blob: b264214c77ead2f8426ae476354231dc7c52169a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Christoph Lameter18004c52012-07-06 15:25:12 -050071 * The global cache-chain is protected by the mutex 'slab_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700118#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
Mel Gorman381760e2012-07-31 16:44:30 -0700120#include <net/sock.h>
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500126#include <trace/events/kmem.h>
127
Mel Gorman072bb0a2012-07-31 16:43:58 -0700128#include "internal.h"
129
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800130#include "slab.h"
131
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
Mel Gorman072bb0a2012-07-31 16:43:58 -0700160/*
161 * true if a page was allocated from pfmemalloc reserves for network-based
162 * swap
163 */
164static bool pfmemalloc_active __read_mostly;
165
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 * struct array_cache
168 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 * Purpose:
170 * - LIFO ordering, to hand out cache-warm objects from _alloc
171 * - reduce the number of linked list operations
172 * - reduce spinlock operations
173 *
174 * The limit is stored in the per-cpu structure to reduce the data cache
175 * footprint.
176 *
177 */
178struct array_cache {
179 unsigned int avail;
180 unsigned int limit;
181 unsigned int batchcount;
182 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700183 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700184 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800185 * Must have this definition in here for the proper
186 * alignment of array_cache. Also simplifies accessing
187 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700188 *
189 * Entries should not be directly dereferenced as
190 * entries belonging to slabs marked pfmemalloc will
191 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800192 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193};
194
Mel Gorman072bb0a2012-07-31 16:43:58 -0700195#define SLAB_OBJ_PFMEMALLOC 1
196static inline bool is_obj_pfmemalloc(void *objp)
197{
198 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
199}
200
201static inline void set_obj_pfmemalloc(void **objp)
202{
203 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
204 return;
205}
206
207static inline void clear_obj_pfmemalloc(void **objp)
208{
209 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
210}
211
Andrew Mortona737b3e2006-03-22 00:08:11 -0800212/*
213 * bootstrap: The caches do not work without cpuarrays anymore, but the
214 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 */
216#define BOOT_CPUCACHE_ENTRIES 1
217struct arraycache_init {
218 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800219 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220};
221
222/*
Christoph Lametere498be72005-09-09 13:03:32 -0700223 * Need this for bootstrapping a per node allocator.
224 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200225#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000226static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700227#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200228#define SIZE_AC MAX_NUMNODES
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000229#define SIZE_NODE (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
Christoph Lametered11d9e2006-06-30 01:55:45 -0700231static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000232 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700233static void free_block(struct kmem_cache *cachep, void **objpp, int len,
234 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300235static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000236static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700237
Ingo Molnare0a42722006-06-23 02:03:46 -0700238static int slab_early_init = 1;
239
Christoph Lametere3366012013-01-10 19:14:18 +0000240#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000241#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700242
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000243static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700244{
245 INIT_LIST_HEAD(&parent->slabs_full);
246 INIT_LIST_HEAD(&parent->slabs_partial);
247 INIT_LIST_HEAD(&parent->slabs_free);
248 parent->shared = NULL;
249 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800250 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700251 spin_lock_init(&parent->list_lock);
252 parent->free_objects = 0;
253 parent->free_touched = 0;
254}
255
Andrew Mortona737b3e2006-03-22 00:08:11 -0800256#define MAKE_LIST(cachep, listp, slab, nodeid) \
257 do { \
258 INIT_LIST_HEAD(listp); \
Christoph Lameter6a673682013-01-10 19:14:19 +0000259 list_splice(&(cachep->node[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700260 } while (0)
261
Andrew Mortona737b3e2006-03-22 00:08:11 -0800262#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
263 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700264 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
265 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
266 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
267 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269#define CFLGS_OFF_SLAB (0x80000000UL)
270#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
271
272#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273/*
274 * Optimization question: fewer reaps means less probability for unnessary
275 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100277 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 * which could lock up otherwise freeable slabs.
279 */
280#define REAPTIMEOUT_CPUC (2*HZ)
281#define REAPTIMEOUT_LIST3 (4*HZ)
282
283#if STATS
284#define STATS_INC_ACTIVE(x) ((x)->num_active++)
285#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
286#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
287#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700288#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800289#define STATS_SET_HIGH(x) \
290 do { \
291 if ((x)->num_active > (x)->high_mark) \
292 (x)->high_mark = (x)->num_active; \
293 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294#define STATS_INC_ERR(x) ((x)->errors++)
295#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700296#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700297#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800298#define STATS_SET_FREEABLE(x, i) \
299 do { \
300 if ((x)->max_freeable < i) \
301 (x)->max_freeable = i; \
302 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
304#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
305#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
306#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
307#else
308#define STATS_INC_ACTIVE(x) do { } while (0)
309#define STATS_DEC_ACTIVE(x) do { } while (0)
310#define STATS_INC_ALLOCED(x) do { } while (0)
311#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700312#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313#define STATS_SET_HIGH(x) do { } while (0)
314#define STATS_INC_ERR(x) do { } while (0)
315#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700316#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700317#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800318#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319#define STATS_INC_ALLOCHIT(x) do { } while (0)
320#define STATS_INC_ALLOCMISS(x) do { } while (0)
321#define STATS_INC_FREEHIT(x) do { } while (0)
322#define STATS_INC_FREEMISS(x) do { } while (0)
323#endif
324
325#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326
Andrew Mortona737b3e2006-03-22 00:08:11 -0800327/*
328 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800330 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 * the end of an object is aligned with the end of the real
332 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800333 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800335 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500336 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
337 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800338 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800340static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800342 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343}
344
David Woodhouseb46b8f12007-05-08 00:22:59 -0700345static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346{
347 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700348 return (unsigned long long*) (objp + obj_offset(cachep) -
349 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350}
351
David Woodhouseb46b8f12007-05-08 00:22:59 -0700352static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
354 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
355 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500356 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700357 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400358 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500359 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700360 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361}
362
Pekka Enberg343e0d72006-02-01 03:05:50 -0800363static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
365 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500366 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367}
368
369#else
370
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800371#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700372#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
373#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
375
376#endif
377
378/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700379 * Do not go above this order unless 0 objects fit into the slab or
380 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 */
David Rientjes543585c2011-10-18 22:09:24 -0700382#define SLAB_MAX_ORDER_HI 1
383#define SLAB_MAX_ORDER_LO 0
384static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700385static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800387static inline struct kmem_cache *virt_to_cache(const void *obj)
388{
Christoph Lameterb49af682007-05-06 14:49:41 -0700389 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500390 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800391}
392
Joonsoo Kim8456a642013-10-24 10:07:49 +0900393static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800394 unsigned int idx)
395{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900396 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800397}
398
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800399/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500400 * We want to avoid an expensive divide : (offset / cache->size)
401 * Using the fact that size is a constant for a particular cache,
402 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800403 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
404 */
405static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900406 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800407{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900408 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800409 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800410}
411
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800413 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
415/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000416static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800417 .batchcount = 1,
418 .limit = BOOT_CPUCACHE_ENTRIES,
419 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500420 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800421 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422};
423
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700424#define BAD_ALIEN_MAGIC 0x01020304ul
425
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200426#ifdef CONFIG_LOCKDEP
427
428/*
429 * Slab sometimes uses the kmalloc slabs to store the slab headers
430 * for other slabs "off slab".
431 * The locking for this is tricky in that it nests within the locks
432 * of all other slabs in a few places; to deal with this special
433 * locking we put on-slab caches into a separate lock-class.
434 *
435 * We set lock class for alien array caches which are up during init.
436 * The lock annotation will be lost if all cpus of a node goes down and
437 * then comes back up during hotplug
438 */
439static struct lock_class_key on_slab_l3_key;
440static struct lock_class_key on_slab_alc_key;
441
Peter Zijlstra83835b32011-07-22 15:26:05 +0200442static struct lock_class_key debugobj_l3_key;
443static struct lock_class_key debugobj_alc_key;
444
445static void slab_set_lock_classes(struct kmem_cache *cachep,
446 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
447 int q)
448{
449 struct array_cache **alc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000450 struct kmem_cache_node *n;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200451 int r;
452
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000453 n = cachep->node[q];
454 if (!n)
Peter Zijlstra83835b32011-07-22 15:26:05 +0200455 return;
456
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000457 lockdep_set_class(&n->list_lock, l3_key);
458 alc = n->alien;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200459 /*
460 * FIXME: This check for BAD_ALIEN_MAGIC
461 * should go away when common slab code is taught to
462 * work even without alien caches.
463 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
464 * for alloc_alien_cache,
465 */
466 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
467 return;
468 for_each_node(r) {
469 if (alc[r])
470 lockdep_set_class(&alc[r]->lock, alc_key);
471 }
472}
473
474static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
475{
476 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
477}
478
479static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
480{
481 int node;
482
483 for_each_online_node(node)
484 slab_set_debugobj_lock_classes_node(cachep, node);
485}
486
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200487static void init_node_lock_keys(int q)
488{
Christoph Lametere3366012013-01-10 19:14:18 +0000489 int i;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200490
Christoph Lameter97d06602012-07-06 15:25:11 -0500491 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200492 return;
493
Christoph Lameter0f8f8092013-07-02 12:12:10 -0700494 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000495 struct kmem_cache_node *n;
Christoph Lametere3366012013-01-10 19:14:18 +0000496 struct kmem_cache *cache = kmalloc_caches[i];
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200497
Christoph Lametere3366012013-01-10 19:14:18 +0000498 if (!cache)
Pekka Enberg00afa752009-12-27 14:33:14 +0200499 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200500
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000501 n = cache->node[q];
502 if (!n || OFF_SLAB(cache))
Christoph Lametere3366012013-01-10 19:14:18 +0000503 continue;
504
505 slab_set_lock_classes(cache, &on_slab_l3_key,
Peter Zijlstra83835b32011-07-22 15:26:05 +0200506 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200507 }
508}
509
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800510static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
511{
Christoph Lameter6a673682013-01-10 19:14:19 +0000512 if (!cachep->node[q])
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800513 return;
514
515 slab_set_lock_classes(cachep, &on_slab_l3_key,
516 &on_slab_alc_key, q);
517}
518
519static inline void on_slab_lock_classes(struct kmem_cache *cachep)
520{
521 int node;
522
523 VM_BUG_ON(OFF_SLAB(cachep));
524 for_each_node(node)
525 on_slab_lock_classes_node(cachep, node);
526}
527
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200528static inline void init_lock_keys(void)
529{
530 int node;
531
532 for_each_node(node)
533 init_node_lock_keys(node);
534}
535#else
536static void init_node_lock_keys(int q)
537{
538}
539
540static inline void init_lock_keys(void)
541{
542}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200543
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800544static inline void on_slab_lock_classes(struct kmem_cache *cachep)
545{
546}
547
548static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
549{
550}
551
Peter Zijlstra83835b32011-07-22 15:26:05 +0200552static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
553{
554}
555
556static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
557{
558}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200559#endif
560
Tejun Heo1871e522009-10-29 22:34:13 +0900561static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562
Pekka Enberg343e0d72006-02-01 03:05:50 -0800563static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564{
565 return cachep->array[smp_processor_id()];
566}
567
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800568static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900570 return ALIGN(nr_objs * sizeof(unsigned int), align);
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800571}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572
Andrew Mortona737b3e2006-03-22 00:08:11 -0800573/*
574 * Calculate the number of objects and left-over bytes for a given buffer size.
575 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800576static void cache_estimate(unsigned long gfporder, size_t buffer_size,
577 size_t align, int flags, size_t *left_over,
578 unsigned int *num)
579{
580 int nr_objs;
581 size_t mgmt_size;
582 size_t slab_size = PAGE_SIZE << gfporder;
583
584 /*
585 * The slab management structure can be either off the slab or
586 * on it. For the latter case, the memory allocated for a
587 * slab is used for:
588 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900589 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800590 * - Padding to respect alignment of @align
591 * - @buffer_size bytes for each object
592 *
593 * If the slab management structure is off the slab, then the
594 * alignment will already be calculated into the size. Because
595 * the slabs are all pages aligned, the objects will be at the
596 * correct alignment when allocated.
597 */
598 if (flags & CFLGS_OFF_SLAB) {
599 mgmt_size = 0;
600 nr_objs = slab_size / buffer_size;
601
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800602 } else {
603 /*
604 * Ignore padding for the initial guess. The padding
605 * is at most @align-1 bytes, and @buffer_size is at
606 * least @align. In the worst case, this result will
607 * be one greater than the number of objects that fit
608 * into the memory allocation when taking the padding
609 * into account.
610 */
Joonsoo Kim8456a642013-10-24 10:07:49 +0900611 nr_objs = (slab_size) / (buffer_size + sizeof(unsigned int));
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800612
613 /*
614 * This calculated number will be either the right
615 * amount, or one greater than what we want.
