]>
Commit | Line | Data |
---|---|---|
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> |
1da177e4 | 117 | |
1da177e4 LT |
118 | #include <asm/cacheflush.h> |
119 | #include <asm/tlbflush.h> | |
120 | #include <asm/page.h> | |
121 | ||
122 | /* | |
50953fe9 | 123 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
124 | * 0 for faster, smaller code (especially in the critical paths). |
125 | * | |
126 | * STATS - 1 to collect stats for /proc/slabinfo. | |
127 | * 0 for faster, smaller code (especially in the critical paths). | |
128 | * | |
129 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
130 | */ | |
131 | ||
132 | #ifdef CONFIG_DEBUG_SLAB | |
133 | #define DEBUG 1 | |
134 | #define STATS 1 | |
135 | #define FORCED_DEBUG 1 | |
136 | #else | |
137 | #define DEBUG 0 | |
138 | #define STATS 0 | |
139 | #define FORCED_DEBUG 0 | |
140 | #endif | |
141 | ||
1da177e4 LT |
142 | /* Shouldn't this be in a header file somewhere? */ |
143 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 144 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 145 | |
1da177e4 LT |
146 | #ifndef ARCH_KMALLOC_MINALIGN |
147 | /* | |
148 | * Enforce a minimum alignment for the kmalloc caches. | |
149 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
150 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
151 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
b46b8f19 DW |
152 | * alignment larger than the alignment of a 64-bit integer. |
153 | * ARCH_KMALLOC_MINALIGN allows that. | |
154 | * Note that increasing this value may disable some debug features. | |
1da177e4 | 155 | */ |
b46b8f19 | 156 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
1da177e4 LT |
157 | #endif |
158 | ||
159 | #ifndef ARCH_SLAB_MINALIGN | |
160 | /* | |
161 | * Enforce a minimum alignment for all caches. | |
162 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
163 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
164 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
165 | * some debug features. | |
166 | */ | |
167 | #define ARCH_SLAB_MINALIGN 0 | |
168 | #endif | |
169 | ||
170 | #ifndef ARCH_KMALLOC_FLAGS | |
171 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
172 | #endif | |
173 | ||
174 | /* Legal flag mask for kmem_cache_create(). */ | |
175 | #if DEBUG | |
50953fe9 | 176 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 177 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 178 | SLAB_CACHE_DMA | \ |
5af60839 | 179 | SLAB_STORE_USER | \ |
1da177e4 | 180 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 181 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
d5cff635 | 182 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE) |
1da177e4 | 183 | #else |
ac2b898c | 184 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 185 | SLAB_CACHE_DMA | \ |
1da177e4 | 186 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 187 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
d5cff635 | 188 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE) |
1da177e4 LT |
189 | #endif |
190 | ||
191 | /* | |
192 | * kmem_bufctl_t: | |
193 | * | |
194 | * Bufctl's are used for linking objs within a slab | |
195 | * linked offsets. | |
196 | * | |
197 | * This implementation relies on "struct page" for locating the cache & | |
198 | * slab an object belongs to. | |
199 | * This allows the bufctl structure to be small (one int), but limits | |
200 | * the number of objects a slab (not a cache) can contain when off-slab | |
201 | * bufctls are used. The limit is the size of the largest general cache | |
202 | * that does not use off-slab slabs. | |
203 | * For 32bit archs with 4 kB pages, is this 56. | |
204 | * This is not serious, as it is only for large objects, when it is unwise | |
205 | * to have too many per slab. | |
206 | * Note: This limit can be raised by introducing a general cache whose size | |
207 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
208 | */ | |
209 | ||
fa5b08d5 | 210 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
211 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
212 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
213 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
214 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 215 | |
1da177e4 LT |
216 | /* |
217 | * struct slab | |
218 | * | |
219 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
220 | * for a slab, or allocated from an general cache. | |
221 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
222 | */ | |
223 | struct slab { | |
b28a02de PE |
224 | struct list_head list; |
225 | unsigned long colouroff; | |
226 | void *s_mem; /* including colour offset */ | |
227 | unsigned int inuse; /* num of objs active in slab */ | |
228 | kmem_bufctl_t free; | |
229 | unsigned short nodeid; | |
1da177e4 LT |
230 | }; |
231 | ||
232 | /* | |
233 | * struct slab_rcu | |
234 | * | |
235 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
236 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
237 | * we need to approach a kernel structure obliquely, from its address | |
238 | * obtained without the usual locking. We can lock the structure to | |
239 | * stabilize it and check it's still at the given address, only if we | |
240 | * can be sure that the memory has not been meanwhile reused for some | |
241 | * other kind of object (which our subsystem's lock might corrupt). | |
242 | * | |
243 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
244 | * taking the spinlock within the structure expected at that address. | |
245 | * | |
246 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
247 | */ | |
248 | struct slab_rcu { | |
b28a02de | 249 | struct rcu_head head; |
343e0d7a | 250 | struct kmem_cache *cachep; |
b28a02de | 251 | void *addr; |
1da177e4 LT |
252 | }; |
253 | ||
254 | /* | |
255 | * struct array_cache | |
256 | * | |
1da177e4 LT |
257 | * Purpose: |
258 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
259 | * - reduce the number of linked list operations | |
260 | * - reduce spinlock operations | |
261 | * | |
262 | * The limit is stored in the per-cpu structure to reduce the data cache | |
263 | * footprint. | |
264 | * | |
265 | */ | |
266 | struct array_cache { | |
267 | unsigned int avail; | |
268 | unsigned int limit; | |
269 | unsigned int batchcount; | |
270 | unsigned int touched; | |
e498be7d | 271 | spinlock_t lock; |
bda5b655 | 272 | void *entry[]; /* |
a737b3e2 AM |
273 | * Must have this definition in here for the proper |
274 | * alignment of array_cache. Also simplifies accessing | |
275 | * the entries. | |
a737b3e2 | 276 | */ |
1da177e4 LT |
277 | }; |
278 | ||
a737b3e2 AM |
279 | /* |
280 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
281 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
282 | */ |
283 | #define BOOT_CPUCACHE_ENTRIES 1 | |
284 | struct arraycache_init { | |
285 | struct array_cache cache; | |
b28a02de | 286 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
287 | }; |
288 | ||
289 | /* | |
e498be7d | 290 | * The slab lists for all objects. |
1da177e4 LT |
291 | */ |
292 | struct kmem_list3 { | |
b28a02de PE |
293 | struct list_head slabs_partial; /* partial list first, better asm code */ |
294 | struct list_head slabs_full; | |
295 | struct list_head slabs_free; | |
296 | unsigned long free_objects; | |
b28a02de | 297 | unsigned int free_limit; |
2e1217cf | 298 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
299 | spinlock_t list_lock; |
300 | struct array_cache *shared; /* shared per node */ | |
301 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
302 | unsigned long next_reap; /* updated without locking */ |
303 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
304 | }; |
305 | ||
7e85ee0c PE |
306 | /* |
307 | * The slab allocator is initialized with interrupts disabled. Therefore, make | |
308 | * sure early boot allocations don't accidentally enable interrupts. | |
309 | */ | |
310 | static gfp_t slab_gfp_mask __read_mostly = SLAB_GFP_BOOT_MASK; | |
311 | ||
e498be7d CL |
312 | /* |
313 | * Need this for bootstrapping a per node allocator. | |
314 | */ | |
556a169d | 315 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
e498be7d CL |
316 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
317 | #define CACHE_CACHE 0 | |
556a169d PE |
318 | #define SIZE_AC MAX_NUMNODES |
319 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 320 | |
ed11d9eb CL |
321 | static int drain_freelist(struct kmem_cache *cache, |
322 | struct kmem_list3 *l3, int tofree); | |
323 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
324 | int node); | |
83b519e8 | 325 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 326 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 327 | |
e498be7d | 328 | /* |
a737b3e2 AM |
329 | * This function must be completely optimized away if a constant is passed to |
330 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 331 | */ |
7243cc05 | 332 | static __always_inline int index_of(const size_t size) |
e498be7d | 333 | { |
5ec8a847 SR |
334 | extern void __bad_size(void); |
335 | ||
e498be7d CL |
336 | if (__builtin_constant_p(size)) { |
337 | int i = 0; | |
338 | ||
339 | #define CACHE(x) \ | |
340 | if (size <=x) \ | |
341 | return i; \ | |
342 | else \ | |
343 | i++; | |
1c61fc40 | 344 | #include <linux/kmalloc_sizes.h> |
e498be7d | 345 | #undef CACHE |
5ec8a847 | 346 | __bad_size(); |
7243cc05 | 347 | } else |
5ec8a847 | 348 | __bad_size(); |
e498be7d CL |
349 | return 0; |
350 | } | |
351 | ||
e0a42726 IM |
352 | static int slab_early_init = 1; |
353 | ||
e498be7d CL |
354 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
355 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 356 | |
5295a74c | 357 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
358 | { |
359 | INIT_LIST_HEAD(&parent->slabs_full); | |
360 | INIT_LIST_HEAD(&parent->slabs_partial); | |
361 | INIT_LIST_HEAD(&parent->slabs_free); | |
362 | parent->shared = NULL; | |
363 | parent->alien = NULL; | |
2e1217cf | 364 | parent->colour_next = 0; |
e498be7d CL |
365 | spin_lock_init(&parent->list_lock); |
366 | parent->free_objects = 0; | |
367 | parent->free_touched = 0; | |
368 | } | |
369 | ||
a737b3e2 AM |
370 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
371 | do { \ | |
372 | INIT_LIST_HEAD(listp); \ | |
373 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
374 | } while (0) |
375 | ||
a737b3e2 AM |
376 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
377 | do { \ | |
e498be7d CL |
378 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
379 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
380 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
381 | } while (0) | |
1da177e4 LT |
382 | |
383 | /* | |
343e0d7a | 384 | * struct kmem_cache |
1da177e4 LT |
385 | * |
386 | * manages a cache. | |
387 | */ | |
b28a02de | 388 | |
2109a2d1 | 389 | struct kmem_cache { |
1da177e4 | 390 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 391 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 392 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
393 | unsigned int batchcount; |
394 | unsigned int limit; | |
395 | unsigned int shared; | |
b5d8ca7c | 396 | |
3dafccf2 | 397 | unsigned int buffer_size; |
6a2d7a95 | 398 | u32 reciprocal_buffer_size; |
b5d8ca7c | 399 | /* 3) touched by every alloc & free from the backend */ |
b5d8ca7c | 400 | |
a737b3e2 AM |
401 | unsigned int flags; /* constant flags */ |
402 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 403 | |
b5d8ca7c | 404 | /* 4) cache_grow/shrink */ |
1da177e4 | 405 | /* order of pgs per slab (2^n) */ |
b28a02de | 406 | unsigned int gfporder; |
1da177e4 LT |
407 | |
408 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 409 | gfp_t gfpflags; |
1da177e4 | 410 | |
a737b3e2 | 411 | size_t colour; /* cache colouring range */ |
b28a02de | 412 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 413 | struct kmem_cache *slabp_cache; |
b28a02de | 414 | unsigned int slab_size; |
a737b3e2 | 415 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
416 | |
417 | /* constructor func */ | |
51cc5068 | 418 | void (*ctor)(void *obj); |
1da177e4 | 419 | |
b5d8ca7c | 420 | /* 5) cache creation/removal */ |
b28a02de PE |
421 | const char *name; |
422 | struct list_head next; | |
1da177e4 | 423 | |
b5d8ca7c | 424 | /* 6) statistics */ |
1da177e4 | 425 | #if STATS |
b28a02de PE |
426 | unsigned long num_active; |
427 | unsigned long num_allocations; | |
428 | unsigned long high_mark; | |
429 | unsigned long grown; | |
430 | unsigned long reaped; | |
431 | unsigned long errors; | |
432 | unsigned long max_freeable; | |
433 | unsigned long node_allocs; | |
434 | unsigned long node_frees; | |
fb7faf33 | 435 | unsigned long node_overflow; |
b28a02de PE |
436 | atomic_t allochit; |
437 | atomic_t allocmiss; | |
438 | atomic_t freehit; | |
439 | atomic_t freemiss; | |
1da177e4 LT |
440 | #endif |
441 | #if DEBUG | |
3dafccf2 MS |
442 | /* |
443 | * If debugging is enabled, then the allocator can add additional | |
444 | * fields and/or padding to every object. buffer_size contains the total | |
445 | * object size including these internal fields, the following two | |
446 | * variables contain the offset to the user object and its size. | |
447 | */ | |
448 | int obj_offset; | |
449 | int obj_size; | |
1da177e4 | 450 | #endif |
8da3430d ED |
451 | /* |
452 | * We put nodelists[] at the end of kmem_cache, because we want to size | |
453 | * this array to nr_node_ids slots instead of MAX_NUMNODES | |
454 | * (see kmem_cache_init()) | |
455 | * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache | |
456 | * is statically defined, so we reserve the max number of nodes. | |
457 | */ | |
458 | struct kmem_list3 *nodelists[MAX_NUMNODES]; | |
459 | /* | |
460 | * Do not add fields after nodelists[] | |
461 | */ | |
1da177e4 LT |
462 | }; |
463 | ||
464 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
465 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
466 | ||
467 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
468 | /* |
469 | * Optimization question: fewer reaps means less probability for unnessary | |
470 | * cpucache drain/refill cycles. | |
1da177e4 | 471 | * |
dc6f3f27 | 472 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
473 | * which could lock up otherwise freeable slabs. |
474 | */ | |
475 | #define REAPTIMEOUT_CPUC (2*HZ) | |
476 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
477 | ||
478 | #if STATS | |
479 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
480 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
481 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
482 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 483 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
484 | #define STATS_SET_HIGH(x) \ |
485 | do { \ | |
486 | if ((x)->num_active > (x)->high_mark) \ | |
487 | (x)->high_mark = (x)->num_active; \ | |
488 | } while (0) | |
1da177e4 LT |
489 | #define STATS_INC_ERR(x) ((x)->errors++) |
490 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 491 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 492 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
493 | #define STATS_SET_FREEABLE(x, i) \ |
494 | do { \ | |
495 | if ((x)->max_freeable < i) \ | |
496 | (x)->max_freeable = i; \ | |
497 | } while (0) | |
1da177e4 LT |
498 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
499 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
500 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
501 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
502 | #else | |
503 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
504 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
505 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
506 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 507 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
508 | #define STATS_SET_HIGH(x) do { } while (0) |
509 | #define STATS_INC_ERR(x) do { } while (0) | |
510 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 511 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 512 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 513 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
514 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
515 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
516 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
517 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
518 | #endif | |
519 | ||
520 | #if DEBUG | |
1da177e4 | 521 | |
a737b3e2 AM |
522 | /* |
523 | * memory layout of objects: | |
1da177e4 | 524 | * 0 : objp |
3dafccf2 | 525 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
526 | * the end of an object is aligned with the end of the real |
527 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 528 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 529 | * redzone word. |
3dafccf2 MS |
530 | * cachep->obj_offset: The real object. |
531 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
532 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
533 | * [BYTES_PER_WORD long] | |
1da177e4 | 534 | */ |
343e0d7a | 535 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 536 | { |
3dafccf2 | 537 | return cachep->obj_offset; |
1da177e4 LT |
538 | } |
539 | ||
343e0d7a | 540 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 541 | { |
3dafccf2 | 542 | return cachep->obj_size; |
1da177e4 LT |
543 | } |
544 | ||
b46b8f19 | 545 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
546 | { |
547 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
548 | return (unsigned long long*) (objp + obj_offset(cachep) - |
549 | sizeof(unsigned long long)); | |
1da177e4 LT |
550 | } |
551 | ||
b46b8f19 | 552 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
553 | { |
554 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
555 | if (cachep->flags & SLAB_STORE_USER) | |
b46b8f19 DW |
556 | return (unsigned long long *)(objp + cachep->buffer_size - |
557 | sizeof(unsigned long long) - | |
87a927c7 | 558 | REDZONE_ALIGN); |
b46b8f19 DW |
559 | return (unsigned long long *) (objp + cachep->buffer_size - |
560 | sizeof(unsigned long long)); | |
1da177e4 LT |
561 | } |
562 | ||
343e0d7a | 563 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
564 | { |
565 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 566 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
567 | } |
568 | ||
569 | #else | |
570 | ||
3dafccf2 MS |
571 | #define obj_offset(x) 0 |
572 | #define obj_size(cachep) (cachep->buffer_size) | |
b46b8f19 DW |
573 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
574 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
575 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
576 | ||
577 | #endif | |
578 | ||
36555751 EGM |
579 | #ifdef CONFIG_KMEMTRACE |
580 | size_t slab_buffer_size(struct kmem_cache *cachep) | |
581 | { | |
582 | return cachep->buffer_size; | |
583 | } | |
584 | EXPORT_SYMBOL(slab_buffer_size); | |
585 | #endif | |
586 | ||
1da177e4 LT |
587 | /* |
588 | * Do not go above this order unless 0 objects fit into the slab. | |
589 | */ | |
590 | #define BREAK_GFP_ORDER_HI 1 | |
591 | #define BREAK_GFP_ORDER_LO 0 | |
592 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
593 | ||
a737b3e2 AM |
594 | /* |
595 | * Functions for storing/retrieving the cachep and or slab from the page | |
596 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
597 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 598 | */ |
065d41cb PE |
599 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
600 | { | |
601 | page->lru.next = (struct list_head *)cache; | |
602 | } | |
603 | ||
604 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
605 | { | |
d85f3385 | 606 | page = compound_head(page); |
ddc2e812 | 607 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
608 | return (struct kmem_cache *)page->lru.next; |
609 | } | |
610 | ||
611 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
612 | { | |
613 | page->lru.prev = (struct list_head *)slab; | |
614 | } | |
615 | ||
616 | static inline struct slab *page_get_slab(struct page *page) | |
617 | { | |
ddc2e812 | 618 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
619 | return (struct slab *)page->lru.prev; |
620 | } | |
1da177e4 | 621 | |
6ed5eb22 PE |
622 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
623 | { | |
b49af68f | 624 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
625 | return page_get_cache(page); |
626 | } | |
627 | ||
628 | static inline struct slab *virt_to_slab(const void *obj) | |
629 | { | |
b49af68f | 630 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
631 | return page_get_slab(page); |
632 | } | |
633 | ||
8fea4e96 PE |
634 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
