blob: 090fb26b3a39b4feba105650f2a663f5c9972064 [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/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 * struct array_cache
173 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 * Purpose:
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
178 *
179 * The limit is stored in the per-cpu structure to reduce the data cache
180 * footprint.
181 *
182 */
183struct array_cache {
184 unsigned int avail;
185 unsigned int limit;
186 unsigned int batchcount;
187 unsigned int touched;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700188 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
191 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800192 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193};
194
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700195struct alien_cache {
196 spinlock_t lock;
197 struct array_cache ac;
198};
199
Andrew Mortona737b3e2006-03-22 00:08:11 -0800200/*
Christoph Lametere498be72005-09-09 13:03:32 -0700201 * Need this for bootstrapping a per node allocator.
202 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700203#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000204static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700205#define CACHE_CACHE 0
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700206#define SIZE_NODE (MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207
Christoph Lametered11d9e2006-06-30 01:55:45 -0700208static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000209 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700210static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700211 int node, struct list_head *list);
212static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300213static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000214static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700215
Joonsoo Kim76b342b2016-05-19 17:10:26 -0700216static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
217 void **list);
218static inline void fixup_slab_list(struct kmem_cache *cachep,
219 struct kmem_cache_node *n, struct page *page,
220 void **list);
Ingo Molnare0a42722006-06-23 02:03:46 -0700221static int slab_early_init = 1;
222
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000223#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700224
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000225static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700226{
227 INIT_LIST_HEAD(&parent->slabs_full);
228 INIT_LIST_HEAD(&parent->slabs_partial);
229 INIT_LIST_HEAD(&parent->slabs_free);
230 parent->shared = NULL;
231 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800232 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700233 spin_lock_init(&parent->list_lock);
234 parent->free_objects = 0;
235 parent->free_touched = 0;
236}
237
Andrew Mortona737b3e2006-03-22 00:08:11 -0800238#define MAKE_LIST(cachep, listp, slab, nodeid) \
239 do { \
240 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700241 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700242 } while (0)
243
Andrew Mortona737b3e2006-03-22 00:08:11 -0800244#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
245 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700246 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
247 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
249 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700251#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252#define CFLGS_OFF_SLAB (0x80000000UL)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700253#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
255
256#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800257/*
258 * Optimization question: fewer reaps means less probability for unnessary
259 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100261 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 * which could lock up otherwise freeable slabs.
263 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800264#define REAPTIMEOUT_AC (2*HZ)
265#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
267#if STATS
268#define STATS_INC_ACTIVE(x) ((x)->num_active++)
269#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
270#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
271#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700272#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273#define STATS_SET_HIGH(x) \
274 do { \
275 if ((x)->num_active > (x)->high_mark) \
276 (x)->high_mark = (x)->num_active; \
277 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278#define STATS_INC_ERR(x) ((x)->errors++)
279#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700280#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700281#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800282#define STATS_SET_FREEABLE(x, i) \
283 do { \
284 if ((x)->max_freeable < i) \
285 (x)->max_freeable = i; \
286 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
288#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
289#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
290#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
291#else
292#define STATS_INC_ACTIVE(x) do { } while (0)
293#define STATS_DEC_ACTIVE(x) do { } while (0)
294#define STATS_INC_ALLOCED(x) do { } while (0)
295#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700296#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297#define STATS_SET_HIGH(x) do { } while (0)
298#define STATS_INC_ERR(x) do { } while (0)
299#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700300#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700301#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800302#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303#define STATS_INC_ALLOCHIT(x) do { } while (0)
304#define STATS_INC_ALLOCMISS(x) do { } while (0)
305#define STATS_INC_FREEHIT(x) do { } while (0)
306#define STATS_INC_FREEMISS(x) do { } while (0)
307#endif
308
309#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Andrew Mortona737b3e2006-03-22 00:08:11 -0800311/*
312 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800314 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 * the end of an object is aligned with the end of the real
316 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800317 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800319 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500320 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
321 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800324static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800326 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327}
328
David Woodhouseb46b8f12007-05-08 00:22:59 -0700329static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330{
331 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700332 return (unsigned long long*) (objp + obj_offset(cachep) -
333 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334}
335
David Woodhouseb46b8f12007-05-08 00:22:59 -0700336static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337{
338 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
339 if (cachep->flags & SLAB_STORE_USER)
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500340 return (unsigned long long *)(objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700341 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400342 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500343 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700344 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345}
346
Pekka Enberg343e0d72006-02-01 03:05:50 -0800347static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348{
349 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500350 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351}
352
353#else
354
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800355#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700356#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
357#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
359
360#endif
361
Joonsoo Kim03787302014-06-23 13:22:06 -0700362#ifdef CONFIG_DEBUG_SLAB_LEAK
363
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700364static inline bool is_store_user_clean(struct kmem_cache *cachep)
Joonsoo Kim03787302014-06-23 13:22:06 -0700365{
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700366 return atomic_read(&cachep->store_user_clean) == 1;
367}
Joonsoo Kim03787302014-06-23 13:22:06 -0700368
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700369static inline void set_store_user_clean(struct kmem_cache *cachep)
370{
371 atomic_set(&cachep->store_user_clean, 1);
372}
Joonsoo Kim03787302014-06-23 13:22:06 -0700373
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700374static inline void set_store_user_dirty(struct kmem_cache *cachep)
375{
376 if (is_store_user_clean(cachep))
377 atomic_set(&cachep->store_user_clean, 0);
Joonsoo Kim03787302014-06-23 13:22:06 -0700378}
379
380#else
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700381static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
Joonsoo Kim03787302014-06-23 13:22:06 -0700382
383#endif
384
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700386 * Do not go above this order unless 0 objects fit into the slab or
387 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 */
David Rientjes543585c2011-10-18 22:09:24 -0700389#define SLAB_MAX_ORDER_HI 1
390#define SLAB_MAX_ORDER_LO 0
391static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700392static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800394static inline struct kmem_cache *virt_to_cache(const void *obj)
395{
Christoph Lameterb49af682007-05-06 14:49:41 -0700396 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500397 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800398}
399
Joonsoo Kim8456a642013-10-24 10:07:49 +0900400static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800401 unsigned int idx)
402{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900403 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800404}
405
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800406/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500407 * We want to avoid an expensive divide : (offset / cache->size)
408 * Using the fact that size is a constant for a particular cache,
409 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800410 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
411 */
412static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900413 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800414{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900415 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800416 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800417}
418
Joonsoo Kim6fb92432016-03-15 14:54:09 -0700419#define BOOT_CPUCACHE_ENTRIES 1
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000421static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800422 .batchcount = 1,
423 .limit = BOOT_CPUCACHE_ENTRIES,
424 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500425 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800426 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427};
428
Tejun Heo1871e522009-10-29 22:34:13 +0900429static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
Pekka Enberg343e0d72006-02-01 03:05:50 -0800431static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700433 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434}
435
Andrew Mortona737b3e2006-03-22 00:08:11 -0800436/*
437 * Calculate the number of objects and left-over bytes for a given buffer size.
438 */
Joonsoo Kim70f75062016-03-15 14:54:53 -0700439static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
440 unsigned long flags, size_t *left_over)
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800441{
Joonsoo Kim70f75062016-03-15 14:54:53 -0700442 unsigned int num;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800443 size_t slab_size = PAGE_SIZE << gfporder;
444
445 /*
446 * The slab management structure can be either off the slab or
447 * on it. For the latter case, the memory allocated for a
448 * slab is used for:
449 *
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800450 * - @buffer_size bytes for each object
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700451 * - One freelist_idx_t for each object
452 *
453 * We don't need to consider alignment of freelist because
454 * freelist will be at the end of slab page. The objects will be
455 * at the correct alignment.
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800456 *
457 * If the slab management structure is off the slab, then the
458 * alignment will already be calculated into the size. Because
459 * the slabs are all pages aligned, the objects will be at the
460 * correct alignment when allocated.
461 */
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700462 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700463 num = slab_size / buffer_size;
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700464 *left_over = slab_size % buffer_size;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800465 } else {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700466 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700467 *left_over = slab_size %
468 (buffer_size + sizeof(freelist_idx_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 }
Joonsoo Kim70f75062016-03-15 14:54:53 -0700470
471 return num;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472}
473
Christoph Lameterf28510d2012-09-11 19:49:38 +0000474#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700475#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476
Andrew Mortona737b3e2006-03-22 00:08:11 -0800477static void __slab_error(const char *function, struct kmem_cache *cachep,
478 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
Joe Perches11705322016-03-17 14:19:50 -0700480 pr_err("slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800481 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030483 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000485#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
Paul Menage3395ee02006-12-06 20:32:16 -0800487/*
488 * By default on NUMA we use alien caches to stage the freeing of
489 * objects allocated from other nodes. This causes massive memory
490 * inefficiencies when using fake NUMA setup to split memory into a
491 * large number of small nodes, so it can be disabled on the command
492 * line
493 */
494
495static int use_alien_caches __read_mostly = 1;
496static int __init noaliencache_setup(char *s)
497{
498 use_alien_caches = 0;
499 return 1;
500}
501__setup("noaliencache", noaliencache_setup);
502
David Rientjes3df1ccc2011-10-18 22:09:28 -0700503static int __init slab_max_order_setup(char *str)
504{
505 get_option(&str, &slab_max_order);
506 slab_max_order = slab_max_order < 0 ? 0 :
507 min(slab_max_order, MAX_ORDER - 1);
508 slab_max_order_set = true;
509
510 return 1;
511}
512__setup("slab_max_order=", slab_max_order_setup);
513
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800514#ifdef CONFIG_NUMA
515/*
516 * Special reaping functions for NUMA systems called from cache_reap().
517 * These take care of doing round robin flushing of alien caches (containing
518 * objects freed on different nodes from which they were allocated) and the
519 * flushing of remote pcps by calling drain_node_pages.
520 */
Tejun Heo1871e522009-10-29 22:34:13 +0900521static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800522
523static void init_reap_node(int cpu)
524{
Andrew Morton0edaf862016-05-19 17:10:58 -0700525 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
526 node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800527}
528
529static void next_reap_node(void)
530{
Christoph Lameter909ea962010-12-08 16:22:55 +0100531 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800532
Andrew Morton0edaf862016-05-19 17:10:58 -0700533 node = next_node_in(node, node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100534 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800535}
536
537#else
538#define init_reap_node(cpu) do { } while (0)
539#define next_reap_node(void) do { } while (0)
540#endif
541
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542/*
543 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
544 * via the workqueue/eventd.
545 * Add the CPU number into the expiration time to minimize the possibility of
546 * the CPUs getting into lockstep and contending for the global cache chain
547 * lock.
548 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400549static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550{
Tejun Heo1871e522009-10-29 22:34:13 +0900551 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552
553 /*
554 * When this gets called from do_initcalls via cpucache_init(),
555 * init_workqueues() has already run, so keventd will be setup
556 * at that time.
557 */
David Howells52bad642006-11-22 14:54:01 +0000558 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800559 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700560 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800561 schedule_delayed_work_on(cpu, reap_work,
562 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 }
564}
565
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700566static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100568 /*
569 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300570 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100571 * cache the pointers are not cleared and they could be counted as
572 * valid references during a kmemleak scan. Therefore, kmemleak must
573 * not scan such objects.
574 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700575 kmemleak_no_scan(ac);
576 if (ac) {
577 ac->avail = 0;
578 ac->limit = limit;
579 ac->batchcount = batch;
580 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700582}
583
584static struct array_cache *alloc_arraycache(int node, int entries,
585 int batchcount, gfp_t gfp)
586{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700587 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700588 struct array_cache *ac = NULL;
589
590 ac = kmalloc_node(memsize, gfp, node);
591 init_arraycache(ac, entries, batchcount);
592 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593}
594
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700595static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
596 struct page *page, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700597{
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700598 struct kmem_cache_node *n;
599 int page_node;
600 LIST_HEAD(list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700601
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700602 page_node = page_to_nid(page);
603 n = get_node(cachep, page_node);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700604
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700605 spin_lock(&n->list_lock);
606 free_block(cachep, &objp, 1, page_node, &list);
607 spin_unlock(&n->list_lock);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700608
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700609 slabs_destroy(cachep, &list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700610}
611
Christoph Lameter3ded1752006-03-25 03:06:44 -0800612/*
613 * Transfer objects in one arraycache to another.
614 * Locking must be handled by the caller.
615 *
616 * Return the number of entries transferred.
617 */
618static int transfer_objects(struct array_cache *to,
619 struct array_cache *from, unsigned int max)
620{
621 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700622 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800623
624 if (!nr)
625 return 0;
626
627 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
628 sizeof(void *) *nr);
629
630 from->avail -= nr;
631 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800632 return nr;
633}
634
Christoph Lameter765c4502006-09-27 01:50:08 -0700635#ifndef CONFIG_NUMA
636
637#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000638#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700639
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700640static inline struct alien_cache **alloc_alien_cache(int node,
641 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700642{
Joonsoo Kim88881772016-05-19 17:10:05 -0700643 return NULL;
Christoph Lameter765c4502006-09-27 01:50:08 -0700644}
645
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700646static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700647{
648}
649
650static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
651{
652 return 0;
653}
654
655static inline void *alternate_node_alloc(struct kmem_cache *cachep,
656 gfp_t flags)
657{
658 return NULL;
659}
660
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800661static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700662 gfp_t flags, int nodeid)
663{
664 return NULL;
665}
666
David Rientjes4167e9b2015-04-14 15:46:55 -0700667static inline gfp_t gfp_exact_node(gfp_t flags)
668{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700669 return flags & ~__GFP_NOFAIL;
David Rientjes4167e9b2015-04-14 15:46:55 -0700670}
671
Christoph Lameter765c4502006-09-27 01:50:08 -0700672#else /* CONFIG_NUMA */
673
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800674static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800675static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800676
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700677static struct alien_cache *__alloc_alien_cache(int node, int entries,
678 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700679{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700680 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700681 struct alien_cache *alc = NULL;
682
683 alc = kmalloc_node(memsize, gfp, node);
684 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700685 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700686 return alc;
687}
688
689static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
690{
691 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700692 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700693 int i;
694
695 if (limit > 1)
696 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700697 alc_ptr = kzalloc_node(memsize, gfp, node);
698 if (!alc_ptr)
699 return NULL;
700
701 for_each_node(i) {
702 if (i == node || !node_online(i))
703 continue;
704 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
705 if (!alc_ptr[i]) {
706 for (i--; i >= 0; i--)
707 kfree(alc_ptr[i]);
708 kfree(alc_ptr);
709 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700710 }
711 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700712 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700713}
714
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700715static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700716{
717 int i;
718
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700719 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700720 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700721 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700722 kfree(alc_ptr[i]);
723 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700724}
725
Pekka Enberg343e0d72006-02-01 03:05:50 -0800726static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700727 struct array_cache *ac, int node,
728 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700729{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700730 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700731
732 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000733 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800734 /*
735 * Stuff objects into the remote nodes shared array first.
