blob: a467b308c682334254c53fdf12a6114aad1ce1fd [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;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700194 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
197 * the entries.
Mel Gorman072bb0a2012-07-31 16:43:58 -0700198 *
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -0800202 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203};
204
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700205struct alien_cache {
206 spinlock_t lock;
207 struct array_cache ac;
208};
209
Mel Gorman072bb0a2012-07-31 16:43:58 -0700210#define SLAB_OBJ_PFMEMALLOC 1
211static inline bool is_obj_pfmemalloc(void *objp)
212{
213 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
214}
215
216static inline void set_obj_pfmemalloc(void **objp)
217{
218 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
219 return;
220}
221
222static inline void clear_obj_pfmemalloc(void **objp)
223{
224 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
225}
226
Andrew Mortona737b3e2006-03-22 00:08:11 -0800227/*
228 * bootstrap: The caches do not work without cpuarrays anymore, but the
229 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 */
231#define BOOT_CPUCACHE_ENTRIES 1
232struct arraycache_init {
233 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800234 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235};
236
237/*
Christoph Lametere498be72005-09-09 13:03:32 -0700238 * Need this for bootstrapping a per node allocator.
239 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200240#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000241static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700242#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200243#define SIZE_AC MAX_NUMNODES
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000244#define SIZE_NODE (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245
Christoph Lametered11d9e2006-06-30 01:55:45 -0700246static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000247 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700248static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700249 int node, struct list_head *list);
250static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300251static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000252static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700253
Ingo Molnare0a42722006-06-23 02:03:46 -0700254static int slab_early_init = 1;
255
Christoph Lametere3366012013-01-10 19:14:18 +0000256#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000257#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700258
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000259static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700260{
261 INIT_LIST_HEAD(&parent->slabs_full);
262 INIT_LIST_HEAD(&parent->slabs_partial);
263 INIT_LIST_HEAD(&parent->slabs_free);
264 parent->shared = NULL;
265 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800266 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700267 spin_lock_init(&parent->list_lock);
268 parent->free_objects = 0;
269 parent->free_touched = 0;
270}
271
Andrew Mortona737b3e2006-03-22 00:08:11 -0800272#define MAKE_LIST(cachep, listp, slab, nodeid) \
273 do { \
274 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700275 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700276 } while (0)
277
Andrew Mortona737b3e2006-03-22 00:08:11 -0800278#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
279 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700280 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
281 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
282 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
283 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285#define CFLGS_OFF_SLAB (0x80000000UL)
286#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
287
288#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800289/*
290 * Optimization question: fewer reaps means less probability for unnessary
291 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100293 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 * which could lock up otherwise freeable slabs.
295 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800296#define REAPTIMEOUT_AC (2*HZ)
297#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298
299#if STATS
300#define STATS_INC_ACTIVE(x) ((x)->num_active++)
301#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
302#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
303#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700304#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800305#define STATS_SET_HIGH(x) \
306 do { \
307 if ((x)->num_active > (x)->high_mark) \
308 (x)->high_mark = (x)->num_active; \
309 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310#define STATS_INC_ERR(x) ((x)->errors++)
311#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700312#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700313#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800314#define STATS_SET_FREEABLE(x, i) \
315 do { \
316 if ((x)->max_freeable < i) \
317 (x)->max_freeable = i; \
318 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
320#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
321#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
322#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
323#else
324#define STATS_INC_ACTIVE(x) do { } while (0)
325#define STATS_DEC_ACTIVE(x) do { } while (0)
326#define STATS_INC_ALLOCED(x) do { } while (0)
327#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700328#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329#define STATS_SET_HIGH(x) do { } while (0)
330#define STATS_INC_ERR(x) do { } while (0)
331#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700332#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700333#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800334#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335#define STATS_INC_ALLOCHIT(x) do { } while (0)
336#define STATS_INC_ALLOCMISS(x) do { } while (0)
337#define STATS_INC_FREEHIT(x) do { } while (0)
338#define STATS_INC_FREEMISS(x) do { } while (0)
339#endif
340
341#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342
Andrew Mortona737b3e2006-03-22 00:08:11 -0800343/*
344 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800346 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 * the end of an object is aligned with the end of the real
348 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800349 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800351 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500352 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
353 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800354 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800356static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800358 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359}
360
David Woodhouseb46b8f12007-05-08 00:22:59 -0700361static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362{
363 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700364 return (unsigned long long*) (objp + obj_offset(cachep) -
365 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366}
367
David Woodhouseb46b8f12007-05-08 00:22:59 -0700368static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369{
370 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
371 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500372 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700373 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400374 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500375 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700376 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377}
378
Pekka Enberg343e0d72006-02-01 03:05:50 -0800379static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
381 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500382 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383}
384
385#else
386
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800387#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700388#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
391
392#endif
393
Joonsoo Kim03787302014-06-23 13:22:06 -0700394#define OBJECT_FREE (0)
395#define OBJECT_ACTIVE (1)
396
397#ifdef CONFIG_DEBUG_SLAB_LEAK
398
399static void set_obj_status(struct page *page, int idx, int val)
400{
401 int freelist_size;
402 char *status;
403 struct kmem_cache *cachep = page->slab_cache;
404
405 freelist_size = cachep->num * sizeof(freelist_idx_t);
406 status = (char *)page->freelist + freelist_size;
407 status[idx] = val;
408}
409
410static inline unsigned int get_obj_status(struct page *page, int idx)
411{
412 int freelist_size;
413 char *status;
414 struct kmem_cache *cachep = page->slab_cache;
415
416 freelist_size = cachep->num * sizeof(freelist_idx_t);
417 status = (char *)page->freelist + freelist_size;
418
419 return status[idx];
420}
421
422#else
423static inline void set_obj_status(struct page *page, int idx, int val) {}
424
425#endif
426
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700428 * Do not go above this order unless 0 objects fit into the slab or
429 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 */
David Rientjes543585c2011-10-18 22:09:24 -0700431#define SLAB_MAX_ORDER_HI 1
432#define SLAB_MAX_ORDER_LO 0
433static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700434static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800436static inline struct kmem_cache *virt_to_cache(const void *obj)
437{
Christoph Lameterb49af682007-05-06 14:49:41 -0700438 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500439 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800440}
441
Joonsoo Kim8456a642013-10-24 10:07:49 +0900442static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800443 unsigned int idx)
444{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900445 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800446}
447
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800448/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500449 * We want to avoid an expensive divide : (offset / cache->size)
450 * Using the fact that size is a constant for a particular cache,
451 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800452 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
453 */
454static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900455 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800456{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900457 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800458 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800459}
460
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800462 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
464/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000465static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800466 .batchcount = 1,
467 .limit = BOOT_CPUCACHE_ENTRIES,
468 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500469 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800470 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471};
472
Joonsoo Kimedcad252014-08-08 14:19:15 -0700473#define BAD_ALIEN_MAGIC 0x01020304ul
474
Tejun Heo1871e522009-10-29 22:34:13 +0900475static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Pekka Enberg343e0d72006-02-01 03:05:50 -0800477static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478{
479 return cachep->array[smp_processor_id()];
480}
481
Joonsoo Kim03787302014-06-23 13:22:06 -0700482static size_t calculate_freelist_size(int nr_objs, size_t align)
483{
484 size_t freelist_size;
485
486 freelist_size = nr_objs * sizeof(freelist_idx_t);
487 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
488 freelist_size += nr_objs * sizeof(char);
489
490 if (align)
491 freelist_size = ALIGN(freelist_size, align);
492
493 return freelist_size;
494}
495
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900496static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
497 size_t idx_size, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498{
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900499 int nr_objs;
Joonsoo Kim03787302014-06-23 13:22:06 -0700500 size_t remained_size;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900501 size_t freelist_size;
Joonsoo Kim03787302014-06-23 13:22:06 -0700502 int extra_space = 0;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900503
Joonsoo Kim03787302014-06-23 13:22:06 -0700504 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
505 extra_space = sizeof(char);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900506 /*
507 * Ignore padding for the initial guess. The padding
508 * is at most @align-1 bytes, and @buffer_size is at
509 * least @align. In the worst case, this result will
510 * be one greater than the number of objects that fit
511 * into the memory allocation when taking the padding
512 * into account.
513 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700514 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900515
516 /*
517 * This calculated number will be either the right
518 * amount, or one greater than what we want.
519 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700520 remained_size = slab_size - nr_objs * buffer_size;
521 freelist_size = calculate_freelist_size(nr_objs, align);
522 if (remained_size < freelist_size)
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900523 nr_objs--;
524
525 return nr_objs;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800526}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527
Andrew Mortona737b3e2006-03-22 00:08:11 -0800528/*
529 * Calculate the number of objects and left-over bytes for a given buffer size.
530 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800531static void cache_estimate(unsigned long gfporder, size_t buffer_size,
532 size_t align, int flags, size_t *left_over,
533 unsigned int *num)
534{
535 int nr_objs;
536 size_t mgmt_size;
537 size_t slab_size = PAGE_SIZE << gfporder;
538
539 /*
540 * The slab management structure can be either off the slab or
541 * on it. For the latter case, the memory allocated for a
542 * slab is used for:
543 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900544 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800545 * - Padding to respect alignment of @align
546 * - @buffer_size bytes for each object
547 *
548 * If the slab management structure is off the slab, then the
549 * alignment will already be calculated into the size. Because
550 * the slabs are all pages aligned, the objects will be at the
551 * correct alignment when allocated.
552 */
553 if (flags & CFLGS_OFF_SLAB) {
554 mgmt_size = 0;
555 nr_objs = slab_size / buffer_size;
556
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800557 } else {
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900558 nr_objs = calculate_nr_objs(slab_size, buffer_size,
Joonsoo Kima41adfa2013-12-02 17:49:42 +0900559 sizeof(freelist_idx_t), align);
Joonsoo Kim03787302014-06-23 13:22:06 -0700560 mgmt_size = calculate_freelist_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800562 *num = nr_objs;
563 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564}
565
Christoph Lameterf28510d2012-09-11 19:49:38 +0000566#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700567#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568
Andrew Mortona737b3e2006-03-22 00:08:11 -0800569static void __slab_error(const char *function, struct kmem_cache *cachep,
570 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571{
572 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800573 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030575 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000577#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578
Paul Menage3395ee02006-12-06 20:32:16 -0800579/*
580 * By default on NUMA we use alien caches to stage the freeing of
581 * objects allocated from other nodes. This causes massive memory
582 * inefficiencies when using fake NUMA setup to split memory into a
583 * large number of small nodes, so it can be disabled on the command
584 * line
585 */
586
587static int use_alien_caches __read_mostly = 1;
588static int __init noaliencache_setup(char *s)
589{
590 use_alien_caches = 0;
591 return 1;
592}
593__setup("noaliencache", noaliencache_setup);
594
David Rientjes3df1ccc2011-10-18 22:09:28 -0700595static int __init slab_max_order_setup(char *str)
596{
597 get_option(&str, &slab_max_order);
598 slab_max_order = slab_max_order < 0 ? 0 :
599 min(slab_max_order, MAX_ORDER - 1);
600 slab_max_order_set = true;
601
602 return 1;
603}
604__setup("slab_max_order=", slab_max_order_setup);
605
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800606#ifdef CONFIG_NUMA
607/*
608 * Special reaping functions for NUMA systems called from cache_reap().
609 * These take care of doing round robin flushing of alien caches (containing
610 * objects freed on different nodes from which they were allocated) and the
611 * flushing of remote pcps by calling drain_node_pages.
612 */
Tejun Heo1871e522009-10-29 22:34:13 +0900613static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800614
615static void init_reap_node(int cpu)
616{
617 int node;
618
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700619 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800620 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800621 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800622
Tejun Heo1871e522009-10-29 22:34:13 +0900623 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800624}
625
626static void next_reap_node(void)
627{
Christoph Lameter909ea962010-12-08 16:22:55 +0100628 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800629
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800630 node = next_node(node, node_online_map);
631 if (unlikely(node >= MAX_NUMNODES))
632 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100633 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800634}
635
636#else
637#define init_reap_node(cpu) do { } while (0)
638#define next_reap_node(void) do { } while (0)
639#endif
640
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641/*
642 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
643 * via the workqueue/eventd.
644 * Add the CPU number into the expiration time to minimize the possibility of
645 * the CPUs getting into lockstep and contending for the global cache chain
646 * lock.
647 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400648static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649{
Tejun Heo1871e522009-10-29 22:34:13 +0900650 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
652 /*
653 * When this gets called from do_initcalls via cpucache_init(),
654 * init_workqueues() has already run, so keventd will be setup
655 * at that time.
