Merge tag 'driver-core-4.9-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6-block.git] / mm / slab.c
CommitLineData
1da177e4
LT
1/*
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
183ff22b 29 * slabs and you must pass objects with the same initializations to
1da177e4
LT
30 * 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 *
a737b3e2 53 * The c_cpuarray may not be read with enabled local interrupts -
1da177e4
LT
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
343e0d7a 58 * Several members in struct kmem_cache and struct slab never change, they
1da177e4
LT
59 * 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
18004c5d 71 * The global cache-chain is protected by the mutex 'slab_mutex'.
1da177e4
LT
72 * 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 *
e498be7d
CL
78 * 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.
1da177e4
LT
87 */
88
1da177e4
LT
89#include <linux/slab.h>
90#include <linux/mm.h>
c9cf5528 91#include <linux/poison.h>
1da177e4
LT
92#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
101a5001 97#include <linux/cpuset.h>
a0ec95a8 98#include <linux/proc_fs.h>
1da177e4
LT
99#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>
543537bd 106#include <linux/string.h>
138ae663 107#include <linux/uaccess.h>
e498be7d 108#include <linux/nodemask.h>
d5cff635 109#include <linux/kmemleak.h>
dc85da15 110#include <linux/mempolicy.h>
fc0abb14 111#include <linux/mutex.h>
8a8b6502 112#include <linux/fault-inject.h>
e7eebaf6 113#include <linux/rtmutex.h>
6a2d7a95 114#include <linux/reciprocal_div.h>
3ac7fe5a 115#include <linux/debugobjects.h>
c175eea4 116#include <linux/kmemcheck.h>
8f9f8d9e 117#include <linux/memory.h>
268bb0ce 118#include <linux/prefetch.h>
1da177e4 119
381760ea
MG
120#include <net/sock.h>
121
1da177e4
LT
122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
4dee6b64
SR
126#include <trace/events/kmem.h>
127
072bb0aa
MG
128#include "internal.h"
129
b9ce5ef4
GC
130#include "slab.h"
131
1da177e4 132/*
50953fe9 133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
1da177e4
LT
134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
1da177e4
LT
152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
87a927c7 154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
1da177e4 155
1da177e4
LT
156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
f315e3fa
JK
160#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
162
163#if FREELIST_BYTE_INDEX
164typedef unsigned char freelist_idx_t;
165#else
166typedef unsigned short freelist_idx_t;
167#endif
168
30321c7b 169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
f315e3fa 170
1da177e4
LT
171/*
172 * struct array_cache
173 *
1da177e4
LT
174 * Purpose:
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
178 *
179 * The limit is stored in the per-cpu structure to reduce the data cache
180 * footprint.
181 *
182 */
183struct array_cache {
184 unsigned int avail;
185 unsigned int limit;
186 unsigned int batchcount;
187 unsigned int touched;
bda5b655 188 void *entry[]; /*
a737b3e2
AM
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
191 * the entries.
a737b3e2 192 */
1da177e4
LT
193};
194
c8522a3a
JK
195struct alien_cache {
196 spinlock_t lock;
197 struct array_cache ac;
198};
199
e498be7d
CL
200/*
201 * Need this for bootstrapping a per node allocator.
202 */
bf0dea23 203#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
ce8eb6c4 204static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
e498be7d 205#define CACHE_CACHE 0
bf0dea23 206#define SIZE_NODE (MAX_NUMNODES)
e498be7d 207
ed11d9eb 208static int drain_freelist(struct kmem_cache *cache,
ce8eb6c4 209 struct kmem_cache_node *n, int tofree);
ed11d9eb 210static void free_block(struct kmem_cache *cachep, void **objpp, int len,
97654dfa
JK
211 int node, struct list_head *list);
212static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
83b519e8 213static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
65f27f38 214static void cache_reap(struct work_struct *unused);
ed11d9eb 215
76b342bd
JK
216static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
217 void **list);
218static inline void fixup_slab_list(struct kmem_cache *cachep,
219 struct kmem_cache_node *n, struct page *page,
220 void **list);
e0a42726
IM
221static int slab_early_init = 1;
222
ce8eb6c4 223#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
1da177e4 224
ce8eb6c4 225static void kmem_cache_node_init(struct kmem_cache_node *parent)
e498be7d
CL
226{
227 INIT_LIST_HEAD(&parent->slabs_full);
228 INIT_LIST_HEAD(&parent->slabs_partial);
229 INIT_LIST_HEAD(&parent->slabs_free);
230 parent->shared = NULL;
231 parent->alien = NULL;
2e1217cf 232 parent->colour_next = 0;
e498be7d
CL
233 spin_lock_init(&parent->list_lock);
234 parent->free_objects = 0;
235 parent->free_touched = 0;
07a63c41 236 parent->num_slabs = 0;
e498be7d
CL
237}
238
a737b3e2
AM
239#define MAKE_LIST(cachep, listp, slab, nodeid) \
240 do { \
241 INIT_LIST_HEAD(listp); \
18bf8541 242 list_splice(&get_node(cachep, nodeid)->slab, listp); \
e498be7d
CL
243 } while (0)
244
a737b3e2
AM
245#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
246 do { \
e498be7d
CL
247 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
249 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
250 } while (0)
1da177e4 251
b03a017b 252#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
1da177e4 253#define CFLGS_OFF_SLAB (0x80000000UL)
b03a017b 254#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
1da177e4
LT
255#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
256
257#define BATCHREFILL_LIMIT 16
a737b3e2
AM
258/*
259 * Optimization question: fewer reaps means less probability for unnessary
260 * cpucache drain/refill cycles.
1da177e4 261 *
dc6f3f27 262 * OTOH the cpuarrays can contain lots of objects,
1da177e4
LT
263 * which could lock up otherwise freeable slabs.
264 */
5f0985bb
JZ
265#define REAPTIMEOUT_AC (2*HZ)
266#define REAPTIMEOUT_NODE (4*HZ)
1da177e4
LT
267
268#if STATS
269#define STATS_INC_ACTIVE(x) ((x)->num_active++)
270#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
271#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
272#define STATS_INC_GROWN(x) ((x)->grown++)
ed11d9eb 273#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
a737b3e2
AM
274#define STATS_SET_HIGH(x) \
275 do { \
276 if ((x)->num_active > (x)->high_mark) \
277 (x)->high_mark = (x)->num_active; \
278 } while (0)
1da177e4
LT
279#define STATS_INC_ERR(x) ((x)->errors++)
280#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
e498be7d 281#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
fb7faf33 282#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
a737b3e2
AM
283#define STATS_SET_FREEABLE(x, i) \
284 do { \
285 if ((x)->max_freeable < i) \
286 (x)->max_freeable = i; \
287 } while (0)
1da177e4
LT
288#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
289#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
290#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
291#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
292#else
293#define STATS_INC_ACTIVE(x) do { } while (0)
294#define STATS_DEC_ACTIVE(x) do { } while (0)
295#define STATS_INC_ALLOCED(x) do { } while (0)
296#define STATS_INC_GROWN(x) do { } while (0)
4e60c86b 297#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
1da177e4
LT
298#define STATS_SET_HIGH(x) do { } while (0)
299#define STATS_INC_ERR(x) do { } while (0)
300#define STATS_INC_NODEALLOCS(x) do { } while (0)
e498be7d 301#define STATS_INC_NODEFREES(x) do { } while (0)
fb7faf33 302#define STATS_INC_ACOVERFLOW(x) do { } while (0)
a737b3e2 303#define STATS_SET_FREEABLE(x, i) do { } while (0)
1da177e4
LT
304#define STATS_INC_ALLOCHIT(x) do { } while (0)
305#define STATS_INC_ALLOCMISS(x) do { } while (0)
306#define STATS_INC_FREEHIT(x) do { } while (0)
307#define STATS_INC_FREEMISS(x) do { } while (0)
308#endif
309
310#if DEBUG
1da177e4 311
a737b3e2
AM
312/*
313 * memory layout of objects:
1da177e4 314 * 0 : objp
3dafccf2 315 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
1da177e4
LT
316 * the end of an object is aligned with the end of the real
317 * allocation. Catches writes behind the end of the allocation.
3dafccf2 318 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
1da177e4 319 * redzone word.
3dafccf2 320 * cachep->obj_offset: The real object.
3b0efdfa
CL
321 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
322 * cachep->size - 1* BYTES_PER_WORD: last caller address
a737b3e2 323 * [BYTES_PER_WORD long]
1da177e4 324 */
343e0d7a 325static int obj_offset(struct kmem_cache *cachep)
1da177e4 326{
3dafccf2 327 return cachep->obj_offset;
1da177e4
LT
328}
329
b46b8f19 330static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
1da177e4
LT
331{
332 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
b46b8f19
DW
333 return (unsigned long long*) (objp + obj_offset(cachep) -
334 sizeof(unsigned long long));
1da177e4
LT
335}
336
b46b8f19 337static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
1da177e4
LT
338{
339 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
340 if (cachep->flags & SLAB_STORE_USER)
3b0efdfa 341 return (unsigned long long *)(objp + cachep->size -
b46b8f19 342 sizeof(unsigned long long) -
87a927c7 343 REDZONE_ALIGN);
3b0efdfa 344 return (unsigned long long *) (objp + cachep->size -
b46b8f19 345 sizeof(unsigned long long));
1da177e4
LT
346}
347
343e0d7a 348static void **dbg_userword(struct kmem_cache *cachep, void *objp)
1da177e4
LT
349{
350 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
3b0efdfa 351 return (void **)(objp + cachep->size - BYTES_PER_WORD);
1da177e4
LT
352}
353
354#else
355
3dafccf2 356#define obj_offset(x) 0
b46b8f19
DW
357#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
358#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
1da177e4
LT
359#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
360
361#endif
362
03787301
JK
363#ifdef CONFIG_DEBUG_SLAB_LEAK
364
d31676df 365static inline bool is_store_user_clean(struct kmem_cache *cachep)
03787301 366{
d31676df
JK
367 return atomic_read(&cachep->store_user_clean) == 1;
368}
03787301 369
d31676df
JK
370static inline void set_store_user_clean(struct kmem_cache *cachep)
371{
372 atomic_set(&cachep->store_user_clean, 1);
373}
03787301 374
d31676df
JK
375static inline void set_store_user_dirty(struct kmem_cache *cachep)
376{
377 if (is_store_user_clean(cachep))
378 atomic_set(&cachep->store_user_clean, 0);
03787301
JK
379}
380
381#else
d31676df 382static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
03787301
JK
383
384#endif
385
1da177e4 386/*
3df1cccd
DR
387 * Do not go above this order unless 0 objects fit into the slab or
388 * overridden on the command line.
1da177e4 389 */
543585cc
DR
390#define SLAB_MAX_ORDER_HI 1
391#define SLAB_MAX_ORDER_LO 0
392static int slab_max_order = SLAB_MAX_ORDER_LO;
3df1cccd 393static bool slab_max_order_set __initdata;
1da177e4 394
6ed5eb22
PE
395static inline struct kmem_cache *virt_to_cache(const void *obj)
396{
b49af68f 397 struct page *page = virt_to_head_page(obj);
35026088 398 return page->slab_cache;
6ed5eb22
PE
399}
400
8456a648 401static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
8fea4e96
PE
402 unsigned int idx)
403{
8456a648 404 return page->s_mem + cache->size * idx;
8fea4e96
PE
405}
406
6a2d7a95 407/*
3b0efdfa
CL
408 * We want to avoid an expensive divide : (offset / cache->size)
409 * Using the fact that size is a constant for a particular cache,
410 * we can replace (offset / cache->size) by
6a2d7a95
ED
411 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
412 */
413static inline unsigned int obj_to_index(const struct kmem_cache *cache,
8456a648 414 const struct page *page, void *obj)
8fea4e96 415{
8456a648 416 u32 offset = (obj - page->s_mem);
6a2d7a95 417 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
8fea4e96
PE
418}
419
6fb92430 420#define BOOT_CPUCACHE_ENTRIES 1
1da177e4 421/* internal cache of cache description objs */
9b030cb8 422static struct kmem_cache kmem_cache_boot = {
b28a02de
PE
423 .batchcount = 1,
424 .limit = BOOT_CPUCACHE_ENTRIES,
425 .shared = 1,
3b0efdfa 426 .size = sizeof(struct kmem_cache),
b28a02de 427 .name = "kmem_cache",
1da177e4
LT
428};
429
1871e52c 430static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
1da177e4 431
343e0d7a 432static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
1da177e4 433{
bf0dea23 434 return this_cpu_ptr(cachep->cpu_cache);
1da177e4
LT
435}
436
a737b3e2
AM
437/*
438 * Calculate the number of objects and left-over bytes for a given buffer size.
439 */
70f75067
JK
440static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
441 unsigned long flags, size_t *left_over)
fbaccacf 442{
70f75067 443 unsigned int num;
fbaccacf 444 size_t slab_size = PAGE_SIZE << gfporder;
1da177e4 445
fbaccacf
SR
446 /*
447 * The slab management structure can be either off the slab or
448 * on it. For the latter case, the memory allocated for a
449 * slab is used for:
450 *
fbaccacf 451 * - @buffer_size bytes for each object
2e6b3602
JK
452 * - One freelist_idx_t for each object
453 *
454 * We don't need to consider alignment of freelist because
455 * freelist will be at the end of slab page. The objects will be
456 * at the correct alignment.
fbaccacf
SR
457 *
458 * If the slab management structure is off the slab, then the
459 * alignment will already be calculated into the size. Because
460 * the slabs are all pages aligned, the objects will be at the
461 * correct alignment when allocated.
462 */
b03a017b 463 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
70f75067 464 num = slab_size / buffer_size;
2e6b3602 465 *left_over = slab_size % buffer_size;
fbaccacf 466 } else {
70f75067 467 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
2e6b3602
JK
468 *left_over = slab_size %
469 (buffer_size + sizeof(freelist_idx_t));
fbaccacf 470 }
70f75067
JK
471
472 return num;
1da177e4
LT
473}
474
f28510d3 475#if DEBUG
d40cee24 476#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
1da177e4 477
a737b3e2
AM
478static void __slab_error(const char *function, struct kmem_cache *cachep,
479 char *msg)
1da177e4 480{
1170532b 481 pr_err("slab error in %s(): cache `%s': %s\n",
b28a02de 482 function, cachep->name, msg);
1da177e4 483 dump_stack();
373d4d09 484 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1da177e4 485}
f28510d3 486#endif
1da177e4 487
3395ee05
PM
488/*
489 * By default on NUMA we use alien caches to stage the freeing of
490 * objects allocated from other nodes. This causes massive memory
491 * inefficiencies when using fake NUMA setup to split memory into a
492 * large number of small nodes, so it can be disabled on the command
493 * line
494 */
495
496static int use_alien_caches __read_mostly = 1;
497static int __init noaliencache_setup(char *s)
498{
499 use_alien_caches = 0;
500 return 1;
501}
502__setup("noaliencache", noaliencache_setup);
503
3df1cccd
DR
504static int __init slab_max_order_setup(char *str)
505{
506 get_option(&str, &slab_max_order);
507 slab_max_order = slab_max_order < 0 ? 0 :
508 min(slab_max_order, MAX_ORDER - 1);
509 slab_max_order_set = true;
510
511 return 1;
512}
513__setup("slab_max_order=", slab_max_order_setup);
514
8fce4d8e
CL
515#ifdef CONFIG_NUMA
516/*
517 * Special reaping functions for NUMA systems called from cache_reap().
518 * These take care of doing round robin flushing of alien caches (containing
519 * objects freed on different nodes from which they were allocated) and the
520 * flushing of remote pcps by calling drain_node_pages.
521 */
1871e52c 522static DEFINE_PER_CPU(unsigned long, slab_reap_node);
8fce4d8e
CL
523
524static void init_reap_node(int cpu)
525{
0edaf86c
AM
526 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
527 node_online_map);
8fce4d8e
CL
528}
529
530static void next_reap_node(void)
531{
909ea964 532 int node = __this_cpu_read(slab_reap_node);
8fce4d8e 533
0edaf86c 534 node = next_node_in(node, node_online_map);
909ea964 535 __this_cpu_write(slab_reap_node, node);
8fce4d8e
CL
536}
537
538#else
539#define init_reap_node(cpu) do { } while (0)
540#define next_reap_node(void) do { } while (0)
541#endif
542
1da177e4
LT
543/*
544 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
545 * via the workqueue/eventd.
546 * Add the CPU number into the expiration time to minimize the possibility of
547 * the CPUs getting into lockstep and contending for the global cache chain
548 * lock.
549 */
0db0628d 550static void start_cpu_timer(int cpu)
1da177e4 551{
1871e52c 552 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
1da177e4
LT
553
554 /*
555 * When this gets called from do_initcalls via cpucache_init(),
556 * init_workqueues() has already run, so keventd will be setup
557 * at that time.
558 */
52bad64d 559 if (keventd_up() && reap_work->work.func == NULL) {
8fce4d8e 560 init_reap_node(cpu);
203b42f7 561 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
2b284214
AV
562 schedule_delayed_work_on(cpu, reap_work,
563 __round_jiffies_relative(HZ, cpu));
1da177e4
LT
564 }
565}
566
1fe00d50 567static void init_arraycache(struct array_cache *ac, int limit, int batch)
1da177e4 568{
d5cff635
CM
569 /*
570 * The array_cache structures contain pointers to free object.
25985edc 571 * However, when such objects are allocated or transferred to another
d5cff635
CM
572 * cache the pointers are not cleared and they could be counted as
573 * valid references during a kmemleak scan. Therefore, kmemleak must
574 * not scan such objects.
575 */
1fe00d50
JK
576 kmemleak_no_scan(ac);
577 if (ac) {
578 ac->avail = 0;
579 ac->limit = limit;
580 ac->batchcount = batch;
581 ac->touched = 0;
1da177e4 582 }
1fe00d50
JK
583}
584
585static struct array_cache *alloc_arraycache(int node, int entries,
586 int batchcount, gfp_t gfp)
587{
5e804789 588 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
1fe00d50
JK
589 struct array_cache *ac = NULL;
590
591 ac = kmalloc_node(memsize, gfp, node);
592 init_arraycache(ac, entries, batchcount);
593 return ac;
1da177e4
LT
594}
595
f68f8ddd
JK
596static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
597 struct page *page, void *objp)
072bb0aa 598{
f68f8ddd
JK
599 struct kmem_cache_node *n;
600 int page_node;
601 LIST_HEAD(list);
072bb0aa 602
f68f8ddd
JK
603 page_node = page_to_nid(page);
604 n = get_node(cachep, page_node);
381760ea 605
f68f8ddd
JK
606 spin_lock(&n->list_lock);
607 free_block(cachep, &objp, 1, page_node, &list);
608 spin_unlock(&n->list_lock);
381760ea 609
f68f8ddd 610 slabs_destroy(cachep, &list);
072bb0aa
MG
611}
612
3ded175a
CL
613/*
614 * Transfer objects in one arraycache to another.
615 * Locking must be handled by the caller.
616 *
617 * Return the number of entries transferred.
