blob: 4fcc5dd8d5a6c2776ac2f88d9011ac23c78fb769 [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)
285
286#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800287/*
288 * Optimization question: fewer reaps means less probability for unnessary
289 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100291 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 * which could lock up otherwise freeable slabs.
293 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800294#define REAPTIMEOUT_AC (2*HZ)
295#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296
297#if STATS
298#define STATS_INC_ACTIVE(x) ((x)->num_active++)
299#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
300#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
301#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700302#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800303#define STATS_SET_HIGH(x) \
304 do { \
305 if ((x)->num_active > (x)->high_mark) \
306 (x)->high_mark = (x)->num_active; \
307 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308#define STATS_INC_ERR(x) ((x)->errors++)
309#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700310#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700311#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800312#define STATS_SET_FREEABLE(x, i) \
313 do { \
314 if ((x)->max_freeable < i) \
315 (x)->max_freeable = i; \
316 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
318#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
319#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
320#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
321#else
322#define STATS_INC_ACTIVE(x) do { } while (0)
323#define STATS_DEC_ACTIVE(x) do { } while (0)
324#define STATS_INC_ALLOCED(x) do { } while (0)
325#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700326#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327#define STATS_SET_HIGH(x) do { } while (0)
328#define STATS_INC_ERR(x) do { } while (0)
329#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700330#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700331#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800332#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333#define STATS_INC_ALLOCHIT(x) do { } while (0)
334#define STATS_INC_ALLOCMISS(x) do { } while (0)
335#define STATS_INC_FREEHIT(x) do { } while (0)
336#define STATS_INC_FREEMISS(x) do { } while (0)
337#endif
338
339#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Andrew Mortona737b3e2006-03-22 00:08:11 -0800341/*
342 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800344 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 * the end of an object is aligned with the end of the real
346 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800347 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800349 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500350 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
351 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800352 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800354static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800356 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357}
358
David Woodhouseb46b8f12007-05-08 00:22:59 -0700359static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360{
361 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700362 return (unsigned long long*) (objp + obj_offset(cachep) -
363 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364}
365
David Woodhouseb46b8f12007-05-08 00:22:59 -0700366static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367{
368 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
369 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500370 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700371 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400372 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500373 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700374 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375}
376
Pekka Enberg343e0d72006-02-01 03:05:50 -0800377static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378{
379 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500380 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381}
382
383#else
384
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800385#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700386#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
387#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
389
390#endif
391
Joonsoo Kim03787302014-06-23 13:22:06 -0700392#define OBJECT_FREE (0)
393#define OBJECT_ACTIVE (1)
394
395#ifdef CONFIG_DEBUG_SLAB_LEAK
396
397static void set_obj_status(struct page *page, int idx, int val)
398{
399 int freelist_size;
400 char *status;
401 struct kmem_cache *cachep = page->slab_cache;
402
403 freelist_size = cachep->num * sizeof(freelist_idx_t);
404 status = (char *)page->freelist + freelist_size;
405 status[idx] = val;
406}
407
408static inline unsigned int get_obj_status(struct page *page, int idx)
409{
410 int freelist_size;
411 char *status;
412 struct kmem_cache *cachep = page->slab_cache;
413
414 freelist_size = cachep->num * sizeof(freelist_idx_t);
415 status = (char *)page->freelist + freelist_size;
416
417 return status[idx];
418}
419
420#else
421static inline void set_obj_status(struct page *page, int idx, int val) {}
422
423#endif
424
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700426 * Do not go above this order unless 0 objects fit into the slab or
427 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 */
David Rientjes543585c2011-10-18 22:09:24 -0700429#define SLAB_MAX_ORDER_HI 1
430#define SLAB_MAX_ORDER_LO 0
431static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700432static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800434static inline struct kmem_cache *virt_to_cache(const void *obj)
435{
Christoph Lameterb49af682007-05-06 14:49:41 -0700436 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500437 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800438}
439
Joonsoo Kim8456a642013-10-24 10:07:49 +0900440static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800441 unsigned int idx)
442{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900443 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800444}
445
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800446/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500447 * We want to avoid an expensive divide : (offset / cache->size)
448 * Using the fact that size is a constant for a particular cache,
449 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800450 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
451 */
452static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900453 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800454{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900455 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800456 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800457}
458
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000460static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800461 .batchcount = 1,
462 .limit = BOOT_CPUCACHE_ENTRIES,
463 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500464 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800465 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466};
467
Joonsoo Kimedcad252014-08-08 14:19:15 -0700468#define BAD_ALIEN_MAGIC 0x01020304ul
469
Tejun Heo1871e522009-10-29 22:34:13 +0900470static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471
Pekka Enberg343e0d72006-02-01 03:05:50 -0800472static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700474 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Joonsoo Kim03787302014-06-23 13:22:06 -0700477static size_t calculate_freelist_size(int nr_objs, size_t align)
478{
479 size_t freelist_size;
480
481 freelist_size = nr_objs * sizeof(freelist_idx_t);
482 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
483 freelist_size += nr_objs * sizeof(char);
484
485 if (align)
486 freelist_size = ALIGN(freelist_size, align);
487
488 return freelist_size;
489}
490
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900491static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
492 size_t idx_size, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493{
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900494 int nr_objs;
Joonsoo Kim03787302014-06-23 13:22:06 -0700495 size_t remained_size;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900496 size_t freelist_size;
Joonsoo Kim03787302014-06-23 13:22:06 -0700497 int extra_space = 0;
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900498
Joonsoo Kim03787302014-06-23 13:22:06 -0700499 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
500 extra_space = sizeof(char);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900501 /*
502 * Ignore padding for the initial guess. The padding
503 * is at most @align-1 bytes, and @buffer_size is at
504 * least @align. In the worst case, this result will
505 * be one greater than the number of objects that fit
506 * into the memory allocation when taking the padding
507 * into account.
508 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700509 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900510
511 /*
512 * This calculated number will be either the right
513 * amount, or one greater than what we want.
514 */
Joonsoo Kim03787302014-06-23 13:22:06 -0700515 remained_size = slab_size - nr_objs * buffer_size;
516 freelist_size = calculate_freelist_size(nr_objs, align);
517 if (remained_size < freelist_size)
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900518 nr_objs--;
519
520 return nr_objs;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800521}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522
Andrew Mortona737b3e2006-03-22 00:08:11 -0800523/*
524 * Calculate the number of objects and left-over bytes for a given buffer size.
525 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800526static void cache_estimate(unsigned long gfporder, size_t buffer_size,
527 size_t align, int flags, size_t *left_over,
528 unsigned int *num)
529{
530 int nr_objs;
531 size_t mgmt_size;
532 size_t slab_size = PAGE_SIZE << gfporder;
533
534 /*
535 * The slab management structure can be either off the slab or
536 * on it. For the latter case, the memory allocated for a
537 * slab is used for:
538 *
Joonsoo Kim16025172013-10-24 10:07:46 +0900539 * - One unsigned int for each object
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800540 * - Padding to respect alignment of @align
541 * - @buffer_size bytes for each object
542 *
543 * If the slab management structure is off the slab, then the
544 * alignment will already be calculated into the size. Because
545 * the slabs are all pages aligned, the objects will be at the
546 * correct alignment when allocated.
547 */
548 if (flags & CFLGS_OFF_SLAB) {
549 mgmt_size = 0;
550 nr_objs = slab_size / buffer_size;
551
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800552 } else {
Joonsoo Kim9cef2e22013-12-02 17:49:39 +0900553 nr_objs = calculate_nr_objs(slab_size, buffer_size,
Joonsoo Kima41adfa2013-12-02 17:49:42 +0900554 sizeof(freelist_idx_t), align);
Joonsoo Kim03787302014-06-23 13:22:06 -0700555 mgmt_size = calculate_freelist_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800557 *num = nr_objs;
558 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
Christoph Lameterf28510d2012-09-11 19:49:38 +0000561#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700562#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
Andrew Mortona737b3e2006-03-22 00:08:11 -0800564static void __slab_error(const char *function, struct kmem_cache *cachep,
565 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566{
567 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800568 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030570 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000572#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
Paul Menage3395ee02006-12-06 20:32:16 -0800574/*
575 * By default on NUMA we use alien caches to stage the freeing of
576 * objects allocated from other nodes. This causes massive memory
577 * inefficiencies when using fake NUMA setup to split memory into a
578 * large number of small nodes, so it can be disabled on the command
579 * line
580 */
581
582static int use_alien_caches __read_mostly = 1;
583static int __init noaliencache_setup(char *s)
584{
585 use_alien_caches = 0;
586 return 1;
587}
588__setup("noaliencache", noaliencache_setup);
589
David Rientjes3df1ccc2011-10-18 22:09:28 -0700590static int __init slab_max_order_setup(char *str)
591{
592 get_option(&str, &slab_max_order);
593 slab_max_order = slab_max_order < 0 ? 0 :
594 min(slab_max_order, MAX_ORDER - 1);
595 slab_max_order_set = true;
596
597 return 1;
598}
599__setup("slab_max_order=", slab_max_order_setup);
600
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800601#ifdef CONFIG_NUMA
602/*
603 * Special reaping functions for NUMA systems called from cache_reap().
604 * These take care of doing round robin flushing of alien caches (containing
605 * objects freed on different nodes from which they were allocated) and the
606 * flushing of remote pcps by calling drain_node_pages.
607 */
Tejun Heo1871e522009-10-29 22:34:13 +0900608static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800609
610static void init_reap_node(int cpu)
611{
612 int node;
613
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700614 node = next_node(cpu_to_mem(cpu), node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800615 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800616 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800617
Tejun Heo1871e522009-10-29 22:34:13 +0900618 per_cpu(slab_reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800619}
620
621static void next_reap_node(void)
622{
Christoph Lameter909ea962010-12-08 16:22:55 +0100623 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800624
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800625 node = next_node(node, node_online_map);
626 if (unlikely(node >= MAX_NUMNODES))
627 node = first_node(node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100628 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800629}
630
631#else
632#define init_reap_node(cpu) do { } while (0)
633#define next_reap_node(void) do { } while (0)
634#endif
635
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636/*
637 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
638 * via the workqueue/eventd.
639 * Add the CPU number into the expiration time to minimize the possibility of
640 * the CPUs getting into lockstep and contending for the global cache chain
641 * lock.
642 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400643static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644{
Tejun Heo1871e522009-10-29 22:34:13 +0900645 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
647 /*
648 * When this gets called from do_initcalls via cpucache_init(),
649 * init_workqueues() has already run, so keventd will be setup
650 * at that time.
651 */
David Howells52bad642006-11-22 14:54:01 +0000652 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800653 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700654 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800655 schedule_delayed_work_on(cpu, reap_work,
656 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 }
658}
659
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700660static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100662 /*
663 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300664 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100665 * cache the pointers are not cleared and they could be counted as
666 * valid references during a kmemleak scan. Therefore, kmemleak must
667 * not scan such objects.
668 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700669 kmemleak_no_scan(ac);
670 if (ac) {
671 ac->avail = 0;
672 ac->limit = limit;
673 ac->batchcount = batch;
674 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700676}
677
678static struct array_cache *alloc_arraycache(int node, int entries,
679 int batchcount, gfp_t gfp)
680{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700681 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700682 struct array_cache *ac = NULL;
683
684 ac = kmalloc_node(memsize, gfp, node);
685 init_arraycache(ac, entries, batchcount);
686 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687}
688
Joonsoo Kim8456a642013-10-24 10:07:49 +0900689static inline bool is_slab_pfmemalloc(struct page *page)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700690{
Mel Gorman072bb0a2012-07-31 16:43:58 -0700691 return PageSlabPfmemalloc(page);
692}
693
694/* Clears pfmemalloc_active if no slabs have pfmalloc set */
695static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
696 struct array_cache *ac)
697{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700698 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
Joonsoo Kim8456a642013-10-24 10:07:49 +0900699 struct page *page;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700700 unsigned long flags;
701
702 if (!pfmemalloc_active)
703 return;
704
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000705 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +0900706 list_for_each_entry(page, &n->slabs_full, lru)
707 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700708 goto out;
709
Joonsoo Kim8456a642013-10-24 10:07:49 +0900710 list_for_each_entry(page, &n->slabs_partial, lru)
711 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700712 goto out;
713
Joonsoo Kim8456a642013-10-24 10:07:49 +0900714 list_for_each_entry(page, &n->slabs_free, lru)
715 if (is_slab_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -0700716 goto out;
717
718 pfmemalloc_active = false;
719out:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000720 spin_unlock_irqrestore(&n->list_lock, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700721}
722
Mel Gorman381760e2012-07-31 16:44:30 -0700723static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
Mel Gorman072bb0a2012-07-31 16:43:58 -0700724 gfp_t flags, bool force_refill)
725{
726 int i;
727 void *objp = ac->entry[--ac->avail];
728
729 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
730 if (unlikely(is_obj_pfmemalloc(objp))) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000731 struct kmem_cache_node *n;
Mel Gorman072bb0a2012-07-31 16:43:58 -0700732
733 if (gfp_pfmemalloc_allowed(flags)) {
734 clear_obj_pfmemalloc(&objp);
735 return objp;
736 }
737
738 /* The caller cannot use PFMEMALLOC objects, find another one */
Joonsoo Kimd014dc22012-09-17 14:09:06 -0700739 for (i = 0; i < ac->avail; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -0700740 /* If a !PFMEMALLOC object is found, swap them */
741 if (!is_obj_pfmemalloc(ac->entry[i])) {
742 objp = ac->entry[i];
743 ac->entry[i] = ac->entry[ac->avail];
744 ac->entry[ac->avail] = objp;
745 return objp;
746 }
747 }
748
749 /*
750 * If there are empty slabs on the slabs_free list and we are
751 * being forced to refill the cache, mark this one !pfmemalloc.
