blob: 19d92181ce24917f799f5f6bbd46dada145d32d8 [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
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900160#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
162
163#if FREELIST_BYTE_INDEX
164typedef unsigned char freelist_idx_t;
165#else
166typedef unsigned short freelist_idx_t;
167#endif
168
David Miller30321c72014-05-05 16:20:04 -0400169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900170
Mel Gorman072bb0a2012-07-31 16:43:58 -0700171/*
172 * true if a page was allocated from pfmemalloc reserves for network-based
173 * swap
174 */
175static bool pfmemalloc_active __read_mostly;
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 * struct array_cache
179 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 * Purpose:
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
184 *
185 * The limit is stored in the per-cpu structure to reduce the data cache
186 * footprint.
187 *
188 */
189struct array_cache {
190 unsigned int avail;
191 unsigned int limit;
192 unsigned int batchcount;
193 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700194 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700195 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800196 * Must have this definition in here for the proper
197 * alignment of array_cache. Also simplifies accessing
198 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700199 *
200 * Entries should not be directly dereferenced as
201 * entries belonging to slabs marked pfmemalloc will
202 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800203 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204};
205
Mel Gorman072bb0a2012-07-31 16:43:58 -0700206#define SLAB_OBJ_PFMEMALLOC 1
207static inline bool is_obj_pfmemalloc(void *objp)
208{
209 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
210}
211
212static inline void set_obj_pfmemalloc(void **objp)
213{
214 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
215 return;
216}
217
218static inline void clear_obj_pfmemalloc(void **objp)
219{
220 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
221}
222
Andrew Mortona737b3e2006-03-22 00:08:11 -0800223/*
224 * bootstrap: The caches do not work without cpuarrays anymore, but the
225 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 */
227#define BOOT_CPUCACHE_ENTRIES 1
228struct arraycache_init {
229 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800230 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231};
232
233/*
Christoph Lametere498be72005-09-09 13:03:32 -0700234 * Need this for bootstrapping a per node allocator.
235 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200236#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000237static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700238#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200239#define SIZE_AC MAX_NUMNODES
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000240#define SIZE_NODE (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241
Christoph Lametered11d9e2006-06-30 01:55:45 -0700242static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000243 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700244static void free_block(struct kmem_cache *cachep, void **objpp, int len,
245 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300246static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000247static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700248
Ingo Molnare0a42722006-06-23 02:03:46 -0700249static int slab_early_init = 1;
250
Christoph Lametere3366012013-01-10 19:14:18 +0000251#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000252#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700253
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000254static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700255{
256 INIT_LIST_HEAD(&parent->slabs_full);
257 INIT_LIST_HEAD(&parent->slabs_partial);
258 INIT_LIST_HEAD(&parent->slabs_free);
259 parent->shared = NULL;
260 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800261 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700262 spin_lock_init(&parent->list_lock);
263 parent->free_objects = 0;
264 parent->free_touched = 0;
265}
266
Andrew Mortona737b3e2006-03-22 00:08:11 -0800267#define MAKE_LIST(cachep, listp, slab, nodeid) \
268 do { \
269 INIT_LIST_HEAD(listp); \
Christoph Lameter6a673682013-01-10 19:14:19 +0000270 list_splice(&(cachep->node[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700271 } while (0)
272
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
274 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700275 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
276 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
277 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
278 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280#define CFLGS_OFF_SLAB (0x80000000UL)
281#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
282
283#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800284/*
285 * Optimization question: fewer reaps means less probability for unnessary
286 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100288 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 * which could lock up otherwise freeable slabs.
290 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800291#define REAPTIMEOUT_AC (2*HZ)
292#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293
294#if STATS
295#define STATS_INC_ACTIVE(x) ((x)->num_active++)
296#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
297#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
298#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700299#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800300#define STATS_SET_HIGH(x) \
301 do { \
302 if ((x)->num_active > (x)->high_mark) \
303 (x)->high_mark = (x)->num_active; \
304 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305#define STATS_INC_ERR(x) ((x)->errors++)
306#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700307#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700308#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800309#define STATS_SET_FREEABLE(x, i) \
310 do { \
311 if ((x)->max_freeable < i) \
312 (x)->max_freeable = i; \
313 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
315#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
316#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
317#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
318#else
319#define STATS_INC_ACTIVE(x) do { } while (0)
320#define STATS_DEC_ACTIVE(x) do { } while (0)
321#define STATS_INC_ALLOCED(x) do { } while (0)
322#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700323#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324#define STATS_SET_HIGH(x) do { } while (0)
325#define STATS_INC_ERR(x) do { } while (0)
326#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700327#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700328#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800329#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330#define STATS_INC_ALLOCHIT(x) do { } while (0)
331#define STATS_INC_ALLOCMISS(x) do { } while (0)
332#define STATS_INC_FREEHIT(x) do { } while (0)
333#define STATS_INC_FREEMISS(x) do { } while (0)
334#endif
335
336#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
Andrew Mortona737b3e2006-03-22 00:08:11 -0800338/*
339 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800341 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 * the end of an object is aligned with the end of the real
343 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800344 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800346 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500347 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
348 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800349 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800351static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800353 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354}
355
David Woodhouseb46b8f12007-05-08 00:22:59 -0700356static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357{
358 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700359 return (unsigned long long*) (objp + obj_offset(cachep) -
360 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361}
362
David Woodhouseb46b8f12007-05-08 00:22:59 -0700363static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
365 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
366 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500367 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700368 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400369 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500370 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700371 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372}
373
Pekka Enberg343e0d72006-02-01 03:05:50 -0800374static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375{
376 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500377 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378}
379
380#else
381
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800382#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700383#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
384#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
386
387#endif
388
389/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700390 * Do not go above this order unless 0 objects fit into the slab or
391 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 */
David Rientjes543585c2011-10-18 22:09:24 -0700393#define SLAB_MAX_ORDER_HI 1
394#define SLAB_MAX_ORDER_LO 0
395static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700396static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800398static inline struct kmem_cache *virt_to_cache(const void *obj)
399{
Christoph Lameterb49af682007-05-06 14:49:41 -0700400 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500401 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800402}
403
Joonsoo Kim8456a642013-10-24 10:07:49 +0900404static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800405 unsigned int idx)
406{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900407 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800408}
409
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800410/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500411 * We want to avoid an expensive divide : (offset / cache->size)
412 * Using the fact that size is a constant for a particular cache,
413 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800414 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
415 */
416static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900417 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800418{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900419 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800420 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800421}
422
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800424 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425
426/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000427static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800428 .batchcount = 1,
429 .limit = BOOT_CPUCACHE_ENTRIES,
430 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500431 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800432 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433};
434
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700435#define BAD_ALIEN_MAGIC 0x01020304ul
436
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200437#ifdef CONFIG_LOCKDEP
438
439/*
440 * Slab sometimes uses the kmalloc slabs to store the slab headers
441 * for other slabs "off slab".
442 * The locking for this is tricky in that it nests within the locks
443 * of all other slabs in a few places; to deal with this special
444 * locking we put on-slab caches into a separate lock-class.
445 *
446 * We set lock class for alien array caches which are up during init.
447 * The lock annotation will be lost if all cpus of a node goes down and
448 * then comes back up during hotplug
449 */
450static struct lock_class_key on_slab_l3_key;
451static struct lock_class_key on_slab_alc_key;
452
Peter Zijlstra83835b32011-07-22 15:26:05 +0200453static struct lock_class_key debugobj_l3_key;
454static struct lock_class_key debugobj_alc_key;
455
456static void slab_set_lock_classes(struct kmem_cache *cachep,
457 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
458 int q)
459{
460 struct array_cache **alc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000461 struct kmem_cache_node *n;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200462 int r;
463
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000464 n = cachep->node[q];
465 if (!n)
Peter Zijlstra83835b32011-07-22 15:26:05 +0200466 return;
467
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000468 lockdep_set_class(&n->list_lock, l3_key);
469 alc = n->alien;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200470 /*
471 * FIXME: This check for BAD_ALIEN_MAGIC
472 * should go away when common slab code is taught to
473 * work even without alien caches.
474 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
475 * for alloc_alien_cache,
476 */
477 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
478 return;
479 for_each_node(r) {
480 if (alc[r])
481 lockdep_set_class(&alc[r]->lock, alc_key);
482 }
483}
484
485static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
486{
487 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
488}
489
490static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
491{
492 int node;
493
494 for_each_online_node(node)
495 slab_set_debugobj_lock_classes_node(cachep, node);
496}
497
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200498static void init_node_lock_keys(int q)
499{
Christoph Lametere3366012013-01-10 19:14:18 +0000500 int i;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200501
Christoph Lameter97d06602012-07-06 15:25:11 -0500502 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200503 return;
504
Christoph Lameter0f8f8092013-07-02 12:12:10 -0700505 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000506 struct kmem_cache_node *n;
Christoph Lametere3366012013-01-10 19:14:18 +0000507 struct kmem_cache *cache = kmalloc_caches[i];
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200508
Christoph Lametere3366012013-01-10 19:14:18 +0000509 if (!cache)
Pekka Enberg00afa752009-12-27 14:33:14 +0200510 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200511
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000512 n = cache->node[q];
513 if (!n || OFF_SLAB(cache))
Christoph Lametere3366012013-01-10 19:14:18 +0000514 continue;
515
516 slab_set_lock_classes(cache, &on_slab_l3_key,
Peter Zijlstra83835b32011-07-22 15:26:05 +0200517 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200518 }
519}
520
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800521static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
522{
Christoph Lameter6a673682013-01-10 19:14:19 +0000523 if (!cachep->node[q])
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800524 return;
525
526 slab_set_lock_classes(cachep, &on_slab_l3_key,
527 &on_slab_alc_key, q);
528}
529
530static inline void on_slab_lock_classes(struct kmem_cache *cachep)
531{
532 int node;
533
534 VM_BUG_ON(OFF_SLAB(cachep));
535 for_each_node(node)
536 on_slab_lock_classes_node(cachep, node);
537}
538
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200539static inline void init_lock_keys(void)
540{
541 int node;
542
543 for_each_node(node)
544 init_node_lock_keys(node);
545}
546#else
547static void init_node_lock_keys(int q)
548{
549}
550
551static inline void init_lock_keys(void)
552{
553}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200554
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800555static inline void on_slab_lock_classes(struct kmem_cache *cachep)
556{
557}
558
559static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
560{
561}
562
Peter Zijlstra83835b32011-07-22 15:26:05 +0200563static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
564{
565}
566
567static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
568{
569}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200570#endif
571
Tejun Heo1871e522009-10-29 22:34:13 +0900572static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
Pekka Enberg343e0d72006-02-01 03:05:50 -0800574static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575{
576 return cachep->array[smp_processor_id()];
577}
578
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900579static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
580 size_t idx_size, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581{
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900582 int nr_objs;
583 size_t freelist_size;
584
585 /*
586 * Ignore padding for the initial guess. The padding
587 * is at most @align-1 bytes, and @buffer_size is at
588 * least @align. In the worst case, this result will
589 * be one greater than the number of objects that fit
590 * into the memory allocation when taking the padding
591 * into account.
592 */
593 nr_objs = slab_size / (buffer_size + idx_size);
594
595 /*
596 * This calculated number will be either the right
597 * amount, or one greater than what we want.
598 */
599 freelist_size = slab_size - nr_objs * buffer_size;
600 if (freelist_size < ALIGN(nr_objs * idx_size, align))
601 nr_objs--;
602
603 return nr_objs;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800604}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Andrew Mortona737b3e2006-03-22 00:08:11 -0800606/*
607 * Calculate the number of objects and left-over bytes for a given buffer size.
608 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800609static void cache_estimate(unsigned long gfporder, size_t buffer_size,
610 size_t align, int flags, size_t *left_over,
611 unsigned int *num)
612{
613 int nr_objs;
614 size_t mgmt_size;
615 size_t slab_size = PAGE_SIZE << gfporder;
616
617 /*
618 * The slab management structure can be either off the slab or
619 * on it. For the latter case, the memory allocated for a
620 * slab is used for:
621 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900622 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800623 * - Padding to respect alignment of @align
624 * - @buffer_size bytes for each object
625 *
626 * If the slab management structure is off the slab, then the
627 * alignment will already be calculated into the size. Because
628 * the slabs are all pages aligned, the objects will be at the
629 * correct alignment when allocated.
