blob: 9ca3b87edabc699a7da3ceea71b61ee34a300e03 [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{
David Rientjes9a02d692014-06-04 16:06:36 -07001624#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001625 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001626 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001627 unsigned long flags;
1628 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001629 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1630 DEFAULT_RATELIMIT_BURST);
1631
1632 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1633 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001634
1635 printk(KERN_WARNING
1636 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1637 nodeid, gfpflags);
1638 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001639 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001640
1641 for_each_online_node(node) {
1642 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1643 unsigned long active_slabs = 0, num_slabs = 0;
1644
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001645 n = cachep->node[node];
1646 if (!n)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001647 continue;
1648
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001649 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001650 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001651 active_objs += cachep->num;
1652 active_slabs++;
1653 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001654 list_for_each_entry(page, &n->slabs_partial, lru) {
1655 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001656 active_slabs++;
1657 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001658 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001659 num_slabs++;
1660
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001661 free_objects += n->free_objects;
1662 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001663
1664 num_slabs += active_slabs;
1665 num_objs = num_slabs * cachep->num;
1666 printk(KERN_WARNING
1667 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1668 node, active_slabs, num_slabs, active_objs, num_objs,
1669 free_objects);
1670 }
David Rientjes9a02d692014-06-04 16:06:36 -07001671#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001672}
1673
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674/*
1675 * Interface to system's page allocator. No need to hold the cache-lock.
1676 *
1677 * If we requested dmaable memory, we will get it. Even if we
1678 * did not request dmaable memory, we might get it, but that
1679 * would be relatively rare and ignorable.
1680 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001681static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1682 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683{
1684 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001685 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001686
Glauber Costaa618e892012-06-14 16:17:21 +04001687 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001688 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1689 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001690
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001691 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1692 return NULL;
1693
Linus Torvalds517d0862009-06-16 19:50:13 -07001694 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001695 if (!page) {
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001696 memcg_uncharge_slab(cachep, cachep->gfporder);
David Rientjes9a02d692014-06-04 16:06:36 -07001697 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001699 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001701 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001702 if (unlikely(page->pfmemalloc))
1703 pfmemalloc_active = true;
1704
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001705 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001707 add_zone_page_state(page_zone(page),
1708 NR_SLAB_RECLAIMABLE, nr_pages);
1709 else
1710 add_zone_page_state(page_zone(page),
1711 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001712 __SetPageSlab(page);
1713 if (page->pfmemalloc)
1714 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001715
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001716 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1717 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1718
1719 if (cachep->ctor)
1720 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1721 else
1722 kmemcheck_mark_unallocated_pages(page, nr_pages);
1723 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001724
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001725 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726}
1727
1728/*
1729 * Interface to system's page release.
1730 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001731static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001733 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001735 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001736
Christoph Lameter972d1a72006-09-25 23:31:51 -07001737 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1738 sub_zone_page_state(page_zone(page),
1739 NR_SLAB_RECLAIMABLE, nr_freed);
1740 else
1741 sub_zone_page_state(page_zone(page),
1742 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001743
Joonsoo Kima57a4982013-10-24 10:07:44 +09001744 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001745 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001746 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001747 page_mapcount_reset(page);
1748 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001749
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 if (current->reclaim_state)
1751 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001752 __free_pages(page, cachep->gfporder);
1753 memcg_uncharge_slab(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754}
1755
1756static void kmem_rcu_free(struct rcu_head *head)
1757{
Joonsoo Kim68126702013-10-24 10:07:42 +09001758 struct kmem_cache *cachep;
1759 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
Joonsoo Kim68126702013-10-24 10:07:42 +09001761 page = container_of(head, struct page, rcu_head);
1762 cachep = page->slab_cache;
1763
1764 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765}
1766
1767#if DEBUG
1768
1769#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001770static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001773 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001775 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001777 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 return;
1779
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001780 *addr++ = 0x12345678;
1781 *addr++ = caller;
1782 *addr++ = smp_processor_id();
1783 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 {
1785 unsigned long *sptr = &caller;
1786 unsigned long svalue;
1787
1788 while (!kstack_end(sptr)) {
1789 svalue = *sptr++;
1790 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 size -= sizeof(unsigned long);
1793 if (size <= sizeof(unsigned long))
1794 break;
1795 }
1796 }
1797
1798 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001799 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800}
1801#endif
1802
Pekka Enberg343e0d72006-02-01 03:05:50 -08001803static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001805 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001806 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807
1808 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001809 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810}
1811
1812static void dump_line(char *data, int offset, int limit)
1813{
1814 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001815 unsigned char error = 0;
1816 int bad_count = 0;
1817
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001818 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001819 for (i = 0; i < limit; i++) {
1820 if (data[offset + i] != POISON_FREE) {
1821 error = data[offset + i];
1822 bad_count++;
1823 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001824 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001825 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1826 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001827
1828 if (bad_count == 1) {
1829 error ^= POISON_FREE;
1830 if (!(error & (error - 1))) {
1831 printk(KERN_ERR "Single bit error detected. Probably "
1832 "bad RAM.\n");
1833#ifdef CONFIG_X86
1834 printk(KERN_ERR "Run memtest86+ or a similar memory "
1835 "test tool.\n");
1836#else
1837 printk(KERN_ERR "Run a memory test tool.\n");
1838#endif
1839 }
1840 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841}
1842#endif
1843
1844#if DEBUG
1845
Pekka Enberg343e0d72006-02-01 03:05:50 -08001846static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847{
1848 int i, size;
1849 char *realobj;
1850
1851 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001852 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001853 *dbg_redzone1(cachep, objp),
1854 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 }
1856
1857 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001858 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1859 *dbg_userword(cachep, objp),
1860 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001862 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001863 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001864 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 int limit;
1866 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001867 if (i + limit > size)
1868 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 dump_line(realobj, i, limit);
1870 }
1871}
1872
Pekka Enberg343e0d72006-02-01 03:05:50 -08001873static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874{
1875 char *realobj;
1876 int size, i;
1877 int lines = 0;
1878
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001879 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001880 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001882 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001884 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 exp = POISON_END;
1886 if (realobj[i] != exp) {
1887 int limit;
1888 /* Mismatch ! */
1889 /* Print header */
1890 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001891 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001892 "Slab corruption (%s): %s start=%p, len=%d\n",
1893 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 print_objinfo(cachep, objp, 0);
1895 }
1896 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001897 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001899 if (i + limit > size)
1900 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 dump_line(realobj, i, limit);
1902 i += 16;
1903 lines++;
1904 /* Limit to 5 lines */
1905 if (lines > 5)
1906 break;
1907 }
1908 }
1909 if (lines != 0) {
1910 /* Print some data about the neighboring objects, if they
1911 * exist:
1912 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001913 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001914 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915
Joonsoo Kim8456a642013-10-24 10:07:49 +09001916 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001918 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001919 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001921 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 print_objinfo(cachep, objp, 2);
1923 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001925 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001926 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001928 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 print_objinfo(cachep, objp, 2);
1930 }
1931 }
1932}
1933#endif
1934
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001936static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1937 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001938{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 int i;
1940 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001941 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942
1943 if (cachep->flags & SLAB_POISON) {
1944#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001945 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001946 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001947 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001948 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 else
1950 check_poison_obj(cachep, objp);
1951#else
1952 check_poison_obj(cachep, objp);
1953#endif
1954 }
1955 if (cachep->flags & SLAB_RED_ZONE) {
1956 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1957 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001958 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1960 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001961 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001964}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001966static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1967 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001968{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001969}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970#endif
1971
Randy Dunlap911851e2006-03-22 00:08:14 -08001972/**
1973 * slab_destroy - destroy and release all objects in a slab
1974 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001975 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001976 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001977 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001978 * Before calling the slab must have been unlinked from the cache. The
1979 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001980 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001981static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001982{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001983 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001984
Joonsoo Kim8456a642013-10-24 10:07:49 +09001985 freelist = page->freelist;
1986 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001988 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989
Joonsoo Kim68126702013-10-24 10:07:42 +09001990 /*
1991 * RCU free overloads the RCU head over the LRU.