752 */
Christoph Lameter18bf8542014-08-06 16:04:11 -0700753 n = get_node(cachep, numa_mem_id());
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000754 if (!list_empty(&n->slabs_free) && force_refill) {
Joonsoo Kim8456a642013-10-24 10:07:49 +0900755 struct page *page = virt_to_head_page(objp);
Joonsoo Kim7ecccf92013-10-24 10:07:50 +0900756 ClearPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700757 clear_obj_pfmemalloc(&objp);
758 recheck_pfmemalloc_active(cachep, ac);
759 return objp;
760 }
761
762 /* No !PFMEMALLOC objects available */
763 ac->avail++;
764 objp = NULL;
765 }
766
767 return objp;
768}
769
Mel Gorman381760e2012-07-31 16:44:30 -0700770static inline void *ac_get_obj(struct kmem_cache *cachep,
771 struct array_cache *ac, gfp_t flags, bool force_refill)
772{
773 void *objp;
774
775 if (unlikely(sk_memalloc_socks()))
776 objp = __ac_get_obj(cachep, ac, flags, force_refill);
777 else
778 objp = ac->entry[--ac->avail];
779
780 return objp;
781}
782
Joonsoo Kimd3aec342014-10-09 15:26:06 -0700783static noinline void *__ac_put_obj(struct kmem_cache *cachep,
784 struct array_cache *ac, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700785{
786 if (unlikely(pfmemalloc_active)) {
787 /* Some pfmemalloc slabs exist, check if this is one */
Mel Gorman30c29be2012-09-17 14:09:03 -0700788 struct page *page = virt_to_head_page(objp);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700789 if (PageSlabPfmemalloc(page))
790 set_obj_pfmemalloc(&objp);
791 }
792
Mel Gorman381760e2012-07-31 16:44:30 -0700793 return objp;
794}
795
796static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
797 void *objp)
798{
799 if (unlikely(sk_memalloc_socks()))
800 objp = __ac_put_obj(cachep, ac, objp);
801
Mel Gorman072bb0a2012-07-31 16:43:58 -0700802 ac->entry[ac->avail++] = objp;
803}
804
Christoph Lameter3ded1752006-03-25 03:06:44 -0800805/*
806 * Transfer objects in one arraycache to another.
807 * Locking must be handled by the caller.
808 *
809 * Return the number of entries transferred.
810 */
811static int transfer_objects(struct array_cache *to,
812 struct array_cache *from, unsigned int max)
813{
814 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700815 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800816
817 if (!nr)
818 return 0;
819
820 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
821 sizeof(void *) *nr);
822
823 from->avail -= nr;
824 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800825 return nr;
826}
827
Christoph Lameter765c4502006-09-27 01:50:08 -0700828#ifndef CONFIG_NUMA
829
830#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000831#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700832
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700833static inline struct alien_cache **alloc_alien_cache(int node,
834 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700835{
Joonsoo Kimedcad252014-08-08 14:19:15 -0700836 return (struct alien_cache **)BAD_ALIEN_MAGIC;
Christoph Lameter765c4502006-09-27 01:50:08 -0700837}
838
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700839static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700840{
841}
842
843static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
844{
845 return 0;
846}
847
848static inline void *alternate_node_alloc(struct kmem_cache *cachep,
849 gfp_t flags)
850{
851 return NULL;
852}
853
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800854static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700855 gfp_t flags, int nodeid)
856{
857 return NULL;
858}
859
David Rientjes4167e9b2015-04-14 15:46:55 -0700860static inline gfp_t gfp_exact_node(gfp_t flags)
861{
862 return flags;
863}
864
Christoph Lameter765c4502006-09-27 01:50:08 -0700865#else /* CONFIG_NUMA */
866
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800867static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800868static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800869
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700870static struct alien_cache *__alloc_alien_cache(int node, int entries,
871 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700872{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700873 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700874 struct alien_cache *alc = NULL;
875
876 alc = kmalloc_node(memsize, gfp, node);
877 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700878 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700879 return alc;
880}
881
882static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
883{
884 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700885 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700886 int i;
887
888 if (limit > 1)
889 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700890 alc_ptr = kzalloc_node(memsize, gfp, node);
891 if (!alc_ptr)
892 return NULL;
893
894 for_each_node(i) {
895 if (i == node || !node_online(i))
896 continue;
897 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
898 if (!alc_ptr[i]) {
899 for (i--; i >= 0; i--)
900 kfree(alc_ptr[i]);
901 kfree(alc_ptr);
902 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700903 }
904 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700905 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700906}
907
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700908static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700909{
910 int i;
911
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700912 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700913 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700914 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700915 kfree(alc_ptr[i]);
916 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700917}
918
Pekka Enberg343e0d72006-02-01 03:05:50 -0800919static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700920 struct array_cache *ac, int node,
921 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700922{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700923 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700924
925 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000926 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800927 /*
928 * Stuff objects into the remote nodes shared array first.
929 * That way we could avoid the overhead of putting the objects
930 * into the free lists and getting them back later.
931 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000932 if (n->shared)
933 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800934
Joonsoo Kim833b7062014-08-06 16:04:33 -0700935 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700936 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000937 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700938 }
939}
940
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800941/*
942 * Called from cache_reap() to regularly drain alien caches round robin.
943 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000944static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800945{
Christoph Lameter909ea962010-12-08 16:22:55 +0100946 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800947
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000948 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700949 struct alien_cache *alc = n->alien[node];
950 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800951
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700952 if (alc) {
953 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700954 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700955 LIST_HEAD(list);
956
957 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700958 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700959 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700960 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800961 }
962 }
963}
964
Andrew Mortona737b3e2006-03-22 00:08:11 -0800965static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700966 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700967{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800968 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700969 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700970 struct array_cache *ac;
971 unsigned long flags;
972
973 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700974 alc = alien[i];
975 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700976 LIST_HEAD(list);
977
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700978 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700979 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700980 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700981 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700982 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700983 }
984 }
985}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700986
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700987static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
988 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700989{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000990 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700991 struct alien_cache *alien = NULL;
992 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700993 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700994
Christoph Lameter18bf8542014-08-06 16:04:11 -0700995 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700996 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700997 if (n->alien && n->alien[page_node]) {
998 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700999 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001000 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001001 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001002 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001003 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001004 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001005 ac_put_obj(cachep, ac, objp);
Joonsoo Kim49dfc302014-08-06 16:04:31 -07001006 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -07001007 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001008 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001009 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001010 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001011 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07001012 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07001013 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001014 }
1015 return 1;
1016}
Joonsoo Kim25c4f302014-10-09 15:26:09 -07001017
1018static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1019{
1020 int page_node = page_to_nid(virt_to_page(objp));
1021 int node = numa_mem_id();
1022 /*
1023 * Make sure we are not freeing a object from another node to the array
1024 * cache on this cpu.
1025 */
1026 if (likely(node == page_node))
1027 return 0;
1028
1029 return __cache_free_alien(cachep, objp, node, page_node);
1030}
David Rientjes4167e9b2015-04-14 15:46:55 -07001031
1032/*
1033 * Construct gfp mask to allocate from a specific node but do not invoke reclaim
1034 * or warn about failures.
1035 */
1036static inline gfp_t gfp_exact_node(gfp_t flags)
1037{
1038 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_WAIT;
1039}
Christoph Lametere498be72005-09-09 13:03:32 -07001040#endif
1041
David Rientjes8f9f8d92010-03-27 19:40:47 -07001042/*
Christoph Lameter6a673682013-01-10 19:14:19 +00001043 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001044 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -07001045 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +00001046 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -07001047 * already in use.
1048 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001049 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001050 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001051static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001052{
1053 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001054 struct kmem_cache_node *n;
Joonsoo Kim5e804782014-08-06 16:04:40 -07001055 const size_t memsize = sizeof(struct kmem_cache_node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001056
Christoph Lameter18004c52012-07-06 15:25:12 -05001057 list_for_each_entry(cachep, &slab_caches, list) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001058 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001059 * Set up the kmem_cache_node for cpu before we can
David Rientjes8f9f8d92010-03-27 19:40:47 -07001060 * begin anything. Make sure some other cpu on this
1061 * node has not already allocated this
1062 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07001063 n = get_node(cachep, node);
1064 if (!n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001065 n = kmalloc_node(memsize, GFP_KERNEL, node);
1066 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001067 return -ENOMEM;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001068 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001069 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1070 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001071
1072 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001073 * The kmem_cache_nodes don't come and go as CPUs
1074 * come and go. slab_mutex is sufficient
David Rientjes8f9f8d92010-03-27 19:40:47 -07001075 * protection here.
1076 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001077 cachep->node[node] = n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001078 }
1079
Christoph Lameter18bf8542014-08-06 16:04:11 -07001080 spin_lock_irq(&n->list_lock);
1081 n->free_limit =
David Rientjes8f9f8d92010-03-27 19:40:47 -07001082 (1 + nr_cpus_node(node)) *
1083 cachep->batchcount + cachep->num;
Christoph Lameter18bf8542014-08-06 16:04:11 -07001084 spin_unlock_irq(&n->list_lock);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001085 }
1086 return 0;
1087}
1088
Wanpeng Li0fa81032013-07-04 08:33:22 +08001089static inline int slabs_tofree(struct kmem_cache *cachep,
1090 struct kmem_cache_node *n)
1091{
1092 return (n->free_objects + cachep->num - 1) / cachep->num;
1093}
1094
Paul Gortmaker0db06282013-06-19 14:53:51 -04001095static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001097 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001098 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001099 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301100 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001101
Christoph Lameter18004c52012-07-06 15:25:12 -05001102 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001103 struct array_cache *nc;
1104 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001105 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -07001106 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001107
Christoph Lameter18bf8542014-08-06 16:04:11 -07001108 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001109 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001110 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001111
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001112 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001113
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001114 /* Free limit for this kmem_cache_node */
1115 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001116
1117 /* cpu is dead; no one can alloc from it. */
1118 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1119 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001120 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001121 nc->avail = 0;
1122 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001123
Rusty Russell58463c12009-12-17 11:43:12 -06001124 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001125 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001126 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001127 }
1128
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001129 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001130 if (shared) {
1131 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001132 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001133 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001134 }
1135
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001136 alien = n->alien;
1137 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001138
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001139 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001140
1141 kfree(shared);
1142 if (alien) {
1143 drain_alien_cache(cachep, alien);
1144 free_alien_cache(alien);
1145 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001146
1147free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001148 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001149 }
1150 /*
1151 * In the previous loop, all the objects were freed to
1152 * the respective cache's slabs, now we can go ahead and
1153 * shrink each nodelist to its limit.
1154 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001155 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001156 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001157 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001158 continue;
Wanpeng Li0fa81032013-07-04 08:33:22 +08001159 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001160 }
1161}
1162
Paul Gortmaker0db06282013-06-19 14:53:51 -04001163static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001164{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001165 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001166 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001167 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001168 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001170 /*
1171 * We need to do this right in the beginning since
1172 * alloc_arraycache's are going to use this list.
1173 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001174 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001175 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001176 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001177 if (err < 0)
1178 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001179
1180 /*
1181 * Now we can go ahead with allocating the shared arrays and
1182 * array caches
1183 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001184 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 struct array_cache *shared = NULL;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07001186 struct alien_cache **alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001187
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001188 if (cachep->shared) {
1189 shared = alloc_arraycache(node,
1190 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001191 0xbaadf00d, GFP_KERNEL);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001192 if (!shared)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001193 goto bad;
1194 }
1195 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001196 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001197 if (!alien) {
1198 kfree(shared);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001199 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001200 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001201 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07001202 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001203 BUG_ON(!n);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001204
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001205 spin_lock_irq(&n->list_lock);
1206 if (!n->shared) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001207 /*
1208 * We are serialised from CPU_DEAD or
1209 * CPU_UP_CANCELLED by the cpucontrol lock
1210 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001211 n->shared = shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001212 shared = NULL;
1213 }
1214#ifdef CONFIG_NUMA
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001215 if (!n->alien) {
1216 n->alien = alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001217 alien = NULL;
1218 }
1219#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001220 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001221 kfree(shared);
1222 free_alien_cache(alien);
1223 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001224
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001225 return 0;
1226bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001227 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001228 return -ENOMEM;
1229}
1230
Paul Gortmaker0db06282013-06-19 14:53:51 -04001231static int cpuup_callback(struct notifier_block *nfb,
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001232 unsigned long action, void *hcpu)
1233{
1234 long cpu = (long)hcpu;
1235 int err = 0;
1236
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001238 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001239 case CPU_UP_PREPARE_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001240 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001241 err = cpuup_prepare(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001242 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 break;
1244 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001245 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 start_cpu_timer(cpu);
1247 break;
1248#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001249 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001250 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001251 /*
Christoph Lameter18004c52012-07-06 15:25:12 -05001252 * Shutdown cache reaper. Note that the slab_mutex is
Christoph Lameter5830c592007-05-09 02:34:22 -07001253 * held so that if cache_reap() is invoked it cannot do
1254 * anything expensive but will only modify reap_work
1255 * and reschedule the timer.