630 */
631 if (flags & CFLGS_OFF_SLAB) {
632 mgmt_size = 0;
633 nr_objs = slab_size / buffer_size;
634
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800635 } else {
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900636 nr_objs = calculate_nr_objs(slab_size, buffer_size,
Joonsoo Kima41adfa2013-12-02 17:49:42 +0900637 sizeof(freelist_idx_t), align);
638 mgmt_size = ALIGN(nr_objs * sizeof(freelist_idx_t), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800640 *num = nr_objs;
641 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642}
643
Christoph Lameterf28510d2012-09-11 19:49:38 +0000644#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700645#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
Andrew Mortona737b3e2006-03-22 00:08:11 -0800647static void __slab_error(const char *function, struct kmem_cache *cachep,
648 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649{
650 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800651 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030653 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000655#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
Paul Menage3395ee02006-12-06 20:32:16 -0800657/*
658 * By default on NUMA we use alien caches to stage the freeing of
659 * objects allocated from other nodes. This causes massive memory
660 * inefficiencies when using fake NUMA setup to split memory into a
661 * large number of small nodes, so it can be disabled on the command
662 * line
663 */
664
665static int use_alien_caches __read_mostly = 1;
666static int __init noaliencache_setup(char *s)
667{
668 use_alien_caches = 0;
669 return 1;
670}
671__setup("noaliencache", noaliencache_setup);
672
David Rientjes3df1ccc2011-10-18 22:09:28 -0700673static int __init slab_max_order_setup(char *str)
674{
675 get_option(&str, &slab_max_order);
676 slab_max_order = slab_max_order < 0 ? 0 :
677 min(slab_max_order, MAX_ORDER - 1);
678 slab_max_order_set = true;
679
680 return 1;
681}
682__setup("slab_max_order=", slab_max_order_setup);
683
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800684#ifdef CONFIG_NUMA
685/*
686 * Special reaping functions for NUMA systems called from cache_reap().
687 * These take care of doing round robin flushing of alien caches (containing
688 * objects freed on different nodes from which they were allocated) and the
689 * flushing of remote pcps by calling drain_node_pages.
690 */
Tejun Heo1871e522009-10-29 22:34:13 +0900691static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800692
693static void init_reap_node(int cpu)
694{
695 int node;
696
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700697 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800698 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800699 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800700
Tejun Heo1871e522009-10-29 22:34:13 +0900701 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800702}
703
704static void next_reap_node(void)
705{
Christoph Lameter909ea962010-12-08 16:22:55 +0100706 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800707
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800708 node = next_node(node, node_online_map);
709 if (unlikely(node >= MAX_NUMNODES))
710 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100711 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800712}
713
714#else
715#define init_reap_node(cpu) do { } while (0)
716#define next_reap_node(void) do { } while (0)
717#endif
718
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719/*
720 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
721 * via the workqueue/eventd.
722 * Add the CPU number into the expiration time to minimize the possibility of
723 * the CPUs getting into lockstep and contending for the global cache chain
724 * lock.
725 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400726static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727{
Tejun Heo1871e522009-10-29 22:34:13 +0900728 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729
730 /*
731 * When this gets called from do_initcalls via cpucache_init(),
732 * init_workqueues() has already run, so keventd will be setup
733 * at that time.
734 */
David Howells52bad642006-11-22 14:54:01 +0000735 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800736 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700737 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800738 schedule_delayed_work_on(cpu, reap_work,
739 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 }
741}
742
Christoph Lametere498be72005-09-09 13:03:32 -0700743static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300744 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800746 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 struct array_cache *nc = NULL;
748
Pekka Enberg83b519e2009-06-10 19:40:04 +0300749 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100750 /*
751 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300752 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100753 * cache the pointers are not cleared and they could be counted as
754 * valid references during a kmemleak scan. Therefore, kmemleak must
755 * not scan such objects.
756 */
757 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 if (nc) {
759 nc->avail = 0;
760 nc->limit = entries;
761 nc->batchcount = batchcount;
762 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700763 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 }
765 return nc;
766}
767
Joonsoo Kim8456a642013-10-24 10:07:49 +0900768static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700769{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700770 return PageSlabPfmemalloc(page);
771}
772
773/* Clears pfmemalloc_active if no slabs have pfmalloc set */
774static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
775 struct array_cache *ac)
776{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000777 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
Joonsoo Kim8456a642013-10-24 10:07:49 +0900778 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700779 unsigned long flags;
780
781 if (!pfmemalloc_active)
782 return;
783
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000784 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900785 list_for_each_entry(page, &n->slabs_full, lru)
786 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700787 goto out;
788
Joonsoo Kim8456a642013-10-24 10:07:49 +0900789 list_for_each_entry(page, &n->slabs_partial, lru)
790 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700791 goto out;
792
Joonsoo Kim8456a642013-10-24 10:07:49 +0900793 list_for_each_entry(page, &n->slabs_free, lru)
794 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700795 goto out;
796
797 pfmemalloc_active = false;
798out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000799 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700800}
801
Mel Gorman381760e2012-07-31 16:44:30 -0700802static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700803 gfp_t flags, bool force_refill)
804{
805 int i;
806 void *objp = ac->entry[--ac->avail];
807
808 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
809 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000810 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700811
812 if (gfp_pfmemalloc_allowed(flags)) {
813 clear_obj_pfmemalloc(&objp);
814 return objp;
815 }
816
817 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700818 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700819 /* If a !PFMEMALLOC object is found, swap them */
820 if (!is_obj_pfmemalloc(ac->entry[i])) {
821 objp = ac->entry[i];
822 ac->entry[i] = ac->entry[ac->avail];
823 ac->entry[ac->avail] = objp;
824 return objp;
825 }
826 }
827
828 /*
829 * If there are empty slabs on the slabs_free list and we are
830 * being forced to refill the cache, mark this one !pfmemalloc.
831 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000832 n = cachep->node[numa_mem_id()];
833 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900834 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900835 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700836 clear_obj_pfmemalloc(&objp);
837 recheck_pfmemalloc_active(cachep, ac);
838 return objp;
839 }
840
841 /* No !PFMEMALLOC objects available */
842 ac->avail++;
843 objp = NULL;
844 }
845
846 return objp;
847}
848
Mel Gorman381760e2012-07-31 16:44:30 -0700849static inline void *ac_get_obj(struct kmem_cache *cachep,
850 struct array_cache *ac, gfp_t flags, bool force_refill)
851{
852 void *objp;
853
854 if (unlikely(sk_memalloc_socks()))
855 objp = __ac_get_obj(cachep, ac, flags, force_refill);
856 else
857 objp = ac->entry[--ac->avail];
858
859 return objp;
860}
861
862static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700863 void *objp)
864{
865 if (unlikely(pfmemalloc_active)) {
866 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700867 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700868 if (PageSlabPfmemalloc(page))
869 set_obj_pfmemalloc(&objp);
870 }
871
Mel Gorman381760e2012-07-31 16:44:30 -0700872 return objp;
873}
874
875static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
876 void *objp)
877{
878 if (unlikely(sk_memalloc_socks()))
879 objp = __ac_put_obj(cachep, ac, objp);
880
Mel Gorman072bb0a2012-07-31 16:43:58 -0700881 ac->entry[ac->avail++] = objp;
882}
883
Christoph Lameter3ded1752006-03-25 03:06:44 -0800884/*
885 * Transfer objects in one arraycache to another.
886 * Locking must be handled by the caller.
887 *
888 * Return the number of entries transferred.
889 */
890static int transfer_objects(struct array_cache *to,
891 struct array_cache *from, unsigned int max)
892{
893 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700894 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800895
896 if (!nr)
897 return 0;
898
899 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
900 sizeof(void *) *nr);
901
902 from->avail -= nr;
903 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800904 return nr;
905}
906
Christoph Lameter765c4502006-09-27 01:50:08 -0700907#ifndef CONFIG_NUMA
908
909#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000910#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700911
Pekka Enberg83b519e2009-06-10 19:40:04 +0300912static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700913{
914 return (struct array_cache **)BAD_ALIEN_MAGIC;
915}
916
917static inline void free_alien_cache(struct array_cache **ac_ptr)
918{
919}
920
921static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
922{
923 return 0;
924}
925
926static inline void *alternate_node_alloc(struct kmem_cache *cachep,
927 gfp_t flags)
928{
929 return NULL;
930}
931
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800932static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700933 gfp_t flags, int nodeid)
934{
935 return NULL;
936}
937
938#else /* CONFIG_NUMA */
939
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800940static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800941static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800942
Pekka Enberg83b519e2009-06-10 19:40:04 +0300943static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700944{
945 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800946 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700947 int i;
948
949 if (limit > 1)
950 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800951 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700952 if (ac_ptr) {
953 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800954 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700955 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300956 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700957 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800958 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700959 kfree(ac_ptr[i]);
960 kfree(ac_ptr);
961 return NULL;
962 }
963 }
964 }
965 return ac_ptr;
966}
967
Pekka Enberg5295a742006-02-01 03:05:48 -0800968static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700969{
970 int i;
971
972 if (!ac_ptr)
973 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700974 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800975 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700976 kfree(ac_ptr);
977}
978
Pekka Enberg343e0d72006-02-01 03:05:50 -0800979static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800980 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700981{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000982 struct kmem_cache_node *n = cachep->node[node];
Christoph Lametere498be72005-09-09 13:03:32 -0700983
984 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000985 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800986 /*
987 * Stuff objects into the remote nodes shared array first.
988 * That way we could avoid the overhead of putting the objects
989 * into the free lists and getting them back later.
990 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000991 if (n->shared)
992 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800993
Christoph Lameterff694162005-09-22 21:44:02 -0700994 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700995 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000996 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700997 }
998}
999
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001000/*
1001 * Called from cache_reap() to regularly drain alien caches round robin.
1002 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001003static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001004{
Christoph Lameter909ea962010-12-08 16:22:55 +01001005 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001006
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001007 if (n->alien) {
1008 struct array_cache *ac = n->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001009
1010 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001011 __drain_alien_cache(cachep, ac, node);
1012 spin_unlock_irq(&ac->lock);
1013 }
1014 }
1015}
1016
Andrew Mortona737b3e2006-03-22 00:08:11 -08001017static void drain_alien_cache(struct kmem_cache *cachep,
1018 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001019{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001020 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001021 struct array_cache *ac;
1022 unsigned long flags;
1023
1024 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001025 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001026 if (ac) {
1027 spin_lock_irqsave(&ac->lock, flags);
1028 __drain_alien_cache(cachep, ac, i);
1029 spin_unlock_irqrestore(&ac->lock, flags);
1030 }
1031 }
1032}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001033
Ingo Molnar873623d2006-07-13 14:44:38 +02001034static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001035{
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001036 int nodeid = page_to_nid(virt_to_page(objp));
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001037 struct kmem_cache_node *n;
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001038 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001039 int node;
1040
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001041 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001042
1043 /*
1044 * Make sure we are not freeing a object from another node to the array
1045 * cache on this cpu.
1046 */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001047 if (likely(nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001048 return 0;
1049
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001050 n = cachep->node[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001051 STATS_INC_NODEFREES(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001052 if (n->alien && n->alien[nodeid]) {
1053 alien = n->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001054 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001055 if (unlikely(alien->avail == alien->limit)) {
1056 STATS_INC_ACOVERFLOW(cachep);
1057 __drain_alien_cache(cachep, alien, nodeid);
1058 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001059 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001060 spin_unlock(&alien->lock);
1061 } else {
Christoph Lameter6a673682013-01-10 19:14:19 +00001062 spin_lock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001063 free_block(cachep, &objp, 1, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001064 spin_unlock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001065 }
1066 return 1;
1067}
Christoph Lametere498be72005-09-09 13:03:32 -07001068#endif
1069
David Rientjes8f9f8d92010-03-27 19:40:47 -07001070/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001071 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001072 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001073 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001074 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001075 * already in use.
1076 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001077 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001078 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001079static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001080{
1081 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001082 struct kmem_cache_node *n;
Christoph Lameter6744f082013-01-10 19:12:17 +00001083 const int memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001084
Christoph Lameter18004c52012-07-06 15:25:12 -05001085 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001086 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001087 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001088 * begin anything. Make sure some other cpu on this
1089 * node has not already allocated this
1090 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001091 if (!cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001092 n = kmalloc_node(memsize, GFP_KERNEL, node);
1093 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001094 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001095 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001096 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1097 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001098
1099 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001100 * The kmem_cache_nodes don't come and go as CPUs
1101 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001102 * protection here.