618 */
619static int transfer_objects(struct array_cache *to,
620 struct array_cache *from, unsigned int max)
621{
622 /* Figure out how many entries to transfer */
732eacc0 623 int nr = min3(from->avail, max, to->limit - to->avail);
3ded175a
CL
624
625 if (!nr)
626 return 0;
627
628 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
629 sizeof(void *) *nr);
630
631 from->avail -= nr;
632 to->avail += nr;
3ded175a
CL
633 return nr;
634}
635
765c4507
CL
636#ifndef CONFIG_NUMA
637
638#define drain_alien_cache(cachep, alien) do { } while (0)
ce8eb6c4 639#define reap_alien(cachep, n) do { } while (0)
765c4507 640
c8522a3a
JK
641static inline struct alien_cache **alloc_alien_cache(int node,
642 int limit, gfp_t gfp)
765c4507 643{
8888177e 644 return NULL;
765c4507
CL
645}
646
c8522a3a 647static inline void free_alien_cache(struct alien_cache **ac_ptr)
765c4507
CL
648{
649}
650
651static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
652{
653 return 0;
654}
655
656static inline void *alternate_node_alloc(struct kmem_cache *cachep,
657 gfp_t flags)
658{
659 return NULL;
660}
661
8b98c169 662static inline void *____cache_alloc_node(struct kmem_cache *cachep,
765c4507
CL
663 gfp_t flags, int nodeid)
664{
665 return NULL;
666}
667
4167e9b2
DR
668static inline gfp_t gfp_exact_node(gfp_t flags)
669{
444eb2a4 670 return flags & ~__GFP_NOFAIL;
4167e9b2
DR
671}
672
765c4507
CL
673#else /* CONFIG_NUMA */
674
8b98c169 675static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
c61afb18 676static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
dc85da15 677
c8522a3a
JK
678static struct alien_cache *__alloc_alien_cache(int node, int entries,
679 int batch, gfp_t gfp)
680{
5e804789 681 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
c8522a3a
JK
682 struct alien_cache *alc = NULL;
683
684 alc = kmalloc_node(memsize, gfp, node);
685 init_arraycache(&alc->ac, entries, batch);
49dfc304 686 spin_lock_init(&alc->lock);
c8522a3a
JK
687 return alc;
688}
689
690static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
e498be7d 691{
c8522a3a 692 struct alien_cache **alc_ptr;
5e804789 693 size_t memsize = sizeof(void *) * nr_node_ids;
e498be7d
CL
694 int i;
695
696 if (limit > 1)
697 limit = 12;
c8522a3a
JK
698 alc_ptr = kzalloc_node(memsize, gfp, node);
699 if (!alc_ptr)
700 return NULL;
701
702 for_each_node(i) {
703 if (i == node || !node_online(i))
704 continue;
705 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
706 if (!alc_ptr[i]) {
707 for (i--; i >= 0; i--)
708 kfree(alc_ptr[i]);
709 kfree(alc_ptr);
710 return NULL;
e498be7d
CL
711 }
712 }
c8522a3a 713 return alc_ptr;
e498be7d
CL
714}
715
c8522a3a 716static void free_alien_cache(struct alien_cache **alc_ptr)
e498be7d
CL
717{
718 int i;
719
c8522a3a 720 if (!alc_ptr)
e498be7d 721 return;
e498be7d 722 for_each_node(i)
c8522a3a
JK
723 kfree(alc_ptr[i]);
724 kfree(alc_ptr);
e498be7d
CL
725}
726
343e0d7a 727static void __drain_alien_cache(struct kmem_cache *cachep,
833b706c
JK
728 struct array_cache *ac, int node,
729 struct list_head *list)
e498be7d 730{
18bf8541 731 struct kmem_cache_node *n = get_node(cachep, node);
e498be7d
CL
732
733 if (ac->avail) {
ce8eb6c4 734 spin_lock(&n->list_lock);
e00946fe
CL
735 /*
736 * Stuff objects into the remote nodes shared array first.
737 * That way we could avoid the overhead of putting the objects
738 * into the free lists and getting them back later.
739 */
ce8eb6c4
CL
740 if (n->shared)
741 transfer_objects(n->shared, ac, ac->limit);
e00946fe 742
833b706c 743 free_block(cachep, ac->entry, ac->avail, node, list);
e498be7d 744 ac->avail = 0;
ce8eb6c4 745 spin_unlock(&n->list_lock);
e498be7d
CL
746 }
747}
748
8fce4d8e
CL
749/*
750 * Called from cache_reap() to regularly drain alien caches round robin.
751 */
ce8eb6c4 752static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
8fce4d8e 753{
909ea964 754 int node = __this_cpu_read(slab_reap_node);
8fce4d8e 755
ce8eb6c4 756 if (n->alien) {
c8522a3a
JK
757 struct alien_cache *alc = n->alien[node];
758 struct array_cache *ac;
759
760 if (alc) {
761 ac = &alc->ac;
49dfc304 762 if (ac->avail && spin_trylock_irq(&alc->lock)) {
833b706c
JK
763 LIST_HEAD(list);
764
765 __drain_alien_cache(cachep, ac, node, &list);
49dfc304 766 spin_unlock_irq(&alc->lock);
833b706c 767 slabs_destroy(cachep, &list);
c8522a3a 768 }
8fce4d8e
CL
769 }
770 }
771}
772
a737b3e2 773static void drain_alien_cache(struct kmem_cache *cachep,
c8522a3a 774 struct alien_cache **alien)
e498be7d 775{
b28a02de 776 int i = 0;
c8522a3a 777 struct alien_cache *alc;
e498be7d
CL
778 struct array_cache *ac;
779 unsigned long flags;
780
781 for_each_online_node(i) {
c8522a3a
JK
782 alc = alien[i];
783 if (alc) {
833b706c
JK
784 LIST_HEAD(list);
785
c8522a3a 786 ac = &alc->ac;
49dfc304 787 spin_lock_irqsave(&alc->lock, flags);
833b706c 788 __drain_alien_cache(cachep, ac, i, &list);
49dfc304 789 spin_unlock_irqrestore(&alc->lock, flags);
833b706c 790 slabs_destroy(cachep, &list);
e498be7d
CL
791 }
792 }
793}
729bd0b7 794
25c4f304
JK
795static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
796 int node, int page_node)
729bd0b7 797{
ce8eb6c4 798 struct kmem_cache_node *n;
c8522a3a
JK
799 struct alien_cache *alien = NULL;
800 struct array_cache *ac;
97654dfa 801 LIST_HEAD(list);
1ca4cb24 802
18bf8541 803 n = get_node(cachep, node);
729bd0b7 804 STATS_INC_NODEFREES(cachep);
25c4f304
JK
805 if (n->alien && n->alien[page_node]) {
806 alien = n->alien[page_node];
c8522a3a 807 ac = &alien->ac;
49dfc304 808 spin_lock(&alien->lock);
c8522a3a 809 if (unlikely(ac->avail == ac->limit)) {
729bd0b7 810 STATS_INC_ACOVERFLOW(cachep);
25c4f304 811 __drain_alien_cache(cachep, ac, page_node, &list);
729bd0b7 812 }
f68f8ddd 813 ac->entry[ac->avail++] = objp;
49dfc304 814 spin_unlock(&alien->lock);
833b706c 815 slabs_destroy(cachep, &list);
729bd0b7 816 } else {
25c4f304 817 n = get_node(cachep, page_node);
18bf8541 818 spin_lock(&n->list_lock);
25c4f304 819 free_block(cachep, &objp, 1, page_node, &list);
18bf8541 820 spin_unlock(&n->list_lock);
97654dfa 821 slabs_destroy(cachep, &list);
729bd0b7
PE
822 }
823 return 1;
824}
25c4f304
JK
825
826static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
827{
828 int page_node = page_to_nid(virt_to_page(objp));
829 int node = numa_mem_id();
830 /*
831 * Make sure we are not freeing a object from another node to the array
832 * cache on this cpu.
833 */
834 if (likely(node == page_node))
835 return 0;
836
837 return __cache_free_alien(cachep, objp, node, page_node);
838}
4167e9b2
DR
839
840/*
444eb2a4
MG
841 * Construct gfp mask to allocate from a specific node but do not reclaim or
842 * warn about failures.
4167e9b2
DR
843 */
844static inline gfp_t gfp_exact_node(gfp_t flags)
845{
444eb2a4 846 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
4167e9b2 847}
e498be7d
CL
848#endif
849
ded0ecf6
JK
850static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
851{
852 struct kmem_cache_node *n;
853
854 /*
855 * Set up the kmem_cache_node for cpu before we can
856 * begin anything. Make sure some other cpu on this
857 * node has not already allocated this
858 */
859 n = get_node(cachep, node);
860 if (n) {
861 spin_lock_irq(&n->list_lock);
862 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
863 cachep->num;
864 spin_unlock_irq(&n->list_lock);
865
866 return 0;
867 }
868
869 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
870 if (!n)
871 return -ENOMEM;
872
873 kmem_cache_node_init(n);
874 n->next_reap = jiffies + REAPTIMEOUT_NODE +
875 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
876
877 n->free_limit =
878 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
879
880 /*
881 * The kmem_cache_nodes don't come and go as CPUs
882 * come and go. slab_mutex is sufficient
883 * protection here.
884 */
885 cachep->node[node] = n;
886
887 return 0;
888}
889
6731d4f1 890#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
8f9f8d9e 891/*
6a67368c 892 * Allocates and initializes node for a node on each slab cache, used for
ce8eb6c4 893 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
8f9f8d9e 894 * will be allocated off-node since memory is not yet online for the new node.
6a67368c 895 * When hotplugging memory or a cpu, existing node are not replaced if
8f9f8d9e
DR
896 * already in use.
897 *
18004c5d 898 * Must hold slab_mutex.
8f9f8d9e 899 */
6a67368c 900static int init_cache_node_node(int node)
8f9f8d9e 901{
ded0ecf6 902 int ret;
8f9f8d9e 903 struct kmem_cache *cachep;
8f9f8d9e 904
18004c5d 905 list_for_each_entry(cachep, &slab_caches, list) {
ded0ecf6
JK
906 ret = init_cache_node(cachep, node, GFP_KERNEL);
907 if (ret)
908 return ret;
8f9f8d9e 909 }
ded0ecf6 910
8f9f8d9e
DR
911 return 0;
912}
6731d4f1 913#endif
8f9f8d9e 914
c3d332b6
JK
915static int setup_kmem_cache_node(struct kmem_cache *cachep,
916 int node, gfp_t gfp, bool force_change)
917{
918 int ret = -ENOMEM;
919 struct kmem_cache_node *n;
920 struct array_cache *old_shared = NULL;
921 struct array_cache *new_shared = NULL;
922 struct alien_cache **new_alien = NULL;
923 LIST_HEAD(list);
924
925 if (use_alien_caches) {
926 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
927 if (!new_alien)
928 goto fail;
929 }
930
931 if (cachep->shared) {
932 new_shared = alloc_arraycache(node,
933 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
934 if (!new_shared)
935 goto fail;
936 }
937
938 ret = init_cache_node(cachep, node, gfp);
939 if (ret)
940 goto fail;
941
942 n = get_node(cachep, node);
943 spin_lock_irq(&n->list_lock);
944 if (n->shared && force_change) {
945 free_block(cachep, n->shared->entry,
946 n->shared->avail, node, &list);
947 n->shared->avail = 0;
948 }
949
950 if (!n->shared || force_change) {
951 old_shared = n->shared;
952 n->shared = new_shared;
953 new_shared = NULL;
954 }
955
956 if (!n->alien) {
957 n->alien = new_alien;
958 new_alien = NULL;
959 }
960
961 spin_unlock_irq(&n->list_lock);
962 slabs_destroy(cachep, &list);
963
801faf0d
JK
964 /*
965 * To protect lockless access to n->shared during irq disabled context.
966 * If n->shared isn't NULL in irq disabled context, accessing to it is
967 * guaranteed to be valid until irq is re-enabled, because it will be
968 * freed after synchronize_sched().
969 */
86d9f485 970 if (old_shared && force_change)
801faf0d
JK
971 synchronize_sched();
972
c3d332b6
JK
973fail:
974 kfree(old_shared);
975 kfree(new_shared);
976 free_alien_cache(new_alien);
977
978 return ret;
979}
980
6731d4f1
SAS
981#ifdef CONFIG_SMP
982
0db0628d 983static void cpuup_canceled(long cpu)
fbf1e473
AM
984{
985 struct kmem_cache *cachep;
ce8eb6c4 986 struct kmem_cache_node *n = NULL;
7d6e6d09 987 int node = cpu_to_mem(cpu);
a70f7302 988 const struct cpumask *mask = cpumask_of_node(node);
fbf1e473 989
18004c5d 990 list_for_each_entry(cachep, &slab_caches, list) {
fbf1e473
AM
991 struct array_cache *nc;
992 struct array_cache *shared;
c8522a3a 993 struct alien_cache **alien;
97654dfa 994 LIST_HEAD(list);
fbf1e473 995
18bf8541 996 n = get_node(cachep, node);
ce8eb6c4 997 if (!n)
bf0dea23 998 continue;
fbf1e473 999
ce8eb6c4 1000 spin_lock_irq(&n->list_lock);
fbf1e473 1001
ce8eb6c4
CL
1002 /* Free limit for this kmem_cache_node */
1003 n->free_limit -= cachep->batchcount;
bf0dea23
JK
1004
1005 /* cpu is dead; no one can alloc from it. */
1006 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1007 if (nc) {
97654dfa 1008 free_block(cachep, nc->entry, nc->avail, node, &list);
bf0dea23
JK
1009 nc->avail = 0;
1010 }
fbf1e473 1011
58463c1f 1012 if (!cpumask_empty(mask)) {
ce8eb6c4 1013 spin_unlock_irq(&n->list_lock);
bf0dea23 1014 goto free_slab;
fbf1e473
AM
1015 }
1016
ce8eb6c4 1017 shared = n->shared;
fbf1e473
AM
1018 if (shared) {
1019 free_block(cachep, shared->entry,
97654dfa 1020 shared->avail, node, &list);
ce8eb6c4 1021 n->shared = NULL;
fbf1e473
AM
1022 }
1023
ce8eb6c4
CL
1024 alien = n->alien;
1025 n->alien = NULL;
fbf1e473 1026
ce8eb6c4 1027 spin_unlock_irq(&n->list_lock);
fbf1e473
AM
1028
1029 kfree(shared);
1030 if (alien) {
1031 drain_alien_cache(cachep, alien);
1032 free_alien_cache(alien);
1033 }
bf0dea23
JK
1034
1035free_slab:
97654dfa 1036 slabs_destroy(cachep, &list);
fbf1e473
AM
1037 }
1038 /*
1039 * In the previous loop, all the objects were freed to
1040 * the respective cache's slabs, now we can go ahead and
1041 * shrink each nodelist to its limit.
1042 */
18004c5d 1043 list_for_each_entry(cachep, &slab_caches, list) {
18bf8541 1044 n = get_node(cachep, node);
ce8eb6c4 1045 if (!n)
fbf1e473 1046 continue;
a5aa63a5 1047 drain_freelist(cachep, n, INT_MAX);
fbf1e473
AM
1048 }
1049}
1050
0db0628d 1051static int cpuup_prepare(long cpu)
1da177e4 1052{
343e0d7a 1053 struct kmem_cache *cachep;
7d6e6d09 1054 int node = cpu_to_mem(cpu);
8f9f8d9e 1055 int err;
1da177e4 1056
fbf1e473
AM
1057 /*
1058 * We need to do this right in the beginning since
1059 * alloc_arraycache's are going to use this list.
1060 * kmalloc_node allows us to add the slab to the right
ce8eb6c4 1061 * kmem_cache_node and not this cpu's kmem_cache_node
fbf1e473 1062 */
6a67368c 1063 err = init_cache_node_node(node);
8f9f8d9e
DR
1064 if (err < 0)
1065 goto bad;
fbf1e473
AM
1066
1067 /*
1068 * Now we can go ahead with allocating the shared arrays and
1069 * array caches
1070 */
18004c5d 1071 list_for_each_entry(cachep, &slab_caches, list) {
c3d332b6
JK
1072 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1073 if (err)
1074 goto bad;
fbf1e473 1075 }
ce79ddc8 1076
fbf1e473
AM
1077 return 0;
1078bad:
12d00f6a 1079 cpuup_canceled(cpu);
fbf1e473
AM
1080 return -ENOMEM;
1081}
1082
6731d4f1 1083int slab_prepare_cpu(unsigned int cpu)
fbf1e473 1084{
6731d4f1 1085 int err;
fbf1e473 1086
6731d4f1
SAS
1087 mutex_lock(&slab_mutex);
1088 err = cpuup_prepare(cpu);
1089 mutex_unlock(&slab_mutex);
1090 return err;
1091}
1092
1093/*
1094 * This is called for a failed online attempt and for a successful
1095 * offline.
1096 *
1097 * Even if all the cpus of a node are down, we don't free the
1098 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1099 * a kmalloc allocation from another cpu for memory from the node of
1100 * the cpu going down. The list3 structure is usually allocated from
1101 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1102 */
1103int slab_dead_cpu(unsigned int cpu)
1104{
1105 mutex_lock(&slab_mutex);
1106 cpuup_canceled(cpu);
1107 mutex_unlock(&slab_mutex);
1108 return 0;
1109}
8f5be20b 1110#endif
6731d4f1
SAS
1111
1112static int slab_online_cpu(unsigned int cpu)
1113{
1114 start_cpu_timer(cpu);
1115 return 0;
1da177e4
LT
1116}
1117
6731d4f1
SAS
1118static int slab_offline_cpu(unsigned int cpu)
1119{
1120 /*
1121 * Shutdown cache reaper. Note that the slab_mutex is held so
1122 * that if cache_reap() is invoked it cannot do anything
1123 * expensive but will only modify reap_work and reschedule the
1124 * timer.
1125 */
1126 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1127 /* Now the cache_reaper is guaranteed to be not running. */
1128 per_cpu(slab_reap_work, cpu).work.func = NULL;
1129 return 0;
1130}
1da177e4 1131
8f9f8d9e
DR
1132#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1133/*
1134 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1135 * Returns -EBUSY if all objects cannot be drained so that the node is not
1136 * removed.
1137 *
18004c5d 1138 * Must hold slab_mutex.
8f9f8d9e 1139 */
6a67368c 1140static int __meminit drain_cache_node_node(int node)
8f9f8d9e
DR
1141{
1142 struct kmem_cache *cachep;
1143 int ret = 0;
1144
18004c5d 1145 list_for_each_entry(cachep, &slab_caches, list) {
ce8eb6c4 1146 struct kmem_cache_node *n;
8f9f8d9e 1147
18bf8541 1148 n = get_node(cachep, node);
ce8eb6c4 1149 if (!n)
8f9f8d9e
DR
1150 continue;
1151
a5aa63a5 1152 drain_freelist(cachep, n, INT_MAX);
8f9f8d9e 1153
ce8eb6c4
CL
1154 if (!list_empty(&n->slabs_full) ||
1155 !list_empty(&n->slabs_partial)) {
8f9f8d9e
DR
1156 ret = -EBUSY;
1157 break;
1158 }
1159 }
1160 return ret;
1161}
1162
1163static int __meminit slab_memory_callback(struct notifier_block *self,
1164 unsigned long action, void *arg)
1165{
1166 struct memory_notify *mnb = arg;
1167 int ret = 0;
1168 int nid;
1169
1170 nid = mnb->status_change_nid;
1171 if (nid < 0)
1172 goto out;
1173
1174 switch (action) {
1175 case MEM_GOING_ONLINE:
18004c5d 1176 mutex_lock(&slab_mutex);
6a67368c 1177 ret = init_cache_node_node(nid);
18004c5d 1178 mutex_unlock(&slab_mutex);
8f9f8d9e
DR
1179 break;
1180 case MEM_GOING_OFFLINE:
18004c5d 1181 mutex_lock(&slab_mutex);
6a67368c 1182 ret = drain_cache_node_node(nid);
18004c5d 1183 mutex_unlock(&slab_mutex);
8f9f8d9e
DR
1184 break;
1185 case MEM_ONLINE:
1186 case MEM_OFFLINE:
1187 case MEM_CANCEL_ONLINE:
1188 case MEM_CANCEL_OFFLINE:
1189 break;
1190 }
1191out:
5fda1bd5 1192 return notifier_from_errno(ret);
8f9f8d9e
DR
1193}
1194#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1195
e498be7d 1196/*
ce8eb6c4 1197 * swap the static kmem_cache_node with kmalloced memory
e498be7d 1198 */
6744f087 1199static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
8f9f8d9e 1200 int nodeid)
e498be7d 1201{
6744f087 1202 struct kmem_cache_node *ptr;
e498be7d 1203
6744f087 1204 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
e498be7d
CL
1205 BUG_ON(!ptr);
1206
6744f087 1207 memcpy(ptr, list, sizeof(struct kmem_cache_node));
2b2d5493
IM
1208 /*
1209 * Do not assume that spinlocks can be initialized via memcpy:
1210 */
1211 spin_lock_init(&ptr->list_lock);
1212
e498be7d 1213 MAKE_ALL_LISTS(cachep, ptr, nodeid);
6a67368c 1214 cachep->node[nodeid] = ptr;
e498be7d
CL
1215}
1216
556a169d 1217/*
ce8eb6c4
CL
1218 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1219 * size of kmem_cache_node.