635 | unsigned int idx) | |
636 | { | |
637 | return slab->s_mem + cache->buffer_size * idx; | |
638 | } | |
639 | ||
6a2d7a95 ED |
640 | /* |
641 | * We want to avoid an expensive divide : (offset / cache->buffer_size) | |
642 | * Using the fact that buffer_size is a constant for a particular cache, | |
643 | * we can replace (offset / cache->buffer_size) by | |
644 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
645 | */ | |
646 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
647 | const struct slab *slab, void *obj) | |
8fea4e96 | 648 | { |
6a2d7a95 ED |
649 | u32 offset = (obj - slab->s_mem); |
650 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
651 | } |
652 | ||
a737b3e2 AM |
653 | /* |
654 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
655 | */ | |
1da177e4 LT |
656 | struct cache_sizes malloc_sizes[] = { |
657 | #define CACHE(x) { .cs_size = (x) }, | |
658 | #include <linux/kmalloc_sizes.h> | |
659 | CACHE(ULONG_MAX) | |
660 | #undef CACHE | |
661 | }; | |
662 | EXPORT_SYMBOL(malloc_sizes); | |
663 | ||
664 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
665 | struct cache_names { | |
666 | char *name; | |
667 | char *name_dma; | |
668 | }; | |
669 | ||
670 | static struct cache_names __initdata cache_names[] = { | |
671 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
672 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 673 | {NULL,} |
1da177e4 LT |
674 | #undef CACHE |
675 | }; | |
676 | ||
677 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 678 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 679 | static struct arraycache_init initarray_generic = |
b28a02de | 680 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
681 | |
682 | /* internal cache of cache description objs */ | |
343e0d7a | 683 | static struct kmem_cache cache_cache = { |
b28a02de PE |
684 | .batchcount = 1, |
685 | .limit = BOOT_CPUCACHE_ENTRIES, | |
686 | .shared = 1, | |
343e0d7a | 687 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 688 | .name = "kmem_cache", |
1da177e4 LT |
689 | }; |
690 | ||
056c6241 RT |
691 | #define BAD_ALIEN_MAGIC 0x01020304ul |
692 | ||
f1aaee53 AV |
693 | #ifdef CONFIG_LOCKDEP |
694 | ||
695 | /* | |
696 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
697 | * for other slabs "off slab". | |
698 | * The locking for this is tricky in that it nests within the locks | |
699 | * of all other slabs in a few places; to deal with this special | |
700 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
701 | * |
702 | * We set lock class for alien array caches which are up during init. | |
703 | * The lock annotation will be lost if all cpus of a node goes down and | |
704 | * then comes back up during hotplug | |
f1aaee53 | 705 | */ |
056c6241 RT |
706 | static struct lock_class_key on_slab_l3_key; |
707 | static struct lock_class_key on_slab_alc_key; | |
708 | ||
709 | static inline void init_lock_keys(void) | |
f1aaee53 | 710 | |
f1aaee53 AV |
711 | { |
712 | int q; | |
056c6241 RT |
713 | struct cache_sizes *s = malloc_sizes; |
714 | ||
715 | while (s->cs_size != ULONG_MAX) { | |
716 | for_each_node(q) { | |
717 | struct array_cache **alc; | |
718 | int r; | |
719 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
720 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
721 | continue; | |
722 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
723 | alc = l3->alien; | |
724 | /* | |
725 | * FIXME: This check for BAD_ALIEN_MAGIC | |
726 | * should go away when common slab code is taught to | |
727 | * work even without alien caches. | |
728 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
729 | * for alloc_alien_cache, | |
730 | */ | |
731 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
732 | continue; | |
733 | for_each_node(r) { | |
734 | if (alc[r]) | |
735 | lockdep_set_class(&alc[r]->lock, | |
736 | &on_slab_alc_key); | |
737 | } | |
738 | } | |
739 | s++; | |
f1aaee53 AV |
740 | } |
741 | } | |
f1aaee53 | 742 | #else |
056c6241 | 743 | static inline void init_lock_keys(void) |
f1aaee53 AV |
744 | { |
745 | } | |
746 | #endif | |
747 | ||
8f5be20b | 748 | /* |
95402b38 | 749 | * Guard access to the cache-chain. |
8f5be20b | 750 | */ |
fc0abb14 | 751 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
752 | static struct list_head cache_chain; |
753 | ||
1da177e4 LT |
754 | /* |
755 | * chicken and egg problem: delay the per-cpu array allocation | |
756 | * until the general caches are up. | |
757 | */ | |
758 | static enum { | |
759 | NONE, | |
e498be7d CL |
760 | PARTIAL_AC, |
761 | PARTIAL_L3, | |
8429db5c | 762 | EARLY, |
1da177e4 LT |
763 | FULL |
764 | } g_cpucache_up; | |
765 | ||
39d24e64 MK |
766 | /* |
767 | * used by boot code to determine if it can use slab based allocator | |
768 | */ | |
769 | int slab_is_available(void) | |
770 | { | |
8429db5c | 771 | return g_cpucache_up >= EARLY; |
39d24e64 MK |
772 | } |
773 | ||
52bad64d | 774 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 775 | |
343e0d7a | 776 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
777 | { |
778 | return cachep->array[smp_processor_id()]; | |
779 | } | |
780 | ||
a737b3e2 AM |
781 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
782 | gfp_t gfpflags) | |
1da177e4 LT |
783 | { |
784 | struct cache_sizes *csizep = malloc_sizes; | |
785 | ||
786 | #if DEBUG | |
787 | /* This happens if someone tries to call | |
b28a02de PE |
788 | * kmem_cache_create(), or __kmalloc(), before |
789 | * the generic caches are initialized. | |
790 | */ | |
c7e43c78 | 791 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 792 | #endif |
6cb8f913 CL |
793 | if (!size) |
794 | return ZERO_SIZE_PTR; | |
795 | ||
1da177e4 LT |
796 | while (size > csizep->cs_size) |
797 | csizep++; | |
798 | ||
799 | /* | |
0abf40c1 | 800 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
801 | * has cs_{dma,}cachep==NULL. Thus no special case |
802 | * for large kmalloc calls required. | |
803 | */ | |
4b51d669 | 804 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
805 | if (unlikely(gfpflags & GFP_DMA)) |
806 | return csizep->cs_dmacachep; | |
4b51d669 | 807 | #endif |
1da177e4 LT |
808 | return csizep->cs_cachep; |
809 | } | |
810 | ||
b221385b | 811 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
812 | { |
813 | return __find_general_cachep(size, gfpflags); | |
814 | } | |
97e2bde4 | 815 | |
fbaccacf | 816 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 817 | { |
fbaccacf SR |
818 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
819 | } | |
1da177e4 | 820 | |
a737b3e2 AM |
821 | /* |
822 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
823 | */ | |
fbaccacf SR |
824 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
825 | size_t align, int flags, size_t *left_over, | |
826 | unsigned int *num) | |
827 | { | |
828 | int nr_objs; | |
829 | size_t mgmt_size; | |
830 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 831 | |
fbaccacf SR |
832 | /* |
833 | * The slab management structure can be either off the slab or | |
834 | * on it. For the latter case, the memory allocated for a | |
835 | * slab is used for: | |
836 | * | |
837 | * - The struct slab | |
838 | * - One kmem_bufctl_t for each object | |
839 | * - Padding to respect alignment of @align | |
840 | * - @buffer_size bytes for each object | |
841 | * | |
842 | * If the slab management structure is off the slab, then the | |
843 | * alignment will already be calculated into the size. Because | |
844 | * the slabs are all pages aligned, the objects will be at the | |
845 | * correct alignment when allocated. | |
846 | */ | |
847 | if (flags & CFLGS_OFF_SLAB) { | |
848 | mgmt_size = 0; | |
849 | nr_objs = slab_size / buffer_size; | |
850 | ||
851 | if (nr_objs > SLAB_LIMIT) | |
852 | nr_objs = SLAB_LIMIT; | |
853 | } else { | |
854 | /* | |
855 | * Ignore padding for the initial guess. The padding | |
856 | * is at most @align-1 bytes, and @buffer_size is at | |
857 | * least @align. In the worst case, this result will | |
858 | * be one greater than the number of objects that fit | |
859 | * into the memory allocation when taking the padding | |
860 | * into account. | |
861 | */ | |
862 | nr_objs = (slab_size - sizeof(struct slab)) / | |
863 | (buffer_size + sizeof(kmem_bufctl_t)); | |
864 | ||
865 | /* | |
866 | * This calculated number will be either the right | |
867 | * amount, or one greater than what we want. | |
868 | */ | |
869 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
870 | > slab_size) | |
871 | nr_objs--; | |
872 | ||
873 | if (nr_objs > SLAB_LIMIT) | |
874 | nr_objs = SLAB_LIMIT; | |
875 | ||
876 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
877 | } | |
878 | *num = nr_objs; | |
879 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
880 | } |
881 | ||
d40cee24 | 882 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 883 | |
a737b3e2 AM |
884 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
885 | char *msg) | |
1da177e4 LT |
886 | { |
887 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 888 | function, cachep->name, msg); |
1da177e4 LT |
889 | dump_stack(); |
890 | } | |
891 | ||
3395ee05 PM |
892 | /* |
893 | * By default on NUMA we use alien caches to stage the freeing of | |
894 | * objects allocated from other nodes. This causes massive memory | |
895 | * inefficiencies when using fake NUMA setup to split memory into a | |
896 | * large number of small nodes, so it can be disabled on the command | |
897 | * line | |
898 | */ | |
899 | ||
900 | static int use_alien_caches __read_mostly = 1; | |
1807a1aa | 901 | static int numa_platform __read_mostly = 1; |
3395ee05 PM |
902 | static int __init noaliencache_setup(char *s) |
903 | { | |
904 | use_alien_caches = 0; | |
905 | return 1; | |
906 | } | |
907 | __setup("noaliencache", noaliencache_setup); | |
908 | ||
8fce4d8e CL |
909 | #ifdef CONFIG_NUMA |
910 | /* | |
911 | * Special reaping functions for NUMA systems called from cache_reap(). | |
912 | * These take care of doing round robin flushing of alien caches (containing | |
913 | * objects freed on different nodes from which they were allocated) and the | |
914 | * flushing of remote pcps by calling drain_node_pages. | |
915 | */ | |
916 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
917 | ||
918 | static void init_reap_node(int cpu) | |
919 | { | |
920 | int node; | |
921 | ||
922 | node = next_node(cpu_to_node(cpu), node_online_map); | |
923 | if (node == MAX_NUMNODES) | |
442295c9 | 924 | node = first_node(node_online_map); |
8fce4d8e | 925 | |
7f6b8876 | 926 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
927 | } |
928 | ||
929 | static void next_reap_node(void) | |
930 | { | |
931 | int node = __get_cpu_var(reap_node); | |
932 | ||
8fce4d8e CL |
933 | node = next_node(node, node_online_map); |
934 | if (unlikely(node >= MAX_NUMNODES)) | |
935 | node = first_node(node_online_map); | |
936 | __get_cpu_var(reap_node) = node; | |
937 | } | |
938 | ||
939 | #else | |
940 | #define init_reap_node(cpu) do { } while (0) | |
941 | #define next_reap_node(void) do { } while (0) | |
942 | #endif | |
943 | ||
1da177e4 LT |
944 | /* |
945 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
946 | * via the workqueue/eventd. | |
947 | * Add the CPU number into the expiration time to minimize the possibility of | |
948 | * the CPUs getting into lockstep and contending for the global cache chain | |
949 | * lock. | |
950 | */ | |
897e679b | 951 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 952 | { |
52bad64d | 953 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
954 | |
955 | /* | |
956 | * When this gets called from do_initcalls via cpucache_init(), | |
957 | * init_workqueues() has already run, so keventd will be setup | |
958 | * at that time. | |
959 | */ | |
52bad64d | 960 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 961 | init_reap_node(cpu); |
65f27f38 | 962 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
963 | schedule_delayed_work_on(cpu, reap_work, |
964 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
965 | } |
966 | } | |
967 | ||
e498be7d | 968 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 969 | int batchcount, gfp_t gfp) |
1da177e4 | 970 | { |
b28a02de | 971 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
972 | struct array_cache *nc = NULL; |
973 | ||
83b519e8 | 974 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
975 | /* |
976 | * The array_cache structures contain pointers to free object. | |
977 | * However, when such objects are allocated or transfered to another | |
978 | * cache the pointers are not cleared and they could be counted as | |
979 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
980 | * not scan such objects. | |
981 | */ | |
982 | kmemleak_no_scan(nc); | |
1da177e4 LT |
983 | if (nc) { |
984 | nc->avail = 0; | |
985 | nc->limit = entries; | |
986 | nc->batchcount = batchcount; | |
987 | nc->touched = 0; | |
e498be7d | 988 | spin_lock_init(&nc->lock); |
1da177e4 LT |
989 | } |
990 | return nc; | |
991 | } | |
992 | ||
3ded175a CL |
993 | /* |
994 | * Transfer objects in one arraycache to another. | |
995 | * Locking must be handled by the caller. | |
996 | * | |
997 | * Return the number of entries transferred. | |
998 | */ | |
999 | static int transfer_objects(struct array_cache *to, | |
1000 | struct array_cache *from, unsigned int max) | |
1001 | { | |
1002 | /* Figure out how many entries to transfer */ | |
1003 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
1004 | ||
1005 | if (!nr) | |
1006 | return 0; | |
1007 | ||
1008 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1009 | sizeof(void *) *nr); | |
1010 | ||
1011 | from->avail -= nr; | |
1012 | to->avail += nr; | |
1013 | to->touched = 1; | |
1014 | return nr; | |
1015 | } | |
1016 | ||
765c4507 CL |
1017 | #ifndef CONFIG_NUMA |
1018 | ||
1019 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1020 | #define reap_alien(cachep, l3) do { } while (0) | |
1021 | ||
83b519e8 | 1022 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1023 | { |
1024 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1025 | } | |
1026 | ||
1027 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1028 | { | |
1029 | } | |
1030 | ||
1031 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1032 | { | |
1033 | return 0; | |
1034 | } | |
1035 | ||
1036 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1037 | gfp_t flags) | |
1038 | { | |
1039 | return NULL; | |
1040 | } | |
1041 | ||
8b98c169 | 1042 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1043 | gfp_t flags, int nodeid) |
1044 | { | |
1045 | return NULL; | |
1046 | } | |
1047 | ||
1048 | #else /* CONFIG_NUMA */ | |
1049 | ||
8b98c169 | 1050 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1051 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1052 | |
83b519e8 | 1053 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1054 | { |
1055 | struct array_cache **ac_ptr; | |
8ef82866 | 1056 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1057 | int i; |
1058 | ||
1059 | if (limit > 1) | |
1060 | limit = 12; | |
83b519e8 | 1061 | ac_ptr = kmalloc_node(memsize, gfp, node); |
e498be7d CL |
1062 | if (ac_ptr) { |
1063 | for_each_node(i) { | |
1064 | if (i == node || !node_online(i)) { | |
1065 | ac_ptr[i] = NULL; | |
1066 | continue; | |
1067 | } | |
83b519e8 | 1068 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1069 | if (!ac_ptr[i]) { |
cc550def | 1070 | for (i--; i >= 0; i--) |
e498be7d CL |
1071 | kfree(ac_ptr[i]); |
1072 | kfree(ac_ptr); | |
1073 | return NULL; | |
1074 | } | |
1075 | } | |
1076 | } | |
1077 | return ac_ptr; | |
1078 | } | |
1079 | ||
5295a74c | 1080 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1081 | { |
1082 | int i; | |
1083 | ||
1084 | if (!ac_ptr) | |
1085 | return; | |
e498be7d | 1086 | for_each_node(i) |
b28a02de | 1087 | kfree(ac_ptr[i]); |
e498be7d CL |
1088 | kfree(ac_ptr); |
1089 | } | |
1090 | ||
343e0d7a | 1091 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1092 | struct array_cache *ac, int node) |
e498be7d CL |
1093 | { |
1094 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1095 | ||
1096 | if (ac->avail) { | |
1097 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1098 | /* |
1099 | * Stuff objects into the remote nodes shared array first. | |
1100 | * That way we could avoid the overhead of putting the objects | |
1101 | * into the free lists and getting them back later. | |
1102 | */ | |
693f7d36 JS |
1103 | if (rl3->shared) |
1104 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1105 | |
ff69416e | 1106 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1107 | ac->avail = 0; |
1108 | spin_unlock(&rl3->list_lock); | |
1109 | } | |
1110 | } | |
1111 | ||
8fce4d8e CL |
1112 | /* |
1113 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1114 | */ | |
1115 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1116 | { | |
1117 | int node = __get_cpu_var(reap_node); | |
1118 | ||
1119 | if (l3->alien) { | |
1120 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1121 | |
1122 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1123 | __drain_alien_cache(cachep, ac, node); |
1124 | spin_unlock_irq(&ac->lock); | |
1125 | } | |
1126 | } | |
1127 | } | |
1128 | ||
a737b3e2 AM |
1129 | static void drain_alien_cache(struct kmem_cache *cachep, |
1130 | struct array_cache **alien) | |
e498be7d | 1131 | { |
b28a02de | 1132 | int i = 0; |
e498be7d CL |
1133 | struct array_cache *ac; |
1134 | unsigned long flags; | |
1135 | ||
1136 | for_each_online_node(i) { | |
4484ebf1 | 1137 | ac = alien[i]; |
e498be7d CL |
1138 | if (ac) { |
1139 | spin_lock_irqsave(&ac->lock, flags); | |
1140 | __drain_alien_cache(cachep, ac, i); | |
1141 | spin_unlock_irqrestore(&ac->lock, flags); | |
1142 | } | |
1143 | } | |
1144 | } | |
729bd0b7 | 1145 | |
873623df | 1146 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1147 | { |
1148 | struct slab *slabp = virt_to_slab(objp); | |
1149 | int nodeid = slabp->nodeid; | |
1150 | struct kmem_list3 *l3; | |
1151 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1152 | int node; |
1153 | ||
1154 | node = numa_node_id(); | |
729bd0b7 PE |
1155 | |
1156 | /* | |
1157 | * Make sure we are not freeing a object from another node to the array | |
1158 | * cache on this cpu. | |
1159 | */ | |
62918a03 | 1160 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1161 | return 0; |
1162 | ||
1ca4cb24 | 1163 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1164 | STATS_INC_NODEFREES(cachep); |
1165 | if (l3->alien && l3->alien[nodeid]) { | |
1166 | alien = l3->alien[nodeid]; | |
873623df | 1167 | spin_lock(&alien->lock); |
729bd0b7 PE |
1168 | if (unlikely(alien->avail == alien->limit)) { |
1169 | STATS_INC_ACOVERFLOW(cachep); | |
1170 | __drain_alien_cache(cachep, alien, nodeid); | |
1171 | } | |
1172 | alien->entry[alien->avail++] = objp; | |
1173 | spin_unlock(&alien->lock); | |
1174 | } else { | |
1175 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1176 | free_block(cachep, &objp, 1, nodeid); | |
1177 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1178 | } | |
1179 | return 1; | |
1180 | } | |
e498be7d CL |
1181 | #endif |
1182 | ||
fbf1e473 AM |
1183 | static void __cpuinit cpuup_canceled(long cpu) |
1184 | { | |
1185 | struct kmem_cache *cachep; | |
1186 | struct kmem_list3 *l3 = NULL; | |
1187 | int node = cpu_to_node(cpu); | |
a70f7302 | 1188 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 AM |
1189 | |
1190 | list_for_each_entry(cachep, &cache_chain, next) { | |
1191 | struct array_cache *nc; | |
1192 | struct array_cache *shared; | |
1193 | struct array_cache **alien; | |
fbf1e473 | 1194 | |
fbf1e473 AM |
1195 | /* cpu is dead; no one can alloc from it. */ |
1196 | nc = cachep->array[cpu]; | |
1197 | cachep->array[cpu] = NULL; | |
1198 | l3 = cachep->nodelists[node]; | |
1199 | ||
1200 | if (!l3) | |
1201 | goto free_array_cache; | |
1202 | ||
1203 | spin_lock_irq(&l3->list_lock); | |
1204 | ||
1205 | /* Free limit for this kmem_list3 */ | |
1206 | l3->free_limit -= cachep->batchcount; | |
1207 | if (nc) | |
1208 | free_block(cachep, nc->entry, nc->avail, node); | |
1209 | ||
c5f59f08 | 1210 | if (!cpus_empty(*mask)) { |
fbf1e473 AM |
1211 | spin_unlock_irq(&l3->list_lock); |
1212 | goto free_array_cache; | |
1213 | } | |
1214 | ||
1215 | shared = l3->shared; | |
1216 | if (shared) { | |
1217 | free_block(cachep, shared->entry, | |
1218 | shared->avail, node); | |
1219 | l3->shared = NULL; | |
1220 | } | |
1221 | ||
1222 | alien = l3->alien; | |
1223 | l3->alien = NULL; | |
1224 | ||
1225 | spin_unlock_irq(&l3->list_lock); | |
1226 | ||
1227 | kfree(shared); | |
1228 | if (alien) { | |
1229 | drain_alien_cache(cachep, alien); | |
1230 | free_alien_cache(alien); | |