1103 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001104 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001105 }
1106
Christoph Lameter6a673682013-01-10 19:14:19 +00001107 spin_lock_irq(&cachep->node[node]->list_lock);
1108 cachep->node[node]->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001109 (1 + nr_cpus_node(node)) *
1110 cachep->batchcount + cachep->num;
Christoph Lameter6a673682013-01-10 19:14:19 +00001111 spin_unlock_irq(&cachep->node[node]->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001112 }
1113 return 0;
1114}
1115
Wanpeng Li0fa81032013-07-04 08:33:22 +08001116static inline int slabs_tofree(struct kmem_cache *cachep,
1117 struct kmem_cache_node *n)
1118{
1119 return (n->free_objects + cachep->num - 1) / cachep->num;
1120}
1121
Paul Gortmaker0db06282013-06-19 14:53:51 -04001122static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001124 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001126 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301127 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001128
Christoph Lameter18004c52012-07-06 15:25:12 -05001129 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001130 struct array_cache *nc;
1131 struct array_cache *shared;
1132 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001133
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001134 /* cpu is dead; no one can alloc from it. */
1135 nc = cachep->array[cpu];
1136 cachep->array[cpu] = NULL;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001137 n = cachep->node[node];
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001138
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001139 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001140 goto free_array_cache;
1141
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001142 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001143
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001144 /* Free limit for this kmem_cache_node */
1145 n->free_limit -= cachep->batchcount;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001146 if (nc)
1147 free_block(cachep, nc->entry, nc->avail, node);
1148
Rusty Russell58463c12009-12-17 11:43:12 -06001149 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001150 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001151 goto free_array_cache;
1152 }
1153
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001154 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001155 if (shared) {
1156 free_block(cachep, shared->entry,
1157 shared->avail, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001158 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001159 }
1160
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001161 alien = n->alien;
1162 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001163
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001164 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001165
1166 kfree(shared);
1167 if (alien) {
1168 drain_alien_cache(cachep, alien);
1169 free_alien_cache(alien);
1170 }
1171free_array_cache:
1172 kfree(nc);
1173 }
1174 /*
1175 * In the previous loop, all the objects were freed to
1176 * the respective cache's slabs, now we can go ahead and
1177 * shrink each nodelist to its limit.
1178 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001179 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001180 n = cachep->node[node];
1181 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001182 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001183 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001184 }
1185}
1186
Paul Gortmaker0db06282013-06-19 14:53:51 -04001187static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001188{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001189 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001190 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001191 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001192 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001194 /*
1195 * We need to do this right in the beginning since
1196 * alloc_arraycache's are going to use this list.
1197 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001198 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001200 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001201 if (err < 0)
1202 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001203
1204 /*
1205 * Now we can go ahead with allocating the shared arrays and
1206 * array caches
1207 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001208 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001209 struct array_cache *nc;
1210 struct array_cache *shared = NULL;
1211 struct array_cache **alien = NULL;
1212
1213 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001214 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001215 if (!nc)
1216 goto bad;
1217 if (cachep->shared) {
1218 shared = alloc_arraycache(node,
1219 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001220 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001221 if (!shared) {
1222 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001223 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001224 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001225 }
1226 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001227 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001228 if (!alien) {
1229 kfree(shared);
1230 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001231 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001232 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001233 }
1234 cachep->array[cpu] = nc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001235 n = cachep->node[node];
1236 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001237
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001238 spin_lock_irq(&n->list_lock);
1239 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001240 /*
1241 * We are serialised from CPU_DEAD or
1242 * CPU_UP_CANCELLED by the cpucontrol lock
1243 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001244 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001245 shared = NULL;
1246 }
1247#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001248 if (!n->alien) {
1249 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001250 alien = NULL;
1251 }
1252#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001253 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001254 kfree(shared);
1255 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001256 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1257 slab_set_debugobj_lock_classes_node(cachep, node);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08001258 else if (!OFF_SLAB(cachep) &&
1259 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1260 on_slab_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001261 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001262 init_node_lock_keys(node);
1263
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001264 return 0;
1265bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001266 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001267 return -ENOMEM;
1268}
1269
Paul Gortmaker0db06282013-06-19 14:53:51 -04001270static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001271 unsigned long action, void *hcpu)
1272{
1273 long cpu = (long)hcpu;
1274 int err = 0;
1275
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001277 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001278 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001279 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001281 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 break;
1283 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001284 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 start_cpu_timer(cpu);
1286 break;
1287#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001288 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001289 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001290 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001291 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001292 * held so that if cache_reap() is invoked it cannot do
1293 * anything expensive but will only modify reap_work
1294 * and reschedule the timer.
1295 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001296 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001297 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001298 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001299 break;
1300 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001301 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001302 start_cpu_timer(cpu);
1303 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001305 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001306 /*
1307 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001308 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001309 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001310 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001311 * structure is usually allocated from kmem_cache_create() and
1312 * gets destroyed at kmem_cache_destroy().
1313 */
Simon Arlott183ff222007-10-20 01:27:18 +02001314 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001315#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001317 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001318 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001319 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001320 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001323 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324}
1325
Paul Gortmaker0db06282013-06-19 14:53:51 -04001326static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001327 &cpuup_callback, NULL, 0
1328};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329
David Rientjes8f9f8d92010-03-27 19:40:47 -07001330#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1331/*
1332 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1333 * Returns -EBUSY if all objects cannot be drained so that the node is not
1334 * removed.
1335 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001336 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001337 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001338static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001339{
1340 struct kmem_cache *cachep;
1341 int ret = 0;
1342
Christoph Lameter18004c52012-07-06 15:25:12 -05001343 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001344 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001345
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001346 n = cachep->node[node];
1347 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001348 continue;
1349
Wanpeng Li0fa81032013-07-04 08:33:22 +08001350 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001351
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001352 if (!list_empty(&n->slabs_full) ||
1353 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001354 ret = -EBUSY;
1355 break;
1356 }
1357 }
1358 return ret;
1359}
1360
1361static int __meminit slab_memory_callback(struct notifier_block *self,
1362 unsigned long action, void *arg)
1363{
1364 struct memory_notify *mnb = arg;
1365 int ret = 0;
1366 int nid;
1367
1368 nid = mnb->status_change_nid;
1369 if (nid < 0)
1370 goto out;
1371
1372 switch (action) {
1373 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001374 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001375 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001376 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001377 break;
1378 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001379 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001380 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001381 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001382 break;
1383 case MEM_ONLINE:
1384 case MEM_OFFLINE:
1385 case MEM_CANCEL_ONLINE:
1386 case MEM_CANCEL_OFFLINE:
1387 break;
1388 }
1389out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001390 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001391}
1392#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1393
Christoph Lametere498be72005-09-09 13:03:32 -07001394/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001395 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001396 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001397static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001398 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001399{
Christoph Lameter6744f082013-01-10 19:12:17 +00001400 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001401
Christoph Lameter6744f082013-01-10 19:12:17 +00001402 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001403 BUG_ON(!ptr);
1404
Christoph Lameter6744f082013-01-10 19:12:17 +00001405 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001406 /*
1407 * Do not assume that spinlocks can be initialized via memcpy:
1408 */
1409 spin_lock_init(&ptr->list_lock);
1410
Christoph Lametere498be72005-09-09 13:03:32 -07001411 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001412 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001413}
1414
Andrew Mortona737b3e2006-03-22 00:08:11 -08001415/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001416 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1417 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001418 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001419static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001420{
1421 int node;
1422
1423 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001424 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001425 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001426 REAPTIMEOUT_NODE +
1427 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001428 }
1429}
1430
1431/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001432 * The memory after the last cpu cache pointer is used for the
Christoph Lameter6a673682013-01-10 19:14:19 +00001433 * the node pointer.
Christoph Lameter3c583462012-11-28 16:23:01 +00001434 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001435static void setup_node_pointer(struct kmem_cache *cachep)
Christoph Lameter3c583462012-11-28 16:23:01 +00001436{
Christoph Lameter6a673682013-01-10 19:14:19 +00001437 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
Christoph Lameter3c583462012-11-28 16:23:01 +00001438}
1439
1440/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001441 * Initialisation. Called after the page allocator have been initialised and
1442 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 */
1444void __init kmem_cache_init(void)
1445{
Christoph Lametere498be72005-09-09 13:03:32 -07001446 int i;
1447
Joonsoo Kim68126702013-10-24 10:07:42 +09001448 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1449 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001450 kmem_cache = &kmem_cache_boot;
Christoph Lameter6a673682013-01-10 19:14:19 +00001451 setup_node_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001452
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001453 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001454 use_alien_caches = 0;
1455
Christoph Lameter3c583462012-11-28 16:23:01 +00001456 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001457 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001458
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001459 set_up_node(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460
1461 /*
1462 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001463 * page orders on machines with more than 32MB of memory if
1464 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001466 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001467 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 /* Bootstrap is tricky, because several objects are allocated
1470 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001471 * 1) initialize the kmem_cache cache: it contains the struct
1472 * kmem_cache structures of all caches, except kmem_cache itself:
1473 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001474 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001475 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001476 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001478 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001479 * An __init data area is used for the head array.
1480 * 3) Create the remaining kmalloc caches, with minimally sized
1481 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001482 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001484 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001485 * the other cache's with kmalloc allocated memory.
1486 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 */
1488
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001489 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490
Eric Dumazet8da34302007-05-06 14:49:29 -07001491 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001492 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001493 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001494 create_boot_cache(kmem_cache, "kmem_cache",
1495 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001496 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001497 SLAB_HWCACHE_ALIGN);
1498 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
1500 /* 2+3) create the kmalloc caches */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501
Andrew Mortona737b3e2006-03-22 00:08:11 -08001502 /*
1503 * Initialize the caches that provide memory for the array cache and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001504 * kmem_cache_node structures first. Without this, further allocations will
Andrew Mortona737b3e2006-03-22 00:08:11 -08001505 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001506 */
1507
Christoph Lametere3366012013-01-10 19:14:18 +00001508 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1509 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001510
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001511 if (INDEX_AC != INDEX_NODE)
1512 kmalloc_caches[INDEX_NODE] =
1513 create_kmalloc_cache("kmalloc-node",
1514 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001515
Ingo Molnare0a42722006-06-23 02:03:46 -07001516 slab_early_init = 0;
1517
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 /* 4) Replace the bootstrap head arrays */
1519 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001520 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001521
Pekka Enberg83b519e2009-06-10 19:40:04 +03001522 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001523
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001524 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001525 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001526 /*
1527 * Do not assume that spinlocks can be initialized via memcpy:
1528 */
1529 spin_lock_init(&ptr->lock);
1530
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001531 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001532
Pekka Enberg83b519e2009-06-10 19:40:04 +03001533 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001534
Christoph Lametere3366012013-01-10 19:14:18 +00001535 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001536 != &initarray_generic.cache);
Christoph Lametere3366012013-01-10 19:14:18 +00001537 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001538 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001539 /*
1540 * Do not assume that spinlocks can be initialized via memcpy:
1541 */
1542 spin_lock_init(&ptr->lock);
1543
Christoph Lametere3366012013-01-10 19:14:18 +00001544 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001546 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001547 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001548 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549
Mel Gorman9c09a952008-01-24 05:49:54 -08001550 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001551 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001552
Christoph Lametere3366012013-01-10 19:14:18 +00001553 init_list(kmalloc_caches[INDEX_AC],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001554 &init_kmem_cache_node[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001555
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001556 if (INDEX_AC != INDEX_NODE) {
1557 init_list(kmalloc_caches[INDEX_NODE],
1558 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001559 }
1560 }
1561 }
1562
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001563 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001564}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001565
Pekka Enberg8429db52009-06-12 15:58:59 +03001566void __init kmem_cache_init_late(void)
1567{
1568 struct kmem_cache *cachep;
1569
Christoph Lameter97d06602012-07-06 15:25:11 -05001570 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001571
Pekka Enberg8429db52009-06-12 15:58:59 +03001572 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001573 mutex_lock(&slab_mutex);
1574 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001575 if (enable_cpucache(cachep, GFP_NOWAIT))
1576 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001577 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001578
Michael Wang947ca182012-09-05 10:33:18 +08001579 /* Annotate slab for lockdep -- annotate the malloc caches */
1580 init_lock_keys();
1581
Christoph Lameter97d06602012-07-06 15:25:11 -05001582 /* Done! */
1583 slab_state = FULL;
1584
Andrew Mortona737b3e2006-03-22 00:08:11 -08001585 /*
1586 * Register a cpu startup notifier callback that initializes
1587 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 */
1589 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590
David Rientjes8f9f8d92010-03-27 19:40:47 -07001591#ifdef CONFIG_NUMA
1592 /*
1593 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001594 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001595 */
1596 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1597#endif
1598
Andrew Mortona737b3e2006-03-22 00:08:11 -08001599 /*
1600 * The reap timers are started later, with a module init call: That part
1601 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 */
1603}
1604
1605static int __init cpucache_init(void)
1606{
1607 int cpu;
1608
Andrew Mortona737b3e2006-03-22 00:08:11 -08001609 /*
1610 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 */
Christoph Lametere498be72005-09-09 13:03:32 -07001612 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001613 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001614
1615 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001616 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 return 0;
1618}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619__initcall(cpucache_init);
1620
Rafael Aquini8bdec192012-03-09 17:27:27 -03001621static noinline void
1622slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1623{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001624 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001625 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001626 unsigned long flags;
1627 int node;
1628
1629 printk(KERN_WARNING
1630 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1631 nodeid, gfpflags);
1632 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001633 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001634
1635 for_each_online_node(node) {
1636 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1637 unsigned long active_slabs = 0, num_slabs = 0;
1638
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001639 n = cachep->node[node];
1640 if (!n)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001641 continue;
1642
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001643 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001644 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001645 active_objs += cachep->num;
1646 active_slabs++;
1647 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001648 list_for_each_entry(page, &n->slabs_partial, lru) {
1649 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001650 active_slabs++;
1651 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001652 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001653 num_slabs++;
1654
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001655 free_objects += n->free_objects;
1656 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001657
1658 num_slabs += active_slabs;
1659 num_objs = num_slabs * cachep->num;
1660 printk(KERN_WARNING
1661 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1662 node, active_slabs, num_slabs, active_objs, num_objs,
1663 free_objects);
1664 }
1665}
1666
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667/*
1668 * Interface to system's page allocator. No need to hold the cache-lock.