556a169d 1220 */
ce8eb6c4 1221static void __init set_up_node(struct kmem_cache *cachep, int index)
556a169d
PE
1222{
1223 int node;
1224
1225 for_each_online_node(node) {
ce8eb6c4 1226 cachep->node[node] = &init_kmem_cache_node[index + node];
6a67368c 1227 cachep->node[node]->next_reap = jiffies +
5f0985bb
JZ
1228 REAPTIMEOUT_NODE +
1229 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
556a169d
PE
1230 }
1231}
1232
a737b3e2
AM
1233/*
1234 * Initialisation. Called after the page allocator have been initialised and
1235 * before smp_init().
1da177e4
LT
1236 */
1237void __init kmem_cache_init(void)
1238{
e498be7d
CL
1239 int i;
1240
68126702
JK
1241 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1242 sizeof(struct rcu_head));
9b030cb8
CL
1243 kmem_cache = &kmem_cache_boot;
1244
8888177e 1245 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
62918a03
SS
1246 use_alien_caches = 0;
1247
3c583465 1248 for (i = 0; i < NUM_INIT_LISTS; i++)
ce8eb6c4 1249 kmem_cache_node_init(&init_kmem_cache_node[i]);
3c583465 1250
1da177e4
LT
1251 /*
1252 * Fragmentation resistance on low memory - only use bigger
3df1cccd
DR
1253 * page orders on machines with more than 32MB of memory if
1254 * not overridden on the command line.
1da177e4 1255 */
3df1cccd 1256 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
543585cc 1257 slab_max_order = SLAB_MAX_ORDER_HI;
1da177e4 1258
1da177e4
LT
1259 /* Bootstrap is tricky, because several objects are allocated
1260 * from caches that do not exist yet:
9b030cb8
CL
1261 * 1) initialize the kmem_cache cache: it contains the struct
1262 * kmem_cache structures of all caches, except kmem_cache itself:
1263 * kmem_cache is statically allocated.
e498be7d 1264 * Initially an __init data area is used for the head array and the
ce8eb6c4 1265 * kmem_cache_node structures, it's replaced with a kmalloc allocated
e498be7d 1266 * array at the end of the bootstrap.
1da177e4 1267 * 2) Create the first kmalloc cache.
343e0d7a 1268 * The struct kmem_cache for the new cache is allocated normally.
e498be7d
CL
1269 * An __init data area is used for the head array.
1270 * 3) Create the remaining kmalloc caches, with minimally sized
1271 * head arrays.
9b030cb8 1272 * 4) Replace the __init data head arrays for kmem_cache and the first
1da177e4 1273 * kmalloc cache with kmalloc allocated arrays.
ce8eb6c4 1274 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
e498be7d
CL
1275 * the other cache's with kmalloc allocated memory.
1276 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1da177e4
LT
1277 */
1278
9b030cb8 1279 /* 1) create the kmem_cache */
1da177e4 1280
8da3430d 1281 /*
b56efcf0 1282 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
8da3430d 1283 */
2f9baa9f 1284 create_boot_cache(kmem_cache, "kmem_cache",
bf0dea23 1285 offsetof(struct kmem_cache, node) +
6744f087 1286 nr_node_ids * sizeof(struct kmem_cache_node *),
2f9baa9f
CL
1287 SLAB_HWCACHE_ALIGN);
1288 list_add(&kmem_cache->list, &slab_caches);
bf0dea23 1289 slab_state = PARTIAL;
1da177e4 1290
a737b3e2 1291 /*
bf0dea23
JK
1292 * Initialize the caches that provide memory for the kmem_cache_node
1293 * structures first. Without this, further allocations will bug.
e498be7d 1294 */
bf0dea23 1295 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
ce8eb6c4 1296 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
bf0dea23 1297 slab_state = PARTIAL_NODE;
34cc6990 1298 setup_kmalloc_cache_index_table();
e498be7d 1299
e0a42726
IM
1300 slab_early_init = 0;
1301
ce8eb6c4 1302 /* 5) Replace the bootstrap kmem_cache_node */
e498be7d 1303 {
1ca4cb24
PE
1304 int nid;
1305
9c09a95c 1306 for_each_online_node(nid) {
ce8eb6c4 1307 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
556a169d 1308
bf0dea23 1309 init_list(kmalloc_caches[INDEX_NODE],
ce8eb6c4 1310 &init_kmem_cache_node[SIZE_NODE + nid], nid);
e498be7d
CL
1311 }
1312 }
1da177e4 1313
f97d5f63 1314 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
8429db5c
PE
1315}
1316
1317void __init kmem_cache_init_late(void)
1318{
1319 struct kmem_cache *cachep;
1320
97d06609 1321 slab_state = UP;
52cef189 1322
8429db5c 1323 /* 6) resize the head arrays to their final sizes */
18004c5d
CL
1324 mutex_lock(&slab_mutex);
1325 list_for_each_entry(cachep, &slab_caches, list)
8429db5c
PE
1326 if (enable_cpucache(cachep, GFP_NOWAIT))
1327 BUG();
18004c5d 1328 mutex_unlock(&slab_mutex);
056c6241 1329
97d06609
CL
1330 /* Done! */
1331 slab_state = FULL;
1332
8f9f8d9e
DR
1333#ifdef CONFIG_NUMA
1334 /*
1335 * Register a memory hotplug callback that initializes and frees
6a67368c 1336 * node.
8f9f8d9e
DR
1337 */
1338 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1339#endif
1340
a737b3e2
AM
1341 /*
1342 * The reap timers are started later, with a module init call: That part
1343 * of the kernel is not yet operational.
1da177e4
LT
1344 */
1345}
1346
1347static int __init cpucache_init(void)
1348{
6731d4f1 1349 int ret;
1da177e4 1350
a737b3e2
AM
1351 /*
1352 * Register the timers that return unneeded pages to the page allocator
1da177e4 1353 */
6731d4f1
SAS
1354 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1355 slab_online_cpu, slab_offline_cpu);
1356 WARN_ON(ret < 0);
a164f896
GC
1357
1358 /* Done! */
97d06609 1359 slab_state = FULL;
1da177e4
LT
1360 return 0;
1361}
1da177e4
LT
1362__initcall(cpucache_init);
1363
8bdec192
RA
1364static noinline void
1365slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1366{
9a02d699 1367#if DEBUG
ce8eb6c4 1368 struct kmem_cache_node *n;
8456a648 1369 struct page *page;
8bdec192
RA
1370 unsigned long flags;
1371 int node;
9a02d699
DR
1372 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1373 DEFAULT_RATELIMIT_BURST);
1374
1375 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1376 return;
8bdec192 1377
5b3810e5
VB
1378 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1379 nodeid, gfpflags, &gfpflags);
1380 pr_warn(" cache: %s, object size: %d, order: %d\n",
3b0efdfa 1381 cachep->name, cachep->size, cachep->gfporder);
8bdec192 1382
18bf8541 1383 for_each_kmem_cache_node(cachep, node, n) {
8bdec192
RA
1384 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1385 unsigned long active_slabs = 0, num_slabs = 0;
07a63c41
AR
1386 unsigned long num_slabs_partial = 0, num_slabs_free = 0;
1387 unsigned long num_slabs_full;
8bdec192 1388
ce8eb6c4 1389 spin_lock_irqsave(&n->list_lock, flags);
07a63c41 1390 num_slabs = n->num_slabs;
8456a648
JK
1391 list_for_each_entry(page, &n->slabs_partial, lru) {
1392 active_objs += page->active;
07a63c41 1393 num_slabs_partial++;
8bdec192 1394 }
8456a648 1395 list_for_each_entry(page, &n->slabs_free, lru)
07a63c41 1396 num_slabs_free++;
8bdec192 1397
ce8eb6c4
CL
1398 free_objects += n->free_objects;
1399 spin_unlock_irqrestore(&n->list_lock, flags);
8bdec192 1400
8bdec192 1401 num_objs = num_slabs * cachep->num;
07a63c41
AR
1402 active_slabs = num_slabs - num_slabs_free;
1403 num_slabs_full = num_slabs -
1404 (num_slabs_partial + num_slabs_free);
1405 active_objs += (num_slabs_full * cachep->num);
1406
5b3810e5 1407 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
8bdec192
RA
1408 node, active_slabs, num_slabs, active_objs, num_objs,
1409 free_objects);
1410 }
9a02d699 1411#endif
8bdec192
RA
1412}
1413
1da177e4 1414/*
8a7d9b43
WSH
1415 * Interface to system's page allocator. No need to hold the
1416 * kmem_cache_node ->list_lock.
1da177e4
LT
1417 *
1418 * If we requested dmaable memory, we will get it. Even if we
1419 * did not request dmaable memory, we might get it, but that
1420 * would be relatively rare and ignorable.
1421 */
0c3aa83e
JK
1422static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1423 int nodeid)
1da177e4
LT
1424{
1425 struct page *page;
e1b6aa6f 1426 int nr_pages;
765c4507 1427
a618e89f 1428 flags |= cachep->allocflags;
e12ba74d
MG
1429 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1430 flags |= __GFP_RECLAIMABLE;
e1b6aa6f 1431
96db800f 1432 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
8bdec192 1433 if (!page) {
9a02d699 1434 slab_out_of_memory(cachep, flags, nodeid);
1da177e4 1435 return NULL;
8bdec192 1436 }
1da177e4 1437
f3ccb2c4
VD
1438 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1439 __free_pages(page, cachep->gfporder);
1440 return NULL;
1441 }
1442
e1b6aa6f 1443 nr_pages = (1 << cachep->gfporder);
1da177e4 1444 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
972d1a7b
CL
1445 add_zone_page_state(page_zone(page),
1446 NR_SLAB_RECLAIMABLE, nr_pages);
1447 else
1448 add_zone_page_state(page_zone(page),
1449 NR_SLAB_UNRECLAIMABLE, nr_pages);
f68f8ddd 1450
a57a4988 1451 __SetPageSlab(page);
f68f8ddd
JK
1452 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1453 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
a57a4988 1454 SetPageSlabPfmemalloc(page);
072bb0aa 1455
b1eeab67
VN
1456 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1457 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1458
1459 if (cachep->ctor)
1460 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1461 else
1462 kmemcheck_mark_unallocated_pages(page, nr_pages);
1463 }
c175eea4 1464
0c3aa83e 1465 return page;
1da177e4
LT
1466}
1467
1468/*
1469 * Interface to system's page release.
1470 */
0c3aa83e 1471static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1da177e4 1472{
27ee57c9
VD
1473 int order = cachep->gfporder;
1474 unsigned long nr_freed = (1 << order);
1da177e4 1475
27ee57c9 1476 kmemcheck_free_shadow(page, order);
c175eea4 1477
972d1a7b
CL
1478 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1479 sub_zone_page_state(page_zone(page),
1480 NR_SLAB_RECLAIMABLE, nr_freed);
1481 else
1482 sub_zone_page_state(page_zone(page),
1483 NR_SLAB_UNRECLAIMABLE, nr_freed);
73293c2f 1484
a57a4988 1485 BUG_ON(!PageSlab(page));
73293c2f 1486 __ClearPageSlabPfmemalloc(page);
a57a4988 1487 __ClearPageSlab(page);
8456a648
JK
1488 page_mapcount_reset(page);
1489 page->mapping = NULL;
1f458cbf 1490
1da177e4
LT
1491 if (current->reclaim_state)
1492 current->reclaim_state->reclaimed_slab += nr_freed;
27ee57c9
VD
1493 memcg_uncharge_slab(page, order, cachep);
1494 __free_pages(page, order);
1da177e4
LT
1495}
1496
1497static void kmem_rcu_free(struct rcu_head *head)
1498{
68126702
JK
1499 struct kmem_cache *cachep;
1500 struct page *page;
1da177e4 1501
68126702
JK
1502 page = container_of(head, struct page, rcu_head);
1503 cachep = page->slab_cache;
1504
1505 kmem_freepages(cachep, page);
1da177e4
LT
1506}
1507
1508#if DEBUG
40b44137
JK
1509static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1510{
1511 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1512 (cachep->size % PAGE_SIZE) == 0)
1513 return true;
1514
1515 return false;
1516}
1da177e4
LT
1517
1518#ifdef CONFIG_DEBUG_PAGEALLOC
343e0d7a 1519static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
b28a02de 1520 unsigned long caller)
1da177e4 1521{
8c138bc0 1522 int size = cachep->object_size;
1da177e4 1523
3dafccf2 1524 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1da177e4 1525
b28a02de 1526 if (size < 5 * sizeof(unsigned long))
1da177e4
LT
1527 return;
1528
b28a02de
PE
1529 *addr++ = 0x12345678;
1530 *addr++ = caller;
1531 *addr++ = smp_processor_id();
1532 size -= 3 * sizeof(unsigned long);
1da177e4
LT
1533 {
1534 unsigned long *sptr = &caller;
1535 unsigned long svalue;
1536
1537 while (!kstack_end(sptr)) {
1538 svalue = *sptr++;
1539 if (kernel_text_address(svalue)) {
b28a02de 1540 *addr++ = svalue;
1da177e4
LT
1541 size -= sizeof(unsigned long);
1542 if (size <= sizeof(unsigned long))
1543 break;
1544 }
1545 }
1546
1547 }
b28a02de 1548 *addr++ = 0x87654321;
1da177e4 1549}
40b44137
JK
1550
1551static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1552 int map, unsigned long caller)
1553{
1554 if (!is_debug_pagealloc_cache(cachep))
1555 return;
1556
1557 if (caller)
1558 store_stackinfo(cachep, objp, caller);
1559
1560 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1561}
1562
1563#else
1564static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1565 int map, unsigned long caller) {}
1566
1da177e4
LT
1567#endif
1568
343e0d7a 1569static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1da177e4 1570{
8c138bc0 1571 int size = cachep->object_size;
3dafccf2 1572 addr = &((char *)addr)[obj_offset(cachep)];
1da177e4
LT
1573
1574 memset(addr, val, size);
b28a02de 1575 *(unsigned char *)(addr + size - 1) = POISON_END;
1da177e4
LT
1576}
1577
1578static void dump_line(char *data, int offset, int limit)
1579{
1580 int i;
aa83aa40
DJ
1581 unsigned char error = 0;
1582 int bad_count = 0;
1583
1170532b 1584 pr_err("%03x: ", offset);
aa83aa40
DJ
1585 for (i = 0; i < limit; i++) {
1586 if (data[offset + i] != POISON_FREE) {
1587 error = data[offset + i];
1588 bad_count++;
1589 }
aa83aa40 1590 }
fdde6abb
SAS
1591 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1592 &data[offset], limit, 1);
aa83aa40
DJ
1593
1594 if (bad_count == 1) {
1595 error ^= POISON_FREE;
1596 if (!(error & (error - 1))) {
1170532b 1597 pr_err("Single bit error detected. Probably bad RAM.\n");
aa83aa40 1598#ifdef CONFIG_X86
1170532b 1599 pr_err("Run memtest86+ or a similar memory test tool.\n");
aa83aa40 1600#else
1170532b 1601 pr_err("Run a memory test tool.\n");
aa83aa40
DJ
1602#endif
1603 }
1604 }
1da177e4
LT
1605}
1606#endif
1607
1608#if DEBUG
1609
343e0d7a 1610static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1da177e4
LT
1611{
1612 int i, size;
1613 char *realobj;
1614
1615 if (cachep->flags & SLAB_RED_ZONE) {
1170532b
JP
1616 pr_err("Redzone: 0x%llx/0x%llx\n",
1617 *dbg_redzone1(cachep, objp),
1618 *dbg_redzone2(cachep, objp));
1da177e4
LT
1619 }
1620
1621 if (cachep->flags & SLAB_STORE_USER) {
1170532b 1622 pr_err("Last user: [<%p>](%pSR)\n",
071361d3
JP
1623 *dbg_userword(cachep, objp),
1624 *dbg_userword(cachep, objp));
1da177e4 1625 }
3dafccf2 1626 realobj = (char *)objp + obj_offset(cachep);
8c138bc0 1627 size = cachep->object_size;
b28a02de 1628 for (i = 0; i < size && lines; i += 16, lines--) {
1da177e4
LT
1629 int limit;
1630 limit = 16;
b28a02de
PE
1631 if (i + limit > size)
1632 limit = size - i;
1da177e4
LT
1633 dump_line(realobj, i, limit);
1634 }
1635}
1636
343e0d7a 1637static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1da177e4
LT
1638{
1639 char *realobj;
1640 int size, i;
1641 int lines = 0;
1642
40b44137
JK
1643 if (is_debug_pagealloc_cache(cachep))
1644 return;
1645
3dafccf2 1646 realobj = (char *)objp + obj_offset(cachep);
8c138bc0 1647 size = cachep->object_size;
1da177e4 1648
b28a02de 1649 for (i = 0; i < size; i++) {
1da177e4 1650 char exp = POISON_FREE;
b28a02de 1651 if (i == size - 1)
1da177e4
LT
1652 exp = POISON_END;
1653 if (realobj[i] != exp) {
1654 int limit;
1655 /* Mismatch ! */
1656 /* Print header */
1657 if (lines == 0) {
1170532b
JP
1658 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1659 print_tainted(), cachep->name,
1660 realobj, size);
1da177e4
LT
1661 print_objinfo(cachep, objp, 0);
1662 }
1663 /* Hexdump the affected line */
b28a02de 1664 i = (i / 16) * 16;
1da177e4 1665 limit = 16;
b28a02de
PE
1666 if (i + limit > size)
1667 limit = size - i;
1da177e4
LT
1668 dump_line(realobj, i, limit);
1669 i += 16;
1670 lines++;
1671 /* Limit to 5 lines */
1672 if (lines > 5)
1673 break;
1674 }
1675 }
1676 if (lines != 0) {
1677 /* Print some data about the neighboring objects, if they
1678 * exist:
1679 */
8456a648 1680 struct page *page = virt_to_head_page(objp);
8fea4e96 1681 unsigned int objnr;
1da177e4 1682
8456a648 1683 objnr = obj_to_index(cachep, page, objp);
1da177e4 1684 if (objnr) {
8456a648 1685 objp = index_to_obj(cachep, page, objnr - 1);
3dafccf2 1686 realobj = (char *)objp + obj_offset(cachep);
1170532b 1687 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1da177e4
LT
1688 print_objinfo(cachep, objp, 2);
1689 }
b28a02de 1690 if (objnr + 1 < cachep->num) {
8456a648 1691 objp = index_to_obj(cachep, page, objnr + 1);
3dafccf2 1692 realobj = (char *)objp + obj_offset(cachep);
1170532b 1693 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1da177e4
LT
1694 print_objinfo(cachep, objp, 2);
1695 }
1696 }
1697}
1698#endif
1699
12dd36fa 1700#if DEBUG
8456a648
JK
1701static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1702 struct page *page)
1da177e4 1703{
1da177e4 1704 int i;
b03a017b
JK
1705
1706 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1707 poison_obj(cachep, page->freelist - obj_offset(cachep),
1708 POISON_FREE);
1709 }
1710
1da177e4 1711 for (i = 0; i < cachep->num; i++) {
8456a648 1712 void *objp = index_to_obj(cachep, page, i);
1da177e4
LT
1713
1714 if (cachep->flags & SLAB_POISON) {
1da177e4 1715 check_poison_obj(cachep, objp);
40b44137 1716 slab_kernel_map(cachep, objp, 1, 0);
1da177e4
LT
1717 }
1718 if (cachep->flags & SLAB_RED_ZONE) {
1719 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
756a025f 1720 slab_error(cachep, "start of a freed object was overwritten");
1da177e4 1721 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
756a025f 1722 slab_error(cachep, "end of a freed object was overwritten");
1da177e4 1723 }
1da177e4 1724 }
12dd36fa 1725}
1da177e4 1726#else
8456a648
JK
1727static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1728 struct page *page)
12dd36fa 1729{
12dd36fa 1730}
1da177e4
LT
1731#endif
1732
911851e6
RD
1733/**
1734 * slab_destroy - destroy and release all objects in a slab
1735 * @cachep: cache pointer being destroyed
cb8ee1a3 1736 * @page: page pointer being destroyed
911851e6 1737 *
8a7d9b43
WSH
1738 * Destroy all the objs in a slab page, and release the mem back to the system.
1739 * Before calling the slab page must have been unlinked from the cache. The
1740 * kmem_cache_node ->list_lock is not held/needed.
