blob: 4765c97ce6900d98b8ce2968cf9ff62a176f6e42 [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 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700240#define NUM_INIT_LISTS (2 * 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
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700243#define SIZE_NODE (MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Christoph Lametered11d9e2006-06-30 01:55:45 -0700245static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000246 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700247static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700248 int node, struct list_head *list);
249static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300250static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000251static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700252
Ingo Molnare0a42722006-06-23 02:03:46 -0700253static int slab_early_init = 1;
254
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000255#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700256
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000257static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700258{
259 INIT_LIST_HEAD(&parent->slabs_full);
260 INIT_LIST_HEAD(&parent->slabs_partial);
261 INIT_LIST_HEAD(&parent->slabs_free);
262 parent->shared = NULL;
263 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800264 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700265 spin_lock_init(&parent->list_lock);
266 parent->free_objects = 0;
267 parent->free_touched = 0;
268}
269
Andrew Mortona737b3e2006-03-22 00:08:11 -0800270#define MAKE_LIST(cachep, listp, slab, nodeid) \
271 do { \
272 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700273 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700274 } while (0)
275
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
277 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700278 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
279 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
280 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
281 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283#define CFLGS_OFF_SLAB (0x80000000UL)
284#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
Catalin Marinasd4322d82015-11-05 18:45:54 -0800285#define OFF_SLAB_MIN_SIZE (max_t(size_t, PAGE_SIZE >> 5, KMALLOC_MIN_SIZE + 1))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286
287#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800288/*
289 * Optimization question: fewer reaps means less probability for unnessary
290 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100292 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 * which could lock up otherwise freeable slabs.
294 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800295#define REAPTIMEOUT_AC (2*HZ)
296#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297
298#if STATS
299#define STATS_INC_ACTIVE(x) ((x)->num_active++)
300#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
301#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
302#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700303#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800304#define STATS_SET_HIGH(x) \
305 do { \
306 if ((x)->num_active > (x)->high_mark) \
307 (x)->high_mark = (x)->num_active; \
308 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309#define STATS_INC_ERR(x) ((x)->errors++)
310#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700311#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700312#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800313#define STATS_SET_FREEABLE(x, i) \
314 do { \
315 if ((x)->max_freeable < i) \
316 (x)->max_freeable = i; \
317 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
319#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
320#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
321#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
322#else
323#define STATS_INC_ACTIVE(x) do { } while (0)
324#define STATS_DEC_ACTIVE(x) do { } while (0)
325#define STATS_INC_ALLOCED(x) do { } while (0)
326#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700327#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328#define STATS_SET_HIGH(x) do { } while (0)
329#define STATS_INC_ERR(x) do { } while (0)
330#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700331#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700332#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800333#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334#define STATS_INC_ALLOCHIT(x) do { } while (0)
335#define STATS_INC_ALLOCMISS(x) do { } while (0)
336#define STATS_INC_FREEHIT(x) do { } while (0)
337#define STATS_INC_FREEMISS(x) do { } while (0)
338#endif
339
340#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341
Andrew Mortona737b3e2006-03-22 00:08:11 -0800342/*
343 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800345 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 * the end of an object is aligned with the end of the real
347 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800348 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800350 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500351 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
352 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800353 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800355static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800357 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358}
359
David Woodhouseb46b8f12007-05-08 00:22:59 -0700360static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361{
362 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700363 return (unsigned long long*) (objp + obj_offset(cachep) -
364 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365}
366
David Woodhouseb46b8f12007-05-08 00:22:59 -0700367static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368{
369 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
370 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500371 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700372 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400373 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500374 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700375 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376}
377
Pekka Enberg343e0d72006-02-01 03:05:50 -0800378static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379{
380 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500381 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382}
383
384#else
385
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800386#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700387#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
388#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
390
391#endif
392
Joonsoo Kim03787302014-06-23 13:22:06 -0700393#define OBJECT_FREE (0)
394#define OBJECT_ACTIVE (1)
395
396#ifdef CONFIG_DEBUG_SLAB_LEAK
397
398static void set_obj_status(struct page *page, int idx, int val)
399{
400 int freelist_size;
401 char *status;
402 struct kmem_cache *cachep = page->slab_cache;
403
404 freelist_size = cachep->num * sizeof(freelist_idx_t);
405 status = (char *)page->freelist + freelist_size;
406 status[idx] = val;
407}
408
409static inline unsigned int get_obj_status(struct page *page, int idx)
410{
411 int freelist_size;
412 char *status;
413 struct kmem_cache *cachep = page->slab_cache;
414
415 freelist_size = cachep->num * sizeof(freelist_idx_t);
416 status = (char *)page->freelist + freelist_size;
417
418 return status[idx];
419}
420
421#else
422static inline void set_obj_status(struct page *page, int idx, int val) {}
423
424#endif
425
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700427 * Do not go above this order unless 0 objects fit into the slab or
428 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 */
David Rientjes543585c2011-10-18 22:09:24 -0700430#define SLAB_MAX_ORDER_HI 1
431#define SLAB_MAX_ORDER_LO 0
432static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700433static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800435static inline struct kmem_cache *virt_to_cache(const void *obj)
436{
Christoph Lameterb49af682007-05-06 14:49:41 -0700437 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500438 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800439}
440
Joonsoo Kim8456a642013-10-24 10:07:49 +0900441static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800442 unsigned int idx)
443{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900444 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800445}
446
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800447/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500448 * We want to avoid an expensive divide : (offset / cache->size)
449 * Using the fact that size is a constant for a particular cache,
450 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800451 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
452 */
453static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900454 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800455{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900456 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800457 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800458}
459
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000461static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800462 .batchcount = 1,
463 .limit = BOOT_CPUCACHE_ENTRIES,
464 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500465 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800466 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467};
468
Joonsoo Kimedcad252014-08-08 14:19:15 -0700469#define BAD_ALIEN_MAGIC 0x01020304ul
470
Tejun Heo1871e522009-10-29 22:34:13 +0900471static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472
Pekka Enberg343e0d72006-02-01 03:05:50 -0800473static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700475 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476}
477
Joonsoo Kim03787302014-06-23 13:22:06 -0700478static size_t calculate_freelist_size(int nr_objs, size_t align)
479{
480 size_t freelist_size;
481
482 freelist_size = nr_objs * sizeof(freelist_idx_t);
483 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
484 freelist_size += nr_objs * sizeof(char);
485
486 if (align)
487 freelist_size = ALIGN(freelist_size, align);
488
489 return freelist_size;
490}
491
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900492static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
493 size_t idx_size, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494{
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900495 int nr_objs;
Joonsoo Kim03787302014-06-23 13:22:06 -0700496 size_t remained_size;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900497 size_t freelist_size;
Joonsoo Kim03787302014-06-23 13:22:06 -0700498 int extra_space = 0;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900499
Joonsoo Kim03787302014-06-23 13:22:06 -0700500 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
501 extra_space = sizeof(char);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900502 /*
503 * Ignore padding for the initial guess. The padding
504 * is at most @align-1 bytes, and @buffer_size is at
505 * least @align. In the worst case, this result will
506 * be one greater than the number of objects that fit
507 * into the memory allocation when taking the padding
508 * into account.
509 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700510 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900511
512 /*
513 * This calculated number will be either the right
514 * amount, or one greater than what we want.
515 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700516 remained_size = slab_size - nr_objs * buffer_size;
517 freelist_size = calculate_freelist_size(nr_objs, align);
518 if (remained_size < freelist_size)
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900519 nr_objs--;
520
521 return nr_objs;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800522}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523
Andrew Mortona737b3e2006-03-22 00:08:11 -0800524/*
525 * Calculate the number of objects and left-over bytes for a given buffer size.
526 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800527static void cache_estimate(unsigned long gfporder, size_t buffer_size,
528 size_t align, int flags, size_t *left_over,
529 unsigned int *num)
530{
531 int nr_objs;
532 size_t mgmt_size;
533 size_t slab_size = PAGE_SIZE << gfporder;
534
535 /*
536 * The slab management structure can be either off the slab or
537 * on it. For the latter case, the memory allocated for a
538 * slab is used for:
539 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900540 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800541 * - Padding to respect alignment of @align
542 * - @buffer_size bytes for each object
543 *
544 * If the slab management structure is off the slab, then the
545 * alignment will already be calculated into the size. Because
546 * the slabs are all pages aligned, the objects will be at the
547 * correct alignment when allocated.
548 */
549 if (flags & CFLGS_OFF_SLAB) {
550 mgmt_size = 0;
551 nr_objs = slab_size / buffer_size;
552
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800553 } else {
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900554 nr_objs = calculate_nr_objs(slab_size, buffer_size,
Joonsoo Kima41adfa2013-12-02 17:49:42 +0900555 sizeof(freelist_idx_t), align);
Joonsoo Kim03787302014-06-23 13:22:06 -0700556 mgmt_size = calculate_freelist_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800558 *num = nr_objs;
559 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
Christoph Lameterf28510d2012-09-11 19:49:38 +0000562#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700563#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564
Andrew Mortona737b3e2006-03-22 00:08:11 -0800565static void __slab_error(const char *function, struct kmem_cache *cachep,
566 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
568 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800569 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030571 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000573#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574
Paul Menage3395ee02006-12-06 20:32:16 -0800575/*
576 * By default on NUMA we use alien caches to stage the freeing of
577 * objects allocated from other nodes. This causes massive memory
578 * inefficiencies when using fake NUMA setup to split memory into a
579 * large number of small nodes, so it can be disabled on the command
580 * line
581 */
582
583static int use_alien_caches __read_mostly = 1;
584static int __init noaliencache_setup(char *s)
585{
586 use_alien_caches = 0;
587 return 1;
588}
589__setup("noaliencache", noaliencache_setup);
590
David Rientjes3df1ccc2011-10-18 22:09:28 -0700591static int __init slab_max_order_setup(char *str)
592{
593 get_option(&str, &slab_max_order);
594 slab_max_order = slab_max_order < 0 ? 0 :
595 min(slab_max_order, MAX_ORDER - 1);
596 slab_max_order_set = true;
597
598 return 1;
599}
600__setup("slab_max_order=", slab_max_order_setup);
601
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800602#ifdef CONFIG_NUMA
603/*
604 * Special reaping functions for NUMA systems called from cache_reap().
605 * These take care of doing round robin flushing of alien caches (containing
606 * objects freed on different nodes from which they were allocated) and the
607 * flushing of remote pcps by calling drain_node_pages.
608 */
Tejun Heo1871e522009-10-29 22:34:13 +0900609static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800610
611static void init_reap_node(int cpu)
612{
613 int node;
614
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700615 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800616 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800617 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800618
Tejun Heo1871e522009-10-29 22:34:13 +0900619 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800620}
621
622static void next_reap_node(void)
623{
Christoph Lameter909ea962010-12-08 16:22:55 +0100624 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800625
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800626 node = next_node(node, node_online_map);
627 if (unlikely(node >= MAX_NUMNODES))
628 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100629 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800630}
631
632#else
633#define init_reap_node(cpu) do { } while (0)
634#define next_reap_node(void) do { } while (0)
635#endif
636
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637/*
638 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
639 * via the workqueue/eventd.
640 * Add the CPU number into the expiration time to minimize the possibility of
641 * the CPUs getting into lockstep and contending for the global cache chain
642 * lock.
643 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400644static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645{
Tejun Heo1871e522009-10-29 22:34:13 +0900646 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
648 /*
649 * When this gets called from do_initcalls via cpucache_init(),
650 * init_workqueues() has already run, so keventd will be setup
651 * at that time.
652 */
David Howells52bad642006-11-22 14:54:01 +0000653 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800654 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700655 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800656 schedule_delayed_work_on(cpu, reap_work,
657 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 }
659}
660
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700661static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100663 /*
664 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300665 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100666 * cache the pointers are not cleared and they could be counted as
667 * valid references during a kmemleak scan. Therefore, kmemleak must
668 * not scan such objects.
