blob: 3070b929a1bfa67778e415525403e9b36e392344 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Christoph Lameter18004c52012-07-06 15:25:12 -050071 * The global cache-chain is protected by the mutex 'slab_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100109#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200116#include <linux/kmemcheck.h>
David Rientjes8f9f8d92010-03-27 19:40:47 -0700117#include <linux/memory.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -0700118#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
Mel Gorman381760e2012-07-31 16:44:30 -0700120#include <net/sock.h>
121
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500126#include <trace/events/kmem.h>
127
Mel Gorman072bb0a2012-07-31 16:43:58 -0700128#include "internal.h"
129
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800130#include "slab.h"
131
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900160#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
162
163#if FREELIST_BYTE_INDEX
164typedef unsigned char freelist_idx_t;
165#else
166typedef unsigned short freelist_idx_t;
167#endif
168
David Miller30321c72014-05-05 16:20:04 -0400169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900170
Mel Gorman072bb0a2012-07-31 16:43:58 -0700171/*
172 * true if a page was allocated from pfmemalloc reserves for network-based
173 * swap
174 */
175static bool pfmemalloc_active __read_mostly;
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 * struct array_cache
179 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 * Purpose:
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
184 *
185 * The limit is stored in the per-cpu structure to reduce the data cache
186 * footprint.
187 *
188 */
189struct array_cache {
190 unsigned int avail;
191 unsigned int limit;
192 unsigned int batchcount;
193 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700194 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700195 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800196 * Must have this definition in here for the proper
197 * alignment of array_cache. Also simplifies accessing
198 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700199 *
200 * Entries should not be directly dereferenced as
201 * entries belonging to slabs marked pfmemalloc will
202 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800203 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204};
205
Mel Gorman072bb0a2012-07-31 16:43:58 -0700206#define SLAB_OBJ_PFMEMALLOC 1
207static inline bool is_obj_pfmemalloc(void *objp)
208{
209 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
210}
211
212static inline void set_obj_pfmemalloc(void **objp)
213{
214 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
215 return;
216}
217
218static inline void clear_obj_pfmemalloc(void **objp)
219{
220 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
221}
222
Andrew Mortona737b3e2006-03-22 00:08:11 -0800223/*
224 * bootstrap: The caches do not work without cpuarrays anymore, but the
225 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 */
227#define BOOT_CPUCACHE_ENTRIES 1
228struct arraycache_init {
229 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800230 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231};
232
233/*
Christoph Lametere498be72005-09-09 13:03:32 -0700234 * Need this for bootstrapping a per node allocator.
235 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200236#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000237static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700238#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200239#define SIZE_AC MAX_NUMNODES
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000240#define SIZE_NODE (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241
Christoph Lametered11d9e2006-06-30 01:55:45 -0700242static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000243 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700244static void free_block(struct kmem_cache *cachep, void **objpp, int len,
245 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300246static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000247static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700248
Ingo Molnare0a42722006-06-23 02:03:46 -0700249static int slab_early_init = 1;
250
Christoph Lametere3366012013-01-10 19:14:18 +0000251#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000252#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700253
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000254static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700255{
256 INIT_LIST_HEAD(&parent->slabs_full);
257 INIT_LIST_HEAD(&parent->slabs_partial);
258 INIT_LIST_HEAD(&parent->slabs_free);
259 parent->shared = NULL;
260 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800261 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700262 spin_lock_init(&parent->list_lock);
263 parent->free_objects = 0;
264 parent->free_touched = 0;
265}
266
Andrew Mortona737b3e2006-03-22 00:08:11 -0800267#define MAKE_LIST(cachep, listp, slab, nodeid) \
268 do { \
269 INIT_LIST_HEAD(listp); \
Christoph Lameter6a673682013-01-10 19:14:19 +0000270 list_splice(&(cachep->node[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700271 } while (0)
272
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
274 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700275 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
276 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
277 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
278 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280#define CFLGS_OFF_SLAB (0x80000000UL)
281#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
282
283#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800284/*
285 * Optimization question: fewer reaps means less probability for unnessary
286 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100288 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 * which could lock up otherwise freeable slabs.
290 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800291#define REAPTIMEOUT_AC (2*HZ)
292#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293
294#if STATS
295#define STATS_INC_ACTIVE(x) ((x)->num_active++)
296#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
297#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
298#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700299#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800300#define STATS_SET_HIGH(x) \
301 do { \
302 if ((x)->num_active > (x)->high_mark) \
303 (x)->high_mark = (x)->num_active; \
304 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305#define STATS_INC_ERR(x) ((x)->errors++)
306#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700307#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700308#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800309#define STATS_SET_FREEABLE(x, i) \
310 do { \
311 if ((x)->max_freeable < i) \
312 (x)->max_freeable = i; \
313 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
315#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
316#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
317#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
318#else
319#define STATS_INC_ACTIVE(x) do { } while (0)
320#define STATS_DEC_ACTIVE(x) do { } while (0)
321#define STATS_INC_ALLOCED(x) do { } while (0)
322#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700323#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324#define STATS_SET_HIGH(x) do { } while (0)
325#define STATS_INC_ERR(x) do { } while (0)
326#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700327#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700328#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800329#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330#define STATS_INC_ALLOCHIT(x) do { } while (0)
331#define STATS_INC_ALLOCMISS(x) do { } while (0)
332#define STATS_INC_FREEHIT(x) do { } while (0)
333#define STATS_INC_FREEMISS(x) do { } while (0)
334#endif
335
336#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
Andrew Mortona737b3e2006-03-22 00:08:11 -0800338/*
339 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800341 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 * the end of an object is aligned with the end of the real
343 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800344 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800346 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500347 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
348 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800349 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800351static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800353 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354}
355
David Woodhouseb46b8f12007-05-08 00:22:59 -0700356static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357{
358 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700359 return (unsigned long long*) (objp + obj_offset(cachep) -
360 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361}
362
David Woodhouseb46b8f12007-05-08 00:22:59 -0700363static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364{
365 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
366 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500367 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700368 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400369 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500370 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700371 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372}
373
Pekka Enberg343e0d72006-02-01 03:05:50 -0800374static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375{
376 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500377 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378}
379
380#else
381
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800382#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700383#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
384#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
386
387#endif
388
Joonsoo Kim03787302014-06-23 13:22:06 -0700389#define OBJECT_FREE (0)
390#define OBJECT_ACTIVE (1)
391
392#ifdef CONFIG_DEBUG_SLAB_LEAK
393
394static void set_obj_status(struct page *page, int idx, int val)
395{
396 int freelist_size;
397 char *status;
398 struct kmem_cache *cachep = page->slab_cache;
399
400 freelist_size = cachep->num * sizeof(freelist_idx_t);
401 status = (char *)page->freelist + freelist_size;
402 status[idx] = val;
403}
404
405static inline unsigned int get_obj_status(struct page *page, int idx)
406{
407 int freelist_size;
408 char *status;
409 struct kmem_cache *cachep = page->slab_cache;
410
411 freelist_size = cachep->num * sizeof(freelist_idx_t);
412 status = (char *)page->freelist + freelist_size;
413
414 return status[idx];
415}
416
417#else
418static inline void set_obj_status(struct page *page, int idx, int val) {}
419
420#endif
421
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700423 * Do not go above this order unless 0 objects fit into the slab or
424 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425 */
David Rientjes543585c2011-10-18 22:09:24 -0700426#define SLAB_MAX_ORDER_HI 1
427#define SLAB_MAX_ORDER_LO 0
428static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700429static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800431static inline struct kmem_cache *virt_to_cache(const void *obj)
432{
Christoph Lameterb49af682007-05-06 14:49:41 -0700433 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500434 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800435}
436
Joonsoo Kim8456a642013-10-24 10:07:49 +0900437static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800438 unsigned int idx)
439{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900440 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800441}
442
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800443/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500444 * We want to avoid an expensive divide : (offset / cache->size)
445 * Using the fact that size is a constant for a particular cache,
446 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800447 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
448 */
449static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900450 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800451{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900452 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800453 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800454}
455
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800457 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458
459/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000460static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800461 .batchcount = 1,
462 .limit = BOOT_CPUCACHE_ENTRIES,
463 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500464 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800465 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466};
467
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700468#define BAD_ALIEN_MAGIC 0x01020304ul
469
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200470#ifdef CONFIG_LOCKDEP
471
472/*
473 * Slab sometimes uses the kmalloc slabs to store the slab headers
474 * for other slabs "off slab".
475 * The locking for this is tricky in that it nests within the locks
476 * of all other slabs in a few places; to deal with this special
477 * locking we put on-slab caches into a separate lock-class.
478 *
479 * We set lock class for alien array caches which are up during init.
480 * The lock annotation will be lost if all cpus of a node goes down and
481 * then comes back up during hotplug
482 */
483static struct lock_class_key on_slab_l3_key;
484static struct lock_class_key on_slab_alc_key;
485
Peter Zijlstra83835b32011-07-22 15:26:05 +0200486static struct lock_class_key debugobj_l3_key;
487static struct lock_class_key debugobj_alc_key;
488
489static void slab_set_lock_classes(struct kmem_cache *cachep,
490 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
491 int q)
492{
493 struct array_cache **alc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000494 struct kmem_cache_node *n;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200495 int r;
496
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000497 n = cachep->node[q];
498 if (!n)
Peter Zijlstra83835b32011-07-22 15:26:05 +0200499 return;
500
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000501 lockdep_set_class(&n->list_lock, l3_key);
502 alc = n->alien;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200503 /*
504 * FIXME: This check for BAD_ALIEN_MAGIC
505 * should go away when common slab code is taught to
506 * work even without alien caches.
507 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
508 * for alloc_alien_cache,
509 */
510 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
511 return;
512 for_each_node(r) {
513 if (alc[r])
514 lockdep_set_class(&alc[r]->lock, alc_key);
515 }
516}
517
518static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
519{
520 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
521}
522
523static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
524{
525 int node;
526
527 for_each_online_node(node)
528 slab_set_debugobj_lock_classes_node(cachep, node);
529}
530
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200531static void init_node_lock_keys(int q)
532{
Christoph Lametere3366012013-01-10 19:14:18 +0000533 int i;
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200534
Christoph Lameter97d06602012-07-06 15:25:11 -0500535 if (slab_state < UP)
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200536 return;
537
Christoph Lameter0f8f8092013-07-02 12:12:10 -0700538 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000539 struct kmem_cache_node *n;
Christoph Lametere3366012013-01-10 19:14:18 +0000540 struct kmem_cache *cache = kmalloc_caches[i];
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200541
Christoph Lametere3366012013-01-10 19:14:18 +0000542 if (!cache)
Pekka Enberg00afa752009-12-27 14:33:14 +0200543 continue;
Peter Zijlstra83835b32011-07-22 15:26:05 +0200544
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000545 n = cache->node[q];
546 if (!n || OFF_SLAB(cache))
Christoph Lametere3366012013-01-10 19:14:18 +0000547 continue;
548
549 slab_set_lock_classes(cache, &on_slab_l3_key,
Peter Zijlstra83835b32011-07-22 15:26:05 +0200550 &on_slab_alc_key, q);
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200551 }
552}
553
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800554static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
555{
Christoph Lameter6a673682013-01-10 19:14:19 +0000556 if (!cachep->node[q])
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800557 return;
558
559 slab_set_lock_classes(cachep, &on_slab_l3_key,
560 &on_slab_alc_key, q);
561}
562
563static inline void on_slab_lock_classes(struct kmem_cache *cachep)
564{
565 int node;
566
567 VM_BUG_ON(OFF_SLAB(cachep));
568 for_each_node(node)
569 on_slab_lock_classes_node(cachep, node);
570}
571
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200572static inline void init_lock_keys(void)
573{
574 int node;
575
576 for_each_node(node)
577 init_node_lock_keys(node);
578}
579#else
580static void init_node_lock_keys(int q)
581{
582}
583
584static inline void init_lock_keys(void)
585{
586}
Peter Zijlstra83835b32011-07-22 15:26:05 +0200587
Glauber Costa6ccfb5b2012-12-18 14:22:31 -0800588static inline void on_slab_lock_classes(struct kmem_cache *cachep)
589{
590}
591
592static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
593{
594}
595
Peter Zijlstra83835b32011-07-22 15:26:05 +0200596static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
597{
598}
599
600static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
601{
602}
Pekka Enbergce79ddc2009-11-23 22:01:15 +0200603#endif
604
Tejun Heo1871e522009-10-29 22:34:13 +0900605static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Pekka Enberg343e0d72006-02-01 03:05:50 -0800607static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608{
609 return cachep->array[smp_processor_id()];
610}
611
Joonsoo Kim03787302014-06-23 13:22:06 -0700612static size_t calculate_freelist_size(int nr_objs, size_t align)
613{
614 size_t freelist_size;
615
616 freelist_size = nr_objs * sizeof(freelist_idx_t);
617 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
618 freelist_size += nr_objs * sizeof(char);
619
620 if (align)
621 freelist_size = ALIGN(freelist_size, align);
622
623 return freelist_size;
624}
625
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900626static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
627 size_t idx_size, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628{
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900629 int nr_objs;
Joonsoo Kim03787302014-06-23 13:22:06 -0700630 size_t remained_size;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900631 size_t freelist_size;
Joonsoo Kim03787302014-06-23 13:22:06 -0700632 int extra_space = 0;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900633
Joonsoo Kim03787302014-06-23 13:22:06 -0700634 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
635 extra_space = sizeof(char);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900636 /*
637 * Ignore padding for the initial guess. The padding
638 * is at most @align-1 bytes, and @buffer_size is at
639 * least @align. In the worst case, this result will
640 * be one greater than the number of objects that fit
641 * into the memory allocation when taking the padding
642 * into account.