616 */
617 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
618 > slab_size)
619 nr_objs--;
620
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800621 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800623 *num = nr_objs;
624 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
626
Christoph Lameterf28510d2012-09-11 19:49:38 +0000627#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700628#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
Andrew Mortona737b3e2006-03-22 00:08:11 -0800630static void __slab_error(const char *function, struct kmem_cache *cachep,
631 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632{
633 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800634 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030636 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000638#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Paul Menage3395ee02006-12-06 20:32:16 -0800640/*
641 * By default on NUMA we use alien caches to stage the freeing of
642 * objects allocated from other nodes. This causes massive memory
643 * inefficiencies when using fake NUMA setup to split memory into a
644 * large number of small nodes, so it can be disabled on the command
645 * line
646 */
647
648static int use_alien_caches __read_mostly = 1;
649static int __init noaliencache_setup(char *s)
650{
651 use_alien_caches = 0;
652 return 1;
653}
654__setup("noaliencache", noaliencache_setup);
655
David Rientjes3df1ccc2011-10-18 22:09:28 -0700656static int __init slab_max_order_setup(char *str)
657{
658 get_option(&str, &slab_max_order);
659 slab_max_order = slab_max_order < 0 ? 0 :
660 min(slab_max_order, MAX_ORDER - 1);
661 slab_max_order_set = true;
662
663 return 1;
664}
665__setup("slab_max_order=", slab_max_order_setup);
666
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800667#ifdef CONFIG_NUMA
668/*
669 * Special reaping functions for NUMA systems called from cache_reap().
670 * These take care of doing round robin flushing of alien caches (containing
671 * objects freed on different nodes from which they were allocated) and the
672 * flushing of remote pcps by calling drain_node_pages.
673 */
Tejun Heo1871e522009-10-29 22:34:13 +0900674static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800675
676static void init_reap_node(int cpu)
677{
678 int node;
679
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700680 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800681 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800682 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800683
Tejun Heo1871e522009-10-29 22:34:13 +0900684 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800685}
686
687static void next_reap_node(void)
688{
Christoph Lameter909ea962010-12-08 16:22:55 +0100689 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800690
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800691 node = next_node(node, node_online_map);
692 if (unlikely(node >= MAX_NUMNODES))
693 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100694 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800695}
696
697#else
698#define init_reap_node(cpu) do { } while (0)
699#define next_reap_node(void) do { } while (0)
700#endif
701
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702/*
703 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
704 * via the workqueue/eventd.
705 * Add the CPU number into the expiration time to minimize the possibility of
706 * the CPUs getting into lockstep and contending for the global cache chain
707 * lock.
708 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400709static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710{
Tejun Heo1871e522009-10-29 22:34:13 +0900711 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712
713 /*
714 * When this gets called from do_initcalls via cpucache_init(),
715 * init_workqueues() has already run, so keventd will be setup
716 * at that time.
717 */
David Howells52bad642006-11-22 14:54:01 +0000718 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800719 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700720 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800721 schedule_delayed_work_on(cpu, reap_work,
722 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 }
724}
725
Christoph Lametere498be72005-09-09 13:03:32 -0700726static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300727 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800729 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 struct array_cache *nc = NULL;
731
Pekka Enberg83b519e2009-06-10 19:40:04 +0300732 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100733 /*
734 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300735 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100736 * cache the pointers are not cleared and they could be counted as
737 * valid references during a kmemleak scan. Therefore, kmemleak must
738 * not scan such objects.
739 */
740 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 if (nc) {
742 nc->avail = 0;
743 nc->limit = entries;
744 nc->batchcount = batchcount;
745 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700746 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 }
748 return nc;
749}
750
Joonsoo Kim8456a642013-10-24 10:07:49 +0900751static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700752{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700753 return PageSlabPfmemalloc(page);
754}
755
756/* Clears pfmemalloc_active if no slabs have pfmalloc set */
757static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
758 struct array_cache *ac)
759{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000760 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
Joonsoo Kim8456a642013-10-24 10:07:49 +0900761 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700762 unsigned long flags;
763
764 if (!pfmemalloc_active)
765 return;
766
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000767 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900768 list_for_each_entry(page, &n->slabs_full, lru)
769 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700770 goto out;
771
Joonsoo Kim8456a642013-10-24 10:07:49 +0900772 list_for_each_entry(page, &n->slabs_partial, lru)
773 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700774 goto out;
775
Joonsoo Kim8456a642013-10-24 10:07:49 +0900776 list_for_each_entry(page, &n->slabs_free, lru)
777 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700778 goto out;
779
780 pfmemalloc_active = false;
781out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000782 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700783}
784
Mel Gorman381760e2012-07-31 16:44:30 -0700785static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700786 gfp_t flags, bool force_refill)
787{
788 int i;
789 void *objp = ac->entry[--ac->avail];
790
791 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
792 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000793 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700794
795 if (gfp_pfmemalloc_allowed(flags)) {
796 clear_obj_pfmemalloc(&objp);
797 return objp;
798 }
799
800 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700801 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700802 /* If a !PFMEMALLOC object is found, swap them */
803 if (!is_obj_pfmemalloc(ac->entry[i])) {
804 objp = ac->entry[i];
805 ac->entry[i] = ac->entry[ac->avail];
806 ac->entry[ac->avail] = objp;
807 return objp;
808 }
809 }
810
811 /*
812 * If there are empty slabs on the slabs_free list and we are
813 * being forced to refill the cache, mark this one !pfmemalloc.
814 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000815 n = cachep->node[numa_mem_id()];
816 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900817 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900818 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700819 clear_obj_pfmemalloc(&objp);
820 recheck_pfmemalloc_active(cachep, ac);
821 return objp;
822 }
823
824 /* No !PFMEMALLOC objects available */
825 ac->avail++;
826 objp = NULL;
827 }
828
829 return objp;
830}
831
Mel Gorman381760e2012-07-31 16:44:30 -0700832static inline void *ac_get_obj(struct kmem_cache *cachep,
833 struct array_cache *ac, gfp_t flags, bool force_refill)
834{
835 void *objp;
836
837 if (unlikely(sk_memalloc_socks()))
838 objp = __ac_get_obj(cachep, ac, flags, force_refill);
839 else
840 objp = ac->entry[--ac->avail];
841
842 return objp;
843}
844
845static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700846 void *objp)
847{
848 if (unlikely(pfmemalloc_active)) {
849 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700850 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700851 if (PageSlabPfmemalloc(page))
852 set_obj_pfmemalloc(&objp);
853 }
854
Mel Gorman381760e2012-07-31 16:44:30 -0700855 return objp;
856}
857
858static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
859 void *objp)
860{
861 if (unlikely(sk_memalloc_socks()))
862 objp = __ac_put_obj(cachep, ac, objp);
863
Mel Gorman072bb0a2012-07-31 16:43:58 -0700864 ac->entry[ac->avail++] = objp;
865}
866
Christoph Lameter3ded1752006-03-25 03:06:44 -0800867/*
868 * Transfer objects in one arraycache to another.
869 * Locking must be handled by the caller.
870 *
871 * Return the number of entries transferred.
872 */
873static int transfer_objects(struct array_cache *to,
874 struct array_cache *from, unsigned int max)
875{
876 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700877 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800878
879 if (!nr)
880 return 0;
881
882 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
883 sizeof(void *) *nr);
884
885 from->avail -= nr;
886 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800887 return nr;
888}
889
Christoph Lameter765c4502006-09-27 01:50:08 -0700890#ifndef CONFIG_NUMA
891
892#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000893#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700894
Pekka Enberg83b519e2009-06-10 19:40:04 +0300895static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700896{
897 return (struct array_cache **)BAD_ALIEN_MAGIC;
898}
899
900static inline void free_alien_cache(struct array_cache **ac_ptr)
901{
902}
903
904static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
905{
906 return 0;
907}
908
909static inline void *alternate_node_alloc(struct kmem_cache *cachep,
910 gfp_t flags)
911{
912 return NULL;
913}
914
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800915static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700916 gfp_t flags, int nodeid)
917{
918 return NULL;
919}
920
921#else /* CONFIG_NUMA */
922
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800923static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800924static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800925
Pekka Enberg83b519e2009-06-10 19:40:04 +0300926static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700927{
928 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800929 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700930 int i;
931
932 if (limit > 1)
933 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +0800934 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700935 if (ac_ptr) {
936 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +0800937 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -0700938 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300939 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700940 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800941 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700942 kfree(ac_ptr[i]);
943 kfree(ac_ptr);
944 return NULL;
945 }
946 }
947 }
948 return ac_ptr;
949}
950
Pekka Enberg5295a742006-02-01 03:05:48 -0800951static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700952{
953 int i;
954
955 if (!ac_ptr)
956 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700957 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800958 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700959 kfree(ac_ptr);
960}
961
Pekka Enberg343e0d72006-02-01 03:05:50 -0800962static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800963 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700964{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000965 struct kmem_cache_node *n = cachep->node[node];
Christoph Lametere498be72005-09-09 13:03:32 -0700966
967 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000968 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800969 /*
970 * Stuff objects into the remote nodes shared array first.
971 * That way we could avoid the overhead of putting the objects
972 * into the free lists and getting them back later.
973 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000974 if (n->shared)
975 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800976
Christoph Lameterff694162005-09-22 21:44:02 -0700977 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700978 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000979 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700980 }
981}
982
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800983/*
984 * Called from cache_reap() to regularly drain alien caches round robin.
985 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000986static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800987{
Christoph Lameter909ea962010-12-08 16:22:55 +0100988 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800989
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000990 if (n->alien) {
991 struct array_cache *ac = n->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -0800992
993 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800994 __drain_alien_cache(cachep, ac, node);
995 spin_unlock_irq(&ac->lock);
996 }
997 }
998}
999
Andrew Mortona737b3e2006-03-22 00:08:11 -08001000static void drain_alien_cache(struct kmem_cache *cachep,
1001 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001002{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001003 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001004 struct array_cache *ac;
1005 unsigned long flags;
1006
1007 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001008 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001009 if (ac) {
1010 spin_lock_irqsave(&ac->lock, flags);
1011 __drain_alien_cache(cachep, ac, i);
1012 spin_unlock_irqrestore(&ac->lock, flags);
1013 }
1014 }
1015}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001016
Ingo Molnar873623d2006-07-13 14:44:38 +02001017static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001018{
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001019 int nodeid = page_to_nid(virt_to_page(objp));
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001020 struct kmem_cache_node *n;
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001021 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001022 int node;
1023
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001024 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001025
1026 /*
1027 * Make sure we are not freeing a object from another node to the array
1028 * cache on this cpu.
1029 */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001030 if (likely(nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001031 return 0;
1032
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001033 n = cachep->node[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001034 STATS_INC_NODEFREES(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001035 if (n->alien && n->alien[nodeid]) {
1036 alien = n->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001037 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001038 if (unlikely(alien->avail == alien->limit)) {
1039 STATS_INC_ACOVERFLOW(cachep);
1040 __drain_alien_cache(cachep, alien, nodeid);
1041 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001042 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001043 spin_unlock(&alien->lock);
1044 } else {
Christoph Lameter6a673682013-01-10 19:14:19 +00001045 spin_lock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001046 free_block(cachep, &objp, 1, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001047 spin_unlock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001048 }
1049 return 1;
1050}
Christoph Lametere498be72005-09-09 13:03:32 -07001051#endif
1052
David Rientjes8f9f8d92010-03-27 19:40:47 -07001053/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001054 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001055 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001056 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001057 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001058 * already in use.
1059 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001060 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001061 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001062static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001063{
1064 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001065 struct kmem_cache_node *n;
Christoph Lameter6744f082013-01-10 19:12:17 +00001066 const int memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001067
Christoph Lameter18004c52012-07-06 15:25:12 -05001068 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001069 /*
1070 * Set up the size64 kmemlist for cpu before we can
1071 * begin anything. Make sure some other cpu on this
1072 * node has not already allocated this
1073 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001074 if (!cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001075 n = kmalloc_node(memsize, GFP_KERNEL, node);
1076 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001077 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001078 kmem_cache_node_init(n);
1079 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
David Rientjes8f9f8d92010-03-27 19:40:47 -07001080 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1081
1082 /*
1083 * The l3s don't come and go as CPUs come and
Christoph Lameter18004c52012-07-06 15:25:12 -05001084 * go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001085 * protection here.