1992 * slab_page has been overloeaded over the LRU,
1993 * however it is not used from now on so that
1994 * we can use it safely.
1995 */
1996 head = (void *)&page->rcu_head;
1997 call_rcu(head, kmem_rcu_free);
1998
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002000 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 }
Joonsoo Kim68126702013-10-24 10:07:42 +09002002
2003 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09002004 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09002005 * although actual page can be freed in rcu context
2006 */
2007 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09002008 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009}
2010
2011/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002012 * calculate_slab_order - calculate size (page order) of slabs
2013 * @cachep: pointer to the cache that is being created
2014 * @size: size of objects to be created in this cache.
2015 * @align: required alignment for the objects.
2016 * @flags: slab allocation flags
2017 *
2018 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002019 *
2020 * This could be made much more intelligent. For now, try to avoid using
2021 * high order pages for slabs. When the gfp() functions are more friendly
2022 * towards high-order requests, this should be changed.
2023 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002024static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002025 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002026{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002027 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002028 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002029 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002030
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002031 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002032 unsigned int num;
2033 size_t remainder;
2034
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002035 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002036 if (!num)
2037 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002038
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002039 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2040 if (num > SLAB_OBJ_MAX_NUM)
2041 break;
2042
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002043 if (flags & CFLGS_OFF_SLAB) {
2044 /*
2045 * Max number of objs-per-slab for caches which
2046 * use off-slab slabs. Needed to avoid a possible
2047 * looping condition in cache_grow().
2048 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002049 offslab_limit = size;
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002050 offslab_limit /= sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002051
2052 if (num > offslab_limit)
2053 break;
2054 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002055
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002056 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002057 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002058 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002059 left_over = remainder;
2060
2061 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002062 * A VFS-reclaimable slab tends to have most allocations
2063 * as GFP_NOFS and we really don't want to have to be allocating
2064 * higher-order pages when we are unable to shrink dcache.
2065 */
2066 if (flags & SLAB_RECLAIM_ACCOUNT)
2067 break;
2068
2069 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002070 * Large number of objects is good, but very large slabs are
2071 * currently bad for the gfp()s.
2072 */
David Rientjes543585c2011-10-18 22:09:24 -07002073 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002074 break;
2075
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002076 /*
2077 * Acceptable internal fragmentation?
2078 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002079 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002080 break;
2081 }
2082 return left_over;
2083}
2084
Pekka Enberg83b519e2009-06-10 19:40:04 +03002085static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002086{
Christoph Lameter97d06602012-07-06 15:25:11 -05002087 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002088 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002089
Christoph Lameter97d06602012-07-06 15:25:11 -05002090 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002091 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002092 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002093 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002094 * of by the caller of __kmem_cache_create
2095 */
2096 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2097 slab_state = PARTIAL;
2098 } else if (slab_state == PARTIAL) {
2099 /*
2100 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002101 * that's used by kmalloc(24), otherwise the creation of
2102 * further caches will BUG().
2103 */
2104 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2105
2106 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002107 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2108 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002109 * otherwise the creation of further caches will BUG().
2110 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002111 set_up_node(cachep, SIZE_AC);
2112 if (INDEX_AC == INDEX_NODE)
2113 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002114 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002115 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002116 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002117 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002118 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002119 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002120
Christoph Lameter97d06602012-07-06 15:25:11 -05002121 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002122 set_up_node(cachep, SIZE_NODE);
2123 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002124 } else {
2125 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002126 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002127 cachep->node[node] =
Christoph Lameter6744f082013-01-10 19:12:17 +00002128 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002129 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002130 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002131 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002132 }
2133 }
2134 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002135 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002136 jiffies + REAPTIMEOUT_NODE +
2137 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002138
2139 cpu_cache_get(cachep)->avail = 0;
2140 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2141 cpu_cache_get(cachep)->batchcount = 1;
2142 cpu_cache_get(cachep)->touched = 0;
2143 cachep->batchcount = 1;
2144 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002145 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002146}
2147
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002148/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002149 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002150 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 *
2153 * Returns a ptr to the cache on success, NULL on failure.
2154 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002155 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 * The flags are
2158 *
2159 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2160 * to catch references to uninitialised memory.
2161 *
2162 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2163 * for buffer overruns.
2164 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2166 * cacheline. This can be beneficial if you're counting cycles as closely
2167 * as davem.
2168 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002169int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002170__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002172 size_t left_over, freelist_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002173 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002174 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002175 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178#if FORCED_DEBUG
2179 /*
2180 * Enable redzoning and last user accounting, except for caches with
2181 * large objects, if the increased size would increase the object size
2182 * above the next power of two: caches with object sizes just above a
2183 * power of two have a significant amount of internal fragmentation.
2184 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002185 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2186 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002187 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 if (!(flags & SLAB_DESTROY_BY_RCU))
2189 flags |= SLAB_POISON;
2190#endif
2191 if (flags & SLAB_DESTROY_BY_RCU)
2192 BUG_ON(flags & SLAB_POISON);
2193#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194
Andrew Mortona737b3e2006-03-22 00:08:11 -08002195 /*
2196 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197 * unaligned accesses for some archs when redzoning is used, and makes
2198 * sure any on-slab bufctl's are also correctly aligned.
2199 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002200 if (size & (BYTES_PER_WORD - 1)) {
2201 size += (BYTES_PER_WORD - 1);
2202 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 }
2204
Pekka Enbergca5f9702006-09-25 23:31:25 -07002205 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002206 * Redzoning and user store require word alignment or possibly larger.
2207 * Note this will be overridden by architecture or caller mandated
2208 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002209 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002210 if (flags & SLAB_STORE_USER)
2211 ralign = BYTES_PER_WORD;
2212
2213 if (flags & SLAB_RED_ZONE) {
2214 ralign = REDZONE_ALIGN;
2215 /* If redzoning, ensure that the second redzone is suitably
2216 * aligned, by adjusting the object size accordingly. */
2217 size += REDZONE_ALIGN - 1;
2218 size &= ~(REDZONE_ALIGN - 1);
2219 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002220
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002221 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002222 if (ralign < cachep->align) {
2223 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002225 /* disable debug if necessary */
2226 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002227 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002228 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002229 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002231 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232
Pekka Enberg83b519e2009-06-10 19:40:04 +03002233 if (slab_is_available())
2234 gfp = GFP_KERNEL;
2235 else
2236 gfp = GFP_NOWAIT;
2237
Christoph Lameter6a673682013-01-10 19:14:19 +00002238 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240
Pekka Enbergca5f9702006-09-25 23:31:25 -07002241 /*
2242 * Both debugging options require word-alignment which is calculated
2243 * into align above.
2244 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002247 cachep->obj_offset += sizeof(unsigned long long);
2248 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 }
2250 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002251 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002252 * the real object. But if the second red zone needs to be
2253 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002254 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002255 if (flags & SLAB_RED_ZONE)
2256 size += REDZONE_ALIGN;
2257 else
2258 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 }
2260#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002261 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002262 && cachep->object_size > cache_line_size()
2263 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2264 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 size = PAGE_SIZE;
2266 }
2267#endif
2268#endif
2269
Ingo Molnare0a42722006-06-23 02:03:46 -07002270 /*
2271 * Determine if the slab management is 'on' or 'off' slab.
2272 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002273 * it too early on. Always use on-slab management when
2274 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002275 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002276 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002277 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 /*
2279 * Size is large, assume best to place the slab management obj
2280 * off-slab (should allow better packing of objs).
2281 */
2282 flags |= CFLGS_OFF_SLAB;
2283
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002284 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002285 /*
2286 * We should restrict the number of objects in a slab to implement
2287 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2288 */
2289 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2290 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002292 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002294 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002295 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002296
Joonsoo Kim8456a642013-10-24 10:07:49 +09002297 freelist_size =
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002298 ALIGN(cachep->num * sizeof(freelist_idx_t), cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299
2300 /*
2301 * If the slab has been placed off-slab, and we have enough space then
2302 * move it on-slab. This is at the expense of any extra colouring.