12dd36fa 1741 */
8456a648 1742static void slab_destroy(struct kmem_cache *cachep, struct page *page)
12dd36fa 1743{
7e007355 1744 void *freelist;
12dd36fa 1745
8456a648
JK
1746 freelist = page->freelist;
1747 slab_destroy_debugcheck(cachep, page);
bc4f610d
KS
1748 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1749 call_rcu(&page->rcu_head, kmem_rcu_free);
1750 else
0c3aa83e 1751 kmem_freepages(cachep, page);
68126702
JK
1752
1753 /*
8456a648 1754 * From now on, we don't use freelist
68126702
JK
1755 * although actual page can be freed in rcu context
1756 */
1757 if (OFF_SLAB(cachep))
8456a648 1758 kmem_cache_free(cachep->freelist_cache, freelist);
1da177e4
LT
1759}
1760
97654dfa
JK
1761static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1762{
1763 struct page *page, *n;
1764
1765 list_for_each_entry_safe(page, n, list, lru) {
1766 list_del(&page->lru);
1767 slab_destroy(cachep, page);
1768 }
1769}
1770
4d268eba 1771/**
a70773dd
RD
1772 * calculate_slab_order - calculate size (page order) of slabs
1773 * @cachep: pointer to the cache that is being created
1774 * @size: size of objects to be created in this cache.
a70773dd
RD
1775 * @flags: slab allocation flags
1776 *
1777 * Also calculates the number of objects per slab.
4d268eba
PE
1778 *
1779 * This could be made much more intelligent. For now, try to avoid using
1780 * high order pages for slabs. When the gfp() functions are more friendly
1781 * towards high-order requests, this should be changed.
1782 */
a737b3e2 1783static size_t calculate_slab_order(struct kmem_cache *cachep,
2e6b3602 1784 size_t size, unsigned long flags)
4d268eba
PE
1785{
1786 size_t left_over = 0;
9888e6fa 1787 int gfporder;
4d268eba 1788
0aa817f0 1789 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
4d268eba
PE
1790 unsigned int num;
1791 size_t remainder;
1792
70f75067 1793 num = cache_estimate(gfporder, size, flags, &remainder);
4d268eba
PE
1794 if (!num)
1795 continue;
9888e6fa 1796
f315e3fa
JK
1797 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1798 if (num > SLAB_OBJ_MAX_NUM)
1799 break;
1800
b1ab41c4 1801 if (flags & CFLGS_OFF_SLAB) {
3217fd9b
JK
1802 struct kmem_cache *freelist_cache;
1803 size_t freelist_size;
1804
1805 freelist_size = num * sizeof(freelist_idx_t);
1806 freelist_cache = kmalloc_slab(freelist_size, 0u);
1807 if (!freelist_cache)
1808 continue;
1809
b1ab41c4 1810 /*
3217fd9b 1811 * Needed to avoid possible looping condition
76b342bd 1812 * in cache_grow_begin()
b1ab41c4 1813 */
3217fd9b
JK
1814 if (OFF_SLAB(freelist_cache))
1815 continue;
b1ab41c4 1816
3217fd9b
JK
1817 /* check if off slab has enough benefit */
1818 if (freelist_cache->size > cachep->size / 2)
1819 continue;
b1ab41c4 1820 }
4d268eba 1821
9888e6fa 1822 /* Found something acceptable - save it away */
4d268eba 1823 cachep->num = num;
9888e6fa 1824 cachep->gfporder = gfporder;
4d268eba
PE
1825 left_over = remainder;
1826
f78bb8ad
LT
1827 /*
1828 * A VFS-reclaimable slab tends to have most allocations
1829 * as GFP_NOFS and we really don't want to have to be allocating
1830 * higher-order pages when we are unable to shrink dcache.
1831 */
1832 if (flags & SLAB_RECLAIM_ACCOUNT)
1833 break;
1834
4d268eba
PE
1835 /*
1836 * Large number of objects is good, but very large slabs are
1837 * currently bad for the gfp()s.
1838 */
543585cc 1839 if (gfporder >= slab_max_order)
4d268eba
PE
1840 break;
1841
9888e6fa
LT
1842 /*
1843 * Acceptable internal fragmentation?
1844 */
a737b3e2 1845 if (left_over * 8 <= (PAGE_SIZE << gfporder))
4d268eba
PE
1846 break;
1847 }
1848 return left_over;
1849}
1850
bf0dea23
JK
1851static struct array_cache __percpu *alloc_kmem_cache_cpus(
1852 struct kmem_cache *cachep, int entries, int batchcount)
1853{
1854 int cpu;
1855 size_t size;
1856 struct array_cache __percpu *cpu_cache;
1857
1858 size = sizeof(void *) * entries + sizeof(struct array_cache);
85c9f4b0 1859 cpu_cache = __alloc_percpu(size, sizeof(void *));
bf0dea23
JK
1860
1861 if (!cpu_cache)
1862 return NULL;
1863
1864 for_each_possible_cpu(cpu) {
1865 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1866 entries, batchcount);
1867 }
1868
1869 return cpu_cache;
1870}
1871
bd721ea7 1872static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
f30cf7d1 1873{
97d06609 1874 if (slab_state >= FULL)
83b519e8 1875 return enable_cpucache(cachep, gfp);
2ed3a4ef 1876
bf0dea23
JK
1877 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1878 if (!cachep->cpu_cache)
1879 return 1;
1880
97d06609 1881 if (slab_state == DOWN) {
bf0dea23
JK
1882 /* Creation of first cache (kmem_cache). */
1883 set_up_node(kmem_cache, CACHE_CACHE);
2f9baa9f 1884 } else if (slab_state == PARTIAL) {
bf0dea23
JK
1885 /* For kmem_cache_node */
1886 set_up_node(cachep, SIZE_NODE);
f30cf7d1 1887 } else {
bf0dea23 1888 int node;
f30cf7d1 1889
bf0dea23
JK
1890 for_each_online_node(node) {
1891 cachep->node[node] = kmalloc_node(
1892 sizeof(struct kmem_cache_node), gfp, node);
1893 BUG_ON(!cachep->node[node]);
1894 kmem_cache_node_init(cachep->node[node]);
f30cf7d1
PE
1895 }
1896 }
bf0dea23 1897
6a67368c 1898 cachep->node[numa_mem_id()]->next_reap =
5f0985bb
JZ
1899 jiffies + REAPTIMEOUT_NODE +
1900 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
f30cf7d1
PE
1901
1902 cpu_cache_get(cachep)->avail = 0;
1903 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1904 cpu_cache_get(cachep)->batchcount = 1;
1905 cpu_cache_get(cachep)->touched = 0;
1906 cachep->batchcount = 1;
1907 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2ed3a4ef 1908 return 0;
f30cf7d1
PE
1909}
1910
12220dea
JK
1911unsigned long kmem_cache_flags(unsigned long object_size,
1912 unsigned long flags, const char *name,
1913 void (*ctor)(void *))
1914{
1915 return flags;
1916}
1917
1918struct kmem_cache *
1919__kmem_cache_alias(const char *name, size_t size, size_t align,
1920 unsigned long flags, void (*ctor)(void *))
1921{
1922 struct kmem_cache *cachep;
1923
1924 cachep = find_mergeable(size, align, flags, name, ctor);
1925 if (cachep) {
1926 cachep->refcount++;
1927
1928 /*
1929 * Adjust the object sizes so that we clear
1930 * the complete object on kzalloc.
1931 */
1932 cachep->object_size = max_t(int, cachep->object_size, size);
1933 }
1934 return cachep;
1935}
1936
b03a017b
JK
1937static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1938 size_t size, unsigned long flags)
1939{
1940 size_t left;
1941
1942 cachep->num = 0;
1943
1944 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1945 return false;
1946
1947 left = calculate_slab_order(cachep, size,
1948 flags | CFLGS_OBJFREELIST_SLAB);
1949 if (!cachep->num)
1950 return false;
1951
1952 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1953 return false;
1954
1955 cachep->colour = left / cachep->colour_off;
1956
1957 return true;
1958}
1959
158e319b
JK
1960static bool set_off_slab_cache(struct kmem_cache *cachep,
1961 size_t size, unsigned long flags)
1962{
1963 size_t left;
1964
1965 cachep->num = 0;
1966
1967 /*
3217fd9b
JK
1968 * Always use on-slab management when SLAB_NOLEAKTRACE
1969 * to avoid recursive calls into kmemleak.
158e319b 1970 */
158e319b
JK
1971 if (flags & SLAB_NOLEAKTRACE)
1972 return false;
1973
1974 /*
1975 * Size is large, assume best to place the slab management obj
1976 * off-slab (should allow better packing of objs).
1977 */
1978 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1979 if (!cachep->num)
1980 return false;
1981
1982 /*
1983 * If the slab has been placed off-slab, and we have enough space then
1984 * move it on-slab. This is at the expense of any extra colouring.
1985 */
1986 if (left >= cachep->num * sizeof(freelist_idx_t))
1987 return false;
1988
1989 cachep->colour = left / cachep->colour_off;
1990
1991 return true;
1992}
1993
1994static bool set_on_slab_cache(struct kmem_cache *cachep,
1995 size_t size, unsigned long flags)
1996{
1997 size_t left;
1998
1999 cachep->num = 0;
2000
2001 left = calculate_slab_order(cachep, size, flags);
2002 if (!cachep->num)
2003 return false;
2004
2005 cachep->colour = left / cachep->colour_off;
2006
2007 return true;
2008}
2009
1da177e4 2010/**
039363f3 2011 * __kmem_cache_create - Create a cache.
a755b76a 2012 * @cachep: cache management descriptor
1da177e4 2013 * @flags: SLAB flags
1da177e4
LT
2014 *
2015 * Returns a ptr to the cache on success, NULL on failure.
2016 * Cannot be called within a int, but can be interrupted.
20c2df83 2017 * The @ctor is run when new pages are allocated by the cache.
1da177e4 2018 *
1da177e4
LT
2019 * The flags are
2020 *
2021 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2022 * to catch references to uninitialised memory.
2023 *
2024 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2025 * for buffer overruns.
2026 *
1da177e4
LT
2027 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2028 * cacheline. This can be beneficial if you're counting cycles as closely
2029 * as davem.
2030 */
278b1bb1 2031int
8a13a4cc 2032__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
1da177e4 2033{
d4a5fca5 2034 size_t ralign = BYTES_PER_WORD;
83b519e8 2035 gfp_t gfp;
278b1bb1 2036 int err;
8a13a4cc 2037 size_t size = cachep->size;
1da177e4 2038
1da177e4 2039#if DEBUG
1da177e4
LT
2040#if FORCED_DEBUG
2041 /*
2042 * Enable redzoning and last user accounting, except for caches with
2043 * large objects, if the increased size would increase the object size
2044 * above the next power of two: caches with object sizes just above a
2045 * power of two have a significant amount of internal fragmentation.
2046 */
87a927c7
DW
2047 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2048 2 * sizeof(unsigned long long)))
b28a02de 2049 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1da177e4
LT
2050 if (!(flags & SLAB_DESTROY_BY_RCU))
2051 flags |= SLAB_POISON;
2052#endif
1da177e4 2053#endif
1da177e4 2054
a737b3e2
AM
2055 /*
2056 * Check that size is in terms of words. This is needed to avoid
1da177e4
LT
2057 * unaligned accesses for some archs when redzoning is used, and makes
2058 * sure any on-slab bufctl's are also correctly aligned.
2059 */
b28a02de
PE
2060 if (size & (BYTES_PER_WORD - 1)) {
2061 size += (BYTES_PER_WORD - 1);
2062 size &= ~(BYTES_PER_WORD - 1);
1da177e4
LT
2063 }
2064
87a927c7
DW
2065 if (flags & SLAB_RED_ZONE) {
2066 ralign = REDZONE_ALIGN;
2067 /* If redzoning, ensure that the second redzone is suitably
2068 * aligned, by adjusting the object size accordingly. */
2069 size += REDZONE_ALIGN - 1;
2070 size &= ~(REDZONE_ALIGN - 1);
2071 }
ca5f9703 2072
a44b56d3 2073 /* 3) caller mandated alignment */
8a13a4cc
CL
2074 if (ralign < cachep->align) {
2075 ralign = cachep->align;
1da177e4 2076 }
3ff84a7f
PE
2077 /* disable debug if necessary */
2078 if (ralign > __alignof__(unsigned long long))
a44b56d3 2079 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
a737b3e2 2080 /*
ca5f9703 2081 * 4) Store it.
1da177e4 2082 */
8a13a4cc 2083 cachep->align = ralign;
158e319b
JK
2084 cachep->colour_off = cache_line_size();
2085 /* Offset must be a multiple of the alignment. */
2086 if (cachep->colour_off < cachep->align)
2087 cachep->colour_off = cachep->align;
1da177e4 2088
83b519e8
PE
2089 if (slab_is_available())
2090 gfp = GFP_KERNEL;
2091 else
2092 gfp = GFP_NOWAIT;
2093
1da177e4 2094#if DEBUG
1da177e4 2095
ca5f9703
PE
2096 /*
2097 * Both debugging options require word-alignment which is calculated
2098 * into align above.
2099 */
1da177e4 2100 if (flags & SLAB_RED_ZONE) {
1da177e4 2101 /* add space for red zone words */
3ff84a7f
PE
2102 cachep->obj_offset += sizeof(unsigned long long);
2103 size += 2 * sizeof(unsigned long long);
1da177e4
LT
2104 }
2105 if (flags & SLAB_STORE_USER) {
ca5f9703 2106 /* user store requires one word storage behind the end of
87a927c7
DW
2107 * the real object. But if the second red zone needs to be
2108 * aligned to 64 bits, we must allow that much space.
1da177e4 2109 */
87a927c7
DW
2110 if (flags & SLAB_RED_ZONE)
2111 size += REDZONE_ALIGN;
2112 else
2113 size += BYTES_PER_WORD;
1da177e4 2114 }
832a15d2
JK
2115#endif
2116
7ed2f9e6
AP
2117 kasan_cache_create(cachep, &size, &flags);
2118
832a15d2
JK
2119 size = ALIGN(size, cachep->align);
2120 /*
2121 * We should restrict the number of objects in a slab to implement
2122 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2123 */
2124 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2125 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2126
2127#if DEBUG
03a2d2a3
JK
2128 /*
2129 * To activate debug pagealloc, off-slab management is necessary
2130 * requirement. In early phase of initialization, small sized slab
2131 * doesn't get initialized so it would not be possible. So, we need
2132 * to check size >= 256. It guarantees that all necessary small
2133 * sized slab is initialized in current slab initialization sequence.
2134 */
40323278 2135 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
f3a3c320
JK
2136 size >= 256 && cachep->object_size > cache_line_size()) {
2137 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2138 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2139
2140 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2141 flags |= CFLGS_OFF_SLAB;
2142 cachep->obj_offset += tmp_size - size;
2143 size = tmp_size;
2144 goto done;
2145 }
2146 }
1da177e4 2147 }
1da177e4
LT
2148#endif
2149
b03a017b
JK
2150 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2151 flags |= CFLGS_OBJFREELIST_SLAB;
2152 goto done;
2153 }
2154
158e319b 2155 if (set_off_slab_cache(cachep, size, flags)) {
1da177e4 2156 flags |= CFLGS_OFF_SLAB;
158e319b 2157 goto done;
832a15d2 2158 }
1da177e4 2159
158e319b
JK
2160 if (set_on_slab_cache(cachep, size, flags))
2161 goto done;
1da177e4 2162
158e319b 2163 return -E2BIG;
1da177e4 2164
158e319b
JK
2165done:
2166 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
1da177e4 2167 cachep->flags = flags;
a57a4988 2168 cachep->allocflags = __GFP_COMP;
a3187e43 2169 if (flags & SLAB_CACHE_DMA)
a618e89f 2170 cachep->allocflags |= GFP_DMA;
3b0efdfa 2171 cachep->size = size;
6a2d7a95 2172 cachep->reciprocal_buffer_size = reciprocal_value(size);
1da177e4 2173
40b44137
JK
2174#if DEBUG
2175 /*
2176 * If we're going to use the generic kernel_map_pages()
2177 * poisoning, then it's going to smash the contents of
2178 * the redzone and userword anyhow, so switch them off.
2179 */
2180 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2181 (cachep->flags & SLAB_POISON) &&
2182 is_debug_pagealloc_cache(cachep))
2183 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2184#endif
2185
2186 if (OFF_SLAB(cachep)) {
158e319b
JK
2187 cachep->freelist_cache =
2188 kmalloc_slab(cachep->freelist_size, 0u);
e5ac9c5a 2189 }
1da177e4 2190
278b1bb1
CL
2191 err = setup_cpu_cache(cachep, gfp);
2192 if (err) {
52b4b950 2193 __kmem_cache_release(cachep);
278b1bb1 2194 return err;
2ed3a4ef 2195 }
1da177e4 2196
278b1bb1 2197 return 0;
1da177e4 2198}
1da177e4
LT
2199
2200#if DEBUG
2201static void check_irq_off(void)
2202{
2203 BUG_ON(!irqs_disabled());
2204}
2205
2206static void check_irq_on(void)
2207{
2208 BUG_ON(irqs_disabled());
2209}
2210
18726ca8
JK
2211static void check_mutex_acquired(void)
2212{
2213 BUG_ON(!mutex_is_locked(&slab_mutex));
2214}
2215
343e0d7a 2216static void check_spinlock_acquired(struct kmem_cache *cachep)
1da177e4
LT
2217{
2218#ifdef CONFIG_SMP
2219 check_irq_off();
18bf8541 2220 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
1da177e4
LT
2221#endif
2222}
e498be7d 2223
343e0d7a 2224static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
e498be7d
CL
2225{
2226#ifdef CONFIG_SMP
2227 check_irq_off();
18bf8541 2228 assert_spin_locked(&get_node(cachep, node)->list_lock);
e498be7d
CL
2229#endif
2230}
2231
1da177e4
LT
2232#else
2233#define check_irq_off() do { } while(0)
2234#define check_irq_on() do { } while(0)
18726ca8 2235#define check_mutex_acquired() do { } while(0)
1da177e4 2236#define check_spinlock_acquired(x) do { } while(0)
e498be7d 2237#define check_spinlock_acquired_node(x, y) do { } while(0)
1da177e4
LT
2238#endif
2239
18726ca8
JK
2240static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2241 int node, bool free_all, struct list_head *list)
2242{
2243 int tofree;
2244
2245 if (!ac || !ac->avail)
2246 return;
2247
2248 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2249 if (tofree > ac->avail)
2250 tofree = (ac->avail + 1) / 2;
2251
2252 free_block(cachep, ac->entry, tofree, node, list);
2253 ac->avail -= tofree;
2254 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2255}
aab2207c 2256
1da177e4
LT
2257static void do_drain(void *arg)
2258{
a737b3e2 2259 struct kmem_cache *cachep = arg;
1da177e4 2260 struct array_cache *ac;
7d6e6d09 2261 int node = numa_mem_id();
18bf8541 2262 struct kmem_cache_node *n;
97654dfa 2263 LIST_HEAD(list);
1da177e4
LT
2264
2265 check_irq_off();
9a2dba4b 2266 ac = cpu_cache_get(cachep);
18bf8541
CL
2267 n = get_node(cachep, node);
2268 spin_lock(&n->list_lock);
97654dfa 2269 free_block(cachep, ac->entry, ac->avail, node, &list);
18bf8541 2270 spin_unlock(&n->list_lock);
97654dfa 2271 slabs_destroy(cachep, &list);
1da177e4
LT
2272 ac->avail = 0;
2273}
2274
343e0d7a 2275static void drain_cpu_caches(struct kmem_cache *cachep)
1da177e4 2276{
ce8eb6c4 2277 struct kmem_cache_node *n;
e498be7d 2278 int node;
18726ca8 2279 LIST_HEAD(list);
e498be7d 2280
15c8b6c1 2281 on_each_cpu(do_drain, cachep, 1);
1da177e4 2282 check_irq_on();
18bf8541
CL
2283 for_each_kmem_cache_node(cachep, node, n)
2284 if (n->alien)
ce8eb6c4 2285 drain_alien_cache(cachep, n->alien);
a4523a8b 2286
18726ca8
JK
2287 for_each_kmem_cache_node(cachep, node, n) {
2288 spin_lock_irq(&n->list_lock);
2289 drain_array_locked(cachep, n->shared, node, true, &list);
2290 spin_unlock_irq(&n->list_lock);
2291
2292 slabs_destroy(cachep, &list);
2293 }
1da177e4
LT
2294}
2295
ed11d9eb
CL
2296/*
2297 * Remove slabs from the list of free slabs.