656 */
David Howells52bad642006-11-22 14:54:01 +0000657 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800658 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700659 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800660 schedule_delayed_work_on(cpu, reap_work,
661 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662 }
663}
664
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700665static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100667 /*
668 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300669 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100670 * cache the pointers are not cleared and they could be counted as
671 * valid references during a kmemleak scan. Therefore, kmemleak must
672 * not scan such objects.
673 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700674 kmemleak_no_scan(ac);
675 if (ac) {
676 ac->avail = 0;
677 ac->limit = limit;
678 ac->batchcount = batch;
679 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700681}
682
683static struct array_cache *alloc_arraycache(int node, int entries,
684 int batchcount, gfp_t gfp)
685{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700686 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700687 struct array_cache *ac = NULL;
688
689 ac = kmalloc_node(memsize, gfp, node);
690 init_arraycache(ac, entries, batchcount);
691 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692}
693
Joonsoo Kim8456a642013-10-24 10:07:49 +0900694static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700695{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700696 return PageSlabPfmemalloc(page);
697}
698
699/* Clears pfmemalloc_active if no slabs have pfmalloc set */
700static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
701 struct array_cache *ac)
702{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700703 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
Joonsoo Kim8456a642013-10-24 10:07:49 +0900704 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700705 unsigned long flags;
706
707 if (!pfmemalloc_active)
708 return;
709
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000710 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900711 list_for_each_entry(page, &n->slabs_full, lru)
712 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700713 goto out;
714
Joonsoo Kim8456a642013-10-24 10:07:49 +0900715 list_for_each_entry(page, &n->slabs_partial, lru)
716 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700717 goto out;
718
Joonsoo Kim8456a642013-10-24 10:07:49 +0900719 list_for_each_entry(page, &n->slabs_free, lru)
720 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700721 goto out;
722
723 pfmemalloc_active = false;
724out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000725 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700726}
727
Mel Gorman381760e2012-07-31 16:44:30 -0700728static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700729 gfp_t flags, bool force_refill)
730{
731 int i;
732 void *objp = ac->entry[--ac->avail];
733
734 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
735 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000736 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700737
738 if (gfp_pfmemalloc_allowed(flags)) {
739 clear_obj_pfmemalloc(&objp);
740 return objp;
741 }
742
743 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700744 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700745 /* If a !PFMEMALLOC object is found, swap them */
746 if (!is_obj_pfmemalloc(ac->entry[i])) {
747 objp = ac->entry[i];
748 ac->entry[i] = ac->entry[ac->avail];
749 ac->entry[ac->avail] = objp;
750 return objp;
751 }
752 }
753
754 /*
755 * If there are empty slabs on the slabs_free list and we are
756 * being forced to refill the cache, mark this one !pfmemalloc.
757 */
Christoph Lameter18bf8542014-08-06 16:04:11 -0700758 n = get_node(cachep, numa_mem_id());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000759 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900760 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900761 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700762 clear_obj_pfmemalloc(&objp);
763 recheck_pfmemalloc_active(cachep, ac);
764 return objp;
765 }
766
767 /* No !PFMEMALLOC objects available */
768 ac->avail++;
769 objp = NULL;
770 }
771
772 return objp;
773}
774
Mel Gorman381760e2012-07-31 16:44:30 -0700775static inline void *ac_get_obj(struct kmem_cache *cachep,
776 struct array_cache *ac, gfp_t flags, bool force_refill)
777{
778 void *objp;
779
780 if (unlikely(sk_memalloc_socks()))
781 objp = __ac_get_obj(cachep, ac, flags, force_refill);
782 else
783 objp = ac->entry[--ac->avail];
784
785 return objp;
786}
787
788static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700789 void *objp)
790{
791 if (unlikely(pfmemalloc_active)) {
792 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700793 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700794 if (PageSlabPfmemalloc(page))
795 set_obj_pfmemalloc(&objp);
796 }
797
Mel Gorman381760e2012-07-31 16:44:30 -0700798 return objp;
799}
800
801static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
802 void *objp)
803{
804 if (unlikely(sk_memalloc_socks()))
805 objp = __ac_put_obj(cachep, ac, objp);
806
Mel Gorman072bb0a2012-07-31 16:43:58 -0700807 ac->entry[ac->avail++] = objp;
808}
809
Christoph Lameter3ded1752006-03-25 03:06:44 -0800810/*
811 * Transfer objects in one arraycache to another.
812 * Locking must be handled by the caller.
813 *
814 * Return the number of entries transferred.
815 */
816static int transfer_objects(struct array_cache *to,
817 struct array_cache *from, unsigned int max)
818{
819 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700820 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800821
822 if (!nr)
823 return 0;
824
825 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
826 sizeof(void *) *nr);
827
828 from->avail -= nr;
829 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800830 return nr;
831}
832
Christoph Lameter765c4502006-09-27 01:50:08 -0700833#ifndef CONFIG_NUMA
834
835#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000836#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700837
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700838static inline struct alien_cache **alloc_alien_cache(int node,
839 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700840{
Joonsoo Kimedcad252014-08-08 14:19:15 -0700841 return (struct alien_cache **)BAD_ALIEN_MAGIC;
Christoph Lameter765c4502006-09-27 01:50:08 -0700842}
843
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700844static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700845{
846}
847
848static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
849{
850 return 0;
851}
852
853static inline void *alternate_node_alloc(struct kmem_cache *cachep,
854 gfp_t flags)
855{
856 return NULL;
857}
858
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800859static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700860 gfp_t flags, int nodeid)
861{
862 return NULL;
863}
864
865#else /* CONFIG_NUMA */
866
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800867static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800868static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800869
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700870static struct alien_cache *__alloc_alien_cache(int node, int entries,
871 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700872{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700873 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700874 struct alien_cache *alc = NULL;
875
876 alc = kmalloc_node(memsize, gfp, node);
877 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700878 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700879 return alc;
880}
881
882static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
883{
884 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700885 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700886 int i;
887
888 if (limit > 1)
889 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700890 alc_ptr = kzalloc_node(memsize, gfp, node);
891 if (!alc_ptr)
892 return NULL;
893
894 for_each_node(i) {
895 if (i == node || !node_online(i))
896 continue;
897 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
898 if (!alc_ptr[i]) {
899 for (i--; i >= 0; i--)
900 kfree(alc_ptr[i]);
901 kfree(alc_ptr);
902 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700903 }
904 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700905 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700906}
907
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700908static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700909{
910 int i;
911
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700912 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700913 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700914 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700915 kfree(alc_ptr[i]);
916 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700917}
918
Pekka Enberg343e0d72006-02-01 03:05:50 -0800919static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700920 struct array_cache *ac, int node,
921 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700922{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700923 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700924
925 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000926 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800927 /*
928 * Stuff objects into the remote nodes shared array first.
929 * That way we could avoid the overhead of putting the objects
930 * into the free lists and getting them back later.
931 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000932 if (n->shared)
933 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800934
Joonsoo Kim833b7062014-08-06 16:04:33 -0700935 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700936 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000937 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700938 }
939}
940
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800941/*
942 * Called from cache_reap() to regularly drain alien caches round robin.
943 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000944static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800945{
Christoph Lameter909ea962010-12-08 16:22:55 +0100946 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800947
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000948 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700949 struct alien_cache *alc = n->alien[node];
950 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800951
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700952 if (alc) {
953 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700954 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700955 LIST_HEAD(list);
956
957 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700958 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700959 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700960 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800961 }
962 }
963}
964
Andrew Mortona737b3e2006-03-22 00:08:11 -0800965static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700966 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700967{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800968 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700969 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700970 struct array_cache *ac;
971 unsigned long flags;
972
973 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700974 alc = alien[i];
975 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700976 LIST_HEAD(list);
977
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700978 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700979 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700980 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700981 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700982 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700983 }
984 }
985}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700986
Ingo Molnar873623d2006-07-13 14:44:38 +0200987static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700988{
Joonsoo Kim1ea991b2013-10-24 10:07:40 +0900989 int nodeid = page_to_nid(virt_to_page(objp));
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000990 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700991 struct alien_cache *alien = NULL;
992 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700993 int node;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700994 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700995
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700996 node = numa_mem_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700997
998 /*
999 * Make sure we are not freeing a object from another node to the array
1000 * cache on this cpu.
1001 */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09001002 if (likely(nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001003 return 0;
1004
Christoph Lameter18bf8542014-08-06 16:04:11 -07001005 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001006 STATS_INC_NODEFREES(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001007 if (n->alien && n->alien[nodeid]) {
1008 alien = n->alien[nodeid];
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001009 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001010 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001011 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001012 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim833b7062014-08-06 16:04:33 -07001013 __drain_alien_cache(cachep, ac, nodeid, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001014 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001015 ac_put_obj(cachep, ac, objp);
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001016 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -07001017 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001018 } else {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001019 n = get_node(cachep, nodeid);
1020 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07001021 free_block(cachep, &objp, 1, nodeid, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001022 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07001023 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001024 }
1025 return 1;
1026}
Christoph Lametere498be72005-09-09 13:03:32 -07001027#endif
1028
David Rientjes8f9f8d92010-03-27 19:40:47 -07001029/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001030 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001031 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001032 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001033 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001034 * already in use.
1035 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001036 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001037 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001038static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001039{
1040 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001041 struct kmem_cache_node *n;
Joonsoo Kim5e804782014-08-06 16:04:40 -07001042 const size_t memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001043
Christoph Lameter18004c52012-07-06 15:25:12 -05001044 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001045 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001046 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001047 * begin anything. Make sure some other cpu on this
1048 * node has not already allocated this
1049 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07001050 n = get_node(cachep, node);
1051 if (!n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001052 n = kmalloc_node(memsize, GFP_KERNEL, node);
1053 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001054 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001055 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001056 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1057 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001058
1059 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001060 * The kmem_cache_nodes don't come and go as CPUs
1061 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001062 * protection here.
1063 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001064 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001065 }
1066
Christoph Lameter18bf8542014-08-06 16:04:11 -07001067 spin_lock_irq(&n->list_lock);
1068 n->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001069 (1 + nr_cpus_node(node)) *
1070 cachep->batchcount + cachep->num;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001071 spin_unlock_irq(&n->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001072 }
1073 return 0;
1074}
1075
Wanpeng Li0fa81032013-07-04 08:33:22 +08001076static inline int slabs_tofree(struct kmem_cache *cachep,
1077 struct kmem_cache_node *n)
1078{
1079 return (n->free_objects + cachep->num - 1) / cachep->num;
1080}
1081
Paul Gortmaker0db06282013-06-19 14:53:51 -04001082static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001084 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001085 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001086 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301087 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001088
Christoph Lameter18004c52012-07-06 15:25:12 -05001089 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001090 struct array_cache *nc;
1091 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001092 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -07001093 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001094
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001095 /* cpu is dead; no one can alloc from it. */
1096 nc = cachep->array[cpu];
1097 cachep->array[cpu] = NULL;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001098 n = get_node(cachep, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001099
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001100 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001101 goto free_array_cache;
1102
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001103 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001104
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001105 /* Free limit for this kmem_cache_node */
1106 n->free_limit -= cachep->batchcount;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001107 if (nc)
Joonsoo Kim97654df2014-08-06 16:04:25 -07001108 free_block(cachep, nc->entry, nc->avail, node, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001109
Rusty Russell58463c12009-12-17 11:43:12 -06001110 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001111 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001112 goto free_array_cache;
1113 }
1114
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001115 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001116 if (shared) {
1117 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001118 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001119 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001120 }
1121
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001122 alien = n->alien;
1123 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001124
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001126
1127 kfree(shared);
1128 if (alien) {
1129 drain_alien_cache(cachep, alien);
1130 free_alien_cache(alien);
1131 }
1132free_array_cache:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001133 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001134 kfree(nc);
1135 }
1136 /*
1137 * In the previous loop, all the objects were freed to
1138 * the respective cache's slabs, now we can go ahead and
1139 * shrink each nodelist to its limit.
1140 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001141 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001142 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001143 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001144 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001145 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001146 }
1147}
1148
Paul Gortmaker0db06282013-06-19 14:53:51 -04001149static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001150{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001151 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001152 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001153 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001154 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001156 /*
1157 * We need to do this right in the beginning since
1158 * alloc_arraycache's are going to use this list.
