if __GFP_THISNODE is not set.
3650 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003651void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3652{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003653 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003654
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003655 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003656 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003657 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003658
3659 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003660}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003661EXPORT_SYMBOL(kmem_cache_alloc_node);
3662
Li Zefan0f24f122009-12-11 15:45:30 +08003663#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003664void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003665 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003666 int nodeid,
3667 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003668{
Steven Rostedt85beb582010-11-24 16:23:34 -05003669 void *ret;
3670
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003671 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003672
Steven Rostedt85beb582010-11-24 16:23:34 -05003673 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003674 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003675 flags, nodeid);
3676 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003677}
Steven Rostedt85beb582010-11-24 16:23:34 -05003678EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003679#endif
3680
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003681static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003682__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003683{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003684 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003685
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003686 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003687 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3688 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003689 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003690}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003691
Li Zefan0bb38a52009-12-11 15:45:50 +08003692#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003693void *__kmalloc_node(size_t size, gfp_t flags, int node)
3694{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003695 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003696}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003697EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003698
3699void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003700 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003701{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003702 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003703}
3704EXPORT_SYMBOL(__kmalloc_node_track_caller);
3705#else
3706void *__kmalloc_node(size_t size, gfp_t flags, int node)
3707{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003708 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003709}
3710EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003711#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003712#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713
3714/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003715 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003717 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003718 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003720static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003721 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003723 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003724 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003726 /* If you want to save a few bytes .text space: replace
3727 * __ with kmem_.
3728 * Then kmalloc uses the uninlined functions instead of the inline
3729 * functions.
3730 */
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003731 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003732 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3733 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003734 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003735
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003736 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003737 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003738
3739 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003740}
3741
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003742
Li Zefan0bb38a52009-12-11 15:45:50 +08003743#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003744void *__kmalloc(size_t size, gfp_t flags)
3745{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003746 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747}
3748EXPORT_SYMBOL(__kmalloc);
3749
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003750void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003751{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003752 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003753}
3754EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003755
3756#else
3757void *__kmalloc(size_t size, gfp_t flags)
3758{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003759 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003760}
3761EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003762#endif
3763
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764/**
3765 * kmem_cache_free - Deallocate an object
3766 * @cachep: The cache the allocation was from.
3767 * @objp: The previously allocated object.
3768 *
3769 * Free an object which was previously allocated from this
3770 * cache.
3771 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003772void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773{
3774 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003775 cachep = cache_from_obj(cachep, objp);
3776 if (!cachep)
3777 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778
3779 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003780 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003781 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003782 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003783 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003785
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003786 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787}
3788EXPORT_SYMBOL(kmem_cache_free);
3789
3790/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 * kfree - free previously allocated memory
3792 * @objp: pointer returned by kmalloc.
3793 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003794 * If @objp is NULL, no operation is performed.
3795 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 * Don't free memory not originally allocated by kmalloc()
3797 * or you will run into trouble.
