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 | |
29 | * slabs and you must pass objects with the same intializations to | |
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 | ||
89 | #include <linux/config.h> | |
90 | #include <linux/slab.h> | |
91 | #include <linux/mm.h> | |
c9cf5528 | 92 | #include <linux/poison.h> |
1da177e4 LT |
93 | #include <linux/swap.h> |
94 | #include <linux/cache.h> | |
95 | #include <linux/interrupt.h> | |
96 | #include <linux/init.h> | |
97 | #include <linux/compiler.h> | |
101a5001 | 98 | #include <linux/cpuset.h> |
1da177e4 LT |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> | |
101 | #include <linux/kallsyms.h> | |
102 | #include <linux/cpu.h> | |
103 | #include <linux/sysctl.h> | |
104 | #include <linux/module.h> | |
105 | #include <linux/rcupdate.h> | |
543537bd | 106 | #include <linux/string.h> |
e498be7d | 107 | #include <linux/nodemask.h> |
dc85da15 | 108 | #include <linux/mempolicy.h> |
fc0abb14 | 109 | #include <linux/mutex.h> |
e7eebaf6 | 110 | #include <linux/rtmutex.h> |
1da177e4 LT |
111 | |
112 | #include <asm/uaccess.h> | |
113 | #include <asm/cacheflush.h> | |
114 | #include <asm/tlbflush.h> | |
115 | #include <asm/page.h> | |
116 | ||
117 | /* | |
118 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
119 | * SLAB_RED_ZONE & SLAB_POISON. | |
120 | * 0 for faster, smaller code (especially in the critical paths). | |
121 | * | |
122 | * STATS - 1 to collect stats for /proc/slabinfo. | |
123 | * 0 for faster, smaller code (especially in the critical paths). | |
124 | * | |
125 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
126 | */ | |
127 | ||
128 | #ifdef CONFIG_DEBUG_SLAB | |
129 | #define DEBUG 1 | |
130 | #define STATS 1 | |
131 | #define FORCED_DEBUG 1 | |
132 | #else | |
133 | #define DEBUG 0 | |
134 | #define STATS 0 | |
135 | #define FORCED_DEBUG 0 | |
136 | #endif | |
137 | ||
1da177e4 LT |
138 | /* Shouldn't this be in a header file somewhere? */ |
139 | #define BYTES_PER_WORD sizeof(void *) | |
140 | ||
141 | #ifndef cache_line_size | |
142 | #define cache_line_size() L1_CACHE_BYTES | |
143 | #endif | |
144 | ||
145 | #ifndef ARCH_KMALLOC_MINALIGN | |
146 | /* | |
147 | * Enforce a minimum alignment for the kmalloc caches. | |
148 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
149 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
150 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
151 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
152 | * Note that this flag disables some debug features. | |
153 | */ | |
154 | #define ARCH_KMALLOC_MINALIGN 0 | |
155 | #endif | |
156 | ||
157 | #ifndef ARCH_SLAB_MINALIGN | |
158 | /* | |
159 | * Enforce a minimum alignment for all caches. | |
160 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
161 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
162 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
163 | * some debug features. | |
164 | */ | |
165 | #define ARCH_SLAB_MINALIGN 0 | |
166 | #endif | |
167 | ||
168 | #ifndef ARCH_KMALLOC_FLAGS | |
169 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
170 | #endif | |
171 | ||
172 | /* Legal flag mask for kmem_cache_create(). */ | |
173 | #if DEBUG | |
174 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
175 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
ac2b898c | 176 | SLAB_CACHE_DMA | \ |
1da177e4 LT |
177 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ |
178 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 179 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 180 | #else |
ac2b898c | 181 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
1da177e4 LT |
182 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ |
183 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 184 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 LT |
185 | #endif |
186 | ||
187 | /* | |
188 | * kmem_bufctl_t: | |
189 | * | |
190 | * Bufctl's are used for linking objs within a slab | |
191 | * linked offsets. | |
192 | * | |
193 | * This implementation relies on "struct page" for locating the cache & | |
194 | * slab an object belongs to. | |
195 | * This allows the bufctl structure to be small (one int), but limits | |
196 | * the number of objects a slab (not a cache) can contain when off-slab | |
197 | * bufctls are used. The limit is the size of the largest general cache | |
198 | * that does not use off-slab slabs. | |
199 | * For 32bit archs with 4 kB pages, is this 56. | |
200 | * This is not serious, as it is only for large objects, when it is unwise | |
201 | * to have too many per slab. | |
202 | * Note: This limit can be raised by introducing a general cache whose size | |
203 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
204 | */ | |
205 | ||
fa5b08d5 | 206 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
207 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
208 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
209 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
210 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 211 | |
1da177e4 LT |
212 | /* |
213 | * struct slab | |
214 | * | |
215 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
216 | * for a slab, or allocated from an general cache. | |
217 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
218 | */ | |
219 | struct slab { | |
b28a02de PE |
220 | struct list_head list; |
221 | unsigned long colouroff; | |
222 | void *s_mem; /* including colour offset */ | |
223 | unsigned int inuse; /* num of objs active in slab */ | |
224 | kmem_bufctl_t free; | |
225 | unsigned short nodeid; | |
1da177e4 LT |
226 | }; |
227 | ||
228 | /* | |
229 | * struct slab_rcu | |
230 | * | |
231 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
232 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
233 | * we need to approach a kernel structure obliquely, from its address | |
234 | * obtained without the usual locking. We can lock the structure to | |
235 | * stabilize it and check it's still at the given address, only if we | |
236 | * can be sure that the memory has not been meanwhile reused for some | |
237 | * other kind of object (which our subsystem's lock might corrupt). | |
238 | * | |
239 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
240 | * taking the spinlock within the structure expected at that address. | |
241 | * | |
242 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
243 | */ | |
244 | struct slab_rcu { | |
b28a02de | 245 | struct rcu_head head; |
343e0d7a | 246 | struct kmem_cache *cachep; |
b28a02de | 247 | void *addr; |
1da177e4 LT |
248 | }; |
249 | ||
250 | /* | |
251 | * struct array_cache | |
252 | * | |
1da177e4 LT |
253 | * Purpose: |
254 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
255 | * - reduce the number of linked list operations | |
256 | * - reduce spinlock operations | |
257 | * | |
258 | * The limit is stored in the per-cpu structure to reduce the data cache | |
259 | * footprint. | |
260 | * | |
261 | */ | |
262 | struct array_cache { | |
263 | unsigned int avail; | |
264 | unsigned int limit; | |
265 | unsigned int batchcount; | |
266 | unsigned int touched; | |
e498be7d | 267 | spinlock_t lock; |
a737b3e2 AM |
268 | void *entry[0]; /* |
269 | * Must have this definition in here for the proper | |
270 | * alignment of array_cache. Also simplifies accessing | |
271 | * the entries. | |
272 | * [0] is for gcc 2.95. It should really be []. | |
273 | */ | |
1da177e4 LT |
274 | }; |
275 | ||
a737b3e2 AM |
276 | /* |
277 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
278 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
279 | */ |
280 | #define BOOT_CPUCACHE_ENTRIES 1 | |
281 | struct arraycache_init { | |
282 | struct array_cache cache; | |
b28a02de | 283 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
284 | }; |
285 | ||
286 | /* | |
e498be7d | 287 | * The slab lists for all objects. |
1da177e4 LT |
288 | */ |
289 | struct kmem_list3 { | |
b28a02de PE |
290 | struct list_head slabs_partial; /* partial list first, better asm code */ |
291 | struct list_head slabs_full; | |
292 | struct list_head slabs_free; | |
293 | unsigned long free_objects; | |
b28a02de | 294 | unsigned int free_limit; |
2e1217cf | 295 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
296 | spinlock_t list_lock; |
297 | struct array_cache *shared; /* shared per node */ | |
298 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
299 | unsigned long next_reap; /* updated without locking */ |
300 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
301 | }; |
302 | ||
e498be7d CL |
303 | /* |
304 | * Need this for bootstrapping a per node allocator. | |
305 | */ | |
306 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
307 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
308 | #define CACHE_CACHE 0 | |
309 | #define SIZE_AC 1 | |
310 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
311 | ||
ed11d9eb CL |
312 | static int drain_freelist(struct kmem_cache *cache, |
313 | struct kmem_list3 *l3, int tofree); | |
314 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
315 | int node); | |
2ed3a4ef | 316 | static int enable_cpucache(struct kmem_cache *cachep); |
ed11d9eb CL |
317 | static void cache_reap(void *unused); |
318 | ||
e498be7d | 319 | /* |
a737b3e2 AM |
320 | * This function must be completely optimized away if a constant is passed to |
321 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 322 | */ |
7243cc05 | 323 | static __always_inline int index_of(const size_t size) |
e498be7d | 324 | { |
5ec8a847 SR |
325 | extern void __bad_size(void); |
326 | ||
e498be7d CL |
327 | if (__builtin_constant_p(size)) { |
328 | int i = 0; | |
329 | ||
330 | #define CACHE(x) \ | |
331 | if (size <=x) \ | |
332 | return i; \ | |
333 | else \ | |
334 | i++; | |
335 | #include "linux/kmalloc_sizes.h" | |
336 | #undef CACHE | |
5ec8a847 | 337 | __bad_size(); |
7243cc05 | 338 | } else |
5ec8a847 | 339 | __bad_size(); |
e498be7d CL |
340 | return 0; |
341 | } | |
342 | ||
e0a42726 IM |
343 | static int slab_early_init = 1; |
344 | ||
e498be7d CL |
345 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
346 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 347 | |
5295a74c | 348 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
349 | { |
350 | INIT_LIST_HEAD(&parent->slabs_full); | |
351 | INIT_LIST_HEAD(&parent->slabs_partial); | |
352 | INIT_LIST_HEAD(&parent->slabs_free); | |
353 | parent->shared = NULL; | |
354 | parent->alien = NULL; | |
2e1217cf | 355 | parent->colour_next = 0; |
e498be7d CL |
356 | spin_lock_init(&parent->list_lock); |
357 | parent->free_objects = 0; | |
358 | parent->free_touched = 0; | |
359 | } | |
360 | ||
a737b3e2 AM |
361 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
362 | do { \ | |
363 | INIT_LIST_HEAD(listp); \ | |
364 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
365 | } while (0) |
366 | ||
a737b3e2 AM |
367 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
368 | do { \ | |
e498be7d CL |
369 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
370 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
371 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
372 | } while (0) | |
1da177e4 LT |
373 | |
374 | /* | |
343e0d7a | 375 | * struct kmem_cache |
1da177e4 LT |
376 | * |
377 | * manages a cache. | |
378 | */ | |
b28a02de | 379 | |
2109a2d1 | 380 | struct kmem_cache { |
1da177e4 | 381 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 382 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 383 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
384 | unsigned int batchcount; |
385 | unsigned int limit; | |
386 | unsigned int shared; | |
b5d8ca7c | 387 | |
3dafccf2 | 388 | unsigned int buffer_size; |
b5d8ca7c | 389 | /* 3) touched by every alloc & free from the backend */ |
b28a02de | 390 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
b5d8ca7c | 391 | |
a737b3e2 AM |
392 | unsigned int flags; /* constant flags */ |
393 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 394 | |
b5d8ca7c | 395 | /* 4) cache_grow/shrink */ |
1da177e4 | 396 | /* order of pgs per slab (2^n) */ |
b28a02de | 397 | unsigned int gfporder; |
1da177e4 LT |
398 | |
399 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 400 | gfp_t gfpflags; |
1da177e4 | 401 | |
a737b3e2 | 402 | size_t colour; /* cache colouring range */ |
b28a02de | 403 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 404 | struct kmem_cache *slabp_cache; |
b28a02de | 405 | unsigned int slab_size; |
a737b3e2 | 406 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
407 | |
408 | /* constructor func */ | |
343e0d7a | 409 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
410 | |
411 | /* de-constructor func */ | |
343e0d7a | 412 | void (*dtor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 | 413 | |
b5d8ca7c | 414 | /* 5) cache creation/removal */ |
b28a02de PE |
415 | const char *name; |
416 | struct list_head next; | |
1da177e4 | 417 | |
b5d8ca7c | 418 | /* 6) statistics */ |
1da177e4 | 419 | #if STATS |
b28a02de PE |
420 | unsigned long num_active; |
421 | unsigned long num_allocations; | |
422 | unsigned long high_mark; | |
423 | unsigned long grown; | |
424 | unsigned long reaped; | |
425 | unsigned long errors; | |
426 | unsigned long max_freeable; | |
427 | unsigned long node_allocs; | |
428 | unsigned long node_frees; | |
fb7faf33 | 429 | unsigned long node_overflow; |
b28a02de PE |
430 | atomic_t allochit; |
431 | atomic_t allocmiss; | |
432 | atomic_t freehit; | |
433 | atomic_t freemiss; | |
1da177e4 LT |
434 | #endif |
435 | #if DEBUG | |
3dafccf2 MS |
436 | /* |
437 | * If debugging is enabled, then the allocator can add additional | |
438 | * fields and/or padding to every object. buffer_size contains the total | |
439 | * object size including these internal fields, the following two | |
440 | * variables contain the offset to the user object and its size. | |
441 | */ | |
442 | int obj_offset; | |
443 | int obj_size; | |
1da177e4 LT |
444 | #endif |
445 | }; | |
446 | ||
447 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
448 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
449 | ||
450 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
451 | /* |
452 | * Optimization question: fewer reaps means less probability for unnessary | |
453 | * cpucache drain/refill cycles. | |
1da177e4 | 454 | * |
dc6f3f27 | 455 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
456 | * which could lock up otherwise freeable slabs. |
457 | */ | |
458 | #define REAPTIMEOUT_CPUC (2*HZ) | |
459 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
460 | ||
461 | #if STATS | |
462 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
463 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
464 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
465 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 466 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
467 | #define STATS_SET_HIGH(x) \ |
468 | do { \ | |
469 | if ((x)->num_active > (x)->high_mark) \ | |
470 | (x)->high_mark = (x)->num_active; \ | |
471 | } while (0) | |
1da177e4 LT |
472 | #define STATS_INC_ERR(x) ((x)->errors++) |
473 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 474 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 475 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
476 | #define STATS_SET_FREEABLE(x, i) \ |
477 | do { \ | |
478 | if ((x)->max_freeable < i) \ | |
479 | (x)->max_freeable = i; \ | |
480 | } while (0) | |
1da177e4 LT |
481 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
482 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
483 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