1159 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001160 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001161 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001162 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001163 if (err < 0)
1164 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001165
1166 /*
1167 * Now we can go ahead with allocating the shared arrays and
1168 * array caches
1169 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001170 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001171 struct array_cache *nc;
1172 struct array_cache *shared = NULL;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001173 struct alien_cache **alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001174
1175 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001176 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001177 if (!nc)
1178 goto bad;
1179 if (cachep->shared) {
1180 shared = alloc_arraycache(node,
1181 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001182 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001183 if (!shared) {
1184 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001186 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001187 }
1188 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001189 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001190 if (!alien) {
1191 kfree(shared);
1192 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001193 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001194 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001195 }
1196 cachep->array[cpu] = nc;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001197 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001198 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001200 spin_lock_irq(&n->list_lock);
1201 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001202 /*
1203 * We are serialised from CPU_DEAD or
1204 * CPU_UP_CANCELLED by the cpucontrol lock
1205 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001206 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001207 shared = NULL;
1208 }
1209#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001210 if (!n->alien) {
1211 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001212 alien = NULL;
1213 }
1214#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001215 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001216 kfree(shared);
1217 free_alien_cache(alien);
1218 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001219
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001220 return 0;
1221bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001222 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001223 return -ENOMEM;
1224}
1225
Paul Gortmaker0db06282013-06-19 14:53:51 -04001226static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001227 unsigned long action, void *hcpu)
1228{
1229 long cpu = (long)hcpu;
1230 int err = 0;
1231
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001233 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001234 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001235 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001236 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001237 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 break;
1239 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001240 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 start_cpu_timer(cpu);
1242 break;
1243#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001244 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001245 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001246 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001247 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001248 * held so that if cache_reap() is invoked it cannot do
1249 * anything expensive but will only modify reap_work
1250 * and reschedule the timer.
1251 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001252 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001253 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001254 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001255 break;
1256 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001257 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001258 start_cpu_timer(cpu);
1259 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001261 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001262 /*
1263 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001264 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001265 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001266 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001267 * structure is usually allocated from kmem_cache_create() and
1268 * gets destroyed at kmem_cache_destroy().
1269 */
Simon Arlott183ff222007-10-20 01:27:18 +02001270 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001271#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001273 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001274 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001275 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001276 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001279 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280}
1281
Paul Gortmaker0db06282013-06-19 14:53:51 -04001282static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001283 &cpuup_callback, NULL, 0
1284};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285
David Rientjes8f9f8d92010-03-27 19:40:47 -07001286#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1287/*
1288 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1289 * Returns -EBUSY if all objects cannot be drained so that the node is not
1290 * removed.
1291 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001292 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001293 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001294static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001295{
1296 struct kmem_cache *cachep;
1297 int ret = 0;
1298
Christoph Lameter18004c52012-07-06 15:25:12 -05001299 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001300 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001301
Christoph Lameter18bf8542014-08-06 16:04:11 -07001302 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001303 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001304 continue;
1305
Wanpeng Li0fa81032013-07-04 08:33:22 +08001306 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001307
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001308 if (!list_empty(&n->slabs_full) ||
1309 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001310 ret = -EBUSY;
1311 break;
1312 }
1313 }
1314 return ret;
1315}
1316
1317static int __meminit slab_memory_callback(struct notifier_block *self,
1318 unsigned long action, void *arg)
1319{
1320 struct memory_notify *mnb = arg;
1321 int ret = 0;
1322 int nid;
1323
1324 nid = mnb->status_change_nid;
1325 if (nid < 0)
1326 goto out;
1327
1328 switch (action) {
1329 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001330 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001331 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001332 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001333 break;
1334 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001335 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001336 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001337 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001338 break;
1339 case MEM_ONLINE:
1340 case MEM_OFFLINE:
1341 case MEM_CANCEL_ONLINE:
1342 case MEM_CANCEL_OFFLINE:
1343 break;
1344 }
1345out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001346 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001347}
1348#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1349
Christoph Lametere498be72005-09-09 13:03:32 -07001350/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001351 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001352 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001353static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001354 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001355{
Christoph Lameter6744f082013-01-10 19:12:17 +00001356 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001357
Christoph Lameter6744f082013-01-10 19:12:17 +00001358 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001359 BUG_ON(!ptr);
1360
Christoph Lameter6744f082013-01-10 19:12:17 +00001361 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001362 /*
1363 * Do not assume that spinlocks can be initialized via memcpy:
1364 */
1365 spin_lock_init(&ptr->list_lock);
1366
Christoph Lametere498be72005-09-09 13:03:32 -07001367 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001368 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001369}
1370
Andrew Mortona737b3e2006-03-22 00:08:11 -08001371/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001372 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1373 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001374 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001375static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001376{
1377 int node;
1378
1379 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001380 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001381 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001382 REAPTIMEOUT_NODE +
1383 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001384 }
1385}
1386
1387/*
Christoph Lameter3c583462012-11-28 16:23:01 +00001388 * The memory after the last cpu cache pointer is used for the
Christoph Lameter6a673682013-01-10 19:14:19 +00001389 * the node pointer.
Christoph Lameter3c583462012-11-28 16:23:01 +00001390 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001391static void setup_node_pointer(struct kmem_cache *cachep)
Christoph Lameter3c583462012-11-28 16:23:01 +00001392{
Christoph Lameter6a673682013-01-10 19:14:19 +00001393 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
Christoph Lameter3c583462012-11-28 16:23:01 +00001394}
1395
1396/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001397 * Initialisation. Called after the page allocator have been initialised and
1398 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 */
1400void __init kmem_cache_init(void)
1401{
Christoph Lametere498be72005-09-09 13:03:32 -07001402 int i;
1403
Joonsoo Kim68126702013-10-24 10:07:42 +09001404 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1405 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001406 kmem_cache = &kmem_cache_boot;
Christoph Lameter6a673682013-01-10 19:14:19 +00001407 setup_node_pointer(kmem_cache);
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001408
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001409 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001410 use_alien_caches = 0;
1411
Christoph Lameter3c583462012-11-28 16:23:01 +00001412 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001413 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001414
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001415 set_up_node(kmem_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
1417 /*
1418 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001419 * page orders on machines with more than 32MB of memory if
1420 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001422 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001423 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 /* Bootstrap is tricky, because several objects are allocated
1426 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001427 * 1) initialize the kmem_cache cache: it contains the struct
1428 * kmem_cache structures of all caches, except kmem_cache itself:
1429 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001430 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001431 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001432 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001434 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001435 * An __init data area is used for the head array.
1436 * 3) Create the remaining kmalloc caches, with minimally sized
1437 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001438 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001440 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001441 * the other cache's with kmalloc allocated memory.
1442 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 */
1444
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001445 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
Eric Dumazet8da34302007-05-06 14:49:29 -07001447 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001448 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001449 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001450 create_boot_cache(kmem_cache, "kmem_cache",
1451 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001452 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001453 SLAB_HWCACHE_ALIGN);
1454 list_add(&kmem_cache->list, &slab_caches);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455
1456 /* 2+3) create the kmalloc caches */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457
Andrew Mortona737b3e2006-03-22 00:08:11 -08001458 /*
1459 * Initialize the caches that provide memory for the array cache and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001460 * kmem_cache_node structures first. Without this, further allocations will
Andrew Mortona737b3e2006-03-22 00:08:11 -08001461 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001462 */
1463
Christoph Lametere3366012013-01-10 19:14:18 +00001464 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1465 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001466
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001467 if (INDEX_AC != INDEX_NODE)
1468 kmalloc_caches[INDEX_NODE] =
1469 create_kmalloc_cache("kmalloc-node",
1470 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Christoph Lametere498be72005-09-09 13:03:32 -07001471
Ingo Molnare0a42722006-06-23 02:03:46 -07001472 slab_early_init = 0;
1473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 /* 4) Replace the bootstrap head arrays */
1475 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001476 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001477
Pekka Enberg83b519e2009-06-10 19:40:04 +03001478 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001479
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001480 memcpy(ptr, cpu_cache_get(kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001481 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001482
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001483 kmem_cache->array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001484
Pekka Enberg83b519e2009-06-10 19:40:04 +03001485 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001486
Christoph Lametere3366012013-01-10 19:14:18 +00001487 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001488 != &initarray_generic.cache);
Christoph Lametere3366012013-01-10 19:14:18 +00001489 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001490 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001491
Christoph Lametere3366012013-01-10 19:14:18 +00001492 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001494 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001495 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001496 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497
Mel Gorman9c09a952008-01-24 05:49:54 -08001498 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001499 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001500
Christoph Lametere3366012013-01-10 19:14:18 +00001501 init_list(kmalloc_caches[INDEX_AC],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001502 &init_kmem_cache_node[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001503
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001504 if (INDEX_AC != INDEX_NODE) {
1505 init_list(kmalloc_caches[INDEX_NODE],
1506 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001507 }
1508 }
1509 }
1510
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001511 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001512}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001513
Pekka Enberg8429db52009-06-12 15:58:59 +03001514void __init kmem_cache_init_late(void)
1515{
1516 struct kmem_cache *cachep;
1517
Christoph Lameter97d06602012-07-06 15:25:11 -05001518 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001519
Pekka Enberg8429db52009-06-12 15:58:59 +03001520 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001521 mutex_lock(&slab_mutex);
1522 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001523 if (enable_cpucache(cachep, GFP_NOWAIT))
1524 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001525 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001526
Christoph Lameter97d06602012-07-06 15:25:11 -05001527 /* Done! */
1528 slab_state = FULL;
1529
Andrew Mortona737b3e2006-03-22 00:08:11 -08001530 /*
1531 * Register a cpu startup notifier callback that initializes
1532 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 */
1534 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001535
David Rientjes8f9f8d92010-03-27 19:40:47 -07001536#ifdef CONFIG_NUMA
1537 /*
1538 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001539 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001540 */
1541 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1542#endif
1543
Andrew Mortona737b3e2006-03-22 00:08:11 -08001544 /*
1545 * The reap timers are started later, with a module init call: That part
1546 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 */
1548}
1549
1550static int __init cpucache_init(void)
1551{
1552 int cpu;
1553
Andrew Mortona737b3e2006-03-22 00:08:11 -08001554 /*
1555 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 */
Christoph Lametere498be72005-09-09 13:03:32 -07001557 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001558 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001559
1560 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001561 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 return 0;
1563}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564__initcall(cpucache_init);
1565
Rafael Aquini8bdec192012-03-09 17:27:27 -03001566static noinline void
1567slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1568{
David Rientjes9a02d692014-06-04 16:06:36 -07001569#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001570 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001571 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001572 unsigned long flags;
1573 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001574 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1575 DEFAULT_RATELIMIT_BURST);
1576
1577 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1578 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001579
1580 printk(KERN_WARNING
1581 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1582 nodeid, gfpflags);
1583 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001584 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001585
Christoph Lameter18bf8542014-08-06 16:04:11 -07001586 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001587 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1588 unsigned long active_slabs = 0, num_slabs = 0;
1589
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001590 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001591 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001592 active_objs += cachep->num;
1593 active_slabs++;
1594 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001595 list_for_each_entry(page, &n->slabs_partial, lru) {
1596 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001597 active_slabs++;
1598 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001599 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001600 num_slabs++;
1601
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001602 free_objects += n->free_objects;
1603 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001604
1605 num_slabs += active_slabs;
1606 num_objs = num_slabs * cachep->num;
1607 printk(KERN_WARNING
1608 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1609 node, active_slabs, num_slabs, active_objs, num_objs,
1610 free_objects);
1611 }
David Rientjes9a02d692014-06-04 16:06:36 -07001612#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001613}
1614
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001616 * Interface to system's page allocator. No need to hold the
1617 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 *
1619 * If we requested dmaable memory, we will get it. Even if we
1620 * did not request dmaable memory, we might get it, but that
1621 * would be relatively rare and ignorable.
1622 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001623static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1624 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625{
1626 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001627 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001628
Glauber Costaa618e892012-06-14 16:17:21 +04001629 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001630 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1631 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001632
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001633 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1634 return NULL;
1635
Linus Torvalds517d0862009-06-16 19:50:13 -07001636 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001637 if (!page) {
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001638 memcg_uncharge_slab(cachep, cachep->gfporder);
David Rientjes9a02d692014-06-04 16:06:36 -07001639 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001641 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001643 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Mel Gorman072bb0a2012-07-31 16:43:58 -07001644 if (unlikely(page->pfmemalloc))
1645 pfmemalloc_active = true;
1646
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001647 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001649 add_zone_page_state(page_zone(page),
1650 NR_SLAB_RECLAIMABLE, nr_pages);
1651 else
1652 add_zone_page_state(page_zone(page),
1653 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001654 __SetPageSlab(page);
1655 if (page->pfmemalloc)
1656 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001657
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001658 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1659 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1660
1661 if (cachep->ctor)
1662 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1663 else
1664 kmemcheck_mark_unallocated_pages(page, nr_pages);
1665 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001666
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001667 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668}
1669
1670/*
1671 * Interface to system's page release.