1086 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001087 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001088 }
1089
Christoph Lameter6a673682013-01-10 19:14:19 +00001090 spin_lock_irq(&cachep->node[node]->list_lock);
1091 cachep->node[node]->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001092 (1 + nr_cpus_node(node)) *
1093 cachep->batchcount + cachep->num;
Christoph Lameter6a673682013-01-10 19:14:19 +00001094 spin_unlock_irq(&cachep->node[node]->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001095 }
1096 return 0;
1097}
1098
Wanpeng Li0fa81032013-07-04 08:33:22 +08001099static inline int slabs_tofree(struct kmem_cache *cachep,
1100 struct kmem_cache_node *n)
1101{
1102 return (n->free_objects + cachep->num - 1) / cachep->num;
1103}
1104
Paul Gortmaker0db06282013-06-19 14:53:51 -04001105static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001107 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001108 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001109 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301110 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001111
Christoph Lameter18004c52012-07-06 15:25:12 -05001112 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001113 struct array_cache *nc;
1114 struct array_cache *shared;
1115 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001116
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001117 /* cpu is dead; no one can alloc from it. */
1118 nc = cachep->array[cpu];
1119 cachep->array[cpu] = NULL;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001120 n = cachep->node[node];
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001121
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001122 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001123 goto free_array_cache;
1124
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001126
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001127 /* Free limit for this kmem_cache_node */
1128 n->free_limit -= cachep->batchcount;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001129 if (nc)
1130 free_block(cachep, nc->entry, nc->avail, node);
1131
Rusty Russell58463c12009-12-17 11:43:12 -06001132 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001133 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001134 goto free_array_cache;
1135 }
1136
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001137 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001138 if (shared) {
1139 free_block(cachep, shared->entry,
1140 shared->avail, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001141 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001142 }
1143
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001144 alien = n->alien;
1145 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001146
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001147 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001148
1149 kfree(shared);
1150 if (alien) {
1151 drain_alien_cache(cachep, alien);
1152 free_alien_cache(alien);
1153 }
1154free_array_cache:
1155 kfree(nc);
1156 }
1157 /*
1158 * In the previous loop, all the objects were freed to
1159 * the respective cache's slabs, now we can go ahead and
1160 * shrink each nodelist to its limit.
1161 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001162 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001163 n = cachep->node[node];
1164 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001165 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001166 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001167 }
1168}
1169
Paul Gortmaker0db06282013-06-19 14:53:51 -04001170static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001171{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001172 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001173 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001174 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001175 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001177 /*
1178 * We need to do this right in the beginning since
1179 * alloc_arraycache's are going to use this list.
1180 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001181 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001182 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001183 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001184 if (err < 0)
1185 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001186
1187 /*
1188 * Now we can go ahead with allocating the shared arrays and
1189 * array caches
1190 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001191 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001192 struct array_cache *nc;
1193 struct array_cache *shared = NULL;
1194 struct array_cache **alien = NULL;
1195
1196 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001197 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001198 if (!nc)
1199 goto bad;
1200 if (cachep->shared) {
1201 shared = alloc_arraycache(node,
1202 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001203 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001204 if (!shared) {
1205 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001206 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001207 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001208 }
1209 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001210 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001211 if (!alien) {
1212 kfree(shared);
1213 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001214 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001215 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001216 }
1217 cachep->array[cpu] = nc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001218 n = cachep->node[node];
1219 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001220
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001221 spin_lock_irq(&n->list_lock);
1222 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001223 /*
1224 * We are serialised from CPU_DEAD or
1225 * CPU_UP_CANCELLED by the cpucontrol lock
1226 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001227 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001228 shared = NULL;
1229 }
1230#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001231 if (!n->alien) {
1232 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001233 alien = NULL;
1234 }
1235#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001236 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001237 kfree(shared);
1238 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001239 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1240 slab_set_debugobj_lock_classes_node(cachep, node);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08001241 else if (!OFF_SLAB(cachep) &&
1242 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1243 on_slab_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001244 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001245 init_node_lock_keys(node);
1246
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001247 return 0;
1248bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001249 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001250 return -ENOMEM;
1251}
1252
Paul Gortmaker0db06282013-06-19 14:53:51 -04001253static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001254 unsigned long action, void *hcpu)
1255{
1256 long cpu = (long)hcpu;
1257 int err = 0;
1258
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001260 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001261 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001262 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001263 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001264 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 break;
1266 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001267 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 start_cpu_timer(cpu);
1269 break;
1270#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001271 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001272 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001273 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001274 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001275 * held so that if cache_reap() is invoked it cannot do
1276 * anything expensive but will only modify reap_work
1277 * and reschedule the timer.
1278 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001279 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001280 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001281 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001282 break;
1283 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001284 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001285 start_cpu_timer(cpu);
1286 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001288 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001289 /*
1290 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001291 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001292 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001293 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001294 * structure is usually allocated from kmem_cache_create() and
1295 * gets destroyed at kmem_cache_destroy().
1296 */
Simon Arlott183ff222007-10-20 01:27:18 +02001297 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001298#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001300 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001301 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001302 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001303 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001306 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307}
1308
Paul Gortmaker0db06282013-06-19 14:53:51 -04001309static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001310 &cpuup_callback, NULL, 0
1311};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312
David Rientjes8f9f8d92010-03-27 19:40:47 -07001313#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1314/*
1315 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1316 * Returns -EBUSY if all objects cannot be drained so that the node is not
1317 * removed.
1318 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001319 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001320 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001321static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001322{
1323 struct kmem_cache *cachep;
1324 int ret = 0;
1325
Christoph Lameter18004c52012-07-06 15:25:12 -05001326 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001327 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001328
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001329 n = cachep->node[node];
1330 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001331 continue;
1332
Wanpeng Li0fa81032013-07-04 08:33:22 +08001333 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001334
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001335 if (!list_empty(&n->slabs_full) ||
1336 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001337 ret = -EBUSY;
1338 break;
1339 }
1340 }
1341 return ret;
1342}
1343
1344static int __meminit slab_memory_callback(struct notifier_block *self,
1345 unsigned long action, void *arg)
1346{
1347 struct memory_notify *mnb = arg;
1348 int ret = 0;
1349 int nid;
1350
1351 nid = mnb->status_change_nid;
1352 if (nid < 0)
1353 goto out;
1354
1355 switch (action) {
1356 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001357 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001358 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001359 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001360 break;
1361 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001362 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001363 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001364 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001365 break;
1366 case MEM_ONLINE:
1367 case MEM_OFFLINE:
1368 case MEM_CANCEL_ONLINE:
1369 case MEM_CANCEL_OFFLINE:
1370 break;
1371 }
1372out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001373 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001374}
1375#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1376
Christoph Lametere498be72005-09-09 13:03:32 -07001377/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001378 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001379 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001380static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001381 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001382{
Christoph Lameter6744f082013-01-10 19:12:17 +00001383 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001384
Christoph Lameter6744f082013-01-10 19:12:17 +00001385 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001386 BUG_ON(!ptr);
1387
Christoph Lameter6744f082013-01-10 19:12:17 +00001388 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001389 /*
1390 * Do not assume that spinlocks can be initialized via memcpy:
1391 */
1392 spin_lock_init(&ptr->list_lock);
1393
Christoph Lametere498be72005-09-09 13:03:32 -07001394 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001395 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001396}
1397
Andrew Mortona737b3e2006-03-22 00:08:11 -08001398/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001399 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1400 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001401 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001402static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001403{
1404 int node;
1405
1406 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001407 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001408 cachep->node[node]->next_reap = jiffies +
Pekka Enberg556a1692008-01-25 08:20:51 +02001409 REAPTIMEOUT_LIST3 +
1410 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1411 }
1412}
1413
1414/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001415 * The memory after the last cpu cache pointer is used for the
Christoph Lameter6a673682013-01-10 19:14:19 +00001416 * the node pointer.
Christoph Lameter3c583462012-11-28 16:23:01 +00001417 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001418static void setup_node_pointer(struct kmem_cache *cachep)
Christoph Lameter3c583462012-11-28 16:23:01 +00001419{
Christoph Lameter6a673682013-01-10 19:14:19 +00001420 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
Christoph Lameter3c583462012-11-28 16:23:01 +00001421}
1422
1423/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001424 * Initialisation. Called after the page allocator have been initialised and
1425 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 */
1427void __init kmem_cache_init(void)
1428{
Christoph Lametere498be72005-09-09 13:03:32 -07001429 int i;
1430
Joonsoo Kim68126702013-10-24 10:07:42 +09001431 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1432 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001433 kmem_cache = &kmem_cache_boot;
Christoph Lameter6a673682013-01-10 19:14:19 +00001434 setup_node_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001435
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001436 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001437 use_alien_caches = 0;
1438
Christoph Lameter3c583462012-11-28 16:23:01 +00001439 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001440 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001441
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001442 set_up_node(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443
1444 /*
1445 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001446 * page orders on machines with more than 32MB of memory if
1447 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001449 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001450 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 /* Bootstrap is tricky, because several objects are allocated
1453 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001454 * 1) initialize the kmem_cache cache: it contains the struct
1455 * kmem_cache structures of all caches, except kmem_cache itself:
1456 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001457 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001458 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001459 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001461 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001462 * An __init data area is used for the head array.
1463 * 3) Create the remaining kmalloc caches, with minimally sized
1464 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001465 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001467 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001468 * the other cache's with kmalloc allocated memory.
1469 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 */
1471
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001472 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473
Eric Dumazet8da34302007-05-06 14:49:29 -07001474 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001475 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001476 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001477 create_boot_cache(kmem_cache, "kmem_cache",
1478 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001479 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001480 SLAB_HWCACHE_ALIGN);
1481 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482
1483 /* 2+3) create the kmalloc caches */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
Andrew Mortona737b3e2006-03-22 00:08:11 -08001485 /*
1486 * Initialize the caches that provide memory for the array cache and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001487 * kmem_cache_node structures first. Without this, further allocations will
Andrew Mortona737b3e2006-03-22 00:08:11 -08001488 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001489 */
1490
Christoph Lametere3366012013-01-10 19:14:18 +00001491 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1492 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001493
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001494 if (INDEX_AC != INDEX_NODE)
1495 kmalloc_caches[INDEX_NODE] =
1496 create_kmalloc_cache("kmalloc-node",
1497 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001498
Ingo Molnare0a42722006-06-23 02:03:46 -07001499 slab_early_init = 0;
1500
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 /* 4) Replace the bootstrap head arrays */
1502 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001503 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001504
Pekka Enberg83b519e2009-06-10 19:40:04 +03001505 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001506
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001507 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001508 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001509 /*
1510 * Do not assume that spinlocks can be initialized via memcpy:
1511 */
1512 spin_lock_init(&ptr->lock);
1513
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001514 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001515
Pekka Enberg83b519e2009-06-10 19:40:04 +03001516 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001517
Christoph Lametere3366012013-01-10 19:14:18 +00001518 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001519 != &initarray_generic.cache);
Christoph Lametere3366012013-01-10 19:14:18 +00001520 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001521 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001522 /*
1523 * Do not assume that spinlocks can be initialized via memcpy:
1524 */
1525 spin_lock_init(&ptr->lock);
1526
Christoph Lametere3366012013-01-10 19:14:18 +00001527 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001529 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001530 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001531 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
Mel Gorman9c09a952008-01-24 05:49:54 -08001533 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001534 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001535
Christoph Lametere3366012013-01-10 19:14:18 +00001536 init_list(kmalloc_caches[INDEX_AC],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001537 &init_kmem_cache_node[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001538
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001539 if (INDEX_AC != INDEX_NODE) {
1540 init_list(kmalloc_caches[INDEX_NODE],
1541 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001542 }
1543 }
1544 }
1545
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001546 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001547}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001548
Pekka Enberg8429db52009-06-12 15:58:59 +03001549void __init kmem_cache_init_late(void)
1550{
1551 struct kmem_cache *cachep;
1552
Christoph Lameter97d06602012-07-06 15:25:11 -05001553 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001554
Pekka Enberg8429db52009-06-12 15:58:59 +03001555 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001556 mutex_lock(&slab_mutex);
1557 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001558 if (enable_cpucache(cachep, GFP_NOWAIT))
1559 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001560 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001561
Michael Wang947ca182012-09-05 10:33:18 +08001562 /* Annotate slab for lockdep -- annotate the malloc caches */
1563 init_lock_keys();
1564
Christoph Lameter97d06602012-07-06 15:25:11 -05001565 /* Done! */
1566 slab_state = FULL;
1567
Andrew Mortona737b3e2006-03-22 00:08:11 -08001568 /*
1569 * Register a cpu startup notifier callback that initializes
1570 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 */
1572 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573
David Rientjes8f9f8d92010-03-27 19:40:47 -07001574#ifdef CONFIG_NUMA
1575 /*
1576 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001577 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001578 */
1579 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1580#endif
1581
Andrew Mortona737b3e2006-03-22 00:08:11 -08001582 /*
1583 * The reap timers are started later, with a module init call: That part
1584 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 */
1586}
1587
1588static int __init cpucache_init(void)
1589{
1590 int cpu;
1591
Andrew Mortona737b3e2006-03-22 00:08:11 -08001592 /*
1593 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 */
Christoph Lametere498be72005-09-09 13:03:32 -07001595 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001597
1598 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001599 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 return 0;
1601}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602__initcall(cpucache_init);
1603
Rafael Aquini8bdec192012-03-09 17:27:27 -03001604static noinline void
1605slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1606{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001607 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001608 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001609 unsigned long flags;
1610 int node;
1611
1612 printk(KERN_WARNING
1613 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1614 nodeid, gfpflags);
1615 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001616 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001617
1618 for_each_online_node(node) {
1619 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1620 unsigned long active_slabs = 0, num_slabs = 0;
1621
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001622 n = cachep->node[node];
1623 if (!n)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001624 continue;
1625
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001626 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001627 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001628 active_objs += cachep->num;
1629 active_slabs++;
1630 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001631 list_for_each_entry(page, &n->slabs_partial, lru) {
1632 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001633 active_slabs++;
1634 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001635 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001636 num_slabs++;
1637
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001638 free_objects += n->free_objects;
1639 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001640
1641 num_slabs += active_slabs;
1642 num_objs = num_slabs * cachep->num;
1643 printk(KERN_WARNING
1644 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1645 node, active_slabs, num_slabs, active_objs, num_objs,
1646 free_objects);
1647 }
1648}
1649
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650/*
1651 * Interface to system's page allocator. No need to hold the cache-lock.