1231 | } | |
1232 | free_array_cache: | |
1233 | kfree(nc); | |
1234 | } | |
1235 | /* | |
1236 | * In the previous loop, all the objects were freed to | |
1237 | * the respective cache's slabs, now we can go ahead and | |
1238 | * shrink each nodelist to its limit. | |
1239 | */ | |
1240 | list_for_each_entry(cachep, &cache_chain, next) { | |
1241 | l3 = cachep->nodelists[node]; | |
1242 | if (!l3) | |
1243 | continue; | |
1244 | drain_freelist(cachep, l3, l3->free_objects); | |
1245 | } | |
1246 | } | |
1247 | ||
1248 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1249 | { |
343e0d7a | 1250 | struct kmem_cache *cachep; |
e498be7d CL |
1251 | struct kmem_list3 *l3 = NULL; |
1252 | int node = cpu_to_node(cpu); | |
ea02e3dd | 1253 | const int memsize = sizeof(struct kmem_list3); |
1da177e4 | 1254 | |
fbf1e473 AM |
1255 | /* |
1256 | * We need to do this right in the beginning since | |
1257 | * alloc_arraycache's are going to use this list. | |
1258 | * kmalloc_node allows us to add the slab to the right | |
1259 | * kmem_list3 and not this cpu's kmem_list3 | |
1260 | */ | |
1261 | ||
1262 | list_for_each_entry(cachep, &cache_chain, next) { | |
a737b3e2 | 1263 | /* |
fbf1e473 AM |
1264 | * Set up the size64 kmemlist for cpu before we can |
1265 | * begin anything. Make sure some other cpu on this | |
1266 | * node has not already allocated this | |
e498be7d | 1267 | */ |
fbf1e473 AM |
1268 | if (!cachep->nodelists[node]) { |
1269 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1270 | if (!l3) | |
1271 | goto bad; | |
1272 | kmem_list3_init(l3); | |
1273 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1274 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1275 | |
a737b3e2 | 1276 | /* |
fbf1e473 AM |
1277 | * The l3s don't come and go as CPUs come and |
1278 | * go. cache_chain_mutex is sufficient | |
1279 | * protection here. | |
e498be7d | 1280 | */ |
fbf1e473 | 1281 | cachep->nodelists[node] = l3; |
e498be7d CL |
1282 | } |
1283 | ||
fbf1e473 AM |
1284 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1285 | cachep->nodelists[node]->free_limit = | |
1286 | (1 + nr_cpus_node(node)) * | |
1287 | cachep->batchcount + cachep->num; | |
1288 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1289 | } | |
1290 | ||
1291 | /* | |
1292 | * Now we can go ahead with allocating the shared arrays and | |
1293 | * array caches | |
1294 | */ | |
1295 | list_for_each_entry(cachep, &cache_chain, next) { | |
1296 | struct array_cache *nc; | |
1297 | struct array_cache *shared = NULL; | |
1298 | struct array_cache **alien = NULL; | |
1299 | ||
1300 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1301 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1302 | if (!nc) |
1303 | goto bad; | |
1304 | if (cachep->shared) { | |
1305 | shared = alloc_arraycache(node, | |
1306 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1307 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1308 | if (!shared) { |
1309 | kfree(nc); | |
1da177e4 | 1310 | goto bad; |
12d00f6a | 1311 | } |
fbf1e473 AM |
1312 | } |
1313 | if (use_alien_caches) { | |
83b519e8 | 1314 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1315 | if (!alien) { |
1316 | kfree(shared); | |
1317 | kfree(nc); | |
fbf1e473 | 1318 | goto bad; |
12d00f6a | 1319 | } |
fbf1e473 AM |
1320 | } |
1321 | cachep->array[cpu] = nc; | |
1322 | l3 = cachep->nodelists[node]; | |
1323 | BUG_ON(!l3); | |
1324 | ||
1325 | spin_lock_irq(&l3->list_lock); | |
1326 | if (!l3->shared) { | |
1327 | /* | |
1328 | * We are serialised from CPU_DEAD or | |
1329 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1330 | */ | |
1331 | l3->shared = shared; | |
1332 | shared = NULL; | |
1333 | } | |
4484ebf1 | 1334 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1335 | if (!l3->alien) { |
1336 | l3->alien = alien; | |
1337 | alien = NULL; | |
1da177e4 | 1338 | } |
fbf1e473 AM |
1339 | #endif |
1340 | spin_unlock_irq(&l3->list_lock); | |
1341 | kfree(shared); | |
1342 | free_alien_cache(alien); | |
1343 | } | |
1344 | return 0; | |
1345 | bad: | |
12d00f6a | 1346 | cpuup_canceled(cpu); |
fbf1e473 AM |
1347 | return -ENOMEM; |
1348 | } | |
1349 | ||
1350 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1351 | unsigned long action, void *hcpu) | |
1352 | { | |
1353 | long cpu = (long)hcpu; | |
1354 | int err = 0; | |
1355 | ||
1356 | switch (action) { | |
fbf1e473 AM |
1357 | case CPU_UP_PREPARE: |
1358 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1359 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1360 | err = cpuup_prepare(cpu); |
95402b38 | 1361 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1362 | break; |
1363 | case CPU_ONLINE: | |
8bb78442 | 1364 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1365 | start_cpu_timer(cpu); |
1366 | break; | |
1367 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1368 | case CPU_DOWN_PREPARE: |
8bb78442 | 1369 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1370 | /* |
1371 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1372 | * held so that if cache_reap() is invoked it cannot do | |
1373 | * anything expensive but will only modify reap_work | |
1374 | * and reschedule the timer. | |
1375 | */ | |
1376 | cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); | |
1377 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1378 | per_cpu(reap_work, cpu).work.func = NULL; | |
1379 | break; | |
1380 | case CPU_DOWN_FAILED: | |
8bb78442 | 1381 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1382 | start_cpu_timer(cpu); |
1383 | break; | |
1da177e4 | 1384 | case CPU_DEAD: |
8bb78442 | 1385 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1386 | /* |
1387 | * Even if all the cpus of a node are down, we don't free the | |
1388 | * kmem_list3 of any cache. This to avoid a race between | |
1389 | * cpu_down, and a kmalloc allocation from another cpu for | |
1390 | * memory from the node of the cpu going down. The list3 | |
1391 | * structure is usually allocated from kmem_cache_create() and | |
1392 | * gets destroyed at kmem_cache_destroy(). | |
1393 | */ | |
183ff22b | 1394 | /* fall through */ |
8f5be20b | 1395 | #endif |
1da177e4 | 1396 | case CPU_UP_CANCELED: |
8bb78442 | 1397 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1398 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1399 | cpuup_canceled(cpu); |
fc0abb14 | 1400 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1401 | break; |
1da177e4 | 1402 | } |
fbf1e473 | 1403 | return err ? NOTIFY_BAD : NOTIFY_OK; |
1da177e4 LT |
1404 | } |
1405 | ||
74b85f37 CS |
1406 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1407 | &cpuup_callback, NULL, 0 | |
1408 | }; | |
1da177e4 | 1409 | |
e498be7d CL |
1410 | /* |
1411 | * swap the static kmem_list3 with kmalloced memory | |
1412 | */ | |
a737b3e2 AM |
1413 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1414 | int nodeid) | |
e498be7d CL |
1415 | { |
1416 | struct kmem_list3 *ptr; | |
1417 | ||
83b519e8 | 1418 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1419 | BUG_ON(!ptr); |
1420 | ||
e498be7d | 1421 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1422 | /* |
1423 | * Do not assume that spinlocks can be initialized via memcpy: | |
1424 | */ | |
1425 | spin_lock_init(&ptr->list_lock); | |
1426 | ||
e498be7d CL |
1427 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1428 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1429 | } |
1430 | ||
556a169d PE |
1431 | /* |
1432 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1433 | * size of kmem_list3. | |
1434 | */ | |
1435 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1436 | { | |
1437 | int node; | |
1438 | ||
1439 | for_each_online_node(node) { | |
1440 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1441 | cachep->nodelists[node]->next_reap = jiffies + | |
1442 | REAPTIMEOUT_LIST3 + | |
1443 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1444 | } | |
1445 | } | |
1446 | ||
a737b3e2 AM |
1447 | /* |
1448 | * Initialisation. Called after the page allocator have been initialised and | |
1449 | * before smp_init(). | |
1da177e4 LT |
1450 | */ |
1451 | void __init kmem_cache_init(void) | |
1452 | { | |
1453 | size_t left_over; | |
1454 | struct cache_sizes *sizes; | |
1455 | struct cache_names *names; | |
e498be7d | 1456 | int i; |
07ed76b2 | 1457 | int order; |
1ca4cb24 | 1458 | int node; |
e498be7d | 1459 | |
1807a1aa | 1460 | if (num_possible_nodes() == 1) { |
62918a03 | 1461 | use_alien_caches = 0; |
1807a1aa SS |
1462 | numa_platform = 0; |
1463 | } | |
62918a03 | 1464 | |
e498be7d CL |
1465 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1466 | kmem_list3_init(&initkmem_list3[i]); | |
1467 | if (i < MAX_NUMNODES) | |
1468 | cache_cache.nodelists[i] = NULL; | |
1469 | } | |
556a169d | 1470 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1471 | |
1472 | /* | |
1473 | * Fragmentation resistance on low memory - only use bigger | |
1474 | * page orders on machines with more than 32MB of memory. | |
1475 | */ | |
1476 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1477 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1478 | ||
1da177e4 LT |
1479 | /* Bootstrap is tricky, because several objects are allocated |
1480 | * from caches that do not exist yet: | |
a737b3e2 AM |
1481 | * 1) initialize the cache_cache cache: it contains the struct |
1482 | * kmem_cache structures of all caches, except cache_cache itself: | |
1483 | * cache_cache is statically allocated. | |
e498be7d CL |
1484 | * Initially an __init data area is used for the head array and the |
1485 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1486 | * array at the end of the bootstrap. | |
1da177e4 | 1487 | * 2) Create the first kmalloc cache. |
343e0d7a | 1488 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1489 | * An __init data area is used for the head array. |
1490 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1491 | * head arrays. | |
1da177e4 LT |
1492 | * 4) Replace the __init data head arrays for cache_cache and the first |
1493 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1494 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1495 | * the other cache's with kmalloc allocated memory. | |
1496 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1497 | */ |
1498 | ||
1ca4cb24 PE |
1499 | node = numa_node_id(); |
1500 | ||
1da177e4 | 1501 | /* 1) create the cache_cache */ |
1da177e4 LT |
1502 | INIT_LIST_HEAD(&cache_chain); |
1503 | list_add(&cache_cache.next, &cache_chain); | |
1504 | cache_cache.colour_off = cache_line_size(); | |
1505 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1506 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1507 | |
8da3430d ED |
1508 | /* |
1509 | * struct kmem_cache size depends on nr_node_ids, which | |
1510 | * can be less than MAX_NUMNODES. | |
1511 | */ | |
1512 | cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + | |
1513 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1514 | #if DEBUG | |
1515 | cache_cache.obj_size = cache_cache.buffer_size; | |
1516 | #endif | |
a737b3e2 AM |
1517 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1518 | cache_line_size()); | |
6a2d7a95 ED |
1519 | cache_cache.reciprocal_buffer_size = |
1520 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1521 | |
07ed76b2 JS |
1522 | for (order = 0; order < MAX_ORDER; order++) { |
1523 | cache_estimate(order, cache_cache.buffer_size, | |
1524 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1525 | if (cache_cache.num) | |
1526 | break; | |
1527 | } | |
40094fa6 | 1528 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1529 | cache_cache.gfporder = order; |
b28a02de | 1530 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1531 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1532 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1533 | |
1534 | /* 2+3) create the kmalloc caches */ | |
1535 | sizes = malloc_sizes; | |
1536 | names = cache_names; | |
1537 | ||
a737b3e2 AM |
1538 | /* |
1539 | * Initialize the caches that provide memory for the array cache and the | |
1540 | * kmem_list3 structures first. Without this, further allocations will | |
1541 | * bug. | |
e498be7d CL |
1542 | */ |
1543 | ||
1544 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1545 | sizes[INDEX_AC].cs_size, |
1546 | ARCH_KMALLOC_MINALIGN, | |
1547 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1548 | NULL); |
e498be7d | 1549 | |
a737b3e2 | 1550 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1551 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1552 | kmem_cache_create(names[INDEX_L3].name, |
1553 | sizes[INDEX_L3].cs_size, | |
1554 | ARCH_KMALLOC_MINALIGN, | |
1555 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1556 | NULL); |
a737b3e2 | 1557 | } |
e498be7d | 1558 | |
e0a42726 IM |
1559 | slab_early_init = 0; |
1560 | ||
1da177e4 | 1561 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1562 | /* |
1563 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1564 | * This should be particularly beneficial on SMP boxes, as it |
1565 | * eliminates "false sharing". | |
1566 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1567 | * allow tighter packing of the smaller caches. |
1568 | */ | |
a737b3e2 | 1569 | if (!sizes->cs_cachep) { |
e498be7d | 1570 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1571 | sizes->cs_size, |
1572 | ARCH_KMALLOC_MINALIGN, | |
1573 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1574 | NULL); |
a737b3e2 | 1575 | } |
4b51d669 CL |
1576 | #ifdef CONFIG_ZONE_DMA |
1577 | sizes->cs_dmacachep = kmem_cache_create( | |
1578 | names->name_dma, | |
a737b3e2 AM |
1579 | sizes->cs_size, |
1580 | ARCH_KMALLOC_MINALIGN, | |
1581 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1582 | SLAB_PANIC, | |
20c2df83 | 1583 | NULL); |
4b51d669 | 1584 | #endif |
1da177e4 LT |
1585 | sizes++; |
1586 | names++; | |
1587 | } | |
1588 | /* 4) Replace the bootstrap head arrays */ | |
1589 | { | |
2b2d5493 | 1590 | struct array_cache *ptr; |
e498be7d | 1591 | |
83b519e8 | 1592 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1593 | |
9a2dba4b PE |
1594 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1595 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1596 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1597 | /* |
1598 | * Do not assume that spinlocks can be initialized via memcpy: | |
1599 | */ | |
1600 | spin_lock_init(&ptr->lock); | |
1601 | ||
1da177e4 | 1602 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1603 | |
83b519e8 | 1604 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1605 | |
9a2dba4b | 1606 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1607 | != &initarray_generic.cache); |
9a2dba4b | 1608 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1609 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1610 | /* |
1611 | * Do not assume that spinlocks can be initialized via memcpy: | |
1612 | */ | |
1613 | spin_lock_init(&ptr->lock); | |
1614 | ||
e498be7d | 1615 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1616 | ptr; |
1da177e4 | 1617 | } |
e498be7d CL |
1618 | /* 5) Replace the bootstrap kmem_list3's */ |
1619 | { | |
1ca4cb24 PE |
1620 | int nid; |
1621 | ||
9c09a95c | 1622 | for_each_online_node(nid) { |
ec1f5eee | 1623 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1624 | |
e498be7d | 1625 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1626 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1627 | |
1628 | if (INDEX_AC != INDEX_L3) { | |
1629 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1630 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1631 | } |
1632 | } | |
1633 | } | |
1da177e4 | 1634 | |
8429db5c | 1635 | g_cpucache_up = EARLY; |
1da177e4 | 1636 | |
056c6241 RT |
1637 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1638 | init_lock_keys(); | |
8429db5c PE |
1639 | } |
1640 | ||
1641 | void __init kmem_cache_init_late(void) | |
1642 | { | |
1643 | struct kmem_cache *cachep; | |
1644 | ||
1645 | /* | |
1646 | * Interrupts are enabled now so all GFP allocations are safe. | |
1647 | */ | |
1648 | slab_gfp_mask = __GFP_BITS_MASK; | |
056c6241 | 1649 | |
8429db5c PE |
1650 | /* 6) resize the head arrays to their final sizes */ |
1651 | mutex_lock(&cache_chain_mutex); | |
1652 | list_for_each_entry(cachep, &cache_chain, next) | |
1653 | if (enable_cpucache(cachep, GFP_NOWAIT)) | |
1654 | BUG(); | |
1655 | mutex_unlock(&cache_chain_mutex); | |
056c6241 | 1656 | |
1da177e4 LT |
1657 | /* Done! */ |
1658 | g_cpucache_up = FULL; | |
1659 | ||
a737b3e2 AM |
1660 | /* |
1661 | * Register a cpu startup notifier callback that initializes | |
1662 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1663 | */ |
1664 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1665 | |
a737b3e2 AM |
1666 | /* |
1667 | * The reap timers are started later, with a module init call: That part | |
1668 | * of the kernel is not yet operational. | |
1da177e4 LT |
1669 | */ |
1670 | } | |
1671 | ||
1672 | static int __init cpucache_init(void) | |
1673 | { | |
1674 | int cpu; | |
1675 | ||
a737b3e2 AM |
1676 | /* |
1677 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1678 | */ |
e498be7d | 1679 | for_each_online_cpu(cpu) |
a737b3e2 | 1680 | start_cpu_timer(cpu); |
1da177e4 LT |
1681 | return 0; |
1682 | } | |
1da177e4 LT |
1683 | __initcall(cpucache_init); |
1684 | ||
1685 | /* | |
1686 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1687 | * | |
1688 | * If we requested dmaable memory, we will get it. Even if we | |
1689 | * did not request dmaable memory, we might get it, but that | |
1690 | * would be relatively rare and ignorable. | |
1691 | */ | |
343e0d7a | 1692 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1693 | { |
1694 | struct page *page; | |
e1b6aa6f | 1695 | int nr_pages; |
1da177e4 LT |
1696 | int i; |
1697 | ||
d6fef9da | 1698 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1699 | /* |
1700 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1701 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1702 | */ |
e1b6aa6f | 1703 | flags |= __GFP_COMP; |
d6fef9da | 1704 | #endif |
765c4507 | 1705 | |
3c517a61 | 1706 | flags |= cachep->gfpflags; |
e12ba74d MG |
1707 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1708 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1709 | |
6484eb3e | 1710 | page = alloc_pages_exact_node(nodeid, flags, cachep->gfporder); |
1da177e4 LT |
1711 | if (!page) |
1712 | return NULL; | |
1da177e4 | 1713 | |
e1b6aa6f | 1714 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1715 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1716 | add_zone_page_state(page_zone(page), |
1717 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1718 | else | |
1719 | add_zone_page_state(page_zone(page), | |
1720 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1721 | for (i = 0; i < nr_pages; i++) |
1722 | __SetPageSlab(page + i); | |
1723 | return page_address(page); | |
1da177e4 LT |
1724 | } |
1725 | ||
1726 | /* | |
1727 | * Interface to system's page release. | |
1728 | */ | |
343e0d7a | 1729 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1730 | { |
b28a02de | 1731 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1732 | struct page *page = virt_to_page(addr); |
1733 | const unsigned long nr_freed = i; | |
1734 | ||
972d1a7b CL |
1735 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1736 | sub_zone_page_state(page_zone(page), | |
1737 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1738 | else | |
1739 | sub_zone_page_state(page_zone(page), | |
1740 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1741 | while (i--) { |
f205b2fe NP |
1742 | BUG_ON(!PageSlab(page)); |
1743 | __ClearPageSlab(page); | |
1da177e4 LT |
1744 | page++; |
1745 | } | |
1da177e4 LT |
1746 | if (current->reclaim_state) |
1747 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1748 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1749 | } |
1750 | ||
1751 | static void kmem_rcu_free(struct rcu_head *head) | |
1752 | { | |
b28a02de | 1753 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1754 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1755 | |
1756 | kmem_freepages(cachep, slab_rcu->addr); | |
1757 | if (OFF_SLAB(cachep)) | |
1758 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1759 | } | |
1760 | ||
1761 | #if DEBUG | |
1762 | ||
1763 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1764 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1765 | unsigned long caller) |
1da177e4 | 1766 | { |
3dafccf2 | 1767 | int size = obj_size(cachep); |
1da177e4 | 1768 | |
3dafccf2 | 1769 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1770 | |
b28a02de | 1771 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1772 | return; |