1672 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001673static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001675 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001677 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001678
Christoph Lameter972d1a72006-09-25 23:31:51 -07001679 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1680 sub_zone_page_state(page_zone(page),
1681 NR_SLAB_RECLAIMABLE, nr_freed);
1682 else
1683 sub_zone_page_state(page_zone(page),
1684 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001685
Joonsoo Kima57a4982013-10-24 10:07:44 +09001686 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001687 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001688 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001689 page_mapcount_reset(page);
1690 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001691
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 if (current->reclaim_state)
1693 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001694 __free_pages(page, cachep->gfporder);
1695 memcg_uncharge_slab(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696}
1697
1698static void kmem_rcu_free(struct rcu_head *head)
1699{
Joonsoo Kim68126702013-10-24 10:07:42 +09001700 struct kmem_cache *cachep;
1701 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Joonsoo Kim68126702013-10-24 10:07:42 +09001703 page = container_of(head, struct page, rcu_head);
1704 cachep = page->slab_cache;
1705
1706 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707}
1708
1709#if DEBUG
1710
1711#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001712static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001713 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001715 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001717 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001719 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 return;
1721
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001722 *addr++ = 0x12345678;
1723 *addr++ = caller;
1724 *addr++ = smp_processor_id();
1725 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 {
1727 unsigned long *sptr = &caller;
1728 unsigned long svalue;
1729
1730 while (!kstack_end(sptr)) {
1731 svalue = *sptr++;
1732 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001733 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 size -= sizeof(unsigned long);
1735 if (size <= sizeof(unsigned long))
1736 break;
1737 }
1738 }
1739
1740 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001741 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742}
1743#endif
1744
Pekka Enberg343e0d72006-02-01 03:05:50 -08001745static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001747 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001748 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749
1750 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001751 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752}
1753
1754static void dump_line(char *data, int offset, int limit)
1755{
1756 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001757 unsigned char error = 0;
1758 int bad_count = 0;
1759
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001760 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001761 for (i = 0; i < limit; i++) {
1762 if (data[offset + i] != POISON_FREE) {
1763 error = data[offset + i];
1764 bad_count++;
1765 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001766 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001767 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1768 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001769
1770 if (bad_count == 1) {
1771 error ^= POISON_FREE;
1772 if (!(error & (error - 1))) {
1773 printk(KERN_ERR "Single bit error detected. Probably "
1774 "bad RAM.\n");
1775#ifdef CONFIG_X86
1776 printk(KERN_ERR "Run memtest86+ or a similar memory "
1777 "test tool.\n");
1778#else
1779 printk(KERN_ERR "Run a memory test tool.\n");
1780#endif
1781 }
1782 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783}
1784#endif
1785
1786#if DEBUG
1787
Pekka Enberg343e0d72006-02-01 03:05:50 -08001788static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789{
1790 int i, size;
1791 char *realobj;
1792
1793 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001794 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001795 *dbg_redzone1(cachep, objp),
1796 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 }
1798
1799 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001800 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1801 *dbg_userword(cachep, objp),
1802 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001804 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001805 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001806 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 int limit;
1808 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001809 if (i + limit > size)
1810 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 dump_line(realobj, i, limit);
1812 }
1813}
1814
Pekka Enberg343e0d72006-02-01 03:05:50 -08001815static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816{
1817 char *realobj;
1818 int size, i;
1819 int lines = 0;
1820
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001821 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001822 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001824 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001826 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 exp = POISON_END;
1828 if (realobj[i] != exp) {
1829 int limit;
1830 /* Mismatch ! */
1831 /* Print header */
1832 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001833 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001834 "Slab corruption (%s): %s start=%p, len=%d\n",
1835 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 print_objinfo(cachep, objp, 0);
1837 }
1838 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001839 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001841 if (i + limit > size)
1842 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 dump_line(realobj, i, limit);
1844 i += 16;
1845 lines++;
1846 /* Limit to 5 lines */
1847 if (lines > 5)
1848 break;
1849 }
1850 }
1851 if (lines != 0) {
1852 /* Print some data about the neighboring objects, if they
1853 * exist:
1854 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001855 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001856 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857
Joonsoo Kim8456a642013-10-24 10:07:49 +09001858 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001860 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001861 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 print_objinfo(cachep, objp, 2);
1865 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001866 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001867 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001868 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001870 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 print_objinfo(cachep, objp, 2);
1872 }
1873 }
1874}
1875#endif
1876
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001878static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1879 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001880{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 int i;
1882 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001883 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884
1885 if (cachep->flags & SLAB_POISON) {
1886#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001887 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001888 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001889 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001890 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 else
1892 check_poison_obj(cachep, objp);
1893#else
1894 check_poison_obj(cachep, objp);
1895#endif
1896 }
1897 if (cachep->flags & SLAB_RED_ZONE) {
1898 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1899 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001900 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1902 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001903 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001906}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001908static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1909 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001910{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001911}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912#endif
1913
Randy Dunlap911851e2006-03-22 00:08:14 -08001914/**
1915 * slab_destroy - destroy and release all objects in a slab
1916 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001917 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001918 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001919 * Destroy all the objs in a slab page, and release the mem back to the system.
1920 * Before calling the slab page must have been unlinked from the cache. The
1921 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001922 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001923static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001924{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001925 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001926
Joonsoo Kim8456a642013-10-24 10:07:49 +09001927 freelist = page->freelist;
1928 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001930 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931
Joonsoo Kim68126702013-10-24 10:07:42 +09001932 /*
1933 * RCU free overloads the RCU head over the LRU.
1934 * slab_page has been overloeaded over the LRU,
1935 * however it is not used from now on so that
1936 * we can use it safely.
1937 */
1938 head = (void *)&page->rcu_head;
1939 call_rcu(head, kmem_rcu_free);
1940
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001942 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 }
Joonsoo Kim68126702013-10-24 10:07:42 +09001944
1945 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001946 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001947 * although actual page can be freed in rcu context
1948 */
1949 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001950 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951}
1952
Joonsoo Kim97654df2014-08-06 16:04:25 -07001953static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1954{
1955 struct page *page, *n;
1956
1957 list_for_each_entry_safe(page, n, list, lru) {
1958 list_del(&page->lru);
1959 slab_destroy(cachep, page);
1960 }
1961}
1962
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001964 * calculate_slab_order - calculate size (page order) of slabs
1965 * @cachep: pointer to the cache that is being created
1966 * @size: size of objects to be created in this cache.
1967 * @align: required alignment for the objects.
1968 * @flags: slab allocation flags
1969 *
1970 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001971 *
1972 * This could be made much more intelligent. For now, try to avoid using
1973 * high order pages for slabs. When the gfp() functions are more friendly
1974 * towards high-order requests, this should be changed.
1975 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001976static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001977 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001978{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001979 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001980 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001981 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001982
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001983 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001984 unsigned int num;
1985 size_t remainder;
1986
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001987 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001988 if (!num)
1989 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001990
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001991 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1992 if (num > SLAB_OBJ_MAX_NUM)
1993 break;
1994
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001995 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim03787302014-06-23 13:22:06 -07001996 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001997 /*
1998 * Max number of objs-per-slab for caches which
1999 * use off-slab slabs. Needed to avoid a possible
2000 * looping condition in cache_grow().
2001 */
Joonsoo Kim03787302014-06-23 13:22:06 -07002002 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
2003 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002004 offslab_limit = size;
Joonsoo Kim03787302014-06-23 13:22:06 -07002005 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002006
2007 if (num > offslab_limit)
2008 break;
2009 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002010
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002011 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002013 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002014 left_over = remainder;
2015
2016 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002017 * A VFS-reclaimable slab tends to have most allocations
2018 * as GFP_NOFS and we really don't want to have to be allocating
2019 * higher-order pages when we are unable to shrink dcache.
2020 */
2021 if (flags & SLAB_RECLAIM_ACCOUNT)
2022 break;
2023
2024 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002025 * Large number of objects is good, but very large slabs are
2026 * currently bad for the gfp()s.
2027 */
David Rientjes543585c2011-10-18 22:09:24 -07002028 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002029 break;
2030
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002031 /*
2032 * Acceptable internal fragmentation?
2033 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002034 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002035 break;
2036 }
2037 return left_over;
2038}
2039
Pekka Enberg83b519e2009-06-10 19:40:04 +03002040static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002041{
Christoph Lameter97d06602012-07-06 15:25:11 -05002042 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002043 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002044
Christoph Lameter97d06602012-07-06 15:25:11 -05002045 if (slab_state == DOWN) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002046 /*
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002047 * Note: Creation of first cache (kmem_cache).
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002048 * The setup_node is taken care
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002049 * of by the caller of __kmem_cache_create
2050 */
2051 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2052 slab_state = PARTIAL;
2053 } else if (slab_state == PARTIAL) {
2054 /*
2055 * Note: the second kmem_cache_create must create the cache
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002056 * that's used by kmalloc(24), otherwise the creation of
2057 * further caches will BUG().
2058 */
2059 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2060
2061 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002062 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2063 * the second cache, then we need to set up all its node/,
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002064 * otherwise the creation of further caches will BUG().
2065 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002066 set_up_node(cachep, SIZE_AC);
2067 if (INDEX_AC == INDEX_NODE)
2068 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002069 else
Christoph Lameter97d06602012-07-06 15:25:11 -05002070 slab_state = PARTIAL_ARRAYCACHE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002071 } else {
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002072 /* Remaining boot caches */
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002073 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002074 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002075
Christoph Lameter97d06602012-07-06 15:25:11 -05002076 if (slab_state == PARTIAL_ARRAYCACHE) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002077 set_up_node(cachep, SIZE_NODE);
2078 slab_state = PARTIAL_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002079 } else {
2080 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002081 for_each_online_node(node) {
Christoph Lameter6a673682013-01-10 19:14:19 +00002082 cachep->node[node] =
Christoph Lameter6744f082013-01-10 19:12:17 +00002083 kmalloc_node(sizeof(struct kmem_cache_node),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002084 gfp, node);
Christoph Lameter6a673682013-01-10 19:14:19 +00002085 BUG_ON(!cachep->node[node]);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002086 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002087 }
2088 }
2089 }
Christoph Lameter6a673682013-01-10 19:14:19 +00002090 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002091 jiffies + REAPTIMEOUT_NODE +
2092 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002093
2094 cpu_cache_get(cachep)->avail = 0;
2095 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2096 cpu_cache_get(cachep)->batchcount = 1;
2097 cpu_cache_get(cachep)->touched = 0;
2098 cachep->batchcount = 1;
2099 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002100 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002101}
2102
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002103/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002104 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002105 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 *
2108 * Returns a ptr to the cache on success, NULL on failure.
2109 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002110 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 * The flags are
2113 *
2114 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2115 * to catch references to uninitialised memory.
2116 *
2117 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2118 * for buffer overruns.
2119 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2121 * cacheline. This can be beneficial if you're counting cycles as closely
2122 * as davem.
2123 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002124int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002125__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002127 size_t left_over, freelist_size, ralign;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002128 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002129 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002130 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133#if FORCED_DEBUG
2134 /*
2135 * Enable redzoning and last user accounting, except for caches with
2136 * large objects, if the increased size would increase the object size
2137 * above the next power of two: caches with object sizes just above a
2138 * power of two have a significant amount of internal fragmentation.
2139 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002140 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2141 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002142 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 if (!(flags & SLAB_DESTROY_BY_RCU))
2144 flags |= SLAB_POISON;
2145#endif
2146 if (flags & SLAB_DESTROY_BY_RCU)
2147 BUG_ON(flags & SLAB_POISON);
2148#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149
Andrew Mortona737b3e2006-03-22 00:08:11 -08002150 /*
2151 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 * unaligned accesses for some archs when redzoning is used, and makes
2153 * sure any on-slab bufctl's are also correctly aligned.
2154 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002155 if (size & (BYTES_PER_WORD - 1)) {
2156 size += (BYTES_PER_WORD - 1);
2157 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 }
2159
Pekka Enbergca5f9702006-09-25 23:31:25 -07002160 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002161 * Redzoning and user store require word alignment or possibly larger.
