net: ipv6 bind to device issue
[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
fc0abb14 71 * The global cache-chain is protected by the mutex 'cache_chain_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>
02af61bb 105#include <linux/kmemtrace.h>
1da177e4 106#include <linux/rcupdate.h>
543537bd 107#include <linux/string.h>
138ae663 108#include <linux/uaccess.h>
e498be7d 109#include <linux/nodemask.h>
d5cff635 110#include <linux/kmemleak.h>
dc85da15 111#include <linux/mempolicy.h>
fc0abb14 112#include <linux/mutex.h>
8a8b6502 113#include <linux/fault-inject.h>
e7eebaf6 114#include <linux/rtmutex.h>
6a2d7a95 115#include <linux/reciprocal_div.h>
3ac7fe5a 116#include <linux/debugobjects.h>
c175eea4 117#include <linux/kmemcheck.h>
1da177e4 118
1da177e4
LT
119#include <asm/cacheflush.h>
120#include <asm/tlbflush.h>
121#include <asm/page.h>
122
123/*
50953fe9 124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
1da177e4
LT
125 * 0 for faster, smaller code (especially in the critical paths).
126 *
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 */
132
133#ifdef CONFIG_DEBUG_SLAB
134#define DEBUG 1
135#define STATS 1
136#define FORCED_DEBUG 1
137#else
138#define DEBUG 0
139#define STATS 0
140#define FORCED_DEBUG 0
141#endif
142
1da177e4
LT
143/* Shouldn't this be in a header file somewhere? */
144#define BYTES_PER_WORD sizeof(void *)
87a927c7 145#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
1da177e4 146
1da177e4
LT
147#ifndef ARCH_KMALLOC_MINALIGN
148/*
149 * Enforce a minimum alignment for the kmalloc caches.
150 * Usually, the kmalloc caches are cache_line_size() aligned, except when
151 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
152 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
b46b8f19
DW
153 * alignment larger than the alignment of a 64-bit integer.
154 * ARCH_KMALLOC_MINALIGN allows that.
155 * Note that increasing this value may disable some debug features.
1da177e4 156 */
b46b8f19 157#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
1da177e4
LT
158#endif
159
160#ifndef ARCH_SLAB_MINALIGN
161/*
162 * Enforce a minimum alignment for all caches.
163 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
164 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
165 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
166 * some debug features.
167 */
168#define ARCH_SLAB_MINALIGN 0
169#endif
170
171#ifndef ARCH_KMALLOC_FLAGS
172#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173#endif
174
175/* Legal flag mask for kmem_cache_create(). */
176#if DEBUG
50953fe9 177# define CREATE_MASK (SLAB_RED_ZONE | \
1da177e4 178 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
ac2b898c 179 SLAB_CACHE_DMA | \
5af60839 180 SLAB_STORE_USER | \
1da177e4 181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
3ac7fe5a 182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
c175eea4 183 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
1da177e4 184#else
ac2b898c 185# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
5af60839 186 SLAB_CACHE_DMA | \
1da177e4 187 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
3ac7fe5a 188 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
c175eea4 189 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
1da177e4
LT
190#endif
191
192/*
193 * kmem_bufctl_t:
194 *
195 * Bufctl's are used for linking objs within a slab
196 * linked offsets.
197 *
198 * This implementation relies on "struct page" for locating the cache &
199 * slab an object belongs to.
200 * This allows the bufctl structure to be small (one int), but limits
201 * the number of objects a slab (not a cache) can contain when off-slab
202 * bufctls are used. The limit is the size of the largest general cache
203 * that does not use off-slab slabs.
204 * For 32bit archs with 4 kB pages, is this 56.
205 * This is not serious, as it is only for large objects, when it is unwise
206 * to have too many per slab.
207 * Note: This limit can be raised by introducing a general cache whose size
208 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
209 */
210
fa5b08d5 211typedef unsigned int kmem_bufctl_t;
1da177e4
LT
212#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
213#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
871751e2
AV
214#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
215#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
1da177e4 216
1da177e4
LT
217/*
218 * struct slab
219 *
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 */
224struct slab {
b28a02de
PE
225 struct list_head list;
226 unsigned long colouroff;
227 void *s_mem; /* including colour offset */
228 unsigned int inuse; /* num of objs active in slab */
229 kmem_bufctl_t free;
230 unsigned short nodeid;
1da177e4
LT
231};
232
233/*
234 * struct slab_rcu
235 *
236 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
237 * arrange for kmem_freepages to be called via RCU. This is useful if
238 * we need to approach a kernel structure obliquely, from its address
239 * obtained without the usual locking. We can lock the structure to
240 * stabilize it and check it's still at the given address, only if we
241 * can be sure that the memory has not been meanwhile reused for some
242 * other kind of object (which our subsystem's lock might corrupt).
243 *
244 * rcu_read_lock before reading the address, then rcu_read_unlock after
245 * taking the spinlock within the structure expected at that address.
246 *
247 * We assume struct slab_rcu can overlay struct slab when destroying.
248 */
249struct slab_rcu {
b28a02de 250 struct rcu_head head;
343e0d7a 251 struct kmem_cache *cachep;
b28a02de 252 void *addr;
1da177e4
LT
253};
254
255/*
256 * struct array_cache
257 *
1da177e4
LT
258 * Purpose:
259 * - LIFO ordering, to hand out cache-warm objects from _alloc
260 * - reduce the number of linked list operations
261 * - reduce spinlock operations
262 *
263 * The limit is stored in the per-cpu structure to reduce the data cache
264 * footprint.
265 *
266 */
267struct array_cache {
268 unsigned int avail;
269 unsigned int limit;
270 unsigned int batchcount;
271 unsigned int touched;
e498be7d 272 spinlock_t lock;
bda5b655 273 void *entry[]; /*
a737b3e2
AM
274 * Must have this definition in here for the proper
275 * alignment of array_cache. Also simplifies accessing
276 * the entries.
a737b3e2 277 */
1da177e4
LT
278};
279
a737b3e2
AM
280/*
281 * bootstrap: The caches do not work without cpuarrays anymore, but the
282 * cpuarrays are allocated from the generic caches...
1da177e4
LT
283 */
284#define BOOT_CPUCACHE_ENTRIES 1
285struct arraycache_init {
286 struct array_cache cache;
b28a02de 287 void *entries[BOOT_CPUCACHE_ENTRIES];
1da177e4
LT
288};
289
290/*
e498be7d 291 * The slab lists for all objects.
1da177e4
LT
292 */
293struct kmem_list3 {
b28a02de
PE
294 struct list_head slabs_partial; /* partial list first, better asm code */
295 struct list_head slabs_full;
296 struct list_head slabs_free;
297 unsigned long free_objects;
b28a02de 298 unsigned int free_limit;
2e1217cf 299 unsigned int colour_next; /* Per-node cache coloring */
b28a02de
PE
300 spinlock_t list_lock;
301 struct array_cache *shared; /* shared per node */
302 struct array_cache **alien; /* on other nodes */
35386e3b
CL
303 unsigned long next_reap; /* updated without locking */
304 int free_touched; /* updated without locking */
1da177e4
LT
305};
306
e498be7d
CL
307/*
308 * Need this for bootstrapping a per node allocator.
309 */
556a169d 310#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
e498be7d
CL
311struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
312#define CACHE_CACHE 0
556a169d
PE
313#define SIZE_AC MAX_NUMNODES
314#define SIZE_L3 (2 * MAX_NUMNODES)
e498be7d 315
ed11d9eb
CL
316static int drain_freelist(struct kmem_cache *cache,
317 struct kmem_list3 *l3, int tofree);
318static void free_block(struct kmem_cache *cachep, void **objpp, int len,
319 int node);
83b519e8 320static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
65f27f38 321static void cache_reap(struct work_struct *unused);
ed11d9eb 322
e498be7d 323/*
a737b3e2
AM
324 * This function must be completely optimized away if a constant is passed to
325 * it. Mostly the same as what is in linux/slab.h except it returns an index.
e498be7d 326 */
7243cc05 327static __always_inline int index_of(const size_t size)
e498be7d 328{
5ec8a847
SR
329 extern void __bad_size(void);
330
e498be7d
CL
331 if (__builtin_constant_p(size)) {
332 int i = 0;
333
334#define CACHE(x) \
335 if (size <=x) \
336 return i; \
337 else \
338 i++;
1c61fc40 339#include <linux/kmalloc_sizes.h>
e498be7d 340#undef CACHE
5ec8a847 341 __bad_size();
7243cc05 342 } else
5ec8a847 343 __bad_size();
e498be7d
CL
344 return 0;
345}
346
e0a42726
IM
347static int slab_early_init = 1;
348
e498be7d
CL
349#define INDEX_AC index_of(sizeof(struct arraycache_init))
350#define INDEX_L3 index_of(sizeof(struct kmem_list3))
1da177e4 351
5295a74c 352static void kmem_list3_init(struct kmem_list3 *parent)
e498be7d
CL
353{
354 INIT_LIST_HEAD(&parent->slabs_full);
355 INIT_LIST_HEAD(&parent->slabs_partial);
356 INIT_LIST_HEAD(&parent->slabs_free);
357 parent->shared = NULL;
358 parent->alien = NULL;
2e1217cf 359 parent->colour_next = 0;
e498be7d
CL
360 spin_lock_init(&parent->list_lock);
361 parent->free_objects = 0;
362 parent->free_touched = 0;
363}
364
a737b3e2
AM
365#define MAKE_LIST(cachep, listp, slab, nodeid) \
366 do { \
367 INIT_LIST_HEAD(listp); \
368 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
e498be7d
CL
369 } while (0)
370
a737b3e2
AM
371#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
372 do { \
e498be7d
CL
373 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
375 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
376 } while (0)
1da177e4 377
1da177e4
LT
378#define CFLGS_OFF_SLAB (0x80000000UL)
379#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
380
381#define BATCHREFILL_LIMIT 16
a737b3e2
AM
382/*
383 * Optimization question: fewer reaps means less probability for unnessary
384 * cpucache drain/refill cycles.
1da177e4 385 *
dc6f3f27 386 * OTOH the cpuarrays can contain lots of objects,
1da177e4
LT
387 * which could lock up otherwise freeable slabs.
388 */
389#define REAPTIMEOUT_CPUC (2*HZ)
390#define REAPTIMEOUT_LIST3 (4*HZ)
391
392#if STATS
393#define STATS_INC_ACTIVE(x) ((x)->num_active++)
394#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
395#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
396#define STATS_INC_GROWN(x) ((x)->grown++)
ed11d9eb 397#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
a737b3e2
AM
398#define STATS_SET_HIGH(x) \
399 do { \
400 if ((x)->num_active > (x)->high_mark) \
401 (x)->high_mark = (x)->num_active; \
402 } while (0)
1da177e4
LT
403#define STATS_INC_ERR(x) ((x)->errors++)
404#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
e498be7d 405#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
fb7faf33 406#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
a737b3e2
AM
407#define STATS_SET_FREEABLE(x, i) \
408 do { \
409 if ((x)->max_freeable < i) \
410 (x)->max_freeable = i; \
411 } while (0)
1da177e4
LT
412#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
413#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
414#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
415#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
416#else
417#define STATS_INC_ACTIVE(x) do { } while (0)
418#define STATS_DEC_ACTIVE(x) do { } while (0)
419#define STATS_INC_ALLOCED(x) do { } while (0)
420#define STATS_INC_GROWN(x) do { } while (0)
ed11d9eb 421#define STATS_ADD_REAPED(x,y) do { } while (0)
1da177e4
LT
422#define STATS_SET_HIGH(x) do { } while (0)
423#define STATS_INC_ERR(x) do { } while (0)
424#define STATS_INC_NODEALLOCS(x) do { } while (0)
e498be7d 425#define STATS_INC_NODEFREES(x) do { } while (0)
fb7faf33 426#define STATS_INC_ACOVERFLOW(x) do { } while (0)
a737b3e2 427#define STATS_SET_FREEABLE(x, i) do { } while (0)
1da177e4
LT
428#define STATS_INC_ALLOCHIT(x) do { } while (0)
429#define STATS_INC_ALLOCMISS(x) do { } while (0)
430#define STATS_INC_FREEHIT(x) do { } while (0)
431#define STATS_INC_FREEMISS(x) do { } while (0)
432#endif
433
434#if DEBUG
1da177e4 435
a737b3e2
AM
436/*
437 * memory layout of objects:
1da177e4 438 * 0 : objp
3dafccf2 439 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
1da177e4
LT
440 * the end of an object is aligned with the end of the real
441 * allocation. Catches writes behind the end of the allocation.
3dafccf2 442 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
1da177e4 443 * redzone word.
3dafccf2
MS
444 * cachep->obj_offset: The real object.
445 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
a737b3e2
AM
446 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
447 * [BYTES_PER_WORD long]
1da177e4 448 */
343e0d7a 449static int obj_offset(struct kmem_cache *cachep)
1da177e4 450{
3dafccf2 451 return cachep->obj_offset;
1da177e4
LT
452}
453
343e0d7a 454static int obj_size(struct kmem_cache *cachep)
1da177e4 455{
3dafccf2 456 return cachep->obj_size;
1da177e4
LT
457}
458
b46b8f19 459static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
1da177e4
LT
460{
461 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
b46b8f19
DW
462 return (unsigned long long*) (objp + obj_offset(cachep) -
463 sizeof(unsigned long long));
1da177e4
LT
464}
465
b46b8f19 466static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
1da177e4
LT
467{
468 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
469 if (cachep->flags & SLAB_STORE_USER)
b46b8f19
DW
470 return (unsigned long long *)(objp + cachep->buffer_size -
471 sizeof(unsigned long long) -
87a927c7 472 REDZONE_ALIGN);
b46b8f19
DW
473 return (unsigned long long *) (objp + cachep->buffer_size -
474 sizeof(unsigned long long));
1da177e4
LT
475}
476
343e0d7a 477static void **dbg_userword(struct kmem_cache *cachep, void *objp)
1da177e4
LT
478{
479 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
3dafccf2 480 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
1da177e4
LT
481}
482
483#else
484
3dafccf2
MS
485#define obj_offset(x) 0
486#define obj_size(cachep) (cachep->buffer_size)
b46b8f19
DW
487#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
488#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
1da177e4
LT
489#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
490
491#endif
492
0f24f128 493#ifdef CONFIG_TRACING
36555751
EGM
494size_t slab_buffer_size(struct kmem_cache *cachep)
495{
496 return cachep->buffer_size;
497}
498EXPORT_SYMBOL(slab_buffer_size);
499#endif
500
1da177e4
LT
501/*
502 * Do not go above this order unless 0 objects fit into the slab.
503 */
504#define BREAK_GFP_ORDER_HI 1
505#define BREAK_GFP_ORDER_LO 0
506static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
507
a737b3e2
AM
508/*
509 * Functions for storing/retrieving the cachep and or slab from the page
510 * allocator. These are used to find the slab an obj belongs to. With kfree(),
511 * these are used to find the cache which an obj belongs to.
1da177e4 512 */
065d41cb
PE
513static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
514{
515 page->lru.next = (struct list_head *)cache;
516}
517
518static inline struct kmem_cache *page_get_cache(struct page *page)
519{
d85f3385 520 page = compound_head(page);
ddc2e812 521 BUG_ON(!PageSlab(page));
065d41cb
PE
522 return (struct kmem_cache *)page->lru.next;
523}
524
525static inline void page_set_slab(struct page *page, struct slab *slab)
526{
527 page->lru.prev = (struct list_head *)slab;
528}
529
530static inline struct slab *page_get_slab(struct page *page)
531{
ddc2e812 532 BUG_ON(!PageSlab(page));
065d41cb
PE
533 return (struct slab *)page->lru.prev;
534}
1da177e4 535
6ed5eb22
PE
536static inline struct kmem_cache *virt_to_cache(const void *obj)
537{
b49af68f 538 struct page *page = virt_to_head_page(obj);
6ed5eb22
PE
539 return page_get_cache(page);
540}
541
542static inline struct slab *virt_to_slab(const void *obj)
543{
b49af68f 544 struct page *page = virt_to_head_page(obj);
6ed5eb22
PE
545 return page_get_slab(page);
546}
547
8fea4e96
PE
548static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
549 unsigned int idx)
550{
551 return slab->s_mem + cache->buffer_size * idx;
552}
553
6a2d7a95
ED
554/*
555 * We want to avoid an expensive divide : (offset / cache->buffer_size)
556 * Using the fact that buffer_size is a constant for a particular cache,
557 * we can replace (offset / cache->buffer_size) by
558 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
559 */
560static inline unsigned int obj_to_index(const struct kmem_cache *cache,
561 const struct slab *slab, void *obj)
8fea4e96 562{
6a2d7a95
ED
563 u32 offset = (obj - slab->s_mem);
564 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
8fea4e96
PE
565}
566
a737b3e2
AM
567/*
568 * These are the default caches for kmalloc. Custom caches can have other sizes.
569 */
1da177e4
LT
570struct cache_sizes malloc_sizes[] = {
571#define CACHE(x) { .cs_size = (x) },
572#include <linux/kmalloc_sizes.h>
573 CACHE(ULONG_MAX)
574#undef CACHE
575};
576EXPORT_SYMBOL(malloc_sizes);
577
578/* Must match cache_sizes above. Out of line to keep cache footprint low. */
579struct cache_names {
580 char *name;
581 char *name_dma;
582};
583
584static struct cache_names __initdata cache_names[] = {
585#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
586#include <linux/kmalloc_sizes.h>
b28a02de 587 {NULL,}
1da177e4
LT
588#undef CACHE
589};
590
591static struct arraycache_init initarray_cache __initdata =
b28a02de 592 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
1da177e4 593static struct arraycache_init initarray_generic =
b28a02de 594 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
1da177e4
LT
595
596/* internal cache of cache description objs */
343e0d7a 597static struct kmem_cache cache_cache = {
b28a02de
PE
598 .batchcount = 1,
599 .limit = BOOT_CPUCACHE_ENTRIES,
600 .shared = 1,
343e0d7a 601 .buffer_size = sizeof(struct kmem_cache),
b28a02de 602 .name = "kmem_cache",
1da177e4
LT
603};
604
056c6241
RT
605#define BAD_ALIEN_MAGIC 0x01020304ul
606
ce79ddc8
PE
607/*
608 * chicken and egg problem: delay the per-cpu array allocation
609 * until the general caches are up.
610 */
611static enum {
612 NONE,
613 PARTIAL_AC,
614 PARTIAL_L3,
615 EARLY,
616 FULL
617} g_cpucache_up;
618
619/*
620 * used by boot code to determine if it can use slab based allocator
621 */
622int slab_is_available(void)
623{
624 return g_cpucache_up >= EARLY;
625}
626
f1aaee53
AV
627#ifdef CONFIG_LOCKDEP
628
629/*
630 * Slab sometimes uses the kmalloc slabs to store the slab headers
631 * for other slabs "off slab".
632 * The locking for this is tricky in that it nests within the locks
633 * of all other slabs in a few places; to deal with this special
634 * locking we put on-slab caches into a separate lock-class.
056c6241
RT
635 *
636 * We set lock class for alien array caches which are up during init.
637 * The lock annotation will be lost if all cpus of a node goes down and
638 * then comes back up during hotplug
f1aaee53 639 */
056c6241
RT
640static struct lock_class_key on_slab_l3_key;
641static struct lock_class_key on_slab_alc_key;
642
ce79ddc8 643static void init_node_lock_keys(int q)
f1aaee53 644{
056c6241
RT
645 struct cache_sizes *s = malloc_sizes;
646
ce79ddc8
PE
647 if (g_cpucache_up != FULL)
648 return;
649
650 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
651 struct array_cache **alc;
652 struct kmem_list3 *l3;
653 int r;
654
655 l3 = s->cs_cachep->nodelists[q];
656 if (!l3 || OFF_SLAB(s->cs_cachep))
00afa758 657 continue;
ce79ddc8
PE
658 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
659 alc = l3->alien;
660 /*
661 * FIXME: This check for BAD_ALIEN_MAGIC
662 * should go away when common slab code is taught to
663 * work even without alien caches.
664 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
665 * for alloc_alien_cache,
666 */
667 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
00afa758 668 continue;
ce79ddc8
PE
669 for_each_node(r) {
670 if (alc[r])
671 lockdep_set_class(&alc[r]->lock,
672 &on_slab_alc_key);
056c6241 673 }
f1aaee53
AV
674 }
675}
ce79ddc8
PE
676
677static inline void init_lock_keys(void)
678{
679 int node;
680
681 for_each_node(node)
682 init_node_lock_keys(node);
683}
f1aaee53 684#else
ce79ddc8
PE
685static void init_node_lock_keys(int q)
686{
687}
688
056c6241 689static inline void init_lock_keys(void)
f1aaee53
AV
690{
691}
692#endif
693
8f5be20b 694/*
95402b38 695 * Guard access to the cache-chain.