1256 */
Tejun Heoafe2c512010-12-14 16:21:17 +01001257 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
Christoph Lameter5830c592007-05-09 02:34:22 -07001258 /* Now the cache_reaper is guaranteed to be not running. */
Tejun Heo1871e522009-10-29 22:34:13 +09001259 per_cpu(slab_reap_work, cpu).work.func = NULL;
Christoph Lameter5830c592007-05-09 02:34:22 -07001260 break;
1261 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001262 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001263 start_cpu_timer(cpu);
1264 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001266 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001267 /*
1268 * Even if all the cpus of a node are down, we don't free the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001269 * kmem_cache_node of any cache. This to avoid a race between
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001270 * cpu_down, and a kmalloc allocation from another cpu for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001271 * memory from the node of the cpu going down. The node
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001272 * structure is usually allocated from kmem_cache_create() and
1273 * gets destroyed at kmem_cache_destroy().
1274 */
Simon Arlott183ff222007-10-20 01:27:18 +02001275 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001276#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001278 case CPU_UP_CANCELED_FROZEN:
Christoph Lameter18004c52012-07-06 15:25:12 -05001279 mutex_lock(&slab_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001280 cpuup_canceled(cpu);
Christoph Lameter18004c52012-07-06 15:25:12 -05001281 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283 }
Akinobu Mitaeac40682010-05-26 14:43:32 -07001284 return notifier_from_errno(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285}
1286
Paul Gortmaker0db06282013-06-19 14:53:51 -04001287static struct notifier_block cpucache_notifier = {
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001288 &cpuup_callback, NULL, 0
1289};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
David Rientjes8f9f8d92010-03-27 19:40:47 -07001291#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1292/*
1293 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1294 * Returns -EBUSY if all objects cannot be drained so that the node is not
1295 * removed.
1296 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001297 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001298 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001299static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001300{
1301 struct kmem_cache *cachep;
1302 int ret = 0;
1303
Christoph Lameter18004c52012-07-06 15:25:12 -05001304 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001305 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001306
Christoph Lameter18bf8542014-08-06 16:04:11 -07001307 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001308 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001309 continue;
1310
Wanpeng Li0fa81032013-07-04 08:33:22 +08001311 drain_freelist(cachep, n, slabs_tofree(cachep, n));
David Rientjes8f9f8d92010-03-27 19:40:47 -07001312
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001313 if (!list_empty(&n->slabs_full) ||
1314 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001315 ret = -EBUSY;
1316 break;
1317 }
1318 }
1319 return ret;
1320}
1321
1322static int __meminit slab_memory_callback(struct notifier_block *self,
1323 unsigned long action, void *arg)
1324{
1325 struct memory_notify *mnb = arg;
1326 int ret = 0;
1327 int nid;
1328
1329 nid = mnb->status_change_nid;
1330 if (nid < 0)
1331 goto out;
1332
1333 switch (action) {
1334 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001335 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001336 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001337 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001338 break;
1339 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001340 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001341 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001342 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001343 break;
1344 case MEM_ONLINE:
1345 case MEM_OFFLINE:
1346 case MEM_CANCEL_ONLINE:
1347 case MEM_CANCEL_OFFLINE:
1348 break;
1349 }
1350out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001351 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001352}
1353#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1354
Christoph Lametere498be72005-09-09 13:03:32 -07001355/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001356 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001357 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001358static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001359 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001360{
Christoph Lameter6744f082013-01-10 19:12:17 +00001361 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001362
Christoph Lameter6744f082013-01-10 19:12:17 +00001363 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001364 BUG_ON(!ptr);
1365
Christoph Lameter6744f082013-01-10 19:12:17 +00001366 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001367 /*
1368 * Do not assume that spinlocks can be initialized via memcpy:
1369 */
1370 spin_lock_init(&ptr->list_lock);
1371
Christoph Lametere498be72005-09-09 13:03:32 -07001372 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001373 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001374}
1375
Andrew Mortona737b3e2006-03-22 00:08:11 -08001376/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001377 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1378 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001379 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001380static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001381{
1382 int node;
1383
1384 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001385 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001386 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001387 REAPTIMEOUT_NODE +
1388 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001389 }
1390}
1391
1392/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001393 * Initialisation. Called after the page allocator have been initialised and
1394 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 */
1396void __init kmem_cache_init(void)
1397{
Christoph Lametere498be72005-09-09 13:03:32 -07001398 int i;
1399
Joonsoo Kim68126702013-10-24 10:07:42 +09001400 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1401 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001402 kmem_cache = &kmem_cache_boot;
1403
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001404 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001405 use_alien_caches = 0;
1406
Christoph Lameter3c583462012-11-28 16:23:01 +00001407 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001408 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001409
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 /*
1411 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001412 * page orders on machines with more than 32MB of memory if
1413 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001415 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001416 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001417
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 /* Bootstrap is tricky, because several objects are allocated
1419 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001420 * 1) initialize the kmem_cache cache: it contains the struct
1421 * kmem_cache structures of all caches, except kmem_cache itself:
1422 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001423 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001424 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001425 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001427 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001428 * An __init data area is used for the head array.
1429 * 3) Create the remaining kmalloc caches, with minimally sized
1430 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001431 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001433 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001434 * the other cache's with kmalloc allocated memory.
1435 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 */
1437
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001438 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439
Eric Dumazet8da34302007-05-06 14:49:29 -07001440 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001441 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001442 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001443 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001444 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001445 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001446 SLAB_HWCACHE_ALIGN);
1447 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001448 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
Andrew Mortona737b3e2006-03-22 00:08:11 -08001450 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001451 * Initialize the caches that provide memory for the kmem_cache_node
1452 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001453 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001454 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001455 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001456 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001457 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001458
Ingo Molnare0a42722006-06-23 02:03:46 -07001459 slab_early_init = 0;
1460
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001461 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001462 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001463 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464
Mel Gorman9c09a952008-01-24 05:49:54 -08001465 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001466 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001467
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001468 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001469 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001470 }
1471 }
1472
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001473 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001474}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001475
Pekka Enberg8429db52009-06-12 15:58:59 +03001476void __init kmem_cache_init_late(void)
1477{
1478 struct kmem_cache *cachep;
1479
Christoph Lameter97d06602012-07-06 15:25:11 -05001480 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001481
Pekka Enberg8429db52009-06-12 15:58:59 +03001482 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001483 mutex_lock(&slab_mutex);
1484 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001485 if (enable_cpucache(cachep, GFP_NOWAIT))
1486 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001487 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001488
Christoph Lameter97d06602012-07-06 15:25:11 -05001489 /* Done! */
1490 slab_state = FULL;
1491
Andrew Mortona737b3e2006-03-22 00:08:11 -08001492 /*
1493 * Register a cpu startup notifier callback that initializes
1494 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 */
1496 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497
David Rientjes8f9f8d92010-03-27 19:40:47 -07001498#ifdef CONFIG_NUMA
1499 /*
1500 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001501 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001502 */
1503 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1504#endif
1505
Andrew Mortona737b3e2006-03-22 00:08:11 -08001506 /*
1507 * The reap timers are started later, with a module init call: That part
1508 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 */
1510}
1511
1512static int __init cpucache_init(void)
1513{
1514 int cpu;
1515
Andrew Mortona737b3e2006-03-22 00:08:11 -08001516 /*
1517 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 */
Christoph Lametere498be72005-09-09 13:03:32 -07001519 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001520 start_cpu_timer(cpu);
Glauber Costaa164f8962012-06-21 00:59:18 +04001521
1522 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001523 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 return 0;
1525}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526__initcall(cpucache_init);
1527
Rafael Aquini8bdec192012-03-09 17:27:27 -03001528static noinline void
1529slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1530{
David Rientjes9a02d692014-06-04 16:06:36 -07001531#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001532 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001533 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001534 unsigned long flags;
1535 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001536 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1537 DEFAULT_RATELIMIT_BURST);
1538
1539 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1540 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001541
1542 printk(KERN_WARNING
1543 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1544 nodeid, gfpflags);
1545 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001546 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001547
Christoph Lameter18bf8542014-08-06 16:04:11 -07001548 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001549 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1550 unsigned long active_slabs = 0, num_slabs = 0;
1551
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001552 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001553 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001554 active_objs += cachep->num;
1555 active_slabs++;
1556 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001557 list_for_each_entry(page, &n->slabs_partial, lru) {
1558 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001559 active_slabs++;
1560 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001561 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001562 num_slabs++;
1563
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001564 free_objects += n->free_objects;
1565 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001566
1567 num_slabs += active_slabs;
1568 num_objs = num_slabs * cachep->num;
1569 printk(KERN_WARNING
1570 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1571 node, active_slabs, num_slabs, active_objs, num_objs,
1572 free_objects);
1573 }
David Rientjes9a02d692014-06-04 16:06:36 -07001574#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001575}
1576
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001578 * Interface to system's page allocator. No need to hold the
1579 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 *
1581 * If we requested dmaable memory, we will get it. Even if we
1582 * did not request dmaable memory, we might get it, but that
1583 * would be relatively rare and ignorable.
1584 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001585static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1586 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587{
1588 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001589 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001590
Glauber Costaa618e892012-06-14 16:17:21 +04001591 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001592 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1593 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001594
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001595 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1596 return NULL;
1597
Vlastimil Babka96db8002015-09-08 15:03:50 -07001598 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001599 if (!page) {
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001600 memcg_uncharge_slab(cachep, cachep->gfporder);
David Rientjes9a02d692014-06-04 16:06:36 -07001601 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001603 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604
Mel Gormanb37f1dd2012-07-31 16:44:03 -07001605 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
Michal Hocko2f064f32015-08-21 14:11:51 -07001606 if (page_is_pfmemalloc(page))
Mel Gorman072bb0a2012-07-31 16:43:58 -07001607 pfmemalloc_active = true;
1608
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001609 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001611 add_zone_page_state(page_zone(page),
1612 NR_SLAB_RECLAIMABLE, nr_pages);
1613 else
1614 add_zone_page_state(page_zone(page),
1615 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001616 __SetPageSlab(page);
Michal Hocko2f064f32015-08-21 14:11:51 -07001617 if (page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001618 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001619
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001620 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1621 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1622
1623 if (cachep->ctor)
1624 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1625 else
1626 kmemcheck_mark_unallocated_pages(page, nr_pages);
1627 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001628
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001629 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630}
1631
1632/*
1633 * Interface to system's page release.
