blob: 2a31ee3c5814f192234385303a0d86cc77790580 [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>
Ingo Molnar3f8c2452017-02-05 14:31:22 +0100119#include <linux/sched/task_stack.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120
Mel Gorman381760e2012-07-31 16:44:30 -0700121#include <net/sock.h>
122
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#include <asm/cacheflush.h>
124#include <asm/tlbflush.h>
125#include <asm/page.h>
126
Steven Rostedt4dee6b62012-01-09 17:15:42 -0500127#include <trace/events/kmem.h>
128
Mel Gorman072bb0a2012-07-31 16:43:58 -0700129#include "internal.h"
130
Glauber Costab9ce5ef2012-12-18 14:22:46 -0800131#include "slab.h"
132
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700134 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 * 0 for faster, smaller code (especially in the critical paths).
136 *
137 * STATS - 1 to collect stats for /proc/slabinfo.
138 * 0 for faster, smaller code (especially in the critical paths).
139 *
140 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 */
142
143#ifdef CONFIG_DEBUG_SLAB
144#define DEBUG 1
145#define STATS 1
146#define FORCED_DEBUG 1
147#else
148#define DEBUG 0
149#define STATS 0
150#define FORCED_DEBUG 0
151#endif
152
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153/* Shouldn't this be in a header file somewhere? */
154#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400155#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#ifndef ARCH_KMALLOC_FLAGS
158#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
159#endif
160
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900161#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
162 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163
164#if FREELIST_BYTE_INDEX
165typedef unsigned char freelist_idx_t;
166#else
167typedef unsigned short freelist_idx_t;
168#endif
169
David Miller30321c72014-05-05 16:20:04 -0400170#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
Joonsoo Kimf315e3f2013-12-02 17:49:41 +0900171
Mel Gorman072bb0a2012-07-31 16:43:58 -0700172/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 * struct array_cache
174 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 * Purpose:
176 * - LIFO ordering, to hand out cache-warm objects from _alloc
177 * - reduce the number of linked list operations
178 * - reduce spinlock operations
179 *
180 * The limit is stored in the per-cpu structure to reduce the data cache
181 * footprint.
182 *
183 */
184struct array_cache {
185 unsigned int avail;
186 unsigned int limit;
187 unsigned int batchcount;
188 unsigned int touched;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700189 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800190 * Must have this definition in here for the proper
191 * alignment of array_cache. Also simplifies accessing
192 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800193 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194};
195
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700196struct alien_cache {
197 spinlock_t lock;
198 struct array_cache ac;
199};
200
Andrew Mortona737b3e2006-03-22 00:08:11 -0800201/*
Christoph Lametere498be72005-09-09 13:03:32 -0700202 * Need this for bootstrapping a per node allocator.
203 */
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700204#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000205static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
Christoph Lametere498be72005-09-09 13:03:32 -0700206#define CACHE_CACHE 0
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700207#define SIZE_NODE (MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208
Christoph Lametered11d9e2006-06-30 01:55:45 -0700209static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000210 struct kmem_cache_node *n, int tofree);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700211static void free_block(struct kmem_cache *cachep, void **objpp, int len,
Joonsoo Kim97654df2014-08-06 16:04:25 -0700212 int node, struct list_head *list);
213static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300214static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000215static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700216
Joonsoo Kim76b342b2016-05-19 17:10:26 -0700217static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
218 void **list);
219static inline void fixup_slab_list(struct kmem_cache *cachep,
220 struct kmem_cache_node *n, struct page *page,
221 void **list);
Ingo Molnare0a42722006-06-23 02:03:46 -0700222static int slab_early_init = 1;
223
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000224#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
Christoph Lametere498be72005-09-09 13:03:32 -0700225
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000226static void kmem_cache_node_init(struct kmem_cache_node *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700227{
228 INIT_LIST_HEAD(&parent->slabs_full);
229 INIT_LIST_HEAD(&parent->slabs_partial);
230 INIT_LIST_HEAD(&parent->slabs_free);
David Rientjesbf00bd32016-12-12 16:41:44 -0800231 parent->total_slabs = 0;
Greg Thelenf728b0a2016-12-12 16:41:41 -0800232 parent->free_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700233 parent->shared = NULL;
234 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800235 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700236 spin_lock_init(&parent->list_lock);
237 parent->free_objects = 0;
238 parent->free_touched = 0;
239}
240
Andrew Mortona737b3e2006-03-22 00:08:11 -0800241#define MAKE_LIST(cachep, listp, slab, nodeid) \
242 do { \
243 INIT_LIST_HEAD(listp); \
Christoph Lameter18bf8542014-08-06 16:04:11 -0700244 list_splice(&get_node(cachep, nodeid)->slab, listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700245 } while (0)
246
Andrew Mortona737b3e2006-03-22 00:08:11 -0800247#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
248 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700249 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
250 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
251 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
252 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700254#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255#define CFLGS_OFF_SLAB (0x80000000UL)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700256#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
258
259#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800260/*
261 * Optimization question: fewer reaps means less probability for unnessary
262 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100264 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 * which could lock up otherwise freeable slabs.
266 */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +0800267#define REAPTIMEOUT_AC (2*HZ)
268#define REAPTIMEOUT_NODE (4*HZ)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
270#if STATS
271#define STATS_INC_ACTIVE(x) ((x)->num_active++)
272#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
273#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
274#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700275#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276#define STATS_SET_HIGH(x) \
277 do { \
278 if ((x)->num_active > (x)->high_mark) \
279 (x)->high_mark = (x)->num_active; \
280 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281#define STATS_INC_ERR(x) ((x)->errors++)
282#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700283#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700284#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800285#define STATS_SET_FREEABLE(x, i) \
286 do { \
287 if ((x)->max_freeable < i) \
288 (x)->max_freeable = i; \
289 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
291#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
292#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
293#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
294#else
295#define STATS_INC_ACTIVE(x) do { } while (0)
296#define STATS_DEC_ACTIVE(x) do { } while (0)
297#define STATS_INC_ALLOCED(x) do { } while (0)
298#define STATS_INC_GROWN(x) do { } while (0)
Andi Kleen4e60c862010-08-09 17:19:03 -0700299#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300#define STATS_SET_HIGH(x) do { } while (0)
301#define STATS_INC_ERR(x) do { } while (0)
302#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700303#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700304#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800305#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306#define STATS_INC_ALLOCHIT(x) do { } while (0)
307#define STATS_INC_ALLOCMISS(x) do { } while (0)
308#define STATS_INC_FREEHIT(x) do { } while (0)
309#define STATS_INC_FREEMISS(x) do { } while (0)
310#endif
311
312#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Andrew Mortona737b3e2006-03-22 00:08:11 -0800314/*
315 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800317 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 * the end of an object is aligned with the end of the real
319 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800320 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800322 * cachep->obj_offset: The real object.
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500323 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
324 * cachep->size - 1* BYTES_PER_WORD: last caller address
Andrew Mortona737b3e2006-03-22 00:08:11 -0800325 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800327static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800329 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330}
331
David Woodhouseb46b8f12007-05-08 00:22:59 -0700332static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333{
334 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700335 return (unsigned long long*) (objp + obj_offset(cachep) -
336 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337}
338
David Woodhouseb46b8f12007-05-08 00:22:59 -0700339static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340{
341 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
342 if (cachep->flags & SLAB_STORE_USER)
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) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400345 REDZONE_ALIGN);
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500346 return (unsigned long long *) (objp + cachep->size -
David Woodhouseb46b8f12007-05-08 00:22:59 -0700347 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348}
349
Pekka Enberg343e0d72006-02-01 03:05:50 -0800350static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700351{
352 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500353 return (void **)(objp + cachep->size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354}
355
356#else
357
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800358#define obj_offset(x) 0
David Woodhouseb46b8f12007-05-08 00:22:59 -0700359#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
360#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
362
363#endif
364
Joonsoo Kim03787302014-06-23 13:22:06 -0700365#ifdef CONFIG_DEBUG_SLAB_LEAK
366
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700367static inline bool is_store_user_clean(struct kmem_cache *cachep)
Joonsoo Kim03787302014-06-23 13:22:06 -0700368{
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700369 return atomic_read(&cachep->store_user_clean) == 1;
370}
Joonsoo Kim03787302014-06-23 13:22:06 -0700371
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700372static inline void set_store_user_clean(struct kmem_cache *cachep)
373{
374 atomic_set(&cachep->store_user_clean, 1);
375}
Joonsoo Kim03787302014-06-23 13:22:06 -0700376
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700377static inline void set_store_user_dirty(struct kmem_cache *cachep)
378{
379 if (is_store_user_clean(cachep))
380 atomic_set(&cachep->store_user_clean, 0);
Joonsoo Kim03787302014-06-23 13:22:06 -0700381}
382
383#else
Joonsoo Kimd31676d2016-03-15 14:54:24 -0700384static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
Joonsoo Kim03787302014-06-23 13:22:06 -0700385
386#endif
387
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388/*
David Rientjes3df1ccc2011-10-18 22:09:28 -0700389 * Do not go above this order unless 0 objects fit into the slab or
390 * overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 */
David Rientjes543585c2011-10-18 22:09:24 -0700392#define SLAB_MAX_ORDER_HI 1
393#define SLAB_MAX_ORDER_LO 0
394static int slab_max_order = SLAB_MAX_ORDER_LO;
David Rientjes3df1ccc2011-10-18 22:09:28 -0700395static bool slab_max_order_set __initdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800397static inline struct kmem_cache *virt_to_cache(const void *obj)
398{
Christoph Lameterb49af682007-05-06 14:49:41 -0700399 struct page *page = virt_to_head_page(obj);
Christoph Lameter35026082012-06-13 10:24:56 -0500400 return page->slab_cache;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800401}
402
Joonsoo Kim8456a642013-10-24 10:07:49 +0900403static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800404 unsigned int idx)
405{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900406 return page->s_mem + cache->size * idx;
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800407}
408
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800409/*
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500410 * We want to avoid an expensive divide : (offset / cache->size)
411 * Using the fact that size is a constant for a particular cache,
412 * we can replace (offset / cache->size) by
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800413 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
414 */
415static inline unsigned int obj_to_index(const struct kmem_cache *cache,
Joonsoo Kim8456a642013-10-24 10:07:49 +0900416 const struct page *page, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800417{
Joonsoo Kim8456a642013-10-24 10:07:49 +0900418 u32 offset = (obj - page->s_mem);
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800419 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800420}
421
Joonsoo Kim6fb92432016-03-15 14:54:09 -0700422#define BOOT_CPUCACHE_ENTRIES 1
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423/* internal cache of cache description objs */
Christoph Lameter9b030cb2012-09-05 00:20:33 +0000424static struct kmem_cache kmem_cache_boot = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800425 .batchcount = 1,
426 .limit = BOOT_CPUCACHE_ENTRIES,
427 .shared = 1,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -0500428 .size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800429 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430};
431
Tejun Heo1871e522009-10-29 22:34:13 +0900432static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
Pekka Enberg343e0d72006-02-01 03:05:50 -0800434static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700436 return this_cpu_ptr(cachep->cpu_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437}
438
Andrew Mortona737b3e2006-03-22 00:08:11 -0800439/*
440 * Calculate the number of objects and left-over bytes for a given buffer size.
441 */
Joonsoo Kim70f75062016-03-15 14:54:53 -0700442static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
443 unsigned long flags, size_t *left_over)
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800444{
Joonsoo Kim70f75062016-03-15 14:54:53 -0700445 unsigned int num;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800446 size_t slab_size = PAGE_SIZE << gfporder;
447
448 /*
449 * The slab management structure can be either off the slab or
450 * on it. For the latter case, the memory allocated for a
451 * slab is used for:
452 *
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800453 * - @buffer_size bytes for each object
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700454 * - One freelist_idx_t for each object
455 *
456 * We don't need to consider alignment of freelist because
457 * freelist will be at the end of slab page. The objects will be
458 * at the correct alignment.
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800459 *
460 * If the slab management structure is off the slab, then the
461 * alignment will already be calculated into the size. Because
462 * the slabs are all pages aligned, the objects will be at the
463 * correct alignment when allocated.
464 */
Joonsoo Kimb03a017b2016-03-15 14:54:50 -0700465 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700466 num = slab_size / buffer_size;
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700467 *left_over = slab_size % buffer_size;
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800468 } else {
Joonsoo Kim70f75062016-03-15 14:54:53 -0700469 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
Joonsoo Kim2e6b3602016-03-15 14:54:30 -0700470 *left_over = slab_size %
471 (buffer_size + sizeof(freelist_idx_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 }
Joonsoo Kim70f75062016-03-15 14:54:53 -0700473
474 return num;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Christoph Lameterf28510d2012-09-11 19:49:38 +0000477#if DEBUG
Harvey Harrisond40cee22008-04-30 00:55:07 -0700478#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479
Andrew Mortona737b3e2006-03-22 00:08:11 -0800480static void __slab_error(const char *function, struct kmem_cache *cachep,
481 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482{
Joe Perches11705322016-03-17 14:19:50 -0700483 pr_err("slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800484 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030486 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487}
Christoph Lameterf28510d2012-09-11 19:49:38 +0000488#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
Paul Menage3395ee02006-12-06 20:32:16 -0800490/*
491 * By default on NUMA we use alien caches to stage the freeing of
492 * objects allocated from other nodes. This causes massive memory
493 * inefficiencies when using fake NUMA setup to split memory into a
494 * large number of small nodes, so it can be disabled on the command
495 * line
496 */
497
498static int use_alien_caches __read_mostly = 1;
499static int __init noaliencache_setup(char *s)
500{
501 use_alien_caches = 0;
502 return 1;
503}
504__setup("noaliencache", noaliencache_setup);
505
David Rientjes3df1ccc2011-10-18 22:09:28 -0700506static int __init slab_max_order_setup(char *str)
507{
508 get_option(&str, &slab_max_order);
509 slab_max_order = slab_max_order < 0 ? 0 :
510 min(slab_max_order, MAX_ORDER - 1);
511 slab_max_order_set = true;
512
513 return 1;
514}
515__setup("slab_max_order=", slab_max_order_setup);
516
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800517#ifdef CONFIG_NUMA
518/*
519 * Special reaping functions for NUMA systems called from cache_reap().
520 * These take care of doing round robin flushing of alien caches (containing
521 * objects freed on different nodes from which they were allocated) and the
522 * flushing of remote pcps by calling drain_node_pages.
523 */
Tejun Heo1871e522009-10-29 22:34:13 +0900524static DEFINE_PER_CPU(unsigned long, slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800525
526static void init_reap_node(int cpu)
527{
Andrew Morton0edaf862016-05-19 17:10:58 -0700528 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
529 node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800530}
531
532static void next_reap_node(void)
533{
Christoph Lameter909ea962010-12-08 16:22:55 +0100534 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800535
Andrew Morton0edaf862016-05-19 17:10:58 -0700536 node = next_node_in(node, node_online_map);
Christoph Lameter909ea962010-12-08 16:22:55 +0100537 __this_cpu_write(slab_reap_node, node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800538}
539
540#else
541#define init_reap_node(cpu) do { } while (0)
542#define next_reap_node(void) do { } while (0)
543#endif
544
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545/*
546 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
547 * via the workqueue/eventd.
548 * Add the CPU number into the expiration time to minimize the possibility of
549 * the CPUs getting into lockstep and contending for the global cache chain
550 * lock.
551 */
Paul Gortmaker0db06282013-06-19 14:53:51 -0400552static void start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553{
Tejun Heo1871e522009-10-29 22:34:13 +0900554 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
Tejun Heoeac03372016-09-16 15:49:34 -0400556 if (reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800557 init_reap_node(cpu);
Tejun Heo203b42f2012-08-21 13:18:23 -0700558 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800559 schedule_delayed_work_on(cpu, reap_work,
560 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 }
562}
563
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700564static void init_arraycache(struct array_cache *ac, int limit, int batch)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565{
Catalin Marinasd5cff632009-06-11 13:22:40 +0100566 /*
567 * The array_cache structures contain pointers to free object.
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300568 * However, when such objects are allocated or transferred to another
Catalin Marinasd5cff632009-06-11 13:22:40 +0100569 * cache the pointers are not cleared and they could be counted as
570 * valid references during a kmemleak scan. Therefore, kmemleak must
571 * not scan such objects.
572 */
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700573 kmemleak_no_scan(ac);
574 if (ac) {
575 ac->avail = 0;
576 ac->limit = limit;
577 ac->batchcount = batch;
578 ac->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 }
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700580}
581
582static struct array_cache *alloc_arraycache(int node, int entries,
583 int batchcount, gfp_t gfp)
584{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700585 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim1fe00d52014-08-06 16:04:27 -0700586 struct array_cache *ac = NULL;
587
588 ac = kmalloc_node(memsize, gfp, node);
589 init_arraycache(ac, entries, batchcount);
590 return ac;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591}
592
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700593static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
594 struct page *page, void *objp)
Mel Gorman072bb0a2012-07-31 16:43:58 -0700595{
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700596 struct kmem_cache_node *n;
597 int page_node;
598 LIST_HEAD(list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700599
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700600 page_node = page_to_nid(page);
601 n = get_node(cachep, page_node);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700602
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700603 spin_lock(&n->list_lock);
604 free_block(cachep, &objp, 1, page_node, &list);
605 spin_unlock(&n->list_lock);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700606
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700607 slabs_destroy(cachep, &list);
Mel Gorman072bb0a2012-07-31 16:43:58 -0700608}
609
Christoph Lameter3ded1752006-03-25 03:06:44 -0800610/*
611 * Transfer objects in one arraycache to another.
612 * Locking must be handled by the caller.
613 *
614 * Return the number of entries transferred.
615 */
616static int transfer_objects(struct array_cache *to,
617 struct array_cache *from, unsigned int max)
618{
619 /* Figure out how many entries to transfer */
Hagen Paul Pfeifer732eacc2010-10-26 14:22:23 -0700620 int nr = min3(from->avail, max, to->limit - to->avail);
Christoph Lameter3ded1752006-03-25 03:06:44 -0800621
622 if (!nr)
623 return 0;
624
625 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
626 sizeof(void *) *nr);
627
628 from->avail -= nr;
629 to->avail += nr;
Christoph Lameter3ded1752006-03-25 03:06:44 -0800630 return nr;
631}
632
Christoph Lameter765c4502006-09-27 01:50:08 -0700633#ifndef CONFIG_NUMA
634
635#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000636#define reap_alien(cachep, n) do { } while (0)
Christoph Lameter765c4502006-09-27 01:50:08 -0700637
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700638static inline struct alien_cache **alloc_alien_cache(int node,
639 int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700640{
Joonsoo Kim88881772016-05-19 17:10:05 -0700641 return NULL;
Christoph Lameter765c4502006-09-27 01:50:08 -0700642}
643
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700644static inline void free_alien_cache(struct alien_cache **ac_ptr)
Christoph Lameter765c4502006-09-27 01:50:08 -0700645{
646}
647
648static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
649{
650 return 0;
651}
652
653static inline void *alternate_node_alloc(struct kmem_cache *cachep,
654 gfp_t flags)
655{
656 return NULL;
657}
658
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800659static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700660 gfp_t flags, int nodeid)
661{
662 return NULL;
663}
664
David Rientjes4167e9b2015-04-14 15:46:55 -0700665static inline gfp_t gfp_exact_node(gfp_t flags)
666{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700667 return flags & ~__GFP_NOFAIL;
David Rientjes4167e9b2015-04-14 15:46:55 -0700668}
669
Christoph Lameter765c4502006-09-27 01:50:08 -0700670#else /* CONFIG_NUMA */
671
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800672static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800673static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800674
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700675static struct alien_cache *__alloc_alien_cache(int node, int entries,
676 int batch, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700677{
Joonsoo Kim5e804782014-08-06 16:04:40 -0700678 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700679 struct alien_cache *alc = NULL;
680
681 alc = kmalloc_node(memsize, gfp, node);
682 init_arraycache(&alc->ac, entries, batch);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700683 spin_lock_init(&alc->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700684 return alc;
685}
686
687static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
688{
689 struct alien_cache **alc_ptr;
Joonsoo Kim5e804782014-08-06 16:04:40 -0700690 size_t memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700691 int i;
692
693 if (limit > 1)
694 limit = 12;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700695 alc_ptr = kzalloc_node(memsize, gfp, node);
696 if (!alc_ptr)
697 return NULL;
698
699 for_each_node(i) {
700 if (i == node || !node_online(i))
701 continue;
702 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
703 if (!alc_ptr[i]) {
704 for (i--; i >= 0; i--)
705 kfree(alc_ptr[i]);
706 kfree(alc_ptr);
707 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700708 }
709 }
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700710 return alc_ptr;
Christoph Lametere498be72005-09-09 13:03:32 -0700711}
712
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700713static void free_alien_cache(struct alien_cache **alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700714{
715 int i;
716
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700717 if (!alc_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700718 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700719 for_each_node(i)
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700720 kfree(alc_ptr[i]);
721 kfree(alc_ptr);
Christoph Lametere498be72005-09-09 13:03:32 -0700722}
723
Pekka Enberg343e0d72006-02-01 03:05:50 -0800724static void __drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kim833b7062014-08-06 16:04:33 -0700725 struct array_cache *ac, int node,
726 struct list_head *list)
Christoph Lametere498be72005-09-09 13:03:32 -0700727{
Christoph Lameter18bf8542014-08-06 16:04:11 -0700728 struct kmem_cache_node *n = get_node(cachep, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700729
730 if (ac->avail) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000731 spin_lock(&n->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800732 /*
733 * Stuff objects into the remote nodes shared array first.