669 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700670 kmemleak_no_scan(ac);
671 if (ac) {
672 ac->avail = 0;
673 ac->limit = limit;
674 ac->batchcount = batch;
675 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700677}
678
679static struct array_cache *alloc_arraycache(int node, int entries,
680 int batchcount, gfp_t gfp)
681{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700682 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700683 struct array_cache *ac = NULL;
684
685 ac = kmalloc_node(memsize, gfp, node);
686 init_arraycache(ac, entries, batchcount);
687 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
689
Joonsoo Kim8456a642013-10-24 10:07:49 +0900690static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700691{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700692 return PageSlabPfmemalloc(page);
693}
694
695/* Clears pfmemalloc_active if no slabs have pfmalloc set */
696static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
697 struct array_cache *ac)
698{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700699 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
Joonsoo Kim8456a642013-10-24 10:07:49 +0900700 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700701 unsigned long flags;
702
703 if (!pfmemalloc_active)
704 return;
705
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000706 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900707 list_for_each_entry(page, &n->slabs_full, lru)
708 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700709 goto out;
710
Joonsoo Kim8456a642013-10-24 10:07:49 +0900711 list_for_each_entry(page, &n->slabs_partial, 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_free, lru)
716 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700717 goto out;
718
719 pfmemalloc_active = false;
720out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000721 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700722}
723
Mel Gorman381760e2012-07-31 16:44:30 -0700724static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700725 gfp_t flags, bool force_refill)
726{
727 int i;
728 void *objp = ac->entry[--ac->avail];
729
730 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
731 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000732 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700733
734 if (gfp_pfmemalloc_allowed(flags)) {
735 clear_obj_pfmemalloc(&objp);
736 return objp;
737 }
738
739 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700740 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700741 /* If a !PFMEMALLOC object is found, swap them */
742 if (!is_obj_pfmemalloc(ac->entry[i])) {
743 objp = ac->entry[i];
744 ac->entry[i] = ac->entry[ac->avail];
745 ac->entry[ac->avail] = objp;
746 return objp;
747 }
748 }
749
750 /*
751 * If there are empty slabs on the slabs_free list and we are
752 * being forced to refill the cache, mark this one !pfmemalloc.
753 */
Christoph Lameter18bf8542014-08-06 16:04:11 -0700754 n = get_node(cachep, numa_mem_id());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000755 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900756 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900757 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700758 clear_obj_pfmemalloc(&objp);
759 recheck_pfmemalloc_active(cachep, ac);
760 return objp;
761 }
762
763 /* No !PFMEMALLOC objects available */
764 ac->avail++;
765 objp = NULL;
766 }
767
768 return objp;
769}
770
Mel Gorman381760e2012-07-31 16:44:30 -0700771static inline void *ac_get_obj(struct kmem_cache *cachep,
772 struct array_cache *ac, gfp_t flags, bool force_refill)
773{
774 void *objp;
775
776 if (unlikely(sk_memalloc_socks()))
777 objp = __ac_get_obj(cachep, ac, flags, force_refill);
778 else
779 objp = ac->entry[--ac->avail];
780
781 return objp;
782}
783
Joonsoo Kimd3aec342014-10-09 15:26:06 -0700784static noinline void *__ac_put_obj(struct kmem_cache *cachep,
785 struct array_cache *ac, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700786{
787 if (unlikely(pfmemalloc_active)) {
788 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700789 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700790 if (PageSlabPfmemalloc(page))
791 set_obj_pfmemalloc(&objp);
792 }
793
Mel Gorman381760e2012-07-31 16:44:30 -0700794 return objp;
795}
796
797static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
798 void *objp)
799{
800 if (unlikely(sk_memalloc_socks()))
801 objp = __ac_put_obj(cachep, ac, objp);
802
Mel Gorman072bb0a2012-07-31 16:43:58 -0700803 ac->entry[ac->avail++] = objp;
804}
805
Christoph Lameter3ded1752006-03-25 03:06:44 -0800806/*
807 * Transfer objects in one arraycache to another.
808 * Locking must be handled by the caller.
809 *
810 * Return the number of entries transferred.
811 */
812static int transfer_objects(struct array_cache *to,
813 struct array_cache *from, unsigned int max)
814{
815 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700816 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800817
818 if (!nr)
819 return 0;
820
821 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
822 sizeof(void *) *nr);
823
824 from->avail -= nr;
825 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800826 return nr;
827}
828
Christoph Lameter765c4502006-09-27 01:50:08 -0700829#ifndef CONFIG_NUMA
830
831#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000832#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700833
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700834static inline struct alien_cache **alloc_alien_cache(int node,
835 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700836{
Joonsoo Kimedcad252014-08-08 14:19:15 -0700837 return (struct alien_cache **)BAD_ALIEN_MAGIC;
Christoph Lameter765c4502006-09-27 01:50:08 -0700838}
839
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700840static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700841{
842}
843
844static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
845{
846 return 0;
847}
848
849static inline void *alternate_node_alloc(struct kmem_cache *cachep,
850 gfp_t flags)
851{
852 return NULL;
853}
854
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800855static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700856 gfp_t flags, int nodeid)
857{
858 return NULL;
859}
860
David Rientjes4167e9b2015-04-14 15:46:55 -0700861static inline gfp_t gfp_exact_node(gfp_t flags)
862{
863 return flags;
864}
865
Christoph Lameter765c4502006-09-27 01:50:08 -0700866#else /* CONFIG_NUMA */
867
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800868static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800869static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800870
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700871static struct alien_cache *__alloc_alien_cache(int node, int entries,
872 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700873{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700874 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700875 struct alien_cache *alc = NULL;
876
877 alc = kmalloc_node(memsize, gfp, node);
878 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700879 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700880 return alc;
881}
882
883static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
884{
885 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700886 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700887 int i;
888
889 if (limit > 1)
890 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700891 alc_ptr = kzalloc_node(memsize, gfp, node);
892 if (!alc_ptr)
893 return NULL;
894
895 for_each_node(i) {
896 if (i == node || !node_online(i))
897 continue;
898 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
899 if (!alc_ptr[i]) {
900 for (i--; i >= 0; i--)
901 kfree(alc_ptr[i]);
902 kfree(alc_ptr);
903 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700904 }
905 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700906 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700907}
908
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700909static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700910{
911 int i;
912
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700913 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700914 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700915 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700916 kfree(alc_ptr[i]);
917 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700918}
919
Pekka Enberg343e0d72006-02-01 03:05:50 -0800920static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700921 struct array_cache *ac, int node,
922 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700923{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700924 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700925
926 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000927 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800928 /*
929 * Stuff objects into the remote nodes shared array first.
930 * That way we could avoid the overhead of putting the objects
931 * into the free lists and getting them back later.
932 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000933 if (n->shared)
934 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800935
Joonsoo Kim833b7062014-08-06 16:04:33 -0700936 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700937 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000938 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700939 }
940}
941
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800942/*
943 * Called from cache_reap() to regularly drain alien caches round robin.
944 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000945static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800946{
Christoph Lameter909ea962010-12-08 16:22:55 +0100947 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800948
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000949 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700950 struct alien_cache *alc = n->alien[node];
951 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800952
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700953 if (alc) {
954 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700955 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700956 LIST_HEAD(list);
957
958 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700959 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700960 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700961 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800962 }
963 }
964}
965
Andrew Mortona737b3e2006-03-22 00:08:11 -0800966static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700967 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700968{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800969 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700970 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700971 struct array_cache *ac;
972 unsigned long flags;
973
974 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700975 alc = alien[i];
976 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700977 LIST_HEAD(list);
978
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700979 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700980 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700981 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700982 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700983 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700984 }
985 }
986}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700987
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700988static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
989 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700990{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000991 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700992 struct alien_cache *alien = NULL;
993 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700994 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700995
Christoph Lameter18bf8542014-08-06 16:04:11 -0700996 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700997 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700998 if (n->alien && n->alien[page_node]) {
999 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001000 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001001 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001002 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001003 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001004 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001005 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001006 ac_put_obj(cachep, ac, objp);
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001007 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -07001008 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001009 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001010 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001011 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001012 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001013 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07001014 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001015 }
1016 return 1;
1017}
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001018
1019static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1020{
1021 int page_node = page_to_nid(virt_to_page(objp));
1022 int node = numa_mem_id();
1023 /*
1024 * Make sure we are not freeing a object from another node to the array
1025 * cache on this cpu.
1026 */
1027 if (likely(node == page_node))
1028 return 0;
1029
1030 return __cache_free_alien(cachep, objp, node, page_node);
1031}
David Rientjes4167e9b2015-04-14 15:46:55 -07001032
1033/*
Mel Gormand0164ad2015-11-06 16:28:21 -08001034 * Construct gfp mask to allocate from a specific node but do not direct reclaim
1035 * or warn about failures. kswapd may still wake to reclaim in the background.
David Rientjes4167e9b2015-04-14 15:46:55 -07001036 */
1037static inline gfp_t gfp_exact_node(gfp_t flags)
1038{
Mel Gormand0164ad2015-11-06 16:28:21 -08001039 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_DIRECT_RECLAIM;
David Rientjes4167e9b2015-04-14 15:46:55 -07001040}
Christoph Lametere498be72005-09-09 13:03:32 -07001041#endif
1042
David Rientjes8f9f8d92010-03-27 19:40:47 -07001043/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001044 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001045 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001046 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001047 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001048 * already in use.
1049 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001050 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001051 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001052static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001053{
1054 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001055 struct kmem_cache_node *n;
Joonsoo Kim5e804782014-08-06 16:04:40 -07001056 const size_t memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001057
Christoph Lameter18004c52012-07-06 15:25:12 -05001058 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001059 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001060 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001061 * begin anything. Make sure some other cpu on this
1062 * node has not already allocated this
1063 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07001064 n = get_node(cachep, node);
1065 if (!n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001066 n = kmalloc_node(memsize, GFP_KERNEL, node);
1067 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001068 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001069 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001070 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1071 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001072
1073 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001074 * The kmem_cache_nodes don't come and go as CPUs
1075 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001076 * protection here.
1077 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001078 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001079 }
1080
Christoph Lameter18bf8542014-08-06 16:04:11 -07001081 spin_lock_irq(&n->list_lock);
1082 n->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001083 (1 + nr_cpus_node(node)) *
1084 cachep->batchcount + cachep->num;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001085 spin_unlock_irq(&n->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001086 }
1087 return 0;
1088}
1089
Wanpeng Li0fa81032013-07-04 08:33:22 +08001090static inline int slabs_tofree(struct kmem_cache *cachep,
1091 struct kmem_cache_node *n)
1092{
1093 return (n->free_objects + cachep->num - 1) / cachep->num;
1094}
1095
Paul Gortmaker0db06282013-06-19 14:53:51 -04001096static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001098 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001099 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001100 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301101 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001102
Christoph Lameter18004c52012-07-06 15:25:12 -05001103 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001104 struct array_cache *nc;
1105 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001106 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -07001107 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001108
Christoph Lameter18bf8542014-08-06 16:04:11 -07001109 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001110 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001111 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001112
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001113 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001114
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001115 /* Free limit for this kmem_cache_node */
1116 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001117
1118 /* cpu is dead; no one can alloc from it. */
1119 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1120 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001121 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001122 nc->avail = 0;
1123 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001124
Rusty Russell58463c12009-12-17 11:43:12 -06001125 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001126 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001127 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001128 }
1129
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001130 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001131 if (shared) {
1132 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001133 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001134 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001135 }
1136
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001137 alien = n->alien;
1138 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001139
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001140 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001141
1142 kfree(shared);
1143 if (alien) {
1144 drain_alien_cache(cachep, alien);
1145 free_alien_cache(alien);
1146 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001147
1148free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001149 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001150 }
1151 /*
1152 * In the previous loop, all the objects were freed to
1153 * the respective cache's slabs, now we can go ahead and
1154 * shrink each nodelist to its limit.
1155 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001156 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001157 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001158 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001159 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001160 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001161 }
1162}
1163
Paul Gortmaker0db06282013-06-19 14:53:51 -04001164static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001165{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001166 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001167 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001168 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001169 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001171 /*
1172 * We need to do this right in the beginning since
1173 * alloc_arraycache's are going to use this list.