1634 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001635static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636{
Joonsoo Kima57a4982013-10-24 10:07:44 +09001637 const unsigned long nr_freed = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001639 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001640
Christoph Lameter972d1a72006-09-25 23:31:51 -07001641 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1642 sub_zone_page_state(page_zone(page),
1643 NR_SLAB_RECLAIMABLE, nr_freed);
1644 else
1645 sub_zone_page_state(page_zone(page),
1646 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001647
Joonsoo Kima57a4982013-10-24 10:07:44 +09001648 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001649 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001650 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001651 page_mapcount_reset(page);
1652 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001653
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 if (current->reclaim_state)
1655 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov5dfb4172014-06-04 16:06:38 -07001656 __free_pages(page, cachep->gfporder);
1657 memcg_uncharge_slab(cachep, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658}
1659
1660static void kmem_rcu_free(struct rcu_head *head)
1661{
Joonsoo Kim68126702013-10-24 10:07:42 +09001662 struct kmem_cache *cachep;
1663 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664
Joonsoo Kim68126702013-10-24 10:07:42 +09001665 page = container_of(head, struct page, rcu_head);
1666 cachep = page->slab_cache;
1667
1668 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669}
1670
1671#if DEBUG
1672
1673#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001674static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001675 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001677 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001679 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001681 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 return;
1683
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001684 *addr++ = 0x12345678;
1685 *addr++ = caller;
1686 *addr++ = smp_processor_id();
1687 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 {
1689 unsigned long *sptr = &caller;
1690 unsigned long svalue;
1691
1692 while (!kstack_end(sptr)) {
1693 svalue = *sptr++;
1694 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001695 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 size -= sizeof(unsigned long);
1697 if (size <= sizeof(unsigned long))
1698 break;
1699 }
1700 }
1701
1702 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001703 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704}
1705#endif
1706
Pekka Enberg343e0d72006-02-01 03:05:50 -08001707static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001709 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001710 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711
1712 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001713 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714}
1715
1716static void dump_line(char *data, int offset, int limit)
1717{
1718 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001719 unsigned char error = 0;
1720 int bad_count = 0;
1721
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001722 printk(KERN_ERR "%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001723 for (i = 0; i < limit; i++) {
1724 if (data[offset + i] != POISON_FREE) {
1725 error = data[offset + i];
1726 bad_count++;
1727 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001728 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001729 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1730 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001731
1732 if (bad_count == 1) {
1733 error ^= POISON_FREE;
1734 if (!(error & (error - 1))) {
1735 printk(KERN_ERR "Single bit error detected. Probably "
1736 "bad RAM.\n");
1737#ifdef CONFIG_X86
1738 printk(KERN_ERR "Run memtest86+ or a similar memory "
1739 "test tool.\n");
1740#else
1741 printk(KERN_ERR "Run a memory test tool.\n");
1742#endif
1743 }
1744 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745}
1746#endif
1747
1748#if DEBUG
1749
Pekka Enberg343e0d72006-02-01 03:05:50 -08001750static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751{
1752 int i, size;
1753 char *realobj;
1754
1755 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001756 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001757 *dbg_redzone1(cachep, objp),
1758 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 }
1760
1761 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches071361d2012-12-12 10:19:12 -08001762 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1763 *dbg_userword(cachep, objp),
1764 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001766 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001767 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001768 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 int limit;
1770 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 if (i + limit > size)
1772 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 dump_line(realobj, i, limit);
1774 }
1775}
1776
Pekka Enberg343e0d72006-02-01 03:05:50 -08001777static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778{
1779 char *realobj;
1780 int size, i;
1781 int lines = 0;
1782
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001783 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001784 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001786 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001788 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 exp = POISON_END;
1790 if (realobj[i] != exp) {
1791 int limit;
1792 /* Mismatch ! */
1793 /* Print header */
1794 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001795 printk(KERN_ERR
Dave Jonesface37f2011-11-15 15:03:52 -08001796 "Slab corruption (%s): %s start=%p, len=%d\n",
1797 print_tainted(), cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798 print_objinfo(cachep, objp, 0);
1799 }
1800 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001801 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001803 if (i + limit > size)
1804 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 dump_line(realobj, i, limit);
1806 i += 16;
1807 lines++;
1808 /* Limit to 5 lines */
1809 if (lines > 5)
1810 break;
1811 }
1812 }
1813 if (lines != 0) {
1814 /* Print some data about the neighboring objects, if they
1815 * exist:
1816 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001817 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001818 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819
Joonsoo Kim8456a642013-10-24 10:07:49 +09001820 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001822 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001823 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001825 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 print_objinfo(cachep, objp, 2);
1827 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001828 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001829 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001830 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001832 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 print_objinfo(cachep, objp, 2);
1834 }
1835 }
1836}
1837#endif
1838
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001840static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1841 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001842{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 int i;
1844 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001845 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846
1847 if (cachep->flags & SLAB_POISON) {
1848#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001849 if (cachep->size % PAGE_SIZE == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08001850 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001852 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 else
1854 check_poison_obj(cachep, objp);
1855#else
1856 check_poison_obj(cachep, objp);
1857#endif
1858 }
1859 if (cachep->flags & SLAB_RED_ZONE) {
1860 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1861 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001862 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1864 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001865 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001868}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001870static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1871 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001872{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001873}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874#endif
1875
Randy Dunlap911851e2006-03-22 00:08:14 -08001876/**
1877 * slab_destroy - destroy and release all objects in a slab
1878 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001879 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001880 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001881 * Destroy all the objs in a slab page, and release the mem back to the system.
1882 * Before calling the slab page must have been unlinked from the cache. The
1883 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001884 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001885static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001886{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001887 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001888
Joonsoo Kim8456a642013-10-24 10:07:49 +09001889 freelist = page->freelist;
1890 slab_destroy_debugcheck(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
Joonsoo Kim68126702013-10-24 10:07:42 +09001892 struct rcu_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893
Joonsoo Kim68126702013-10-24 10:07:42 +09001894 /*
1895 * RCU free overloads the RCU head over the LRU.
1896 * slab_page has been overloeaded over the LRU,
1897 * however it is not used from now on so that
1898 * we can use it safely.
1899 */
1900 head = (void *)&page->rcu_head;
1901 call_rcu(head, kmem_rcu_free);
1902
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 } else {
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001904 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 }
Joonsoo Kim68126702013-10-24 10:07:42 +09001906
1907 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001908 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001909 * although actual page can be freed in rcu context
1910 */
1911 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001912 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913}
1914
Joonsoo Kim97654df2014-08-06 16:04:25 -07001915static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1916{
1917 struct page *page, *n;
1918
1919 list_for_each_entry_safe(page, n, list, lru) {
1920 list_del(&page->lru);
1921 slab_destroy(cachep, page);
1922 }
1923}
1924
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001926 * calculate_slab_order - calculate size (page order) of slabs
1927 * @cachep: pointer to the cache that is being created
1928 * @size: size of objects to be created in this cache.
1929 * @align: required alignment for the objects.
1930 * @flags: slab allocation flags
1931 *
1932 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001933 *
1934 * This could be made much more intelligent. For now, try to avoid using
1935 * high order pages for slabs. When the gfp() functions are more friendly
1936 * towards high-order requests, this should be changed.
1937 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001938static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001939 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001940{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001941 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001942 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001943 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001944
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001945 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001946 unsigned int num;
1947 size_t remainder;
1948
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001949 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001950 if (!num)
1951 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001952
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001953 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1954 if (num > SLAB_OBJ_MAX_NUM)
1955 break;
1956
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001957 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim03787302014-06-23 13:22:06 -07001958 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001959 /*
1960 * Max number of objs-per-slab for caches which
1961 * use off-slab slabs. Needed to avoid a possible
1962 * looping condition in cache_grow().
1963 */
Joonsoo Kim03787302014-06-23 13:22:06 -07001964 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1965 freelist_size_per_obj += sizeof(char);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001966 offslab_limit = size;
Joonsoo Kim03787302014-06-23 13:22:06 -07001967 offslab_limit /= freelist_size_per_obj;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001968
1969 if (num > offslab_limit)
1970 break;
1971 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001972
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001973 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001974 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001975 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001976 left_over = remainder;
1977
1978 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001979 * A VFS-reclaimable slab tends to have most allocations
1980 * as GFP_NOFS and we really don't want to have to be allocating
1981 * higher-order pages when we are unable to shrink dcache.
1982 */
1983 if (flags & SLAB_RECLAIM_ACCOUNT)
1984 break;
1985
1986 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001987 * Large number of objects is good, but very large slabs are
1988 * currently bad for the gfp()s.
1989 */
David Rientjes543585c2011-10-18 22:09:24 -07001990 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001991 break;
1992
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001993 /*
1994 * Acceptable internal fragmentation?
1995 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001996 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001997 break;
1998 }
1999 return left_over;
2000}
2001
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002002static struct array_cache __percpu *alloc_kmem_cache_cpus(
2003 struct kmem_cache *cachep, int entries, int batchcount)
2004{
2005 int cpu;
2006 size_t size;
2007 struct array_cache __percpu *cpu_cache;
2008
2009 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07002010 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002011
2012 if (!cpu_cache)
2013 return NULL;
2014
2015 for_each_possible_cpu(cpu) {
2016 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2017 entries, batchcount);
2018 }
2019
2020 return cpu_cache;
2021}
2022
Pekka Enberg83b519e2009-06-10 19:40:04 +03002023static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002024{
Christoph Lameter97d06602012-07-06 15:25:11 -05002025 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002026 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002027
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002028 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2029 if (!cachep->cpu_cache)
2030 return 1;
2031
Christoph Lameter97d06602012-07-06 15:25:11 -05002032 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002033 /* Creation of first cache (kmem_cache). */
2034 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00002035 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002036 /* For kmem_cache_node */
2037 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002038 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002039 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002040
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002041 for_each_online_node(node) {
2042 cachep->node[node] = kmalloc_node(
2043 sizeof(struct kmem_cache_node), gfp, node);
2044 BUG_ON(!cachep->node[node]);
2045 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002046 }
2047 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002048
Christoph Lameter6a673682013-01-10 19:14:19 +00002049 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002050 jiffies + REAPTIMEOUT_NODE +
2051 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002052
2053 cpu_cache_get(cachep)->avail = 0;
2054 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2055 cpu_cache_get(cachep)->batchcount = 1;
2056 cpu_cache_get(cachep)->touched = 0;
2057 cachep->batchcount = 1;
2058 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002059 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002060}
2061
Joonsoo Kim12220de2014-10-09 15:26:24 -07002062unsigned long kmem_cache_flags(unsigned long object_size,
2063 unsigned long flags, const char *name,
2064 void (*ctor)(void *))
2065{
2066 return flags;
2067}
2068
2069struct kmem_cache *
2070__kmem_cache_alias(const char *name, size_t size, size_t align,
2071 unsigned long flags, void (*ctor)(void *))
2072{
2073 struct kmem_cache *cachep;
2074
2075 cachep = find_mergeable(size, align, flags, name, ctor);
2076 if (cachep) {
2077 cachep->refcount++;
2078
2079 /*
2080 * Adjust the object sizes so that we clear
2081 * the complete object on kzalloc.
2082 */
2083 cachep->object_size = max_t(int, cachep->object_size, size);
2084 }
2085 return cachep;
2086}
2087
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002088/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002089 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002090 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 *
2093 * Returns a ptr to the cache on success, NULL on failure.
2094 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002095 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 * The flags are
2098 *
2099 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2100 * to catch references to uninitialised memory.
2101 *
2102 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2103 * for buffer overruns.
2104 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2106 * cacheline. This can be beneficial if you're counting cycles as closely
2107 * as davem.
2108 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002109int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002110__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002112 size_t left_over, freelist_size;
2113 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002114 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002115 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002116 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119#if FORCED_DEBUG
2120 /*
2121 * Enable redzoning and last user accounting, except for caches with
2122 * large objects, if the increased size would increase the object size
2123 * above the next power of two: caches with object sizes just above a
2124 * power of two have a significant amount of internal fragmentation.
2125 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002126 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2127 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002128 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 if (!(flags & SLAB_DESTROY_BY_RCU))
2130 flags |= SLAB_POISON;
2131#endif
2132 if (flags & SLAB_DESTROY_BY_RCU)
2133 BUG_ON(flags & SLAB_POISON);
2134#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135
Andrew Mortona737b3e2006-03-22 00:08:11 -08002136 /*
2137 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 * unaligned accesses for some archs when redzoning is used, and makes
2139 * sure any on-slab bufctl's are also correctly aligned.
2140 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002141 if (size & (BYTES_PER_WORD - 1)) {
2142 size += (BYTES_PER_WORD - 1);
2143 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 }
2145
David Woodhouse87a927c2007-07-04 21:26:44 -04002146 if (flags & SLAB_RED_ZONE) {
2147 ralign = REDZONE_ALIGN;
2148 /* If redzoning, ensure that the second redzone is suitably
2149 * aligned, by adjusting the object size accordingly. */
2150 size += REDZONE_ALIGN - 1;
2151 size &= ~(REDZONE_ALIGN - 1);
2152 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002153
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002154 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002155 if (ralign < cachep->align) {
2156 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002158 /* disable debug if necessary */
2159 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002160 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002161 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002162 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002164 cachep->align = ralign;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
Pekka Enberg83b519e2009-06-10 19:40:04 +03002166 if (slab_is_available())
2167 gfp = GFP_KERNEL;
2168 else
2169 gfp = GFP_NOWAIT;
2170
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
Pekka Enbergca5f9702006-09-25 23:31:25 -07002173 /*
2174 * Both debugging options require word-alignment which is calculated
2175 * into align above.
2176 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002179 cachep->obj_offset += sizeof(unsigned long long);
2180 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 }
2182 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002183 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002184 * the real object. But if the second red zone needs to be
2185 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002187 if (flags & SLAB_RED_ZONE)
2188 size += REDZONE_ALIGN;
2189 else
2190 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 }
2192#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002193 /*
2194 * To activate debug pagealloc, off-slab management is necessary
2195 * requirement. In early phase of initialization, small sized slab
2196 * doesn't get initialized so it would not be possible. So, we need
2197 * to check size >= 256. It guarantees that all necessary small
2198 * sized slab is initialized in current slab initialization sequence.
2199 */
2200 if (!slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2201 size >= 256 && cachep->object_size > cache_line_size() &&
2202 ALIGN(size, cachep->align) < PAGE_SIZE) {
Tetsuo Handa608da7e2012-09-30 17:28:25 +09002203 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 size = PAGE_SIZE;
2205 }
2206#endif
2207#endif
2208
Ingo Molnare0a42722006-06-23 02:03:46 -07002209 /*
2210 * Determine if the slab management is 'on' or 'off' slab.
2211 * (bootstrapping cannot cope with offslab caches so don't do
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002212 * it too early on. Always use on-slab management when
2213 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
Ingo Molnare0a42722006-06-23 02:03:46 -07002214 */
Joonsoo Kim8fc9cf42013-12-02 17:49:43 +09002215 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
Catalin Marinase7cb55b2009-10-28 13:33:08 +00002216 !(flags & SLAB_NOLEAKTRACE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 /*
2218 * Size is large, assume best to place the slab management obj
2219 * off-slab (should allow better packing of objs).