734 * That way we could avoid the overhead of putting the objects
735 * into the free lists and getting them back later.
736 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000737 if (n->shared)
738 transfer_objects(n->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800739
Joonsoo Kim833b7062014-08-06 16:04:33 -0700740 free_block(cachep, ac->entry, ac->avail, node, list);
Christoph Lametere498be72005-09-09 13:03:32 -0700741 ac->avail = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000742 spin_unlock(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -0700743 }
744}
745
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800746/*
747 * Called from cache_reap() to regularly drain alien caches round robin.
748 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000749static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800750{
Christoph Lameter909ea962010-12-08 16:22:55 +0100751 int node = __this_cpu_read(slab_reap_node);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800752
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000753 if (n->alien) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700754 struct alien_cache *alc = n->alien[node];
755 struct array_cache *ac;
Christoph Lametere00946f2006-03-25 03:06:45 -0800756
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700757 if (alc) {
758 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700759 if (ac->avail && spin_trylock_irq(&alc->lock)) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700760 LIST_HEAD(list);
761
762 __drain_alien_cache(cachep, ac, node, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700763 spin_unlock_irq(&alc->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700764 slabs_destroy(cachep, &list);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700765 }
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800766 }
767 }
768}
769
Andrew Mortona737b3e2006-03-22 00:08:11 -0800770static void drain_alien_cache(struct kmem_cache *cachep,
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700771 struct alien_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700772{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800773 int i = 0;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700774 struct alien_cache *alc;
Christoph Lametere498be72005-09-09 13:03:32 -0700775 struct array_cache *ac;
776 unsigned long flags;
777
778 for_each_online_node(i) {
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700779 alc = alien[i];
780 if (alc) {
Joonsoo Kim833b7062014-08-06 16:04:33 -0700781 LIST_HEAD(list);
782
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700783 ac = &alc->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700784 spin_lock_irqsave(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700785 __drain_alien_cache(cachep, ac, i, &list);
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700786 spin_unlock_irqrestore(&alc->lock, flags);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700787 slabs_destroy(cachep, &list);
Christoph Lametere498be72005-09-09 13:03:32 -0700788 }
789 }
790}
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700791
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700792static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
793 int node, int page_node)
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700794{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000795 struct kmem_cache_node *n;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700796 struct alien_cache *alien = NULL;
797 struct array_cache *ac;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700798 LIST_HEAD(list);
Pekka Enberg1ca4cb22006-10-06 00:43:52 -0700799
Christoph Lameter18bf8542014-08-06 16:04:11 -0700800 n = get_node(cachep, node);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700801 STATS_INC_NODEFREES(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700802 if (n->alien && n->alien[page_node]) {
803 alien = n->alien[page_node];
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700804 ac = &alien->ac;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700805 spin_lock(&alien->lock);
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700806 if (unlikely(ac->avail == ac->limit)) {
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700807 STATS_INC_ACOVERFLOW(cachep);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700808 __drain_alien_cache(cachep, ac, page_node, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700809 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -0700810 ac->entry[ac->avail++] = objp;
Joonsoo Kim49dfc302014-08-06 16:04:31 -0700811 spin_unlock(&alien->lock);
Joonsoo Kim833b7062014-08-06 16:04:33 -0700812 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700813 } else {
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700814 n = get_node(cachep, page_node);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700815 spin_lock(&n->list_lock);
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700816 free_block(cachep, &objp, 1, page_node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -0700817 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -0700818 slabs_destroy(cachep, &list);
Pekka Enberg729bd0b2006-06-23 02:03:05 -0700819 }
820 return 1;
821}
Joonsoo Kim25c4f302014-10-09 15:26:09 -0700822
823static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
824{
825 int page_node = page_to_nid(virt_to_page(objp));
826 int node = numa_mem_id();
827 /*
828 * Make sure we are not freeing a object from another node to the array
829 * cache on this cpu.
830 */
831 if (likely(node == page_node))
832 return 0;
833
834 return __cache_free_alien(cachep, objp, node, page_node);
835}
David Rientjes4167e9b2015-04-14 15:46:55 -0700836
837/*
Mel Gorman444eb2a42016-03-17 14:19:23 -0700838 * Construct gfp mask to allocate from a specific node but do not reclaim or
839 * warn about failures.
David Rientjes4167e9b2015-04-14 15:46:55 -0700840 */
841static inline gfp_t gfp_exact_node(gfp_t flags)
842{
Mel Gorman444eb2a42016-03-17 14:19:23 -0700843 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
David Rientjes4167e9b2015-04-14 15:46:55 -0700844}
Christoph Lametere498be72005-09-09 13:03:32 -0700845#endif
846
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700847static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
848{
849 struct kmem_cache_node *n;
850
851 /*
852 * Set up the kmem_cache_node for cpu before we can
853 * begin anything. Make sure some other cpu on this
854 * node has not already allocated this
855 */
856 n = get_node(cachep, node);
857 if (n) {
858 spin_lock_irq(&n->list_lock);
859 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
860 cachep->num;
861 spin_unlock_irq(&n->list_lock);
862
863 return 0;
864 }
865
866 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
867 if (!n)
868 return -ENOMEM;
869
870 kmem_cache_node_init(n);
871 n->next_reap = jiffies + REAPTIMEOUT_NODE +
872 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
873
874 n->free_limit =
875 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
876
877 /*
878 * The kmem_cache_nodes don't come and go as CPUs
879 * come and go. slab_mutex is sufficient
880 * protection here.
881 */
882 cachep->node[node] = n;
883
884 return 0;
885}
886
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200887#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700888/*
Christoph Lameter6a673682013-01-10 19:14:19 +0000889 * Allocates and initializes node for a node on each slab cache, used for
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000890 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
David Rientjes8f9f8d92010-03-27 19:40:47 -0700891 * will be allocated off-node since memory is not yet online for the new node.
Christoph Lameter6a673682013-01-10 19:14:19 +0000892 * When hotplugging memory or a cpu, existing node are not replaced if
David Rientjes8f9f8d92010-03-27 19:40:47 -0700893 * already in use.
894 *
Christoph Lameter18004c52012-07-06 15:25:12 -0500895 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -0700896 */
Christoph Lameter6a673682013-01-10 19:14:19 +0000897static int init_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -0700898{
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700899 int ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700900 struct kmem_cache *cachep;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700901
Christoph Lameter18004c52012-07-06 15:25:12 -0500902 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700903 ret = init_cache_node(cachep, node, GFP_KERNEL);
904 if (ret)
905 return ret;
David Rientjes8f9f8d92010-03-27 19:40:47 -0700906 }
Joonsoo Kimded0ecf2016-05-19 17:10:11 -0700907
David Rientjes8f9f8d92010-03-27 19:40:47 -0700908 return 0;
909}
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200910#endif
David Rientjes8f9f8d92010-03-27 19:40:47 -0700911
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700912static int setup_kmem_cache_node(struct kmem_cache *cachep,
913 int node, gfp_t gfp, bool force_change)
914{
915 int ret = -ENOMEM;
916 struct kmem_cache_node *n;
917 struct array_cache *old_shared = NULL;
918 struct array_cache *new_shared = NULL;
919 struct alien_cache **new_alien = NULL;
920 LIST_HEAD(list);
921
922 if (use_alien_caches) {
923 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
924 if (!new_alien)
925 goto fail;
926 }
927
928 if (cachep->shared) {
929 new_shared = alloc_arraycache(node,
930 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
931 if (!new_shared)
932 goto fail;
933 }
934
935 ret = init_cache_node(cachep, node, gfp);
936 if (ret)
937 goto fail;
938
939 n = get_node(cachep, node);
940 spin_lock_irq(&n->list_lock);
941 if (n->shared && force_change) {
942 free_block(cachep, n->shared->entry,
943 n->shared->avail, node, &list);
944 n->shared->avail = 0;
945 }
946
947 if (!n->shared || force_change) {
948 old_shared = n->shared;
949 n->shared = new_shared;
950 new_shared = NULL;
951 }
952
953 if (!n->alien) {
954 n->alien = new_alien;
955 new_alien = NULL;
956 }
957
958 spin_unlock_irq(&n->list_lock);
959 slabs_destroy(cachep, &list);
960
Joonsoo Kim801faf02016-05-19 17:10:31 -0700961 /*
962 * To protect lockless access to n->shared during irq disabled context.
963 * If n->shared isn't NULL in irq disabled context, accessing to it is
964 * guaranteed to be valid until irq is re-enabled, because it will be
965 * freed after synchronize_sched().
966 */
Joonsoo Kim86d9f482016-10-27 17:46:18 -0700967 if (old_shared && force_change)
Joonsoo Kim801faf02016-05-19 17:10:31 -0700968 synchronize_sched();
969
Joonsoo Kimc3d332b2016-05-19 17:10:14 -0700970fail:
971 kfree(old_shared);
972 kfree(new_shared);
973 free_alien_cache(new_alien);
974
975 return ret;
976}
977
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +0200978#ifdef CONFIG_SMP
979
Paul Gortmaker0db06282013-06-19 14:53:51 -0400980static void cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981{
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700982 struct kmem_cache *cachep;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000983 struct kmem_cache_node *n = NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -0700984 int node = cpu_to_mem(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +1030985 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700986
Christoph Lameter18004c52012-07-06 15:25:12 -0500987 list_for_each_entry(cachep, &slab_caches, list) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700988 struct array_cache *nc;
989 struct array_cache *shared;
Joonsoo Kimc8522a32014-08-06 16:04:29 -0700990 struct alien_cache **alien;
Joonsoo Kim97654df2014-08-06 16:04:25 -0700991 LIST_HEAD(list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700992
Christoph Lameter18bf8542014-08-06 16:04:11 -0700993 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000994 if (!n)
Joonsoo Kimbf0dea22014-10-09 15:26:27 -0700995 continue;
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700996
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000997 spin_lock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -0700998
Christoph Lameterce8eb6c2013-01-10 19:14:19 +0000999 /* Free limit for this kmem_cache_node */
1000 n->free_limit -= cachep->batchcount;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001001
1002 /* cpu is dead; no one can alloc from it. */
1003 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1004 if (nc) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07001005 free_block(cachep, nc->entry, nc->avail, node, &list);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001006 nc->avail = 0;
1007 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001008
Rusty Russell58463c12009-12-17 11:43:12 -06001009 if (!cpumask_empty(mask)) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001010 spin_unlock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001011 goto free_slab;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001012 }
1013
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001014 shared = n->shared;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001015 if (shared) {
1016 free_block(cachep, shared->entry,
Joonsoo Kim97654df2014-08-06 16:04:25 -07001017 shared->avail, node, &list);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001018 n->shared = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001019 }
1020
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001021 alien = n->alien;
1022 n->alien = NULL;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001023
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001024 spin_unlock_irq(&n->list_lock);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001025
1026 kfree(shared);
1027 if (alien) {
1028 drain_alien_cache(cachep, alien);
1029 free_alien_cache(alien);
1030 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001031
1032free_slab:
Joonsoo Kim97654df2014-08-06 16:04:25 -07001033 slabs_destroy(cachep, &list);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001034 }
1035 /*
1036 * In the previous loop, all the objects were freed to
1037 * the respective cache's slabs, now we can go ahead and
1038 * shrink each nodelist to its limit.
1039 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001040 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07001041 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001042 if (!n)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001043 continue;
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001044 drain_freelist(cachep, n, INT_MAX);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001045 }
1046}
1047
Paul Gortmaker0db06282013-06-19 14:53:51 -04001048static int cpuup_prepare(long cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001049{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001050 struct kmem_cache *cachep;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07001051 int node = cpu_to_mem(cpu);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001052 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001054 /*
1055 * We need to do this right in the beginning since
1056 * alloc_arraycache's are going to use this list.
1057 * kmalloc_node allows us to add the slab to the right
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001058 * kmem_cache_node and not this cpu's kmem_cache_node
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001059 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001060 err = init_cache_node_node(node);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001061 if (err < 0)
1062 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001063
1064 /*
1065 * Now we can go ahead with allocating the shared arrays and
1066 * array caches
1067 */
Christoph Lameter18004c52012-07-06 15:25:12 -05001068 list_for_each_entry(cachep, &slab_caches, list) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07001069 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1070 if (err)
1071 goto bad;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001072 }
Pekka Enbergce79ddc2009-11-23 22:01:15 +02001073
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001074 return 0;
1075bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001076 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001077 return -ENOMEM;
1078}
1079
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001080int slab_prepare_cpu(unsigned int cpu)
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001081{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001082 int err;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001083
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001084 mutex_lock(&slab_mutex);
1085 err = cpuup_prepare(cpu);
1086 mutex_unlock(&slab_mutex);
1087 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088}
1089
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001090/*
1091 * This is called for a failed online attempt and for a successful
1092 * offline.
1093 *
1094 * Even if all the cpus of a node are down, we don't free the
1095 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1096 * a kmalloc allocation from another cpu for memory from the node of
1097 * the cpu going down. The list3 structure is usually allocated from
1098 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1099 */
1100int slab_dead_cpu(unsigned int cpu)
1101{
1102 mutex_lock(&slab_mutex);
1103 cpuup_canceled(cpu);
1104 mutex_unlock(&slab_mutex);
1105 return 0;
1106}
1107#endif
1108
1109static int slab_online_cpu(unsigned int cpu)
1110{
1111 start_cpu_timer(cpu);
1112 return 0;
1113}
1114
1115static int slab_offline_cpu(unsigned int cpu)
1116{
1117 /*
1118 * Shutdown cache reaper. Note that the slab_mutex is held so
1119 * that if cache_reap() is invoked it cannot do anything
1120 * expensive but will only modify reap_work and reschedule the
1121 * timer.
1122 */
1123 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1124 /* Now the cache_reaper is guaranteed to be not running. */
1125 per_cpu(slab_reap_work, cpu).work.func = NULL;
1126 return 0;
1127}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
David Rientjes8f9f8d92010-03-27 19:40:47 -07001129#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1130/*
1131 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1132 * Returns -EBUSY if all objects cannot be drained so that the node is not
1133 * removed.
1134 *
Christoph Lameter18004c52012-07-06 15:25:12 -05001135 * Must hold slab_mutex.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001136 */
Christoph Lameter6a673682013-01-10 19:14:19 +00001137static int __meminit drain_cache_node_node(int node)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001138{
1139 struct kmem_cache *cachep;
1140 int ret = 0;
1141
Christoph Lameter18004c52012-07-06 15:25:12 -05001142 list_for_each_entry(cachep, &slab_caches, list) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001143 struct kmem_cache_node *n;
David Rientjes8f9f8d92010-03-27 19:40:47 -07001144
Christoph Lameter18bf8542014-08-06 16:04:11 -07001145 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001146 if (!n)
David Rientjes8f9f8d92010-03-27 19:40:47 -07001147 continue;
1148
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07001149 drain_freelist(cachep, n, INT_MAX);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001150
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001151 if (!list_empty(&n->slabs_full) ||
1152 !list_empty(&n->slabs_partial)) {
David Rientjes8f9f8d92010-03-27 19:40:47 -07001153 ret = -EBUSY;
1154 break;
1155 }
1156 }
1157 return ret;
1158}
1159
1160static int __meminit slab_memory_callback(struct notifier_block *self,
1161 unsigned long action, void *arg)
1162{
1163 struct memory_notify *mnb = arg;
1164 int ret = 0;
1165 int nid;
1166
1167 nid = mnb->status_change_nid;
1168 if (nid < 0)
1169 goto out;
1170
1171 switch (action) {
1172 case MEM_GOING_ONLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001173 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001174 ret = init_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001175 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001176 break;
1177 case MEM_GOING_OFFLINE:
Christoph Lameter18004c52012-07-06 15:25:12 -05001178 mutex_lock(&slab_mutex);
Christoph Lameter6a673682013-01-10 19:14:19 +00001179 ret = drain_cache_node_node(nid);
Christoph Lameter18004c52012-07-06 15:25:12 -05001180 mutex_unlock(&slab_mutex);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001181 break;
1182 case MEM_ONLINE:
1183 case MEM_OFFLINE:
1184 case MEM_CANCEL_ONLINE:
1185 case MEM_CANCEL_OFFLINE:
1186 break;
1187 }
1188out:
Prarit Bhargava5fda1bd2011-03-22 16:30:49 -07001189 return notifier_from_errno(ret);
David Rientjes8f9f8d92010-03-27 19:40:47 -07001190}
1191#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1192
Christoph Lametere498be72005-09-09 13:03:32 -07001193/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001194 * swap the static kmem_cache_node with kmalloced memory
Christoph Lametere498be72005-09-09 13:03:32 -07001195 */
Christoph Lameter6744f082013-01-10 19:12:17 +00001196static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
David Rientjes8f9f8d92010-03-27 19:40:47 -07001197 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001198{
Christoph Lameter6744f082013-01-10 19:12:17 +00001199 struct kmem_cache_node *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001200
Christoph Lameter6744f082013-01-10 19:12:17 +00001201 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001202 BUG_ON(!ptr);
1203
Christoph Lameter6744f082013-01-10 19:12:17 +00001204 memcpy(ptr, list, sizeof(struct kmem_cache_node));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001205 /*
1206 * Do not assume that spinlocks can be initialized via memcpy:
1207 */
1208 spin_lock_init(&ptr->list_lock);
1209
Christoph Lametere498be72005-09-09 13:03:32 -07001210 MAKE_ALL_LISTS(cachep, ptr, nodeid);
Christoph Lameter6a673682013-01-10 19:14:19 +00001211 cachep->node[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001212}
1213
Andrew Mortona737b3e2006-03-22 00:08:11 -08001214/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001215 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1216 * size of kmem_cache_node.