1773 | ||
b28a02de PE |
1774 | *addr++ = 0x12345678; |
1775 | *addr++ = caller; | |
1776 | *addr++ = smp_processor_id(); | |
1777 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1778 | { |
1779 | unsigned long *sptr = &caller; | |
1780 | unsigned long svalue; | |
1781 | ||
1782 | while (!kstack_end(sptr)) { | |
1783 | svalue = *sptr++; | |
1784 | if (kernel_text_address(svalue)) { | |
b28a02de | 1785 | *addr++ = svalue; |
1da177e4 LT |
1786 | size -= sizeof(unsigned long); |
1787 | if (size <= sizeof(unsigned long)) | |
1788 | break; | |
1789 | } | |
1790 | } | |
1791 | ||
1792 | } | |
b28a02de | 1793 | *addr++ = 0x87654321; |
1da177e4 LT |
1794 | } |
1795 | #endif | |
1796 | ||
343e0d7a | 1797 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1798 | { |
3dafccf2 MS |
1799 | int size = obj_size(cachep); |
1800 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1801 | |
1802 | memset(addr, val, size); | |
b28a02de | 1803 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1804 | } |
1805 | ||
1806 | static void dump_line(char *data, int offset, int limit) | |
1807 | { | |
1808 | int i; | |
aa83aa40 DJ |
1809 | unsigned char error = 0; |
1810 | int bad_count = 0; | |
1811 | ||
1da177e4 | 1812 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1813 | for (i = 0; i < limit; i++) { |
1814 | if (data[offset + i] != POISON_FREE) { | |
1815 | error = data[offset + i]; | |
1816 | bad_count++; | |
1817 | } | |
b28a02de | 1818 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1819 | } |
1da177e4 | 1820 | printk("\n"); |
aa83aa40 DJ |
1821 | |
1822 | if (bad_count == 1) { | |
1823 | error ^= POISON_FREE; | |
1824 | if (!(error & (error - 1))) { | |
1825 | printk(KERN_ERR "Single bit error detected. Probably " | |
1826 | "bad RAM.\n"); | |
1827 | #ifdef CONFIG_X86 | |
1828 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1829 | "test tool.\n"); | |
1830 | #else | |
1831 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1832 | #endif | |
1833 | } | |
1834 | } | |
1da177e4 LT |
1835 | } |
1836 | #endif | |
1837 | ||
1838 | #if DEBUG | |
1839 | ||
343e0d7a | 1840 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1841 | { |
1842 | int i, size; | |
1843 | char *realobj; | |
1844 | ||
1845 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1846 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1847 | *dbg_redzone1(cachep, objp), |
1848 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1849 | } |
1850 | ||
1851 | if (cachep->flags & SLAB_STORE_USER) { | |
1852 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1853 | *dbg_userword(cachep, objp)); |
1da177e4 | 1854 | print_symbol("(%s)", |
a737b3e2 | 1855 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1856 | printk("\n"); |
1857 | } | |
3dafccf2 MS |
1858 | realobj = (char *)objp + obj_offset(cachep); |
1859 | size = obj_size(cachep); | |
b28a02de | 1860 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1861 | int limit; |
1862 | limit = 16; | |
b28a02de PE |
1863 | if (i + limit > size) |
1864 | limit = size - i; | |
1da177e4 LT |
1865 | dump_line(realobj, i, limit); |
1866 | } | |
1867 | } | |
1868 | ||
343e0d7a | 1869 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1870 | { |
1871 | char *realobj; | |
1872 | int size, i; | |
1873 | int lines = 0; | |
1874 | ||
3dafccf2 MS |
1875 | realobj = (char *)objp + obj_offset(cachep); |
1876 | size = obj_size(cachep); | |
1da177e4 | 1877 | |
b28a02de | 1878 | for (i = 0; i < size; i++) { |
1da177e4 | 1879 | char exp = POISON_FREE; |
b28a02de | 1880 | if (i == size - 1) |
1da177e4 LT |
1881 | exp = POISON_END; |
1882 | if (realobj[i] != exp) { | |
1883 | int limit; | |
1884 | /* Mismatch ! */ | |
1885 | /* Print header */ | |
1886 | if (lines == 0) { | |
b28a02de | 1887 | printk(KERN_ERR |
e94a40c5 DH |
1888 | "Slab corruption: %s start=%p, len=%d\n", |
1889 | cachep->name, realobj, size); | |
1da177e4 LT |
1890 | print_objinfo(cachep, objp, 0); |
1891 | } | |
1892 | /* Hexdump the affected line */ | |
b28a02de | 1893 | i = (i / 16) * 16; |
1da177e4 | 1894 | limit = 16; |
b28a02de PE |
1895 | if (i + limit > size) |
1896 | limit = size - i; | |
1da177e4 LT |
1897 | dump_line(realobj, i, limit); |
1898 | i += 16; | |
1899 | lines++; | |
1900 | /* Limit to 5 lines */ | |
1901 | if (lines > 5) | |
1902 | break; | |
1903 | } | |
1904 | } | |
1905 | if (lines != 0) { | |
1906 | /* Print some data about the neighboring objects, if they | |
1907 | * exist: | |
1908 | */ | |
6ed5eb22 | 1909 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1910 | unsigned int objnr; |
1da177e4 | 1911 | |
8fea4e96 | 1912 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1913 | if (objnr) { |
8fea4e96 | 1914 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1915 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1916 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1917 | realobj, size); |
1da177e4 LT |
1918 | print_objinfo(cachep, objp, 2); |
1919 | } | |
b28a02de | 1920 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1921 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1922 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1923 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1924 | realobj, size); |
1da177e4 LT |
1925 | print_objinfo(cachep, objp, 2); |
1926 | } | |
1927 | } | |
1928 | } | |
1929 | #endif | |
1930 | ||
12dd36fa | 1931 | #if DEBUG |
e79aec29 | 1932 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1933 | { |
1da177e4 LT |
1934 | int i; |
1935 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1936 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1937 | |
1938 | if (cachep->flags & SLAB_POISON) { | |
1939 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1940 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1941 | OFF_SLAB(cachep)) | |
b28a02de | 1942 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1943 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1944 | else |
1945 | check_poison_obj(cachep, objp); | |
1946 | #else | |
1947 | check_poison_obj(cachep, objp); | |
1948 | #endif | |
1949 | } | |
1950 | if (cachep->flags & SLAB_RED_ZONE) { | |
1951 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1952 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1953 | "was overwritten"); |
1da177e4 LT |
1954 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1955 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1956 | "was overwritten"); |
1da177e4 | 1957 | } |
1da177e4 | 1958 | } |
12dd36fa | 1959 | } |
1da177e4 | 1960 | #else |
e79aec29 | 1961 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1962 | { |
12dd36fa | 1963 | } |
1da177e4 LT |
1964 | #endif |
1965 | ||
911851e6 RD |
1966 | /** |
1967 | * slab_destroy - destroy and release all objects in a slab | |
1968 | * @cachep: cache pointer being destroyed | |
1969 | * @slabp: slab pointer being destroyed | |
1970 | * | |
12dd36fa | 1971 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1972 | * Before calling the slab must have been unlinked from the cache. The |
1973 | * cache-lock is not held/needed. | |
12dd36fa | 1974 | */ |
343e0d7a | 1975 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1976 | { |
1977 | void *addr = slabp->s_mem - slabp->colouroff; | |
1978 | ||
e79aec29 | 1979 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
1980 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1981 | struct slab_rcu *slab_rcu; | |
1982 | ||
b28a02de | 1983 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1984 | slab_rcu->cachep = cachep; |
1985 | slab_rcu->addr = addr; | |
1986 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1987 | } else { | |
1988 | kmem_freepages(cachep, addr); | |
873623df IM |
1989 | if (OFF_SLAB(cachep)) |
1990 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1991 | } |
1992 | } | |
1993 | ||
117f6eb1 CL |
1994 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1995 | { | |
1996 | int i; | |
1997 | struct kmem_list3 *l3; | |
1998 | ||
1999 | for_each_online_cpu(i) | |
2000 | kfree(cachep->array[i]); | |
2001 | ||
2002 | /* NUMA: free the list3 structures */ | |
2003 | for_each_online_node(i) { | |
2004 | l3 = cachep->nodelists[i]; | |
2005 | if (l3) { | |
2006 | kfree(l3->shared); | |
2007 | free_alien_cache(l3->alien); | |
2008 | kfree(l3); | |
2009 | } | |
2010 | } | |
2011 | kmem_cache_free(&cache_cache, cachep); | |
2012 | } | |
2013 | ||
2014 | ||
4d268eba | 2015 | /** |
a70773dd RD |
2016 | * calculate_slab_order - calculate size (page order) of slabs |
2017 | * @cachep: pointer to the cache that is being created | |
2018 | * @size: size of objects to be created in this cache. | |
2019 | * @align: required alignment for the objects. | |
2020 | * @flags: slab allocation flags | |
2021 | * | |
2022 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2023 | * |
2024 | * This could be made much more intelligent. For now, try to avoid using | |
2025 | * high order pages for slabs. When the gfp() functions are more friendly | |
2026 | * towards high-order requests, this should be changed. | |
2027 | */ | |
a737b3e2 | 2028 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2029 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2030 | { |
b1ab41c4 | 2031 | unsigned long offslab_limit; |
4d268eba | 2032 | size_t left_over = 0; |
9888e6fa | 2033 | int gfporder; |
4d268eba | 2034 | |
0aa817f0 | 2035 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2036 | unsigned int num; |
2037 | size_t remainder; | |
2038 | ||
9888e6fa | 2039 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2040 | if (!num) |
2041 | continue; | |
9888e6fa | 2042 | |
b1ab41c4 IM |
2043 | if (flags & CFLGS_OFF_SLAB) { |
2044 | /* | |
2045 | * Max number of objs-per-slab for caches which | |
2046 | * use off-slab slabs. Needed to avoid a possible | |
2047 | * looping condition in cache_grow(). | |
2048 | */ | |
2049 | offslab_limit = size - sizeof(struct slab); | |
2050 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2051 | ||
2052 | if (num > offslab_limit) | |
2053 | break; | |
2054 | } | |
4d268eba | 2055 | |
9888e6fa | 2056 | /* Found something acceptable - save it away */ |
4d268eba | 2057 | cachep->num = num; |
9888e6fa | 2058 | cachep->gfporder = gfporder; |
4d268eba PE |
2059 | left_over = remainder; |
2060 | ||
f78bb8ad LT |
2061 | /* |
2062 | * A VFS-reclaimable slab tends to have most allocations | |
2063 | * as GFP_NOFS and we really don't want to have to be allocating | |
2064 | * higher-order pages when we are unable to shrink dcache. | |
2065 | */ | |
2066 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2067 | break; | |
2068 | ||
4d268eba PE |
2069 | /* |
2070 | * Large number of objects is good, but very large slabs are | |
2071 | * currently bad for the gfp()s. | |
2072 | */ | |
9888e6fa | 2073 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2074 | break; |
2075 | ||
9888e6fa LT |
2076 | /* |
2077 | * Acceptable internal fragmentation? | |
2078 | */ | |
a737b3e2 | 2079 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2080 | break; |
2081 | } | |
2082 | return left_over; | |
2083 | } | |
2084 | ||
83b519e8 | 2085 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2086 | { |
2ed3a4ef | 2087 | if (g_cpucache_up == FULL) |
83b519e8 | 2088 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2089 | |
f30cf7d1 PE |
2090 | if (g_cpucache_up == NONE) { |
2091 | /* | |
2092 | * Note: the first kmem_cache_create must create the cache | |
2093 | * that's used by kmalloc(24), otherwise the creation of | |
2094 | * further caches will BUG(). | |
2095 | */ | |
2096 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2097 | ||
2098 | /* | |
2099 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2100 | * the first cache, then we need to set up all its list3s, | |
2101 | * otherwise the creation of further caches will BUG(). | |
2102 | */ | |
2103 | set_up_list3s(cachep, SIZE_AC); | |
2104 | if (INDEX_AC == INDEX_L3) | |
2105 | g_cpucache_up = PARTIAL_L3; | |
2106 | else | |
2107 | g_cpucache_up = PARTIAL_AC; | |
2108 | } else { | |
2109 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2110 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 PE |
2111 | |
2112 | if (g_cpucache_up == PARTIAL_AC) { | |
2113 | set_up_list3s(cachep, SIZE_L3); | |
2114 | g_cpucache_up = PARTIAL_L3; | |
2115 | } else { | |
2116 | int node; | |
556a169d | 2117 | for_each_online_node(node) { |
f30cf7d1 PE |
2118 | cachep->nodelists[node] = |
2119 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2120 | gfp, node); |
f30cf7d1 PE |
2121 | BUG_ON(!cachep->nodelists[node]); |
2122 | kmem_list3_init(cachep->nodelists[node]); | |
2123 | } | |
2124 | } | |
2125 | } | |
2126 | cachep->nodelists[numa_node_id()]->next_reap = | |
2127 | jiffies + REAPTIMEOUT_LIST3 + | |
2128 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2129 | ||
2130 | cpu_cache_get(cachep)->avail = 0; | |
2131 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2132 | cpu_cache_get(cachep)->batchcount = 1; | |
2133 | cpu_cache_get(cachep)->touched = 0; | |
2134 | cachep->batchcount = 1; | |
2135 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2136 | return 0; |
f30cf7d1 PE |
2137 | } |
2138 | ||
1da177e4 LT |
2139 | /** |
2140 | * kmem_cache_create - Create a cache. | |
2141 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2142 | * @size: The size of objects to be created in this cache. | |
2143 | * @align: The required alignment for the objects. | |
2144 | * @flags: SLAB flags | |
2145 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2146 | * |
2147 | * Returns a ptr to the cache on success, NULL on failure. | |
2148 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2149 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2150 | * |
2151 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 | 2152 | * the module calling this has to destroy the cache before getting unloaded. |
249da166 CM |
2153 | * Note that kmem_cache_name() is not guaranteed to return the same pointer, |
2154 | * therefore applications must manage it themselves. | |
a737b3e2 | 2155 | * |
1da177e4 LT |
2156 | * The flags are |
2157 | * | |
2158 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2159 | * to catch references to uninitialised memory. | |
2160 | * | |
2161 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2162 | * for buffer overruns. | |
2163 | * | |
1da177e4 LT |
2164 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2165 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2166 | * as davem. | |
2167 | */ | |
343e0d7a | 2168 | struct kmem_cache * |
1da177e4 | 2169 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2170 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2171 | { |
2172 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2173 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2174 | gfp_t gfp; |
1da177e4 LT |
2175 | |
2176 | /* | |
2177 | * Sanity checks... these are all serious usage bugs. | |
2178 | */ | |
a737b3e2 | 2179 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2180 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2181 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2182 | name); |
b28a02de PE |
2183 | BUG(); |
2184 | } | |
1da177e4 | 2185 | |
f0188f47 | 2186 | /* |
8f5be20b | 2187 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2188 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2189 | */ |
83b519e8 PE |
2190 | if (slab_is_available()) { |
2191 | get_online_cpus(); | |
2192 | mutex_lock(&cache_chain_mutex); | |
2193 | } | |
4f12bb4f | 2194 | |
7a7c381d | 2195 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2196 | char tmp; |
2197 | int res; | |
2198 | ||
2199 | /* | |
2200 | * This happens when the module gets unloaded and doesn't | |
2201 | * destroy its slab cache and no-one else reuses the vmalloc | |
2202 | * area of the module. Print a warning. | |
2203 | */ | |
138ae663 | 2204 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2205 | if (res) { |
b4169525 | 2206 | printk(KERN_ERR |
2207 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2208 | pc->buffer_size); |
4f12bb4f AM |
2209 | continue; |
2210 | } | |
2211 | ||
b28a02de | 2212 | if (!strcmp(pc->name, name)) { |
b4169525 | 2213 | printk(KERN_ERR |
2214 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2215 | dump_stack(); |
2216 | goto oops; | |
2217 | } | |
2218 | } | |
2219 | ||
1da177e4 LT |
2220 | #if DEBUG |
2221 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2222 | #if FORCED_DEBUG |
2223 | /* | |
2224 | * Enable redzoning and last user accounting, except for caches with | |
2225 | * large objects, if the increased size would increase the object size | |
2226 | * above the next power of two: caches with object sizes just above a | |
2227 | * power of two have a significant amount of internal fragmentation. | |
2228 | */ | |
87a927c7 DW |
2229 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2230 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2231 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2232 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2233 | flags |= SLAB_POISON; | |
2234 | #endif | |
2235 | if (flags & SLAB_DESTROY_BY_RCU) | |
2236 | BUG_ON(flags & SLAB_POISON); | |
2237 | #endif | |
1da177e4 | 2238 | /* |
a737b3e2 AM |
2239 | * Always checks flags, a caller might be expecting debug support which |
2240 | * isn't available. | |
1da177e4 | 2241 | */ |
40094fa6 | 2242 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2243 | |
a737b3e2 AM |
2244 | /* |
2245 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2246 | * unaligned accesses for some archs when redzoning is used, and makes |
2247 | * sure any on-slab bufctl's are also correctly aligned. | |
2248 | */ | |
b28a02de PE |
2249 | if (size & (BYTES_PER_WORD - 1)) { |
2250 | size += (BYTES_PER_WORD - 1); | |
2251 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2252 | } |
2253 | ||
a737b3e2 AM |
2254 | /* calculate the final buffer alignment: */ |
2255 | ||
1da177e4 LT |
2256 | /* 1) arch recommendation: can be overridden for debug */ |
2257 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2258 | /* |
2259 | * Default alignment: as specified by the arch code. Except if | |
2260 | * an object is really small, then squeeze multiple objects into | |
2261 | * one cacheline. | |
1da177e4 LT |
2262 | */ |
2263 | ralign = cache_line_size(); | |
b28a02de | 2264 | while (size <= ralign / 2) |
1da177e4 LT |
2265 | ralign /= 2; |
2266 | } else { | |
2267 | ralign = BYTES_PER_WORD; | |
2268 | } | |
ca5f9703 PE |
2269 | |
2270 | /* | |
87a927c7 DW |
2271 | * Redzoning and user store require word alignment or possibly larger. |
2272 | * Note this will be overridden by architecture or caller mandated | |
2273 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2274 | */ |
87a927c7 DW |
2275 | if (flags & SLAB_STORE_USER) |
2276 | ralign = BYTES_PER_WORD; | |
2277 | ||
2278 | if (flags & SLAB_RED_ZONE) { | |
2279 | ralign = REDZONE_ALIGN; | |
2280 | /* If redzoning, ensure that the second redzone is suitably | |
2281 | * aligned, by adjusting the object size accordingly. */ | |
2282 | size += REDZONE_ALIGN - 1; | |
2283 | size &= ~(REDZONE_ALIGN - 1); | |
2284 | } | |
ca5f9703 | 2285 | |
a44b56d3 | 2286 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2287 | if (ralign < ARCH_SLAB_MINALIGN) { |
2288 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2289 | } |
a44b56d3 | 2290 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2291 | if (ralign < align) { |
2292 | ralign = align; | |
1da177e4 | 2293 | } |
a44b56d3 | 2294 | /* disable debug if necessary */ |
b46b8f19 | 2295 | if (ralign > __alignof__(unsigned long long)) |
a44b56d3 | 2296 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2297 | /* |
ca5f9703 | 2298 | * 4) Store it. |
1da177e4 LT |
2299 | */ |
2300 | align = ralign; | |
2301 | ||
83b519e8 PE |
2302 | if (slab_is_available()) |
2303 | gfp = GFP_KERNEL; | |
2304 | else | |
2305 | gfp = GFP_NOWAIT; | |
2306 | ||
1da177e4 | 2307 | /* Get cache's description obj. */ |
83b519e8 | 2308 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2309 | if (!cachep) |
4f12bb4f | 2310 | goto oops; |
1da177e4 LT |
2311 | |
2312 | #if DEBUG | |
3dafccf2 | 2313 | cachep->obj_size = size; |
1da177e4 | 2314 | |
ca5f9703 PE |
2315 | /* |
2316 | * Both debugging options require word-alignment which is calculated | |
2317 | * into align above. | |
2318 | */ | |
1da177e4 | 2319 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2320 | /* add space for red zone words */ |
b46b8f19 DW |
2321 | cachep->obj_offset += sizeof(unsigned long long); |
2322 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2323 | } |
2324 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2325 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2326 | * the real object. But if the second red zone needs to be |