8f5be20b 696 */
fc0abb14 697static DEFINE_MUTEX(cache_chain_mutex);
1da177e4
LT
698static struct list_head cache_chain;
699
1871e52c 700static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
1da177e4 701
343e0d7a 702static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
1da177e4
LT
703{
704 return cachep->array[smp_processor_id()];
705}
706
a737b3e2
AM
707static inline struct kmem_cache *__find_general_cachep(size_t size,
708 gfp_t gfpflags)
1da177e4
LT
709{
710 struct cache_sizes *csizep = malloc_sizes;
711
712#if DEBUG
713 /* This happens if someone tries to call
b28a02de
PE
714 * kmem_cache_create(), or __kmalloc(), before
715 * the generic caches are initialized.
716 */
c7e43c78 717 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
1da177e4 718#endif
6cb8f913
CL
719 if (!size)
720 return ZERO_SIZE_PTR;
721
1da177e4
LT
722 while (size > csizep->cs_size)
723 csizep++;
724
725 /*
0abf40c1 726 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
1da177e4
LT
727 * has cs_{dma,}cachep==NULL. Thus no special case
728 * for large kmalloc calls required.
729 */
4b51d669 730#ifdef CONFIG_ZONE_DMA
1da177e4
LT
731 if (unlikely(gfpflags & GFP_DMA))
732 return csizep->cs_dmacachep;
4b51d669 733#endif
1da177e4
LT
734 return csizep->cs_cachep;
735}
736
b221385b 737static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
97e2bde4
MS
738{
739 return __find_general_cachep(size, gfpflags);
740}
97e2bde4 741
fbaccacf 742static size_t slab_mgmt_size(size_t nr_objs, size_t align)
1da177e4 743{
fbaccacf
SR
744 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
745}
1da177e4 746
a737b3e2
AM
747/*
748 * Calculate the number of objects and left-over bytes for a given buffer size.
749 */
fbaccacf
SR
750static void cache_estimate(unsigned long gfporder, size_t buffer_size,
751 size_t align, int flags, size_t *left_over,
752 unsigned int *num)
753{
754 int nr_objs;
755 size_t mgmt_size;
756 size_t slab_size = PAGE_SIZE << gfporder;
1da177e4 757
fbaccacf
SR
758 /*
759 * The slab management structure can be either off the slab or
760 * on it. For the latter case, the memory allocated for a
761 * slab is used for:
762 *
763 * - The struct slab
764 * - One kmem_bufctl_t for each object
765 * - Padding to respect alignment of @align
766 * - @buffer_size bytes for each object
767 *
768 * If the slab management structure is off the slab, then the
769 * alignment will already be calculated into the size. Because
770 * the slabs are all pages aligned, the objects will be at the
771 * correct alignment when allocated.
772 */
773 if (flags & CFLGS_OFF_SLAB) {
774 mgmt_size = 0;
775 nr_objs = slab_size / buffer_size;
776
777 if (nr_objs > SLAB_LIMIT)
778 nr_objs = SLAB_LIMIT;
779 } else {
780 /*
781 * Ignore padding for the initial guess. The padding
782 * is at most @align-1 bytes, and @buffer_size is at
783 * least @align. In the worst case, this result will
784 * be one greater than the number of objects that fit
785 * into the memory allocation when taking the padding
786 * into account.
787 */
788 nr_objs = (slab_size - sizeof(struct slab)) /
789 (buffer_size + sizeof(kmem_bufctl_t));
790
791 /*
792 * This calculated number will be either the right
793 * amount, or one greater than what we want.
794 */
795 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
796 > slab_size)
797 nr_objs--;
798
799 if (nr_objs > SLAB_LIMIT)
800 nr_objs = SLAB_LIMIT;
801
802 mgmt_size = slab_mgmt_size(nr_objs, align);
803 }
804 *num = nr_objs;
805 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
1da177e4
LT
806}
807
d40cee24 808#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
1da177e4 809
a737b3e2
AM
810static void __slab_error(const char *function, struct kmem_cache *cachep,
811 char *msg)
1da177e4
LT
812{
813 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
b28a02de 814 function, cachep->name, msg);
1da177e4
LT
815 dump_stack();
816}
817
3395ee05
PM
818/*
819 * By default on NUMA we use alien caches to stage the freeing of
820 * objects allocated from other nodes. This causes massive memory
821 * inefficiencies when using fake NUMA setup to split memory into a
822 * large number of small nodes, so it can be disabled on the command
823 * line
824 */
825
826static int use_alien_caches __read_mostly = 1;
827static int __init noaliencache_setup(char *s)
828{
829 use_alien_caches = 0;
830 return 1;
831}
832__setup("noaliencache", noaliencache_setup);
833
8fce4d8e
CL
834#ifdef CONFIG_NUMA
835/*
836 * Special reaping functions for NUMA systems called from cache_reap().
837 * These take care of doing round robin flushing of alien caches (containing
838 * objects freed on different nodes from which they were allocated) and the
839 * flushing of remote pcps by calling drain_node_pages.
840 */
1871e52c 841static DEFINE_PER_CPU(unsigned long, slab_reap_node);
8fce4d8e
CL
842
843static void init_reap_node(int cpu)
844{
845 int node;
846
847 node = next_node(cpu_to_node(cpu), node_online_map);
848 if (node == MAX_NUMNODES)
442295c9 849 node = first_node(node_online_map);
8fce4d8e 850
1871e52c 851 per_cpu(slab_reap_node, cpu) = node;
8fce4d8e
CL
852}
853
854static void next_reap_node(void)
855{
1871e52c 856 int node = __get_cpu_var(slab_reap_node);
8fce4d8e 857
8fce4d8e
CL
858 node = next_node(node, node_online_map);
859 if (unlikely(node >= MAX_NUMNODES))
860 node = first_node(node_online_map);
1871e52c 861 __get_cpu_var(slab_reap_node) = node;
8fce4d8e
CL
862}
863
864#else
865#define init_reap_node(cpu) do { } while (0)
866#define next_reap_node(void) do { } while (0)
867#endif
868
1da177e4
LT
869/*
870 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
871 * via the workqueue/eventd.
872 * Add the CPU number into the expiration time to minimize the possibility of
873 * the CPUs getting into lockstep and contending for the global cache chain
874 * lock.
875 */
897e679b 876static void __cpuinit start_cpu_timer(int cpu)
1da177e4 877{
1871e52c 878 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
1da177e4
LT
879
880 /*
881 * When this gets called from do_initcalls via cpucache_init(),
882 * init_workqueues() has already run, so keventd will be setup
883 * at that time.
884 */
52bad64d 885 if (keventd_up() && reap_work->work.func == NULL) {
8fce4d8e 886 init_reap_node(cpu);
65f27f38 887 INIT_DELAYED_WORK(reap_work, cache_reap);
2b284214
AV
888 schedule_delayed_work_on(cpu, reap_work,
889 __round_jiffies_relative(HZ, cpu));
1da177e4
LT
890 }
891}
892
e498be7d 893static struct array_cache *alloc_arraycache(int node, int entries,
83b519e8 894 int batchcount, gfp_t gfp)
1da177e4 895{
b28a02de 896 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
1da177e4
LT
897 struct array_cache *nc = NULL;
898
83b519e8 899 nc = kmalloc_node(memsize, gfp, node);
d5cff635
CM
900 /*
901 * The array_cache structures contain pointers to free object.
902 * However, when such objects are allocated or transfered to another
903 * cache the pointers are not cleared and they could be counted as
904 * valid references during a kmemleak scan. Therefore, kmemleak must
905 * not scan such objects.
906 */
907 kmemleak_no_scan(nc);
1da177e4
LT
908 if (nc) {
909 nc->avail = 0;
910 nc->limit = entries;
911 nc->batchcount = batchcount;
912 nc->touched = 0;
e498be7d 913 spin_lock_init(&nc->lock);
1da177e4
LT
914 }
915 return nc;
916}
917
3ded175a
CL
918/*
919 * Transfer objects in one arraycache to another.
920 * Locking must be handled by the caller.
921 *
922 * Return the number of entries transferred.
923 */
924static int transfer_objects(struct array_cache *to,
925 struct array_cache *from, unsigned int max)
926{
927 /* Figure out how many entries to transfer */
928 int nr = min(min(from->avail, max), to->limit - to->avail);
929
930 if (!nr)
931 return 0;
932
933 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
934 sizeof(void *) *nr);
935
936 from->avail -= nr;
937 to->avail += nr;
3ded175a
CL
938 return nr;
939}
940
765c4507
CL
941#ifndef CONFIG_NUMA
942
943#define drain_alien_cache(cachep, alien) do { } while (0)
944#define reap_alien(cachep, l3) do { } while (0)
945
83b519e8 946static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
765c4507
CL
947{
948 return (struct array_cache **)BAD_ALIEN_MAGIC;
949}
950
951static inline void free_alien_cache(struct array_cache **ac_ptr)
952{
953}
954
955static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
956{
957 return 0;
958}
959
960static inline void *alternate_node_alloc(struct kmem_cache *cachep,
961 gfp_t flags)
962{
963 return NULL;
964}
965
8b98c169 966static inline void *____cache_alloc_node(struct kmem_cache *cachep,
765c4507
CL
967 gfp_t flags, int nodeid)
968{
969 return NULL;
970}
971
972#else /* CONFIG_NUMA */
973
8b98c169 974static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
c61afb18 975static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
dc85da15 976
83b519e8 977static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
e498be7d
CL
978{
979 struct array_cache **ac_ptr;
8ef82866 980 int memsize = sizeof(void *) * nr_node_ids;
e498be7d
CL
981 int i;
982
983 if (limit > 1)
984 limit = 12;
f3186a9c 985 ac_ptr = kzalloc_node(memsize, gfp, node);
e498be7d
CL
986 if (ac_ptr) {
987 for_each_node(i) {
f3186a9c 988 if (i == node || !node_online(i))
e498be7d 989 continue;
83b519e8 990 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
e498be7d 991 if (!ac_ptr[i]) {
cc550def 992 for (i--; i >= 0; i--)
e498be7d
CL
993 kfree(ac_ptr[i]);
994 kfree(ac_ptr);
995 return NULL;
996 }
997 }
998 }
999 return ac_ptr;
1000}
1001
5295a74c 1002static void free_alien_cache(struct array_cache **ac_ptr)
e498be7d
CL
1003{
1004 int i;
1005
1006 if (!ac_ptr)
1007 return;
e498be7d 1008 for_each_node(i)
b28a02de 1009 kfree(ac_ptr[i]);
e498be7d
CL
1010 kfree(ac_ptr);
1011}
1012
343e0d7a 1013static void __drain_alien_cache(struct kmem_cache *cachep,
5295a74c 1014 struct array_cache *ac, int node)
e498be7d
CL
1015{
1016 struct kmem_list3 *rl3 = cachep->nodelists[node];
1017
1018 if (ac->avail) {
1019 spin_lock(&rl3->list_lock);
e00946fe
CL
1020 /*
1021 * Stuff objects into the remote nodes shared array first.
1022 * That way we could avoid the overhead of putting the objects
1023 * into the free lists and getting them back later.
1024 */
693f7d36 1025 if (rl3->shared)
1026 transfer_objects(rl3->shared, ac, ac->limit);
e00946fe 1027
ff69416e 1028 free_block(cachep, ac->entry, ac->avail, node);
e498be7d
CL
1029 ac->avail = 0;
1030 spin_unlock(&rl3->list_lock);
1031 }
1032}
1033
8fce4d8e
CL
1034/*
1035 * Called from cache_reap() to regularly drain alien caches round robin.
1036 */
1037static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1038{
1871e52c 1039 int node = __get_cpu_var(slab_reap_node);
8fce4d8e
CL
1040
1041 if (l3->alien) {
1042 struct array_cache *ac = l3->alien[node];
e00946fe
CL
1043
1044 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
8fce4d8e
CL
1045 __drain_alien_cache(cachep, ac, node);
1046 spin_unlock_irq(&ac->lock);
1047 }
1048 }
1049}
1050
a737b3e2
AM
1051static void drain_alien_cache(struct kmem_cache *cachep,
1052 struct array_cache **alien)
e498be7d 1053{
b28a02de 1054 int i = 0;
e498be7d
CL
1055 struct array_cache *ac;
1056 unsigned long flags;
1057
1058 for_each_online_node(i) {
4484ebf1 1059 ac = alien[i];
e498be7d
CL
1060 if (ac) {
1061 spin_lock_irqsave(&ac->lock, flags);
1062 __drain_alien_cache(cachep, ac, i);
1063 spin_unlock_irqrestore(&ac->lock, flags);
1064 }
1065 }
1066}
729bd0b7 1067
873623df 1068static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
729bd0b7
PE
1069{
1070 struct slab *slabp = virt_to_slab(objp);
1071 int nodeid = slabp->nodeid;
1072 struct kmem_list3 *l3;
1073 struct array_cache *alien = NULL;
1ca4cb24
PE
1074 int node;
1075
1076 node = numa_node_id();
729bd0b7
PE
1077
1078 /*
1079 * Make sure we are not freeing a object from another node to the array
1080 * cache on this cpu.
1081 */
62918a03 1082 if (likely(slabp->nodeid == node))
729bd0b7
PE
1083 return 0;
1084
1ca4cb24 1085 l3 = cachep->nodelists[node];
729bd0b7
PE
1086 STATS_INC_NODEFREES(cachep);
1087 if (l3->alien && l3->alien[nodeid]) {
1088 alien = l3->alien[nodeid];
873623df 1089 spin_lock(&alien->lock);
729bd0b7
PE
1090 if (unlikely(alien->avail == alien->limit)) {
1091 STATS_INC_ACOVERFLOW(cachep);
1092 __drain_alien_cache(cachep, alien, nodeid);
1093 }
1094 alien->entry[alien->avail++] = objp;
1095 spin_unlock(&alien->lock);
1096 } else {
1097 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1098 free_block(cachep, &objp, 1, nodeid);
1099 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1100 }
1101 return 1;
1102}
e498be7d
CL
1103#endif
1104
fbf1e473
AM
1105static void __cpuinit cpuup_canceled(long cpu)
1106{
1107 struct kmem_cache *cachep;
1108 struct kmem_list3 *l3 = NULL;
1109 int node = cpu_to_node(cpu);
a70f7302 1110 const struct cpumask *mask = cpumask_of_node(node);
fbf1e473
AM
1111
1112 list_for_each_entry(cachep, &cache_chain, next) {
1113 struct array_cache *nc;
1114 struct array_cache *shared;
1115 struct array_cache **alien;
fbf1e473 1116
fbf1e473
AM
1117 /* cpu is dead; no one can alloc from it. */
1118 nc = cachep->array[cpu];
1119 cachep->array[cpu] = NULL;
1120 l3 = cachep->nodelists[node];
1121
1122 if (!l3)
1123 goto free_array_cache;
1124
1125 spin_lock_irq(&l3->list_lock);
1126
1127 /* Free limit for this kmem_list3 */
1128 l3->free_limit -= cachep->batchcount;
1129 if (nc)
1130 free_block(cachep, nc->entry, nc->avail, node);
1131
58463c1f 1132 if (!cpumask_empty(mask)) {
fbf1e473
AM
1133 spin_unlock_irq(&l3->list_lock);
1134 goto free_array_cache;
1135 }
1136
1137 shared = l3->shared;
1138 if (shared) {
1139 free_block(cachep, shared->entry,
1140 shared->avail, node);
1141 l3->shared = NULL;
1142 }
1143
1144 alien = l3->alien;
1145 l3->alien = NULL;
1146
1147 spin_unlock_irq(&l3->list_lock);
1148
1149 kfree(shared);
1150 if (alien) {
1151 drain_alien_cache(cachep, alien);
1152 free_alien_cache(alien);
1153 }
1154free_array_cache:
1155 kfree(nc);
1156 }
1157 /*
1158 * In the previous loop, all the objects were freed to
1159 * the respective cache's slabs, now we can go ahead and
1160 * shrink each nodelist to its limit.
1161 */
1162 list_for_each_entry(cachep, &cache_chain, next) {
1163 l3 = cachep->nodelists[node];
1164 if (!l3)
1165 continue;
1166 drain_freelist(cachep, l3, l3->free_objects);
1167 }
1168}
1169
1170static int __cpuinit cpuup_prepare(long cpu)
1da177e4 1171{
343e0d7a 1172 struct kmem_cache *cachep;
e498be7d
CL
1173 struct kmem_list3 *l3 = NULL;
1174 int node = cpu_to_node(cpu);
ea02e3dd 1175 const int memsize = sizeof(struct kmem_list3);
1da177e4 1176
fbf1e473
AM
1177 /*
1178 * We need to do this right in the beginning since
1179 * alloc_arraycache's are going to use this list.
1180 * kmalloc_node allows us to add the slab to the right
1181 * kmem_list3 and not this cpu's kmem_list3
1182 */
1183
1184 list_for_each_entry(cachep, &cache_chain, next) {
a737b3e2 1185 /*
fbf1e473
AM
1186 * Set up the size64 kmemlist for cpu before we can
1187 * begin anything. Make sure some other cpu on this
1188 * node has not already allocated this
e498be7d 1189 */
fbf1e473
AM
1190 if (!cachep->nodelists[node]) {
1191 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1192 if (!l3)
1193 goto bad;
1194 kmem_list3_init(l3);
1195 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1196 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
e498be7d 1197
a737b3e2 1198 /*
fbf1e473
AM
1199 * The l3s don't come and go as CPUs come and
1200 * go. cache_chain_mutex is sufficient
1201 * protection here.
e498be7d 1202 */
fbf1e473 1203 cachep->nodelists[node] = l3;
e498be7d
CL
1204 }
1205
fbf1e473
AM
1206 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1207 cachep->nodelists[node]->free_limit =
1208 (1 + nr_cpus_node(node)) *
1209 cachep->batchcount + cachep->num;
1210 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1211 }
1212
1213 /*
1214 * Now we can go ahead with allocating the shared arrays and
1215 * array caches
1216 */
1217 list_for_each_entry(cachep, &cache_chain, next) {
1218 struct array_cache *nc;
1219 struct array_cache *shared = NULL;
1220 struct array_cache **alien = NULL;
1221
1222 nc = alloc_arraycache(node, cachep->limit,
83b519e8 1223 cachep->batchcount, GFP_KERNEL);
fbf1e473
AM
1224 if (!nc)
1225 goto bad;
1226 if (cachep->shared) {
1227 shared = alloc_arraycache(node,
1228 cachep->shared * cachep->batchcount,
83b519e8 1229 0xbaadf00d, GFP_KERNEL);
12d00f6a
AM
1230 if (!shared) {
1231 kfree(nc);
1da177e4 1232 goto bad;
12d00f6a 1233 }
fbf1e473
AM
1234 }
1235 if (use_alien_caches) {
83b519e8 1236 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
12d00f6a
AM
1237 if (!alien) {
1238 kfree(shared);
1239 kfree(nc);
fbf1e473 1240 goto bad;
12d00f6a 1241 }
fbf1e473
AM
1242 }
1243 cachep->array[cpu] = nc;
1244 l3 = cachep->nodelists[node];
1245 BUG_ON(!l3);
1246
1247 spin_lock_irq(&l3->list_lock);
1248 if (!l3->shared) {
1249 /*
1250 * We are serialised from CPU_DEAD or
1251 * CPU_UP_CANCELLED by the cpucontrol lock
1252 */
1253 l3->shared = shared;
1254 shared = NULL;
1255 }
4484ebf1 1256#ifdef CONFIG_NUMA
fbf1e473
AM
1257 if (!l3->alien) {
1258 l3->alien = alien;
1259 alien = NULL;
1da177e4 1260 }
fbf1e473
AM
1261#endif
1262 spin_unlock_irq(&l3->list_lock);
1263 kfree(shared);
1264 free_alien_cache(alien);
1265 }
ce79ddc8
PE
1266 init_node_lock_keys(node);
1267
fbf1e473
AM
1268 return 0;
1269bad:
12d00f6a 1270 cpuup_canceled(cpu);
fbf1e473
AM
1271 return -ENOMEM;
1272}
1273
1274static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1275 unsigned long action, void *hcpu)
1276{
1277 long cpu = (long)hcpu;
1278 int err = 0;
1279
1280 switch (action) {
fbf1e473
AM
1281 case CPU_UP_PREPARE:
1282 case CPU_UP_PREPARE_FROZEN:
95402b38 1283 mutex_lock(&cache_chain_mutex);
fbf1e473 1284 err = cpuup_prepare(cpu);
95402b38 1285 mutex_unlock(&cache_chain_mutex);
1da177e4
LT
1286 break;
1287 case CPU_ONLINE:
8bb78442 1288 case CPU_ONLINE_FROZEN:
1da177e4
LT
1289 start_cpu_timer(cpu);
1290 break;
1291#ifdef CONFIG_HOTPLUG_CPU
5830c590 1292 case CPU_DOWN_PREPARE:
8bb78442 1293 case CPU_DOWN_PREPARE_FROZEN:
5830c590
CL
1294 /*
1295 * Shutdown cache reaper. Note that the cache_chain_mutex is
1296 * held so that if cache_reap() is invoked it cannot do
1297 * anything expensive but will only modify reap_work
1298 * and reschedule the timer.