1174 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001175 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001176 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001177 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001178 if (err < 0)
1179 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001180
1181 /*
1182 * Now we can go ahead with allocating the shared arrays and
1183 * array caches
1184 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001185 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001186 struct array_cache *shared = NULL;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001187 struct alien_cache **alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001188
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001189 if (cachep->shared) {
1190 shared = alloc_arraycache(node,
1191 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001192 0xbaadf00d, GFP_KERNEL);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001193 if (!shared)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001194 goto bad;
1195 }
1196 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001197 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001198 if (!alien) {
1199 kfree(shared);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001200 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001201 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001202 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07001203 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001204 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001205
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001206 spin_lock_irq(&n->list_lock);
1207 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001208 /*
1209 * We are serialised from CPU_DEAD or
1210 * CPU_UP_CANCELLED by the cpucontrol lock
1211 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001212 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001213 shared = NULL;
1214 }
1215#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001216 if (!n->alien) {
1217 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001218 alien = NULL;
1219 }
1220#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001221 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001222 kfree(shared);
1223 free_alien_cache(alien);
1224 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001225
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001226 return 0;
1227bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001228 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001229 return -ENOMEM;
1230}
1231
Paul Gortmaker0db06282013-06-19 14:53:51 -04001232static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001233 unsigned long action, void *hcpu)
1234{
1235 long cpu = (long)hcpu;
1236 int err = 0;
1237
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001239 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001240 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001241 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001242 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001243 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 break;
1245 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001246 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 start_cpu_timer(cpu);
1248 break;
1249#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001250 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001251 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001252 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001253 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001254 * held so that if cache_reap() is invoked it cannot do
1255 * anything expensive but will only modify reap_work
1256 * and reschedule the timer.
1257 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001258 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001259 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001260 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001261 break;
1262 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001263 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001264 start_cpu_timer(cpu);
1265 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001267 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001268 /*
1269 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001270 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001271 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001272 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001273 * structure is usually allocated from kmem_cache_create() and
1274 * gets destroyed at kmem_cache_destroy().
1275 */
Simon Arlott183ff222007-10-20 01:27:18 +02001276 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001277#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001279 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001280 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001281 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001282 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001285 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286}
1287
Paul Gortmaker0db06282013-06-19 14:53:51 -04001288static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001289 &cpuup_callback, NULL, 0
1290};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291
David Rientjes8f9f8d92010-03-27 19:40:47 -07001292#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1293/*
1294 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1295 * Returns -EBUSY if all objects cannot be drained so that the node is not
1296 * removed.
1297 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001298 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001299 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001300static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001301{
1302 struct kmem_cache *cachep;
1303 int ret = 0;
1304
Christoph Lameter18004c52012-07-06 15:25:12 -05001305 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001306 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001307
Christoph Lameter18bf8542014-08-06 16:04:11 -07001308 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001309 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001310 continue;
1311
Wanpeng Li0fa81032013-07-04 08:33:22 +08001312 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001313
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001314 if (!list_empty(&n->slabs_full) ||
1315 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001316 ret = -EBUSY;
1317 break;
1318 }
1319 }
1320 return ret;
1321}
1322
1323static int __meminit slab_memory_callback(struct notifier_block *self,
1324 unsigned long action, void *arg)
1325{
1326 struct memory_notify *mnb = arg;
1327 int ret = 0;
1328 int nid;
1329
1330 nid = mnb->status_change_nid;
1331 if (nid < 0)
1332 goto out;
1333
1334 switch (action) {
1335 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001336 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001337 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001338 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001339 break;
1340 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001341 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001342 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001343 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001344 break;
1345 case MEM_ONLINE:
1346 case MEM_OFFLINE:
1347 case MEM_CANCEL_ONLINE:
1348 case MEM_CANCEL_OFFLINE:
1349 break;
1350 }
1351out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001352 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001353}
1354#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1355
Christoph Lametere498be72005-09-09 13:03:32 -07001356/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001357 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001358 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001359static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001360 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001361{
Christoph Lameter6744f082013-01-10 19:12:17 +00001362 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001363
Christoph Lameter6744f082013-01-10 19:12:17 +00001364 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001365 BUG_ON(!ptr);
1366
Christoph Lameter6744f082013-01-10 19:12:17 +00001367 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001368 /*
1369 * Do not assume that spinlocks can be initialized via memcpy:
1370 */
1371 spin_lock_init(&ptr->list_lock);
1372
Christoph Lametere498be72005-09-09 13:03:32 -07001373 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001374 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001375}
1376
Andrew Mortona737b3e2006-03-22 00:08:11 -08001377/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001378 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1379 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001380 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001381static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001382{
1383 int node;
1384
1385 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001386 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001387 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001388 REAPTIMEOUT_NODE +
1389 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001390 }
1391}
1392
1393/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001394 * Initialisation. Called after the page allocator have been initialised and
1395 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 */
1397void __init kmem_cache_init(void)
1398{
Christoph Lametere498be72005-09-09 13:03:32 -07001399 int i;
1400
Joonsoo Kim68126702013-10-24 10:07:42 +09001401 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1402 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001403 kmem_cache = &kmem_cache_boot;
1404
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001405 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001406 use_alien_caches = 0;
1407
Christoph Lameter3c583462012-11-28 16:23:01 +00001408 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001409 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001410
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 /*
1412 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001413 * page orders on machines with more than 32MB of memory if
1414 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001416 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001417 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419 /* Bootstrap is tricky, because several objects are allocated
1420 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001421 * 1) initialize the kmem_cache cache: it contains the struct
1422 * kmem_cache structures of all caches, except kmem_cache itself:
1423 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001424 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001425 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001426 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001428 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001429 * An __init data area is used for the head array.
1430 * 3) Create the remaining kmalloc caches, with minimally sized
1431 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001432 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001434 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001435 * the other cache's with kmalloc allocated memory.
1436 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 */
1438
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001439 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440
Eric Dumazet8da34302007-05-06 14:49:29 -07001441 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001442 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001443 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001444 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001445 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001446 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001447 SLAB_HWCACHE_ALIGN);
1448 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001449 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450
Andrew Mortona737b3e2006-03-22 00:08:11 -08001451 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001452 * Initialize the caches that provide memory for the kmem_cache_node
1453 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001454 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001455 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001456 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001457 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001458 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001459
Ingo Molnare0a42722006-06-23 02:03:46 -07001460 slab_early_init = 0;
1461
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001462 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001463 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001464 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
Mel Gorman9c09a952008-01-24 05:49:54 -08001466 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001467 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001468
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001469 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001470 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001471 }
1472 }
1473
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001474 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001475}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001476
Pekka Enberg8429db52009-06-12 15:58:59 +03001477void __init kmem_cache_init_late(void)
1478{
1479 struct kmem_cache *cachep;
1480
Christoph Lameter97d06602012-07-06 15:25:11 -05001481 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001482
Pekka Enberg8429db52009-06-12 15:58:59 +03001483 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001484 mutex_lock(&slab_mutex);
1485 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001486 if (enable_cpucache(cachep, GFP_NOWAIT))
1487 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001488 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001489
Christoph Lameter97d06602012-07-06 15:25:11 -05001490 /* Done! */
1491 slab_state = FULL;
1492
Andrew Mortona737b3e2006-03-22 00:08:11 -08001493 /*
1494 * Register a cpu startup notifier callback that initializes
1495 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 */
1497 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
David Rientjes8f9f8d92010-03-27 19:40:47 -07001499#ifdef CONFIG_NUMA
1500 /*
1501 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001502 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001503 */
1504 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1505#endif
1506
Andrew Mortona737b3e2006-03-22 00:08:11 -08001507 /*
1508 * The reap timers are started later, with a module init call: That part
1509 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 */
1511}
1512
1513static int __init cpucache_init(void)
1514{
1515 int cpu;
1516
Andrew Mortona737b3e2006-03-22 00:08:11 -08001517 /*
1518 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 */
Christoph Lametere498be72005-09-09 13:03:32 -07001520 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001521 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001522
1523 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001524 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 return 0;
1526}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527__initcall(cpucache_init);
1528
Rafael Aquini8bdec192012-03-09 17:27:27 -03001529static noinline void
1530slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1531{
David Rientjes9a02d692014-06-04 16:06:36 -07001532#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001533 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001534 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001535 unsigned long flags;
1536 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001537 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1538 DEFAULT_RATELIMIT_BURST);
1539
1540 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1541 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001542
1543 printk(KERN_WARNING
1544 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1545 nodeid, gfpflags);
1546 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001547 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001548
Christoph Lameter18bf8542014-08-06 16:04:11 -07001549 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001550 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1551 unsigned long active_slabs = 0, num_slabs = 0;
1552
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001553 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001554 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001555 active_objs += cachep->num;
1556 active_slabs++;
1557 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001558 list_for_each_entry(page, &n->slabs_partial, lru) {
1559 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001560 active_slabs++;
1561 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001562 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001563 num_slabs++;
1564
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001565 free_objects += n->free_objects;
1566 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001567
1568 num_slabs += active_slabs;
1569 num_objs = num_slabs * cachep->num;
1570 printk(KERN_WARNING
1571 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1572 node, active_slabs, num_slabs, active_objs, num_objs,
1573 free_objects);
1574 }
David Rientjes9a02d692014-06-04 16:06:36 -07001575#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001576}
1577
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001579 * Interface to system's page allocator. No need to hold the
1580 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581 *
1582 * If we requested dmaable memory, we will get it. Even if we
1583 * did not request dmaable memory, we might get it, but that
1584 * would be relatively rare and ignorable.
1585 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001586static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1587 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588{
1589 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001590 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001591
Glauber Costaa618e892012-06-14 16:17:21 +04001592 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001593 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1594 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001595
Vlastimil Babka96db8002015-09-08 15:03:50 -07001596 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001597 if (!page) {
David Rientjes9a02d692014-06-04 16:06:36 -07001598 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001600 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001602 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1603 __free_pages(page, cachep->gfporder);
1604 return NULL;
1605 }
1606
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001607 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Michal Hocko2f064f32015-08-21 14:11:51 -07001608 if (page_is_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -07001609 pfmemalloc_active = true;
1610
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001611 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001613 add_zone_page_state(page_zone(page),
1614 NR_SLAB_RECLAIMABLE, nr_pages);
1615 else
1616 add_zone_page_state(page_zone(page),
1617 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001618 __SetPageSlab(page);
Michal Hocko2f064f32015-08-21 14:11:51 -07001619 if (page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001620 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001621
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001622 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1623 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1624
1625 if (cachep->ctor)
1626 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1627 else
1628 kmemcheck_mark_unallocated_pages(page, nr_pages);
1629 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001630
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001631 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632}
1633
1634/*
1635 * Interface to system's page release.