2298 * Specify the number of slabs to drain in tofree.
2299 *
2300 * Returns the actual number of slabs released.
2301 */
2302static int drain_freelist(struct kmem_cache *cache,
ce8eb6c4 2303 struct kmem_cache_node *n, int tofree)
1da177e4 2304{
ed11d9eb
CL
2305 struct list_head *p;
2306 int nr_freed;
8456a648 2307 struct page *page;
1da177e4 2308
ed11d9eb 2309 nr_freed = 0;
ce8eb6c4 2310 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
1da177e4 2311
ce8eb6c4
CL
2312 spin_lock_irq(&n->list_lock);
2313 p = n->slabs_free.prev;
2314 if (p == &n->slabs_free) {
2315 spin_unlock_irq(&n->list_lock);
ed11d9eb
CL
2316 goto out;
2317 }
1da177e4 2318
8456a648 2319 page = list_entry(p, struct page, lru);
8456a648 2320 list_del(&page->lru);
07a63c41 2321 n->num_slabs--;
ed11d9eb
CL
2322 /*
2323 * Safe to drop the lock. The slab is no longer linked
2324 * to the cache.
2325 */
ce8eb6c4
CL
2326 n->free_objects -= cache->num;
2327 spin_unlock_irq(&n->list_lock);
8456a648 2328 slab_destroy(cache, page);
ed11d9eb 2329 nr_freed++;
1da177e4 2330 }
ed11d9eb
CL
2331out:
2332 return nr_freed;
1da177e4
LT
2333}
2334
d6e0b7fa 2335int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
e498be7d 2336{
18bf8541
CL
2337 int ret = 0;
2338 int node;
ce8eb6c4 2339 struct kmem_cache_node *n;
e498be7d
CL
2340
2341 drain_cpu_caches(cachep);
2342
2343 check_irq_on();
18bf8541 2344 for_each_kmem_cache_node(cachep, node, n) {
a5aa63a5 2345 drain_freelist(cachep, n, INT_MAX);
ed11d9eb 2346
ce8eb6c4
CL
2347 ret += !list_empty(&n->slabs_full) ||
2348 !list_empty(&n->slabs_partial);
e498be7d
CL
2349 }
2350 return (ret ? 1 : 0);
2351}
2352
945cf2b6 2353int __kmem_cache_shutdown(struct kmem_cache *cachep)
52b4b950
DS
2354{
2355 return __kmem_cache_shrink(cachep, false);
2356}
2357
2358void __kmem_cache_release(struct kmem_cache *cachep)
1da177e4 2359{
12c3667f 2360 int i;
ce8eb6c4 2361 struct kmem_cache_node *n;
1da177e4 2362
c7ce4f60
TG
2363 cache_random_seq_destroy(cachep);
2364
bf0dea23 2365 free_percpu(cachep->cpu_cache);
1da177e4 2366
ce8eb6c4 2367 /* NUMA: free the node structures */
18bf8541
CL
2368 for_each_kmem_cache_node(cachep, i, n) {
2369 kfree(n->shared);
2370 free_alien_cache(n->alien);
2371 kfree(n);
2372 cachep->node[i] = NULL;
12c3667f 2373 }
1da177e4 2374}
1da177e4 2375
e5ac9c5a
RT
2376/*
2377 * Get the memory for a slab management obj.
5f0985bb
JZ
2378 *
2379 * For a slab cache when the slab descriptor is off-slab, the
2380 * slab descriptor can't come from the same cache which is being created,
2381 * Because if it is the case, that means we defer the creation of
2382 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2383 * And we eventually call down to __kmem_cache_create(), which
2384 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2385 * This is a "chicken-and-egg" problem.
2386 *
2387 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2388 * which are all initialized during kmem_cache_init().
e5ac9c5a 2389 */
7e007355 2390static void *alloc_slabmgmt(struct kmem_cache *cachep,
0c3aa83e
JK
2391 struct page *page, int colour_off,
2392 gfp_t local_flags, int nodeid)
1da177e4 2393{
7e007355 2394 void *freelist;
0c3aa83e 2395 void *addr = page_address(page);
b28a02de 2396
2e6b3602
JK
2397 page->s_mem = addr + colour_off;
2398 page->active = 0;
2399
b03a017b
JK
2400 if (OBJFREELIST_SLAB(cachep))
2401 freelist = NULL;
2402 else if (OFF_SLAB(cachep)) {
1da177e4 2403 /* Slab management obj is off-slab. */
8456a648 2404 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
8759ec50 2405 local_flags, nodeid);
8456a648 2406 if (!freelist)
1da177e4
LT
2407 return NULL;
2408 } else {
2e6b3602
JK
2409 /* We will use last bytes at the slab for freelist */
2410 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2411 cachep->freelist_size;
1da177e4 2412 }
2e6b3602 2413
8456a648 2414 return freelist;
1da177e4
LT
2415}
2416
7cc68973 2417static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
1da177e4 2418{
a41adfaa 2419 return ((freelist_idx_t *)page->freelist)[idx];
e5c58dfd
JK
2420}
2421
2422static inline void set_free_obj(struct page *page,
7cc68973 2423 unsigned int idx, freelist_idx_t val)
e5c58dfd 2424{
a41adfaa 2425 ((freelist_idx_t *)(page->freelist))[idx] = val;
1da177e4
LT
2426}
2427
10b2e9e8 2428static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
1da177e4 2429{
10b2e9e8 2430#if DEBUG
1da177e4
LT
2431 int i;
2432
2433 for (i = 0; i < cachep->num; i++) {
8456a648 2434 void *objp = index_to_obj(cachep, page, i);
10b2e9e8 2435
1da177e4
LT
2436 if (cachep->flags & SLAB_STORE_USER)
2437 *dbg_userword(cachep, objp) = NULL;
2438
2439 if (cachep->flags & SLAB_RED_ZONE) {
2440 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2441 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2442 }
2443 /*
a737b3e2
AM
2444 * Constructors are not allowed to allocate memory from the same
2445 * cache which they are a constructor for. Otherwise, deadlock.
2446 * They must also be threaded.
1da177e4 2447 */
7ed2f9e6
AP
2448 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2449 kasan_unpoison_object_data(cachep,
2450 objp + obj_offset(cachep));
51cc5068 2451 cachep->ctor(objp + obj_offset(cachep));
7ed2f9e6
AP
2452 kasan_poison_object_data(
2453 cachep, objp + obj_offset(cachep));
2454 }
1da177e4
LT
2455
2456 if (cachep->flags & SLAB_RED_ZONE) {
2457 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
756a025f 2458 slab_error(cachep, "constructor overwrote the end of an object");
1da177e4 2459 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
756a025f 2460 slab_error(cachep, "constructor overwrote the start of an object");
1da177e4 2461 }
40b44137
JK
2462 /* need to poison the objs? */
2463 if (cachep->flags & SLAB_POISON) {
2464 poison_obj(cachep, objp, POISON_FREE);
2465 slab_kernel_map(cachep, objp, 0, 0);
2466 }
10b2e9e8 2467 }
1da177e4 2468#endif
10b2e9e8
JK
2469}
2470
c7ce4f60
TG
2471#ifdef CONFIG_SLAB_FREELIST_RANDOM
2472/* Hold information during a freelist initialization */
2473union freelist_init_state {
2474 struct {
2475 unsigned int pos;
7c00fce9 2476 unsigned int *list;
c7ce4f60
TG
2477 unsigned int count;
2478 unsigned int rand;
2479 };
2480 struct rnd_state rnd_state;
2481};
2482
2483/*
2484 * Initialize the state based on the randomization methode available.
2485 * return true if the pre-computed list is available, false otherwize.
2486 */
2487static bool freelist_state_initialize(union freelist_init_state *state,
2488 struct kmem_cache *cachep,
2489 unsigned int count)
2490{
2491 bool ret;
2492 unsigned int rand;
2493
2494 /* Use best entropy available to define a random shift */
7c00fce9 2495 rand = get_random_int();
c7ce4f60
TG
2496
2497 /* Use a random state if the pre-computed list is not available */
2498 if (!cachep->random_seq) {
2499 prandom_seed_state(&state->rnd_state, rand);
2500 ret = false;
2501 } else {
2502 state->list = cachep->random_seq;
2503 state->count = count;
2504 state->pos = 0;
2505 state->rand = rand;
2506 ret = true;
2507 }
2508 return ret;
2509}
2510
2511/* Get the next entry on the list and randomize it using a random shift */
2512static freelist_idx_t next_random_slot(union freelist_init_state *state)
2513{
2514 return (state->list[state->pos++] + state->rand) % state->count;
2515}
2516
7c00fce9
TG
2517/* Swap two freelist entries */
2518static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2519{
2520 swap(((freelist_idx_t *)page->freelist)[a],
2521 ((freelist_idx_t *)page->freelist)[b]);
2522}
2523
c7ce4f60
TG
2524/*
2525 * Shuffle the freelist initialization state based on pre-computed lists.
2526 * return true if the list was successfully shuffled, false otherwise.
2527 */
2528static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2529{
7c00fce9 2530 unsigned int objfreelist = 0, i, rand, count = cachep->num;
c7ce4f60
TG
2531 union freelist_init_state state;
2532 bool precomputed;
2533
2534 if (count < 2)
2535 return false;
2536
2537 precomputed = freelist_state_initialize(&state, cachep, count);
2538
2539 /* Take a random entry as the objfreelist */
2540 if (OBJFREELIST_SLAB(cachep)) {
2541 if (!precomputed)
2542 objfreelist = count - 1;
2543 else
2544 objfreelist = next_random_slot(&state);
2545 page->freelist = index_to_obj(cachep, page, objfreelist) +
2546 obj_offset(cachep);
2547 count--;
2548 }
2549
2550 /*
2551 * On early boot, generate the list dynamically.
2552 * Later use a pre-computed list for speed.
2553 */
2554 if (!precomputed) {
7c00fce9
TG
2555 for (i = 0; i < count; i++)
2556 set_free_obj(page, i, i);
2557
2558 /* Fisher-Yates shuffle */
2559 for (i = count - 1; i > 0; i--) {
2560 rand = prandom_u32_state(&state.rnd_state);
2561 rand %= (i + 1);
2562 swap_free_obj(page, i, rand);
2563 }
c7ce4f60
TG
2564 } else {
2565 for (i = 0; i < count; i++)
2566 set_free_obj(page, i, next_random_slot(&state));
2567 }
2568
2569 if (OBJFREELIST_SLAB(cachep))
2570 set_free_obj(page, cachep->num - 1, objfreelist);
2571
2572 return true;
2573}
2574#else
2575static inline bool shuffle_freelist(struct kmem_cache *cachep,
2576 struct page *page)
2577{
2578 return false;
2579}
2580#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2581
10b2e9e8
JK
2582static void cache_init_objs(struct kmem_cache *cachep,
2583 struct page *page)
2584{
2585 int i;
7ed2f9e6 2586 void *objp;
c7ce4f60 2587 bool shuffled;
10b2e9e8
JK
2588
2589 cache_init_objs_debug(cachep, page);
2590
c7ce4f60
TG
2591 /* Try to randomize the freelist if enabled */
2592 shuffled = shuffle_freelist(cachep, page);
2593
2594 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
b03a017b
JK
2595 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2596 obj_offset(cachep);
2597 }
2598
10b2e9e8 2599 for (i = 0; i < cachep->num; i++) {
b3cbd9bf
AR
2600 objp = index_to_obj(cachep, page, i);
2601 kasan_init_slab_obj(cachep, objp);
2602
10b2e9e8 2603 /* constructor could break poison info */
7ed2f9e6 2604 if (DEBUG == 0 && cachep->ctor) {
7ed2f9e6
AP
2605 kasan_unpoison_object_data(cachep, objp);
2606 cachep->ctor(objp);
2607 kasan_poison_object_data(cachep, objp);
2608 }
10b2e9e8 2609
c7ce4f60
TG
2610 if (!shuffled)
2611 set_free_obj(page, i, i);
1da177e4 2612 }
1da177e4
LT
2613}
2614
260b61dd 2615static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
78d382d7 2616{
b1cb0982 2617 void *objp;
78d382d7 2618
e5c58dfd 2619 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
8456a648 2620 page->active++;
78d382d7 2621
d31676df
JK
2622#if DEBUG
2623 if (cachep->flags & SLAB_STORE_USER)
2624 set_store_user_dirty(cachep);
2625#endif
2626
78d382d7
MD
2627 return objp;
2628}
2629
260b61dd
JK
2630static void slab_put_obj(struct kmem_cache *cachep,
2631 struct page *page, void *objp)
78d382d7 2632{
8456a648 2633 unsigned int objnr = obj_to_index(cachep, page, objp);
78d382d7 2634#if DEBUG
16025177 2635 unsigned int i;
b1cb0982 2636
b1cb0982 2637 /* Verify double free bug */
8456a648 2638 for (i = page->active; i < cachep->num; i++) {
e5c58dfd 2639 if (get_free_obj(page, i) == objnr) {
1170532b 2640 pr_err("slab: double free detected in cache '%s', objp %p\n",
756a025f 2641 cachep->name, objp);
b1cb0982
JK
2642 BUG();
2643 }
78d382d7
MD
2644 }
2645#endif
8456a648 2646 page->active--;
b03a017b
JK
2647 if (!page->freelist)
2648 page->freelist = objp + obj_offset(cachep);
2649
e5c58dfd 2650 set_free_obj(page, page->active, objnr);
78d382d7
MD
2651}
2652
4776874f
PE
2653/*
2654 * Map pages beginning at addr to the given cache and slab. This is required
2655 * for the slab allocator to be able to lookup the cache and slab of a
ccd35fb9 2656 * virtual address for kfree, ksize, and slab debugging.
4776874f 2657 */
8456a648 2658static void slab_map_pages(struct kmem_cache *cache, struct page *page,
7e007355 2659 void *freelist)
1da177e4 2660{
a57a4988 2661 page->slab_cache = cache;
8456a648 2662 page->freelist = freelist;
1da177e4
LT
2663}
2664
2665/*
2666 * Grow (by 1) the number of slabs within a cache. This is called by
2667 * kmem_cache_alloc() when there are no active objs left in a cache.
2668 */
76b342bd
JK
2669static struct page *cache_grow_begin(struct kmem_cache *cachep,
2670 gfp_t flags, int nodeid)
1da177e4 2671{
7e007355 2672 void *freelist;
b28a02de
PE
2673 size_t offset;
2674 gfp_t local_flags;
511e3a05 2675 int page_node;
ce8eb6c4 2676 struct kmem_cache_node *n;
511e3a05 2677 struct page *page;
1da177e4 2678
a737b3e2
AM
2679 /*
2680 * Be lazy and only check for valid flags here, keeping it out of the
2681 * critical path in kmem_cache_alloc().
1da177e4 2682 */
c871ac4e 2683 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
bacdcb34 2684 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
72baeef0
MH
2685 flags &= ~GFP_SLAB_BUG_MASK;
2686 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2687 invalid_mask, &invalid_mask, flags, &flags);
2688 dump_stack();
c871ac4e 2689 }
6cb06229 2690 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
1da177e4 2691
1da177e4 2692 check_irq_off();
d0164adc 2693 if (gfpflags_allow_blocking(local_flags))
1da177e4
LT
2694 local_irq_enable();
2695
a737b3e2
AM
2696 /*
2697 * Get mem for the objs. Attempt to allocate a physical page from
2698 * 'nodeid'.
e498be7d 2699 */
511e3a05 2700 page = kmem_getpages(cachep, local_flags, nodeid);
0c3aa83e 2701 if (!page)
1da177e4
LT
2702 goto failed;
2703
511e3a05
JK
2704 page_node = page_to_nid(page);
2705 n = get_node(cachep, page_node);
03d1d43a
JK
2706
2707 /* Get colour for the slab, and cal the next value. */
2708 n->colour_next++;
2709 if (n->colour_next >= cachep->colour)
2710 n->colour_next = 0;
2711
2712 offset = n->colour_next;
2713 if (offset >= cachep->colour)
2714 offset = 0;
2715
2716 offset *= cachep->colour_off;
2717
1da177e4 2718 /* Get slab management. */
8456a648 2719 freelist = alloc_slabmgmt(cachep, page, offset,
511e3a05 2720 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
b03a017b 2721 if (OFF_SLAB(cachep) && !freelist)
1da177e4
LT
2722 goto opps1;
2723
8456a648 2724 slab_map_pages(cachep, page, freelist);
1da177e4 2725
7ed2f9e6 2726 kasan_poison_slab(page);
8456a648 2727 cache_init_objs(cachep, page);
1da177e4 2728
d0164adc 2729 if (gfpflags_allow_blocking(local_flags))
1da177e4 2730 local_irq_disable();
1da177e4 2731
76b342bd
JK
2732 return page;
2733
a737b3e2 2734opps1:
0c3aa83e 2735 kmem_freepages(cachep, page);
a737b3e2 2736failed:
d0164adc 2737 if (gfpflags_allow_blocking(local_flags))
1da177e4 2738 local_irq_disable();
76b342bd
JK
2739 return NULL;
2740}
2741
2742static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2743{
2744 struct kmem_cache_node *n;
2745 void *list = NULL;
2746
2747 check_irq_off();
2748
2749 if (!page)
2750 return;
2751
2752 INIT_LIST_HEAD(&page->lru);
2753 n = get_node(cachep, page_to_nid(page));
2754
2755 spin_lock(&n->list_lock);
2756 if (!page->active)
2757 list_add_tail(&page->lru, &(n->slabs_free));
2758 else
2759 fixup_slab_list(cachep, n, page, &list);
07a63c41
AR
2760
2761 n->num_slabs++;
76b342bd
JK
2762 STATS_INC_GROWN(cachep);
2763 n->free_objects += cachep->num - page->active;
2764 spin_unlock(&n->list_lock);
2765
2766 fixup_objfreelist_debug(cachep, &list);
1da177e4
LT
2767}
2768
2769#if DEBUG
2770
2771/*
2772 * Perform extra freeing checks:
2773 * - detect bad pointers.
2774 * - POISON/RED_ZONE checking
1da177e4
LT
2775 */
2776static void kfree_debugcheck(const void *objp)
2777{
1da177e4 2778 if (!virt_addr_valid(objp)) {
1170532b 2779 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
b28a02de
PE
2780 (unsigned long)objp);
2781 BUG();
1da177e4 2782 }
1da177e4
LT
2783}
2784
58ce1fd5
PE
2785static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2786{
b46b8f19 2787 unsigned long long redzone1, redzone2;
58ce1fd5
PE
2788
2789 redzone1 = *dbg_redzone1(cache, obj);
2790 redzone2 = *dbg_redzone2(cache, obj);
2791
2792 /*
2793 * Redzone is ok.