2220 */
2221 flags |= CFLGS_OFF_SLAB;
2222
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002223 size = ALIGN(size, cachep->align);
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09002224 /*
2225 * We should restrict the number of objects in a slab to implement
2226 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2227 */
2228 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2229 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002231 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002233 if (!cachep->num)
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002234 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002235
Joonsoo Kim03787302014-06-23 13:22:06 -07002236 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237
2238 /*
2239 * If the slab has been placed off-slab, and we have enough space then
2240 * move it on-slab. This is at the expense of any extra colouring.
2241 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002242 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 flags &= ~CFLGS_OFF_SLAB;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002244 left_over -= freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 }
2246
2247 if (flags & CFLGS_OFF_SLAB) {
2248 /* really off slab. No need for manual alignment */
Joonsoo Kim03787302014-06-23 13:22:06 -07002249 freelist_size = calculate_freelist_size(cachep->num, 0);
Ron Lee67461362009-05-22 04:58:22 +09302250
2251#ifdef CONFIG_PAGE_POISONING
2252 /* If we're going to use the generic kernel_map_pages()
2253 * poisoning, then it's going to smash the contents of
2254 * the redzone and userword anyhow, so switch them off.
2255 */
2256 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2257 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2258#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 }
2260
2261 cachep->colour_off = cache_line_size();
2262 /* Offset must be a multiple of the alignment. */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002263 if (cachep->colour_off < cachep->align)
2264 cachep->colour_off = cachep->align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002265 cachep->colour = left_over / cachep->colour_off;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002266 cachep->freelist_size = freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002268 cachep->allocflags = __GFP_COMP;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002269 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Glauber Costaa618e892012-06-14 16:17:21 +04002270 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002271 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002272 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002274 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002275 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002276 /*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002277 * This is a possibility for one of the kmalloc_{dma,}_caches.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002278 * But since we go off slab only for object size greater than
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002279 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2280 * in ascending order,this should not happen at all.
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002281 * But leave a BUG_ON for some lucky dude.
2282 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002283 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002284 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002286 err = setup_cpu_cache(cachep, gfp);
2287 if (err) {
Christoph Lameter12c36672012-09-04 23:38:33 +00002288 __kmem_cache_shutdown(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002289 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002290 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002292 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
2295#if DEBUG
2296static void check_irq_off(void)
2297{
2298 BUG_ON(!irqs_disabled());
2299}
2300
2301static void check_irq_on(void)
2302{
2303 BUG_ON(irqs_disabled());
2304}
2305
Pekka Enberg343e0d72006-02-01 03:05:50 -08002306static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307{
2308#ifdef CONFIG_SMP
2309 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002310 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311#endif
2312}
Christoph Lametere498be72005-09-09 13:03:32 -07002313
Pekka Enberg343e0d72006-02-01 03:05:50 -08002314static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002315{
2316#ifdef CONFIG_SMP
2317 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002318 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002319#endif
2320}
2321
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322#else
2323#define check_irq_off() do { } while(0)
2324#define check_irq_on() do { } while(0)
2325#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002326#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327#endif
2328
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002329static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameteraab22072006-03-22 00:09:06 -08002330 struct array_cache *ac,
2331 int force, int node);
2332
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333static void do_drain(void *arg)
2334{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002335 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002337 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002338 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002339 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340
2341 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002342 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002343 n = get_node(cachep, node);
2344 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002345 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002346 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002347 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 ac->avail = 0;
2349}
2350
Pekka Enberg343e0d72006-02-01 03:05:50 -08002351static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002353 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002354 int node;
2355
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002356 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002358 for_each_kmem_cache_node(cachep, node, n)
2359 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002360 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002361
Christoph Lameter18bf8542014-08-06 16:04:11 -07002362 for_each_kmem_cache_node(cachep, node, n)
2363 drain_array(cachep, n, n->shared, 1, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364}
2365
Christoph Lametered11d9e2006-06-30 01:55:45 -07002366/*
2367 * Remove slabs from the list of free slabs.
2368 * Specify the number of slabs to drain in tofree.
2369 *
2370 * Returns the actual number of slabs released.
2371 */
2372static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002373 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002375 struct list_head *p;
2376 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002377 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378
Christoph Lametered11d9e2006-06-30 01:55:45 -07002379 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002380 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002382 spin_lock_irq(&n->list_lock);
2383 p = n->slabs_free.prev;
2384 if (p == &n->slabs_free) {
2385 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002386 goto out;
2387 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
Joonsoo Kim8456a642013-10-24 10:07:49 +09002389 page = list_entry(p, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09002391 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002393 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002394 /*
2395 * Safe to drop the lock. The slab is no longer linked
2396 * to the cache.
2397 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002398 n->free_objects -= cache->num;
2399 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002400 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002401 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002403out:
2404 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405}
2406
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002407int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
Christoph Lametere498be72005-09-09 13:03:32 -07002408{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002409 int ret = 0;
2410 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002411 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002412
2413 drain_cpu_caches(cachep);
2414
2415 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002416 for_each_kmem_cache_node(cachep, node, n) {
Wanpeng Li0fa81032013-07-04 08:33:22 +08002417 drain_freelist(cachep, n, slabs_tofree(cachep, n));
Christoph Lametered11d9e2006-06-30 01:55:45 -07002418
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002419 ret += !list_empty(&n->slabs_full) ||
2420 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002421 }
2422 return (ret ? 1 : 0);
2423}
2424
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002425int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426{
Christoph Lameter12c36672012-09-04 23:38:33 +00002427 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002428 struct kmem_cache_node *n;
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002429 int rc = __kmem_cache_shrink(cachep, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430
Christoph Lameter12c36672012-09-04 23:38:33 +00002431 if (rc)
2432 return rc;
2433
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002434 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002435
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002436 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002437 for_each_kmem_cache_node(cachep, i, n) {
2438 kfree(n->shared);
2439 free_alien_cache(n->alien);
2440 kfree(n);
2441 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 }
Christoph Lameter12c36672012-09-04 23:38:33 +00002443 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002446/*
2447 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002448 *
2449 * For a slab cache when the slab descriptor is off-slab, the
2450 * slab descriptor can't come from the same cache which is being created,
2451 * Because if it is the case, that means we defer the creation of
2452 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2453 * And we eventually call down to __kmem_cache_create(), which
2454 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2455 * This is a "chicken-and-egg" problem.
2456 *
2457 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2458 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002459 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002460static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002461 struct page *page, int colour_off,
2462 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002464 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002465 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002466
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 if (OFF_SLAB(cachep)) {
2468 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002469 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002470 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002471 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 return NULL;
2473 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002474 freelist = addr + colour_off;
2475 colour_off += cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09002477 page->active = 0;
2478 page->s_mem = addr + colour_off;
2479 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480}
2481
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002482static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002484 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002485}
2486
2487static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002488 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002489{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002490 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491}
2492
Pekka Enberg343e0d72006-02-01 03:05:50 -08002493static void cache_init_objs(struct kmem_cache *cachep,
Joonsoo Kim8456a642013-10-24 10:07:49 +09002494 struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495{
2496 int i;
2497
2498 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002499 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500#if DEBUG
2501 /* need to poison the objs? */
2502 if (cachep->flags & SLAB_POISON)
2503 poison_obj(cachep, objp, POISON_FREE);
2504 if (cachep->flags & SLAB_STORE_USER)
2505 *dbg_userword(cachep, objp) = NULL;
2506
2507 if (cachep->flags & SLAB_RED_ZONE) {
2508 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2509 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2510 }
2511 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002512 * Constructors are not allowed to allocate memory from the same
2513 * cache which they are a constructor for. Otherwise, deadlock.
2514 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515 */
2516 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002517 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
2519 if (cachep->flags & SLAB_RED_ZONE) {
2520 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2521 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002522 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2524 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002525 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 }
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002527 if ((cachep->size % PAGE_SIZE) == 0 &&
Andrew Mortona737b3e2006-03-22 00:08:11 -08002528 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002529 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002530 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531#else
2532 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002533 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534#endif
Joonsoo Kim03787302014-06-23 13:22:06 -07002535 set_obj_status(page, i, OBJECT_FREE);
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002536 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538}
2539
Pekka Enberg343e0d72006-02-01 03:05:50 -08002540static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002542 if (CONFIG_ZONE_DMA_FLAG) {
2543 if (flags & GFP_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002544 BUG_ON(!(cachep->allocflags & GFP_DMA));
Christoph Lameter4b51d662007-02-10 01:43:10 -08002545 else
Glauber Costaa618e892012-06-14 16:17:21 +04002546 BUG_ON(cachep->allocflags & GFP_DMA);
Christoph Lameter4b51d662007-02-10 01:43:10 -08002547 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548}
2549
Joonsoo Kim8456a642013-10-24 10:07:49 +09002550static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002551 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002552{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002553 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002554
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002555 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002556 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002557#if DEBUG
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002558 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002559#endif
Matthew Dobson78d382d2006-02-01 03:05:47 -08002560
2561 return objp;
2562}
2563
Joonsoo Kim8456a642013-10-24 10:07:49 +09002564static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002565 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002566{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002567 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002568#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002569 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002570
Matthew Dobson78d382d2006-02-01 03:05:47 -08002571 /* Verify that the slab belongs to the intended node */
Joonsoo Kim1ea991b2013-10-24 10:07:40 +09002572 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002573
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002574 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002575 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002576 if (get_free_obj(page, i) == objnr) {
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002577 printk(KERN_ERR "slab: double free detected in cache "
2578 "'%s', objp %p\n", cachep->name, objp);
2579 BUG();
2580 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002581 }
2582#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002583 page->active--;
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002584 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002585}
2586
Pekka Enberg47768742006-06-23 02:03:07 -07002587/*
2588 * Map pages beginning at addr to the given cache and slab. This is required
2589 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002590 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002591 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002592static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002593 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002595 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002596 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597}
2598
2599/*
2600 * Grow (by 1) the number of slabs within a cache. This is called by
2601 * kmem_cache_alloc() when there are no active objs left in a cache.
2602 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002603static int cache_grow(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002604 gfp_t flags, int nodeid, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002606 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002607 size_t offset;
2608 gfp_t local_flags;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002609 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610
Andrew Mortona737b3e2006-03-22 00:08:11 -08002611 /*
2612 * Be lazy and only check for valid flags here, keeping it out of the
2613 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002615 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2616 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2617 BUG();
2618 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002619 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002621 /* Take the node list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002623 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002624 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625
2626 /* Get colour for the slab, and cal the next value. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002627 offset = n->colour_next;
2628 n->colour_next++;
2629 if (n->colour_next >= cachep->colour)
2630 n->colour_next = 0;
2631 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002633 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
2635 if (local_flags & __GFP_WAIT)
2636 local_irq_enable();
2637
2638 /*
2639 * The test for missing atomic flag is performed here, rather than
2640 * the more obvious place, simply to reduce the critical path length
2641 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2642 * will eventually be caught here (where it matters).
2643 */
2644 kmem_flagcheck(cachep, flags);
2645
Andrew Mortona737b3e2006-03-22 00:08:11 -08002646 /*
2647 * Get mem for the objs. Attempt to allocate a physical page from
2648 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002649 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002650 if (!page)
2651 page = kmem_getpages(cachep, local_flags, nodeid);
2652 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 goto failed;
2654
2655 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002656 freelist = alloc_slabmgmt(cachep, page, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002657 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002658 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 goto opps1;
2660
Joonsoo Kim8456a642013-10-24 10:07:49 +09002661 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662
Joonsoo Kim8456a642013-10-24 10:07:49 +09002663 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664
2665 if (local_flags & __GFP_WAIT)
2666 local_irq_disable();
2667 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002668 spin_lock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669
2670 /* Make slab active. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002671 list_add_tail(&page->lru, &(n->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672 STATS_INC_GROWN(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002673 n->free_objects += cachep->num;
2674 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002676opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002677 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002678failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 if (local_flags & __GFP_WAIT)
2680 local_irq_disable();
2681 return 0;
2682}
2683
2684#if DEBUG
2685
2686/*
2687 * Perform extra freeing checks:
2688 * - detect bad pointers.
2689 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 */
2691static void kfree_debugcheck(const void *objp)
2692{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 if (!virt_addr_valid(objp)) {
2694 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002695 (unsigned long)objp);
2696 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698}
2699
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002700static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2701{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002702 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002703
2704 redzone1 = *dbg_redzone1(cache, obj);
2705 redzone2 = *dbg_redzone2(cache, obj);
2706
2707 /*
2708 * Redzone is ok.