1669 *
1670 * If we requested dmaable memory, we will get it. Even if we
1671 * did not request dmaable memory, we might get it, but that
1672 * would be relatively rare and ignorable.
1673 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001674static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1675 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676{
1677 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001678 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001679
Glauber Costaa618e892012-06-14 16:17:21 +04001680 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001681 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1682 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001683
Linus Torvalds517d0862009-06-16 19:50:13 -07001684 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001685 if (!page) {
1686 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1687 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001689 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001691 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001692 if (unlikely(page->pfmemalloc))
1693 pfmemalloc_active = true;
1694
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001695 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001697 add_zone_page_state(page_zone(page),
1698 NR_SLAB_RECLAIMABLE, nr_pages);
1699 else
1700 add_zone_page_state(page_zone(page),
1701 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001702 __SetPageSlab(page);
1703 if (page->pfmemalloc)
1704 SetPageSlabPfmemalloc(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001705 memcg_bind_pages(cachep, cachep->gfporder);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001706
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001707 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1708 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1709
1710 if (cachep->ctor)
1711 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1712 else
1713 kmemcheck_mark_unallocated_pages(page, nr_pages);
1714 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001715
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001716 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717}
1718
1719/*
1720 * Interface to system's page release.
1721 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001722static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001724 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001726 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001727
Christoph Lameter972d1a72006-09-25 23:31:51 -07001728 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1729 sub_zone_page_state(page_zone(page),
1730 NR_SLAB_RECLAIMABLE, nr_freed);
1731 else
1732 sub_zone_page_state(page_zone(page),
1733 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001734
Joonsoo Kima57a4982013-10-24 10:07:44 +09001735 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001736 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001737 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001738 page_mapcount_reset(page);
1739 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001740
1741 memcg_release_pages(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 if (current->reclaim_state)
1743 current->reclaim_state->reclaimed_slab += nr_freed;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001744 __free_memcg_kmem_pages(page, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745}
1746
1747static void kmem_rcu_free(struct rcu_head *head)
1748{
Joonsoo Kim68126702013-10-24 10:07:42 +09001749 struct kmem_cache *cachep;
1750 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
Joonsoo Kim68126702013-10-24 10:07:42 +09001752 page = container_of(head, struct page, rcu_head);
1753 cachep = page->slab_cache;
1754
1755 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756}
1757
1758#if DEBUG
1759
1760#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001761static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001762 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001764 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001766 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001768 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 return;
1770
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 *addr++ = 0x12345678;
1772 *addr++ = caller;
1773 *addr++ = smp_processor_id();
1774 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 {
1776 unsigned long *sptr = &caller;
1777 unsigned long svalue;
1778
1779 while (!kstack_end(sptr)) {
1780 svalue = *sptr++;
1781 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001782 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 size -= sizeof(unsigned long);
1784 if (size <= sizeof(unsigned long))
1785 break;
1786 }
1787 }
1788
1789 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001790 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791}
1792#endif
1793
Pekka Enberg343e0d72006-02-01 03:05:50 -08001794static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001796 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001797 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
1799 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801}
1802
1803static void dump_line(char *data, int offset, int limit)
1804{
1805 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001806 unsigned char error = 0;
1807 int bad_count = 0;
1808
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001809 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001810 for (i = 0; i < limit; i++) {
1811 if (data[offset + i] != POISON_FREE) {
1812 error = data[offset + i];
1813 bad_count++;
1814 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001815 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001816 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1817 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001818
1819 if (bad_count == 1) {
1820 error ^= POISON_FREE;
1821 if (!(error & (error - 1))) {
1822 printk(KERN_ERR "Single bit error detected. Probably "
1823 "bad RAM.\n");
1824#ifdef CONFIG_X86
1825 printk(KERN_ERR "Run memtest86+ or a similar memory "
1826 "test tool.\n");
1827#else
1828 printk(KERN_ERR "Run a memory test tool.\n");
1829#endif
1830 }
1831 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832}
1833#endif
1834
1835#if DEBUG
1836
Pekka Enberg343e0d72006-02-01 03:05:50 -08001837static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838{
1839 int i, size;
1840 char *realobj;
1841
1842 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001843 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001844 *dbg_redzone1(cachep, objp),
1845 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 }
1847
1848 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001849 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1850 *dbg_userword(cachep, objp),
1851 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001853 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001854 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001855 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 int limit;
1857 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 if (i + limit > size)
1859 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 dump_line(realobj, i, limit);
1861 }
1862}
1863
Pekka Enberg343e0d72006-02-01 03:05:50 -08001864static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865{
1866 char *realobj;
1867 int size, i;
1868 int lines = 0;
1869
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001870 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001871 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001873 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 exp = POISON_END;
1877 if (realobj[i] != exp) {
1878 int limit;
1879 /* Mismatch ! */
1880 /* Print header */
1881 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001882 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001883 "Slab corruption (%s): %s start=%p, len=%d\n",
1884 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 print_objinfo(cachep, objp, 0);
1886 }
1887 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001888 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001890 if (i + limit > size)
1891 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 dump_line(realobj, i, limit);
1893 i += 16;
1894 lines++;
1895 /* Limit to 5 lines */
1896 if (lines > 5)
1897 break;
1898 }
1899 }
1900 if (lines != 0) {
1901 /* Print some data about the neighboring objects, if they
1902 * exist:
1903 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001904 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001905 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906
Joonsoo Kim8456a642013-10-24 10:07:49 +09001907 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001909 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001910 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001912 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 print_objinfo(cachep, objp, 2);
1914 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001915 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001916 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001917 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001919 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 print_objinfo(cachep, objp, 2);
1921 }
1922 }
1923}
1924#endif
1925
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001927static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1928 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001929{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 int i;
1931 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001932 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933
1934 if (cachep->flags & SLAB_POISON) {
1935#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001936 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001937 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001938 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001939 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 else
1941 check_poison_obj(cachep, objp);
1942#else
1943 check_poison_obj(cachep, objp);
1944#endif
1945 }
1946 if (cachep->flags & SLAB_RED_ZONE) {
1947 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1948 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001949 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1951 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001952 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001955}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001957static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1958 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001959{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001960}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961#endif
1962
Randy Dunlap911851e2006-03-22 00:08:14 -08001963/**
1964 * slab_destroy - destroy and release all objects in a slab
1965 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001966 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001967 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001968 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001969 * Before calling the slab must have been unlinked from the cache. The
1970 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001971 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001972static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001973{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001974 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001975
Joonsoo Kim8456a642013-10-24 10:07:49 +09001976 freelist = page->freelist;
1977 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001979 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980
Joonsoo Kim68126702013-10-24 10:07:42 +09001981 /*
1982 * RCU free overloads the RCU head over the LRU.
1983 * slab_page has been overloeaded over the LRU,
1984 * however it is not used from now on so that
1985 * we can use it safely.
1986 */
1987 head = (void *)&page->rcu_head;
1988 call_rcu(head, kmem_rcu_free);
1989
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001991 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 }
Joonsoo Kim68126702013-10-24 10:07:42 +09001993
1994 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001995 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001996 * although actual page can be freed in rcu context
1997 */
1998 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001999 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000}
2001
2002/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002003 * calculate_slab_order - calculate size (page order) of slabs
2004 * @cachep: pointer to the cache that is being created
2005 * @size: size of objects to be created in this cache.
2006 * @align: required alignment for the objects.
2007 * @flags: slab allocation flags
2008 *
2009 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010 *
2011 * This could be made much more intelligent. For now, try to avoid using
2012 * high order pages for slabs. When the gfp() functions are more friendly
2013 * towards high-order requests, this should be changed.
2014 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002015static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002016 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002017{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002018 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002019 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002020 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002021
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002022 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002023 unsigned int num;
2024 size_t remainder;
2025
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002026 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002027 if (!num)
2028 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002029
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002030 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2031 if (num > SLAB_OBJ_MAX_NUM)
2032 break;
2033
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002034 if (flags & CFLGS_OFF_SLAB) {
2035 /*
2036 * Max number of objs-per-slab for caches which
2037 * use off-slab slabs. Needed to avoid a possible
2038 * looping condition in cache_grow().
2039 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002040 offslab_limit = size;
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002041 offslab_limit /= sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002042
2043 if (num > offslab_limit)
2044 break;
2045 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002046
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002047 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002048 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002049 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002050 left_over = remainder;
2051
2052 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002053 * A VFS-reclaimable slab tends to have most allocations
2054 * as GFP_NOFS and we really don't want to have to be allocating
2055 * higher-order pages when we are unable to shrink dcache.
2056 */
2057 if (flags & SLAB_RECLAIM_ACCOUNT)
2058 break;
2059
2060 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002061 * Large number of objects is good, but very large slabs are
2062 * currently bad for the gfp()s.
2063 */
David Rientjes543585c2011-10-18 22:09:24 -07002064 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002065 break;
2066
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002067 /*
2068 * Acceptable internal fragmentation?
2069 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002070 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002071 break;
2072 }
2073 return left_over;
2074}
2075
Pekka Enberg83b519e2009-06-10 19:40:04 +03002076static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002077{
Christoph Lameter97d06602012-07-06 15:25:11 -05002078 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002079 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002080
Christoph Lameter97d06602012-07-06 15:25:11 -05002081 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002082 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002083 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002084 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002085 * of by the caller of __kmem_cache_create
2086 */
2087 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2088 slab_state = PARTIAL;
2089 } else if (slab_state == PARTIAL) {
2090 /*
2091 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002092 * that's used by kmalloc(24), otherwise the creation of
2093 * further caches will BUG().
2094 */
2095 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2096
2097 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002098 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2099 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002100 * otherwise the creation of further caches will BUG().
2101 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002102 set_up_node(cachep, SIZE_AC);
2103 if (INDEX_AC == INDEX_NODE)
2104 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002105 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002106 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002107 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002108 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002109 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002110 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002111
Christoph Lameter97d06602012-07-06 15:25:11 -05002112 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002113 set_up_node(cachep, SIZE_NODE);
2114 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002115 } else {
2116 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002117 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002118 cachep->node[node] =
Christoph Lameter6744f082013-01-10 19:12:17 +00002119 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002120 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002121 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002122 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002123 }
2124 }
2125 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002126 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002127 jiffies + REAPTIMEOUT_NODE +
2128 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002129
2130 cpu_cache_get(cachep)->avail = 0;
2131 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2132 cpu_cache_get(cachep)->batchcount = 1;
2133 cpu_cache_get(cachep)->touched = 0;
2134 cachep->batchcount = 1;
2135 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002136 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002137}
2138
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002139/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002140 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002141 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 *
2144 * Returns a ptr to the cache on success, NULL on failure.
2145 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002146 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 * The flags are
2149 *
2150 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2151 * to catch references to uninitialised memory.
2152 *
2153 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2154 * for buffer overruns.
2155 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2157 * cacheline. This can be beneficial if you're counting cycles as closely
2158 * as davem.
2159 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002160int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002161__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002163 size_t left_over, freelist_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002164 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002165 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002166 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169#if FORCED_DEBUG
2170 /*
2171 * Enable redzoning and last user accounting, except for caches with
2172 * large objects, if the increased size would increase the object size
2173 * above the next power of two: caches with object sizes just above a
2174 * power of two have a significant amount of internal fragmentation.
2175 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002176 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2177 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002178 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 if (!(flags & SLAB_DESTROY_BY_RCU))
2180 flags |= SLAB_POISON;
2181#endif
2182 if (flags & SLAB_DESTROY_BY_RCU)
2183 BUG_ON(flags & SLAB_POISON);
2184#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
Andrew Mortona737b3e2006-03-22 00:08:11 -08002186 /*
2187 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 * unaligned accesses for some archs when redzoning is used, and makes
2189 * sure any on-slab bufctl's are also correctly aligned.
2190 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002191 if (size & (BYTES_PER_WORD - 1)) {
2192 size += (BYTES_PER_WORD - 1);
2193 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 }
2195
Pekka Enbergca5f9702006-09-25 23:31:25 -07002196 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002197 * Redzoning and user store require word alignment or possibly larger.