2794 */
2795 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2796 return;
2797
2798 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2799 slab_error(cache, "double free detected");
2800 else
2801 slab_error(cache, "memory outside object was overwritten");
2802
1170532b
JP
2803 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2804 obj, redzone1, redzone2);
58ce1fd5
PE
2805}
2806
343e0d7a 2807static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
7c0cb9c6 2808 unsigned long caller)
1da177e4 2809{
1da177e4 2810 unsigned int objnr;
8456a648 2811 struct page *page;
1da177e4 2812
80cbd911
MW
2813 BUG_ON(virt_to_cache(objp) != cachep);
2814
3dafccf2 2815 objp -= obj_offset(cachep);
1da177e4 2816 kfree_debugcheck(objp);
b49af68f 2817 page = virt_to_head_page(objp);
1da177e4 2818
1da177e4 2819 if (cachep->flags & SLAB_RED_ZONE) {
58ce1fd5 2820 verify_redzone_free(cachep, objp);
1da177e4
LT
2821 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2822 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2823 }
d31676df
JK
2824 if (cachep->flags & SLAB_STORE_USER) {
2825 set_store_user_dirty(cachep);
7c0cb9c6 2826 *dbg_userword(cachep, objp) = (void *)caller;
d31676df 2827 }
1da177e4 2828
8456a648 2829 objnr = obj_to_index(cachep, page, objp);
1da177e4
LT
2830
2831 BUG_ON(objnr >= cachep->num);
8456a648 2832 BUG_ON(objp != index_to_obj(cachep, page, objnr));
1da177e4 2833
1da177e4 2834 if (cachep->flags & SLAB_POISON) {
1da177e4 2835 poison_obj(cachep, objp, POISON_FREE);
40b44137 2836 slab_kernel_map(cachep, objp, 0, caller);
1da177e4
LT
2837 }
2838 return objp;
2839}
2840
1da177e4
LT
2841#else
2842#define kfree_debugcheck(x) do { } while(0)
2843#define cache_free_debugcheck(x,objp,z) (objp)
1da177e4
LT
2844#endif
2845
b03a017b
JK
2846static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2847 void **list)
2848{
2849#if DEBUG
2850 void *next = *list;
2851 void *objp;
2852
2853 while (next) {
2854 objp = next - obj_offset(cachep);
2855 next = *(void **)next;
2856 poison_obj(cachep, objp, POISON_FREE);
2857 }
2858#endif
2859}
2860
d8410234 2861static inline void fixup_slab_list(struct kmem_cache *cachep,
b03a017b
JK
2862 struct kmem_cache_node *n, struct page *page,
2863 void **list)
d8410234
JK
2864{
2865 /* move slabp to correct slabp list: */
2866 list_del(&page->lru);
b03a017b 2867 if (page->active == cachep->num) {
d8410234 2868 list_add(&page->lru, &n->slabs_full);
b03a017b
JK
2869 if (OBJFREELIST_SLAB(cachep)) {
2870#if DEBUG
2871 /* Poisoning will be done without holding the lock */
2872 if (cachep->flags & SLAB_POISON) {
2873 void **objp = page->freelist;
2874
2875 *objp = *list;
2876 *list = objp;
2877 }
2878#endif
2879 page->freelist = NULL;
2880 }
2881 } else
d8410234
JK
2882 list_add(&page->lru, &n->slabs_partial);
2883}
2884
f68f8ddd
JK
2885/* Try to find non-pfmemalloc slab if needed */
2886static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2887 struct page *page, bool pfmemalloc)
2888{
2889 if (!page)
2890 return NULL;
2891
2892 if (pfmemalloc)
2893 return page;
2894
2895 if (!PageSlabPfmemalloc(page))
2896 return page;
2897
2898 /* No need to keep pfmemalloc slab if we have enough free objects */
2899 if (n->free_objects > n->free_limit) {
2900 ClearPageSlabPfmemalloc(page);
2901 return page;
2902 }
2903
2904 /* Move pfmemalloc slab to the end of list to speed up next search */
2905 list_del(&page->lru);
2906 if (!page->active)
2907 list_add_tail(&page->lru, &n->slabs_free);
2908 else
2909 list_add_tail(&page->lru, &n->slabs_partial);
2910
2911 list_for_each_entry(page, &n->slabs_partial, lru) {
2912 if (!PageSlabPfmemalloc(page))
2913 return page;
2914 }
2915
2916 list_for_each_entry(page, &n->slabs_free, lru) {
2917 if (!PageSlabPfmemalloc(page))
2918 return page;
2919 }
2920
2921 return NULL;
2922}
2923
2924static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
7aa0d227
GT
2925{
2926 struct page *page;
2927
2928 page = list_first_entry_or_null(&n->slabs_partial,
2929 struct page, lru);
2930 if (!page) {
2931 n->free_touched = 1;
2932 page = list_first_entry_or_null(&n->slabs_free,
2933 struct page, lru);
2934 }
2935
f68f8ddd
JK
2936 if (sk_memalloc_socks())
2937 return get_valid_first_slab(n, page, pfmemalloc);
2938
7aa0d227
GT
2939 return page;
2940}
2941
f68f8ddd
JK
2942static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2943 struct kmem_cache_node *n, gfp_t flags)
2944{
2945 struct page *page;
2946 void *obj;
2947 void *list = NULL;
2948
2949 if (!gfp_pfmemalloc_allowed(flags))
2950 return NULL;
2951
2952 spin_lock(&n->list_lock);
2953 page = get_first_slab(n, true);
2954 if (!page) {
2955 spin_unlock(&n->list_lock);
2956 return NULL;
2957 }
2958
2959 obj = slab_get_obj(cachep, page);
2960 n->free_objects--;
2961
2962 fixup_slab_list(cachep, n, page, &list);
2963
2964 spin_unlock(&n->list_lock);
2965 fixup_objfreelist_debug(cachep, &list);
2966
2967 return obj;
2968}
2969
213b4695
JK
2970/*
2971 * Slab list should be fixed up by fixup_slab_list() for existing slab
2972 * or cache_grow_end() for new slab
2973 */
2974static __always_inline int alloc_block(struct kmem_cache *cachep,
2975 struct array_cache *ac, struct page *page, int batchcount)
2976{
2977 /*
2978 * There must be at least one object available for
2979 * allocation.
2980 */
2981 BUG_ON(page->active >= cachep->num);
2982
2983 while (page->active < cachep->num && batchcount--) {
2984 STATS_INC_ALLOCED(cachep);
2985 STATS_INC_ACTIVE(cachep);
2986 STATS_SET_HIGH(cachep);
2987
2988 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2989 }
2990
2991 return batchcount;
2992}
2993
f68f8ddd 2994static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
1da177e4
LT
2995{
2996 int batchcount;
ce8eb6c4 2997 struct kmem_cache_node *n;
801faf0d 2998 struct array_cache *ac, *shared;
1ca4cb24 2999 int node;
b03a017b 3000 void *list = NULL;
76b342bd 3001 struct page *page;
1ca4cb24 3002
1da177e4 3003 check_irq_off();
7d6e6d09 3004 node = numa_mem_id();
f68f8ddd 3005
9a2dba4b 3006 ac = cpu_cache_get(cachep);
1da177e4
LT
3007 batchcount = ac->batchcount;
3008 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
a737b3e2
AM
3009 /*
3010 * If there was little recent activity on this cache, then
3011 * perform only a partial refill. Otherwise we could generate
3012 * refill bouncing.
1da177e4
LT
3013 */
3014 batchcount = BATCHREFILL_LIMIT;
3015 }
18bf8541 3016 n = get_node(cachep, node);
e498be7d 3017
ce8eb6c4 3018 BUG_ON(ac->avail > 0 || !n);
801faf0d
JK
3019 shared = READ_ONCE(n->shared);
3020 if (!n->free_objects && (!shared || !shared->avail))
3021 goto direct_grow;
3022
ce8eb6c4 3023 spin_lock(&n->list_lock);
801faf0d 3024 shared = READ_ONCE(n->shared);
1da177e4 3025
3ded175a 3026 /* See if we can refill from the shared array */
801faf0d
JK
3027 if (shared && transfer_objects(ac, shared, batchcount)) {
3028 shared->touched = 1;
3ded175a 3029 goto alloc_done;
44b57f1c 3030 }
3ded175a 3031
1da177e4 3032 while (batchcount > 0) {
1da177e4 3033 /* Get slab alloc is to come from. */
f68f8ddd 3034 page = get_first_slab(n, false);
7aa0d227
GT
3035 if (!page)
3036 goto must_grow;
1da177e4 3037
1da177e4 3038 check_spinlock_acquired(cachep);
714b8171 3039
213b4695 3040 batchcount = alloc_block(cachep, ac, page, batchcount);
b03a017b 3041 fixup_slab_list(cachep, n, page, &list);
1da177e4
LT
3042 }
3043
a737b3e2 3044must_grow:
ce8eb6c4 3045 n->free_objects -= ac->avail;
a737b3e2 3046alloc_done:
ce8eb6c4 3047 spin_unlock(&n->list_lock);
b03a017b 3048 fixup_objfreelist_debug(cachep, &list);
1da177e4 3049
801faf0d 3050direct_grow:
1da177e4 3051 if (unlikely(!ac->avail)) {
f68f8ddd
JK
3052 /* Check if we can use obj in pfmemalloc slab */
3053 if (sk_memalloc_socks()) {
3054 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3055
3056 if (obj)
3057 return obj;
3058 }
3059
76b342bd 3060 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
e498be7d 3061
76b342bd
JK
3062 /*
3063 * cache_grow_begin() can reenable interrupts,
3064 * then ac could change.
3065 */
9a2dba4b 3066 ac = cpu_cache_get(cachep);
213b4695
JK
3067 if (!ac->avail && page)
3068 alloc_block(cachep, ac, page, batchcount);
3069 cache_grow_end(cachep, page);
072bb0aa 3070
213b4695 3071 if (!ac->avail)
1da177e4 3072 return NULL;
1da177e4
LT
3073 }
3074 ac->touched = 1;
072bb0aa 3075
f68f8ddd 3076 return ac->entry[--ac->avail];
1da177e4
LT
3077}
3078
a737b3e2
AM
3079static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3080 gfp_t flags)
1da177e4 3081{
d0164adc 3082 might_sleep_if(gfpflags_allow_blocking(flags));
1da177e4
LT
3083}
3084
3085#if DEBUG
a737b3e2 3086static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
7c0cb9c6 3087 gfp_t flags, void *objp, unsigned long caller)
1da177e4 3088{
b28a02de 3089 if (!objp)
1da177e4 3090 return objp;
b28a02de 3091 if (cachep->flags & SLAB_POISON) {
1da177e4 3092 check_poison_obj(cachep, objp);
40b44137 3093 slab_kernel_map(cachep, objp, 1, 0);
1da177e4
LT
3094 poison_obj(cachep, objp, POISON_INUSE);
3095 }
3096 if (cachep->flags & SLAB_STORE_USER)
7c0cb9c6 3097 *dbg_userword(cachep, objp) = (void *)caller;
1da177e4
LT
3098
3099 if (cachep->flags & SLAB_RED_ZONE) {
a737b3e2
AM
3100 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3101 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
756a025f 3102 slab_error(cachep, "double free, or memory outside object was overwritten");
1170532b
JP
3103 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3104 objp, *dbg_redzone1(cachep, objp),
3105 *dbg_redzone2(cachep, objp));
1da177e4
LT
3106 }
3107 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3108 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3109 }
03787301 3110
3dafccf2 3111 objp += obj_offset(cachep);
4f104934 3112 if (cachep->ctor && cachep->flags & SLAB_POISON)
51cc5068 3113 cachep->ctor(objp);
7ea466f2
TH
3114 if (ARCH_SLAB_MINALIGN &&
3115 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
1170532b 3116 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
c225150b 3117 objp, (int)ARCH_SLAB_MINALIGN);
a44b56d3 3118 }
1da177e4
LT
3119 return objp;
3120}
3121#else
3122#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3123#endif
3124
343e0d7a 3125static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
1da177e4 3126{
b28a02de 3127 void *objp;
1da177e4
LT
3128 struct array_cache *ac;
3129
5c382300 3130 check_irq_off();
8a8b6502 3131
9a2dba4b 3132 ac = cpu_cache_get(cachep);
1da177e4 3133 if (likely(ac->avail)) {
1da177e4 3134 ac->touched = 1;
f68f8ddd 3135 objp = ac->entry[--ac->avail];
072bb0aa 3136
f68f8ddd
JK
3137 STATS_INC_ALLOCHIT(cachep);
3138 goto out;
1da177e4 3139 }
072bb0aa
MG
3140
3141 STATS_INC_ALLOCMISS(cachep);
f68f8ddd 3142 objp = cache_alloc_refill(cachep, flags);
072bb0aa
MG
3143 /*
3144 * the 'ac' may be updated by cache_alloc_refill(),
3145 * and kmemleak_erase() requires its correct value.
3146 */
3147 ac = cpu_cache_get(cachep);
3148
3149out:
d5cff635
CM
3150 /*
3151 * To avoid a false negative, if an object that is in one of the
3152 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3153 * treat the array pointers as a reference to the object.
3154 */
f3d8b53a
O
3155 if (objp)
3156 kmemleak_erase(&ac->entry[ac->avail]);
5c382300
AK
3157 return objp;
3158}
3159
e498be7d 3160#ifdef CONFIG_NUMA
c61afb18 3161/*
2ad654bc 3162 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
c61afb18
PJ
3163 *
3164 * If we are in_interrupt, then process context, including cpusets and
3165 * mempolicy, may not apply and should not be used for allocation policy.
3166 */
3167static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3168{
3169 int nid_alloc, nid_here;
3170
765c4507 3171 if (in_interrupt() || (flags & __GFP_THISNODE))
c61afb18 3172 return NULL;
7d6e6d09 3173 nid_alloc = nid_here = numa_mem_id();
c61afb18 3174 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
6adef3eb 3175 nid_alloc = cpuset_slab_spread_node();
c61afb18 3176 else if (current->mempolicy)
2a389610 3177 nid_alloc = mempolicy_slab_node();
c61afb18 3178 if (nid_alloc != nid_here)
8b98c169 3179 return ____cache_alloc_node(cachep, flags, nid_alloc);
c61afb18
PJ
3180 return NULL;
3181}
3182
765c4507
CL
3183/*
3184 * Fallback function if there was no memory available and no objects on a
3c517a61 3185 * certain node and fall back is permitted. First we scan all the
6a67368c 3186 * available node for available objects. If that fails then we
3c517a61
CL
3187 * perform an allocation without specifying a node. This allows the page
3188 * allocator to do its reclaim / fallback magic. We then insert the
3189 * slab into the proper nodelist and then allocate from it.
765c4507 3190 */
8c8cc2c1 3191static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
765c4507 3192{
8c8cc2c1 3193 struct zonelist *zonelist;
dd1a239f 3194 struct zoneref *z;
54a6eb5c
MG
3195 struct zone *zone;
3196 enum zone_type high_zoneidx = gfp_zone(flags);
765c4507 3197 void *obj = NULL;
76b342bd 3198 struct page *page;
3c517a61 3199 int nid;
cc9a6c87 3200 unsigned int cpuset_mems_cookie;
8c8cc2c1
PE
3201
3202 if (flags & __GFP_THISNODE)
3203 return NULL;
3204
cc9a6c87 3205retry_cpuset:
d26914d1 3206 cpuset_mems_cookie = read_mems_allowed_begin();
2a389610 3207 zonelist = node_zonelist(mempolicy_slab_node(), flags);
cc9a6c87 3208
3c517a61
CL
3209retry:
3210 /*
3211 * Look through allowed nodes for objects available
3212 * from existing per node queues.
3213 */
54a6eb5c
MG
3214 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3215 nid = zone_to_nid(zone);
aedb0eb1 3216
061d7074 3217 if (cpuset_zone_allowed(zone, flags) &&
18bf8541
CL
3218 get_node(cache, nid) &&
3219 get_node(cache, nid)->free_objects) {
3c517a61 3220 obj = ____cache_alloc_node(cache,
4167e9b2 3221 gfp_exact_node(flags), nid);
481c5346
CL
3222 if (obj)
3223 break;
3224 }
3c517a61
CL
3225 }
3226
cfce6604 3227 if (!obj) {
3c517a61
CL
3228 /*
3229 * This allocation will be performed within the constraints
3230 * of the current cpuset / memory policy requirements.
3231 * We may trigger various forms of reclaim on the allowed
3232 * set and go into memory reserves if necessary.
3233 */
76b342bd
JK
3234 page = cache_grow_begin(cache, flags, numa_mem_id());
3235 cache_grow_end(cache, page);
3236 if (page) {
3237 nid = page_to_nid(page);
511e3a05
JK
3238 obj = ____cache_alloc_node(cache,
3239 gfp_exact_node(flags), nid);
0c3aa83e 3240
3c517a61 3241 /*
511e3a05
JK
3242 * Another processor may allocate the objects in
3243 * the slab since we are not holding any locks.
3c517a61 3244 */
511e3a05
JK
3245 if (!obj)
3246 goto retry;
3c517a61 3247 }
aedb0eb1 3248 }
cc9a6c87 3249
d26914d1 3250 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
cc9a6c87 3251 goto retry_cpuset;
765c4507
CL
3252 return obj;
3253}
3254
e498be7d
CL
3255/*
3256 * A interface to enable slab creation on nodeid
1da177e4 3257 */
8b98c169 3258static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
a737b3e2 3259 int nodeid)
e498be7d 3260{
8456a648 3261 struct page *page;
ce8eb6c4 3262 struct kmem_cache_node *n;
213b4695 3263 void *obj = NULL;
b03a017b 3264 void *list = NULL;
b28a02de 3265
7c3fbbdd 3266 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
18bf8541 3267 n = get_node(cachep, nodeid);
ce8eb6c4 3268 BUG_ON(!n);
b28a02de 3269
ca3b9b91 3270 check_irq_off();
ce8eb6c4 3271 spin_lock(&n->list_lock);
f68f8ddd 3272 page = get_first_slab(n, false);
7aa0d227
GT
3273 if (!page)
3274 goto must_grow;
b28a02de 3275
b28a02de 3276 check_spinlock_acquired_node(cachep, nodeid);
b28a02de
PE
3277
3278 STATS_INC_NODEALLOCS(cachep);
3279 STATS_INC_ACTIVE(cachep);
3280 STATS_SET_HIGH(cachep);
3281
8456a648 3282 BUG_ON(page->active == cachep->num);
b28a02de 3283
260b61dd 3284 obj = slab_get_obj(cachep, page);
ce8eb6c4 3285 n->free_objects--;
b28a02de 3286
b03a017b 3287 fixup_slab_list(cachep, n, page, &list);
e498be7d 3288
ce8eb6c4 3289 spin_unlock(&n->list_lock);
b03a017b 3290 fixup_objfreelist_debug(cachep, &list);
213b4695 3291 return obj;
e498be7d 3292
a737b3e2 3293must_grow:
ce8eb6c4 3294 spin_unlock(&n->list_lock);
76b342bd 3295 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
213b4695
JK
3296 if (page) {
3297 /* This slab isn't counted yet so don't update free_objects */
3298 obj = slab_get_obj(cachep, page);
3299 }
76b342bd 3300 cache_grow_end(cachep, page);
1da177e4 3301
213b4695 3302 return obj ? obj : fallback_alloc(cachep, flags);
e498be7d 3303}
8c8cc2c1 3304
8c8cc2c1 3305static __always_inline void *
48356303 3306slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
7c0cb9c6 3307 unsigned long caller)
8c8cc2c1
PE
3308{
3309 unsigned long save_flags;
3310 void *ptr;
7d6e6d09 3311 int slab_node = numa_mem_id();
8c8cc2c1 3312
dcce284a 3313 flags &= gfp_allowed_mask;
011eceaf
JDB
3314 cachep = slab_pre_alloc_hook(cachep, flags);
3315 if (unlikely(!cachep))
824ebef1
AM
3316 return NULL;
3317
8c8cc2c1
PE
3318 cache_alloc_debugcheck_before(cachep, flags);
3319 local_irq_save(save_flags);
3320
eacbbae3 3321 if (nodeid == NUMA_NO_NODE)
7d6e6d09 3322 nodeid = slab_node;
8c8cc2c1 3323
18bf8541 3324 if (unlikely(!get_node(cachep, nodeid))) {
8c8cc2c1
PE
3325 /* Node not bootstrapped yet */
3326 ptr = fallback_alloc(cachep, flags);
3327 goto out;
3328 }
3329
7d6e6d09 3330 if (nodeid == slab_node) {
8c8cc2c1
PE
3331 /*
3332 * Use the locally cached objects if possible.
3333 * However ____cache_alloc does not allow fallback
3334 * to other nodes. It may fail while we still have
3335 * objects on other nodes available.
3336 */
3337 ptr = ____cache_alloc(cachep, flags);
3338 if (ptr)
3339 goto out;
3340 }
3341 /* ___cache_alloc_node can fall back to other nodes */
3342 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3343 out:
3344 local_irq_restore(save_flags);
3345 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3346
d5e3ed66
JDB
3347 if (unlikely(flags & __GFP_ZERO) && ptr)
3348 memset(ptr, 0, cachep->object_size);
d07dbea4 3349
d5e3ed66 3350 slab_post_alloc_hook(cachep, flags, 1, &ptr);
8c8cc2c1
PE
3351 return ptr;
3352}
3353
3354static __always_inline void *
3355__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3356{
3357 void *objp;
3358
2ad654bc 3359 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
8c8cc2c1
PE
3360 objp = alternate_node_alloc(cache, flags);
3361 if (objp)
3362 goto out;
3363 }
3364 objp = ____cache_alloc(cache, flags);
3365
3366 /*
3367 * We may just have run out of memory on the local node.