484 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
485 | #else | |
486 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
487 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
488 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
489 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 490 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
491 | #define STATS_SET_HIGH(x) do { } while (0) |
492 | #define STATS_INC_ERR(x) do { } while (0) | |
493 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 494 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 495 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 496 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
497 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
498 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
499 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
500 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
501 | #endif | |
502 | ||
503 | #if DEBUG | |
1da177e4 | 504 | |
a737b3e2 AM |
505 | /* |
506 | * memory layout of objects: | |
1da177e4 | 507 | * 0 : objp |
3dafccf2 | 508 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
509 | * the end of an object is aligned with the end of the real |
510 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 511 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 512 | * redzone word. |
3dafccf2 MS |
513 | * cachep->obj_offset: The real object. |
514 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
515 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
516 | * [BYTES_PER_WORD long] | |
1da177e4 | 517 | */ |
343e0d7a | 518 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 519 | { |
3dafccf2 | 520 | return cachep->obj_offset; |
1da177e4 LT |
521 | } |
522 | ||
343e0d7a | 523 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 524 | { |
3dafccf2 | 525 | return cachep->obj_size; |
1da177e4 LT |
526 | } |
527 | ||
343e0d7a | 528 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
529 | { |
530 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 531 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
532 | } |
533 | ||
343e0d7a | 534 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
535 | { |
536 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
537 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 538 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 539 | 2 * BYTES_PER_WORD); |
3dafccf2 | 540 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
541 | } |
542 | ||
343e0d7a | 543 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
544 | { |
545 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 546 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
547 | } |
548 | ||
549 | #else | |
550 | ||
3dafccf2 MS |
551 | #define obj_offset(x) 0 |
552 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
553 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
554 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
555 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
556 | ||
557 | #endif | |
558 | ||
559 | /* | |
a737b3e2 AM |
560 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
561 | * order. | |
1da177e4 LT |
562 | */ |
563 | #if defined(CONFIG_LARGE_ALLOCS) | |
564 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
565 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
566 | #elif defined(CONFIG_MMU) | |
567 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
568 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
569 | #else | |
570 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
571 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
572 | #endif | |
573 | ||
574 | /* | |
575 | * Do not go above this order unless 0 objects fit into the slab. | |
576 | */ | |
577 | #define BREAK_GFP_ORDER_HI 1 | |
578 | #define BREAK_GFP_ORDER_LO 0 | |
579 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
580 | ||
a737b3e2 AM |
581 | /* |
582 | * Functions for storing/retrieving the cachep and or slab from the page | |
583 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
584 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 585 | */ |
065d41cb PE |
586 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
587 | { | |
588 | page->lru.next = (struct list_head *)cache; | |
589 | } | |
590 | ||
591 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
592 | { | |
84097518 NP |
593 | if (unlikely(PageCompound(page))) |
594 | page = (struct page *)page_private(page); | |
ddc2e812 | 595 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
596 | return (struct kmem_cache *)page->lru.next; |
597 | } | |
598 | ||
599 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
600 | { | |
601 | page->lru.prev = (struct list_head *)slab; | |
602 | } | |
603 | ||
604 | static inline struct slab *page_get_slab(struct page *page) | |
605 | { | |
84097518 NP |
606 | if (unlikely(PageCompound(page))) |
607 | page = (struct page *)page_private(page); | |
ddc2e812 | 608 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
609 | return (struct slab *)page->lru.prev; |
610 | } | |
1da177e4 | 611 | |
6ed5eb22 PE |
612 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
613 | { | |
614 | struct page *page = virt_to_page(obj); | |
615 | return page_get_cache(page); | |
616 | } | |
617 | ||
618 | static inline struct slab *virt_to_slab(const void *obj) | |
619 | { | |
620 | struct page *page = virt_to_page(obj); | |
621 | return page_get_slab(page); | |
622 | } | |
623 | ||
8fea4e96 PE |
624 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
625 | unsigned int idx) | |
626 | { | |
627 | return slab->s_mem + cache->buffer_size * idx; | |
628 | } | |
629 | ||
630 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
631 | struct slab *slab, void *obj) | |
632 | { | |
633 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
634 | } | |
635 | ||
a737b3e2 AM |
636 | /* |
637 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
638 | */ | |
1da177e4 LT |
639 | struct cache_sizes malloc_sizes[] = { |
640 | #define CACHE(x) { .cs_size = (x) }, | |
641 | #include <linux/kmalloc_sizes.h> | |
642 | CACHE(ULONG_MAX) | |
643 | #undef CACHE | |
644 | }; | |
645 | EXPORT_SYMBOL(malloc_sizes); | |
646 | ||
647 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
648 | struct cache_names { | |
649 | char *name; | |
650 | char *name_dma; | |
651 | }; | |
652 | ||
653 | static struct cache_names __initdata cache_names[] = { | |
654 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
655 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 656 | {NULL,} |
1da177e4 LT |
657 | #undef CACHE |
658 | }; | |
659 | ||
660 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 661 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 662 | static struct arraycache_init initarray_generic = |
b28a02de | 663 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
664 | |
665 | /* internal cache of cache description objs */ | |
343e0d7a | 666 | static struct kmem_cache cache_cache = { |
b28a02de PE |
667 | .batchcount = 1, |
668 | .limit = BOOT_CPUCACHE_ENTRIES, | |
669 | .shared = 1, | |
343e0d7a | 670 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 671 | .name = "kmem_cache", |
1da177e4 | 672 | #if DEBUG |
343e0d7a | 673 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
674 | #endif |
675 | }; | |
676 | ||
f1aaee53 AV |
677 | #ifdef CONFIG_LOCKDEP |
678 | ||
679 | /* | |
680 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
681 | * for other slabs "off slab". | |
682 | * The locking for this is tricky in that it nests within the locks | |
683 | * of all other slabs in a few places; to deal with this special | |
684 | * locking we put on-slab caches into a separate lock-class. | |
685 | */ | |
686 | static struct lock_class_key on_slab_key; | |
687 | ||
688 | static inline void init_lock_keys(struct cache_sizes *s) | |
689 | { | |
690 | int q; | |
691 | ||
692 | for (q = 0; q < MAX_NUMNODES; q++) { | |
693 | if (!s->cs_cachep->nodelists[q] || OFF_SLAB(s->cs_cachep)) | |
694 | continue; | |
695 | lockdep_set_class(&s->cs_cachep->nodelists[q]->list_lock, | |
696 | &on_slab_key); | |
697 | } | |
698 | } | |
699 | ||
700 | #else | |
701 | static inline void init_lock_keys(struct cache_sizes *s) | |
702 | { | |
703 | } | |
704 | #endif | |
705 | ||
706 | ||
707 | ||
1da177e4 | 708 | /* Guard access to the cache-chain. */ |
fc0abb14 | 709 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
710 | static struct list_head cache_chain; |
711 | ||
712 | /* | |
a737b3e2 AM |
713 | * vm_enough_memory() looks at this to determine how many slab-allocated pages |
714 | * are possibly freeable under pressure | |
1da177e4 LT |
715 | * |
716 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
717 | */ | |
718 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
719 | |
720 | /* | |
721 | * chicken and egg problem: delay the per-cpu array allocation | |
722 | * until the general caches are up. | |
723 | */ | |
724 | static enum { | |
725 | NONE, | |
e498be7d CL |
726 | PARTIAL_AC, |
727 | PARTIAL_L3, | |
1da177e4 LT |
728 | FULL |
729 | } g_cpucache_up; | |
730 | ||
39d24e64 MK |
731 | /* |
732 | * used by boot code to determine if it can use slab based allocator | |
733 | */ | |
734 | int slab_is_available(void) | |
735 | { | |
736 | return g_cpucache_up == FULL; | |
737 | } | |
738 | ||
1da177e4 LT |
739 | static DEFINE_PER_CPU(struct work_struct, reap_work); |
740 | ||
343e0d7a | 741 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
742 | { |
743 | return cachep->array[smp_processor_id()]; | |
744 | } | |
745 | ||
a737b3e2 AM |
746 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
747 | gfp_t gfpflags) | |
1da177e4 LT |
748 | { |
749 | struct cache_sizes *csizep = malloc_sizes; | |
750 | ||
751 | #if DEBUG | |
752 | /* This happens if someone tries to call | |
b28a02de PE |
753 | * kmem_cache_create(), or __kmalloc(), before |
754 | * the generic caches are initialized. | |
755 | */ | |
c7e43c78 | 756 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
757 | #endif |
758 | while (size > csizep->cs_size) | |
759 | csizep++; | |
760 | ||
761 | /* | |
0abf40c1 | 762 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
763 | * has cs_{dma,}cachep==NULL. Thus no special case |
764 | * for large kmalloc calls required. | |
765 | */ | |
766 | if (unlikely(gfpflags & GFP_DMA)) | |
767 | return csizep->cs_dmacachep; | |
768 | return csizep->cs_cachep; | |
769 | } | |
770 | ||
b221385b | 771 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
772 | { |
773 | return __find_general_cachep(size, gfpflags); | |
774 | } | |
97e2bde4 | 775 | |
fbaccacf | 776 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 777 | { |
fbaccacf SR |
778 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
779 | } | |
1da177e4 | 780 | |
a737b3e2 AM |
781 | /* |
782 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
783 | */ | |
fbaccacf SR |
784 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
785 | size_t align, int flags, size_t *left_over, | |
786 | unsigned int *num) | |
787 | { | |
788 | int nr_objs; | |
789 | size_t mgmt_size; | |
790 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 791 | |
fbaccacf SR |
792 | /* |
793 | * The slab management structure can be either off the slab or | |
794 | * on it. For the latter case, the memory allocated for a | |
795 | * slab is used for: | |
796 | * | |
797 | * - The struct slab | |
798 | * - One kmem_bufctl_t for each object | |
799 | * - Padding to respect alignment of @align | |
800 | * - @buffer_size bytes for each object | |
801 | * | |
802 | * If the slab management structure is off the slab, then the | |
803 | * alignment will already be calculated into the size. Because | |
804 | * the slabs are all pages aligned, the objects will be at the | |
805 | * correct alignment when allocated. | |
806 | */ | |
807 | if (flags & CFLGS_OFF_SLAB) { | |
808 | mgmt_size = 0; | |
809 | nr_objs = slab_size / buffer_size; | |
810 | ||
811 | if (nr_objs > SLAB_LIMIT) | |
812 | nr_objs = SLAB_LIMIT; | |
813 | } else { | |
814 | /* | |
815 | * Ignore padding for the initial guess. The padding | |
816 | * is at most @align-1 bytes, and @buffer_size is at | |
817 | * least @align. In the worst case, this result will | |
818 | * be one greater than the number of objects that fit | |
819 | * into the memory allocation when taking the padding | |
820 | * into account. | |
821 | */ | |
822 | nr_objs = (slab_size - sizeof(struct slab)) / | |
823 | (buffer_size + sizeof(kmem_bufctl_t)); | |
824 | ||
825 | /* | |
826 | * This calculated number will be either the right | |
827 | * amount, or one greater than what we want. | |
828 | */ | |
829 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
830 | > slab_size) | |
831 | nr_objs--; | |
832 | ||
833 | if (nr_objs > SLAB_LIMIT) | |
834 | nr_objs = SLAB_LIMIT; | |
835 | ||
836 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
837 | } | |
838 | *num = nr_objs; | |
839 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
840 | } |
841 | ||
842 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
843 | ||
a737b3e2 AM |
844 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
845 | char *msg) | |
1da177e4 LT |
846 | { |
847 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 848 | function, cachep->name, msg); |
1da177e4 LT |
849 | dump_stack(); |
850 | } | |
851 | ||
8fce4d8e CL |
852 | #ifdef CONFIG_NUMA |
853 | /* | |
854 | * Special reaping functions for NUMA systems called from cache_reap(). | |
855 | * These take care of doing round robin flushing of alien caches (containing | |
856 | * objects freed on different nodes from which they were allocated) and the | |
857 | * flushing of remote pcps by calling drain_node_pages. | |
858 | */ | |
859 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
860 | ||
861 | static void init_reap_node(int cpu) | |
862 | { | |
863 | int node; | |
864 | ||
865 | node = next_node(cpu_to_node(cpu), node_online_map); | |
866 | if (node == MAX_NUMNODES) | |
442295c9 | 867 | node = first_node(node_online_map); |
8fce4d8e CL |
868 | |
869 | __get_cpu_var(reap_node) = node; | |
870 | } | |
871 | ||
872 | static void next_reap_node(void) | |
873 | { | |
874 | int node = __get_cpu_var(reap_node); | |
875 | ||
876 | /* | |
877 | * Also drain per cpu pages on remote zones | |
878 | */ | |
879 | if (node != numa_node_id()) | |
880 | drain_node_pages(node); | |
881 | ||
882 | node = next_node(node, node_online_map); | |
883 | if (unlikely(node >= MAX_NUMNODES)) | |
884 | node = first_node(node_online_map); | |
885 | __get_cpu_var(reap_node) = node; | |
886 | } | |
887 | ||
888 | #else | |
889 | #define init_reap_node(cpu) do { } while (0) | |
890 | #define next_reap_node(void) do { } while (0) | |
891 | #endif | |
892 | ||
1da177e4 LT |
893 | /* |
894 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
895 | * via the workqueue/eventd. | |
896 | * Add the CPU number into the expiration time to minimize the possibility of | |
897 | * the CPUs getting into lockstep and contending for the global cache chain | |
898 | * lock. | |
899 | */ | |
900 | static void __devinit start_cpu_timer(int cpu) | |
901 | { | |
902 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
903 | ||
904 | /* | |
905 | * When this gets called from do_initcalls via cpucache_init(), | |
906 | * init_workqueues() has already run, so keventd will be setup | |
907 | * at that time. | |
908 | */ | |
909 | if (keventd_up() && reap_work->func == NULL) { | |
8fce4d8e | 910 | init_reap_node(cpu); |
1da177e4 LT |
911 | INIT_WORK(reap_work, cache_reap, NULL); |
912 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
913 | } | |