Pekka Enberg556a1692008-01-25 08:20:51 +02001217 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001218static void __init set_up_node(struct kmem_cache *cachep, int index)
Pekka Enberg556a1692008-01-25 08:20:51 +02001219{
1220 int node;
1221
1222 for_each_online_node(node) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001223 cachep->node[node] = &init_kmem_cache_node[index + node];
Christoph Lameter6a673682013-01-10 19:14:19 +00001224 cachep->node[node]->next_reap = jiffies +
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001225 REAPTIMEOUT_NODE +
1226 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enberg556a1692008-01-25 08:20:51 +02001227 }
1228}
1229
1230/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001231 * Initialisation. Called after the page allocator have been initialised and
1232 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 */
1234void __init kmem_cache_init(void)
1235{
Christoph Lametere498be72005-09-09 13:03:32 -07001236 int i;
1237
Joonsoo Kim68126702013-10-24 10:07:42 +09001238 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1239 sizeof(struct rcu_head));
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001240 kmem_cache = &kmem_cache_boot;
1241
Joonsoo Kim88881772016-05-19 17:10:05 -07001242 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001243 use_alien_caches = 0;
1244
Christoph Lameter3c583462012-11-28 16:23:01 +00001245 for (i = 0; i < NUM_INIT_LISTS; i++)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001246 kmem_cache_node_init(&init_kmem_cache_node[i]);
Christoph Lameter3c583462012-11-28 16:23:01 +00001247
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 /*
1249 * Fragmentation resistance on low memory - only use bigger
David Rientjes3df1ccc2011-10-18 22:09:28 -07001250 * page orders on machines with more than 32MB of memory if
1251 * not overridden on the command line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 */
David Rientjes3df1ccc2011-10-18 22:09:28 -07001253 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
David Rientjes543585c2011-10-18 22:09:24 -07001254 slab_max_order = SLAB_MAX_ORDER_HI;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 /* Bootstrap is tricky, because several objects are allocated
1257 * from caches that do not exist yet:
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001258 * 1) initialize the kmem_cache cache: it contains the struct
1259 * kmem_cache structures of all caches, except kmem_cache itself:
1260 * kmem_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001261 * Initially an __init data area is used for the head array and the
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001262 * kmem_cache_node structures, it's replaced with a kmalloc allocated
Christoph Lametere498be72005-09-09 13:03:32 -07001263 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001265 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001266 * An __init data area is used for the head array.
1267 * 3) Create the remaining kmalloc caches, with minimally sized
1268 * head arrays.
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001269 * 4) Replace the __init data head arrays for kmem_cache and the first
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 * kmalloc cache with kmalloc allocated arrays.
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001271 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
Christoph Lametere498be72005-09-09 13:03:32 -07001272 * the other cache's with kmalloc allocated memory.
1273 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 */
1275
Christoph Lameter9b030cb2012-09-05 00:20:33 +00001276 /* 1) create the kmem_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
Eric Dumazet8da34302007-05-06 14:49:29 -07001278 /*
Eric Dumazetb56efcf2011-07-20 19:04:23 +02001279 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
Eric Dumazet8da34302007-05-06 14:49:29 -07001280 */
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001281 create_boot_cache(kmem_cache, "kmem_cache",
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001282 offsetof(struct kmem_cache, node) +
Christoph Lameter6744f082013-01-10 19:12:17 +00001283 nr_node_ids * sizeof(struct kmem_cache_node *),
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001284 SLAB_HWCACHE_ALIGN);
1285 list_add(&kmem_cache->list, &slab_caches);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001286 slab_state = PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287
Andrew Mortona737b3e2006-03-22 00:08:11 -08001288 /*
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001289 * Initialize the caches that provide memory for the kmem_cache_node
1290 * structures first. Without this, further allocations will bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001291 */
Vlastimil Babkaaf3b5f82017-02-22 15:41:05 -08001292 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache(
1293 kmalloc_info[INDEX_NODE].name,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001294 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001295 slab_state = PARTIAL_NODE;
Daniel Sanders34cc6992015-06-24 16:55:57 -07001296 setup_kmalloc_cache_index_table();
Christoph Lametere498be72005-09-09 13:03:32 -07001297
Ingo Molnare0a42722006-06-23 02:03:46 -07001298 slab_early_init = 0;
1299
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001300 /* 5) Replace the bootstrap kmem_cache_node */
Christoph Lametere498be72005-09-09 13:03:32 -07001301 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001302 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
Mel Gorman9c09a952008-01-24 05:49:54 -08001304 for_each_online_node(nid) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001305 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001306
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001307 init_list(kmalloc_caches[INDEX_NODE],
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001308 &init_kmem_cache_node[SIZE_NODE + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001309 }
1310 }
1311
Christoph Lameterf97d5f62013-01-10 19:12:17 +00001312 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
Pekka Enberg8429db52009-06-12 15:58:59 +03001313}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001314
Pekka Enberg8429db52009-06-12 15:58:59 +03001315void __init kmem_cache_init_late(void)
1316{
1317 struct kmem_cache *cachep;
1318
Christoph Lameter97d06602012-07-06 15:25:11 -05001319 slab_state = UP;
Peter Zijlstra52cef182011-11-28 21:12:40 +01001320
Pekka Enberg8429db52009-06-12 15:58:59 +03001321 /* 6) resize the head arrays to their final sizes */
Christoph Lameter18004c52012-07-06 15:25:12 -05001322 mutex_lock(&slab_mutex);
1323 list_for_each_entry(cachep, &slab_caches, list)
Pekka Enberg8429db52009-06-12 15:58:59 +03001324 if (enable_cpucache(cachep, GFP_NOWAIT))
1325 BUG();
Christoph Lameter18004c52012-07-06 15:25:12 -05001326 mutex_unlock(&slab_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001327
Christoph Lameter97d06602012-07-06 15:25:11 -05001328 /* Done! */
1329 slab_state = FULL;
1330
David Rientjes8f9f8d92010-03-27 19:40:47 -07001331#ifdef CONFIG_NUMA
1332 /*
1333 * Register a memory hotplug callback that initializes and frees
Christoph Lameter6a673682013-01-10 19:14:19 +00001334 * node.
David Rientjes8f9f8d92010-03-27 19:40:47 -07001335 */
1336 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1337#endif
1338
Andrew Mortona737b3e2006-03-22 00:08:11 -08001339 /*
1340 * The reap timers are started later, with a module init call: That part
1341 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 */
1343}
1344
1345static int __init cpucache_init(void)
1346{
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001347 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
Andrew Mortona737b3e2006-03-22 00:08:11 -08001349 /*
1350 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 */
Sebastian Andrzej Siewior6731d4f2016-08-23 14:53:19 +02001352 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1353 slab_online_cpu, slab_offline_cpu);
1354 WARN_ON(ret < 0);
Glauber Costaa164f8962012-06-21 00:59:18 +04001355
1356 /* Done! */
Christoph Lameter97d06602012-07-06 15:25:11 -05001357 slab_state = FULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 return 0;
1359}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360__initcall(cpucache_init);
1361
Rafael Aquini8bdec192012-03-09 17:27:27 -03001362static noinline void
1363slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1364{
David Rientjes9a02d692014-06-04 16:06:36 -07001365#if DEBUG
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001366 struct kmem_cache_node *n;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001367 unsigned long flags;
1368 int node;
David Rientjes9a02d692014-06-04 16:06:36 -07001369 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1370 DEFAULT_RATELIMIT_BURST);
1371
1372 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1373 return;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001374
Vlastimil Babka5b3810e2016-03-15 14:56:33 -07001375 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1376 nodeid, gfpflags, &gfpflags);
1377 pr_warn(" cache: %s, object size: %d, order: %d\n",
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05001378 cachep->name, cachep->size, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001379
Christoph Lameter18bf8542014-08-06 16:04:11 -07001380 for_each_kmem_cache_node(cachep, node, n) {
David Rientjesbf00bd32016-12-12 16:41:44 -08001381 unsigned long total_slabs, free_slabs, free_objs;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001382
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001383 spin_lock_irqsave(&n->list_lock, flags);
David Rientjesbf00bd32016-12-12 16:41:44 -08001384 total_slabs = n->total_slabs;
1385 free_slabs = n->free_slabs;
1386 free_objs = n->free_objects;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00001387 spin_unlock_irqrestore(&n->list_lock, flags);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001388
David Rientjesbf00bd32016-12-12 16:41:44 -08001389 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1390 node, total_slabs - free_slabs, total_slabs,
1391 (total_slabs * cachep->num) - free_objs,
1392 total_slabs * cachep->num);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001393 }
David Rientjes9a02d692014-06-04 16:06:36 -07001394#endif
Rafael Aquini8bdec192012-03-09 17:27:27 -03001395}
1396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397/*
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001398 * Interface to system's page allocator. No need to hold the
1399 * kmem_cache_node ->list_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 *
1401 * If we requested dmaable memory, we will get it. Even if we
1402 * did not request dmaable memory, we might get it, but that
1403 * would be relatively rare and ignorable.
1404 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001405static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1406 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407{
1408 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001409 int nr_pages;
Christoph Lameter765c4502006-09-27 01:50:08 -07001410
Glauber Costaa618e892012-06-14 16:17:21 +04001411 flags |= cachep->allocflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001412 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1413 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001414
Vlastimil Babka96db8002015-09-08 15:03:50 -07001415 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Rafael Aquini8bdec192012-03-09 17:27:27 -03001416 if (!page) {
David Rientjes9a02d692014-06-04 16:06:36 -07001417 slab_out_of_memory(cachep, flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 return NULL;
Rafael Aquini8bdec192012-03-09 17:27:27 -03001419 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420
Vladimir Davydovf3ccb2c42015-11-05 18:49:01 -08001421 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1422 __free_pages(page, cachep->gfporder);
1423 return NULL;
1424 }
1425
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001426 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001428 add_zone_page_state(page_zone(page),
1429 NR_SLAB_RECLAIMABLE, nr_pages);
1430 else
1431 add_zone_page_state(page_zone(page),
1432 NR_SLAB_UNRECLAIMABLE, nr_pages);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001433
Joonsoo Kima57a4982013-10-24 10:07:44 +09001434 __SetPageSlab(page);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07001435 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1436 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
Joonsoo Kima57a4982013-10-24 10:07:44 +09001437 SetPageSlabPfmemalloc(page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07001438
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001439 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1440 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1441
1442 if (cachep->ctor)
1443 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1444 else
1445 kmemcheck_mark_unallocated_pages(page, nr_pages);
1446 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001447
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001448 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449}
1450
1451/*
1452 * Interface to system's page release.
1453 */
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001454static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455{
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001456 int order = cachep->gfporder;
1457 unsigned long nr_freed = (1 << order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001459 kmemcheck_free_shadow(page, order);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001460
Christoph Lameter972d1a72006-09-25 23:31:51 -07001461 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1462 sub_zone_page_state(page_zone(page),
1463 NR_SLAB_RECLAIMABLE, nr_freed);
1464 else
1465 sub_zone_page_state(page_zone(page),
1466 NR_SLAB_UNRECLAIMABLE, nr_freed);
Joonsoo Kim73293c22013-10-24 10:07:37 +09001467
Joonsoo Kima57a4982013-10-24 10:07:44 +09001468 BUG_ON(!PageSlab(page));
Joonsoo Kim73293c22013-10-24 10:07:37 +09001469 __ClearPageSlabPfmemalloc(page);
Joonsoo Kima57a4982013-10-24 10:07:44 +09001470 __ClearPageSlab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09001471 page_mapcount_reset(page);
1472 page->mapping = NULL;
Glauber Costa1f458cb2012-12-18 14:22:50 -08001473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 if (current->reclaim_state)
1475 current->reclaim_state->reclaimed_slab += nr_freed;
Vladimir Davydov27ee57c2016-03-17 14:17:35 -07001476 memcg_uncharge_slab(page, order, cachep);
1477 __free_pages(page, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478}
1479
1480static void kmem_rcu_free(struct rcu_head *head)
1481{
Joonsoo Kim68126702013-10-24 10:07:42 +09001482 struct kmem_cache *cachep;
1483 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
Joonsoo Kim68126702013-10-24 10:07:42 +09001485 page = container_of(head, struct page, rcu_head);
1486 cachep = page->slab_cache;
1487
1488 kmem_freepages(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489}
1490
1491#if DEBUG
Joonsoo Kim40b44132016-03-15 14:54:21 -07001492static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1493{
1494 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1495 (cachep->size % PAGE_SIZE) == 0)
1496 return true;
1497
1498 return false;
1499}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500
1501#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001502static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001503 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001505 int size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001507 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001509 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 return;
1511
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001512 *addr++ = 0x12345678;
1513 *addr++ = caller;
1514 *addr++ = smp_processor_id();
1515 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 {
1517 unsigned long *sptr = &caller;
1518 unsigned long svalue;
1519
1520 while (!kstack_end(sptr)) {
1521 svalue = *sptr++;
1522 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001523 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 size -= sizeof(unsigned long);
1525 if (size <= sizeof(unsigned long))
1526 break;
1527 }
1528 }
1529
1530 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001531 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532}
Joonsoo Kim40b44132016-03-15 14:54:21 -07001533
1534static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1535 int map, unsigned long caller)
1536{
1537 if (!is_debug_pagealloc_cache(cachep))
1538 return;
1539
1540 if (caller)
1541 store_stackinfo(cachep, objp, caller);
1542
1543 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1544}
1545
1546#else
1547static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1548 int map, unsigned long caller) {}
1549
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550#endif
1551
Pekka Enberg343e0d72006-02-01 03:05:50 -08001552static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553{
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001554 int size = cachep->object_size;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001555 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556
1557 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001558 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559}
1560
1561static void dump_line(char *data, int offset, int limit)
1562{
1563 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001564 unsigned char error = 0;
1565 int bad_count = 0;
1566
Joe Perches11705322016-03-17 14:19:50 -07001567 pr_err("%03x: ", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001568 for (i = 0; i < limit; i++) {
1569 if (data[offset + i] != POISON_FREE) {
1570 error = data[offset + i];
1571 bad_count++;
1572 }
Dave Jonesaa83aa42006-09-29 01:59:51 -07001573 }
Sebastian Andrzej Siewiorfdde6ab2011-07-29 18:22:13 +02001574 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1575 &data[offset], limit, 1);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001576
1577 if (bad_count == 1) {
1578 error ^= POISON_FREE;
1579 if (!(error & (error - 1))) {
Joe Perches11705322016-03-17 14:19:50 -07001580 pr_err("Single bit error detected. Probably bad RAM.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001581#ifdef CONFIG_X86
Joe Perches11705322016-03-17 14:19:50 -07001582 pr_err("Run memtest86+ or a similar memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001583#else
Joe Perches11705322016-03-17 14:19:50 -07001584 pr_err("Run a memory test tool.\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001585#endif
1586 }
1587 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588}
1589#endif
1590
1591#if DEBUG
1592
Pekka Enberg343e0d72006-02-01 03:05:50 -08001593static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594{
1595 int i, size;
1596 char *realobj;
1597
1598 if (cachep->flags & SLAB_RED_ZONE) {
Joe Perches11705322016-03-17 14:19:50 -07001599 pr_err("Redzone: 0x%llx/0x%llx\n",
1600 *dbg_redzone1(cachep, objp),
1601 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 }
1603
1604 if (cachep->flags & SLAB_STORE_USER) {
Joe Perches11705322016-03-17 14:19:50 -07001605 pr_err("Last user: [<%p>](%pSR)\n",
Joe Perches071361d2012-12-12 10:19:12 -08001606 *dbg_userword(cachep, objp),
1607 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001609 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001610 size = cachep->object_size;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001611 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 int limit;
1613 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001614 if (i + limit > size)
1615 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 dump_line(realobj, i, limit);
1617 }
1618}
1619
Pekka Enberg343e0d72006-02-01 03:05:50 -08001620static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621{
1622 char *realobj;
1623 int size, i;
1624 int lines = 0;
1625
Joonsoo Kim40b44132016-03-15 14:54:21 -07001626 if (is_debug_pagealloc_cache(cachep))
1627 return;
1628
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001629 realobj = (char *)objp + obj_offset(cachep);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05001630 size = cachep->object_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001632 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001634 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 exp = POISON_END;
1636 if (realobj[i] != exp) {
1637 int limit;
1638 /* Mismatch ! */
1639 /* Print header */
1640 if (lines == 0) {
Joe Perches11705322016-03-17 14:19:50 -07001641 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1642 print_tainted(), cachep->name,
1643 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 print_objinfo(cachep, objp, 0);
1645 }
1646 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001647 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001649 if (i + limit > size)
1650 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651 dump_line(realobj, i, limit);
1652 i += 16;
1653 lines++;
1654 /* Limit to 5 lines */
1655 if (lines > 5)
1656 break;
1657 }
1658 }
1659 if (lines != 0) {
1660 /* Print some data about the neighboring objects, if they
1661 * exist:
1662 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001663 struct page *page = virt_to_head_page(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001664 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Joonsoo Kim8456a642013-10-24 10:07:49 +09001666 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 if (objnr) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001668 objp = index_to_obj(cachep, page, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001669 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001670 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 print_objinfo(cachep, objp, 2);
1672 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001673 if (objnr + 1 < cachep->num) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001674 objp = index_to_obj(cachep, page, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001675 realobj = (char *)objp + obj_offset(cachep);
Joe Perches11705322016-03-17 14:19:50 -07001676 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 print_objinfo(cachep, objp, 2);
1678 }
1679 }
1680}
1681#endif
1682
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683#if DEBUG
Joonsoo Kim8456a642013-10-24 10:07:49 +09001684static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1685 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001686{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 int i;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001688
1689 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1690 poison_obj(cachep, page->freelist - obj_offset(cachep),
1691 POISON_FREE);
1692 }
1693
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09001695 void *objp = index_to_obj(cachep, page, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696
1697 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07001699 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 }
1701 if (cachep->flags & SLAB_RED_ZONE) {
1702 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07001703 slab_error(cachep, "start of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07001705 slab_error(cachep, "end of a freed object was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001708}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709#else
Joonsoo Kim8456a642013-10-24 10:07:49 +09001710static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1711 struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001712{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001713}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714#endif
1715
Randy Dunlap911851e2006-03-22 00:08:14 -08001716/**
1717 * slab_destroy - destroy and release all objects in a slab
1718 * @cachep: cache pointer being destroyed
Masanari Iidacb8ee1a2014-01-28 02:57:08 +09001719 * @page: page pointer being destroyed
Randy Dunlap911851e2006-03-22 00:08:14 -08001720 *
Wang Sheng-Hui8a7d9b42014-08-06 16:04:46 -07001721 * Destroy all the objs in a slab page, and release the mem back to the system.
1722 * Before calling the slab page must have been unlinked from the cache. The
1723 * kmem_cache_node ->list_lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001724 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09001725static void slab_destroy(struct kmem_cache *cachep, struct page *page)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001726{
Joonsoo Kim7e007352013-10-30 19:04:01 +09001727 void *freelist;
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001728
Joonsoo Kim8456a642013-10-24 10:07:49 +09001729 freelist = page->freelist;
1730 slab_destroy_debugcheck(cachep, page);
Paul E. McKenney5f0d5a32017-01-18 02:53:44 -08001731 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
Kirill A. Shutemovbc4f6102015-11-06 16:29:44 -08001732 call_rcu(&page->rcu_head, kmem_rcu_free);
1733 else
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09001734 kmem_freepages(cachep, page);
Joonsoo Kim68126702013-10-24 10:07:42 +09001735
1736 /*
Joonsoo Kim8456a642013-10-24 10:07:49 +09001737 * From now on, we don't use freelist
Joonsoo Kim68126702013-10-24 10:07:42 +09001738 * although actual page can be freed in rcu context
1739 */
1740 if (OFF_SLAB(cachep))
Joonsoo Kim8456a642013-10-24 10:07:49 +09001741 kmem_cache_free(cachep->freelist_cache, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742}
1743
Joonsoo Kim97654df2014-08-06 16:04:25 -07001744static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1745{
1746 struct page *page, *n;
1747
1748 list_for_each_entry_safe(page, n, list, lru) {
1749 list_del(&page->lru);
1750 slab_destroy(cachep, page);
1751 }
1752}
1753
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001755 * calculate_slab_order - calculate size (page order) of slabs
1756 * @cachep: pointer to the cache that is being created
1757 * @size: size of objects to be created in this cache.