1636 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001637static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001639 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001641 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001642
Christoph Lameter972d1a72006-09-25 23:31:51 -07001643 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1644 sub_zone_page_state(page_zone(page),
1645 NR_SLAB_RECLAIMABLE, nr_freed);
1646 else
1647 sub_zone_page_state(page_zone(page),
1648 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001649
Joonsoo Kima57a4982013-10-24 10:07:44 +09001650 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001651 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001652 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001653 page_mapcount_reset(page);
1654 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001655
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 if (current->reclaim_state)
1657 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001658 __free_kmem_pages(page, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659}
1660
1661static void kmem_rcu_free(struct rcu_head *head)
1662{
Joonsoo Kim68126702013-10-24 10:07:42 +09001663 struct kmem_cache *cachep;
1664 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Joonsoo Kim68126702013-10-24 10:07:42 +09001666 page = container_of(head, struct page, rcu_head);
1667 cachep = page->slab_cache;
1668
1669 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670}
1671
1672#if DEBUG
1673
1674#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001675static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001676 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001678 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001680 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001682 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 return;
1684
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001685 *addr++ = 0x12345678;
1686 *addr++ = caller;
1687 *addr++ = smp_processor_id();
1688 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 {
1690 unsigned long *sptr = &caller;
1691 unsigned long svalue;
1692
1693 while (!kstack_end(sptr)) {
1694 svalue = *sptr++;
1695 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001696 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 size -= sizeof(unsigned long);
1698 if (size <= sizeof(unsigned long))
1699 break;
1700 }
1701 }
1702
1703 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001704 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705}
1706#endif
1707
Pekka Enberg343e0d72006-02-01 03:05:50 -08001708static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001710 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001711 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712
1713 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001714 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715}
1716
1717static void dump_line(char *data, int offset, int limit)
1718{
1719 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001720 unsigned char error = 0;
1721 int bad_count = 0;
1722
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001723 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001724 for (i = 0; i < limit; i++) {
1725 if (data[offset + i] != POISON_FREE) {
1726 error = data[offset + i];
1727 bad_count++;
1728 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001729 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001730 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1731 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001732
1733 if (bad_count == 1) {
1734 error ^= POISON_FREE;
1735 if (!(error & (error - 1))) {
1736 printk(KERN_ERR "Single bit error detected. Probably "
1737 "bad RAM.\n");
1738#ifdef CONFIG_X86
1739 printk(KERN_ERR "Run memtest86+ or a similar memory "
1740 "test tool.\n");
1741#else
1742 printk(KERN_ERR "Run a memory test tool.\n");
1743#endif
1744 }
1745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746}
1747#endif
1748
1749#if DEBUG
1750
Pekka Enberg343e0d72006-02-01 03:05:50 -08001751static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752{
1753 int i, size;
1754 char *realobj;
1755
1756 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001757 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001758 *dbg_redzone1(cachep, objp),
1759 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 }
1761
1762 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001763 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1764 *dbg_userword(cachep, objp),
1765 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001767 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001768 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001769 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 int limit;
1771 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001772 if (i + limit > size)
1773 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 dump_line(realobj, i, limit);
1775 }
1776}
1777
Pekka Enberg343e0d72006-02-01 03:05:50 -08001778static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779{
1780 char *realobj;
1781 int size, i;
1782 int lines = 0;
1783
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001784 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001785 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001787 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001789 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 exp = POISON_END;
1791 if (realobj[i] != exp) {
1792 int limit;
1793 /* Mismatch ! */
1794 /* Print header */
1795 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001796 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001797 "Slab corruption (%s): %s start=%p, len=%d\n",
1798 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 print_objinfo(cachep, objp, 0);
1800 }
1801 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001802 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001804 if (i + limit > size)
1805 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806 dump_line(realobj, i, limit);
1807 i += 16;
1808 lines++;
1809 /* Limit to 5 lines */
1810 if (lines > 5)
1811 break;
1812 }
1813 }
1814 if (lines != 0) {
1815 /* Print some data about the neighboring objects, if they
1816 * exist:
1817 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001818 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001819 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820
Joonsoo Kim8456a642013-10-24 10:07:49 +09001821 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001823 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001824 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001826 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 print_objinfo(cachep, objp, 2);
1828 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001829 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001830 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001831 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001833 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 print_objinfo(cachep, objp, 2);
1835 }
1836 }
1837}
1838#endif
1839
Linus Torvalds1da177e2005-04-16 15:20:36 -07001840#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001841static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1842 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001843{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 int i;
1845 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001846 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847
1848 if (cachep->flags & SLAB_POISON) {
1849#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001850 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001851 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001852 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001853 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 else
1855 check_poison_obj(cachep, objp);
1856#else
1857 check_poison_obj(cachep, objp);
1858#endif
1859 }
1860 if (cachep->flags & SLAB_RED_ZONE) {
1861 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1862 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1865 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001866 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001869}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001871static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1872 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001873{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001874}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875#endif
1876
Randy Dunlap911851e2006-03-22 00:08:14 -08001877/**
1878 * slab_destroy - destroy and release all objects in a slab
1879 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001880 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001881 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001882 * Destroy all the objs in a slab page, and release the mem back to the system.
1883 * Before calling the slab page must have been unlinked from the cache. The
1884 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001885 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001886static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001887{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001888 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001889
Joonsoo Kim8456a642013-10-24 10:07:49 +09001890 freelist = page->freelist;
1891 slab_destroy_debugcheck(cachep, page);
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001892 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1893 call_rcu(&page->rcu_head, kmem_rcu_free);
1894 else
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001895 kmem_freepages(cachep, page);
Joonsoo Kim68126702013-10-24 10:07:42 +09001896
1897 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001898 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001899 * although actual page can be freed in rcu context
1900 */
1901 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001902 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903}
1904
Joonsoo Kim97654df2014-08-06 16:04:25 -07001905static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1906{
1907 struct page *page, *n;
1908
1909 list_for_each_entry_safe(page, n, list, lru) {
1910 list_del(&page->lru);
1911 slab_destroy(cachep, page);
1912 }
1913}
1914
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001916 * calculate_slab_order - calculate size (page order) of slabs
1917 * @cachep: pointer to the cache that is being created
1918 * @size: size of objects to be created in this cache.
1919 * @align: required alignment for the objects.
1920 * @flags: slab allocation flags
1921 *
1922 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001923 *
1924 * This could be made much more intelligent. For now, try to avoid using
1925 * high order pages for slabs. When the gfp() functions are more friendly
1926 * towards high-order requests, this should be changed.
1927 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001928static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001929 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001930{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001931 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001932 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001933 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001934
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001935 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001936 unsigned int num;
1937 size_t remainder;
1938
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001939 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001940 if (!num)
1941 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001942
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001943 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1944 if (num > SLAB_OBJ_MAX_NUM)
1945 break;
1946
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001947 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim03787302014-06-23 13:22:06 -07001948 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001949 /*
1950 * Max number of objs-per-slab for caches which
1951 * use off-slab slabs. Needed to avoid a possible
1952 * looping condition in cache_grow().
1953 */
Joonsoo Kim03787302014-06-23 13:22:06 -07001954 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1955 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001956 offslab_limit = size;
Joonsoo Kim03787302014-06-23 13:22:06 -07001957 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001958
1959 if (num > offslab_limit)
1960 break;
1961 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001962
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001963 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001964 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001965 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001966 left_over = remainder;
1967
1968 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001969 * A VFS-reclaimable slab tends to have most allocations
1970 * as GFP_NOFS and we really don't want to have to be allocating
1971 * higher-order pages when we are unable to shrink dcache.
1972 */
1973 if (flags & SLAB_RECLAIM_ACCOUNT)
1974 break;
1975
1976 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001977 * Large number of objects is good, but very large slabs are
1978 * currently bad for the gfp()s.
1979 */
David Rientjes543585c2011-10-18 22:09:24 -07001980 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001981 break;
1982
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001983 /*
1984 * Acceptable internal fragmentation?
1985 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001986 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001987 break;
1988 }
1989 return left_over;
1990}
1991
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001992static struct array_cache __percpu *alloc_kmem_cache_cpus(
1993 struct kmem_cache *cachep, int entries, int batchcount)
1994{
1995 int cpu;
1996 size_t size;
1997 struct array_cache __percpu *cpu_cache;
1998
1999 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07002000 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002001
2002 if (!cpu_cache)
2003 return NULL;
2004
2005 for_each_possible_cpu(cpu) {
2006 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2007 entries, batchcount);
2008 }
2009
2010 return cpu_cache;
2011}
2012
Pekka Enberg83b519e2009-06-10 19:40:04 +03002013static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002014{
Christoph Lameter97d06602012-07-06 15:25:11 -05002015 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002016 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002017
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002018 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2019 if (!cachep->cpu_cache)
2020 return 1;
2021
Christoph Lameter97d06602012-07-06 15:25:11 -05002022 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002023 /* Creation of first cache (kmem_cache). */
2024 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002025 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002026 /* For kmem_cache_node */
2027 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002028 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002029 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002030
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002031 for_each_online_node(node) {
2032 cachep->node[node] = kmalloc_node(
2033 sizeof(struct kmem_cache_node), gfp, node);
2034 BUG_ON(!cachep->node[node]);
2035 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002036 }
2037 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002038
Christoph Lameter6a673682013-01-10 19:14:19 +00002039 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002040 jiffies + REAPTIMEOUT_NODE +
2041 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002042
2043 cpu_cache_get(cachep)->avail = 0;
2044 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2045 cpu_cache_get(cachep)->batchcount = 1;
2046 cpu_cache_get(cachep)->touched = 0;
2047 cachep->batchcount = 1;
2048 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002049 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002050}
2051
Joonsoo Kim12220de2014-10-09 15:26:24 -07002052unsigned long kmem_cache_flags(unsigned long object_size,
2053 unsigned long flags, const char *name,
2054 void (*ctor)(void *))
2055{
2056 return flags;
2057}
2058
2059struct kmem_cache *
2060__kmem_cache_alias(const char *name, size_t size, size_t align,
2061 unsigned long flags, void (*ctor)(void *))
2062{
2063 struct kmem_cache *cachep;
2064
2065 cachep = find_mergeable(size, align, flags, name, ctor);
2066 if (cachep) {
2067 cachep->refcount++;
2068
2069 /*
2070 * Adjust the object sizes so that we clear
2071 * the complete object on kzalloc.
2072 */
2073 cachep->object_size = max_t(int, cachep->object_size, size);
2074 }
2075 return cachep;
2076}
2077
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002078/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002079 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002080 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 *
2083 * Returns a ptr to the cache on success, NULL on failure.
2084 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002085 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 * The flags are
2088 *
2089 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2090 * to catch references to uninitialised memory.
2091 *
2092 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2093 * for buffer overruns.
2094 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2096 * cacheline. This can be beneficial if you're counting cycles as closely
2097 * as davem.
2098 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002099int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002100__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002102 size_t left_over, freelist_size;
2103 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002104 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002105 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002106 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109#if FORCED_DEBUG
2110 /*
2111 * Enable redzoning and last user accounting, except for caches with
2112 * large objects, if the increased size would increase the object size
2113 * above the next power of two: caches with object sizes just above a
2114 * power of two have a significant amount of internal fragmentation.
2115 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002116 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2117 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002118 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 if (!(flags & SLAB_DESTROY_BY_RCU))
2120 flags |= SLAB_POISON;
2121#endif
2122 if (flags & SLAB_DESTROY_BY_RCU)
2123 BUG_ON(flags & SLAB_POISON);
2124#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
Andrew Mortona737b3e2006-03-22 00:08:11 -08002126 /*
2127 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 * unaligned accesses for some archs when redzoning is used, and makes
2129 * sure any on-slab bufctl's are also correctly aligned.
2130 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002131 if (size & (BYTES_PER_WORD - 1)) {
2132 size += (BYTES_PER_WORD - 1);
2133 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 }
2135
David Woodhouse87a927c2007-07-04 21:26:44 -04002136 if (flags & SLAB_RED_ZONE) {
2137 ralign = REDZONE_ALIGN;
2138 /* If redzoning, ensure that the second redzone is suitably
2139 * aligned, by adjusting the object size accordingly. */
2140 size += REDZONE_ALIGN - 1;
2141 size &= ~(REDZONE_ALIGN - 1);
2142 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002143
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002144 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002145 if (ralign < cachep->align) {
2146 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002148 /* disable debug if necessary */
2149 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002150 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002151 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002152 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002154 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155
Pekka Enberg83b519e2009-06-10 19:40:04 +03002156 if (slab_is_available())
2157 gfp = GFP_KERNEL;
2158 else
2159 gfp = GFP_NOWAIT;
2160
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162
Pekka Enbergca5f9702006-09-25 23:31:25 -07002163 /*
2164 * Both debugging options require word-alignment which is calculated
2165 * into align above.
2166 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002169 cachep->obj_offset += sizeof(unsigned long long);
2170 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 }
2172 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002173 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002174 * the real object. But if the second red zone needs to be
2175 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002177 if (flags & SLAB_RED_ZONE)
2178 size += REDZONE_ALIGN;
2179 else
2180 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 }
2182#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002183 /*
2184 * To activate debug pagealloc, off-slab management is necessary
2185 * requirement. In early phase of initialization, small sized slab
2186 * doesn't get initialized so it would not be possible. So, we need
2187 * to check size >= 256. It guarantees that all necessary small
2188 * sized slab is initialized in current slab initialization sequence.