736 * That way we could avoid the overhead of putting the objects
737 * into the free lists and getting them back later.
738 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000739 if (n->shared)
740 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800741
Joonsoo Kim833b7062014-08-06 16:04:33 -0700742 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700743 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000744 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700745 }
746}
747
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800748/*
749 * Called from cache_reap() to regularly drain alien caches round robin.
750 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000751static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800752{
Christoph Lameter909ea962010-12-08 16:22:55 +0100753 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800754
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000755 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700756 struct alien_cache *alc = n->alien[node];
757 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800758
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700759 if (alc) {
760 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700761 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700762 LIST_HEAD(list);
763
764 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700765 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700766 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700767 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800768 }
769 }
770}
771
Andrew Mortona737b3e2006-03-22 00:08:11 -0800772static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700773 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700774{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800775 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700776 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700777 struct array_cache *ac;
778 unsigned long flags;
779
780 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700781 alc = alien[i];
782 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700783 LIST_HEAD(list);
784
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700785 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700786 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700787 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700788 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700789 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700790 }
791 }
792}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700793
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700794static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
795 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700796{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000797 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700798 struct alien_cache *alien = NULL;
799 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700800 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700801
Christoph Lameter18bf8542014-08-06 16:04:11 -0700802 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700803 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700804 if (n->alien && n->alien[page_node]) {
805 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700806 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700807 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700808 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700809 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700810 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700811 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700812 ac->entry[ac->avail++] = objp;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700813 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700814 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700815 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700816 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700817 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700818 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700819 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -0700820 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700821 }
822 return 1;
823}
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700824
825static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
826{
827 int page_node = page_to_nid(virt_to_page(objp));
828 int node = numa_mem_id();
829 /*
830 * Make sure we are not freeing a object from another node to the array
831 * cache on this cpu.
832 */
833 if (likely(node == page_node))
834 return 0;
835
836 return __cache_free_alien(cachep, objp, node, page_node);
837}
David Rientjes4167e9b2015-04-14 15:46:55 -0700838
839/*
Mel Gorman444eb2a42016-03-17 14:19:23 -0700840 * Construct gfp mask to allocate from a specific node but do not reclaim or
841 * warn about failures.
David Rientjes4167e9b2015-04-14 15:46:55 -0700842 */
843static inline gfp_t gfp_exact_node(gfp_t flags)
844{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700845 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
David Rientjes4167e9b2015-04-14 15:46:55 -0700846}
Christoph Lametere498be72005-09-09 13:03:32 -0700847#endif
848
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700849static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
850{
851 struct kmem_cache_node *n;
852
853 /*
854 * Set up the kmem_cache_node for cpu before we can
855 * begin anything. Make sure some other cpu on this
856 * node has not already allocated this
857 */
858 n = get_node(cachep, node);
859 if (n) {
860 spin_lock_irq(&n->list_lock);
861 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
862 cachep->num;
863 spin_unlock_irq(&n->list_lock);
864
865 return 0;
866 }
867
868 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
869 if (!n)
870 return -ENOMEM;
871
872 kmem_cache_node_init(n);
873 n->next_reap = jiffies + REAPTIMEOUT_NODE +
874 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
875
876 n->free_limit =
877 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
878
879 /*
880 * The kmem_cache_nodes don't come and go as CPUs
881 * come and go. slab_mutex is sufficient
882 * protection here.
883 */
884 cachep->node[node] = n;
885
886 return 0;
887}
888
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200889#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700890/*
Christoph Lameter6a673682013-01-10 19:14:19 +0000891 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000892 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -0700893 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +0000894 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -0700895 * already in use.
896 *
Christoph Lameter18004c52012-07-06 15:25:12 -0500897 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -0700898 */
Christoph Lameter6a673682013-01-10 19:14:19 +0000899static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700900{
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700901 int ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700902 struct kmem_cache *cachep;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700903
Christoph Lameter18004c52012-07-06 15:25:12 -0500904 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700905 ret = init_cache_node(cachep, node, GFP_KERNEL);
906 if (ret)
907 return ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700908 }
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700909
David Rientjes8f9f8d92010-03-27 19:40:47 -0700910 return 0;
911}
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200912#endif
David Rientjes8f9f8d92010-03-27 19:40:47 -0700913
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700914static int setup_kmem_cache_node(struct kmem_cache *cachep,
915 int node, gfp_t gfp, bool force_change)
916{
917 int ret = -ENOMEM;
918 struct kmem_cache_node *n;
919 struct array_cache *old_shared = NULL;
920 struct array_cache *new_shared = NULL;
921 struct alien_cache **new_alien = NULL;
922 LIST_HEAD(list);
923
924 if (use_alien_caches) {
925 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
926 if (!new_alien)
927 goto fail;
928 }
929
930 if (cachep->shared) {
931 new_shared = alloc_arraycache(node,
932 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
933 if (!new_shared)
934 goto fail;
935 }
936
937 ret = init_cache_node(cachep, node, gfp);
938 if (ret)
939 goto fail;
940
941 n = get_node(cachep, node);
942 spin_lock_irq(&n->list_lock);
943 if (n->shared && force_change) {
944 free_block(cachep, n->shared->entry,
945 n->shared->avail, node, &list);
946 n->shared->avail = 0;
947 }
948
949 if (!n->shared || force_change) {
950 old_shared = n->shared;
951 n->shared = new_shared;
952 new_shared = NULL;
953 }
954
955 if (!n->alien) {
956 n->alien = new_alien;
957 new_alien = NULL;
958 }
959
960 spin_unlock_irq(&n->list_lock);
961 slabs_destroy(cachep, &list);
962
Joonsoo Kim801faf02016-05-19 17:10:31 -0700963 /*
964 * To protect lockless access to n->shared during irq disabled context.
965 * If n->shared isn't NULL in irq disabled context, accessing to it is
966 * guaranteed to be valid until irq is re-enabled, because it will be
967 * freed after synchronize_sched().
968 */
969 if (force_change)
970 synchronize_sched();
971
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700972fail:
973 kfree(old_shared);
974 kfree(new_shared);
975 free_alien_cache(new_alien);
976
977 return ret;
978}
979
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200980#ifdef CONFIG_SMP
981
Paul Gortmaker0db06282013-06-19 14:53:51 -0400982static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983{
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700984 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000985 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700986 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +1030987 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700988
Christoph Lameter18004c52012-07-06 15:25:12 -0500989 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700990 struct array_cache *nc;
991 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700992 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700993 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700994
Christoph Lameter18bf8542014-08-06 16:04:11 -0700995 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000996 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700997 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700998
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000999 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001000
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001001 /* Free limit for this kmem_cache_node */
1002 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001003
1004 /* cpu is dead; no one can alloc from it. */
1005 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1006 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001007 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001008 nc->avail = 0;
1009 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001010
Rusty Russell58463c12009-12-17 11:43:12 -06001011 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001012 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001013 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001014 }
1015
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001016 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001017 if (shared) {
1018 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001019 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001020 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001021 }
1022
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001023 alien = n->alien;
1024 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001025
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001026 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001027
1028 kfree(shared);
1029 if (alien) {
1030 drain_alien_cache(cachep, alien);
1031 free_alien_cache(alien);
1032 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001033
1034free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001035 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001036 }
1037 /*
1038 * In the previous loop, all the objects were freed to
1039 * the respective cache's slabs, now we can go ahead and
1040 * shrink each nodelist to its limit.
1041 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001042 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001043 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001044 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001045 continue;
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001046 drain_freelist(cachep, n, INT_MAX);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001047 }
1048}
1049
Paul Gortmaker0db06282013-06-19 14:53:51 -04001050static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001051{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001052 struct kmem_cache *cachep;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001053 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001054 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001056 /*
1057 * We need to do this right in the beginning since
1058 * alloc_arraycache's are going to use this list.
1059 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001060 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001061 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001062 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001063 if (err < 0)
1064 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001065
1066 /*
1067 * Now we can go ahead with allocating the shared arrays and
1068 * array caches
1069 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001070 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07001071 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1072 if (err)
1073 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001074 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001075
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001076 return 0;
1077bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001078 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001079 return -ENOMEM;
1080}
1081
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001082int slab_prepare_cpu(unsigned int cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001083{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001084 int err;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001085
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001086 mutex_lock(&slab_mutex);
1087 err = cpuup_prepare(cpu);
1088 mutex_unlock(&slab_mutex);
1089 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090}
1091
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001092/*
1093 * This is called for a failed online attempt and for a successful
1094 * offline.
1095 *
1096 * Even if all the cpus of a node are down, we don't free the
1097 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1098 * a kmalloc allocation from another cpu for memory from the node of
1099 * the cpu going down. The list3 structure is usually allocated from
1100 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1101 */
1102int slab_dead_cpu(unsigned int cpu)
1103{
1104 mutex_lock(&slab_mutex);
1105 cpuup_canceled(cpu);
1106 mutex_unlock(&slab_mutex);
1107 return 0;
1108}
1109#endif
1110
1111static int slab_online_cpu(unsigned int cpu)
1112{
1113 start_cpu_timer(cpu);
1114 return 0;
1115}
1116
1117static int slab_offline_cpu(unsigned int cpu)
1118{
1119 /*
1120 * Shutdown cache reaper. Note that the slab_mutex is held so
1121 * that if cache_reap() is invoked it cannot do anything
1122 * expensive but will only modify reap_work and reschedule the
1123 * timer.
1124 */
1125 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1126 /* Now the cache_reaper is guaranteed to be not running. */
1127 per_cpu(slab_reap_work, cpu).work.func = NULL;
1128 return 0;
1129}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
David Rientjes8f9f8d92010-03-27 19:40:47 -07001131#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1132/*
1133 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1134 * Returns -EBUSY if all objects cannot be drained so that the node is not
1135 * removed.
1136 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001137 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001138 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001139static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001140{
1141 struct kmem_cache *cachep;
1142 int ret = 0;
1143
Christoph Lameter18004c52012-07-06 15:25:12 -05001144 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001145 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001146
Christoph Lameter18bf8542014-08-06 16:04:11 -07001147 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001148 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001149 continue;
1150
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001151 drain_freelist(cachep, n, INT_MAX);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001152
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001153 if (!list_empty(&n->slabs_full) ||
1154 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001155 ret = -EBUSY;
1156 break;
1157 }
1158 }
1159 return ret;
1160}
1161
1162static int __meminit slab_memory_callback(struct notifier_block *self,
1163 unsigned long action, void *arg)
1164{
1165 struct memory_notify *mnb = arg;
1166 int ret = 0;
1167 int nid;
1168
1169 nid = mnb->status_change_nid;
1170 if (nid < 0)
1171 goto out;
1172
1173 switch (action) {
1174 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001175 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001176 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001177 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001178 break;
1179 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001180 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001181 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001182 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001183 break;
1184 case MEM_ONLINE:
1185 case MEM_OFFLINE:
1186 case MEM_CANCEL_ONLINE:
1187 case MEM_CANCEL_OFFLINE:
1188 break;
1189 }
1190out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001191 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001192}
1193#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1194
Christoph Lametere498be72005-09-09 13:03:32 -07001195/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001196 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001197 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001198static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001199 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001200{
Christoph Lameter6744f082013-01-10 19:12:17 +00001201 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001202
Christoph Lameter6744f082013-01-10 19:12:17 +00001203 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001204 BUG_ON(!ptr);
1205
Christoph Lameter6744f082013-01-10 19:12:17 +00001206 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001207 /*
1208 * Do not assume that spinlocks can be initialized via memcpy:
1209 */
1210 spin_lock_init(&ptr->list_lock);
1211
Christoph Lametere498be72005-09-09 13:03:32 -07001212 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001213 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001214}
1215
Andrew Mortona737b3e2006-03-22 00:08:11 -08001216/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001217 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1218 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001219 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001220static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001221{
1222 int node;
1223
1224 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001225 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001226 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001227 REAPTIMEOUT_NODE +
1228 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001229 }
1230}
1231
1232/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001233 * Initialisation. Called after the page allocator have been initialised and
1234 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 */
1236void __init kmem_cache_init(void)
1237{
Christoph Lametere498be72005-09-09 13:03:32 -07001238 int i;
1239
Joonsoo Kim68126702013-10-24 10:07:42 +09001240 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1241 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001242 kmem_cache = &kmem_cache_boot;
1243
Joonsoo Kim88881772016-05-19 17:10:05 -07001244 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001245 use_alien_caches = 0;
1246
Christoph Lameter3c583462012-11-28 16:23:01 +00001247 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001248 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001249
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 /*
1251 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001252 * page orders on machines with more than 32MB of memory if
1253 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001255 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001256 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 /* Bootstrap is tricky, because several objects are allocated
1259 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001260 * 1) initialize the kmem_cache cache: it contains the struct
1261 * kmem_cache structures of all caches, except kmem_cache itself:
1262 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001263 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001264 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001265 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001267 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001268 * An __init data area is used for the head array.
1269 * 3) Create the remaining kmalloc caches, with minimally sized
1270 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001271 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001273 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001274 * the other cache's with kmalloc allocated memory.