2162 * Note this will be overridden by architecture or caller mandated
2163 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002164 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002165 if (flags & SLAB_STORE_USER)
2166 ralign = BYTES_PER_WORD;
2167
2168 if (flags & SLAB_RED_ZONE) {
2169 ralign = REDZONE_ALIGN;
2170 /* If redzoning, ensure that the second redzone is suitably
2171 * aligned, by adjusting the object size accordingly. */
2172 size += REDZONE_ALIGN - 1;
2173 size &= ~(REDZONE_ALIGN - 1);
2174 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002175
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002176 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002177 if (ralign < cachep->align) {
2178 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002180 /* disable debug if necessary */
2181 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002182 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002183 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002184 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002186 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187
Pekka Enberg83b519e2009-06-10 19:40:04 +03002188 if (slab_is_available())
2189 gfp = GFP_KERNEL;
2190 else
2191 gfp = GFP_NOWAIT;
2192
Christoph Lameter6a673682013-01-10 19:14:19 +00002193 setup_node_pointer(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195
Pekka Enbergca5f9702006-09-25 23:31:25 -07002196 /*
2197 * Both debugging options require word-alignment which is calculated
2198 * into align above.
2199 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002202 cachep->obj_offset += sizeof(unsigned long long);
2203 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 }
2205 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002206 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002207 * the real object. But if the second red zone needs to be
2208 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002210 if (flags & SLAB_RED_ZONE)
2211 size += REDZONE_ALIGN;
2212 else
2213 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214 }
2215#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002216 if (size >= kmalloc_size(INDEX_NODE + 1)
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002217 && cachep->object_size > cache_line_size()
2218 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2219 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 size = PAGE_SIZE;
2221 }
2222#endif
2223#endif
2224
Ingo Molnare0a42722006-06-23 02:03:46 -07002225 /*
2226 * Determine if the slab management is 'on' or 'off' slab.
2227 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002228 * it too early on. Always use on-slab management when
2229 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002230 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002231 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002232 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 /*
2234 * Size is large, assume best to place the slab management obj
2235 * off-slab (should allow better packing of objs).
2236 */
2237 flags |= CFLGS_OFF_SLAB;
2238
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002239 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002240 /*
2241 * We should restrict the number of objects in a slab to implement
2242 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2243 */
2244 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2245 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002247 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002249 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002250 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002251
Joonsoo Kim03787302014-06-23 13:22:06 -07002252 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253
2254 /*
2255 * If the slab has been placed off-slab, and we have enough space then
2256 * move it on-slab. This is at the expense of any extra colouring.
2257 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002258 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002260 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 }
2262
2263 if (flags & CFLGS_OFF_SLAB) {
2264 /* really off slab. No need for manual alignment */
Joonsoo Kim03787302014-06-23 13:22:06 -07002265 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302266
2267#ifdef CONFIG_PAGE_POISONING
2268 /* If we're going to use the generic kernel_map_pages()
2269 * poisoning, then it's going to smash the contents of
2270 * the redzone and userword anyhow, so switch them off.
2271 */
2272 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2273 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2274#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 }
2276
2277 cachep->colour_off = cache_line_size();
2278 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002279 if (cachep->colour_off < cachep->align)
2280 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002281 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002282 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002284 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002285 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002286 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002287 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002288 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002290 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002291 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002292 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002293 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002294 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002295 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2296 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002297 * But leave a BUG_ON for some lucky dude.
2298 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002299 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002300 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002302 err = setup_cpu_cache(cachep, gfp);
2303 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002304 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002305 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002306 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002308 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310
2311#if DEBUG
2312static void check_irq_off(void)
2313{
2314 BUG_ON(!irqs_disabled());
2315}
2316
2317static void check_irq_on(void)
2318{
2319 BUG_ON(irqs_disabled());
2320}
2321
Pekka Enberg343e0d72006-02-01 03:05:50 -08002322static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323{
2324#ifdef CONFIG_SMP
2325 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002326 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327#endif
2328}
Christoph Lametere498be72005-09-09 13:03:32 -07002329
Pekka Enberg343e0d72006-02-01 03:05:50 -08002330static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002331{
2332#ifdef CONFIG_SMP
2333 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002334 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002335#endif
2336}
2337
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338#else
2339#define check_irq_off() do { } while(0)
2340#define check_irq_on() do { } while(0)
2341#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002342#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343#endif
2344
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002345static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002346 struct array_cache *ac,
2347 int force, int node);
2348
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349static void do_drain(void *arg)
2350{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002351 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002353 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002354 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002355 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356
2357 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002358 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002359 n = get_node(cachep, node);
2360 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002361 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002362 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002363 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 ac->avail = 0;
2365}
2366
Pekka Enberg343e0d72006-02-01 03:05:50 -08002367static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002369 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002370 int node;
2371
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002372 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002374 for_each_kmem_cache_node(cachep, node, n)
2375 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002376 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002377
Christoph Lameter18bf8542014-08-06 16:04:11 -07002378 for_each_kmem_cache_node(cachep, node, n)
2379 drain_array(cachep, n, n->shared, 1, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380}
2381
Christoph Lametered11d9e2006-06-30 01:55:45 -07002382/*
2383 * Remove slabs from the list of free slabs.
2384 * Specify the number of slabs to drain in tofree.
2385 *
2386 * Returns the actual number of slabs released.
2387 */
2388static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002389 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002391 struct list_head *p;
2392 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002393 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394
Christoph Lametered11d9e2006-06-30 01:55:45 -07002395 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002396 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002398 spin_lock_irq(&n->list_lock);
2399 p = n->slabs_free.prev;
2400 if (p == &n->slabs_free) {
2401 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002402 goto out;
2403 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404
Joonsoo Kim8456a642013-10-24 10:07:49 +09002405 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002407 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002409 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002410 /*
2411 * Safe to drop the lock. The slab is no longer linked
2412 * to the cache.
2413 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002414 n->free_objects -= cache->num;
2415 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002416 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002417 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002419out:
2420 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421}
2422
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002423int __kmem_cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002424{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002425 int ret = 0;
2426 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002427 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002428
2429 drain_cpu_caches(cachep);
2430
2431 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002432 for_each_kmem_cache_node(cachep, node, n) {
Wanpeng Li0fa81032013-07-04 08:33:22 +08002433 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002434
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002435 ret += !list_empty(&n->slabs_full) ||
2436 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002437 }
2438 return (ret ? 1 : 0);
2439}
2440
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002441int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442{
Christoph Lameter12c36672012-09-04 23:38:33 +00002443 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002444 struct kmem_cache_node *n;
Vladimir Davydov03afc0e2014-06-04 16:07:20 -07002445 int rc = __kmem_cache_shrink(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446
Christoph Lameter12c36672012-09-04 23:38:33 +00002447 if (rc)
2448 return rc;
2449
2450 for_each_online_cpu(i)
2451 kfree(cachep->array[i]);
2452
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002453 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002454 for_each_kmem_cache_node(cachep, i, n) {
2455 kfree(n->shared);
2456 free_alien_cache(n->alien);
2457 kfree(n);
2458 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002460 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002463/*
2464 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002465 *
2466 * For a slab cache when the slab descriptor is off-slab, the
2467 * slab descriptor can't come from the same cache which is being created,
2468 * Because if it is the case, that means we defer the creation of
2469 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2470 * And we eventually call down to __kmem_cache_create(), which
2471 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2472 * This is a "chicken-and-egg" problem.
2473 *
2474 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2475 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002476 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002477static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002478 struct page *page, int colour_off,
2479 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002481 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002482 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002483
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 if (OFF_SLAB(cachep)) {
2485 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002486 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002487 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002488 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 return NULL;
2490 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002491 freelist = addr + colour_off;
2492 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002494 page->active = 0;
2495 page->s_mem = addr + colour_off;
2496 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497}
2498
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002499static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002501 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002502}
2503
2504static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002505 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002506{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002507 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508}
2509
Pekka Enberg343e0d72006-02-01 03:05:50 -08002510static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002511 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512{
2513 int i;
2514
2515 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002516 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517#if DEBUG
2518 /* need to poison the objs? */
2519 if (cachep->flags & SLAB_POISON)
2520 poison_obj(cachep, objp, POISON_FREE);
2521 if (cachep->flags & SLAB_STORE_USER)
2522 *dbg_userword(cachep, objp) = NULL;
2523
2524 if (cachep->flags & SLAB_RED_ZONE) {
2525 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2526 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2527 }
2528 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002529 * Constructors are not allowed to allocate memory from the same
2530 * cache which they are a constructor for. Otherwise, deadlock.
2531 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532 */
2533 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002534 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535
2536 if (cachep->flags & SLAB_RED_ZONE) {
2537 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2538 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002539 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2541 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002542 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002544 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002545 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002546 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002547 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548#else
2549 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002550 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551#endif
Joonsoo Kim03787302014-06-23 13:22:06 -07002552 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002553 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555}
2556
Pekka Enberg343e0d72006-02-01 03:05:50 -08002557static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002559 if (CONFIG_ZONE_DMA_FLAG) {
2560 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002561 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002562 else
Glauber Costaa618e892012-06-14 16:17:21 +04002563 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002564 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565}
2566
Joonsoo Kim8456a642013-10-24 10:07:49 +09002567static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002568 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002569{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002570 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002571
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002572 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002573 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002574#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002575 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002576#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002577
2578 return objp;
2579}
2580
Joonsoo Kim8456a642013-10-24 10:07:49 +09002581static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002582 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002583{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002584 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002585#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002586 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002587
Matthew Dobson78d382d2006-02-01 03:05:47 -08002588 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002589 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002590
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002591 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002592 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002593 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002594 printk(KERN_ERR "slab: double free detected in cache "
2595 "'%s', objp %p\n", cachep->name, objp);
2596 BUG();
2597 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002598 }
2599#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002600 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002601 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002602}
2603
Pekka Enberg47768742006-06-23 02:03:07 -07002604/*
2605 * Map pages beginning at addr to the given cache and slab. This is required
2606 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002607 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002608 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002609static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002610 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002612 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002613 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614}
2615
2616/*
2617 * Grow (by 1) the number of slabs within a cache. This is called by
2618 * kmem_cache_alloc() when there are no active objs left in a cache.
2619 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002620static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002621 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002623 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002624 size_t offset;
2625 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002626 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627
Andrew Mortona737b3e2006-03-22 00:08:11 -08002628 /*
2629 * Be lazy and only check for valid flags here, keeping it out of the
2630 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002632 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2633 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002635 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002637 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002638 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639
2640 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002641 offset = n->colour_next;
2642 n->colour_next++;
2643 if (n->colour_next >= cachep->colour)
2644 n->colour_next = 0;
2645 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002647 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648
2649 if (local_flags & __GFP_WAIT)
2650 local_irq_enable();
2651
2652 /*
2653 * The test for missing atomic flag is performed here, rather than
2654 * the more obvious place, simply to reduce the critical path length
2655 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2656 * will eventually be caught here (where it matters).
2657 */
2658 kmem_flagcheck(cachep, flags);
2659
Andrew Mortona737b3e2006-03-22 00:08:11 -08002660 /*
2661 * Get mem for the objs. Attempt to allocate a physical page from
2662 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002663 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002664 if (!page)
2665 page = kmem_getpages(cachep, local_flags, nodeid);
2666 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667 goto failed;
2668
2669 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002670 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002671 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002672 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 goto opps1;
2674
Joonsoo Kim8456a642013-10-24 10:07:49 +09002675 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676
Joonsoo Kim8456a642013-10-24 10:07:49 +09002677 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678
2679 if (local_flags & __GFP_WAIT)
2680 local_irq_disable();
2681 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002682 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683
2684 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002685 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002687 n->free_objects += cachep->num;
2688 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002690opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002691 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002692failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 if (local_flags & __GFP_WAIT)
2694 local_irq_disable();
2695 return 0;
2696}
2697
2698#if DEBUG
2699
2700/*
2701 * Perform extra freeing checks:
2702 * - detect bad pointers.
2703 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 */
2705static void kfree_debugcheck(const void *objp)
2706{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 if (!virt_addr_valid(objp)) {
2708 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002709 (unsigned long)objp);
2710 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712}
2713
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002714static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2715{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002716 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002717
2718 redzone1 = *dbg_redzone1(cache, obj);
2719 redzone2 = *dbg_redzone2(cache, obj);
2720
2721 /*
2722 * Redzone is ok.