1652 *
1653 * If we requested dmaable memory, we will get it. Even if we
1654 * did not request dmaable memory, we might get it, but that
1655 * would be relatively rare and ignorable.
1656 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001657static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1658 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659{
1660 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001661 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001662
Glauber Costaa618e892012-06-14 16:17:21 +04001663 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001664 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1665 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001666
Linus Torvalds517d0862009-06-16 19:50:13 -07001667 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001668 if (!page) {
1669 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1670 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001672 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001674 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001675 if (unlikely(page->pfmemalloc))
1676 pfmemalloc_active = true;
1677
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001678 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001680 add_zone_page_state(page_zone(page),
1681 NR_SLAB_RECLAIMABLE, nr_pages);
1682 else
1683 add_zone_page_state(page_zone(page),
1684 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001685 __SetPageSlab(page);
1686 if (page->pfmemalloc)
1687 SetPageSlabPfmemalloc(page);
Glauber Costa1f458cb2012-12-18 14:22:50 -08001688 memcg_bind_pages(cachep, cachep->gfporder);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001689
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001690 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1691 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1692
1693 if (cachep->ctor)
1694 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1695 else
1696 kmemcheck_mark_unallocated_pages(page, nr_pages);
1697 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001698
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001699 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700}
1701
1702/*
1703 * Interface to system's page release.
1704 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001705static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001707 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001709 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001710
Christoph Lameter972d1a72006-09-25 23:31:51 -07001711 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1712 sub_zone_page_state(page_zone(page),
1713 NR_SLAB_RECLAIMABLE, nr_freed);
1714 else
1715 sub_zone_page_state(page_zone(page),
1716 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001717
Joonsoo Kima57a4982013-10-24 10:07:44 +09001718 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001719 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001720 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001721 page_mapcount_reset(page);
1722 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001723
1724 memcg_release_pages(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 if (current->reclaim_state)
1726 current->reclaim_state->reclaimed_slab += nr_freed;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001727 __free_memcg_kmem_pages(page, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728}
1729
1730static void kmem_rcu_free(struct rcu_head *head)
1731{
Joonsoo Kim68126702013-10-24 10:07:42 +09001732 struct kmem_cache *cachep;
1733 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734
Joonsoo Kim68126702013-10-24 10:07:42 +09001735 page = container_of(head, struct page, rcu_head);
1736 cachep = page->slab_cache;
1737
1738 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739}
1740
1741#if DEBUG
1742
1743#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001744static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001745 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001747 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001749 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001751 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 return;
1753
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001754 *addr++ = 0x12345678;
1755 *addr++ = caller;
1756 *addr++ = smp_processor_id();
1757 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 {
1759 unsigned long *sptr = &caller;
1760 unsigned long svalue;
1761
1762 while (!kstack_end(sptr)) {
1763 svalue = *sptr++;
1764 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001765 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 size -= sizeof(unsigned long);
1767 if (size <= sizeof(unsigned long))
1768 break;
1769 }
1770 }
1771
1772 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001773 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774}
1775#endif
1776
Pekka Enberg343e0d72006-02-01 03:05:50 -08001777static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001779 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001780 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781
1782 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001783 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784}
1785
1786static void dump_line(char *data, int offset, int limit)
1787{
1788 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001789 unsigned char error = 0;
1790 int bad_count = 0;
1791
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001792 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001793 for (i = 0; i < limit; i++) {
1794 if (data[offset + i] != POISON_FREE) {
1795 error = data[offset + i];
1796 bad_count++;
1797 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001798 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001799 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1800 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001801
1802 if (bad_count == 1) {
1803 error ^= POISON_FREE;
1804 if (!(error & (error - 1))) {
1805 printk(KERN_ERR "Single bit error detected. Probably "
1806 "bad RAM.\n");
1807#ifdef CONFIG_X86
1808 printk(KERN_ERR "Run memtest86+ or a similar memory "
1809 "test tool.\n");
1810#else
1811 printk(KERN_ERR "Run a memory test tool.\n");
1812#endif
1813 }
1814 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815}
1816#endif
1817
1818#if DEBUG
1819
Pekka Enberg343e0d72006-02-01 03:05:50 -08001820static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821{
1822 int i, size;
1823 char *realobj;
1824
1825 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001826 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001827 *dbg_redzone1(cachep, objp),
1828 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 }
1830
1831 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001832 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1833 *dbg_userword(cachep, objp),
1834 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001836 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001837 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001838 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 int limit;
1840 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001841 if (i + limit > size)
1842 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 dump_line(realobj, i, limit);
1844 }
1845}
1846
Pekka Enberg343e0d72006-02-01 03:05:50 -08001847static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848{
1849 char *realobj;
1850 int size, i;
1851 int lines = 0;
1852
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001853 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001854 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001856 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 exp = POISON_END;
1860 if (realobj[i] != exp) {
1861 int limit;
1862 /* Mismatch ! */
1863 /* Print header */
1864 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001865 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001866 "Slab corruption (%s): %s start=%p, len=%d\n",
1867 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 print_objinfo(cachep, objp, 0);
1869 }
1870 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001871 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001873 if (i + limit > size)
1874 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 dump_line(realobj, i, limit);
1876 i += 16;
1877 lines++;
1878 /* Limit to 5 lines */
1879 if (lines > 5)
1880 break;
1881 }
1882 }
1883 if (lines != 0) {
1884 /* Print some data about the neighboring objects, if they
1885 * exist:
1886 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001887 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001888 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889
Joonsoo Kim8456a642013-10-24 10:07:49 +09001890 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001892 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001893 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 print_objinfo(cachep, objp, 2);
1897 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001898 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001899 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001900 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001902 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 print_objinfo(cachep, objp, 2);
1904 }
1905 }
1906}
1907#endif
1908
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001910static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1911 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001912{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 int i;
1914 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001915 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
1917 if (cachep->flags & SLAB_POISON) {
1918#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001919 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001920 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001921 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001922 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 else
1924 check_poison_obj(cachep, objp);
1925#else
1926 check_poison_obj(cachep, objp);
1927#endif
1928 }
1929 if (cachep->flags & SLAB_RED_ZONE) {
1930 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1931 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001932 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1934 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001935 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001938}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001940static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1941 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001942{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001943}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944#endif
1945
Randy Dunlap911851e2006-03-22 00:08:14 -08001946/**
1947 * slab_destroy - destroy and release all objects in a slab
1948 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001949 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001950 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001951 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001952 * Before calling the slab must have been unlinked from the cache. The
1953 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001954 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001955static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001956{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001957 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001958
Joonsoo Kim8456a642013-10-24 10:07:49 +09001959 freelist = page->freelist;
1960 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001962 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963
Joonsoo Kim68126702013-10-24 10:07:42 +09001964 /*
1965 * RCU free overloads the RCU head over the LRU.
1966 * slab_page has been overloeaded over the LRU,
1967 * however it is not used from now on so that
1968 * we can use it safely.
1969 */
1970 head = (void *)&page->rcu_head;
1971 call_rcu(head, kmem_rcu_free);
1972
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001974 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 }
Joonsoo Kim68126702013-10-24 10:07:42 +09001976
1977 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001978 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001979 * although actual page can be freed in rcu context
1980 */
1981 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001982 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983}
1984
1985/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001986 * calculate_slab_order - calculate size (page order) of slabs
1987 * @cachep: pointer to the cache that is being created
1988 * @size: size of objects to be created in this cache.
1989 * @align: required alignment for the objects.
1990 * @flags: slab allocation flags
1991 *
1992 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001993 *
1994 * This could be made much more intelligent. For now, try to avoid using
1995 * high order pages for slabs. When the gfp() functions are more friendly
1996 * towards high-order requests, this should be changed.
1997 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001998static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001999 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002000{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002001 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002002 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002003 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002004
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002005 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002006 unsigned int num;
2007 size_t remainder;
2008
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002009 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010 if (!num)
2011 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002012
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002013 if (flags & CFLGS_OFF_SLAB) {
2014 /*
2015 * Max number of objs-per-slab for caches which
2016 * use off-slab slabs. Needed to avoid a possible
2017 * looping condition in cache_grow().
2018 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002019 offslab_limit = size;
Joonsoo Kim16025172013-10-24 10:07:46 +09002020 offslab_limit /= sizeof(unsigned int);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002021
2022 if (num > offslab_limit)
2023 break;
2024 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002025
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002026 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002027 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002028 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002029 left_over = remainder;
2030
2031 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002032 * A VFS-reclaimable slab tends to have most allocations
2033 * as GFP_NOFS and we really don't want to have to be allocating
2034 * higher-order pages when we are unable to shrink dcache.
2035 */
2036 if (flags & SLAB_RECLAIM_ACCOUNT)
2037 break;
2038
2039 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002040 * Large number of objects is good, but very large slabs are
2041 * currently bad for the gfp()s.
2042 */
David Rientjes543585c2011-10-18 22:09:24 -07002043 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002044 break;
2045
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002046 /*
2047 * Acceptable internal fragmentation?
2048 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002049 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002050 break;
2051 }
2052 return left_over;
2053}
2054
Pekka Enberg83b519e2009-06-10 19:40:04 +03002055static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002056{
Christoph Lameter97d06602012-07-06 15:25:11 -05002057 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002058 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002059
Christoph Lameter97d06602012-07-06 15:25:11 -05002060 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002061 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002062 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002063 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002064 * of by the caller of __kmem_cache_create
2065 */
2066 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2067 slab_state = PARTIAL;
2068 } else if (slab_state == PARTIAL) {
2069 /*
2070 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002071 * that's used by kmalloc(24), otherwise the creation of
2072 * further caches will BUG().
2073 */
2074 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2075
2076 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002077 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2078 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002079 * otherwise the creation of further caches will BUG().
2080 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002081 set_up_node(cachep, SIZE_AC);
2082 if (INDEX_AC == INDEX_NODE)
2083 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002084 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002085 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002086 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002087 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002088 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002089 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002090
Christoph Lameter97d06602012-07-06 15:25:11 -05002091 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002092 set_up_node(cachep, SIZE_NODE);
2093 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002094 } else {
2095 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002096 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002097 cachep->node[node] =
Christoph Lameter6744f082013-01-10 19:12:17 +00002098 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002099 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002100 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002101 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002102 }
2103 }
2104 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002105 cachep->node[numa_mem_id()]->next_reap =
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002106 jiffies + REAPTIMEOUT_LIST3 +
2107 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2108
2109 cpu_cache_get(cachep)->avail = 0;
2110 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2111 cpu_cache_get(cachep)->batchcount = 1;
2112 cpu_cache_get(cachep)->touched = 0;
2113 cachep->batchcount = 1;
2114 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002115 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002116}
2117
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002118/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002119 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002120 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 *
2123 * Returns a ptr to the cache on success, NULL on failure.
2124 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002125 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 * The flags are
2128 *
2129 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2130 * to catch references to uninitialised memory.
2131 *
2132 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2133 * for buffer overruns.
2134 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2136 * cacheline. This can be beneficial if you're counting cycles as closely
2137 * as davem.
2138 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002139int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002140__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002142 size_t left_over, freelist_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002143 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002144 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002145 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148#if FORCED_DEBUG
2149 /*
2150 * Enable redzoning and last user accounting, except for caches with
2151 * large objects, if the increased size would increase the object size
2152 * above the next power of two: caches with object sizes just above a
2153 * power of two have a significant amount of internal fragmentation.
2154 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002155 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2156 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002157 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 if (!(flags & SLAB_DESTROY_BY_RCU))
2159 flags |= SLAB_POISON;
2160#endif
2161 if (flags & SLAB_DESTROY_BY_RCU)
2162 BUG_ON(flags & SLAB_POISON);
2163#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164
Andrew Mortona737b3e2006-03-22 00:08:11 -08002165 /*
2166 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 * unaligned accesses for some archs when redzoning is used, and makes
2168 * sure any on-slab bufctl's are also correctly aligned.
2169 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002170 if (size & (BYTES_PER_WORD - 1)) {
2171 size += (BYTES_PER_WORD - 1);
2172 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 }
2174
Pekka Enbergca5f9702006-09-25 23:31:25 -07002175 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002176 * Redzoning and user store require word alignment or possibly larger.