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001758 * @flags: slab allocation flags
1759 *
1760 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001761 *
1762 * This could be made much more intelligent. For now, try to avoid using
1763 * high order pages for slabs. When the gfp() functions are more friendly
1764 * towards high-order requests, this should be changed.
1765 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001766static size_t calculate_slab_order(struct kmem_cache *cachep,
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07001767 size_t size, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001768{
1769 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001770 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001771
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001772 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001773 unsigned int num;
1774 size_t remainder;
1775
Joonsoo Kim70f75062016-03-15 14:54:53 -07001776 num = cache_estimate(gfporder, size, flags, &remainder);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001777 if (!num)
1778 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001779
Joonsoo Kimf315e3f2013-12-02 17:49:41 +09001780 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1781 if (num > SLAB_OBJ_MAX_NUM)
1782 break;
1783
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001784 if (flags & CFLGS_OFF_SLAB) {
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001785 struct kmem_cache *freelist_cache;
1786 size_t freelist_size;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001787
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001788 freelist_size = num * sizeof(freelist_idx_t);
1789 freelist_cache = kmalloc_slab(freelist_size, 0u);
1790 if (!freelist_cache)
1791 continue;
1792
1793 /*
1794 * Needed to avoid possible looping condition
Joonsoo Kim76b342b2016-05-19 17:10:26 -07001795 * in cache_grow_begin()
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001796 */
1797 if (OFF_SLAB(freelist_cache))
1798 continue;
1799
1800 /* check if off slab has enough benefit */
1801 if (freelist_cache->size > cachep->size / 2)
1802 continue;
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001803 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001804
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001805 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001806 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001807 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001808 left_over = remainder;
1809
1810 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001811 * A VFS-reclaimable slab tends to have most allocations
1812 * as GFP_NOFS and we really don't want to have to be allocating
1813 * higher-order pages when we are unable to shrink dcache.
1814 */
1815 if (flags & SLAB_RECLAIM_ACCOUNT)
1816 break;
1817
1818 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001819 * Large number of objects is good, but very large slabs are
1820 * currently bad for the gfp()s.
1821 */
David Rientjes543585c2011-10-18 22:09:24 -07001822 if (gfporder >= slab_max_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001823 break;
1824
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001825 /*
1826 * Acceptable internal fragmentation?
1827 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001828 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001829 break;
1830 }
1831 return left_over;
1832}
1833
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001834static struct array_cache __percpu *alloc_kmem_cache_cpus(
1835 struct kmem_cache *cachep, int entries, int batchcount)
1836{
1837 int cpu;
1838 size_t size;
1839 struct array_cache __percpu *cpu_cache;
1840
1841 size = sizeof(void *) * entries + sizeof(struct array_cache);
Joonsoo Kim85c9f4b2014-10-13 15:51:01 -07001842 cpu_cache = __alloc_percpu(size, sizeof(void *));
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001843
1844 if (!cpu_cache)
1845 return NULL;
1846
1847 for_each_possible_cpu(cpu) {
1848 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1849 entries, batchcount);
1850 }
1851
1852 return cpu_cache;
1853}
1854
Fabian Frederickbd721ea2016-08-02 14:03:33 -07001855static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001856{
Christoph Lameter97d06602012-07-06 15:25:11 -05001857 if (slab_state >= FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03001858 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001859
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001860 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1861 if (!cachep->cpu_cache)
1862 return 1;
1863
Christoph Lameter97d06602012-07-06 15:25:11 -05001864 if (slab_state == DOWN) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001865 /* Creation of first cache (kmem_cache). */
1866 set_up_node(kmem_cache, CACHE_CACHE);
Christoph Lameter2f9baa92012-11-28 16:23:09 +00001867 } else if (slab_state == PARTIAL) {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001868 /* For kmem_cache_node */
1869 set_up_node(cachep, SIZE_NODE);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001870 } else {
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001871 int node;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001872
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001873 for_each_online_node(node) {
1874 cachep->node[node] = kmalloc_node(
1875 sizeof(struct kmem_cache_node), gfp, node);
1876 BUG_ON(!cachep->node[node]);
1877 kmem_cache_node_init(cachep->node[node]);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001878 }
1879 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07001880
Christoph Lameter6a673682013-01-10 19:14:19 +00001881 cachep->node[numa_mem_id()]->next_reap =
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08001882 jiffies + REAPTIMEOUT_NODE +
1883 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001884
1885 cpu_cache_get(cachep)->avail = 0;
1886 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1887 cpu_cache_get(cachep)->batchcount = 1;
1888 cpu_cache_get(cachep)->touched = 0;
1889 cachep->batchcount = 1;
1890 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001891 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001892}
1893
Joonsoo Kim12220de2014-10-09 15:26:24 -07001894unsigned long kmem_cache_flags(unsigned long object_size,
1895 unsigned long flags, const char *name,
1896 void (*ctor)(void *))
1897{
1898 return flags;
1899}
1900
1901struct kmem_cache *
1902__kmem_cache_alias(const char *name, size_t size, size_t align,
1903 unsigned long flags, void (*ctor)(void *))
1904{
1905 struct kmem_cache *cachep;
1906
1907 cachep = find_mergeable(size, align, flags, name, ctor);
1908 if (cachep) {
1909 cachep->refcount++;
1910
1911 /*
1912 * Adjust the object sizes so that we clear
1913 * the complete object on kzalloc.
1914 */
1915 cachep->object_size = max_t(int, cachep->object_size, size);
1916 }
1917 return cachep;
1918}
1919
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001920static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1921 size_t size, unsigned long flags)
1922{
1923 size_t left;
1924
1925 cachep->num = 0;
1926
Paul E. McKenney5f0d5a32017-01-18 02:53:44 -08001927 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07001928 return false;
1929
1930 left = calculate_slab_order(cachep, size,
1931 flags | CFLGS_OBJFREELIST_SLAB);
1932 if (!cachep->num)
1933 return false;
1934
1935 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1936 return false;
1937
1938 cachep->colour = left / cachep->colour_off;
1939
1940 return true;
1941}
1942
Joonsoo Kim158e3192016-03-15 14:54:35 -07001943static bool set_off_slab_cache(struct kmem_cache *cachep,
1944 size_t size, unsigned long flags)
1945{
1946 size_t left;
1947
1948 cachep->num = 0;
1949
1950 /*
Joonsoo Kim3217fd92016-03-15 14:54:41 -07001951 * Always use on-slab management when SLAB_NOLEAKTRACE
1952 * to avoid recursive calls into kmemleak.
Joonsoo Kim158e3192016-03-15 14:54:35 -07001953 */
Joonsoo Kim158e3192016-03-15 14:54:35 -07001954 if (flags & SLAB_NOLEAKTRACE)
1955 return false;
1956
1957 /*
1958 * Size is large, assume best to place the slab management obj
1959 * off-slab (should allow better packing of objs).
1960 */
1961 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1962 if (!cachep->num)
1963 return false;
1964
1965 /*
1966 * If the slab has been placed off-slab, and we have enough space then
1967 * move it on-slab. This is at the expense of any extra colouring.
1968 */
1969 if (left >= cachep->num * sizeof(freelist_idx_t))
1970 return false;
1971
1972 cachep->colour = left / cachep->colour_off;
1973
1974 return true;
1975}
1976
1977static bool set_on_slab_cache(struct kmem_cache *cachep,
1978 size_t size, unsigned long flags)
1979{
1980 size_t left;
1981
1982 cachep->num = 0;
1983
1984 left = calculate_slab_order(cachep, size, flags);
1985 if (!cachep->num)
1986 return false;
1987
1988 cachep->colour = left / cachep->colour_off;
1989
1990 return true;
1991}
1992
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001993/**
Christoph Lameter039363f2012-07-06 15:25:10 -05001994 * __kmem_cache_create - Create a cache.
Randy Dunlapa755b762012-11-06 17:10:10 -08001995 * @cachep: cache management descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 * @flags: SLAB flags
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 *
1998 * Returns a ptr to the cache on success, NULL on failure.
1999 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002000 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 * The flags are
2003 *
2004 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2005 * to catch references to uninitialised memory.
2006 *
2007 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2008 * for buffer overruns.
2009 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2011 * cacheline. This can be beneficial if you're counting cycles as closely
2012 * as davem.
2013 */
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002014int
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002015__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016{
David Rientjesd4a5fca2014-09-25 16:05:20 -07002017 size_t ralign = BYTES_PER_WORD;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002018 gfp_t gfp;
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002019 int err;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002020 size_t size = cachep->size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023#if FORCED_DEBUG
2024 /*
2025 * Enable redzoning and last user accounting, except for caches with
2026 * large objects, if the increased size would increase the object size
2027 * above the next power of two: caches with object sizes just above a
2028 * power of two have a significant amount of internal fragmentation.
2029 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002030 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2031 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002032 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Paul E. McKenney5f0d5a32017-01-18 02:53:44 -08002033 if (!(flags & SLAB_TYPESAFE_BY_RCU))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 flags |= SLAB_POISON;
2035#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037
Andrew Mortona737b3e2006-03-22 00:08:11 -08002038 /*
2039 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 * unaligned accesses for some archs when redzoning is used, and makes
2041 * sure any on-slab bufctl's are also correctly aligned.
2042 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002043 if (size & (BYTES_PER_WORD - 1)) {
2044 size += (BYTES_PER_WORD - 1);
2045 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 }
2047
David Woodhouse87a927c2007-07-04 21:26:44 -04002048 if (flags & SLAB_RED_ZONE) {
2049 ralign = REDZONE_ALIGN;
2050 /* If redzoning, ensure that the second redzone is suitably
2051 * aligned, by adjusting the object size accordingly. */
2052 size += REDZONE_ALIGN - 1;
2053 size &= ~(REDZONE_ALIGN - 1);
2054 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002055
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002056 /* 3) caller mandated alignment */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002057 if (ralign < cachep->align) {
2058 ralign = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 }
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002060 /* disable debug if necessary */
2061 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002062 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002063 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002064 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065 */
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002066 cachep->align = ralign;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002067 cachep->colour_off = cache_line_size();
2068 /* Offset must be a multiple of the alignment. */
2069 if (cachep->colour_off < cachep->align)
2070 cachep->colour_off = cachep->align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071
Pekka Enberg83b519e2009-06-10 19:40:04 +03002072 if (slab_is_available())
2073 gfp = GFP_KERNEL;
2074 else
2075 gfp = GFP_NOWAIT;
2076
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078
Pekka Enbergca5f9702006-09-25 23:31:25 -07002079 /*
2080 * Both debugging options require word-alignment which is calculated
2081 * into align above.
2082 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 /* add space for red zone words */
Pekka Enberg3ff84a72011-02-14 17:46:21 +02002085 cachep->obj_offset += sizeof(unsigned long long);
2086 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 }
2088 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002089 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002090 * the real object. But if the second red zone needs to be
2091 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002093 if (flags & SLAB_RED_ZONE)
2094 size += REDZONE_ALIGN;
2095 else
2096 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 }
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002098#endif
2099
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002100 kasan_cache_create(cachep, &size, &flags);
2101
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002102 size = ALIGN(size, cachep->align);
2103 /*
2104 * We should restrict the number of objects in a slab to implement
2105 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2106 */
2107 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2108 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2109
2110#if DEBUG
Joonsoo Kim03a2d2a2015-10-01 15:36:54 -07002111 /*
2112 * To activate debug pagealloc, off-slab management is necessary
2113 * requirement. In early phase of initialization, small sized slab
2114 * doesn't get initialized so it would not be possible. So, we need
2115 * to check size >= 256. It guarantees that all necessary small
2116 * sized slab is initialized in current slab initialization sequence.
2117 */
Joonsoo Kim40323272016-03-15 14:54:18 -07002118 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
Joonsoo Kimf3a3c322016-03-15 14:54:38 -07002119 size >= 256 && cachep->object_size > cache_line_size()) {
2120 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2121 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2122
2123 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2124 flags |= CFLGS_OFF_SLAB;
2125 cachep->obj_offset += tmp_size - size;
2126 size = tmp_size;
2127 goto done;
2128 }
2129 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 }
2131#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002133 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2134 flags |= CFLGS_OBJFREELIST_SLAB;
2135 goto done;
2136 }
2137
Joonsoo Kim158e3192016-03-15 14:54:35 -07002138 if (set_off_slab_cache(cachep, size, flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 flags |= CFLGS_OFF_SLAB;
Joonsoo Kim158e3192016-03-15 14:54:35 -07002140 goto done;
Joonsoo Kim832a15d2016-03-15 14:54:33 -07002141 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142
Joonsoo Kim158e3192016-03-15 14:54:35 -07002143 if (set_on_slab_cache(cachep, size, flags))
2144 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145
Joonsoo Kim158e3192016-03-15 14:54:35 -07002146 return -E2BIG;
Christoph Lameter8a13a4c2012-09-04 23:18:33 +00002147
Joonsoo Kim158e3192016-03-15 14:54:35 -07002148done:
2149 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 cachep->flags = flags;
Joonsoo Kima57a4982013-10-24 10:07:44 +09002151 cachep->allocflags = __GFP_COMP;
Yang Shia3187e42016-05-19 17:10:41 -07002152 if (flags & SLAB_CACHE_DMA)
Glauber Costaa618e892012-06-14 16:17:21 +04002153 cachep->allocflags |= GFP_DMA;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05002154 cachep->size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002155 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
Joonsoo Kim40b44132016-03-15 14:54:21 -07002157#if DEBUG
2158 /*
2159 * If we're going to use the generic kernel_map_pages()
2160 * poisoning, then it's going to smash the contents of
2161 * the redzone and userword anyhow, so switch them off.
2162 */
2163 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2164 (cachep->flags & SLAB_POISON) &&
2165 is_debug_pagealloc_cache(cachep))
2166 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2167#endif
2168
2169 if (OFF_SLAB(cachep)) {
Joonsoo Kim158e3192016-03-15 14:54:35 -07002170 cachep->freelist_cache =
2171 kmalloc_slab(cachep->freelist_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002172 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002174 err = setup_cpu_cache(cachep, gfp);
2175 if (err) {
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002176 __kmem_cache_release(cachep);
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002177 return err;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002178 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179
Christoph Lameter278b1bb2012-09-05 00:20:34 +00002180 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182
2183#if DEBUG
2184static void check_irq_off(void)
2185{
2186 BUG_ON(!irqs_disabled());
2187}
2188
2189static void check_irq_on(void)
2190{
2191 BUG_ON(irqs_disabled());
2192}
2193
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002194static void check_mutex_acquired(void)
2195{
2196 BUG_ON(!mutex_is_locked(&slab_mutex));
2197}
2198
Pekka Enberg343e0d72006-02-01 03:05:50 -08002199static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200{
2201#ifdef CONFIG_SMP
2202 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002203 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204#endif
2205}
Christoph Lametere498be72005-09-09 13:03:32 -07002206
Pekka Enberg343e0d72006-02-01 03:05:50 -08002207static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002208{
2209#ifdef CONFIG_SMP
2210 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002211 assert_spin_locked(&get_node(cachep, node)->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002212#endif
2213}
2214
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215#else
2216#define check_irq_off() do { } while(0)
2217#define check_irq_on() do { } while(0)
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002218#define check_mutex_acquired() do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002220#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221#endif
2222
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002223static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2224 int node, bool free_all, struct list_head *list)
2225{
2226 int tofree;
2227
2228 if (!ac || !ac->avail)
2229 return;
2230
2231 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2232 if (tofree > ac->avail)
2233 tofree = (ac->avail + 1) / 2;
2234
2235 free_block(cachep, ac->entry, tofree, node, list);
2236 ac->avail -= tofree;
2237 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2238}
Christoph Lameteraab22072006-03-22 00:09:06 -08002239
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240static void do_drain(void *arg)
2241{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002242 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 struct array_cache *ac;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07002244 int node = numa_mem_id();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002245 struct kmem_cache_node *n;
Joonsoo Kim97654df2014-08-06 16:04:25 -07002246 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247
2248 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002249 ac = cpu_cache_get(cachep);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002250 n = get_node(cachep, node);
2251 spin_lock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002252 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07002253 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07002254 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 ac->avail = 0;
2256}
2257
Pekka Enberg343e0d72006-02-01 03:05:50 -08002258static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259{
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002260 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002261 int node;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002262 LIST_HEAD(list);
Christoph Lametere498be72005-09-09 13:03:32 -07002263
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002264 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002266 for_each_kmem_cache_node(cachep, node, n)
2267 if (n->alien)
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002268 drain_alien_cache(cachep, n->alien);
Roland Dreiera4523a82006-05-15 11:41:00 -07002269
Joonsoo Kim18726ca2016-05-19 17:10:02 -07002270 for_each_kmem_cache_node(cachep, node, n) {
2271 spin_lock_irq(&n->list_lock);
2272 drain_array_locked(cachep, n->shared, node, true, &list);
2273 spin_unlock_irq(&n->list_lock);
2274
2275 slabs_destroy(cachep, &list);
2276 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277}
2278
Christoph Lametered11d9e2006-06-30 01:55:45 -07002279/*
2280 * Remove slabs from the list of free slabs.
2281 * Specify the number of slabs to drain in tofree.
2282 *
2283 * Returns the actual number of slabs released.
2284 */
2285static int drain_freelist(struct kmem_cache *cache,
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002286 struct kmem_cache_node *n, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002288 struct list_head *p;
2289 int nr_freed;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002290 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
Christoph Lametered11d9e2006-06-30 01:55:45 -07002292 nr_freed = 0;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002293 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002295 spin_lock_irq(&n->list_lock);
2296 p = n->slabs_free.prev;
2297 if (p == &n->slabs_free) {
2298 spin_unlock_irq(&n->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002299 goto out;
2300 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301
Joonsoo Kim8456a642013-10-24 10:07:49 +09002302 page = list_entry(p, struct page, lru);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002303 list_del(&page->lru);
Greg Thelenf728b0a2016-12-12 16:41:41 -08002304 n->free_slabs--;
David Rientjesbf00bd32016-12-12 16:41:44 -08002305 n->total_slabs--;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002306 /*
2307 * Safe to drop the lock. The slab is no longer linked
2308 * to the cache.