1275 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 */
1277
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001278 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
Eric Dumazet8da34302007-05-06 14:49:29 -07001280 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001281 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001282 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001283 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001284 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001285 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001286 SLAB_HWCACHE_ALIGN);
1287 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001288 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289
Andrew Mortona737b3e2006-03-22 00:08:11 -08001290 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001291 * Initialize the caches that provide memory for the kmem_cache_node
1292 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001293 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001294 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001295 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001296 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001297 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001298
Ingo Molnare0a42722006-06-23 02:03:46 -07001299 slab_early_init = 0;
1300
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001301 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001302 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001303 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304
Mel Gorman9c09a952008-01-24 05:49:54 -08001305 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001306 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001307
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001308 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001309 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001310 }
1311 }
1312
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001313 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001314}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001315
Pekka Enberg8429db52009-06-12 15:58:59 +03001316void __init kmem_cache_init_late(void)
1317{
1318 struct kmem_cache *cachep;
1319
Christoph Lameter97d06602012-07-06 15:25:11 -05001320 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001321
Pekka Enberg8429db52009-06-12 15:58:59 +03001322 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001323 mutex_lock(&slab_mutex);
1324 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001325 if (enable_cpucache(cachep, GFP_NOWAIT))
1326 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001327 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001328
Christoph Lameter97d06602012-07-06 15:25:11 -05001329 /* Done! */
1330 slab_state = FULL;
1331
David Rientjes8f9f8d92010-03-27 19:40:47 -07001332#ifdef CONFIG_NUMA
1333 /*
1334 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001335 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001336 */
1337 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1338#endif
1339
Andrew Mortona737b3e2006-03-22 00:08:11 -08001340 /*
1341 * The reap timers are started later, with a module init call: That part
1342 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 */
1344}
1345
1346static int __init cpucache_init(void)
1347{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001348 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349
Andrew Mortona737b3e2006-03-22 00:08:11 -08001350 /*
1351 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 */
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001353 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1354 slab_online_cpu, slab_offline_cpu);
1355 WARN_ON(ret < 0);
Glauber Costaa164f8962012-06-21 00:59:18 +04001356
1357 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001358 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 return 0;
1360}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361__initcall(cpucache_init);
1362
Rafael Aquini8bdec192012-03-09 17:27:27 -03001363static noinline void
1364slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1365{
David Rientjes9a02d692014-06-04 16:06:36 -07001366#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001367 struct kmem_cache_node *n;
Joonsoo Kim8456a642013-10-24 10:07:49 +09001368 struct page *page;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001369 unsigned long flags;
1370 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001371 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1372 DEFAULT_RATELIMIT_BURST);
1373
1374 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1375 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001376
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001377 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1378 nodeid, gfpflags, &gfpflags);
1379 pr_warn(" cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001380 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001381
Christoph Lameter18bf8542014-08-06 16:04:11 -07001382 for_each_kmem_cache_node(cachep, node, n) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001383 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1384 unsigned long active_slabs = 0, num_slabs = 0;
1385
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001386 spin_lock_irqsave(&n->list_lock, flags);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001387 list_for_each_entry(page, &n->slabs_full, lru) {
Rafael Aquini8bdec192012-03-09 17:27:27 -03001388 active_objs += cachep->num;
1389 active_slabs++;
1390 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001391 list_for_each_entry(page, &n->slabs_partial, lru) {
1392 active_objs += page->active;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001393 active_slabs++;
1394 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09001395 list_for_each_entry(page, &n->slabs_free, lru)
Rafael Aquini8bdec192012-03-09 17:27:27 -03001396 num_slabs++;
1397
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001398 free_objects += n->free_objects;
1399 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001400
1401 num_slabs += active_slabs;
1402 num_objs = num_slabs * cachep->num;
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001403 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
Rafael Aquini8bdec192012-03-09 17:27:27 -03001404 node, active_slabs, num_slabs, active_objs, num_objs,
1405 free_objects);
1406 }
David Rientjes9a02d692014-06-04 16:06:36 -07001407#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001408}
1409
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001411 * Interface to system's page allocator. No need to hold the
1412 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 *
1414 * If we requested dmaable memory, we will get it. Even if we
1415 * did not request dmaable memory, we might get it, but that
1416 * would be relatively rare and ignorable.
1417 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001418static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1419 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420{
1421 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001422 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001423
Glauber Costaa618e892012-06-14 16:17:21 +04001424 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001425 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1426 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001427
Vlastimil Babka96db8002015-09-08 15:03:50 -07001428 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001429 if (!page) {
David Rientjes9a02d692014-06-04 16:06:36 -07001430 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001432 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001434 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1435 __free_pages(page, cachep->gfporder);
1436 return NULL;
1437 }
1438
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001439 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001441 add_zone_page_state(page_zone(page),
1442 NR_SLAB_RECLAIMABLE, nr_pages);
1443 else
1444 add_zone_page_state(page_zone(page),
1445 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001446
Joonsoo Kima57a4982013-10-24 10:07:44 +09001447 __SetPageSlab(page);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001448 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1449 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001450 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001451
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001452 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1453 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1454
1455 if (cachep->ctor)
1456 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1457 else
1458 kmemcheck_mark_unallocated_pages(page, nr_pages);
1459 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001460
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001461 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462}
1463
1464/*
1465 * Interface to system's page release.
1466 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001467static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468{
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001469 int order = cachep->gfporder;
1470 unsigned long nr_freed = (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001472 kmemcheck_free_shadow(page, order);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001473
Christoph Lameter972d1a72006-09-25 23:31:51 -07001474 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1475 sub_zone_page_state(page_zone(page),
1476 NR_SLAB_RECLAIMABLE, nr_freed);
1477 else
1478 sub_zone_page_state(page_zone(page),
1479 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001480
Joonsoo Kima57a4982013-10-24 10:07:44 +09001481 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001482 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001483 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001484 page_mapcount_reset(page);
1485 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 if (current->reclaim_state)
1488 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001489 memcg_uncharge_slab(page, order, cachep);
1490 __free_pages(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491}
1492
1493static void kmem_rcu_free(struct rcu_head *head)
1494{
Joonsoo Kim68126702013-10-24 10:07:42 +09001495 struct kmem_cache *cachep;
1496 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497
Joonsoo Kim68126702013-10-24 10:07:42 +09001498 page = container_of(head, struct page, rcu_head);
1499 cachep = page->slab_cache;
1500
1501 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502}
1503
1504#if DEBUG
Joonsoo Kim40b44132016-03-15 14:54:21 -07001505static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1506{
1507 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1508 (cachep->size % PAGE_SIZE) == 0)
1509 return true;
1510
1511 return false;
1512}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513
1514#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001515static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001516 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001518 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001520 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001522 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 return;
1524
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001525 *addr++ = 0x12345678;
1526 *addr++ = caller;
1527 *addr++ = smp_processor_id();
1528 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 {
1530 unsigned long *sptr = &caller;
1531 unsigned long svalue;
1532
1533 while (!kstack_end(sptr)) {
1534 svalue = *sptr++;
1535 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001536 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 size -= sizeof(unsigned long);
1538 if (size <= sizeof(unsigned long))
1539 break;
1540 }
1541 }
1542
1543 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001544 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545}
Joonsoo Kim40b44132016-03-15 14:54:21 -07001546
1547static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1548 int map, unsigned long caller)
1549{
1550 if (!is_debug_pagealloc_cache(cachep))
1551 return;
1552
1553 if (caller)
1554 store_stackinfo(cachep, objp, caller);
1555
1556 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1557}
1558
1559#else
1560static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1561 int map, unsigned long caller) {}
1562
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563#endif
1564
Pekka Enberg343e0d72006-02-01 03:05:50 -08001565static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001567 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001568 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
1570 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001571 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572}
1573
1574static void dump_line(char *data, int offset, int limit)
1575{
1576 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001577 unsigned char error = 0;
1578 int bad_count = 0;
1579
Joe Perches11705322016-03-17 14:19:50 -07001580 pr_err("%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001581 for (i = 0; i < limit; i++) {
1582 if (data[offset + i] != POISON_FREE) {
1583 error = data[offset + i];
1584 bad_count++;
1585 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001586 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001587 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1588 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001589
1590 if (bad_count == 1) {
1591 error ^= POISON_FREE;
1592 if (!(error & (error - 1))) {
Joe Perches11705322016-03-17 14:19:50 -07001593 pr_err("Single bit error detected. Probably bad RAM.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001594#ifdef CONFIG_X86
Joe Perches11705322016-03-17 14:19:50 -07001595 pr_err("Run memtest86+ or a similar memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001596#else
Joe Perches11705322016-03-17 14:19:50 -07001597 pr_err("Run a memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001598#endif
1599 }
1600 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601}
1602#endif
1603
1604#if DEBUG
1605
Pekka Enberg343e0d72006-02-01 03:05:50 -08001606static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607{
1608 int i, size;
1609 char *realobj;
1610
1611 if (cachep->flags & SLAB_RED_ZONE) {
Joe Perches11705322016-03-17 14:19:50 -07001612 pr_err("Redzone: 0x%llx/0x%llx\n",
1613 *dbg_redzone1(cachep, objp),
1614 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615 }
1616
1617 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches11705322016-03-17 14:19:50 -07001618 pr_err("Last user: [<%p>](%pSR)\n",
Joe Perches071361d2012-12-12 10:19:12 -08001619 *dbg_userword(cachep, objp),
1620 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001622 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001623 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001624 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 int limit;
1626 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001627 if (i + limit > size)
1628 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 dump_line(realobj, i, limit);
1630 }
1631}
1632
Pekka Enberg343e0d72006-02-01 03:05:50 -08001633static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634{
1635 char *realobj;
1636 int size, i;
1637 int lines = 0;
1638
Joonsoo Kim40b44132016-03-15 14:54:21 -07001639 if (is_debug_pagealloc_cache(cachep))
1640 return;
1641
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001642 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001643 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001645 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001647 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 exp = POISON_END;
1649 if (realobj[i] != exp) {
1650 int limit;
1651 /* Mismatch ! */
1652 /* Print header */
1653 if (lines == 0) {
Joe Perches11705322016-03-17 14:19:50 -07001654 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1655 print_tainted(), cachep->name,
1656 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 print_objinfo(cachep, objp, 0);
1658 }
1659 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001660 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001662 if (i + limit > size)
1663 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 dump_line(realobj, i, limit);
1665 i += 16;
1666 lines++;
1667 /* Limit to 5 lines */
1668 if (lines > 5)
1669 break;
1670 }
1671 }
1672 if (lines != 0) {
1673 /* Print some data about the neighboring objects, if they
1674 * exist:
1675 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001676 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001677 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678
Joonsoo Kim8456a642013-10-24 10:07:49 +09001679 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001681 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001682 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001683 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 print_objinfo(cachep, objp, 2);
1685 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001686 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001687 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001688 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001689 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 print_objinfo(cachep, objp, 2);
1691 }
1692 }
1693}
1694#endif
1695
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001697static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1698 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001699{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 int i;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001701
1702 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1703 poison_obj(cachep, page->freelist - obj_offset(cachep),
1704 POISON_FREE);
1705 }
1706
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001708 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709
1710 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07001712 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 }
1714 if (cachep->flags & SLAB_RED_ZONE) {
1715 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07001716 slab_error(cachep, "start of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07001718 slab_error(cachep, "end of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001721}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001723static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1724 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001725{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001726}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727#endif
1728
Randy Dunlap911851e2006-03-22 00:08:14 -08001729/**
1730 * slab_destroy - destroy and release all objects in a slab
1731 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001732 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001733 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001734 * Destroy all the objs in a slab page, and release the mem back to the system.
1735 * Before calling the slab page must have been unlinked from the cache. The
1736 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001737 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001738static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001739{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001740 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001741
Joonsoo Kim8456a642013-10-24 10:07:49 +09001742 freelist = page->freelist;
1743 slab_destroy_debugcheck(cachep, page);
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001744 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1745 call_rcu(&page->rcu_head, kmem_rcu_free);
1746 else
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001747 kmem_freepages(cachep, page);
Joonsoo Kim68126702013-10-24 10:07:42 +09001748
1749 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001750 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001751 * although actual page can be freed in rcu context
1752 */
1753 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001754 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755}
1756
Joonsoo Kim97654df2014-08-06 16:04:25 -07001757static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1758{
1759 struct page *page, *n;
1760
1761 list_for_each_entry_safe(page, n, list, lru) {
1762 list_del(&page->lru);
1763 slab_destroy(cachep, page);
1764 }
1765}
1766
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001768 * calculate_slab_order - calculate size (page order) of slabs
1769 * @cachep: pointer to the cache that is being created
1770 * @size: size of objects to be created in this cache.
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001771 * @flags: slab allocation flags
1772 *
1773 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001774 *
1775 * This could be made much more intelligent. For now, try to avoid using
1776 * high order pages for slabs. When the gfp() functions are more friendly
1777 * towards high-order requests, this should be changed.
1778 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001779static size_t calculate_slab_order(struct kmem_cache *cachep,
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07001780 size_t size, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001781{
1782 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001783 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001784
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001785 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001786 unsigned int num;
1787 size_t remainder;
1788
Joonsoo Kim70f75062016-03-15 14:54:53 -07001789 num = cache_estimate(gfporder, size, flags, &remainder);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001790 if (!num)
1791 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001792
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001793 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1794 if (num > SLAB_OBJ_MAX_NUM)
1795 break;
1796
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001797 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001798 struct kmem_cache *freelist_cache;
1799 size_t freelist_size;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001800
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001801 freelist_size = num * sizeof(freelist_idx_t);
1802 freelist_cache = kmalloc_slab(freelist_size, 0u);
1803 if (!freelist_cache)
1804 continue;
1805
1806 /*
1807 * Needed to avoid possible looping condition
Joonsoo Kim76b342b2016-05-19 17:10:26 -07001808 * in cache_grow_begin()
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001809 */
1810 if (OFF_SLAB(freelist_cache))
1811 continue;
1812
1813 /* check if off slab has enough benefit */
1814 if (freelist_cache->size > cachep->size / 2)
1815 continue;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001816 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001817
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001818 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001819 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001820 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001821 left_over = remainder;
1822
1823 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001824 * A VFS-reclaimable slab tends to have most allocations
1825 * as GFP_NOFS and we really don't want to have to be allocating
1826 * higher-order pages when we are unable to shrink dcache.
1827 */
1828 if (flags & SLAB_RECLAIM_ACCOUNT)
1829 break;
1830
1831 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001832 * Large number of objects is good, but very large slabs are
1833 * currently bad for the gfp()s.
1834 */
David Rientjes543585c2011-10-18 22:09:24 -07001835 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001836 break;
1837
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001838 /*
1839 * Acceptable internal fragmentation?