2189 */
2190 if (!slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2191 size >= 256 && cachep->object_size > cache_line_size() &&
2192 ALIGN(size, cachep->align) < PAGE_SIZE) {
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002193 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 size = PAGE_SIZE;
2195 }
2196#endif
2197#endif
2198
Ingo Molnare0a42722006-06-23 02:03:46 -07002199 /*
2200 * Determine if the slab management is 'on' or 'off' slab.
2201 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002202 * it too early on. Always use on-slab management when
2203 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002204 */
Catalin Marinasd4322d82015-11-05 18:45:54 -08002205 if (size >= OFF_SLAB_MIN_SIZE && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002206 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 /*
2208 * Size is large, assume best to place the slab management obj
2209 * off-slab (should allow better packing of objs).
2210 */
2211 flags |= CFLGS_OFF_SLAB;
2212
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002213 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002214 /*
2215 * We should restrict the number of objects in a slab to implement
2216 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2217 */
2218 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2219 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002221 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002223 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002224 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002225
Joonsoo Kim03787302014-06-23 13:22:06 -07002226 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227
2228 /*
2229 * If the slab has been placed off-slab, and we have enough space then
2230 * move it on-slab. This is at the expense of any extra colouring.
2231 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002232 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002234 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 }
2236
2237 if (flags & CFLGS_OFF_SLAB) {
2238 /* really off slab. No need for manual alignment */
Joonsoo Kim03787302014-06-23 13:22:06 -07002239 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302240
2241#ifdef CONFIG_PAGE_POISONING
2242 /* If we're going to use the generic kernel_map_pages()
2243 * poisoning, then it's going to smash the contents of
2244 * the redzone and userword anyhow, so switch them off.
2245 */
2246 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2247 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2248#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 }
2250
2251 cachep->colour_off = cache_line_size();
2252 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002253 if (cachep->colour_off < cachep->align)
2254 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002255 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002256 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002258 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002259 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002260 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002261 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002262 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002264 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002265 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002266 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002267 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002268 * But since we go off slab only for object size greater than
Catalin Marinasd4322d82015-11-05 18:45:54 -08002269 * OFF_SLAB_MIN_SIZE, and kmalloc_{dma,}_caches get created
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002270 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002271 * But leave a BUG_ON for some lucky dude.
2272 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002273 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002274 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002276 err = setup_cpu_cache(cachep, gfp);
2277 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002278 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002279 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002280 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002282 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284
2285#if DEBUG
2286static void check_irq_off(void)
2287{
2288 BUG_ON(!irqs_disabled());
2289}
2290
2291static void check_irq_on(void)
2292{
2293 BUG_ON(irqs_disabled());
2294}
2295
Pekka Enberg343e0d72006-02-01 03:05:50 -08002296static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297{
2298#ifdef CONFIG_SMP
2299 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002300 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301#endif
2302}
Christoph Lametere498be72005-09-09 13:03:32 -07002303
Pekka Enberg343e0d72006-02-01 03:05:50 -08002304static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002305{
2306#ifdef CONFIG_SMP
2307 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002308 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002309#endif
2310}
2311
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312#else
2313#define check_irq_off() do { } while(0)
2314#define check_irq_on() do { } while(0)
2315#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002316#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317#endif
2318
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002319static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002320 struct array_cache *ac,
2321 int force, int node);
2322
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323static void do_drain(void *arg)
2324{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002325 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002327 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002328 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002329 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330
2331 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002332 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002333 n = get_node(cachep, node);
2334 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002335 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002336 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002337 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 ac->avail = 0;
2339}
2340
Pekka Enberg343e0d72006-02-01 03:05:50 -08002341static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002343 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002344 int node;
2345
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002346 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002348 for_each_kmem_cache_node(cachep, node, n)
2349 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002350 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002351
Christoph Lameter18bf8542014-08-06 16:04:11 -07002352 for_each_kmem_cache_node(cachep, node, n)
2353 drain_array(cachep, n, n->shared, 1, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354}
2355
Christoph Lametered11d9e2006-06-30 01:55:45 -07002356/*
2357 * Remove slabs from the list of free slabs.
2358 * Specify the number of slabs to drain in tofree.
2359 *
2360 * Returns the actual number of slabs released.
2361 */
2362static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002363 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002365 struct list_head *p;
2366 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002367 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
Christoph Lametered11d9e2006-06-30 01:55:45 -07002369 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002370 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002372 spin_lock_irq(&n->list_lock);
2373 p = n->slabs_free.prev;
2374 if (p == &n->slabs_free) {
2375 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002376 goto out;
2377 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378
Joonsoo Kim8456a642013-10-24 10:07:49 +09002379 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002381 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002383 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002384 /*
2385 * Safe to drop the lock. The slab is no longer linked
2386 * to the cache.
2387 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002388 n->free_objects -= cache->num;
2389 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002390 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002391 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002393out:
2394 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395}
2396
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002397int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
Christoph Lametere498be72005-09-09 13:03:32 -07002398{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002399 int ret = 0;
2400 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002401 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002402
2403 drain_cpu_caches(cachep);
2404
2405 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002406 for_each_kmem_cache_node(cachep, node, n) {
Wanpeng Li0fa81032013-07-04 08:33:22 +08002407 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002408
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002409 ret += !list_empty(&n->slabs_full) ||
2410 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002411 }
2412 return (ret ? 1 : 0);
2413}
2414
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002415int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416{
Christoph Lameter12c36672012-09-04 23:38:33 +00002417 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002418 struct kmem_cache_node *n;
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002419 int rc = __kmem_cache_shrink(cachep, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420
Christoph Lameter12c36672012-09-04 23:38:33 +00002421 if (rc)
2422 return rc;
2423
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002424 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002425
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002426 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002427 for_each_kmem_cache_node(cachep, i, n) {
2428 kfree(n->shared);
2429 free_alien_cache(n->alien);
2430 kfree(n);
2431 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002433 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002436/*
2437 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002438 *
2439 * For a slab cache when the slab descriptor is off-slab, the
2440 * slab descriptor can't come from the same cache which is being created,
2441 * Because if it is the case, that means we defer the creation of
2442 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2443 * And we eventually call down to __kmem_cache_create(), which
2444 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2445 * This is a "chicken-and-egg" problem.
2446 *
2447 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2448 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002449 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002450static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002451 struct page *page, int colour_off,
2452 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002454 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002455 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002456
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 if (OFF_SLAB(cachep)) {
2458 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002459 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002460 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002461 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 return NULL;
2463 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002464 freelist = addr + colour_off;
2465 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002467 page->active = 0;
2468 page->s_mem = addr + colour_off;
2469 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470}
2471
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002472static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002474 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002475}
2476
2477static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002478 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002479{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002480 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481}
2482
Pekka Enberg343e0d72006-02-01 03:05:50 -08002483static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002484 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485{
2486 int i;
2487
2488 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002489 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490#if DEBUG
2491 /* need to poison the objs? */
2492 if (cachep->flags & SLAB_POISON)
2493 poison_obj(cachep, objp, POISON_FREE);
2494 if (cachep->flags & SLAB_STORE_USER)
2495 *dbg_userword(cachep, objp) = NULL;
2496
2497 if (cachep->flags & SLAB_RED_ZONE) {
2498 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2499 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2500 }
2501 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002502 * Constructors are not allowed to allocate memory from the same
2503 * cache which they are a constructor for. Otherwise, deadlock.
2504 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505 */
2506 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002507 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508
2509 if (cachep->flags & SLAB_RED_ZONE) {
2510 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2511 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002512 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2514 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002515 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002517 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002518 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002519 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002520 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521#else
2522 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002523 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524#endif
Joonsoo Kim03787302014-06-23 13:22:06 -07002525 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002526 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528}
2529
Pekka Enberg343e0d72006-02-01 03:05:50 -08002530static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002532 if (CONFIG_ZONE_DMA_FLAG) {
2533 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002534 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002535 else
Glauber Costaa618e892012-06-14 16:17:21 +04002536 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538}
2539
Joonsoo Kim8456a642013-10-24 10:07:49 +09002540static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002541 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002542{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002543 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002544
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002545 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002546 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002547#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002548 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002549#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002550
2551 return objp;
2552}
2553
Joonsoo Kim8456a642013-10-24 10:07:49 +09002554static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002555 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002556{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002557 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002558#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002559 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002560
Matthew Dobson78d382d2006-02-01 03:05:47 -08002561 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002562 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002563
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002564 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002565 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002566 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002567 printk(KERN_ERR "slab: double free detected in cache "
2568 "'%s', objp %p\n", cachep->name, objp);
2569 BUG();
2570 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002571 }
2572#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002573 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002574 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002575}
2576
Pekka Enberg47768742006-06-23 02:03:07 -07002577/*
2578 * Map pages beginning at addr to the given cache and slab. This is required
2579 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002580 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002581 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002582static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002583 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002585 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002586 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587}
2588
2589/*
2590 * Grow (by 1) the number of slabs within a cache. This is called by
2591 * kmem_cache_alloc() when there are no active objs left in a cache.
2592 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002593static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002594 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002596 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002597 size_t offset;
2598 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002599 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600
Andrew Mortona737b3e2006-03-22 00:08:11 -08002601 /*
2602 * Be lazy and only check for valid flags here, keeping it out of the
2603 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002605 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2606 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2607 BUG();
2608 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002609 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002611 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002613 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002614 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615
2616 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002617 offset = n->colour_next;
2618 n->colour_next++;
2619 if (n->colour_next >= cachep->colour)
2620 n->colour_next = 0;
2621 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002623 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624
Mel Gormand0164ad2015-11-06 16:28:21 -08002625 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 local_irq_enable();
2627
2628 /*
2629 * The test for missing atomic flag is performed here, rather than
2630 * the more obvious place, simply to reduce the critical path length
2631 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2632 * will eventually be caught here (where it matters).
2633 */
2634 kmem_flagcheck(cachep, flags);
2635
Andrew Mortona737b3e2006-03-22 00:08:11 -08002636 /*
2637 * Get mem for the objs. Attempt to allocate a physical page from
2638 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002639 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002640 if (!page)
2641 page = kmem_getpages(cachep, local_flags, nodeid);
2642 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 goto failed;
2644
2645 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002646 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002647 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002648 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 goto opps1;
2650
Joonsoo Kim8456a642013-10-24 10:07:49 +09002651 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652
Joonsoo Kim8456a642013-10-24 10:07:49 +09002653 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654
Mel Gormand0164ad2015-11-06 16:28:21 -08002655 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 local_irq_disable();
2657 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002658 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
2660 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002661 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002663 n->free_objects += cachep->num;
2664 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002666opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002667 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002668failed:
Mel Gormand0164ad2015-11-06 16:28:21 -08002669 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 local_irq_disable();
2671 return 0;
2672}
2673
2674#if DEBUG
2675
2676/*
2677 * Perform extra freeing checks:
2678 * - detect bad pointers.
2679 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 */
2681static void kfree_debugcheck(const void *objp)
2682{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 if (!virt_addr_valid(objp)) {
2684 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002685 (unsigned long)objp);
2686 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688}
2689
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002690static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2691{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002692 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002693
2694 redzone1 = *dbg_redzone1(cache, obj);
2695 redzone2 = *dbg_redzone2(cache, obj);
2696
2697 /*
2698 * Redzone is ok.