1299 */
1871e52c 1300 cancel_rearming_delayed_work(&per_cpu(slab_reap_work, cpu));
5830c590 1301 /* Now the cache_reaper is guaranteed to be not running. */
1871e52c 1302 per_cpu(slab_reap_work, cpu).work.func = NULL;
5830c590
CL
1303 break;
1304 case CPU_DOWN_FAILED:
8bb78442 1305 case CPU_DOWN_FAILED_FROZEN:
5830c590
CL
1306 start_cpu_timer(cpu);
1307 break;
1da177e4 1308 case CPU_DEAD:
8bb78442 1309 case CPU_DEAD_FROZEN:
4484ebf1
RT
1310 /*
1311 * Even if all the cpus of a node are down, we don't free the
1312 * kmem_list3 of any cache. This to avoid a race between
1313 * cpu_down, and a kmalloc allocation from another cpu for
1314 * memory from the node of the cpu going down. The list3
1315 * structure is usually allocated from kmem_cache_create() and
1316 * gets destroyed at kmem_cache_destroy().
1317 */
183ff22b 1318 /* fall through */
8f5be20b 1319#endif
1da177e4 1320 case CPU_UP_CANCELED:
8bb78442 1321 case CPU_UP_CANCELED_FROZEN:
95402b38 1322 mutex_lock(&cache_chain_mutex);
fbf1e473 1323 cpuup_canceled(cpu);
fc0abb14 1324 mutex_unlock(&cache_chain_mutex);
1da177e4 1325 break;
1da177e4 1326 }
fbf1e473 1327 return err ? NOTIFY_BAD : NOTIFY_OK;
1da177e4
LT
1328}
1329
74b85f37
CS
1330static struct notifier_block __cpuinitdata cpucache_notifier = {
1331 &cpuup_callback, NULL, 0
1332};
1da177e4 1333
e498be7d
CL
1334/*
1335 * swap the static kmem_list3 with kmalloced memory
1336 */
a737b3e2
AM
1337static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1338 int nodeid)
e498be7d
CL
1339{
1340 struct kmem_list3 *ptr;
1341
83b519e8 1342 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
e498be7d
CL
1343 BUG_ON(!ptr);
1344
e498be7d 1345 memcpy(ptr, list, sizeof(struct kmem_list3));
2b2d5493
IM
1346 /*
1347 * Do not assume that spinlocks can be initialized via memcpy:
1348 */
1349 spin_lock_init(&ptr->list_lock);
1350
e498be7d
CL
1351 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1352 cachep->nodelists[nodeid] = ptr;
e498be7d
CL
1353}
1354
556a169d
PE
1355/*
1356 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1357 * size of kmem_list3.
1358 */
1359static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1360{
1361 int node;
1362
1363 for_each_online_node(node) {
1364 cachep->nodelists[node] = &initkmem_list3[index + node];
1365 cachep->nodelists[node]->next_reap = jiffies +
1366 REAPTIMEOUT_LIST3 +
1367 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1368 }
1369}
1370
a737b3e2
AM
1371/*
1372 * Initialisation. Called after the page allocator have been initialised and
1373 * before smp_init().
1da177e4
LT
1374 */
1375void __init kmem_cache_init(void)
1376{
1377 size_t left_over;
1378 struct cache_sizes *sizes;
1379 struct cache_names *names;
e498be7d 1380 int i;
07ed76b2 1381 int order;
1ca4cb24 1382 int node;
e498be7d 1383
b6e68bc1 1384 if (num_possible_nodes() == 1)
62918a03
SS
1385 use_alien_caches = 0;
1386
e498be7d
CL
1387 for (i = 0; i < NUM_INIT_LISTS; i++) {
1388 kmem_list3_init(&initkmem_list3[i]);
1389 if (i < MAX_NUMNODES)
1390 cache_cache.nodelists[i] = NULL;
1391 }
556a169d 1392 set_up_list3s(&cache_cache, CACHE_CACHE);
1da177e4
LT
1393
1394 /*
1395 * Fragmentation resistance on low memory - only use bigger
1396 * page orders on machines with more than 32MB of memory.
1397 */
4481374c 1398 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1da177e4
LT
1399 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1400
1da177e4
LT
1401 /* Bootstrap is tricky, because several objects are allocated
1402 * from caches that do not exist yet:
a737b3e2
AM
1403 * 1) initialize the cache_cache cache: it contains the struct
1404 * kmem_cache structures of all caches, except cache_cache itself:
1405 * cache_cache is statically allocated.
e498be7d
CL
1406 * Initially an __init data area is used for the head array and the
1407 * kmem_list3 structures, it's replaced with a kmalloc allocated
1408 * array at the end of the bootstrap.
1da177e4 1409 * 2) Create the first kmalloc cache.
343e0d7a 1410 * The struct kmem_cache for the new cache is allocated normally.
e498be7d
CL
1411 * An __init data area is used for the head array.
1412 * 3) Create the remaining kmalloc caches, with minimally sized
1413 * head arrays.
1da177e4
LT
1414 * 4) Replace the __init data head arrays for cache_cache and the first
1415 * kmalloc cache with kmalloc allocated arrays.
e498be7d
CL
1416 * 5) Replace the __init data for kmem_list3 for cache_cache and
1417 * the other cache's with kmalloc allocated memory.
1418 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1da177e4
LT
1419 */
1420
1ca4cb24
PE
1421 node = numa_node_id();
1422
1da177e4 1423 /* 1) create the cache_cache */
1da177e4
LT
1424 INIT_LIST_HEAD(&cache_chain);
1425 list_add(&cache_cache.next, &cache_chain);
1426 cache_cache.colour_off = cache_line_size();
1427 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
ec1f5eee 1428 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1da177e4 1429
8da3430d
ED
1430 /*
1431 * struct kmem_cache size depends on nr_node_ids, which
1432 * can be less than MAX_NUMNODES.
1433 */
1434 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1435 nr_node_ids * sizeof(struct kmem_list3 *);
1436#if DEBUG
1437 cache_cache.obj_size = cache_cache.buffer_size;
1438#endif
a737b3e2
AM
1439 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1440 cache_line_size());
6a2d7a95
ED
1441 cache_cache.reciprocal_buffer_size =
1442 reciprocal_value(cache_cache.buffer_size);
1da177e4 1443
07ed76b2
JS
1444 for (order = 0; order < MAX_ORDER; order++) {
1445 cache_estimate(order, cache_cache.buffer_size,
1446 cache_line_size(), 0, &left_over, &cache_cache.num);
1447 if (cache_cache.num)
1448 break;
1449 }
40094fa6 1450 BUG_ON(!cache_cache.num);
07ed76b2 1451 cache_cache.gfporder = order;
b28a02de 1452 cache_cache.colour = left_over / cache_cache.colour_off;
b28a02de
PE
1453 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1454 sizeof(struct slab), cache_line_size());
1da177e4
LT
1455
1456 /* 2+3) create the kmalloc caches */
1457 sizes = malloc_sizes;
1458 names = cache_names;
1459
a737b3e2
AM
1460 /*
1461 * Initialize the caches that provide memory for the array cache and the
1462 * kmem_list3 structures first. Without this, further allocations will
1463 * bug.
e498be7d
CL
1464 */
1465
1466 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
a737b3e2
AM
1467 sizes[INDEX_AC].cs_size,
1468 ARCH_KMALLOC_MINALIGN,
1469 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
20c2df83 1470 NULL);
e498be7d 1471
a737b3e2 1472 if (INDEX_AC != INDEX_L3) {
e498be7d 1473 sizes[INDEX_L3].cs_cachep =
a737b3e2
AM
1474 kmem_cache_create(names[INDEX_L3].name,
1475 sizes[INDEX_L3].cs_size,
1476 ARCH_KMALLOC_MINALIGN,
1477 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
20c2df83 1478 NULL);
a737b3e2 1479 }
e498be7d 1480
e0a42726
IM
1481 slab_early_init = 0;
1482
1da177e4 1483 while (sizes->cs_size != ULONG_MAX) {
e498be7d
CL
1484 /*
1485 * For performance, all the general caches are L1 aligned.
1da177e4
LT
1486 * This should be particularly beneficial on SMP boxes, as it
1487 * eliminates "false sharing".
1488 * Note for systems short on memory removing the alignment will
e498be7d
CL
1489 * allow tighter packing of the smaller caches.
1490 */
a737b3e2 1491 if (!sizes->cs_cachep) {
e498be7d 1492 sizes->cs_cachep = kmem_cache_create(names->name,
a737b3e2
AM
1493 sizes->cs_size,
1494 ARCH_KMALLOC_MINALIGN,
1495 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
20c2df83 1496 NULL);
a737b3e2 1497 }
4b51d669
CL
1498#ifdef CONFIG_ZONE_DMA
1499 sizes->cs_dmacachep = kmem_cache_create(
1500 names->name_dma,
a737b3e2
AM
1501 sizes->cs_size,
1502 ARCH_KMALLOC_MINALIGN,
1503 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1504 SLAB_PANIC,
20c2df83 1505 NULL);
4b51d669 1506#endif
1da177e4
LT
1507 sizes++;
1508 names++;
1509 }
1510 /* 4) Replace the bootstrap head arrays */
1511 {
2b2d5493 1512 struct array_cache *ptr;
e498be7d 1513
83b519e8 1514 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
e498be7d 1515
9a2dba4b
PE
1516 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1517 memcpy(ptr, cpu_cache_get(&cache_cache),
b28a02de 1518 sizeof(struct arraycache_init));
2b2d5493
IM
1519 /*
1520 * Do not assume that spinlocks can be initialized via memcpy:
1521 */
1522 spin_lock_init(&ptr->lock);
1523
1da177e4 1524 cache_cache.array[smp_processor_id()] = ptr;
e498be7d 1525
83b519e8 1526 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
e498be7d 1527
9a2dba4b 1528 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
b28a02de 1529 != &initarray_generic.cache);
9a2dba4b 1530 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
b28a02de 1531 sizeof(struct arraycache_init));
2b2d5493
IM
1532 /*
1533 * Do not assume that spinlocks can be initialized via memcpy:
1534 */
1535 spin_lock_init(&ptr->lock);
1536
e498be7d 1537 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
b28a02de 1538 ptr;
1da177e4 1539 }
e498be7d
CL
1540 /* 5) Replace the bootstrap kmem_list3's */
1541 {
1ca4cb24
PE
1542 int nid;
1543
9c09a95c 1544 for_each_online_node(nid) {
ec1f5eee 1545 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
556a169d 1546
e498be7d 1547 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1ca4cb24 1548 &initkmem_list3[SIZE_AC + nid], nid);
e498be7d
CL
1549
1550 if (INDEX_AC != INDEX_L3) {
1551 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1ca4cb24 1552 &initkmem_list3[SIZE_L3 + nid], nid);
e498be7d
CL
1553 }
1554 }
1555 }
1da177e4 1556
8429db5c 1557 g_cpucache_up = EARLY;
8429db5c
PE
1558}
1559
1560void __init kmem_cache_init_late(void)
1561{
1562 struct kmem_cache *cachep;
1563
8429db5c
PE
1564 /* 6) resize the head arrays to their final sizes */
1565 mutex_lock(&cache_chain_mutex);
1566 list_for_each_entry(cachep, &cache_chain, next)
1567 if (enable_cpucache(cachep, GFP_NOWAIT))
1568 BUG();
1569 mutex_unlock(&cache_chain_mutex);
056c6241 1570
1da177e4
LT
1571 /* Done! */
1572 g_cpucache_up = FULL;
1573
ec5a36f9
PE
1574 /* Annotate slab for lockdep -- annotate the malloc caches */
1575 init_lock_keys();
1576
a737b3e2
AM
1577 /*
1578 * Register a cpu startup notifier callback that initializes
1579 * cpu_cache_get for all new cpus
1da177e4
LT
1580 */
1581 register_cpu_notifier(&cpucache_notifier);
1da177e4 1582
a737b3e2
AM
1583 /*
1584 * The reap timers are started later, with a module init call: That part
1585 * of the kernel is not yet operational.
1da177e4
LT
1586 */
1587}
1588
1589static int __init cpucache_init(void)
1590{
1591 int cpu;
1592
a737b3e2
AM
1593 /*
1594 * Register the timers that return unneeded pages to the page allocator
1da177e4 1595 */
e498be7d 1596 for_each_online_cpu(cpu)
a737b3e2 1597 start_cpu_timer(cpu);
1da177e4
LT
1598 return 0;
1599}
1da177e4
LT
1600__initcall(cpucache_init);
1601
1602/*
1603 * Interface to system's page allocator. No need to hold the cache-lock.
1604 *
1605 * If we requested dmaable memory, we will get it. Even if we
1606 * did not request dmaable memory, we might get it, but that
1607 * would be relatively rare and ignorable.
1608 */
343e0d7a 1609static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1da177e4
LT
1610{
1611 struct page *page;
e1b6aa6f 1612 int nr_pages;
1da177e4
LT
1613 int i;
1614
d6fef9da 1615#ifndef CONFIG_MMU
e1b6aa6f
CH
1616 /*
1617 * Nommu uses slab's for process anonymous memory allocations, and thus
1618 * requires __GFP_COMP to properly refcount higher order allocations
d6fef9da 1619 */
e1b6aa6f 1620 flags |= __GFP_COMP;
d6fef9da 1621#endif
765c4507 1622
3c517a61 1623 flags |= cachep->gfpflags;
e12ba74d
MG
1624 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1625 flags |= __GFP_RECLAIMABLE;
e1b6aa6f 1626
517d0869 1627 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1da177e4
LT
1628 if (!page)
1629 return NULL;
1da177e4 1630
e1b6aa6f 1631 nr_pages = (1 << cachep->gfporder);
1da177e4 1632 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
972d1a7b
CL
1633 add_zone_page_state(page_zone(page),
1634 NR_SLAB_RECLAIMABLE, nr_pages);
1635 else
1636 add_zone_page_state(page_zone(page),
1637 NR_SLAB_UNRECLAIMABLE, nr_pages);
e1b6aa6f
CH
1638 for (i = 0; i < nr_pages; i++)
1639 __SetPageSlab(page + i);
c175eea4 1640
b1eeab67
VN
1641 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1642 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1643
1644 if (cachep->ctor)
1645 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1646 else
1647 kmemcheck_mark_unallocated_pages(page, nr_pages);
1648 }
c175eea4 1649
e1b6aa6f 1650 return page_address(page);
1da177e4
LT
1651}
1652
1653/*
1654 * Interface to system's page release.
1655 */
343e0d7a 1656static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1da177e4 1657{
b28a02de 1658 unsigned long i = (1 << cachep->gfporder);
1da177e4
LT
1659 struct page *page = virt_to_page(addr);
1660 const unsigned long nr_freed = i;
1661
b1eeab67 1662 kmemcheck_free_shadow(page, cachep->gfporder);
c175eea4 1663
972d1a7b
CL
1664 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1665 sub_zone_page_state(page_zone(page),
1666 NR_SLAB_RECLAIMABLE, nr_freed);
1667 else
1668 sub_zone_page_state(page_zone(page),
1669 NR_SLAB_UNRECLAIMABLE, nr_freed);
1da177e4 1670 while (i--) {
f205b2fe
NP
1671 BUG_ON(!PageSlab(page));
1672 __ClearPageSlab(page);
1da177e4
LT
1673 page++;
1674 }
1da177e4
LT
1675 if (current->reclaim_state)
1676 current->reclaim_state->reclaimed_slab += nr_freed;
1677 free_pages((unsigned long)addr, cachep->gfporder);
1da177e4
LT
1678}
1679
1680static void kmem_rcu_free(struct rcu_head *head)
1681{
b28a02de 1682 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
343e0d7a 1683 struct kmem_cache *cachep = slab_rcu->cachep;
1da177e4
LT
1684
1685 kmem_freepages(cachep, slab_rcu->addr);
1686 if (OFF_SLAB(cachep))
1687 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1688}
1689
1690#if DEBUG
1691
1692#ifdef CONFIG_DEBUG_PAGEALLOC
343e0d7a 1693static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
b28a02de 1694 unsigned long caller)
1da177e4 1695{
3dafccf2 1696 int size = obj_size(cachep);
1da177e4 1697
3dafccf2 1698 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1da177e4 1699
b28a02de 1700 if (size < 5 * sizeof(unsigned long))
1da177e4
LT
1701 return;
1702
b28a02de
PE
1703 *addr++ = 0x12345678;
1704 *addr++ = caller;
1705 *addr++ = smp_processor_id();
1706 size -= 3 * sizeof(unsigned long);
1da177e4
LT
1707 {
1708 unsigned long *sptr = &caller;
1709 unsigned long svalue;
1710
1711 while (!kstack_end(sptr)) {
1712 svalue = *sptr++;
1713 if (kernel_text_address(svalue)) {
b28a02de 1714 *addr++ = svalue;
1da177e4
LT
1715 size -= sizeof(unsigned long);
1716 if (size <= sizeof(unsigned long))
1717 break;
1718 }
1719 }
1720
1721 }
b28a02de 1722 *addr++ = 0x87654321;
1da177e4
LT
1723}
1724#endif
1725
343e0d7a 1726static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1da177e4 1727{
3dafccf2
MS
1728 int size = obj_size(cachep);
1729 addr = &((char *)addr)[obj_offset(cachep)];
1da177e4
LT
1730
1731 memset(addr, val, size);
b28a02de 1732 *(unsigned char *)(addr + size - 1) = POISON_END;
1da177e4
LT
1733}
1734
1735static void dump_line(char *data, int offset, int limit)
1736{
1737 int i;
aa83aa40
DJ
1738 unsigned char error = 0;
1739 int bad_count = 0;
1740
1da177e4 1741 printk(KERN_ERR "%03x:", offset);
aa83aa40
DJ
1742 for (i = 0; i < limit; i++) {
1743 if (data[offset + i] != POISON_FREE) {
1744 error = data[offset + i];
1745 bad_count++;
1746 }
b28a02de 1747 printk(" %02x", (unsigned char)data[offset + i]);
aa83aa40 1748 }
1da177e4 1749 printk("\n");
aa83aa40
DJ
1750
1751 if (bad_count == 1) {
1752 error ^= POISON_FREE;
1753 if (!(error & (error - 1))) {
1754 printk(KERN_ERR "Single bit error detected. Probably "
1755 "bad RAM.\n");
1756#ifdef CONFIG_X86
1757 printk(KERN_ERR "Run memtest86+ or a similar memory "
1758 "test tool.\n");
1759#else
1760 printk(KERN_ERR "Run a memory test tool.\n");
1761#endif
1762 }
1763 }
1da177e4
LT
1764}
1765#endif
1766
1767#if DEBUG
1768
343e0d7a 1769static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1da177e4
LT
1770{
1771 int i, size;
1772 char *realobj;
1773
1774 if (cachep->flags & SLAB_RED_ZONE) {
b46b8f19 1775 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
a737b3e2
AM
1776 *dbg_redzone1(cachep, objp),
1777 *dbg_redzone2(cachep, objp));
1da177e4
LT
1778 }
1779
1780 if (cachep->flags & SLAB_STORE_USER) {
1781 printk(KERN_ERR "Last user: [<%p>]",
a737b3e2 1782 *dbg_userword(cachep, objp));
1da177e4 1783 print_symbol("(%s)",
a737b3e2 1784 (unsigned long)*dbg_userword(cachep, objp));
1da177e4
LT
1785 printk("\n");
1786 }
3dafccf2
MS
1787 realobj = (char *)objp + obj_offset(cachep);
1788 size = obj_size(cachep);
b28a02de 1789 for (i = 0; i < size && lines; i += 16, lines--) {
1da177e4
LT
1790 int limit;
1791 limit = 16;
b28a02de
PE
1792 if (i + limit > size)
1793 limit = size - i;
1da177e4
LT
1794 dump_line(realobj, i, limit);
1795 }
1796}
1797
343e0d7a 1798static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1da177e4
LT
1799{
1800 char *realobj;
1801 int size, i;
1802 int lines = 0;
1803
3dafccf2
MS
1804 realobj = (char *)objp + obj_offset(cachep);
1805 size = obj_size(cachep);
1da177e4 1806
b28a02de 1807 for (i = 0; i < size; i++) {
1da177e4 1808 char exp = POISON_FREE;
b28a02de 1809 if (i == size - 1)
1da177e4
LT
1810 exp = POISON_END;
1811 if (realobj[i] != exp) {
1812 int limit;
1813 /* Mismatch ! */
1814 /* Print header */
1815 if (lines == 0) {
b28a02de 1816 printk(KERN_ERR
e94a40c5
DH
1817 "Slab corruption: %s start=%p, len=%d\n",
1818 cachep->name, realobj, size);
1da177e4
LT
1819 print_objinfo(cachep, objp, 0);
1820 }
1821 /* Hexdump the affected line */
b28a02de 1822 i = (i / 16) * 16;
1da177e4 1823 limit = 16;
b28a02de
PE
1824 if (i + limit > size)
1825 limit = size - i;
1da177e4
LT
1826 dump_line(realobj, i, limit);
1827 i += 16;
1828 lines++;
1829 /* Limit to 5 lines */
1830 if (lines > 5)
1831 break;
1832 }
1833 }
1834 if (lines != 0) {
1835 /* Print some data about the neighboring objects, if they
1836 * exist:
1837 */
6ed5eb22 1838 struct slab *slabp = virt_to_slab(objp);
8fea4e96 1839 unsigned int objnr;
1da177e4 1840
8fea4e96 1841 objnr = obj_to_index(cachep, slabp, objp);
1da177e4 1842 if (objnr) {
8fea4e96 1843 objp = index_to_obj(cachep, slabp, objnr - 1);
3dafccf2 1844 realobj = (char *)objp + obj_offset(cachep);
1da177e4 1845 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
b28a02de 1846 realobj, size);
1da177e4
LT
1847 print_objinfo(cachep, objp, 2);
1848 }
b28a02de 1849 if (objnr + 1 < cachep->num) {
8fea4e96 1850 objp = index_to_obj(cachep, slabp, objnr + 1);
3dafccf2 1851 realobj = (char *)objp + obj_offset(cachep);
1da177e4 1852 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
b28a02de 1853 realobj, size);
1da177e4
LT
1854 print_objinfo(cachep, objp, 2);
1855 }
1856 }
1857}
1858#endif
1859
12dd36fa 1860#if DEBUG
e79aec29 1861static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1da177e4 1862{
1da177e4
LT
1863 int i;
1864 for (i = 0; i < cachep->num; i++) {
8fea4e96 1865 void *objp = index_to_obj(cachep, slabp, i);
1da177e4
LT
1866
1867 if (cachep->flags & SLAB_POISON) {
1868#ifdef CONFIG_DEBUG_PAGEALLOC
a737b3e2
AM
1869 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1870 OFF_SLAB(cachep))
b28a02de 1871 kernel_map_pages(virt_to_page(objp),
a737b3e2 1872 cachep->buffer_size / PAGE_SIZE, 1);
1da177e4
LT
1873 else
1874 check_poison_obj(cachep, objp);
1875#else
1876 check_poison_obj(cachep, objp);
1877#endif
1878 }
1879 if (cachep->flags & SLAB_RED_ZONE) {
1880 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1881 slab_error(cachep, "start of a freed object "
b28a02de 1882 "was overwritten");
1da177e4
LT
1883 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1884 slab_error(cachep, "end of a freed object "
b28a02de 1885 "was overwritten");
1da177e4 1886 }
1da177e4 1887 }
12dd36fa 1888}
1da177e4 1889#else
e79aec29 1890static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
12dd36fa 1891{
12dd36fa 1892}
1da177e4
LT
1893#endif
1894
911851e6
RD
1895/**
1896 * slab_destroy - destroy and release all objects in a slab
1897 * @cachep: cache pointer being destroyed
1898 * @slabp: slab pointer being destroyed
1899 *
12dd36fa 1900 * Destroy all the objs in a slab, and release the mem back to the system.