2723 */
2724 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2725 return;
2726
2727 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2728 slab_error(cache, "double free detected");
2729 else
2730 slab_error(cache, "memory outside object was overwritten");
2731
David Woodhouseb46b8f12007-05-08 00:22:59 -07002732 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002733 obj, redzone1, redzone2);
2734}
2735
Pekka Enberg343e0d72006-02-01 03:05:50 -08002736static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002737 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002740 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002742 BUG_ON(virt_to_cache(objp) != cachep);
2743
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002744 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002746 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002749 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2751 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2752 }
2753 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002754 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755
Joonsoo Kim8456a642013-10-24 10:07:49 +09002756 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757
2758 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002759 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760
Joonsoo Kim03787302014-06-23 13:22:06 -07002761 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 if (cachep->flags & SLAB_POISON) {
2763#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002764 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002765 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002766 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002767 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 } else {
2769 poison_obj(cachep, objp, POISON_FREE);
2770 }
2771#else
2772 poison_obj(cachep, objp, POISON_FREE);
2773#endif
2774 }
2775 return objp;
2776}
2777
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778#else
2779#define kfree_debugcheck(x) do { } while(0)
2780#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781#endif
2782
Mel Gorman072bb0a2012-07-31 16:43:58 -07002783static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2784 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785{
2786 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002787 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002789 int node;
2790
Joe Korty6d2144d2008-03-05 15:04:59 -08002791 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002792 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002793 if (unlikely(force_refill))
2794 goto force_grow;
2795retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002796 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 batchcount = ac->batchcount;
2798 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002799 /*
2800 * If there was little recent activity on this cache, then
2801 * perform only a partial refill. Otherwise we could generate
2802 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 */
2804 batchcount = BATCHREFILL_LIMIT;
2805 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07002806 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002808 BUG_ON(ac->avail > 0 || !n);
2809 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002810
Christoph Lameter3ded1752006-03-25 03:06:44 -08002811 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002812 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2813 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002814 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002815 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002816
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 while (batchcount > 0) {
2818 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002819 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002821 entry = n->slabs_partial.next;
2822 if (entry == &n->slabs_partial) {
2823 n->free_touched = 1;
2824 entry = n->slabs_free.next;
2825 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 goto must_grow;
2827 }
2828
Joonsoo Kim8456a642013-10-24 10:07:49 +09002829 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002831
2832 /*
2833 * The slab was either on partial or free list so
2834 * there must be at least one object available for
2835 * allocation.
2836 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002837 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002838
Joonsoo Kim8456a642013-10-24 10:07:49 +09002839 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 STATS_INC_ALLOCED(cachep);
2841 STATS_INC_ACTIVE(cachep);
2842 STATS_SET_HIGH(cachep);
2843
Joonsoo Kim8456a642013-10-24 10:07:49 +09002844 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002845 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847
2848 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002849 list_del(&page->lru);
2850 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002851 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002853 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 }
2855
Andrew Mortona737b3e2006-03-22 00:08:11 -08002856must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002857 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002858alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002859 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860
2861 if (unlikely(!ac->avail)) {
2862 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002863force_grow:
Christoph Lameter3c517a62006-12-06 20:33:29 -08002864 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002865
Andrew Mortona737b3e2006-03-22 00:08:11 -08002866 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002867 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002868 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002869
2870 /* no objects in sight? abort */
2871 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 return NULL;
2873
Andrew Mortona737b3e2006-03-22 00:08:11 -08002874 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 goto retry;
2876 }
2877 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002878
2879 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880}
2881
Andrew Mortona737b3e2006-03-22 00:08:11 -08002882static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2883 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884{
2885 might_sleep_if(flags & __GFP_WAIT);
2886#if DEBUG
2887 kmem_flagcheck(cachep, flags);
2888#endif
2889}
2890
2891#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002892static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002893 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894{
Joonsoo Kim03787302014-06-23 13:22:06 -07002895 struct page *page;
2896
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002897 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002899 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002901 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002902 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002903 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 else
2905 check_poison_obj(cachep, objp);
2906#else
2907 check_poison_obj(cachep, objp);
2908#endif
2909 poison_obj(cachep, objp, POISON_INUSE);
2910 }
2911 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002912 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913
2914 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002915 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2916 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2917 slab_error(cachep, "double free, or memory outside"
2918 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002919 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002920 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002921 objp, *dbg_redzone1(cachep, objp),
2922 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 }
2924 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2925 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2926 }
Joonsoo Kim03787302014-06-23 13:22:06 -07002927
2928 page = virt_to_head_page(objp);
2929 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002930 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002931 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002932 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002933 if (ARCH_SLAB_MINALIGN &&
2934 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002935 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002936 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002937 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 return objp;
2939}
2940#else
2941#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2942#endif
2943
Akinobu Mita773ff602008-12-23 19:37:01 +09002944static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002945{
Joonsoo Kim8a9c61d2014-08-06 16:04:20 -07002946 if (unlikely(cachep == kmem_cache))
Akinobu Mita773ff602008-12-23 19:37:01 +09002947 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002948
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002949 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002950}
2951
Pekka Enberg343e0d72006-02-01 03:05:50 -08002952static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002954 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002956 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957
Alok N Kataria5c382302005-09-27 21:45:46 -07002958 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002959
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002960 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002963 objp = ac_get_obj(cachep, ac, flags, false);
2964
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002965 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07002966 * Allow for the possibility all avail objects are not allowed
2967 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002968 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07002969 if (objp) {
2970 STATS_INC_ALLOCHIT(cachep);
2971 goto out;
2972 }
2973 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07002975
2976 STATS_INC_ALLOCMISS(cachep);
2977 objp = cache_alloc_refill(cachep, flags, force_refill);
2978 /*
2979 * the 'ac' may be updated by cache_alloc_refill(),
2980 * and kmemleak_erase() requires its correct value.
2981 */
2982 ac = cpu_cache_get(cachep);
2983
2984out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01002985 /*
2986 * To avoid a false negative, if an object that is in one of the
2987 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2988 * treat the array pointers as a reference to the object.
2989 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09002990 if (objp)
2991 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07002992 return objp;
2993}
2994
Christoph Lametere498be72005-09-09 13:03:32 -07002995#ifdef CONFIG_NUMA
2996/*
David Rientjesf0432d12014-04-07 15:37:30 -07002997 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002998 *
2999 * If we are in_interrupt, then process context, including cpusets and
3000 * mempolicy, may not apply and should not be used for allocation policy.
3001 */
3002static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3003{
3004 int nid_alloc, nid_here;
3005
Christoph Lameter765c4502006-09-27 01:50:08 -07003006 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003007 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003008 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003009 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003010 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003011 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003012 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003013 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003014 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003015 return NULL;
3016}
3017
3018/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003019 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003020 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003021 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003022 * perform an allocation without specifying a node. This allows the page
3023 * allocator to do its reclaim / fallback magic. We then insert the
3024 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003025 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003026static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003027{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003028 struct zonelist *zonelist;
3029 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003030 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003031 struct zone *zone;
3032 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003033 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003034 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003035 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003036
3037 if (flags & __GFP_THISNODE)
3038 return NULL;
3039
Christoph Lameter6cb06222007-10-16 01:25:41 -07003040 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003041
Mel Gormancc9a6c82012-03-21 16:34:11 -07003042retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003043 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003044 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003045
Christoph Lameter3c517a62006-12-06 20:33:29 -08003046retry:
3047 /*
3048 * Look through allowed nodes for objects available
3049 * from existing per node queues.
3050 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003051 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3052 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003053
Mel Gorman54a6eb52008-04-28 02:12:16 -07003054 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003055 get_node(cache, nid) &&
3056 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003057 obj = ____cache_alloc_node(cache,
3058 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003059 if (obj)
3060 break;
3061 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003062 }
3063
Christoph Lametercfce6602007-05-06 14:50:17 -07003064 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003065 /*
3066 * This allocation will be performed within the constraints
3067 * of the current cpuset / memory policy requirements.
3068 * We may trigger various forms of reclaim on the allowed
3069 * set and go into memory reserves if necessary.
3070 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003071 struct page *page;
3072
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003073 if (local_flags & __GFP_WAIT)
3074 local_irq_enable();
3075 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003076 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003077 if (local_flags & __GFP_WAIT)
3078 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003079 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003080 /*
3081 * Insert into the appropriate per node queues
3082 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003083 nid = page_to_nid(page);
3084 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003085 obj = ____cache_alloc_node(cache,
3086 flags | GFP_THISNODE, nid);
3087 if (!obj)
3088 /*
3089 * Another processor may allocate the
3090 * objects in the slab since we are
3091 * not holding any locks.
3092 */
3093 goto retry;
3094 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003095 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003096 obj = NULL;
3097 }
3098 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003099 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003100
Mel Gormand26914d2014-04-03 14:47:24 -07003101 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003102 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003103 return obj;
3104}
3105
3106/*
Christoph Lametere498be72005-09-09 13:03:32 -07003107 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003109static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003110 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003111{
3112 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003113 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003114 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003115 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003116 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117
Aaron Tomlin14e50c62013-04-26 16:15:34 +01003118 VM_BUG_ON(nodeid > num_online_nodes());
Christoph Lameter18bf8542014-08-06 16:04:11 -07003119 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003120 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003121
Andrew Mortona737b3e2006-03-22 00:08:11 -08003122retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003123 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003124 spin_lock(&n->list_lock);
3125 entry = n->slabs_partial.next;
3126 if (entry == &n->slabs_partial) {
3127 n->free_touched = 1;
3128 entry = n->slabs_free.next;
3129 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003130 goto must_grow;
3131 }
Christoph Lametere498be72005-09-09 13:03:32 -07003132
Joonsoo Kim8456a642013-10-24 10:07:49 +09003133 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003134 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003135
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003136 STATS_INC_NODEALLOCS(cachep);
3137 STATS_INC_ACTIVE(cachep);
3138 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003139
Joonsoo Kim8456a642013-10-24 10:07:49 +09003140 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003141
Joonsoo Kim8456a642013-10-24 10:07:49 +09003142 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003143 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003144 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003145 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003146
Joonsoo Kim8456a642013-10-24 10:07:49 +09003147 if (page->active == cachep->num)
3148 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003149 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003150 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003151
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003152 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003153 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003154
Andrew Mortona737b3e2006-03-22 00:08:11 -08003155must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003156 spin_unlock(&n->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003157 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003158 if (x)
3159 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003160
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003161 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003162
Andrew Mortona737b3e2006-03-22 00:08:11 -08003163done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003164 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003165}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003166
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003167static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003168slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003169 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003170{
3171 unsigned long save_flags;
3172 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003173 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003174
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003175 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003176
Nick Piggincf40bd12009-01-21 08:12:39 +01003177 lockdep_trace_alloc(flags);
3178
Akinobu Mita773ff602008-12-23 19:37:01 +09003179 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003180 return NULL;
3181
Glauber Costad79923f2012-12-18 14:22:48 -08003182 cachep = memcg_kmem_get_cache(cachep, flags);
3183
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003184 cache_alloc_debugcheck_before(cachep, flags);
3185 local_irq_save(save_flags);
3186
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003187 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003188 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003189
Christoph Lameter18bf8542014-08-06 16:04:11 -07003190 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003191 /* Node not bootstrapped yet */
3192 ptr = fallback_alloc(cachep, flags);
3193 goto out;
3194 }
3195
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003196 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003197 /*
3198 * Use the locally cached objects if possible.
3199 * However ____cache_alloc does not allow fallback
3200 * to other nodes. It may fail while we still have
3201 * objects on other nodes available.
3202 */
3203 ptr = ____cache_alloc(cachep, flags);
3204 if (ptr)
3205 goto out;
3206 }
3207 /* ___cache_alloc_node can fall back to other nodes */
3208 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3209 out:
3210 local_irq_restore(save_flags);
3211 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003212 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003213 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003214
Joe Perches5087c822013-09-10 17:02:51 -07003215 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003216 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003217 if (unlikely(flags & __GFP_ZERO))
3218 memset(ptr, 0, cachep->object_size);
3219 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003220
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003221 return ptr;
3222}
3223
3224static __always_inline void *
3225__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3226{
3227 void *objp;
3228
David Rientjesf0432d12014-04-07 15:37:30 -07003229 if (current->mempolicy || unlikely(current->flags & PF_SPREAD_SLAB)) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003230 objp = alternate_node_alloc(cache, flags);
3231 if (objp)
3232 goto out;
3233 }
3234 objp = ____cache_alloc(cache, flags);
3235
3236 /*
3237 * We may just have run out of memory on the local node.