1840 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001841 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001842 break;
1843 }
1844 return left_over;
1845}
1846
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001847static struct array_cache __percpu *alloc_kmem_cache_cpus(
1848 struct kmem_cache *cachep, int entries, int batchcount)
1849{
1850 int cpu;
1851 size_t size;
1852 struct array_cache __percpu *cpu_cache;
1853
1854 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07001855 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001856
1857 if (!cpu_cache)
1858 return NULL;
1859
1860 for_each_possible_cpu(cpu) {
1861 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1862 entries, batchcount);
1863 }
1864
1865 return cpu_cache;
1866}
1867
Fabian Frederickbd721ea2016-08-02 14:03:33 -07001868static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001869{
Christoph Lameter97d06602012-07-06 15:25:11 -05001870 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03001871 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001872
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001873 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1874 if (!cachep->cpu_cache)
1875 return 1;
1876
Christoph Lameter97d06602012-07-06 15:25:11 -05001877 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001878 /* Creation of first cache (kmem_cache). */
1879 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001880 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001881 /* For kmem_cache_node */
1882 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001883 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001884 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001885
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001886 for_each_online_node(node) {
1887 cachep->node[node] = kmalloc_node(
1888 sizeof(struct kmem_cache_node), gfp, node);
1889 BUG_ON(!cachep->node[node]);
1890 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001891 }
1892 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001893
Christoph Lameter6a673682013-01-10 19:14:19 +00001894 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001895 jiffies + REAPTIMEOUT_NODE +
1896 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001897
1898 cpu_cache_get(cachep)->avail = 0;
1899 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1900 cpu_cache_get(cachep)->batchcount = 1;
1901 cpu_cache_get(cachep)->touched = 0;
1902 cachep->batchcount = 1;
1903 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001904 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001905}
1906
Joonsoo Kim12220de2014-10-09 15:26:24 -07001907unsigned long kmem_cache_flags(unsigned long object_size,
1908 unsigned long flags, const char *name,
1909 void (*ctor)(void *))
1910{
1911 return flags;
1912}
1913
1914struct kmem_cache *
1915__kmem_cache_alias(const char *name, size_t size, size_t align,
1916 unsigned long flags, void (*ctor)(void *))
1917{
1918 struct kmem_cache *cachep;
1919
1920 cachep = find_mergeable(size, align, flags, name, ctor);
1921 if (cachep) {
1922 cachep->refcount++;
1923
1924 /*
1925 * Adjust the object sizes so that we clear
1926 * the complete object on kzalloc.
1927 */
1928 cachep->object_size = max_t(int, cachep->object_size, size);
1929 }
1930 return cachep;
1931}
1932
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001933static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1934 size_t size, unsigned long flags)
1935{
1936 size_t left;
1937
1938 cachep->num = 0;
1939
1940 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1941 return false;
1942
1943 left = calculate_slab_order(cachep, size,
1944 flags | CFLGS_OBJFREELIST_SLAB);
1945 if (!cachep->num)
1946 return false;
1947
1948 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1949 return false;
1950
1951 cachep->colour = left / cachep->colour_off;
1952
1953 return true;
1954}
1955
Joonsoo Kim158e3192016-03-15 14:54:35 -07001956static bool set_off_slab_cache(struct kmem_cache *cachep,
1957 size_t size, unsigned long flags)
1958{
1959 size_t left;
1960
1961 cachep->num = 0;
1962
1963 /*
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001964 * Always use on-slab management when SLAB_NOLEAKTRACE
1965 * to avoid recursive calls into kmemleak.
Joonsoo Kim158e3192016-03-15 14:54:35 -07001966 */
Joonsoo Kim158e3192016-03-15 14:54:35 -07001967 if (flags & SLAB_NOLEAKTRACE)
1968 return false;
1969
1970 /*
1971 * Size is large, assume best to place the slab management obj
1972 * off-slab (should allow better packing of objs).
1973 */
1974 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1975 if (!cachep->num)
1976 return false;
1977
1978 /*
1979 * If the slab has been placed off-slab, and we have enough space then
1980 * move it on-slab. This is at the expense of any extra colouring.
1981 */
1982 if (left >= cachep->num * sizeof(freelist_idx_t))
1983 return false;
1984
1985 cachep->colour = left / cachep->colour_off;
1986
1987 return true;
1988}
1989
1990static bool set_on_slab_cache(struct kmem_cache *cachep,
1991 size_t size, unsigned long flags)
1992{
1993 size_t left;
1994
1995 cachep->num = 0;
1996
1997 left = calculate_slab_order(cachep, size, flags);
1998 if (!cachep->num)
1999 return false;
2000
2001 cachep->colour = left / cachep->colour_off;
2002
2003 return true;
2004}
2005
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002006/**
Christoph Lameter039363f2012-07-06 15:25:10 -05002007 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08002008 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 *
2011 * Returns a ptr to the cache on success, NULL on failure.
2012 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002013 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 * The flags are
2016 *
2017 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2018 * to catch references to uninitialised memory.
2019 *
2020 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2021 * for buffer overruns.
2022 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2024 * cacheline. This can be beneficial if you're counting cycles as closely
2025 * as davem.
2026 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002027int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002028__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002030 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002031 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002032 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002033 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036#if FORCED_DEBUG
2037 /*
2038 * Enable redzoning and last user accounting, except for caches with
2039 * large objects, if the increased size would increase the object size
2040 * above the next power of two: caches with object sizes just above a
2041 * power of two have a significant amount of internal fragmentation.
2042 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002043 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2044 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002045 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 if (!(flags & SLAB_DESTROY_BY_RCU))
2047 flags |= SLAB_POISON;
2048#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050
Andrew Mortona737b3e2006-03-22 00:08:11 -08002051 /*
2052 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 * unaligned accesses for some archs when redzoning is used, and makes
2054 * sure any on-slab bufctl's are also correctly aligned.
2055 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002056 if (size & (BYTES_PER_WORD - 1)) {
2057 size += (BYTES_PER_WORD - 1);
2058 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 }
2060
David Woodhouse87a927c2007-07-04 21:26:44 -04002061 if (flags & SLAB_RED_ZONE) {
2062 ralign = REDZONE_ALIGN;
2063 /* If redzoning, ensure that the second redzone is suitably
2064 * aligned, by adjusting the object size accordingly. */
2065 size += REDZONE_ALIGN - 1;
2066 size &= ~(REDZONE_ALIGN - 1);
2067 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002068
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002069 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002070 if (ralign < cachep->align) {
2071 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002073 /* disable debug if necessary */
2074 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002075 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002076 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002077 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002079 cachep->align = ralign;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002080 cachep->colour_off = cache_line_size();
2081 /* Offset must be a multiple of the alignment. */
2082 if (cachep->colour_off < cachep->align)
2083 cachep->colour_off = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084
Pekka Enberg83b519e2009-06-10 19:40:04 +03002085 if (slab_is_available())
2086 gfp = GFP_KERNEL;
2087 else
2088 gfp = GFP_NOWAIT;
2089
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091
Pekka Enbergca5f9702006-09-25 23:31:25 -07002092 /*
2093 * Both debugging options require word-alignment which is calculated
2094 * into align above.
2095 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002098 cachep->obj_offset += sizeof(unsigned long long);
2099 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 }
2101 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002102 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002103 * the real object. But if the second red zone needs to be
2104 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002106 if (flags & SLAB_RED_ZONE)
2107 size += REDZONE_ALIGN;
2108 else
2109 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 }
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002111#endif
2112
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002113 kasan_cache_create(cachep, &size, &flags);
2114
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002115 size = ALIGN(size, cachep->align);
2116 /*
2117 * We should restrict the number of objects in a slab to implement
2118 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2119 */
2120 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2121 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2122
2123#if DEBUG
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002124 /*
2125 * To activate debug pagealloc, off-slab management is necessary
2126 * requirement. In early phase of initialization, small sized slab
2127 * doesn't get initialized so it would not be possible. So, we need
2128 * to check size >= 256. It guarantees that all necessary small
2129 * sized slab is initialized in current slab initialization sequence.
2130 */
Joonsoo Kim40323272016-03-15 14:54:18 -07002131 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
Joonsoo Kimf3a3c322016-03-15 14:54:38 -07002132 size >= 256 && cachep->object_size > cache_line_size()) {
2133 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2134 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2135
2136 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2137 flags |= CFLGS_OFF_SLAB;
2138 cachep->obj_offset += tmp_size - size;
2139 size = tmp_size;
2140 goto done;
2141 }
2142 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 }
2144#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002146 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2147 flags |= CFLGS_OBJFREELIST_SLAB;
2148 goto done;
2149 }
2150
Joonsoo Kim158e3192016-03-15 14:54:35 -07002151 if (set_off_slab_cache(cachep, size, flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 flags |= CFLGS_OFF_SLAB;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002153 goto done;
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002154 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155
Joonsoo Kim158e3192016-03-15 14:54:35 -07002156 if (set_on_slab_cache(cachep, size, flags))
2157 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158
Joonsoo Kim158e3192016-03-15 14:54:35 -07002159 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002160
Joonsoo Kim158e3192016-03-15 14:54:35 -07002161done:
2162 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002164 cachep->allocflags = __GFP_COMP;
Yang Shia3187e42016-05-19 17:10:41 -07002165 if (flags & SLAB_CACHE_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002166 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002167 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002168 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169
Joonsoo Kim40b44132016-03-15 14:54:21 -07002170#if DEBUG
2171 /*
2172 * If we're going to use the generic kernel_map_pages()
2173 * poisoning, then it's going to smash the contents of
2174 * the redzone and userword anyhow, so switch them off.
2175 */
2176 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2177 (cachep->flags & SLAB_POISON) &&
2178 is_debug_pagealloc_cache(cachep))
2179 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2180#endif
2181
2182 if (OFF_SLAB(cachep)) {
Joonsoo Kim158e3192016-03-15 14:54:35 -07002183 cachep->freelist_cache =
2184 kmalloc_slab(cachep->freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002185 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002187 err = setup_cpu_cache(cachep, gfp);
2188 if (err) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002189 __kmem_cache_release(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002190 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002191 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002193 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195
2196#if DEBUG
2197static void check_irq_off(void)
2198{
2199 BUG_ON(!irqs_disabled());
2200}
2201
2202static void check_irq_on(void)
2203{
2204 BUG_ON(irqs_disabled());
2205}
2206
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002207static void check_mutex_acquired(void)
2208{
2209 BUG_ON(!mutex_is_locked(&slab_mutex));
2210}
2211
Pekka Enberg343e0d72006-02-01 03:05:50 -08002212static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213{
2214#ifdef CONFIG_SMP
2215 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002216 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217#endif
2218}
Christoph Lametere498be72005-09-09 13:03:32 -07002219
Pekka Enberg343e0d72006-02-01 03:05:50 -08002220static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002221{
2222#ifdef CONFIG_SMP
2223 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002224 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002225#endif
2226}
2227
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228#else
2229#define check_irq_off() do { } while(0)
2230#define check_irq_on() do { } while(0)
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002231#define check_mutex_acquired() do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002233#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234#endif
2235
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002236static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2237 int node, bool free_all, struct list_head *list)
2238{
2239 int tofree;
2240
2241 if (!ac || !ac->avail)
2242 return;
2243
2244 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2245 if (tofree > ac->avail)
2246 tofree = (ac->avail + 1) / 2;
2247
2248 free_block(cachep, ac->entry, tofree, node, list);
2249 ac->avail -= tofree;
2250 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2251}
Christoph Lameteraab22072006-03-22 00:09:06 -08002252
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253static void do_drain(void *arg)
2254{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002255 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002257 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002258 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002259 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260
2261 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002262 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002263 n = get_node(cachep, node);
2264 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002265 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002266 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002267 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 ac->avail = 0;
2269}
2270
Pekka Enberg343e0d72006-02-01 03:05:50 -08002271static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002273 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002274 int node;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002275 LIST_HEAD(list);
Christoph Lametere498be72005-09-09 13:03:32 -07002276
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002277 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002279 for_each_kmem_cache_node(cachep, node, n)
2280 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002281 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002282
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002283 for_each_kmem_cache_node(cachep, node, n) {
2284 spin_lock_irq(&n->list_lock);
2285 drain_array_locked(cachep, n->shared, node, true, &list);
2286 spin_unlock_irq(&n->list_lock);
2287
2288 slabs_destroy(cachep, &list);
2289 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290}
2291
Christoph Lametered11d9e2006-06-30 01:55:45 -07002292/*
2293 * Remove slabs from the list of free slabs.
2294 * Specify the number of slabs to drain in tofree.
2295 *
2296 * Returns the actual number of slabs released.
2297 */
2298static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002299 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002301 struct list_head *p;
2302 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002303 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304
Christoph Lametered11d9e2006-06-30 01:55:45 -07002305 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002306 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002308 spin_lock_irq(&n->list_lock);
2309 p = n->slabs_free.prev;
2310 if (p == &n->slabs_free) {
2311 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002312 goto out;
2313 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314
Joonsoo Kim8456a642013-10-24 10:07:49 +09002315 page = list_entry(p, struct page, lru);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002316 list_del(&page->lru);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002317 /*
2318 * Safe to drop the lock. The slab is no longer linked
2319 * to the cache.
2320 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002321 n->free_objects -= cache->num;
2322 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002323 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002324 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002326out:
2327 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328}
2329
Vladimir Davydovd6e0b7f2015-02-12 14:59:47 -08002330int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
Christoph Lametere498be72005-09-09 13:03:32 -07002331{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002332 int ret = 0;
2333 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002334 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002335
2336 drain_cpu_caches(cachep);
2337
2338 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002339 for_each_kmem_cache_node(cachep, node, n) {
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07002340 drain_freelist(cachep, n, INT_MAX);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002341
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002342 ret += !list_empty(&n->slabs_full) ||
2343 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002344 }
2345 return (ret ? 1 : 0);
2346}
2347
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002348int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349{
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002350 return __kmem_cache_shrink(cachep, false);
2351}
2352
2353void __kmem_cache_release(struct kmem_cache *cachep)
2354{
Christoph Lameter12c36672012-09-04 23:38:33 +00002355 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002356 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002357
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002358 cache_random_seq_destroy(cachep);
2359
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002360 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002361
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002362 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002363 for_each_kmem_cache_node(cachep, i, n) {
2364 kfree(n->shared);
2365 free_alien_cache(n->alien);
2366 kfree(n);
2367 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002371/*
2372 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002373 *
2374 * For a slab cache when the slab descriptor is off-slab, the
2375 * slab descriptor can't come from the same cache which is being created,
2376 * Because if it is the case, that means we defer the creation of
2377 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2378 * And we eventually call down to __kmem_cache_create(), which
2379 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2380 * This is a "chicken-and-egg" problem.