914 | } | |
915 | ||
e498be7d | 916 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 917 | int batchcount) |
1da177e4 | 918 | { |
b28a02de | 919 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
920 | struct array_cache *nc = NULL; |
921 | ||
e498be7d | 922 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
923 | if (nc) { |
924 | nc->avail = 0; | |
925 | nc->limit = entries; | |
926 | nc->batchcount = batchcount; | |
927 | nc->touched = 0; | |
e498be7d | 928 | spin_lock_init(&nc->lock); |
1da177e4 LT |
929 | } |
930 | return nc; | |
931 | } | |
932 | ||
3ded175a CL |
933 | /* |
934 | * Transfer objects in one arraycache to another. | |
935 | * Locking must be handled by the caller. | |
936 | * | |
937 | * Return the number of entries transferred. | |
938 | */ | |
939 | static int transfer_objects(struct array_cache *to, | |
940 | struct array_cache *from, unsigned int max) | |
941 | { | |
942 | /* Figure out how many entries to transfer */ | |
943 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
944 | ||
945 | if (!nr) | |
946 | return 0; | |
947 | ||
948 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
949 | sizeof(void *) *nr); | |
950 | ||
951 | from->avail -= nr; | |
952 | to->avail += nr; | |
953 | to->touched = 1; | |
954 | return nr; | |
955 | } | |
956 | ||
e498be7d | 957 | #ifdef CONFIG_NUMA |
343e0d7a | 958 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 959 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 960 | |
5295a74c | 961 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
962 | { |
963 | struct array_cache **ac_ptr; | |
b28a02de | 964 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
965 | int i; |
966 | ||
967 | if (limit > 1) | |
968 | limit = 12; | |
969 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
970 | if (ac_ptr) { | |
971 | for_each_node(i) { | |
972 | if (i == node || !node_online(i)) { | |
973 | ac_ptr[i] = NULL; | |
974 | continue; | |
975 | } | |
976 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
977 | if (!ac_ptr[i]) { | |
b28a02de | 978 | for (i--; i <= 0; i--) |
e498be7d CL |
979 | kfree(ac_ptr[i]); |
980 | kfree(ac_ptr); | |
981 | return NULL; | |
982 | } | |
983 | } | |
984 | } | |
985 | return ac_ptr; | |
986 | } | |
987 | ||
5295a74c | 988 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
989 | { |
990 | int i; | |
991 | ||
992 | if (!ac_ptr) | |
993 | return; | |
e498be7d | 994 | for_each_node(i) |
b28a02de | 995 | kfree(ac_ptr[i]); |
e498be7d CL |
996 | kfree(ac_ptr); |
997 | } | |
998 | ||
343e0d7a | 999 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1000 | struct array_cache *ac, int node) |
e498be7d CL |
1001 | { |
1002 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1003 | ||
1004 | if (ac->avail) { | |
1005 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1006 | /* |
1007 | * Stuff objects into the remote nodes shared array first. | |
1008 | * That way we could avoid the overhead of putting the objects | |
1009 | * into the free lists and getting them back later. | |
1010 | */ | |
693f7d36 | 1011 | if (rl3->shared) |
1012 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1013 | |
ff69416e | 1014 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1015 | ac->avail = 0; |
1016 | spin_unlock(&rl3->list_lock); | |
1017 | } | |
1018 | } | |
1019 | ||
8fce4d8e CL |
1020 | /* |
1021 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1022 | */ | |
1023 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1024 | { | |
1025 | int node = __get_cpu_var(reap_node); | |
1026 | ||
1027 | if (l3->alien) { | |
1028 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1029 | |
1030 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1031 | __drain_alien_cache(cachep, ac, node); |
1032 | spin_unlock_irq(&ac->lock); | |
1033 | } | |
1034 | } | |
1035 | } | |
1036 | ||
a737b3e2 AM |
1037 | static void drain_alien_cache(struct kmem_cache *cachep, |
1038 | struct array_cache **alien) | |
e498be7d | 1039 | { |
b28a02de | 1040 | int i = 0; |
e498be7d CL |
1041 | struct array_cache *ac; |
1042 | unsigned long flags; | |
1043 | ||
1044 | for_each_online_node(i) { | |
4484ebf1 | 1045 | ac = alien[i]; |
e498be7d CL |
1046 | if (ac) { |
1047 | spin_lock_irqsave(&ac->lock, flags); | |
1048 | __drain_alien_cache(cachep, ac, i); | |
1049 | spin_unlock_irqrestore(&ac->lock, flags); | |
1050 | } | |
1051 | } | |
1052 | } | |
729bd0b7 | 1053 | |
873623df | 1054 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1055 | { |
1056 | struct slab *slabp = virt_to_slab(objp); | |
1057 | int nodeid = slabp->nodeid; | |
1058 | struct kmem_list3 *l3; | |
1059 | struct array_cache *alien = NULL; | |
1060 | ||
1061 | /* | |
1062 | * Make sure we are not freeing a object from another node to the array | |
1063 | * cache on this cpu. | |
1064 | */ | |
1065 | if (likely(slabp->nodeid == numa_node_id())) | |
1066 | return 0; | |
1067 | ||
1068 | l3 = cachep->nodelists[numa_node_id()]; | |
1069 | STATS_INC_NODEFREES(cachep); | |
1070 | if (l3->alien && l3->alien[nodeid]) { | |
1071 | alien = l3->alien[nodeid]; | |
873623df | 1072 | spin_lock(&alien->lock); |
729bd0b7 PE |
1073 | if (unlikely(alien->avail == alien->limit)) { |
1074 | STATS_INC_ACOVERFLOW(cachep); | |
1075 | __drain_alien_cache(cachep, alien, nodeid); | |
1076 | } | |
1077 | alien->entry[alien->avail++] = objp; | |
1078 | spin_unlock(&alien->lock); | |
1079 | } else { | |
1080 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1081 | free_block(cachep, &objp, 1, nodeid); | |
1082 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1083 | } | |
1084 | return 1; | |
1085 | } | |
1086 | ||
e498be7d | 1087 | #else |
7a21ef6f | 1088 | |
4484ebf1 | 1089 | #define drain_alien_cache(cachep, alien) do { } while (0) |
8fce4d8e | 1090 | #define reap_alien(cachep, l3) do { } while (0) |
4484ebf1 | 1091 | |
7a21ef6f LT |
1092 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
1093 | { | |
1094 | return (struct array_cache **) 0x01020304ul; | |
1095 | } | |
1096 | ||
4484ebf1 RT |
1097 | static inline void free_alien_cache(struct array_cache **ac_ptr) |
1098 | { | |
1099 | } | |
7a21ef6f | 1100 | |
873623df | 1101 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1102 | { |
1103 | return 0; | |
1104 | } | |
1105 | ||
e498be7d CL |
1106 | #endif |
1107 | ||
8c78f307 | 1108 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1109 | unsigned long action, void *hcpu) |
1da177e4 LT |
1110 | { |
1111 | long cpu = (long)hcpu; | |
343e0d7a | 1112 | struct kmem_cache *cachep; |
e498be7d CL |
1113 | struct kmem_list3 *l3 = NULL; |
1114 | int node = cpu_to_node(cpu); | |
1115 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1116 | |
1117 | switch (action) { | |
1118 | case CPU_UP_PREPARE: | |
fc0abb14 | 1119 | mutex_lock(&cache_chain_mutex); |
a737b3e2 AM |
1120 | /* |
1121 | * We need to do this right in the beginning since | |
e498be7d CL |
1122 | * alloc_arraycache's are going to use this list. |
1123 | * kmalloc_node allows us to add the slab to the right | |
1124 | * kmem_list3 and not this cpu's kmem_list3 | |
1125 | */ | |
1126 | ||
1da177e4 | 1127 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1128 | /* |
1129 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1130 | * begin anything. Make sure some other cpu on this |
1131 | * node has not already allocated this | |
1132 | */ | |
1133 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1134 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1135 | if (!l3) | |
e498be7d CL |
1136 | goto bad; |
1137 | kmem_list3_init(l3); | |
1138 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1139 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1140 | |
4484ebf1 RT |
1141 | /* |
1142 | * The l3s don't come and go as CPUs come and | |
1143 | * go. cache_chain_mutex is sufficient | |
1144 | * protection here. | |
1145 | */ | |
e498be7d CL |
1146 | cachep->nodelists[node] = l3; |
1147 | } | |
1da177e4 | 1148 | |
e498be7d CL |
1149 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1150 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1151 | (1 + nr_cpus_node(node)) * |
1152 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1153 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1154 | } | |
1155 | ||
a737b3e2 AM |
1156 | /* |
1157 | * Now we can go ahead with allocating the shared arrays and | |
1158 | * array caches | |
1159 | */ | |
e498be7d | 1160 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1161 | struct array_cache *nc; |
4484ebf1 RT |
1162 | struct array_cache *shared; |
1163 | struct array_cache **alien; | |
cd105df4 | 1164 | |
e498be7d | 1165 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1166 | cachep->batchcount); |
1da177e4 LT |
1167 | if (!nc) |
1168 | goto bad; | |
4484ebf1 RT |
1169 | shared = alloc_arraycache(node, |
1170 | cachep->shared * cachep->batchcount, | |
1171 | 0xbaadf00d); | |
1172 | if (!shared) | |
1173 | goto bad; | |
7a21ef6f | 1174 | |
4484ebf1 RT |
1175 | alien = alloc_alien_cache(node, cachep->limit); |
1176 | if (!alien) | |
1177 | goto bad; | |
1da177e4 | 1178 | cachep->array[cpu] = nc; |
e498be7d CL |
1179 | l3 = cachep->nodelists[node]; |
1180 | BUG_ON(!l3); | |
e498be7d | 1181 | |
4484ebf1 RT |
1182 | spin_lock_irq(&l3->list_lock); |
1183 | if (!l3->shared) { | |
1184 | /* | |
1185 | * We are serialised from CPU_DEAD or | |
1186 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1187 | */ | |
1188 | l3->shared = shared; | |
1189 | shared = NULL; | |
e498be7d | 1190 | } |
4484ebf1 RT |
1191 | #ifdef CONFIG_NUMA |
1192 | if (!l3->alien) { | |
1193 | l3->alien = alien; | |
1194 | alien = NULL; | |
1195 | } | |
1196 | #endif | |
1197 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1198 | kfree(shared); |
1199 | free_alien_cache(alien); | |
1da177e4 | 1200 | } |
fc0abb14 | 1201 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1202 | break; |
1203 | case CPU_ONLINE: | |
1204 | start_cpu_timer(cpu); | |
1205 | break; | |
1206 | #ifdef CONFIG_HOTPLUG_CPU | |
1207 | case CPU_DEAD: | |
4484ebf1 RT |
1208 | /* |
1209 | * Even if all the cpus of a node are down, we don't free the | |
1210 | * kmem_list3 of any cache. This to avoid a race between | |
1211 | * cpu_down, and a kmalloc allocation from another cpu for | |
1212 | * memory from the node of the cpu going down. The list3 | |
1213 | * structure is usually allocated from kmem_cache_create() and | |
1214 | * gets destroyed at kmem_cache_destroy(). | |
1215 | */ | |
1da177e4 LT |
1216 | /* fall thru */ |
1217 | case CPU_UP_CANCELED: | |
fc0abb14 | 1218 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1219 | list_for_each_entry(cachep, &cache_chain, next) { |
1220 | struct array_cache *nc; | |
4484ebf1 RT |
1221 | struct array_cache *shared; |
1222 | struct array_cache **alien; | |
e498be7d | 1223 | cpumask_t mask; |
1da177e4 | 1224 | |
e498be7d | 1225 | mask = node_to_cpumask(node); |
1da177e4 LT |
1226 | /* cpu is dead; no one can alloc from it. */ |
1227 | nc = cachep->array[cpu]; | |
1228 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1229 | l3 = cachep->nodelists[node]; |
1230 | ||
1231 | if (!l3) | |
4484ebf1 | 1232 | goto free_array_cache; |
e498be7d | 1233 | |
ca3b9b91 | 1234 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1235 | |
1236 | /* Free limit for this kmem_list3 */ | |
1237 | l3->free_limit -= cachep->batchcount; | |
1238 | if (nc) | |
ff69416e | 1239 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1240 | |
1241 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1242 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1243 | goto free_array_cache; |
b28a02de | 1244 | } |
e498be7d | 1245 | |
4484ebf1 RT |
1246 | shared = l3->shared; |
1247 | if (shared) { | |
e498be7d | 1248 | free_block(cachep, l3->shared->entry, |
b28a02de | 1249 | l3->shared->avail, node); |
e498be7d CL |
1250 | l3->shared = NULL; |
1251 | } | |
e498be7d | 1252 | |
4484ebf1 RT |
1253 | alien = l3->alien; |
1254 | l3->alien = NULL; | |
1255 | ||
1256 | spin_unlock_irq(&l3->list_lock); | |
1257 | ||
1258 | kfree(shared); | |
1259 | if (alien) { | |
1260 | drain_alien_cache(cachep, alien); | |
1261 | free_alien_cache(alien); | |
e498be7d | 1262 | } |
4484ebf1 | 1263 | free_array_cache: |
1da177e4 LT |
1264 | kfree(nc); |
1265 | } | |
4484ebf1 RT |
1266 | /* |
1267 | * In the previous loop, all the objects were freed to | |
1268 | * the respective cache's slabs, now we can go ahead and | |
1269 | * shrink each nodelist to its limit. | |
1270 | */ | |
1271 | list_for_each_entry(cachep, &cache_chain, next) { | |
1272 | l3 = cachep->nodelists[node]; | |
1273 | if (!l3) | |
1274 | continue; | |
ed11d9eb | 1275 | drain_freelist(cachep, l3, l3->free_objects); |
4484ebf1 | 1276 | } |
fc0abb14 | 1277 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1278 | break; |
1279 | #endif | |
1280 | } | |
1281 | return NOTIFY_OK; | |
a737b3e2 | 1282 | bad: |
fc0abb14 | 1283 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1284 | return NOTIFY_BAD; |
1285 | } | |
1286 | ||
74b85f37 CS |
1287 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1288 | &cpuup_callback, NULL, 0 | |
1289 | }; | |
1da177e4 | 1290 | |
e498be7d CL |
1291 | /* |
1292 | * swap the static kmem_list3 with kmalloced memory | |
1293 | */ | |
a737b3e2 AM |
1294 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1295 | int nodeid) | |
e498be7d CL |
1296 | { |
1297 | struct kmem_list3 *ptr; | |
1298 | ||
1299 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1300 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1301 | BUG_ON(!ptr); | |
1302 | ||
1303 | local_irq_disable(); | |
1304 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1305 | /* |
1306 | * Do not assume that spinlocks can be initialized via memcpy: | |
1307 | */ | |
1308 | spin_lock_init(&ptr->list_lock); | |
1309 | ||
e498be7d CL |
1310 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1311 | cachep->nodelists[nodeid] = ptr; | |
1312 | local_irq_enable(); | |
1313 | } | |
1314 | ||
a737b3e2 AM |
1315 | /* |
1316 | * Initialisation. Called after the page allocator have been initialised and | |
1317 | * before smp_init(). | |
1da177e4 LT |
1318 | */ |
1319 | void __init kmem_cache_init(void) | |
1320 | { | |
1321 | size_t left_over; | |
1322 | struct cache_sizes *sizes; | |
1323 | struct cache_names *names; | |
e498be7d | 1324 | int i; |
07ed76b2 | 1325 | int order; |
e498be7d CL |
1326 | |
1327 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1328 | kmem_list3_init(&initkmem_list3[i]); | |
1329 | if (i < MAX_NUMNODES) | |
1330 | cache_cache.nodelists[i] = NULL; | |
1331 | } | |
1da177e4 LT |
1332 | |
1333 | /* | |
1334 | * Fragmentation resistance on low memory - only use bigger | |
1335 | * page orders on machines with more than 32MB of memory. | |
1336 | */ | |
1337 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1338 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1339 | ||
1da177e4 LT |
1340 | /* Bootstrap is tricky, because several objects are allocated |
1341 | * from caches that do not exist yet: | |
a737b3e2 AM |
1342 | * 1) initialize the cache_cache cache: it contains the struct |
1343 | * kmem_cache structures of all caches, except cache_cache itself: | |
1344 | * cache_cache is statically allocated. | |
e498be7d CL |
1345 | * Initially an __init data area is used for the head array and the |
1346 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1347 | * array at the end of the bootstrap. | |
1da177e4 | 1348 | * 2) Create the first kmalloc cache. |
343e0d7a | 1349 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1350 | * An __init data area is used for the head array. |