2198 * Note this will be overridden by architecture or caller mandated
2199 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002200 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002201 if (flags & SLAB_STORE_USER)
2202 ralign = BYTES_PER_WORD;
2203
2204 if (flags & SLAB_RED_ZONE) {
2205 ralign = REDZONE_ALIGN;
2206 /* If redzoning, ensure that the second redzone is suitably
2207 * aligned, by adjusting the object size accordingly. */
2208 size += REDZONE_ALIGN - 1;
2209 size &= ~(REDZONE_ALIGN - 1);
2210 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002211
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002212 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002213 if (ralign < cachep->align) {
2214 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002216 /* disable debug if necessary */
2217 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002218 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002219 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002220 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002222 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223
Pekka Enberg83b519e2009-06-10 19:40:04 +03002224 if (slab_is_available())
2225 gfp = GFP_KERNEL;
2226 else
2227 gfp = GFP_NOWAIT;
2228
Christoph Lameter6a673682013-01-10 19:14:19 +00002229 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231
Pekka Enbergca5f9702006-09-25 23:31:25 -07002232 /*
2233 * Both debugging options require word-alignment which is calculated
2234 * into align above.
2235 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002238 cachep->obj_offset += sizeof(unsigned long long);
2239 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 }
2241 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002242 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002243 * the real object. But if the second red zone needs to be
2244 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002246 if (flags & SLAB_RED_ZONE)
2247 size += REDZONE_ALIGN;
2248 else
2249 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 }
2251#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002252 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002253 && cachep->object_size > cache_line_size()
2254 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2255 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 size = PAGE_SIZE;
2257 }
2258#endif
2259#endif
2260
Ingo Molnare0a42722006-06-23 02:03:46 -07002261 /*
2262 * Determine if the slab management is 'on' or 'off' slab.
2263 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002264 * it too early on. Always use on-slab management when
2265 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002266 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002267 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002268 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 /*
2270 * Size is large, assume best to place the slab management obj
2271 * off-slab (should allow better packing of objs).
2272 */
2273 flags |= CFLGS_OFF_SLAB;
2274
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002275 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002276 /*
2277 * We should restrict the number of objects in a slab to implement
2278 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2279 */
2280 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2281 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002283 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002285 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002286 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002287
Joonsoo Kim8456a642013-10-24 10:07:49 +09002288 freelist_size =
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002289 ALIGN(cachep->num * sizeof(freelist_idx_t), cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290
2291 /*
2292 * If the slab has been placed off-slab, and we have enough space then
2293 * move it on-slab. This is at the expense of any extra colouring.
2294 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002295 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002297 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 }
2299
2300 if (flags & CFLGS_OFF_SLAB) {
2301 /* really off slab. No need for manual alignment */
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002302 freelist_size = cachep->num * sizeof(freelist_idx_t);
Ron Lee67461362009-05-22 04:58:22 +09302303
2304#ifdef CONFIG_PAGE_POISONING
2305 /* If we're going to use the generic kernel_map_pages()
2306 * poisoning, then it's going to smash the contents of
2307 * the redzone and userword anyhow, so switch them off.
2308 */
2309 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2310 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2311#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 }
2313
2314 cachep->colour_off = cache_line_size();
2315 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002316 if (cachep->colour_off < cachep->align)
2317 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002318 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002319 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002321 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002322 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002323 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002324 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002325 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002327 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002328 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002329 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002330 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002331 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002332 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2333 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002334 * But leave a BUG_ON for some lucky dude.
2335 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002336 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002337 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002339 err = setup_cpu_cache(cachep, gfp);
2340 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002341 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002342 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002343 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344
Peter Zijlstra83835b32011-07-22 15:26:05 +02002345 if (flags & SLAB_DEBUG_OBJECTS) {
2346 /*
2347 * Would deadlock through slab_destroy()->call_rcu()->
2348 * debug_object_activate()->kmem_cache_alloc().
2349 */
2350 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2351
2352 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002353 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2354 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002355
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002356 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
2359#if DEBUG
2360static void check_irq_off(void)
2361{
2362 BUG_ON(!irqs_disabled());
2363}
2364
2365static void check_irq_on(void)
2366{
2367 BUG_ON(irqs_disabled());
2368}
2369
Pekka Enberg343e0d72006-02-01 03:05:50 -08002370static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371{
2372#ifdef CONFIG_SMP
2373 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002374 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375#endif
2376}
Christoph Lametere498be72005-09-09 13:03:32 -07002377
Pekka Enberg343e0d72006-02-01 03:05:50 -08002378static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002379{
2380#ifdef CONFIG_SMP
2381 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002382 assert_spin_locked(&cachep->node[node]->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002383#endif
2384}
2385
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386#else
2387#define check_irq_off() do { } while(0)
2388#define check_irq_on() do { } while(0)
2389#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002390#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391#endif
2392
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002393static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002394 struct array_cache *ac,
2395 int force, int node);
2396
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397static void do_drain(void *arg)
2398{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002399 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002401 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402
2403 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002404 ac = cpu_cache_get(cachep);
Christoph Lameter6a673682013-01-10 19:14:19 +00002405 spin_lock(&cachep->node[node]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002406 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002407 spin_unlock(&cachep->node[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 ac->avail = 0;
2409}
2410
Pekka Enberg343e0d72006-02-01 03:05:50 -08002411static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002413 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002414 int node;
2415
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002416 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002418 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002419 n = cachep->node[node];
2420 if (n && n->alien)
2421 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002422 }
2423
2424 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002425 n = cachep->node[node];
2426 if (n)
2427 drain_array(cachep, n, n->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002428 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429}
2430
Christoph Lametered11d9e2006-06-30 01:55:45 -07002431/*
2432 * Remove slabs from the list of free slabs.
2433 * Specify the number of slabs to drain in tofree.
2434 *
2435 * Returns the actual number of slabs released.
2436 */
2437static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002438 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002440 struct list_head *p;
2441 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002442 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443
Christoph Lametered11d9e2006-06-30 01:55:45 -07002444 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002445 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002447 spin_lock_irq(&n->list_lock);
2448 p = n->slabs_free.prev;
2449 if (p == &n->slabs_free) {
2450 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002451 goto out;
2452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453
Joonsoo Kim8456a642013-10-24 10:07:49 +09002454 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002456 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002458 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002459 /*
2460 * Safe to drop the lock. The slab is no longer linked
2461 * to the cache.
2462 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002463 n->free_objects -= cache->num;
2464 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002465 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002466 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002468out:
2469 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470}
2471
Christoph Lameter18004c52012-07-06 15:25:12 -05002472/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002473static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002474{
2475 int ret = 0, i = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002476 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002477
2478 drain_cpu_caches(cachep);
2479
2480 check_irq_on();
2481 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002482 n = cachep->node[i];
2483 if (!n)
Christoph Lametered11d9e2006-06-30 01:55:45 -07002484 continue;
2485
Wanpeng Li0fa81032013-07-04 08:33:22 +08002486 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002487
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002488 ret += !list_empty(&n->slabs_full) ||
2489 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002490 }
2491 return (ret ? 1 : 0);
2492}
2493
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494/**
2495 * kmem_cache_shrink - Shrink a cache.
2496 * @cachep: The cache to shrink.
2497 *
2498 * Releases as many slabs as possible for a cache.
2499 * To help debugging, a zero exit status indicates all slabs were released.
2500 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002501int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002503 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002504 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002506 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002507 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002508 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002509 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002510 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002511 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512}
2513EXPORT_SYMBOL(kmem_cache_shrink);
2514
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002515int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516{
Christoph Lameter12c36672012-09-04 23:38:33 +00002517 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002518 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002519 int rc = __cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
Christoph Lameter12c36672012-09-04 23:38:33 +00002521 if (rc)
2522 return rc;
2523
2524 for_each_online_cpu(i)
2525 kfree(cachep->array[i]);
2526
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002527 /* NUMA: free the node structures */
Christoph Lameter12c36672012-09-04 23:38:33 +00002528 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002529 n = cachep->node[i];
2530 if (n) {
2531 kfree(n->shared);
2532 free_alien_cache(n->alien);
2533 kfree(n);
Christoph Lameter12c36672012-09-04 23:38:33 +00002534 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002536 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002539/*
2540 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002541 *
2542 * For a slab cache when the slab descriptor is off-slab, the
2543 * slab descriptor can't come from the same cache which is being created,
2544 * Because if it is the case, that means we defer the creation of
2545 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2546 * And we eventually call down to __kmem_cache_create(), which
2547 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2548 * This is a "chicken-and-egg" problem.
2549 *
2550 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2551 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002552 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002553static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002554 struct page *page, int colour_off,
2555 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002557 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002558 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002559
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 if (OFF_SLAB(cachep)) {
2561 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002562 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002563 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002564 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 return NULL;
2566 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002567 freelist = addr + colour_off;
2568 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002570 page->active = 0;
2571 page->s_mem = addr + colour_off;
2572 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573}
2574
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002575static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002577 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002578}
2579
2580static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002581 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002582{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002583 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584}
2585
Pekka Enberg343e0d72006-02-01 03:05:50 -08002586static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002587 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588{
2589 int i;
2590
2591 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002592 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593#if DEBUG
2594 /* need to poison the objs? */
2595 if (cachep->flags & SLAB_POISON)
2596 poison_obj(cachep, objp, POISON_FREE);
2597 if (cachep->flags & SLAB_STORE_USER)
2598 *dbg_userword(cachep, objp) = NULL;
2599
2600 if (cachep->flags & SLAB_RED_ZONE) {
2601 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2602 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2603 }
2604 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002605 * Constructors are not allowed to allocate memory from the same
2606 * cache which they are a constructor for. Otherwise, deadlock.
2607 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 */
2609 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002610 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611
2612 if (cachep->flags & SLAB_RED_ZONE) {
2613 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2614 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002615 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2617 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002618 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002620 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002621 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002622 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002623 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624#else
2625 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002626 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627#endif
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002628 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630}
2631
Pekka Enberg343e0d72006-02-01 03:05:50 -08002632static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002634 if (CONFIG_ZONE_DMA_FLAG) {
2635 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002636 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002637 else
Glauber Costaa618e892012-06-14 16:17:21 +04002638 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002639 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640}
2641
Joonsoo Kim8456a642013-10-24 10:07:49 +09002642static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002643 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002644{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002645 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002646
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002647 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002648 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002649#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002650 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002651#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002652
2653 return objp;
2654}
2655
Joonsoo Kim8456a642013-10-24 10:07:49 +09002656static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002657 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002658{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002659 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002660#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002661 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002662
Matthew Dobson78d382d2006-02-01 03:05:47 -08002663 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002664 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002665
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002666 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002667 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002668 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002669 printk(KERN_ERR "slab: double free detected in cache "
2670 "'%s', objp %p\n", cachep->name, objp);
2671 BUG();
2672 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002673 }
2674#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002675 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002676 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002677}
2678
Pekka Enberg47768742006-06-23 02:03:07 -07002679/*
2680 * Map pages beginning at addr to the given cache and slab. This is required
2681 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002682 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002683 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002684static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002685 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002687 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002688 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689}
2690
2691/*
2692 * Grow (by 1) the number of slabs within a cache. This is called by
2693 * kmem_cache_alloc() when there are no active objs left in a cache.
2694 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002695static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002696 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002698 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002699 size_t offset;
2700 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002701 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702
Andrew Mortona737b3e2006-03-22 00:08:11 -08002703 /*
2704 * Be lazy and only check for valid flags here, keeping it out of the
2705 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002707 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2708 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002710 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002712 n = cachep->node[nodeid];
2713 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714
2715 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002716 offset = n->colour_next;
2717 n->colour_next++;
2718 if (n->colour_next >= cachep->colour)
2719 n->colour_next = 0;
2720 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002722 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
2724 if (local_flags & __GFP_WAIT)
2725 local_irq_enable();
2726
2727 /*
2728 * The test for missing atomic flag is performed here, rather than
2729 * the more obvious place, simply to reduce the critical path length
2730 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2731 * will eventually be caught here (where it matters).
2732 */
2733 kmem_flagcheck(cachep, flags);
2734
Andrew Mortona737b3e2006-03-22 00:08:11 -08002735 /*
2736 * Get mem for the objs. Attempt to allocate a physical page from
2737 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002738 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002739 if (!page)
2740 page = kmem_getpages(cachep, local_flags, nodeid);
2741 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 goto failed;
2743
2744 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002745 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002746 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002747 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 goto opps1;
2749
Joonsoo Kim8456a642013-10-24 10:07:49 +09002750 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751
Joonsoo Kim8456a642013-10-24 10:07:49 +09002752 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753
2754 if (local_flags & __GFP_WAIT)
2755 local_irq_disable();
2756 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002757 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758
2759 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002760 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002762 n->free_objects += cachep->num;
2763 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002765opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002766 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002767failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 if (local_flags & __GFP_WAIT)
2769 local_irq_disable();
2770 return 0;
2771}
2772
2773#if DEBUG
2774
2775/*
2776 * Perform extra freeing checks:
2777 * - detect bad pointers.