3798 */
3799void kfree(const void *objp)
3800{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003801 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 unsigned long flags;
3803
Pekka Enberg2121db72009-03-25 11:05:57 +02003804 trace_kfree(_RET_IP_, objp);
3805
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003806 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 return;
3808 local_irq_save(flags);
3809 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003810 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003811 debug_check_no_locks_freed(objp, c->object_size);
3812
3813 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003814 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003815 local_irq_restore(flags);
3816}
3817EXPORT_SYMBOL(kfree);
3818
Christoph Lametere498be72005-09-09 13:03:32 -07003819/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003820 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003821 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003822static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003823{
3824 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003825 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003826 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003827 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003828
Mel Gorman9c09a952008-01-24 05:49:54 -08003829 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003830
Paul Menage3395ee02006-12-06 20:32:16 -08003831 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003832 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003833 if (!new_alien)
3834 goto fail;
3835 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003836
Eric Dumazet63109842007-05-06 14:49:28 -07003837 new_shared = NULL;
3838 if (cachep->shared) {
3839 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003840 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003841 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003842 if (!new_shared) {
3843 free_alien_cache(new_alien);
3844 goto fail;
3845 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003846 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003847
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003848 n = cachep->node[node];
3849 if (n) {
3850 struct array_cache *shared = n->shared;
Christoph Lametercafeb022006-03-25 03:06:46 -08003851
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003852 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003853
Christoph Lametercafeb022006-03-25 03:06:46 -08003854 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003855 free_block(cachep, shared->entry,
3856 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003857
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003858 n->shared = new_shared;
3859 if (!n->alien) {
3860 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003861 new_alien = NULL;
3862 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003863 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003864 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003865 spin_unlock_irq(&n->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003866 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003867 free_alien_cache(new_alien);
3868 continue;
3869 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003870 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3871 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003872 free_alien_cache(new_alien);
3873 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003874 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003875 }
Christoph Lametere498be72005-09-09 13:03:32 -07003876
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003877 kmem_cache_node_init(n);
3878 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003879 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003880 n->shared = new_shared;
3881 n->alien = new_alien;
3882 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003883 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003884 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003885 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003886 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003887
Andrew Mortona737b3e2006-03-22 00:08:11 -08003888fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003889 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003890 /* Cache is not active yet. Roll back what we did */
3891 node--;
3892 while (node >= 0) {
Christoph Lameter6a673682013-01-10 19:14:19 +00003893 if (cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003894 n = cachep->node[node];
Christoph Lameter0718dc22006-03-25 03:06:47 -08003895
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003896 kfree(n->shared);
3897 free_alien_cache(n->alien);
3898 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003899 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003900 }
3901 node--;
3902 }
3903 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003904 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003905}
3906
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003908 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003909 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910};
3911
3912static void do_ccupdate_local(void *info)
3913{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003914 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915 struct array_cache *old;
3916
3917 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003918 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003919
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3921 new->new[smp_processor_id()] = old;
3922}
3923
Christoph Lameter18004c52012-07-06 15:25:12 -05003924/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003925static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003926 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003927{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003928 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003929 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003931 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3932 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003933 if (!new)
3934 return -ENOMEM;
3935
Christoph Lametere498be72005-09-09 13:03:32 -07003936 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003937 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003938 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003939 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003940 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003941 kfree(new->new[i]);
3942 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003943 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 }
3945 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003946 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003948 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003949
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 cachep->batchcount = batchcount;
3952 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003953 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954
Christoph Lametere498be72005-09-09 13:03:32 -07003955 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003956 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 if (!ccold)
3958 continue;
Christoph Lameter6a673682013-01-10 19:14:19 +00003959 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003960 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
Christoph Lameter6a673682013-01-10 19:14:19 +00003961 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 kfree(ccold);
3963 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003964 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003965 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966}
3967
Glauber Costa943a4512012-12-18 14:23:03 -08003968static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3969 int batchcount, int shared, gfp_t gfp)
3970{
3971 int ret;
3972 struct kmem_cache *c = NULL;
3973 int i = 0;
3974
3975 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3976
3977 if (slab_state < FULL)
3978 return ret;
3979
3980 if ((ret < 0) || !is_root_cache(cachep))
3981 return ret;
3982
Glauber Costaebe945c2012-12-18 14:23:10 -08003983 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003984 for_each_memcg_cache_index(i) {
3985 c = cache_from_memcg(cachep, i);
3986 if (c)
3987 /* return value determined by the parent cache only */
3988 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3989 }
3990
3991 return ret;
3992}
3993
Christoph Lameter18004c52012-07-06 15:25:12 -05003994/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003995static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996{
3997 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003998 int limit = 0;
3999 int shared = 0;
4000 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001
Glauber Costa943a4512012-12-18 14:23:03 -08004002 if (!is_root_cache(cachep)) {
4003 struct kmem_cache *root = memcg_root_cache(cachep);
4004 limit = root->limit;
4005 shared = root->shared;
4006 batchcount = root->batchcount;
4007 }
4008
4009 if (limit && shared && batchcount)
4010 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004011 /*
4012 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004013 * - create a LIFO ordering, i.e. return objects that are cache-warm
4014 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004015 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 * bufctl chains: array operations are cheaper.
4017 * The numbers are guessed, we should auto-tune as described by
4018 * Bonwick.