2709 */
2710 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2711 return;
2712
2713 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2714 slab_error(cache, "double free detected");
2715 else
2716 slab_error(cache, "memory outside object was overwritten");
2717
David Woodhouseb46b8f12007-05-08 00:22:59 -07002718 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002719 obj, redzone1, redzone2);
2720}
2721
Pekka Enberg343e0d72006-02-01 03:05:50 -08002722static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002723 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002726 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002728 BUG_ON(virt_to_cache(objp) != cachep);
2729
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002730 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002732 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002735 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2737 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2738 }
2739 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002740 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
Joonsoo Kim8456a642013-10-24 10:07:49 +09002742 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743
2744 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002745 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
Joonsoo Kim03787302014-06-23 13:22:06 -07002747 set_obj_status(page, objnr, OBJECT_FREE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 if (cachep->flags & SLAB_POISON) {
2749#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002750 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002751 store_stackinfo(cachep, objp, caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002752 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002753 cachep->size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 } else {
2755 poison_obj(cachep, objp, POISON_FREE);
2756 }
2757#else
2758 poison_obj(cachep, objp, POISON_FREE);
2759#endif
2760 }
2761 return objp;
2762}
2763
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764#else
2765#define kfree_debugcheck(x) do { } while(0)
2766#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767#endif
2768
Mel Gorman072bb0a2012-07-31 16:43:58 -07002769static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2770 bool force_refill)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771{
2772 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002773 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002775 int node;
2776
Joe Korty6d2144d2008-03-05 15:04:59 -08002777 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002778 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002779 if (unlikely(force_refill))
2780 goto force_grow;
2781retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002782 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 batchcount = ac->batchcount;
2784 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002785 /*
2786 * If there was little recent activity on this cache, then
2787 * perform only a partial refill. Otherwise we could generate
2788 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 */
2790 batchcount = BATCHREFILL_LIMIT;
2791 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07002792 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002794 BUG_ON(ac->avail > 0 || !n);
2795 spin_lock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002796
Christoph Lameter3ded1752006-03-25 03:06:44 -08002797 /* See if we can refill from the shared array */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002798 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2799 n->shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08002800 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11002801 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08002802
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 while (batchcount > 0) {
2804 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002805 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 /* Get slab alloc is to come from. */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002807 entry = n->slabs_partial.next;
2808 if (entry == &n->slabs_partial) {
2809 n->free_touched = 1;
2810 entry = n->slabs_free.next;
2811 if (entry == &n->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 goto must_grow;
2813 }
2814
Joonsoo Kim8456a642013-10-24 10:07:49 +09002815 page = list_entry(entry, struct page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002817
2818 /*
2819 * The slab was either on partial or free list so
2820 * there must be at least one object available for
2821 * allocation.
2822 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002823 BUG_ON(page->active >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002824
Joonsoo Kim8456a642013-10-24 10:07:49 +09002825 while (page->active < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 STATS_INC_ALLOCED(cachep);
2827 STATS_INC_ACTIVE(cachep);
2828 STATS_SET_HIGH(cachep);
2829
Joonsoo Kim8456a642013-10-24 10:07:49 +09002830 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
Mel Gorman072bb0a2012-07-31 16:43:58 -07002831 node));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833
2834 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002835 list_del(&page->lru);
2836 if (page->active == cachep->num)
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002837 list_add(&page->lru, &n->slabs_full);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838 else
Dave Hansen34bf6ef2014-04-08 13:44:27 -07002839 list_add(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 }
2841
Andrew Mortona737b3e2006-03-22 00:08:11 -08002842must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002843 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002844alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002845 spin_unlock(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846
2847 if (unlikely(!ac->avail)) {
2848 int x;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002849force_grow:
David Rientjes4167e9b2015-04-14 15:46:55 -07002850 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07002851
Andrew Mortona737b3e2006-03-22 00:08:11 -08002852 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002853 ac = cpu_cache_get(cachep);
David Rientjes51cd8e62012-08-28 19:57:21 -07002854 node = numa_mem_id();
Mel Gorman072bb0a2012-07-31 16:43:58 -07002855
2856 /* no objects in sight? abort */
2857 if (!x && (ac->avail == 0 || force_refill))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 return NULL;
2859
Andrew Mortona737b3e2006-03-22 00:08:11 -08002860 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 goto retry;
2862 }
2863 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002864
2865 return ac_get_obj(cachep, ac, flags, force_refill);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866}
2867
Andrew Mortona737b3e2006-03-22 00:08:11 -08002868static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2869 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870{
2871 might_sleep_if(flags & __GFP_WAIT);
2872#if DEBUG
2873 kmem_flagcheck(cachep, flags);
2874#endif
2875}
2876
2877#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002878static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002879 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880{
Joonsoo Kim03787302014-06-23 13:22:06 -07002881 struct page *page;
2882
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002885 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886#ifdef CONFIG_DEBUG_PAGEALLOC
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002887 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002888 kernel_map_pages(virt_to_page(objp),
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002889 cachep->size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890 else
2891 check_poison_obj(cachep, objp);
2892#else
2893 check_poison_obj(cachep, objp);
2894#endif
2895 poison_obj(cachep, objp, POISON_INUSE);
2896 }
2897 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002898 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899
2900 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002901 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2902 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2903 slab_error(cachep, "double free, or memory outside"
2904 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002905 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07002906 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08002907 objp, *dbg_redzone1(cachep, objp),
2908 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 }
2910 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2911 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2912 }
Joonsoo Kim03787302014-06-23 13:22:06 -07002913
2914 page = virt_to_head_page(objp);
2915 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002916 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07002917 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002918 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09002919 if (ARCH_SLAB_MINALIGN &&
2920 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002921 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07002922 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002923 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 return objp;
2925}
2926#else
2927#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2928#endif
2929
Akinobu Mita773ff602008-12-23 19:37:01 +09002930static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002931{
Joonsoo Kim8a9c61d2014-08-06 16:04:20 -07002932 if (unlikely(cachep == kmem_cache))
Akinobu Mita773ff602008-12-23 19:37:01 +09002933 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002934
Christoph Lameter8c138bc2012-06-13 10:24:58 -05002935 return should_failslab(cachep->object_size, flags, cachep->flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002936}
2937
Pekka Enberg343e0d72006-02-01 03:05:50 -08002938static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002940 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 struct array_cache *ac;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002942 bool force_refill = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943
Alok N Kataria5c382302005-09-27 21:45:46 -07002944 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08002945
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002946 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07002949 objp = ac_get_obj(cachep, ac, flags, false);
2950
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002951 /*
Mel Gorman072bb0a2012-07-31 16:43:58 -07002952 * Allow for the possibility all avail objects are not allowed
2953 * by the current flags
J. R. Okajimaddbf2e82009-12-02 16:55:50 +09002954 */
Mel Gorman072bb0a2012-07-31 16:43:58 -07002955 if (objp) {
2956 STATS_INC_ALLOCHIT(cachep);
2957 goto out;
2958 }
2959 force_refill = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07002961
2962 STATS_INC_ALLOCMISS(cachep);
2963 objp = cache_alloc_refill(cachep, flags, force_refill);
2964 /*
2965 * the 'ac' may be updated by cache_alloc_refill(),
2966 * and kmemleak_erase() requires its correct value.
2967 */
2968 ac = cpu_cache_get(cachep);
2969
2970out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01002971 /*
2972 * To avoid a false negative, if an object that is in one of the
2973 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2974 * treat the array pointers as a reference to the object.
2975 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09002976 if (objp)
2977 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07002978 return objp;
2979}
2980
Christoph Lametere498be72005-09-09 13:03:32 -07002981#ifdef CONFIG_NUMA
2982/*
Zefan Li2ad654b2014-09-25 09:41:02 +08002983 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002984 *
2985 * If we are in_interrupt, then process context, including cpusets and
2986 * mempolicy, may not apply and should not be used for allocation policy.
2987 */
2988static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2989{
2990 int nid_alloc, nid_here;
2991
Christoph Lameter765c4502006-09-27 01:50:08 -07002992 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08002993 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002994 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002995 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07002996 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002997 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07002998 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08002999 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003000 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003001 return NULL;
3002}
3003
3004/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003005 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003006 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003007 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003008 * perform an allocation without specifying a node. This allows the page
3009 * allocator to do its reclaim / fallback magic. We then insert the
3010 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003011 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003012static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003013{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003014 struct zonelist *zonelist;
3015 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003016 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003017 struct zone *zone;
3018 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003019 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003020 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003021 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003022
3023 if (flags & __GFP_THISNODE)
3024 return NULL;
3025
Christoph Lameter6cb06222007-10-16 01:25:41 -07003026 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003027
Mel Gormancc9a6c82012-03-21 16:34:11 -07003028retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003029 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003030 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003031
Christoph Lameter3c517a62006-12-06 20:33:29 -08003032retry:
3033 /*
3034 * Look through allowed nodes for objects available
3035 * from existing per node queues.
3036 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003037 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3038 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003039
Vladimir Davydov061d7072014-12-12 16:58:25 -08003040 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003041 get_node(cache, nid) &&
3042 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003043 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003044 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003045 if (obj)
3046 break;
3047 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003048 }
3049
Christoph Lametercfce6602007-05-06 14:50:17 -07003050 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003051 /*
3052 * This allocation will be performed within the constraints
3053 * of the current cpuset / memory policy requirements.
3054 * We may trigger various forms of reclaim on the allowed
3055 * set and go into memory reserves if necessary.
3056 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003057 struct page *page;
3058
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003059 if (local_flags & __GFP_WAIT)
3060 local_irq_enable();
3061 kmem_flagcheck(cache, flags);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003062 page = kmem_getpages(cache, local_flags, numa_mem_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003063 if (local_flags & __GFP_WAIT)
3064 local_irq_disable();
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003065 if (page) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003066 /*
3067 * Insert into the appropriate per node queues
3068 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003069 nid = page_to_nid(page);
3070 if (cache_grow(cache, flags, nid, page)) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003071 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003072 gfp_exact_node(flags), nid);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003073 if (!obj)
3074 /*
3075 * Another processor may allocate the
3076 * objects in the slab since we are
3077 * not holding any locks.
3078 */
3079 goto retry;
3080 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003081 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003082 obj = NULL;
3083 }
3084 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003085 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003086
Mel Gormand26914d2014-04-03 14:47:24 -07003087 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003088 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003089 return obj;
3090}
3091
3092/*
Christoph Lametere498be72005-09-09 13:03:32 -07003093 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003095static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003096 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003097{
3098 struct list_head *entry;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003099 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003100 struct kmem_cache_node *n;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003101 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003102 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003104 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003105 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003106 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003107
Andrew Mortona737b3e2006-03-22 00:08:11 -08003108retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003109 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003110 spin_lock(&n->list_lock);
3111 entry = n->slabs_partial.next;
3112 if (entry == &n->slabs_partial) {
3113 n->free_touched = 1;
3114 entry = n->slabs_free.next;
3115 if (entry == &n->slabs_free)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003116 goto must_grow;
3117 }
Christoph Lametere498be72005-09-09 13:03:32 -07003118
Joonsoo Kim8456a642013-10-24 10:07:49 +09003119 page = list_entry(entry, struct page, lru);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003120 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003121
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003122 STATS_INC_NODEALLOCS(cachep);
3123 STATS_INC_ACTIVE(cachep);
3124 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003125
Joonsoo Kim8456a642013-10-24 10:07:49 +09003126 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003127
Joonsoo Kim8456a642013-10-24 10:07:49 +09003128 obj = slab_get_obj(cachep, page, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003129 n->free_objects--;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003130 /* move slabp to correct slabp list: */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003131 list_del(&page->lru);
Christoph Lametere498be72005-09-09 13:03:32 -07003132
Joonsoo Kim8456a642013-10-24 10:07:49 +09003133 if (page->active == cachep->num)
3134 list_add(&page->lru, &n->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003135 else
Joonsoo Kim8456a642013-10-24 10:07:49 +09003136 list_add(&page->lru, &n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003137
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003138 spin_unlock(&n->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003139 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003140
Andrew Mortona737b3e2006-03-22 00:08:11 -08003141must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003142 spin_unlock(&n->list_lock);
David Rientjes4167e9b2015-04-14 15:46:55 -07003143 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003144 if (x)
3145 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003146
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003147 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003148
Andrew Mortona737b3e2006-03-22 00:08:11 -08003149done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003150 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003151}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003152
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003153static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003154slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003155 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003156{
3157 unsigned long save_flags;
3158 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003159 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003160
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003161 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003162
Nick Piggincf40bd12009-01-21 08:12:39 +01003163 lockdep_trace_alloc(flags);
3164
Akinobu Mita773ff602008-12-23 19:37:01 +09003165 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003166 return NULL;
3167
Glauber Costad79923f2012-12-18 14:22:48 -08003168 cachep = memcg_kmem_get_cache(cachep, flags);
3169
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003170 cache_alloc_debugcheck_before(cachep, flags);
3171 local_irq_save(save_flags);
3172
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003173 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003174 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003175
Christoph Lameter18bf8542014-08-06 16:04:11 -07003176 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003177 /* Node not bootstrapped yet */
3178 ptr = fallback_alloc(cachep, flags);
3179 goto out;
3180 }
3181
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003182 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003183 /*
3184 * Use the locally cached objects if possible.
3185 * However ____cache_alloc does not allow fallback
3186 * to other nodes. It may fail while we still have
3187 * objects on other nodes available.