3238 * ____cache_alloc_node() knows how to locate memory on other nodes
3239 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003240 if (!objp)
3241 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003242
3243 out:
3244 return objp;
3245}
3246#else
3247
3248static __always_inline void *
3249__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3250{
3251 return ____cache_alloc(cachep, flags);
3252}
3253
3254#endif /* CONFIG_NUMA */
3255
3256static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003257slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003258{
3259 unsigned long save_flags;
3260 void *objp;
3261
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003262 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003263
Nick Piggincf40bd12009-01-21 08:12:39 +01003264 lockdep_trace_alloc(flags);
3265
Akinobu Mita773ff602008-12-23 19:37:01 +09003266 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003267 return NULL;
3268
Glauber Costad79923f2012-12-18 14:22:48 -08003269 cachep = memcg_kmem_get_cache(cachep, flags);
3270
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003271 cache_alloc_debugcheck_before(cachep, flags);
3272 local_irq_save(save_flags);
3273 objp = __do_cache_alloc(cachep, flags);
3274 local_irq_restore(save_flags);
3275 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003276 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003277 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003278 prefetchw(objp);
3279
Joe Perches5087c822013-09-10 17:02:51 -07003280 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003281 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003282 if (unlikely(flags & __GFP_ZERO))
3283 memset(objp, 0, cachep->object_size);
3284 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003285
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003286 return objp;
3287}
Christoph Lametere498be72005-09-09 13:03:32 -07003288
3289/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003290 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003291 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003292 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003293static void free_block(struct kmem_cache *cachep, void **objpp,
3294 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295{
3296 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003297 struct kmem_cache_node *n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298
3299 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003300 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003301 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302
Mel Gorman072bb0a2012-07-31 16:43:58 -07003303 clear_obj_pfmemalloc(&objpp[i]);
3304 objp = objpp[i];
3305
Joonsoo Kim8456a642013-10-24 10:07:49 +09003306 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003307 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003308 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003309 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003311 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312
3313 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003314 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003315 if (n->free_objects > n->free_limit) {
3316 n->free_objects -= cachep->num;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003317 list_add_tail(&page->lru, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003319 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320 }
3321 } else {
3322 /* Unconditionally move a slab to the end of the
3323 * partial list on free - maximum time for the
3324 * other objects to be freed, too.
3325 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003326 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 }
3328 }
3329}
3330
Pekka Enberg343e0d72006-02-01 03:05:50 -08003331static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332{
3333 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003334 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003335 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003336 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337
3338 batchcount = ac->batchcount;
3339#if DEBUG
3340 BUG_ON(!batchcount || batchcount > ac->avail);
3341#endif
3342 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003343 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003344 spin_lock(&n->list_lock);
3345 if (n->shared) {
3346 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003347 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 if (max) {
3349 if (batchcount > max)
3350 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003351 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003352 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003353 shared_array->avail += batchcount;
3354 goto free_done;
3355 }
3356 }
3357
Joonsoo Kim97654df2014-08-06 16:04:25 -07003358 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003359free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360#if STATS
3361 {
3362 int i = 0;
3363 struct list_head *p;
3364
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003365 p = n->slabs_free.next;
3366 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003367 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368
Joonsoo Kim8456a642013-10-24 10:07:49 +09003369 page = list_entry(p, struct page, lru);
3370 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371
3372 i++;
3373 p = p->next;
3374 }
3375 STATS_SET_FREEABLE(cachep, i);
3376 }
3377#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003378 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003379 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003381 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382}
3383
3384/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003385 * Release an obj back to its cache. If the obj has a constructed state, it must
3386 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003388static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003389 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003391 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392
3393 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003394 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003395 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003397 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003398
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003399 /*
3400 * Skip calling cache_free_alien() when the platform is not numa.
3401 * This will avoid cache misses that happen while accessing slabp (which
3402 * is per page memory reference) to get nodeid. Instead use a global
3403 * variable to skip the call, which is mostly likely to be present in
3404 * the cache.
3405 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003406 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003407 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003408
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 if (likely(ac->avail < ac->limit)) {
3410 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411 } else {
3412 STATS_INC_FREEMISS(cachep);
3413 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003415
Mel Gorman072bb0a2012-07-31 16:43:58 -07003416 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417}
3418
3419/**
3420 * kmem_cache_alloc - Allocate an object
3421 * @cachep: The cache to allocate from.
3422 * @flags: See kmalloc().
3423 *
3424 * Allocate an object from this cache. The flags are only relevant
3425 * if the cache has no available objects.
3426 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003427void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003429 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003430
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003431 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003432 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003433
3434 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435}
3436EXPORT_SYMBOL(kmem_cache_alloc);
3437
Li Zefan0f24f122009-12-11 15:45:30 +08003438#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003439void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003440kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003441{
Steven Rostedt85beb582010-11-24 16:23:34 -05003442 void *ret;
3443
Ezequiel Garcia48356302012-09-08 17:47:57 -03003444 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003445
3446 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003447 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003448 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003449}
Steven Rostedt85beb582010-11-24 16:23:34 -05003450EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003451#endif
3452
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003454/**
3455 * kmem_cache_alloc_node - Allocate an object on the specified node
3456 * @cachep: The cache to allocate from.
3457 * @flags: See kmalloc().
3458 * @nodeid: node number of the target node.
3459 *
3460 * Identical to kmem_cache_alloc but it will allocate memory on the given
3461 * node, which can improve the performance for cpu bound structures.
3462 *
3463 * Fallback to other node is possible if __GFP_THISNODE is not set.
3464 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003465void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3466{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003467 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003468
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003469 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003470 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003471 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003472
3473 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003474}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475EXPORT_SYMBOL(kmem_cache_alloc_node);
3476
Li Zefan0f24f122009-12-11 15:45:30 +08003477#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003478void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003479 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003480 int nodeid,
3481 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003482{
Steven Rostedt85beb582010-11-24 16:23:34 -05003483 void *ret;
3484
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003485 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003486
Steven Rostedt85beb582010-11-24 16:23:34 -05003487 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003488 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003489 flags, nodeid);
3490 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003491}
Steven Rostedt85beb582010-11-24 16:23:34 -05003492EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003493#endif
3494
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003495static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003496__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003497{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003498 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003499
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003500 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003501 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3502 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003503 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003504}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003505
Li Zefan0bb38a52009-12-11 15:45:50 +08003506#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003507void *__kmalloc_node(size_t size, gfp_t flags, int node)
3508{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003509 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003510}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003511EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003512
3513void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003514 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003515{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003516 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003517}
3518EXPORT_SYMBOL(__kmalloc_node_track_caller);
3519#else
3520void *__kmalloc_node(size_t size, gfp_t flags, int node)
3521{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003522 return __do_kmalloc_node(size, flags, node, 0);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003523}
3524EXPORT_SYMBOL(__kmalloc_node);
Li Zefan0bb38a52009-12-11 15:45:50 +08003525#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003526#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527
3528/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003529 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003531 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003532 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003534static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003535 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003537 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003538 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003540 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003541 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3542 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003543 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003544
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003545 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003546 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003547
3548 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003549}
3550
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003551
Li Zefan0bb38a52009-12-11 15:45:50 +08003552#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003553void *__kmalloc(size_t size, gfp_t flags)
3554{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003555 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556}
3557EXPORT_SYMBOL(__kmalloc);
3558
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003559void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003560{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003561 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003562}
3563EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003564
3565#else
3566void *__kmalloc(size_t size, gfp_t flags)
3567{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003568 return __do_kmalloc(size, flags, 0);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003569}
3570EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003571#endif
3572
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573/**
3574 * kmem_cache_free - Deallocate an object
3575 * @cachep: The cache the allocation was from.
3576 * @objp: The previously allocated object.
3577 *
3578 * Free an object which was previously allocated from this
3579 * cache.
3580 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003581void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582{
3583 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003584 cachep = cache_from_obj(cachep, objp);
3585 if (!cachep)
3586 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587
3588 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003589 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003590 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003591 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003592 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003594
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003595 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596}
3597EXPORT_SYMBOL(kmem_cache_free);
3598
3599/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 * kfree - free previously allocated memory
3601 * @objp: pointer returned by kmalloc.
3602 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003603 * If @objp is NULL, no operation is performed.
3604 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605 * Don't free memory not originally allocated by kmalloc()
3606 * or you will run into trouble.
3607 */
3608void kfree(const void *objp)
3609{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003610 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 unsigned long flags;
3612
Pekka Enberg2121db72009-03-25 11:05:57 +02003613 trace_kfree(_RET_IP_, objp);
3614
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003615 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616 return;
3617 local_irq_save(flags);
3618 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003619 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003620 debug_check_no_locks_freed(objp, c->object_size);
3621
3622 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003623 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003624 local_irq_restore(flags);
3625}
3626EXPORT_SYMBOL(kfree);
3627
Christoph Lametere498be72005-09-09 13:03:32 -07003628/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003629 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003630 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003631static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003632{
3633 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003634 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003635 struct array_cache *new_shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07003636 struct alien_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003637
Mel Gorman9c09a952008-01-24 05:49:54 -08003638 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003639
Paul Menage3395ee02006-12-06 20:32:16 -08003640 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003641 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003642 if (!new_alien)
3643 goto fail;
3644 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003645
Eric Dumazet63109842007-05-06 14:49:28 -07003646 new_shared = NULL;
3647 if (cachep->shared) {
3648 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003649 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003650 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003651 if (!new_shared) {
3652 free_alien_cache(new_alien);
3653 goto fail;
3654 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003655 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003656
Christoph Lameter18bf8542014-08-06 16:04:11 -07003657 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003658 if (n) {
3659 struct array_cache *shared = n->shared;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003660 LIST_HEAD(list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003661
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003662 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003663
Christoph Lametercafeb022006-03-25 03:06:46 -08003664 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003665 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07003666 shared->avail, node, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003667
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003668 n->shared = new_shared;
3669 if (!n->alien) {
3670 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003671 new_alien = NULL;
3672 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003673 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003674 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003675 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003676 slabs_destroy(cachep, &list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003677 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003678 free_alien_cache(new_alien);
3679 continue;
3680 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003681 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3682 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003683 free_alien_cache(new_alien);
3684 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003685 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003686 }
Christoph Lametere498be72005-09-09 13:03:32 -07003687
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003688 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003689 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3690 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003691 n->shared = new_shared;
3692 n->alien = new_alien;
3693 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003694 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003695 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003696 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003697 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003698
Andrew Mortona737b3e2006-03-22 00:08:11 -08003699fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003700 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003701 /* Cache is not active yet. Roll back what we did */
3702 node--;
3703 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003704 n = get_node(cachep, node);
3705 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003706 kfree(n->shared);
3707 free_alien_cache(n->alien);
3708 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003709 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003710 }
3711 node--;
3712 }
3713 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003714 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003715}
3716
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003718 struct kmem_cache *cachep;
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003719 struct array_cache *new[0];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720};
3721
3722static void do_ccupdate_local(void *info)
3723{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003724 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 struct array_cache *old;
3726
3727 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003728 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003729
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3731 new->new[smp_processor_id()] = old;
3732}
3733
Christoph Lameter18004c52012-07-06 15:25:12 -05003734/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003735static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003736 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003738 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003739 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740
Eric Dumazetacfe7d72011-07-25 08:55:42 +02003741 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3742 gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003743 if (!new)
3744 return -ENOMEM;
3745
Christoph Lametere498be72005-09-09 13:03:32 -07003746 for_each_online_cpu(i) {
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003747 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003748 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003749 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003750 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003751 kfree(new->new[i]);
3752 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003753 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754 }
3755 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003756 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003758 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003759
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761 cachep->batchcount = batchcount;
3762 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003763 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764
Christoph Lametere498be72005-09-09 13:03:32 -07003765 for_each_online_cpu(i) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003766 LIST_HEAD(list);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003767 struct array_cache *ccold = new->new[i];
Christoph Lameter18bf8542014-08-06 16:04:11 -07003768 int node;
3769 struct kmem_cache_node *n;
3770
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 if (!ccold)
3772 continue;
Christoph Lameter18bf8542014-08-06 16:04:11 -07003773
3774 node = cpu_to_mem(i);
3775 n = get_node(cachep, node);
3776 spin_lock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003777 free_block(cachep, ccold->entry, ccold->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003778 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003779 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780 kfree(ccold);
3781 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003782 kfree(new);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003783 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784}
3785
Glauber Costa943a4512012-12-18 14:23:03 -08003786static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3787 int batchcount, int shared, gfp_t gfp)
3788{
3789 int ret;
3790 struct kmem_cache *c = NULL;
3791 int i = 0;
3792
3793 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3794
3795 if (slab_state < FULL)
3796 return ret;
3797
3798 if ((ret < 0) || !is_root_cache(cachep))
3799 return ret;
3800
Glauber Costaebe945c2012-12-18 14:23:10 -08003801 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
Glauber Costa943a4512012-12-18 14:23:03 -08003802 for_each_memcg_cache_index(i) {
Qiang Huang2ade4de2013-11-12 15:08:23 -08003803 c = cache_from_memcg_idx(cachep, i);
Glauber Costa943a4512012-12-18 14:23:03 -08003804 if (c)
3805 /* return value determined by the parent cache only */
3806 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3807 }
3808
3809 return ret;
3810}
3811
Christoph Lameter18004c52012-07-06 15:25:12 -05003812/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003813static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814{
3815 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003816 int limit = 0;
3817 int shared = 0;
3818 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819
Glauber Costa943a4512012-12-18 14:23:03 -08003820 if (!is_root_cache(cachep)) {
3821 struct kmem_cache *root = memcg_root_cache(cachep);
3822 limit = root->limit;
3823 shared = root->shared;
3824 batchcount = root->batchcount;
3825 }
3826
3827 if (limit && shared && batchcount)
3828 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003829 /*
3830 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831 * - create a LIFO ordering, i.e. return objects that are cache-warm
3832 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003833 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003834 * bufctl chains: array operations are cheaper.