2381 *
2382 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2383 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002384 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002385static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002386 struct page *page, int colour_off,
2387 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002389 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002390 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002391
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002392 page->s_mem = addr + colour_off;
2393 page->active = 0;
2394
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002395 if (OBJFREELIST_SLAB(cachep))
2396 freelist = NULL;
2397 else if (OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002399 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002400 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002401 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 return NULL;
2403 } else {
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002404 /* We will use last bytes at the slab for freelist */
2405 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2406 cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 }
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002408
Joonsoo Kim8456a642013-10-24 10:07:49 +09002409 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410}
2411
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002412static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002414 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002415}
2416
2417static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc689732014-04-18 16:24:09 +09002418 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002419{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002420 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421}
2422
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002423static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424{
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002425#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 int i;
2427
2428 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002429 void *objp = index_to_obj(cachep, page, i);
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002430
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 if (cachep->flags & SLAB_STORE_USER)
2432 *dbg_userword(cachep, objp) = NULL;
2433
2434 if (cachep->flags & SLAB_RED_ZONE) {
2435 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2436 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2437 }
2438 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002439 * Constructors are not allowed to allocate memory from the same
2440 * cache which they are a constructor for. Otherwise, deadlock.
2441 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002443 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2444 kasan_unpoison_object_data(cachep,
2445 objp + obj_offset(cachep));
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002446 cachep->ctor(objp + obj_offset(cachep));
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002447 kasan_poison_object_data(
2448 cachep, objp + obj_offset(cachep));
2449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450
2451 if (cachep->flags & SLAB_RED_ZONE) {
2452 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07002453 slab_error(cachep, "constructor overwrote the end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a025f02016-03-17 14:19:47 -07002455 slab_error(cachep, "constructor overwrote the start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 }
Joonsoo Kim40b44132016-03-15 14:54:21 -07002457 /* need to poison the objs? */
2458 if (cachep->flags & SLAB_POISON) {
2459 poison_obj(cachep, objp, POISON_FREE);
2460 slab_kernel_map(cachep, objp, 0, 0);
2461 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463#endif
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002464}
2465
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002466#ifdef CONFIG_SLAB_FREELIST_RANDOM
2467/* Hold information during a freelist initialization */
2468union freelist_init_state {
2469 struct {
2470 unsigned int pos;
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002471 unsigned int *list;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002472 unsigned int count;
2473 unsigned int rand;
2474 };
2475 struct rnd_state rnd_state;
2476};
2477
2478/*
2479 * Initialize the state based on the randomization methode available.
2480 * return true if the pre-computed list is available, false otherwize.
2481 */
2482static bool freelist_state_initialize(union freelist_init_state *state,
2483 struct kmem_cache *cachep,
2484 unsigned int count)
2485{
2486 bool ret;
2487 unsigned int rand;
2488
2489 /* Use best entropy available to define a random shift */
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002490 rand = get_random_int();
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002491
2492 /* Use a random state if the pre-computed list is not available */
2493 if (!cachep->random_seq) {
2494 prandom_seed_state(&state->rnd_state, rand);
2495 ret = false;
2496 } else {
2497 state->list = cachep->random_seq;
2498 state->count = count;
2499 state->pos = 0;
2500 state->rand = rand;
2501 ret = true;
2502 }
2503 return ret;
2504}
2505
2506/* Get the next entry on the list and randomize it using a random shift */
2507static freelist_idx_t next_random_slot(union freelist_init_state *state)
2508{
2509 return (state->list[state->pos++] + state->rand) % state->count;
2510}
2511
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002512/* Swap two freelist entries */
2513static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2514{
2515 swap(((freelist_idx_t *)page->freelist)[a],
2516 ((freelist_idx_t *)page->freelist)[b]);
2517}
2518
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002519/*
2520 * Shuffle the freelist initialization state based on pre-computed lists.
2521 * return true if the list was successfully shuffled, false otherwise.
2522 */
2523static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2524{
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002525 unsigned int objfreelist = 0, i, rand, count = cachep->num;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002526 union freelist_init_state state;
2527 bool precomputed;
2528
2529 if (count < 2)
2530 return false;
2531
2532 precomputed = freelist_state_initialize(&state, cachep, count);
2533
2534 /* Take a random entry as the objfreelist */
2535 if (OBJFREELIST_SLAB(cachep)) {
2536 if (!precomputed)
2537 objfreelist = count - 1;
2538 else
2539 objfreelist = next_random_slot(&state);
2540 page->freelist = index_to_obj(cachep, page, objfreelist) +
2541 obj_offset(cachep);
2542 count--;
2543 }
2544
2545 /*
2546 * On early boot, generate the list dynamically.
2547 * Later use a pre-computed list for speed.
2548 */
2549 if (!precomputed) {
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002550 for (i = 0; i < count; i++)
2551 set_free_obj(page, i, i);
2552
2553 /* Fisher-Yates shuffle */
2554 for (i = count - 1; i > 0; i--) {
2555 rand = prandom_u32_state(&state.rnd_state);
2556 rand %= (i + 1);
2557 swap_free_obj(page, i, rand);
2558 }
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002559 } else {
2560 for (i = 0; i < count; i++)
2561 set_free_obj(page, i, next_random_slot(&state));
2562 }
2563
2564 if (OBJFREELIST_SLAB(cachep))
2565 set_free_obj(page, cachep->num - 1, objfreelist);
2566
2567 return true;
2568}
2569#else
2570static inline bool shuffle_freelist(struct kmem_cache *cachep,
2571 struct page *page)
2572{
2573 return false;
2574}
2575#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2576
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002577static void cache_init_objs(struct kmem_cache *cachep,
2578 struct page *page)
2579{
2580 int i;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002581 void *objp;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002582 bool shuffled;
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002583
2584 cache_init_objs_debug(cachep, page);
2585
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002586 /* Try to randomize the freelist if enabled */
2587 shuffled = shuffle_freelist(cachep, page);
2588
2589 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002590 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2591 obj_offset(cachep);
2592 }
2593
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002594 for (i = 0; i < cachep->num; i++) {
Andrey Ryabininb3cbd9b2016-08-02 14:02:52 -07002595 objp = index_to_obj(cachep, page, i);
2596 kasan_init_slab_obj(cachep, objp);
2597
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002598 /* constructor could break poison info */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002599 if (DEBUG == 0 && cachep->ctor) {
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002600 kasan_unpoison_object_data(cachep, objp);
2601 cachep->ctor(objp);
2602 kasan_poison_object_data(cachep, objp);
2603 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002604
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002605 if (!shuffled)
2606 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608}
2609
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002610static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002611{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002612 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002613
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002614 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002615 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002616
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002617#if DEBUG
2618 if (cachep->flags & SLAB_STORE_USER)
2619 set_store_user_dirty(cachep);
2620#endif
2621
Matthew Dobson78d382d2006-02-01 03:05:47 -08002622 return objp;
2623}
2624
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002625static void slab_put_obj(struct kmem_cache *cachep,
2626 struct page *page, void *objp)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002627{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002628 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002629#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002630 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002631
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002632 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002633 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002634 if (get_free_obj(page, i) == objnr) {
Joe Perches11705322016-03-17 14:19:50 -07002635 pr_err("slab: double free detected in cache '%s', objp %p\n",
Joe Perches756a025f02016-03-17 14:19:47 -07002636 cachep->name, objp);
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002637 BUG();
2638 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002639 }
2640#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002641 page->active--;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002642 if (!page->freelist)
2643 page->freelist = objp + obj_offset(cachep);
2644
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002645 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002646}
2647
Pekka Enberg47768742006-06-23 02:03:07 -07002648/*
2649 * Map pages beginning at addr to the given cache and slab. This is required
2650 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002651 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002652 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002653static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002654 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002656 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002657 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658}
2659
2660/*
2661 * Grow (by 1) the number of slabs within a cache. This is called by
2662 * kmem_cache_alloc() when there are no active objs left in a cache.
2663 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002664static struct page *cache_grow_begin(struct kmem_cache *cachep,
2665 gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002667 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002668 size_t offset;
2669 gfp_t local_flags;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002670 int page_node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002671 struct kmem_cache_node *n;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002672 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673
Andrew Mortona737b3e2006-03-22 00:08:11 -08002674 /*
2675 * Be lazy and only check for valid flags here, keeping it out of the
2676 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002678 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
Michal Hockobacdcb32016-07-26 15:22:02 -07002679 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
Michal Hocko72baeef0c2016-07-26 15:22:05 -07002680 flags &= ~GFP_SLAB_BUG_MASK;
2681 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2682 invalid_mask, &invalid_mask, flags, &flags);
2683 dump_stack();
Andrew Mortonc871ac42014-12-10 15:42:25 -08002684 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002685 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 check_irq_off();
Mel Gormand0164ad2015-11-06 16:28:21 -08002688 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689 local_irq_enable();
2690
2691 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002692 * Get mem for the objs. Attempt to allocate a physical page from
2693 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002694 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002695 page = kmem_getpages(cachep, local_flags, nodeid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002696 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 goto failed;
2698
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002699 page_node = page_to_nid(page);
2700 n = get_node(cachep, page_node);
Joonsoo Kim03d1d432016-05-19 17:10:20 -07002701
2702 /* Get colour for the slab, and cal the next value. */
2703 n->colour_next++;
2704 if (n->colour_next >= cachep->colour)
2705 n->colour_next = 0;
2706
2707 offset = n->colour_next;
2708 if (offset >= cachep->colour)
2709 offset = 0;
2710
2711 offset *= cachep->colour_off;
2712
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002714 freelist = alloc_slabmgmt(cachep, page, offset,
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002715 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002716 if (OFF_SLAB(cachep) && !freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 goto opps1;
2718
Joonsoo Kim8456a642013-10-24 10:07:49 +09002719 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002721 kasan_poison_slab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002722 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723
Mel Gormand0164ad2015-11-06 16:28:21 -08002724 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 local_irq_disable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002727 return page;
2728
Andrew Mortona737b3e2006-03-22 00:08:11 -08002729opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002730 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002731failed:
Mel Gormand0164ad2015-11-06 16:28:21 -08002732 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 local_irq_disable();
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002734 return NULL;
2735}
2736
2737static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2738{
2739 struct kmem_cache_node *n;
2740 void *list = NULL;
2741
2742 check_irq_off();
2743
2744 if (!page)
2745 return;
2746
2747 INIT_LIST_HEAD(&page->lru);
2748 n = get_node(cachep, page_to_nid(page));
2749
2750 spin_lock(&n->list_lock);
2751 if (!page->active)
2752 list_add_tail(&page->lru, &(n->slabs_free));
2753 else
2754 fixup_slab_list(cachep, n, page, &list);
2755 STATS_INC_GROWN(cachep);
2756 n->free_objects += cachep->num - page->active;
2757 spin_unlock(&n->list_lock);
2758
2759 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760}
2761
2762#if DEBUG
2763
2764/*
2765 * Perform extra freeing checks:
2766 * - detect bad pointers.
2767 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 */
2769static void kfree_debugcheck(const void *objp)
2770{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 if (!virt_addr_valid(objp)) {
Joe Perches11705322016-03-17 14:19:50 -07002772 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002773 (unsigned long)objp);
2774 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776}
2777
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002778static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2779{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002780 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002781
2782 redzone1 = *dbg_redzone1(cache, obj);
2783 redzone2 = *dbg_redzone2(cache, obj);
2784
2785 /*
2786 * Redzone is ok.