2303 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002304 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002306 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 }
2308
2309 if (flags & CFLGS_OFF_SLAB) {
2310 /* really off slab. No need for manual alignment */
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002311 freelist_size = cachep->num * sizeof(freelist_idx_t);
Ron Lee67461362009-05-22 04:58:22 +09302312
2313#ifdef CONFIG_PAGE_POISONING
2314 /* If we're going to use the generic kernel_map_pages()
2315 * poisoning, then it's going to smash the contents of
2316 * the redzone and userword anyhow, so switch them off.
2317 */
2318 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2319 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2320#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 }
2322
2323 cachep->colour_off = cache_line_size();
2324 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002325 if (cachep->colour_off < cachep->align)
2326 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002327 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002328 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002330 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002331 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002332 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002333 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002334 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002336 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002337 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002338 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002339 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002340 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002341 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2342 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002343 * But leave a BUG_ON for some lucky dude.
2344 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002345 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002348 err = setup_cpu_cache(cachep, gfp);
2349 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002350 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002351 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002352 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353
Peter Zijlstra83835b32011-07-22 15:26:05 +02002354 if (flags & SLAB_DEBUG_OBJECTS) {
2355 /*
2356 * Would deadlock through slab_destroy()->call_rcu()->
2357 * debug_object_activate()->kmem_cache_alloc().
2358 */
2359 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2360
2361 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002362 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2363 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002364
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002365 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367
2368#if DEBUG
2369static void check_irq_off(void)
2370{
2371 BUG_ON(!irqs_disabled());
2372}
2373
2374static void check_irq_on(void)
2375{
2376 BUG_ON(irqs_disabled());
2377}
2378
Pekka Enberg343e0d72006-02-01 03:05:50 -08002379static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380{
2381#ifdef CONFIG_SMP
2382 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002383 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384#endif
2385}
Christoph Lametere498be72005-09-09 13:03:32 -07002386
Pekka Enberg343e0d72006-02-01 03:05:50 -08002387static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002388{
2389#ifdef CONFIG_SMP
2390 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002391 assert_spin_locked(&cachep->node[node]->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002392#endif
2393}
2394
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395#else
2396#define check_irq_off() do { } while(0)
2397#define check_irq_on() do { } while(0)
2398#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002399#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400#endif
2401
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002402static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002403 struct array_cache *ac,
2404 int force, int node);
2405
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406static void do_drain(void *arg)
2407{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002408 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002410 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
2412 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002413 ac = cpu_cache_get(cachep);
Christoph Lameter6a673682013-01-10 19:14:19 +00002414 spin_lock(&cachep->node[node]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002415 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002416 spin_unlock(&cachep->node[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 ac->avail = 0;
2418}
2419
Pekka Enberg343e0d72006-02-01 03:05:50 -08002420static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002422 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002423 int node;
2424
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002425 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002427 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002428 n = cachep->node[node];
2429 if (n && n->alien)
2430 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002431 }
2432
2433 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002434 n = cachep->node[node];
2435 if (n)
2436 drain_array(cachep, n, n->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002437 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438}
2439
Christoph Lametered11d9e2006-06-30 01:55:45 -07002440/*
2441 * Remove slabs from the list of free slabs.
2442 * Specify the number of slabs to drain in tofree.
2443 *
2444 * Returns the actual number of slabs released.
2445 */
2446static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002447 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002449 struct list_head *p;
2450 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002451 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452
Christoph Lametered11d9e2006-06-30 01:55:45 -07002453 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002454 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002456 spin_lock_irq(&n->list_lock);
2457 p = n->slabs_free.prev;
2458 if (p == &n->slabs_free) {
2459 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002460 goto out;
2461 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462
Joonsoo Kim8456a642013-10-24 10:07:49 +09002463 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002465 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002467 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002468 /*
2469 * Safe to drop the lock. The slab is no longer linked
2470 * to the cache.
2471 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002472 n->free_objects -= cache->num;
2473 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002474 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002475 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002477out:
2478 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479}
2480
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002481int __kmem_cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002482{
2483 int ret = 0, i = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002484 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002485
2486 drain_cpu_caches(cachep);
2487
2488 check_irq_on();
2489 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002490 n = cachep->node[i];
2491 if (!n)
Christoph Lametered11d9e2006-06-30 01:55:45 -07002492 continue;
2493
Wanpeng Li0fa81032013-07-04 08:33:22 +08002494 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002495
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002496 ret += !list_empty(&n->slabs_full) ||
2497 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002498 }
2499 return (ret ? 1 : 0);
2500}
2501
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002502int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503{
Christoph Lameter12c36672012-09-04 23:38:33 +00002504 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002505 struct kmem_cache_node *n;
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002506 int rc = __kmem_cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507
Christoph Lameter12c36672012-09-04 23:38:33 +00002508 if (rc)
2509 return rc;
2510
2511 for_each_online_cpu(i)
2512 kfree(cachep->array[i]);
2513
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002514 /* NUMA: free the node structures */
Christoph Lameter12c36672012-09-04 23:38:33 +00002515 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002516 n = cachep->node[i];
2517 if (n) {
2518 kfree(n->shared);
2519 free_alien_cache(n->alien);
2520 kfree(n);
Christoph Lameter12c36672012-09-04 23:38:33 +00002521 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002523 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002526/*
2527 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002528 *
2529 * For a slab cache when the slab descriptor is off-slab, the
2530 * slab descriptor can't come from the same cache which is being created,
2531 * Because if it is the case, that means we defer the creation of
2532 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2533 * And we eventually call down to __kmem_cache_create(), which
2534 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2535 * This is a "chicken-and-egg" problem.
2536 *
2537 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2538 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002539 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002540static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002541 struct page *page, int colour_off,
2542 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002544 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002545 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002546
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 if (OFF_SLAB(cachep)) {
2548 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002549 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002550 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002551 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 return NULL;
2553 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002554 freelist = addr + colour_off;
2555 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002557 page->active = 0;
2558 page->s_mem = addr + colour_off;
2559 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560}
2561
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002562static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002564 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002565}
2566
2567static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002568 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002569{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002570 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571}
2572
Pekka Enberg343e0d72006-02-01 03:05:50 -08002573static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002574 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575{
2576 int i;
2577
2578 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002579 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580#if DEBUG
2581 /* need to poison the objs? */
2582 if (cachep->flags & SLAB_POISON)
2583 poison_obj(cachep, objp, POISON_FREE);
2584 if (cachep->flags & SLAB_STORE_USER)
2585 *dbg_userword(cachep, objp) = NULL;
2586
2587 if (cachep->flags & SLAB_RED_ZONE) {
2588 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2589 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2590 }
2591 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002592 * Constructors are not allowed to allocate memory from the same
2593 * cache which they are a constructor for. Otherwise, deadlock.
2594 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 */
2596 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002597 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598
2599 if (cachep->flags & SLAB_RED_ZONE) {
2600 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2601 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002602 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2604 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002605 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002607 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002608 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002609 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002610 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611#else
2612 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002613 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614#endif
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002615 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617}
2618
Pekka Enberg343e0d72006-02-01 03:05:50 -08002619static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002621 if (CONFIG_ZONE_DMA_FLAG) {
2622 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002623 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002624 else
Glauber Costaa618e892012-06-14 16:17:21 +04002625 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002626 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627}
2628
Joonsoo Kim8456a642013-10-24 10:07:49 +09002629static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002630 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002631{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002632 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002633
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002634 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002635 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002636#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002637 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002638#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002639
2640 return objp;
2641}
2642
Joonsoo Kim8456a642013-10-24 10:07:49 +09002643static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002644 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002645{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002646 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002647#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002648 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002649
Matthew Dobson78d382d2006-02-01 03:05:47 -08002650 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002651 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002652
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002653 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002654 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002655 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002656 printk(KERN_ERR "slab: double free detected in cache "
2657 "'%s', objp %p\n", cachep->name, objp);
2658 BUG();
2659 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002660 }
2661#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002662 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002663 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002664}
2665
Pekka Enberg47768742006-06-23 02:03:07 -07002666/*
2667 * Map pages beginning at addr to the given cache and slab. This is required
2668 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002669 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002670 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002671static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002672 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002674 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002675 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676}
2677
2678/*
2679 * Grow (by 1) the number of slabs within a cache. This is called by
2680 * kmem_cache_alloc() when there are no active objs left in a cache.