1351 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1352 | * head arrays. | |
1da177e4 LT |
1353 | * 4) Replace the __init data head arrays for cache_cache and the first |
1354 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1355 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1356 | * the other cache's with kmalloc allocated memory. | |
1357 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1358 | */ |
1359 | ||
1360 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1361 | INIT_LIST_HEAD(&cache_chain); |
1362 | list_add(&cache_cache.next, &cache_chain); | |
1363 | cache_cache.colour_off = cache_line_size(); | |
1364 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1365 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1366 | |
a737b3e2 AM |
1367 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1368 | cache_line_size()); | |
1da177e4 | 1369 | |
07ed76b2 JS |
1370 | for (order = 0; order < MAX_ORDER; order++) { |
1371 | cache_estimate(order, cache_cache.buffer_size, | |
1372 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1373 | if (cache_cache.num) | |
1374 | break; | |
1375 | } | |
40094fa6 | 1376 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1377 | cache_cache.gfporder = order; |
b28a02de | 1378 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1379 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1380 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1381 | |
1382 | /* 2+3) create the kmalloc caches */ | |
1383 | sizes = malloc_sizes; | |
1384 | names = cache_names; | |
1385 | ||
a737b3e2 AM |
1386 | /* |
1387 | * Initialize the caches that provide memory for the array cache and the | |
1388 | * kmem_list3 structures first. Without this, further allocations will | |
1389 | * bug. | |
e498be7d CL |
1390 | */ |
1391 | ||
1392 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1393 | sizes[INDEX_AC].cs_size, |
1394 | ARCH_KMALLOC_MINALIGN, | |
1395 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1396 | NULL, NULL); | |
e498be7d | 1397 | |
a737b3e2 | 1398 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1399 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1400 | kmem_cache_create(names[INDEX_L3].name, |
1401 | sizes[INDEX_L3].cs_size, | |
1402 | ARCH_KMALLOC_MINALIGN, | |
1403 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1404 | NULL, NULL); | |
1405 | } | |
e498be7d | 1406 | |
e0a42726 IM |
1407 | slab_early_init = 0; |
1408 | ||
1da177e4 | 1409 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1410 | /* |
1411 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1412 | * This should be particularly beneficial on SMP boxes, as it |
1413 | * eliminates "false sharing". | |
1414 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1415 | * allow tighter packing of the smaller caches. |
1416 | */ | |
a737b3e2 | 1417 | if (!sizes->cs_cachep) { |
e498be7d | 1418 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1419 | sizes->cs_size, |
1420 | ARCH_KMALLOC_MINALIGN, | |
1421 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1422 | NULL, NULL); | |
1423 | } | |
f1aaee53 | 1424 | init_lock_keys(sizes); |
1da177e4 | 1425 | |
1da177e4 | 1426 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, |
a737b3e2 AM |
1427 | sizes->cs_size, |
1428 | ARCH_KMALLOC_MINALIGN, | |
1429 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1430 | SLAB_PANIC, | |
1431 | NULL, NULL); | |
1da177e4 LT |
1432 | sizes++; |
1433 | names++; | |
1434 | } | |
1435 | /* 4) Replace the bootstrap head arrays */ | |
1436 | { | |
2b2d5493 | 1437 | struct array_cache *ptr; |
e498be7d | 1438 | |
1da177e4 | 1439 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1440 | |
1da177e4 | 1441 | local_irq_disable(); |
9a2dba4b PE |
1442 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1443 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1444 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1445 | /* |
1446 | * Do not assume that spinlocks can be initialized via memcpy: | |
1447 | */ | |
1448 | spin_lock_init(&ptr->lock); | |
1449 | ||
1da177e4 LT |
1450 | cache_cache.array[smp_processor_id()] = ptr; |
1451 | local_irq_enable(); | |
e498be7d | 1452 | |
1da177e4 | 1453 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1454 | |
1da177e4 | 1455 | local_irq_disable(); |
9a2dba4b | 1456 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1457 | != &initarray_generic.cache); |
9a2dba4b | 1458 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1459 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1460 | /* |
1461 | * Do not assume that spinlocks can be initialized via memcpy: | |
1462 | */ | |
1463 | spin_lock_init(&ptr->lock); | |
1464 | ||
e498be7d | 1465 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1466 | ptr; |
1da177e4 LT |
1467 | local_irq_enable(); |
1468 | } | |
e498be7d CL |
1469 | /* 5) Replace the bootstrap kmem_list3's */ |
1470 | { | |
1471 | int node; | |
1472 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1473 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1474 | numa_node_id()); |
e498be7d CL |
1475 | |
1476 | for_each_online_node(node) { | |
1477 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1478 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1479 | |
1480 | if (INDEX_AC != INDEX_L3) { | |
1481 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1482 | &initkmem_list3[SIZE_L3 + node], |
1483 | node); | |
e498be7d CL |
1484 | } |
1485 | } | |
1486 | } | |
1da177e4 | 1487 | |
e498be7d | 1488 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1489 | { |
343e0d7a | 1490 | struct kmem_cache *cachep; |
fc0abb14 | 1491 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1492 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1493 | if (enable_cpucache(cachep)) |
1494 | BUG(); | |
fc0abb14 | 1495 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1496 | } |
1497 | ||
1498 | /* Done! */ | |
1499 | g_cpucache_up = FULL; | |
1500 | ||
a737b3e2 AM |
1501 | /* |
1502 | * Register a cpu startup notifier callback that initializes | |
1503 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1504 | */ |
1505 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1506 | |
a737b3e2 AM |
1507 | /* |
1508 | * The reap timers are started later, with a module init call: That part | |
1509 | * of the kernel is not yet operational. | |
1da177e4 LT |
1510 | */ |
1511 | } | |
1512 | ||
1513 | static int __init cpucache_init(void) | |
1514 | { | |
1515 | int cpu; | |
1516 | ||
a737b3e2 AM |
1517 | /* |
1518 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1519 | */ |
e498be7d | 1520 | for_each_online_cpu(cpu) |
a737b3e2 | 1521 | start_cpu_timer(cpu); |
1da177e4 LT |
1522 | return 0; |
1523 | } | |
1da177e4 LT |
1524 | __initcall(cpucache_init); |
1525 | ||
1526 | /* | |
1527 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1528 | * | |
1529 | * If we requested dmaable memory, we will get it. Even if we | |
1530 | * did not request dmaable memory, we might get it, but that | |
1531 | * would be relatively rare and ignorable. | |
1532 | */ | |
343e0d7a | 1533 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1534 | { |
1535 | struct page *page; | |
e1b6aa6f | 1536 | int nr_pages; |
1da177e4 LT |
1537 | int i; |
1538 | ||
d6fef9da | 1539 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1540 | /* |
1541 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1542 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1543 | */ |
e1b6aa6f | 1544 | flags |= __GFP_COMP; |
d6fef9da | 1545 | #endif |
e1b6aa6f CH |
1546 | flags |= cachep->gfpflags; |
1547 | ||
1548 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1549 | if (!page) |
1550 | return NULL; | |
1da177e4 | 1551 | |
e1b6aa6f | 1552 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1553 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
e1b6aa6f | 1554 | atomic_add(nr_pages, &slab_reclaim_pages); |
9a865ffa | 1555 | add_zone_page_state(page_zone(page), NR_SLAB, nr_pages); |
e1b6aa6f CH |
1556 | for (i = 0; i < nr_pages; i++) |
1557 | __SetPageSlab(page + i); | |
1558 | return page_address(page); | |
1da177e4 LT |
1559 | } |
1560 | ||
1561 | /* | |
1562 | * Interface to system's page release. | |
1563 | */ | |
343e0d7a | 1564 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1565 | { |
b28a02de | 1566 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1567 | struct page *page = virt_to_page(addr); |
1568 | const unsigned long nr_freed = i; | |
1569 | ||
9a865ffa | 1570 | sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed); |
1da177e4 | 1571 | while (i--) { |
f205b2fe NP |
1572 | BUG_ON(!PageSlab(page)); |
1573 | __ClearPageSlab(page); | |
1da177e4 LT |
1574 | page++; |
1575 | } | |
1da177e4 LT |
1576 | if (current->reclaim_state) |
1577 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1578 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1579 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1580 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1581 | } |
1582 | ||
1583 | static void kmem_rcu_free(struct rcu_head *head) | |
1584 | { | |
b28a02de | 1585 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1586 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1587 | |
1588 | kmem_freepages(cachep, slab_rcu->addr); | |
1589 | if (OFF_SLAB(cachep)) | |
1590 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1591 | } | |
1592 | ||
1593 | #if DEBUG | |
1594 | ||
1595 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1596 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1597 | unsigned long caller) |
1da177e4 | 1598 | { |
3dafccf2 | 1599 | int size = obj_size(cachep); |
1da177e4 | 1600 | |
3dafccf2 | 1601 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1602 | |
b28a02de | 1603 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1604 | return; |
1605 | ||
b28a02de PE |
1606 | *addr++ = 0x12345678; |
1607 | *addr++ = caller; | |
1608 | *addr++ = smp_processor_id(); | |
1609 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1610 | { |
1611 | unsigned long *sptr = &caller; | |
1612 | unsigned long svalue; | |
1613 | ||
1614 | while (!kstack_end(sptr)) { | |
1615 | svalue = *sptr++; | |
1616 | if (kernel_text_address(svalue)) { | |
b28a02de | 1617 | *addr++ = svalue; |
1da177e4 LT |
1618 | size -= sizeof(unsigned long); |
1619 | if (size <= sizeof(unsigned long)) | |
1620 | break; | |
1621 | } | |
1622 | } | |
1623 | ||
1624 | } | |
b28a02de | 1625 | *addr++ = 0x87654321; |
1da177e4 LT |
1626 | } |
1627 | #endif | |
1628 | ||
343e0d7a | 1629 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1630 | { |
3dafccf2 MS |
1631 | int size = obj_size(cachep); |
1632 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1633 | |
1634 | memset(addr, val, size); | |
b28a02de | 1635 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1636 | } |
1637 | ||
1638 | static void dump_line(char *data, int offset, int limit) | |
1639 | { | |
1640 | int i; | |
1641 | printk(KERN_ERR "%03x:", offset); | |
a737b3e2 | 1642 | for (i = 0; i < limit; i++) |
b28a02de | 1643 | printk(" %02x", (unsigned char)data[offset + i]); |
1da177e4 LT |
1644 | printk("\n"); |
1645 | } | |
1646 | #endif | |
1647 | ||
1648 | #if DEBUG | |
1649 | ||
343e0d7a | 1650 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1651 | { |
1652 | int i, size; | |
1653 | char *realobj; | |
1654 | ||
1655 | if (cachep->flags & SLAB_RED_ZONE) { | |
1656 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
a737b3e2 AM |
1657 | *dbg_redzone1(cachep, objp), |
1658 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1659 | } |
1660 | ||
1661 | if (cachep->flags & SLAB_STORE_USER) { | |
1662 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1663 | *dbg_userword(cachep, objp)); |
1da177e4 | 1664 | print_symbol("(%s)", |
a737b3e2 | 1665 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1666 | printk("\n"); |
1667 | } | |
3dafccf2 MS |
1668 | realobj = (char *)objp + obj_offset(cachep); |
1669 | size = obj_size(cachep); | |
b28a02de | 1670 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1671 | int limit; |
1672 | limit = 16; | |
b28a02de PE |
1673 | if (i + limit > size) |
1674 | limit = size - i; | |
1da177e4 LT |
1675 | dump_line(realobj, i, limit); |
1676 | } | |
1677 | } | |
1678 | ||
343e0d7a | 1679 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1680 | { |
1681 | char *realobj; | |
1682 | int size, i; | |
1683 | int lines = 0; | |
1684 | ||
3dafccf2 MS |
1685 | realobj = (char *)objp + obj_offset(cachep); |
1686 | size = obj_size(cachep); | |
1da177e4 | 1687 | |
b28a02de | 1688 | for (i = 0; i < size; i++) { |
1da177e4 | 1689 | char exp = POISON_FREE; |
b28a02de | 1690 | if (i == size - 1) |
1da177e4 LT |
1691 | exp = POISON_END; |
1692 | if (realobj[i] != exp) { | |
1693 | int limit; | |
1694 | /* Mismatch ! */ | |
1695 | /* Print header */ | |
1696 | if (lines == 0) { | |
b28a02de | 1697 | printk(KERN_ERR |
a737b3e2 AM |
1698 | "Slab corruption: start=%p, len=%d\n", |
1699 | realobj, size); | |
1da177e4 LT |
1700 | print_objinfo(cachep, objp, 0); |
1701 | } | |
1702 | /* Hexdump the affected line */ | |
b28a02de | 1703 | i = (i / 16) * 16; |
1da177e4 | 1704 | limit = 16; |
b28a02de PE |
1705 | if (i + limit > size) |
1706 | limit = size - i; | |
1da177e4 LT |
1707 | dump_line(realobj, i, limit); |
1708 | i += 16; | |
1709 | lines++; | |
1710 | /* Limit to 5 lines */ | |
1711 | if (lines > 5) | |
1712 | break; | |
1713 | } | |
1714 | } | |
1715 | if (lines != 0) { | |
1716 | /* Print some data about the neighboring objects, if they | |
1717 | * exist: | |
1718 | */ | |
6ed5eb22 | 1719 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1720 | unsigned int objnr; |
1da177e4 | 1721 | |
8fea4e96 | 1722 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1723 | if (objnr) { |
8fea4e96 | 1724 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1725 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1726 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1727 | realobj, size); |
1da177e4 LT |
1728 | print_objinfo(cachep, objp, 2); |
1729 | } | |
b28a02de | 1730 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1731 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1732 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1733 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1734 | realobj, size); |
1da177e4 LT |
1735 | print_objinfo(cachep, objp, 2); |
1736 | } | |
1737 | } | |
1738 | } | |
1739 | #endif | |
1740 | ||
12dd36fa MD |
1741 | #if DEBUG |
1742 | /** | |
911851e6 RD |
1743 | * slab_destroy_objs - destroy a slab and its objects |
1744 | * @cachep: cache pointer being destroyed | |
1745 | * @slabp: slab pointer being destroyed | |
1746 | * | |
1747 | * Call the registered destructor for each object in a slab that is being | |
1748 | * destroyed. | |
1da177e4 | 1749 | */ |
343e0d7a | 1750 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1751 | { |
1da177e4 LT |
1752 | int i; |
1753 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1754 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1755 | |
1756 | if (cachep->flags & SLAB_POISON) { | |
1757 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1758 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1759 | OFF_SLAB(cachep)) | |
b28a02de | 1760 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1761 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1762 | else |
1763 | check_poison_obj(cachep, objp); | |
1764 | #else | |
1765 | check_poison_obj(cachep, objp); | |
1766 | #endif | |
1767 | } | |
1768 | if (cachep->flags & SLAB_RED_ZONE) { | |
1769 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1770 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1771 | "was overwritten"); |