643 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700644 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900645
646 /*
647 * This calculated number will be either the right
648 * amount, or one greater than what we want.
649 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700650 remained_size = slab_size - nr_objs * buffer_size;
651 freelist_size = calculate_freelist_size(nr_objs, align);
652 if (remained_size < freelist_size)
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900653 nr_objs--;
654
655 return nr_objs;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800656}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
Andrew Mortona737b3e2006-03-22 00:08:11 -0800658/*
659 * Calculate the number of objects and left-over bytes for a given buffer size.
660 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800661static void cache_estimate(unsigned long gfporder, size_t buffer_size,
662 size_t align, int flags, size_t *left_over,
663 unsigned int *num)
664{
665 int nr_objs;
666 size_t mgmt_size;
667 size_t slab_size = PAGE_SIZE << gfporder;
668
669 /*
670 * The slab management structure can be either off the slab or
671 * on it. For the latter case, the memory allocated for a
672 * slab is used for:
673 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900674 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800675 * - Padding to respect alignment of @align
676 * - @buffer_size bytes for each object
677 *
678 * If the slab management structure is off the slab, then the
679 * alignment will already be calculated into the size. Because
680 * the slabs are all pages aligned, the objects will be at the
681 * correct alignment when allocated.
682 */
683 if (flags & CFLGS_OFF_SLAB) {
684 mgmt_size = 0;
685 nr_objs = slab_size / buffer_size;
686
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800687 } else {
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900688 nr_objs = calculate_nr_objs(slab_size, buffer_size,
Joonsoo Kima41adfa2013-12-02 17:49:42 +0900689 sizeof(freelist_idx_t), align);
Joonsoo Kim03787302014-06-23 13:22:06 -0700690 mgmt_size = calculate_freelist_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800692 *num = nr_objs;
693 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694}
695
Christoph Lameterf28510d2012-09-11 19:49:38 +0000696#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700697#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
Andrew Mortona737b3e2006-03-22 00:08:11 -0800699static void __slab_error(const char *function, struct kmem_cache *cachep,
700 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701{
702 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800703 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030705 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000707#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
Paul Menage3395ee02006-12-06 20:32:16 -0800709/*
710 * By default on NUMA we use alien caches to stage the freeing of
711 * objects allocated from other nodes. This causes massive memory
712 * inefficiencies when using fake NUMA setup to split memory into a
713 * large number of small nodes, so it can be disabled on the command
714 * line
715 */
716
717static int use_alien_caches __read_mostly = 1;
718static int __init noaliencache_setup(char *s)
719{
720 use_alien_caches = 0;
721 return 1;
722}
723__setup("noaliencache", noaliencache_setup);
724
David Rientjes3df1ccc2011-10-18 22:09:28 -0700725static int __init slab_max_order_setup(char *str)
726{
727 get_option(&str, &slab_max_order);
728 slab_max_order = slab_max_order < 0 ? 0 :
729 min(slab_max_order, MAX_ORDER - 1);
730 slab_max_order_set = true;
731
732 return 1;
733}
734__setup("slab_max_order=", slab_max_order_setup);
735
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800736#ifdef CONFIG_NUMA
737/*
738 * Special reaping functions for NUMA systems called from cache_reap().
739 * These take care of doing round robin flushing of alien caches (containing
740 * objects freed on different nodes from which they were allocated) and the
741 * flushing of remote pcps by calling drain_node_pages.
742 */
Tejun Heo1871e522009-10-29 22:34:13 +0900743static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800744
745static void init_reap_node(int cpu)
746{
747 int node;
748
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700749 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800750 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800751 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800752
Tejun Heo1871e522009-10-29 22:34:13 +0900753 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800754}
755
756static void next_reap_node(void)
757{
Christoph Lameter909ea962010-12-08 16:22:55 +0100758 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800759
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800760 node = next_node(node, node_online_map);
761 if (unlikely(node >= MAX_NUMNODES))
762 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100763 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800764}
765
766#else
767#define init_reap_node(cpu) do { } while (0)
768#define next_reap_node(void) do { } while (0)
769#endif
770
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771/*
772 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
773 * via the workqueue/eventd.
774 * Add the CPU number into the expiration time to minimize the possibility of
775 * the CPUs getting into lockstep and contending for the global cache chain
776 * lock.
777 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400778static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779{
Tejun Heo1871e522009-10-29 22:34:13 +0900780 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781
782 /*
783 * When this gets called from do_initcalls via cpucache_init(),
784 * init_workqueues() has already run, so keventd will be setup
785 * at that time.
786 */
David Howells52bad642006-11-22 14:54:01 +0000787 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800788 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700789 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800790 schedule_delayed_work_on(cpu, reap_work,
791 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 }
793}
794
Christoph Lametere498be72005-09-09 13:03:32 -0700795static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300796 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800798 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 struct array_cache *nc = NULL;
800
Pekka Enberg83b519e2009-06-10 19:40:04 +0300801 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100802 /*
803 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300804 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100805 * cache the pointers are not cleared and they could be counted as
806 * valid references during a kmemleak scan. Therefore, kmemleak must
807 * not scan such objects.
808 */
809 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 if (nc) {
811 nc->avail = 0;
812 nc->limit = entries;
813 nc->batchcount = batchcount;
814 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700815 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 }
817 return nc;
818}
819
Joonsoo Kim8456a642013-10-24 10:07:49 +0900820static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700821{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700822 return PageSlabPfmemalloc(page);
823}
824
825/* Clears pfmemalloc_active if no slabs have pfmalloc set */
826static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
827 struct array_cache *ac)
828{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000829 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
Joonsoo Kim8456a642013-10-24 10:07:49 +0900830 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700831 unsigned long flags;
832
833 if (!pfmemalloc_active)
834 return;
835
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000836 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900837 list_for_each_entry(page, &n->slabs_full, lru)
838 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700839 goto out;
840
Joonsoo Kim8456a642013-10-24 10:07:49 +0900841 list_for_each_entry(page, &n->slabs_partial, lru)
842 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700843 goto out;
844
Joonsoo Kim8456a642013-10-24 10:07:49 +0900845 list_for_each_entry(page, &n->slabs_free, lru)
846 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700847 goto out;
848
849 pfmemalloc_active = false;
850out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000851 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700852}
853
Mel Gorman381760e2012-07-31 16:44:30 -0700854static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700855 gfp_t flags, bool force_refill)
856{
857 int i;
858 void *objp = ac->entry[--ac->avail];
859
860 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
861 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000862 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700863
864 if (gfp_pfmemalloc_allowed(flags)) {
865 clear_obj_pfmemalloc(&objp);
866 return objp;
867 }
868
869 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700870 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700871 /* If a !PFMEMALLOC object is found, swap them */
872 if (!is_obj_pfmemalloc(ac->entry[i])) {
873 objp = ac->entry[i];
874 ac->entry[i] = ac->entry[ac->avail];
875 ac->entry[ac->avail] = objp;
876 return objp;
877 }
878 }
879
880 /*
881 * If there are empty slabs on the slabs_free list and we are
882 * being forced to refill the cache, mark this one !pfmemalloc.
883 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000884 n = cachep->node[numa_mem_id()];
885 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900886 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900887 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700888 clear_obj_pfmemalloc(&objp);
889 recheck_pfmemalloc_active(cachep, ac);
890 return objp;
891 }
892
893 /* No !PFMEMALLOC objects available */
894 ac->avail++;
895 objp = NULL;
896 }
897
898 return objp;
899}
900
Mel Gorman381760e2012-07-31 16:44:30 -0700901static inline void *ac_get_obj(struct kmem_cache *cachep,
902 struct array_cache *ac, gfp_t flags, bool force_refill)
903{
904 void *objp;
905
906 if (unlikely(sk_memalloc_socks()))
907 objp = __ac_get_obj(cachep, ac, flags, force_refill);
908 else
909 objp = ac->entry[--ac->avail];
910
911 return objp;
912}
913
914static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700915 void *objp)
916{
917 if (unlikely(pfmemalloc_active)) {
918 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700919 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700920 if (PageSlabPfmemalloc(page))
921 set_obj_pfmemalloc(&objp);
922 }
923
Mel Gorman381760e2012-07-31 16:44:30 -0700924 return objp;
925}
926
927static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
928 void *objp)
929{
930 if (unlikely(sk_memalloc_socks()))
931 objp = __ac_put_obj(cachep, ac, objp);
932
Mel Gorman072bb0a2012-07-31 16:43:58 -0700933 ac->entry[ac->avail++] = objp;
934}
935
Christoph Lameter3ded1752006-03-25 03:06:44 -0800936/*
937 * Transfer objects in one arraycache to another.
938 * Locking must be handled by the caller.
939 *
940 * Return the number of entries transferred.
941 */
942static int transfer_objects(struct array_cache *to,
943 struct array_cache *from, unsigned int max)
944{
945 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700946 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800947
948 if (!nr)
949 return 0;
950
951 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
952 sizeof(void *) *nr);
953
954 from->avail -= nr;
955 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800956 return nr;
957}
958
Christoph Lameter765c4502006-09-27 01:50:08 -0700959#ifndef CONFIG_NUMA
960
961#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000962#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700963
Pekka Enberg83b519e2009-06-10 19:40:04 +0300964static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700965{
966 return (struct array_cache **)BAD_ALIEN_MAGIC;
967}
968
969static inline void free_alien_cache(struct array_cache **ac_ptr)
970{
971}
972
973static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
974{
975 return 0;
976}
977
978static inline void *alternate_node_alloc(struct kmem_cache *cachep,
979 gfp_t flags)
980{
981 return NULL;
982}
983
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800984static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700985 gfp_t flags, int nodeid)
986{
987 return NULL;
988}
989
990#else /* CONFIG_NUMA */
991
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800992static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800993static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800994
Pekka Enberg83b519e2009-06-10 19:40:04 +0300995static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700996{
997 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800998 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700999 int i;
1000
1001 if (limit > 1)
1002 limit = 12;
Haicheng Lif3186a92010-01-06 15:25:23 +08001003 ac_ptr = kzalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001004 if (ac_ptr) {
1005 for_each_node(i) {
Haicheng Lif3186a92010-01-06 15:25:23 +08001006 if (i == node || !node_online(i))
Christoph Lametere498be72005-09-09 13:03:32 -07001007 continue;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001008 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001009 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001010 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001011 kfree(ac_ptr[i]);
1012 kfree(ac_ptr);
1013 return NULL;
1014 }
1015 }
1016 }
1017 return ac_ptr;
1018}
1019
Pekka Enberg5295a742006-02-01 03:05:48 -08001020static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001021{
1022 int i;
1023
1024 if (!ac_ptr)
1025 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001026 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001027 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001028 kfree(ac_ptr);
1029}
1030
Pekka Enberg343e0d72006-02-01 03:05:50 -08001031static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001032 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001033{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001034 struct kmem_cache_node *n = cachep->node[node];
Christoph Lametere498be72005-09-09 13:03:32 -07001035
1036 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001037 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001038 /*
1039 * Stuff objects into the remote nodes shared array first.
1040 * That way we could avoid the overhead of putting the objects
1041 * into the free lists and getting them back later.
1042 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001043 if (n->shared)
1044 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001045
Christoph Lameterff694162005-09-22 21:44:02 -07001046 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001047 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001048 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001049 }
1050}
1051
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001052/*
1053 * Called from cache_reap() to regularly drain alien caches round robin.
1054 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001055static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001056{
Christoph Lameter909ea962010-12-08 16:22:55 +01001057 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001058
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001059 if (n->alien) {
1060 struct array_cache *ac = n->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001061
1062 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001063 __drain_alien_cache(cachep, ac, node);
1064 spin_unlock_irq(&ac->lock);
1065 }
1066 }
1067}
1068
Andrew Mortona737b3e2006-03-22 00:08:11 -08001069static void drain_alien_cache(struct kmem_cache *cachep,
1070 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001071{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001072 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001073 struct array_cache *ac;
1074 unsigned long flags;
1075
1076 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001077 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001078 if (ac) {
1079 spin_lock_irqsave(&ac->lock, flags);
1080 __drain_alien_cache(cachep, ac, i);
1081 spin_unlock_irqrestore(&ac->lock, flags);
1082 }
1083 }
1084}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001085
Ingo Molnar873623d2006-07-13 14:44:38 +02001086static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001087{
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001088 int nodeid = page_to_nid(virt_to_page(objp));
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001089 struct kmem_cache_node *n;
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001090 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001091 int node;
1092
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001093 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001094
1095 /*
1096 * Make sure we are not freeing a object from another node to the array
1097 * cache on this cpu.