2778 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 */
2780static void kfree_debugcheck(const void *objp)
2781{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 if (!virt_addr_valid(objp)) {
2783 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002784 (unsigned long)objp);
2785 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787}
2788
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002789static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2790{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002791 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002792
2793 redzone1 = *dbg_redzone1(cache, obj);
2794 redzone2 = *dbg_redzone2(cache, obj);
2795
2796 /*
2797 * Redzone is ok.
2798 */
2799 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2800 return;
2801
2802 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2803 slab_error(cache, "double free detected");
2804 else
2805 slab_error(cache, "memory outside object was overwritten");
2806
David Woodhouseb46b8f12007-05-08 00:22:59 -07002807 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002808 obj, redzone1, redzone2);
2809}
2810
Pekka Enberg343e0d72006-02-01 03:05:50 -08002811static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002812 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002815 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002817 BUG_ON(virt_to_cache(objp) != cachep);
2818
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002819 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002821 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002824 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2826 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2827 }
2828 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002829 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830
Joonsoo Kim8456a642013-10-24 10:07:49 +09002831 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832
2833 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002834 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 if (cachep->flags & SLAB_POISON) {
2837#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002838 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002839 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002840 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002841 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 } else {
2843 poison_obj(cachep, objp, POISON_FREE);
2844 }
2845#else
2846 poison_obj(cachep, objp, POISON_FREE);
2847#endif
2848 }
2849 return objp;
2850}
2851
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852#else
2853#define kfree_debugcheck(x) do { } while(0)
2854#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855#endif
2856
Mel Gorman072bb0a2012-07-31 16:43:58 -07002857static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2858 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859{
2860 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002861 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002863 int node;
2864
Joe Korty6d2144d2008-03-05 15:04:59 -08002865 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002866 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002867 if (unlikely(force_refill))
2868 goto force_grow;
2869retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002870 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 batchcount = ac->batchcount;
2872 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002873 /*
2874 * If there was little recent activity on this cache, then
2875 * perform only a partial refill. Otherwise we could generate
2876 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 */
2878 batchcount = BATCHREFILL_LIMIT;
2879 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002880 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002882 BUG_ON(ac->avail > 0 || !n);
2883 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002884
Christoph Lameter3ded1752006-03-25 03:06:44 -08002885 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002886 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2887 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002888 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002889 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002890
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 while (batchcount > 0) {
2892 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002893 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002895 entry = n->slabs_partial.next;
2896 if (entry == &n->slabs_partial) {
2897 n->free_touched = 1;
2898 entry = n->slabs_free.next;
2899 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 goto must_grow;
2901 }
2902
Joonsoo Kim8456a642013-10-24 10:07:49 +09002903 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002905
2906 /*
2907 * The slab was either on partial or free list so
2908 * there must be at least one object available for
2909 * allocation.
2910 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002911 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002912
Joonsoo Kim8456a642013-10-24 10:07:49 +09002913 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 STATS_INC_ALLOCED(cachep);
2915 STATS_INC_ACTIVE(cachep);
2916 STATS_SET_HIGH(cachep);
2917
Joonsoo Kim8456a642013-10-24 10:07:49 +09002918 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002919 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921
2922 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002923 list_del(&page->lru);
2924 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002925 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002927 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 }
2929
Andrew Mortona737b3e2006-03-22 00:08:11 -08002930must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002931 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002932alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002933 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934
2935 if (unlikely(!ac->avail)) {
2936 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002937force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002938 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002939
Andrew Mortona737b3e2006-03-22 00:08:11 -08002940 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002941 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002942 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002943
2944 /* no objects in sight? abort */
2945 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 return NULL;
2947
Andrew Mortona737b3e2006-03-22 00:08:11 -08002948 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949 goto retry;
2950 }
2951 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002952
2953 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954}
2955
Andrew Mortona737b3e2006-03-22 00:08:11 -08002956static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2957 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958{
2959 might_sleep_if(flags & __GFP_WAIT);
2960#if DEBUG
2961 kmem_flagcheck(cachep, flags);
2962#endif
2963}
2964
2965#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002966static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002967 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002969 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002971 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002973 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002974 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002975 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 else
2977 check_poison_obj(cachep, objp);
2978#else
2979 check_poison_obj(cachep, objp);
2980#endif
2981 poison_obj(cachep, objp, POISON_INUSE);
2982 }
2983 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002984 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985
2986 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002987 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2988 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2989 slab_error(cachep, "double free, or memory outside"
2990 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002991 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002992 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002993 objp, *dbg_redzone1(cachep, objp),
2994 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 }
2996 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2997 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2998 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002999 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003000 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003001 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003002 if (ARCH_SLAB_MINALIGN &&
3003 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003004 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003005 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003006 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 return objp;
3008}
3009#else
3010#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3011#endif
3012
Akinobu Mita773ff602008-12-23 19:37:01 +09003013static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003014{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00003015 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003016 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003017
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003018 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003019}
3020
Pekka Enberg343e0d72006-02-01 03:05:50 -08003021static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003023 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003025 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
Alok N Kataria5c382302005-09-27 21:45:46 -07003027 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003028
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003029 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003032 objp = ac_get_obj(cachep, ac, flags, false);
3033
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003034 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003035 * Allow for the possibility all avail objects are not allowed
3036 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003037 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003038 if (objp) {
3039 STATS_INC_ALLOCHIT(cachep);
3040 goto out;
3041 }
3042 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003044
3045 STATS_INC_ALLOCMISS(cachep);
3046 objp = cache_alloc_refill(cachep, flags, force_refill);
3047 /*
3048 * the 'ac' may be updated by cache_alloc_refill(),
3049 * and kmemleak_erase() requires its correct value.
3050 */
3051 ac = cpu_cache_get(cachep);
3052
3053out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003054 /*
3055 * To avoid a false negative, if an object that is in one of the
3056 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3057 * treat the array pointers as a reference to the object.
3058 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003059 if (objp)
3060 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003061 return objp;
3062}
3063
Christoph Lametere498be72005-09-09 13:03:32 -07003064#ifdef CONFIG_NUMA
3065/*
David Rientjesf0432d12014-04-07 15:37:30 -07003066 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003067 *
3068 * If we are in_interrupt, then process context, including cpusets and
3069 * mempolicy, may not apply and should not be used for allocation policy.
3070 */
3071static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3072{
3073 int nid_alloc, nid_here;
3074
Christoph Lameter765c4502006-09-27 01:50:08 -07003075 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003076 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003077 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003078 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003079 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003080 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003081 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003082 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003083 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003084 return NULL;
3085}
3086
3087/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003088 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003089 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003090 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003091 * perform an allocation without specifying a node. This allows the page
3092 * allocator to do its reclaim / fallback magic. We then insert the
3093 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003094 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003095static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003096{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003097 struct zonelist *zonelist;
3098 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003099 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003100 struct zone *zone;
3101 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003102 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003103 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003104 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003105
3106 if (flags & __GFP_THISNODE)
3107 return NULL;
3108
Christoph Lameter6cb06222007-10-16 01:25:41 -07003109 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003110
Mel Gormancc9a6c82012-03-21 16:34:11 -07003111retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003112 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003113 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003114
Christoph Lameter3c517a62006-12-06 20:33:29 -08003115retry:
3116 /*
3117 * Look through allowed nodes for objects available
3118 * from existing per node queues.
3119 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003120 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3121 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003122
Mel Gorman54a6eb52008-04-28 02:12:16 -07003123 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter6a673682013-01-10 19:14:19 +00003124 cache->node[nid] &&
3125 cache->node[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003126 obj = ____cache_alloc_node(cache,
3127 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003128 if (obj)
3129 break;
3130 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003131 }
3132
Christoph Lametercfce6602007-05-06 14:50:17 -07003133 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003134 /*
3135 * This allocation will be performed within the constraints
3136 * of the current cpuset / memory policy requirements.
3137 * We may trigger various forms of reclaim on the allowed
3138 * set and go into memory reserves if necessary.
3139 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003140 struct page *page;
3141
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003142 if (local_flags & __GFP_WAIT)
3143 local_irq_enable();
3144 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003145 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003146 if (local_flags & __GFP_WAIT)
3147 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003148 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003149 /*
3150 * Insert into the appropriate per node queues
3151 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003152 nid = page_to_nid(page);
3153 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003154 obj = ____cache_alloc_node(cache,
3155 flags | GFP_THISNODE, nid);
3156 if (!obj)
3157 /*
3158 * Another processor may allocate the
3159 * objects in the slab since we are
3160 * not holding any locks.
3161 */
3162 goto retry;
3163 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003164 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003165 obj = NULL;
3166 }
3167 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003168 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003169
Mel Gormand26914d2014-04-03 14:47:24 -07003170 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003171 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003172 return obj;
3173}
3174
3175/*
Christoph Lametere498be72005-09-09 13:03:32 -07003176 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003178static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003179 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003180{
3181 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003182 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003183 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003184 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003185 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003187 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003188 n = cachep->node[nodeid];
3189 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003190
Andrew Mortona737b3e2006-03-22 00:08:11 -08003191retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003192 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003193 spin_lock(&n->list_lock);
3194 entry = n->slabs_partial.next;
3195 if (entry == &n->slabs_partial) {
3196 n->free_touched = 1;
3197 entry = n->slabs_free.next;
3198 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003199 goto must_grow;
3200 }
Christoph Lametere498be72005-09-09 13:03:32 -07003201
Joonsoo Kim8456a642013-10-24 10:07:49 +09003202 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003203 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003204
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003205 STATS_INC_NODEALLOCS(cachep);
3206 STATS_INC_ACTIVE(cachep);
3207 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003208
Joonsoo Kim8456a642013-10-24 10:07:49 +09003209 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003210
Joonsoo Kim8456a642013-10-24 10:07:49 +09003211 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003212 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003213 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003214 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003215
Joonsoo Kim8456a642013-10-24 10:07:49 +09003216 if (page->active == cachep->num)
3217 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003218 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003219 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003220
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003221 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003222 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003223
Andrew Mortona737b3e2006-03-22 00:08:11 -08003224must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003225 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003226 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003227 if (x)
3228 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003229
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003230 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003231
Andrew Mortona737b3e2006-03-22 00:08:11 -08003232done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003233 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003234}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003235
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003236static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003237slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003238 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003239{
3240 unsigned long save_flags;
3241 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003242 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003243
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003244 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003245
Nick Piggincf40bd12009-01-21 08:12:39 +01003246 lockdep_trace_alloc(flags);
3247
Akinobu Mita773ff602008-12-23 19:37:01 +09003248 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003249 return NULL;
3250
Glauber Costad79923f2012-12-18 14:22:48 -08003251 cachep = memcg_kmem_get_cache(cachep, flags);
3252
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003253 cache_alloc_debugcheck_before(cachep, flags);
3254 local_irq_save(save_flags);
3255
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003256 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003257 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003258
Christoph Lameter6a673682013-01-10 19:14:19 +00003259 if (unlikely(!cachep->node[nodeid])) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003260 /* Node not bootstrapped yet */
3261 ptr = fallback_alloc(cachep, flags);
3262 goto out;
3263 }
3264
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003265 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003266 /*
3267 * Use the locally cached objects if possible.
3268 * However ____cache_alloc does not allow fallback
3269 * to other nodes. It may fail while we still have
3270 * objects on other nodes available.
3271 */
3272 ptr = ____cache_alloc(cachep, flags);
3273 if (ptr)
3274 goto out;
3275 }
3276 /* ___cache_alloc_node can fall back to other nodes */
3277 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3278 out:
3279 local_irq_restore(save_flags);
3280 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003281 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003282 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003283
Joe Perches5087c822013-09-10 17:02:51 -07003284 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003285 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003286 if (unlikely(flags & __GFP_ZERO))
3287 memset(ptr, 0, cachep->object_size);
3288 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003289
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003290 return ptr;
3291}
3292
3293static __always_inline void *
3294__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3295{
3296 void *objp;
3297
David Rientjesf0432d12014-04-07 15:37:30 -07003298 if (current->mempolicy || unlikely(current->flags & PF_SPREAD_SLAB)) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003299 objp = alternate_node_alloc(cache, flags);
3300 if (objp)
3301 goto out;
3302 }
3303 objp = ____cache_alloc(cache, flags);
3304
3305 /*
3306 * We may just have run out of memory on the local node.