3188 */
3189 ptr = ____cache_alloc(cachep, flags);
3190 if (ptr)
3191 goto out;
3192 }
3193 /* ___cache_alloc_node can fall back to other nodes */
3194 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3195 out:
3196 local_irq_restore(save_flags);
3197 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003198 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003199 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003200
Joe Perches5087c822013-09-10 17:02:51 -07003201 if (likely(ptr)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003202 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003203 if (unlikely(flags & __GFP_ZERO))
3204 memset(ptr, 0, cachep->object_size);
3205 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003206
Vladimir Davydov8135be52014-12-12 16:56:38 -08003207 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003208 return ptr;
3209}
3210
3211static __always_inline void *
3212__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3213{
3214 void *objp;
3215
Zefan Li2ad654b2014-09-25 09:41:02 +08003216 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003217 objp = alternate_node_alloc(cache, flags);
3218 if (objp)
3219 goto out;
3220 }
3221 objp = ____cache_alloc(cache, flags);
3222
3223 /*
3224 * We may just have run out of memory on the local node.
3225 * ____cache_alloc_node() knows how to locate memory on other nodes
3226 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003227 if (!objp)
3228 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003229
3230 out:
3231 return objp;
3232}
3233#else
3234
3235static __always_inline void *
3236__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3237{
3238 return ____cache_alloc(cachep, flags);
3239}
3240
3241#endif /* CONFIG_NUMA */
3242
3243static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003244slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003245{
3246 unsigned long save_flags;
3247 void *objp;
3248
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003249 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003250
Nick Piggincf40bd12009-01-21 08:12:39 +01003251 lockdep_trace_alloc(flags);
3252
Akinobu Mita773ff602008-12-23 19:37:01 +09003253 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003254 return NULL;
3255
Glauber Costad79923f2012-12-18 14:22:48 -08003256 cachep = memcg_kmem_get_cache(cachep, flags);
3257
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003258 cache_alloc_debugcheck_before(cachep, flags);
3259 local_irq_save(save_flags);
3260 objp = __do_cache_alloc(cachep, flags);
3261 local_irq_restore(save_flags);
3262 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003263 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
Catalin Marinasd5cff632009-06-11 13:22:40 +01003264 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003265 prefetchw(objp);
3266
Joe Perches5087c822013-09-10 17:02:51 -07003267 if (likely(objp)) {
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003268 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
Joe Perches5087c822013-09-10 17:02:51 -07003269 if (unlikely(flags & __GFP_ZERO))
3270 memset(objp, 0, cachep->object_size);
3271 }
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003272
Vladimir Davydov8135be52014-12-12 16:56:38 -08003273 memcg_kmem_put_cache(cachep);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003274 return objp;
3275}
Christoph Lametere498be72005-09-09 13:03:32 -07003276
3277/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003278 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003279 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003280 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003281static void free_block(struct kmem_cache *cachep, void **objpp,
3282 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283{
3284 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003285 struct kmem_cache_node *n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286
3287 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003288 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003289 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
Mel Gorman072bb0a2012-07-31 16:43:58 -07003291 clear_obj_pfmemalloc(&objpp[i]);
3292 objp = objpp[i];
3293
Joonsoo Kim8456a642013-10-24 10:07:49 +09003294 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003295 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003296 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003297 slab_put_obj(cachep, page, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 STATS_DEC_ACTIVE(cachep);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003299 n->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300
3301 /* fixup slab chains */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003302 if (page->active == 0) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003303 if (n->free_objects > n->free_limit) {
3304 n->free_objects -= cachep->num;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003305 list_add_tail(&page->lru, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 } else {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003307 list_add(&page->lru, &n->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 }
3309 } else {
3310 /* Unconditionally move a slab to the end of the
3311 * partial list on free - maximum time for the
3312 * other objects to be freed, too.
3313 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003314 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 }
3316 }
3317}
3318
Pekka Enberg343e0d72006-02-01 03:05:50 -08003319static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320{
3321 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003322 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003323 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003324 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325
3326 batchcount = ac->batchcount;
3327#if DEBUG
3328 BUG_ON(!batchcount || batchcount > ac->avail);
3329#endif
3330 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003331 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003332 spin_lock(&n->list_lock);
3333 if (n->shared) {
3334 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003335 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336 if (max) {
3337 if (batchcount > max)
3338 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003339 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003340 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341 shared_array->avail += batchcount;
3342 goto free_done;
3343 }
3344 }
3345
Joonsoo Kim97654df2014-08-06 16:04:25 -07003346 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003347free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348#if STATS
3349 {
3350 int i = 0;
3351 struct list_head *p;
3352
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003353 p = n->slabs_free.next;
3354 while (p != &(n->slabs_free)) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003355 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356
Joonsoo Kim8456a642013-10-24 10:07:49 +09003357 page = list_entry(p, struct page, lru);
3358 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359
3360 i++;
3361 p = p->next;
3362 }
3363 STATS_SET_FREEABLE(cachep, i);
3364 }
3365#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003366 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003367 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003369 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370}
3371
3372/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003373 * Release an obj back to its cache. If the obj has a constructed state, it must
3374 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003376static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003377 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003379 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380
3381 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003382 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003383 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003385 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003386
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003387 /*
3388 * Skip calling cache_free_alien() when the platform is not numa.
3389 * This will avoid cache misses that happen while accessing slabp (which
3390 * is per page memory reference) to get nodeid. Instead use a global
3391 * variable to skip the call, which is mostly likely to be present in
3392 * the cache.
3393 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003394 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003395 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003396
Joonsoo Kim3d880192014-10-09 15:26:04 -07003397 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 } else {
3400 STATS_INC_FREEMISS(cachep);
3401 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003403
Mel Gorman072bb0a2012-07-31 16:43:58 -07003404 ac_put_obj(cachep, ac, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405}
3406
3407/**
3408 * kmem_cache_alloc - Allocate an object
3409 * @cachep: The cache to allocate from.
3410 * @flags: See kmalloc().
3411 *
3412 * Allocate an object from this cache. The flags are only relevant
3413 * if the cache has no available objects.
3414 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003415void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003417 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003418
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003419 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003420 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003421
3422 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423}
3424EXPORT_SYMBOL(kmem_cache_alloc);
3425
Christoph Lameter484748f2015-09-04 15:45:34 -07003426void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3427{
3428 __kmem_cache_free_bulk(s, size, p);
3429}
3430EXPORT_SYMBOL(kmem_cache_free_bulk);
3431
3432bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3433 void **p)
3434{
3435 return __kmem_cache_alloc_bulk(s, flags, size, p);
3436}
3437EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3438
Li Zefan0f24f122009-12-11 15:45:30 +08003439#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003440void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003441kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003442{
Steven Rostedt85beb582010-11-24 16:23:34 -05003443 void *ret;
3444
Ezequiel Garcia48356302012-09-08 17:47:57 -03003445 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003446
3447 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003448 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003449 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003450}
Steven Rostedt85beb582010-11-24 16:23:34 -05003451EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003452#endif
3453
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003455/**
3456 * kmem_cache_alloc_node - Allocate an object on the specified node
3457 * @cachep: The cache to allocate from.
3458 * @flags: See kmalloc().
3459 * @nodeid: node number of the target node.
3460 *
3461 * Identical to kmem_cache_alloc but it will allocate memory on the given
3462 * node, which can improve the performance for cpu bound structures.
3463 *
3464 * Fallback to other node is possible if __GFP_THISNODE is not set.
3465 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003466void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3467{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003468 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003469
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003470 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003471 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003472 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003473
3474 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003475}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476EXPORT_SYMBOL(kmem_cache_alloc_node);
3477
Li Zefan0f24f122009-12-11 15:45:30 +08003478#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003479void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003480 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003481 int nodeid,
3482 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003483{
Steven Rostedt85beb582010-11-24 16:23:34 -05003484 void *ret;
3485
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003486 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003487
Steven Rostedt85beb582010-11-24 16:23:34 -05003488 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003489 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003490 flags, nodeid);
3491 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003492}
Steven Rostedt85beb582010-11-24 16:23:34 -05003493EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003494#endif
3495
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003496static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003497__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003498{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003499 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003500
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003501 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003502 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3503 return cachep;
Ezequiel Garcia40521472012-09-08 17:47:56 -03003504 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003505}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003506
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003507void *__kmalloc_node(size_t size, gfp_t flags, int node)
3508{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003509 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003510}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003511EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003512
3513void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003514 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003515{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003516 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003517}
3518EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003519#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520
3521/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003522 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003524 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003525 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003527static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003528 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003530 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003531 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003533 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003534 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3535 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003536 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003537
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003538 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003539 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003540
3541 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003542}
3543
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003544void *__kmalloc(size_t size, gfp_t flags)
3545{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003546 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547}
3548EXPORT_SYMBOL(__kmalloc);
3549
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003550void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003551{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003552 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003553}
3554EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003555
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556/**
3557 * kmem_cache_free - Deallocate an object
3558 * @cachep: The cache the allocation was from.
3559 * @objp: The previously allocated object.
3560 *
3561 * Free an object which was previously allocated from this
3562 * cache.
3563 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003564void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565{
3566 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003567 cachep = cache_from_obj(cachep, objp);
3568 if (!cachep)
3569 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570
3571 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003572 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003573 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003574 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003575 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003577
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003578 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579}
3580EXPORT_SYMBOL(kmem_cache_free);
3581
3582/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 * kfree - free previously allocated memory
3584 * @objp: pointer returned by kmalloc.
3585 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003586 * If @objp is NULL, no operation is performed.
3587 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 * Don't free memory not originally allocated by kmalloc()
3589 * or you will run into trouble.
3590 */
3591void kfree(const void *objp)
3592{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003593 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 unsigned long flags;
3595
Pekka Enberg2121db72009-03-25 11:05:57 +02003596 trace_kfree(_RET_IP_, objp);
3597
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003598 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 return;
3600 local_irq_save(flags);
3601 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003602 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003603 debug_check_no_locks_freed(objp, c->object_size);
3604
3605 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003606 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 local_irq_restore(flags);
3608}
3609EXPORT_SYMBOL(kfree);
3610
Christoph Lametere498be72005-09-09 13:03:32 -07003611/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003612 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003613 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003614static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003615{
3616 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003617 struct kmem_cache_node *n;
Christoph Lametercafeb022006-03-25 03:06:46 -08003618 struct array_cache *new_shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -07003619 struct alien_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003620
Mel Gorman9c09a952008-01-24 05:49:54 -08003621 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003622
LQYMGTb455def2014-12-10 15:42:13 -08003623 if (use_alien_caches) {
3624 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3625 if (!new_alien)
3626 goto fail;
3627 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003628
Eric Dumazet63109842007-05-06 14:49:28 -07003629 new_shared = NULL;
3630 if (cachep->shared) {
3631 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003632 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003633 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003634 if (!new_shared) {
3635 free_alien_cache(new_alien);
3636 goto fail;
3637 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003638 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003639
Christoph Lameter18bf8542014-08-06 16:04:11 -07003640 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003641 if (n) {
3642 struct array_cache *shared = n->shared;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003643 LIST_HEAD(list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003644
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003645 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003646
Christoph Lametercafeb022006-03-25 03:06:46 -08003647 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003648 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07003649 shared->avail, node, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003650
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003651 n->shared = new_shared;
3652 if (!n->alien) {
3653 n->alien = new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003654 new_alien = NULL;
3655 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003656 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003657 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003658 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003659 slabs_destroy(cachep, &list);
Christoph Lametercafeb022006-03-25 03:06:46 -08003660 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003661 free_alien_cache(new_alien);
3662 continue;
3663 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003664 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3665 if (!n) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003666 free_alien_cache(new_alien);
3667 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003668 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003669 }
Christoph Lametere498be72005-09-09 13:03:32 -07003670
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003671 kmem_cache_node_init(n);
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003672 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3673 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003674 n->shared = new_shared;
3675 n->alien = new_alien;
3676 n->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003677 cachep->batchcount + cachep->num;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003678 cachep->node[node] = n;
Christoph Lametere498be72005-09-09 13:03:32 -07003679 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003680 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003681
Andrew Mortona737b3e2006-03-22 00:08:11 -08003682fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003683 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003684 /* Cache is not active yet. Roll back what we did */
3685 node--;
3686 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003687 n = get_node(cachep, node);
3688 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003689 kfree(n->shared);
3690 free_alien_cache(n->alien);
3691 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003692 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003693 }
3694 node--;
3695 }
3696 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003697 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003698}
3699
Christoph Lameter18004c52012-07-06 15:25:12 -05003700/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003701static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003702 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003704 struct array_cache __percpu *cpu_cache, *prev;
3705 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003707 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3708 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003709 return -ENOMEM;
3710
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003711 prev = cachep->cpu_cache;
3712 cachep->cpu_cache = cpu_cache;
3713 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003714
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716 cachep->batchcount = batchcount;
3717 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003718 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003720 if (!prev)
3721 goto alloc_node;
3722
3723 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003724 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003725 int node;
3726 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003727 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003728
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003729 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003730 n = get_node(cachep, node);
3731 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003732 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003733 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003734 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003736 free_percpu(prev);
3737
3738alloc_node:
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003739 return alloc_kmem_cache_node(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740}
3741
Glauber Costa943a4512012-12-18 14:23:03 -08003742static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3743 int batchcount, int shared, gfp_t gfp)
3744{
3745 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003746 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003747
3748 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3749
3750 if (slab_state < FULL)
3751 return ret;
3752
3753 if ((ret < 0) || !is_root_cache(cachep))
3754 return ret;
3755
Vladimir Davydov426589f2015-02-12 14:59:23 -08003756 lockdep_assert_held(&slab_mutex);
3757 for_each_memcg_cache(c, cachep) {
3758 /* return value determined by the root cache only */
3759 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003760 }
3761
3762 return ret;
3763}
3764
Christoph Lameter18004c52012-07-06 15:25:12 -05003765/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003766static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767{
3768 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003769 int limit = 0;
3770 int shared = 0;
3771 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772
Glauber Costa943a4512012-12-18 14:23:03 -08003773 if (!is_root_cache(cachep)) {
3774 struct kmem_cache *root = memcg_root_cache(cachep);
3775 limit = root->limit;
3776 shared = root->shared;
3777 batchcount = root->batchcount;
3778 }
3779
3780 if (limit && shared && batchcount)
3781 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003782 /*
3783 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784 * - create a LIFO ordering, i.e. return objects that are cache-warm
3785 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003786 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 * bufctl chains: array operations are cheaper.