1098 */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001099 if (likely(nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001100 return 0;
1101
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001102 n = cachep->node[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001103 STATS_INC_NODEFREES(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001104 if (n->alien && n->alien[nodeid]) {
1105 alien = n->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001106 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001107 if (unlikely(alien->avail == alien->limit)) {
1108 STATS_INC_ACOVERFLOW(cachep);
1109 __drain_alien_cache(cachep, alien, nodeid);
1110 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07001111 ac_put_obj(cachep, alien, objp);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001112 spin_unlock(&alien->lock);
1113 } else {
Christoph Lameter6a673682013-01-10 19:14:19 +00001114 spin_lock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001115 free_block(cachep, &objp, 1, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001116 spin_unlock(&(cachep->node[nodeid])->list_lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001117 }
1118 return 1;
1119}
Christoph Lametere498be72005-09-09 13:03:32 -07001120#endif
1121
David Rientjes8f9f8d92010-03-27 19:40:47 -07001122/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001123 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001124 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001125 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001126 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001127 * already in use.
1128 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001129 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001130 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001131static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001132{
1133 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001134 struct kmem_cache_node *n;
Christoph Lameter6744f082013-01-10 19:12:17 +00001135 const int memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001136
Christoph Lameter18004c52012-07-06 15:25:12 -05001137 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001138 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001139 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001140 * begin anything. Make sure some other cpu on this
1141 * node has not already allocated this
1142 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001143 if (!cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001144 n = kmalloc_node(memsize, GFP_KERNEL, node);
1145 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001146 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001147 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001148 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1149 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001150
1151 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001152 * The kmem_cache_nodes don't come and go as CPUs
1153 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001154 * protection here.
1155 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001156 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001157 }
1158
Christoph Lameter6a673682013-01-10 19:14:19 +00001159 spin_lock_irq(&cachep->node[node]->list_lock);
1160 cachep->node[node]->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001161 (1 + nr_cpus_node(node)) *
1162 cachep->batchcount + cachep->num;
Christoph Lameter6a673682013-01-10 19:14:19 +00001163 spin_unlock_irq(&cachep->node[node]->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001164 }
1165 return 0;
1166}
1167
Wanpeng Li0fa81032013-07-04 08:33:22 +08001168static inline int slabs_tofree(struct kmem_cache *cachep,
1169 struct kmem_cache_node *n)
1170{
1171 return (n->free_objects + cachep->num - 1) / cachep->num;
1172}
1173
Paul Gortmaker0db06282013-06-19 14:53:51 -04001174static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001176 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001177 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001178 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301179 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001180
Christoph Lameter18004c52012-07-06 15:25:12 -05001181 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001182 struct array_cache *nc;
1183 struct array_cache *shared;
1184 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001186 /* cpu is dead; no one can alloc from it. */
1187 nc = cachep->array[cpu];
1188 cachep->array[cpu] = NULL;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001189 n = cachep->node[node];
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001190
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001191 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001192 goto free_array_cache;
1193
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001194 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001195
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001196 /* Free limit for this kmem_cache_node */
1197 n->free_limit -= cachep->batchcount;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001198 if (nc)
1199 free_block(cachep, nc->entry, nc->avail, node);
1200
Rusty Russell58463c12009-12-17 11:43:12 -06001201 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001202 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001203 goto free_array_cache;
1204 }
1205
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001206 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001207 if (shared) {
1208 free_block(cachep, shared->entry,
1209 shared->avail, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001210 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001211 }
1212
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001213 alien = n->alien;
1214 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001215
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001216 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001217
1218 kfree(shared);
1219 if (alien) {
1220 drain_alien_cache(cachep, alien);
1221 free_alien_cache(alien);
1222 }
1223free_array_cache:
1224 kfree(nc);
1225 }
1226 /*
1227 * In the previous loop, all the objects were freed to
1228 * the respective cache's slabs, now we can go ahead and
1229 * shrink each nodelist to its limit.
1230 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001231 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001232 n = cachep->node[node];
1233 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001234 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001235 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001236 }
1237}
1238
Paul Gortmaker0db06282013-06-19 14:53:51 -04001239static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001240{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001241 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001242 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001243 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001244 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001246 /*
1247 * We need to do this right in the beginning since
1248 * alloc_arraycache's are going to use this list.
1249 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001250 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001251 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001252 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001253 if (err < 0)
1254 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001255
1256 /*
1257 * Now we can go ahead with allocating the shared arrays and
1258 * array caches
1259 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001260 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001261 struct array_cache *nc;
1262 struct array_cache *shared = NULL;
1263 struct array_cache **alien = NULL;
1264
1265 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001266 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001267 if (!nc)
1268 goto bad;
1269 if (cachep->shared) {
1270 shared = alloc_arraycache(node,
1271 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001272 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001273 if (!shared) {
1274 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001275 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001276 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001277 }
1278 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001279 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001280 if (!alien) {
1281 kfree(shared);
1282 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001283 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001284 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001285 }
1286 cachep->array[cpu] = nc;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001287 n = cachep->node[node];
1288 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001289
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001290 spin_lock_irq(&n->list_lock);
1291 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001292 /*
1293 * We are serialised from CPU_DEAD or
1294 * CPU_UP_CANCELLED by the cpucontrol lock
1295 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001296 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001297 shared = NULL;
1298 }
1299#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001300 if (!n->alien) {
1301 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001302 alien = NULL;
1303 }
1304#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001305 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001306 kfree(shared);
1307 free_alien_cache(alien);
Peter Zijlstra83835b32011-07-22 15:26:05 +02001308 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1309 slab_set_debugobj_lock_classes_node(cachep, node);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08001310 else if (!OFF_SLAB(cachep) &&
1311 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1312 on_slab_lock_classes_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001313 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001314 init_node_lock_keys(node);
1315
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001316 return 0;
1317bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001318 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001319 return -ENOMEM;
1320}
1321
Paul Gortmaker0db06282013-06-19 14:53:51 -04001322static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001323 unsigned long action, void *hcpu)
1324{
1325 long cpu = (long)hcpu;
1326 int err = 0;
1327
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001329 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001330 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001331 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001332 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001333 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 break;
1335 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001336 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 start_cpu_timer(cpu);
1338 break;
1339#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001340 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001341 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001342 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001343 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001344 * held so that if cache_reap() is invoked it cannot do
1345 * anything expensive but will only modify reap_work
1346 * and reschedule the timer.
1347 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001348 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001349 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001350 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001351 break;
1352 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001353 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001354 start_cpu_timer(cpu);
1355 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001357 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001358 /*
1359 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001360 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001361 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001362 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001363 * structure is usually allocated from kmem_cache_create() and
1364 * gets destroyed at kmem_cache_destroy().
1365 */
Simon Arlott183ff222007-10-20 01:27:18 +02001366 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001367#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001369 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001370 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001371 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001372 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001375 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376}
1377
Paul Gortmaker0db06282013-06-19 14:53:51 -04001378static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001379 &cpuup_callback, NULL, 0
1380};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381
David Rientjes8f9f8d92010-03-27 19:40:47 -07001382#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1383/*
1384 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1385 * Returns -EBUSY if all objects cannot be drained so that the node is not
1386 * removed.
1387 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001388 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001389 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001390static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001391{
1392 struct kmem_cache *cachep;
1393 int ret = 0;
1394
Christoph Lameter18004c52012-07-06 15:25:12 -05001395 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001396 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001397
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001398 n = cachep->node[node];
1399 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001400 continue;
1401
Wanpeng Li0fa81032013-07-04 08:33:22 +08001402 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001403
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001404 if (!list_empty(&n->slabs_full) ||
1405 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001406 ret = -EBUSY;
1407 break;
1408 }
1409 }
1410 return ret;
1411}
1412
1413static int __meminit slab_memory_callback(struct notifier_block *self,
1414 unsigned long action, void *arg)
1415{
1416 struct memory_notify *mnb = arg;
1417 int ret = 0;
1418 int nid;
1419
1420 nid = mnb->status_change_nid;
1421 if (nid < 0)
1422 goto out;
1423
1424 switch (action) {
1425 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001426 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001427 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001428 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001429 break;
1430 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001431 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001432 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001433 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001434 break;
1435 case MEM_ONLINE:
1436 case MEM_OFFLINE:
1437 case MEM_CANCEL_ONLINE:
1438 case MEM_CANCEL_OFFLINE:
1439 break;
1440 }
1441out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001442 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001443}
1444#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1445
Christoph Lametere498be72005-09-09 13:03:32 -07001446/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001447 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001448 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001449static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001450 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001451{
Christoph Lameter6744f082013-01-10 19:12:17 +00001452 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001453
Christoph Lameter6744f082013-01-10 19:12:17 +00001454 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001455 BUG_ON(!ptr);
1456
Christoph Lameter6744f082013-01-10 19:12:17 +00001457 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001458 /*
1459 * Do not assume that spinlocks can be initialized via memcpy:
1460 */
1461 spin_lock_init(&ptr->list_lock);
1462
Christoph Lametere498be72005-09-09 13:03:32 -07001463 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001464 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001465}
1466
Andrew Mortona737b3e2006-03-22 00:08:11 -08001467/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001468 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1469 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001470 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001471static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001472{
1473 int node;
1474
1475 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001476 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001477 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001478 REAPTIMEOUT_NODE +
1479 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001480 }
1481}
1482
1483/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001484 * The memory after the last cpu cache pointer is used for the
Christoph Lameter6a673682013-01-10 19:14:19 +00001485 * the node pointer.
Christoph Lameter3c583462012-11-28 16:23:01 +00001486 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001487static void setup_node_pointer(struct kmem_cache *cachep)
Christoph Lameter3c583462012-11-28 16:23:01 +00001488{
Christoph Lameter6a673682013-01-10 19:14:19 +00001489 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
Christoph Lameter3c583462012-11-28 16:23:01 +00001490}
1491
1492/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001493 * Initialisation. Called after the page allocator have been initialised and
1494 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 */
1496void __init kmem_cache_init(void)
1497{
Christoph Lametere498be72005-09-09 13:03:32 -07001498 int i;
1499
Joonsoo Kim68126702013-10-24 10:07:42 +09001500 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1501 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001502 kmem_cache = &kmem_cache_boot;
Christoph Lameter6a673682013-01-10 19:14:19 +00001503 setup_node_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001504
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001505 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001506 use_alien_caches = 0;
1507
Christoph Lameter3c583462012-11-28 16:23:01 +00001508 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001509 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001510
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001511 set_up_node(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
1513 /*
1514 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001515 * page orders on machines with more than 32MB of memory if
1516 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001518 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001519 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 /* Bootstrap is tricky, because several objects are allocated
1522 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001523 * 1) initialize the kmem_cache cache: it contains the struct
1524 * kmem_cache structures of all caches, except kmem_cache itself:
1525 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001526 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001527 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001528 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001530 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001531 * An __init data area is used for the head array.
1532 * 3) Create the remaining kmalloc caches, with minimally sized
1533 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001534 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001536 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001537 * the other cache's with kmalloc allocated memory.