2699 */
2700 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2701 return;
2702
2703 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2704 slab_error(cache, "double free detected");
2705 else
2706 slab_error(cache, "memory outside object was overwritten");
2707
David Woodhouseb46b8f12007-05-08 00:22:59 -07002708 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002709 obj, redzone1, redzone2);
2710}
2711
Pekka Enberg343e0d72006-02-01 03:05:50 -08002712static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002713 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002716 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002718 BUG_ON(virt_to_cache(objp) != cachep);
2719
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002720 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002722 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002725 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2727 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2728 }
2729 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002730 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731
Joonsoo Kim8456a642013-10-24 10:07:49 +09002732 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
2734 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002735 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736
Joonsoo Kim03787302014-06-23 13:22:06 -07002737 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 if (cachep->flags & SLAB_POISON) {
2739#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002740 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002741 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002742 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002743 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 } else {
2745 poison_obj(cachep, objp, POISON_FREE);
2746 }
2747#else
2748 poison_obj(cachep, objp, POISON_FREE);
2749#endif
2750 }
2751 return objp;
2752}
2753
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754#else
2755#define kfree_debugcheck(x) do { } while(0)
2756#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757#endif
2758
Mel Gorman072bb0a2012-07-31 16:43:58 -07002759static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2760 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761{
2762 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002763 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002765 int node;
2766
Joe Korty6d2144d2008-03-05 15:04:59 -08002767 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002768 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002769 if (unlikely(force_refill))
2770 goto force_grow;
2771retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002772 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 batchcount = ac->batchcount;
2774 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002775 /*
2776 * If there was little recent activity on this cache, then
2777 * perform only a partial refill. Otherwise we could generate
2778 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 */
2780 batchcount = BATCHREFILL_LIMIT;
2781 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07002782 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002784 BUG_ON(ac->avail > 0 || !n);
2785 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002786
Christoph Lameter3ded1752006-03-25 03:06:44 -08002787 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002788 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2789 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002790 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002791 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002792
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 while (batchcount > 0) {
2794 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002795 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002797 entry = n->slabs_partial.next;
2798 if (entry == &n->slabs_partial) {
2799 n->free_touched = 1;
2800 entry = n->slabs_free.next;
2801 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 goto must_grow;
2803 }
2804
Joonsoo Kim8456a642013-10-24 10:07:49 +09002805 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002807
2808 /*
2809 * The slab was either on partial or free list so
2810 * there must be at least one object available for
2811 * allocation.
2812 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002813 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002814
Joonsoo Kim8456a642013-10-24 10:07:49 +09002815 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 STATS_INC_ALLOCED(cachep);
2817 STATS_INC_ACTIVE(cachep);
2818 STATS_SET_HIGH(cachep);
2819
Joonsoo Kim8456a642013-10-24 10:07:49 +09002820 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002821 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
2824 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002825 list_del(&page->lru);
2826 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002827 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002829 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 }
2831
Andrew Mortona737b3e2006-03-22 00:08:11 -08002832must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002833 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002834alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002835 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836
2837 if (unlikely(!ac->avail)) {
2838 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002839force_grow:
David Rientjes4167e9b2015-04-14 15:46:55 -07002840 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002841
Andrew Mortona737b3e2006-03-22 00:08:11 -08002842 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002843 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002844 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002845
2846 /* no objects in sight? abort */
2847 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 return NULL;
2849
Andrew Mortona737b3e2006-03-22 00:08:11 -08002850 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 goto retry;
2852 }
2853 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002854
2855 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856}
2857
Andrew Mortona737b3e2006-03-22 00:08:11 -08002858static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2859 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860{
Mel Gormand0164ad2015-11-06 16:28:21 -08002861 might_sleep_if(gfpflags_allow_blocking(flags));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862#if DEBUG
2863 kmem_flagcheck(cachep, flags);
2864#endif
2865}
2866
2867#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002868static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002869 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870{
Joonsoo Kim03787302014-06-23 13:22:06 -07002871 struct page *page;
2872
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002873 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002875 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002877 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002878 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002879 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 else
2881 check_poison_obj(cachep, objp);
2882#else
2883 check_poison_obj(cachep, objp);
2884#endif
2885 poison_obj(cachep, objp, POISON_INUSE);
2886 }
2887 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002888 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889
2890 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002891 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2892 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2893 slab_error(cachep, "double free, or memory outside"
2894 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002895 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002896 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002897 objp, *dbg_redzone1(cachep, objp),
2898 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899 }
2900 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2901 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2902 }
Joonsoo Kim03787302014-06-23 13:22:06 -07002903
2904 page = virt_to_head_page(objp);
2905 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002906 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002907 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002908 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002909 if (ARCH_SLAB_MINALIGN &&
2910 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002911 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002912 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002913 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 return objp;
2915}
2916#else
2917#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2918#endif
2919
Akinobu Mita773ff602008-12-23 19:37:01 +09002920static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002921{
Joonsoo Kim8a9c61d2014-08-06 16:04:20 -07002922 if (unlikely(cachep == kmem_cache))
Akinobu Mita773ff602008-12-23 19:37:01 +09002923 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002924
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002925 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002926}
2927
Pekka Enberg343e0d72006-02-01 03:05:50 -08002928static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002930 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002932 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933
Alok N Kataria5c382302005-09-27 21:45:46 -07002934 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002935
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002936 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002939 objp = ac_get_obj(cachep, ac, flags, false);
2940
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002941 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07002942 * Allow for the possibility all avail objects are not allowed
2943 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002944 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07002945 if (objp) {
2946 STATS_INC_ALLOCHIT(cachep);
2947 goto out;
2948 }
2949 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07002951
2952 STATS_INC_ALLOCMISS(cachep);
2953 objp = cache_alloc_refill(cachep, flags, force_refill);
2954 /*
2955 * the 'ac' may be updated by cache_alloc_refill(),
2956 * and kmemleak_erase() requires its correct value.
2957 */
2958 ac = cpu_cache_get(cachep);
2959
2960out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01002961 /*
2962 * To avoid a false negative, if an object that is in one of the
2963 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2964 * treat the array pointers as a reference to the object.
2965 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09002966 if (objp)
2967 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07002968 return objp;
2969}
2970
Christoph Lametere498be72005-09-09 13:03:32 -07002971#ifdef CONFIG_NUMA
2972/*
Zefan Li2ad654b2014-09-25 09:41:02 +08002973 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002974 *
2975 * If we are in_interrupt, then process context, including cpusets and
2976 * mempolicy, may not apply and should not be used for allocation policy.
2977 */
2978static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2979{
2980 int nid_alloc, nid_here;
2981
Christoph Lameter765c4502006-09-27 01:50:08 -07002982 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08002983 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002984 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002985 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07002986 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002987 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07002988 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002989 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08002990 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08002991 return NULL;
2992}
2993
2994/*
Christoph Lameter765c4502006-09-27 01:50:08 -07002995 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08002996 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00002997 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08002998 * perform an allocation without specifying a node. This allows the page
2999 * allocator to do its reclaim / fallback magic. We then insert the
3000 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003001 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003002static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003003{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003004 struct zonelist *zonelist;
3005 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003006 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003007 struct zone *zone;
3008 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003009 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003010 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003011 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003012
3013 if (flags & __GFP_THISNODE)
3014 return NULL;
3015
Christoph Lameter6cb06222007-10-16 01:25:41 -07003016 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003017
Mel Gormancc9a6c82012-03-21 16:34:11 -07003018retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003019 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003020 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003021
Christoph Lameter3c517a62006-12-06 20:33:29 -08003022retry:
3023 /*
3024 * Look through allowed nodes for objects available
3025 * from existing per node queues.
3026 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003027 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3028 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003029
Vladimir Davydov061d7072014-12-12 16:58:25 -08003030 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003031 get_node(cache, nid) &&
3032 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003033 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003034 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003035 if (obj)
3036 break;
3037 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003038 }
3039
Christoph Lametercfce6602007-05-06 14:50:17 -07003040 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003041 /*
3042 * This allocation will be performed within the constraints
3043 * of the current cpuset / memory policy requirements.
3044 * We may trigger various forms of reclaim on the allowed
3045 * set and go into memory reserves if necessary.
3046 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003047 struct page *page;
3048
Mel Gormand0164ad2015-11-06 16:28:21 -08003049 if (gfpflags_allow_blocking(local_flags))
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003050 local_irq_enable();
3051 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003052 page = kmem_getpages(cache, local_flags, numa_mem_id());
Mel Gormand0164ad2015-11-06 16:28:21 -08003053 if (gfpflags_allow_blocking(local_flags))
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003054 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003055 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003056 /*
3057 * Insert into the appropriate per node queues
3058 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003059 nid = page_to_nid(page);
3060 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003061 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003062 gfp_exact_node(flags), nid);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003063 if (!obj)
3064 /*
3065 * Another processor may allocate the
3066 * objects in the slab since we are
3067 * not holding any locks.
3068 */
3069 goto retry;
3070 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003071 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003072 obj = NULL;
3073 }
3074 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003075 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003076
Mel Gormand26914d2014-04-03 14:47:24 -07003077 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003078 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003079 return obj;
3080}
3081
3082/*
Christoph Lametere498be72005-09-09 13:03:32 -07003083 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003085static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003086 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003087{
3088 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003089 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003090 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003092 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003094 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003095 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003096 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003097
Andrew Mortona737b3e2006-03-22 00:08:11 -08003098retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003099 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003100 spin_lock(&n->list_lock);
3101 entry = n->slabs_partial.next;
3102 if (entry == &n->slabs_partial) {
3103 n->free_touched = 1;
3104 entry = n->slabs_free.next;
3105 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003106 goto must_grow;
3107 }
Christoph Lametere498be72005-09-09 13:03:32 -07003108
Joonsoo Kim8456a642013-10-24 10:07:49 +09003109 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003110 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003111
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003112 STATS_INC_NODEALLOCS(cachep);
3113 STATS_INC_ACTIVE(cachep);
3114 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003115
Joonsoo Kim8456a642013-10-24 10:07:49 +09003116 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003117
Joonsoo Kim8456a642013-10-24 10:07:49 +09003118 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003119 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003120 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003121 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003122
Joonsoo Kim8456a642013-10-24 10:07:49 +09003123 if (page->active == cachep->num)
3124 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003125 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003126 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003127
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003128 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003129 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003130
Andrew Mortona737b3e2006-03-22 00:08:11 -08003131must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003132 spin_unlock(&n->list_lock);
David Rientjes4167e9b2015-04-14 15:46:55 -07003133 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003134 if (x)
3135 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003136
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003137 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003138
Andrew Mortona737b3e2006-03-22 00:08:11 -08003139done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003140 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003141}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003142
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003143static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003144slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003145 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003146{
3147 unsigned long save_flags;
3148 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003149 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003150
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003151 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003152
Nick Piggincf40bd12009-01-21 08:12:39 +01003153 lockdep_trace_alloc(flags);
3154
Akinobu Mita773ff602008-12-23 19:37:01 +09003155 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003156 return NULL;
3157
Glauber Costad79923f2012-12-18 14:22:48 -08003158 cachep = memcg_kmem_get_cache(cachep, flags);
3159
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003160 cache_alloc_debugcheck_before(cachep, flags);
3161 local_irq_save(save_flags);
3162
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003163 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003164 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003165
Christoph Lameter18bf8542014-08-06 16:04:11 -07003166 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003167 /* Node not bootstrapped yet */
3168 ptr = fallback_alloc(cachep, flags);
3169 goto out;
3170 }
3171
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003172 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003173 /*
3174 * Use the locally cached objects if possible.
3175 * However ____cache_alloc does not allow fallback
3176 * to other nodes. It may fail while we still have
3177 * objects on other nodes available.
3178 */
3179 ptr = ____cache_alloc(cachep, flags);
3180 if (ptr)
3181 goto out;
3182 }
3183 /* ___cache_alloc_node can fall back to other nodes */
3184 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3185 out:
3186 local_irq_restore(save_flags);
3187 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003188 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003189 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003190
Joe Perches5087c822013-09-10 17:02:51 -07003191 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003192 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003193 if (unlikely(flags & __GFP_ZERO))
3194 memset(ptr, 0, cachep->object_size);
3195 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003196
Vladimir Davydov8135be52014-12-12 16:56:38 -08003197 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003198 return ptr;
3199}
3200
3201static __always_inline void *
3202__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3203{
3204 void *objp;
3205
Zefan Li2ad654b2014-09-25 09:41:02 +08003206 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003207 objp = alternate_node_alloc(cache, flags);
3208 if (objp)
3209 goto out;
3210 }
3211 objp = ____cache_alloc(cache, flags);
3212
3213 /*
3214 * We may just have run out of memory on the local node.