2681 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002682static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002683 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002685 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002686 size_t offset;
2687 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002688 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689
Andrew Mortona737b3e2006-03-22 00:08:11 -08002690 /*
2691 * Be lazy and only check for valid flags here, keeping it out of the
2692 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002694 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2695 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002697 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002699 n = cachep->node[nodeid];
2700 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701
2702 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002703 offset = n->colour_next;
2704 n->colour_next++;
2705 if (n->colour_next >= cachep->colour)
2706 n->colour_next = 0;
2707 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002709 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710
2711 if (local_flags & __GFP_WAIT)
2712 local_irq_enable();
2713
2714 /*
2715 * The test for missing atomic flag is performed here, rather than
2716 * the more obvious place, simply to reduce the critical path length
2717 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2718 * will eventually be caught here (where it matters).
2719 */
2720 kmem_flagcheck(cachep, flags);
2721
Andrew Mortona737b3e2006-03-22 00:08:11 -08002722 /*
2723 * Get mem for the objs. Attempt to allocate a physical page from
2724 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002725 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002726 if (!page)
2727 page = kmem_getpages(cachep, local_flags, nodeid);
2728 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 goto failed;
2730
2731 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002732 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002733 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002734 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 goto opps1;
2736
Joonsoo Kim8456a642013-10-24 10:07:49 +09002737 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738
Joonsoo Kim8456a642013-10-24 10:07:49 +09002739 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740
2741 if (local_flags & __GFP_WAIT)
2742 local_irq_disable();
2743 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002744 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
2746 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002747 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002749 n->free_objects += cachep->num;
2750 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002752opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002753 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002754failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 if (local_flags & __GFP_WAIT)
2756 local_irq_disable();
2757 return 0;
2758}
2759
2760#if DEBUG
2761
2762/*
2763 * Perform extra freeing checks:
2764 * - detect bad pointers.
2765 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 */
2767static void kfree_debugcheck(const void *objp)
2768{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769 if (!virt_addr_valid(objp)) {
2770 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002771 (unsigned long)objp);
2772 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774}
2775
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002776static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2777{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002778 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002779
2780 redzone1 = *dbg_redzone1(cache, obj);
2781 redzone2 = *dbg_redzone2(cache, obj);
2782
2783 /*
2784 * Redzone is ok.
2785 */
2786 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2787 return;
2788
2789 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2790 slab_error(cache, "double free detected");
2791 else
2792 slab_error(cache, "memory outside object was overwritten");
2793
David Woodhouseb46b8f12007-05-08 00:22:59 -07002794 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002795 obj, redzone1, redzone2);
2796}
2797
Pekka Enberg343e0d72006-02-01 03:05:50 -08002798static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002799 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002802 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002804 BUG_ON(virt_to_cache(objp) != cachep);
2805
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002806 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002808 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002811 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2813 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2814 }
2815 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002816 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817
Joonsoo Kim8456a642013-10-24 10:07:49 +09002818 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
2820 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002821 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 if (cachep->flags & SLAB_POISON) {
2824#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002825 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002826 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002827 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002828 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 } else {
2830 poison_obj(cachep, objp, POISON_FREE);
2831 }
2832#else
2833 poison_obj(cachep, objp, POISON_FREE);
2834#endif
2835 }
2836 return objp;
2837}
2838
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839#else
2840#define kfree_debugcheck(x) do { } while(0)
2841#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842#endif
2843
Mel Gorman072bb0a2012-07-31 16:43:58 -07002844static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2845 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846{
2847 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002848 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002850 int node;
2851
Joe Korty6d2144d2008-03-05 15:04:59 -08002852 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002853 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002854 if (unlikely(force_refill))
2855 goto force_grow;
2856retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002857 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 batchcount = ac->batchcount;
2859 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002860 /*
2861 * If there was little recent activity on this cache, then
2862 * perform only a partial refill. Otherwise we could generate
2863 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 */
2865 batchcount = BATCHREFILL_LIMIT;
2866 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002867 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002869 BUG_ON(ac->avail > 0 || !n);
2870 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002871
Christoph Lameter3ded1752006-03-25 03:06:44 -08002872 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002873 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2874 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002875 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002876 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002877
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 while (batchcount > 0) {
2879 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002880 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002882 entry = n->slabs_partial.next;
2883 if (entry == &n->slabs_partial) {
2884 n->free_touched = 1;
2885 entry = n->slabs_free.next;
2886 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 goto must_grow;
2888 }
2889
Joonsoo Kim8456a642013-10-24 10:07:49 +09002890 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002892
2893 /*
2894 * The slab was either on partial or free list so
2895 * there must be at least one object available for
2896 * allocation.
2897 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002898 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002899
Joonsoo Kim8456a642013-10-24 10:07:49 +09002900 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 STATS_INC_ALLOCED(cachep);
2902 STATS_INC_ACTIVE(cachep);
2903 STATS_SET_HIGH(cachep);
2904
Joonsoo Kim8456a642013-10-24 10:07:49 +09002905 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002906 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908
2909 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002910 list_del(&page->lru);
2911 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002912 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002914 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 }
2916
Andrew Mortona737b3e2006-03-22 00:08:11 -08002917must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002918 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002919alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002920 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921
2922 if (unlikely(!ac->avail)) {
2923 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002924force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002925 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002926
Andrew Mortona737b3e2006-03-22 00:08:11 -08002927 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002928 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002929 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002930
2931 /* no objects in sight? abort */
2932 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 return NULL;
2934
Andrew Mortona737b3e2006-03-22 00:08:11 -08002935 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 goto retry;
2937 }
2938 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002939
2940 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941}
2942
Andrew Mortona737b3e2006-03-22 00:08:11 -08002943static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2944 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945{
2946 might_sleep_if(flags & __GFP_WAIT);
2947#if DEBUG
2948 kmem_flagcheck(cachep, flags);
2949#endif
2950}
2951
2952#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002953static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002954 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002956 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002958 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002960 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002961 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002962 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 else
2964 check_poison_obj(cachep, objp);
2965#else
2966 check_poison_obj(cachep, objp);
2967#endif
2968 poison_obj(cachep, objp, POISON_INUSE);
2969 }
2970 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002971 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972
2973 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002974 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2975 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2976 slab_error(cachep, "double free, or memory outside"
2977 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002978 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002979 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002980 objp, *dbg_redzone1(cachep, objp),
2981 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 }
2983 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2984 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2985 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002986 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002987 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002988 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002989 if (ARCH_SLAB_MINALIGN &&
2990 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002991 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002992 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002993 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 return objp;
2995}
2996#else
2997#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2998#endif
2999
Akinobu Mita773ff602008-12-23 19:37:01 +09003000static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003001{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00003002 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003003 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003004
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003005 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003006}
3007
Pekka Enberg343e0d72006-02-01 03:05:50 -08003008static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003010 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003012 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013
Alok N Kataria5c382302005-09-27 21:45:46 -07003014 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003015
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003016 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003019 objp = ac_get_obj(cachep, ac, flags, false);
3020
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003021 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003022 * Allow for the possibility all avail objects are not allowed
3023 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003024 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003025 if (objp) {
3026 STATS_INC_ALLOCHIT(cachep);
3027 goto out;
3028 }
3029 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003031
3032 STATS_INC_ALLOCMISS(cachep);
3033 objp = cache_alloc_refill(cachep, flags, force_refill);
3034 /*
3035 * the 'ac' may be updated by cache_alloc_refill(),
3036 * and kmemleak_erase() requires its correct value.
3037 */
3038 ac = cpu_cache_get(cachep);
3039
3040out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003041 /*
3042 * To avoid a false negative, if an object that is in one of the
3043 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3044 * treat the array pointers as a reference to the object.