1538 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 */
1540
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001541 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542
Eric Dumazet8da34302007-05-06 14:49:29 -07001543 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001544 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001545 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001546 create_boot_cache(kmem_cache, "kmem_cache",
1547 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001548 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001549 SLAB_HWCACHE_ALIGN);
1550 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001551
1552 /* 2+3) create the kmalloc caches */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553
Andrew Mortona737b3e2006-03-22 00:08:11 -08001554 /*
1555 * Initialize the caches that provide memory for the array cache and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001556 * kmem_cache_node structures first. Without this, further allocations will
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001558 */
1559
Christoph Lametere3366012013-01-10 19:14:18 +00001560 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1561 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001562
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001563 if (INDEX_AC != INDEX_NODE)
1564 kmalloc_caches[INDEX_NODE] =
1565 create_kmalloc_cache("kmalloc-node",
1566 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001567
Ingo Molnare0a42722006-06-23 02:03:46 -07001568 slab_early_init = 0;
1569
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 /* 4) Replace the bootstrap head arrays */
1571 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001572 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001573
Pekka Enberg83b519e2009-06-10 19:40:04 +03001574 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001575
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001576 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001577 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001578 /*
1579 * Do not assume that spinlocks can be initialized via memcpy:
1580 */
1581 spin_lock_init(&ptr->lock);
1582
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001583 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001584
Pekka Enberg83b519e2009-06-10 19:40:04 +03001585 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001586
Christoph Lametere3366012013-01-10 19:14:18 +00001587 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001588 != &initarray_generic.cache);
Christoph Lametere3366012013-01-10 19:14:18 +00001589 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001590 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001591 /*
1592 * Do not assume that spinlocks can be initialized via memcpy:
1593 */
1594 spin_lock_init(&ptr->lock);
1595
Christoph Lametere3366012013-01-10 19:14:18 +00001596 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001598 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001599 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001600 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Mel Gorman9c09a952008-01-24 05:49:54 -08001602 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001603 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001604
Christoph Lametere3366012013-01-10 19:14:18 +00001605 init_list(kmalloc_caches[INDEX_AC],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001606 &init_kmem_cache_node[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001607
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001608 if (INDEX_AC != INDEX_NODE) {
1609 init_list(kmalloc_caches[INDEX_NODE],
1610 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001611 }
1612 }
1613 }
1614
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001615 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001616}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001617
Pekka Enberg8429db52009-06-12 15:58:59 +03001618void __init kmem_cache_init_late(void)
1619{
1620 struct kmem_cache *cachep;
1621
Christoph Lameter97d06602012-07-06 15:25:11 -05001622 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001623
Pekka Enberg8429db52009-06-12 15:58:59 +03001624 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001625 mutex_lock(&slab_mutex);
1626 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001627 if (enable_cpucache(cachep, GFP_NOWAIT))
1628 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001629 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001630
Michael Wang947ca182012-09-05 10:33:18 +08001631 /* Annotate slab for lockdep -- annotate the malloc caches */
1632 init_lock_keys();
1633
Christoph Lameter97d06602012-07-06 15:25:11 -05001634 /* Done! */
1635 slab_state = FULL;
1636
Andrew Mortona737b3e2006-03-22 00:08:11 -08001637 /*
1638 * Register a cpu startup notifier callback that initializes
1639 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 */
1641 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
David Rientjes8f9f8d92010-03-27 19:40:47 -07001643#ifdef CONFIG_NUMA
1644 /*
1645 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001646 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001647 */
1648 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1649#endif
1650
Andrew Mortona737b3e2006-03-22 00:08:11 -08001651 /*
1652 * The reap timers are started later, with a module init call: That part
1653 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 */
1655}
1656
1657static int __init cpucache_init(void)
1658{
1659 int cpu;
1660
Andrew Mortona737b3e2006-03-22 00:08:11 -08001661 /*
1662 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 */
Christoph Lametere498be72005-09-09 13:03:32 -07001664 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001665 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001666
1667 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001668 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 return 0;
1670}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671__initcall(cpucache_init);
1672
Rafael Aquini8bdec192012-03-09 17:27:27 -03001673static noinline void
1674slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1675{
David Rientjes9a02d692014-06-04 16:06:36 -07001676#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001677 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001678 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001679 unsigned long flags;
1680 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001681 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1682 DEFAULT_RATELIMIT_BURST);
1683
1684 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1685 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001686
1687 printk(KERN_WARNING
1688 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1689 nodeid, gfpflags);
1690 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001691 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001692
1693 for_each_online_node(node) {
1694 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1695 unsigned long active_slabs = 0, num_slabs = 0;
1696
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001697 n = cachep->node[node];
1698 if (!n)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001699 continue;
1700
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001701 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001702 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001703 active_objs += cachep->num;
1704 active_slabs++;
1705 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001706 list_for_each_entry(page, &n->slabs_partial, lru) {
1707 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001708 active_slabs++;
1709 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001710 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001711 num_slabs++;
1712
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001713 free_objects += n->free_objects;
1714 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001715
1716 num_slabs += active_slabs;
1717 num_objs = num_slabs * cachep->num;
1718 printk(KERN_WARNING
1719 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1720 node, active_slabs, num_slabs, active_objs, num_objs,
1721 free_objects);
1722 }
David Rientjes9a02d692014-06-04 16:06:36 -07001723#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001724}
1725
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726/*
1727 * Interface to system's page allocator. No need to hold the cache-lock.
1728 *
1729 * If we requested dmaable memory, we will get it. Even if we
1730 * did not request dmaable memory, we might get it, but that
1731 * would be relatively rare and ignorable.
1732 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001733static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1734 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735{
1736 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001737 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001738
Glauber Costaa618e892012-06-14 16:17:21 +04001739 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001740 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1741 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001742
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001743 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1744 return NULL;
1745
Linus Torvalds517d0862009-06-16 19:50:13 -07001746 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001747 if (!page) {
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001748 memcg_uncharge_slab(cachep, cachep->gfporder);
David Rientjes9a02d692014-06-04 16:06:36 -07001749 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001751 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001753 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001754 if (unlikely(page->pfmemalloc))
1755 pfmemalloc_active = true;
1756
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001757 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001758 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001759 add_zone_page_state(page_zone(page),
1760 NR_SLAB_RECLAIMABLE, nr_pages);
1761 else
1762 add_zone_page_state(page_zone(page),
1763 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001764 __SetPageSlab(page);
1765 if (page->pfmemalloc)
1766 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001767
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001768 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1769 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1770
1771 if (cachep->ctor)
1772 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1773 else
1774 kmemcheck_mark_unallocated_pages(page, nr_pages);
1775 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001776
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001777 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778}
1779
1780/*
1781 * Interface to system's page release.
1782 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001783static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001785 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001787 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001788
Christoph Lameter972d1a72006-09-25 23:31:51 -07001789 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1790 sub_zone_page_state(page_zone(page),
1791 NR_SLAB_RECLAIMABLE, nr_freed);
1792 else
1793 sub_zone_page_state(page_zone(page),
1794 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001795
Joonsoo Kima57a4982013-10-24 10:07:44 +09001796 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001797 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001798 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001799 page_mapcount_reset(page);
1800 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001801
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 if (current->reclaim_state)
1803 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001804 __free_pages(page, cachep->gfporder);
1805 memcg_uncharge_slab(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806}
1807
1808static void kmem_rcu_free(struct rcu_head *head)
1809{
Joonsoo Kim68126702013-10-24 10:07:42 +09001810 struct kmem_cache *cachep;
1811 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812
Joonsoo Kim68126702013-10-24 10:07:42 +09001813 page = container_of(head, struct page, rcu_head);
1814 cachep = page->slab_cache;
1815
1816 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817}
1818
1819#if DEBUG
1820
1821#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001822static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001823 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001825 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001827 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001829 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 return;
1831
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001832 *addr++ = 0x12345678;
1833 *addr++ = caller;
1834 *addr++ = smp_processor_id();
1835 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 {
1837 unsigned long *sptr = &caller;
1838 unsigned long svalue;
1839
1840 while (!kstack_end(sptr)) {
1841 svalue = *sptr++;
1842 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001843 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 size -= sizeof(unsigned long);
1845 if (size <= sizeof(unsigned long))
1846 break;
1847 }
1848 }
1849
1850 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852}
1853#endif
1854
Pekka Enberg343e0d72006-02-01 03:05:50 -08001855static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001857 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001858 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859
1860 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862}
1863
1864static void dump_line(char *data, int offset, int limit)
1865{
1866 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001867 unsigned char error = 0;
1868 int bad_count = 0;
1869
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001870 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001871 for (i = 0; i < limit; i++) {
1872 if (data[offset + i] != POISON_FREE) {
1873 error = data[offset + i];
1874 bad_count++;
1875 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001876 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001877 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1878 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001879
1880 if (bad_count == 1) {
1881 error ^= POISON_FREE;
1882 if (!(error & (error - 1))) {
1883 printk(KERN_ERR "Single bit error detected. Probably "
1884 "bad RAM.\n");
1885#ifdef CONFIG_X86
1886 printk(KERN_ERR "Run memtest86+ or a similar memory "
1887 "test tool.\n");
1888#else
1889 printk(KERN_ERR "Run a memory test tool.\n");
1890#endif
1891 }
1892 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893}
1894#endif
1895
1896#if DEBUG
1897
Pekka Enberg343e0d72006-02-01 03:05:50 -08001898static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899{
1900 int i, size;
1901 char *realobj;
1902
1903 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001904 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001905 *dbg_redzone1(cachep, objp),
1906 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 }
1908
1909 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001910 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1911 *dbg_userword(cachep, objp),
1912 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001914 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001915 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001916 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 int limit;
1918 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001919 if (i + limit > size)
1920 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 dump_line(realobj, i, limit);
1922 }
1923}
1924
Pekka Enberg343e0d72006-02-01 03:05:50 -08001925static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926{
1927 char *realobj;
1928 int size, i;
1929 int lines = 0;
1930
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001931 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001932 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001934 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001936 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 exp = POISON_END;
1938 if (realobj[i] != exp) {
1939 int limit;
1940 /* Mismatch ! */
1941 /* Print header */
1942 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001943 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001944 "Slab corruption (%s): %s start=%p, len=%d\n",
1945 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 print_objinfo(cachep, objp, 0);
1947 }
1948 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001949 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001951 if (i + limit > size)
1952 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 dump_line(realobj, i, limit);
1954 i += 16;
1955 lines++;
1956 /* Limit to 5 lines */
1957 if (lines > 5)
1958 break;
1959 }
1960 }
1961 if (lines != 0) {
1962 /* Print some data about the neighboring objects, if they
1963 * exist:
1964 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001965 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001966 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
Joonsoo Kim8456a642013-10-24 10:07:49 +09001968 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001970 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001971 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001973 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 print_objinfo(cachep, objp, 2);
1975 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001976 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001977 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001978 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001980 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 print_objinfo(cachep, objp, 2);
1982 }
1983 }
1984}
1985#endif
1986
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001988static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1989 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001990{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 int i;
1992 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001993 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994
1995 if (cachep->flags & SLAB_POISON) {
1996#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001997 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001998 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001999 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002000 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 else
2002 check_poison_obj(cachep, objp);
2003#else
2004 check_poison_obj(cachep, objp);
2005#endif
2006 }
2007 if (cachep->flags & SLAB_RED_ZONE) {
2008 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2009 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002010 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2012 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002013 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002016}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09002018static void slab_destroy_debugcheck(struct kmem_cache *cachep,
2019 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002020{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002021}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022#endif
2023
Randy Dunlap911851e2006-03-22 00:08:14 -08002024/**
2025 * slab_destroy - destroy and release all objects in a slab
2026 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09002027 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08002028 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002029 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002030 * Before calling the slab must have been unlinked from the cache. The
2031 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002032 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002033static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002034{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002035 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08002036
Joonsoo Kim8456a642013-10-24 10:07:49 +09002037 freelist = page->freelist;
2038 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09002040 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041
Joonsoo Kim68126702013-10-24 10:07:42 +09002042 /*
2043 * RCU free overloads the RCU head over the LRU.
2044 * slab_page has been overloeaded over the LRU,
2045 * however it is not used from now on so that
2046 * we can use it safely.
2047 */
2048 head = (void *)&page->rcu_head;
2049 call_rcu(head, kmem_rcu_free);
2050
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002052 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 }
Joonsoo Kim68126702013-10-24 10:07:42 +09002054
2055 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09002056 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09002057 * although actual page can be freed in rcu context
2058 */
2059 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09002060 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061}
2062
2063/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002064 * calculate_slab_order - calculate size (page order) of slabs
2065 * @cachep: pointer to the cache that is being created
2066 * @size: size of objects to be created in this cache.
2067 * @align: required alignment for the objects.
2068 * @flags: slab allocation flags
2069 *
2070 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002071 *
2072 * This could be made much more intelligent. For now, try to avoid using
2073 * high order pages for slabs. When the gfp() functions are more friendly
2074 * towards high-order requests, this should be changed.
2075 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002076static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002077 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002078{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002079 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002080 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002081 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002082
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002083 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002084 unsigned int num;
2085 size_t remainder;
2086
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002087 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002088 if (!num)
2089 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002090
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002091 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2092 if (num > SLAB_OBJ_MAX_NUM)
2093 break;
2094
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002095 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim03787302014-06-23 13:22:06 -07002096 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002097 /*
2098 * Max number of objs-per-slab for caches which
2099 * use off-slab slabs. Needed to avoid a possible
2100 * looping condition in cache_grow().
2101 */
Joonsoo Kim03787302014-06-23 13:22:06 -07002102 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
2103 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002104 offslab_limit = size;
Joonsoo Kim03787302014-06-23 13:22:06 -07002105 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002106
2107 if (num > offslab_limit)
2108 break;
2109 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002110
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002111 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002112 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002113 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002114 left_over = remainder;
2115
2116 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002117 * A VFS-reclaimable slab tends to have most allocations
2118 * as GFP_NOFS and we really don't want to have to be allocating
2119 * higher-order pages when we are unable to shrink dcache.
2120 */
2121 if (flags & SLAB_RECLAIM_ACCOUNT)
2122 break;
2123
2124 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002125 * Large number of objects is good, but very large slabs are
2126 * currently bad for the gfp()s.
2127 */
David Rientjes543585c2011-10-18 22:09:24 -07002128 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002129 break;
2130
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002131 /*
2132 * Acceptable internal fragmentation?
2133 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002134 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002135 break;
2136 }
2137 return left_over;
2138}
2139
Pekka Enberg83b519e2009-06-10 19:40:04 +03002140static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002141{
Christoph Lameter97d06602012-07-06 15:25:11 -05002142 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002143 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002144
Christoph Lameter97d06602012-07-06 15:25:11 -05002145 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002146 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002147 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002148 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002149 * of by the caller of __kmem_cache_create
2150 */
2151 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2152 slab_state = PARTIAL;
2153 } else if (slab_state == PARTIAL) {
2154 /*
2155 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002156 * that's used by kmalloc(24), otherwise the creation of
2157 * further caches will BUG().
2158 */
2159 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2160
2161 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002162 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2163 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002164 * otherwise the creation of further caches will BUG().