3368 * ____cache_alloc_node() knows how to locate memory on other nodes
3369 */
7d6e6d09
LS
3370 if (!objp)
3371 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
8c8cc2c1
PE
3372
3373 out:
3374 return objp;
3375}
3376#else
3377
3378static __always_inline void *
3379__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3380{
3381 return ____cache_alloc(cachep, flags);
3382}
3383
3384#endif /* CONFIG_NUMA */
3385
3386static __always_inline void *
48356303 3387slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
8c8cc2c1
PE
3388{
3389 unsigned long save_flags;
3390 void *objp;
3391
dcce284a 3392 flags &= gfp_allowed_mask;
011eceaf
JDB
3393 cachep = slab_pre_alloc_hook(cachep, flags);
3394 if (unlikely(!cachep))
824ebef1
AM
3395 return NULL;
3396
8c8cc2c1
PE
3397 cache_alloc_debugcheck_before(cachep, flags);
3398 local_irq_save(save_flags);
3399 objp = __do_cache_alloc(cachep, flags);
3400 local_irq_restore(save_flags);
3401 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3402 prefetchw(objp);
3403
d5e3ed66
JDB
3404 if (unlikely(flags & __GFP_ZERO) && objp)
3405 memset(objp, 0, cachep->object_size);
d07dbea4 3406
d5e3ed66 3407 slab_post_alloc_hook(cachep, flags, 1, &objp);
8c8cc2c1
PE
3408 return objp;
3409}
e498be7d
CL
3410
3411/*
5f0985bb 3412 * Caller needs to acquire correct kmem_cache_node's list_lock
97654dfa 3413 * @list: List of detached free slabs should be freed by caller
e498be7d 3414 */
97654dfa
JK
3415static void free_block(struct kmem_cache *cachep, void **objpp,
3416 int nr_objects, int node, struct list_head *list)
1da177e4
LT
3417{
3418 int i;
25c063fb 3419 struct kmem_cache_node *n = get_node(cachep, node);
6052b788
JK
3420 struct page *page;
3421
3422 n->free_objects += nr_objects;
1da177e4
LT
3423
3424 for (i = 0; i < nr_objects; i++) {
072bb0aa 3425 void *objp;
8456a648 3426 struct page *page;
1da177e4 3427
072bb0aa
MG
3428 objp = objpp[i];
3429
8456a648 3430 page = virt_to_head_page(objp);
8456a648 3431 list_del(&page->lru);
ff69416e 3432 check_spinlock_acquired_node(cachep, node);
260b61dd 3433 slab_put_obj(cachep, page, objp);
1da177e4 3434 STATS_DEC_ACTIVE(cachep);
1da177e4
LT
3435
3436 /* fixup slab chains */
6052b788
JK
3437 if (page->active == 0)
3438 list_add(&page->lru, &n->slabs_free);
3439 else {
1da177e4
LT
3440 /* Unconditionally move a slab to the end of the
3441 * partial list on free - maximum time for the
3442 * other objects to be freed, too.
3443 */
8456a648 3444 list_add_tail(&page->lru, &n->slabs_partial);
1da177e4
LT
3445 }
3446 }
6052b788
JK
3447
3448 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3449 n->free_objects -= cachep->num;
3450
3451 page = list_last_entry(&n->slabs_free, struct page, lru);
de24baec 3452 list_move(&page->lru, list);
07a63c41 3453 n->num_slabs--;
6052b788 3454 }
1da177e4
LT
3455}
3456
343e0d7a 3457static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
1da177e4
LT
3458{
3459 int batchcount;
ce8eb6c4 3460 struct kmem_cache_node *n;
7d6e6d09 3461 int node = numa_mem_id();
97654dfa 3462 LIST_HEAD(list);
1da177e4
LT
3463
3464 batchcount = ac->batchcount;
260b61dd 3465
1da177e4 3466 check_irq_off();
18bf8541 3467 n = get_node(cachep, node);
ce8eb6c4
CL
3468 spin_lock(&n->list_lock);
3469 if (n->shared) {
3470 struct array_cache *shared_array = n->shared;
b28a02de 3471 int max = shared_array->limit - shared_array->avail;
1da177e4
LT
3472 if (max) {
3473 if (batchcount > max)
3474 batchcount = max;
e498be7d 3475 memcpy(&(shared_array->entry[shared_array->avail]),
b28a02de 3476 ac->entry, sizeof(void *) * batchcount);
1da177e4
LT
3477 shared_array->avail += batchcount;
3478 goto free_done;
3479 }
3480 }
3481
97654dfa 3482 free_block(cachep, ac->entry, batchcount, node, &list);
a737b3e2 3483free_done:
1da177e4
LT
3484#if STATS
3485 {
3486 int i = 0;
73c0219d 3487 struct page *page;
1da177e4 3488
73c0219d 3489 list_for_each_entry(page, &n->slabs_free, lru) {
8456a648 3490 BUG_ON(page->active);
1da177e4
LT
3491
3492 i++;
1da177e4
LT
3493 }
3494 STATS_SET_FREEABLE(cachep, i);
3495 }
3496#endif
ce8eb6c4 3497 spin_unlock(&n->list_lock);
97654dfa 3498 slabs_destroy(cachep, &list);
1da177e4 3499 ac->avail -= batchcount;
a737b3e2 3500 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
1da177e4
LT
3501}
3502
3503/*
a737b3e2
AM
3504 * 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.
1da177e4 3506 */
a947eb95 3507static inline void __cache_free(struct kmem_cache *cachep, void *objp,
7c0cb9c6 3508 unsigned long caller)
1da177e4 3509{
55834c59
AP
3510 /* Put the object into the quarantine, don't touch it for now. */
3511 if (kasan_slab_free(cachep, objp))
3512 return;
3513
3514 ___cache_free(cachep, objp, caller);
3515}
1da177e4 3516
55834c59
AP
3517void ___cache_free(struct kmem_cache *cachep, void *objp,
3518 unsigned long caller)
3519{
3520 struct array_cache *ac = cpu_cache_get(cachep);
7ed2f9e6 3521
1da177e4 3522 check_irq_off();
d5cff635 3523 kmemleak_free_recursive(objp, cachep->flags);
a947eb95 3524 objp = cache_free_debugcheck(cachep, objp, caller);
1da177e4 3525
8c138bc0 3526 kmemcheck_slab_free(cachep, objp, cachep->object_size);
c175eea4 3527
1807a1aa
SS
3528 /*
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 */
b6e68bc1 3535 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
729bd0b7
PE
3536 return;
3537
3d880194 3538 if (ac->avail < ac->limit) {
1da177e4 3539 STATS_INC_FREEHIT(cachep);
1da177e4
LT
3540 } else {
3541 STATS_INC_FREEMISS(cachep);
3542 cache_flusharray(cachep, ac);
1da177e4 3543 }
42c8c99c 3544
f68f8ddd
JK
3545 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;
1da177e4
LT
3555}
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 */
343e0d7a 3565void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
1da177e4 3566{
48356303 3567 void *ret = slab_alloc(cachep, flags, _RET_IP_);
36555751 3568
505f5dcb 3569 kasan_slab_alloc(cachep, ret, flags);
ca2b84cb 3570 trace_kmem_cache_alloc(_RET_IP_, ret,
8c138bc0 3571 cachep->object_size, cachep->size, flags);
36555751
EGM
3572
3573 return ret;
1da177e4
LT
3574}
3575EXPORT_SYMBOL(kmem_cache_alloc);
3576
7b0501dd
JDB
3577static __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
865762a8 3587int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2a777eac 3588 void **p)
484748f0 3589{
2a777eac
JDB
3590 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
2a777eac
JDB
3602 if (unlikely(!objp))
3603 goto error;
3604 p[i] = objp;
3605 }
3606 local_irq_enable();
3607
7b0501dd
JDB
3608 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3609
2a777eac
JDB
3610 /* 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();
7b0501dd 3620 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
2a777eac
JDB
3621 slab_post_alloc_hook(s, flags, i, p);
3622 __kmem_cache_free_bulk(s, i, p);
3623 return 0;
484748f0
CL
3624}
3625EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3626
0f24f128 3627#ifdef CONFIG_TRACING
85beb586 3628void *
4052147c 3629kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
36555751 3630{
85beb586
SR
3631 void *ret;
3632
48356303 3633 ret = slab_alloc(cachep, flags, _RET_IP_);
85beb586 3634
505f5dcb 3635 kasan_kmalloc(cachep, ret, size, flags);
85beb586 3636 trace_kmalloc(_RET_IP_, ret,
ff4fcd01 3637 size, cachep->size, flags);
85beb586 3638 return ret;
36555751 3639}
85beb586 3640EXPORT_SYMBOL(kmem_cache_alloc_trace);
36555751
EGM
3641#endif
3642
1da177e4 3643#ifdef CONFIG_NUMA
d0d04b78
ZL
3644/**
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 */
8b98c169
CH
3655void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3656{
48356303 3657 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
36555751 3658
505f5dcb 3659 kasan_slab_alloc(cachep, ret, flags);
ca2b84cb 3660 trace_kmem_cache_alloc_node(_RET_IP_, ret,
8c138bc0 3661 cachep->object_size, cachep->size,
ca2b84cb 3662 flags, nodeid);
36555751
EGM
3663
3664 return ret;
8b98c169 3665}
1da177e4
LT
3666EXPORT_SYMBOL(kmem_cache_alloc_node);
3667
0f24f128 3668#ifdef CONFIG_TRACING
4052147c 3669void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
85beb586 3670 gfp_t flags,
4052147c
EG
3671 int nodeid,
3672 size_t size)
36555751 3673{
85beb586
SR
3674 void *ret;
3675
592f4145 3676 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
505f5dcb
AP
3677
3678 kasan_kmalloc(cachep, ret, size, flags);
85beb586 3679 trace_kmalloc_node(_RET_IP_, ret,
ff4fcd01 3680 size, cachep->size,
85beb586
SR
3681 flags, nodeid);
3682 return ret;
36555751 3683}
85beb586 3684EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
36555751
EGM
3685#endif
3686
8b98c169 3687static __always_inline void *
7c0cb9c6 3688__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
97e2bde4 3689{
343e0d7a 3690 struct kmem_cache *cachep;
7ed2f9e6 3691 void *ret;
97e2bde4 3692
2c59dd65 3693 cachep = kmalloc_slab(size, flags);
6cb8f913
CL
3694 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3695 return cachep;
7ed2f9e6 3696 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
505f5dcb 3697 kasan_kmalloc(cachep, ret, size, flags);
7ed2f9e6
AP
3698
3699 return ret;
97e2bde4 3700}
8b98c169 3701
8b98c169
CH
3702void *__kmalloc_node(size_t size, gfp_t flags, int node)
3703{
7c0cb9c6 3704 return __do_kmalloc_node(size, flags, node, _RET_IP_);
8b98c169 3705}
dbe5e69d 3706EXPORT_SYMBOL(__kmalloc_node);
8b98c169
CH
3707
3708void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
ce71e27c 3709 int node, unsigned long caller)
8b98c169 3710{
7c0cb9c6 3711 return __do_kmalloc_node(size, flags, node, caller);
8b98c169
CH
3712}
3713EXPORT_SYMBOL(__kmalloc_node_track_caller);
8b98c169 3714#endif /* CONFIG_NUMA */
1da177e4
LT
3715
3716/**
800590f5 3717 * __do_kmalloc - allocate memory
1da177e4 3718 * @size: how many bytes of memory are required.
800590f5 3719 * @flags: the type of memory to allocate (see kmalloc).
911851e6 3720 * @caller: function caller for debug tracking of the caller
1da177e4 3721 */
7fd6b141 3722static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
7c0cb9c6 3723 unsigned long caller)
1da177e4 3724{
343e0d7a 3725 struct kmem_cache *cachep;
36555751 3726 void *ret;
1da177e4 3727
2c59dd65 3728 cachep = kmalloc_slab(size, flags);
a5c96d8a
LT
3729 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3730 return cachep;
48356303 3731 ret = slab_alloc(cachep, flags, caller);
36555751 3732
505f5dcb 3733 kasan_kmalloc(cachep, ret, size, flags);
7c0cb9c6 3734 trace_kmalloc(caller, ret,
3b0efdfa 3735 size, cachep->size, flags);
36555751
EGM
3736
3737 return ret;
7fd6b141
PE
3738}
3739
7fd6b141
PE
3740void *__kmalloc(size_t size, gfp_t flags)
3741{
7c0cb9c6 3742 return __do_kmalloc(size, flags, _RET_IP_);
1da177e4
LT
3743}
3744EXPORT_SYMBOL(__kmalloc);
3745
ce71e27c 3746void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
7fd6b141 3747{
7c0cb9c6 3748 return __do_kmalloc(size, flags, caller);
7fd6b141
PE
3749}
3750EXPORT_SYMBOL(__kmalloc_track_caller);
1d2c8eea 3751
1da177e4
LT
3752/**
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 */
343e0d7a 3760void kmem_cache_free(struct kmem_cache *cachep, void *objp)
1da177e4
LT
3761{
3762 unsigned long flags;
b9ce5ef4
GC
3763 cachep = cache_from_obj(cachep, objp);
3764 if (!cachep)
3765 return;
1da177e4
LT
3766
3767 local_irq_save(flags);
d97d476b 3768 debug_check_no_locks_freed(objp, cachep->object_size);
3ac7fe5a 3769 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
8c138bc0 3770 debug_check_no_obj_freed(objp, cachep->object_size);
7c0cb9c6 3771 __cache_free(cachep, objp, _RET_IP_);
1da177e4 3772 local_irq_restore(flags);
36555751 3773
ca2b84cb 3774 trace_kmem_cache_free(_RET_IP_, objp);
1da177e4
LT
3775}
3776EXPORT_SYMBOL(kmem_cache_free);
3777
e6cdb58d
JDB
3778void 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
ca257195
JDB
3787 if (!orig_s) /* called via kfree_bulk */
3788 s = virt_to_cache(objp);
3789 else
3790 s = cache_from_obj(orig_s, objp);
e6cdb58d
JDB
3791
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
1da177e4
LT
3804/**
3805 * kfree - free previously allocated memory
3806 * @objp: pointer returned by kmalloc.
3807 *
80e93eff
PE
3808 * If @objp is NULL, no operation is performed.
3809 *
1da177e4
LT
3810 * Don't free memory not originally allocated by kmalloc()
3811 * or you will run into trouble.
3812 */
3813void kfree(const void *objp)
3814{
343e0d7a 3815 struct kmem_cache *c;
1da177e4
LT
3816 unsigned long flags;
3817
2121db74
PE
3818 trace_kfree(_RET_IP_, objp);
3819
6cb8f913 3820 if (unlikely(ZERO_OR_NULL_PTR(objp)))
1da177e4
LT
3821 return;
3822 local_irq_save(flags);
3823 kfree_debugcheck(objp);
6ed5eb22 3824 c = virt_to_cache(objp);
8c138bc0
CL
3825 debug_check_no_locks_freed(objp, c->object_size);
3826
3827 debug_check_no_obj_freed(objp, c->object_size);
7c0cb9c6 3828 __cache_free(c, (void *)objp, _RET_IP_);
1da177e4
LT
3829 local_irq_restore(flags);
3830}
3831EXPORT_SYMBOL(kfree);
3832
e498be7d 3833/*
ce8eb6c4 3834 * This initializes kmem_cache_node or resizes various caches for all nodes.
e498be7d 3835 */
c3d332b6 3836static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
e498be7d 3837{
c3d332b6 3838 int ret;
e498be7d 3839 int node;
ce8eb6c4 3840 struct kmem_cache_node *n;
e498be7d 3841
9c09a95c 3842 for_each_online_node(node) {
c3d332b6
JK
3843 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3844 if (ret)
e498be7d
CL
3845 goto fail;
3846
e498be7d 3847 }
c3d332b6 3848
cafeb02e 3849 return 0;
0718dc2a 3850
a737b3e2 3851fail:
3b0efdfa 3852 if (!cachep->list.next) {
0718dc2a
CL
3853 /* Cache is not active yet. Roll back what we did */
3854 node--;
3855 while (node >= 0) {
18bf8541
CL
3856 n = get_node(cachep, node);
3857 if (n) {
ce8eb6c4
CL
3858 kfree(n->shared);
3859 free_alien_cache(n->alien);
3860 kfree(n);
6a67368c 3861 cachep->node[node] = NULL;
0718dc2a
CL
3862 }
3863 node--;
3864 }
3865 }
cafeb02e 3866 return -ENOMEM;
e498be7d
CL
3867}
3868
18004c5d 3869/* Always called with the slab_mutex held */
943a451a 3870static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
83b519e8 3871 int batchcount, int shared, gfp_t gfp)
1da177e4 3872{
bf0dea23
JK
3873 struct array_cache __percpu *cpu_cache, *prev;
3874 int cpu;
1da177e4 3875
bf0dea23
JK
3876 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3877 if (!cpu_cache)
d2e7b7d0
SS
3878 return -ENOMEM;
3879
bf0dea23
JK
3880 prev = cachep->cpu_cache;
3881 cachep->cpu_cache = cpu_cache;
3882 kick_all_cpus_sync();
e498be7d 3883
1da177e4 3884 check_irq_on();
1da177e4
LT
3885 cachep->batchcount = batchcount;
3886 cachep->limit = limit;
e498be7d 3887 cachep->shared = shared;
1da177e4 3888
bf0dea23 3889 if (!prev)
c3d332b6 3890 goto setup_node;
bf0dea23
JK
3891
3892 for_each_online_cpu(cpu) {
97654dfa 3893 LIST_HEAD(list);
18bf8541
CL
3894 int node;
3895 struct kmem_cache_node *n;
bf0dea23 3896 struct array_cache *ac = per_cpu_ptr(prev, cpu);
18bf8541 3897
bf0dea23 3898 node = cpu_to_mem(cpu);
18bf8541
CL
3899 n = get_node(cachep, node);
3900 spin_lock_irq(&n->list_lock);
bf0dea23 3901 free_block(cachep, ac->entry, ac->avail, node, &list);
18bf8541 3902 spin_unlock_irq(&n->list_lock);
97654dfa 3903 slabs_destroy(cachep, &list);
1da177e4 3904 }
bf0dea23
JK
3905 free_percpu(prev);
3906
c3d332b6
JK
3907setup_node:
3908 return setup_kmem_cache_nodes(cachep, gfp);
1da177e4
LT
3909}
3910
943a451a
GC
3911static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3912 int batchcount, int shared, gfp_t gfp)
3913{
3914 int ret;
426589f5 3915 struct kmem_cache *c;
943a451a
GC
3916
3917 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3918
3919 if (slab_state < FULL)
3920 return ret;
3921
3922 if ((ret < 0) || !is_root_cache(cachep))
3923 return ret;
3924
426589f5
VD
3925 lockdep_assert_held(&slab_mutex);
3926 for_each_memcg_cache(c, cachep) {
3927 /* return value determined by the root cache only */
3928 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
943a451a
GC
3929 }
3930
3931 return ret;
3932}
3933
18004c5d 3934/* Called with slab_mutex held always */
83b519e8 3935static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
1da177e4
LT
3936{
3937 int err;
943a451a
GC
3938 int limit = 0;
3939 int shared = 0;
3940 int batchcount = 0;
3941
7c00fce9 3942 err = cache_random_seq_create(cachep, cachep->num, gfp);
c7ce4f60
TG
3943 if (err)
3944 goto end;
3945
943a451a
GC
3946 if (!is_root_cache(cachep)) {
3947 struct kmem_cache *root = memcg_root_cache(cachep);
3948 limit = root->limit;
3949 shared = root->shared;
3950 batchcount = root->batchcount;
3951 }
1da177e4 3952
943a451a
GC
3953 if (limit && shared && batchcount)
3954 goto skip_setup;
a737b3e2
AM
3955 /*
3956 * The head array serves three purposes:
1da177e4
LT
3957 * - create a LIFO ordering, i.e. return objects that are cache-warm
3958 * - reduce the number of spinlock operations.
a737b3e2 3959 * - reduce the number of linked list operations on the slab and
1da177e4
LT
3960 * bufctl chains: array operations are cheaper.
3961 * The numbers are guessed, we should auto-tune as described by
3962 * Bonwick.