2787 */
2788 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2789 return;
2790
2791 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2792 slab_error(cache, "double free detected");
2793 else
2794 slab_error(cache, "memory outside object was overwritten");
2795
Joe Perches11705322016-03-17 14:19:50 -07002796 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2797 obj, redzone1, redzone2);
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002798}
2799
Pekka Enberg343e0d72006-02-01 03:05:50 -08002800static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002801 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002804 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002806 BUG_ON(virt_to_cache(objp) != cachep);
2807
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002808 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002810 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002813 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2815 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2816 }
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002817 if (cachep->flags & SLAB_STORE_USER) {
2818 set_store_user_dirty(cachep);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002819 *dbg_userword(cachep, objp) = (void *)caller;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002820 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821
Joonsoo Kim8456a642013-10-24 10:07:49 +09002822 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
2824 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002825 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 poison_obj(cachep, objp, POISON_FREE);
Joonsoo Kim40b44132016-03-15 14:54:21 -07002829 slab_kernel_map(cachep, objp, 0, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 }
2831 return objp;
2832}
2833
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834#else
2835#define kfree_debugcheck(x) do { } while(0)
2836#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002837#endif
2838
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002839static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2840 void **list)
2841{
2842#if DEBUG
2843 void *next = *list;
2844 void *objp;
2845
2846 while (next) {
2847 objp = next - obj_offset(cachep);
2848 next = *(void **)next;
2849 poison_obj(cachep, objp, POISON_FREE);
2850 }
2851#endif
2852}
2853
Joonsoo Kimd8410232016-03-15 14:54:44 -07002854static inline void fixup_slab_list(struct kmem_cache *cachep,
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002855 struct kmem_cache_node *n, struct page *page,
2856 void **list)
Joonsoo Kimd8410232016-03-15 14:54:44 -07002857{
2858 /* move slabp to correct slabp list: */
2859 list_del(&page->lru);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002860 if (page->active == cachep->num) {
Joonsoo Kimd8410232016-03-15 14:54:44 -07002861 list_add(&page->lru, &n->slabs_full);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002862 if (OBJFREELIST_SLAB(cachep)) {
2863#if DEBUG
2864 /* Poisoning will be done without holding the lock */
2865 if (cachep->flags & SLAB_POISON) {
2866 void **objp = page->freelist;
2867
2868 *objp = *list;
2869 *list = objp;
2870 }
2871#endif
2872 page->freelist = NULL;
2873 }
2874 } else
Joonsoo Kimd8410232016-03-15 14:54:44 -07002875 list_add(&page->lru, &n->slabs_partial);
2876}
2877
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002878/* Try to find non-pfmemalloc slab if needed */
2879static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2880 struct page *page, bool pfmemalloc)
2881{
2882 if (!page)
2883 return NULL;
2884
2885 if (pfmemalloc)
2886 return page;
2887
2888 if (!PageSlabPfmemalloc(page))
2889 return page;
2890
2891 /* No need to keep pfmemalloc slab if we have enough free objects */
2892 if (n->free_objects > n->free_limit) {
2893 ClearPageSlabPfmemalloc(page);
2894 return page;
2895 }
2896
2897 /* Move pfmemalloc slab to the end of list to speed up next search */
2898 list_del(&page->lru);
2899 if (!page->active)
2900 list_add_tail(&page->lru, &n->slabs_free);
2901 else
2902 list_add_tail(&page->lru, &n->slabs_partial);
2903
2904 list_for_each_entry(page, &n->slabs_partial, lru) {
2905 if (!PageSlabPfmemalloc(page))
2906 return page;
2907 }
2908
2909 list_for_each_entry(page, &n->slabs_free, lru) {
2910 if (!PageSlabPfmemalloc(page))
2911 return page;
2912 }
2913
2914 return NULL;
2915}
2916
2917static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
Geliang Tang7aa0d222016-01-14 15:18:02 -08002918{
2919 struct page *page;
2920
2921 page = list_first_entry_or_null(&n->slabs_partial,
2922 struct page, lru);
2923 if (!page) {
2924 n->free_touched = 1;
2925 page = list_first_entry_or_null(&n->slabs_free,
2926 struct page, lru);
2927 }
2928
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002929 if (sk_memalloc_socks())
2930 return get_valid_first_slab(n, page, pfmemalloc);
2931
Geliang Tang7aa0d222016-01-14 15:18:02 -08002932 return page;
2933}
2934
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002935static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2936 struct kmem_cache_node *n, gfp_t flags)
2937{
2938 struct page *page;
2939 void *obj;
2940 void *list = NULL;
2941
2942 if (!gfp_pfmemalloc_allowed(flags))
2943 return NULL;
2944
2945 spin_lock(&n->list_lock);
2946 page = get_first_slab(n, true);
2947 if (!page) {
2948 spin_unlock(&n->list_lock);
2949 return NULL;
2950 }
2951
2952 obj = slab_get_obj(cachep, page);
2953 n->free_objects--;
2954
2955 fixup_slab_list(cachep, n, page, &list);
2956
2957 spin_unlock(&n->list_lock);
2958 fixup_objfreelist_debug(cachep, &list);
2959
2960 return obj;
2961}
2962
Joonsoo Kim213b4692016-05-19 17:10:29 -07002963/*
2964 * Slab list should be fixed up by fixup_slab_list() for existing slab
2965 * or cache_grow_end() for new slab
2966 */
2967static __always_inline int alloc_block(struct kmem_cache *cachep,
2968 struct array_cache *ac, struct page *page, int batchcount)
2969{
2970 /*
2971 * There must be at least one object available for
2972 * allocation.
2973 */
2974 BUG_ON(page->active >= cachep->num);
2975
2976 while (page->active < cachep->num && batchcount--) {
2977 STATS_INC_ALLOCED(cachep);
2978 STATS_INC_ACTIVE(cachep);
2979 STATS_SET_HIGH(cachep);
2980
2981 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2982 }
2983
2984 return batchcount;
2985}
2986
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002987static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988{
2989 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002990 struct kmem_cache_node *n;
Joonsoo Kim801faf02016-05-19 17:10:31 -07002991 struct array_cache *ac, *shared;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002992 int node;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002993 void *list = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002994 struct page *page;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002995
Joe Korty6d2144d2008-03-05 15:04:59 -08002996 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002997 node = numa_mem_id();
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002998
Joe Korty6d2144d2008-03-05 15:04:59 -08002999 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 batchcount = ac->batchcount;
3001 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003002 /*
3003 * If there was little recent activity on this cache, then
3004 * perform only a partial refill. Otherwise we could generate
3005 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 */
3007 batchcount = BATCHREFILL_LIMIT;
3008 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07003009 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003011 BUG_ON(ac->avail > 0 || !n);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003012 shared = READ_ONCE(n->shared);
3013 if (!n->free_objects && (!shared || !shared->avail))
3014 goto direct_grow;
3015
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003016 spin_lock(&n->list_lock);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003017 shared = READ_ONCE(n->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003018
Christoph Lameter3ded1752006-03-25 03:06:44 -08003019 /* See if we can refill from the shared array */
Joonsoo Kim801faf02016-05-19 17:10:31 -07003020 if (shared && transfer_objects(ac, shared, batchcount)) {
3021 shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003022 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003023 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003024
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 while (batchcount > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 /* Get slab alloc is to come from. */
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003027 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003028 if (!page)
3029 goto must_grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003032
Joonsoo Kim213b4692016-05-19 17:10:29 -07003033 batchcount = alloc_block(cachep, ac, page, batchcount);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003034 fixup_slab_list(cachep, n, page, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 }
3036
Andrew Mortona737b3e2006-03-22 00:08:11 -08003037must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003038 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003039alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003040 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003041 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042
Joonsoo Kim801faf02016-05-19 17:10:31 -07003043direct_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 if (unlikely(!ac->avail)) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003045 /* Check if we can use obj in pfmemalloc slab */
3046 if (sk_memalloc_socks()) {
3047 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3048
3049 if (obj)
3050 return obj;
3051 }
3052
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003053 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
Christoph Lametere498be72005-09-09 13:03:32 -07003054
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003055 /*
3056 * cache_grow_begin() can reenable interrupts,
3057 * then ac could change.
3058 */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003059 ac = cpu_cache_get(cachep);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003060 if (!ac->avail && page)
3061 alloc_block(cachep, ac, page, batchcount);
3062 cache_grow_end(cachep, page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003063
Joonsoo Kim213b4692016-05-19 17:10:29 -07003064 if (!ac->avail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 }
3067 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003068
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003069 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070}
3071
Andrew Mortona737b3e2006-03-22 00:08:11 -08003072static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3073 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074{
Mel Gormand0164ad2015-11-06 16:28:21 -08003075 might_sleep_if(gfpflags_allow_blocking(flags));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076}
3077
3078#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003079static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003080 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003082 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003084 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07003086 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087 poison_obj(cachep, objp, POISON_INUSE);
3088 }
3089 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003090 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091
3092 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003093 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3094 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
Joe Perches756a025f02016-03-17 14:19:47 -07003095 slab_error(cachep, "double free, or memory outside object was overwritten");
Joe Perches11705322016-03-17 14:19:50 -07003096 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3097 objp, *dbg_redzone1(cachep, objp),
3098 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 }
3100 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3101 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3102 }
Joonsoo Kim03787302014-06-23 13:22:06 -07003103
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003104 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003105 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003106 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003107 if (ARCH_SLAB_MINALIGN &&
3108 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Joe Perches11705322016-03-17 14:19:50 -07003109 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003110 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003111 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 return objp;
3113}
3114#else
3115#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3116#endif
3117
Pekka Enberg343e0d72006-02-01 03:05:50 -08003118static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003120 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121 struct array_cache *ac;
3122
Alok N Kataria5c382302005-09-27 21:45:46 -07003123 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003124
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003125 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 ac->touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003128 objp = ac->entry[--ac->avail];
Mel Gorman072bb0a2012-07-31 16:43:58 -07003129
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003130 STATS_INC_ALLOCHIT(cachep);
3131 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003133
3134 STATS_INC_ALLOCMISS(cachep);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003135 objp = cache_alloc_refill(cachep, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003136 /*
3137 * the 'ac' may be updated by cache_alloc_refill(),
3138 * and kmemleak_erase() requires its correct value.
3139 */
3140 ac = cpu_cache_get(cachep);
3141
3142out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003143 /*
3144 * To avoid a false negative, if an object that is in one of the
3145 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3146 * treat the array pointers as a reference to the object.
3147 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003148 if (objp)
3149 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003150 return objp;
3151}
3152
Christoph Lametere498be72005-09-09 13:03:32 -07003153#ifdef CONFIG_NUMA
3154/*
Zefan Li2ad654b2014-09-25 09:41:02 +08003155 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003156 *
3157 * If we are in_interrupt, then process context, including cpusets and
3158 * mempolicy, may not apply and should not be used for allocation policy.
3159 */
3160static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3161{
3162 int nid_alloc, nid_here;
3163
Christoph Lameter765c4502006-09-27 01:50:08 -07003164 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003165 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003166 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003167 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003168 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003169 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003170 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003171 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003172 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003173 return NULL;
3174}
3175
3176/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003177 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003178 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003179 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003180 * perform an allocation without specifying a node. This allows the page
3181 * allocator to do its reclaim / fallback magic. We then insert the
3182 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003183 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003184static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003185{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003186 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07003187 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003188 struct zone *zone;
3189 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003190 void *obj = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003191 struct page *page;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003192 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003193 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003194
3195 if (flags & __GFP_THISNODE)
3196 return NULL;
3197
Mel Gormancc9a6c82012-03-21 16:34:11 -07003198retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003199 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003200 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003201
Christoph Lameter3c517a62006-12-06 20:33:29 -08003202retry:
3203 /*
3204 * Look through allowed nodes for objects available
3205 * from existing per node queues.
3206 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003207 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3208 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003209
Vladimir Davydov061d7072014-12-12 16:58:25 -08003210 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003211 get_node(cache, nid) &&
3212 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003213 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003214 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003215 if (obj)
3216 break;
3217 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003218 }
3219
Christoph Lametercfce6602007-05-06 14:50:17 -07003220 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003221 /*
3222 * This allocation will be performed within the constraints
3223 * of the current cpuset / memory policy requirements.
3224 * We may trigger various forms of reclaim on the allowed
3225 * set and go into memory reserves if necessary.
3226 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003227 page = cache_grow_begin(cache, flags, numa_mem_id());
3228 cache_grow_end(cache, page);
3229 if (page) {
3230 nid = page_to_nid(page);
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003231 obj = ____cache_alloc_node(cache,
3232 gfp_exact_node(flags), nid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003233
Christoph Lameter3c517a62006-12-06 20:33:29 -08003234 /*
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003235 * Another processor may allocate the objects in
3236 * the slab since we are not holding any locks.
Christoph Lameter3c517a62006-12-06 20:33:29 -08003237 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003238 if (!obj)
3239 goto retry;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003240 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003241 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003242
Mel Gormand26914d2014-04-03 14:47:24 -07003243 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003244 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003245 return obj;
3246}
3247
3248/*
Christoph Lametere498be72005-09-09 13:03:32 -07003249 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003251static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003252 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003253{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003254 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003255 struct kmem_cache_node *n;
Joonsoo Kim213b4692016-05-19 17:10:29 -07003256 void *obj = NULL;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003257 void *list = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003259 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003260 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003261 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003262
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003263 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003264 spin_lock(&n->list_lock);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003265 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003266 if (!page)
3267 goto must_grow;
Christoph Lametere498be72005-09-09 13:03:32 -07003268
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003269 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003270
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003271 STATS_INC_NODEALLOCS(cachep);
3272 STATS_INC_ACTIVE(cachep);
3273 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003274
Joonsoo Kim8456a642013-10-24 10:07:49 +09003275 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003276
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003277 obj = slab_get_obj(cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003278 n->free_objects--;
Christoph Lametere498be72005-09-09 13:03:32 -07003279
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003280 fixup_slab_list(cachep, n, page, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003281
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003282 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003283 fixup_objfreelist_debug(cachep, &list);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003284 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003285
Andrew Mortona737b3e2006-03-22 00:08:11 -08003286must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003287 spin_unlock(&n->list_lock);
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003288 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003289 if (page) {
3290 /* This slab isn't counted yet so don't update free_objects */
3291 obj = slab_get_obj(cachep, page);
3292 }
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003293 cache_grow_end(cachep, page);
Christoph Lametere498be72005-09-09 13:03:32 -07003294
Joonsoo Kim213b4692016-05-19 17:10:29 -07003295 return obj ? obj : fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003296}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003297
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003298static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003299slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003300 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003301{
3302 unsigned long save_flags;
3303 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003304 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003305
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003306 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003307 cachep = slab_pre_alloc_hook(cachep, flags);
3308 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003309 return NULL;
3310
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003311 cache_alloc_debugcheck_before(cachep, flags);
3312 local_irq_save(save_flags);
3313
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003314 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003315 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003316
Christoph Lameter18bf8542014-08-06 16:04:11 -07003317 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003318 /* Node not bootstrapped yet */
3319 ptr = fallback_alloc(cachep, flags);
3320 goto out;
3321 }
3322
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003323 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003324 /*
3325 * Use the locally cached objects if possible.
3326 * However ____cache_alloc does not allow fallback
3327 * to other nodes. It may fail while we still have
3328 * objects on other nodes available.
3329 */
3330 ptr = ____cache_alloc(cachep, flags);
3331 if (ptr)
3332 goto out;
3333 }
3334 /* ___cache_alloc_node can fall back to other nodes */
3335 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3336 out:
3337 local_irq_restore(save_flags);
3338 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3339
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003340 if (unlikely(flags & __GFP_ZERO) && ptr)
3341 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003342
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003343 slab_post_alloc_hook(cachep, flags, 1, &ptr);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003344 return ptr;
3345}
3346
3347static __always_inline void *
3348__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3349{
3350 void *objp;
3351
Zefan Li2ad654b2014-09-25 09:41:02 +08003352 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003353 objp = alternate_node_alloc(cache, flags);
3354 if (objp)
3355 goto out;
3356 }
3357 objp = ____cache_alloc(cache, flags);
3358
3359 /*
3360 * We may just have run out of memory on the local node.