2165 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002166 set_up_node(cachep, SIZE_AC);
2167 if (INDEX_AC == INDEX_NODE)
2168 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002169 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002170 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002171 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002172 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002173 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002174 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002175
Christoph Lameter97d06602012-07-06 15:25:11 -05002176 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002177 set_up_node(cachep, SIZE_NODE);
2178 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002179 } else {
2180 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002181 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002182 cachep->node[node] =
Christoph Lameter6744f082013-01-10 19:12:17 +00002183 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002184 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002185 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002186 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002187 }
2188 }
2189 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002190 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002191 jiffies + REAPTIMEOUT_NODE +
2192 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002193
2194 cpu_cache_get(cachep)->avail = 0;
2195 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2196 cpu_cache_get(cachep)->batchcount = 1;
2197 cpu_cache_get(cachep)->touched = 0;
2198 cachep->batchcount = 1;
2199 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002200 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002201}
2202
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002203/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002204 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002205 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 *
2208 * Returns a ptr to the cache on success, NULL on failure.
2209 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002210 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 * The flags are
2213 *
2214 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2215 * to catch references to uninitialised memory.
2216 *
2217 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2218 * for buffer overruns.
2219 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2221 * cacheline. This can be beneficial if you're counting cycles as closely
2222 * as davem.
2223 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002224int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002225__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002227 size_t left_over, freelist_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002228 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002229 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002230 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233#if FORCED_DEBUG
2234 /*
2235 * Enable redzoning and last user accounting, except for caches with
2236 * large objects, if the increased size would increase the object size
2237 * above the next power of two: caches with object sizes just above a
2238 * power of two have a significant amount of internal fragmentation.
2239 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002240 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2241 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002242 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 if (!(flags & SLAB_DESTROY_BY_RCU))
2244 flags |= SLAB_POISON;
2245#endif
2246 if (flags & SLAB_DESTROY_BY_RCU)
2247 BUG_ON(flags & SLAB_POISON);
2248#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249
Andrew Mortona737b3e2006-03-22 00:08:11 -08002250 /*
2251 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 * unaligned accesses for some archs when redzoning is used, and makes
2253 * sure any on-slab bufctl's are also correctly aligned.
2254 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002255 if (size & (BYTES_PER_WORD - 1)) {
2256 size += (BYTES_PER_WORD - 1);
2257 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 }
2259
Pekka Enbergca5f9702006-09-25 23:31:25 -07002260 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002261 * Redzoning and user store require word alignment or possibly larger.
2262 * Note this will be overridden by architecture or caller mandated
2263 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002264 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002265 if (flags & SLAB_STORE_USER)
2266 ralign = BYTES_PER_WORD;
2267
2268 if (flags & SLAB_RED_ZONE) {
2269 ralign = REDZONE_ALIGN;
2270 /* If redzoning, ensure that the second redzone is suitably
2271 * aligned, by adjusting the object size accordingly. */
2272 size += REDZONE_ALIGN - 1;
2273 size &= ~(REDZONE_ALIGN - 1);
2274 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002275
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002276 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002277 if (ralign < cachep->align) {
2278 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002280 /* disable debug if necessary */
2281 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002282 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002283 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002284 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002286 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
Pekka Enberg83b519e2009-06-10 19:40:04 +03002288 if (slab_is_available())
2289 gfp = GFP_KERNEL;
2290 else
2291 gfp = GFP_NOWAIT;
2292
Christoph Lameter6a673682013-01-10 19:14:19 +00002293 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295
Pekka Enbergca5f9702006-09-25 23:31:25 -07002296 /*
2297 * Both debugging options require word-alignment which is calculated
2298 * into align above.
2299 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002302 cachep->obj_offset += sizeof(unsigned long long);
2303 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 }
2305 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002306 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002307 * the real object. But if the second red zone needs to be
2308 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002310 if (flags & SLAB_RED_ZONE)
2311 size += REDZONE_ALIGN;
2312 else
2313 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 }
2315#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002316 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002317 && cachep->object_size > cache_line_size()
2318 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2319 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 size = PAGE_SIZE;
2321 }
2322#endif
2323#endif
2324
Ingo Molnare0a42722006-06-23 02:03:46 -07002325 /*
2326 * Determine if the slab management is 'on' or 'off' slab.
2327 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002328 * it too early on. Always use on-slab management when
2329 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002330 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002331 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002332 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 /*
2334 * Size is large, assume best to place the slab management obj
2335 * off-slab (should allow better packing of objs).
2336 */
2337 flags |= CFLGS_OFF_SLAB;
2338
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002339 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002340 /*
2341 * We should restrict the number of objects in a slab to implement
2342 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2343 */
2344 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2345 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002347 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002349 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002350 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002351
Joonsoo Kim03787302014-06-23 13:22:06 -07002352 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353
2354 /*
2355 * If the slab has been placed off-slab, and we have enough space then
2356 * move it on-slab. This is at the expense of any extra colouring.
2357 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002358 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002360 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 }
2362
2363 if (flags & CFLGS_OFF_SLAB) {
2364 /* really off slab. No need for manual alignment */
Joonsoo Kim03787302014-06-23 13:22:06 -07002365 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302366
2367#ifdef CONFIG_PAGE_POISONING
2368 /* If we're going to use the generic kernel_map_pages()
2369 * poisoning, then it's going to smash the contents of
2370 * the redzone and userword anyhow, so switch them off.
2371 */
2372 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2373 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2374#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 }
2376
2377 cachep->colour_off = cache_line_size();
2378 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002379 if (cachep->colour_off < cachep->align)
2380 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002381 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002382 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002384 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002385 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002386 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002387 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002388 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002390 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002391 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002392 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002393 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002394 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002395 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2396 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002397 * But leave a BUG_ON for some lucky dude.
2398 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002399 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002400 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002402 err = setup_cpu_cache(cachep, gfp);
2403 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002404 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002405 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002406 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407
Peter Zijlstra83835b32011-07-22 15:26:05 +02002408 if (flags & SLAB_DEBUG_OBJECTS) {
2409 /*
2410 * Would deadlock through slab_destroy()->call_rcu()->
2411 * debug_object_activate()->kmem_cache_alloc().
2412 */
2413 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2414
2415 slab_set_debugobj_lock_classes(cachep);
Glauber Costa6ccfb5b2012-12-18 14:22:31 -08002416 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2417 on_slab_lock_classes(cachep);
Peter Zijlstra83835b32011-07-22 15:26:05 +02002418
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002419 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421
2422#if DEBUG
2423static void check_irq_off(void)
2424{
2425 BUG_ON(!irqs_disabled());
2426}
2427
2428static void check_irq_on(void)
2429{
2430 BUG_ON(irqs_disabled());
2431}
2432
Pekka Enberg343e0d72006-02-01 03:05:50 -08002433static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434{
2435#ifdef CONFIG_SMP
2436 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002437 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438#endif
2439}
Christoph Lametere498be72005-09-09 13:03:32 -07002440
Pekka Enberg343e0d72006-02-01 03:05:50 -08002441static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002442{
2443#ifdef CONFIG_SMP
2444 check_irq_off();
Christoph Lameter6a673682013-01-10 19:14:19 +00002445 assert_spin_locked(&cachep->node[node]->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002446#endif
2447}
2448
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449#else
2450#define check_irq_off() do { } while(0)
2451#define check_irq_on() do { } while(0)
2452#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002453#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454#endif
2455
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002456static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002457 struct array_cache *ac,
2458 int force, int node);
2459
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460static void do_drain(void *arg)
2461{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002462 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002464 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
2466 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002467 ac = cpu_cache_get(cachep);
Christoph Lameter6a673682013-01-10 19:14:19 +00002468 spin_lock(&cachep->node[node]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002469 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002470 spin_unlock(&cachep->node[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 ac->avail = 0;
2472}
2473
Pekka Enberg343e0d72006-02-01 03:05:50 -08002474static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002476 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002477 int node;
2478
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002479 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002481 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002482 n = cachep->node[node];
2483 if (n && n->alien)
2484 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002485 }
2486
2487 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002488 n = cachep->node[node];
2489 if (n)
2490 drain_array(cachep, n, n->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002491 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492}
2493
Christoph Lametered11d9e2006-06-30 01:55:45 -07002494/*
2495 * Remove slabs from the list of free slabs.
2496 * Specify the number of slabs to drain in tofree.
2497 *
2498 * Returns the actual number of slabs released.
2499 */
2500static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002501 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002503 struct list_head *p;
2504 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002505 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
Christoph Lametered11d9e2006-06-30 01:55:45 -07002507 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002508 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002510 spin_lock_irq(&n->list_lock);
2511 p = n->slabs_free.prev;
2512 if (p == &n->slabs_free) {
2513 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002514 goto out;
2515 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516
Joonsoo Kim8456a642013-10-24 10:07:49 +09002517 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002519 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002521 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002522 /*
2523 * Safe to drop the lock. The slab is no longer linked
2524 * to the cache.
2525 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002526 n->free_objects -= cache->num;
2527 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002528 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002529 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002531out:
2532 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533}
2534
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002535int __kmem_cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002536{
2537 int ret = 0, i = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002538 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002539
2540 drain_cpu_caches(cachep);
2541
2542 check_irq_on();
2543 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002544 n = cachep->node[i];
2545 if (!n)
Christoph Lametered11d9e2006-06-30 01:55:45 -07002546 continue;
2547
Wanpeng Li0fa81032013-07-04 08:33:22 +08002548 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002549
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002550 ret += !list_empty(&n->slabs_full) ||
2551 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002552 }
2553 return (ret ? 1 : 0);
2554}
2555
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002556int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557{
Christoph Lameter12c36672012-09-04 23:38:33 +00002558 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002559 struct kmem_cache_node *n;
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002560 int rc = __kmem_cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561
Christoph Lameter12c36672012-09-04 23:38:33 +00002562 if (rc)
2563 return rc;
2564
2565 for_each_online_cpu(i)
2566 kfree(cachep->array[i]);
2567
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002568 /* NUMA: free the node structures */
Christoph Lameter12c36672012-09-04 23:38:33 +00002569 for_each_online_node(i) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002570 n = cachep->node[i];
2571 if (n) {
2572 kfree(n->shared);
2573 free_alien_cache(n->alien);
2574 kfree(n);
Christoph Lameter12c36672012-09-04 23:38:33 +00002575 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002577 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002580/*
2581 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002582 *
2583 * For a slab cache when the slab descriptor is off-slab, the
2584 * slab descriptor can't come from the same cache which is being created,
2585 * Because if it is the case, that means we defer the creation of
2586 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2587 * And we eventually call down to __kmem_cache_create(), which
2588 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2589 * This is a "chicken-and-egg" problem.
2590 *
2591 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2592 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002593 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002594static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002595 struct page *page, int colour_off,
2596 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002598 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002599 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002600
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 if (OFF_SLAB(cachep)) {
2602 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002603 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002604 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002605 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 return NULL;
2607 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002608 freelist = addr + colour_off;
2609 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002611 page->active = 0;
2612 page->s_mem = addr + colour_off;
2613 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614}
2615
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002616static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002618 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002619}
2620
2621static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002622 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002623{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002624 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625}
2626
Pekka Enberg343e0d72006-02-01 03:05:50 -08002627static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002628 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629{
2630 int i;
2631
2632 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002633 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634#if DEBUG
2635 /* need to poison the objs? */
2636 if (cachep->flags & SLAB_POISON)
2637 poison_obj(cachep, objp, POISON_FREE);
2638 if (cachep->flags & SLAB_STORE_USER)
2639 *dbg_userword(cachep, objp) = NULL;
2640
2641 if (cachep->flags & SLAB_RED_ZONE) {
2642 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2643 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2644 }
2645 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002646 * Constructors are not allowed to allocate memory from the same
2647 * cache which they are a constructor for. Otherwise, deadlock.
2648 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 */
2650 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002651 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652
2653 if (cachep->flags & SLAB_RED_ZONE) {
2654 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2655 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002656 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2658 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002659 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002661 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002662 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002663 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002664 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665#else
2666 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002667 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668#endif
Joonsoo Kim03787302014-06-23 13:22:06 -07002669 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002670 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672}
2673
Pekka Enberg343e0d72006-02-01 03:05:50 -08002674static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002676 if (CONFIG_ZONE_DMA_FLAG) {
2677 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002678 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002679 else
Glauber Costaa618e892012-06-14 16:17:21 +04002680 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002681 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682}
2683
Joonsoo Kim8456a642013-10-24 10:07:49 +09002684static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002685 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002686{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002687 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002688
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002689 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002690 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002691#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002692 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002693#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002694
2695 return objp;
2696}
2697
Joonsoo Kim8456a642013-10-24 10:07:49 +09002698static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002699 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002700{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002701 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002702#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002703 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002704
Matthew Dobson78d382d2006-02-01 03:05:47 -08002705 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002706 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002707
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002708 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002709 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002710 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002711 printk(KERN_ERR "slab: double free detected in cache "
2712 "'%s', objp %p\n", cachep->name, objp);
2713 BUG();
2714 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002715 }
2716#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002717 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002718 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002719}
2720
Pekka Enberg47768742006-06-23 02:03:07 -07002721/*
2722 * Map pages beginning at addr to the given cache and slab. This is required
2723 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002724 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002725 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002726static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002727 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002729 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002730 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731}
2732
2733/*
2734 * Grow (by 1) the number of slabs within a cache. This is called by
2735 * kmem_cache_alloc() when there are no active objs left in a cache.