2177 * Note this will be overridden by architecture or caller mandated
2178 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002179 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002180 if (flags & SLAB_STORE_USER)
2181 ralign = BYTES_PER_WORD;
2182
2183 if (flags & SLAB_RED_ZONE) {
2184 ralign = REDZONE_ALIGN;
2185 /* If redzoning, ensure that the second redzone is suitably
2186 * aligned, by adjusting the object size accordingly. */
2187 size += REDZONE_ALIGN - 1;
2188 size &= ~(REDZONE_ALIGN - 1);
2189 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002190
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002191 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002192 if (ralign < cachep->align) {
2193 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002195 /* disable debug if necessary */
2196 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002197 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002198 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002199 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002201 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202
Pekka Enberg83b519e2009-06-10 19:40:04 +03002203 if (slab_is_available())
2204 gfp = GFP_KERNEL;
2205 else
2206 gfp = GFP_NOWAIT;
2207
Christoph Lameter6a673682013-01-10 19:14:19 +00002208 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210
Pekka Enbergca5f9702006-09-25 23:31:25 -07002211 /*
2212 * Both debugging options require word-alignment which is calculated
2213 * into align above.
2214 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002217 cachep->obj_offset += sizeof(unsigned long long);
2218 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 }
2220 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002221 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002222 * the real object. But if the second red zone needs to be
2223 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002225 if (flags & SLAB_RED_ZONE)
2226 size += REDZONE_ALIGN;
2227 else
2228 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 }
2230#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002231 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002232 && cachep->object_size > cache_line_size()
2233 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2234 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 size = PAGE_SIZE;
2236 }
2237#endif
2238#endif
2239
Ingo Molnare0a42722006-06-23 02:03:46 -07002240 /*
2241 * Determine if the slab management is 'on' or 'off' slab.
2242 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002243 * it too early on. Always use on-slab management when
2244 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002245 */
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002246 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2247 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 /*
2249 * Size is large, assume best to place the slab management obj
2250 * off-slab (should allow better packing of objs).
2251 */
2252 flags |= CFLGS_OFF_SLAB;
2253
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002254 size = ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002256 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002258 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002259 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002260
Joonsoo Kim8456a642013-10-24 10:07:49 +09002261 freelist_size =
2262 ALIGN(cachep->num * sizeof(unsigned int), cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
2264 /*
2265 * If the slab has been placed off-slab, and we have enough space then
2266 * move it on-slab. This is at the expense of any extra colouring.
2267 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002268 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002270 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 }
2272
2273 if (flags & CFLGS_OFF_SLAB) {
2274 /* really off slab. No need for manual alignment */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002275 freelist_size = cachep->num * sizeof(unsigned int);
Ron Lee67461362009-05-22 04:58:22 +09302276
2277#ifdef CONFIG_PAGE_POISONING
2278 /* If we're going to use the generic kernel_map_pages()
2279 * poisoning, then it's going to smash the contents of
2280 * the redzone and userword anyhow, so switch them off.
2281 */
2282 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2283 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2284#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 }
2286
2287 cachep->colour_off = cache_line_size();
2288 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002289 if (cachep->colour_off < cachep->align)
2290 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002291 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002292 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002294 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002295 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002296 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002297 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002298 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002300 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002301 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002302 /*
2303 * This is a possibility for one of the malloc_sizes caches.
2304 * But since we go off slab only for object size greater than
2305 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2306 * this should not happen at all.
2307 * But leave a BUG_ON for some lucky dude.
2308 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002309 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002310 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002312 err = setup_cpu_cache(cachep, gfp);
2313 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002314 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002315 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002316 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317
Peter Zijlstra83835b32011-07-22 15:26:05 +02002318 if (flags & SLAB_DEBUG_OBJECTS) {
2319 /*
2320 * Would deadlock through slab_destroy()->call_rcu()->
2321 * debug_object_activate()->kmem_cache_alloc().
2322 */
2323 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2324
2325 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002326 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2327 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002328
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002329 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331
2332#if DEBUG
2333static void check_irq_off(void)
2334{
2335 BUG_ON(!irqs_disabled());
2336}
2337
2338static void check_irq_on(void)
2339{
2340 BUG_ON(irqs_disabled());
2341}
2342
Pekka Enberg343e0d72006-02-01 03:05:50 -08002343static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344{
2345#ifdef CONFIG_SMP
2346 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002347 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348#endif
2349}
Christoph Lametere498be72005-09-09 13:03:32 -07002350
Pekka Enberg343e0d72006-02-01 03:05:50 -08002351static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002352{
2353#ifdef CONFIG_SMP
2354 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002355 assert_spin_locked(&cachep->node[node]->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002356#endif
2357}
2358
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359#else
2360#define check_irq_off() do { } while(0)
2361#define check_irq_on() do { } while(0)
2362#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002363#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364#endif
2365
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002366static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002367 struct array_cache *ac,
2368 int force, int node);
2369
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370static void do_drain(void *arg)
2371{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002372 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002374 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375
2376 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002377 ac = cpu_cache_get(cachep);
Christoph Lameter6a673682013-01-10 19:14:19 +00002378 spin_lock(&cachep->node[node]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002379 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002380 spin_unlock(&cachep->node[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 ac->avail = 0;
2382}
2383
Pekka Enberg343e0d72006-02-01 03:05:50 -08002384static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002386 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002387 int node;
2388
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002389 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002391 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002392 n = cachep->node[node];
2393 if (n && n->alien)
2394 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002395 }
2396
2397 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002398 n = cachep->node[node];
2399 if (n)
2400 drain_array(cachep, n, n->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002401 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402}
2403
Christoph Lametered11d9e2006-06-30 01:55:45 -07002404/*
2405 * Remove slabs from the list of free slabs.
2406 * Specify the number of slabs to drain in tofree.
2407 *
2408 * Returns the actual number of slabs released.
2409 */
2410static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002411 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002413 struct list_head *p;
2414 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002415 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
Christoph Lametered11d9e2006-06-30 01:55:45 -07002417 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002418 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002420 spin_lock_irq(&n->list_lock);
2421 p = n->slabs_free.prev;
2422 if (p == &n->slabs_free) {
2423 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002424 goto out;
2425 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
Joonsoo Kim8456a642013-10-24 10:07:49 +09002427 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002429 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002431 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002432 /*
2433 * Safe to drop the lock. The slab is no longer linked
2434 * to the cache.
2435 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002436 n->free_objects -= cache->num;
2437 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002438 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002439 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002441out:
2442 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443}
2444
Christoph Lameter18004c52012-07-06 15:25:12 -05002445/* Called with slab_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002446static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002447{
2448 int ret = 0, i = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002449 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002450
2451 drain_cpu_caches(cachep);
2452
2453 check_irq_on();
2454 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002455 n = cachep->node[i];
2456 if (!n)
Christoph Lametered11d9e2006-06-30 01:55:45 -07002457 continue;
2458
Wanpeng Li0fa81032013-07-04 08:33:22 +08002459 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002460
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002461 ret += !list_empty(&n->slabs_full) ||
2462 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002463 }
2464 return (ret ? 1 : 0);
2465}
2466
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467/**
2468 * kmem_cache_shrink - Shrink a cache.
2469 * @cachep: The cache to shrink.
2470 *
2471 * Releases as many slabs as possible for a cache.
2472 * To help debugging, a zero exit status indicates all slabs were released.
2473 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002474int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002476 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002477 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002479 get_online_cpus();
Christoph Lameter18004c52012-07-06 15:25:12 -05002480 mutex_lock(&slab_mutex);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002481 ret = __cache_shrink(cachep);
Christoph Lameter18004c52012-07-06 15:25:12 -05002482 mutex_unlock(&slab_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002483 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002484 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485}
2486EXPORT_SYMBOL(kmem_cache_shrink);
2487
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002488int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489{
Christoph Lameter12c36672012-09-04 23:38:33 +00002490 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002491 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002492 int rc = __cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493
Christoph Lameter12c36672012-09-04 23:38:33 +00002494 if (rc)
2495 return rc;
2496
2497 for_each_online_cpu(i)
2498 kfree(cachep->array[i]);
2499
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002500 /* NUMA: free the node structures */
Christoph Lameter12c36672012-09-04 23:38:33 +00002501 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002502 n = cachep->node[i];
2503 if (n) {
2504 kfree(n->shared);
2505 free_alien_cache(n->alien);
2506 kfree(n);
Christoph Lameter12c36672012-09-04 23:38:33 +00002507 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002509 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002512/*
2513 * Get the memory for a slab management obj.
2514 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2515 * always come from malloc_sizes caches. The slab descriptor cannot
2516 * come from the same cache which is getting created because,
2517 * when we are searching for an appropriate cache for these
2518 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2519 * If we are creating a malloc_sizes cache here it would not be visible to
2520 * kmem_find_general_cachep till the initialization is complete.
Joonsoo Kim8456a642013-10-24 10:07:49 +09002521 * Hence we cannot have freelist_cache same as the original cache.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002522 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002523static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002524 struct page *page, int colour_off,
2525 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002527 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002528 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002529
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 if (OFF_SLAB(cachep)) {
2531 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002532 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002533 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002534 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 return NULL;
2536 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002537 freelist = addr + colour_off;
2538 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002540 page->active = 0;
2541 page->s_mem = addr + colour_off;
2542 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543}
2544
Joonsoo Kime7444d92013-10-24 10:07:51 +09002545static inline unsigned int *slab_freelist(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002547 return (unsigned int *)(page->freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548}
2549
Pekka Enberg343e0d72006-02-01 03:05:50 -08002550static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002551 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552{
2553 int i;
2554
2555 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002556 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557#if DEBUG
2558 /* need to poison the objs? */
2559 if (cachep->flags & SLAB_POISON)
2560 poison_obj(cachep, objp, POISON_FREE);
2561 if (cachep->flags & SLAB_STORE_USER)
2562 *dbg_userword(cachep, objp) = NULL;
2563
2564 if (cachep->flags & SLAB_RED_ZONE) {
2565 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2566 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2567 }
2568 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002569 * Constructors are not allowed to allocate memory from the same
2570 * cache which they are a constructor for. Otherwise, deadlock.
2571 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 */
2573 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002574 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575
2576 if (cachep->flags & SLAB_RED_ZONE) {
2577 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2578 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002579 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2581 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002582 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002584 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002585 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002586 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002587 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588#else
2589 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002590 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591#endif
Joonsoo Kime7444d92013-10-24 10:07:51 +09002592 slab_freelist(page)[i] = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594}
2595
Pekka Enberg343e0d72006-02-01 03:05:50 -08002596static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002598 if (CONFIG_ZONE_DMA_FLAG) {
2599 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002600 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002601 else
Glauber Costaa618e892012-06-14 16:17:21 +04002602 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002603 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604}
2605
Joonsoo Kim8456a642013-10-24 10:07:49 +09002606static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002607 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002608{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002609 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002610
Joonsoo Kime7444d92013-10-24 10:07:51 +09002611 objp = index_to_obj(cachep, page, slab_freelist(page)[page->active]);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002612 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002613#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002614 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002615#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002616
2617 return objp;
2618}
2619
Joonsoo Kim8456a642013-10-24 10:07:49 +09002620static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002621 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002622{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002623 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002624#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002625 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002626
Matthew Dobson78d382d2006-02-01 03:05:47 -08002627 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002628 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002629
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002630 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002631 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime7444d92013-10-24 10:07:51 +09002632 if (slab_freelist(page)[i] == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002633 printk(KERN_ERR "slab: double free detected in cache "
2634 "'%s', objp %p\n", cachep->name, objp);
2635 BUG();
2636 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002637 }
2638#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002639 page->active--;
Joonsoo Kime7444d92013-10-24 10:07:51 +09002640 slab_freelist(page)[page->active] = objnr;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002641}
2642
Pekka Enberg47768742006-06-23 02:03:07 -07002643/*
2644 * Map pages beginning at addr to the given cache and slab. This is required
2645 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002646 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002647 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002648static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002649 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002651 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002652 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653}
2654
2655/*
2656 * Grow (by 1) the number of slabs within a cache. This is called by
2657 * kmem_cache_alloc() when there are no active objs left in a cache.
2658 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002659static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002660 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002662 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002663 size_t offset;
2664 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002665 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666
Andrew Mortona737b3e2006-03-22 00:08:11 -08002667 /*
2668 * Be lazy and only check for valid flags here, keeping it out of the
2669 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002671 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2672 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002674 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002676 n = cachep->node[nodeid];
2677 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678
2679 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002680 offset = n->colour_next;
2681 n->colour_next++;
2682 if (n->colour_next >= cachep->colour)
2683 n->colour_next = 0;
2684 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002686 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687
2688 if (local_flags & __GFP_WAIT)
2689 local_irq_enable();
2690
2691 /*
2692 * The test for missing atomic flag is performed here, rather than
2693 * the more obvious place, simply to reduce the critical path length
2694 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2695 * will eventually be caught here (where it matters).