1da177e4 LT |
1772 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1773 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1774 | "was overwritten"); |
1da177e4 LT |
1775 | } |
1776 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1777 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1778 | } |
12dd36fa | 1779 | } |
1da177e4 | 1780 | #else |
343e0d7a | 1781 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1782 | { |
1da177e4 LT |
1783 | if (cachep->dtor) { |
1784 | int i; | |
1785 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1786 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1787 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1788 | } |
1789 | } | |
12dd36fa | 1790 | } |
1da177e4 LT |
1791 | #endif |
1792 | ||
911851e6 RD |
1793 | /** |
1794 | * slab_destroy - destroy and release all objects in a slab | |
1795 | * @cachep: cache pointer being destroyed | |
1796 | * @slabp: slab pointer being destroyed | |
1797 | * | |
12dd36fa | 1798 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1799 | * Before calling the slab must have been unlinked from the cache. The |
1800 | * cache-lock is not held/needed. | |
12dd36fa | 1801 | */ |
343e0d7a | 1802 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1803 | { |
1804 | void *addr = slabp->s_mem - slabp->colouroff; | |
1805 | ||
1806 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1807 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1808 | struct slab_rcu *slab_rcu; | |
1809 | ||
b28a02de | 1810 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1811 | slab_rcu->cachep = cachep; |
1812 | slab_rcu->addr = addr; | |
1813 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1814 | } else { | |
1815 | kmem_freepages(cachep, addr); | |
873623df IM |
1816 | if (OFF_SLAB(cachep)) |
1817 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1818 | } |
1819 | } | |
1820 | ||
a737b3e2 AM |
1821 | /* |
1822 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1823 | * size of kmem_list3. | |
1824 | */ | |
343e0d7a | 1825 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1826 | { |
1827 | int node; | |
1828 | ||
1829 | for_each_online_node(node) { | |
b28a02de | 1830 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1831 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1832 | REAPTIMEOUT_LIST3 + |
1833 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1834 | } |
1835 | } | |
1836 | ||
117f6eb1 CL |
1837 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1838 | { | |
1839 | int i; | |
1840 | struct kmem_list3 *l3; | |
1841 | ||
1842 | for_each_online_cpu(i) | |
1843 | kfree(cachep->array[i]); | |
1844 | ||
1845 | /* NUMA: free the list3 structures */ | |
1846 | for_each_online_node(i) { | |
1847 | l3 = cachep->nodelists[i]; | |
1848 | if (l3) { | |
1849 | kfree(l3->shared); | |
1850 | free_alien_cache(l3->alien); | |
1851 | kfree(l3); | |
1852 | } | |
1853 | } | |
1854 | kmem_cache_free(&cache_cache, cachep); | |
1855 | } | |
1856 | ||
1857 | ||
4d268eba | 1858 | /** |
a70773dd RD |
1859 | * calculate_slab_order - calculate size (page order) of slabs |
1860 | * @cachep: pointer to the cache that is being created | |
1861 | * @size: size of objects to be created in this cache. | |
1862 | * @align: required alignment for the objects. | |
1863 | * @flags: slab allocation flags | |
1864 | * | |
1865 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1866 | * |
1867 | * This could be made much more intelligent. For now, try to avoid using | |
1868 | * high order pages for slabs. When the gfp() functions are more friendly | |
1869 | * towards high-order requests, this should be changed. | |
1870 | */ | |
a737b3e2 | 1871 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1872 | size_t size, size_t align, unsigned long flags) |
4d268eba | 1873 | { |
b1ab41c4 | 1874 | unsigned long offslab_limit; |
4d268eba | 1875 | size_t left_over = 0; |
9888e6fa | 1876 | int gfporder; |
4d268eba | 1877 | |
a737b3e2 | 1878 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1879 | unsigned int num; |
1880 | size_t remainder; | |
1881 | ||
9888e6fa | 1882 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1883 | if (!num) |
1884 | continue; | |
9888e6fa | 1885 | |
b1ab41c4 IM |
1886 | if (flags & CFLGS_OFF_SLAB) { |
1887 | /* | |
1888 | * Max number of objs-per-slab for caches which | |
1889 | * use off-slab slabs. Needed to avoid a possible | |
1890 | * looping condition in cache_grow(). | |
1891 | */ | |
1892 | offslab_limit = size - sizeof(struct slab); | |
1893 | offslab_limit /= sizeof(kmem_bufctl_t); | |
1894 | ||
1895 | if (num > offslab_limit) | |
1896 | break; | |
1897 | } | |
4d268eba | 1898 | |
9888e6fa | 1899 | /* Found something acceptable - save it away */ |
4d268eba | 1900 | cachep->num = num; |
9888e6fa | 1901 | cachep->gfporder = gfporder; |
4d268eba PE |
1902 | left_over = remainder; |
1903 | ||
f78bb8ad LT |
1904 | /* |
1905 | * A VFS-reclaimable slab tends to have most allocations | |
1906 | * as GFP_NOFS and we really don't want to have to be allocating | |
1907 | * higher-order pages when we are unable to shrink dcache. | |
1908 | */ | |
1909 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1910 | break; | |
1911 | ||
4d268eba PE |
1912 | /* |
1913 | * Large number of objects is good, but very large slabs are | |
1914 | * currently bad for the gfp()s. | |
1915 | */ | |
9888e6fa | 1916 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1917 | break; |
1918 | ||
9888e6fa LT |
1919 | /* |
1920 | * Acceptable internal fragmentation? | |
1921 | */ | |
a737b3e2 | 1922 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
1923 | break; |
1924 | } | |
1925 | return left_over; | |
1926 | } | |
1927 | ||
2ed3a4ef | 1928 | static int setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 1929 | { |
2ed3a4ef CL |
1930 | if (g_cpucache_up == FULL) |
1931 | return enable_cpucache(cachep); | |
1932 | ||
f30cf7d1 PE |
1933 | if (g_cpucache_up == NONE) { |
1934 | /* | |
1935 | * Note: the first kmem_cache_create must create the cache | |
1936 | * that's used by kmalloc(24), otherwise the creation of | |
1937 | * further caches will BUG(). | |
1938 | */ | |
1939 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
1940 | ||
1941 | /* | |
1942 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
1943 | * the first cache, then we need to set up all its list3s, | |
1944 | * otherwise the creation of further caches will BUG(). | |
1945 | */ | |
1946 | set_up_list3s(cachep, SIZE_AC); | |
1947 | if (INDEX_AC == INDEX_L3) | |
1948 | g_cpucache_up = PARTIAL_L3; | |
1949 | else | |
1950 | g_cpucache_up = PARTIAL_AC; | |
1951 | } else { | |
1952 | cachep->array[smp_processor_id()] = | |
1953 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
1954 | ||
1955 | if (g_cpucache_up == PARTIAL_AC) { | |
1956 | set_up_list3s(cachep, SIZE_L3); | |
1957 | g_cpucache_up = PARTIAL_L3; | |
1958 | } else { | |
1959 | int node; | |
1960 | for_each_online_node(node) { | |
1961 | cachep->nodelists[node] = | |
1962 | kmalloc_node(sizeof(struct kmem_list3), | |
1963 | GFP_KERNEL, node); | |
1964 | BUG_ON(!cachep->nodelists[node]); | |
1965 | kmem_list3_init(cachep->nodelists[node]); | |
1966 | } | |
1967 | } | |
1968 | } | |
1969 | cachep->nodelists[numa_node_id()]->next_reap = | |
1970 | jiffies + REAPTIMEOUT_LIST3 + | |
1971 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1972 | ||
1973 | cpu_cache_get(cachep)->avail = 0; | |
1974 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1975 | cpu_cache_get(cachep)->batchcount = 1; | |
1976 | cpu_cache_get(cachep)->touched = 0; | |
1977 | cachep->batchcount = 1; | |
1978 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 1979 | return 0; |
f30cf7d1 PE |
1980 | } |
1981 | ||
1da177e4 LT |
1982 | /** |
1983 | * kmem_cache_create - Create a cache. | |
1984 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1985 | * @size: The size of objects to be created in this cache. | |
1986 | * @align: The required alignment for the objects. | |
1987 | * @flags: SLAB flags | |
1988 | * @ctor: A constructor for the objects. | |
1989 | * @dtor: A destructor for the objects. | |
1990 | * | |
1991 | * Returns a ptr to the cache on success, NULL on failure. | |
1992 | * Cannot be called within a int, but can be interrupted. | |
1993 | * The @ctor is run when new pages are allocated by the cache | |
1994 | * and the @dtor is run before the pages are handed back. | |
1995 | * | |
1996 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
1997 | * the module calling this has to destroy the cache before getting unloaded. |
1998 | * | |
1da177e4 LT |
1999 | * The flags are |
2000 | * | |
2001 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2002 | * to catch references to uninitialised memory. | |
2003 | * | |
2004 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2005 | * for buffer overruns. | |
2006 | * | |
1da177e4 LT |
2007 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2008 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2009 | * as davem. | |
2010 | */ | |
343e0d7a | 2011 | struct kmem_cache * |
1da177e4 | 2012 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
2013 | unsigned long flags, |
2014 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 2015 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
2016 | { |
2017 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2018 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2019 | |
2020 | /* | |
2021 | * Sanity checks... these are all serious usage bugs. | |
2022 | */ | |
a737b3e2 | 2023 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
b28a02de | 2024 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { |
a737b3e2 AM |
2025 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
2026 | name); | |
b28a02de PE |
2027 | BUG(); |
2028 | } | |
1da177e4 | 2029 | |
f0188f47 RT |
2030 | /* |
2031 | * Prevent CPUs from coming and going. | |
2032 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
2033 | */ | |
2034 | lock_cpu_hotplug(); | |
2035 | ||
fc0abb14 | 2036 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2037 | |
7a7c381d | 2038 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2039 | mm_segment_t old_fs = get_fs(); |
2040 | char tmp; | |
2041 | int res; | |
2042 | ||
2043 | /* | |
2044 | * This happens when the module gets unloaded and doesn't | |
2045 | * destroy its slab cache and no-one else reuses the vmalloc | |
2046 | * area of the module. Print a warning. | |
2047 | */ | |
2048 | set_fs(KERNEL_DS); | |
2049 | res = __get_user(tmp, pc->name); | |
2050 | set_fs(old_fs); | |
2051 | if (res) { | |
2052 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2053 | pc->buffer_size); |
4f12bb4f AM |
2054 | continue; |
2055 | } | |
2056 | ||
b28a02de | 2057 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
2058 | printk("kmem_cache_create: duplicate cache %s\n", name); |
2059 | dump_stack(); | |
2060 | goto oops; | |
2061 | } | |
2062 | } | |
2063 | ||
1da177e4 LT |
2064 | #if DEBUG |
2065 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
2066 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
2067 | /* No constructor, but inital state check requested */ | |
2068 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 2069 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
2070 | flags &= ~SLAB_DEBUG_INITIAL; |
2071 | } | |
1da177e4 LT |
2072 | #if FORCED_DEBUG |
2073 | /* | |
2074 | * Enable redzoning and last user accounting, except for caches with | |
2075 | * large objects, if the increased size would increase the object size | |
2076 | * above the next power of two: caches with object sizes just above a | |
2077 | * power of two have a significant amount of internal fragmentation. | |
2078 | */ | |
a737b3e2 | 2079 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 2080 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2081 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2082 | flags |= SLAB_POISON; | |
2083 | #endif | |
2084 | if (flags & SLAB_DESTROY_BY_RCU) | |
2085 | BUG_ON(flags & SLAB_POISON); | |
2086 | #endif | |
2087 | if (flags & SLAB_DESTROY_BY_RCU) | |
2088 | BUG_ON(dtor); | |
2089 | ||
2090 | /* | |
a737b3e2 AM |
2091 | * Always checks flags, a caller might be expecting debug support which |
2092 | * isn't available. | |
1da177e4 | 2093 | */ |
40094fa6 | 2094 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2095 | |
a737b3e2 AM |
2096 | /* |
2097 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2098 | * unaligned accesses for some archs when redzoning is used, and makes |
2099 | * sure any on-slab bufctl's are also correctly aligned. | |
2100 | */ | |
b28a02de PE |
2101 | if (size & (BYTES_PER_WORD - 1)) { |
2102 | size += (BYTES_PER_WORD - 1); | |
2103 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2104 | } |
2105 | ||
a737b3e2 AM |
2106 | /* calculate the final buffer alignment: */ |
2107 | ||
1da177e4 LT |
2108 | /* 1) arch recommendation: can be overridden for debug */ |
2109 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2110 | /* |
2111 | * Default alignment: as specified by the arch code. Except if | |
2112 | * an object is really small, then squeeze multiple objects into | |
2113 | * one cacheline. | |
1da177e4 LT |
2114 | */ |
2115 | ralign = cache_line_size(); | |
b28a02de | 2116 | while (size <= ralign / 2) |
1da177e4 LT |
2117 | ralign /= 2; |
2118 | } else { | |
2119 | ralign = BYTES_PER_WORD; | |
2120 | } | |
ca5f9703 PE |
2121 | |
2122 | /* | |
2123 | * Redzoning and user store require word alignment. Note this will be | |
2124 | * overridden by architecture or caller mandated alignment if either | |
2125 | * is greater than BYTES_PER_WORD. | |
2126 | */ | |
2127 | if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) | |
2128 | ralign = BYTES_PER_WORD; | |
2129 | ||
1da177e4 LT |
2130 | /* 2) arch mandated alignment: disables debug if necessary */ |
2131 | if (ralign < ARCH_SLAB_MINALIGN) { | |
2132 | ralign = ARCH_SLAB_MINALIGN; | |
2133 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2134 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
2135 | } |
2136 | /* 3) caller mandated alignment: disables debug if necessary */ | |
2137 | if (ralign < align) { | |
2138 | ralign = align; | |
2139 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2140 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 | 2141 | } |
a737b3e2 | 2142 | /* |
ca5f9703 | 2143 | * 4) Store it. |
1da177e4 LT |
2144 | */ |
2145 | align = ralign; | |
2146 | ||
2147 | /* Get cache's description obj. */ | |
c5e3b83e | 2148 | cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 2149 | if (!cachep) |
4f12bb4f | 2150 | goto oops; |
1da177e4 LT |
2151 | |
2152 | #if DEBUG | |
3dafccf2 | 2153 | cachep->obj_size = size; |
1da177e4 | 2154 | |
ca5f9703 PE |
2155 | /* |
2156 | * Both debugging options require word-alignment which is calculated | |
2157 | * into align above. | |
2158 | */ | |
1da177e4 | 2159 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2160 | /* add space for red zone words */ |
3dafccf2 | 2161 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 2162 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
2163 | } |
2164 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 PE |
2165 | /* user store requires one word storage behind the end of |
2166 | * the real object. | |
1da177e4 | 2167 | */ |
1da177e4 LT |
2168 | size += BYTES_PER_WORD; |
2169 | } | |
2170 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2171 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2172 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2173 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2174 | size = PAGE_SIZE; |