2736 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002737static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002738 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002740 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002741 size_t offset;
2742 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002743 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744
Andrew Mortona737b3e2006-03-22 00:08:11 -08002745 /*
2746 * Be lazy and only check for valid flags here, keeping it out of the
2747 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002749 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2750 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002752 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002754 n = cachep->node[nodeid];
2755 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756
2757 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002758 offset = n->colour_next;
2759 n->colour_next++;
2760 if (n->colour_next >= cachep->colour)
2761 n->colour_next = 0;
2762 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002764 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
2766 if (local_flags & __GFP_WAIT)
2767 local_irq_enable();
2768
2769 /*
2770 * The test for missing atomic flag is performed here, rather than
2771 * the more obvious place, simply to reduce the critical path length
2772 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2773 * will eventually be caught here (where it matters).
2774 */
2775 kmem_flagcheck(cachep, flags);
2776
Andrew Mortona737b3e2006-03-22 00:08:11 -08002777 /*
2778 * Get mem for the objs. Attempt to allocate a physical page from
2779 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002780 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002781 if (!page)
2782 page = kmem_getpages(cachep, local_flags, nodeid);
2783 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 goto failed;
2785
2786 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002787 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002788 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002789 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 goto opps1;
2791
Joonsoo Kim8456a642013-10-24 10:07:49 +09002792 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
Joonsoo Kim8456a642013-10-24 10:07:49 +09002794 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795
2796 if (local_flags & __GFP_WAIT)
2797 local_irq_disable();
2798 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002799 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800
2801 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002802 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002804 n->free_objects += cachep->num;
2805 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002807opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002808 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002809failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 if (local_flags & __GFP_WAIT)
2811 local_irq_disable();
2812 return 0;
2813}
2814
2815#if DEBUG
2816
2817/*
2818 * Perform extra freeing checks:
2819 * - detect bad pointers.
2820 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 */
2822static void kfree_debugcheck(const void *objp)
2823{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 if (!virt_addr_valid(objp)) {
2825 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002826 (unsigned long)objp);
2827 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829}
2830
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002831static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2832{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002833 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002834
2835 redzone1 = *dbg_redzone1(cache, obj);
2836 redzone2 = *dbg_redzone2(cache, obj);
2837
2838 /*
2839 * Redzone is ok.
2840 */
2841 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2842 return;
2843
2844 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2845 slab_error(cache, "double free detected");
2846 else
2847 slab_error(cache, "memory outside object was overwritten");
2848
David Woodhouseb46b8f12007-05-08 00:22:59 -07002849 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002850 obj, redzone1, redzone2);
2851}
2852
Pekka Enberg343e0d72006-02-01 03:05:50 -08002853static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002854 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002857 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002859 BUG_ON(virt_to_cache(objp) != cachep);
2860
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002861 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002863 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002866 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2868 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2869 }
2870 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002871 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872
Joonsoo Kim8456a642013-10-24 10:07:49 +09002873 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874
2875 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002876 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877
Joonsoo Kim03787302014-06-23 13:22:06 -07002878 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 if (cachep->flags & SLAB_POISON) {
2880#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002881 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002882 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002884 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 } else {
2886 poison_obj(cachep, objp, POISON_FREE);
2887 }
2888#else
2889 poison_obj(cachep, objp, POISON_FREE);
2890#endif
2891 }
2892 return objp;
2893}
2894
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895#else
2896#define kfree_debugcheck(x) do { } while(0)
2897#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898#endif
2899
Mel Gorman072bb0a2012-07-31 16:43:58 -07002900static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2901 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902{
2903 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002904 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002906 int node;
2907
Joe Korty6d2144d2008-03-05 15:04:59 -08002908 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002909 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002910 if (unlikely(force_refill))
2911 goto force_grow;
2912retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002913 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 batchcount = ac->batchcount;
2915 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002916 /*
2917 * If there was little recent activity on this cache, then
2918 * perform only a partial refill. Otherwise we could generate
2919 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 */
2921 batchcount = BATCHREFILL_LIMIT;
2922 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002923 n = cachep->node[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002925 BUG_ON(ac->avail > 0 || !n);
2926 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002927
Christoph Lameter3ded1752006-03-25 03:06:44 -08002928 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002929 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2930 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002931 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002932 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002933
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 while (batchcount > 0) {
2935 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002936 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002938 entry = n->slabs_partial.next;
2939 if (entry == &n->slabs_partial) {
2940 n->free_touched = 1;
2941 entry = n->slabs_free.next;
2942 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 goto must_grow;
2944 }
2945
Joonsoo Kim8456a642013-10-24 10:07:49 +09002946 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002948
2949 /*
2950 * The slab was either on partial or free list so
2951 * there must be at least one object available for
2952 * allocation.
2953 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002954 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002955
Joonsoo Kim8456a642013-10-24 10:07:49 +09002956 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 STATS_INC_ALLOCED(cachep);
2958 STATS_INC_ACTIVE(cachep);
2959 STATS_SET_HIGH(cachep);
2960
Joonsoo Kim8456a642013-10-24 10:07:49 +09002961 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002962 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964
2965 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002966 list_del(&page->lru);
2967 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002968 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002970 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 }
2972
Andrew Mortona737b3e2006-03-22 00:08:11 -08002973must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002974 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002975alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002976 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977
2978 if (unlikely(!ac->avail)) {
2979 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002980force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002981 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002982
Andrew Mortona737b3e2006-03-22 00:08:11 -08002983 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002984 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002985 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002986
2987 /* no objects in sight? abort */
2988 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 return NULL;
2990
Andrew Mortona737b3e2006-03-22 00:08:11 -08002991 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992 goto retry;
2993 }
2994 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002995
2996 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997}
2998
Andrew Mortona737b3e2006-03-22 00:08:11 -08002999static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3000 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001{
3002 might_sleep_if(flags & __GFP_WAIT);
3003#if DEBUG
3004 kmem_flagcheck(cachep, flags);
3005#endif
3006}
3007
3008#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003009static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003010 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011{
Joonsoo Kim03787302014-06-23 13:22:06 -07003012 struct page *page;
3013
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003014 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003016 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003018 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003019 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003020 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021 else
3022 check_poison_obj(cachep, objp);
3023#else
3024 check_poison_obj(cachep, objp);
3025#endif
3026 poison_obj(cachep, objp, POISON_INUSE);
3027 }
3028 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003029 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030
3031 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003032 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3033 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3034 slab_error(cachep, "double free, or memory outside"
3035 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003036 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003037 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003038 objp, *dbg_redzone1(cachep, objp),
3039 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 }
3041 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3042 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3043 }
Joonsoo Kim03787302014-06-23 13:22:06 -07003044
3045 page = virt_to_head_page(objp);
3046 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003047 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003048 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003049 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003050 if (ARCH_SLAB_MINALIGN &&
3051 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003052 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003053 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003054 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 return objp;
3056}
3057#else
3058#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3059#endif
3060
Akinobu Mita773ff602008-12-23 19:37:01 +09003061static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003062{
Christoph Lameter9b030cb2012-09-05 00:20:33 +00003063 if (cachep == kmem_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003064 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003065
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003066 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003067}
3068
Pekka Enberg343e0d72006-02-01 03:05:50 -08003069static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003071 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003073 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074
Alok N Kataria5c382302005-09-27 21:45:46 -07003075 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003076
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003077 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003080 objp = ac_get_obj(cachep, ac, flags, false);
3081
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003082 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07003083 * Allow for the possibility all avail objects are not allowed
3084 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09003085 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07003086 if (objp) {
3087 STATS_INC_ALLOCHIT(cachep);
3088 goto out;
3089 }
3090 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003092
3093 STATS_INC_ALLOCMISS(cachep);
3094 objp = cache_alloc_refill(cachep, flags, force_refill);
3095 /*
3096 * the 'ac' may be updated by cache_alloc_refill(),
3097 * and kmemleak_erase() requires its correct value.
3098 */
3099 ac = cpu_cache_get(cachep);
3100
3101out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003102 /*
3103 * To avoid a false negative, if an object that is in one of the
3104 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3105 * treat the array pointers as a reference to the object.
3106 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003107 if (objp)
3108 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003109 return objp;
3110}
3111
Christoph Lametere498be72005-09-09 13:03:32 -07003112#ifdef CONFIG_NUMA
3113/*
David Rientjesf0432d12014-04-07 15:37:30 -07003114 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003115 *
3116 * If we are in_interrupt, then process context, including cpusets and
3117 * mempolicy, may not apply and should not be used for allocation policy.
3118 */
3119static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3120{
3121 int nid_alloc, nid_here;
3122
Christoph Lameter765c4502006-09-27 01:50:08 -07003123 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003124 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003125 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003126 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003127 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003128 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003129 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003130 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003131 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003132 return NULL;
3133}
3134
3135/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003136 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003137 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003138 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003139 * perform an allocation without specifying a node. This allows the page
3140 * allocator to do its reclaim / fallback magic. We then insert the
3141 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003142 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003143static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003144{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003145 struct zonelist *zonelist;
3146 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003147 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003148 struct zone *zone;
3149 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003150 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003151 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003152 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003153
3154 if (flags & __GFP_THISNODE)
3155 return NULL;
3156
Christoph Lameter6cb06222007-10-16 01:25:41 -07003157 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003158
Mel Gormancc9a6c82012-03-21 16:34:11 -07003159retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003160 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003161 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003162
Christoph Lameter3c517a62006-12-06 20:33:29 -08003163retry:
3164 /*
3165 * Look through allowed nodes for objects available
3166 * from existing per node queues.
3167 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003168 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3169 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003170
Mel Gorman54a6eb52008-04-28 02:12:16 -07003171 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter6a673682013-01-10 19:14:19 +00003172 cache->node[nid] &&
3173 cache->node[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003174 obj = ____cache_alloc_node(cache,
3175 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003176 if (obj)
3177 break;
3178 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003179 }
3180
Christoph Lametercfce6602007-05-06 14:50:17 -07003181 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003182 /*
3183 * This allocation will be performed within the constraints
3184 * of the current cpuset / memory policy requirements.
3185 * We may trigger various forms of reclaim on the allowed
3186 * set and go into memory reserves if necessary.
3187 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003188 struct page *page;
3189
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003190 if (local_flags & __GFP_WAIT)
3191 local_irq_enable();
3192 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003193 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003194 if (local_flags & __GFP_WAIT)
3195 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003196 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003197 /*
3198 * Insert into the appropriate per node queues
3199 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003200 nid = page_to_nid(page);
3201 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003202 obj = ____cache_alloc_node(cache,
3203 flags | GFP_THISNODE, nid);
3204 if (!obj)
3205 /*
3206 * Another processor may allocate the
3207 * objects in the slab since we are
3208 * not holding any locks.
3209 */
3210 goto retry;
3211 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003212 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003213 obj = NULL;
3214 }
3215 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003216 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003217
Mel Gormand26914d2014-04-03 14:47:24 -07003218 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003219 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003220 return obj;
3221}
3222
3223/*
Christoph Lametere498be72005-09-09 13:03:32 -07003224 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003226static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003227 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003228{
3229 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003230 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003231 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003232 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003233 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003235 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003236 n = cachep->node[nodeid];
3237 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003238
Andrew Mortona737b3e2006-03-22 00:08:11 -08003239retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003240 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003241 spin_lock(&n->list_lock);
3242 entry = n->slabs_partial.next;
3243 if (entry == &n->slabs_partial) {
3244 n->free_touched = 1;
3245 entry = n->slabs_free.next;
3246 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003247 goto must_grow;
3248 }
Christoph Lametere498be72005-09-09 13:03:32 -07003249
Joonsoo Kim8456a642013-10-24 10:07:49 +09003250 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003251 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003252
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003253 STATS_INC_NODEALLOCS(cachep);
3254 STATS_INC_ACTIVE(cachep);
3255 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003256
Joonsoo Kim8456a642013-10-24 10:07:49 +09003257 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003258
Joonsoo Kim8456a642013-10-24 10:07:49 +09003259 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003260 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003261 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003262 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003263
Joonsoo Kim8456a642013-10-24 10:07:49 +09003264 if (page->active == cachep->num)
3265 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003266 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003267 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003268
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003269 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003270 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003271
Andrew Mortona737b3e2006-03-22 00:08:11 -08003272must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003273 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003274 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003275 if (x)
3276 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003277
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003278 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003279
Andrew Mortona737b3e2006-03-22 00:08:11 -08003280done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003281 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003282}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003283
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003284static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003285slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003286 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003287{
3288 unsigned long save_flags;
3289 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003290 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003291
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003292 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003293
Nick Piggincf40bd12009-01-21 08:12:39 +01003294 lockdep_trace_alloc(flags);
3295
Akinobu Mita773ff602008-12-23 19:37:01 +09003296 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003297 return NULL;
3298
Glauber Costad79923f2012-12-18 14:22:48 -08003299 cachep = memcg_kmem_get_cache(cachep, flags);
3300
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003301 cache_alloc_debugcheck_before(cachep, flags);
3302 local_irq_save(save_flags);
3303
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003304 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003305 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003306
Christoph Lameter6a673682013-01-10 19:14:19 +00003307 if (unlikely(!cachep->node[nodeid])) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003308 /* Node not bootstrapped yet */
3309 ptr = fallback_alloc(cachep, flags);
3310 goto out;
3311 }
3312
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003313 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003314 /*
3315 * Use the locally cached objects if possible.
3316 * However ____cache_alloc does not allow fallback
3317 * to other nodes. It may fail while we still have
3318 * objects on other nodes available.