4019 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004020 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004022 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004024 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004026 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027 limit = 54;
4028 else
4029 limit = 120;
4030
Andrew Mortona737b3e2006-03-22 00:08:11 -08004031 /*
4032 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 * allocation behaviour: Most allocs on one cpu, most free operations
4034 * on another cpu. For these cases, an efficient object passing between
4035 * cpus is necessary. This is provided by a shared array. The array
4036 * replaces Bonwick's magazine layer.
4037 * On uniprocessor, it's functionally equivalent (but less efficient)
4038 * to a larger limit. Thus disabled by default.
4039 */
4040 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004041 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043
4044#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004045 /*
4046 * With debugging enabled, large batchcount lead to excessively long
4047 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 */
4049 if (limit > 32)
4050 limit = 32;
4051#endif
Glauber Costa943a4512012-12-18 14:23:03 -08004052 batchcount = (limit + 1) / 2;
4053skip_setup:
4054 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 if (err)
4056 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004057 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004058 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059}
4060
Christoph Lameter1b552532006-03-22 00:09:07 -08004061/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004062 * Drain an array if it contains any elements taking the node lock only if
4063 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004064 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004065 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004066static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08004067 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068{
4069 int tofree;
4070
Christoph Lameter1b552532006-03-22 00:09:07 -08004071 if (!ac || !ac->avail)
4072 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073 if (ac->touched && !force) {
4074 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004075 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004076 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004077 if (ac->avail) {
4078 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4079 if (tofree > ac->avail)
4080 tofree = (ac->avail + 1) / 2;
4081 free_block(cachep, ac->entry, tofree, node);
4082 ac->avail -= tofree;
4083 memmove(ac->entry, &(ac->entry[tofree]),
4084 sizeof(void *) * ac->avail);
4085 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004086 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 }
4088}
4089
4090/**
4091 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004092 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004093 *
4094 * Called from workqueue/eventd every few seconds.
4095 * Purpose:
4096 * - clear the per-cpu caches for this CPU.
4097 * - return freeable pages to the main free memory pool.
4098 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004099 * If we cannot acquire the cache chain mutex then just give up - we'll try
4100 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004102static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004104 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004105 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004106 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004107 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108
Christoph Lameter18004c52012-07-06 15:25:12 -05004109 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004111 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112
Christoph Lameter18004c52012-07-06 15:25:12 -05004113 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 check_irq_on();
4115
Christoph Lameter35386e32006-03-22 00:09:05 -08004116 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004117 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004118 * have established with reasonable certainty that
4119 * we can do some work if the lock was obtained.
4120 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004121 n = searchp->node[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004122
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004123 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004125 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126
Christoph Lameter35386e32006-03-22 00:09:05 -08004127 /*
4128 * These are racy checks but it does not matter
4129 * if we skip one check or scan twice.
4130 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004131 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004132 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004134 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004136 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004138 if (n->free_touched)
4139 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004140 else {
4141 int freed;
4142
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004143 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004144 5 * searchp->num - 1) / (5 * searchp->num));
4145 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004147next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 cond_resched();
4149 }
4150 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004151 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004152 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004153out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004154 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004155 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156}
4157
Linus Torvalds158a9622008-01-02 13:04:48 -08004158#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004159void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004161 struct slab *slabp;
4162 unsigned long active_objs;
4163 unsigned long num_objs;
4164 unsigned long active_slabs = 0;
4165 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004166 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004168 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004169 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171 active_objs = 0;