2696 */
2697 kmem_flagcheck(cachep, flags);
2698
Andrew Mortona737b3e2006-03-22 00:08:11 -08002699 /*
2700 * Get mem for the objs. Attempt to allocate a physical page from
2701 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002702 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002703 if (!page)
2704 page = kmem_getpages(cachep, local_flags, nodeid);
2705 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 goto failed;
2707
2708 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002709 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002710 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002711 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 goto opps1;
2713
Joonsoo Kim8456a642013-10-24 10:07:49 +09002714 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715
Joonsoo Kim8456a642013-10-24 10:07:49 +09002716 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717
2718 if (local_flags & __GFP_WAIT)
2719 local_irq_disable();
2720 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002721 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
2723 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002724 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002726 n->free_objects += cachep->num;
2727 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002729opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002730 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002731failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 if (local_flags & __GFP_WAIT)
2733 local_irq_disable();
2734 return 0;
2735}
2736
2737#if DEBUG
2738
2739/*
2740 * Perform extra freeing checks:
2741 * - detect bad pointers.
2742 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 */
2744static void kfree_debugcheck(const void *objp)
2745{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 if (!virt_addr_valid(objp)) {
2747 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002748 (unsigned long)objp);
2749 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751}
2752
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002753static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2754{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002755 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002756
2757 redzone1 = *dbg_redzone1(cache, obj);
2758 redzone2 = *dbg_redzone2(cache, obj);
2759
2760 /*
2761 * Redzone is ok.
2762 */
2763 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2764 return;
2765
2766 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2767 slab_error(cache, "double free detected");
2768 else
2769 slab_error(cache, "memory outside object was overwritten");
2770
David Woodhouseb46b8f12007-05-08 00:22:59 -07002771 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002772 obj, redzone1, redzone2);
2773}
2774
Pekka Enberg343e0d72006-02-01 03:05:50 -08002775static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002776 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002779 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002781 BUG_ON(virt_to_cache(objp) != cachep);
2782
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002783 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002785 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002788 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2790 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2791 }
2792 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002793 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794
Joonsoo Kim8456a642013-10-24 10:07:49 +09002795 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796
2797 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002798 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 if (cachep->flags & SLAB_POISON) {
2801#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002802 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002803 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002804 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002805 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 } else {
2807 poison_obj(cachep, objp, POISON_FREE);
2808 }
2809#else
2810 poison_obj(cachep, objp, POISON_FREE);
2811#endif
2812 }
2813 return objp;
2814}
2815
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816#else
2817#define kfree_debugcheck(x) do { } while(0)
2818#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819#endif
2820
Mel Gorman072bb0a2012-07-31 16:43:58 -07002821static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2822 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823{
2824 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002825 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002827 int node;
2828
Joe Korty6d2144d2008-03-05 15:04:59 -08002829 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002830 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002831 if (unlikely(force_refill))
2832 goto force_grow;
2833retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002834 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 batchcount = ac->batchcount;
2836 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002837 /*
2838 * If there was little recent activity on this cache, then
2839 * perform only a partial refill. Otherwise we could generate
2840 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 */
2842 batchcount = BATCHREFILL_LIMIT;
2843 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002844 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002846 BUG_ON(ac->avail > 0 || !n);
2847 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002848
Christoph Lameter3ded1752006-03-25 03:06:44 -08002849 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002850 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2851 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002852 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002853 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002854
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 while (batchcount > 0) {
2856 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002857 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002859 entry = n->slabs_partial.next;
2860 if (entry == &n->slabs_partial) {
2861 n->free_touched = 1;
2862 entry = n->slabs_free.next;
2863 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 goto must_grow;
2865 }
2866
Joonsoo Kim8456a642013-10-24 10:07:49 +09002867 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002869
2870 /*
2871 * The slab was either on partial or free list so
2872 * there must be at least one object available for
2873 * allocation.
2874 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002875 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002876
Joonsoo Kim8456a642013-10-24 10:07:49 +09002877 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 STATS_INC_ALLOCED(cachep);
2879 STATS_INC_ACTIVE(cachep);
2880 STATS_SET_HIGH(cachep);
2881
Joonsoo Kim8456a642013-10-24 10:07:49 +09002882 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002883 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885
2886 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002887 list_del(&page->lru);
2888 if (page->active == cachep->num)
2889 list_add(&page->list, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09002891 list_add(&page->list, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 }
2893
Andrew Mortona737b3e2006-03-22 00:08:11 -08002894must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002895 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002896alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002897 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898
2899 if (unlikely(!ac->avail)) {
2900 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002901force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002902 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002903
Andrew Mortona737b3e2006-03-22 00:08:11 -08002904 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002905 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002906 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002907
2908 /* no objects in sight? abort */
2909 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 return NULL;
2911
Andrew Mortona737b3e2006-03-22 00:08:11 -08002912 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 goto retry;
2914 }
2915 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002916
2917 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918}
2919
Andrew Mortona737b3e2006-03-22 00:08:11 -08002920static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2921 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922{
2923 might_sleep_if(flags & __GFP_WAIT);
2924#if DEBUG
2925 kmem_flagcheck(cachep, flags);
2926#endif
2927}
2928
2929#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002930static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002931 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002932{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002933 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002935 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002937 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002938 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002939 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940 else
2941 check_poison_obj(cachep, objp);
2942#else
2943 check_poison_obj(cachep, objp);
2944#endif
2945 poison_obj(cachep, objp, POISON_INUSE);
2946 }
2947 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002948 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949
2950 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002951 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2952 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2953 slab_error(cachep, "double free, or memory outside"
2954 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002955 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002956 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002957 objp, *dbg_redzone1(cachep, objp),
2958 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 }
2960 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2961 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2962 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002963 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002964 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002965 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002966 if (ARCH_SLAB_MINALIGN &&
2967 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002968 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002969 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002970 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 return objp;
2972}
2973#else
2974#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2975#endif
2976
Akinobu Mita773ff602008-12-23 19:37:01 +09002977static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002978{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00002979 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09002980 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002981
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002982 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002983}
2984
Pekka Enberg343e0d72006-02-01 03:05:50 -08002985static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002987 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002989 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990
Alok N Kataria5c382302005-09-27 21:45:46 -07002991 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002992
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002993 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002996 objp = ac_get_obj(cachep, ac, flags, false);
2997
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002998 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07002999 * Allow for the possibility all avail objects are not allowed
3000 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003001 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003002 if (objp) {
3003 STATS_INC_ALLOCHIT(cachep);
3004 goto out;
3005 }
3006 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003008
3009 STATS_INC_ALLOCMISS(cachep);
3010 objp = cache_alloc_refill(cachep, flags, force_refill);
3011 /*
3012 * the 'ac' may be updated by cache_alloc_refill(),
3013 * and kmemleak_erase() requires its correct value.
3014 */
3015 ac = cpu_cache_get(cachep);
3016
3017out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003018 /*
3019 * To avoid a false negative, if an object that is in one of the
3020 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3021 * treat the array pointers as a reference to the object.
3022 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003023 if (objp)
3024 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003025 return objp;
3026}
3027
Christoph Lametere498be72005-09-09 13:03:32 -07003028#ifdef CONFIG_NUMA
3029/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003030 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003031 *
3032 * If we are in_interrupt, then process context, including cpusets and
3033 * mempolicy, may not apply and should not be used for allocation policy.
3034 */
3035static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3036{
3037 int nid_alloc, nid_here;
3038
Christoph Lameter765c4502006-09-27 01:50:08 -07003039 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003040 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003041 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003042 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003043 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003044 else if (current->mempolicy)
Andi Kleene7b691b2012-06-09 02:40:03 -07003045 nid_alloc = slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003046 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003047 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003048 return NULL;
3049}
3050
3051/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003052 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003053 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003054 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003055 * perform an allocation without specifying a node. This allows the page
3056 * allocator to do its reclaim / fallback magic. We then insert the
3057 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003058 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003059static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003060{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003061 struct zonelist *zonelist;
3062 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003063 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003064 struct zone *zone;
3065 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003066 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003067 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003068 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003069
3070 if (flags & __GFP_THISNODE)
3071 return NULL;
3072
Christoph Lameter6cb06222007-10-16 01:25:41 -07003073 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003074
Mel Gormancc9a6c82012-03-21 16:34:11 -07003075retry_cpuset:
3076 cpuset_mems_cookie = get_mems_allowed();
Andi Kleene7b691b2012-06-09 02:40:03 -07003077 zonelist = node_zonelist(slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003078
Christoph Lameter3c517a62006-12-06 20:33:29 -08003079retry:
3080 /*
3081 * Look through allowed nodes for objects available
3082 * from existing per node queues.
3083 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003084 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3085 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003086
Mel Gorman54a6eb52008-04-28 02:12:16 -07003087 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter6a673682013-01-10 19:14:19 +00003088 cache->node[nid] &&
3089 cache->node[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003090 obj = ____cache_alloc_node(cache,
3091 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003092 if (obj)
3093 break;
3094 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003095 }
3096
Christoph Lametercfce6602007-05-06 14:50:17 -07003097 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003098 /*
3099 * This allocation will be performed within the constraints
3100 * of the current cpuset / memory policy requirements.
3101 * We may trigger various forms of reclaim on the allowed
3102 * set and go into memory reserves if necessary.
3103 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003104 struct page *page;
3105
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003106 if (local_flags & __GFP_WAIT)
3107 local_irq_enable();
3108 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003109 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003110 if (local_flags & __GFP_WAIT)
3111 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003112 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003113 /*
3114 * Insert into the appropriate per node queues
3115 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003116 nid = page_to_nid(page);
3117 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003118 obj = ____cache_alloc_node(cache,
3119 flags | GFP_THISNODE, nid);
3120 if (!obj)
3121 /*
3122 * Another processor may allocate the
3123 * objects in the slab since we are
3124 * not holding any locks.
3125 */
3126 goto retry;
3127 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003128 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003129 obj = NULL;
3130 }
3131 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003132 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003133
3134 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3135 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003136 return obj;
3137}
3138
3139/*
Christoph Lametere498be72005-09-09 13:03:32 -07003140 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003142static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003143 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003144{
3145 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003146 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003147 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003148 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003149 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003151 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003152 n = cachep->node[nodeid];
3153 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003154
Andrew Mortona737b3e2006-03-22 00:08:11 -08003155retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003156 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003157 spin_lock(&n->list_lock);
3158 entry = n->slabs_partial.next;
3159 if (entry == &n->slabs_partial) {
3160 n->free_touched = 1;
3161 entry = n->slabs_free.next;
3162 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003163 goto must_grow;
3164 }
Christoph Lametere498be72005-09-09 13:03:32 -07003165
Joonsoo Kim8456a642013-10-24 10:07:49 +09003166 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003167 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003168
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003169 STATS_INC_NODEALLOCS(cachep);
3170 STATS_INC_ACTIVE(cachep);
3171 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003172
Joonsoo Kim8456a642013-10-24 10:07:49 +09003173 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003174
Joonsoo Kim8456a642013-10-24 10:07:49 +09003175 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003176 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003177 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003178 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003179
Joonsoo Kim8456a642013-10-24 10:07:49 +09003180 if (page->active == cachep->num)
3181 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003182 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003183 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003184
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003185 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003186 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003187
Andrew Mortona737b3e2006-03-22 00:08:11 -08003188must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003189 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003190 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003191 if (x)
3192 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003193
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003194 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003195
Andrew Mortona737b3e2006-03-22 00:08:11 -08003196done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003197 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003198}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003199
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003200static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003201slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003202 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003203{
3204 unsigned long save_flags;
3205 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003206 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003207
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003208 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003209
Nick Piggincf40bd12009-01-21 08:12:39 +01003210 lockdep_trace_alloc(flags);
3211
Akinobu Mita773ff602008-12-23 19:37:01 +09003212 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003213 return NULL;
3214
Glauber Costad79923f2012-12-18 14:22:48 -08003215 cachep = memcg_kmem_get_cache(cachep, flags);
3216
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003217 cache_alloc_debugcheck_before(cachep, flags);
3218 local_irq_save(save_flags);
3219
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003220 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003221 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003222
Christoph Lameter6a673682013-01-10 19:14:19 +00003223 if (unlikely(!cachep->node[nodeid])) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003224 /* Node not bootstrapped yet */
3225 ptr = fallback_alloc(cachep, flags);
3226 goto out;
3227 }
3228
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003229 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003230 /*
3231 * Use the locally cached objects if possible.
3232 * However ____cache_alloc does not allow fallback
3233 * to other nodes. It may fail while we still have
3234 * objects on other nodes available.
3235 */
3236 ptr = ____cache_alloc(cachep, flags);
3237 if (ptr)
3238 goto out;
3239 }
3240 /* ___cache_alloc_node can fall back to other nodes */
3241 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3242 out:
3243 local_irq_restore(save_flags);
3244 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003245 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003246 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003247
Pekka Enbergc175eea2008-05-09 20:35:53 +02003248 if (likely(ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003249 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003250
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003251 if (unlikely((flags & __GFP_ZERO) && ptr))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003252 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003253
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003254 return ptr;
3255}
3256
3257static __always_inline void *
3258__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3259{
3260 void *objp;
3261
3262 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3263 objp = alternate_node_alloc(cache, flags);
3264 if (objp)
3265 goto out;
3266 }
3267 objp = ____cache_alloc(cache, flags);
3268
3269 /*
3270 * We may just have run out of memory on the local node.