3788 * The numbers are guessed, we should auto-tune as described by
3789 * Bonwick.
3790 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003791 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003793 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003795 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003797 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 limit = 54;
3799 else
3800 limit = 120;
3801
Andrew Mortona737b3e2006-03-22 00:08:11 -08003802 /*
3803 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804 * allocation behaviour: Most allocs on one cpu, most free operations
3805 * on another cpu. For these cases, an efficient object passing between
3806 * cpus is necessary. This is provided by a shared array. The array
3807 * replaces Bonwick's magazine layer.
3808 * On uniprocessor, it's functionally equivalent (but less efficient)
3809 * to a larger limit. Thus disabled by default.
3810 */
3811 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003812 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814
3815#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003816 /*
3817 * With debugging enabled, large batchcount lead to excessively long
3818 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 */
3820 if (limit > 32)
3821 limit = 32;
3822#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003823 batchcount = (limit + 1) / 2;
3824skip_setup:
3825 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 if (err)
3827 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003828 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003829 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830}
3831
Christoph Lameter1b552532006-03-22 00:09:07 -08003832/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003833 * Drain an array if it contains any elements taking the node lock only if
3834 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003835 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003836 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003837static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Christoph Lameter1b552532006-03-22 00:09:07 -08003838 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839{
Joonsoo Kim97654df2014-08-06 16:04:25 -07003840 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003841 int tofree;
3842
Christoph Lameter1b552532006-03-22 00:09:07 -08003843 if (!ac || !ac->avail)
3844 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 if (ac->touched && !force) {
3846 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003847 } else {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003848 spin_lock_irq(&n->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003849 if (ac->avail) {
3850 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3851 if (tofree > ac->avail)
3852 tofree = (ac->avail + 1) / 2;
Joonsoo Kim97654df2014-08-06 16:04:25 -07003853 free_block(cachep, ac->entry, tofree, node, &list);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003854 ac->avail -= tofree;
3855 memmove(ac->entry, &(ac->entry[tofree]),
3856 sizeof(void *) * ac->avail);
3857 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003858 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003859 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 }
3861}
3862
3863/**
3864 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003865 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003866 *
3867 * Called from workqueue/eventd every few seconds.
3868 * Purpose:
3869 * - clear the per-cpu caches for this CPU.
3870 * - return freeable pages to the main free memory pool.
3871 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003872 * If we cannot acquire the cache chain mutex then just give up - we'll try
3873 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003875static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003877 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003878 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003879 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07003880 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881
Christoph Lameter18004c52012-07-06 15:25:12 -05003882 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003884 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885
Christoph Lameter18004c52012-07-06 15:25:12 -05003886 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 check_irq_on();
3888
Christoph Lameter35386e32006-03-22 00:09:05 -08003889 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003890 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08003891 * have established with reasonable certainty that
3892 * we can do some work if the lock was obtained.
3893 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07003894 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08003895
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003896 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003898 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899
Christoph Lameter35386e32006-03-22 00:09:05 -08003900 /*
3901 * These are racy checks but it does not matter
3902 * if we skip one check or scan twice.
3903 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003904 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003905 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003907 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003909 drain_array(searchp, n, n->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003911 if (n->free_touched)
3912 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003913 else {
3914 int freed;
3915
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003916 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07003917 5 * searchp->num - 1) / (5 * searchp->num));
3918 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003920next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921 cond_resched();
3922 }
3923 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05003924 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003925 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003926out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08003927 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003928 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929}
3930
Linus Torvalds158a9622008-01-02 13:04:48 -08003931#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04003932void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003934 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003935 unsigned long active_objs;
3936 unsigned long num_objs;
3937 unsigned long active_slabs = 0;
3938 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003939 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003941 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003942 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 active_objs = 0;
3945 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07003946 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07003947
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003948 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003949 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003950
Joonsoo Kim8456a642013-10-24 10:07:49 +09003951 list_for_each_entry(page, &n->slabs_full, lru) {
3952 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07003953 error = "slabs_full accounting error";
3954 active_objs += cachep->num;
3955 active_slabs++;
3956 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003957 list_for_each_entry(page, &n->slabs_partial, lru) {
3958 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003959 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003960 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003961 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09003962 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07003963 active_slabs++;
3964 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09003965 list_for_each_entry(page, &n->slabs_free, lru) {
3966 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09003967 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07003968 num_slabs++;
3969 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003970 free_objects += n->free_objects;
3971 if (n->shared)
3972 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003973
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003974 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003976 num_slabs += active_slabs;
3977 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003978 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 error = "free_objects accounting error";
3980
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003981 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 if (error)
3983 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3984
Glauber Costa0d7561c2012-10-19 18:20:27 +04003985 sinfo->active_objs = active_objs;
3986 sinfo->num_objs = num_objs;
3987 sinfo->active_slabs = active_slabs;
3988 sinfo->num_slabs = num_slabs;
3989 sinfo->shared_avail = shared_avail;
3990 sinfo->limit = cachep->limit;
3991 sinfo->batchcount = cachep->batchcount;
3992 sinfo->shared = cachep->shared;
3993 sinfo->objects_per_slab = cachep->num;
3994 sinfo->cache_order = cachep->gfporder;
3995}
3996
3997void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3998{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004000 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004001 unsigned long high = cachep->high_mark;
4002 unsigned long allocs = cachep->num_allocations;
4003 unsigned long grown = cachep->grown;
4004 unsigned long reaped = cachep->reaped;
4005 unsigned long errors = cachep->errors;
4006 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004008 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004009 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010
Joe Perchese92dd4f2010-03-26 19:27:58 -07004011 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4012 "%4lu %4lu %4lu %4lu %4lu",
4013 allocs, high, grown,
4014 reaped, errors, max_freeable, node_allocs,
4015 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 }
4017 /* cpu stats */
4018 {
4019 unsigned long allochit = atomic_read(&cachep->allochit);
4020 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4021 unsigned long freehit = atomic_read(&cachep->freehit);
4022 unsigned long freemiss = atomic_read(&cachep->freemiss);
4023
4024 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004025 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 }
4027#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028}
4029
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030#define MAX_SLABINFO_WRITE 128
4031/**
4032 * slabinfo_write - Tuning for the slab allocator
4033 * @file: unused
4034 * @buffer: user buffer
4035 * @count: data length
4036 * @ppos: unused
4037 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004038ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004039 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004041 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004043 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004044
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045 if (count > MAX_SLABINFO_WRITE)
4046 return -EINVAL;
4047 if (copy_from_user(&kbuf, buffer, count))
4048 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004049 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050
4051 tmp = strchr(kbuf, ' ');
4052 if (!tmp)
4053 return -EINVAL;
4054 *tmp = '\0';
4055 tmp++;
4056 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4057 return -EINVAL;
4058
4059 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004060 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004062 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004064 if (limit < 1 || batchcount < 1 ||
4065 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004066 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004068 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004069 batchcount, shared,
4070 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071 }
4072 break;
4073 }
4074 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004075 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 if (res >= 0)
4077 res = count;
4078 return res;
4079}
Al Viro871751e2006-03-25 03:06:39 -08004080
4081#ifdef CONFIG_DEBUG_SLAB_LEAK
4082
Al Viro871751e2006-03-25 03:06:39 -08004083static inline int add_caller(unsigned long *n, unsigned long v)
4084{
4085 unsigned long *p;
4086 int l;
4087 if (!v)
4088 return 1;
4089 l = n[1];
4090 p = n + 2;
4091 while (l) {
4092 int i = l/2;
4093 unsigned long *q = p + 2 * i;
4094 if (*q == v) {
4095 q[1]++;
4096 return 1;
4097 }
4098 if (*q > v) {
4099 l = i;
4100 } else {
4101 p = q + 2;
4102 l -= i + 1;
4103 }
4104 }
4105 if (++n[1] == n[0])
4106 return 0;
4107 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4108 p[0] = v;
4109 p[1] = 1;
4110 return 1;
4111}
4112
Joonsoo Kim8456a642013-10-24 10:07:49 +09004113static void handle_slab(unsigned long *n, struct kmem_cache *c,
4114 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004115{
4116 void *p;
Joonsoo Kim03787302014-06-23 13:22:06 -07004117 int i;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004118
Al Viro871751e2006-03-25 03:06:39 -08004119 if (n[0] == n[1])
4120 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004121 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kim03787302014-06-23 13:22:06 -07004122 if (get_obj_status(page, i) != OBJECT_ACTIVE)
Al Viro871751e2006-03-25 03:06:39 -08004123 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004124
Al Viro871751e2006-03-25 03:06:39 -08004125 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4126 return;
4127 }
4128}
4129
4130static void show_symbol(struct seq_file *m, unsigned long address)
4131{
4132#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004133 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004134 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004135
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004136 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004137 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004138 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004139 seq_printf(m, " [%s]", modname);
4140 return;
4141 }
4142#endif
4143 seq_printf(m, "%p", (void *)address);
4144}
4145
4146static int leaks_show(struct seq_file *m, void *p)
4147{
Thierry Reding0672aa72012-06-22 19:42:49 +02004148 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004149 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004150 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004151 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004152 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004153 int node;
4154 int i;
4155
4156 if (!(cachep->flags & SLAB_STORE_USER))
4157 return 0;
4158 if (!(cachep->flags & SLAB_RED_ZONE))
4159 return 0;
4160
4161 /* OK, we can do it */
4162
Christoph Lameterdb845062013-02-05 18:45:23 +00004163 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004164
Christoph Lameter18bf8542014-08-06 16:04:11 -07004165 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004166
4167 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004168 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004169
Joonsoo Kim8456a642013-10-24 10:07:49 +09004170 list_for_each_entry(page, &n->slabs_full, lru)
4171 handle_slab(x, cachep, page);
4172 list_for_each_entry(page, &n->slabs_partial, lru)
4173 handle_slab(x, cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004174 spin_unlock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004175 }
4176 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004177 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004178 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004179 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004180 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004181 if (!m->private) {
4182 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004183 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004184 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004185 return -ENOMEM;
4186 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004187 *(unsigned long *)m->private = x[0] * 2;
4188 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004189 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004190 /* Now make sure this entry will be retried */
4191 m->count = m->size;
4192 return 0;
4193 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004194 for (i = 0; i < x[1]; i++) {
4195 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4196 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004197 seq_putc(m, '\n');
4198 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004199
Al Viro871751e2006-03-25 03:06:39 -08004200 return 0;
4201}
4202
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004203static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004204 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004205 .next = slab_next,
4206 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004207 .show = leaks_show,
4208};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004209
4210static int slabstats_open(struct inode *inode, struct file *file)
4211{
Rob Jonesb208ce32014-10-09 15:28:03 -07004212 unsigned long *n;
4213
4214 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4215 if (!n)
4216 return -ENOMEM;
4217
4218 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4219
4220 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004221}
4222
4223static const struct file_operations proc_slabstats_operations = {
4224 .open = slabstats_open,
4225 .read = seq_read,
4226 .llseek = seq_lseek,
4227 .release = seq_release_private,
4228};
Al Viro871751e2006-03-25 03:06:39 -08004229#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004230
4231static int __init slab_proc_init(void)
4232{
4233#ifdef CONFIG_DEBUG_SLAB_LEAK
4234 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4235#endif
4236 return 0;
4237}
4238module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239#endif
4240
Manfred Spraul00e145b2005-09-03 15:55:07 -07004241/**
4242 * ksize - get the actual amount of memory allocated for a given object
4243 * @objp: Pointer to the object
4244 *
4245 * kmalloc may internally round up allocations and return more memory
4246 * than requested. ksize() can be used to determine the actual amount of
4247 * memory allocated. The caller may use this additional memory, even though
4248 * a smaller amount of memory was initially specified with the kmalloc call.
4249 * The caller must guarantee that objp points to a valid object previously
4250 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4251 * must not be freed during the duration of the call.
4252 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004253size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004255 BUG_ON(!objp);
4256 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004257 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258
Christoph Lameter8c138bc2012-06-13 10:24:58 -05004259 return virt_to_cache(objp)->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004261EXPORT_SYMBOL(ksize);