2327 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2328 | */ |
87a927c7 DW |
2329 | if (flags & SLAB_RED_ZONE) |
2330 | size += REDZONE_ALIGN; | |
2331 | else | |
2332 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2333 | } |
2334 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2335 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2336 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2337 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2338 | size = PAGE_SIZE; |
2339 | } | |
2340 | #endif | |
2341 | #endif | |
2342 | ||
e0a42726 IM |
2343 | /* |
2344 | * Determine if the slab management is 'on' or 'off' slab. | |
2345 | * (bootstrapping cannot cope with offslab caches so don't do | |
2346 | * it too early on.) | |
2347 | */ | |
2348 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2349 | /* |
2350 | * Size is large, assume best to place the slab management obj | |
2351 | * off-slab (should allow better packing of objs). | |
2352 | */ | |
2353 | flags |= CFLGS_OFF_SLAB; | |
2354 | ||
2355 | size = ALIGN(size, align); | |
2356 | ||
f78bb8ad | 2357 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2358 | |
2359 | if (!cachep->num) { | |
b4169525 | 2360 | printk(KERN_ERR |
2361 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2362 | kmem_cache_free(&cache_cache, cachep); |
2363 | cachep = NULL; | |
4f12bb4f | 2364 | goto oops; |
1da177e4 | 2365 | } |
b28a02de PE |
2366 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2367 | + sizeof(struct slab), align); | |
1da177e4 LT |
2368 | |
2369 | /* | |
2370 | * If the slab has been placed off-slab, and we have enough space then | |
2371 | * move it on-slab. This is at the expense of any extra colouring. | |
2372 | */ | |
2373 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2374 | flags &= ~CFLGS_OFF_SLAB; | |
2375 | left_over -= slab_size; | |
2376 | } | |
2377 | ||
2378 | if (flags & CFLGS_OFF_SLAB) { | |
2379 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2380 | slab_size = |
2381 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2382 | } |
2383 | ||
2384 | cachep->colour_off = cache_line_size(); | |
2385 | /* Offset must be a multiple of the alignment. */ | |
2386 | if (cachep->colour_off < align) | |
2387 | cachep->colour_off = align; | |
b28a02de | 2388 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2389 | cachep->slab_size = slab_size; |
2390 | cachep->flags = flags; | |
2391 | cachep->gfpflags = 0; | |
4b51d669 | 2392 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2393 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2394 | cachep->buffer_size = size; |
6a2d7a95 | 2395 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2396 | |
e5ac9c5a | 2397 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2398 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2399 | /* |
2400 | * This is a possibility for one of the malloc_sizes caches. | |
2401 | * But since we go off slab only for object size greater than | |
2402 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2403 | * this should not happen at all. | |
2404 | * But leave a BUG_ON for some lucky dude. | |
2405 | */ | |
6cb8f913 | 2406 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2407 | } |
1da177e4 | 2408 | cachep->ctor = ctor; |
1da177e4 LT |
2409 | cachep->name = name; |
2410 | ||
83b519e8 | 2411 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef CL |
2412 | __kmem_cache_destroy(cachep); |
2413 | cachep = NULL; | |
2414 | goto oops; | |
2415 | } | |
1da177e4 | 2416 | |
1da177e4 LT |
2417 | /* cache setup completed, link it into the list */ |
2418 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2419 | oops: |
1da177e4 LT |
2420 | if (!cachep && (flags & SLAB_PANIC)) |
2421 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2422 | name); |
83b519e8 PE |
2423 | if (slab_is_available()) { |
2424 | mutex_unlock(&cache_chain_mutex); | |
2425 | put_online_cpus(); | |
2426 | } | |
1da177e4 LT |
2427 | return cachep; |
2428 | } | |
2429 | EXPORT_SYMBOL(kmem_cache_create); | |
2430 | ||
2431 | #if DEBUG | |
2432 | static void check_irq_off(void) | |
2433 | { | |
2434 | BUG_ON(!irqs_disabled()); | |
2435 | } | |
2436 | ||
2437 | static void check_irq_on(void) | |
2438 | { | |
2439 | BUG_ON(irqs_disabled()); | |
2440 | } | |
2441 | ||
343e0d7a | 2442 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2443 | { |
2444 | #ifdef CONFIG_SMP | |
2445 | check_irq_off(); | |
e498be7d | 2446 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2447 | #endif |
2448 | } | |
e498be7d | 2449 | |
343e0d7a | 2450 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2451 | { |
2452 | #ifdef CONFIG_SMP | |
2453 | check_irq_off(); | |
2454 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2455 | #endif | |
2456 | } | |
2457 | ||
1da177e4 LT |
2458 | #else |
2459 | #define check_irq_off() do { } while(0) | |
2460 | #define check_irq_on() do { } while(0) | |
2461 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2462 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2463 | #endif |
2464 | ||
aab2207c CL |
2465 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2466 | struct array_cache *ac, | |
2467 | int force, int node); | |
2468 | ||
1da177e4 LT |
2469 | static void do_drain(void *arg) |
2470 | { | |
a737b3e2 | 2471 | struct kmem_cache *cachep = arg; |
1da177e4 | 2472 | struct array_cache *ac; |
ff69416e | 2473 | int node = numa_node_id(); |
1da177e4 LT |
2474 | |
2475 | check_irq_off(); | |
9a2dba4b | 2476 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2477 | spin_lock(&cachep->nodelists[node]->list_lock); |
2478 | free_block(cachep, ac->entry, ac->avail, node); | |
2479 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2480 | ac->avail = 0; |
2481 | } | |
2482 | ||
343e0d7a | 2483 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2484 | { |
e498be7d CL |
2485 | struct kmem_list3 *l3; |
2486 | int node; | |
2487 | ||
15c8b6c1 | 2488 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2489 | check_irq_on(); |
b28a02de | 2490 | for_each_online_node(node) { |
e498be7d | 2491 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2492 | if (l3 && l3->alien) |
2493 | drain_alien_cache(cachep, l3->alien); | |
2494 | } | |
2495 | ||
2496 | for_each_online_node(node) { | |
2497 | l3 = cachep->nodelists[node]; | |
2498 | if (l3) | |
aab2207c | 2499 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2500 | } |
1da177e4 LT |
2501 | } |
2502 | ||
ed11d9eb CL |
2503 | /* |
2504 | * Remove slabs from the list of free slabs. | |
2505 | * Specify the number of slabs to drain in tofree. | |
2506 | * | |
2507 | * Returns the actual number of slabs released. | |
2508 | */ | |
2509 | static int drain_freelist(struct kmem_cache *cache, | |
2510 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2511 | { |
ed11d9eb CL |
2512 | struct list_head *p; |
2513 | int nr_freed; | |
1da177e4 | 2514 | struct slab *slabp; |
1da177e4 | 2515 | |
ed11d9eb CL |
2516 | nr_freed = 0; |
2517 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2518 | |
ed11d9eb | 2519 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2520 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2521 | if (p == &l3->slabs_free) { |
2522 | spin_unlock_irq(&l3->list_lock); | |
2523 | goto out; | |
2524 | } | |
1da177e4 | 2525 | |
ed11d9eb | 2526 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2527 | #if DEBUG |
40094fa6 | 2528 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2529 | #endif |
2530 | list_del(&slabp->list); | |
ed11d9eb CL |
2531 | /* |
2532 | * Safe to drop the lock. The slab is no longer linked | |
2533 | * to the cache. | |
2534 | */ | |
2535 | l3->free_objects -= cache->num; | |
e498be7d | 2536 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2537 | slab_destroy(cache, slabp); |
2538 | nr_freed++; | |
1da177e4 | 2539 | } |
ed11d9eb CL |
2540 | out: |
2541 | return nr_freed; | |
1da177e4 LT |
2542 | } |
2543 | ||
8f5be20b | 2544 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2545 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2546 | { |
2547 | int ret = 0, i = 0; | |
2548 | struct kmem_list3 *l3; | |
2549 | ||
2550 | drain_cpu_caches(cachep); | |
2551 | ||
2552 | check_irq_on(); | |
2553 | for_each_online_node(i) { | |
2554 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2555 | if (!l3) |
2556 | continue; | |
2557 | ||
2558 | drain_freelist(cachep, l3, l3->free_objects); | |
2559 | ||
2560 | ret += !list_empty(&l3->slabs_full) || | |
2561 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2562 | } |
2563 | return (ret ? 1 : 0); | |
2564 | } | |
2565 | ||
1da177e4 LT |
2566 | /** |
2567 | * kmem_cache_shrink - Shrink a cache. | |
2568 | * @cachep: The cache to shrink. | |
2569 | * | |
2570 | * Releases as many slabs as possible for a cache. | |
2571 | * To help debugging, a zero exit status indicates all slabs were released. | |
2572 | */ | |
343e0d7a | 2573 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2574 | { |
8f5be20b | 2575 | int ret; |
40094fa6 | 2576 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2577 | |
95402b38 | 2578 | get_online_cpus(); |
8f5be20b RT |
2579 | mutex_lock(&cache_chain_mutex); |
2580 | ret = __cache_shrink(cachep); | |
2581 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2582 | put_online_cpus(); |
8f5be20b | 2583 | return ret; |
1da177e4 LT |
2584 | } |
2585 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2586 | ||
2587 | /** | |
2588 | * kmem_cache_destroy - delete a cache | |
2589 | * @cachep: the cache to destroy | |
2590 | * | |
72fd4a35 | 2591 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2592 | * |
2593 | * It is expected this function will be called by a module when it is | |
2594 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2595 | * cache being allocated each time a module is loaded and unloaded, if the | |
2596 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2597 | * | |
2598 | * The cache must be empty before calling this function. | |
2599 | * | |
2600 | * The caller must guarantee that noone will allocate memory from the cache | |
2601 | * during the kmem_cache_destroy(). | |
2602 | */ | |
133d205a | 2603 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2604 | { |
40094fa6 | 2605 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2606 | |
1da177e4 | 2607 | /* Find the cache in the chain of caches. */ |
95402b38 | 2608 | get_online_cpus(); |
fc0abb14 | 2609 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2610 | /* |
2611 | * the chain is never empty, cache_cache is never destroyed | |
2612 | */ | |
2613 | list_del(&cachep->next); | |
1da177e4 LT |
2614 | if (__cache_shrink(cachep)) { |
2615 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2616 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2617 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2618 | put_online_cpus(); |
133d205a | 2619 | return; |
1da177e4 LT |
2620 | } |
2621 | ||
2622 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2623 | synchronize_rcu(); |
1da177e4 | 2624 | |
117f6eb1 | 2625 | __kmem_cache_destroy(cachep); |
8f5be20b | 2626 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2627 | put_online_cpus(); |
1da177e4 LT |
2628 | } |
2629 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2630 | ||
e5ac9c5a RT |
2631 | /* |
2632 | * Get the memory for a slab management obj. | |
2633 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2634 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2635 | * come from the same cache which is getting created because, | |
2636 | * when we are searching for an appropriate cache for these | |
2637 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2638 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2639 | * kmem_find_general_cachep till the initialization is complete. | |
2640 | * Hence we cannot have slabp_cache same as the original cache. | |
2641 | */ | |
343e0d7a | 2642 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2643 | int colour_off, gfp_t local_flags, |
2644 | int nodeid) | |
1da177e4 LT |
2645 | { |
2646 | struct slab *slabp; | |
b28a02de | 2647 | |
1da177e4 LT |
2648 | if (OFF_SLAB(cachep)) { |
2649 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2650 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2651 | local_flags, nodeid); |
d5cff635 CM |
2652 | /* |
2653 | * If the first object in the slab is leaked (it's allocated | |
2654 | * but no one has a reference to it), we want to make sure | |
2655 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2656 | * to the object. Otherwise we will not report the leak. | |
2657 | */ | |
2658 | kmemleak_scan_area(slabp, offsetof(struct slab, list), | |
2659 | sizeof(struct list_head), local_flags); | |
1da177e4 LT |
2660 | if (!slabp) |
2661 | return NULL; | |
2662 | } else { | |
b28a02de | 2663 | slabp = objp + colour_off; |
1da177e4 LT |
2664 | colour_off += cachep->slab_size; |
2665 | } | |
2666 | slabp->inuse = 0; | |
2667 | slabp->colouroff = colour_off; | |
b28a02de | 2668 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2669 | slabp->nodeid = nodeid; |
e51bfd0a | 2670 | slabp->free = 0; |
1da177e4 LT |
2671 | return slabp; |
2672 | } | |
2673 | ||
2674 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2675 | { | |
b28a02de | 2676 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2677 | } |
2678 | ||
343e0d7a | 2679 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2680 | struct slab *slabp) |
1da177e4 LT |
2681 | { |
2682 | int i; | |
2683 | ||
2684 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2685 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2686 | #if DEBUG |
2687 | /* need to poison the objs? */ | |
2688 | if (cachep->flags & SLAB_POISON) | |
2689 | poison_obj(cachep, objp, POISON_FREE); | |
2690 | if (cachep->flags & SLAB_STORE_USER) | |
2691 | *dbg_userword(cachep, objp) = NULL; | |
2692 | ||
2693 | if (cachep->flags & SLAB_RED_ZONE) { | |
2694 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2695 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2696 | } | |
2697 | /* | |
a737b3e2 AM |
2698 | * Constructors are not allowed to allocate memory from the same |
2699 | * cache which they are a constructor for. Otherwise, deadlock. | |
2700 | * They must also be threaded. | |
1da177e4 LT |
2701 | */ |
2702 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2703 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2704 | |
2705 | if (cachep->flags & SLAB_RED_ZONE) { | |
2706 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2707 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2708 | " end of an object"); |
1da177e4 LT |
2709 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2710 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2711 | " start of an object"); |
1da177e4 | 2712 | } |
a737b3e2 AM |
2713 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2714 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2715 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2716 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2717 | #else |
2718 | if (cachep->ctor) | |
51cc5068 | 2719 | cachep->ctor(objp); |
1da177e4 | 2720 | #endif |
b28a02de | 2721 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2722 | } |
b28a02de | 2723 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2724 | } |
2725 | ||
343e0d7a | 2726 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2727 | { |
4b51d669 CL |
2728 | if (CONFIG_ZONE_DMA_FLAG) { |
2729 | if (flags & GFP_DMA) | |
2730 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2731 | else | |
2732 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2733 | } | |
1da177e4 LT |
2734 | } |
2735 | ||
a737b3e2 AM |
2736 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2737 | int nodeid) | |
78d382d7 | 2738 | { |
8fea4e96 | 2739 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2740 | kmem_bufctl_t next; |
2741 | ||
2742 | slabp->inuse++; | |
2743 | next = slab_bufctl(slabp)[slabp->free]; | |
2744 | #if DEBUG | |
2745 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2746 | WARN_ON(slabp->nodeid != nodeid); | |
2747 | #endif | |
2748 | slabp->free = next; | |
2749 | ||
2750 | return objp; | |
2751 | } | |
2752 | ||
a737b3e2 AM |
2753 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2754 | void *objp, int nodeid) | |
78d382d7 | 2755 | { |
8fea4e96 | 2756 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2757 | |
2758 | #if DEBUG | |
2759 | /* Verify that the slab belongs to the intended node */ | |
2760 | WARN_ON(slabp->nodeid != nodeid); | |
2761 | ||
871751e2 | 2762 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2763 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2764 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2765 | BUG(); |
2766 | } | |
2767 | #endif | |
2768 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2769 | slabp->free = objnr; | |
2770 | slabp->inuse--; | |
2771 | } | |
2772 | ||
4776874f PE |
2773 | /* |
2774 | * Map pages beginning at addr to the given cache and slab. This is required | |
2775 | * for the slab allocator to be able to lookup the cache and slab of a | |
2776 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2777 | */ | |
2778 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2779 | void *addr) | |
1da177e4 | 2780 | { |
4776874f | 2781 | int nr_pages; |
1da177e4 LT |
2782 | struct page *page; |
2783 | ||
4776874f | 2784 | page = virt_to_page(addr); |
84097518 | 2785 | |
4776874f | 2786 | nr_pages = 1; |
84097518 | 2787 | if (likely(!PageCompound(page))) |
4776874f PE |
2788 | nr_pages <<= cache->gfporder; |
2789 | ||
1da177e4 | 2790 | do { |
4776874f PE |
2791 | page_set_cache(page, cache); |
2792 | page_set_slab(page, slab); | |
1da177e4 | 2793 | page++; |
4776874f | 2794 | } while (--nr_pages); |
1da177e4 LT |
2795 | } |
2796 | ||
2797 | /* | |
2798 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2799 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2800 | */ | |
3c517a61 CL |
2801 | static int cache_grow(struct kmem_cache *cachep, |
2802 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2803 | { |
b28a02de | 2804 | struct slab *slabp; |
b28a02de PE |
2805 | size_t offset; |
2806 | gfp_t local_flags; | |
e498be7d | 2807 | struct kmem_list3 *l3; |
1da177e4 | 2808 | |
a737b3e2 AM |
2809 | /* |
2810 | * Be lazy and only check for valid flags here, keeping it out of the | |
2811 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2812 | */ |
6cb06229 CL |
2813 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2814 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2815 | |
2e1217cf | 2816 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2817 | check_irq_off(); |
2e1217cf RT |
2818 | l3 = cachep->nodelists[nodeid]; |
2819 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2820 | |
2821 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2822 | offset = l3->colour_next; |
2823 | l3->colour_next++; | |
2824 | if (l3->colour_next >= cachep->colour) | |
2825 | l3->colour_next = 0; | |
2826 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2827 | |
2e1217cf | 2828 | offset *= cachep->colour_off; |
1da177e4 LT |
2829 | |
2830 | if (local_flags & __GFP_WAIT) | |
2831 | local_irq_enable(); | |
2832 | ||
2833 | /* | |
2834 | * The test for missing atomic flag is performed here, rather than | |
2835 | * the more obvious place, simply to reduce the critical path length | |
2836 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2837 | * will eventually be caught here (where it matters). | |
2838 | */ | |
2839 | kmem_flagcheck(cachep, flags); | |
2840 | ||
a737b3e2 AM |
2841 | /* |
2842 | * Get mem for the objs. Attempt to allocate a physical page from | |
2843 | * 'nodeid'. | |
e498be7d | 2844 | */ |
3c517a61 | 2845 | if (!objp) |
b8c1c5da | 2846 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2847 | if (!objp) |
1da177e4 LT |
2848 | goto failed; |
2849 | ||
2850 | /* Get slab management. */ | |
3c517a61 | 2851 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2852 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2853 | if (!slabp) |
1da177e4 LT |
2854 | goto opps1; |
2855 | ||
4776874f | 2856 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2857 | |
a35afb83 | 2858 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2859 | |
2860 | if (local_flags & __GFP_WAIT) | |
2861 | local_irq_disable(); | |
2862 | check_irq_off(); | |
e498be7d | 2863 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2864 | |
2865 | /* Make slab active. */ | |
e498be7d | 2866 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2867 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2868 | l3->free_objects += cachep->num; |
2869 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2870 | return 1; |