a737b3e2
AM
1901 * Before calling the slab must have been unlinked from the cache. The
1902 * cache-lock is not held/needed.
12dd36fa 1903 */
343e0d7a 1904static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
12dd36fa
MD
1905{
1906 void *addr = slabp->s_mem - slabp->colouroff;
1907
e79aec29 1908 slab_destroy_debugcheck(cachep, slabp);
1da177e4
LT
1909 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1910 struct slab_rcu *slab_rcu;
1911
b28a02de 1912 slab_rcu = (struct slab_rcu *)slabp;
1da177e4
LT
1913 slab_rcu->cachep = cachep;
1914 slab_rcu->addr = addr;
1915 call_rcu(&slab_rcu->head, kmem_rcu_free);
1916 } else {
1917 kmem_freepages(cachep, addr);
873623df
IM
1918 if (OFF_SLAB(cachep))
1919 kmem_cache_free(cachep->slabp_cache, slabp);
1da177e4
LT
1920 }
1921}
1922
117f6eb1
CL
1923static void __kmem_cache_destroy(struct kmem_cache *cachep)
1924{
1925 int i;
1926 struct kmem_list3 *l3;
1927
1928 for_each_online_cpu(i)
1929 kfree(cachep->array[i]);
1930
1931 /* NUMA: free the list3 structures */
1932 for_each_online_node(i) {
1933 l3 = cachep->nodelists[i];
1934 if (l3) {
1935 kfree(l3->shared);
1936 free_alien_cache(l3->alien);
1937 kfree(l3);
1938 }
1939 }
1940 kmem_cache_free(&cache_cache, cachep);
1941}
1942
1943
4d268eba 1944/**
a70773dd
RD
1945 * calculate_slab_order - calculate size (page order) of slabs
1946 * @cachep: pointer to the cache that is being created
1947 * @size: size of objects to be created in this cache.
1948 * @align: required alignment for the objects.
1949 * @flags: slab allocation flags
1950 *
1951 * Also calculates the number of objects per slab.
4d268eba
PE
1952 *
1953 * This could be made much more intelligent. For now, try to avoid using
1954 * high order pages for slabs. When the gfp() functions are more friendly
1955 * towards high-order requests, this should be changed.
1956 */
a737b3e2 1957static size_t calculate_slab_order(struct kmem_cache *cachep,
ee13d785 1958 size_t size, size_t align, unsigned long flags)
4d268eba 1959{
b1ab41c4 1960 unsigned long offslab_limit;
4d268eba 1961 size_t left_over = 0;
9888e6fa 1962 int gfporder;
4d268eba 1963
0aa817f0 1964 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
4d268eba
PE
1965 unsigned int num;
1966 size_t remainder;
1967
9888e6fa 1968 cache_estimate(gfporder, size, align, flags, &remainder, &num);
4d268eba
PE
1969 if (!num)
1970 continue;
9888e6fa 1971
b1ab41c4
IM
1972 if (flags & CFLGS_OFF_SLAB) {
1973 /*
1974 * Max number of objs-per-slab for caches which
1975 * use off-slab slabs. Needed to avoid a possible
1976 * looping condition in cache_grow().
1977 */
1978 offslab_limit = size - sizeof(struct slab);
1979 offslab_limit /= sizeof(kmem_bufctl_t);
1980
1981 if (num > offslab_limit)
1982 break;
1983 }
4d268eba 1984
9888e6fa 1985 /* Found something acceptable - save it away */
4d268eba 1986 cachep->num = num;
9888e6fa 1987 cachep->gfporder = gfporder;
4d268eba
PE
1988 left_over = remainder;
1989
f78bb8ad
LT
1990 /*
1991 * A VFS-reclaimable slab tends to have most allocations
1992 * as GFP_NOFS and we really don't want to have to be allocating
1993 * higher-order pages when we are unable to shrink dcache.
1994 */
1995 if (flags & SLAB_RECLAIM_ACCOUNT)
1996 break;
1997
4d268eba
PE
1998 /*
1999 * Large number of objects is good, but very large slabs are
2000 * currently bad for the gfp()s.
2001 */
9888e6fa 2002 if (gfporder >= slab_break_gfp_order)
4d268eba
PE
2003 break;
2004
9888e6fa
LT
2005 /*
2006 * Acceptable internal fragmentation?
2007 */
a737b3e2 2008 if (left_over * 8 <= (PAGE_SIZE << gfporder))
4d268eba
PE
2009 break;
2010 }
2011 return left_over;
2012}
2013
83b519e8 2014static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
f30cf7d1 2015{
2ed3a4ef 2016 if (g_cpucache_up == FULL)
83b519e8 2017 return enable_cpucache(cachep, gfp);
2ed3a4ef 2018
f30cf7d1
PE
2019 if (g_cpucache_up == NONE) {
2020 /*
2021 * Note: the first kmem_cache_create must create the cache
2022 * that's used by kmalloc(24), otherwise the creation of
2023 * further caches will BUG().
2024 */
2025 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2026
2027 /*
2028 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2029 * the first cache, then we need to set up all its list3s,
2030 * otherwise the creation of further caches will BUG().
2031 */
2032 set_up_list3s(cachep, SIZE_AC);
2033 if (INDEX_AC == INDEX_L3)
2034 g_cpucache_up = PARTIAL_L3;
2035 else
2036 g_cpucache_up = PARTIAL_AC;
2037 } else {
2038 cachep->array[smp_processor_id()] =
83b519e8 2039 kmalloc(sizeof(struct arraycache_init), gfp);
f30cf7d1
PE
2040
2041 if (g_cpucache_up == PARTIAL_AC) {
2042 set_up_list3s(cachep, SIZE_L3);
2043 g_cpucache_up = PARTIAL_L3;
2044 } else {
2045 int node;
556a169d 2046 for_each_online_node(node) {
f30cf7d1
PE
2047 cachep->nodelists[node] =
2048 kmalloc_node(sizeof(struct kmem_list3),
eb91f1d0 2049 gfp, node);
f30cf7d1
PE
2050 BUG_ON(!cachep->nodelists[node]);
2051 kmem_list3_init(cachep->nodelists[node]);
2052 }
2053 }
2054 }
2055 cachep->nodelists[numa_node_id()]->next_reap =
2056 jiffies + REAPTIMEOUT_LIST3 +
2057 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2058
2059 cpu_cache_get(cachep)->avail = 0;
2060 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2061 cpu_cache_get(cachep)->batchcount = 1;
2062 cpu_cache_get(cachep)->touched = 0;
2063 cachep->batchcount = 1;
2064 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2ed3a4ef 2065 return 0;
f30cf7d1
PE
2066}
2067
1da177e4
LT
2068/**
2069 * kmem_cache_create - Create a cache.
2070 * @name: A string which is used in /proc/slabinfo to identify this cache.
2071 * @size: The size of objects to be created in this cache.
2072 * @align: The required alignment for the objects.
2073 * @flags: SLAB flags
2074 * @ctor: A constructor for the objects.
1da177e4
LT
2075 *
2076 * Returns a ptr to the cache on success, NULL on failure.
2077 * Cannot be called within a int, but can be interrupted.
20c2df83 2078 * The @ctor is run when new pages are allocated by the cache.
1da177e4
LT
2079 *
2080 * @name must be valid until the cache is destroyed. This implies that
a737b3e2 2081 * the module calling this has to destroy the cache before getting unloaded.
249da166
CM
2082 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2083 * therefore applications must manage it themselves.
a737b3e2 2084 *
1da177e4
LT
2085 * The flags are
2086 *
2087 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2088 * to catch references to uninitialised memory.
2089 *
2090 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2091 * for buffer overruns.
2092 *
1da177e4
LT
2093 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2094 * cacheline. This can be beneficial if you're counting cycles as closely
2095 * as davem.
2096 */
343e0d7a 2097struct kmem_cache *
1da177e4 2098kmem_cache_create (const char *name, size_t size, size_t align,
51cc5068 2099 unsigned long flags, void (*ctor)(void *))
1da177e4
LT
2100{
2101 size_t left_over, slab_size, ralign;
7a7c381d 2102 struct kmem_cache *cachep = NULL, *pc;
83b519e8 2103 gfp_t gfp;
1da177e4
LT
2104
2105 /*
2106 * Sanity checks... these are all serious usage bugs.
2107 */
a737b3e2 2108 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
20c2df83 2109 size > KMALLOC_MAX_SIZE) {
d40cee24 2110 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
a737b3e2 2111 name);
b28a02de
PE
2112 BUG();
2113 }
1da177e4 2114
f0188f47 2115 /*
8f5be20b 2116 * We use cache_chain_mutex to ensure a consistent view of
174596a0 2117 * cpu_online_mask as well. Please see cpuup_callback
f0188f47 2118 */
83b519e8
PE
2119 if (slab_is_available()) {
2120 get_online_cpus();
2121 mutex_lock(&cache_chain_mutex);
2122 }
4f12bb4f 2123
7a7c381d 2124 list_for_each_entry(pc, &cache_chain, next) {
4f12bb4f
AM
2125 char tmp;
2126 int res;
2127
2128 /*
2129 * This happens when the module gets unloaded and doesn't
2130 * destroy its slab cache and no-one else reuses the vmalloc
2131 * area of the module. Print a warning.
2132 */
138ae663 2133 res = probe_kernel_address(pc->name, tmp);
4f12bb4f 2134 if (res) {
b4169525 2135 printk(KERN_ERR
2136 "SLAB: cache with size %d has lost its name\n",
3dafccf2 2137 pc->buffer_size);
4f12bb4f
AM
2138 continue;
2139 }
2140
b28a02de 2141 if (!strcmp(pc->name, name)) {
b4169525 2142 printk(KERN_ERR
2143 "kmem_cache_create: duplicate cache %s\n", name);
4f12bb4f
AM
2144 dump_stack();
2145 goto oops;
2146 }
2147 }
2148
1da177e4
LT
2149#if DEBUG
2150 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1da177e4
LT
2151#if FORCED_DEBUG
2152 /*
2153 * Enable redzoning and last user accounting, except for caches with
2154 * large objects, if the increased size would increase the object size
2155 * above the next power of two: caches with object sizes just above a
2156 * power of two have a significant amount of internal fragmentation.
2157 */
87a927c7
DW
2158 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2159 2 * sizeof(unsigned long long)))
b28a02de 2160 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1da177e4
LT
2161 if (!(flags & SLAB_DESTROY_BY_RCU))
2162 flags |= SLAB_POISON;
2163#endif
2164 if (flags & SLAB_DESTROY_BY_RCU)
2165 BUG_ON(flags & SLAB_POISON);
2166#endif
1da177e4 2167 /*
a737b3e2
AM
2168 * Always checks flags, a caller might be expecting debug support which
2169 * isn't available.
1da177e4 2170 */
40094fa6 2171 BUG_ON(flags & ~CREATE_MASK);
1da177e4 2172
a737b3e2
AM
2173 /*
2174 * Check that size is in terms of words. This is needed to avoid
1da177e4
LT
2175 * unaligned accesses for some archs when redzoning is used, and makes
2176 * sure any on-slab bufctl's are also correctly aligned.
2177 */
b28a02de
PE
2178 if (size & (BYTES_PER_WORD - 1)) {
2179 size += (BYTES_PER_WORD - 1);
2180 size &= ~(BYTES_PER_WORD - 1);
1da177e4
LT
2181 }
2182
a737b3e2
AM
2183 /* calculate the final buffer alignment: */
2184
1da177e4
LT
2185 /* 1) arch recommendation: can be overridden for debug */
2186 if (flags & SLAB_HWCACHE_ALIGN) {
a737b3e2
AM
2187 /*
2188 * Default alignment: as specified by the arch code. Except if
2189 * an object is really small, then squeeze multiple objects into
2190 * one cacheline.
1da177e4
LT
2191 */
2192 ralign = cache_line_size();
b28a02de 2193 while (size <= ralign / 2)
1da177e4
LT
2194 ralign /= 2;
2195 } else {
2196 ralign = BYTES_PER_WORD;
2197 }
ca5f9703
PE
2198
2199 /*
87a927c7
DW
2200 * Redzoning and user store require word alignment or possibly larger.
2201 * Note this will be overridden by architecture or caller mandated
2202 * alignment if either is greater than BYTES_PER_WORD.
ca5f9703 2203 */
87a927c7
DW
2204 if (flags & SLAB_STORE_USER)
2205 ralign = BYTES_PER_WORD;
2206
2207 if (flags & SLAB_RED_ZONE) {
2208 ralign = REDZONE_ALIGN;
2209 /* If redzoning, ensure that the second redzone is suitably
2210 * aligned, by adjusting the object size accordingly. */
2211 size += REDZONE_ALIGN - 1;
2212 size &= ~(REDZONE_ALIGN - 1);
2213 }
ca5f9703 2214
a44b56d3 2215 /* 2) arch mandated alignment */
1da177e4
LT
2216 if (ralign < ARCH_SLAB_MINALIGN) {
2217 ralign = ARCH_SLAB_MINALIGN;
1da177e4 2218 }
a44b56d3 2219 /* 3) caller mandated alignment */
1da177e4
LT
2220 if (ralign < align) {
2221 ralign = align;
1da177e4 2222 }
a44b56d3 2223 /* disable debug if necessary */
b46b8f19 2224 if (ralign > __alignof__(unsigned long long))
a44b56d3 2225 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
a737b3e2 2226 /*
ca5f9703 2227 * 4) Store it.
1da177e4
LT
2228 */
2229 align = ralign;
2230
83b519e8
PE
2231 if (slab_is_available())
2232 gfp = GFP_KERNEL;
2233 else
2234 gfp = GFP_NOWAIT;
2235
1da177e4 2236 /* Get cache's description obj. */
83b519e8 2237 cachep = kmem_cache_zalloc(&cache_cache, gfp);
1da177e4 2238 if (!cachep)
4f12bb4f 2239 goto oops;
1da177e4
LT
2240
2241#if DEBUG
3dafccf2 2242 cachep->obj_size = size;
1da177e4 2243
ca5f9703
PE
2244 /*
2245 * Both debugging options require word-alignment which is calculated
2246 * into align above.
2247 */
1da177e4 2248 if (flags & SLAB_RED_ZONE) {
1da177e4 2249 /* add space for red zone words */
b46b8f19
DW
2250 cachep->obj_offset += sizeof(unsigned long long);
2251 size += 2 * sizeof(unsigned long long);
1da177e4
LT
2252 }
2253 if (flags & SLAB_STORE_USER) {
ca5f9703 2254 /* user store requires one word storage behind the end of
87a927c7
DW
2255 * the real object. But if the second red zone needs to be
2256 * aligned to 64 bits, we must allow that much space.
1da177e4 2257 */
87a927c7
DW
2258 if (flags & SLAB_RED_ZONE)
2259 size += REDZONE_ALIGN;
2260 else
2261 size += BYTES_PER_WORD;
1da177e4
LT
2262 }
2263#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
b28a02de 2264 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
3dafccf2
MS
2265 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2266 cachep->obj_offset += PAGE_SIZE - size;
1da177e4
LT
2267 size = PAGE_SIZE;
2268 }
2269#endif
2270#endif
2271
e0a42726
IM
2272 /*
2273 * Determine if the slab management is 'on' or 'off' slab.
2274 * (bootstrapping cannot cope with offslab caches so don't do
e7cb55b9
CM
2275 * it too early on. Always use on-slab management when
2276 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
e0a42726 2277 */
e7cb55b9
CM
2278 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2279 !(flags & SLAB_NOLEAKTRACE))
1da177e4
LT
2280 /*
2281 * Size is large, assume best to place the slab management obj
2282 * off-slab (should allow better packing of objs).
2283 */
2284 flags |= CFLGS_OFF_SLAB;
2285
2286 size = ALIGN(size, align);
2287
f78bb8ad 2288 left_over = calculate_slab_order(cachep, size, align, flags);
1da177e4
LT
2289
2290 if (!cachep->num) {
b4169525 2291 printk(KERN_ERR
2292 "kmem_cache_create: couldn't create cache %s.\n", name);
1da177e4
LT
2293 kmem_cache_free(&cache_cache, cachep);
2294 cachep = NULL;
4f12bb4f 2295 goto oops;
1da177e4 2296 }
b28a02de
PE
2297 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2298 + sizeof(struct slab), align);
1da177e4
LT
2299
2300 /*
2301 * If the slab has been placed off-slab, and we have enough space then
2302 * move it on-slab. This is at the expense of any extra colouring.
2303 */
2304 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2305 flags &= ~CFLGS_OFF_SLAB;
2306 left_over -= slab_size;
2307 }
2308
2309 if (flags & CFLGS_OFF_SLAB) {
2310 /* really off slab. No need for manual alignment */
b28a02de
PE
2311 slab_size =
2312 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
67461365
RL
2313
2314#ifdef CONFIG_PAGE_POISONING
2315 /* If we're going to use the generic kernel_map_pages()
2316 * poisoning, then it's going to smash the contents of
2317 * the redzone and userword anyhow, so switch them off.