2309 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002310 n->free_objects -= cache->num;
2311 spin_unlock_irq(&n->list_lock);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002312 slab_destroy(cache, page);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002313 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002315out:
2316 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317}
2318
Tejun Heoc9fc5862017-02-22 15:41:27 -08002319int __kmem_cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002320{
Christoph Lameter18bf8542014-08-06 16:04:11 -07002321 int ret = 0;
2322 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002323 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07002324
2325 drain_cpu_caches(cachep);
2326
2327 check_irq_on();
Christoph Lameter18bf8542014-08-06 16:04:11 -07002328 for_each_kmem_cache_node(cachep, node, n) {
Joonsoo Kima5aa63a2016-05-19 17:10:08 -07002329 drain_freelist(cachep, n, INT_MAX);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002330
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002331 ret += !list_empty(&n->slabs_full) ||
2332 !list_empty(&n->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002333 }
2334 return (ret ? 1 : 0);
2335}
2336
Tejun Heoc9fc5862017-02-22 15:41:27 -08002337#ifdef CONFIG_MEMCG
2338void __kmemcg_cache_deactivate(struct kmem_cache *cachep)
2339{
2340 __kmem_cache_shrink(cachep);
2341}
2342#endif
2343
Christoph Lameter945cf2b2012-09-04 23:18:33 +00002344int __kmem_cache_shutdown(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345{
Tejun Heoc9fc5862017-02-22 15:41:27 -08002346 return __kmem_cache_shrink(cachep);
Dmitry Safonov52b4b952016-02-17 13:11:37 -08002347}
2348
2349void __kmem_cache_release(struct kmem_cache *cachep)
2350{
Christoph Lameter12c36672012-09-04 23:38:33 +00002351 int i;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002352 struct kmem_cache_node *n;
Christoph Lameter12c36672012-09-04 23:38:33 +00002353
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002354 cache_random_seq_destroy(cachep);
2355
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07002356 free_percpu(cachep->cpu_cache);
Christoph Lameter12c36672012-09-04 23:38:33 +00002357
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002358 /* NUMA: free the node structures */
Christoph Lameter18bf8542014-08-06 16:04:11 -07002359 for_each_kmem_cache_node(cachep, i, n) {
2360 kfree(n->shared);
2361 free_alien_cache(n->alien);
2362 kfree(n);
2363 cachep->node[i] = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002367/*
2368 * Get the memory for a slab management obj.
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08002369 *
2370 * For a slab cache when the slab descriptor is off-slab, the
2371 * slab descriptor can't come from the same cache which is being created,
2372 * Because if it is the case, that means we defer the creation of
2373 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2374 * And we eventually call down to __kmem_cache_create(), which
2375 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2376 * This is a "chicken-and-egg" problem.
2377 *
2378 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2379 * which are all initialized during kmem_cache_init().
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002380 */
Joonsoo Kim7e007352013-10-30 19:04:01 +09002381static void *alloc_slabmgmt(struct kmem_cache *cachep,
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002382 struct page *page, int colour_off,
2383 gfp_t local_flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002385 void *freelist;
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002386 void *addr = page_address(page);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002387
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002388 page->s_mem = addr + colour_off;
2389 page->active = 0;
2390
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002391 if (OBJFREELIST_SLAB(cachep))
2392 freelist = NULL;
2393 else if (OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 /* Slab management obj is off-slab. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002395 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002396 local_flags, nodeid);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002397 if (!freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 return NULL;
2399 } else {
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002400 /* We will use last bytes at the slab for freelist */
2401 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2402 cachep->freelist_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 }
Joonsoo Kim2e6b3602016-03-15 14:54:30 -07002404
Joonsoo Kim8456a642013-10-24 10:07:49 +09002405 return freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406}
2407
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002408static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002410 return ((freelist_idx_t *)page->freelist)[idx];
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002411}
2412
2413static inline void set_free_obj(struct page *page,
Joonsoo Kim7cc68972014-04-18 16:24:09 +09002414 unsigned int idx, freelist_idx_t val)
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002415{
Joonsoo Kima41adfa2013-12-02 17:49:42 +09002416 ((freelist_idx_t *)(page->freelist))[idx] = val;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417}
2418
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002419static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420{
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002421#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422 int i;
2423
2424 for (i = 0; i < cachep->num; i++) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09002425 void *objp = index_to_obj(cachep, page, i);
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002426
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 if (cachep->flags & SLAB_STORE_USER)
2428 *dbg_userword(cachep, objp) = NULL;
2429
2430 if (cachep->flags & SLAB_RED_ZONE) {
2431 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2432 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2433 }
2434 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002435 * Constructors are not allowed to allocate memory from the same
2436 * cache which they are a constructor for. Otherwise, deadlock.
2437 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002439 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2440 kasan_unpoison_object_data(cachep,
2441 objp + obj_offset(cachep));
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002442 cachep->ctor(objp + obj_offset(cachep));
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002443 kasan_poison_object_data(
2444 cachep, objp + obj_offset(cachep));
2445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446
2447 if (cachep->flags & SLAB_RED_ZONE) {
2448 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07002449 slab_error(cachep, "constructor overwrote the end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
Joe Perches756a0252016-03-17 14:19:47 -07002451 slab_error(cachep, "constructor overwrote the start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 }
Joonsoo Kim40b44132016-03-15 14:54:21 -07002453 /* need to poison the objs? */
2454 if (cachep->flags & SLAB_POISON) {
2455 poison_obj(cachep, objp, POISON_FREE);
2456 slab_kernel_map(cachep, objp, 0, 0);
2457 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002458 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459#endif
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002460}
2461
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002462#ifdef CONFIG_SLAB_FREELIST_RANDOM
2463/* Hold information during a freelist initialization */
2464union freelist_init_state {
2465 struct {
2466 unsigned int pos;
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002467 unsigned int *list;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002468 unsigned int count;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002469 };
2470 struct rnd_state rnd_state;
2471};
2472
2473/*
2474 * Initialize the state based on the randomization methode available.
2475 * return true if the pre-computed list is available, false otherwize.
2476 */
2477static bool freelist_state_initialize(union freelist_init_state *state,
2478 struct kmem_cache *cachep,
2479 unsigned int count)
2480{
2481 bool ret;
2482 unsigned int rand;
2483
2484 /* Use best entropy available to define a random shift */
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002485 rand = get_random_int();
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002486
2487 /* Use a random state if the pre-computed list is not available */
2488 if (!cachep->random_seq) {
2489 prandom_seed_state(&state->rnd_state, rand);
2490 ret = false;
2491 } else {
2492 state->list = cachep->random_seq;
2493 state->count = count;
John Sperbeckc4e490c2017-01-10 16:58:24 -08002494 state->pos = rand % count;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002495 ret = true;
2496 }
2497 return ret;
2498}
2499
2500/* Get the next entry on the list and randomize it using a random shift */
2501static freelist_idx_t next_random_slot(union freelist_init_state *state)
2502{
John Sperbeckc4e490c2017-01-10 16:58:24 -08002503 if (state->pos >= state->count)
2504 state->pos = 0;
2505 return state->list[state->pos++];
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002506}
2507
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002508/* Swap two freelist entries */
2509static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2510{
2511 swap(((freelist_idx_t *)page->freelist)[a],
2512 ((freelist_idx_t *)page->freelist)[b]);
2513}
2514
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002515/*
2516 * Shuffle the freelist initialization state based on pre-computed lists.
2517 * return true if the list was successfully shuffled, false otherwise.
2518 */
2519static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2520{
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002521 unsigned int objfreelist = 0, i, rand, count = cachep->num;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002522 union freelist_init_state state;
2523 bool precomputed;
2524
2525 if (count < 2)
2526 return false;
2527
2528 precomputed = freelist_state_initialize(&state, cachep, count);
2529
2530 /* Take a random entry as the objfreelist */
2531 if (OBJFREELIST_SLAB(cachep)) {
2532 if (!precomputed)
2533 objfreelist = count - 1;
2534 else
2535 objfreelist = next_random_slot(&state);
2536 page->freelist = index_to_obj(cachep, page, objfreelist) +
2537 obj_offset(cachep);
2538 count--;
2539 }
2540
2541 /*
2542 * On early boot, generate the list dynamically.
2543 * Later use a pre-computed list for speed.
2544 */
2545 if (!precomputed) {
Thomas Garnier7c00fce2016-07-26 15:21:56 -07002546 for (i = 0; i < count; i++)
2547 set_free_obj(page, i, i);
2548
2549 /* Fisher-Yates shuffle */
2550 for (i = count - 1; i > 0; i--) {
2551 rand = prandom_u32_state(&state.rnd_state);
2552 rand %= (i + 1);
2553 swap_free_obj(page, i, rand);
2554 }
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002555 } else {
2556 for (i = 0; i < count; i++)
2557 set_free_obj(page, i, next_random_slot(&state));
2558 }
2559
2560 if (OBJFREELIST_SLAB(cachep))
2561 set_free_obj(page, cachep->num - 1, objfreelist);
2562
2563 return true;
2564}
2565#else
2566static inline bool shuffle_freelist(struct kmem_cache *cachep,
2567 struct page *page)
2568{
2569 return false;
2570}
2571#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2572
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002573static void cache_init_objs(struct kmem_cache *cachep,
2574 struct page *page)
2575{
2576 int i;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002577 void *objp;
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002578 bool shuffled;
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002579
2580 cache_init_objs_debug(cachep, page);
2581
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002582 /* Try to randomize the freelist if enabled */
2583 shuffled = shuffle_freelist(cachep, page);
2584
2585 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002586 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2587 obj_offset(cachep);
2588 }
2589
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002590 for (i = 0; i < cachep->num; i++) {
Andrey Ryabininb3cbd9b2016-08-02 14:02:52 -07002591 objp = index_to_obj(cachep, page, i);
2592 kasan_init_slab_obj(cachep, objp);
2593
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002594 /* constructor could break poison info */
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002595 if (DEBUG == 0 && cachep->ctor) {
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002596 kasan_unpoison_object_data(cachep, objp);
2597 cachep->ctor(objp);
2598 kasan_poison_object_data(cachep, objp);
2599 }
Joonsoo Kim10b2e9e2016-03-15 14:54:47 -07002600
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07002601 if (!shuffled)
2602 set_free_obj(page, i, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604}
2605
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002606static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002607{
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002608 void *objp;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002609
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002610 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
Joonsoo Kim8456a642013-10-24 10:07:49 +09002611 page->active++;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002612
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002613#if DEBUG
2614 if (cachep->flags & SLAB_STORE_USER)
2615 set_store_user_dirty(cachep);
2616#endif
2617
Matthew Dobson78d382d2006-02-01 03:05:47 -08002618 return objp;
2619}
2620
Joonsoo Kim260b61d2016-03-15 14:54:12 -07002621static void slab_put_obj(struct kmem_cache *cachep,
2622 struct page *page, void *objp)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002623{
Joonsoo Kim8456a642013-10-24 10:07:49 +09002624 unsigned int objnr = obj_to_index(cachep, page, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002625#if DEBUG
Joonsoo Kim16025172013-10-24 10:07:46 +09002626 unsigned int i;
Matthew Dobson78d382d2006-02-01 03:05:47 -08002627
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002628 /* Verify double free bug */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002629 for (i = page->active; i < cachep->num; i++) {
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002630 if (get_free_obj(page, i) == objnr) {
Joe Perches11705322016-03-17 14:19:50 -07002631 pr_err("slab: double free detected in cache '%s', objp %p\n",
Joe Perches756a0252016-03-17 14:19:47 -07002632 cachep->name, objp);
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09002633 BUG();
2634 }
Matthew Dobson78d382d2006-02-01 03:05:47 -08002635 }
2636#endif
Joonsoo Kim8456a642013-10-24 10:07:49 +09002637 page->active--;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002638 if (!page->freelist)
2639 page->freelist = objp + obj_offset(cachep);
2640
Joonsoo Kime5c58df2013-12-02 17:49:40 +09002641 set_free_obj(page, page->active, objnr);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002642}
2643
Pekka Enberg47768742006-06-23 02:03:07 -07002644/*
2645 * Map pages beginning at addr to the given cache and slab. This is required
2646 * for the slab allocator to be able to lookup the cache and slab of a
Nick Pigginccd35fb2011-01-07 17:49:17 +11002647 * virtual address for kfree, ksize, and slab debugging.
Pekka Enberg47768742006-06-23 02:03:07 -07002648 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002649static void slab_map_pages(struct kmem_cache *cache, struct page *page,
Joonsoo Kim7e007352013-10-30 19:04:01 +09002650 void *freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651{
Joonsoo Kima57a4982013-10-24 10:07:44 +09002652 page->slab_cache = cache;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002653 page->freelist = freelist;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654}
2655
2656/*
2657 * Grow (by 1) the number of slabs within a cache. This is called by
2658 * kmem_cache_alloc() when there are no active objs left in a cache.
2659 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002660static struct page *cache_grow_begin(struct kmem_cache *cachep,
2661 gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662{
Joonsoo Kim7e007352013-10-30 19:04:01 +09002663 void *freelist;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002664 size_t offset;
2665 gfp_t local_flags;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002666 int page_node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002667 struct kmem_cache_node *n;
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002668 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669
Andrew Mortona737b3e2006-03-22 00:08:11 -08002670 /*
2671 * Be lazy and only check for valid flags here, keeping it out of the
2672 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 */
Andrew Mortonc871ac42014-12-10 15:42:25 -08002674 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
Michal Hockobacdcb32016-07-26 15:22:02 -07002675 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
Michal Hocko72baeef0c2016-07-26 15:22:05 -07002676 flags &= ~GFP_SLAB_BUG_MASK;
2677 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2678 invalid_mask, &invalid_mask, flags, &flags);
2679 dump_stack();
Andrew Mortonc871ac42014-12-10 15:42:25 -08002680 }
Christoph Lameter6cb06222007-10-16 01:25:41 -07002681 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 check_irq_off();
Mel Gormand0164ad2015-11-06 16:28:21 -08002684 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 local_irq_enable();
2686
2687 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002688 * Get mem for the objs. Attempt to allocate a physical page from
2689 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002690 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002691 page = kmem_getpages(cachep, local_flags, nodeid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002692 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 goto failed;
2694
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002695 page_node = page_to_nid(page);
2696 n = get_node(cachep, page_node);
Joonsoo Kim03d1d432016-05-19 17:10:20 -07002697
2698 /* Get colour for the slab, and cal the next value. */
2699 n->colour_next++;
2700 if (n->colour_next >= cachep->colour)
2701 n->colour_next = 0;
2702
2703 offset = n->colour_next;
2704 if (offset >= cachep->colour)
2705 offset = 0;
2706
2707 offset *= cachep->colour_off;
2708
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 /* Get slab management. */
Joonsoo Kim8456a642013-10-24 10:07:49 +09002710 freelist = alloc_slabmgmt(cachep, page, offset,
Joonsoo Kim511e3a052016-05-19 17:10:23 -07002711 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002712 if (OFF_SLAB(cachep) && !freelist)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 goto opps1;
2714
Joonsoo Kim8456a642013-10-24 10:07:49 +09002715 slab_map_pages(cachep, page, freelist);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07002717 kasan_poison_slab(page);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002718 cache_init_objs(cachep, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719
Mel Gormand0164ad2015-11-06 16:28:21 -08002720 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721 local_irq_disable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002723 return page;
2724
Andrew Mortona737b3e2006-03-22 00:08:11 -08002725opps1:
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09002726 kmem_freepages(cachep, page);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002727failed:
Mel Gormand0164ad2015-11-06 16:28:21 -08002728 if (gfpflags_allow_blocking(local_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 local_irq_disable();
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002730 return NULL;
2731}
2732
2733static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2734{
2735 struct kmem_cache_node *n;
2736 void *list = NULL;
2737
2738 check_irq_off();
2739
2740 if (!page)
2741 return;
2742
2743 INIT_LIST_HEAD(&page->lru);
2744 n = get_node(cachep, page_to_nid(page));
2745
2746 spin_lock(&n->list_lock);
David Rientjesbf00bd32016-12-12 16:41:44 -08002747 n->total_slabs++;
Greg Thelenf728b0a2016-12-12 16:41:41 -08002748 if (!page->active) {
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002749 list_add_tail(&page->lru, &(n->slabs_free));
Greg Thelenf728b0a2016-12-12 16:41:41 -08002750 n->free_slabs++;
David Rientjesbf00bd32016-12-12 16:41:44 -08002751 } else
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002752 fixup_slab_list(cachep, n, page, &list);
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07002753
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002754 STATS_INC_GROWN(cachep);
2755 n->free_objects += cachep->num - page->active;
2756 spin_unlock(&n->list_lock);
2757
2758 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759}
2760
2761#if DEBUG
2762
2763/*
2764 * Perform extra freeing checks:
2765 * - detect bad pointers.
2766 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 */
2768static void kfree_debugcheck(const void *objp)
2769{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 if (!virt_addr_valid(objp)) {
Joe Perches11705322016-03-17 14:19:50 -07002771 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002772 (unsigned long)objp);
2773 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775}
2776
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002777static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2778{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002779 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002780
2781 redzone1 = *dbg_redzone1(cache, obj);
2782 redzone2 = *dbg_redzone2(cache, obj);
2783
2784 /*
2785 * Redzone is ok.
2786 */
2787 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2788 return;
2789
2790 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2791 slab_error(cache, "double free detected");
2792 else
2793 slab_error(cache, "memory outside object was overwritten");
2794
Joe Perches11705322016-03-17 14:19:50 -07002795 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2796 obj, redzone1, redzone2);
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002797}
2798
Pekka Enberg343e0d72006-02-01 03:05:50 -08002799static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002800 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 unsigned int objnr;
Joonsoo Kim8456a642013-10-24 10:07:49 +09002803 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002805 BUG_ON(virt_to_cache(objp) != cachep);
2806
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002807 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002809 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002812 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2814 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2815 }
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002816 if (cachep->flags & SLAB_STORE_USER) {
2817 set_store_user_dirty(cachep);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03002818 *dbg_userword(cachep, objp) = (void *)caller;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07002819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
Joonsoo Kim8456a642013-10-24 10:07:49 +09002821 objnr = obj_to_index(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822
2823 BUG_ON(objnr >= cachep->num);
Joonsoo Kim8456a642013-10-24 10:07:49 +09002824 BUG_ON(objp != index_to_obj(cachep, page, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 poison_obj(cachep, objp, POISON_FREE);
Joonsoo Kim40b44132016-03-15 14:54:21 -07002828 slab_kernel_map(cachep, objp, 0, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 }
2830 return objp;
2831}
2832
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833#else
2834#define kfree_debugcheck(x) do { } while(0)
2835#define cache_free_debugcheck(x,objp,z) (objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836#endif
2837
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002838static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2839 void **list)
2840{
2841#if DEBUG
2842 void *next = *list;
2843 void *objp;
2844
2845 while (next) {
2846 objp = next - obj_offset(cachep);
2847 next = *(void **)next;
2848 poison_obj(cachep, objp, POISON_FREE);
2849 }
2850#endif
2851}
2852
Joonsoo Kimd8410232016-03-15 14:54:44 -07002853static inline void fixup_slab_list(struct kmem_cache *cachep,
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002854 struct kmem_cache_node *n, struct page *page,
2855 void **list)
Joonsoo Kimd8410232016-03-15 14:54:44 -07002856{
2857 /* move slabp to correct slabp list: */
2858 list_del(&page->lru);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002859 if (page->active == cachep->num) {
Joonsoo Kimd8410232016-03-15 14:54:44 -07002860 list_add(&page->lru, &n->slabs_full);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002861 if (OBJFREELIST_SLAB(cachep)) {
2862#if DEBUG
2863 /* Poisoning will be done without holding the lock */
2864 if (cachep->flags & SLAB_POISON) {
2865 void **objp = page->freelist;
2866
2867 *objp = *list;
2868 *list = objp;
2869 }
2870#endif
2871 page->freelist = NULL;
2872 }
2873 } else
Joonsoo Kimd8410232016-03-15 14:54:44 -07002874 list_add(&page->lru, &n->slabs_partial);
2875}
2876
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002877/* Try to find non-pfmemalloc slab if needed */
2878static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
David Rientjesbf00bd32016-12-12 16:41:44 -08002879 struct page *page, bool pfmemalloc)
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002880{
2881 if (!page)
2882 return NULL;
2883
2884 if (pfmemalloc)
2885 return page;
2886
2887 if (!PageSlabPfmemalloc(page))
2888 return page;
2889
2890 /* No need to keep pfmemalloc slab if we have enough free objects */
2891 if (n->free_objects > n->free_limit) {
2892 ClearPageSlabPfmemalloc(page);
2893 return page;
2894 }
2895
2896 /* Move pfmemalloc slab to the end of list to speed up next search */
2897 list_del(&page->lru);
David Rientjesbf00bd32016-12-12 16:41:44 -08002898 if (!page->active) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002899 list_add_tail(&page->lru, &n->slabs_free);
David Rientjesbf00bd32016-12-12 16:41:44 -08002900 n->free_slabs++;
Greg Thelenf728b0a2016-12-12 16:41:41 -08002901 } else
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002902 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
Greg Thelenf728b0a2016-12-12 16:41:41 -08002909 n->free_touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002910 list_for_each_entry(page, &n->slabs_free, lru) {
Greg Thelenf728b0a2016-12-12 16:41:41 -08002911 if (!PageSlabPfmemalloc(page)) {
David Rientjesbf00bd32016-12-12 16:41:44 -08002912 n->free_slabs--;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002913 return page;
Greg Thelenf728b0a2016-12-12 16:41:41 -08002914 }
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002915 }
2916
2917 return NULL;
2918}
2919
2920static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
Geliang Tang7aa0d222016-01-14 15:18:02 -08002921{
2922 struct page *page;
2923
Greg Thelenf728b0a2016-12-12 16:41:41 -08002924 assert_spin_locked(&n->list_lock);
David Rientjesbf00bd32016-12-12 16:41:44 -08002925 page = list_first_entry_or_null(&n->slabs_partial, struct page, lru);
Geliang Tang7aa0d222016-01-14 15:18:02 -08002926 if (!page) {
2927 n->free_touched = 1;
David Rientjesbf00bd32016-12-12 16:41:44 -08002928 page = list_first_entry_or_null(&n->slabs_free, struct page,
2929 lru);
Greg Thelenf728b0a2016-12-12 16:41:41 -08002930 if (page)
David Rientjesbf00bd32016-12-12 16:41:44 -08002931 n->free_slabs--;
Geliang Tang7aa0d222016-01-14 15:18:02 -08002932 }
2933
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002934 if (sk_memalloc_socks())
David Rientjesbf00bd32016-12-12 16:41:44 -08002935 page = get_valid_first_slab(n, page, pfmemalloc);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002936
Geliang Tang7aa0d222016-01-14 15:18:02 -08002937 return page;
2938}
2939
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002940static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2941 struct kmem_cache_node *n, gfp_t flags)
2942{
2943 struct page *page;
2944 void *obj;
2945 void *list = NULL;
2946
2947 if (!gfp_pfmemalloc_allowed(flags))
2948 return NULL;
2949
2950 spin_lock(&n->list_lock);
2951 page = get_first_slab(n, true);
2952 if (!page) {
2953 spin_unlock(&n->list_lock);
2954 return NULL;
2955 }
2956
2957 obj = slab_get_obj(cachep, page);
2958 n->free_objects--;
2959
2960 fixup_slab_list(cachep, n, page, &list);
2961
2962 spin_unlock(&n->list_lock);
2963 fixup_objfreelist_debug(cachep, &list);
2964
2965 return obj;
2966}
2967
Joonsoo Kim213b4692016-05-19 17:10:29 -07002968/*
2969 * Slab list should be fixed up by fixup_slab_list() for existing slab
2970 * or cache_grow_end() for new slab
2971 */
2972static __always_inline int alloc_block(struct kmem_cache *cachep,
2973 struct array_cache *ac, struct page *page, int batchcount)
2974{
2975 /*
2976 * There must be at least one object available for
2977 * allocation.