3045 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003046 if (objp)
3047 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003048 return objp;
3049}
3050
Christoph Lametere498be72005-09-09 13:03:32 -07003051#ifdef CONFIG_NUMA
3052/*
David Rientjesf0432d12014-04-07 15:37:30 -07003053 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003054 *
3055 * If we are in_interrupt, then process context, including cpusets and
3056 * mempolicy, may not apply and should not be used for allocation policy.
3057 */
3058static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3059{
3060 int nid_alloc, nid_here;
3061
Christoph Lameter765c4502006-09-27 01:50:08 -07003062 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003063 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003064 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003065 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003066 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003067 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003068 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003069 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003070 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003071 return NULL;
3072}
3073
3074/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003075 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003076 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003077 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003078 * perform an allocation without specifying a node. This allows the page
3079 * allocator to do its reclaim / fallback magic. We then insert the
3080 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003081 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003082static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003083{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003084 struct zonelist *zonelist;
3085 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003086 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003087 struct zone *zone;
3088 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003089 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003090 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003091 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003092
3093 if (flags & __GFP_THISNODE)
3094 return NULL;
3095
Christoph Lameter6cb06222007-10-16 01:25:41 -07003096 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003097
Mel Gormancc9a6c82012-03-21 16:34:11 -07003098retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003099 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003100 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003101
Christoph Lameter3c517a62006-12-06 20:33:29 -08003102retry:
3103 /*
3104 * Look through allowed nodes for objects available
3105 * from existing per node queues.
3106 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003107 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3108 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003109
Mel Gorman54a6eb52008-04-28 02:12:16 -07003110 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter6a673682013-01-10 19:14:19 +00003111 cache->node[nid] &&
3112 cache->node[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003113 obj = ____cache_alloc_node(cache,
3114 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003115 if (obj)
3116 break;
3117 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003118 }
3119
Christoph Lametercfce6602007-05-06 14:50:17 -07003120 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003121 /*
3122 * This allocation will be performed within the constraints
3123 * of the current cpuset / memory policy requirements.
3124 * We may trigger various forms of reclaim on the allowed
3125 * set and go into memory reserves if necessary.
3126 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003127 struct page *page;
3128
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003129 if (local_flags & __GFP_WAIT)
3130 local_irq_enable();
3131 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003132 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003133 if (local_flags & __GFP_WAIT)
3134 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003135 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003136 /*
3137 * Insert into the appropriate per node queues
3138 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003139 nid = page_to_nid(page);
3140 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003141 obj = ____cache_alloc_node(cache,
3142 flags | GFP_THISNODE, nid);
3143 if (!obj)
3144 /*
3145 * Another processor may allocate the
3146 * objects in the slab since we are
3147 * not holding any locks.
3148 */
3149 goto retry;
3150 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003151 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003152 obj = NULL;
3153 }
3154 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003155 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003156
Mel Gormand26914d2014-04-03 14:47:24 -07003157 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003158 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003159 return obj;
3160}
3161
3162/*
Christoph Lametere498be72005-09-09 13:03:32 -07003163 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003165static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003166 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003167{
3168 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003169 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003170 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003171 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003172 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003174 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003175 n = cachep->node[nodeid];
3176 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003177
Andrew Mortona737b3e2006-03-22 00:08:11 -08003178retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003179 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003180 spin_lock(&n->list_lock);
3181 entry = n->slabs_partial.next;
3182 if (entry == &n->slabs_partial) {
3183 n->free_touched = 1;
3184 entry = n->slabs_free.next;
3185 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003186 goto must_grow;
3187 }
Christoph Lametere498be72005-09-09 13:03:32 -07003188
Joonsoo Kim8456a642013-10-24 10:07:49 +09003189 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003190 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003191
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003192 STATS_INC_NODEALLOCS(cachep);
3193 STATS_INC_ACTIVE(cachep);
3194 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003195
Joonsoo Kim8456a642013-10-24 10:07:49 +09003196 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003197
Joonsoo Kim8456a642013-10-24 10:07:49 +09003198 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003199 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003200 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003201 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003202
Joonsoo Kim8456a642013-10-24 10:07:49 +09003203 if (page->active == cachep->num)
3204 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003205 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003206 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003207
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003208 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003209 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003210
Andrew Mortona737b3e2006-03-22 00:08:11 -08003211must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003212 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003213 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003214 if (x)
3215 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003216
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003217 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003218
Andrew Mortona737b3e2006-03-22 00:08:11 -08003219done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003220 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003221}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003222
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003223static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003224slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003225 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003226{
3227 unsigned long save_flags;
3228 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003229 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003230
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003231 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003232
Nick Piggincf40bd12009-01-21 08:12:39 +01003233 lockdep_trace_alloc(flags);
3234
Akinobu Mita773ff602008-12-23 19:37:01 +09003235 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003236 return NULL;
3237
Glauber Costad79923f2012-12-18 14:22:48 -08003238 cachep = memcg_kmem_get_cache(cachep, flags);
3239
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003240 cache_alloc_debugcheck_before(cachep, flags);
3241 local_irq_save(save_flags);
3242
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003243 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003244 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003245
Christoph Lameter6a673682013-01-10 19:14:19 +00003246 if (unlikely(!cachep->node[nodeid])) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003247 /* Node not bootstrapped yet */
3248 ptr = fallback_alloc(cachep, flags);
3249 goto out;
3250 }
3251
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003252 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003253 /*
3254 * Use the locally cached objects if possible.
3255 * However ____cache_alloc does not allow fallback
3256 * to other nodes. It may fail while we still have
3257 * objects on other nodes available.
3258 */
3259 ptr = ____cache_alloc(cachep, flags);
3260 if (ptr)
3261 goto out;
3262 }
3263 /* ___cache_alloc_node can fall back to other nodes */
3264 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3265 out:
3266 local_irq_restore(save_flags);
3267 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003268 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003269 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003270
Joe Perches5087c822013-09-10 17:02:51 -07003271 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003272 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003273 if (unlikely(flags & __GFP_ZERO))
3274 memset(ptr, 0, cachep->object_size);
3275 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003276
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003277 return ptr;
3278}
3279
3280static __always_inline void *
3281__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3282{
3283 void *objp;
3284
David Rientjesf0432d12014-04-07 15:37:30 -07003285 if (current->mempolicy || unlikely(current->flags & PF_SPREAD_SLAB)) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003286 objp = alternate_node_alloc(cache, flags);
3287 if (objp)
3288 goto out;
3289 }
3290 objp = ____cache_alloc(cache, flags);
3291
3292 /*
3293 * We may just have run out of memory on the local node.
3294 * ____cache_alloc_node() knows how to locate memory on other nodes
3295 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003296 if (!objp)
3297 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003298
3299 out:
3300 return objp;
3301}
3302#else
3303
3304static __always_inline void *
3305__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3306{
3307 return ____cache_alloc(cachep, flags);
3308}
3309
3310#endif /* CONFIG_NUMA */
3311
3312static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003313slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003314{
3315 unsigned long save_flags;
3316 void *objp;
3317
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003318 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003319
Nick Piggincf40bd12009-01-21 08:12:39 +01003320 lockdep_trace_alloc(flags);
3321
Akinobu Mita773ff602008-12-23 19:37:01 +09003322 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003323 return NULL;
3324
Glauber Costad79923f2012-12-18 14:22:48 -08003325 cachep = memcg_kmem_get_cache(cachep, flags);
3326
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003327 cache_alloc_debugcheck_before(cachep, flags);
3328 local_irq_save(save_flags);
3329 objp = __do_cache_alloc(cachep, flags);
3330 local_irq_restore(save_flags);
3331 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003332 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003333 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003334 prefetchw(objp);
3335
Joe Perches5087c822013-09-10 17:02:51 -07003336 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003337 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003338 if (unlikely(flags & __GFP_ZERO))
3339 memset(objp, 0, cachep->object_size);
3340 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003341
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003342 return objp;
3343}
Christoph Lametere498be72005-09-09 13:03:32 -07003344
3345/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003346 * Caller needs to acquire correct kmem_cache_node's list_lock
Christoph Lametere498be72005-09-09 13:03:32 -07003347 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003348static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003349 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350{
3351 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003352 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353
3354 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003355 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003356 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357
Mel Gorman072bb0a2012-07-31 16:43:58 -07003358 clear_obj_pfmemalloc(&objpp[i]);
3359 objp = objpp[i];
3360
Joonsoo Kim8456a642013-10-24 10:07:49 +09003361 page = virt_to_head_page(objp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003362 n = cachep->node[node];
Joonsoo Kim8456a642013-10-24 10:07:49 +09003363 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003364 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003365 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003367 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368
3369 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003370 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003371 if (n->free_objects > n->free_limit) {
3372 n->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003373 /* No need to drop any previously held
3374 * lock here, even if we have a off-slab slab
3375 * descriptor it is guaranteed to come from
3376 * a different cache, refer to comments before
3377 * alloc_slabmgmt.