3215 * ____cache_alloc_node() knows how to locate memory on other nodes
3216 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003217 if (!objp)
3218 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003219
3220 out:
3221 return objp;
3222}
3223#else
3224
3225static __always_inline void *
3226__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3227{
3228 return ____cache_alloc(cachep, flags);
3229}
3230
3231#endif /* CONFIG_NUMA */
3232
3233static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003234slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003235{
3236 unsigned long save_flags;
3237 void *objp;
3238
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003239 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003240
Nick Piggincf40bd12009-01-21 08:12:39 +01003241 lockdep_trace_alloc(flags);
3242
Akinobu Mita773ff602008-12-23 19:37:01 +09003243 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003244 return NULL;
3245
Glauber Costad79923f2012-12-18 14:22:48 -08003246 cachep = memcg_kmem_get_cache(cachep, flags);
3247
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003248 cache_alloc_debugcheck_before(cachep, flags);
3249 local_irq_save(save_flags);
3250 objp = __do_cache_alloc(cachep, flags);
3251 local_irq_restore(save_flags);
3252 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003253 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003254 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003255 prefetchw(objp);
3256
Joe Perches5087c822013-09-10 17:02:51 -07003257 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003258 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003259 if (unlikely(flags & __GFP_ZERO))
3260 memset(objp, 0, cachep->object_size);
3261 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003262
Vladimir Davydov8135be52014-12-12 16:56:38 -08003263 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003264 return objp;
3265}
Christoph Lametere498be72005-09-09 13:03:32 -07003266
3267/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003268 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003269 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003270 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003271static void free_block(struct kmem_cache *cachep, void **objpp,
3272 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273{
3274 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003275 struct kmem_cache_node *n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276
3277 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003278 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003279 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280
Mel Gorman072bb0a2012-07-31 16:43:58 -07003281 clear_obj_pfmemalloc(&objpp[i]);
3282 objp = objpp[i];
3283
Joonsoo Kim8456a642013-10-24 10:07:49 +09003284 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003285 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003286 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003287 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003289 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
3291 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003292 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003293 if (n->free_objects > n->free_limit) {
3294 n->free_objects -= cachep->num;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003295 list_add_tail(&page->lru, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003297 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 }
3299 } else {
3300 /* Unconditionally move a slab to the end of the
3301 * partial list on free - maximum time for the
3302 * other objects to be freed, too.
3303 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003304 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 }
3306 }
3307}
3308
Pekka Enberg343e0d72006-02-01 03:05:50 -08003309static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310{
3311 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003312 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003313 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003314 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315
3316 batchcount = ac->batchcount;
3317#if DEBUG
3318 BUG_ON(!batchcount || batchcount > ac->avail);
3319#endif
3320 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003321 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003322 spin_lock(&n->list_lock);
3323 if (n->shared) {
3324 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003325 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 if (max) {
3327 if (batchcount > max)
3328 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003329 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003330 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 shared_array->avail += batchcount;
3332 goto free_done;
3333 }
3334 }
3335
Joonsoo Kim97654df2014-08-06 16:04:25 -07003336 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003337free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003338#if STATS
3339 {
3340 int i = 0;
3341 struct list_head *p;
3342
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003343 p = n->slabs_free.next;
3344 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003345 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346
Joonsoo Kim8456a642013-10-24 10:07:49 +09003347 page = list_entry(p, struct page, lru);
3348 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349
3350 i++;
3351 p = p->next;
3352 }
3353 STATS_SET_FREEABLE(cachep, i);
3354 }
3355#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003356 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003357 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003359 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360}
3361
3362/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003363 * Release an obj back to its cache. If the obj has a constructed state, it must
3364 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003366static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003367 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003369 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
3371 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003372 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003373 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003375 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003376
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003377 /*
3378 * Skip calling cache_free_alien() when the platform is not numa.
3379 * This will avoid cache misses that happen while accessing slabp (which
3380 * is per page memory reference) to get nodeid. Instead use a global
3381 * variable to skip the call, which is mostly likely to be present in
3382 * the cache.
3383 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003384 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003385 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003386
Joonsoo Kim3d880192014-10-09 15:26:04 -07003387 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389 } else {
3390 STATS_INC_FREEMISS(cachep);
3391 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003393
Mel Gorman072bb0a2012-07-31 16:43:58 -07003394 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395}
3396
3397/**
3398 * kmem_cache_alloc - Allocate an object
3399 * @cachep: The cache to allocate from.
3400 * @flags: See kmalloc().
3401 *
3402 * Allocate an object from this cache. The flags are only relevant
3403 * if the cache has no available objects.
3404 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003405void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003407 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003408
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003409 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003410 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003411
3412 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413}
3414EXPORT_SYMBOL(kmem_cache_alloc);
3415
Christoph Lameter484748f2015-09-04 15:45:34 -07003416void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3417{
3418 __kmem_cache_free_bulk(s, size, p);
3419}
3420EXPORT_SYMBOL(kmem_cache_free_bulk);
3421
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08003422int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
Christoph Lameter484748f2015-09-04 15:45:34 -07003423 void **p)
3424{
3425 return __kmem_cache_alloc_bulk(s, flags, size, p);
3426}
3427EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3428
Li Zefan0f24f122009-12-11 15:45:30 +08003429#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003430void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003431kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003432{
Steven Rostedt85beb582010-11-24 16:23:34 -05003433 void *ret;
3434
Ezequiel Garcia48356302012-09-08 17:47:57 -03003435 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003436
3437 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003438 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003439 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003440}
Steven Rostedt85beb582010-11-24 16:23:34 -05003441EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003442#endif
3443
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003445/**
3446 * kmem_cache_alloc_node - Allocate an object on the specified node
3447 * @cachep: The cache to allocate from.
3448 * @flags: See kmalloc().
3449 * @nodeid: node number of the target node.
3450 *
3451 * Identical to kmem_cache_alloc but it will allocate memory on the given
3452 * node, which can improve the performance for cpu bound structures.
3453 *
3454 * Fallback to other node is possible if __GFP_THISNODE is not set.
3455 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003456void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3457{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003458 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003459
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003460 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003461 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003462 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003463
3464 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003465}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466EXPORT_SYMBOL(kmem_cache_alloc_node);
3467
Li Zefan0f24f122009-12-11 15:45:30 +08003468#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003469void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003470 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003471 int nodeid,
3472 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003473{
Steven Rostedt85beb582010-11-24 16:23:34 -05003474 void *ret;
3475
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003476 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003477
Steven Rostedt85beb582010-11-24 16:23:34 -05003478 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003479 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003480 flags, nodeid);
3481 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003482}
Steven Rostedt85beb582010-11-24 16:23:34 -05003483EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003484#endif
3485
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003486static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003487__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003488{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003489 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003490
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003491 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003492 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3493 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003494 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003495}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003496
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003497void *__kmalloc_node(size_t size, gfp_t flags, int node)
3498{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003499 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003500}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003501EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003502
3503void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003504 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003505{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003506 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003507}
3508EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003509#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510
3511/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003512 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003514 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003515 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003517static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003518 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003520 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003521 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003523 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003524 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3525 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003526 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003527
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003528 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003529 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003530
3531 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003532}
3533
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003534void *__kmalloc(size_t size, gfp_t flags)
3535{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003536 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537}
3538EXPORT_SYMBOL(__kmalloc);
3539
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003540void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003541{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003542 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003543}
3544EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003545
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546/**
3547 * kmem_cache_free - Deallocate an object
3548 * @cachep: The cache the allocation was from.
3549 * @objp: The previously allocated object.
3550 *
3551 * Free an object which was previously allocated from this
3552 * cache.
3553 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003554void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555{
3556 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003557 cachep = cache_from_obj(cachep, objp);
3558 if (!cachep)
3559 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560
3561 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003562 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003563 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003564 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003565 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003567
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003568 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569}
3570EXPORT_SYMBOL(kmem_cache_free);
3571
3572/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 * kfree - free previously allocated memory
3574 * @objp: pointer returned by kmalloc.
3575 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003576 * If @objp is NULL, no operation is performed.
3577 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 * Don't free memory not originally allocated by kmalloc()
3579 * or you will run into trouble.
3580 */
3581void kfree(const void *objp)
3582{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003583 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 unsigned long flags;
3585
Pekka Enberg2121db72009-03-25 11:05:57 +02003586 trace_kfree(_RET_IP_, objp);
3587
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003588 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 return;
3590 local_irq_save(flags);
3591 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003592 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003593 debug_check_no_locks_freed(objp, c->object_size);
3594
3595 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003596 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597 local_irq_restore(flags);
3598}
3599EXPORT_SYMBOL(kfree);
3600
Christoph Lametere498be72005-09-09 13:03:32 -07003601/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003602 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003603 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003604static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003605{
3606 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003607 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003608 struct array_cache *new_shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07003609 struct alien_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003610
Mel Gorman9c09a952008-01-24 05:49:54 -08003611 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003612
LQYMGTb455def2014-12-10 15:42:13 -08003613 if (use_alien_caches) {
3614 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3615 if (!new_alien)
3616 goto fail;
3617 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003618
Eric Dumazet63109842007-05-06 14:49:28 -07003619 new_shared = NULL;
3620 if (cachep->shared) {
3621 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003622 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003623 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003624 if (!new_shared) {
3625 free_alien_cache(new_alien);
3626 goto fail;
3627 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003628 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003629
Christoph Lameter18bf8542014-08-06 16:04:11 -07003630 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003631 if (n) {
3632 struct array_cache *shared = n->shared;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003633 LIST_HEAD(list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003634
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003635 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003636
Christoph Lametercafeb022006-03-25 03:06:46 -08003637 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003638 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07003639 shared->avail, node, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003640
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003641 n->shared = new_shared;
3642 if (!n->alien) {
3643 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003644 new_alien = NULL;
3645 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003646 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003647 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003648 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003649 slabs_destroy(cachep, &list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003650 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003651 free_alien_cache(new_alien);
3652 continue;
3653 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003654 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3655 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003656 free_alien_cache(new_alien);
3657 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003658 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003659 }
Christoph Lametere498be72005-09-09 13:03:32 -07003660
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003661 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003662 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3663 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003664 n->shared = new_shared;
3665 n->alien = new_alien;
3666 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003667 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003668 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003669 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003670 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003671
Andrew Mortona737b3e2006-03-22 00:08:11 -08003672fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003673 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003674 /* Cache is not active yet. Roll back what we did */
3675 node--;
3676 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003677 n = get_node(cachep, node);
3678 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003679 kfree(n->shared);
3680 free_alien_cache(n->alien);
3681 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003682 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003683 }
3684 node--;
3685 }
3686 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003687 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003688}
3689
Christoph Lameter18004c52012-07-06 15:25:12 -05003690/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003691static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003692 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003694 struct array_cache __percpu *cpu_cache, *prev;
3695 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003697 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3698 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003699 return -ENOMEM;
3700
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003701 prev = cachep->cpu_cache;
3702 cachep->cpu_cache = cpu_cache;
3703 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003704
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 cachep->batchcount = batchcount;
3707 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003708 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003710 if (!prev)
3711 goto alloc_node;
3712
3713 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003714 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003715 int node;
3716 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003717 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003718
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003719 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003720 n = get_node(cachep, node);
3721 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003722 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003723 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003724 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003726 free_percpu(prev);
3727
3728alloc_node:
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003729 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730}
3731
Glauber Costa943a4512012-12-18 14:23:03 -08003732static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3733 int batchcount, int shared, gfp_t gfp)
3734{
3735 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003736 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003737
3738 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3739
3740 if (slab_state < FULL)
3741 return ret;
3742
3743 if ((ret < 0) || !is_root_cache(cachep))
3744 return ret;
3745
Vladimir Davydov426589f2015-02-12 14:59:23 -08003746 lockdep_assert_held(&slab_mutex);
3747 for_each_memcg_cache(c, cachep) {
3748 /* return value determined by the root cache only */
3749 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003750 }
3751
3752 return ret;
3753}
3754
Christoph Lameter18004c52012-07-06 15:25:12 -05003755/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003756static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757{
3758 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003759 int limit = 0;
3760 int shared = 0;
3761 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762
Glauber Costa943a4512012-12-18 14:23:03 -08003763 if (!is_root_cache(cachep)) {
3764 struct kmem_cache *root = memcg_root_cache(cachep);
3765 limit = root->limit;
3766 shared = root->shared;
3767 batchcount = root->batchcount;
3768 }
3769
3770 if (limit && shared && batchcount)
3771 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003772 /*
3773 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003774 * - create a LIFO ordering, i.e. return objects that are cache-warm
3775 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003776 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 * bufctl chains: array operations are cheaper.