2978 */
2979 BUG_ON(page->active >= cachep->num);
2980
2981 while (page->active < cachep->num && batchcount--) {
2982 STATS_INC_ALLOCED(cachep);
2983 STATS_INC_ACTIVE(cachep);
2984 STATS_SET_HIGH(cachep);
2985
2986 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2987 }
2988
2989 return batchcount;
2990}
2991
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07002992static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993{
2994 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00002995 struct kmem_cache_node *n;
Joonsoo Kim801faf02016-05-19 17:10:31 -07002996 struct array_cache *ac, *shared;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002997 int node;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07002998 void *list = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07002999 struct page *page;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003000
Joe Korty6d2144d2008-03-05 15:04:59 -08003001 check_irq_off();
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003002 node = numa_mem_id();
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003003
Joe Korty6d2144d2008-03-05 15:04:59 -08003004 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005 batchcount = ac->batchcount;
3006 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003007 /*
3008 * If there was little recent activity on this cache, then
3009 * perform only a partial refill. Otherwise we could generate
3010 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 */
3012 batchcount = BATCHREFILL_LIMIT;
3013 }
Christoph Lameter18bf8542014-08-06 16:04:11 -07003014 n = get_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003016 BUG_ON(ac->avail > 0 || !n);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003017 shared = READ_ONCE(n->shared);
3018 if (!n->free_objects && (!shared || !shared->avail))
3019 goto direct_grow;
3020
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003021 spin_lock(&n->list_lock);
Joonsoo Kim801faf02016-05-19 17:10:31 -07003022 shared = READ_ONCE(n->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003023
Christoph Lameter3ded1752006-03-25 03:06:44 -08003024 /* See if we can refill from the shared array */
Joonsoo Kim801faf02016-05-19 17:10:31 -07003025 if (shared && transfer_objects(ac, shared, batchcount)) {
3026 shared->touched = 1;
Christoph Lameter3ded1752006-03-25 03:06:44 -08003027 goto alloc_done;
Nick Piggin44b57f12010-01-27 22:27:40 +11003028 }
Christoph Lameter3ded1752006-03-25 03:06:44 -08003029
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 while (batchcount > 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 /* Get slab alloc is to come from. */
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003032 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003033 if (!page)
3034 goto must_grow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07003037
Joonsoo Kim213b4692016-05-19 17:10:29 -07003038 batchcount = alloc_block(cachep, ac, page, batchcount);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003039 fixup_slab_list(cachep, n, page, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 }
3041
Andrew Mortona737b3e2006-03-22 00:08:11 -08003042must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003043 n->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003044alloc_done:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003045 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003046 fixup_objfreelist_debug(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047
Joonsoo Kim801faf02016-05-19 17:10:31 -07003048direct_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049 if (unlikely(!ac->avail)) {
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003050 /* Check if we can use obj in pfmemalloc slab */
3051 if (sk_memalloc_socks()) {
3052 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3053
3054 if (obj)
3055 return obj;
3056 }
3057
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003058 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
Christoph Lametere498be72005-09-09 13:03:32 -07003059
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003060 /*
3061 * cache_grow_begin() can reenable interrupts,
3062 * then ac could change.
3063 */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003064 ac = cpu_cache_get(cachep);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003065 if (!ac->avail && page)
3066 alloc_block(cachep, ac, page, batchcount);
3067 cache_grow_end(cachep, page);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003068
Joonsoo Kim213b4692016-05-19 17:10:29 -07003069 if (!ac->avail)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 }
3072 ac->touched = 1;
Mel Gorman072bb0a2012-07-31 16:43:58 -07003073
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003074 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075}
3076
Andrew Mortona737b3e2006-03-22 00:08:11 -08003077static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3078 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079{
Mel Gormand0164ad2015-11-06 16:28:21 -08003080 might_sleep_if(gfpflags_allow_blocking(flags));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081}
3082
3083#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003084static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003085 gfp_t flags, void *objp, unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003087 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003089 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 check_poison_obj(cachep, objp);
Joonsoo Kim40b44132016-03-15 14:54:21 -07003091 slab_kernel_map(cachep, objp, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 poison_obj(cachep, objp, POISON_INUSE);
3093 }
3094 if (cachep->flags & SLAB_STORE_USER)
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003095 *dbg_userword(cachep, objp) = (void *)caller;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096
3097 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003098 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3099 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
Joe Perches756a0252016-03-17 14:19:47 -07003100 slab_error(cachep, "double free, or memory outside object was overwritten");
Joe Perches11705322016-03-17 14:19:50 -07003101 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3102 objp, *dbg_redzone1(cachep, objp),
3103 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 }
3105 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3106 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3107 }
Joonsoo Kim03787302014-06-23 13:22:06 -07003108
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003109 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003110 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003111 cachep->ctor(objp);
Tetsuo Handa7ea466f2011-07-21 09:42:45 +09003112 if (ARCH_SLAB_MINALIGN &&
3113 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
Joe Perches11705322016-03-17 14:19:50 -07003114 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
Hugh Dickinsc2251502011-07-11 13:35:08 -07003115 objp, (int)ARCH_SLAB_MINALIGN);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003116 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 return objp;
3118}
3119#else
3120#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3121#endif
3122
Pekka Enberg343e0d72006-02-01 03:05:50 -08003123static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003125 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 struct array_cache *ac;
3127
Alok N Kataria5c382302005-09-27 21:45:46 -07003128 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003129
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003130 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 if (likely(ac->avail)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132 ac->touched = 1;
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003133 objp = ac->entry[--ac->avail];
Mel Gorman072bb0a2012-07-31 16:43:58 -07003134
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003135 STATS_INC_ALLOCHIT(cachep);
3136 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 }
Mel Gorman072bb0a2012-07-31 16:43:58 -07003138
3139 STATS_INC_ALLOCMISS(cachep);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003140 objp = cache_alloc_refill(cachep, flags);
Mel Gorman072bb0a2012-07-31 16:43:58 -07003141 /*
3142 * the 'ac' may be updated by cache_alloc_refill(),
3143 * and kmemleak_erase() requires its correct value.
3144 */
3145 ac = cpu_cache_get(cachep);
3146
3147out:
Catalin Marinasd5cff632009-06-11 13:22:40 +01003148 /*
3149 * To avoid a false negative, if an object that is in one of the
3150 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3151 * treat the array pointers as a reference to the object.
3152 */
J. R. Okajimaf3d8b532009-12-02 16:55:49 +09003153 if (objp)
3154 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003155 return objp;
3156}
3157
Christoph Lametere498be72005-09-09 13:03:32 -07003158#ifdef CONFIG_NUMA
3159/*
Zefan Li2ad654b2014-09-25 09:41:02 +08003160 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003161 *
3162 * If we are in_interrupt, then process context, including cpusets and
3163 * mempolicy, may not apply and should not be used for allocation policy.
3164 */
3165static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3166{
3167 int nid_alloc, nid_here;
3168
Christoph Lameter765c4502006-09-27 01:50:08 -07003169 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003170 return NULL;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003171 nid_alloc = nid_here = numa_mem_id();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003172 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
Jack Steiner6adef3e2010-05-26 14:42:49 -07003173 nid_alloc = cpuset_slab_spread_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003174 else if (current->mempolicy)
David Rientjes2a389612014-04-07 15:37:29 -07003175 nid_alloc = mempolicy_slab_node();
Paul Jacksonc61afb12006-03-24 03:16:08 -08003176 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003177 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003178 return NULL;
3179}
3180
3181/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003182 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003183 * certain node and fall back is permitted. First we scan all the
Christoph Lameter6a673682013-01-10 19:14:19 +00003184 * available node for available objects. If that fails then we
Christoph Lameter3c517a62006-12-06 20:33:29 -08003185 * perform an allocation without specifying a node. This allows the page
3186 * allocator to do its reclaim / fallback magic. We then insert the
3187 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003188 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003189static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003190{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003191 struct zonelist *zonelist;
Mel Gormandd1a2392008-04-28 02:12:17 -07003192 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003193 struct zone *zone;
3194 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003195 void *obj = NULL;
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003196 struct page *page;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003197 int nid;
Mel Gormancc9a6c82012-03-21 16:34:11 -07003198 unsigned int cpuset_mems_cookie;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003199
3200 if (flags & __GFP_THISNODE)
3201 return NULL;
3202
Mel Gormancc9a6c82012-03-21 16:34:11 -07003203retry_cpuset:
Mel Gormand26914d2014-04-03 14:47:24 -07003204 cpuset_mems_cookie = read_mems_allowed_begin();
David Rientjes2a389612014-04-07 15:37:29 -07003205 zonelist = node_zonelist(mempolicy_slab_node(), flags);
Mel Gormancc9a6c82012-03-21 16:34:11 -07003206
Christoph Lameter3c517a62006-12-06 20:33:29 -08003207retry:
3208 /*
3209 * Look through allowed nodes for objects available
3210 * from existing per node queues.
3211 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003212 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3213 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003214
Vladimir Davydov061d7072014-12-12 16:58:25 -08003215 if (cpuset_zone_allowed(zone, flags) &&
Christoph Lameter18bf8542014-08-06 16:04:11 -07003216 get_node(cache, nid) &&
3217 get_node(cache, nid)->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003218 obj = ____cache_alloc_node(cache,
David Rientjes4167e9b2015-04-14 15:46:55 -07003219 gfp_exact_node(flags), nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003220 if (obj)
3221 break;
3222 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003223 }
3224
Christoph Lametercfce6602007-05-06 14:50:17 -07003225 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003226 /*
3227 * This allocation will be performed within the constraints
3228 * of the current cpuset / memory policy requirements.
3229 * We may trigger various forms of reclaim on the allowed
3230 * set and go into memory reserves if necessary.
3231 */
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003232 page = cache_grow_begin(cache, flags, numa_mem_id());
3233 cache_grow_end(cache, page);
3234 if (page) {
3235 nid = page_to_nid(page);
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003236 obj = ____cache_alloc_node(cache,
3237 gfp_exact_node(flags), nid);
Joonsoo Kim0c3aa832013-10-24 10:07:38 +09003238
Christoph Lameter3c517a62006-12-06 20:33:29 -08003239 /*
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003240 * Another processor may allocate the objects in
3241 * the slab since we are not holding any locks.
Christoph Lameter3c517a62006-12-06 20:33:29 -08003242 */
Joonsoo Kim511e3a052016-05-19 17:10:23 -07003243 if (!obj)
3244 goto retry;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003245 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003246 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07003247
Mel Gormand26914d2014-04-03 14:47:24 -07003248 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
Mel Gormancc9a6c82012-03-21 16:34:11 -07003249 goto retry_cpuset;
Christoph Lameter765c4502006-09-27 01:50:08 -07003250 return obj;
3251}
3252
3253/*
Christoph Lametere498be72005-09-09 13:03:32 -07003254 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003256static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003257 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003258{
Joonsoo Kim8456a642013-10-24 10:07:49 +09003259 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003260 struct kmem_cache_node *n;
Joonsoo Kim213b4692016-05-19 17:10:29 -07003261 void *obj = NULL;
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003262 void *list = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263
Paul Mackerras7c3fbbd2014-12-02 15:59:48 -08003264 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003265 n = get_node(cachep, nodeid);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003266 BUG_ON(!n);
Christoph Lametere498be72005-09-09 13:03:32 -07003267
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003268 check_irq_off();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003269 spin_lock(&n->list_lock);
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003270 page = get_first_slab(n, false);
Geliang Tang7aa0d222016-01-14 15:18:02 -08003271 if (!page)
3272 goto must_grow;
Christoph Lametere498be72005-09-09 13:03:32 -07003273
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003274 check_spinlock_acquired_node(cachep, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003275
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003276 STATS_INC_NODEALLOCS(cachep);
3277 STATS_INC_ACTIVE(cachep);
3278 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003279
Joonsoo Kim8456a642013-10-24 10:07:49 +09003280 BUG_ON(page->active == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003281
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003282 obj = slab_get_obj(cachep, page);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003283 n->free_objects--;
Christoph Lametere498be72005-09-09 13:03:32 -07003284
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003285 fixup_slab_list(cachep, n, page, &list);
Christoph Lametere498be72005-09-09 13:03:32 -07003286
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003287 spin_unlock(&n->list_lock);
Joonsoo Kimb03a017b2016-03-15 14:54:50 -07003288 fixup_objfreelist_debug(cachep, &list);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003289 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003290
Andrew Mortona737b3e2006-03-22 00:08:11 -08003291must_grow:
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003292 spin_unlock(&n->list_lock);
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003293 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
Joonsoo Kim213b4692016-05-19 17:10:29 -07003294 if (page) {
3295 /* This slab isn't counted yet so don't update free_objects */
3296 obj = slab_get_obj(cachep, page);
3297 }
Joonsoo Kim76b342b2016-05-19 17:10:26 -07003298 cache_grow_end(cachep, page);
Christoph Lametere498be72005-09-09 13:03:32 -07003299
Joonsoo Kim213b4692016-05-19 17:10:29 -07003300 return obj ? obj : fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003301}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003302
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003303static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003304slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003305 unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003306{
3307 unsigned long save_flags;
3308 void *ptr;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003309 int slab_node = numa_mem_id();
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003310
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003311 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003312 cachep = slab_pre_alloc_hook(cachep, flags);
3313 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003314 return NULL;
3315
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003316 cache_alloc_debugcheck_before(cachep, flags);
3317 local_irq_save(save_flags);
3318
Andrew Mortoneacbbae2011-07-28 13:59:49 -07003319 if (nodeid == NUMA_NO_NODE)
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003320 nodeid = slab_node;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003321
Christoph Lameter18bf8542014-08-06 16:04:11 -07003322 if (unlikely(!get_node(cachep, nodeid))) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003323 /* Node not bootstrapped yet */
3324 ptr = fallback_alloc(cachep, flags);
3325 goto out;
3326 }
3327
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003328 if (nodeid == slab_node) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003329 /*
3330 * Use the locally cached objects if possible.
3331 * However ____cache_alloc does not allow fallback
3332 * to other nodes. It may fail while we still have
3333 * objects on other nodes available.
3334 */
3335 ptr = ____cache_alloc(cachep, flags);
3336 if (ptr)
3337 goto out;
3338 }
3339 /* ___cache_alloc_node can fall back to other nodes */
3340 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3341 out:
3342 local_irq_restore(save_flags);
3343 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3344
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003345 if (unlikely(flags & __GFP_ZERO) && ptr)
3346 memset(ptr, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003347
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003348 slab_post_alloc_hook(cachep, flags, 1, &ptr);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003349 return ptr;
3350}
3351
3352static __always_inline void *
3353__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3354{
3355 void *objp;
3356
Zefan Li2ad654b2014-09-25 09:41:02 +08003357 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003358 objp = alternate_node_alloc(cache, flags);
3359 if (objp)
3360 goto out;
3361 }
3362 objp = ____cache_alloc(cache, flags);
3363
3364 /*
3365 * We may just have run out of memory on the local node.