2318 */
2319 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2320 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2321#endif
1da177e4
LT
2322 }
2323
2324 cachep->colour_off = cache_line_size();
2325 /* Offset must be a multiple of the alignment. */
2326 if (cachep->colour_off < align)
2327 cachep->colour_off = align;
b28a02de 2328 cachep->colour = left_over / cachep->colour_off;
1da177e4
LT
2329 cachep->slab_size = slab_size;
2330 cachep->flags = flags;
2331 cachep->gfpflags = 0;
4b51d669 2332 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
1da177e4 2333 cachep->gfpflags |= GFP_DMA;
3dafccf2 2334 cachep->buffer_size = size;
6a2d7a95 2335 cachep->reciprocal_buffer_size = reciprocal_value(size);
1da177e4 2336
e5ac9c5a 2337 if (flags & CFLGS_OFF_SLAB) {
b2d55073 2338 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
e5ac9c5a
RT
2339 /*
2340 * This is a possibility for one of the malloc_sizes caches.
2341 * But since we go off slab only for object size greater than
2342 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2343 * this should not happen at all.
2344 * But leave a BUG_ON for some lucky dude.
2345 */
6cb8f913 2346 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
e5ac9c5a 2347 }
1da177e4 2348 cachep->ctor = ctor;
1da177e4
LT
2349 cachep->name = name;
2350
83b519e8 2351 if (setup_cpu_cache(cachep, gfp)) {
2ed3a4ef
CL
2352 __kmem_cache_destroy(cachep);
2353 cachep = NULL;
2354 goto oops;
2355 }
1da177e4 2356
1da177e4
LT
2357 /* cache setup completed, link it into the list */
2358 list_add(&cachep->next, &cache_chain);
a737b3e2 2359oops:
1da177e4
LT
2360 if (!cachep && (flags & SLAB_PANIC))
2361 panic("kmem_cache_create(): failed to create slab `%s'\n",
b28a02de 2362 name);
83b519e8
PE
2363 if (slab_is_available()) {
2364 mutex_unlock(&cache_chain_mutex);
2365 put_online_cpus();
2366 }
1da177e4
LT
2367 return cachep;
2368}
2369EXPORT_SYMBOL(kmem_cache_create);
2370
2371#if DEBUG
2372static void check_irq_off(void)
2373{
2374 BUG_ON(!irqs_disabled());
2375}
2376
2377static void check_irq_on(void)
2378{
2379 BUG_ON(irqs_disabled());
2380}
2381
343e0d7a 2382static void check_spinlock_acquired(struct kmem_cache *cachep)
1da177e4
LT
2383{
2384#ifdef CONFIG_SMP
2385 check_irq_off();
e498be7d 2386 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
1da177e4
LT
2387#endif
2388}
e498be7d 2389
343e0d7a 2390static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
e498be7d
CL
2391{
2392#ifdef CONFIG_SMP
2393 check_irq_off();
2394 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2395#endif
2396}
2397
1da177e4
LT
2398#else
2399#define check_irq_off() do { } while(0)
2400#define check_irq_on() do { } while(0)
2401#define check_spinlock_acquired(x) do { } while(0)
e498be7d 2402#define check_spinlock_acquired_node(x, y) do { } while(0)
1da177e4
LT
2403#endif
2404
aab2207c
CL
2405static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2406 struct array_cache *ac,
2407 int force, int node);
2408
1da177e4
LT
2409static void do_drain(void *arg)
2410{
a737b3e2 2411 struct kmem_cache *cachep = arg;
1da177e4 2412 struct array_cache *ac;
ff69416e 2413 int node = numa_node_id();
1da177e4
LT
2414
2415 check_irq_off();
9a2dba4b 2416 ac = cpu_cache_get(cachep);
ff69416e
CL
2417 spin_lock(&cachep->nodelists[node]->list_lock);
2418 free_block(cachep, ac->entry, ac->avail, node);
2419 spin_unlock(&cachep->nodelists[node]->list_lock);
1da177e4
LT
2420 ac->avail = 0;
2421}
2422
343e0d7a 2423static void drain_cpu_caches(struct kmem_cache *cachep)
1da177e4 2424{
e498be7d
CL
2425 struct kmem_list3 *l3;
2426 int node;
2427
15c8b6c1 2428 on_each_cpu(do_drain, cachep, 1);
1da177e4 2429 check_irq_on();
b28a02de 2430 for_each_online_node(node) {
e498be7d 2431 l3 = cachep->nodelists[node];
a4523a8b
RD
2432 if (l3 && l3->alien)
2433 drain_alien_cache(cachep, l3->alien);
2434 }
2435
2436 for_each_online_node(node) {
2437 l3 = cachep->nodelists[node];
2438 if (l3)
aab2207c 2439 drain_array(cachep, l3, l3->shared, 1, node);
e498be7d 2440 }
1da177e4
LT
2441}
2442
ed11d9eb
CL
2443/*
2444 * Remove slabs from the list of free slabs.
2445 * Specify the number of slabs to drain in tofree.
2446 *
2447 * Returns the actual number of slabs released.
2448 */
2449static int drain_freelist(struct kmem_cache *cache,
2450 struct kmem_list3 *l3, int tofree)
1da177e4 2451{
ed11d9eb
CL
2452 struct list_head *p;
2453 int nr_freed;
1da177e4 2454 struct slab *slabp;
1da177e4 2455
ed11d9eb
CL
2456 nr_freed = 0;
2457 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
1da177e4 2458
ed11d9eb 2459 spin_lock_irq(&l3->list_lock);
e498be7d 2460 p = l3->slabs_free.prev;
ed11d9eb
CL
2461 if (p == &l3->slabs_free) {
2462 spin_unlock_irq(&l3->list_lock);
2463 goto out;
2464 }
1da177e4 2465
ed11d9eb 2466 slabp = list_entry(p, struct slab, list);
1da177e4 2467#if DEBUG
40094fa6 2468 BUG_ON(slabp->inuse);
1da177e4
LT
2469#endif
2470 list_del(&slabp->list);
ed11d9eb
CL
2471 /*
2472 * Safe to drop the lock. The slab is no longer linked
2473 * to the cache.
2474 */
2475 l3->free_objects -= cache->num;
e498be7d 2476 spin_unlock_irq(&l3->list_lock);
ed11d9eb
CL
2477 slab_destroy(cache, slabp);
2478 nr_freed++;
1da177e4 2479 }
ed11d9eb
CL
2480out:
2481 return nr_freed;
1da177e4
LT
2482}
2483
8f5be20b 2484/* Called with cache_chain_mutex held to protect against cpu hotplug */
343e0d7a 2485static int __cache_shrink(struct kmem_cache *cachep)
e498be7d
CL
2486{
2487 int ret = 0, i = 0;
2488 struct kmem_list3 *l3;
2489
2490 drain_cpu_caches(cachep);
2491
2492 check_irq_on();
2493 for_each_online_node(i) {
2494 l3 = cachep->nodelists[i];
ed11d9eb
CL
2495 if (!l3)
2496 continue;
2497
2498 drain_freelist(cachep, l3, l3->free_objects);
2499
2500 ret += !list_empty(&l3->slabs_full) ||
2501 !list_empty(&l3->slabs_partial);
e498be7d
CL
2502 }
2503 return (ret ? 1 : 0);
2504}
2505
1da177e4
LT
2506/**
2507 * kmem_cache_shrink - Shrink a cache.
2508 * @cachep: The cache to shrink.
2509 *
2510 * Releases as many slabs as possible for a cache.
2511 * To help debugging, a zero exit status indicates all slabs were released.
2512 */
343e0d7a 2513int kmem_cache_shrink(struct kmem_cache *cachep)
1da177e4 2514{
8f5be20b 2515 int ret;
40094fa6 2516 BUG_ON(!cachep || in_interrupt());
1da177e4 2517
95402b38 2518 get_online_cpus();
8f5be20b
RT
2519 mutex_lock(&cache_chain_mutex);
2520 ret = __cache_shrink(cachep);
2521 mutex_unlock(&cache_chain_mutex);
95402b38 2522 put_online_cpus();
8f5be20b 2523 return ret;
1da177e4
LT
2524}
2525EXPORT_SYMBOL(kmem_cache_shrink);
2526
2527/**
2528 * kmem_cache_destroy - delete a cache
2529 * @cachep: the cache to destroy
2530 *
72fd4a35 2531 * Remove a &struct kmem_cache object from the slab cache.
1da177e4
LT
2532 *
2533 * It is expected this function will be called by a module when it is
2534 * unloaded. This will remove the cache completely, and avoid a duplicate
2535 * cache being allocated each time a module is loaded and unloaded, if the
2536 * module doesn't have persistent in-kernel storage across loads and unloads.
2537 *
2538 * The cache must be empty before calling this function.
2539 *
2540 * The caller must guarantee that noone will allocate memory from the cache
2541 * during the kmem_cache_destroy().
2542 */
133d205a 2543void kmem_cache_destroy(struct kmem_cache *cachep)
1da177e4 2544{
40094fa6 2545 BUG_ON(!cachep || in_interrupt());
1da177e4 2546
1da177e4 2547 /* Find the cache in the chain of caches. */
95402b38 2548 get_online_cpus();
fc0abb14 2549 mutex_lock(&cache_chain_mutex);
1da177e4
LT
2550 /*
2551 * the chain is never empty, cache_cache is never destroyed
2552 */
2553 list_del(&cachep->next);
1da177e4
LT
2554 if (__cache_shrink(cachep)) {
2555 slab_error(cachep, "Can't free all objects");
b28a02de 2556 list_add(&cachep->next, &cache_chain);
fc0abb14 2557 mutex_unlock(&cache_chain_mutex);
95402b38 2558 put_online_cpus();
133d205a 2559 return;
1da177e4
LT
2560 }
2561
2562 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
7ed9f7e5 2563 rcu_barrier();
1da177e4 2564
117f6eb1 2565 __kmem_cache_destroy(cachep);
8f5be20b 2566 mutex_unlock(&cache_chain_mutex);
95402b38 2567 put_online_cpus();
1da177e4
LT
2568}
2569EXPORT_SYMBOL(kmem_cache_destroy);
2570
e5ac9c5a
RT
2571/*
2572 * Get the memory for a slab management obj.
2573 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2574 * always come from malloc_sizes caches. The slab descriptor cannot
2575 * come from the same cache which is getting created because,
2576 * when we are searching for an appropriate cache for these
2577 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2578 * If we are creating a malloc_sizes cache here it would not be visible to
2579 * kmem_find_general_cachep till the initialization is complete.
2580 * Hence we cannot have slabp_cache same as the original cache.
2581 */
343e0d7a 2582static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
5b74ada7
RT
2583 int colour_off, gfp_t local_flags,
2584 int nodeid)
1da177e4
LT
2585{
2586 struct slab *slabp;
b28a02de 2587
1da177e4
LT
2588 if (OFF_SLAB(cachep)) {
2589 /* Slab management obj is off-slab. */
5b74ada7 2590 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
8759ec50 2591 local_flags, nodeid);
d5cff635
CM
2592 /*
2593 * If the first object in the slab is leaked (it's allocated
2594 * but no one has a reference to it), we want to make sure
2595 * kmemleak does not treat the ->s_mem pointer as a reference
2596 * to the object. Otherwise we will not report the leak.
2597 */
c017b4be
CM
2598 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2599 local_flags);
1da177e4
LT
2600 if (!slabp)
2601 return NULL;
2602 } else {
b28a02de 2603 slabp = objp + colour_off;
1da177e4
LT
2604 colour_off += cachep->slab_size;
2605 }
2606 slabp->inuse = 0;
2607 slabp->colouroff = colour_off;
b28a02de 2608 slabp->s_mem = objp + colour_off;
5b74ada7 2609 slabp->nodeid = nodeid;
e51bfd0a 2610 slabp->free = 0;
1da177e4
LT
2611 return slabp;
2612}
2613
2614static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2615{
b28a02de 2616 return (kmem_bufctl_t *) (slabp + 1);
1da177e4
LT
2617}
2618
343e0d7a 2619static void cache_init_objs(struct kmem_cache *cachep,
a35afb83 2620 struct slab *slabp)
1da177e4
LT
2621{
2622 int i;
2623
2624 for (i = 0; i < cachep->num; i++) {
8fea4e96 2625 void *objp = index_to_obj(cachep, slabp, i);
1da177e4
LT
2626#if DEBUG
2627 /* need to poison the objs? */
2628 if (cachep->flags & SLAB_POISON)
2629 poison_obj(cachep, objp, POISON_FREE);
2630 if (cachep->flags & SLAB_STORE_USER)
2631 *dbg_userword(cachep, objp) = NULL;
2632
2633 if (cachep->flags & SLAB_RED_ZONE) {
2634 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2635 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2636 }
2637 /*
a737b3e2
AM
2638 * Constructors are not allowed to allocate memory from the same
2639 * cache which they are a constructor for. Otherwise, deadlock.
2640 * They must also be threaded.
1da177e4
LT
2641 */
2642 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
51cc5068 2643 cachep->ctor(objp + obj_offset(cachep));
1da177e4
LT
2644
2645 if (cachep->flags & SLAB_RED_ZONE) {
2646 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2647 slab_error(cachep, "constructor overwrote the"
b28a02de 2648 " end of an object");
1da177e4
LT
2649 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2650 slab_error(cachep, "constructor overwrote the"
b28a02de 2651 " start of an object");
1da177e4 2652 }
a737b3e2
AM
2653 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2654 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
b28a02de 2655 kernel_map_pages(virt_to_page(objp),
3dafccf2 2656 cachep->buffer_size / PAGE_SIZE, 0);
1da177e4
LT
2657#else
2658 if (cachep->ctor)
51cc5068 2659 cachep->ctor(objp);
1da177e4 2660#endif
b28a02de 2661 slab_bufctl(slabp)[i] = i + 1;
1da177e4 2662 }
b28a02de 2663 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
1da177e4
LT
2664}
2665
343e0d7a 2666static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
1da177e4 2667{
4b51d669
CL
2668 if (CONFIG_ZONE_DMA_FLAG) {
2669 if (flags & GFP_DMA)
2670 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2671 else
2672 BUG_ON(cachep->gfpflags & GFP_DMA);
2673 }
1da177e4
LT
2674}
2675
a737b3e2
AM
2676static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2677 int nodeid)
78d382d7 2678{
8fea4e96 2679 void *objp = index_to_obj(cachep, slabp, slabp->free);
78d382d7
MD
2680 kmem_bufctl_t next;
2681
2682 slabp->inuse++;
2683 next = slab_bufctl(slabp)[slabp->free];
2684#if DEBUG
2685 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2686 WARN_ON(slabp->nodeid != nodeid);
2687#endif
2688 slabp->free = next;
2689
2690 return objp;
2691}
2692
a737b3e2
AM
2693static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2694 void *objp, int nodeid)
78d382d7 2695{
8fea4e96 2696 unsigned int objnr = obj_to_index(cachep, slabp, objp);
78d382d7
MD
2697
2698#if DEBUG
2699 /* Verify that the slab belongs to the intended node */
2700 WARN_ON(slabp->nodeid != nodeid);
2701
871751e2 2702 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
78d382d7 2703 printk(KERN_ERR "slab: double free detected in cache "
a737b3e2 2704 "'%s', objp %p\n", cachep->name, objp);
78d382d7
MD
2705 BUG();
2706 }
2707#endif
2708 slab_bufctl(slabp)[objnr] = slabp->free;
2709 slabp->free = objnr;
2710 slabp->inuse--;
2711}
2712
4776874f
PE
2713/*
2714 * Map pages beginning at addr to the given cache and slab. This is required
2715 * for the slab allocator to be able to lookup the cache and slab of a
2716 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2717 */
2718static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2719 void *addr)
1da177e4 2720{
4776874f 2721 int nr_pages;
1da177e4
LT
2722 struct page *page;
2723
4776874f 2724 page = virt_to_page(addr);
84097518 2725
4776874f 2726 nr_pages = 1;
84097518 2727 if (likely(!PageCompound(page)))
4776874f
PE
2728 nr_pages <<= cache->gfporder;
2729
1da177e4 2730 do {
4776874f
PE
2731 page_set_cache(page, cache);
2732 page_set_slab(page, slab);
1da177e4 2733 page++;
4776874f 2734 } while (--nr_pages);
1da177e4
LT
2735}
2736
2737/*
2738 * Grow (by 1) the number of slabs within a cache. This is called by
2739 * kmem_cache_alloc() when there are no active objs left in a cache.
2740 */
3c517a61
CL
2741static int cache_grow(struct kmem_cache *cachep,
2742 gfp_t flags, int nodeid, void *objp)
1da177e4 2743{
b28a02de 2744 struct slab *slabp;
b28a02de
PE
2745 size_t offset;
2746 gfp_t local_flags;
e498be7d 2747 struct kmem_list3 *l3;
1da177e4 2748
a737b3e2
AM
2749 /*
2750 * Be lazy and only check for valid flags here, keeping it out of the
2751 * critical path in kmem_cache_alloc().
1da177e4 2752 */
6cb06229
CL
2753 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2754 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
1da177e4 2755
2e1217cf 2756 /* Take the l3 list lock to change the colour_next on this node */
1da177e4 2757 check_irq_off();
2e1217cf
RT
2758 l3 = cachep->nodelists[nodeid];
2759 spin_lock(&l3->list_lock);
1da177e4
LT
2760
2761 /* Get colour for the slab, and cal the next value. */
2e1217cf
RT
2762 offset = l3->colour_next;
2763 l3->colour_next++;
2764 if (l3->colour_next >= cachep->colour)
2765 l3->colour_next = 0;
2766 spin_unlock(&l3->list_lock);
1da177e4 2767
2e1217cf 2768 offset *= cachep->colour_off;
1da177e4
LT
2769
2770 if (local_flags & __GFP_WAIT)
2771 local_irq_enable();
2772
2773 /*
2774 * The test for missing atomic flag is performed here, rather than
2775 * the more obvious place, simply to reduce the critical path length
2776 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2777 * will eventually be caught here (where it matters).
2778 */
2779 kmem_flagcheck(cachep, flags);
2780
a737b3e2
AM
2781 /*
2782 * Get mem for the objs. Attempt to allocate a physical page from
2783 * 'nodeid'.
e498be7d 2784 */
3c517a61 2785 if (!objp)
b8c1c5da 2786 objp = kmem_getpages(cachep, local_flags, nodeid);
a737b3e2 2787 if (!objp)
1da177e4
LT
2788 goto failed;
2789
2790 /* Get slab management. */
3c517a61 2791 slabp = alloc_slabmgmt(cachep, objp, offset,
6cb06229 2792 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
a737b3e2 2793 if (!slabp)
1da177e4
LT
2794 goto opps1;
2795
4776874f 2796 slab_map_pages(cachep, slabp, objp);
1da177e4 2797
a35afb83 2798 cache_init_objs(cachep, slabp);
1da177e4
LT
2799
2800 if (local_flags & __GFP_WAIT)
2801 local_irq_disable();
2802 check_irq_off();
e498be7d 2803 spin_lock(&l3->list_lock);
1da177e4
LT
2804
2805 /* Make slab active. */
e498be7d 2806 list_add_tail(&slabp->list, &(l3->slabs_free));
1da177e4 2807 STATS_INC_GROWN(cachep);
e498be7d
CL
2808 l3->free_objects += cachep->num;
2809 spin_unlock(&l3->list_lock);
1da177e4 2810 return 1;
a737b3e2 2811opps1:
1da177e4 2812 kmem_freepages(cachep, objp);
a737b3e2 2813failed:
1da177e4
LT
2814 if (local_flags & __GFP_WAIT)
2815 local_irq_disable();
2816 return 0;
2817}
2818
2819#if DEBUG
2820
2821/*
2822 * Perform extra freeing checks:
2823 * - detect bad pointers.
2824 * - POISON/RED_ZONE checking
1da177e4
LT
2825 */
2826static void kfree_debugcheck(const void *objp)
2827{
1da177e4
LT
2828 if (!virt_addr_valid(objp)) {
2829 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
b28a02de
PE
2830 (unsigned long)objp);
2831 BUG();
1da177e4 2832 }
1da177e4
LT
2833}
2834
58ce1fd5
PE
2835static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2836{
b46b8f19 2837 unsigned long long redzone1, redzone2;
58ce1fd5
PE
2838
2839 redzone1 = *dbg_redzone1(cache, obj);
2840 redzone2 = *dbg_redzone2(cache, obj);
2841
2842 /*
2843 * Redzone is ok.