3319 */
3320 ptr = ____cache_alloc(cachep, flags);
3321 if (ptr)
3322 goto out;
3323 }
3324 /* ___cache_alloc_node can fall back to other nodes */
3325 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3326 out:
3327 local_irq_restore(save_flags);
3328 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003329 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003330 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003331
Joe Perches5087c822013-09-10 17:02:51 -07003332 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003333 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003334 if (unlikely(flags & __GFP_ZERO))
3335 memset(ptr, 0, cachep->object_size);
3336 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003337
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003338 return ptr;
3339}
3340
3341static __always_inline void *
3342__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3343{
3344 void *objp;
3345
David Rientjesf0432d12014-04-07 15:37:30 -07003346 if (current->mempolicy || unlikely(current->flags & PF_SPREAD_SLAB)) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003347 objp = alternate_node_alloc(cache, flags);
3348 if (objp)
3349 goto out;
3350 }
3351 objp = ____cache_alloc(cache, flags);
3352
3353 /*
3354 * We may just have run out of memory on the local node.
3355 * ____cache_alloc_node() knows how to locate memory on other nodes
3356 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003357 if (!objp)
3358 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003359
3360 out:
3361 return objp;
3362}
3363#else
3364
3365static __always_inline void *
3366__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3367{
3368 return ____cache_alloc(cachep, flags);
3369}
3370
3371#endif /* CONFIG_NUMA */
3372
3373static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003374slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003375{
3376 unsigned long save_flags;
3377 void *objp;
3378
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003379 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003380
Nick Piggincf40bd12009-01-21 08:12:39 +01003381 lockdep_trace_alloc(flags);
3382
Akinobu Mita773ff602008-12-23 19:37:01 +09003383 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003384 return NULL;
3385
Glauber Costad79923f2012-12-18 14:22:48 -08003386 cachep = memcg_kmem_get_cache(cachep, flags);
3387
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003388 cache_alloc_debugcheck_before(cachep, flags);
3389 local_irq_save(save_flags);
3390 objp = __do_cache_alloc(cachep, flags);
3391 local_irq_restore(save_flags);
3392 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003393 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003394 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003395 prefetchw(objp);
3396
Joe Perches5087c822013-09-10 17:02:51 -07003397 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003398 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003399 if (unlikely(flags & __GFP_ZERO))
3400 memset(objp, 0, cachep->object_size);
3401 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003402
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003403 return objp;
3404}
Christoph Lametere498be72005-09-09 13:03:32 -07003405
3406/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003407 * Caller needs to acquire correct kmem_cache_node's list_lock
Christoph Lametere498be72005-09-09 13:03:32 -07003408 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003409static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003410 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411{
3412 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003413 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414
3415 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003416 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003417 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418
Mel Gorman072bb0a2012-07-31 16:43:58 -07003419 clear_obj_pfmemalloc(&objpp[i]);
3420 objp = objpp[i];
3421
Joonsoo Kim8456a642013-10-24 10:07:49 +09003422 page = virt_to_head_page(objp);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003423 n = cachep->node[node];
Joonsoo Kim8456a642013-10-24 10:07:49 +09003424 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003425 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003426 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003428 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429
3430 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003431 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003432 if (n->free_objects > n->free_limit) {
3433 n->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003434 /* No need to drop any previously held
3435 * lock here, even if we have a off-slab slab
3436 * descriptor it is guaranteed to come from
3437 * a different cache, refer to comments before
3438 * alloc_slabmgmt.
3439 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003440 slab_destroy(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003442 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443 }
3444 } else {
3445 /* Unconditionally move a slab to the end of the
3446 * partial list on free - maximum time for the
3447 * other objects to be freed, too.
3448 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003449 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450 }
3451 }
3452}
3453
Pekka Enberg343e0d72006-02-01 03:05:50 -08003454static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
3456 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003457 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003458 int node = numa_mem_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459
3460 batchcount = ac->batchcount;
3461#if DEBUG
3462 BUG_ON(!batchcount || batchcount > ac->avail);
3463#endif
3464 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003465 n = cachep->node[node];
3466 spin_lock(&n->list_lock);
3467 if (n->shared) {
3468 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003469 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 if (max) {
3471 if (batchcount > max)
3472 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003473 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003474 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 shared_array->avail += batchcount;
3476 goto free_done;
3477 }
3478 }
3479
Christoph Lameterff694162005-09-22 21:44:02 -07003480 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003481free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482#if STATS
3483 {
3484 int i = 0;
3485 struct list_head *p;
3486
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003487 p = n->slabs_free.next;
3488 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003489 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490
Joonsoo Kim8456a642013-10-24 10:07:49 +09003491 page = list_entry(p, struct page, lru);
3492 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493
3494 i++;
3495 p = p->next;
3496 }
3497 STATS_SET_FREEABLE(cachep, i);
3498 }
3499#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003500 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003502 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503}
3504
3505/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003506 * Release an obj back to its cache. If the obj has a constructed state, it must
3507 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003509static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003510 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003512 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513
3514 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003515 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003516 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003518 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003519
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003520 /*
3521 * Skip calling cache_free_alien() when the platform is not numa.
3522 * This will avoid cache misses that happen while accessing slabp (which
3523 * is per page memory reference) to get nodeid. Instead use a global
3524 * variable to skip the call, which is mostly likely to be present in
3525 * the cache.
3526 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003527 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003528 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003529
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530 if (likely(ac->avail < ac->limit)) {
3531 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 } else {
3533 STATS_INC_FREEMISS(cachep);
3534 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003536
Mel Gorman072bb0a2012-07-31 16:43:58 -07003537 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538}
3539
3540/**
3541 * kmem_cache_alloc - Allocate an object
3542 * @cachep: The cache to allocate from.
3543 * @flags: See kmalloc().
3544 *
3545 * Allocate an object from this cache. The flags are only relevant
3546 * if the cache has no available objects.
3547 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003548void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003550 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003551
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003552 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003553 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003554
3555 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556}
3557EXPORT_SYMBOL(kmem_cache_alloc);
3558
Li Zefan0f24f122009-12-11 15:45:30 +08003559#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003560void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003561kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003562{
Steven Rostedt85beb582010-11-24 16:23:34 -05003563 void *ret;
3564
Ezequiel Garcia48356302012-09-08 17:47:57 -03003565 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003566
3567 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003568 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003569 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003570}
Steven Rostedt85beb582010-11-24 16:23:34 -05003571EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003572#endif
3573
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003575/**
3576 * kmem_cache_alloc_node - Allocate an object on the specified node
3577 * @cachep: The cache to allocate from.
3578 * @flags: See kmalloc().
3579 * @nodeid: node number of the target node.
3580 *
3581 * Identical to kmem_cache_alloc but it will allocate memory on the given
3582 * node, which can improve the performance for cpu bound structures.
3583 *
3584 * Fallback to other node is possible if __GFP_THISNODE is not set.
3585 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003586void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3587{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003588 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003589
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003590 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003591 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003592 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003593
3594 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003595}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596EXPORT_SYMBOL(kmem_cache_alloc_node);
3597
Li Zefan0f24f122009-12-11 15:45:30 +08003598#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003599void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003600 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003601 int nodeid,
3602 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003603{
Steven Rostedt85beb582010-11-24 16:23:34 -05003604 void *ret;
3605
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003606 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003607
Steven Rostedt85beb582010-11-24 16:23:34 -05003608 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003609 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003610 flags, nodeid);
3611 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003612}
Steven Rostedt85beb582010-11-24 16:23:34 -05003613EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003614#endif
3615
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003616static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003617__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003618{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003619 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003620
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003621 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003622 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3623 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003624 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003625}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003626
Li Zefan0bb38a52009-12-11 15:45:50 +08003627#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003628void *__kmalloc_node(size_t size, gfp_t flags, int node)
3629{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003630 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003631}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003632EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003633
3634void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003635 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003636{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003637 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003638}
3639EXPORT_SYMBOL(__kmalloc_node_track_caller);
3640#else
3641void *__kmalloc_node(size_t size, gfp_t flags, int node)
3642{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003643 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003644}
3645EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003646#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003647#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648
3649/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003650 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003652 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003653 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003655static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003656 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003658 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003659 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003661 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003662 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3663 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003664 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003665
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003666 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003667 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003668
3669 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003670}
3671
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003672
Li Zefan0bb38a52009-12-11 15:45:50 +08003673#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003674void *__kmalloc(size_t size, gfp_t flags)
3675{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003676 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677}
3678EXPORT_SYMBOL(__kmalloc);
3679
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003680void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003681{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003682 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003683}
3684EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003685
3686#else
3687void *__kmalloc(size_t size, gfp_t flags)
3688{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003689 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003690}
3691EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003692#endif
3693
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694/**
3695 * kmem_cache_free - Deallocate an object
3696 * @cachep: The cache the allocation was from.
3697 * @objp: The previously allocated object.
3698 *
3699 * Free an object which was previously allocated from this
3700 * cache.
3701 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003702void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703{
3704 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003705 cachep = cache_from_obj(cachep, objp);
3706 if (!cachep)
3707 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708
3709 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003710 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003711 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003712 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003713 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003715
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003716 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717}
3718EXPORT_SYMBOL(kmem_cache_free);
3719
3720/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 * kfree - free previously allocated memory
3722 * @objp: pointer returned by kmalloc.
3723 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003724 * If @objp is NULL, no operation is performed.
3725 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 * Don't free memory not originally allocated by kmalloc()
3727 * or you will run into trouble.
3728 */
3729void kfree(const void *objp)
3730{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003731 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732 unsigned long flags;
3733
Pekka Enberg2121db72009-03-25 11:05:57 +02003734 trace_kfree(_RET_IP_, objp);
3735
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003736 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737 return;
3738 local_irq_save(flags);
3739 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003740 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003741 debug_check_no_locks_freed(objp, c->object_size);
3742
3743 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003744 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 local_irq_restore(flags);
3746}
3747EXPORT_SYMBOL(kfree);
3748
Christoph Lametere498be72005-09-09 13:03:32 -07003749/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003750 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003751 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003752static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003753{
3754 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003755 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003756 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003757 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003758
Mel Gorman9c09a952008-01-24 05:49:54 -08003759 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003760
Paul Menage3395ee02006-12-06 20:32:16 -08003761 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003762 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003763 if (!new_alien)
3764 goto fail;
3765 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003766
Eric Dumazet63109842007-05-06 14:49:28 -07003767 new_shared = NULL;
3768 if (cachep->shared) {
3769 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003770 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003771 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003772 if (!new_shared) {
3773 free_alien_cache(new_alien);
3774 goto fail;
3775 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003776 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003777
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003778 n = cachep->node[node];
3779 if (n) {
3780 struct array_cache *shared = n->shared;
Christoph Lametercafeb022006-03-25 03:06:46 -08003781
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003782 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003783
Christoph Lametercafeb022006-03-25 03:06:46 -08003784 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003785 free_block(cachep, shared->entry,
3786 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003787
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003788 n->shared = new_shared;
3789 if (!n->alien) {
3790 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003791 new_alien = NULL;
3792 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003793 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003794 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003795 spin_unlock_irq(&n->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003796 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003797 free_alien_cache(new_alien);
3798 continue;
3799 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003800 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3801 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003802 free_alien_cache(new_alien);
3803 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003804 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003805 }
Christoph Lametere498be72005-09-09 13:03:32 -07003806
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003807 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003808 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3809 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003810 n->shared = new_shared;
3811 n->alien = new_alien;
3812 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003813 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003814 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003815 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003816 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003817
Andrew Mortona737b3e2006-03-22 00:08:11 -08003818fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003819 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003820 /* Cache is not active yet. Roll back what we did */
3821 node--;
3822 while (node >= 0) {
Christoph Lameter6a673682013-01-10 19:14:19 +00003823 if (cachep->node[node]) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003824 n = cachep->node[node];
Christoph Lameter0718dc22006-03-25 03:06:47 -08003825
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003826 kfree(n->shared);
3827 free_alien_cache(n->alien);
3828 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003829 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003830 }
3831 node--;
3832 }
3833 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003834 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003835}
3836
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003838 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003839 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840};
3841
3842static void do_ccupdate_local(void *info)
3843{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003844 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 struct array_cache *old;
3846
3847 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003848 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003849
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3851 new->new[smp_processor_id()] = old;
3852}
3853
Christoph Lameter18004c52012-07-06 15:25:12 -05003854/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003855static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003856 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003857{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003858 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003859 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003861 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3862 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003863 if (!new)
3864 return -ENOMEM;
3865
Christoph Lametere498be72005-09-09 13:03:32 -07003866 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003867 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003868 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003869 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003870 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003871 kfree(new->new[i]);
3872 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003873 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 }
3875 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003876 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003878 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003879
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881 cachep->batchcount = batchcount;
3882 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003883 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884
Christoph Lametere498be72005-09-09 13:03:32 -07003885 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003886 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 if (!ccold)
3888 continue;
Christoph Lameter6a673682013-01-10 19:14:19 +00003889 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003890 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
Christoph Lameter6a673682013-01-10 19:14:19 +00003891 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892 kfree(ccold);
3893 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003894 kfree(new);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003895 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896}
3897
Glauber Costa943a4512012-12-18 14:23:03 -08003898static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3899 int batchcount, int shared, gfp_t gfp)
3900{
3901 int ret;
3902 struct kmem_cache *c = NULL;
3903 int i = 0;
3904
3905 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3906
3907 if (slab_state < FULL)
3908 return ret;
3909
3910 if ((ret < 0) || !is_root_cache(cachep))
3911 return ret;
3912
Glauber Costaebe945c2012-12-18 14:23:10 -08003913 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003914 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003915 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003916 if (c)
3917 /* return value determined by the parent cache only */
3918 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3919 }
3920
3921 return ret;
3922}
3923
Christoph Lameter18004c52012-07-06 15:25:12 -05003924/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003925static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926{
3927 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003928 int limit = 0;
3929 int shared = 0;
3930 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003931
Glauber Costa943a4512012-12-18 14:23:03 -08003932 if (!is_root_cache(cachep)) {
3933 struct kmem_cache *root = memcg_root_cache(cachep);
3934 limit = root->limit;
3935 shared = root->shared;
3936 batchcount = root->batchcount;
3937 }
3938
3939 if (limit && shared && batchcount)
3940 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003941 /*
3942 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 * - create a LIFO ordering, i.e. return objects that are cache-warm
3944 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003945 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946 * bufctl chains: array operations are cheaper.