3307 * ____cache_alloc_node() knows how to locate memory on other nodes
3308 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003309 if (!objp)
3310 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003311
3312 out:
3313 return objp;
3314}
3315#else
3316
3317static __always_inline void *
3318__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3319{
3320 return ____cache_alloc(cachep, flags);
3321}
3322
3323#endif /* CONFIG_NUMA */
3324
3325static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003326slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003327{
3328 unsigned long save_flags;
3329 void *objp;
3330
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003331 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003332
Nick Piggincf40bd12009-01-21 08:12:39 +01003333 lockdep_trace_alloc(flags);
3334
Akinobu Mita773ff602008-12-23 19:37:01 +09003335 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003336 return NULL;
3337
Glauber Costad79923f2012-12-18 14:22:48 -08003338 cachep = memcg_kmem_get_cache(cachep, flags);
3339
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003340 cache_alloc_debugcheck_before(cachep, flags);
3341 local_irq_save(save_flags);
3342 objp = __do_cache_alloc(cachep, flags);
3343 local_irq_restore(save_flags);
3344 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003345 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003346 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003347 prefetchw(objp);
3348
Joe Perches5087c822013-09-10 17:02:51 -07003349 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003350 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003351 if (unlikely(flags & __GFP_ZERO))
3352 memset(objp, 0, cachep->object_size);
3353 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003354
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003355 return objp;
3356}
Christoph Lametere498be72005-09-09 13:03:32 -07003357
3358/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003359 * Caller needs to acquire correct kmem_cache_node's list_lock
Christoph Lametere498be72005-09-09 13:03:32 -07003360 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003361static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003362 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363{
3364 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003365 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366
3367 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003368 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003369 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
Mel Gorman072bb0a2012-07-31 16:43:58 -07003371 clear_obj_pfmemalloc(&objpp[i]);
3372 objp = objpp[i];
3373
Joonsoo Kim8456a642013-10-24 10:07:49 +09003374 page = virt_to_head_page(objp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003375 n = cachep->node[node];
Joonsoo Kim8456a642013-10-24 10:07:49 +09003376 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003377 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003378 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003380 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381
3382 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003383 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003384 if (n->free_objects > n->free_limit) {
3385 n->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003386 /* No need to drop any previously held
3387 * lock here, even if we have a off-slab slab
3388 * descriptor it is guaranteed to come from
3389 * a different cache, refer to comments before
3390 * alloc_slabmgmt.
3391 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003392 slab_destroy(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003394 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395 }
3396 } else {
3397 /* Unconditionally move a slab to the end of the
3398 * partial list on free - maximum time for the
3399 * other objects to be freed, too.
3400 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003401 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 }
3403 }
3404}
3405
Pekka Enberg343e0d72006-02-01 03:05:50 -08003406static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407{
3408 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003409 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003410 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411
3412 batchcount = ac->batchcount;
3413#if DEBUG
3414 BUG_ON(!batchcount || batchcount > ac->avail);
3415#endif
3416 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003417 n = cachep->node[node];
3418 spin_lock(&n->list_lock);
3419 if (n->shared) {
3420 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003421 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 if (max) {
3423 if (batchcount > max)
3424 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003425 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003426 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 shared_array->avail += batchcount;
3428 goto free_done;
3429 }
3430 }
3431
Christoph Lameterff694162005-09-22 21:44:02 -07003432 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003433free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434#if STATS
3435 {
3436 int i = 0;
3437 struct list_head *p;
3438
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003439 p = n->slabs_free.next;
3440 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003441 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442
Joonsoo Kim8456a642013-10-24 10:07:49 +09003443 page = list_entry(p, struct page, lru);
3444 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445
3446 i++;
3447 p = p->next;
3448 }
3449 STATS_SET_FREEABLE(cachep, i);
3450 }
3451#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003452 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003454 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455}
3456
3457/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003458 * Release an obj back to its cache. If the obj has a constructed state, it must
3459 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003461static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003462 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003464 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
3466 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003467 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003468 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003470 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003471
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003472 /*
3473 * Skip calling cache_free_alien() when the platform is not numa.
3474 * This will avoid cache misses that happen while accessing slabp (which
3475 * is per page memory reference) to get nodeid. Instead use a global
3476 * variable to skip the call, which is mostly likely to be present in
3477 * the cache.
3478 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003479 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003480 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003481
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 if (likely(ac->avail < ac->limit)) {
3483 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 } else {
3485 STATS_INC_FREEMISS(cachep);
3486 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003488
Mel Gorman072bb0a2012-07-31 16:43:58 -07003489 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490}
3491
3492/**
3493 * kmem_cache_alloc - Allocate an object
3494 * @cachep: The cache to allocate from.
3495 * @flags: See kmalloc().
3496 *
3497 * Allocate an object from this cache. The flags are only relevant
3498 * if the cache has no available objects.
3499 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003500void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003502 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003503
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003504 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003505 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003506
3507 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508}
3509EXPORT_SYMBOL(kmem_cache_alloc);
3510
Li Zefan0f24f122009-12-11 15:45:30 +08003511#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003512void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003513kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003514{
Steven Rostedt85beb582010-11-24 16:23:34 -05003515 void *ret;
3516
Ezequiel Garcia48356302012-09-08 17:47:57 -03003517 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003518
3519 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003520 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003521 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003522}
Steven Rostedt85beb582010-11-24 16:23:34 -05003523EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003524#endif
3525
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003527/**
3528 * kmem_cache_alloc_node - Allocate an object on the specified node
3529 * @cachep: The cache to allocate from.
3530 * @flags: See kmalloc().
3531 * @nodeid: node number of the target node.
3532 *
3533 * Identical to kmem_cache_alloc but it will allocate memory on the given
3534 * node, which can improve the performance for cpu bound structures.
3535 *
3536 * Fallback to other node is possible if __GFP_THISNODE is not set.
3537 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003538void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3539{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003540 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003541
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003542 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003543 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003544 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003545
3546 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003547}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548EXPORT_SYMBOL(kmem_cache_alloc_node);
3549
Li Zefan0f24f122009-12-11 15:45:30 +08003550#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003551void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003552 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003553 int nodeid,
3554 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003555{
Steven Rostedt85beb582010-11-24 16:23:34 -05003556 void *ret;
3557
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003558 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003559
Steven Rostedt85beb582010-11-24 16:23:34 -05003560 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003561 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003562 flags, nodeid);
3563 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003564}
Steven Rostedt85beb582010-11-24 16:23:34 -05003565EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003566#endif
3567
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003568static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003569__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003570{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003571 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003572
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003573 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003574 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3575 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003576 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003577}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003578
Li Zefan0bb38a52009-12-11 15:45:50 +08003579#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003580void *__kmalloc_node(size_t size, gfp_t flags, int node)
3581{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003582 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003583}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003584EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003585
3586void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003587 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003588{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003589 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003590}
3591EXPORT_SYMBOL(__kmalloc_node_track_caller);
3592#else
3593void *__kmalloc_node(size_t size, gfp_t flags, int node)
3594{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003595 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003596}
3597EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003598#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003599#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
3601/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003602 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003604 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003605 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003607static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003608 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003610 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003611 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003613 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003614 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3615 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003616 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003617
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003618 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003619 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003620
3621 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003622}
3623
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003624
Li Zefan0bb38a52009-12-11 15:45:50 +08003625#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003626void *__kmalloc(size_t size, gfp_t flags)
3627{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003628 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629}
3630EXPORT_SYMBOL(__kmalloc);
3631
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003632void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003633{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003634 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003635}
3636EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003637
3638#else
3639void *__kmalloc(size_t size, gfp_t flags)
3640{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003641 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003642}
3643EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003644#endif
3645
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646/**
3647 * kmem_cache_free - Deallocate an object
3648 * @cachep: The cache the allocation was from.
3649 * @objp: The previously allocated object.
3650 *
3651 * Free an object which was previously allocated from this
3652 * cache.
3653 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003654void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655{
3656 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003657 cachep = cache_from_obj(cachep, objp);
3658 if (!cachep)
3659 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660
3661 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003662 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003663 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003664 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003665 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003667
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003668 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669}
3670EXPORT_SYMBOL(kmem_cache_free);
3671
3672/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 * kfree - free previously allocated memory
3674 * @objp: pointer returned by kmalloc.
3675 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003676 * If @objp is NULL, no operation is performed.
3677 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 * Don't free memory not originally allocated by kmalloc()
3679 * or you will run into trouble.
3680 */
3681void kfree(const void *objp)
3682{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003683 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684 unsigned long flags;
3685
Pekka Enberg2121db72009-03-25 11:05:57 +02003686 trace_kfree(_RET_IP_, objp);
3687
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003688 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689 return;
3690 local_irq_save(flags);
3691 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003692 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003693 debug_check_no_locks_freed(objp, c->object_size);
3694
3695 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003696 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697 local_irq_restore(flags);
3698}
3699EXPORT_SYMBOL(kfree);
3700
Christoph Lametere498be72005-09-09 13:03:32 -07003701/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003702 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003703 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003704static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003705{
3706 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003707 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003708 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003709 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003710
Mel Gorman9c09a952008-01-24 05:49:54 -08003711 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003712
Paul Menage3395ee02006-12-06 20:32:16 -08003713 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003714 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003715 if (!new_alien)
3716 goto fail;
3717 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003718
Eric Dumazet63109842007-05-06 14:49:28 -07003719 new_shared = NULL;
3720 if (cachep->shared) {
3721 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003722 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003723 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003724 if (!new_shared) {
3725 free_alien_cache(new_alien);
3726 goto fail;
3727 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003728 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003729
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003730 n = cachep->node[node];
3731 if (n) {
3732 struct array_cache *shared = n->shared;
Christoph Lametercafeb022006-03-25 03:06:46 -08003733
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003734 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003735
Christoph Lametercafeb022006-03-25 03:06:46 -08003736 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003737 free_block(cachep, shared->entry,
3738 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003739
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003740 n->shared = new_shared;
3741 if (!n->alien) {
3742 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003743 new_alien = NULL;
3744 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003745 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003746 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003747 spin_unlock_irq(&n->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003748 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003749 free_alien_cache(new_alien);
3750 continue;
3751 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003752 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3753 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003754 free_alien_cache(new_alien);
3755 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003756 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003757 }
Christoph Lametere498be72005-09-09 13:03:32 -07003758
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003759 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003760 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3761 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003762 n->shared = new_shared;
3763 n->alien = new_alien;
3764 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003765 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003766 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003767 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003768 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003769
Andrew Mortona737b3e2006-03-22 00:08:11 -08003770fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003771 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003772 /* Cache is not active yet. Roll back what we did */
3773 node--;
3774 while (node >= 0) {
Christoph Lameter6a673682013-01-10 19:14:19 +00003775 if (cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003776 n = cachep->node[node];
Christoph Lameter0718dc22006-03-25 03:06:47 -08003777
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003778 kfree(n->shared);
3779 free_alien_cache(n->alien);
3780 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003781 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003782 }
3783 node--;
3784 }
3785 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003786 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003787}
3788
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003790 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003791 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792};
3793
3794static void do_ccupdate_local(void *info)
3795{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003796 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 struct array_cache *old;
3798
3799 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003800 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003801
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3803 new->new[smp_processor_id()] = old;
3804}
3805
Christoph Lameter18004c52012-07-06 15:25:12 -05003806/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003807static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003808 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003810 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003811 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003812
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003813 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3814 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003815 if (!new)
3816 return -ENOMEM;
3817
Christoph Lametere498be72005-09-09 13:03:32 -07003818 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003819 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003820 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003821 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003822 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003823 kfree(new->new[i]);
3824 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003825 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 }
3827 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003828 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003830 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003831
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 cachep->batchcount = batchcount;
3834 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003835 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836
Christoph Lametere498be72005-09-09 13:03:32 -07003837 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003838 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839 if (!ccold)
3840 continue;
Christoph Lameter6a673682013-01-10 19:14:19 +00003841 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003842 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
Christoph Lameter6a673682013-01-10 19:14:19 +00003843 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844 kfree(ccold);
3845 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003846 kfree(new);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003847 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848}
3849
Glauber Costa943a4512012-12-18 14:23:03 -08003850static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3851 int batchcount, int shared, gfp_t gfp)
3852{
3853 int ret;
3854 struct kmem_cache *c = NULL;
3855 int i = 0;
3856
3857 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3858
3859 if (slab_state < FULL)
3860 return ret;
3861
3862 if ((ret < 0) || !is_root_cache(cachep))
3863 return ret;
3864
Glauber Costaebe945c2012-12-18 14:23:10 -08003865 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003866 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003867 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003868 if (c)
3869 /* return value determined by the parent cache only */
3870 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3871 }
3872
3873 return ret;
3874}
3875
Christoph Lameter18004c52012-07-06 15:25:12 -05003876/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003877static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878{
3879 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003880 int limit = 0;
3881 int shared = 0;
3882 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883
Glauber Costa943a4512012-12-18 14:23:03 -08003884 if (!is_root_cache(cachep)) {
3885 struct kmem_cache *root = memcg_root_cache(cachep);
3886 limit = root->limit;
3887 shared = root->shared;
3888 batchcount = root->batchcount;
3889 }
3890
3891 if (limit && shared && batchcount)
3892 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003893 /*
3894 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003895 * - create a LIFO ordering, i.e. return objects that are cache-warm
3896 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003897 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 * bufctl chains: array operations are cheaper.