4172 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004173 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004174 n = cachep->node[node];
4175 if (!n)
Christoph Lametere498be72005-09-09 13:03:32 -07004176 continue;
4177
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004178 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004179 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004180
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004181 list_for_each_entry(slabp, &n->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004182 if (slabp->inuse != cachep->num && !error)
4183 error = "slabs_full accounting error";
4184 active_objs += cachep->num;
4185 active_slabs++;
4186 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004187 list_for_each_entry(slabp, &n->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004188 if (slabp->inuse == cachep->num && !error)
4189 error = "slabs_partial inuse accounting error";
4190 if (!slabp->inuse && !error)
4191 error = "slabs_partial/inuse accounting error";
4192 active_objs += slabp->inuse;
4193 active_slabs++;
4194 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004195 list_for_each_entry(slabp, &n->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004196 if (slabp->inuse && !error)
4197 error = "slabs_free/inuse accounting error";
4198 num_slabs++;
4199 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004200 free_objects += n->free_objects;
4201 if (n->shared)
4202 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004203
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004204 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004206 num_slabs += active_slabs;
4207 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004208 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209 error = "free_objects accounting error";
4210
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004211 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212 if (error)
4213 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4214
Glauber Costa0d7561c2012-10-19 18:20:27 +04004215 sinfo->active_objs = active_objs;
4216 sinfo->num_objs = num_objs;
4217 sinfo->active_slabs = active_slabs;
4218 sinfo->num_slabs = num_slabs;
4219 sinfo->shared_avail = shared_avail;
4220 sinfo->limit = cachep->limit;
4221 sinfo->batchcount = cachep->batchcount;
4222 sinfo->shared = cachep->shared;
4223 sinfo->objects_per_slab = cachep->num;
4224 sinfo->cache_order = cachep->gfporder;
4225}
4226
4227void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4228{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004230 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231 unsigned long high = cachep->high_mark;
4232 unsigned long allocs = cachep->num_allocations;
4233 unsigned long grown = cachep->grown;
4234 unsigned long reaped = cachep->reaped;
4235 unsigned long errors = cachep->errors;
4236 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004237 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004238 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004239 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240
Joe Perchese92dd4f2010-03-26 19:27:58 -07004241 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4242 "%4lu %4lu %4lu %4lu %4lu",
4243 allocs, high, grown,
4244 reaped, errors, max_freeable, node_allocs,
4245 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004246 }
4247 /* cpu stats */
4248 {
4249 unsigned long allochit = atomic_read(&cachep->allochit);
4250 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4251 unsigned long freehit = atomic_read(&cachep->freehit);
4252 unsigned long freemiss = atomic_read(&cachep->freemiss);
4253
4254 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004255 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004256 }
4257#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258}
4259
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260#define MAX_SLABINFO_WRITE 128
4261/**
4262 * slabinfo_write - Tuning for the slab allocator
4263 * @file: unused
4264 * @buffer: user buffer
4265 * @count: data length
4266 * @ppos: unused
4267 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004268ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004269 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004270{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004271 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004273 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004274
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275 if (count > MAX_SLABINFO_WRITE)
4276 return -EINVAL;
4277 if (copy_from_user(&kbuf, buffer, count))
4278 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004279 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280
4281 tmp = strchr(kbuf, ' ');
4282 if (!tmp)
4283 return -EINVAL;
4284 *tmp = '\0';
4285 tmp++;
4286 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4287 return -EINVAL;
4288
4289 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004290 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004292 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004293 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004294 if (limit < 1 || batchcount < 1 ||
4295 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004296 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004297 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004298 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004299 batchcount, shared,
4300 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 }
4302 break;
4303 }
4304 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004305 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306 if (res >= 0)
4307 res = count;
4308 return res;
4309}
Al Viro871751e2006-03-25 03:06:39 -08004310
4311#ifdef CONFIG_DEBUG_SLAB_LEAK
4312
4313static void *leaks_start(struct seq_file *m, loff_t *pos)
4314{
Christoph Lameter18004c52012-07-06 15:25:12 -05004315 mutex_lock(&slab_mutex);
4316 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004317}
4318
4319static inline int add_caller(unsigned long *n, unsigned long v)
4320{
4321 unsigned long *p;
4322 int l;
4323 if (!v)
4324 return 1;
4325 l = n[1];
4326 p = n + 2;
4327 while (l) {
4328 int i = l/2;
4329 unsigned long *q = p + 2 * i;
4330 if (*q == v) {
4331 q[1]++;
4332 return 1;
4333 }
4334 if (*q > v) {
4335 l = i;
4336 } else {
4337 p = q + 2;
4338 l -= i + 1;
4339 }
4340 }
4341 if (++n[1] == n[0])
4342 return 0;