2844 */
2845 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2846 return;
2847
2848 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2849 slab_error(cache, "double free detected");
2850 else
2851 slab_error(cache, "memory outside object was overwritten");
2852
b46b8f19 2853 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
58ce1fd5
PE
2854 obj, redzone1, redzone2);
2855}
2856
343e0d7a 2857static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
b28a02de 2858 void *caller)
1da177e4
LT
2859{
2860 struct page *page;
2861 unsigned int objnr;
2862 struct slab *slabp;
2863
80cbd911
MW
2864 BUG_ON(virt_to_cache(objp) != cachep);
2865
3dafccf2 2866 objp -= obj_offset(cachep);
1da177e4 2867 kfree_debugcheck(objp);
b49af68f 2868 page = virt_to_head_page(objp);
1da177e4 2869
065d41cb 2870 slabp = page_get_slab(page);
1da177e4
LT
2871
2872 if (cachep->flags & SLAB_RED_ZONE) {
58ce1fd5 2873 verify_redzone_free(cachep, objp);
1da177e4
LT
2874 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2875 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2876 }
2877 if (cachep->flags & SLAB_STORE_USER)
2878 *dbg_userword(cachep, objp) = caller;
2879
8fea4e96 2880 objnr = obj_to_index(cachep, slabp, objp);
1da177e4
LT
2881
2882 BUG_ON(objnr >= cachep->num);
8fea4e96 2883 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
1da177e4 2884
871751e2
AV
2885#ifdef CONFIG_DEBUG_SLAB_LEAK
2886 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2887#endif
1da177e4
LT
2888 if (cachep->flags & SLAB_POISON) {
2889#ifdef CONFIG_DEBUG_PAGEALLOC
a737b3e2 2890 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
1da177e4 2891 store_stackinfo(cachep, objp, (unsigned long)caller);
b28a02de 2892 kernel_map_pages(virt_to_page(objp),
3dafccf2 2893 cachep->buffer_size / PAGE_SIZE, 0);
1da177e4
LT
2894 } else {
2895 poison_obj(cachep, objp, POISON_FREE);
2896 }
2897#else
2898 poison_obj(cachep, objp, POISON_FREE);
2899#endif
2900 }
2901 return objp;
2902}
2903
343e0d7a 2904static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
1da177e4
LT
2905{
2906 kmem_bufctl_t i;
2907 int entries = 0;
b28a02de 2908
1da177e4
LT
2909 /* Check slab's freelist to see if this obj is there. */
2910 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2911 entries++;
2912 if (entries > cachep->num || i >= cachep->num)
2913 goto bad;
2914 }
2915 if (entries != cachep->num - slabp->inuse) {
a737b3e2
AM
2916bad:
2917 printk(KERN_ERR "slab: Internal list corruption detected in "
2918 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2919 cachep->name, cachep->num, slabp, slabp->inuse);
b28a02de 2920 for (i = 0;
264132bc 2921 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
b28a02de 2922 i++) {
a737b3e2 2923 if (i % 16 == 0)
1da177e4 2924 printk("\n%03x:", i);
b28a02de 2925 printk(" %02x", ((unsigned char *)slabp)[i]);
1da177e4
LT
2926 }
2927 printk("\n");
2928 BUG();
2929 }
2930}
2931#else
2932#define kfree_debugcheck(x) do { } while(0)
2933#define cache_free_debugcheck(x,objp,z) (objp)
2934#define check_slabp(x,y) do { } while(0)
2935#endif
2936
343e0d7a 2937static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
1da177e4
LT
2938{
2939 int batchcount;
2940 struct kmem_list3 *l3;
2941 struct array_cache *ac;
1ca4cb24
PE
2942 int node;
2943
6d2144d3 2944retry:
1da177e4 2945 check_irq_off();
6d2144d3 2946 node = numa_node_id();
9a2dba4b 2947 ac = cpu_cache_get(cachep);
1da177e4
LT
2948 batchcount = ac->batchcount;
2949 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
a737b3e2
AM
2950 /*
2951 * If there was little recent activity on this cache, then
2952 * perform only a partial refill. Otherwise we could generate
2953 * refill bouncing.
1da177e4
LT
2954 */
2955 batchcount = BATCHREFILL_LIMIT;
2956 }
1ca4cb24 2957 l3 = cachep->nodelists[node];
e498be7d
CL
2958
2959 BUG_ON(ac->avail > 0 || !l3);
2960 spin_lock(&l3->list_lock);
1da177e4 2961
3ded175a 2962 /* See if we can refill from the shared array */
44b57f1c
NP
2963 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
2964 l3->shared->touched = 1;
3ded175a 2965 goto alloc_done;
44b57f1c 2966 }
3ded175a 2967
1da177e4
LT
2968 while (batchcount > 0) {
2969 struct list_head *entry;
2970 struct slab *slabp;
2971 /* Get slab alloc is to come from. */
2972 entry = l3->slabs_partial.next;
2973 if (entry == &l3->slabs_partial) {
2974 l3->free_touched = 1;
2975 entry = l3->slabs_free.next;
2976 if (entry == &l3->slabs_free)
2977 goto must_grow;
2978 }
2979
2980 slabp = list_entry(entry, struct slab, list);
2981 check_slabp(cachep, slabp);
2982 check_spinlock_acquired(cachep);
714b8171
PE
2983
2984 /*
2985 * The slab was either on partial or free list so
2986 * there must be at least one object available for
2987 * allocation.
2988 */
249b9f33 2989 BUG_ON(slabp->inuse >= cachep->num);
714b8171 2990
1da177e4 2991 while (slabp->inuse < cachep->num && batchcount--) {
1da177e4
LT
2992 STATS_INC_ALLOCED(cachep);
2993 STATS_INC_ACTIVE(cachep);
2994 STATS_SET_HIGH(cachep);
2995
78d382d7 2996 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
1ca4cb24 2997 node);
1da177e4
LT
2998 }
2999 check_slabp(cachep, slabp);
3000
3001 /* move slabp to correct slabp list: */
3002 list_del(&slabp->list);
3003 if (slabp->free == BUFCTL_END)
3004 list_add(&slabp->list, &l3->slabs_full);
3005 else
3006 list_add(&slabp->list, &l3->slabs_partial);
3007 }
3008
a737b3e2 3009must_grow:
1da177e4 3010 l3->free_objects -= ac->avail;
a737b3e2 3011alloc_done:
e498be7d 3012 spin_unlock(&l3->list_lock);
1da177e4
LT
3013
3014 if (unlikely(!ac->avail)) {
3015 int x;
3c517a61 3016 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
e498be7d 3017
a737b3e2 3018 /* cache_grow can reenable interrupts, then ac could change. */
9a2dba4b 3019 ac = cpu_cache_get(cachep);
a737b3e2 3020 if (!x && ac->avail == 0) /* no objects in sight? abort */
1da177e4
LT
3021 return NULL;
3022
a737b3e2 3023 if (!ac->avail) /* objects refilled by interrupt? */
1da177e4
LT
3024 goto retry;
3025 }
3026 ac->touched = 1;
e498be7d 3027 return ac->entry[--ac->avail];
1da177e4
LT
3028}
3029
a737b3e2
AM
3030static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3031 gfp_t flags)
1da177e4
LT
3032{
3033 might_sleep_if(flags & __GFP_WAIT);
3034#if DEBUG
3035 kmem_flagcheck(cachep, flags);
3036#endif
3037}
3038
3039#if DEBUG
a737b3e2
AM
3040static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3041 gfp_t flags, void *objp, void *caller)
1da177e4 3042{
b28a02de 3043 if (!objp)
1da177e4 3044 return objp;
b28a02de 3045 if (cachep->flags & SLAB_POISON) {
1da177e4 3046#ifdef CONFIG_DEBUG_PAGEALLOC
3dafccf2 3047 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
b28a02de 3048 kernel_map_pages(virt_to_page(objp),
3dafccf2 3049 cachep->buffer_size / PAGE_SIZE, 1);
1da177e4
LT
3050 else
3051 check_poison_obj(cachep, objp);
3052#else
3053 check_poison_obj(cachep, objp);
3054#endif
3055 poison_obj(cachep, objp, POISON_INUSE);
3056 }
3057 if (cachep->flags & SLAB_STORE_USER)
3058 *dbg_userword(cachep, objp) = caller;
3059
3060 if (cachep->flags & SLAB_RED_ZONE) {
a737b3e2
AM
3061 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3062 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3063 slab_error(cachep, "double free, or memory outside"
3064 " object was overwritten");
b28a02de 3065 printk(KERN_ERR
b46b8f19 3066 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
a737b3e2
AM
3067 objp, *dbg_redzone1(cachep, objp),
3068 *dbg_redzone2(cachep, objp));
1da177e4
LT
3069 }
3070 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3071 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3072 }
871751e2
AV
3073#ifdef CONFIG_DEBUG_SLAB_LEAK
3074 {
3075 struct slab *slabp;
3076 unsigned objnr;
3077
b49af68f 3078 slabp = page_get_slab(virt_to_head_page(objp));
871751e2
AV
3079 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3080 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3081 }
3082#endif
3dafccf2 3083 objp += obj_offset(cachep);
4f104934 3084 if (cachep->ctor && cachep->flags & SLAB_POISON)
51cc5068 3085 cachep->ctor(objp);
a44b56d3
KH
3086#if ARCH_SLAB_MINALIGN
3087 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3088 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3089 objp, ARCH_SLAB_MINALIGN);
3090 }
3091#endif
1da177e4
LT
3092 return objp;
3093}
3094#else
3095#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3096#endif
3097
773ff60e 3098static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
8a8b6502
AM
3099{
3100 if (cachep == &cache_cache)
773ff60e 3101 return false;
8a8b6502 3102
4c13dd3b 3103 return should_failslab(obj_size(cachep), flags, cachep->flags);
8a8b6502
AM
3104}
3105
343e0d7a 3106static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
1da177e4 3107{
b28a02de 3108 void *objp;
1da177e4
LT
3109 struct array_cache *ac;
3110
5c382300 3111 check_irq_off();
8a8b6502 3112
9a2dba4b 3113 ac = cpu_cache_get(cachep);
1da177e4
LT
3114 if (likely(ac->avail)) {
3115 STATS_INC_ALLOCHIT(cachep);
3116 ac->touched = 1;
e498be7d 3117 objp = ac->entry[--ac->avail];
1da177e4
LT
3118 } else {
3119 STATS_INC_ALLOCMISS(cachep);
3120 objp = cache_alloc_refill(cachep, flags);
ddbf2e83
O
3121 /*
3122 * the 'ac' may be updated by cache_alloc_refill(),
3123 * and kmemleak_erase() requires its correct value.
3124 */
3125 ac = cpu_cache_get(cachep);
1da177e4 3126 }
d5cff635
CM
3127 /*
3128 * To avoid a false negative, if an object that is in one of the
3129 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3130 * treat the array pointers as a reference to the object.
3131 */
f3d8b53a
O
3132 if (objp)
3133 kmemleak_erase(&ac->entry[ac->avail]);
5c382300
AK
3134 return objp;
3135}
3136
e498be7d 3137#ifdef CONFIG_NUMA
c61afb18 3138/*
b2455396 3139 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
c61afb18
PJ
3140 *
3141 * If we are in_interrupt, then process context, including cpusets and
3142 * mempolicy, may not apply and should not be used for allocation policy.
3143 */
3144static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3145{
3146 int nid_alloc, nid_here;
3147
765c4507 3148 if (in_interrupt() || (flags & __GFP_THISNODE))
c61afb18
PJ
3149 return NULL;
3150 nid_alloc = nid_here = numa_node_id();
3151 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3152 nid_alloc = cpuset_mem_spread_node();
3153 else if (current->mempolicy)
3154 nid_alloc = slab_node(current->mempolicy);
3155 if (nid_alloc != nid_here)
8b98c169 3156 return ____cache_alloc_node(cachep, flags, nid_alloc);
c61afb18
PJ
3157 return NULL;
3158}
3159
765c4507
CL
3160/*
3161 * Fallback function if there was no memory available and no objects on a
3c517a61
CL
3162 * certain node and fall back is permitted. First we scan all the
3163 * available nodelists for available objects. If that fails then we
3164 * perform an allocation without specifying a node. This allows the page
3165 * allocator to do its reclaim / fallback magic. We then insert the
3166 * slab into the proper nodelist and then allocate from it.
765c4507 3167 */
8c8cc2c1 3168static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
765c4507 3169{
8c8cc2c1
PE
3170 struct zonelist *zonelist;
3171 gfp_t local_flags;
dd1a239f 3172 struct zoneref *z;
54a6eb5c
MG
3173 struct zone *zone;
3174 enum zone_type high_zoneidx = gfp_zone(flags);
765c4507 3175 void *obj = NULL;
3c517a61 3176 int nid;
8c8cc2c1
PE
3177
3178 if (flags & __GFP_THISNODE)
3179 return NULL;
3180
0e88460d 3181 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
6cb06229 3182 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
765c4507 3183
3c517a61
CL
3184retry:
3185 /*
3186 * Look through allowed nodes for objects available
3187 * from existing per node queues.
3188 */
54a6eb5c
MG
3189 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3190 nid = zone_to_nid(zone);
aedb0eb1 3191
54a6eb5c 3192 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3c517a61 3193 cache->nodelists[nid] &&
481c5346 3194 cache->nodelists[nid]->free_objects) {
3c517a61
CL
3195 obj = ____cache_alloc_node(cache,
3196 flags | GFP_THISNODE, nid);
481c5346
CL
3197 if (obj)
3198 break;
3199 }
3c517a61
CL
3200 }
3201
cfce6604 3202 if (!obj) {
3c517a61
CL
3203 /*
3204 * This allocation will be performed within the constraints
3205 * of the current cpuset / memory policy requirements.
3206 * We may trigger various forms of reclaim on the allowed
3207 * set and go into memory reserves if necessary.
3208 */
dd47ea75
CL
3209 if (local_flags & __GFP_WAIT)
3210 local_irq_enable();
3211 kmem_flagcheck(cache, flags);
6484eb3e 3212 obj = kmem_getpages(cache, local_flags, numa_node_id());
dd47ea75
CL
3213 if (local_flags & __GFP_WAIT)
3214 local_irq_disable();
3c517a61
CL
3215 if (obj) {
3216 /*
3217 * Insert into the appropriate per node queues
3218 */
3219 nid = page_to_nid(virt_to_page(obj));
3220 if (cache_grow(cache, flags, nid, obj)) {
3221 obj = ____cache_alloc_node(cache,
3222 flags | GFP_THISNODE, nid);
3223 if (!obj)
3224 /*
3225 * Another processor may allocate the
3226 * objects in the slab since we are
3227 * not holding any locks.
3228 */
3229 goto retry;
3230 } else {
b6a60451 3231 /* cache_grow already freed obj */
3c517a61
CL
3232 obj = NULL;
3233 }
3234 }
aedb0eb1 3235 }
765c4507
CL
3236 return obj;
3237}
3238
e498be7d
CL
3239/*
3240 * A interface to enable slab creation on nodeid
1da177e4 3241 */
8b98c169 3242static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
a737b3e2 3243 int nodeid)
e498be7d
CL
3244{
3245 struct list_head *entry;
b28a02de
PE
3246 struct slab *slabp;
3247 struct kmem_list3 *l3;
3248 void *obj;
b28a02de
PE
3249 int x;
3250
3251 l3 = cachep->nodelists[nodeid];
3252 BUG_ON(!l3);
3253
a737b3e2 3254retry:
ca3b9b91 3255 check_irq_off();
b28a02de
PE
3256 spin_lock(&l3->list_lock);
3257 entry = l3->slabs_partial.next;
3258 if (entry == &l3->slabs_partial) {
3259 l3->free_touched = 1;
3260 entry = l3->slabs_free.next;
3261 if (entry == &l3->slabs_free)
3262 goto must_grow;
3263 }
3264
3265 slabp = list_entry(entry, struct slab, list);
3266 check_spinlock_acquired_node(cachep, nodeid);
3267 check_slabp(cachep, slabp);
3268
3269 STATS_INC_NODEALLOCS(cachep);
3270 STATS_INC_ACTIVE(cachep);
3271 STATS_SET_HIGH(cachep);
3272
3273 BUG_ON(slabp->inuse == cachep->num);
3274
78d382d7 3275 obj = slab_get_obj(cachep, slabp, nodeid);
b28a02de
PE
3276 check_slabp(cachep, slabp);
3277 l3->free_objects--;
3278 /* move slabp to correct slabp list: */
3279 list_del(&slabp->list);
3280
a737b3e2 3281 if (slabp->free == BUFCTL_END)
b28a02de 3282 list_add(&slabp->list, &l3->slabs_full);
a737b3e2 3283 else
b28a02de 3284 list_add(&slabp->list, &l3->slabs_partial);
e498be7d 3285
b28a02de
PE
3286 spin_unlock(&l3->list_lock);
3287 goto done;
e498be7d 3288
a737b3e2 3289must_grow:
b28a02de 3290 spin_unlock(&l3->list_lock);
3c517a61 3291 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
765c4507
CL
3292 if (x)
3293 goto retry;
1da177e4 3294
8c8cc2c1 3295 return fallback_alloc(cachep, flags);
e498be7d 3296
a737b3e2 3297done:
b28a02de 3298 return obj;
e498be7d 3299}
8c8cc2c1
PE
3300
3301/**
3302 * kmem_cache_alloc_node - Allocate an object on the specified node
3303 * @cachep: The cache to allocate from.
3304 * @flags: See kmalloc().
3305 * @nodeid: node number of the target node.
3306 * @caller: return address of caller, used for debug information
3307 *
3308 * Identical to kmem_cache_alloc but it will allocate memory on the given
3309 * node, which can improve the performance for cpu bound structures.
3310 *
3311 * Fallback to other node is possible if __GFP_THISNODE is not set.
3312 */
3313static __always_inline void *
3314__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3315 void *caller)
3316{
3317 unsigned long save_flags;
3318 void *ptr;
3319
dcce284a 3320 flags &= gfp_allowed_mask;
7e85ee0c 3321
cf40bd16
NP
3322 lockdep_trace_alloc(flags);
3323
773ff60e 3324 if (slab_should_failslab(cachep, flags))
824ebef1
AM
3325 return NULL;
3326
8c8cc2c1
PE
3327 cache_alloc_debugcheck_before(cachep, flags);
3328 local_irq_save(save_flags);
3329
8e15b79c 3330 if (nodeid == -1)
8c8cc2c1
PE
3331 nodeid = numa_node_id();
3332
3333 if (unlikely(!cachep->nodelists[nodeid])) {
3334 /* Node not bootstrapped yet */
3335 ptr = fallback_alloc(cachep, flags);
3336 goto out;
3337 }
3338
3339 if (nodeid == numa_node_id()) {
3340 /*
3341 * Use the locally cached objects if possible.
3342 * However ____cache_alloc does not allow fallback
3343 * to other nodes. It may fail while we still have
3344 * objects on other nodes available.
3345 */
3346 ptr = ____cache_alloc(cachep, flags);
3347 if (ptr)
3348 goto out;
3349 }
3350 /* ___cache_alloc_node can fall back to other nodes */
3351 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3352 out:
3353 local_irq_restore(save_flags);
3354 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
d5cff635
CM
3355 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3356 flags);
8c8cc2c1 3357
c175eea4
PE
3358 if (likely(ptr))
3359 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3360
d07dbea4
CL
3361 if (unlikely((flags & __GFP_ZERO) && ptr))
3362 memset(ptr, 0, obj_size(cachep));
3363
8c8cc2c1
PE
3364 return ptr;
3365}
3366
3367static __always_inline void *
3368__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3369{
3370 void *objp;
3371
3372 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3373 objp = alternate_node_alloc(cache, flags);
3374 if (objp)
3375 goto out;
3376 }
3377 objp = ____cache_alloc(cache, flags);
3378
3379 /*
3380 * We may just have run out of memory on the local node.