a737b3e2 | 2871 | opps1: |
1da177e4 | 2872 | kmem_freepages(cachep, objp); |
a737b3e2 | 2873 | failed: |
1da177e4 LT |
2874 | if (local_flags & __GFP_WAIT) |
2875 | local_irq_disable(); | |
2876 | return 0; | |
2877 | } | |
2878 | ||
2879 | #if DEBUG | |
2880 | ||
2881 | /* | |
2882 | * Perform extra freeing checks: | |
2883 | * - detect bad pointers. | |
2884 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2885 | */ |
2886 | static void kfree_debugcheck(const void *objp) | |
2887 | { | |
1da177e4 LT |
2888 | if (!virt_addr_valid(objp)) { |
2889 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2890 | (unsigned long)objp); |
2891 | BUG(); | |
1da177e4 | 2892 | } |
1da177e4 LT |
2893 | } |
2894 | ||
58ce1fd5 PE |
2895 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2896 | { | |
b46b8f19 | 2897 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2898 | |
2899 | redzone1 = *dbg_redzone1(cache, obj); | |
2900 | redzone2 = *dbg_redzone2(cache, obj); | |
2901 | ||
2902 | /* | |
2903 | * Redzone is ok. | |
2904 | */ | |
2905 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2906 | return; | |
2907 | ||
2908 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2909 | slab_error(cache, "double free detected"); | |
2910 | else | |
2911 | slab_error(cache, "memory outside object was overwritten"); | |
2912 | ||
b46b8f19 | 2913 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2914 | obj, redzone1, redzone2); |
2915 | } | |
2916 | ||
343e0d7a | 2917 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2918 | void *caller) |
1da177e4 LT |
2919 | { |
2920 | struct page *page; | |
2921 | unsigned int objnr; | |
2922 | struct slab *slabp; | |
2923 | ||
80cbd911 MW |
2924 | BUG_ON(virt_to_cache(objp) != cachep); |
2925 | ||
3dafccf2 | 2926 | objp -= obj_offset(cachep); |
1da177e4 | 2927 | kfree_debugcheck(objp); |
b49af68f | 2928 | page = virt_to_head_page(objp); |
1da177e4 | 2929 | |
065d41cb | 2930 | slabp = page_get_slab(page); |
1da177e4 LT |
2931 | |
2932 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2933 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2934 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2935 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2936 | } | |
2937 | if (cachep->flags & SLAB_STORE_USER) | |
2938 | *dbg_userword(cachep, objp) = caller; | |
2939 | ||
8fea4e96 | 2940 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2941 | |
2942 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2943 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2944 | |
871751e2 AV |
2945 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2946 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2947 | #endif | |
1da177e4 LT |
2948 | if (cachep->flags & SLAB_POISON) { |
2949 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2950 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2951 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2952 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2953 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2954 | } else { |
2955 | poison_obj(cachep, objp, POISON_FREE); | |
2956 | } | |
2957 | #else | |
2958 | poison_obj(cachep, objp, POISON_FREE); | |
2959 | #endif | |
2960 | } | |
2961 | return objp; | |
2962 | } | |
2963 | ||
343e0d7a | 2964 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2965 | { |
2966 | kmem_bufctl_t i; | |
2967 | int entries = 0; | |
b28a02de | 2968 | |
1da177e4 LT |
2969 | /* Check slab's freelist to see if this obj is there. */ |
2970 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2971 | entries++; | |
2972 | if (entries > cachep->num || i >= cachep->num) | |
2973 | goto bad; | |
2974 | } | |
2975 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2976 | bad: |
2977 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2978 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2979 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2980 | for (i = 0; |
264132bc | 2981 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2982 | i++) { |
a737b3e2 | 2983 | if (i % 16 == 0) |
1da177e4 | 2984 | printk("\n%03x:", i); |
b28a02de | 2985 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2986 | } |
2987 | printk("\n"); | |
2988 | BUG(); | |
2989 | } | |
2990 | } | |
2991 | #else | |
2992 | #define kfree_debugcheck(x) do { } while(0) | |
2993 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2994 | #define check_slabp(x,y) do { } while(0) | |
2995 | #endif | |
2996 | ||
343e0d7a | 2997 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2998 | { |
2999 | int batchcount; | |
3000 | struct kmem_list3 *l3; | |
3001 | struct array_cache *ac; | |
1ca4cb24 PE |
3002 | int node; |
3003 | ||
6d2144d3 | 3004 | retry: |
1da177e4 | 3005 | check_irq_off(); |
6d2144d3 | 3006 | node = numa_node_id(); |
9a2dba4b | 3007 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3008 | batchcount = ac->batchcount; |
3009 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3010 | /* |
3011 | * If there was little recent activity on this cache, then | |
3012 | * perform only a partial refill. Otherwise we could generate | |
3013 | * refill bouncing. | |
1da177e4 LT |
3014 | */ |
3015 | batchcount = BATCHREFILL_LIMIT; | |
3016 | } | |
1ca4cb24 | 3017 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3018 | |
3019 | BUG_ON(ac->avail > 0 || !l3); | |
3020 | spin_lock(&l3->list_lock); | |
1da177e4 | 3021 | |
3ded175a CL |
3022 | /* See if we can refill from the shared array */ |
3023 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
3024 | goto alloc_done; | |
3025 | ||
1da177e4 LT |
3026 | while (batchcount > 0) { |
3027 | struct list_head *entry; | |
3028 | struct slab *slabp; | |
3029 | /* Get slab alloc is to come from. */ | |
3030 | entry = l3->slabs_partial.next; | |
3031 | if (entry == &l3->slabs_partial) { | |
3032 | l3->free_touched = 1; | |
3033 | entry = l3->slabs_free.next; | |
3034 | if (entry == &l3->slabs_free) | |
3035 | goto must_grow; | |
3036 | } | |
3037 | ||
3038 | slabp = list_entry(entry, struct slab, list); | |
3039 | check_slabp(cachep, slabp); | |
3040 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3041 | |
3042 | /* | |
3043 | * The slab was either on partial or free list so | |
3044 | * there must be at least one object available for | |
3045 | * allocation. | |
3046 | */ | |
249b9f33 | 3047 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3048 | |
1da177e4 | 3049 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3050 | STATS_INC_ALLOCED(cachep); |
3051 | STATS_INC_ACTIVE(cachep); | |
3052 | STATS_SET_HIGH(cachep); | |
3053 | ||
78d382d7 | 3054 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3055 | node); |
1da177e4 LT |
3056 | } |
3057 | check_slabp(cachep, slabp); | |
3058 | ||
3059 | /* move slabp to correct slabp list: */ | |
3060 | list_del(&slabp->list); | |
3061 | if (slabp->free == BUFCTL_END) | |
3062 | list_add(&slabp->list, &l3->slabs_full); | |
3063 | else | |
3064 | list_add(&slabp->list, &l3->slabs_partial); | |
3065 | } | |
3066 | ||
a737b3e2 | 3067 | must_grow: |
1da177e4 | 3068 | l3->free_objects -= ac->avail; |
a737b3e2 | 3069 | alloc_done: |
e498be7d | 3070 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3071 | |
3072 | if (unlikely(!ac->avail)) { | |
3073 | int x; | |
3c517a61 | 3074 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3075 | |
a737b3e2 | 3076 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3077 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3078 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3079 | return NULL; |
3080 | ||
a737b3e2 | 3081 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3082 | goto retry; |
3083 | } | |
3084 | ac->touched = 1; | |
e498be7d | 3085 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3086 | } |
3087 | ||
a737b3e2 AM |
3088 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3089 | gfp_t flags) | |
1da177e4 LT |
3090 | { |
3091 | might_sleep_if(flags & __GFP_WAIT); | |
3092 | #if DEBUG | |
3093 | kmem_flagcheck(cachep, flags); | |
3094 | #endif | |
3095 | } | |
3096 | ||
3097 | #if DEBUG | |
a737b3e2 AM |
3098 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3099 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3100 | { |
b28a02de | 3101 | if (!objp) |
1da177e4 | 3102 | return objp; |
b28a02de | 3103 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3104 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3105 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3106 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3107 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3108 | else |
3109 | check_poison_obj(cachep, objp); | |
3110 | #else | |
3111 | check_poison_obj(cachep, objp); | |
3112 | #endif | |
3113 | poison_obj(cachep, objp, POISON_INUSE); | |
3114 | } | |
3115 | if (cachep->flags & SLAB_STORE_USER) | |
3116 | *dbg_userword(cachep, objp) = caller; | |
3117 | ||
3118 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3119 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3120 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3121 | slab_error(cachep, "double free, or memory outside" | |
3122 | " object was overwritten"); | |
b28a02de | 3123 | printk(KERN_ERR |
b46b8f19 | 3124 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3125 | objp, *dbg_redzone1(cachep, objp), |
3126 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3127 | } |
3128 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3129 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3130 | } | |
871751e2 AV |
3131 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3132 | { | |
3133 | struct slab *slabp; | |
3134 | unsigned objnr; | |
3135 | ||
b49af68f | 3136 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3137 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3138 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3139 | } | |
3140 | #endif | |
3dafccf2 | 3141 | objp += obj_offset(cachep); |
4f104934 | 3142 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3143 | cachep->ctor(objp); |
a44b56d3 KH |
3144 | #if ARCH_SLAB_MINALIGN |
3145 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3146 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3147 | objp, ARCH_SLAB_MINALIGN); | |
3148 | } | |
3149 | #endif | |
1da177e4 LT |
3150 | return objp; |
3151 | } | |
3152 | #else | |
3153 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3154 | #endif | |
3155 | ||
773ff60e | 3156 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3157 | { |
3158 | if (cachep == &cache_cache) | |
773ff60e | 3159 | return false; |
8a8b6502 | 3160 | |
773ff60e | 3161 | return should_failslab(obj_size(cachep), flags); |
8a8b6502 AM |
3162 | } |
3163 | ||
343e0d7a | 3164 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3165 | { |
b28a02de | 3166 | void *objp; |
1da177e4 LT |
3167 | struct array_cache *ac; |
3168 | ||
5c382300 | 3169 | check_irq_off(); |
8a8b6502 | 3170 | |
9a2dba4b | 3171 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3172 | if (likely(ac->avail)) { |
3173 | STATS_INC_ALLOCHIT(cachep); | |
3174 | ac->touched = 1; | |
e498be7d | 3175 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3176 | } else { |
3177 | STATS_INC_ALLOCMISS(cachep); | |
3178 | objp = cache_alloc_refill(cachep, flags); | |
3179 | } | |
d5cff635 CM |
3180 | /* |
3181 | * To avoid a false negative, if an object that is in one of the | |
3182 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3183 | * treat the array pointers as a reference to the object. | |
3184 | */ | |
3185 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3186 | return objp; |
3187 | } | |
3188 | ||
e498be7d | 3189 | #ifdef CONFIG_NUMA |
c61afb18 | 3190 | /* |
b2455396 | 3191 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3192 | * |
3193 | * If we are in_interrupt, then process context, including cpusets and | |
3194 | * mempolicy, may not apply and should not be used for allocation policy. | |
3195 | */ | |
3196 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3197 | { | |
3198 | int nid_alloc, nid_here; | |
3199 | ||
765c4507 | 3200 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3201 | return NULL; |
3202 | nid_alloc = nid_here = numa_node_id(); | |
3203 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3204 | nid_alloc = cpuset_mem_spread_node(); | |
3205 | else if (current->mempolicy) | |
3206 | nid_alloc = slab_node(current->mempolicy); | |
3207 | if (nid_alloc != nid_here) | |
8b98c169 | 3208 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3209 | return NULL; |
3210 | } | |
3211 | ||
765c4507 CL |
3212 | /* |
3213 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3214 | * certain node and fall back is permitted. First we scan all the |
3215 | * available nodelists for available objects. If that fails then we | |
3216 | * perform an allocation without specifying a node. This allows the page | |
3217 | * allocator to do its reclaim / fallback magic. We then insert the | |
3218 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3219 | */ |
8c8cc2c1 | 3220 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3221 | { |
8c8cc2c1 PE |
3222 | struct zonelist *zonelist; |
3223 | gfp_t local_flags; | |
dd1a239f | 3224 | struct zoneref *z; |
54a6eb5c MG |
3225 | struct zone *zone; |
3226 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3227 | void *obj = NULL; |
3c517a61 | 3228 | int nid; |
8c8cc2c1 PE |
3229 | |
3230 | if (flags & __GFP_THISNODE) | |
3231 | return NULL; | |
3232 | ||
0e88460d | 3233 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
6cb06229 | 3234 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3235 | |
3c517a61 CL |
3236 | retry: |
3237 | /* | |
3238 | * Look through allowed nodes for objects available | |
3239 | * from existing per node queues. | |
3240 | */ | |
54a6eb5c MG |
3241 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3242 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3243 | |
54a6eb5c | 3244 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3245 | cache->nodelists[nid] && |
481c5346 | 3246 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3247 | obj = ____cache_alloc_node(cache, |
3248 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3249 | if (obj) |
3250 | break; | |
3251 | } | |
3c517a61 CL |
3252 | } |
3253 | ||
cfce6604 | 3254 | if (!obj) { |
3c517a61 CL |
3255 | /* |
3256 | * This allocation will be performed within the constraints | |
3257 | * of the current cpuset / memory policy requirements. | |
3258 | * We may trigger various forms of reclaim on the allowed | |
3259 | * set and go into memory reserves if necessary. | |
3260 | */ | |
dd47ea75 CL |
3261 | if (local_flags & __GFP_WAIT) |
3262 | local_irq_enable(); | |
3263 | kmem_flagcheck(cache, flags); | |
6484eb3e | 3264 | obj = kmem_getpages(cache, local_flags, numa_node_id()); |
dd47ea75 CL |
3265 | if (local_flags & __GFP_WAIT) |
3266 | local_irq_disable(); | |
3c517a61 CL |
3267 | if (obj) { |
3268 | /* | |
3269 | * Insert into the appropriate per node queues | |
3270 | */ | |
3271 | nid = page_to_nid(virt_to_page(obj)); | |
3272 | if (cache_grow(cache, flags, nid, obj)) { | |
3273 | obj = ____cache_alloc_node(cache, | |
3274 | flags | GFP_THISNODE, nid); | |
3275 | if (!obj) | |
3276 | /* | |
3277 | * Another processor may allocate the | |
3278 | * objects in the slab since we are | |
3279 | * not holding any locks. | |
3280 | */ | |
3281 | goto retry; | |
3282 | } else { | |
b6a60451 | 3283 | /* cache_grow already freed obj */ |
3c517a61 CL |
3284 | obj = NULL; |
3285 | } | |
3286 | } | |
aedb0eb1 | 3287 | } |
765c4507 CL |
3288 | return obj; |
3289 | } | |
3290 | ||
e498be7d CL |
3291 | /* |
3292 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3293 | */ |
8b98c169 | 3294 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3295 | int nodeid) |
e498be7d CL |
3296 | { |
3297 | struct list_head *entry; | |
b28a02de PE |
3298 | struct slab *slabp; |
3299 | struct kmem_list3 *l3; | |
3300 | void *obj; | |
b28a02de PE |
3301 | int x; |
3302 | ||
3303 | l3 = cachep->nodelists[nodeid]; | |
3304 | BUG_ON(!l3); | |
3305 | ||
a737b3e2 | 3306 | retry: |
ca3b9b91 | 3307 | check_irq_off(); |
b28a02de PE |
3308 | spin_lock(&l3->list_lock); |
3309 | entry = l3->slabs_partial.next; | |
3310 | if (entry == &l3->slabs_partial) { | |
3311 | l3->free_touched = 1; | |
3312 | entry = l3->slabs_free.next; | |
3313 | if (entry == &l3->slabs_free) | |
3314 | goto must_grow; | |
3315 | } | |
3316 | ||
3317 | slabp = list_entry(entry, struct slab, list); | |
3318 | check_spinlock_acquired_node(cachep, nodeid); | |
3319 | check_slabp(cachep, slabp); | |
3320 | ||
3321 | STATS_INC_NODEALLOCS(cachep); | |
3322 | STATS_INC_ACTIVE(cachep); | |
3323 | STATS_SET_HIGH(cachep); | |
3324 | ||
3325 | BUG_ON(slabp->inuse == cachep->num); | |
3326 | ||
78d382d7 | 3327 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3328 | check_slabp(cachep, slabp); |
3329 | l3->free_objects--; | |
3330 | /* move slabp to correct slabp list: */ | |
3331 | list_del(&slabp->list); | |
3332 | ||
a737b3e2 | 3333 | if (slabp->free == BUFCTL_END) |
b28a02de | 3334 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3335 | else |
b28a02de | 3336 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3337 | |
b28a02de PE |
3338 | spin_unlock(&l3->list_lock); |
3339 | goto done; | |
e498be7d | 3340 | |
a737b3e2 | 3341 | must_grow: |
b28a02de | 3342 | spin_unlock(&l3->list_lock); |
3c517a61 | 3343 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3344 | if (x) |
3345 | goto retry; | |
1da177e4 | 3346 | |
8c8cc2c1 | 3347 | return fallback_alloc(cachep, flags); |
e498be7d | 3348 | |
a737b3e2 | 3349 | done: |
b28a02de | 3350 | return obj; |
e498be7d | 3351 | } |
8c8cc2c1 PE |
3352 | |
3353 | /** | |
3354 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3355 | * @cachep: The cache to allocate from. | |
3356 | * @flags: See kmalloc(). | |
3357 | * @nodeid: node number of the target node. | |
3358 | * @caller: return address of caller, used for debug information | |
3359 | * | |
3360 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3361 | * node, which can improve the performance for cpu bound structures. | |
3362 | * | |
3363 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3364 | */ | |
3365 | static __always_inline void * | |
3366 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3367 | void *caller) | |
3368 | { | |
3369 | unsigned long save_flags; | |
3370 | void *ptr; | |
3371 | ||
7e85ee0c PE |
3372 | flags &= slab_gfp_mask; |
3373 | ||
cf40bd16 NP |
3374 | lockdep_trace_alloc(flags); |
3375 | ||
773ff60e | 3376 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3377 | return NULL; |
3378 | ||
8c8cc2c1 PE |
3379 | cache_alloc_debugcheck_before(cachep, flags); |
3380 | local_irq_save(save_flags); | |
3381 | ||
3382 | if (unlikely(nodeid == -1)) | |
3383 | nodeid = numa_node_id(); | |
3384 | ||
3385 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3386 | /* Node not bootstrapped yet */ | |
3387 | ptr = fallback_alloc(cachep, flags); | |
3388 | goto out; | |
3389 | } | |
3390 | ||
3391 | if (nodeid == numa_node_id()) { | |
3392 | /* | |
3393 | * Use the locally cached objects if possible. | |
3394 | * However ____cache_alloc does not allow fallback | |
3395 | * to other nodes. It may fail while we still have | |
3396 | * objects on other nodes available. | |
3397 | */ | |
3398 | ptr = ____cache_alloc(cachep, flags); | |
3399 | if (ptr) | |
3400 | goto out; | |
3401 | } | |
3402 | /* ___cache_alloc_node can fall back to other nodes */ | |
3403 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3404 | out: | |
3405 | local_irq_restore(save_flags); | |
3406 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
d5cff635 CM |
3407 | kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags, |
3408 | flags); | |
8c8cc2c1 | 3409 | |
d07dbea4 CL |
3410 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3411 | memset(ptr, 0, obj_size(cachep)); | |
3412 | ||
8c8cc2c1 PE |
3413 | return ptr; |
3414 | } | |
3415 | ||
3416 | static __always_inline void * | |
3417 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3418 | { | |
3419 | void *objp; | |
3420 | ||
3421 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3422 | objp = alternate_node_alloc(cache, flags); | |
3423 | if (objp) | |
3424 | goto out; | |
3425 | } | |
3426 | objp = ____cache_alloc(cache, flags); | |
3427 | ||
3428 | /* | |
3429 | * We may just have run out of memory on the local node. | |
3430 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3431 | */ | |
3432 | if (!objp) | |
3433 | objp = ____cache_alloc_node(cache, flags, numa_node_id()); | |
3434 | ||
3435 | out: | |
3436 | return objp; | |
3437 | } | |
3438 | #else | |
3439 | ||
3440 | static __always_inline void * | |