4343 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4344 p[0] = v;
4345 p[1] = 1;
4346 return 1;
4347}
4348
4349static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4350{
4351 void *p;
4352 int i;
4353 if (n[0] == n[1])
4354 return;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05004355 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
Al Viro871751e2006-03-25 03:06:39 -08004356 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4357 continue;
4358 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4359 return;
4360 }
4361}
4362
4363static void show_symbol(struct seq_file *m, unsigned long address)
4364{
4365#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004366 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004367 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004368
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004369 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004370 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004371 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004372 seq_printf(m, " [%s]", modname);
4373 return;
4374 }
4375#endif
4376 seq_printf(m, "%p", (void *)address);
4377}
4378
4379static int leaks_show(struct seq_file *m, void *p)
4380{
Thierry Reding0672aa72012-06-22 19:42:49 +02004381 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Al Viro871751e2006-03-25 03:06:39 -08004382 struct slab *slabp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004383 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004384 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004385 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004386 int node;
4387 int i;
4388
4389 if (!(cachep->flags & SLAB_STORE_USER))
4390 return 0;
4391 if (!(cachep->flags & SLAB_RED_ZONE))
4392 return 0;
4393
4394 /* OK, we can do it */
4395
Christoph Lameterdb845062013-02-05 18:45:23 +00004396 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004397
4398 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004399 n = cachep->node[node];
4400 if (!n)
Al Viro871751e2006-03-25 03:06:39 -08004401 continue;
4402
4403 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004404 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004405
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004406 list_for_each_entry(slabp, &n->slabs_full, list)
Christoph Lameterdb845062013-02-05 18:45:23 +00004407 handle_slab(x, cachep, slabp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004408 list_for_each_entry(slabp, &n->slabs_partial, list)
Christoph Lameterdb845062013-02-05 18:45:23 +00004409 handle_slab(x, cachep, slabp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004410 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004411 }
4412 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004413 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004414 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004415 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004416 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004417 if (!m->private) {
4418 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004419 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004420 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004421 return -ENOMEM;
4422 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004423 *(unsigned long *)m->private = x[0] * 2;
4424 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004425 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004426 /* Now make sure this entry will be retried */
4427 m->count = m->size;
4428 return 0;
4429 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004430 for (i = 0; i < x[1]; i++) {
4431 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4432 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004433 seq_putc(m, '\n');
4434 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004435
Al Viro871751e2006-03-25 03:06:39 -08004436 return 0;
4437}
4438
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004439static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004440 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004441 .next = slab_next,
4442 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004443 .show = leaks_show,
4444};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004445
4446static int slabstats_open(struct inode *inode, struct file *file)
4447{
4448 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4449 int ret = -ENOMEM;
4450 if (n) {
4451 ret = seq_open(file, &slabstats_op);
4452 if (!ret) {
4453 struct seq_file *m = file->private_data;
4454 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4455 m->private = n;
4456 n = NULL;
4457 }
4458 kfree(n);
4459 }
4460 return ret;
4461}
4462
4463static const struct file_operations proc_slabstats_operations = {
4464 .open = slabstats_open,
4465 .read = seq_read,
4466 .llseek = seq_lseek,
4467 .release = seq_release_private,
4468};
Al Viro871751e2006-03-25 03:06:39 -08004469#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004470
4471static int __init slab_proc_init(void)
4472{
4473#ifdef CONFIG_DEBUG_SLAB_LEAK
4474 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4475#endif
4476 return 0;
4477}
4478module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479#endif
4480
Manfred Spraul00e145b2005-09-03 15:55:07 -07004481/**
4482 * ksize - get the actual amount of memory allocated for a given object
4483 * @objp: Pointer to the object
4484 *
4485 * kmalloc may internally round up allocations and return more memory
4486 * than requested. ksize() can be used to determine the actual amount of
4487 * memory allocated. The caller may use this additional memory, even though
4488 * a smaller amount of memory was initially specified with the kmalloc call.
4489 * The caller must guarantee that objp points to a valid object previously
4490 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4491 * must not be freed during the duration of the call.
4492 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004493size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004494{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004495 BUG_ON(!objp);
4496 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004497 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004498
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004499 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004500}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004501EXPORT_SYMBOL(ksize);