3378 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003379 slab_destroy(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003381 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 }
3383 } else {
3384 /* Unconditionally move a slab to the end of the
3385 * partial list on free - maximum time for the
3386 * other objects to be freed, too.
3387 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003388 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389 }
3390 }
3391}
3392
Pekka Enberg343e0d72006-02-01 03:05:50 -08003393static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394{
3395 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003396 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003397 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398
3399 batchcount = ac->batchcount;
3400#if DEBUG
3401 BUG_ON(!batchcount || batchcount > ac->avail);
3402#endif
3403 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003404 n = cachep->node[node];
3405 spin_lock(&n->list_lock);
3406 if (n->shared) {
3407 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003408 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 if (max) {
3410 if (batchcount > max)
3411 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003412 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003413 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 shared_array->avail += batchcount;
3415 goto free_done;
3416 }
3417 }
3418
Christoph Lameterff694162005-09-22 21:44:02 -07003419 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003420free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421#if STATS
3422 {
3423 int i = 0;
3424 struct list_head *p;
3425
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003426 p = n->slabs_free.next;
3427 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003428 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429
Joonsoo Kim8456a642013-10-24 10:07:49 +09003430 page = list_entry(p, struct page, lru);
3431 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432
3433 i++;
3434 p = p->next;
3435 }
3436 STATS_SET_FREEABLE(cachep, i);
3437 }
3438#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003439 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003441 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442}
3443
3444/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003445 * Release an obj back to its cache. If the obj has a constructed state, it must
3446 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003448static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003449 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003451 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452
3453 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003454 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003455 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003457 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003458
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003459 /*
3460 * Skip calling cache_free_alien() when the platform is not numa.
3461 * This will avoid cache misses that happen while accessing slabp (which
3462 * is per page memory reference) to get nodeid. Instead use a global
3463 * variable to skip the call, which is mostly likely to be present in
3464 * the cache.
3465 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003466 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003467 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003468
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 if (likely(ac->avail < ac->limit)) {
3470 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 } else {
3472 STATS_INC_FREEMISS(cachep);
3473 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003475
Mel Gorman072bb0a2012-07-31 16:43:58 -07003476 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477}
3478
3479/**
3480 * kmem_cache_alloc - Allocate an object
3481 * @cachep: The cache to allocate from.
3482 * @flags: See kmalloc().
3483 *
3484 * Allocate an object from this cache. The flags are only relevant
3485 * if the cache has no available objects.
3486 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003487void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003489 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003490
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003491 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003492 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003493
3494 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495}
3496EXPORT_SYMBOL(kmem_cache_alloc);
3497
Li Zefan0f24f122009-12-11 15:45:30 +08003498#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003499void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003500kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003501{
Steven Rostedt85beb582010-11-24 16:23:34 -05003502 void *ret;
3503
Ezequiel Garcia48356302012-09-08 17:47:57 -03003504 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003505
3506 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003507 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003508 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003509}
Steven Rostedt85beb582010-11-24 16:23:34 -05003510EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003511#endif
3512
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003514/**
3515 * kmem_cache_alloc_node - Allocate an object on the specified node
3516 * @cachep: The cache to allocate from.
3517 * @flags: See kmalloc().
3518 * @nodeid: node number of the target node.
3519 *
3520 * Identical to kmem_cache_alloc but it will allocate memory on the given
3521 * node, which can improve the performance for cpu bound structures.
3522 *
3523 * Fallback to other node is possible if __GFP_THISNODE is not set.
3524 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003525void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3526{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003527 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003528
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003529 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003530 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003531 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003532
3533 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003534}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535EXPORT_SYMBOL(kmem_cache_alloc_node);
3536
Li Zefan0f24f122009-12-11 15:45:30 +08003537#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003538void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003539 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003540 int nodeid,
3541 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003542{
Steven Rostedt85beb582010-11-24 16:23:34 -05003543 void *ret;
3544
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003545 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003546
Steven Rostedt85beb582010-11-24 16:23:34 -05003547 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003548 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003549 flags, nodeid);
3550 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003551}
Steven Rostedt85beb582010-11-24 16:23:34 -05003552EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003553#endif
3554
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003555static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003556__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003557{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003558 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003559
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003560 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003561 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3562 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003563 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003564}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003565
Li Zefan0bb38a52009-12-11 15:45:50 +08003566#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003567void *__kmalloc_node(size_t size, gfp_t flags, int node)
3568{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003569 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003570}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003571EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003572
3573void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003574 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003575{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003576 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003577}
3578EXPORT_SYMBOL(__kmalloc_node_track_caller);
3579#else
3580void *__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, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003583}
3584EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003585#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003586#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
3588/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003589 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003591 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003592 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003594static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003595 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003597 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003598 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003600 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003601 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3602 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003603 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003604
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003605 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003606 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003607
3608 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003609}
3610
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003611
Li Zefan0bb38a52009-12-11 15:45:50 +08003612#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003613void *__kmalloc(size_t size, gfp_t flags)
3614{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003615 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616}
3617EXPORT_SYMBOL(__kmalloc);
3618
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003619void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003620{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003621 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003622}
3623EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003624
3625#else
3626void *__kmalloc(size_t size, gfp_t flags)
3627{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003628 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003629}
3630EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003631#endif
3632
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633/**
3634 * kmem_cache_free - Deallocate an object
3635 * @cachep: The cache the allocation was from.
3636 * @objp: The previously allocated object.
3637 *
3638 * Free an object which was previously allocated from this
3639 * cache.
3640 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003641void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642{
3643 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003644 cachep = cache_from_obj(cachep, objp);
3645 if (!cachep)
3646 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647
3648 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003649 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003650 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003651 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003652 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003654
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003655 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656}
3657EXPORT_SYMBOL(kmem_cache_free);
3658
3659/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 * kfree - free previously allocated memory
3661 * @objp: pointer returned by kmalloc.
3662 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003663 * If @objp is NULL, no operation is performed.
3664 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 * Don't free memory not originally allocated by kmalloc()
3666 * or you will run into trouble.