3441 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3442 | { | |
3443 | return ____cache_alloc(cachep, flags); | |
3444 | } | |
3445 | ||
3446 | #endif /* CONFIG_NUMA */ | |
3447 | ||
3448 | static __always_inline void * | |
3449 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3450 | { | |
3451 | unsigned long save_flags; | |
3452 | void *objp; | |
3453 | ||
7e85ee0c PE |
3454 | flags &= slab_gfp_mask; |
3455 | ||
cf40bd16 NP |
3456 | lockdep_trace_alloc(flags); |
3457 | ||
773ff60e | 3458 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3459 | return NULL; |
3460 | ||
8c8cc2c1 PE |
3461 | cache_alloc_debugcheck_before(cachep, flags); |
3462 | local_irq_save(save_flags); | |
3463 | objp = __do_cache_alloc(cachep, flags); | |
3464 | local_irq_restore(save_flags); | |
3465 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
d5cff635 CM |
3466 | kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags, |
3467 | flags); | |
8c8cc2c1 PE |
3468 | prefetchw(objp); |
3469 | ||
d07dbea4 CL |
3470 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3471 | memset(objp, 0, obj_size(cachep)); | |
3472 | ||
8c8cc2c1 PE |
3473 | return objp; |
3474 | } | |
e498be7d CL |
3475 | |
3476 | /* | |
3477 | * Caller needs to acquire correct kmem_list's list_lock | |
3478 | */ | |
343e0d7a | 3479 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3480 | int node) |
1da177e4 LT |
3481 | { |
3482 | int i; | |
e498be7d | 3483 | struct kmem_list3 *l3; |
1da177e4 LT |
3484 | |
3485 | for (i = 0; i < nr_objects; i++) { | |
3486 | void *objp = objpp[i]; | |
3487 | struct slab *slabp; | |
1da177e4 | 3488 | |
6ed5eb22 | 3489 | slabp = virt_to_slab(objp); |
ff69416e | 3490 | l3 = cachep->nodelists[node]; |
1da177e4 | 3491 | list_del(&slabp->list); |
ff69416e | 3492 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3493 | check_slabp(cachep, slabp); |
78d382d7 | 3494 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3495 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3496 | l3->free_objects++; |
1da177e4 LT |
3497 | check_slabp(cachep, slabp); |
3498 | ||
3499 | /* fixup slab chains */ | |
3500 | if (slabp->inuse == 0) { | |
e498be7d CL |
3501 | if (l3->free_objects > l3->free_limit) { |
3502 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3503 | /* No need to drop any previously held |
3504 | * lock here, even if we have a off-slab slab | |
3505 | * descriptor it is guaranteed to come from | |
3506 | * a different cache, refer to comments before | |
3507 | * alloc_slabmgmt. | |
3508 | */ | |
1da177e4 LT |
3509 | slab_destroy(cachep, slabp); |
3510 | } else { | |
e498be7d | 3511 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3512 | } |
3513 | } else { | |
3514 | /* Unconditionally move a slab to the end of the | |
3515 | * partial list on free - maximum time for the | |
3516 | * other objects to be freed, too. | |
3517 | */ | |
e498be7d | 3518 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3519 | } |
3520 | } | |
3521 | } | |
3522 | ||
343e0d7a | 3523 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3524 | { |
3525 | int batchcount; | |
e498be7d | 3526 | struct kmem_list3 *l3; |
ff69416e | 3527 | int node = numa_node_id(); |
1da177e4 LT |
3528 | |
3529 | batchcount = ac->batchcount; | |
3530 | #if DEBUG | |
3531 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3532 | #endif | |
3533 | check_irq_off(); | |
ff69416e | 3534 | l3 = cachep->nodelists[node]; |
873623df | 3535 | spin_lock(&l3->list_lock); |
e498be7d CL |
3536 | if (l3->shared) { |
3537 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3538 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3539 | if (max) { |
3540 | if (batchcount > max) | |
3541 | batchcount = max; | |
e498be7d | 3542 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3543 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3544 | shared_array->avail += batchcount; |
3545 | goto free_done; | |
3546 | } | |
3547 | } | |
3548 | ||
ff69416e | 3549 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3550 | free_done: |
1da177e4 LT |
3551 | #if STATS |
3552 | { | |
3553 | int i = 0; | |
3554 | struct list_head *p; | |
3555 | ||
e498be7d CL |
3556 | p = l3->slabs_free.next; |
3557 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3558 | struct slab *slabp; |
3559 | ||
3560 | slabp = list_entry(p, struct slab, list); | |
3561 | BUG_ON(slabp->inuse); | |
3562 | ||
3563 | i++; | |
3564 | p = p->next; | |
3565 | } | |
3566 | STATS_SET_FREEABLE(cachep, i); | |
3567 | } | |
3568 | #endif | |
e498be7d | 3569 | spin_unlock(&l3->list_lock); |
1da177e4 | 3570 | ac->avail -= batchcount; |
a737b3e2 | 3571 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3572 | } |
3573 | ||
3574 | /* | |
a737b3e2 AM |
3575 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3576 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3577 | */ |
873623df | 3578 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3579 | { |
9a2dba4b | 3580 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3581 | |
3582 | check_irq_off(); | |
d5cff635 | 3583 | kmemleak_free_recursive(objp, cachep->flags); |
1da177e4 LT |
3584 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); |
3585 | ||
1807a1aa SS |
3586 | /* |
3587 | * Skip calling cache_free_alien() when the platform is not numa. | |
3588 | * This will avoid cache misses that happen while accessing slabp (which | |
3589 | * is per page memory reference) to get nodeid. Instead use a global | |
3590 | * variable to skip the call, which is mostly likely to be present in | |
3591 | * the cache. | |
3592 | */ | |
3593 | if (numa_platform && cache_free_alien(cachep, objp)) | |
729bd0b7 PE |
3594 | return; |
3595 | ||
1da177e4 LT |
3596 | if (likely(ac->avail < ac->limit)) { |
3597 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3598 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3599 | return; |
3600 | } else { | |
3601 | STATS_INC_FREEMISS(cachep); | |
3602 | cache_flusharray(cachep, ac); | |
e498be7d | 3603 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3604 | } |
3605 | } | |
3606 | ||
3607 | /** | |
3608 | * kmem_cache_alloc - Allocate an object | |
3609 | * @cachep: The cache to allocate from. | |
3610 | * @flags: See kmalloc(). | |
3611 | * | |
3612 | * Allocate an object from this cache. The flags are only relevant | |
3613 | * if the cache has no available objects. | |
3614 | */ | |
343e0d7a | 3615 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3616 | { |
36555751 EGM |
3617 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3618 | ||
ca2b84cb EGM |
3619 | trace_kmem_cache_alloc(_RET_IP_, ret, |
3620 | obj_size(cachep), cachep->buffer_size, flags); | |
36555751 EGM |
3621 | |
3622 | return ret; | |
1da177e4 LT |
3623 | } |
3624 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3625 | ||
36555751 EGM |
3626 | #ifdef CONFIG_KMEMTRACE |
3627 | void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags) | |
3628 | { | |
3629 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3630 | } | |
3631 | EXPORT_SYMBOL(kmem_cache_alloc_notrace); | |
3632 | #endif | |
3633 | ||
1da177e4 | 3634 | /** |
7682486b | 3635 | * kmem_ptr_validate - check if an untrusted pointer might be a slab entry. |
1da177e4 LT |
3636 | * @cachep: the cache we're checking against |
3637 | * @ptr: pointer to validate | |
3638 | * | |
7682486b | 3639 | * This verifies that the untrusted pointer looks sane; |
1da177e4 LT |
3640 | * it is _not_ a guarantee that the pointer is actually |
3641 | * part of the slab cache in question, but it at least | |
3642 | * validates that the pointer can be dereferenced and | |
3643 | * looks half-way sane. | |
3644 | * | |
3645 | * Currently only used for dentry validation. | |
3646 | */ | |
b7f869a2 | 3647 | int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) |
1da177e4 | 3648 | { |
b28a02de | 3649 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3650 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3651 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3652 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3653 | struct page *page; |
3654 | ||
3655 | if (unlikely(addr < min_addr)) | |
3656 | goto out; | |
3657 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3658 | goto out; | |
3659 | if (unlikely(addr & align_mask)) | |
3660 | goto out; | |
3661 | if (unlikely(!kern_addr_valid(addr))) | |
3662 | goto out; | |
3663 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3664 | goto out; | |
3665 | page = virt_to_page(ptr); | |
3666 | if (unlikely(!PageSlab(page))) | |
3667 | goto out; | |
065d41cb | 3668 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3669 | goto out; |
3670 | return 1; | |
a737b3e2 | 3671 | out: |
1da177e4 LT |
3672 | return 0; |
3673 | } | |
3674 | ||
3675 | #ifdef CONFIG_NUMA | |
8b98c169 CH |
3676 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3677 | { | |
36555751 EGM |
3678 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3679 | __builtin_return_address(0)); | |
3680 | ||
ca2b84cb EGM |
3681 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
3682 | obj_size(cachep), cachep->buffer_size, | |
3683 | flags, nodeid); | |
36555751 EGM |
3684 | |
3685 | return ret; | |
8b98c169 | 3686 | } |
1da177e4 LT |
3687 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3688 | ||
36555751 EGM |
3689 | #ifdef CONFIG_KMEMTRACE |
3690 | void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep, | |
3691 | gfp_t flags, | |
3692 | int nodeid) | |
3693 | { | |
3694 | return __cache_alloc_node(cachep, flags, nodeid, | |
3695 | __builtin_return_address(0)); | |
3696 | } | |
3697 | EXPORT_SYMBOL(kmem_cache_alloc_node_notrace); | |
3698 | #endif | |
3699 | ||
8b98c169 CH |
3700 | static __always_inline void * |
3701 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3702 | { |
343e0d7a | 3703 | struct kmem_cache *cachep; |
36555751 | 3704 | void *ret; |
97e2bde4 MS |
3705 | |
3706 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3707 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3708 | return cachep; | |
36555751 EGM |
3709 | ret = kmem_cache_alloc_node_notrace(cachep, flags, node); |
3710 | ||
ca2b84cb EGM |
3711 | trace_kmalloc_node((unsigned long) caller, ret, |
3712 | size, cachep->buffer_size, flags, node); | |
36555751 EGM |
3713 | |
3714 | return ret; | |
97e2bde4 | 3715 | } |
8b98c169 | 3716 | |
36555751 | 3717 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE) |
8b98c169 CH |
3718 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3719 | { | |
3720 | return __do_kmalloc_node(size, flags, node, | |
3721 | __builtin_return_address(0)); | |
3722 | } | |
dbe5e69d | 3723 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3724 | |
3725 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3726 | int node, unsigned long caller) |
8b98c169 | 3727 | { |
ce71e27c | 3728 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3729 | } |
3730 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3731 | #else | |
3732 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3733 | { | |
3734 | return __do_kmalloc_node(size, flags, node, NULL); | |
3735 | } | |
3736 | EXPORT_SYMBOL(__kmalloc_node); | |
3737 | #endif /* CONFIG_DEBUG_SLAB */ | |
3738 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3739 | |
3740 | /** | |
800590f5 | 3741 | * __do_kmalloc - allocate memory |
1da177e4 | 3742 | * @size: how many bytes of memory are required. |
800590f5 | 3743 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3744 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3745 | */ |
7fd6b141 PE |
3746 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3747 | void *caller) | |
1da177e4 | 3748 | { |
343e0d7a | 3749 | struct kmem_cache *cachep; |
36555751 | 3750 | void *ret; |
1da177e4 | 3751 | |
97e2bde4 MS |
3752 | /* If you want to save a few bytes .text space: replace |
3753 | * __ with kmem_. | |
3754 | * Then kmalloc uses the uninlined functions instead of the inline | |
3755 | * functions. | |
3756 | */ | |
3757 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3758 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3759 | return cachep; | |
36555751 EGM |
3760 | ret = __cache_alloc(cachep, flags, caller); |
3761 | ||
ca2b84cb EGM |
3762 | trace_kmalloc((unsigned long) caller, ret, |
3763 | size, cachep->buffer_size, flags); | |
36555751 EGM |
3764 | |
3765 | return ret; | |
7fd6b141 PE |
3766 | } |
3767 | ||
7fd6b141 | 3768 | |
36555751 | 3769 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE) |
7fd6b141 PE |
3770 | void *__kmalloc(size_t size, gfp_t flags) |
3771 | { | |
871751e2 | 3772 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3773 | } |
3774 | EXPORT_SYMBOL(__kmalloc); | |
3775 | ||
ce71e27c | 3776 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3777 | { |
ce71e27c | 3778 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3779 | } |
3780 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3781 | |
3782 | #else | |
3783 | void *__kmalloc(size_t size, gfp_t flags) | |
3784 | { | |
3785 | return __do_kmalloc(size, flags, NULL); | |
3786 | } | |
3787 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3788 | #endif |
3789 | ||
1da177e4 LT |
3790 | /** |
3791 | * kmem_cache_free - Deallocate an object | |
3792 | * @cachep: The cache the allocation was from. | |
3793 | * @objp: The previously allocated object. | |
3794 | * | |
3795 | * Free an object which was previously allocated from this | |
3796 | * cache. | |
3797 | */ | |
343e0d7a | 3798 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3799 | { |
3800 | unsigned long flags; | |
3801 | ||
3802 | local_irq_save(flags); | |
898552c9 | 3803 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
3ac7fe5a TG |
3804 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3805 | debug_check_no_obj_freed(objp, obj_size(cachep)); | |
873623df | 3806 | __cache_free(cachep, objp); |
1da177e4 | 3807 | local_irq_restore(flags); |
36555751 | 3808 | |
ca2b84cb | 3809 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3810 | } |
3811 | EXPORT_SYMBOL(kmem_cache_free); | |
3812 | ||
1da177e4 LT |
3813 | /** |
3814 | * kfree - free previously allocated memory | |
3815 | * @objp: pointer returned by kmalloc. | |
3816 | * | |
80e93eff PE |
3817 | * If @objp is NULL, no operation is performed. |
3818 | * | |
1da177e4 LT |
3819 | * Don't free memory not originally allocated by kmalloc() |
3820 | * or you will run into trouble. | |
3821 | */ | |
3822 | void kfree(const void *objp) | |
3823 | { | |
343e0d7a | 3824 | struct kmem_cache *c; |
1da177e4 LT |
3825 | unsigned long flags; |
3826 | ||
2121db74 PE |
3827 | trace_kfree(_RET_IP_, objp); |
3828 | ||
6cb8f913 | 3829 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3830 | return; |
3831 | local_irq_save(flags); | |
3832 | kfree_debugcheck(objp); | |
6ed5eb22 | 3833 | c = virt_to_cache(objp); |
f9b8404c | 3834 | debug_check_no_locks_freed(objp, obj_size(c)); |
3ac7fe5a | 3835 | debug_check_no_obj_freed(objp, obj_size(c)); |
873623df | 3836 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3837 | local_irq_restore(flags); |
3838 | } | |
3839 | EXPORT_SYMBOL(kfree); | |
3840 | ||
343e0d7a | 3841 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3842 | { |
3dafccf2 | 3843 | return obj_size(cachep); |
1da177e4 LT |
3844 | } |
3845 | EXPORT_SYMBOL(kmem_cache_size); | |
3846 | ||
343e0d7a | 3847 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3848 | { |
3849 | return cachep->name; | |
3850 | } | |
3851 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3852 | ||
e498be7d | 3853 | /* |
183ff22b | 3854 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3855 | */ |
83b519e8 | 3856 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3857 | { |
3858 | int node; | |
3859 | struct kmem_list3 *l3; | |
cafeb02e | 3860 | struct array_cache *new_shared; |
3395ee05 | 3861 | struct array_cache **new_alien = NULL; |
e498be7d | 3862 | |
9c09a95c | 3863 | for_each_online_node(node) { |
cafeb02e | 3864 | |
3395ee05 | 3865 | if (use_alien_caches) { |
83b519e8 | 3866 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3867 | if (!new_alien) |
3868 | goto fail; | |
3869 | } | |
cafeb02e | 3870 | |
63109846 ED |
3871 | new_shared = NULL; |
3872 | if (cachep->shared) { | |
3873 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3874 | cachep->shared*cachep->batchcount, |
83b519e8 | 3875 | 0xbaadf00d, gfp); |
63109846 ED |
3876 | if (!new_shared) { |
3877 | free_alien_cache(new_alien); | |
3878 | goto fail; | |
3879 | } | |
0718dc2a | 3880 | } |
cafeb02e | 3881 | |
a737b3e2 AM |
3882 | l3 = cachep->nodelists[node]; |
3883 | if (l3) { | |
cafeb02e CL |
3884 | struct array_cache *shared = l3->shared; |
3885 | ||
e498be7d CL |
3886 | spin_lock_irq(&l3->list_lock); |
3887 | ||
cafeb02e | 3888 | if (shared) |
0718dc2a CL |
3889 | free_block(cachep, shared->entry, |
3890 | shared->avail, node); | |
e498be7d | 3891 | |
cafeb02e CL |
3892 | l3->shared = new_shared; |
3893 | if (!l3->alien) { | |
e498be7d CL |
3894 | l3->alien = new_alien; |
3895 | new_alien = NULL; | |
3896 | } | |
b28a02de | 3897 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3898 | cachep->batchcount + cachep->num; |
e498be7d | 3899 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3900 | kfree(shared); |
e498be7d CL |
3901 | free_alien_cache(new_alien); |
3902 | continue; | |
3903 | } | |
83b519e8 | 3904 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3905 | if (!l3) { |
3906 | free_alien_cache(new_alien); | |
3907 | kfree(new_shared); | |
e498be7d | 3908 | goto fail; |
0718dc2a | 3909 | } |
e498be7d CL |
3910 | |
3911 | kmem_list3_init(l3); | |
3912 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3913 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3914 | l3->shared = new_shared; |
e498be7d | 3915 | l3->alien = new_alien; |
b28a02de | 3916 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3917 | cachep->batchcount + cachep->num; |
e498be7d CL |
3918 | cachep->nodelists[node] = l3; |
3919 | } | |
cafeb02e | 3920 | return 0; |
0718dc2a | 3921 | |
a737b3e2 | 3922 | fail: |
0718dc2a CL |
3923 | if (!cachep->next.next) { |
3924 | /* Cache is not active yet. Roll back what we did */ | |
3925 | node--; | |
3926 | while (node >= 0) { | |
3927 | if (cachep->nodelists[node]) { | |
3928 | l3 = cachep->nodelists[node]; | |
3929 | ||
3930 | kfree(l3->shared); | |
3931 | free_alien_cache(l3->alien); | |
3932 | kfree(l3); | |
3933 | cachep->nodelists[node] = NULL; | |
3934 | } | |
3935 | node--; | |
3936 | } | |
3937 | } | |
cafeb02e | 3938 | return -ENOMEM; |
e498be7d CL |
3939 | } |
3940 | ||
1da177e4 | 3941 | struct ccupdate_struct { |
343e0d7a | 3942 | struct kmem_cache *cachep; |
1da177e4 LT |
3943 | struct array_cache *new[NR_CPUS]; |
3944 | }; | |
3945 | ||
3946 | static void do_ccupdate_local(void *info) | |
3947 | { | |
a737b3e2 | 3948 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3949 | struct array_cache *old; |
3950 | ||
3951 | check_irq_off(); | |
9a2dba4b | 3952 | old = cpu_cache_get(new->cachep); |
e498be7d | 3953 | |
1da177e4 LT |
3954 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3955 | new->new[smp_processor_id()] = old; | |
3956 | } | |
3957 | ||
b5d8ca7c | 3958 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 3959 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 3960 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 3961 | { |
d2e7b7d0 | 3962 | struct ccupdate_struct *new; |
2ed3a4ef | 3963 | int i; |
1da177e4 | 3964 | |
83b519e8 | 3965 | new = kzalloc(sizeof(*new), gfp); |
d2e7b7d0 SS |
3966 | if (!new) |
3967 | return -ENOMEM; | |
3968 | ||
e498be7d | 3969 | for_each_online_cpu(i) { |
d2e7b7d0 | 3970 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
83b519e8 | 3971 | batchcount, gfp); |
d2e7b7d0 | 3972 | if (!new->new[i]) { |
b28a02de | 3973 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3974 | kfree(new->new[i]); |
3975 | kfree(new); | |
e498be7d | 3976 | return -ENOMEM; |
1da177e4 LT |
3977 | } |
3978 | } | |
d2e7b7d0 | 3979 | new->cachep = cachep; |
1da177e4 | 3980 | |
15c8b6c1 | 3981 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 3982 | |
1da177e4 | 3983 | check_irq_on(); |
1da177e4 LT |
3984 | cachep->batchcount = batchcount; |
3985 | cachep->limit = limit; | |
e498be7d | 3986 | cachep->shared = shared; |
1da177e4 | 3987 | |
e498be7d | 3988 | for_each_online_cpu(i) { |
d2e7b7d0 | 3989 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3990 | if (!ccold) |
3991 | continue; | |
e498be7d | 3992 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3993 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3994 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3995 | kfree(ccold); |
3996 | } | |
d2e7b7d0 | 3997 | kfree(new); |
83b519e8 | 3998 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