3271 * ____cache_alloc_node() knows how to locate memory on other nodes
3272 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003273 if (!objp)
3274 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003275
3276 out:
3277 return objp;
3278}
3279#else
3280
3281static __always_inline void *
3282__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3283{
3284 return ____cache_alloc(cachep, flags);
3285}
3286
3287#endif /* CONFIG_NUMA */
3288
3289static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003290slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003291{
3292 unsigned long save_flags;
3293 void *objp;
3294
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003295 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003296
Nick Piggincf40bd12009-01-21 08:12:39 +01003297 lockdep_trace_alloc(flags);
3298
Akinobu Mita773ff602008-12-23 19:37:01 +09003299 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003300 return NULL;
3301
Glauber Costad79923f2012-12-18 14:22:48 -08003302 cachep = memcg_kmem_get_cache(cachep, flags);
3303
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003304 cache_alloc_debugcheck_before(cachep, flags);
3305 local_irq_save(save_flags);
3306 objp = __do_cache_alloc(cachep, flags);
3307 local_irq_restore(save_flags);
3308 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003309 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003310 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003311 prefetchw(objp);
3312
Pekka Enbergc175eea2008-05-09 20:35:53 +02003313 if (likely(objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003314 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003315
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003316 if (unlikely((flags & __GFP_ZERO) && objp))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003317 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003318
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003319 return objp;
3320}
Christoph Lametere498be72005-09-09 13:03:32 -07003321
3322/*
3323 * Caller needs to acquire correct kmem_list's list_lock
3324 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003325static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003326 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327{
3328 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003329 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330
3331 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003332 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003333 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003334
Mel Gorman072bb0a2012-07-31 16:43:58 -07003335 clear_obj_pfmemalloc(&objpp[i]);
3336 objp = objpp[i];
3337
Joonsoo Kim8456a642013-10-24 10:07:49 +09003338 page = virt_to_head_page(objp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003339 n = cachep->node[node];
Joonsoo Kim8456a642013-10-24 10:07:49 +09003340 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003341 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003342 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003344 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345
3346 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003347 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003348 if (n->free_objects > n->free_limit) {
3349 n->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003350 /* No need to drop any previously held
3351 * lock here, even if we have a off-slab slab
3352 * descriptor it is guaranteed to come from
3353 * a different cache, refer to comments before
3354 * alloc_slabmgmt.
3355 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003356 slab_destroy(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003358 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359 }
3360 } else {
3361 /* Unconditionally move a slab to the end of the
3362 * partial list on free - maximum time for the
3363 * other objects to be freed, too.
3364 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003365 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 }
3367 }
3368}
3369
Pekka Enberg343e0d72006-02-01 03:05:50 -08003370static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371{
3372 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003373 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003374 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375
3376 batchcount = ac->batchcount;
3377#if DEBUG
3378 BUG_ON(!batchcount || batchcount > ac->avail);
3379#endif
3380 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003381 n = cachep->node[node];
3382 spin_lock(&n->list_lock);
3383 if (n->shared) {
3384 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003385 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386 if (max) {
3387 if (batchcount > max)
3388 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003389 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003390 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 shared_array->avail += batchcount;
3392 goto free_done;
3393 }
3394 }
3395
Christoph Lameterff694162005-09-22 21:44:02 -07003396 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003397free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398#if STATS
3399 {
3400 int i = 0;
3401 struct list_head *p;
3402
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003403 p = n->slabs_free.next;
3404 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003405 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406
Joonsoo Kim8456a642013-10-24 10:07:49 +09003407 page = list_entry(p, struct page, lru);
3408 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409
3410 i++;
3411 p = p->next;
3412 }
3413 STATS_SET_FREEABLE(cachep, i);
3414 }
3415#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003416 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003418 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419}
3420
3421/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003422 * Release an obj back to its cache. If the obj has a constructed state, it must
3423 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003425static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003426 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003428 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429
3430 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003431 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003432 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003434 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003435
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003436 /*
3437 * Skip calling cache_free_alien() when the platform is not numa.
3438 * This will avoid cache misses that happen while accessing slabp (which
3439 * is per page memory reference) to get nodeid. Instead use a global
3440 * variable to skip the call, which is mostly likely to be present in
3441 * the cache.
3442 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003443 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003444 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003445
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 if (likely(ac->avail < ac->limit)) {
3447 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 } else {
3449 STATS_INC_FREEMISS(cachep);
3450 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003452
Mel Gorman072bb0a2012-07-31 16:43:58 -07003453 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454}
3455
3456/**
3457 * kmem_cache_alloc - Allocate an object
3458 * @cachep: The cache to allocate from.
3459 * @flags: See kmalloc().
3460 *
3461 * Allocate an object from this cache. The flags are only relevant
3462 * if the cache has no available objects.
3463 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003464void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003466 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003467
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003468 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003469 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003470
3471 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472}
3473EXPORT_SYMBOL(kmem_cache_alloc);
3474
Li Zefan0f24f122009-12-11 15:45:30 +08003475#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003476void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003477kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003478{
Steven Rostedt85beb582010-11-24 16:23:34 -05003479 void *ret;
3480
Ezequiel Garcia48356302012-09-08 17:47:57 -03003481 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003482
3483 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003484 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003485 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003486}
Steven Rostedt85beb582010-11-24 16:23:34 -05003487EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003488#endif
3489
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003491/**
3492 * kmem_cache_alloc_node - Allocate an object on the specified node
3493 * @cachep: The cache to allocate from.
3494 * @flags: See kmalloc().
3495 * @nodeid: node number of the target node.
3496 *
3497 * Identical to kmem_cache_alloc but it will allocate memory on the given
3498 * node, which can improve the performance for cpu bound structures.
3499 *
3500 * Fallback to other node is possible if __GFP_THISNODE is not set.
3501 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003502void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3503{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003504 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003505
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003506 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003507 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003508 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003509
3510 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003511}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512EXPORT_SYMBOL(kmem_cache_alloc_node);
3513
Li Zefan0f24f122009-12-11 15:45:30 +08003514#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003515void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003516 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003517 int nodeid,
3518 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003519{
Steven Rostedt85beb582010-11-24 16:23:34 -05003520 void *ret;
3521
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003522 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003523
Steven Rostedt85beb582010-11-24 16:23:34 -05003524 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003525 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003526 flags, nodeid);
3527 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003528}
Steven Rostedt85beb582010-11-24 16:23:34 -05003529EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003530#endif
3531
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003532static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003533__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003534{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003535 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003536
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003537 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003538 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3539 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003540 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003541}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003542
Li Zefan0bb38a52009-12-11 15:45:50 +08003543#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003544void *__kmalloc_node(size_t size, gfp_t flags, int node)
3545{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003546 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003547}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003548EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003549
3550void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003551 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003552{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003553 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003554}
3555EXPORT_SYMBOL(__kmalloc_node_track_caller);
3556#else
3557void *__kmalloc_node(size_t size, gfp_t flags, int node)
3558{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003559 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003560}
3561EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003562#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003563#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564
3565/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003566 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003568 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003569 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003571static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003572 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003574 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003575 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003577 /* If you want to save a few bytes .text space: replace
3578 * __ with kmem_.
3579 * Then kmalloc uses the uninlined functions instead of the inline
3580 * functions.
3581 */
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003582 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003583 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3584 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003585 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003586
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003587 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003588 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003589
3590 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003591}
3592
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003593
Li Zefan0bb38a52009-12-11 15:45:50 +08003594#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003595void *__kmalloc(size_t size, gfp_t flags)
3596{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003597 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598}
3599EXPORT_SYMBOL(__kmalloc);
3600
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003601void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003602{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003603 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003604}
3605EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003606
3607#else
3608void *__kmalloc(size_t size, gfp_t flags)
3609{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003610 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003611}
3612EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003613#endif
3614
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615/**
3616 * kmem_cache_free - Deallocate an object
3617 * @cachep: The cache the allocation was from.
3618 * @objp: The previously allocated object.
3619 *
3620 * Free an object which was previously allocated from this
3621 * cache.
3622 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003623void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624{
3625 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003626 cachep = cache_from_obj(cachep, objp);
3627 if (!cachep)
3628 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629
3630 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003631 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003632 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003633 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003634 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003636
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003637 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638}
3639EXPORT_SYMBOL(kmem_cache_free);
3640
3641/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 * kfree - free previously allocated memory
3643 * @objp: pointer returned by kmalloc.
3644 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003645 * If @objp is NULL, no operation is performed.
3646 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647 * Don't free memory not originally allocated by kmalloc()
3648 * or you will run into trouble.
3649 */
3650void kfree(const void *objp)
3651{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003652 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 unsigned long flags;
3654
Pekka Enberg2121db72009-03-25 11:05:57 +02003655 trace_kfree(_RET_IP_, objp);
3656
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003657 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658 return;
3659 local_irq_save(flags);
3660 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003661 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003662 debug_check_no_locks_freed(objp, c->object_size);
3663
3664 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003665 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666 local_irq_restore(flags);
3667}
3668EXPORT_SYMBOL(kfree);
3669
Christoph Lametere498be72005-09-09 13:03:32 -07003670/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003671 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003672 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003673static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003674{
3675 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003676 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003677 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003678 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003679
Mel Gorman9c09a952008-01-24 05:49:54 -08003680 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003681
Paul Menage3395ee02006-12-06 20:32:16 -08003682 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003683 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003684 if (!new_alien)
3685 goto fail;
3686 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003687
Eric Dumazet63109842007-05-06 14:49:28 -07003688 new_shared = NULL;
3689 if (cachep->shared) {
3690 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003691 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003692 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003693 if (!new_shared) {
3694 free_alien_cache(new_alien);
3695 goto fail;
3696 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003697 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003698
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003699 n = cachep->node[node];
3700 if (n) {
3701 struct array_cache *shared = n->shared;
Christoph Lametercafeb022006-03-25 03:06:46 -08003702
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003703 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003704
Christoph Lametercafeb022006-03-25 03:06:46 -08003705 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003706 free_block(cachep, shared->entry,
3707 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003708
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003709 n->shared = new_shared;
3710 if (!n->alien) {
3711 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003712 new_alien = NULL;
3713 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003714 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003715 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003716 spin_unlock_irq(&n->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003717 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003718 free_alien_cache(new_alien);
3719 continue;
3720 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003721 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3722 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003723 free_alien_cache(new_alien);
3724 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003725 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003726 }
Christoph Lametere498be72005-09-09 13:03:32 -07003727
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003728 kmem_cache_node_init(n);
3729 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003730 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003731 n->shared = new_shared;
3732 n->alien = new_alien;
3733 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003734 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003735 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003736 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003737 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003738
Andrew Mortona737b3e2006-03-22 00:08:11 -08003739fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003740 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003741 /* Cache is not active yet. Roll back what we did */
3742 node--;
3743 while (node >= 0) {
Christoph Lameter6a673682013-01-10 19:14:19 +00003744 if (cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003745 n = cachep->node[node];
Christoph Lameter0718dc22006-03-25 03:06:47 -08003746
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003747 kfree(n->shared);
3748 free_alien_cache(n->alien);
3749 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003750 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003751 }
3752 node--;
3753 }
3754 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003755 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003756}
3757
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003759 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003760 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761};
3762
3763static void do_ccupdate_local(void *info)
3764{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003765 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 struct array_cache *old;
3767
3768 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003769 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003770
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3772 new->new[smp_processor_id()] = old;
3773}
3774
Christoph Lameter18004c52012-07-06 15:25:12 -05003775/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003776static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003777 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003779 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003780 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003782 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3783 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003784 if (!new)
3785 return -ENOMEM;
3786
Christoph Lametere498be72005-09-09 13:03:32 -07003787 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003788 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003789 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003790 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003791 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003792 kfree(new->new[i]);
3793 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003794 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795 }
3796 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003797 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003799 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003800
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 cachep->batchcount = batchcount;
3803 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003804 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805
Christoph Lametere498be72005-09-09 13:03:32 -07003806 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003807 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 if (!ccold)
3809 continue;
Christoph Lameter6a673682013-01-10 19:14:19 +00003810 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003811 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
Christoph Lameter6a673682013-01-10 19:14:19 +00003812 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 kfree(ccold);
3814 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003815 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003816 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817}
3818
Glauber Costa943a4512012-12-18 14:23:03 -08003819static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3820 int batchcount, int shared, gfp_t gfp)
3821{
3822 int ret;
3823 struct kmem_cache *c = NULL;
3824 int i = 0;
3825
3826 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3827
3828 if (slab_state < FULL)
3829 return ret;
3830
3831 if ((ret < 0) || !is_root_cache(cachep))
3832 return ret;
3833
Glauber Costaebe945c2012-12-18 14:23:10 -08003834 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003835 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003836 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003837 if (c)
3838 /* return value determined by the parent cache only */
3839 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3840 }
3841
3842 return ret;
3843}
3844
Christoph Lameter18004c52012-07-06 15:25:12 -05003845/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003846static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847{
3848 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003849 int limit = 0;
3850 int shared = 0;
3851 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852
Glauber Costa943a4512012-12-18 14:23:03 -08003853 if (!is_root_cache(cachep)) {
3854 struct kmem_cache *root = memcg_root_cache(cachep);
3855 limit = root->limit;
3856 shared = root->shared;
3857 batchcount = root->batchcount;
3858 }
3859
3860 if (limit && shared && batchcount)
3861 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003862 /*
3863 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 * - create a LIFO ordering, i.e. return objects that are cache-warm
3865 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003866 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 * bufctl chains: array operations are cheaper.