3667 */
3668void kfree(const void *objp)
3669{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003670 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671 unsigned long flags;
3672
Pekka Enberg2121db72009-03-25 11:05:57 +02003673 trace_kfree(_RET_IP_, objp);
3674
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003675 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 return;
3677 local_irq_save(flags);
3678 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003679 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003680 debug_check_no_locks_freed(objp, c->object_size);
3681
3682 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003683 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684 local_irq_restore(flags);
3685}
3686EXPORT_SYMBOL(kfree);
3687
Christoph Lametere498be72005-09-09 13:03:32 -07003688/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003689 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003690 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003691static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003692{
3693 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003694 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003695 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003696 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003697
Mel Gorman9c09a952008-01-24 05:49:54 -08003698 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003699
Paul Menage3395ee02006-12-06 20:32:16 -08003700 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003701 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003702 if (!new_alien)
3703 goto fail;
3704 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003705
Eric Dumazet63109842007-05-06 14:49:28 -07003706 new_shared = NULL;
3707 if (cachep->shared) {
3708 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003709 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003710 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003711 if (!new_shared) {
3712 free_alien_cache(new_alien);
3713 goto fail;
3714 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003715 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003716
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003717 n = cachep->node[node];
3718 if (n) {
3719 struct array_cache *shared = n->shared;
Christoph Lametercafeb022006-03-25 03:06:46 -08003720
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003721 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003722
Christoph Lametercafeb022006-03-25 03:06:46 -08003723 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003724 free_block(cachep, shared->entry,
3725 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003726
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003727 n->shared = new_shared;
3728 if (!n->alien) {
3729 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003730 new_alien = NULL;
3731 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003732 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003733 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003734 spin_unlock_irq(&n->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003735 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003736 free_alien_cache(new_alien);
3737 continue;
3738 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003739 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3740 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003741 free_alien_cache(new_alien);
3742 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003743 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003744 }
Christoph Lametere498be72005-09-09 13:03:32 -07003745
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003746 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003747 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3748 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003749 n->shared = new_shared;
3750 n->alien = new_alien;
3751 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003752 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003753 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003754 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003755 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003756
Andrew Mortona737b3e2006-03-22 00:08:11 -08003757fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003758 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003759 /* Cache is not active yet. Roll back what we did */
3760 node--;
3761 while (node >= 0) {
Christoph Lameter6a673682013-01-10 19:14:19 +00003762 if (cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003763 n = cachep->node[node];
Christoph Lameter0718dc22006-03-25 03:06:47 -08003764
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003765 kfree(n->shared);
3766 free_alien_cache(n->alien);
3767 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003768 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003769 }
3770 node--;
3771 }
3772 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003773 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003774}
3775
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003777 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003778 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779};
3780
3781static void do_ccupdate_local(void *info)
3782{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003783 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784 struct array_cache *old;
3785
3786 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003787 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003788
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3790 new->new[smp_processor_id()] = old;
3791}
3792
Christoph Lameter18004c52012-07-06 15:25:12 -05003793/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003794static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003795 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003797 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003798 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003800 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3801 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003802 if (!new)
3803 return -ENOMEM;
3804
Christoph Lametere498be72005-09-09 13:03:32 -07003805 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003806 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003807 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003808 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003809 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003810 kfree(new->new[i]);
3811 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003812 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 }
3814 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003815 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003817 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003818
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003820 cachep->batchcount = batchcount;
3821 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003822 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823
Christoph Lametere498be72005-09-09 13:03:32 -07003824 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003825 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 if (!ccold)
3827 continue;
Christoph Lameter6a673682013-01-10 19:14:19 +00003828 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003829 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
Christoph Lameter6a673682013-01-10 19:14:19 +00003830 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831 kfree(ccold);
3832 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003833 kfree(new);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003834 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835}
3836
Glauber Costa943a4512012-12-18 14:23:03 -08003837static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3838 int batchcount, int shared, gfp_t gfp)
3839{
3840 int ret;
3841 struct kmem_cache *c = NULL;
3842 int i = 0;
3843
3844 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3845
3846 if (slab_state < FULL)
3847 return ret;
3848
3849 if ((ret < 0) || !is_root_cache(cachep))
3850 return ret;
3851
Glauber Costaebe945c2012-12-18 14:23:10 -08003852 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003853 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003854 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003855 if (c)
3856 /* return value determined by the parent cache only */
3857 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3858 }
3859
3860 return ret;
3861}
3862
Christoph Lameter18004c52012-07-06 15:25:12 -05003863/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003864static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865{
3866 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003867 int limit = 0;
3868 int shared = 0;
3869 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870
Glauber Costa943a4512012-12-18 14:23:03 -08003871 if (!is_root_cache(cachep)) {
3872 struct kmem_cache *root = memcg_root_cache(cachep);
3873 limit = root->limit;
3874 shared = root->shared;
3875 batchcount = root->batchcount;
3876 }
3877
3878 if (limit && shared && batchcount)
3879 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003880 /*
3881 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882 * - create a LIFO ordering, i.e. return objects that are cache-warm
3883 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003884 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885 * bufctl chains: array operations are cheaper.
3886 * The numbers are guessed, we should auto-tune as described by
3887 * Bonwick.
3888 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003889 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003891 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003893 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003895 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896 limit = 54;
3897 else
3898 limit = 120;
3899
Andrew Mortona737b3e2006-03-22 00:08:11 -08003900 /*
3901 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 * allocation behaviour: Most allocs on one cpu, most free operations
3903 * on another cpu. For these cases, an efficient object passing between
3904 * cpus is necessary. This is provided by a shared array. The array
3905 * replaces Bonwick's magazine layer.
3906 * On uniprocessor, it's functionally equivalent (but less efficient)
3907 * to a larger limit. Thus disabled by default.
3908 */
3909 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003910 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912
3913#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003914 /*
3915 * With debugging enabled, large batchcount lead to excessively long
3916 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 */
3918 if (limit > 32)
3919 limit = 32;
3920#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003921 batchcount = (limit + 1) / 2;
3922skip_setup:
3923 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 if (err)
3925 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003926 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003927 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928}
3929
Christoph Lameter1b552532006-03-22 00:09:07 -08003930/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003931 * Drain an array if it contains any elements taking the node lock only if
3932 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003933 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003934 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003935static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003936 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937{
3938 int tofree;
3939
Christoph Lameter1b552532006-03-22 00:09:07 -08003940 if (!ac || !ac->avail)
3941 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 if (ac->touched && !force) {
3943 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003944 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003945 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003946 if (ac->avail) {
3947 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3948 if (tofree > ac->avail)
3949 tofree = (ac->avail + 1) / 2;
3950 free_block(cachep, ac->entry, tofree, node);
3951 ac->avail -= tofree;
3952 memmove(ac->entry, &(ac->entry[tofree]),
3953 sizeof(void *) * ac->avail);
3954 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003955 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 }
3957}
3958
3959/**
3960 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003961 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 *
3963 * Called from workqueue/eventd every few seconds.
3964 * Purpose:
3965 * - clear the per-cpu caches for this CPU.
3966 * - return freeable pages to the main free memory pool.
3967 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003968 * If we cannot acquire the cache chain mutex then just give up - we'll try
3969 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003971static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003973 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003974 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003975 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003976 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977
Christoph Lameter18004c52012-07-06 15:25:12 -05003978 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003980 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981
Christoph Lameter18004c52012-07-06 15:25:12 -05003982 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 check_irq_on();
3984
Christoph Lameter35386e32006-03-22 00:09:05 -08003985 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003986 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003987 * have established with reasonable certainty that
3988 * we can do some work if the lock was obtained.
3989 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003990 n = searchp->node[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003991
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003992 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003994 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995
Christoph Lameter35386e32006-03-22 00:09:05 -08003996 /*
3997 * These are racy checks but it does not matter
3998 * if we skip one check or scan twice.