3361 * ____cache_alloc_node() knows how to locate memory on other nodes
3362 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003363 if (!objp)
3364 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003365
3366 out:
3367 return objp;
3368}
3369#else
3370
3371static __always_inline void *
3372__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3373{
3374 return ____cache_alloc(cachep, flags);
3375}
3376
3377#endif /* CONFIG_NUMA */
3378
3379static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003380slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003381{
3382 unsigned long save_flags;
3383 void *objp;
3384
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003385 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003386 cachep = slab_pre_alloc_hook(cachep, flags);
3387 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003388 return NULL;
3389
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003390 cache_alloc_debugcheck_before(cachep, flags);
3391 local_irq_save(save_flags);
3392 objp = __do_cache_alloc(cachep, flags);
3393 local_irq_restore(save_flags);
3394 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3395 prefetchw(objp);
3396
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003397 if (unlikely(flags & __GFP_ZERO) && objp)
3398 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003399
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003400 slab_post_alloc_hook(cachep, flags, 1, &objp);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003401 return objp;
3402}
Christoph Lametere498be72005-09-09 13:03:32 -07003403
3404/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003405 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003406 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003407 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003408static void free_block(struct kmem_cache *cachep, void **objpp,
3409 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410{
3411 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003412 struct kmem_cache_node *n = get_node(cachep, node);
Joonsoo Kim6052b782016-05-19 17:10:17 -07003413 struct page *page;
3414
3415 n->free_objects += nr_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416
3417 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003418 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003419 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003420
Mel Gorman072bb0a2012-07-31 16:43:58 -07003421 objp = objpp[i];
3422
Joonsoo Kim8456a642013-10-24 10:07:49 +09003423 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003424 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003425 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003426 slab_put_obj(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 STATS_DEC_ACTIVE(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
3429 /* fixup slab chains */
Joonsoo Kim6052b782016-05-19 17:10:17 -07003430 if (page->active == 0)
3431 list_add(&page->lru, &n->slabs_free);
3432 else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 /* Unconditionally move a slab to the end of the
3434 * partial list on free - maximum time for the
3435 * other objects to be freed, too.
3436 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003437 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438 }
3439 }
Joonsoo Kim6052b782016-05-19 17:10:17 -07003440
3441 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3442 n->free_objects -= cachep->num;
3443
3444 page = list_last_entry(&n->slabs_free, struct page, lru);
Wei Yongjunde24bae2016-07-26 15:22:11 -07003445 list_move(&page->lru, list);
Joonsoo Kim6052b782016-05-19 17:10:17 -07003446 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447}
3448
Pekka Enberg343e0d72006-02-01 03:05:50 -08003449static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450{
3451 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003452 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003453 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003454 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455
3456 batchcount = ac->batchcount;
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003457
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003459 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003460 spin_lock(&n->list_lock);
3461 if (n->shared) {
3462 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003463 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 if (max) {
3465 if (batchcount > max)
3466 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003467 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003468 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 shared_array->avail += batchcount;
3470 goto free_done;
3471 }
3472 }
3473
Joonsoo Kim97654df2014-08-06 16:04:25 -07003474 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003475free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476#if STATS
3477 {
3478 int i = 0;
Geliang Tang73c02192016-01-14 15:17:59 -08003479 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480
Geliang Tang73c02192016-01-14 15:17:59 -08003481 list_for_each_entry(page, &n->slabs_free, lru) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003482 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483
3484 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 }
3486 STATS_SET_FREEABLE(cachep, i);
3487 }
3488#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003489 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003490 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003492 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493}
3494
3495/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003496 * Release an obj back to its cache. If the obj has a constructed state, it must
3497 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003499static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003500 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501{
Alexander Potapenko55834c52016-05-20 16:59:11 -07003502 /* Put the object into the quarantine, don't touch it for now. */
3503 if (kasan_slab_free(cachep, objp))
3504 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505
Alexander Potapenko55834c52016-05-20 16:59:11 -07003506 ___cache_free(cachep, objp, caller);
3507}
3508
3509void ___cache_free(struct kmem_cache *cachep, void *objp,
3510 unsigned long caller)
3511{
3512 struct array_cache *ac = cpu_cache_get(cachep);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003513
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003515 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003516 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003518 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003519
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003520 /*
3521 * Skip calling cache_free_alien() when the platform is not numa.
3522 * This will avoid cache misses that happen while accessing slabp (which
3523 * is per page memory reference) to get nodeid. Instead use a global
3524 * variable to skip the call, which is mostly likely to be present in
3525 * the cache.
3526 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003527 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003528 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003529
Joonsoo Kim3d880192014-10-09 15:26:04 -07003530 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 } else {
3533 STATS_INC_FREEMISS(cachep);
3534 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003536
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003537 if (sk_memalloc_socks()) {
3538 struct page *page = virt_to_head_page(objp);
3539
3540 if (unlikely(PageSlabPfmemalloc(page))) {
3541 cache_free_pfmemalloc(cachep, page, objp);
3542 return;
3543 }
3544 }
3545
3546 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547}
3548
3549/**
3550 * kmem_cache_alloc - Allocate an object
3551 * @cachep: The cache to allocate from.
3552 * @flags: See kmalloc().
3553 *
3554 * Allocate an object from this cache. The flags are only relevant
3555 * if the cache has no available objects.
3556 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003557void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003559 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003560
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003561 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003562 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003563 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003564
3565 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566}
3567EXPORT_SYMBOL(kmem_cache_alloc);
3568
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003569static __always_inline void
3570cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3571 size_t size, void **p, unsigned long caller)
3572{
3573 size_t i;
3574
3575 for (i = 0; i < size; i++)
3576 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3577}
3578
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08003579int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003580 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07003581{
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003582 size_t i;
3583
3584 s = slab_pre_alloc_hook(s, flags);
3585 if (!s)
3586 return 0;
3587
3588 cache_alloc_debugcheck_before(s, flags);
3589
3590 local_irq_disable();
3591 for (i = 0; i < size; i++) {
3592 void *objp = __do_cache_alloc(s, flags);
3593
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003594 if (unlikely(!objp))
3595 goto error;
3596 p[i] = objp;
3597 }
3598 local_irq_enable();
3599
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003600 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3601
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003602 /* Clear memory outside IRQ disabled section */
3603 if (unlikely(flags & __GFP_ZERO))
3604 for (i = 0; i < size; i++)
3605 memset(p[i], 0, s->object_size);
3606
3607 slab_post_alloc_hook(s, flags, size, p);
3608 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3609 return size;
3610error:
3611 local_irq_enable();
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003612 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003613 slab_post_alloc_hook(s, flags, i, p);
3614 __kmem_cache_free_bulk(s, i, p);
3615 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07003616}
3617EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3618
Li Zefan0f24f122009-12-11 15:45:30 +08003619#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003620void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003621kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003622{
Steven Rostedt85beb582010-11-24 16:23:34 -05003623 void *ret;
3624
Ezequiel Garcia48356302012-09-08 17:47:57 -03003625 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003626
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003627 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003628 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003629 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003630 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003631}
Steven Rostedt85beb582010-11-24 16:23:34 -05003632EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003633#endif
3634
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003636/**
3637 * kmem_cache_alloc_node - Allocate an object on the specified node
3638 * @cachep: The cache to allocate from.
3639 * @flags: See kmalloc().
3640 * @nodeid: node number of the target node.
3641 *
3642 * Identical to kmem_cache_alloc but it will allocate memory on the given
3643 * node, which can improve the performance for cpu bound structures.
3644 *
3645 * Fallback to other node is possible if __GFP_THISNODE is not set.
3646 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003647void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3648{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003649 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003650
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003651 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003652 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003653 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003654 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003655
3656 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003657}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003658EXPORT_SYMBOL(kmem_cache_alloc_node);
3659
Li Zefan0f24f122009-12-11 15:45:30 +08003660#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003661void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003662 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003663 int nodeid,
3664 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003665{
Steven Rostedt85beb582010-11-24 16:23:34 -05003666 void *ret;
3667
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003668 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003669
3670 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003671 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003672 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003673 flags, nodeid);
3674 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003675}
Steven Rostedt85beb582010-11-24 16:23:34 -05003676EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003677#endif
3678
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003679static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003680__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003681{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003682 struct kmem_cache *cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003683 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003684
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003685 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003686 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3687 return cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003688 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003689 kasan_kmalloc(cachep, ret, size, flags);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003690
3691 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003692}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003693
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003694void *__kmalloc_node(size_t size, gfp_t flags, int node)
3695{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003696 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003697}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003698EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003699
3700void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003701 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003702{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003703 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003704}
3705EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003706#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707
3708/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003709 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003711 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003712 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003714static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003715 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003717 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003718 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003720 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003721 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3722 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003723 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003724
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003725 kasan_kmalloc(cachep, ret, size, flags);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003726 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003727 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003728
3729 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003730}
3731
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003732void *__kmalloc(size_t size, gfp_t flags)
3733{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003734 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735}
3736EXPORT_SYMBOL(__kmalloc);
3737
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003738void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003739{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003740 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003741}
3742EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003743
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744/**
3745 * kmem_cache_free - Deallocate an object
3746 * @cachep: The cache the allocation was from.
3747 * @objp: The previously allocated object.
3748 *
3749 * Free an object which was previously allocated from this
3750 * cache.
3751 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003752void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753{
3754 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003755 cachep = cache_from_obj(cachep, objp);
3756 if (!cachep)
3757 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758
3759 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003760 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003761 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003762 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003763 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003765
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003766 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767}
3768EXPORT_SYMBOL(kmem_cache_free);
3769
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003770void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3771{
3772 struct kmem_cache *s;
3773 size_t i;
3774
3775 local_irq_disable();
3776 for (i = 0; i < size; i++) {
3777 void *objp = p[i];
3778
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07003779 if (!orig_s) /* called via kfree_bulk */
3780 s = virt_to_cache(objp);
3781 else
3782 s = cache_from_obj(orig_s, objp);
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003783
3784 debug_check_no_locks_freed(objp, s->object_size);
3785 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3786 debug_check_no_obj_freed(objp, s->object_size);
3787
3788 __cache_free(s, objp, _RET_IP_);
3789 }
3790 local_irq_enable();
3791
3792 /* FIXME: add tracing */
3793}
3794EXPORT_SYMBOL(kmem_cache_free_bulk);
3795
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 * kfree - free previously allocated memory
3798 * @objp: pointer returned by kmalloc.
3799 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003800 * If @objp is NULL, no operation is performed.
3801 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003802 * Don't free memory not originally allocated by kmalloc()
3803 * or you will run into trouble.
3804 */
3805void kfree(const void *objp)
3806{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003807 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 unsigned long flags;
3809
Pekka Enberg2121db72009-03-25 11:05:57 +02003810 trace_kfree(_RET_IP_, objp);
3811
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003812 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 return;
3814 local_irq_save(flags);
3815 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003816 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003817 debug_check_no_locks_freed(objp, c->object_size);
3818
3819 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003820 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 local_irq_restore(flags);
3822}
3823EXPORT_SYMBOL(kfree);
3824
Christoph Lametere498be72005-09-09 13:03:32 -07003825/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003826 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003827 */
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003828static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003829{
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003830 int ret;
Christoph Lametere498be72005-09-09 13:03:32 -07003831 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003832 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07003833
Mel Gorman9c09a952008-01-24 05:49:54 -08003834 for_each_online_node(node) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003835 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3836 if (ret)
Christoph Lametere498be72005-09-09 13:03:32 -07003837 goto fail;
3838
Christoph Lametere498be72005-09-09 13:03:32 -07003839 }
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003840
Christoph Lametercafeb022006-03-25 03:06:46 -08003841 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003842
Andrew Mortona737b3e2006-03-22 00:08:11 -08003843fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003844 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003845 /* Cache is not active yet. Roll back what we did */
3846 node--;
3847 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003848 n = get_node(cachep, node);
3849 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003850 kfree(n->shared);
3851 free_alien_cache(n->alien);
3852 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003853 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003854 }
3855 node--;
3856 }
3857 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003858 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003859}
3860
Christoph Lameter18004c52012-07-06 15:25:12 -05003861/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003862static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003863 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003865 struct array_cache __percpu *cpu_cache, *prev;
3866 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003868 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3869 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003870 return -ENOMEM;
3871
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003872 prev = cachep->cpu_cache;
3873 cachep->cpu_cache = cpu_cache;
3874 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003875
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 cachep->batchcount = batchcount;
3878 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003879 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003881 if (!prev)
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003882 goto setup_node;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003883
3884 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003885 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003886 int node;
3887 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003888 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003889
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003890 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003891 n = get_node(cachep, node);
3892 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003893 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003894 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003895 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003897 free_percpu(prev);
3898
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003899setup_node:
3900 return setup_kmem_cache_nodes(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901}
3902
Glauber Costa943a4512012-12-18 14:23:03 -08003903static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3904 int batchcount, int shared, gfp_t gfp)
3905{
3906 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003907 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003908
3909 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3910
3911 if (slab_state < FULL)
3912 return ret;
3913
3914 if ((ret < 0) || !is_root_cache(cachep))
3915 return ret;
3916
Vladimir Davydov426589f2015-02-12 14:59:23 -08003917 lockdep_assert_held(&slab_mutex);
3918 for_each_memcg_cache(c, cachep) {
3919 /* return value determined by the root cache only */
3920 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003921 }
3922
3923 return ret;
3924}
3925
Christoph Lameter18004c52012-07-06 15:25:12 -05003926/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003927static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928{
3929 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003930 int limit = 0;
3931 int shared = 0;
3932 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933
Thomas Garnier7c00fce2016-07-26 15:21:56 -07003934 err = cache_random_seq_create(cachep, cachep->num, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003935 if (err)
3936 goto end;
3937
Glauber Costa943a4512012-12-18 14:23:03 -08003938 if (!is_root_cache(cachep)) {
3939 struct kmem_cache *root = memcg_root_cache(cachep);
3940 limit = root->limit;
3941 shared = root->shared;
3942 batchcount = root->batchcount;
3943 }
3944
3945 if (limit && shared && batchcount)
3946 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003947 /*
3948 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 * - create a LIFO ordering, i.e. return objects that are cache-warm
3950 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003951 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 * bufctl chains: array operations are cheaper.
3953 * The numbers are guessed, we should auto-tune as described by
3954 * Bonwick.
3955 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003956 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003958 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003960 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003962 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963 limit = 54;
3964 else
3965 limit = 120;
3966
Andrew Mortona737b3e2006-03-22 00:08:11 -08003967 /*
3968 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 * allocation behaviour: Most allocs on one cpu, most free operations
3970 * on another cpu. For these cases, an efficient object passing between
3971 * cpus is necessary. This is provided by a shared array. The array
3972 * replaces Bonwick's magazine layer.