3868 * The numbers are guessed, we should auto-tune as described by
3869 * Bonwick.
3870 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003871 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003873 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003875 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003877 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 limit = 54;
3879 else
3880 limit = 120;
3881
Andrew Mortona737b3e2006-03-22 00:08:11 -08003882 /*
3883 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 * allocation behaviour: Most allocs on one cpu, most free operations
3885 * on another cpu. For these cases, an efficient object passing between
3886 * cpus is necessary. This is provided by a shared array. The array
3887 * replaces Bonwick's magazine layer.
3888 * On uniprocessor, it's functionally equivalent (but less efficient)
3889 * to a larger limit. Thus disabled by default.
3890 */
3891 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003892 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894
3895#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003896 /*
3897 * With debugging enabled, large batchcount lead to excessively long
3898 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899 */
3900 if (limit > 32)
3901 limit = 32;
3902#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003903 batchcount = (limit + 1) / 2;
3904skip_setup:
3905 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906 if (err)
3907 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003908 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003909 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910}
3911
Christoph Lameter1b552532006-03-22 00:09:07 -08003912/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003913 * Drain an array if it contains any elements taking the node lock only if
3914 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003915 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003916 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003917static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003918 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919{
3920 int tofree;
3921
Christoph Lameter1b552532006-03-22 00:09:07 -08003922 if (!ac || !ac->avail)
3923 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 if (ac->touched && !force) {
3925 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003926 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003927 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003928 if (ac->avail) {
3929 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3930 if (tofree > ac->avail)
3931 tofree = (ac->avail + 1) / 2;
3932 free_block(cachep, ac->entry, tofree, node);
3933 ac->avail -= tofree;
3934 memmove(ac->entry, &(ac->entry[tofree]),
3935 sizeof(void *) * ac->avail);
3936 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003937 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938 }
3939}
3940
3941/**
3942 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003943 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 *
3945 * Called from workqueue/eventd every few seconds.
3946 * Purpose:
3947 * - clear the per-cpu caches for this CPU.
3948 * - return freeable pages to the main free memory pool.
3949 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003950 * If we cannot acquire the cache chain mutex then just give up - we'll try
3951 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003953static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003955 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003956 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003957 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003958 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959
Christoph Lameter18004c52012-07-06 15:25:12 -05003960 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003962 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963
Christoph Lameter18004c52012-07-06 15:25:12 -05003964 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 check_irq_on();
3966
Christoph Lameter35386e32006-03-22 00:09:05 -08003967 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003968 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003969 * have established with reasonable certainty that
3970 * we can do some work if the lock was obtained.
3971 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003972 n = searchp->node[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003973
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003974 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003976 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977
Christoph Lameter35386e32006-03-22 00:09:05 -08003978 /*
3979 * These are racy checks but it does not matter
3980 * if we skip one check or scan twice.
3981 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003982 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003983 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003985 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003987 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003989 if (n->free_touched)
3990 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003991 else {
3992 int freed;
3993
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003994 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07003995 5 * searchp->num - 1) / (5 * searchp->num));
3996 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003998next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 cond_resched();
4000 }
4001 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004002 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004003 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004004out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004005 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004006 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007}
4008
Linus Torvalds158a9622008-01-02 13:04:48 -08004009#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004010void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004012 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004013 unsigned long active_objs;
4014 unsigned long num_objs;
4015 unsigned long active_slabs = 0;
4016 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004017 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004019 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004020 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 active_objs = 0;
4023 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004024 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004025 n = cachep->node[node];
4026 if (!n)
Christoph Lametere498be72005-09-09 13:03:32 -07004027 continue;
4028
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004029 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004030 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004031
Joonsoo Kim8456a642013-10-24 10:07:49 +09004032 list_for_each_entry(page, &n->slabs_full, lru) {
4033 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004034 error = "slabs_full accounting error";
4035 active_objs += cachep->num;
4036 active_slabs++;
4037 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004038 list_for_each_entry(page, &n->slabs_partial, lru) {
4039 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004040 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004041 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004042 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004043 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004044 active_slabs++;
4045 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004046 list_for_each_entry(page, &n->slabs_free, lru) {
4047 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004048 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004049 num_slabs++;
4050 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004051 free_objects += n->free_objects;
4052 if (n->shared)
4053 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004054
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004055 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004057 num_slabs += active_slabs;
4058 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004059 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 error = "free_objects accounting error";
4061
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004062 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 if (error)
4064 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4065
Glauber Costa0d7561c2012-10-19 18:20:27 +04004066 sinfo->active_objs = active_objs;
4067 sinfo->num_objs = num_objs;
4068 sinfo->active_slabs = active_slabs;
4069 sinfo->num_slabs = num_slabs;
4070 sinfo->shared_avail = shared_avail;
4071 sinfo->limit = cachep->limit;
4072 sinfo->batchcount = cachep->batchcount;
4073 sinfo->shared = cachep->shared;
4074 sinfo->objects_per_slab = cachep->num;
4075 sinfo->cache_order = cachep->gfporder;
4076}
4077
4078void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4079{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004081 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082 unsigned long high = cachep->high_mark;
4083 unsigned long allocs = cachep->num_allocations;
4084 unsigned long grown = cachep->grown;
4085 unsigned long reaped = cachep->reaped;
4086 unsigned long errors = cachep->errors;
4087 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004089 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004090 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091
Joe Perchese92dd4f2010-03-26 19:27:58 -07004092 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4093 "%4lu %4lu %4lu %4lu %4lu",
4094 allocs, high, grown,
4095 reaped, errors, max_freeable, node_allocs,
4096 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 }
4098 /* cpu stats */
4099 {
4100 unsigned long allochit = atomic_read(&cachep->allochit);
4101 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4102 unsigned long freehit = atomic_read(&cachep->freehit);
4103 unsigned long freemiss = atomic_read(&cachep->freemiss);
4104
4105 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004106 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 }
4108#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109}
4110
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111#define MAX_SLABINFO_WRITE 128
4112/**
4113 * slabinfo_write - Tuning for the slab allocator
4114 * @file: unused
4115 * @buffer: user buffer
4116 * @count: data length
4117 * @ppos: unused
4118 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004119ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004120 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004122 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004124 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004125
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126 if (count > MAX_SLABINFO_WRITE)
4127 return -EINVAL;
4128 if (copy_from_user(&kbuf, buffer, count))
4129 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004130 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131
4132 tmp = strchr(kbuf, ' ');
4133 if (!tmp)
4134 return -EINVAL;
4135 *tmp = '\0';
4136 tmp++;
4137 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4138 return -EINVAL;
4139
4140 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004141 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004143 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004145 if (limit < 1 || batchcount < 1 ||
4146 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004147 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004149 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004150 batchcount, shared,
4151 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152 }
4153 break;
4154 }
4155 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004156 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157 if (res >= 0)
4158 res = count;
4159 return res;
4160}
Al Viro871751e2006-03-25 03:06:39 -08004161
4162#ifdef CONFIG_DEBUG_SLAB_LEAK
4163
4164static void *leaks_start(struct seq_file *m, loff_t *pos)
4165{
Christoph Lameter18004c52012-07-06 15:25:12 -05004166 mutex_lock(&slab_mutex);
4167 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004168}
4169
4170static inline int add_caller(unsigned long *n, unsigned long v)
4171{
4172 unsigned long *p;
4173 int l;
4174 if (!v)
4175 return 1;
4176 l = n[1];
4177 p = n + 2;
4178 while (l) {
4179 int i = l/2;
4180 unsigned long *q = p + 2 * i;
4181 if (*q == v) {
4182 q[1]++;
4183 return 1;
4184 }
4185 if (*q > v) {
4186 l = i;
4187 } else {
4188 p = q + 2;
4189 l -= i + 1;
4190 }
4191 }
4192 if (++n[1] == n[0])
4193 return 0;
4194 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4195 p[0] = v;
4196 p[1] = 1;
4197 return 1;
4198}
4199
Joonsoo Kim8456a642013-10-24 10:07:49 +09004200static void handle_slab(unsigned long *n, struct kmem_cache *c,
4201 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004202{
4203 void *p;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004204 int i, j;
4205
Al Viro871751e2006-03-25 03:06:39 -08004206 if (n[0] == n[1])
4207 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004208 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004209 bool active = true;
4210
Joonsoo Kim8456a642013-10-24 10:07:49 +09004211 for (j = page->active; j < c->num; j++) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004212 /* Skip freed item */
Joonsoo Kime7444d92013-10-24 10:07:51 +09004213 if (slab_freelist(page)[j] == i) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004214 active = false;
4215 break;
4216 }
4217 }
4218 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004219 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004220
Al Viro871751e2006-03-25 03:06:39 -08004221 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4222 return;
4223 }
4224}
4225
4226static void show_symbol(struct seq_file *m, unsigned long address)
4227{
4228#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004229 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004230 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004231
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004232 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004233 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004234 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004235 seq_printf(m, " [%s]", modname);
4236 return;
4237 }
4238#endif
4239 seq_printf(m, "%p", (void *)address);
4240}
4241
4242static int leaks_show(struct seq_file *m, void *p)
4243{
Thierry Reding0672aa72012-06-22 19:42:49 +02004244 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004245 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004246 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004247 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004248 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004249 int node;
4250 int i;
4251
4252 if (!(cachep->flags & SLAB_STORE_USER))
4253 return 0;
4254 if (!(cachep->flags & SLAB_RED_ZONE))
4255 return 0;
4256
4257 /* OK, we can do it */
4258
Christoph Lameterdb845062013-02-05 18:45:23 +00004259 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004260
4261 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004262 n = cachep->node[node];
4263 if (!n)
Al Viro871751e2006-03-25 03:06:39 -08004264 continue;
4265
4266 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004267 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004268
Joonsoo Kim8456a642013-10-24 10:07:49 +09004269 list_for_each_entry(page, &n->slabs_full, lru)
4270 handle_slab(x, cachep, page);
4271 list_for_each_entry(page, &n->slabs_partial, lru)
4272 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004273 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004274 }
4275 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004276 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004277 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004278 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004279 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004280 if (!m->private) {
4281 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004282 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004283 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004284 return -ENOMEM;
4285 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004286 *(unsigned long *)m->private = x[0] * 2;
4287 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004288 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004289 /* Now make sure this entry will be retried */
4290 m->count = m->size;
4291 return 0;
4292 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004293 for (i = 0; i < x[1]; i++) {
4294 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4295 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004296 seq_putc(m, '\n');
4297 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004298
Al Viro871751e2006-03-25 03:06:39 -08004299 return 0;
4300}
4301
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004302static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004303 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004304 .next = slab_next,
4305 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004306 .show = leaks_show,
4307};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004308
4309static int slabstats_open(struct inode *inode, struct file *file)
4310{
4311 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4312 int ret = -ENOMEM;
4313 if (n) {
4314 ret = seq_open(file, &slabstats_op);
4315 if (!ret) {
4316 struct seq_file *m = file->private_data;
4317 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4318 m->private = n;
4319 n = NULL;
4320 }
4321 kfree(n);
4322 }
4323 return ret;
4324}
4325
4326static const struct file_operations proc_slabstats_operations = {
4327 .open = slabstats_open,
4328 .read = seq_read,
4329 .llseek = seq_lseek,
4330 .release = seq_release_private,
4331};
Al Viro871751e2006-03-25 03:06:39 -08004332#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004333
4334static int __init slab_proc_init(void)
4335{
4336#ifdef CONFIG_DEBUG_SLAB_LEAK
4337 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4338#endif
4339 return 0;
4340}
4341module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004342#endif
4343
Manfred Spraul00e145b2005-09-03 15:55:07 -07004344/**
4345 * ksize - get the actual amount of memory allocated for a given object
4346 * @objp: Pointer to the object
4347 *
4348 * kmalloc may internally round up allocations and return more memory
4349 * than requested. ksize() can be used to determine the actual amount of
4350 * memory allocated. The caller may use this additional memory, even though
4351 * a smaller amount of memory was initially specified with the kmalloc call.
4352 * The caller must guarantee that objp points to a valid object previously
4353 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4354 * must not be freed during the duration of the call.
4355 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004356size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004357{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004358 BUG_ON(!objp);
4359 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004360 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004361
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004362 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004363}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004364EXPORT_SYMBOL(ksize);