3835 * The numbers are guessed, we should auto-tune as described by
3836 * Bonwick.
3837 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003838 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003840 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003842 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003843 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003844 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 limit = 54;
3846 else
3847 limit = 120;
3848
Andrew Mortona737b3e2006-03-22 00:08:11 -08003849 /*
3850 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851 * allocation behaviour: Most allocs on one cpu, most free operations
3852 * on another cpu. For these cases, an efficient object passing between
3853 * cpus is necessary. This is provided by a shared array. The array
3854 * replaces Bonwick's magazine layer.
3855 * On uniprocessor, it's functionally equivalent (but less efficient)
3856 * to a larger limit. Thus disabled by default.
3857 */
3858 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003859 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861
3862#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003863 /*
3864 * With debugging enabled, large batchcount lead to excessively long
3865 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 */
3867 if (limit > 32)
3868 limit = 32;
3869#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003870 batchcount = (limit + 1) / 2;
3871skip_setup:
3872 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 if (err)
3874 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003875 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003876 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877}
3878
Christoph Lameter1b552532006-03-22 00:09:07 -08003879/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003880 * Drain an array if it contains any elements taking the node lock only if
3881 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003882 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003883 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003884static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003885 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886{
Joonsoo Kim97654df2014-08-06 16:04:25 -07003887 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888 int tofree;
3889
Christoph Lameter1b552532006-03-22 00:09:07 -08003890 if (!ac || !ac->avail)
3891 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892 if (ac->touched && !force) {
3893 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003894 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003895 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003896 if (ac->avail) {
3897 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3898 if (tofree > ac->avail)
3899 tofree = (ac->avail + 1) / 2;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003900 free_block(cachep, ac->entry, tofree, node, &list);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003901 ac->avail -= tofree;
3902 memmove(ac->entry, &(ac->entry[tofree]),
3903 sizeof(void *) * ac->avail);
3904 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003905 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003906 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907 }
3908}
3909
3910/**
3911 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003912 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 *
3914 * Called from workqueue/eventd every few seconds.
3915 * Purpose:
3916 * - clear the per-cpu caches for this CPU.
3917 * - return freeable pages to the main free memory pool.
3918 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003919 * If we cannot acquire the cache chain mutex then just give up - we'll try
3920 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003922static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003924 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003925 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003926 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003927 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928
Christoph Lameter18004c52012-07-06 15:25:12 -05003929 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003931 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932
Christoph Lameter18004c52012-07-06 15:25:12 -05003933 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 check_irq_on();
3935
Christoph Lameter35386e32006-03-22 00:09:05 -08003936 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003937 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003938 * have established with reasonable certainty that
3939 * we can do some work if the lock was obtained.
3940 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07003941 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08003942
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003943 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003945 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946
Christoph Lameter35386e32006-03-22 00:09:05 -08003947 /*
3948 * These are racy checks but it does not matter
3949 * if we skip one check or scan twice.
3950 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003951 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003952 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003953
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003954 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003956 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003958 if (n->free_touched)
3959 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003960 else {
3961 int freed;
3962
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003963 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07003964 5 * searchp->num - 1) / (5 * searchp->num));
3965 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003967next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 cond_resched();
3969 }
3970 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05003971 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003972 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003973out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08003974 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003975 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976}
3977
Linus Torvalds158a9622008-01-02 13:04:48 -08003978#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04003979void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003981 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003982 unsigned long active_objs;
3983 unsigned long num_objs;
3984 unsigned long active_slabs = 0;
3985 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003986 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003988 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003989 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 active_objs = 0;
3992 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07003993 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07003994
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003995 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003996 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003997
Joonsoo Kim8456a642013-10-24 10:07:49 +09003998 list_for_each_entry(page, &n->slabs_full, lru) {
3999 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004000 error = "slabs_full accounting error";
4001 active_objs += cachep->num;
4002 active_slabs++;
4003 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004004 list_for_each_entry(page, &n->slabs_partial, lru) {
4005 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004006 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004007 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004008 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004009 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004010 active_slabs++;
4011 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004012 list_for_each_entry(page, &n->slabs_free, lru) {
4013 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004014 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004015 num_slabs++;
4016 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004017 free_objects += n->free_objects;
4018 if (n->shared)
4019 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004020
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004021 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004023 num_slabs += active_slabs;
4024 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004025 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 error = "free_objects accounting error";
4027
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004028 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 if (error)
4030 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4031
Glauber Costa0d7561c2012-10-19 18:20:27 +04004032 sinfo->active_objs = active_objs;
4033 sinfo->num_objs = num_objs;
4034 sinfo->active_slabs = active_slabs;
4035 sinfo->num_slabs = num_slabs;
4036 sinfo->shared_avail = shared_avail;
4037 sinfo->limit = cachep->limit;
4038 sinfo->batchcount = cachep->batchcount;
4039 sinfo->shared = cachep->shared;
4040 sinfo->objects_per_slab = cachep->num;
4041 sinfo->cache_order = cachep->gfporder;
4042}
4043
4044void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4045{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004047 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 unsigned long high = cachep->high_mark;
4049 unsigned long allocs = cachep->num_allocations;
4050 unsigned long grown = cachep->grown;
4051 unsigned long reaped = cachep->reaped;
4052 unsigned long errors = cachep->errors;
4053 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004055 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004056 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057
Joe Perchese92dd4f2010-03-26 19:27:58 -07004058 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4059 "%4lu %4lu %4lu %4lu %4lu",
4060 allocs, high, grown,
4061 reaped, errors, max_freeable, node_allocs,
4062 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 }
4064 /* cpu stats */
4065 {
4066 unsigned long allochit = atomic_read(&cachep->allochit);
4067 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4068 unsigned long freehit = atomic_read(&cachep->freehit);
4069 unsigned long freemiss = atomic_read(&cachep->freemiss);
4070
4071 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004072 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073 }
4074#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075}
4076
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077#define MAX_SLABINFO_WRITE 128
4078/**
4079 * slabinfo_write - Tuning for the slab allocator
4080 * @file: unused
4081 * @buffer: user buffer
4082 * @count: data length
4083 * @ppos: unused
4084 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004085ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004086 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004088 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004090 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004091
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092 if (count > MAX_SLABINFO_WRITE)
4093 return -EINVAL;
4094 if (copy_from_user(&kbuf, buffer, count))
4095 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004096 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097
4098 tmp = strchr(kbuf, ' ');
4099 if (!tmp)
4100 return -EINVAL;
4101 *tmp = '\0';
4102 tmp++;
4103 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4104 return -EINVAL;
4105
4106 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004107 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004109 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004111 if (limit < 1 || batchcount < 1 ||
4112 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004113 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004115 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004116 batchcount, shared,
4117 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004118 }
4119 break;
4120 }
4121 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004122 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 if (res >= 0)
4124 res = count;
4125 return res;
4126}
Al Viro871751e2006-03-25 03:06:39 -08004127
4128#ifdef CONFIG_DEBUG_SLAB_LEAK
4129
4130static void *leaks_start(struct seq_file *m, loff_t *pos)
4131{
Christoph Lameter18004c52012-07-06 15:25:12 -05004132 mutex_lock(&slab_mutex);
4133 return seq_list_start(&slab_caches, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004134}
4135
4136static inline int add_caller(unsigned long *n, unsigned long v)
4137{
4138 unsigned long *p;
4139 int l;
4140 if (!v)
4141 return 1;
4142 l = n[1];
4143 p = n + 2;
4144 while (l) {
4145 int i = l/2;
4146 unsigned long *q = p + 2 * i;
4147 if (*q == v) {
4148 q[1]++;
4149 return 1;
4150 }
4151 if (*q > v) {
4152 l = i;
4153 } else {
4154 p = q + 2;
4155 l -= i + 1;
4156 }
4157 }
4158 if (++n[1] == n[0])
4159 return 0;
4160 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4161 p[0] = v;
4162 p[1] = 1;
4163 return 1;
4164}
4165
Joonsoo Kim8456a642013-10-24 10:07:49 +09004166static void handle_slab(unsigned long *n, struct kmem_cache *c,
4167 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004168{
4169 void *p;
Joonsoo Kim03787302014-06-23 13:22:06 -07004170 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004171
Al Viro871751e2006-03-25 03:06:39 -08004172 if (n[0] == n[1])
4173 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004174 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kim03787302014-06-23 13:22:06 -07004175 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004176 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004177
Al Viro871751e2006-03-25 03:06:39 -08004178 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4179 return;
4180 }
4181}
4182
4183static void show_symbol(struct seq_file *m, unsigned long address)
4184{
4185#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004186 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004187 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004188
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004189 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004190 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004191 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004192 seq_printf(m, " [%s]", modname);
4193 return;
4194 }
4195#endif
4196 seq_printf(m, "%p", (void *)address);
4197}
4198
4199static int leaks_show(struct seq_file *m, void *p)
4200{
Thierry Reding0672aa72012-06-22 19:42:49 +02004201 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004202 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004203 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004204 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004205 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004206 int node;
4207 int i;
4208
4209 if (!(cachep->flags & SLAB_STORE_USER))
4210 return 0;
4211 if (!(cachep->flags & SLAB_RED_ZONE))
4212 return 0;
4213
4214 /* OK, we can do it */
4215
Christoph Lameterdb845062013-02-05 18:45:23 +00004216 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004217
Christoph Lameter18bf8542014-08-06 16:04:11 -07004218 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004219
4220 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004221 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004222
Joonsoo Kim8456a642013-10-24 10:07:49 +09004223 list_for_each_entry(page, &n->slabs_full, lru)
4224 handle_slab(x, cachep, page);
4225 list_for_each_entry(page, &n->slabs_partial, lru)
4226 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004227 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004228 }
4229 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004230 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004231 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004232 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004233 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004234 if (!m->private) {
4235 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004236 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004237 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004238 return -ENOMEM;
4239 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004240 *(unsigned long *)m->private = x[0] * 2;
4241 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004242 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004243 /* Now make sure this entry will be retried */
4244 m->count = m->size;
4245 return 0;
4246 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004247 for (i = 0; i < x[1]; i++) {
4248 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4249 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004250 seq_putc(m, '\n');
4251 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004252
Al Viro871751e2006-03-25 03:06:39 -08004253 return 0;
4254}
4255
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004256static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004257 .start = leaks_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004258 .next = slab_next,
4259 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004260 .show = leaks_show,
4261};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004262
4263static int slabstats_open(struct inode *inode, struct file *file)
4264{
4265 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4266 int ret = -ENOMEM;
4267 if (n) {
4268 ret = seq_open(file, &slabstats_op);
4269 if (!ret) {
4270 struct seq_file *m = file->private_data;
4271 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4272 m->private = n;
4273 n = NULL;
4274 }
4275 kfree(n);
4276 }
4277 return ret;
4278}
4279
4280static const struct file_operations proc_slabstats_operations = {
4281 .open = slabstats_open,
4282 .read = seq_read,
4283 .llseek = seq_lseek,
4284 .release = seq_release_private,
4285};
Al Viro871751e2006-03-25 03:06:39 -08004286#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004287
4288static int __init slab_proc_init(void)
4289{
4290#ifdef CONFIG_DEBUG_SLAB_LEAK
4291 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4292#endif
4293 return 0;
4294}
4295module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004296#endif
4297
Manfred Spraul00e145b2005-09-03 15:55:07 -07004298/**
4299 * ksize - get the actual amount of memory allocated for a given object
4300 * @objp: Pointer to the object
4301 *
4302 * kmalloc may internally round up allocations and return more memory
4303 * than requested. ksize() can be used to determine the actual amount of
4304 * memory allocated. The caller may use this additional memory, even though
4305 * a smaller amount of memory was initially specified with the kmalloc call.
4306 * The caller must guarantee that objp points to a valid object previously
4307 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4308 * must not be freed during the duration of the call.
4309 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004310size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004312 BUG_ON(!objp);
4313 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004314 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004315
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004316 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004317}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004318EXPORT_SYMBOL(ksize);