3381 * ____cache_alloc_node() knows how to locate memory on other nodes
3382 */
3383 if (!objp)
3384 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3385
3386 out:
3387 return objp;
3388}
3389#else
3390
3391static __always_inline void *
3392__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3393{
3394 return ____cache_alloc(cachep, flags);
3395}
3396
3397#endif /* CONFIG_NUMA */
3398
3399static __always_inline void *
3400__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3401{
3402 unsigned long save_flags;
3403 void *objp;
3404
dcce284a 3405 flags &= gfp_allowed_mask;
7e85ee0c 3406
cf40bd16
NP
3407 lockdep_trace_alloc(flags);
3408
773ff60e 3409 if (slab_should_failslab(cachep, flags))
824ebef1
AM
3410 return NULL;
3411
8c8cc2c1
PE
3412 cache_alloc_debugcheck_before(cachep, flags);
3413 local_irq_save(save_flags);
3414 objp = __do_cache_alloc(cachep, flags);
3415 local_irq_restore(save_flags);
3416 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
d5cff635
CM
3417 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3418 flags);
8c8cc2c1
PE
3419 prefetchw(objp);
3420
c175eea4
PE
3421 if (likely(objp))
3422 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3423
d07dbea4
CL
3424 if (unlikely((flags & __GFP_ZERO) && objp))
3425 memset(objp, 0, obj_size(cachep));
3426
8c8cc2c1
PE
3427 return objp;
3428}
e498be7d
CL
3429
3430/*
3431 * Caller needs to acquire correct kmem_list's list_lock
3432 */
343e0d7a 3433static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
b28a02de 3434 int node)
1da177e4
LT
3435{
3436 int i;
e498be7d 3437 struct kmem_list3 *l3;
1da177e4
LT
3438
3439 for (i = 0; i < nr_objects; i++) {
3440 void *objp = objpp[i];
3441 struct slab *slabp;
1da177e4 3442
6ed5eb22 3443 slabp = virt_to_slab(objp);
ff69416e 3444 l3 = cachep->nodelists[node];
1da177e4 3445 list_del(&slabp->list);
ff69416e 3446 check_spinlock_acquired_node(cachep, node);
1da177e4 3447 check_slabp(cachep, slabp);
78d382d7 3448 slab_put_obj(cachep, slabp, objp, node);
1da177e4 3449 STATS_DEC_ACTIVE(cachep);
e498be7d 3450 l3->free_objects++;
1da177e4
LT
3451 check_slabp(cachep, slabp);
3452
3453 /* fixup slab chains */
3454 if (slabp->inuse == 0) {
e498be7d
CL
3455 if (l3->free_objects > l3->free_limit) {
3456 l3->free_objects -= cachep->num;
e5ac9c5a
RT
3457 /* No need to drop any previously held
3458 * lock here, even if we have a off-slab slab
3459 * descriptor it is guaranteed to come from
3460 * a different cache, refer to comments before
3461 * alloc_slabmgmt.
3462 */
1da177e4
LT
3463 slab_destroy(cachep, slabp);
3464 } else {
e498be7d 3465 list_add(&slabp->list, &l3->slabs_free);
1da177e4
LT
3466 }
3467 } else {
3468 /* Unconditionally move a slab to the end of the
3469 * partial list on free - maximum time for the
3470 * other objects to be freed, too.
3471 */
e498be7d 3472 list_add_tail(&slabp->list, &l3->slabs_partial);
1da177e4
LT
3473 }
3474 }
3475}
3476
343e0d7a 3477static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
1da177e4
LT
3478{
3479 int batchcount;
e498be7d 3480 struct kmem_list3 *l3;
ff69416e 3481 int node = numa_node_id();
1da177e4
LT
3482
3483 batchcount = ac->batchcount;
3484#if DEBUG
3485 BUG_ON(!batchcount || batchcount > ac->avail);
3486#endif
3487 check_irq_off();
ff69416e 3488 l3 = cachep->nodelists[node];
873623df 3489 spin_lock(&l3->list_lock);
e498be7d
CL
3490 if (l3->shared) {
3491 struct array_cache *shared_array = l3->shared;
b28a02de 3492 int max = shared_array->limit - shared_array->avail;
1da177e4
LT
3493 if (max) {
3494 if (batchcount > max)
3495 batchcount = max;
e498be7d 3496 memcpy(&(shared_array->entry[shared_array->avail]),
b28a02de 3497 ac->entry, sizeof(void *) * batchcount);
1da177e4
LT
3498 shared_array->avail += batchcount;
3499 goto free_done;
3500 }
3501 }
3502
ff69416e 3503 free_block(cachep, ac->entry, batchcount, node);
a737b3e2 3504free_done:
1da177e4
LT
3505#if STATS
3506 {
3507 int i = 0;
3508 struct list_head *p;
3509
e498be7d
CL
3510 p = l3->slabs_free.next;
3511 while (p != &(l3->slabs_free)) {
1da177e4
LT
3512 struct slab *slabp;
3513
3514 slabp = list_entry(p, struct slab, list);
3515 BUG_ON(slabp->inuse);
3516
3517 i++;
3518 p = p->next;
3519 }
3520 STATS_SET_FREEABLE(cachep, i);
3521 }
3522#endif
e498be7d 3523 spin_unlock(&l3->list_lock);
1da177e4 3524 ac->avail -= batchcount;
a737b3e2 3525 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
1da177e4
LT
3526}
3527
3528/*
a737b3e2
AM
3529 * Release an obj back to its cache. If the obj has a constructed state, it must
3530 * be in this state _before_ it is released. Called with disabled ints.
1da177e4 3531 */
873623df 3532static inline void __cache_free(struct kmem_cache *cachep, void *objp)
1da177e4 3533{
9a2dba4b 3534 struct array_cache *ac = cpu_cache_get(cachep);
1da177e4
LT
3535
3536 check_irq_off();
d5cff635 3537 kmemleak_free_recursive(objp, cachep->flags);
1da177e4
LT
3538 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3539
c175eea4
PE
3540 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3541
1807a1aa
SS
3542 /*
3543 * Skip calling cache_free_alien() when the platform is not numa.
3544 * This will avoid cache misses that happen while accessing slabp (which
3545 * is per page memory reference) to get nodeid. Instead use a global
3546 * variable to skip the call, which is mostly likely to be present in
3547 * the cache.
3548 */
b6e68bc1 3549 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
729bd0b7
PE
3550 return;
3551
1da177e4
LT
3552 if (likely(ac->avail < ac->limit)) {
3553 STATS_INC_FREEHIT(cachep);
e498be7d 3554 ac->entry[ac->avail++] = objp;
1da177e4
LT
3555 return;
3556 } else {
3557 STATS_INC_FREEMISS(cachep);
3558 cache_flusharray(cachep, ac);
e498be7d 3559 ac->entry[ac->avail++] = objp;
1da177e4
LT
3560 }
3561}
3562
3563/**
3564 * kmem_cache_alloc - Allocate an object
3565 * @cachep: The cache to allocate from.
3566 * @flags: See kmalloc().
3567 *
3568 * Allocate an object from this cache. The flags are only relevant
3569 * if the cache has no available objects.
3570 */
343e0d7a 3571void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
1da177e4 3572{
36555751
EGM
3573 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3574
ca2b84cb
EGM
3575 trace_kmem_cache_alloc(_RET_IP_, ret,
3576 obj_size(cachep), cachep->buffer_size, flags);
36555751
EGM
3577
3578 return ret;
1da177e4
LT
3579}
3580EXPORT_SYMBOL(kmem_cache_alloc);
3581
0f24f128 3582#ifdef CONFIG_TRACING
36555751
EGM
3583void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3584{
3585 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3586}
3587EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3588#endif
3589
1da177e4 3590/**
7682486b 3591 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
1da177e4
LT
3592 * @cachep: the cache we're checking against
3593 * @ptr: pointer to validate
3594 *
7682486b 3595 * This verifies that the untrusted pointer looks sane;
1da177e4
LT
3596 * it is _not_ a guarantee that the pointer is actually
3597 * part of the slab cache in question, but it at least
3598 * validates that the pointer can be dereferenced and
3599 * looks half-way sane.
3600 *
3601 * Currently only used for dentry validation.
3602 */
b7f869a2 3603int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
1da177e4 3604{
3dafccf2 3605 unsigned long size = cachep->buffer_size;
1da177e4
LT
3606 struct page *page;
3607
fc1c1833 3608 if (unlikely(!kern_ptr_validate(ptr, size)))
1da177e4
LT
3609 goto out;
3610 page = virt_to_page(ptr);
3611 if (unlikely(!PageSlab(page)))
3612 goto out;
065d41cb 3613 if (unlikely(page_get_cache(page) != cachep))
1da177e4
LT
3614 goto out;
3615 return 1;
a737b3e2 3616out:
1da177e4
LT
3617 return 0;
3618}
3619
3620#ifdef CONFIG_NUMA
8b98c169
CH
3621void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3622{
36555751
EGM
3623 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3624 __builtin_return_address(0));
3625
ca2b84cb
EGM
3626 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3627 obj_size(cachep), cachep->buffer_size,
3628 flags, nodeid);
36555751
EGM
3629
3630 return ret;
8b98c169 3631}
1da177e4
LT
3632EXPORT_SYMBOL(kmem_cache_alloc_node);
3633
0f24f128 3634#ifdef CONFIG_TRACING
36555751
EGM
3635void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3636 gfp_t flags,
3637 int nodeid)
3638{
3639 return __cache_alloc_node(cachep, flags, nodeid,
3640 __builtin_return_address(0));
3641}
3642EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3643#endif
3644
8b98c169
CH
3645static __always_inline void *
3646__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
97e2bde4 3647{
343e0d7a 3648 struct kmem_cache *cachep;
36555751 3649 void *ret;
97e2bde4
MS
3650
3651 cachep = kmem_find_general_cachep(size, flags);
6cb8f913
CL
3652 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3653 return cachep;
36555751
EGM
3654 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3655
ca2b84cb
EGM
3656 trace_kmalloc_node((unsigned long) caller, ret,
3657 size, cachep->buffer_size, flags, node);
36555751
EGM
3658
3659 return ret;
97e2bde4 3660}
8b98c169 3661
0bb38a5c 3662#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
8b98c169
CH
3663void *__kmalloc_node(size_t size, gfp_t flags, int node)
3664{
3665 return __do_kmalloc_node(size, flags, node,
3666 __builtin_return_address(0));
3667}
dbe5e69d 3668EXPORT_SYMBOL(__kmalloc_node);
8b98c169
CH
3669
3670void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
ce71e27c 3671 int node, unsigned long caller)
8b98c169 3672{
ce71e27c 3673 return __do_kmalloc_node(size, flags, node, (void *)caller);
8b98c169
CH
3674}
3675EXPORT_SYMBOL(__kmalloc_node_track_caller);
3676#else
3677void *__kmalloc_node(size_t size, gfp_t flags, int node)
3678{
3679 return __do_kmalloc_node(size, flags, node, NULL);
3680}
3681EXPORT_SYMBOL(__kmalloc_node);
0bb38a5c 3682#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
8b98c169 3683#endif /* CONFIG_NUMA */
1da177e4
LT
3684
3685/**
800590f5 3686 * __do_kmalloc - allocate memory
1da177e4 3687 * @size: how many bytes of memory are required.
800590f5 3688 * @flags: the type of memory to allocate (see kmalloc).
911851e6 3689 * @caller: function caller for debug tracking of the caller
1da177e4 3690 */
7fd6b141
PE
3691static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3692 void *caller)
1da177e4 3693{
343e0d7a 3694 struct kmem_cache *cachep;
36555751 3695 void *ret;
1da177e4 3696
97e2bde4
MS
3697 /* If you want to save a few bytes .text space: replace
3698 * __ with kmem_.
3699 * Then kmalloc uses the uninlined functions instead of the inline
3700 * functions.
3701 */
3702 cachep = __find_general_cachep(size, flags);
a5c96d8a
LT
3703 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3704 return cachep;
36555751
EGM
3705 ret = __cache_alloc(cachep, flags, caller);
3706
ca2b84cb
EGM
3707 trace_kmalloc((unsigned long) caller, ret,
3708 size, cachep->buffer_size, flags);
36555751
EGM
3709
3710 return ret;
7fd6b141
PE
3711}
3712
7fd6b141 3713
0bb38a5c 3714#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
7fd6b141
PE
3715void *__kmalloc(size_t size, gfp_t flags)
3716{
871751e2 3717 return __do_kmalloc(size, flags, __builtin_return_address(0));
1da177e4
LT
3718}
3719EXPORT_SYMBOL(__kmalloc);
3720
ce71e27c 3721void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
7fd6b141 3722{
ce71e27c 3723 return __do_kmalloc(size, flags, (void *)caller);
7fd6b141
PE
3724}
3725EXPORT_SYMBOL(__kmalloc_track_caller);
1d2c8eea
CH
3726
3727#else
3728void *__kmalloc(size_t size, gfp_t flags)
3729{
3730 return __do_kmalloc(size, flags, NULL);
3731}
3732EXPORT_SYMBOL(__kmalloc);
7fd6b141
PE
3733#endif
3734
1da177e4
LT
3735/**
3736 * kmem_cache_free - Deallocate an object
3737 * @cachep: The cache the allocation was from.
3738 * @objp: The previously allocated object.
3739 *
3740 * Free an object which was previously allocated from this
3741 * cache.
3742 */
343e0d7a 3743void kmem_cache_free(struct kmem_cache *cachep, void *objp)
1da177e4
LT
3744{
3745 unsigned long flags;
3746
3747 local_irq_save(flags);
898552c9 3748 debug_check_no_locks_freed(objp, obj_size(cachep));
3ac7fe5a
TG
3749 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3750 debug_check_no_obj_freed(objp, obj_size(cachep));
873623df 3751 __cache_free(cachep, objp);
1da177e4 3752 local_irq_restore(flags);
36555751 3753
ca2b84cb 3754 trace_kmem_cache_free(_RET_IP_, objp);
1da177e4
LT
3755}
3756EXPORT_SYMBOL(kmem_cache_free);
3757
1da177e4
LT
3758/**
3759 * kfree - free previously allocated memory
3760 * @objp: pointer returned by kmalloc.
3761 *
80e93eff
PE
3762 * If @objp is NULL, no operation is performed.
3763 *
1da177e4
LT
3764 * Don't free memory not originally allocated by kmalloc()
3765 * or you will run into trouble.
3766 */
3767void kfree(const void *objp)
3768{
343e0d7a 3769 struct kmem_cache *c;
1da177e4
LT
3770 unsigned long flags;
3771
2121db74
PE
3772 trace_kfree(_RET_IP_, objp);
3773
6cb8f913 3774 if (unlikely(ZERO_OR_NULL_PTR(objp)))
1da177e4
LT
3775 return;
3776 local_irq_save(flags);
3777 kfree_debugcheck(objp);
6ed5eb22 3778 c = virt_to_cache(objp);
f9b8404c 3779 debug_check_no_locks_freed(objp, obj_size(c));
3ac7fe5a 3780 debug_check_no_obj_freed(objp, obj_size(c));
873623df 3781 __cache_free(c, (void *)objp);
1da177e4
LT
3782 local_irq_restore(flags);
3783}
3784EXPORT_SYMBOL(kfree);
3785
343e0d7a 3786unsigned int kmem_cache_size(struct kmem_cache *cachep)
1da177e4 3787{
3dafccf2 3788 return obj_size(cachep);
1da177e4
LT
3789}
3790EXPORT_SYMBOL(kmem_cache_size);
3791
343e0d7a 3792const char *kmem_cache_name(struct kmem_cache *cachep)
1944972d
ACM
3793{
3794 return cachep->name;
3795}
3796EXPORT_SYMBOL_GPL(kmem_cache_name);
3797
e498be7d 3798/*
183ff22b 3799 * This initializes kmem_list3 or resizes various caches for all nodes.
e498be7d 3800 */
83b519e8 3801static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
e498be7d
CL
3802{
3803 int node;
3804 struct kmem_list3 *l3;
cafeb02e 3805 struct array_cache *new_shared;
3395ee05 3806 struct array_cache **new_alien = NULL;
e498be7d 3807
9c09a95c 3808 for_each_online_node(node) {
cafeb02e 3809
3395ee05 3810 if (use_alien_caches) {
83b519e8 3811 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3395ee05
PM
3812 if (!new_alien)
3813 goto fail;
3814 }
cafeb02e 3815
63109846
ED
3816 new_shared = NULL;
3817 if (cachep->shared) {
3818 new_shared = alloc_arraycache(node,
0718dc2a 3819 cachep->shared*cachep->batchcount,
83b519e8 3820 0xbaadf00d, gfp);
63109846
ED
3821 if (!new_shared) {
3822 free_alien_cache(new_alien);
3823 goto fail;
3824 }
0718dc2a 3825 }
cafeb02e 3826
a737b3e2
AM
3827 l3 = cachep->nodelists[node];
3828 if (l3) {
cafeb02e
CL
3829 struct array_cache *shared = l3->shared;
3830
e498be7d
CL
3831 spin_lock_irq(&l3->list_lock);
3832
cafeb02e 3833 if (shared)
0718dc2a
CL
3834 free_block(cachep, shared->entry,
3835 shared->avail, node);
e498be7d 3836
cafeb02e
CL
3837 l3->shared = new_shared;
3838 if (!l3->alien) {
e498be7d
CL
3839 l3->alien = new_alien;
3840 new_alien = NULL;
3841 }
b28a02de 3842 l3->free_limit = (1 + nr_cpus_node(node)) *
a737b3e2 3843 cachep->batchcount + cachep->num;
e498be7d 3844 spin_unlock_irq(&l3->list_lock);
cafeb02e 3845 kfree(shared);
e498be7d
CL
3846 free_alien_cache(new_alien);
3847 continue;
3848 }
83b519e8 3849 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
0718dc2a
CL
3850 if (!l3) {
3851 free_alien_cache(new_alien);
3852 kfree(new_shared);
e498be7d 3853 goto fail;
0718dc2a 3854 }
e498be7d
CL
3855
3856 kmem_list3_init(l3);
3857 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
a737b3e2 3858 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
cafeb02e 3859 l3->shared = new_shared;
e498be7d 3860 l3->alien = new_alien;
b28a02de 3861 l3->free_limit = (1 + nr_cpus_node(node)) *
a737b3e2 3862 cachep->batchcount + cachep->num;
e498be7d
CL
3863 cachep->nodelists[node] = l3;
3864 }
cafeb02e 3865 return 0;
0718dc2a 3866
a737b3e2 3867fail:
0718dc2a
CL
3868 if (!cachep->next.next) {
3869 /* Cache is not active yet. Roll back what we did */
3870 node--;
3871 while (node >= 0) {
3872 if (cachep->nodelists[node]) {
3873 l3 = cachep->nodelists[node];
3874
3875 kfree(l3->shared);
3876 free_alien_cache(l3->alien);
3877 kfree(l3);
3878 cachep->nodelists[node] = NULL;
3879 }
3880 node--;
3881 }
3882 }
cafeb02e 3883 return -ENOMEM;
e498be7d
CL
3884}
3885
1da177e4 3886struct ccupdate_struct {
343e0d7a 3887 struct kmem_cache *cachep;
1da177e4
LT
3888 struct array_cache *new[NR_CPUS];
3889};
3890
3891static void do_ccupdate_local(void *info)
3892{
a737b3e2 3893 struct ccupdate_struct *new = info;
1da177e4
LT
3894 struct array_cache *old;
3895
3896 check_irq_off();
9a2dba4b 3897 old = cpu_cache_get(new->cachep);
e498be7d 3898
1da177e4
LT
3899 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3900 new->new[smp_processor_id()] = old;
3901}
3902
b5d8ca7c 3903/* Always called with the cache_chain_mutex held */
a737b3e2 3904static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
83b519e8 3905 int batchcount, int shared, gfp_t gfp)
1da177e4 3906{
d2e7b7d0 3907 struct ccupdate_struct *new;
2ed3a4ef 3908 int i;
1da177e4 3909
83b519e8 3910 new = kzalloc(sizeof(*new), gfp);
d2e7b7d0
SS
3911 if (!new)
3912 return -ENOMEM;
3913
e498be7d 3914 for_each_online_cpu(i) {
d2e7b7d0 3915 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
83b519e8 3916 batchcount, gfp);
d2e7b7d0 3917 if (!new->new[i]) {
b28a02de 3918 for (i--; i >= 0; i--)
d2e7b7d0
SS
3919 kfree(new->new[i]);
3920 kfree(new);
e498be7d 3921 return -ENOMEM;
1da177e4
LT
3922 }
3923 }
d2e7b7d0 3924 new->cachep = cachep;
1da177e4 3925
15c8b6c1 3926 on_each_cpu(do_ccupdate_local, (void *)new, 1);
e498be7d 3927
1da177e4 3928 check_irq_on();
1da177e4
LT
3929 cachep->batchcount = batchcount;
3930 cachep->limit = limit;
e498be7d 3931 cachep->shared = shared;
1da177e4 3932
e498be7d 3933 for_each_online_cpu(i) {
d2e7b7d0 3934 struct array_cache *ccold = new->new[i];
1da177e4
LT
3935 if (!ccold)
3936 continue;
e498be7d 3937 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
ff69416e 3938 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
e498be7d 3939 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
1da177e4
LT
3940 kfree(ccold);
3941 }
d2e7b7d0 3942 kfree(new);
83b519e8 3943 return alloc_kmemlist(cachep, gfp);
1da177e4
LT
3944}
3945
b5d8ca7c 3946/* Called with cache_chain_mutex held always */
83b519e8 3947static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
1da177e4
LT
3948{
3949 int err;
3950 int limit, shared;
3951
a737b3e2
AM
3952 /*
3953 * The head array serves three purposes:
1da177e4
LT
3954 * - create a LIFO ordering, i.e. return objects that are cache-warm
3955 * - reduce the number of spinlock operations.
a737b3e2 3956 * - reduce the number of linked list operations on the slab and
1da177e4
LT
3957 * bufctl chains: array operations are cheaper.