3899 * The numbers are guessed, we should auto-tune as described by
3900 * Bonwick.
3901 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003902 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003904 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003906 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003908 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 limit = 54;
3910 else
3911 limit = 120;
3912
Andrew Mortona737b3e2006-03-22 00:08:11 -08003913 /*
3914 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915 * allocation behaviour: Most allocs on one cpu, most free operations
3916 * on another cpu. For these cases, an efficient object passing between
3917 * cpus is necessary. This is provided by a shared array. The array
3918 * replaces Bonwick's magazine layer.
3919 * On uniprocessor, it's functionally equivalent (but less efficient)
3920 * to a larger limit. Thus disabled by default.
3921 */
3922 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003923 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925
3926#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003927 /*
3928 * With debugging enabled, large batchcount lead to excessively long
3929 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 */
3931 if (limit > 32)
3932 limit = 32;
3933#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003934 batchcount = (limit + 1) / 2;
3935skip_setup:
3936 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937 if (err)
3938 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003939 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003940 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941}
3942
Christoph Lameter1b552532006-03-22 00:09:07 -08003943/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003944 * Drain an array if it contains any elements taking the node lock only if
3945 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003946 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003947 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003948static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003949 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950{
3951 int tofree;
3952
Christoph Lameter1b552532006-03-22 00:09:07 -08003953 if (!ac || !ac->avail)
3954 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955 if (ac->touched && !force) {
3956 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003957 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003958 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003959 if (ac->avail) {
3960 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3961 if (tofree > ac->avail)
3962 tofree = (ac->avail + 1) / 2;
3963 free_block(cachep, ac->entry, tofree, node);
3964 ac->avail -= tofree;
3965 memmove(ac->entry, &(ac->entry[tofree]),
3966 sizeof(void *) * ac->avail);
3967 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003968 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 }
3970}
3971
3972/**
3973 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003974 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 *
3976 * Called from workqueue/eventd every few seconds.
3977 * Purpose:
3978 * - clear the per-cpu caches for this CPU.
3979 * - return freeable pages to the main free memory pool.
3980 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003981 * If we cannot acquire the cache chain mutex then just give up - we'll try
3982 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003984static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003986 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003987 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003988 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003989 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990
Christoph Lameter18004c52012-07-06 15:25:12 -05003991 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003993 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994
Christoph Lameter18004c52012-07-06 15:25:12 -05003995 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996 check_irq_on();
3997
Christoph Lameter35386e32006-03-22 00:09:05 -08003998 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003999 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004000 * have established with reasonable certainty that
4001 * we can do some work if the lock was obtained.
4002 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004003 n = searchp->node[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004004
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004005 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004007 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008
Christoph Lameter35386e32006-03-22 00:09:05 -08004009 /*
4010 * These are racy checks but it does not matter
4011 * if we skip one check or scan twice.
4012 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004013 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004014 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004016 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004018 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004020 if (n->free_touched)
4021 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004022 else {
4023 int freed;
4024
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004025 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004026 5 * searchp->num - 1) / (5 * searchp->num));
4027 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004029next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 cond_resched();
4031 }
4032 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004033 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004034 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004035out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004036 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004037 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038}
4039
Linus Torvalds158a9622008-01-02 13:04:48 -08004040#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004041void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004043 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004044 unsigned long active_objs;
4045 unsigned long num_objs;
4046 unsigned long active_slabs = 0;
4047 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004048 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004050 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004051 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053 active_objs = 0;
4054 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004055 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004056 n = cachep->node[node];
4057 if (!n)
Christoph Lametere498be72005-09-09 13:03:32 -07004058 continue;
4059
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004060 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004061 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004062
Joonsoo Kim8456a642013-10-24 10:07:49 +09004063 list_for_each_entry(page, &n->slabs_full, lru) {
4064 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004065 error = "slabs_full accounting error";
4066 active_objs += cachep->num;
4067 active_slabs++;
4068 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004069 list_for_each_entry(page, &n->slabs_partial, lru) {
4070 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004071 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004072 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004073 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004074 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004075 active_slabs++;
4076 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004077 list_for_each_entry(page, &n->slabs_free, lru) {
4078 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004079 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004080 num_slabs++;
4081 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004082 free_objects += n->free_objects;
4083 if (n->shared)
4084 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004085
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004086 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004088 num_slabs += active_slabs;
4089 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004090 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091 error = "free_objects accounting error";
4092
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004093 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094 if (error)
4095 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4096
Glauber Costa0d7561c2012-10-19 18:20:27 +04004097 sinfo->active_objs = active_objs;
4098 sinfo->num_objs = num_objs;
4099 sinfo->active_slabs = active_slabs;
4100 sinfo->num_slabs = num_slabs;
4101 sinfo->shared_avail = shared_avail;
4102 sinfo->limit = cachep->limit;
4103 sinfo->batchcount = cachep->batchcount;
4104 sinfo->shared = cachep->shared;
4105 sinfo->objects_per_slab = cachep->num;
4106 sinfo->cache_order = cachep->gfporder;
4107}
4108
4109void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4110{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004112 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 unsigned long high = cachep->high_mark;
4114 unsigned long allocs = cachep->num_allocations;
4115 unsigned long grown = cachep->grown;
4116 unsigned long reaped = cachep->reaped;
4117 unsigned long errors = cachep->errors;
4118 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004120 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004121 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004122
Joe Perchese92dd4f2010-03-26 19:27:58 -07004123 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4124 "%4lu %4lu %4lu %4lu %4lu",
4125 allocs, high, grown,
4126 reaped, errors, max_freeable, node_allocs,
4127 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 }
4129 /* cpu stats */
4130 {
4131 unsigned long allochit = atomic_read(&cachep->allochit);
4132 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4133 unsigned long freehit = atomic_read(&cachep->freehit);
4134 unsigned long freemiss = atomic_read(&cachep->freemiss);
4135
4136 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004137 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 }
4139#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140}
4141
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142#define MAX_SLABINFO_WRITE 128
4143/**
4144 * slabinfo_write - Tuning for the slab allocator
4145 * @file: unused
4146 * @buffer: user buffer
4147 * @count: data length
4148 * @ppos: unused
4149 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004150ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004151 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004153 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004155 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004156
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157 if (count > MAX_SLABINFO_WRITE)
4158 return -EINVAL;
4159 if (copy_from_user(&kbuf, buffer, count))
4160 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004161 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162
4163 tmp = strchr(kbuf, ' ');
4164 if (!tmp)
4165 return -EINVAL;
4166 *tmp = '\0';
4167 tmp++;
4168 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4169 return -EINVAL;
4170
4171 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004172 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004174 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004176 if (limit < 1 || batchcount < 1 ||
4177 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004178 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004180 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004181 batchcount, shared,
4182 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 }
4184 break;
4185 }
4186 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004187 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188 if (res >= 0)
4189 res = count;
4190 return res;
4191}
Al Viro871751e2006-03-25 03:06:39 -08004192
4193#ifdef CONFIG_DEBUG_SLAB_LEAK
4194
4195static void *leaks_start(struct seq_file *m, loff_t *pos)
4196{
Christoph Lameter18004c52012-07-06 15:25:12 -05004197 mutex_lock(&slab_mutex);
4198 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004199}
4200
4201static inline int add_caller(unsigned long *n, unsigned long v)
4202{
4203 unsigned long *p;
4204 int l;
4205 if (!v)
4206 return 1;
4207 l = n[1];
4208 p = n + 2;
4209 while (l) {
4210 int i = l/2;
4211 unsigned long *q = p + 2 * i;
4212 if (*q == v) {
4213 q[1]++;
4214 return 1;
4215 }
4216 if (*q > v) {
4217 l = i;
4218 } else {
4219 p = q + 2;
4220 l -= i + 1;
4221 }
4222 }
4223 if (++n[1] == n[0])
4224 return 0;
4225 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4226 p[0] = v;
4227 p[1] = 1;
4228 return 1;
4229}
4230
Joonsoo Kim8456a642013-10-24 10:07:49 +09004231static void handle_slab(unsigned long *n, struct kmem_cache *c,
4232 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004233{
4234 void *p;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004235 int i, j;
4236
Al Viro871751e2006-03-25 03:06:39 -08004237 if (n[0] == n[1])
4238 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004239 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004240 bool active = true;
4241
Joonsoo Kim8456a642013-10-24 10:07:49 +09004242 for (j = page->active; j < c->num; j++) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004243 /* Skip freed item */
Joonsoo Kime5c58df2013-12-02 17:49:40 +09004244 if (get_free_obj(page, j) == i) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004245 active = false;
4246 break;
4247 }
4248 }
4249 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004250 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004251
Al Viro871751e2006-03-25 03:06:39 -08004252 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4253 return;
4254 }
4255}
4256
4257static void show_symbol(struct seq_file *m, unsigned long address)
4258{
4259#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004260 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004261 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004262
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004263 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004264 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004265 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004266 seq_printf(m, " [%s]", modname);
4267 return;
4268 }
4269#endif
4270 seq_printf(m, "%p", (void *)address);
4271}
4272
4273static int leaks_show(struct seq_file *m, void *p)
4274{
Thierry Reding0672aa72012-06-22 19:42:49 +02004275 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004276 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004277 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004278 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004279 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004280 int node;
4281 int i;
4282
4283 if (!(cachep->flags & SLAB_STORE_USER))
4284 return 0;
4285 if (!(cachep->flags & SLAB_RED_ZONE))
4286 return 0;
4287
4288 /* OK, we can do it */
4289
Christoph Lameterdb845062013-02-05 18:45:23 +00004290 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004291
4292 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004293 n = cachep->node[node];
4294 if (!n)
Al Viro871751e2006-03-25 03:06:39 -08004295 continue;
4296
4297 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004298 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004299
Joonsoo Kim8456a642013-10-24 10:07:49 +09004300 list_for_each_entry(page, &n->slabs_full, lru)
4301 handle_slab(x, cachep, page);
4302 list_for_each_entry(page, &n->slabs_partial, lru)
4303 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004304 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004305 }
4306 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004307 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004308 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004309 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004310 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004311 if (!m->private) {
4312 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004313 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004314 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004315 return -ENOMEM;
4316 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004317 *(unsigned long *)m->private = x[0] * 2;
4318 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004319 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004320 /* Now make sure this entry will be retried */
4321 m->count = m->size;
4322 return 0;
4323 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004324 for (i = 0; i < x[1]; i++) {
4325 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4326 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004327 seq_putc(m, '\n');
4328 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004329
Al Viro871751e2006-03-25 03:06:39 -08004330 return 0;
4331}
4332
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004333static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004334 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004335 .next = slab_next,
4336 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004337 .show = leaks_show,
4338};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004339
4340static int slabstats_open(struct inode *inode, struct file *file)
4341{
4342 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4343 int ret = -ENOMEM;
4344 if (n) {
4345 ret = seq_open(file, &slabstats_op);
4346 if (!ret) {
4347 struct seq_file *m = file->private_data;
4348 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4349 m->private = n;
4350 n = NULL;
4351 }
4352 kfree(n);
4353 }
4354 return ret;
4355}
4356
4357static const struct file_operations proc_slabstats_operations = {
4358 .open = slabstats_open,
4359 .read = seq_read,
4360 .llseek = seq_lseek,
4361 .release = seq_release_private,
4362};
Al Viro871751e2006-03-25 03:06:39 -08004363#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004364
4365static int __init slab_proc_init(void)
4366{
4367#ifdef CONFIG_DEBUG_SLAB_LEAK
4368 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4369#endif
4370 return 0;
4371}
4372module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373#endif
4374
Manfred Spraul00e145b2005-09-03 15:55:07 -07004375/**
4376 * ksize - get the actual amount of memory allocated for a given object
4377 * @objp: Pointer to the object
4378 *
4379 * kmalloc may internally round up allocations and return more memory
4380 * than requested. ksize() can be used to determine the actual amount of
4381 * memory allocated. The caller may use this additional memory, even though
4382 * a smaller amount of memory was initially specified with the kmalloc call.
4383 * The caller must guarantee that objp points to a valid object previously
4384 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4385 * must not be freed during the duration of the call.
4386 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004387size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004388{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004389 BUG_ON(!objp);
4390 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004391 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004392
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004393 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004394}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004395EXPORT_SYMBOL(ksize);