3963 */
3b0efdfa 3964 if (cachep->size > 131072)
1da177e4 3965 limit = 1;
3b0efdfa 3966 else if (cachep->size > PAGE_SIZE)
1da177e4 3967 limit = 8;
3b0efdfa 3968 else if (cachep->size > 1024)
1da177e4 3969 limit = 24;
3b0efdfa 3970 else if (cachep->size > 256)
1da177e4
LT
3971 limit = 54;
3972 else
3973 limit = 120;
3974
a737b3e2
AM
3975 /*
3976 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
1da177e4
LT
3977 * allocation behaviour: Most allocs on one cpu, most free operations
3978 * on another cpu. For these cases, an efficient object passing between
3979 * cpus is necessary. This is provided by a shared array. The array
3980 * replaces Bonwick's magazine layer.
3981 * On uniprocessor, it's functionally equivalent (but less efficient)
3982 * to a larger limit. Thus disabled by default.
3983 */
3984 shared = 0;
3b0efdfa 3985 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
1da177e4 3986 shared = 8;
1da177e4
LT
3987
3988#if DEBUG
a737b3e2
AM
3989 /*
3990 * With debugging enabled, large batchcount lead to excessively long
3991 * periods with disabled local interrupts. Limit the batchcount
1da177e4
LT
3992 */
3993 if (limit > 32)
3994 limit = 32;
3995#endif
943a451a
GC
3996 batchcount = (limit + 1) / 2;
3997skip_setup:
3998 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
c7ce4f60 3999end:
1da177e4 4000 if (err)
1170532b 4001 pr_err("enable_cpucache failed for %s, error %d\n",
b28a02de 4002 cachep->name, -err);
2ed3a4ef 4003 return err;
1da177e4
LT
4004}
4005
1b55253a 4006/*
ce8eb6c4
CL
4007 * Drain an array if it contains any elements taking the node lock only if
4008 * necessary. Note that the node listlock also protects the array_cache
b18e7e65 4009 * if drain_array() is used on the shared array.
1b55253a 4010 */
ce8eb6c4 4011static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
18726ca8 4012 struct array_cache *ac, int node)
1da177e4 4013{
97654dfa 4014 LIST_HEAD(list);
18726ca8
JK
4015
4016 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4017 check_mutex_acquired();
1da177e4 4018
1b55253a
CL
4019 if (!ac || !ac->avail)
4020 return;
18726ca8
JK
4021
4022 if (ac->touched) {
1da177e4 4023 ac->touched = 0;
18726ca8 4024 return;
1da177e4 4025 }
18726ca8
JK
4026
4027 spin_lock_irq(&n->list_lock);
4028 drain_array_locked(cachep, ac, node, false, &list);
4029 spin_unlock_irq(&n->list_lock);
4030
4031 slabs_destroy(cachep, &list);
1da177e4
LT
4032}
4033
4034/**
4035 * cache_reap - Reclaim memory from caches.
05fb6bf0 4036 * @w: work descriptor
1da177e4
LT
4037 *
4038 * Called from workqueue/eventd every few seconds.
4039 * Purpose:
4040 * - clear the per-cpu caches for this CPU.
4041 * - return freeable pages to the main free memory pool.
4042 *
a737b3e2
AM
4043 * If we cannot acquire the cache chain mutex then just give up - we'll try
4044 * again on the next iteration.
1da177e4 4045 */
7c5cae36 4046static void cache_reap(struct work_struct *w)
1da177e4 4047{
7a7c381d 4048 struct kmem_cache *searchp;
ce8eb6c4 4049 struct kmem_cache_node *n;
7d6e6d09 4050 int node = numa_mem_id();
bf6aede7 4051 struct delayed_work *work = to_delayed_work(w);
1da177e4 4052
18004c5d 4053 if (!mutex_trylock(&slab_mutex))
1da177e4 4054 /* Give up. Setup the next iteration. */
7c5cae36 4055 goto out;
1da177e4 4056
18004c5d 4057 list_for_each_entry(searchp, &slab_caches, list) {
1da177e4
LT
4058 check_irq_on();
4059
35386e3b 4060 /*
ce8eb6c4 4061 * We only take the node lock if absolutely necessary and we
35386e3b
CL
4062 * have established with reasonable certainty that
4063 * we can do some work if the lock was obtained.
4064 */
18bf8541 4065 n = get_node(searchp, node);
35386e3b 4066
ce8eb6c4 4067 reap_alien(searchp, n);
1da177e4 4068
18726ca8 4069 drain_array(searchp, n, cpu_cache_get(searchp), node);
1da177e4 4070
35386e3b
CL
4071 /*
4072 * These are racy checks but it does not matter
4073 * if we skip one check or scan twice.
4074 */
ce8eb6c4 4075 if (time_after(n->next_reap, jiffies))
35386e3b 4076 goto next;
1da177e4 4077
5f0985bb 4078 n->next_reap = jiffies + REAPTIMEOUT_NODE;
1da177e4 4079
18726ca8 4080 drain_array(searchp, n, n->shared, node);
1da177e4 4081
ce8eb6c4
CL
4082 if (n->free_touched)
4083 n->free_touched = 0;
ed11d9eb
CL
4084 else {
4085 int freed;
1da177e4 4086
ce8eb6c4 4087 freed = drain_freelist(searchp, n, (n->free_limit +
ed11d9eb
CL
4088 5 * searchp->num - 1) / (5 * searchp->num));
4089 STATS_ADD_REAPED(searchp, freed);
4090 }
35386e3b 4091next:
1da177e4
LT
4092 cond_resched();
4093 }
4094 check_irq_on();
18004c5d 4095 mutex_unlock(&slab_mutex);
8fce4d8e 4096 next_reap_node();
7c5cae36 4097out:
a737b3e2 4098 /* Set up the next iteration */
5f0985bb 4099 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
1da177e4
LT
4100}
4101
158a9624 4102#ifdef CONFIG_SLABINFO
0d7561c6 4103void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
1da177e4 4104{
8456a648 4105 struct page *page;
b28a02de
PE
4106 unsigned long active_objs;
4107 unsigned long num_objs;
4108 unsigned long active_slabs = 0;
4109 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
07a63c41
AR
4110 unsigned long num_slabs_partial = 0, num_slabs_free = 0;
4111 unsigned long num_slabs_full = 0;
e498be7d 4112 const char *name;
1da177e4 4113 char *error = NULL;
e498be7d 4114 int node;
ce8eb6c4 4115 struct kmem_cache_node *n;
1da177e4 4116
1da177e4
LT
4117 active_objs = 0;
4118 num_slabs = 0;
18bf8541 4119 for_each_kmem_cache_node(cachep, node, n) {
e498be7d 4120
ca3b9b91 4121 check_irq_on();
ce8eb6c4 4122 spin_lock_irq(&n->list_lock);
e498be7d 4123
07a63c41
AR
4124 num_slabs += n->num_slabs;
4125
8456a648
JK
4126 list_for_each_entry(page, &n->slabs_partial, lru) {
4127 if (page->active == cachep->num && !error)
106a74e1 4128 error = "slabs_partial accounting error";
8456a648 4129 if (!page->active && !error)
106a74e1 4130 error = "slabs_partial accounting error";
8456a648 4131 active_objs += page->active;
07a63c41 4132 num_slabs_partial++;
e498be7d 4133 }
07a63c41 4134
8456a648
JK
4135 list_for_each_entry(page, &n->slabs_free, lru) {
4136 if (page->active && !error)
106a74e1 4137 error = "slabs_free accounting error";
07a63c41 4138 num_slabs_free++;
e498be7d 4139 }
07a63c41 4140
ce8eb6c4
CL
4141 free_objects += n->free_objects;
4142 if (n->shared)
4143 shared_avail += n->shared->avail;
e498be7d 4144
ce8eb6c4 4145 spin_unlock_irq(&n->list_lock);
1da177e4 4146 }
b28a02de 4147 num_objs = num_slabs * cachep->num;
07a63c41
AR
4148 active_slabs = num_slabs - num_slabs_free;
4149 num_slabs_full = num_slabs - (num_slabs_partial + num_slabs_free);
4150 active_objs += (num_slabs_full * cachep->num);
4151
e498be7d 4152 if (num_objs - active_objs != free_objects && !error)
1da177e4
LT
4153 error = "free_objects accounting error";
4154
b28a02de 4155 name = cachep->name;
1da177e4 4156 if (error)
1170532b 4157 pr_err("slab: cache %s error: %s\n", name, error);
1da177e4 4158
0d7561c6
GC
4159 sinfo->active_objs = active_objs;
4160 sinfo->num_objs = num_objs;
4161 sinfo->active_slabs = active_slabs;
4162 sinfo->num_slabs = num_slabs;
4163 sinfo->shared_avail = shared_avail;
4164 sinfo->limit = cachep->limit;
4165 sinfo->batchcount = cachep->batchcount;
4166 sinfo->shared = cachep->shared;
4167 sinfo->objects_per_slab = cachep->num;
4168 sinfo->cache_order = cachep->gfporder;
4169}
4170
4171void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4172{
1da177e4 4173#if STATS
ce8eb6c4 4174 { /* node stats */
1da177e4
LT
4175 unsigned long high = cachep->high_mark;
4176 unsigned long allocs = cachep->num_allocations;
4177 unsigned long grown = cachep->grown;
4178 unsigned long reaped = cachep->reaped;
4179 unsigned long errors = cachep->errors;
4180 unsigned long max_freeable = cachep->max_freeable;
1da177e4 4181 unsigned long node_allocs = cachep->node_allocs;
e498be7d 4182 unsigned long node_frees = cachep->node_frees;
fb7faf33 4183 unsigned long overflows = cachep->node_overflow;
1da177e4 4184
756a025f 4185 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
e92dd4fd
JP
4186 allocs, high, grown,
4187 reaped, errors, max_freeable, node_allocs,
4188 node_frees, overflows);
1da177e4
LT
4189 }
4190 /* cpu stats */
4191 {
4192 unsigned long allochit = atomic_read(&cachep->allochit);
4193 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4194 unsigned long freehit = atomic_read(&cachep->freehit);
4195 unsigned long freemiss = atomic_read(&cachep->freemiss);
4196
4197 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
b28a02de 4198 allochit, allocmiss, freehit, freemiss);
1da177e4
LT
4199 }
4200#endif
1da177e4
LT
4201}
4202
1da177e4
LT
4203#define MAX_SLABINFO_WRITE 128
4204/**
4205 * slabinfo_write - Tuning for the slab allocator
4206 * @file: unused
4207 * @buffer: user buffer
4208 * @count: data length
4209 * @ppos: unused
4210 */
b7454ad3 4211ssize_t slabinfo_write(struct file *file, const char __user *buffer,
b28a02de 4212 size_t count, loff_t *ppos)
1da177e4 4213{
b28a02de 4214 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
1da177e4 4215 int limit, batchcount, shared, res;
7a7c381d 4216 struct kmem_cache *cachep;
b28a02de 4217
1da177e4
LT
4218 if (count > MAX_SLABINFO_WRITE)
4219 return -EINVAL;
4220 if (copy_from_user(&kbuf, buffer, count))
4221 return -EFAULT;
b28a02de 4222 kbuf[MAX_SLABINFO_WRITE] = '\0';
1da177e4
LT
4223
4224 tmp = strchr(kbuf, ' ');
4225 if (!tmp)
4226 return -EINVAL;
4227 *tmp = '\0';
4228 tmp++;
4229 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4230 return -EINVAL;
4231
4232 /* Find the cache in the chain of caches. */
18004c5d 4233 mutex_lock(&slab_mutex);
1da177e4 4234 res = -EINVAL;
18004c5d 4235 list_for_each_entry(cachep, &slab_caches, list) {
1da177e4 4236 if (!strcmp(cachep->name, kbuf)) {
a737b3e2
AM
4237 if (limit < 1 || batchcount < 1 ||
4238 batchcount > limit || shared < 0) {
e498be7d 4239 res = 0;
1da177e4 4240 } else {
e498be7d 4241 res = do_tune_cpucache(cachep, limit,
83b519e8
PE
4242 batchcount, shared,
4243 GFP_KERNEL);
1da177e4
LT
4244 }
4245 break;
4246 }
4247 }
18004c5d 4248 mutex_unlock(&slab_mutex);
1da177e4
LT
4249 if (res >= 0)
4250 res = count;
4251 return res;
4252}
871751e2
AV
4253
4254#ifdef CONFIG_DEBUG_SLAB_LEAK
4255
871751e2
AV
4256static inline int add_caller(unsigned long *n, unsigned long v)
4257{
4258 unsigned long *p;
4259 int l;
4260 if (!v)
4261 return 1;
4262 l = n[1];
4263 p = n + 2;
4264 while (l) {
4265 int i = l/2;
4266 unsigned long *q = p + 2 * i;
4267 if (*q == v) {
4268 q[1]++;
4269 return 1;
4270 }
4271 if (*q > v) {
4272 l = i;
4273 } else {
4274 p = q + 2;
4275 l -= i + 1;
4276 }
4277 }
4278 if (++n[1] == n[0])
4279 return 0;
4280 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4281 p[0] = v;
4282 p[1] = 1;
4283 return 1;
4284}
4285
8456a648
JK
4286static void handle_slab(unsigned long *n, struct kmem_cache *c,
4287 struct page *page)
871751e2
AV
4288{
4289 void *p;
d31676df
JK
4290 int i, j;
4291 unsigned long v;
b1cb0982 4292
871751e2
AV
4293 if (n[0] == n[1])
4294 return;
8456a648 4295 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
d31676df
JK
4296 bool active = true;
4297
4298 for (j = page->active; j < c->num; j++) {
4299 if (get_free_obj(page, j) == i) {
4300 active = false;
4301 break;
4302 }
4303 }
4304
4305 if (!active)
871751e2 4306 continue;
b1cb0982 4307
d31676df
JK
4308 /*
4309 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4310 * mapping is established when actual object allocation and
4311 * we could mistakenly access the unmapped object in the cpu
4312 * cache.
4313 */
4314 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4315 continue;
4316
4317 if (!add_caller(n, v))
871751e2
AV
4318 return;
4319 }
4320}
4321
4322static void show_symbol(struct seq_file *m, unsigned long address)
4323{
4324#ifdef CONFIG_KALLSYMS
871751e2 4325 unsigned long offset, size;
9281acea 4326 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
871751e2 4327
a5c43dae 4328 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
871751e2 4329 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
a5c43dae 4330 if (modname[0])
871751e2
AV
4331 seq_printf(m, " [%s]", modname);
4332 return;
4333 }
4334#endif
4335 seq_printf(m, "%p", (void *)address);
4336}
4337
4338static int leaks_show(struct seq_file *m, void *p)
4339{
0672aa7c 4340 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
8456a648 4341 struct page *page;
ce8eb6c4 4342 struct kmem_cache_node *n;
871751e2 4343 const char *name;
db845067 4344 unsigned long *x = m->private;
871751e2
AV
4345 int node;
4346 int i;
4347
4348 if (!(cachep->flags & SLAB_STORE_USER))
4349 return 0;
4350 if (!(cachep->flags & SLAB_RED_ZONE))
4351 return 0;
4352
d31676df
JK
4353 /*
4354 * Set store_user_clean and start to grab stored user information
4355 * for all objects on this cache. If some alloc/free requests comes
4356 * during the processing, information would be wrong so restart
4357 * whole processing.
4358 */
4359 do {
4360 set_store_user_clean(cachep);
4361 drain_cpu_caches(cachep);
4362
4363 x[1] = 0;
871751e2 4364
d31676df 4365 for_each_kmem_cache_node(cachep, node, n) {
871751e2 4366
d31676df
JK
4367 check_irq_on();
4368 spin_lock_irq(&n->list_lock);
871751e2 4369
d31676df
JK
4370 list_for_each_entry(page, &n->slabs_full, lru)
4371 handle_slab(x, cachep, page);
4372 list_for_each_entry(page, &n->slabs_partial, lru)
4373 handle_slab(x, cachep, page);
4374 spin_unlock_irq(&n->list_lock);
4375 }
4376 } while (!is_store_user_clean(cachep));
871751e2 4377
871751e2 4378 name = cachep->name;
db845067 4379 if (x[0] == x[1]) {
871751e2 4380 /* Increase the buffer size */
18004c5d 4381 mutex_unlock(&slab_mutex);
db845067 4382 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
871751e2
AV
4383 if (!m->private) {
4384 /* Too bad, we are really out */
db845067 4385 m->private = x;
18004c5d 4386 mutex_lock(&slab_mutex);
871751e2
AV
4387 return -ENOMEM;
4388 }
db845067
CL
4389 *(unsigned long *)m->private = x[0] * 2;
4390 kfree(x);
18004c5d 4391 mutex_lock(&slab_mutex);
871751e2
AV
4392 /* Now make sure this entry will be retried */
4393 m->count = m->size;
4394 return 0;
4395 }
db845067
CL
4396 for (i = 0; i < x[1]; i++) {
4397 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4398 show_symbol(m, x[2*i+2]);
871751e2
AV
4399 seq_putc(m, '\n');
4400 }
d2e7b7d0 4401
871751e2
AV
4402 return 0;
4403}
4404
a0ec95a8 4405static const struct seq_operations slabstats_op = {
1df3b26f 4406 .start = slab_start,
276a2439
WL
4407 .next = slab_next,
4408 .stop = slab_stop,
871751e2
AV
4409 .show = leaks_show,
4410};
a0ec95a8
AD
4411
4412static int slabstats_open(struct inode *inode, struct file *file)
4413{
b208ce32
RJ
4414 unsigned long *n;
4415
4416 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4417 if (!n)
4418 return -ENOMEM;
4419
4420 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4421
4422 return 0;
a0ec95a8
AD
4423}
4424
4425static const struct file_operations proc_slabstats_operations = {
4426 .open = slabstats_open,
4427 .read = seq_read,
4428 .llseek = seq_lseek,
4429 .release = seq_release_private,
4430};
4431#endif
4432
4433static int __init slab_proc_init(void)
4434{
4435#ifdef CONFIG_DEBUG_SLAB_LEAK
4436 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
871751e2 4437#endif
a0ec95a8
AD
4438 return 0;
4439}
4440module_init(slab_proc_init);
1da177e4
LT
4441#endif
4442
04385fc5
KC
4443#ifdef CONFIG_HARDENED_USERCOPY
4444/*
4445 * Rejects objects that are incorrectly sized.
4446 *
4447 * Returns NULL if check passes, otherwise const char * to name of cache
4448 * to indicate an error.
4449 */
4450const char *__check_heap_object(const void *ptr, unsigned long n,
4451 struct page *page)
4452{
4453 struct kmem_cache *cachep;
4454 unsigned int objnr;
4455 unsigned long offset;
4456
4457 /* Find and validate object. */
4458 cachep = page->slab_cache;
4459 objnr = obj_to_index(cachep, page, (void *)ptr);
4460 BUG_ON(objnr >= cachep->num);
4461
4462 /* Find offset within object. */
4463 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4464
4465 /* Allow address range falling entirely within object size. */
4466 if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4467 return NULL;
4468
4469 return cachep->name;
4470}
4471#endif /* CONFIG_HARDENED_USERCOPY */
4472
00e145b6
MS
4473/**
4474 * ksize - get the actual amount of memory allocated for a given object
4475 * @objp: Pointer to the object
4476 *
4477 * kmalloc may internally round up allocations and return more memory
4478 * than requested. ksize() can be used to determine the actual amount of
4479 * memory allocated. The caller may use this additional memory, even though
4480 * a smaller amount of memory was initially specified with the kmalloc call.
4481 * The caller must guarantee that objp points to a valid object previously
4482 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4483 * must not be freed during the duration of the call.
4484 */
fd76bab2 4485size_t ksize(const void *objp)
1da177e4 4486{
7ed2f9e6
AP
4487 size_t size;
4488
ef8b4520
CL
4489 BUG_ON(!objp);
4490 if (unlikely(objp == ZERO_SIZE_PTR))
00e145b6 4491 return 0;
1da177e4 4492
7ed2f9e6
AP
4493 size = virt_to_cache(objp)->object_size;
4494 /* We assume that ksize callers could use the whole allocated area,
4495 * so we need to unpoison this area.
4496 */
4ebb31a4 4497 kasan_unpoison_shadow(objp, size);
7ed2f9e6
AP
4498
4499 return size;
1da177e4 4500}
b1aabecd 4501EXPORT_SYMBOL(ksize);