3778 * The numbers are guessed, we should auto-tune as described by
3779 * Bonwick.
3780 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003781 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003782 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003783 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003785 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003786 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003787 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 limit = 54;
3789 else
3790 limit = 120;
3791
Andrew Mortona737b3e2006-03-22 00:08:11 -08003792 /*
3793 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 * allocation behaviour: Most allocs on one cpu, most free operations
3795 * on another cpu. For these cases, an efficient object passing between
3796 * cpus is necessary. This is provided by a shared array. The array
3797 * replaces Bonwick's magazine layer.
3798 * On uniprocessor, it's functionally equivalent (but less efficient)
3799 * to a larger limit. Thus disabled by default.
3800 */
3801 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003802 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804
3805#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003806 /*
3807 * With debugging enabled, large batchcount lead to excessively long
3808 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809 */
3810 if (limit > 32)
3811 limit = 32;
3812#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003813 batchcount = (limit + 1) / 2;
3814skip_setup:
3815 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816 if (err)
3817 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003818 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003819 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003820}
3821
Christoph Lameter1b552532006-03-22 00:09:07 -08003822/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003823 * Drain an array if it contains any elements taking the node lock only if
3824 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003825 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003826 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003827static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003828 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829{
Joonsoo Kim97654df2014-08-06 16:04:25 -07003830 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831 int tofree;
3832
Christoph Lameter1b552532006-03-22 00:09:07 -08003833 if (!ac || !ac->avail)
3834 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835 if (ac->touched && !force) {
3836 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003837 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003838 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003839 if (ac->avail) {
3840 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3841 if (tofree > ac->avail)
3842 tofree = (ac->avail + 1) / 2;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003843 free_block(cachep, ac->entry, tofree, node, &list);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003844 ac->avail -= tofree;
3845 memmove(ac->entry, &(ac->entry[tofree]),
3846 sizeof(void *) * ac->avail);
3847 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003848 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003849 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850 }
3851}
3852
3853/**
3854 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003855 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856 *
3857 * Called from workqueue/eventd every few seconds.
3858 * Purpose:
3859 * - clear the per-cpu caches for this CPU.
3860 * - return freeable pages to the main free memory pool.
3861 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003862 * If we cannot acquire the cache chain mutex then just give up - we'll try
3863 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003865static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003867 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003868 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003869 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003870 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871
Christoph Lameter18004c52012-07-06 15:25:12 -05003872 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003874 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875
Christoph Lameter18004c52012-07-06 15:25:12 -05003876 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 check_irq_on();
3878
Christoph Lameter35386e32006-03-22 00:09:05 -08003879 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003880 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003881 * have established with reasonable certainty that
3882 * we can do some work if the lock was obtained.
3883 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07003884 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08003885
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003886 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003888 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889
Christoph Lameter35386e32006-03-22 00:09:05 -08003890 /*
3891 * These are racy checks but it does not matter
3892 * if we skip one check or scan twice.
3893 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003894 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003895 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003897 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003899 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003901 if (n->free_touched)
3902 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003903 else {
3904 int freed;
3905
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003906 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07003907 5 * searchp->num - 1) / (5 * searchp->num));
3908 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003910next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 cond_resched();
3912 }
3913 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05003914 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003915 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003916out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08003917 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003918 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919}
3920
Linus Torvalds158a9622008-01-02 13:04:48 -08003921#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04003922void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003924 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003925 unsigned long active_objs;
3926 unsigned long num_objs;
3927 unsigned long active_slabs = 0;
3928 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003929 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003931 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003932 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 active_objs = 0;
3935 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07003936 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07003937
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003938 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003939 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003940
Joonsoo Kim8456a642013-10-24 10:07:49 +09003941 list_for_each_entry(page, &n->slabs_full, lru) {
3942 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07003943 error = "slabs_full accounting error";
3944 active_objs += cachep->num;
3945 active_slabs++;
3946 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003947 list_for_each_entry(page, &n->slabs_partial, lru) {
3948 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003949 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003950 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003951 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003952 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07003953 active_slabs++;
3954 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003955 list_for_each_entry(page, &n->slabs_free, lru) {
3956 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003957 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07003958 num_slabs++;
3959 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003960 free_objects += n->free_objects;
3961 if (n->shared)
3962 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003963
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003964 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003966 num_slabs += active_slabs;
3967 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003968 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 error = "free_objects accounting error";
3970
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003971 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 if (error)
3973 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3974
Glauber Costa0d7561c2012-10-19 18:20:27 +04003975 sinfo->active_objs = active_objs;
3976 sinfo->num_objs = num_objs;
3977 sinfo->active_slabs = active_slabs;
3978 sinfo->num_slabs = num_slabs;
3979 sinfo->shared_avail = shared_avail;
3980 sinfo->limit = cachep->limit;
3981 sinfo->batchcount = cachep->batchcount;
3982 sinfo->shared = cachep->shared;
3983 sinfo->objects_per_slab = cachep->num;
3984 sinfo->cache_order = cachep->gfporder;
3985}
3986
3987void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3988{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003990 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 unsigned long high = cachep->high_mark;
3992 unsigned long allocs = cachep->num_allocations;
3993 unsigned long grown = cachep->grown;
3994 unsigned long reaped = cachep->reaped;
3995 unsigned long errors = cachep->errors;
3996 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003998 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003999 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000
Joe Perchese92dd4f2010-03-26 19:27:58 -07004001 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4002 "%4lu %4lu %4lu %4lu %4lu",
4003 allocs, high, grown,
4004 reaped, errors, max_freeable, node_allocs,
4005 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006 }
4007 /* cpu stats */
4008 {
4009 unsigned long allochit = atomic_read(&cachep->allochit);
4010 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4011 unsigned long freehit = atomic_read(&cachep->freehit);
4012 unsigned long freemiss = atomic_read(&cachep->freemiss);
4013
4014 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004015 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 }
4017#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018}
4019
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020#define MAX_SLABINFO_WRITE 128
4021/**
4022 * slabinfo_write - Tuning for the slab allocator
4023 * @file: unused
4024 * @buffer: user buffer
4025 * @count: data length
4026 * @ppos: unused
4027 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004028ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004029 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004031 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004033 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004034
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 if (count > MAX_SLABINFO_WRITE)
4036 return -EINVAL;
4037 if (copy_from_user(&kbuf, buffer, count))
4038 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004039 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040
4041 tmp = strchr(kbuf, ' ');
4042 if (!tmp)
4043 return -EINVAL;
4044 *tmp = '\0';
4045 tmp++;
4046 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4047 return -EINVAL;
4048
4049 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004050 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004052 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004054 if (limit < 1 || batchcount < 1 ||
4055 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004056 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004058 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004059 batchcount, shared,
4060 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 }
4062 break;
4063 }
4064 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004065 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066 if (res >= 0)
4067 res = count;
4068 return res;
4069}
Al Viro871751e2006-03-25 03:06:39 -08004070
4071#ifdef CONFIG_DEBUG_SLAB_LEAK
4072
Al Viro871751e2006-03-25 03:06:39 -08004073static inline int add_caller(unsigned long *n, unsigned long v)
4074{
4075 unsigned long *p;
4076 int l;
4077 if (!v)
4078 return 1;
4079 l = n[1];
4080 p = n + 2;
4081 while (l) {
4082 int i = l/2;
4083 unsigned long *q = p + 2 * i;
4084 if (*q == v) {
4085 q[1]++;
4086 return 1;
4087 }
4088 if (*q > v) {
4089 l = i;
4090 } else {
4091 p = q + 2;
4092 l -= i + 1;
4093 }
4094 }
4095 if (++n[1] == n[0])
4096 return 0;
4097 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4098 p[0] = v;
4099 p[1] = 1;
4100 return 1;
4101}
4102
Joonsoo Kim8456a642013-10-24 10:07:49 +09004103static void handle_slab(unsigned long *n, struct kmem_cache *c,
4104 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004105{
4106 void *p;
Joonsoo Kim03787302014-06-23 13:22:06 -07004107 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004108
Al Viro871751e2006-03-25 03:06:39 -08004109 if (n[0] == n[1])
4110 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004111 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kim03787302014-06-23 13:22:06 -07004112 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004113 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004114
Al Viro871751e2006-03-25 03:06:39 -08004115 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4116 return;
4117 }
4118}
4119
4120static void show_symbol(struct seq_file *m, unsigned long address)
4121{
4122#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004123 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004124 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004125
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004126 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004127 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004128 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004129 seq_printf(m, " [%s]", modname);
4130 return;
4131 }
4132#endif
4133 seq_printf(m, "%p", (void *)address);
4134}
4135
4136static int leaks_show(struct seq_file *m, void *p)
4137{
Thierry Reding0672aa72012-06-22 19:42:49 +02004138 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004139 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004140 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004141 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004142 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004143 int node;
4144 int i;
4145
4146 if (!(cachep->flags & SLAB_STORE_USER))
4147 return 0;
4148 if (!(cachep->flags & SLAB_RED_ZONE))
4149 return 0;
4150
4151 /* OK, we can do it */
4152
Christoph Lameterdb845062013-02-05 18:45:23 +00004153 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004154
Christoph Lameter18bf8542014-08-06 16:04:11 -07004155 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004156
4157 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004158 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004159
Joonsoo Kim8456a642013-10-24 10:07:49 +09004160 list_for_each_entry(page, &n->slabs_full, lru)
4161 handle_slab(x, cachep, page);
4162 list_for_each_entry(page, &n->slabs_partial, lru)
4163 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004164 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004165 }
4166 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004167 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004168 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004169 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004170 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004171 if (!m->private) {
4172 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004173 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004174 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004175 return -ENOMEM;
4176 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004177 *(unsigned long *)m->private = x[0] * 2;
4178 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004179 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004180 /* Now make sure this entry will be retried */
4181 m->count = m->size;
4182 return 0;
4183 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004184 for (i = 0; i < x[1]; i++) {
4185 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4186 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004187 seq_putc(m, '\n');
4188 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004189
Al Viro871751e2006-03-25 03:06:39 -08004190 return 0;
4191}
4192
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004193static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004194 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004195 .next = slab_next,
4196 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004197 .show = leaks_show,
4198};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004199
4200static int slabstats_open(struct inode *inode, struct file *file)
4201{
Rob Jonesb208ce32014-10-09 15:28:03 -07004202 unsigned long *n;
4203
4204 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4205 if (!n)
4206 return -ENOMEM;
4207
4208 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4209
4210 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004211}
4212
4213static const struct file_operations proc_slabstats_operations = {
4214 .open = slabstats_open,
4215 .read = seq_read,
4216 .llseek = seq_lseek,
4217 .release = seq_release_private,
4218};
Al Viro871751e2006-03-25 03:06:39 -08004219#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004220
4221static int __init slab_proc_init(void)
4222{
4223#ifdef CONFIG_DEBUG_SLAB_LEAK
4224 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4225#endif
4226 return 0;
4227}
4228module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229#endif
4230
Manfred Spraul00e145b2005-09-03 15:55:07 -07004231/**
4232 * ksize - get the actual amount of memory allocated for a given object
4233 * @objp: Pointer to the object
4234 *
4235 * kmalloc may internally round up allocations and return more memory
4236 * than requested. ksize() can be used to determine the actual amount of
4237 * memory allocated. The caller may use this additional memory, even though
4238 * a smaller amount of memory was initially specified with the kmalloc call.
4239 * The caller must guarantee that objp points to a valid object previously
4240 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4241 * must not be freed during the duration of the call.
4242 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004243size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004244{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004245 BUG_ON(!objp);
4246 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004247 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004249 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004251EXPORT_SYMBOL(ksize);