2175 | } | |
2176 | #endif | |
2177 | #endif | |
2178 | ||
e0a42726 IM |
2179 | /* |
2180 | * Determine if the slab management is 'on' or 'off' slab. | |
2181 | * (bootstrapping cannot cope with offslab caches so don't do | |
2182 | * it too early on.) | |
2183 | */ | |
2184 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2185 | /* |
2186 | * Size is large, assume best to place the slab management obj | |
2187 | * off-slab (should allow better packing of objs). | |
2188 | */ | |
2189 | flags |= CFLGS_OFF_SLAB; | |
2190 | ||
2191 | size = ALIGN(size, align); | |
2192 | ||
f78bb8ad | 2193 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2194 | |
2195 | if (!cachep->num) { | |
2196 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
2197 | kmem_cache_free(&cache_cache, cachep); | |
2198 | cachep = NULL; | |
4f12bb4f | 2199 | goto oops; |
1da177e4 | 2200 | } |
b28a02de PE |
2201 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2202 | + sizeof(struct slab), align); | |
1da177e4 LT |
2203 | |
2204 | /* | |
2205 | * If the slab has been placed off-slab, and we have enough space then | |
2206 | * move it on-slab. This is at the expense of any extra colouring. | |
2207 | */ | |
2208 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2209 | flags &= ~CFLGS_OFF_SLAB; | |
2210 | left_over -= slab_size; | |
2211 | } | |
2212 | ||
2213 | if (flags & CFLGS_OFF_SLAB) { | |
2214 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2215 | slab_size = |
2216 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2217 | } |
2218 | ||
2219 | cachep->colour_off = cache_line_size(); | |
2220 | /* Offset must be a multiple of the alignment. */ | |
2221 | if (cachep->colour_off < align) | |
2222 | cachep->colour_off = align; | |
b28a02de | 2223 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2224 | cachep->slab_size = slab_size; |
2225 | cachep->flags = flags; | |
2226 | cachep->gfpflags = 0; | |
2227 | if (flags & SLAB_CACHE_DMA) | |
2228 | cachep->gfpflags |= GFP_DMA; | |
3dafccf2 | 2229 | cachep->buffer_size = size; |
1da177e4 | 2230 | |
e5ac9c5a | 2231 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2232 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2233 | /* |
2234 | * This is a possibility for one of the malloc_sizes caches. | |
2235 | * But since we go off slab only for object size greater than | |
2236 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2237 | * this should not happen at all. | |
2238 | * But leave a BUG_ON for some lucky dude. | |
2239 | */ | |
2240 | BUG_ON(!cachep->slabp_cache); | |
2241 | } | |
1da177e4 LT |
2242 | cachep->ctor = ctor; |
2243 | cachep->dtor = dtor; | |
2244 | cachep->name = name; | |
2245 | ||
2ed3a4ef CL |
2246 | if (setup_cpu_cache(cachep)) { |
2247 | __kmem_cache_destroy(cachep); | |
2248 | cachep = NULL; | |
2249 | goto oops; | |
2250 | } | |
1da177e4 | 2251 | |
1da177e4 LT |
2252 | /* cache setup completed, link it into the list */ |
2253 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2254 | oops: |
1da177e4 LT |
2255 | if (!cachep && (flags & SLAB_PANIC)) |
2256 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2257 | name); |
fc0abb14 | 2258 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 2259 | unlock_cpu_hotplug(); |
1da177e4 LT |
2260 | return cachep; |
2261 | } | |
2262 | EXPORT_SYMBOL(kmem_cache_create); | |
2263 | ||
2264 | #if DEBUG | |
2265 | static void check_irq_off(void) | |
2266 | { | |
2267 | BUG_ON(!irqs_disabled()); | |
2268 | } | |
2269 | ||
2270 | static void check_irq_on(void) | |
2271 | { | |
2272 | BUG_ON(irqs_disabled()); | |
2273 | } | |
2274 | ||
343e0d7a | 2275 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2276 | { |
2277 | #ifdef CONFIG_SMP | |
2278 | check_irq_off(); | |
e498be7d | 2279 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2280 | #endif |
2281 | } | |
e498be7d | 2282 | |
343e0d7a | 2283 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2284 | { |
2285 | #ifdef CONFIG_SMP | |
2286 | check_irq_off(); | |
2287 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2288 | #endif | |
2289 | } | |
2290 | ||
1da177e4 LT |
2291 | #else |
2292 | #define check_irq_off() do { } while(0) | |
2293 | #define check_irq_on() do { } while(0) | |
2294 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2295 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2296 | #endif |
2297 | ||
aab2207c CL |
2298 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2299 | struct array_cache *ac, | |
2300 | int force, int node); | |
2301 | ||
1da177e4 LT |
2302 | static void do_drain(void *arg) |
2303 | { | |
a737b3e2 | 2304 | struct kmem_cache *cachep = arg; |
1da177e4 | 2305 | struct array_cache *ac; |
ff69416e | 2306 | int node = numa_node_id(); |
1da177e4 LT |
2307 | |
2308 | check_irq_off(); | |
9a2dba4b | 2309 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2310 | spin_lock(&cachep->nodelists[node]->list_lock); |
2311 | free_block(cachep, ac->entry, ac->avail, node); | |
2312 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2313 | ac->avail = 0; |
2314 | } | |
2315 | ||
343e0d7a | 2316 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2317 | { |
e498be7d CL |
2318 | struct kmem_list3 *l3; |
2319 | int node; | |
2320 | ||
a07fa394 | 2321 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2322 | check_irq_on(); |
b28a02de | 2323 | for_each_online_node(node) { |
e498be7d | 2324 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2325 | if (l3 && l3->alien) |
2326 | drain_alien_cache(cachep, l3->alien); | |
2327 | } | |
2328 | ||
2329 | for_each_online_node(node) { | |
2330 | l3 = cachep->nodelists[node]; | |
2331 | if (l3) | |
aab2207c | 2332 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2333 | } |
1da177e4 LT |
2334 | } |
2335 | ||
ed11d9eb CL |
2336 | /* |
2337 | * Remove slabs from the list of free slabs. | |
2338 | * Specify the number of slabs to drain in tofree. | |
2339 | * | |
2340 | * Returns the actual number of slabs released. | |
2341 | */ | |
2342 | static int drain_freelist(struct kmem_cache *cache, | |
2343 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2344 | { |
ed11d9eb CL |
2345 | struct list_head *p; |
2346 | int nr_freed; | |
1da177e4 | 2347 | struct slab *slabp; |
1da177e4 | 2348 | |
ed11d9eb CL |
2349 | nr_freed = 0; |
2350 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2351 | |
ed11d9eb | 2352 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2353 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2354 | if (p == &l3->slabs_free) { |
2355 | spin_unlock_irq(&l3->list_lock); | |
2356 | goto out; | |
2357 | } | |
1da177e4 | 2358 | |
ed11d9eb | 2359 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2360 | #if DEBUG |
40094fa6 | 2361 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2362 | #endif |
2363 | list_del(&slabp->list); | |
ed11d9eb CL |
2364 | /* |
2365 | * Safe to drop the lock. The slab is no longer linked | |
2366 | * to the cache. | |
2367 | */ | |
2368 | l3->free_objects -= cache->num; | |
e498be7d | 2369 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2370 | slab_destroy(cache, slabp); |
2371 | nr_freed++; | |
1da177e4 | 2372 | } |
ed11d9eb CL |
2373 | out: |
2374 | return nr_freed; | |
1da177e4 LT |
2375 | } |
2376 | ||
343e0d7a | 2377 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2378 | { |
2379 | int ret = 0, i = 0; | |
2380 | struct kmem_list3 *l3; | |
2381 | ||
2382 | drain_cpu_caches(cachep); | |
2383 | ||
2384 | check_irq_on(); | |
2385 | for_each_online_node(i) { | |
2386 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2387 | if (!l3) |
2388 | continue; | |
2389 | ||
2390 | drain_freelist(cachep, l3, l3->free_objects); | |
2391 | ||
2392 | ret += !list_empty(&l3->slabs_full) || | |
2393 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2394 | } |
2395 | return (ret ? 1 : 0); | |
2396 | } | |
2397 | ||
1da177e4 LT |
2398 | /** |
2399 | * kmem_cache_shrink - Shrink a cache. | |
2400 | * @cachep: The cache to shrink. | |
2401 | * | |
2402 | * Releases as many slabs as possible for a cache. | |
2403 | * To help debugging, a zero exit status indicates all slabs were released. | |
2404 | */ | |
343e0d7a | 2405 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2406 | { |
40094fa6 | 2407 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2408 | |
2409 | return __cache_shrink(cachep); | |
2410 | } | |
2411 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2412 | ||
2413 | /** | |
2414 | * kmem_cache_destroy - delete a cache | |
2415 | * @cachep: the cache to destroy | |
2416 | * | |
343e0d7a | 2417 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2418 | * Returns 0 on success. |
2419 | * | |
2420 | * It is expected this function will be called by a module when it is | |
2421 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2422 | * cache being allocated each time a module is loaded and unloaded, if the | |
2423 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2424 | * | |
2425 | * The cache must be empty before calling this function. | |
2426 | * | |
2427 | * The caller must guarantee that noone will allocate memory from the cache | |
2428 | * during the kmem_cache_destroy(). | |
2429 | */ | |
343e0d7a | 2430 | int kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2431 | { |
40094fa6 | 2432 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2433 | |
2434 | /* Don't let CPUs to come and go */ | |
2435 | lock_cpu_hotplug(); | |
2436 | ||
2437 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2438 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2439 | /* |
2440 | * the chain is never empty, cache_cache is never destroyed | |
2441 | */ | |
2442 | list_del(&cachep->next); | |
fc0abb14 | 2443 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2444 | |
2445 | if (__cache_shrink(cachep)) { | |
2446 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2447 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2448 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2449 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2450 | unlock_cpu_hotplug(); |
2451 | return 1; | |
2452 | } | |
2453 | ||
2454 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2455 | synchronize_rcu(); |
1da177e4 | 2456 | |
117f6eb1 | 2457 | __kmem_cache_destroy(cachep); |
1da177e4 | 2458 | unlock_cpu_hotplug(); |
1da177e4 LT |
2459 | return 0; |
2460 | } | |
2461 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2462 | ||
e5ac9c5a RT |
2463 | /* |
2464 | * Get the memory for a slab management obj. | |
2465 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2466 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2467 | * come from the same cache which is getting created because, | |
2468 | * when we are searching for an appropriate cache for these | |
2469 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2470 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2471 | * kmem_find_general_cachep till the initialization is complete. | |
2472 | * Hence we cannot have slabp_cache same as the original cache. | |
2473 | */ | |
343e0d7a | 2474 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2475 | int colour_off, gfp_t local_flags, |
2476 | int nodeid) | |
1da177e4 LT |
2477 | { |
2478 | struct slab *slabp; | |
b28a02de | 2479 | |
1da177e4 LT |
2480 | if (OFF_SLAB(cachep)) { |
2481 | /* Slab management obj is off-slab. */ | |
5b74ada7 RT |
2482 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
2483 | local_flags, nodeid); | |
1da177e4 LT |
2484 | if (!slabp) |
2485 | return NULL; | |
2486 | } else { | |
b28a02de | 2487 | slabp = objp + colour_off; |
1da177e4 LT |
2488 | colour_off += cachep->slab_size; |
2489 | } | |
2490 | slabp->inuse = 0; | |
2491 | slabp->colouroff = colour_off; | |
b28a02de | 2492 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2493 | slabp->nodeid = nodeid; |
1da177e4 LT |
2494 | return slabp; |
2495 | } | |
2496 | ||
2497 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2498 | { | |
b28a02de | 2499 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2500 | } |
2501 | ||
343e0d7a | 2502 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2503 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2504 | { |
2505 | int i; | |
2506 | ||
2507 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2508 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2509 | #if DEBUG |
2510 | /* need to poison the objs? */ | |
2511 | if (cachep->flags & SLAB_POISON) | |
2512 | poison_obj(cachep, objp, POISON_FREE); | |
2513 | if (cachep->flags & SLAB_STORE_USER) | |
2514 | *dbg_userword(cachep, objp) = NULL; | |
2515 | ||
2516 | if (cachep->flags & SLAB_RED_ZONE) { | |
2517 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2518 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2519 | } | |
2520 | /* | |
a737b3e2 AM |
2521 | * Constructors are not allowed to allocate memory from the same |
2522 | * cache which they are a constructor for. Otherwise, deadlock. | |
2523 | * They must also be threaded. | |
1da177e4 LT |
2524 | */ |
2525 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2526 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2527 | ctor_flags); |
1da177e4 LT |
2528 | |
2529 | if (cachep->flags & SLAB_RED_ZONE) { | |
2530 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2531 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2532 | " end of an object"); |
1da177e4 LT |
2533 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2534 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2535 | " start of an object"); |
1da177e4 | 2536 | } |
a737b3e2 AM |
2537 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2538 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2539 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2540 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2541 | #else |
2542 | if (cachep->ctor) | |
2543 | cachep->ctor(objp, cachep, ctor_flags); | |
2544 | #endif | |
b28a02de | 2545 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2546 | } |
b28a02de | 2547 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2548 | slabp->free = 0; |
2549 | } | |
2550 | ||
343e0d7a | 2551 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2552 | { |
a737b3e2 AM |
2553 | if (flags & SLAB_DMA) |
2554 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2555 | else | |
2556 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
1da177e4 LT |
2557 | } |
2558 | ||
a737b3e2 AM |
2559 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2560 | int nodeid) | |
78d382d7 | 2561 | { |
8fea4e96 | 2562 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2563 | kmem_bufctl_t next; |
2564 | ||
2565 | slabp->inuse++; | |
2566 | next = slab_bufctl(slabp)[slabp->free]; | |
2567 | #if DEBUG | |
2568 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2569 | WARN_ON(slabp->nodeid != nodeid); | |
2570 | #endif | |
2571 | slabp->free = next; | |
2572 | ||
2573 | return objp; | |
2574 | } | |
2575 | ||
a737b3e2 AM |
2576 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2577 | void *objp, int nodeid) | |
78d382d7 | 2578 | { |
8fea4e96 | 2579 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2580 | |
2581 | #if DEBUG | |
2582 | /* Verify that the slab belongs to the intended node */ | |
2583 | WARN_ON(slabp->nodeid != nodeid); | |
2584 | ||
871751e2 | 2585 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2586 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2587 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2588 | BUG(); |
2589 | } | |
2590 | #endif | |
2591 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2592 | slabp->free = objnr; | |
2593 | slabp->inuse--; | |
2594 | } | |
2595 | ||
4776874f PE |
2596 | /* |