3958 * The numbers are guessed, we should auto-tune as described by
3959 * Bonwick.
3960 */
3dafccf2 3961 if (cachep->buffer_size > 131072)
1da177e4 3962 limit = 1;
3dafccf2 3963 else if (cachep->buffer_size > PAGE_SIZE)
1da177e4 3964 limit = 8;
3dafccf2 3965 else if (cachep->buffer_size > 1024)
1da177e4 3966 limit = 24;
3dafccf2 3967 else if (cachep->buffer_size > 256)
1da177e4
LT
3968 limit = 54;
3969 else
3970 limit = 120;
3971
a737b3e2
AM
3972 /*
3973 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
1da177e4
LT
3974 * allocation behaviour: Most allocs on one cpu, most free operations
3975 * on another cpu. For these cases, an efficient object passing between
3976 * cpus is necessary. This is provided by a shared array. The array
3977 * replaces Bonwick's magazine layer.
3978 * On uniprocessor, it's functionally equivalent (but less efficient)
3979 * to a larger limit. Thus disabled by default.
3980 */
3981 shared = 0;
364fbb29 3982 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
1da177e4 3983 shared = 8;
1da177e4
LT
3984
3985#if DEBUG
a737b3e2
AM
3986 /*
3987 * With debugging enabled, large batchcount lead to excessively long
3988 * periods with disabled local interrupts. Limit the batchcount
1da177e4
LT
3989 */
3990 if (limit > 32)
3991 limit = 32;
3992#endif
83b519e8 3993 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
1da177e4
LT
3994 if (err)
3995 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
b28a02de 3996 cachep->name, -err);
2ed3a4ef 3997 return err;
1da177e4
LT
3998}
3999
1b55253a
CL
4000/*
4001 * Drain an array if it contains any elements taking the l3 lock only if
b18e7e65
CL
4002 * necessary. Note that the l3 listlock also protects the array_cache
4003 * if drain_array() is used on the shared array.
1b55253a
CL
4004 */
4005void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4006 struct array_cache *ac, int force, int node)
1da177e4
LT
4007{
4008 int tofree;
4009
1b55253a
CL
4010 if (!ac || !ac->avail)
4011 return;
1da177e4
LT
4012 if (ac->touched && !force) {
4013 ac->touched = 0;
b18e7e65 4014 } else {
1b55253a 4015 spin_lock_irq(&l3->list_lock);
b18e7e65
CL
4016 if (ac->avail) {
4017 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4018 if (tofree > ac->avail)
4019 tofree = (ac->avail + 1) / 2;
4020 free_block(cachep, ac->entry, tofree, node);
4021 ac->avail -= tofree;
4022 memmove(ac->entry, &(ac->entry[tofree]),
4023 sizeof(void *) * ac->avail);
4024 }
1b55253a 4025 spin_unlock_irq(&l3->list_lock);
1da177e4
LT
4026 }
4027}
4028
4029/**
4030 * cache_reap - Reclaim memory from caches.
05fb6bf0 4031 * @w: work descriptor
1da177e4
LT
4032 *
4033 * Called from workqueue/eventd every few seconds.
4034 * Purpose:
4035 * - clear the per-cpu caches for this CPU.
4036 * - return freeable pages to the main free memory pool.
4037 *
a737b3e2
AM
4038 * If we cannot acquire the cache chain mutex then just give up - we'll try
4039 * again on the next iteration.
1da177e4 4040 */
7c5cae36 4041static void cache_reap(struct work_struct *w)
1da177e4 4042{
7a7c381d 4043 struct kmem_cache *searchp;
e498be7d 4044 struct kmem_list3 *l3;
aab2207c 4045 int node = numa_node_id();
bf6aede7 4046 struct delayed_work *work = to_delayed_work(w);
1da177e4 4047
7c5cae36 4048 if (!mutex_trylock(&cache_chain_mutex))
1da177e4 4049 /* Give up. Setup the next iteration. */
7c5cae36 4050 goto out;
1da177e4 4051
7a7c381d 4052 list_for_each_entry(searchp, &cache_chain, next) {
1da177e4
LT
4053 check_irq_on();
4054
35386e3b
CL
4055 /*
4056 * We only take the l3 lock if absolutely necessary and we
4057 * have established with reasonable certainty that
4058 * we can do some work if the lock was obtained.
4059 */
aab2207c 4060 l3 = searchp->nodelists[node];
35386e3b 4061
8fce4d8e 4062 reap_alien(searchp, l3);
1da177e4 4063
aab2207c 4064 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
1da177e4 4065
35386e3b
CL
4066 /*
4067 * These are racy checks but it does not matter
4068 * if we skip one check or scan twice.
4069 */
e498be7d 4070 if (time_after(l3->next_reap, jiffies))
35386e3b 4071 goto next;
1da177e4 4072
e498be7d 4073 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
1da177e4 4074
aab2207c 4075 drain_array(searchp, l3, l3->shared, 0, node);
1da177e4 4076
ed11d9eb 4077 if (l3->free_touched)
e498be7d 4078 l3->free_touched = 0;
ed11d9eb
CL
4079 else {
4080 int freed;
1da177e4 4081
ed11d9eb
CL
4082 freed = drain_freelist(searchp, l3, (l3->free_limit +
4083 5 * searchp->num - 1) / (5 * searchp->num));
4084 STATS_ADD_REAPED(searchp, freed);
4085 }
35386e3b 4086next:
1da177e4
LT
4087 cond_resched();
4088 }
4089 check_irq_on();
fc0abb14 4090 mutex_unlock(&cache_chain_mutex);
8fce4d8e 4091 next_reap_node();
7c5cae36 4092out:
a737b3e2 4093 /* Set up the next iteration */
7c5cae36 4094 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
1da177e4
LT
4095}
4096
158a9624 4097#ifdef CONFIG_SLABINFO
1da177e4 4098
85289f98 4099static void print_slabinfo_header(struct seq_file *m)
1da177e4 4100{
85289f98
PE
4101 /*
4102 * Output format version, so at least we can change it
4103 * without _too_ many complaints.
4104 */
1da177e4 4105#if STATS
85289f98 4106 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1da177e4 4107#else
85289f98 4108 seq_puts(m, "slabinfo - version: 2.1\n");
1da177e4 4109#endif
85289f98
PE
4110 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4111 "<objperslab> <pagesperslab>");
4112 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4113 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1da177e4 4114#if STATS
85289f98 4115 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
fb7faf33 4116 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
85289f98 4117 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1da177e4 4118#endif
85289f98
PE
4119 seq_putc(m, '\n');
4120}
4121
4122static void *s_start(struct seq_file *m, loff_t *pos)
4123{
4124 loff_t n = *pos;
85289f98 4125
fc0abb14 4126 mutex_lock(&cache_chain_mutex);
85289f98
PE
4127 if (!n)
4128 print_slabinfo_header(m);
b92151ba
PE
4129
4130 return seq_list_start(&cache_chain, *pos);
1da177e4
LT
4131}
4132
4133static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4134{
b92151ba 4135 return seq_list_next(p, &cache_chain, pos);
1da177e4
LT
4136}
4137
4138static void s_stop(struct seq_file *m, void *p)
4139{
fc0abb14 4140 mutex_unlock(&cache_chain_mutex);
1da177e4
LT
4141}
4142
4143static int s_show(struct seq_file *m, void *p)
4144{
b92151ba 4145 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
b28a02de
PE
4146 struct slab *slabp;
4147 unsigned long active_objs;
4148 unsigned long num_objs;
4149 unsigned long active_slabs = 0;
4150 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
e498be7d 4151 const char *name;
1da177e4 4152 char *error = NULL;
e498be7d
CL
4153 int node;
4154 struct kmem_list3 *l3;
1da177e4 4155
1da177e4
LT
4156 active_objs = 0;
4157 num_slabs = 0;
e498be7d
CL
4158 for_each_online_node(node) {
4159 l3 = cachep->nodelists[node];
4160 if (!l3)
4161 continue;
4162
ca3b9b91
RT
4163 check_irq_on();
4164 spin_lock_irq(&l3->list_lock);
e498be7d 4165
7a7c381d 4166 list_for_each_entry(slabp, &l3->slabs_full, list) {
e498be7d
CL
4167 if (slabp->inuse != cachep->num && !error)
4168 error = "slabs_full accounting error";
4169 active_objs += cachep->num;
4170 active_slabs++;
4171 }
7a7c381d 4172 list_for_each_entry(slabp, &l3->slabs_partial, list) {
e498be7d
CL
4173 if (slabp->inuse == cachep->num && !error)
4174 error = "slabs_partial inuse accounting error";
4175 if (!slabp->inuse && !error)
4176 error = "slabs_partial/inuse accounting error";
4177 active_objs += slabp->inuse;
4178 active_slabs++;
4179 }
7a7c381d 4180 list_for_each_entry(slabp, &l3->slabs_free, list) {
e498be7d
CL
4181 if (slabp->inuse && !error)
4182 error = "slabs_free/inuse accounting error";
4183 num_slabs++;
4184 }
4185 free_objects += l3->free_objects;
4484ebf1
RT
4186 if (l3->shared)
4187 shared_avail += l3->shared->avail;
e498be7d 4188
ca3b9b91 4189 spin_unlock_irq(&l3->list_lock);
1da177e4 4190 }
b28a02de
PE
4191 num_slabs += active_slabs;
4192 num_objs = num_slabs * cachep->num;
e498be7d 4193 if (num_objs - active_objs != free_objects && !error)
1da177e4
LT
4194 error = "free_objects accounting error";
4195
b28a02de 4196 name = cachep->name;
1da177e4
LT
4197 if (error)
4198 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4199
4200 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3dafccf2 4201 name, active_objs, num_objs, cachep->buffer_size,
b28a02de 4202 cachep->num, (1 << cachep->gfporder));
1da177e4 4203 seq_printf(m, " : tunables %4u %4u %4u",
b28a02de 4204 cachep->limit, cachep->batchcount, cachep->shared);
e498be7d 4205 seq_printf(m, " : slabdata %6lu %6lu %6lu",
b28a02de 4206 active_slabs, num_slabs, shared_avail);
1da177e4 4207#if STATS
b28a02de 4208 { /* list3 stats */
1da177e4
LT
4209 unsigned long high = cachep->high_mark;
4210 unsigned long allocs = cachep->num_allocations;
4211 unsigned long grown = cachep->grown;
4212 unsigned long reaped = cachep->reaped;
4213 unsigned long errors = cachep->errors;
4214 unsigned long max_freeable = cachep->max_freeable;
1da177e4 4215 unsigned long node_allocs = cachep->node_allocs;
e498be7d 4216 unsigned long node_frees = cachep->node_frees;
fb7faf33 4217 unsigned long overflows = cachep->node_overflow;
1da177e4 4218
e498be7d 4219 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
fb7faf33 4220 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
a737b3e2 4221 reaped, errors, max_freeable, node_allocs,
fb7faf33 4222 node_frees, overflows);
1da177e4
LT
4223 }
4224 /* cpu stats */
4225 {
4226 unsigned long allochit = atomic_read(&cachep->allochit);
4227 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4228 unsigned long freehit = atomic_read(&cachep->freehit);
4229 unsigned long freemiss = atomic_read(&cachep->freemiss);
4230
4231 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
b28a02de 4232 allochit, allocmiss, freehit, freemiss);
1da177e4
LT
4233 }
4234#endif
4235 seq_putc(m, '\n');
1da177e4
LT
4236 return 0;
4237}
4238
4239/*
4240 * slabinfo_op - iterator that generates /proc/slabinfo
4241 *
4242 * Output layout:
4243 * cache-name
4244 * num-active-objs
4245 * total-objs
4246 * object size
4247 * num-active-slabs
4248 * total-slabs
4249 * num-pages-per-slab
4250 * + further values on SMP and with statistics enabled
4251 */
4252
7b3c3a50 4253static const struct seq_operations slabinfo_op = {
b28a02de
PE
4254 .start = s_start,
4255 .next = s_next,
4256 .stop = s_stop,
4257 .show = s_show,
1da177e4
LT
4258};
4259
4260#define MAX_SLABINFO_WRITE 128
4261/**
4262 * slabinfo_write - Tuning for the slab allocator
4263 * @file: unused
4264 * @buffer: user buffer
4265 * @count: data length
4266 * @ppos: unused
4267 */
b28a02de
PE
4268ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4269 size_t count, loff_t *ppos)
1da177e4 4270{
b28a02de 4271 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
1da177e4 4272 int limit, batchcount, shared, res;
7a7c381d 4273 struct kmem_cache *cachep;
b28a02de 4274
1da177e4
LT
4275 if (count > MAX_SLABINFO_WRITE)
4276 return -EINVAL;
4277 if (copy_from_user(&kbuf, buffer, count))
4278 return -EFAULT;
b28a02de 4279 kbuf[MAX_SLABINFO_WRITE] = '\0';
1da177e4
LT
4280
4281 tmp = strchr(kbuf, ' ');
4282 if (!tmp)
4283 return -EINVAL;
4284 *tmp = '\0';
4285 tmp++;
4286 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4287 return -EINVAL;
4288
4289 /* Find the cache in the chain of caches. */
fc0abb14 4290 mutex_lock(&cache_chain_mutex);
1da177e4 4291 res = -EINVAL;
7a7c381d 4292 list_for_each_entry(cachep, &cache_chain, next) {
1da177e4 4293 if (!strcmp(cachep->name, kbuf)) {
a737b3e2
AM
4294 if (limit < 1 || batchcount < 1 ||
4295 batchcount > limit || shared < 0) {
e498be7d 4296 res = 0;
1da177e4 4297 } else {
e498be7d 4298 res = do_tune_cpucache(cachep, limit,
83b519e8
PE
4299 batchcount, shared,
4300 GFP_KERNEL);
1da177e4
LT
4301 }
4302 break;
4303 }
4304 }
fc0abb14 4305 mutex_unlock(&cache_chain_mutex);
1da177e4
LT
4306 if (res >= 0)
4307 res = count;
4308 return res;
4309}
871751e2 4310
7b3c3a50
AD
4311static int slabinfo_open(struct inode *inode, struct file *file)
4312{
4313 return seq_open(file, &slabinfo_op);
4314}
4315
4316static const struct file_operations proc_slabinfo_operations = {
4317 .open = slabinfo_open,
4318 .read = seq_read,
4319 .write = slabinfo_write,
4320 .llseek = seq_lseek,
4321 .release = seq_release,
4322};
4323
871751e2
AV
4324#ifdef CONFIG_DEBUG_SLAB_LEAK
4325
4326static void *leaks_start(struct seq_file *m, loff_t *pos)
4327{
871751e2 4328 mutex_lock(&cache_chain_mutex);
b92151ba 4329 return seq_list_start(&cache_chain, *pos);
871751e2
AV
4330}
4331
4332static inline int add_caller(unsigned long *n, unsigned long v)
4333{
4334 unsigned long *p;
4335 int l;
4336 if (!v)
4337 return 1;
4338 l = n[1];
4339 p = n + 2;
4340 while (l) {
4341 int i = l/2;
4342 unsigned long *q = p + 2 * i;
4343 if (*q == v) {
4344 q[1]++;
4345 return 1;
4346 }
4347 if (*q > v) {
4348 l = i;
4349 } else {
4350 p = q + 2;
4351 l -= i + 1;
4352 }
4353 }
4354 if (++n[1] == n[0])
4355 return 0;
4356 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4357 p[0] = v;
4358 p[1] = 1;
4359 return 1;
4360}
4361
4362static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4363{
4364 void *p;
4365 int i;
4366 if (n[0] == n[1])
4367 return;
4368 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4369 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4370 continue;
4371 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4372 return;
4373 }
4374}
4375
4376static void show_symbol(struct seq_file *m, unsigned long address)
4377{
4378#ifdef CONFIG_KALLSYMS
871751e2 4379 unsigned long offset, size;
9281acea 4380 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
871751e2 4381
a5c43dae 4382 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
871751e2 4383 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
a5c43dae 4384 if (modname[0])
871751e2
AV
4385 seq_printf(m, " [%s]", modname);
4386 return;
4387 }
4388#endif
4389 seq_printf(m, "%p", (void *)address);
4390}
4391
4392static int leaks_show(struct seq_file *m, void *p)
4393{
b92151ba 4394 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
871751e2
AV
4395 struct slab *slabp;
4396 struct kmem_list3 *l3;
4397 const char *name;
4398 unsigned long *n = m->private;
4399 int node;
4400 int i;
4401
4402 if (!(cachep->flags & SLAB_STORE_USER))
4403 return 0;
4404 if (!(cachep->flags & SLAB_RED_ZONE))
4405 return 0;
4406
4407 /* OK, we can do it */
4408
4409 n[1] = 0;
4410
4411 for_each_online_node(node) {
4412 l3 = cachep->nodelists[node];
4413 if (!l3)
4414 continue;
4415
4416 check_irq_on();
4417 spin_lock_irq(&l3->list_lock);
4418
7a7c381d 4419 list_for_each_entry(slabp, &l3->slabs_full, list)
871751e2 4420 handle_slab(n, cachep, slabp);
7a7c381d 4421 list_for_each_entry(slabp, &l3->slabs_partial, list)
871751e2 4422 handle_slab(n, cachep, slabp);
871751e2
AV
4423 spin_unlock_irq(&l3->list_lock);
4424 }
4425 name = cachep->name;
4426 if (n[0] == n[1]) {
4427 /* Increase the buffer size */
4428 mutex_unlock(&cache_chain_mutex);
4429 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4430 if (!m->private) {
4431 /* Too bad, we are really out */
4432 m->private = n;
4433 mutex_lock(&cache_chain_mutex);
4434 return -ENOMEM;
4435 }
4436 *(unsigned long *)m->private = n[0] * 2;
4437 kfree(n);
4438 mutex_lock(&cache_chain_mutex);
4439 /* Now make sure this entry will be retried */
4440 m->count = m->size;
4441 return 0;
4442 }
4443 for (i = 0; i < n[1]; i++) {
4444 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4445 show_symbol(m, n[2*i+2]);
4446 seq_putc(m, '\n');
4447 }
d2e7b7d0 4448
871751e2
AV
4449 return 0;
4450}
4451
a0ec95a8 4452static const struct seq_operations slabstats_op = {
871751e2
AV
4453 .start = leaks_start,
4454 .next = s_next,
4455 .stop = s_stop,
4456 .show = leaks_show,
4457};
a0ec95a8
AD
4458
4459static int slabstats_open(struct inode *inode, struct file *file)
4460{
4461 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4462 int ret = -ENOMEM;
4463 if (n) {
4464 ret = seq_open(file, &slabstats_op);
4465 if (!ret) {
4466 struct seq_file *m = file->private_data;
4467 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4468 m->private = n;
4469 n = NULL;
4470 }
4471 kfree(n);
4472 }
4473 return ret;
4474}
4475
4476static const struct file_operations proc_slabstats_operations = {
4477 .open = slabstats_open,
4478 .read = seq_read,
4479 .llseek = seq_lseek,
4480 .release = seq_release_private,
4481};
4482#endif
4483
4484static int __init slab_proc_init(void)
4485{
7b3c3a50 4486 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
a0ec95a8
AD
4487#ifdef CONFIG_DEBUG_SLAB_LEAK
4488 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
871751e2 4489#endif
a0ec95a8
AD
4490 return 0;
4491}
4492module_init(slab_proc_init);
1da177e4
LT
4493#endif
4494
00e145b6
MS
4495/**
4496 * ksize - get the actual amount of memory allocated for a given object
4497 * @objp: Pointer to the object
4498 *
4499 * kmalloc may internally round up allocations and return more memory
4500 * than requested. ksize() can be used to determine the actual amount of
4501 * memory allocated. The caller may use this additional memory, even though
4502 * a smaller amount of memory was initially specified with the kmalloc call.
4503 * The caller must guarantee that objp points to a valid object previously
4504 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4505 * must not be freed during the duration of the call.
4506 */
fd76bab2 4507size_t ksize(const void *objp)
1da177e4 4508{
ef8b4520
CL
4509 BUG_ON(!objp);
4510 if (unlikely(objp == ZERO_SIZE_PTR))
00e145b6 4511 return 0;
1da177e4 4512
6ed5eb22 4513 return obj_size(virt_to_cache(objp));
1da177e4 4514}
b1aabecd 4515EXPORT_SYMBOL(ksize);