3366 * ____cache_alloc_node() knows how to locate memory on other nodes
3367 */
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003368 if (!objp)
3369 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003370
3371 out:
3372 return objp;
3373}
3374#else
3375
3376static __always_inline void *
3377__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3378{
3379 return ____cache_alloc(cachep, flags);
3380}
3381
3382#endif /* CONFIG_NUMA */
3383
3384static __always_inline void *
Ezequiel Garcia48356302012-09-08 17:47:57 -03003385slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003386{
3387 unsigned long save_flags;
3388 void *objp;
3389
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003390 flags &= gfp_allowed_mask;
Jesper Dangaard Brouer011ecea2016-03-15 14:53:41 -07003391 cachep = slab_pre_alloc_hook(cachep, flags);
3392 if (unlikely(!cachep))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003393 return NULL;
3394
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003395 cache_alloc_debugcheck_before(cachep, flags);
3396 local_irq_save(save_flags);
3397 objp = __do_cache_alloc(cachep, flags);
3398 local_irq_restore(save_flags);
3399 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3400 prefetchw(objp);
3401
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003402 if (unlikely(flags & __GFP_ZERO) && objp)
3403 memset(objp, 0, cachep->object_size);
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003404
Jesper Dangaard Brouerd5e3ed62016-03-15 14:53:47 -07003405 slab_post_alloc_hook(cachep, flags, 1, &objp);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003406 return objp;
3407}
Christoph Lametere498be72005-09-09 13:03:32 -07003408
3409/*
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08003410 * Caller needs to acquire correct kmem_cache_node's list_lock
Joonsoo Kim97654df2014-08-06 16:04:25 -07003411 * @list: List of detached free slabs should be freed by caller
Christoph Lametere498be72005-09-09 13:03:32 -07003412 */
Joonsoo Kim97654df2014-08-06 16:04:25 -07003413static void free_block(struct kmem_cache *cachep, void **objpp,
3414 int nr_objects, int node, struct list_head *list)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415{
3416 int i;
Joonsoo Kim25c063f2014-08-06 16:04:22 -07003417 struct kmem_cache_node *n = get_node(cachep, node);
Joonsoo Kim6052b782016-05-19 17:10:17 -07003418 struct page *page;
3419
3420 n->free_objects += nr_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421
3422 for (i = 0; i < nr_objects; i++) {
Mel Gorman072bb0a2012-07-31 16:43:58 -07003423 void *objp;
Joonsoo Kim8456a642013-10-24 10:07:49 +09003424 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425
Mel Gorman072bb0a2012-07-31 16:43:58 -07003426 objp = objpp[i];
3427
Joonsoo Kim8456a642013-10-24 10:07:49 +09003428 page = virt_to_head_page(objp);
Joonsoo Kim8456a642013-10-24 10:07:49 +09003429 list_del(&page->lru);
Christoph Lameterff694162005-09-22 21:44:02 -07003430 check_spinlock_acquired_node(cachep, node);
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003431 slab_put_obj(cachep, page, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432 STATS_DEC_ACTIVE(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
3434 /* fixup slab chains */
Greg Thelenf728b0a2016-12-12 16:41:41 -08003435 if (page->active == 0) {
Joonsoo Kim6052b782016-05-19 17:10:17 -07003436 list_add(&page->lru, &n->slabs_free);
Greg Thelenf728b0a2016-12-12 16:41:41 -08003437 n->free_slabs++;
Greg Thelenf728b0a2016-12-12 16:41:41 -08003438 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003439 /* Unconditionally move a slab to the end of the
3440 * partial list on free - maximum time for the
3441 * other objects to be freed, too.
3442 */
Joonsoo Kim8456a642013-10-24 10:07:49 +09003443 list_add_tail(&page->lru, &n->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 }
3445 }
Joonsoo Kim6052b782016-05-19 17:10:17 -07003446
3447 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3448 n->free_objects -= cachep->num;
3449
3450 page = list_last_entry(&n->slabs_free, struct page, lru);
Wei Yongjunde24bae2016-07-26 15:22:11 -07003451 list_move(&page->lru, list);
Greg Thelenf728b0a2016-12-12 16:41:41 -08003452 n->free_slabs--;
David Rientjesbf00bd32016-12-12 16:41:44 -08003453 n->total_slabs--;
Joonsoo Kim6052b782016-05-19 17:10:17 -07003454 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455}
3456
Pekka Enberg343e0d72006-02-01 03:05:50 -08003457static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458{
3459 int batchcount;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003460 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07003461 int node = numa_mem_id();
Joonsoo Kim97654df2014-08-06 16:04:25 -07003462 LIST_HEAD(list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463
3464 batchcount = ac->batchcount;
Joonsoo Kim260b61d2016-03-15 14:54:12 -07003465
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 check_irq_off();
Christoph Lameter18bf8542014-08-06 16:04:11 -07003467 n = get_node(cachep, node);
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003468 spin_lock(&n->list_lock);
3469 if (n->shared) {
3470 struct array_cache *shared_array = n->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003471 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 if (max) {
3473 if (batchcount > max)
3474 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003475 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003476 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 shared_array->avail += batchcount;
3478 goto free_done;
3479 }
3480 }
3481
Joonsoo Kim97654df2014-08-06 16:04:25 -07003482 free_block(cachep, ac->entry, batchcount, node, &list);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003483free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484#if STATS
3485 {
3486 int i = 0;
Geliang Tang73c02192016-01-14 15:17:59 -08003487 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488
Geliang Tang73c02192016-01-14 15:17:59 -08003489 list_for_each_entry(page, &n->slabs_free, lru) {
Joonsoo Kim8456a642013-10-24 10:07:49 +09003490 BUG_ON(page->active);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491
3492 i++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 }
3494 STATS_SET_FREEABLE(cachep, i);
3495 }
3496#endif
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003497 spin_unlock(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003498 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003500 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501}
3502
3503/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003504 * Release an obj back to its cache. If the obj has a constructed state, it must
3505 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506 */
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003507static inline void __cache_free(struct kmem_cache *cachep, void *objp,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003508 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509{
Alexander Potapenko55834c52016-05-20 16:59:11 -07003510 /* Put the object into the quarantine, don't touch it for now. */
3511 if (kasan_slab_free(cachep, objp))
3512 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513
Alexander Potapenko55834c52016-05-20 16:59:11 -07003514 ___cache_free(cachep, objp, caller);
3515}
3516
3517void ___cache_free(struct kmem_cache *cachep, void *objp,
3518 unsigned long caller)
3519{
3520 struct array_cache *ac = cpu_cache_get(cachep);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003521
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003523 kmemleak_free_recursive(objp, cachep->flags);
Suleiman Souhlala947eb92011-06-02 00:16:42 -07003524 objp = cache_free_debugcheck(cachep, objp, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003526 kmemcheck_slab_free(cachep, objp, cachep->object_size);
Pekka Enbergc175eea2008-05-09 20:35:53 +02003527
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003528 /*
3529 * Skip calling cache_free_alien() when the platform is not numa.
3530 * This will avoid cache misses that happen while accessing slabp (which
3531 * is per page memory reference) to get nodeid. Instead use a global
3532 * variable to skip the call, which is mostly likely to be present in
3533 * the cache.
3534 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003535 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003536 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003537
Joonsoo Kim3d880192014-10-09 15:26:04 -07003538 if (ac->avail < ac->limit) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 STATS_INC_FREEHIT(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 } else {
3541 STATS_INC_FREEMISS(cachep);
3542 cache_flusharray(cachep, ac);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543 }
Zhao Jin42c8c992011-08-27 00:26:17 +08003544
Joonsoo Kimf68f8dd2016-03-15 14:54:56 -07003545 if (sk_memalloc_socks()) {
3546 struct page *page = virt_to_head_page(objp);
3547
3548 if (unlikely(PageSlabPfmemalloc(page))) {
3549 cache_free_pfmemalloc(cachep, page, objp);
3550 return;
3551 }
3552 }
3553
3554 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555}
3556
3557/**
3558 * kmem_cache_alloc - Allocate an object
3559 * @cachep: The cache to allocate from.
3560 * @flags: See kmalloc().
3561 *
3562 * Allocate an object from this cache. The flags are only relevant
3563 * if the cache has no available objects.
3564 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003565void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003567 void *ret = slab_alloc(cachep, flags, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003568
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003569 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003570 trace_kmem_cache_alloc(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003571 cachep->object_size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003572
3573 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574}
3575EXPORT_SYMBOL(kmem_cache_alloc);
3576
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003577static __always_inline void
3578cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3579 size_t size, void **p, unsigned long caller)
3580{
3581 size_t i;
3582
3583 for (i = 0; i < size; i++)
3584 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3585}
3586
Jesper Dangaard Brouer865762a2015-11-20 15:57:58 -08003587int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003588 void **p)
Christoph Lameter484748f2015-09-04 15:45:34 -07003589{
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003590 size_t i;
3591
3592 s = slab_pre_alloc_hook(s, flags);
3593 if (!s)
3594 return 0;
3595
3596 cache_alloc_debugcheck_before(s, flags);
3597
3598 local_irq_disable();
3599 for (i = 0; i < size; i++) {
3600 void *objp = __do_cache_alloc(s, flags);
3601
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003602 if (unlikely(!objp))
3603 goto error;
3604 p[i] = objp;
3605 }
3606 local_irq_enable();
3607
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003608 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3609
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003610 /* Clear memory outside IRQ disabled section */
3611 if (unlikely(flags & __GFP_ZERO))
3612 for (i = 0; i < size; i++)
3613 memset(p[i], 0, s->object_size);
3614
3615 slab_post_alloc_hook(s, flags, size, p);
3616 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3617 return size;
3618error:
3619 local_irq_enable();
Jesper Dangaard Brouer7b0501d2016-03-15 14:53:53 -07003620 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
Jesper Dangaard Brouer2a777ea2016-03-15 14:53:50 -07003621 slab_post_alloc_hook(s, flags, i, p);
3622 __kmem_cache_free_bulk(s, i, p);
3623 return 0;
Christoph Lameter484748f2015-09-04 15:45:34 -07003624}
3625EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3626
Li Zefan0f24f122009-12-11 15:45:30 +08003627#ifdef CONFIG_TRACING
Steven Rostedt85beb582010-11-24 16:23:34 -05003628void *
Ezequiel Garcia40521472012-09-08 17:47:56 -03003629kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003630{
Steven Rostedt85beb582010-11-24 16:23:34 -05003631 void *ret;
3632
Ezequiel Garcia48356302012-09-08 17:47:57 -03003633 ret = slab_alloc(cachep, flags, _RET_IP_);
Steven Rostedt85beb582010-11-24 16:23:34 -05003634
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003635 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003636 trace_kmalloc(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003637 size, cachep->size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003638 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003639}
Steven Rostedt85beb582010-11-24 16:23:34 -05003640EXPORT_SYMBOL(kmem_cache_alloc_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003641#endif
3642
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643#ifdef CONFIG_NUMA
Zhouping Liud0d04b72013-05-16 11:36:23 +08003644/**
3645 * kmem_cache_alloc_node - Allocate an object on the specified node
3646 * @cachep: The cache to allocate from.
3647 * @flags: See kmalloc().
3648 * @nodeid: node number of the target node.
3649 *
3650 * Identical to kmem_cache_alloc but it will allocate memory on the given
3651 * node, which can improve the performance for cpu bound structures.
3652 *
3653 * Fallback to other node is possible if __GFP_THISNODE is not set.
3654 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003655void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3656{
Ezequiel Garcia48356302012-09-08 17:47:57 -03003657 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003658
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003659 kasan_slab_alloc(cachep, ret, flags);
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003660 trace_kmem_cache_alloc_node(_RET_IP_, ret,
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003661 cachep->object_size, cachep->size,
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003662 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003663
3664 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003665}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666EXPORT_SYMBOL(kmem_cache_alloc_node);
3667
Li Zefan0f24f122009-12-11 15:45:30 +08003668#ifdef CONFIG_TRACING
Ezequiel Garcia40521472012-09-08 17:47:56 -03003669void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
Steven Rostedt85beb582010-11-24 16:23:34 -05003670 gfp_t flags,
Ezequiel Garcia40521472012-09-08 17:47:56 -03003671 int nodeid,
3672 size_t size)
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003673{
Steven Rostedt85beb582010-11-24 16:23:34 -05003674 void *ret;
3675
Ezequiel Garcia592f4142012-09-25 08:07:08 -03003676 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003677
3678 kasan_kmalloc(cachep, ret, size, flags);
Steven Rostedt85beb582010-11-24 16:23:34 -05003679 trace_kmalloc_node(_RET_IP_, ret,
Ezequiel Garciaff4fcd02012-09-08 17:47:52 -03003680 size, cachep->size,
Steven Rostedt85beb582010-11-24 16:23:34 -05003681 flags, nodeid);
3682 return ret;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003683}
Steven Rostedt85beb582010-11-24 16:23:34 -05003684EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003685#endif
3686
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003687static __always_inline void *
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003688__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003689{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003690 struct kmem_cache *cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003691 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003692
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003693 cachep = kmalloc_slab(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003694 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3695 return cachep;
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003696 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003697 kasan_kmalloc(cachep, ret, size, flags);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07003698
3699 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003700}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003701
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003702void *__kmalloc_node(size_t size, gfp_t flags, int node)
3703{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003704 return __do_kmalloc_node(size, flags, node, _RET_IP_);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003705}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003706EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003707
3708void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003709 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003710{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003711 return __do_kmalloc_node(size, flags, node, caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003712}
3713EXPORT_SYMBOL(__kmalloc_node_track_caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003714#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715
3716/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003717 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003719 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003720 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003721 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003722static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003723 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003725 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003726 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727
Christoph Lameter2c59dd62013-01-10 19:14:19 +00003728 cachep = kmalloc_slab(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003729 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3730 return cachep;
Ezequiel Garcia48356302012-09-08 17:47:57 -03003731 ret = slab_alloc(cachep, flags, caller);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003732
Alexander Potapenko505f5dc2016-03-25 14:22:02 -07003733 kasan_kmalloc(cachep, ret, size, flags);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003734 trace_kmalloc(caller, ret,
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003735 size, cachep->size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003736
3737 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003738}
3739
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003740void *__kmalloc(size_t size, gfp_t flags)
3741{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003742 return __do_kmalloc(size, flags, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743}
3744EXPORT_SYMBOL(__kmalloc);
3745
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003746void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003747{
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003748 return __do_kmalloc(size, flags, caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003749}
3750EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003751
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752/**
3753 * kmem_cache_free - Deallocate an object
3754 * @cachep: The cache the allocation was from.
3755 * @objp: The previously allocated object.
3756 *
3757 * Free an object which was previously allocated from this
3758 * cache.
3759 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003760void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761{
3762 unsigned long flags;
Glauber Costab9ce5ef2012-12-18 14:22:46 -08003763 cachep = cache_from_obj(cachep, objp);
3764 if (!cachep)
3765 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766
3767 local_irq_save(flags);
Feng Tangd97d4762012-07-02 14:29:10 +08003768 debug_check_no_locks_freed(objp, cachep->object_size);
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003769 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003770 debug_check_no_obj_freed(objp, cachep->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003771 __cache_free(cachep, objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003773
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003774 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775}
3776EXPORT_SYMBOL(kmem_cache_free);
3777
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003778void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3779{
3780 struct kmem_cache *s;
3781 size_t i;
3782
3783 local_irq_disable();
3784 for (i = 0; i < size; i++) {
3785 void *objp = p[i];
3786
Jesper Dangaard Brouerca257192016-03-15 14:54:00 -07003787 if (!orig_s) /* called via kfree_bulk */
3788 s = virt_to_cache(objp);
3789 else
3790 s = cache_from_obj(orig_s, objp);
Jesper Dangaard Brouere6cdb582016-03-15 14:53:56 -07003791
3792 debug_check_no_locks_freed(objp, s->object_size);
3793 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3794 debug_check_no_obj_freed(objp, s->object_size);
3795
3796 __cache_free(s, objp, _RET_IP_);
3797 }
3798 local_irq_enable();
3799
3800 /* FIXME: add tracing */
3801}
3802EXPORT_SYMBOL(kmem_cache_free_bulk);
3803
Linus Torvalds1da177e2005-04-16 15:20:36 -07003804/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805 * kfree - free previously allocated memory
3806 * @objp: pointer returned by kmalloc.
3807 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003808 * If @objp is NULL, no operation is performed.
3809 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810 * Don't free memory not originally allocated by kmalloc()
3811 * or you will run into trouble.
3812 */
3813void kfree(const void *objp)
3814{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003815 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816 unsigned long flags;
3817
Pekka Enberg2121db72009-03-25 11:05:57 +02003818 trace_kfree(_RET_IP_, objp);
3819
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003820 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821 return;
3822 local_irq_save(flags);
3823 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003824 c = virt_to_cache(objp);
Christoph Lameter8c138bc2012-06-13 10:24:58 -05003825 debug_check_no_locks_freed(objp, c->object_size);
3826
3827 debug_check_no_obj_freed(objp, c->object_size);
Ezequiel Garcia7c0cb9c2012-09-08 17:47:55 -03003828 __cache_free(c, (void *)objp, _RET_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 local_irq_restore(flags);
3830}
3831EXPORT_SYMBOL(kfree);
3832
Christoph Lametere498be72005-09-09 13:03:32 -07003833/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003834 * This initializes kmem_cache_node or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003835 */
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003836static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003837{
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003838 int ret;
Christoph Lametere498be72005-09-09 13:03:32 -07003839 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003840 struct kmem_cache_node *n;
Christoph Lametere498be72005-09-09 13:03:32 -07003841
Mel Gorman9c09a952008-01-24 05:49:54 -08003842 for_each_online_node(node) {
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003843 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3844 if (ret)
Christoph Lametere498be72005-09-09 13:03:32 -07003845 goto fail;
3846
Christoph Lametere498be72005-09-09 13:03:32 -07003847 }
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003848
Christoph Lametercafeb022006-03-25 03:06:46 -08003849 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003850
Andrew Mortona737b3e2006-03-22 00:08:11 -08003851fail:
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003852 if (!cachep->list.next) {
Christoph Lameter0718dc22006-03-25 03:06:47 -08003853 /* Cache is not active yet. Roll back what we did */
3854 node--;
3855 while (node >= 0) {
Christoph Lameter18bf8542014-08-06 16:04:11 -07003856 n = get_node(cachep, node);
3857 if (n) {
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00003858 kfree(n->shared);
3859 free_alien_cache(n->alien);
3860 kfree(n);
Christoph Lameter6a673682013-01-10 19:14:19 +00003861 cachep->node[node] = NULL;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003862 }
3863 node--;
3864 }
3865 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003866 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003867}
3868
Christoph Lameter18004c52012-07-06 15:25:12 -05003869/* Always called with the slab_mutex held */
Glauber Costa943a4512012-12-18 14:23:03 -08003870static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003871 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872{
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003873 struct array_cache __percpu *cpu_cache, *prev;
3874 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003876 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3877 if (!cpu_cache)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003878 return -ENOMEM;
3879
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003880 prev = cachep->cpu_cache;
3881 cachep->cpu_cache = cpu_cache;
Greg Thelena87c75f2017-05-03 14:51:47 -07003882 /*
3883 * Without a previous cpu_cache there's no need to synchronize remote
3884 * cpus, so skip the IPIs.
3885 */
3886 if (prev)
3887 kick_all_cpus_sync();
Christoph Lametere498be72005-09-09 13:03:32 -07003888
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890 cachep->batchcount = batchcount;
3891 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003892 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003894 if (!prev)
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003895 goto setup_node;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003896
3897 for_each_online_cpu(cpu) {
Joonsoo Kim97654df2014-08-06 16:04:25 -07003898 LIST_HEAD(list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003899 int node;
3900 struct kmem_cache_node *n;
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003901 struct array_cache *ac = per_cpu_ptr(prev, cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003902
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003903 node = cpu_to_mem(cpu);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003904 n = get_node(cachep, node);
3905 spin_lock_irq(&n->list_lock);
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003906 free_block(cachep, ac->entry, ac->avail, node, &list);
Christoph Lameter18bf8542014-08-06 16:04:11 -07003907 spin_unlock_irq(&n->list_lock);
Joonsoo Kim97654df2014-08-06 16:04:25 -07003908 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 }
Joonsoo Kimbf0dea22014-10-09 15:26:27 -07003910 free_percpu(prev);
3911
Joonsoo Kimc3d332b2016-05-19 17:10:14 -07003912setup_node:
3913 return setup_kmem_cache_nodes(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914}
3915
Glauber Costa943a4512012-12-18 14:23:03 -08003916static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3917 int batchcount, int shared, gfp_t gfp)
3918{
3919 int ret;
Vladimir Davydov426589f2015-02-12 14:59:23 -08003920 struct kmem_cache *c;
Glauber Costa943a4512012-12-18 14:23:03 -08003921
3922 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3923
3924 if (slab_state < FULL)
3925 return ret;
3926
3927 if ((ret < 0) || !is_root_cache(cachep))
3928 return ret;
3929
Vladimir Davydov426589f2015-02-12 14:59:23 -08003930 lockdep_assert_held(&slab_mutex);
3931 for_each_memcg_cache(c, cachep) {
3932 /* return value determined by the root cache only */
3933 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
Glauber Costa943a4512012-12-18 14:23:03 -08003934 }
3935
3936 return ret;
3937}
3938
Christoph Lameter18004c52012-07-06 15:25:12 -05003939/* Called with slab_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003940static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941{
3942 int err;
Glauber Costa943a4512012-12-18 14:23:03 -08003943 int limit = 0;
3944 int shared = 0;
3945 int batchcount = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946
Thomas Garnier7c00fce2016-07-26 15:21:56 -07003947 err = cache_random_seq_create(cachep, cachep->num, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07003948 if (err)
3949 goto end;
3950
Glauber Costa943a4512012-12-18 14:23:03 -08003951 if (!is_root_cache(cachep)) {
3952 struct kmem_cache *root = memcg_root_cache(cachep);
3953 limit = root->limit;
3954 shared = root->shared;
3955 batchcount = root->batchcount;
3956 }
3957
3958 if (limit && shared && batchcount)
3959 goto skip_setup;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003960 /*
3961 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962 * - create a LIFO ordering, i.e. return objects that are cache-warm
3963 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003964 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 * bufctl chains: array operations are cheaper.