2597 | * Map pages beginning at addr to the given cache and slab. This is required | |
2598 | * for the slab allocator to be able to lookup the cache and slab of a | |
2599 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2600 | */ | |
2601 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2602 | void *addr) | |
1da177e4 | 2603 | { |
4776874f | 2604 | int nr_pages; |
1da177e4 LT |
2605 | struct page *page; |
2606 | ||
4776874f | 2607 | page = virt_to_page(addr); |
84097518 | 2608 | |
4776874f | 2609 | nr_pages = 1; |
84097518 | 2610 | if (likely(!PageCompound(page))) |
4776874f PE |
2611 | nr_pages <<= cache->gfporder; |
2612 | ||
1da177e4 | 2613 | do { |
4776874f PE |
2614 | page_set_cache(page, cache); |
2615 | page_set_slab(page, slab); | |
1da177e4 | 2616 | page++; |
4776874f | 2617 | } while (--nr_pages); |
1da177e4 LT |
2618 | } |
2619 | ||
2620 | /* | |
2621 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2622 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2623 | */ | |
343e0d7a | 2624 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2625 | { |
b28a02de PE |
2626 | struct slab *slabp; |
2627 | void *objp; | |
2628 | size_t offset; | |
2629 | gfp_t local_flags; | |
2630 | unsigned long ctor_flags; | |
e498be7d | 2631 | struct kmem_list3 *l3; |
1da177e4 | 2632 | |
a737b3e2 AM |
2633 | /* |
2634 | * Be lazy and only check for valid flags here, keeping it out of the | |
2635 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2636 | */ |
40094fa6 | 2637 | BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); |
1da177e4 LT |
2638 | if (flags & SLAB_NO_GROW) |
2639 | return 0; | |
2640 | ||
2641 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2642 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2643 | if (!(local_flags & __GFP_WAIT)) | |
2644 | /* | |
2645 | * Not allowed to sleep. Need to tell a constructor about | |
2646 | * this - it might need to know... | |
2647 | */ | |
2648 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2649 | ||
2e1217cf | 2650 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2651 | check_irq_off(); |
2e1217cf RT |
2652 | l3 = cachep->nodelists[nodeid]; |
2653 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2654 | |
2655 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2656 | offset = l3->colour_next; |
2657 | l3->colour_next++; | |
2658 | if (l3->colour_next >= cachep->colour) | |
2659 | l3->colour_next = 0; | |
2660 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2661 | |
2e1217cf | 2662 | offset *= cachep->colour_off; |
1da177e4 LT |
2663 | |
2664 | if (local_flags & __GFP_WAIT) | |
2665 | local_irq_enable(); | |
2666 | ||
2667 | /* | |
2668 | * The test for missing atomic flag is performed here, rather than | |
2669 | * the more obvious place, simply to reduce the critical path length | |
2670 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2671 | * will eventually be caught here (where it matters). | |
2672 | */ | |
2673 | kmem_flagcheck(cachep, flags); | |
2674 | ||
a737b3e2 AM |
2675 | /* |
2676 | * Get mem for the objs. Attempt to allocate a physical page from | |
2677 | * 'nodeid'. | |
e498be7d | 2678 | */ |
a737b3e2 AM |
2679 | objp = kmem_getpages(cachep, flags, nodeid); |
2680 | if (!objp) | |
1da177e4 LT |
2681 | goto failed; |
2682 | ||
2683 | /* Get slab management. */ | |
5b74ada7 | 2684 | slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); |
a737b3e2 | 2685 | if (!slabp) |
1da177e4 LT |
2686 | goto opps1; |
2687 | ||
e498be7d | 2688 | slabp->nodeid = nodeid; |
4776874f | 2689 | slab_map_pages(cachep, slabp, objp); |
1da177e4 LT |
2690 | |
2691 | cache_init_objs(cachep, slabp, ctor_flags); | |
2692 | ||
2693 | if (local_flags & __GFP_WAIT) | |
2694 | local_irq_disable(); | |
2695 | check_irq_off(); | |
e498be7d | 2696 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2697 | |
2698 | /* Make slab active. */ | |
e498be7d | 2699 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2700 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2701 | l3->free_objects += cachep->num; |
2702 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2703 | return 1; |
a737b3e2 | 2704 | opps1: |
1da177e4 | 2705 | kmem_freepages(cachep, objp); |
a737b3e2 | 2706 | failed: |
1da177e4 LT |
2707 | if (local_flags & __GFP_WAIT) |
2708 | local_irq_disable(); | |
2709 | return 0; | |
2710 | } | |
2711 | ||
2712 | #if DEBUG | |
2713 | ||
2714 | /* | |
2715 | * Perform extra freeing checks: | |
2716 | * - detect bad pointers. | |
2717 | * - POISON/RED_ZONE checking | |
2718 | * - destructor calls, for caches with POISON+dtor | |
2719 | */ | |
2720 | static void kfree_debugcheck(const void *objp) | |
2721 | { | |
2722 | struct page *page; | |
2723 | ||
2724 | if (!virt_addr_valid(objp)) { | |
2725 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2726 | (unsigned long)objp); |
2727 | BUG(); | |
1da177e4 LT |
2728 | } |
2729 | page = virt_to_page(objp); | |
2730 | if (!PageSlab(page)) { | |
b28a02de PE |
2731 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2732 | (unsigned long)objp); | |
1da177e4 LT |
2733 | BUG(); |
2734 | } | |
2735 | } | |
2736 | ||
58ce1fd5 PE |
2737 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2738 | { | |
2739 | unsigned long redzone1, redzone2; | |
2740 | ||
2741 | redzone1 = *dbg_redzone1(cache, obj); | |
2742 | redzone2 = *dbg_redzone2(cache, obj); | |
2743 | ||
2744 | /* | |
2745 | * Redzone is ok. | |
2746 | */ | |
2747 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2748 | return; | |
2749 | ||
2750 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2751 | slab_error(cache, "double free detected"); | |
2752 | else | |
2753 | slab_error(cache, "memory outside object was overwritten"); | |
2754 | ||
2755 | printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n", | |
2756 | obj, redzone1, redzone2); | |
2757 | } | |
2758 | ||
343e0d7a | 2759 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2760 | void *caller) |
1da177e4 LT |
2761 | { |
2762 | struct page *page; | |
2763 | unsigned int objnr; | |
2764 | struct slab *slabp; | |
2765 | ||
3dafccf2 | 2766 | objp -= obj_offset(cachep); |
1da177e4 LT |
2767 | kfree_debugcheck(objp); |
2768 | page = virt_to_page(objp); | |
2769 | ||
065d41cb | 2770 | slabp = page_get_slab(page); |
1da177e4 LT |
2771 | |
2772 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2773 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2774 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2775 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2776 | } | |
2777 | if (cachep->flags & SLAB_STORE_USER) | |
2778 | *dbg_userword(cachep, objp) = caller; | |
2779 | ||
8fea4e96 | 2780 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2781 | |
2782 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2783 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2784 | |
2785 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
a737b3e2 AM |
2786 | /* |
2787 | * Need to call the slab's constructor so the caller can | |
2788 | * perform a verify of its state (debugging). Called without | |
2789 | * the cache-lock held. | |
1da177e4 | 2790 | */ |
3dafccf2 | 2791 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2792 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2793 | } |
2794 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2795 | /* we want to cache poison the object, | |
2796 | * call the destruction callback | |
2797 | */ | |
3dafccf2 | 2798 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 2799 | } |
871751e2 AV |
2800 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2801 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2802 | #endif | |
1da177e4 LT |
2803 | if (cachep->flags & SLAB_POISON) { |
2804 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2805 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2806 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2807 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2808 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2809 | } else { |
2810 | poison_obj(cachep, objp, POISON_FREE); | |
2811 | } | |
2812 | #else | |
2813 | poison_obj(cachep, objp, POISON_FREE); | |
2814 | #endif | |
2815 | } | |
2816 | return objp; | |
2817 | } | |
2818 | ||
343e0d7a | 2819 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2820 | { |
2821 | kmem_bufctl_t i; | |
2822 | int entries = 0; | |
b28a02de | 2823 | |
1da177e4 LT |
2824 | /* Check slab's freelist to see if this obj is there. */ |
2825 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2826 | entries++; | |
2827 | if (entries > cachep->num || i >= cachep->num) | |
2828 | goto bad; | |
2829 | } | |
2830 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2831 | bad: |
2832 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2833 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2834 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2835 | for (i = 0; |
264132bc | 2836 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2837 | i++) { |
a737b3e2 | 2838 | if (i % 16 == 0) |
1da177e4 | 2839 | printk("\n%03x:", i); |
b28a02de | 2840 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2841 | } |
2842 | printk("\n"); | |
2843 | BUG(); | |
2844 | } | |
2845 | } | |
2846 | #else | |
2847 | #define kfree_debugcheck(x) do { } while(0) | |
2848 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2849 | #define check_slabp(x,y) do { } while(0) | |
2850 | #endif | |
2851 | ||
343e0d7a | 2852 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2853 | { |
2854 | int batchcount; | |
2855 | struct kmem_list3 *l3; | |
2856 | struct array_cache *ac; | |
2857 | ||
2858 | check_irq_off(); | |
9a2dba4b | 2859 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2860 | retry: |
1da177e4 LT |
2861 | batchcount = ac->batchcount; |
2862 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2863 | /* |
2864 | * If there was little recent activity on this cache, then | |
2865 | * perform only a partial refill. Otherwise we could generate | |
2866 | * refill bouncing. | |
1da177e4 LT |
2867 | */ |
2868 | batchcount = BATCHREFILL_LIMIT; | |
2869 | } | |
e498be7d CL |
2870 | l3 = cachep->nodelists[numa_node_id()]; |
2871 | ||
2872 | BUG_ON(ac->avail > 0 || !l3); | |
2873 | spin_lock(&l3->list_lock); | |
1da177e4 | 2874 | |
3ded175a CL |
2875 | /* See if we can refill from the shared array */ |
2876 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2877 | goto alloc_done; | |
2878 | ||
1da177e4 LT |
2879 | while (batchcount > 0) { |
2880 | struct list_head *entry; | |
2881 | struct slab *slabp; | |
2882 | /* Get slab alloc is to come from. */ | |
2883 | entry = l3->slabs_partial.next; | |
2884 | if (entry == &l3->slabs_partial) { | |
2885 | l3->free_touched = 1; | |
2886 | entry = l3->slabs_free.next; | |
2887 | if (entry == &l3->slabs_free) | |
2888 | goto must_grow; | |
2889 | } | |
2890 | ||
2891 | slabp = list_entry(entry, struct slab, list); | |
2892 | check_slabp(cachep, slabp); | |
2893 | check_spinlock_acquired(cachep); | |
2894 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2895 | STATS_INC_ALLOCED(cachep); |
2896 | STATS_INC_ACTIVE(cachep); | |
2897 | STATS_SET_HIGH(cachep); | |
2898 | ||
78d382d7 MD |
2899 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2900 | numa_node_id()); | |
1da177e4 LT |
2901 | } |
2902 | check_slabp(cachep, slabp); | |
2903 | ||
2904 | /* move slabp to correct slabp list: */ | |
2905 | list_del(&slabp->list); | |
2906 | if (slabp->free == BUFCTL_END) | |
2907 | list_add(&slabp->list, &l3->slabs_full); | |
2908 | else | |
2909 | list_add(&slabp->list, &l3->slabs_partial); | |
2910 | } | |
2911 | ||
a737b3e2 | 2912 | must_grow: |
1da177e4 | 2913 | l3->free_objects -= ac->avail; |
a737b3e2 | 2914 | alloc_done: |
e498be7d | 2915 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2916 | |
2917 | if (unlikely(!ac->avail)) { | |
2918 | int x; | |
e498be7d CL |
2919 | x = cache_grow(cachep, flags, numa_node_id()); |
2920 | ||
a737b3e2 | 2921 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 2922 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2923 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
2924 | return NULL; |
2925 | ||
a737b3e2 | 2926 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
2927 | goto retry; |
2928 | } | |
2929 | ac->touched = 1; | |
e498be7d | 2930 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2931 | } |
2932 | ||
a737b3e2 AM |
2933 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
2934 | gfp_t flags) | |
1da177e4 LT |
2935 | { |
2936 | might_sleep_if(flags & __GFP_WAIT); | |
2937 | #if DEBUG | |
2938 | kmem_flagcheck(cachep, flags); | |
2939 | #endif | |
2940 | } | |
2941 | ||
2942 | #if DEBUG | |
a737b3e2 AM |
2943 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
2944 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 2945 | { |
b28a02de | 2946 | if (!objp) |
1da177e4 | 2947 | return objp; |
b28a02de | 2948 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2949 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2950 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2951 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2952 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2953 | else |
2954 | check_poison_obj(cachep, objp); | |
2955 | #else | |
2956 | check_poison_obj(cachep, objp); | |
2957 | #endif | |
2958 | poison_obj(cachep, objp, POISON_INUSE); | |
2959 | } | |
2960 | if (cachep->flags & SLAB_STORE_USER) | |
2961 | *dbg_userword(cachep, objp) = caller; | |
2962 | ||
2963 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2964 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
2965 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2966 | slab_error(cachep, "double free, or memory outside" | |
2967 | " object was overwritten"); | |
b28a02de | 2968 | printk(KERN_ERR |
a737b3e2 AM |
2969 | "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", |
2970 | objp, *dbg_redzone1(cachep, objp), | |
2971 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2972 | } |
2973 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2974 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2975 | } | |
871751e2 AV |
2976 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2977 | { | |
2978 | struct slab *slabp; | |
2979 | unsigned objnr; | |
2980 | ||
2981 | slabp = page_get_slab(virt_to_page(objp)); | |
2982 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; | |
2983 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
2984 | } | |
2985 | #endif | |
3dafccf2 | 2986 | objp += obj_offset(cachep); |
1da177e4 | 2987 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2988 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2989 | |
2990 | if (!(flags & __GFP_WAIT)) | |
2991 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2992 | ||
2993 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2994 | } |
1da177e4 LT |
2995 | return objp; |
2996 | } | |
2997 | #else | |
2998 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2999 | #endif | |
3000 | ||
343e0d7a | 3001 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3002 | { |
b28a02de | 3003 | void *objp; |
1da177e4 LT |
3004 | struct array_cache *ac; |
3005 | ||
dc85da15 | 3006 | #ifdef CONFIG_NUMA |
b2455396 | 3007 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { |
c61afb18 PJ |
3008 | objp = alternate_node_alloc(cachep, flags); |
3009 | if (objp != NULL) | |
3010 | return objp; | |
dc85da15 CL |
3011 | } |
3012 | #endif | |
3013 | ||
5c382300 | 3014 | check_irq_off(); |
9a2dba4b | 3015 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3016 | if (likely(ac->avail)) { |
3017 | STATS_INC_ALLOCHIT(cachep); | |
3018 | ac->touched = 1; | |