3999 | } |
4000 | ||
b5d8ca7c | 4001 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 4002 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4003 | { |
4004 | int err; | |
4005 | int limit, shared; | |
4006 | ||
a737b3e2 AM |
4007 | /* |
4008 | * The head array serves three purposes: | |
1da177e4 LT |
4009 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4010 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4011 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4012 | * bufctl chains: array operations are cheaper. |
4013 | * The numbers are guessed, we should auto-tune as described by | |
4014 | * Bonwick. | |
4015 | */ | |
3dafccf2 | 4016 | if (cachep->buffer_size > 131072) |
1da177e4 | 4017 | limit = 1; |
3dafccf2 | 4018 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 4019 | limit = 8; |
3dafccf2 | 4020 | else if (cachep->buffer_size > 1024) |
1da177e4 | 4021 | limit = 24; |
3dafccf2 | 4022 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
4023 | limit = 54; |
4024 | else | |
4025 | limit = 120; | |
4026 | ||
a737b3e2 AM |
4027 | /* |
4028 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4029 | * allocation behaviour: Most allocs on one cpu, most free operations |
4030 | * on another cpu. For these cases, an efficient object passing between | |
4031 | * cpus is necessary. This is provided by a shared array. The array | |
4032 | * replaces Bonwick's magazine layer. | |
4033 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4034 | * to a larger limit. Thus disabled by default. | |
4035 | */ | |
4036 | shared = 0; | |
364fbb29 | 4037 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4038 | shared = 8; |
1da177e4 LT |
4039 | |
4040 | #if DEBUG | |
a737b3e2 AM |
4041 | /* |
4042 | * With debugging enabled, large batchcount lead to excessively long | |
4043 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4044 | */ |
4045 | if (limit > 32) | |
4046 | limit = 32; | |
4047 | #endif | |
83b519e8 | 4048 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4049 | if (err) |
4050 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4051 | cachep->name, -err); |
2ed3a4ef | 4052 | return err; |
1da177e4 LT |
4053 | } |
4054 | ||
1b55253a CL |
4055 | /* |
4056 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4057 | * necessary. Note that the l3 listlock also protects the array_cache |
4058 | * if drain_array() is used on the shared array. | |
1b55253a CL |
4059 | */ |
4060 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4061 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
4062 | { |
4063 | int tofree; | |
4064 | ||
1b55253a CL |
4065 | if (!ac || !ac->avail) |
4066 | return; | |
1da177e4 LT |
4067 | if (ac->touched && !force) { |
4068 | ac->touched = 0; | |
b18e7e65 | 4069 | } else { |
1b55253a | 4070 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4071 | if (ac->avail) { |
4072 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4073 | if (tofree > ac->avail) | |
4074 | tofree = (ac->avail + 1) / 2; | |
4075 | free_block(cachep, ac->entry, tofree, node); | |
4076 | ac->avail -= tofree; | |
4077 | memmove(ac->entry, &(ac->entry[tofree]), | |
4078 | sizeof(void *) * ac->avail); | |
4079 | } | |
1b55253a | 4080 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4081 | } |
4082 | } | |
4083 | ||
4084 | /** | |
4085 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4086 | * @w: work descriptor |
1da177e4 LT |
4087 | * |
4088 | * Called from workqueue/eventd every few seconds. | |
4089 | * Purpose: | |
4090 | * - clear the per-cpu caches for this CPU. | |
4091 | * - return freeable pages to the main free memory pool. | |
4092 | * | |
a737b3e2 AM |
4093 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4094 | * again on the next iteration. | |
1da177e4 | 4095 | */ |
7c5cae36 | 4096 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4097 | { |
7a7c381d | 4098 | struct kmem_cache *searchp; |
e498be7d | 4099 | struct kmem_list3 *l3; |
aab2207c | 4100 | int node = numa_node_id(); |
bf6aede7 | 4101 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4102 | |
7c5cae36 | 4103 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4104 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4105 | goto out; |
1da177e4 | 4106 | |
7a7c381d | 4107 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4108 | check_irq_on(); |
4109 | ||
35386e3b CL |
4110 | /* |
4111 | * We only take the l3 lock if absolutely necessary and we | |
4112 | * have established with reasonable certainty that | |
4113 | * we can do some work if the lock was obtained. | |
4114 | */ | |
aab2207c | 4115 | l3 = searchp->nodelists[node]; |
35386e3b | 4116 | |
8fce4d8e | 4117 | reap_alien(searchp, l3); |
1da177e4 | 4118 | |
aab2207c | 4119 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4120 | |
35386e3b CL |
4121 | /* |
4122 | * These are racy checks but it does not matter | |
4123 | * if we skip one check or scan twice. | |
4124 | */ | |
e498be7d | 4125 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4126 | goto next; |
1da177e4 | 4127 | |
e498be7d | 4128 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4129 | |
aab2207c | 4130 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4131 | |
ed11d9eb | 4132 | if (l3->free_touched) |
e498be7d | 4133 | l3->free_touched = 0; |
ed11d9eb CL |
4134 | else { |
4135 | int freed; | |
1da177e4 | 4136 | |
ed11d9eb CL |
4137 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4138 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4139 | STATS_ADD_REAPED(searchp, freed); | |
4140 | } | |
35386e3b | 4141 | next: |
1da177e4 LT |
4142 | cond_resched(); |
4143 | } | |
4144 | check_irq_on(); | |
fc0abb14 | 4145 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4146 | next_reap_node(); |
7c5cae36 | 4147 | out: |
a737b3e2 | 4148 | /* Set up the next iteration */ |
7c5cae36 | 4149 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4150 | } |
4151 | ||
158a9624 | 4152 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4153 | |
85289f98 | 4154 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4155 | { |
85289f98 PE |
4156 | /* |
4157 | * Output format version, so at least we can change it | |
4158 | * without _too_ many complaints. | |
4159 | */ | |
1da177e4 | 4160 | #if STATS |
85289f98 | 4161 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4162 | #else |
85289f98 | 4163 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4164 | #endif |
85289f98 PE |
4165 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4166 | "<objperslab> <pagesperslab>"); | |
4167 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4168 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4169 | #if STATS |
85289f98 | 4170 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4171 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4172 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4173 | #endif |
85289f98 PE |
4174 | seq_putc(m, '\n'); |
4175 | } | |
4176 | ||
4177 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4178 | { | |
4179 | loff_t n = *pos; | |
85289f98 | 4180 | |
fc0abb14 | 4181 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4182 | if (!n) |
4183 | print_slabinfo_header(m); | |
b92151ba PE |
4184 | |
4185 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4186 | } |
4187 | ||
4188 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4189 | { | |
b92151ba | 4190 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4191 | } |
4192 | ||
4193 | static void s_stop(struct seq_file *m, void *p) | |
4194 | { | |
fc0abb14 | 4195 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4196 | } |
4197 | ||
4198 | static int s_show(struct seq_file *m, void *p) | |
4199 | { | |
b92151ba | 4200 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
b28a02de PE |
4201 | struct slab *slabp; |
4202 | unsigned long active_objs; | |
4203 | unsigned long num_objs; | |
4204 | unsigned long active_slabs = 0; | |
4205 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4206 | const char *name; |
1da177e4 | 4207 | char *error = NULL; |
e498be7d CL |
4208 | int node; |
4209 | struct kmem_list3 *l3; | |
1da177e4 | 4210 | |
1da177e4 LT |
4211 | active_objs = 0; |
4212 | num_slabs = 0; | |
e498be7d CL |
4213 | for_each_online_node(node) { |
4214 | l3 = cachep->nodelists[node]; | |
4215 | if (!l3) | |
4216 | continue; | |
4217 | ||
ca3b9b91 RT |
4218 | check_irq_on(); |
4219 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4220 | |
7a7c381d | 4221 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4222 | if (slabp->inuse != cachep->num && !error) |
4223 | error = "slabs_full accounting error"; | |
4224 | active_objs += cachep->num; | |
4225 | active_slabs++; | |
4226 | } | |
7a7c381d | 4227 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4228 | if (slabp->inuse == cachep->num && !error) |
4229 | error = "slabs_partial inuse accounting error"; | |
4230 | if (!slabp->inuse && !error) | |
4231 | error = "slabs_partial/inuse accounting error"; | |
4232 | active_objs += slabp->inuse; | |
4233 | active_slabs++; | |
4234 | } | |
7a7c381d | 4235 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4236 | if (slabp->inuse && !error) |
4237 | error = "slabs_free/inuse accounting error"; | |
4238 | num_slabs++; | |
4239 | } | |
4240 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4241 | if (l3->shared) |
4242 | shared_avail += l3->shared->avail; | |
e498be7d | 4243 | |
ca3b9b91 | 4244 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4245 | } |
b28a02de PE |
4246 | num_slabs += active_slabs; |
4247 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4248 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4249 | error = "free_objects accounting error"; |
4250 | ||
b28a02de | 4251 | name = cachep->name; |
1da177e4 LT |
4252 | if (error) |
4253 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4254 | ||
4255 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4256 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4257 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4258 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4259 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4260 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4261 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4262 | #if STATS |
b28a02de | 4263 | { /* list3 stats */ |
1da177e4 LT |
4264 | unsigned long high = cachep->high_mark; |
4265 | unsigned long allocs = cachep->num_allocations; | |
4266 | unsigned long grown = cachep->grown; | |
4267 | unsigned long reaped = cachep->reaped; | |
4268 | unsigned long errors = cachep->errors; | |
4269 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4270 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4271 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4272 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4273 | |
e498be7d | 4274 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4275 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4276 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4277 | node_frees, overflows); |
1da177e4 LT |
4278 | } |
4279 | /* cpu stats */ | |
4280 | { | |
4281 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4282 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4283 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4284 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4285 | ||
4286 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4287 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4288 | } |
4289 | #endif | |
4290 | seq_putc(m, '\n'); | |
1da177e4 LT |
4291 | return 0; |
4292 | } | |
4293 | ||
4294 | /* | |
4295 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4296 | * | |
4297 | * Output layout: | |
4298 | * cache-name | |
4299 | * num-active-objs | |
4300 | * total-objs | |
4301 | * object size | |
4302 | * num-active-slabs | |
4303 | * total-slabs | |
4304 | * num-pages-per-slab | |
4305 | * + further values on SMP and with statistics enabled | |
4306 | */ | |
4307 | ||
7b3c3a50 | 4308 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4309 | .start = s_start, |
4310 | .next = s_next, | |
4311 | .stop = s_stop, | |
4312 | .show = s_show, | |
1da177e4 LT |
4313 | }; |
4314 | ||
4315 | #define MAX_SLABINFO_WRITE 128 | |
4316 | /** | |
4317 | * slabinfo_write - Tuning for the slab allocator | |
4318 | * @file: unused | |
4319 | * @buffer: user buffer | |
4320 | * @count: data length | |
4321 | * @ppos: unused | |
4322 | */ | |
b28a02de PE |
4323 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4324 | size_t count, loff_t *ppos) | |
1da177e4 | 4325 | { |
b28a02de | 4326 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4327 | int limit, batchcount, shared, res; |
7a7c381d | 4328 | struct kmem_cache *cachep; |
b28a02de | 4329 | |
1da177e4 LT |
4330 | if (count > MAX_SLABINFO_WRITE) |
4331 | return -EINVAL; | |
4332 | if (copy_from_user(&kbuf, buffer, count)) | |
4333 | return -EFAULT; | |
b28a02de | 4334 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4335 | |
4336 | tmp = strchr(kbuf, ' '); | |
4337 | if (!tmp) | |
4338 | return -EINVAL; | |
4339 | *tmp = '\0'; | |
4340 | tmp++; | |
4341 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4342 | return -EINVAL; | |
4343 | ||
4344 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4345 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4346 | res = -EINVAL; |
7a7c381d | 4347 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4348 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4349 | if (limit < 1 || batchcount < 1 || |
4350 | batchcount > limit || shared < 0) { | |
e498be7d | 4351 | res = 0; |
1da177e4 | 4352 | } else { |
e498be7d | 4353 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4354 | batchcount, shared, |
4355 | GFP_KERNEL); | |
1da177e4 LT |
4356 | } |
4357 | break; | |
4358 | } | |
4359 | } | |
fc0abb14 | 4360 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4361 | if (res >= 0) |
4362 | res = count; | |
4363 | return res; | |
4364 | } | |
871751e2 | 4365 | |
7b3c3a50 AD |
4366 | static int slabinfo_open(struct inode *inode, struct file *file) |
4367 | { | |
4368 | return seq_open(file, &slabinfo_op); | |
4369 | } | |
4370 | ||
4371 | static const struct file_operations proc_slabinfo_operations = { | |
4372 | .open = slabinfo_open, | |
4373 | .read = seq_read, | |
4374 | .write = slabinfo_write, | |
4375 | .llseek = seq_lseek, | |
4376 | .release = seq_release, | |
4377 | }; | |
4378 | ||
871751e2 AV |
4379 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4380 | ||
4381 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4382 | { | |
871751e2 | 4383 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4384 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4385 | } |
4386 | ||
4387 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4388 | { | |
4389 | unsigned long *p; | |
4390 | int l; | |
4391 | if (!v) | |
4392 | return 1; | |
4393 | l = n[1]; | |
4394 | p = n + 2; | |
4395 | while (l) { | |
4396 | int i = l/2; | |
4397 | unsigned long *q = p + 2 * i; | |
4398 | if (*q == v) { | |
4399 | q[1]++; | |
4400 | return 1; | |
4401 | } | |
4402 | if (*q > v) { | |
4403 | l = i; | |
4404 | } else { | |
4405 | p = q + 2; | |
4406 | l -= i + 1; | |
4407 | } | |
4408 | } | |
4409 | if (++n[1] == n[0]) | |
4410 | return 0; | |
4411 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4412 | p[0] = v; | |
4413 | p[1] = 1; | |
4414 | return 1; | |
4415 | } | |
4416 | ||
4417 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4418 | { | |
4419 | void *p; | |
4420 | int i; | |
4421 | if (n[0] == n[1]) | |
4422 | return; | |
4423 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4424 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4425 | continue; | |
4426 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4427 | return; | |
4428 | } | |
4429 | } | |
4430 | ||
4431 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4432 | { | |
4433 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4434 | unsigned long offset, size; |
9281acea | 4435 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4436 | |
a5c43dae | 4437 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4438 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4439 | if (modname[0]) |
871751e2 AV |
4440 | seq_printf(m, " [%s]", modname); |
4441 | return; | |
4442 | } | |
4443 | #endif | |
4444 | seq_printf(m, "%p", (void *)address); | |
4445 | } | |
4446 | ||
4447 | static int leaks_show(struct seq_file *m, void *p) | |
4448 | { | |
b92151ba | 4449 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4450 | struct slab *slabp; |
4451 | struct kmem_list3 *l3; | |
4452 | const char *name; | |
4453 | unsigned long *n = m->private; | |
4454 | int node; | |
4455 | int i; | |
4456 | ||
4457 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4458 | return 0; | |
4459 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4460 | return 0; | |
4461 | ||
4462 | /* OK, we can do it */ | |
4463 | ||
4464 | n[1] = 0; | |
4465 | ||
4466 | for_each_online_node(node) { | |
4467 | l3 = cachep->nodelists[node]; | |
4468 | if (!l3) | |
4469 | continue; | |
4470 | ||
4471 | check_irq_on(); | |
4472 | spin_lock_irq(&l3->list_lock); | |
4473 | ||
7a7c381d | 4474 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4475 | handle_slab(n, cachep, slabp); |
7a7c381d | 4476 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4477 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4478 | spin_unlock_irq(&l3->list_lock); |
4479 | } | |
4480 | name = cachep->name; | |
4481 | if (n[0] == n[1]) { | |
4482 | /* Increase the buffer size */ | |
4483 | mutex_unlock(&cache_chain_mutex); | |
4484 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4485 | if (!m->private) { | |
4486 | /* Too bad, we are really out */ | |
4487 | m->private = n; | |
4488 | mutex_lock(&cache_chain_mutex); | |
4489 | return -ENOMEM; | |
4490 | } | |
4491 | *(unsigned long *)m->private = n[0] * 2; | |
4492 | kfree(n); | |
4493 | mutex_lock(&cache_chain_mutex); | |
4494 | /* Now make sure this entry will be retried */ | |
4495 | m->count = m->size; | |
4496 | return 0; | |
4497 | } | |
4498 | for (i = 0; i < n[1]; i++) { | |
4499 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4500 | show_symbol(m, n[2*i+2]); | |
4501 | seq_putc(m, '\n'); | |
4502 | } | |
d2e7b7d0 | 4503 | |
871751e2 AV |
4504 | return 0; |
4505 | } | |
4506 | ||
a0ec95a8 | 4507 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4508 | .start = leaks_start, |
4509 | .next = s_next, | |
4510 | .stop = s_stop, | |
4511 | .show = leaks_show, | |
4512 | }; | |
a0ec95a8 AD |
4513 | |
4514 | static int slabstats_open(struct inode *inode, struct file *file) | |
4515 | { | |
4516 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4517 | int ret = -ENOMEM; | |
4518 | if (n) { | |
4519 | ret = seq_open(file, &slabstats_op); | |
4520 | if (!ret) { | |
4521 | struct seq_file *m = file->private_data; | |
4522 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4523 | m->private = n; | |
4524 | n = NULL; | |
4525 | } | |
4526 | kfree(n); | |
4527 | } | |
4528 | return ret; | |
4529 | } | |
4530 | ||
4531 | static const struct file_operations proc_slabstats_operations = { | |
4532 | .open = slabstats_open, | |
4533 | .read = seq_read, | |
4534 | .llseek = seq_lseek, | |
4535 | .release = seq_release_private, | |
4536 | }; | |
4537 | #endif | |
4538 | ||
4539 | static int __init slab_proc_init(void) | |
4540 | { | |
7b3c3a50 | 4541 | proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4542 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4543 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4544 | #endif |
a0ec95a8 AD |
4545 | return 0; |
4546 | } | |
4547 | module_init(slab_proc_init); | |
1da177e4 LT |
4548 | #endif |
4549 | ||
00e145b6 MS |
4550 | /** |
4551 | * ksize - get the actual amount of memory allocated for a given object | |
4552 | * @objp: Pointer to the object | |
4553 | * | |
4554 | * kmalloc may internally round up allocations and return more memory | |
4555 | * than requested. ksize() can be used to determine the actual amount of | |
4556 | * memory allocated. The caller may use this additional memory, even though | |
4557 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4558 | * The caller must guarantee that objp points to a valid object previously | |
4559 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4560 | * must not be freed during the duration of the call. | |
4561 | */ | |
fd76bab2 | 4562 | size_t ksize(const void *objp) |
1da177e4 | 4563 | { |
ef8b4520 CL |
4564 | BUG_ON(!objp); |
4565 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4566 | return 0; |
1da177e4 | 4567 | |
6ed5eb22 | 4568 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4569 | } |
b1aabecd | 4570 | EXPORT_SYMBOL(ksize); |