3999 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004000 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004001 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004003 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004005 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004007 if (n->free_touched)
4008 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004009 else {
4010 int freed;
4011
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004012 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004013 5 * searchp->num - 1) / (5 * searchp->num));
4014 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004016next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 cond_resched();
4018 }
4019 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004020 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004021 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004022out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004023 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004024 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025}
4026
Linus Torvalds158a9622008-01-02 13:04:48 -08004027#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004028void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004030 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004031 unsigned long active_objs;
4032 unsigned long num_objs;
4033 unsigned long active_slabs = 0;
4034 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004035 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004037 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004038 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 active_objs = 0;
4041 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004042 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004043 n = cachep->node[node];
4044 if (!n)
Christoph Lametere498be72005-09-09 13:03:32 -07004045 continue;
4046
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004047 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004048 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004049
Joonsoo Kim8456a642013-10-24 10:07:49 +09004050 list_for_each_entry(page, &n->slabs_full, lru) {
4051 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004052 error = "slabs_full accounting error";
4053 active_objs += cachep->num;
4054 active_slabs++;
4055 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004056 list_for_each_entry(page, &n->slabs_partial, lru) {
4057 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004058 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004059 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004060 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004061 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004062 active_slabs++;
4063 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004064 list_for_each_entry(page, &n->slabs_free, lru) {
4065 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004066 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004067 num_slabs++;
4068 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004069 free_objects += n->free_objects;
4070 if (n->shared)
4071 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004072
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004073 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004075 num_slabs += active_slabs;
4076 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004077 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 error = "free_objects accounting error";
4079
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004080 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 if (error)
4082 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4083
Glauber Costa0d7561c2012-10-19 18:20:27 +04004084 sinfo->active_objs = active_objs;
4085 sinfo->num_objs = num_objs;
4086 sinfo->active_slabs = active_slabs;
4087 sinfo->num_slabs = num_slabs;
4088 sinfo->shared_avail = shared_avail;
4089 sinfo->limit = cachep->limit;
4090 sinfo->batchcount = cachep->batchcount;
4091 sinfo->shared = cachep->shared;
4092 sinfo->objects_per_slab = cachep->num;
4093 sinfo->cache_order = cachep->gfporder;
4094}
4095
4096void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4097{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004099 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100 unsigned long high = cachep->high_mark;
4101 unsigned long allocs = cachep->num_allocations;
4102 unsigned long grown = cachep->grown;
4103 unsigned long reaped = cachep->reaped;
4104 unsigned long errors = cachep->errors;
4105 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004107 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004108 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109
Joe Perchese92dd4f2010-03-26 19:27:58 -07004110 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4111 "%4lu %4lu %4lu %4lu %4lu",
4112 allocs, high, grown,
4113 reaped, errors, max_freeable, node_allocs,
4114 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115 }
4116 /* cpu stats */
4117 {
4118 unsigned long allochit = atomic_read(&cachep->allochit);
4119 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4120 unsigned long freehit = atomic_read(&cachep->freehit);
4121 unsigned long freemiss = atomic_read(&cachep->freemiss);
4122
4123 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004124 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125 }
4126#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127}
4128
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129#define MAX_SLABINFO_WRITE 128
4130/**
4131 * slabinfo_write - Tuning for the slab allocator
4132 * @file: unused
4133 * @buffer: user buffer
4134 * @count: data length
4135 * @ppos: unused
4136 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004137ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004138 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004140 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004141 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004142 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004143
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144 if (count > MAX_SLABINFO_WRITE)
4145 return -EINVAL;
4146 if (copy_from_user(&kbuf, buffer, count))
4147 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004148 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149
4150 tmp = strchr(kbuf, ' ');
4151 if (!tmp)
4152 return -EINVAL;
4153 *tmp = '\0';
4154 tmp++;
4155 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4156 return -EINVAL;
4157
4158 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004159 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004161 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004163 if (limit < 1 || batchcount < 1 ||
4164 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004165 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004167 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004168 batchcount, shared,
4169 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170 }
4171 break;
4172 }
4173 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004174 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175 if (res >= 0)
4176 res = count;
4177 return res;
4178}
Al Viro871751e2006-03-25 03:06:39 -08004179
4180#ifdef CONFIG_DEBUG_SLAB_LEAK
4181
4182static void *leaks_start(struct seq_file *m, loff_t *pos)
4183{
Christoph Lameter18004c52012-07-06 15:25:12 -05004184 mutex_lock(&slab_mutex);
4185 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004186}
4187
4188static inline int add_caller(unsigned long *n, unsigned long v)
4189{
4190 unsigned long *p;
4191 int l;
4192 if (!v)
4193 return 1;
4194 l = n[1];
4195 p = n + 2;
4196 while (l) {
4197 int i = l/2;
4198 unsigned long *q = p + 2 * i;
4199 if (*q == v) {
4200 q[1]++;
4201 return 1;
4202 }
4203 if (*q > v) {
4204 l = i;
4205 } else {
4206 p = q + 2;
4207 l -= i + 1;
4208 }
4209 }
4210 if (++n[1] == n[0])
4211 return 0;
4212 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4213 p[0] = v;
4214 p[1] = 1;
4215 return 1;
4216}
4217
Joonsoo Kim8456a642013-10-24 10:07:49 +09004218static void handle_slab(unsigned long *n, struct kmem_cache *c,
4219 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004220{
4221 void *p;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004222 int i, j;
4223
Al Viro871751e2006-03-25 03:06:39 -08004224 if (n[0] == n[1])
4225 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004226 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004227 bool active = true;
4228
Joonsoo Kim8456a642013-10-24 10:07:49 +09004229 for (j = page->active; j < c->num; j++) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004230 /* Skip freed item */
Joonsoo Kime5c58df2013-12-02 17:49:40 +09004231 if (get_free_obj(page, j) == i) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004232 active = false;
4233 break;
4234 }
4235 }
4236 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004237 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004238
Al Viro871751e2006-03-25 03:06:39 -08004239 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4240 return;
4241 }
4242}
4243
4244static void show_symbol(struct seq_file *m, unsigned long address)
4245{
4246#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004247 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004248 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004249
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004250 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004251 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004252 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004253 seq_printf(m, " [%s]", modname);
4254 return;
4255 }
4256#endif
4257 seq_printf(m, "%p", (void *)address);
4258}
4259
4260static int leaks_show(struct seq_file *m, void *p)
4261{
Thierry Reding0672aa72012-06-22 19:42:49 +02004262 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004263 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004264 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004265 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004266 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004267 int node;
4268 int i;
4269
4270 if (!(cachep->flags & SLAB_STORE_USER))
4271 return 0;
4272 if (!(cachep->flags & SLAB_RED_ZONE))
4273 return 0;
4274
4275 /* OK, we can do it */
4276
Christoph Lameterdb845062013-02-05 18:45:23 +00004277 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004278
4279 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004280 n = cachep->node[node];
4281 if (!n)
Al Viro871751e2006-03-25 03:06:39 -08004282 continue;
4283
4284 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004285 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004286
Joonsoo Kim8456a642013-10-24 10:07:49 +09004287 list_for_each_entry(page, &n->slabs_full, lru)
4288 handle_slab(x, cachep, page);
4289 list_for_each_entry(page, &n->slabs_partial, lru)
4290 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004291 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004292 }
4293 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004294 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004295 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004296 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004297 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004298 if (!m->private) {
4299 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004300 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004301 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004302 return -ENOMEM;
4303 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004304 *(unsigned long *)m->private = x[0] * 2;
4305 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004306 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004307 /* Now make sure this entry will be retried */
4308 m->count = m->size;
4309 return 0;
4310 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004311 for (i = 0; i < x[1]; i++) {
4312 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4313 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004314 seq_putc(m, '\n');
4315 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004316
Al Viro871751e2006-03-25 03:06:39 -08004317 return 0;
4318}
4319
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004320static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004321 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004322 .next = slab_next,
4323 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004324 .show = leaks_show,
4325};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004326
4327static int slabstats_open(struct inode *inode, struct file *file)
4328{
4329 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4330 int ret = -ENOMEM;
4331 if (n) {
4332 ret = seq_open(file, &slabstats_op);
4333 if (!ret) {
4334 struct seq_file *m = file->private_data;
4335 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4336 m->private = n;
4337 n = NULL;
4338 }
4339 kfree(n);
4340 }
4341 return ret;
4342}
4343
4344static const struct file_operations proc_slabstats_operations = {
4345 .open = slabstats_open,
4346 .read = seq_read,
4347 .llseek = seq_lseek,
4348 .release = seq_release_private,
4349};
Al Viro871751e2006-03-25 03:06:39 -08004350#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004351
4352static int __init slab_proc_init(void)
4353{
4354#ifdef CONFIG_DEBUG_SLAB_LEAK
4355 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4356#endif
4357 return 0;
4358}
4359module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004360#endif
4361
Manfred Spraul00e145b2005-09-03 15:55:07 -07004362/**
4363 * ksize - get the actual amount of memory allocated for a given object
4364 * @objp: Pointer to the object
4365 *
4366 * kmalloc may internally round up allocations and return more memory
4367 * than requested. ksize() can be used to determine the actual amount of
4368 * memory allocated. The caller may use this additional memory, even though
4369 * a smaller amount of memory was initially specified with the kmalloc call.
4370 * The caller must guarantee that objp points to a valid object previously
4371 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4372 * must not be freed during the duration of the call.
4373 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004374size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004376 BUG_ON(!objp);
4377 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004378 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004379
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004380 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004382EXPORT_SYMBOL(ksize);