3947 * The numbers are guessed, we should auto-tune as described by
3948 * Bonwick.
3949 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003950 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003952 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003954 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003956 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 limit = 54;
3958 else
3959 limit = 120;
3960
Andrew Mortona737b3e2006-03-22 00:08:11 -08003961 /*
3962 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963 * allocation behaviour: Most allocs on one cpu, most free operations
3964 * on another cpu. For these cases, an efficient object passing between
3965 * cpus is necessary. This is provided by a shared array. The array
3966 * replaces Bonwick's magazine layer.
3967 * On uniprocessor, it's functionally equivalent (but less efficient)
3968 * to a larger limit. Thus disabled by default.
3969 */
3970 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003971 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973
3974#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003975 /*
3976 * With debugging enabled, large batchcount lead to excessively long
3977 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 */
3979 if (limit > 32)
3980 limit = 32;
3981#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003982 batchcount = (limit + 1) / 2;
3983skip_setup:
3984 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 if (err)
3986 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003987 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003988 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989}
3990
Christoph Lameter1b552532006-03-22 00:09:07 -08003991/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003992 * Drain an array if it contains any elements taking the node lock only if
3993 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003994 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003995 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003996static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003997 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998{
3999 int tofree;
4000
Christoph Lameter1b552532006-03-22 00:09:07 -08004001 if (!ac || !ac->avail)
4002 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 if (ac->touched && !force) {
4004 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004005 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004006 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004007 if (ac->avail) {
4008 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4009 if (tofree > ac->avail)
4010 tofree = (ac->avail + 1) / 2;
4011 free_block(cachep, ac->entry, tofree, node);
4012 ac->avail -= tofree;
4013 memmove(ac->entry, &(ac->entry[tofree]),
4014 sizeof(void *) * ac->avail);
4015 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004016 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 }
4018}
4019
4020/**
4021 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004022 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 *
4024 * Called from workqueue/eventd every few seconds.
4025 * Purpose:
4026 * - clear the per-cpu caches for this CPU.
4027 * - return freeable pages to the main free memory pool.
4028 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004029 * If we cannot acquire the cache chain mutex then just give up - we'll try
4030 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004032static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004034 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004035 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004036 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004037 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038
Christoph Lameter18004c52012-07-06 15:25:12 -05004039 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004041 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042
Christoph Lameter18004c52012-07-06 15:25:12 -05004043 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 check_irq_on();
4045
Christoph Lameter35386e32006-03-22 00:09:05 -08004046 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004047 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004048 * have established with reasonable certainty that
4049 * we can do some work if the lock was obtained.
4050 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004051 n = searchp->node[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004052
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004053 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004055 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056
Christoph Lameter35386e32006-03-22 00:09:05 -08004057 /*
4058 * These are racy checks but it does not matter
4059 * if we skip one check or scan twice.
4060 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004061 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004062 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004064 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004066 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004068 if (n->free_touched)
4069 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004070 else {
4071 int freed;
4072
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004073 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004074 5 * searchp->num - 1) / (5 * searchp->num));
4075 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004077next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 cond_resched();
4079 }
4080 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004081 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004082 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004083out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004084 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004085 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086}
4087
Linus Torvalds158a9622008-01-02 13:04:48 -08004088#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004089void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004090{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004091 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004092 unsigned long active_objs;
4093 unsigned long num_objs;
4094 unsigned long active_slabs = 0;
4095 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004096 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004098 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004099 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 active_objs = 0;
4102 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004103 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004104 n = cachep->node[node];
4105 if (!n)
Christoph Lametere498be72005-09-09 13:03:32 -07004106 continue;
4107
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004108 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004109 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004110
Joonsoo Kim8456a642013-10-24 10:07:49 +09004111 list_for_each_entry(page, &n->slabs_full, lru) {
4112 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004113 error = "slabs_full accounting error";
4114 active_objs += cachep->num;
4115 active_slabs++;
4116 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004117 list_for_each_entry(page, &n->slabs_partial, lru) {
4118 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004119 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004120 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004121 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004122 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004123 active_slabs++;
4124 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004125 list_for_each_entry(page, &n->slabs_free, lru) {
4126 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004127 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004128 num_slabs++;
4129 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004130 free_objects += n->free_objects;
4131 if (n->shared)
4132 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004133
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004134 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004136 num_slabs += active_slabs;
4137 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004138 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004139 error = "free_objects accounting error";
4140
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004141 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 if (error)
4143 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4144
Glauber Costa0d7561c2012-10-19 18:20:27 +04004145 sinfo->active_objs = active_objs;
4146 sinfo->num_objs = num_objs;
4147 sinfo->active_slabs = active_slabs;
4148 sinfo->num_slabs = num_slabs;
4149 sinfo->shared_avail = shared_avail;
4150 sinfo->limit = cachep->limit;
4151 sinfo->batchcount = cachep->batchcount;
4152 sinfo->shared = cachep->shared;
4153 sinfo->objects_per_slab = cachep->num;
4154 sinfo->cache_order = cachep->gfporder;
4155}
4156
4157void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4158{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004160 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 unsigned long high = cachep->high_mark;
4162 unsigned long allocs = cachep->num_allocations;
4163 unsigned long grown = cachep->grown;
4164 unsigned long reaped = cachep->reaped;
4165 unsigned long errors = cachep->errors;
4166 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004168 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004169 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170
Joe Perchese92dd4f2010-03-26 19:27:58 -07004171 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4172 "%4lu %4lu %4lu %4lu %4lu",
4173 allocs, high, grown,
4174 reaped, errors, max_freeable, node_allocs,
4175 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004176 }
4177 /* cpu stats */
4178 {
4179 unsigned long allochit = atomic_read(&cachep->allochit);
4180 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4181 unsigned long freehit = atomic_read(&cachep->freehit);
4182 unsigned long freemiss = atomic_read(&cachep->freemiss);
4183
4184 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004185 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186 }
4187#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188}
4189
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190#define MAX_SLABINFO_WRITE 128
4191/**
4192 * slabinfo_write - Tuning for the slab allocator
4193 * @file: unused
4194 * @buffer: user buffer
4195 * @count: data length
4196 * @ppos: unused
4197 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004198ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004199 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004200{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004201 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004203 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004204
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 if (count > MAX_SLABINFO_WRITE)
4206 return -EINVAL;
4207 if (copy_from_user(&kbuf, buffer, count))
4208 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004209 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210
4211 tmp = strchr(kbuf, ' ');
4212 if (!tmp)
4213 return -EINVAL;
4214 *tmp = '\0';
4215 tmp++;
4216 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4217 return -EINVAL;
4218
4219 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004220 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004222 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004224 if (limit < 1 || batchcount < 1 ||
4225 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004226 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004228 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004229 batchcount, shared,
4230 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004231 }
4232 break;
4233 }
4234 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004235 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 if (res >= 0)
4237 res = count;
4238 return res;
4239}
Al Viro871751e2006-03-25 03:06:39 -08004240
4241#ifdef CONFIG_DEBUG_SLAB_LEAK
4242
4243static void *leaks_start(struct seq_file *m, loff_t *pos)
4244{
Christoph Lameter18004c52012-07-06 15:25:12 -05004245 mutex_lock(&slab_mutex);
4246 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004247}
4248
4249static inline int add_caller(unsigned long *n, unsigned long v)
4250{
4251 unsigned long *p;
4252 int l;
4253 if (!v)
4254 return 1;
4255 l = n[1];
4256 p = n + 2;
4257 while (l) {
4258 int i = l/2;
4259 unsigned long *q = p + 2 * i;
4260 if (*q == v) {
4261 q[1]++;
4262 return 1;
4263 }
4264 if (*q > v) {
4265 l = i;
4266 } else {
4267 p = q + 2;
4268 l -= i + 1;
4269 }
4270 }
4271 if (++n[1] == n[0])
4272 return 0;
4273 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4274 p[0] = v;
4275 p[1] = 1;
4276 return 1;
4277}
4278
Joonsoo Kim8456a642013-10-24 10:07:49 +09004279static void handle_slab(unsigned long *n, struct kmem_cache *c,
4280 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004281{
4282 void *p;
Joonsoo Kim03787302014-06-23 13:22:06 -07004283 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004284
Al Viro871751e2006-03-25 03:06:39 -08004285 if (n[0] == n[1])
4286 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004287 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kim03787302014-06-23 13:22:06 -07004288 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004289 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004290
Al Viro871751e2006-03-25 03:06:39 -08004291 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4292 return;
4293 }
4294}
4295
4296static void show_symbol(struct seq_file *m, unsigned long address)
4297{
4298#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004299 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004300 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004301
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004302 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004303 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004304 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004305 seq_printf(m, " [%s]", modname);
4306 return;
4307 }
4308#endif
4309 seq_printf(m, "%p", (void *)address);
4310}
4311
4312static int leaks_show(struct seq_file *m, void *p)
4313{
Thierry Reding0672aa72012-06-22 19:42:49 +02004314 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004315 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004316 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004317 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004318 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004319 int node;
4320 int i;
4321
4322 if (!(cachep->flags & SLAB_STORE_USER))
4323 return 0;
4324 if (!(cachep->flags & SLAB_RED_ZONE))
4325 return 0;
4326
4327 /* OK, we can do it */
4328
Christoph Lameterdb845062013-02-05 18:45:23 +00004329 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004330
4331 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004332 n = cachep->node[node];
4333 if (!n)
Al Viro871751e2006-03-25 03:06:39 -08004334 continue;
4335
4336 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004337 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004338
Joonsoo Kim8456a642013-10-24 10:07:49 +09004339 list_for_each_entry(page, &n->slabs_full, lru)
4340 handle_slab(x, cachep, page);
4341 list_for_each_entry(page, &n->slabs_partial, lru)
4342 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004343 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004344 }
4345 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004346 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004347 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004348 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004349 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004350 if (!m->private) {
4351 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004352 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004353 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004354 return -ENOMEM;
4355 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004356 *(unsigned long *)m->private = x[0] * 2;
4357 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004358 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004359 /* Now make sure this entry will be retried */
4360 m->count = m->size;
4361 return 0;
4362 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004363 for (i = 0; i < x[1]; i++) {
4364 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4365 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004366 seq_putc(m, '\n');
4367 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004368
Al Viro871751e2006-03-25 03:06:39 -08004369 return 0;
4370}
4371
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004372static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004373 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004374 .next = slab_next,
4375 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004376 .show = leaks_show,
4377};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004378
4379static int slabstats_open(struct inode *inode, struct file *file)
4380{
4381 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4382 int ret = -ENOMEM;
4383 if (n) {
4384 ret = seq_open(file, &slabstats_op);
4385 if (!ret) {
4386 struct seq_file *m = file->private_data;
4387 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4388 m->private = n;
4389 n = NULL;
4390 }
4391 kfree(n);
4392 }
4393 return ret;
4394}
4395
4396static const struct file_operations proc_slabstats_operations = {
4397 .open = slabstats_open,
4398 .read = seq_read,
4399 .llseek = seq_lseek,
4400 .release = seq_release_private,
4401};
Al Viro871751e2006-03-25 03:06:39 -08004402#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004403
4404static int __init slab_proc_init(void)
4405{
4406#ifdef CONFIG_DEBUG_SLAB_LEAK
4407 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4408#endif
4409 return 0;
4410}
4411module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004412#endif
4413
Manfred Spraul00e145b2005-09-03 15:55:07 -07004414/**
4415 * ksize - get the actual amount of memory allocated for a given object
4416 * @objp: Pointer to the object
4417 *
4418 * kmalloc may internally round up allocations and return more memory
4419 * than requested. ksize() can be used to determine the actual amount of
4420 * memory allocated. The caller may use this additional memory, even though
4421 * a smaller amount of memory was initially specified with the kmalloc call.
4422 * The caller must guarantee that objp points to a valid object previously
4423 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4424 * must not be freed during the duration of the call.
4425 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004426size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004428 BUG_ON(!objp);
4429 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004430 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004431
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004432 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004433}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004434EXPORT_SYMBOL(ksize);