3973 * On uniprocessor, it's functionally equivalent (but less efficient)
3974 * to a larger limit. Thus disabled by default.
3975 */
3976 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003977 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979
3980#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003981 /*
3982 * With debugging enabled, large batchcount lead to excessively long
3983 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 */
3985 if (limit > 32)
3986 limit = 32;
3987#endif
Glauber Costa943a4512012-12-18 14:23:03 -08003988 batchcount = (limit + 1) / 2;
3989skip_setup:
3990 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003991end:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992 if (err)
Joe Perches11705322016-03-17 14:19:50 -07003993 pr_err("enable_cpucache failed for %s, error %d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003994 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003995 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996}
3997
Christoph Lameter1b552532006-03-22 00:09:07 -08003998/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003999 * Drain an array if it contains any elements taking the node lock only if
4000 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004001 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004002 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004003static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004004 struct array_cache *ac, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005{
Joonsoo Kim97654df2014-08-06 16:04:25 -07004006 LIST_HEAD(list);
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004007
4008 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4009 check_mutex_acquired();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010
Christoph Lameter1b552532006-03-22 00:09:07 -08004011 if (!ac || !ac->avail)
4012 return;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004013
4014 if (ac->touched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015 ac->touched = 0;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004016 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 }
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004018
4019 spin_lock_irq(&n->list_lock);
4020 drain_array_locked(cachep, ac, node, false, &list);
4021 spin_unlock_irq(&n->list_lock);
4022
4023 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024}
4025
4026/**
4027 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004028 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 *
4030 * Called from workqueue/eventd every few seconds.
4031 * Purpose:
4032 * - clear the per-cpu caches for this CPU.
4033 * - return freeable pages to the main free memory pool.
4034 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004035 * If we cannot acquire the cache chain mutex then just give up - we'll try
4036 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004038static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004040 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004041 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004042 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004043 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044
Christoph Lameter18004c52012-07-06 15:25:12 -05004045 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004047 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048
Christoph Lameter18004c52012-07-06 15:25:12 -05004049 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 check_irq_on();
4051
Christoph Lameter35386e32006-03-22 00:09:05 -08004052 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004053 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004054 * have established with reasonable certainty that
4055 * we can do some work if the lock was obtained.
4056 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07004057 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08004058
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004059 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004061 drain_array(searchp, n, cpu_cache_get(searchp), node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062
Christoph Lameter35386e32006-03-22 00:09:05 -08004063 /*
4064 * These are racy checks but it does not matter
4065 * if we skip one check or scan twice.
4066 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004067 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004068 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004070 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004072 drain_array(searchp, n, n->shared, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004074 if (n->free_touched)
4075 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004076 else {
4077 int freed;
4078
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004079 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004080 5 * searchp->num - 1) / (5 * searchp->num));
4081 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004083next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084 cond_resched();
4085 }
4086 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004087 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004088 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004089out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004090 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004091 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004092}
4093
Linus Torvalds158a9622008-01-02 13:04:48 -08004094#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004095void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096{
Joonsoo Kim8456a642013-10-24 10:07:49 +09004097 struct page *page;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004098 unsigned long active_objs;
4099 unsigned long num_objs;
4100 unsigned long active_slabs = 0;
4101 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004102 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004104 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004105 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 active_objs = 0;
4108 num_slabs = 0;
Christoph Lameter18bf8542014-08-06 16:04:11 -07004109 for_each_kmem_cache_node(cachep, node, n) {
Christoph Lametere498be72005-09-09 13:03:32 -07004110
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004111 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004112 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004113
Joonsoo Kim8456a642013-10-24 10:07:49 +09004114 list_for_each_entry(page, &n->slabs_full, lru) {
4115 if (page->active != cachep->num && !error)
Christoph Lametere498be72005-09-09 13:03:32 -07004116 error = "slabs_full accounting error";
4117 active_objs += cachep->num;
4118 active_slabs++;
4119 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004120 list_for_each_entry(page, &n->slabs_partial, lru) {
4121 if (page->active == cachep->num && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004122 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004123 if (!page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004124 error = "slabs_partial accounting error";
Joonsoo Kim8456a642013-10-24 10:07:49 +09004125 active_objs += page->active;
Christoph Lametere498be72005-09-09 13:03:32 -07004126 active_slabs++;
4127 }
Joonsoo Kim8456a642013-10-24 10:07:49 +09004128 list_for_each_entry(page, &n->slabs_free, lru) {
4129 if (page->active && !error)
Joonsoo Kim106a74e2013-10-24 10:07:48 +09004130 error = "slabs_free accounting error";
Christoph Lametere498be72005-09-09 13:03:32 -07004131 num_slabs++;
4132 }
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004133 free_objects += n->free_objects;
4134 if (n->shared)
4135 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004136
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004137 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004138 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004139 num_slabs += active_slabs;
4140 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004141 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 error = "free_objects accounting error";
4143
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004144 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145 if (error)
Joe Perches11705322016-03-17 14:19:50 -07004146 pr_err("slab: cache %s error: %s\n", name, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147
Glauber Costa0d7561c2012-10-19 18:20:27 +04004148 sinfo->active_objs = active_objs;
4149 sinfo->num_objs = num_objs;
4150 sinfo->active_slabs = active_slabs;
4151 sinfo->num_slabs = num_slabs;
4152 sinfo->shared_avail = shared_avail;
4153 sinfo->limit = cachep->limit;
4154 sinfo->batchcount = cachep->batchcount;
4155 sinfo->shared = cachep->shared;
4156 sinfo->objects_per_slab = cachep->num;
4157 sinfo->cache_order = cachep->gfporder;
4158}
4159
4160void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4161{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004163 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164 unsigned long high = cachep->high_mark;
4165 unsigned long allocs = cachep->num_allocations;
4166 unsigned long grown = cachep->grown;
4167 unsigned long reaped = cachep->reaped;
4168 unsigned long errors = cachep->errors;
4169 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004171 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004172 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173
Joe Perches756a025f02016-03-17 14:19:47 -07004174 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
Joe Perchese92dd4f2010-03-26 19:27:58 -07004175 allocs, high, grown,
4176 reaped, errors, max_freeable, node_allocs,
4177 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178 }
4179 /* cpu stats */
4180 {
4181 unsigned long allochit = atomic_read(&cachep->allochit);
4182 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4183 unsigned long freehit = atomic_read(&cachep->freehit);
4184 unsigned long freemiss = atomic_read(&cachep->freemiss);
4185
4186 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004187 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188 }
4189#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190}
4191
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192#define MAX_SLABINFO_WRITE 128
4193/**
4194 * slabinfo_write - Tuning for the slab allocator
4195 * @file: unused
4196 * @buffer: user buffer
4197 * @count: data length
4198 * @ppos: unused
4199 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004200ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004201 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004203 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004205 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004206
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207 if (count > MAX_SLABINFO_WRITE)
4208 return -EINVAL;
4209 if (copy_from_user(&kbuf, buffer, count))
4210 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004211 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212
4213 tmp = strchr(kbuf, ' ');
4214 if (!tmp)
4215 return -EINVAL;
4216 *tmp = '\0';
4217 tmp++;
4218 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4219 return -EINVAL;
4220
4221 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004222 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004224 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004225 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004226 if (limit < 1 || batchcount < 1 ||
4227 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004228 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004230 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004231 batchcount, shared,
4232 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233 }
4234 break;
4235 }
4236 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004237 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 if (res >= 0)
4239 res = count;
4240 return res;
4241}
Al Viro871751e2006-03-25 03:06:39 -08004242
4243#ifdef CONFIG_DEBUG_SLAB_LEAK
4244
Al Viro871751e2006-03-25 03:06:39 -08004245static inline int add_caller(unsigned long *n, unsigned long v)
4246{
4247 unsigned long *p;
4248 int l;
4249 if (!v)
4250 return 1;
4251 l = n[1];
4252 p = n + 2;
4253 while (l) {
4254 int i = l/2;
4255 unsigned long *q = p + 2 * i;
4256 if (*q == v) {
4257 q[1]++;
4258 return 1;
4259 }
4260 if (*q > v) {
4261 l = i;
4262 } else {
4263 p = q + 2;
4264 l -= i + 1;
4265 }
4266 }
4267 if (++n[1] == n[0])
4268 return 0;
4269 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4270 p[0] = v;
4271 p[1] = 1;
4272 return 1;
4273}
4274
Joonsoo Kim8456a642013-10-24 10:07:49 +09004275static void handle_slab(unsigned long *n, struct kmem_cache *c,
4276 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004277{
4278 void *p;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004279 int i, j;
4280 unsigned long v;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004281
Al Viro871751e2006-03-25 03:06:39 -08004282 if (n[0] == n[1])
4283 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004284 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004285 bool active = true;
4286
4287 for (j = page->active; j < c->num; j++) {
4288 if (get_free_obj(page, j) == i) {
4289 active = false;
4290 break;
4291 }
4292 }
4293
4294 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004295 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004296
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004297 /*
4298 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4299 * mapping is established when actual object allocation and
4300 * we could mistakenly access the unmapped object in the cpu
4301 * cache.
4302 */
4303 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4304 continue;
4305
4306 if (!add_caller(n, v))
Al Viro871751e2006-03-25 03:06:39 -08004307 return;
4308 }
4309}
4310
4311static void show_symbol(struct seq_file *m, unsigned long address)
4312{
4313#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004314 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004315 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004316
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004317 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004318 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004319 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004320 seq_printf(m, " [%s]", modname);
4321 return;
4322 }
4323#endif
4324 seq_printf(m, "%p", (void *)address);
4325}
4326
4327static int leaks_show(struct seq_file *m, void *p)
4328{
Thierry Reding0672aa72012-06-22 19:42:49 +02004329 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004330 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004331 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004332 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004333 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004334 int node;
4335 int i;
4336
4337 if (!(cachep->flags & SLAB_STORE_USER))
4338 return 0;
4339 if (!(cachep->flags & SLAB_RED_ZONE))
4340 return 0;
4341
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004342 /*
4343 * Set store_user_clean and start to grab stored user information
4344 * for all objects on this cache. If some alloc/free requests comes
4345 * during the processing, information would be wrong so restart
4346 * whole processing.
4347 */
4348 do {
4349 set_store_user_clean(cachep);
4350 drain_cpu_caches(cachep);
Al Viro871751e2006-03-25 03:06:39 -08004351
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004352 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004353
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004354 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004355
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004356 check_irq_on();
4357 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004358
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004359 list_for_each_entry(page, &n->slabs_full, lru)
4360 handle_slab(x, cachep, page);
4361 list_for_each_entry(page, &n->slabs_partial, lru)
4362 handle_slab(x, cachep, page);
4363 spin_unlock_irq(&n->list_lock);
4364 }
4365 } while (!is_store_user_clean(cachep));
4366
Al Viro871751e2006-03-25 03:06:39 -08004367 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004368 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004369 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004370 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004371 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004372 if (!m->private) {
4373 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004374 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004375 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004376 return -ENOMEM;
4377 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004378 *(unsigned long *)m->private = x[0] * 2;
4379 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004380 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004381 /* Now make sure this entry will be retried */
4382 m->count = m->size;
4383 return 0;
4384 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004385 for (i = 0; i < x[1]; i++) {
4386 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4387 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004388 seq_putc(m, '\n');
4389 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004390
Al Viro871751e2006-03-25 03:06:39 -08004391 return 0;
4392}
4393
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004394static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004395 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004396 .next = slab_next,
4397 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004398 .show = leaks_show,
4399};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004400
4401static int slabstats_open(struct inode *inode, struct file *file)
4402{
Rob Jonesb208ce32014-10-09 15:28:03 -07004403 unsigned long *n;
4404
4405 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4406 if (!n)
4407 return -ENOMEM;
4408
4409 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4410
4411 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004412}
4413
4414static const struct file_operations proc_slabstats_operations = {
4415 .open = slabstats_open,
4416 .read = seq_read,
4417 .llseek = seq_lseek,
4418 .release = seq_release_private,
4419};
Al Viro871751e2006-03-25 03:06:39 -08004420#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004421
4422static int __init slab_proc_init(void)
4423{
4424#ifdef CONFIG_DEBUG_SLAB_LEAK
4425 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4426#endif
4427 return 0;
4428}
4429module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004430#endif
4431
Kees Cook04385fc2016-06-23 15:20:59 -07004432#ifdef CONFIG_HARDENED_USERCOPY
4433/*
4434 * Rejects objects that are incorrectly sized.
4435 *
4436 * Returns NULL if check passes, otherwise const char * to name of cache
4437 * to indicate an error.
4438 */
4439const char *__check_heap_object(const void *ptr, unsigned long n,
4440 struct page *page)
4441{
4442 struct kmem_cache *cachep;
4443 unsigned int objnr;
4444 unsigned long offset;
4445
4446 /* Find and validate object. */
4447 cachep = page->slab_cache;
4448 objnr = obj_to_index(cachep, page, (void *)ptr);
4449 BUG_ON(objnr >= cachep->num);
4450
4451 /* Find offset within object. */
4452 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4453
4454 /* Allow address range falling entirely within object size. */
4455 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4456 return NULL;
4457
4458 return cachep->name;
4459}
4460#endif /* CONFIG_HARDENED_USERCOPY */
4461
Manfred Spraul00e145b2005-09-03 15:55:07 -07004462/**
4463 * ksize - get the actual amount of memory allocated for a given object
4464 * @objp: Pointer to the object
4465 *
4466 * kmalloc may internally round up allocations and return more memory
4467 * than requested. ksize() can be used to determine the actual amount of
4468 * memory allocated. The caller may use this additional memory, even though
4469 * a smaller amount of memory was initially specified with the kmalloc call.
4470 * The caller must guarantee that objp points to a valid object previously
4471 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4472 * must not be freed during the duration of the call.
4473 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004474size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004475{
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004476 size_t size;
4477
Christoph Lameteref8b4522007-10-16 01:24:46 -07004478 BUG_ON(!objp);
4479 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004480 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004482 size = virt_to_cache(objp)->object_size;
4483 /* We assume that ksize callers could use the whole allocated area,
4484 * so we need to unpoison this area.
4485 */
Alexander Potapenko4ebb31a42016-05-20 16:59:14 -07004486 kasan_unpoison_shadow(objp, size);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004487
4488 return size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004489}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004490EXPORT_SYMBOL(ksize);