3966 * The numbers are guessed, we should auto-tune as described by
3967 * Bonwick.
3968 */
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003969 if (cachep->size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 limit = 1;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003971 else if (cachep->size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 limit = 8;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003973 else if (cachep->size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 limit = 24;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003975 else if (cachep->size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 limit = 54;
3977 else
3978 limit = 120;
3979
Andrew Mortona737b3e2006-03-22 00:08:11 -08003980 /*
3981 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 * allocation behaviour: Most allocs on one cpu, most free operations
3983 * on another cpu. For these cases, an efficient object passing between
3984 * cpus is necessary. This is provided by a shared array. The array
3985 * replaces Bonwick's magazine layer.
3986 * On uniprocessor, it's functionally equivalent (but less efficient)
3987 * to a larger limit. Thus disabled by default.
3988 */
3989 shared = 0;
Christoph Lameter3b0efdf2012-06-13 10:24:57 -05003990 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003991 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992
3993#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003994 /*
3995 * With debugging enabled, large batchcount lead to excessively long
3996 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 */
3998 if (limit > 32)
3999 limit = 32;
4000#endif
Glauber Costa943a4512012-12-18 14:23:03 -08004001 batchcount = (limit + 1) / 2;
4002skip_setup:
4003 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
Thomas Garnierc7ce4f602016-05-19 17:10:37 -07004004end:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 if (err)
Joe Perches11705322016-03-17 14:19:50 -07004006 pr_err("enable_cpucache failed for %s, error %d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004007 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004008 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009}
4010
Christoph Lameter1b552532006-03-22 00:09:07 -08004011/*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004012 * Drain an array if it contains any elements taking the node lock only if
4013 * necessary. Note that the node listlock also protects the array_cache
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004014 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004015 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004016static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004017 struct array_cache *ac, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018{
Joonsoo Kim97654df2014-08-06 16:04:25 -07004019 LIST_HEAD(list);
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004020
4021 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4022 check_mutex_acquired();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023
Christoph Lameter1b552532006-03-22 00:09:07 -08004024 if (!ac || !ac->avail)
4025 return;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004026
4027 if (ac->touched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028 ac->touched = 0;
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004029 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 }
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004031
4032 spin_lock_irq(&n->list_lock);
4033 drain_array_locked(cachep, ac, node, false, &list);
4034 spin_unlock_irq(&n->list_lock);
4035
4036 slabs_destroy(cachep, &list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037}
4038
4039/**
4040 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004041 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 *
4043 * Called from workqueue/eventd every few seconds.
4044 * Purpose:
4045 * - clear the per-cpu caches for this CPU.
4046 * - return freeable pages to the main free memory pool.
4047 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004048 * If we cannot acquire the cache chain mutex then just give up - we'll try
4049 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004051static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004053 struct kmem_cache *searchp;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004054 struct kmem_cache_node *n;
Lee Schermerhorn7d6e6d02010-05-26 14:45:03 -07004055 int node = numa_mem_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004056 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057
Christoph Lameter18004c52012-07-06 15:25:12 -05004058 if (!mutex_trylock(&slab_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004060 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004061
Christoph Lameter18004c52012-07-06 15:25:12 -05004062 list_for_each_entry(searchp, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 check_irq_on();
4064
Christoph Lameter35386e32006-03-22 00:09:05 -08004065 /*
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004066 * We only take the node lock if absolutely necessary and we
Christoph Lameter35386e32006-03-22 00:09:05 -08004067 * have established with reasonable certainty that
4068 * we can do some work if the lock was obtained.
4069 */
Christoph Lameter18bf8542014-08-06 16:04:11 -07004070 n = get_node(searchp, node);
Christoph Lameter35386e32006-03-22 00:09:05 -08004071
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004072 reap_alien(searchp, n);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004074 drain_array(searchp, n, cpu_cache_get(searchp), node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075
Christoph Lameter35386e32006-03-22 00:09:05 -08004076 /*
4077 * These are racy checks but it does not matter
4078 * if we skip one check or scan twice.
4079 */
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004080 if (time_after(n->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004081 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004083 n->next_reap = jiffies + REAPTIMEOUT_NODE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084
Joonsoo Kim18726ca2016-05-19 17:10:02 -07004085 drain_array(searchp, n, n->shared, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004087 if (n->free_touched)
4088 n->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004089 else {
4090 int freed;
4091
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004092 freed = drain_freelist(searchp, n, (n->free_limit +
Christoph Lametered11d9e2006-06-30 01:55:45 -07004093 5 * searchp->num - 1) / (5 * searchp->num));
4094 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004096next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 cond_resched();
4098 }
4099 check_irq_on();
Christoph Lameter18004c52012-07-06 15:25:12 -05004100 mutex_unlock(&slab_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004101 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004102out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004103 /* Set up the next iteration */
Jianyu Zhan5f0985b2014-03-30 17:02:20 +08004104 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105}
4106
Linus Torvalds158a9622008-01-02 13:04:48 -08004107#ifdef CONFIG_SLABINFO
Glauber Costa0d7561c2012-10-19 18:20:27 +04004108void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109{
Greg Thelenf728b0a2016-12-12 16:41:41 -08004110 unsigned long active_objs, num_objs, active_slabs;
David Rientjesbf00bd32016-12-12 16:41:44 -08004111 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4112 unsigned long free_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004113 int node;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004114 struct kmem_cache_node *n;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115
Christoph Lameter18bf8542014-08-06 16:04:11 -07004116 for_each_kmem_cache_node(cachep, node, n) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004117 check_irq_on();
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004118 spin_lock_irq(&n->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004119
David Rientjesbf00bd32016-12-12 16:41:44 -08004120 total_slabs += n->total_slabs;
4121 free_slabs += n->free_slabs;
Greg Thelenf728b0a2016-12-12 16:41:41 -08004122 free_objs += n->free_objects;
Aruna Ramakrishna07a63c42016-10-27 17:46:32 -07004123
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004124 if (n->shared)
4125 shared_avail += n->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004126
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004127 spin_unlock_irq(&n->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004128 }
David Rientjesbf00bd32016-12-12 16:41:44 -08004129 num_objs = total_slabs * cachep->num;
4130 active_slabs = total_slabs - free_slabs;
Greg Thelenf728b0a2016-12-12 16:41:41 -08004131 active_objs = num_objs - free_objs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132
Glauber Costa0d7561c2012-10-19 18:20:27 +04004133 sinfo->active_objs = active_objs;
4134 sinfo->num_objs = num_objs;
4135 sinfo->active_slabs = active_slabs;
David Rientjesbf00bd32016-12-12 16:41:44 -08004136 sinfo->num_slabs = total_slabs;
Glauber Costa0d7561c2012-10-19 18:20:27 +04004137 sinfo->shared_avail = shared_avail;
4138 sinfo->limit = cachep->limit;
4139 sinfo->batchcount = cachep->batchcount;
4140 sinfo->shared = cachep->shared;
4141 sinfo->objects_per_slab = cachep->num;
4142 sinfo->cache_order = cachep->gfporder;
4143}
4144
4145void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4146{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147#if STATS
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004148 { /* node stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149 unsigned long high = cachep->high_mark;
4150 unsigned long allocs = cachep->num_allocations;
4151 unsigned long grown = cachep->grown;
4152 unsigned long reaped = cachep->reaped;
4153 unsigned long errors = cachep->errors;
4154 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004156 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004157 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158
Joe Perches756a0252016-03-17 14:19:47 -07004159 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
Joe Perchese92dd4f2010-03-26 19:27:58 -07004160 allocs, high, grown,
4161 reaped, errors, max_freeable, node_allocs,
4162 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 }
4164 /* cpu stats */
4165 {
4166 unsigned long allochit = atomic_read(&cachep->allochit);
4167 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4168 unsigned long freehit = atomic_read(&cachep->freehit);
4169 unsigned long freemiss = atomic_read(&cachep->freemiss);
4170
4171 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004172 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 }
4174#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175}
4176
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177#define MAX_SLABINFO_WRITE 128
4178/**
4179 * slabinfo_write - Tuning for the slab allocator
4180 * @file: unused
4181 * @buffer: user buffer
4182 * @count: data length
4183 * @ppos: unused
4184 */
Glauber Costab7454ad2012-10-19 18:20:25 +04004185ssize_t slabinfo_write(struct file *file, const char __user *buffer,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004186 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004188 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004190 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004191
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 if (count > MAX_SLABINFO_WRITE)
4193 return -EINVAL;
4194 if (copy_from_user(&kbuf, buffer, count))
4195 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004196 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197
4198 tmp = strchr(kbuf, ' ');
4199 if (!tmp)
4200 return -EINVAL;
4201 *tmp = '\0';
4202 tmp++;
4203 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4204 return -EINVAL;
4205
4206 /* Find the cache in the chain of caches. */
Christoph Lameter18004c52012-07-06 15:25:12 -05004207 mutex_lock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 res = -EINVAL;
Christoph Lameter18004c52012-07-06 15:25:12 -05004209 list_for_each_entry(cachep, &slab_caches, list) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004211 if (limit < 1 || batchcount < 1 ||
4212 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004213 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004215 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004216 batchcount, shared,
4217 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 }
4219 break;
4220 }
4221 }
Christoph Lameter18004c52012-07-06 15:25:12 -05004222 mutex_unlock(&slab_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 if (res >= 0)
4224 res = count;
4225 return res;
4226}
Al Viro871751e2006-03-25 03:06:39 -08004227
4228#ifdef CONFIG_DEBUG_SLAB_LEAK
4229
Al Viro871751e2006-03-25 03:06:39 -08004230static inline int add_caller(unsigned long *n, unsigned long v)
4231{
4232 unsigned long *p;
4233 int l;
4234 if (!v)
4235 return 1;
4236 l = n[1];
4237 p = n + 2;
4238 while (l) {
4239 int i = l/2;
4240 unsigned long *q = p + 2 * i;
4241 if (*q == v) {
4242 q[1]++;
4243 return 1;
4244 }
4245 if (*q > v) {
4246 l = i;
4247 } else {
4248 p = q + 2;
4249 l -= i + 1;
4250 }
4251 }
4252 if (++n[1] == n[0])
4253 return 0;
4254 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4255 p[0] = v;
4256 p[1] = 1;
4257 return 1;
4258}
4259
Joonsoo Kim8456a642013-10-24 10:07:49 +09004260static void handle_slab(unsigned long *n, struct kmem_cache *c,
4261 struct page *page)
Al Viro871751e2006-03-25 03:06:39 -08004262{
4263 void *p;
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004264 int i, j;
4265 unsigned long v;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004266
Al Viro871751e2006-03-25 03:06:39 -08004267 if (n[0] == n[1])
4268 return;
Joonsoo Kim8456a642013-10-24 10:07:49 +09004269 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004270 bool active = true;
4271
4272 for (j = page->active; j < c->num; j++) {
4273 if (get_free_obj(page, j) == i) {
4274 active = false;
4275 break;
4276 }
4277 }
4278
4279 if (!active)
Al Viro871751e2006-03-25 03:06:39 -08004280 continue;
Joonsoo Kimb1cb0982013-10-24 10:07:45 +09004281
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004282 /*
4283 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4284 * mapping is established when actual object allocation and
4285 * we could mistakenly access the unmapped object in the cpu
4286 * cache.
4287 */
4288 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4289 continue;
4290
4291 if (!add_caller(n, v))
Al Viro871751e2006-03-25 03:06:39 -08004292 return;
4293 }
4294}
4295
4296static void show_symbol(struct seq_file *m, unsigned long address)
4297{
4298#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004299 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004300 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004301
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004302 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004303 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004304 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004305 seq_printf(m, " [%s]", modname);
4306 return;
4307 }
4308#endif
4309 seq_printf(m, "%p", (void *)address);
4310}
4311
4312static int leaks_show(struct seq_file *m, void *p)
4313{
Thierry Reding0672aa72012-06-22 19:42:49 +02004314 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
Joonsoo Kim8456a642013-10-24 10:07:49 +09004315 struct page *page;
Christoph Lameterce8eb6c2013-01-10 19:14:19 +00004316 struct kmem_cache_node *n;
Al Viro871751e2006-03-25 03:06:39 -08004317 const char *name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004318 unsigned long *x = m->private;
Al Viro871751e2006-03-25 03:06:39 -08004319 int node;
4320 int i;
4321
4322 if (!(cachep->flags & SLAB_STORE_USER))
4323 return 0;
4324 if (!(cachep->flags & SLAB_RED_ZONE))
4325 return 0;
4326
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004327 /*
4328 * Set store_user_clean and start to grab stored user information
4329 * for all objects on this cache. If some alloc/free requests comes
4330 * during the processing, information would be wrong so restart
4331 * whole processing.
4332 */
4333 do {
4334 set_store_user_clean(cachep);
4335 drain_cpu_caches(cachep);
Al Viro871751e2006-03-25 03:06:39 -08004336
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004337 x[1] = 0;
Al Viro871751e2006-03-25 03:06:39 -08004338
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004339 for_each_kmem_cache_node(cachep, node, n) {
Al Viro871751e2006-03-25 03:06:39 -08004340
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004341 check_irq_on();
4342 spin_lock_irq(&n->list_lock);
Al Viro871751e2006-03-25 03:06:39 -08004343
Joonsoo Kimd31676d2016-03-15 14:54:24 -07004344 list_for_each_entry(page, &n->slabs_full, lru)
4345 handle_slab(x, cachep, page);
4346 list_for_each_entry(page, &n->slabs_partial, lru)
4347 handle_slab(x, cachep, page);
4348 spin_unlock_irq(&n->list_lock);
4349 }
4350 } while (!is_store_user_clean(cachep));
4351
Al Viro871751e2006-03-25 03:06:39 -08004352 name = cachep->name;
Christoph Lameterdb845062013-02-05 18:45:23 +00004353 if (x[0] == x[1]) {
Al Viro871751e2006-03-25 03:06:39 -08004354 /* Increase the buffer size */
Christoph Lameter18004c52012-07-06 15:25:12 -05004355 mutex_unlock(&slab_mutex);
Christoph Lameterdb845062013-02-05 18:45:23 +00004356 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
Al Viro871751e2006-03-25 03:06:39 -08004357 if (!m->private) {
4358 /* Too bad, we are really out */
Christoph Lameterdb845062013-02-05 18:45:23 +00004359 m->private = x;
Christoph Lameter18004c52012-07-06 15:25:12 -05004360 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004361 return -ENOMEM;
4362 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004363 *(unsigned long *)m->private = x[0] * 2;
4364 kfree(x);
Christoph Lameter18004c52012-07-06 15:25:12 -05004365 mutex_lock(&slab_mutex);
Al Viro871751e2006-03-25 03:06:39 -08004366 /* Now make sure this entry will be retried */
4367 m->count = m->size;
4368 return 0;
4369 }
Christoph Lameterdb845062013-02-05 18:45:23 +00004370 for (i = 0; i < x[1]; i++) {
4371 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4372 show_symbol(m, x[2*i+2]);
Al Viro871751e2006-03-25 03:06:39 -08004373 seq_putc(m, '\n');
4374 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004375
Al Viro871751e2006-03-25 03:06:39 -08004376 return 0;
4377}
4378
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004379static const struct seq_operations slabstats_op = {
Vladimir Davydov1df3b262014-12-10 15:42:16 -08004380 .start = slab_start,
Wanpeng Li276a2432013-07-08 08:08:28 +08004381 .next = slab_next,
4382 .stop = slab_stop,
Al Viro871751e2006-03-25 03:06:39 -08004383 .show = leaks_show,
4384};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004385
4386static int slabstats_open(struct inode *inode, struct file *file)
4387{
Rob Jonesb208ce32014-10-09 15:28:03 -07004388 unsigned long *n;
4389
4390 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4391 if (!n)
4392 return -ENOMEM;
4393
4394 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4395
4396 return 0;
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004397}
4398
4399static const struct file_operations proc_slabstats_operations = {
4400 .open = slabstats_open,
4401 .read = seq_read,
4402 .llseek = seq_lseek,
4403 .release = seq_release_private,
4404};
Al Viro871751e2006-03-25 03:06:39 -08004405#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004406
4407static int __init slab_proc_init(void)
4408{
4409#ifdef CONFIG_DEBUG_SLAB_LEAK
4410 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4411#endif
4412 return 0;
4413}
4414module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415#endif
4416
Kees Cook04385fc2016-06-23 15:20:59 -07004417#ifdef CONFIG_HARDENED_USERCOPY
4418/*
4419 * Rejects objects that are incorrectly sized.
4420 *
4421 * Returns NULL if check passes, otherwise const char * to name of cache
4422 * to indicate an error.
4423 */
4424const char *__check_heap_object(const void *ptr, unsigned long n,
4425 struct page *page)
4426{
4427 struct kmem_cache *cachep;
4428 unsigned int objnr;
4429 unsigned long offset;
4430
4431 /* Find and validate object. */
4432 cachep = page->slab_cache;
4433 objnr = obj_to_index(cachep, page, (void *)ptr);
4434 BUG_ON(objnr >= cachep->num);
4435
4436 /* Find offset within object. */
4437 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4438
4439 /* Allow address range falling entirely within object size. */
4440 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4441 return NULL;
4442
4443 return cachep->name;
4444}
4445#endif /* CONFIG_HARDENED_USERCOPY */
4446
Manfred Spraul00e145b2005-09-03 15:55:07 -07004447/**
4448 * ksize - get the actual amount of memory allocated for a given object
4449 * @objp: Pointer to the object
4450 *
4451 * kmalloc may internally round up allocations and return more memory
4452 * than requested. ksize() can be used to determine the actual amount of
4453 * memory allocated. The caller may use this additional memory, even though
4454 * a smaller amount of memory was initially specified with the kmalloc call.
4455 * The caller must guarantee that objp points to a valid object previously
4456 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4457 * must not be freed during the duration of the call.
4458 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004459size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460{
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004461 size_t size;
4462
Christoph Lameteref8b4522007-10-16 01:24:46 -07004463 BUG_ON(!objp);
4464 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004465 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004466
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004467 size = virt_to_cache(objp)->object_size;
4468 /* We assume that ksize callers could use the whole allocated area,
4469 * so we need to unpoison this area.
4470 */
Alexander Potapenko4ebb31a42016-05-20 16:59:14 -07004471 kasan_unpoison_shadow(objp, size);
Alexander Potapenko7ed2f9e2016-03-25 14:21:59 -07004472
4473 return size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004474}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004475EXPORT_SYMBOL(ksize);