e498be7d | 3019 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3020 | } else { |
3021 | STATS_INC_ALLOCMISS(cachep); | |
3022 | objp = cache_alloc_refill(cachep, flags); | |
3023 | } | |
5c382300 AK |
3024 | return objp; |
3025 | } | |
3026 | ||
a737b3e2 AM |
3027 | static __always_inline void *__cache_alloc(struct kmem_cache *cachep, |
3028 | gfp_t flags, void *caller) | |
5c382300 AK |
3029 | { |
3030 | unsigned long save_flags; | |
b28a02de | 3031 | void *objp; |
5c382300 AK |
3032 | |
3033 | cache_alloc_debugcheck_before(cachep, flags); | |
3034 | ||
3035 | local_irq_save(save_flags); | |
3036 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 3037 | local_irq_restore(save_flags); |
34342e86 | 3038 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 3039 | caller); |
34342e86 | 3040 | prefetchw(objp); |
1da177e4 LT |
3041 | return objp; |
3042 | } | |
3043 | ||
e498be7d | 3044 | #ifdef CONFIG_NUMA |
c61afb18 | 3045 | /* |
b2455396 | 3046 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3047 | * |
3048 | * If we are in_interrupt, then process context, including cpusets and | |
3049 | * mempolicy, may not apply and should not be used for allocation policy. | |
3050 | */ | |
3051 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3052 | { | |
3053 | int nid_alloc, nid_here; | |
3054 | ||
3055 | if (in_interrupt()) | |
3056 | return NULL; | |
3057 | nid_alloc = nid_here = numa_node_id(); | |
3058 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3059 | nid_alloc = cpuset_mem_spread_node(); | |
3060 | else if (current->mempolicy) | |
3061 | nid_alloc = slab_node(current->mempolicy); | |
3062 | if (nid_alloc != nid_here) | |
3063 | return __cache_alloc_node(cachep, flags, nid_alloc); | |
3064 | return NULL; | |
3065 | } | |
3066 | ||
e498be7d CL |
3067 | /* |
3068 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3069 | */ |
a737b3e2 AM |
3070 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
3071 | int nodeid) | |
e498be7d CL |
3072 | { |
3073 | struct list_head *entry; | |
b28a02de PE |
3074 | struct slab *slabp; |
3075 | struct kmem_list3 *l3; | |
3076 | void *obj; | |
b28a02de PE |
3077 | int x; |
3078 | ||
3079 | l3 = cachep->nodelists[nodeid]; | |
3080 | BUG_ON(!l3); | |
3081 | ||
a737b3e2 | 3082 | retry: |
ca3b9b91 | 3083 | check_irq_off(); |
b28a02de PE |
3084 | spin_lock(&l3->list_lock); |
3085 | entry = l3->slabs_partial.next; | |
3086 | if (entry == &l3->slabs_partial) { | |
3087 | l3->free_touched = 1; | |
3088 | entry = l3->slabs_free.next; | |
3089 | if (entry == &l3->slabs_free) | |
3090 | goto must_grow; | |
3091 | } | |
3092 | ||
3093 | slabp = list_entry(entry, struct slab, list); | |
3094 | check_spinlock_acquired_node(cachep, nodeid); | |
3095 | check_slabp(cachep, slabp); | |
3096 | ||
3097 | STATS_INC_NODEALLOCS(cachep); | |
3098 | STATS_INC_ACTIVE(cachep); | |
3099 | STATS_SET_HIGH(cachep); | |
3100 | ||
3101 | BUG_ON(slabp->inuse == cachep->num); | |
3102 | ||
78d382d7 | 3103 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3104 | check_slabp(cachep, slabp); |
3105 | l3->free_objects--; | |
3106 | /* move slabp to correct slabp list: */ | |
3107 | list_del(&slabp->list); | |
3108 | ||
a737b3e2 | 3109 | if (slabp->free == BUFCTL_END) |
b28a02de | 3110 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3111 | else |
b28a02de | 3112 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3113 | |
b28a02de PE |
3114 | spin_unlock(&l3->list_lock); |
3115 | goto done; | |
e498be7d | 3116 | |
a737b3e2 | 3117 | must_grow: |
b28a02de PE |
3118 | spin_unlock(&l3->list_lock); |
3119 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 3120 | |
b28a02de PE |
3121 | if (!x) |
3122 | return NULL; | |
e498be7d | 3123 | |
b28a02de | 3124 | goto retry; |
a737b3e2 | 3125 | done: |
b28a02de | 3126 | return obj; |
e498be7d CL |
3127 | } |
3128 | #endif | |
3129 | ||
3130 | /* | |
3131 | * Caller needs to acquire correct kmem_list's list_lock | |
3132 | */ | |
343e0d7a | 3133 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3134 | int node) |
1da177e4 LT |
3135 | { |
3136 | int i; | |
e498be7d | 3137 | struct kmem_list3 *l3; |
1da177e4 LT |
3138 | |
3139 | for (i = 0; i < nr_objects; i++) { | |
3140 | void *objp = objpp[i]; | |
3141 | struct slab *slabp; | |
1da177e4 | 3142 | |
6ed5eb22 | 3143 | slabp = virt_to_slab(objp); |
ff69416e | 3144 | l3 = cachep->nodelists[node]; |
1da177e4 | 3145 | list_del(&slabp->list); |
ff69416e | 3146 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3147 | check_slabp(cachep, slabp); |
78d382d7 | 3148 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3149 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3150 | l3->free_objects++; |
1da177e4 LT |
3151 | check_slabp(cachep, slabp); |
3152 | ||
3153 | /* fixup slab chains */ | |
3154 | if (slabp->inuse == 0) { | |
e498be7d CL |
3155 | if (l3->free_objects > l3->free_limit) { |
3156 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3157 | /* No need to drop any previously held |
3158 | * lock here, even if we have a off-slab slab | |
3159 | * descriptor it is guaranteed to come from | |
3160 | * a different cache, refer to comments before | |
3161 | * alloc_slabmgmt. | |
3162 | */ | |
1da177e4 LT |
3163 | slab_destroy(cachep, slabp); |
3164 | } else { | |
e498be7d | 3165 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3166 | } |
3167 | } else { | |
3168 | /* Unconditionally move a slab to the end of the | |
3169 | * partial list on free - maximum time for the | |
3170 | * other objects to be freed, too. | |
3171 | */ | |
e498be7d | 3172 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3173 | } |
3174 | } | |
3175 | } | |
3176 | ||
343e0d7a | 3177 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3178 | { |
3179 | int batchcount; | |
e498be7d | 3180 | struct kmem_list3 *l3; |
ff69416e | 3181 | int node = numa_node_id(); |
1da177e4 LT |
3182 | |
3183 | batchcount = ac->batchcount; | |
3184 | #if DEBUG | |
3185 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3186 | #endif | |
3187 | check_irq_off(); | |
ff69416e | 3188 | l3 = cachep->nodelists[node]; |
873623df | 3189 | spin_lock(&l3->list_lock); |
e498be7d CL |
3190 | if (l3->shared) { |
3191 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3192 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3193 | if (max) { |
3194 | if (batchcount > max) | |
3195 | batchcount = max; | |
e498be7d | 3196 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3197 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3198 | shared_array->avail += batchcount; |
3199 | goto free_done; | |
3200 | } | |
3201 | } | |
3202 | ||
ff69416e | 3203 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3204 | free_done: |
1da177e4 LT |
3205 | #if STATS |
3206 | { | |
3207 | int i = 0; | |
3208 | struct list_head *p; | |
3209 | ||
e498be7d CL |
3210 | p = l3->slabs_free.next; |
3211 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3212 | struct slab *slabp; |
3213 | ||
3214 | slabp = list_entry(p, struct slab, list); | |
3215 | BUG_ON(slabp->inuse); | |
3216 | ||
3217 | i++; | |
3218 | p = p->next; | |
3219 | } | |
3220 | STATS_SET_FREEABLE(cachep, i); | |
3221 | } | |
3222 | #endif | |
e498be7d | 3223 | spin_unlock(&l3->list_lock); |
1da177e4 | 3224 | ac->avail -= batchcount; |
a737b3e2 | 3225 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3226 | } |
3227 | ||
3228 | /* | |
a737b3e2 AM |
3229 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3230 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3231 | */ |
873623df | 3232 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3233 | { |
9a2dba4b | 3234 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3235 | |
3236 | check_irq_off(); | |
3237 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3238 | ||
873623df | 3239 | if (cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3240 | return; |
3241 | ||
1da177e4 LT |
3242 | if (likely(ac->avail < ac->limit)) { |
3243 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3244 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3245 | return; |
3246 | } else { | |
3247 | STATS_INC_FREEMISS(cachep); | |
3248 | cache_flusharray(cachep, ac); | |
e498be7d | 3249 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3250 | } |
3251 | } | |
3252 | ||
3253 | /** | |
3254 | * kmem_cache_alloc - Allocate an object | |
3255 | * @cachep: The cache to allocate from. | |
3256 | * @flags: See kmalloc(). | |
3257 | * | |
3258 | * Allocate an object from this cache. The flags are only relevant | |
3259 | * if the cache has no available objects. | |
3260 | */ | |
343e0d7a | 3261 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3262 | { |
7fd6b141 | 3263 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3264 | } |
3265 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3266 | ||
a8c0f9a4 | 3267 | /** |
b8008b2b | 3268 | * kmem_cache_zalloc - Allocate an object. The memory is set to zero. |
a8c0f9a4 PE |
3269 | * @cache: The cache to allocate from. |
3270 | * @flags: See kmalloc(). | |
3271 | * | |
3272 | * Allocate an object from this cache and set the allocated memory to zero. | |
3273 | * The flags are only relevant if the cache has no available objects. | |
3274 | */ | |
3275 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3276 | { | |
3277 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3278 | if (ret) | |
3279 | memset(ret, 0, obj_size(cache)); | |
3280 | return ret; | |
3281 | } | |
3282 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3283 | ||
1da177e4 LT |
3284 | /** |
3285 | * kmem_ptr_validate - check if an untrusted pointer might | |
3286 | * be a slab entry. | |
3287 | * @cachep: the cache we're checking against | |
3288 | * @ptr: pointer to validate | |
3289 | * | |
3290 | * This verifies that the untrusted pointer looks sane: | |
3291 | * it is _not_ a guarantee that the pointer is actually | |
3292 | * part of the slab cache in question, but it at least | |
3293 | * validates that the pointer can be dereferenced and | |
3294 | * looks half-way sane. | |
3295 | * | |
3296 | * Currently only used for dentry validation. | |
3297 | */ | |
343e0d7a | 3298 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3299 | { |
b28a02de | 3300 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3301 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3302 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3303 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3304 | struct page *page; |
3305 | ||
3306 | if (unlikely(addr < min_addr)) | |
3307 | goto out; | |
3308 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3309 | goto out; | |
3310 | if (unlikely(addr & align_mask)) | |
3311 | goto out; | |
3312 | if (unlikely(!kern_addr_valid(addr))) | |
3313 | goto out; | |
3314 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3315 | goto out; | |
3316 | page = virt_to_page(ptr); | |
3317 | if (unlikely(!PageSlab(page))) | |
3318 | goto out; | |
065d41cb | 3319 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3320 | goto out; |
3321 | return 1; | |
a737b3e2 | 3322 | out: |
1da177e4 LT |
3323 | return 0; |
3324 | } | |
3325 | ||
3326 | #ifdef CONFIG_NUMA | |
3327 | /** | |
3328 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3329 | * @cachep: The cache to allocate from. | |
3330 | * @flags: See kmalloc(). | |
3331 | * @nodeid: node number of the target node. | |
3332 | * | |
3333 | * Identical to kmem_cache_alloc, except that this function is slow | |
3334 | * and can sleep. And it will allocate memory on the given node, which | |
3335 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3336 | * New and improved: it will now make sure that the object gets |
3337 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3338 | */ |
343e0d7a | 3339 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3340 | { |
e498be7d CL |
3341 | unsigned long save_flags; |
3342 | void *ptr; | |
1da177e4 | 3343 | |
e498be7d CL |
3344 | cache_alloc_debugcheck_before(cachep, flags); |
3345 | local_irq_save(save_flags); | |
18f820f6 CL |
3346 | |
3347 | if (nodeid == -1 || nodeid == numa_node_id() || | |
a737b3e2 | 3348 | !cachep->nodelists[nodeid]) |
5c382300 AK |
3349 | ptr = ____cache_alloc(cachep, flags); |
3350 | else | |
3351 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3352 | local_irq_restore(save_flags); |
18f820f6 CL |
3353 | |
3354 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3355 | __builtin_return_address(0)); | |
1da177e4 | 3356 | |
e498be7d | 3357 | return ptr; |
1da177e4 LT |
3358 | } |
3359 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3360 | ||
dbe5e69d | 3361 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3362 | { |
343e0d7a | 3363 | struct kmem_cache *cachep; |
97e2bde4 MS |
3364 | |
3365 | cachep = kmem_find_general_cachep(size, flags); | |
3366 | if (unlikely(cachep == NULL)) | |
3367 | return NULL; | |
3368 | return kmem_cache_alloc_node(cachep, flags, node); | |
3369 | } | |
dbe5e69d | 3370 | EXPORT_SYMBOL(__kmalloc_node); |
1da177e4 LT |
3371 | #endif |
3372 | ||
3373 | /** | |
800590f5 | 3374 | * __do_kmalloc - allocate memory |
1da177e4 | 3375 | * @size: how many bytes of memory are required. |
800590f5 | 3376 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3377 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3378 | */ |
7fd6b141 PE |
3379 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3380 | void *caller) | |
1da177e4 | 3381 | { |
343e0d7a | 3382 | struct kmem_cache *cachep; |
1da177e4 | 3383 | |
97e2bde4 MS |
3384 | /* If you want to save a few bytes .text space: replace |
3385 | * __ with kmem_. | |
3386 | * Then kmalloc uses the uninlined functions instead of the inline | |
3387 | * functions. | |
3388 | */ | |
3389 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3390 | if (unlikely(cachep == NULL)) |
3391 | return NULL; | |
7fd6b141 PE |
3392 | return __cache_alloc(cachep, flags, caller); |
3393 | } | |
3394 | ||
7fd6b141 PE |
3395 | |
3396 | void *__kmalloc(size_t size, gfp_t flags) | |
3397 | { | |
871751e2 | 3398 | #ifndef CONFIG_DEBUG_SLAB |
7fd6b141 | 3399 | return __do_kmalloc(size, flags, NULL); |
871751e2 AV |
3400 | #else |
3401 | return __do_kmalloc(size, flags, __builtin_return_address(0)); | |
3402 | #endif | |
1da177e4 LT |
3403 | } |
3404 | EXPORT_SYMBOL(__kmalloc); | |
3405 | ||
871751e2 | 3406 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3407 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3408 | { | |
3409 | return __do_kmalloc(size, flags, caller); | |
3410 | } | |
3411 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
7fd6b141 PE |
3412 | #endif |
3413 | ||
1da177e4 LT |
3414 | #ifdef CONFIG_SMP |
3415 | /** | |
7ff6f082 MP |
3416 | * percpu_depopulate - depopulate per-cpu data for given cpu |
3417 | * @__pdata: per-cpu data to depopulate | |
3418 | * @cpu: depopulate per-cpu data for this cpu | |
1da177e4 | 3419 | * |
7ff6f082 MP |
3420 | * Depopulating per-cpu data for a cpu going offline would be a typical |
3421 | * use case. You need to register a cpu hotplug handler for that purpose. | |
1da177e4 | 3422 | */ |
7ff6f082 | 3423 | void percpu_depopulate(void *__pdata, int cpu) |
1da177e4 | 3424 | { |
7ff6f082 MP |
3425 | struct percpu_data *pdata = __percpu_disguise(__pdata); |
3426 | if (pdata->ptrs[cpu]) { | |
3427 | kfree(pdata->ptrs[cpu]); | |
3428 | pdata->ptrs[cpu] = NULL; | |
3429 | } | |
3430 | } | |
3431 | EXPORT_SYMBOL_GPL(percpu_depopulate); | |
1da177e4 | 3432 | |
7ff6f082 MP |
3433 | /** |
3434 | * percpu_depopulate_mask - depopulate per-cpu data for some cpu's | |
3435 | * @__pdata: per-cpu data to depopulate | |
3436 | * @mask: depopulate per-cpu data for cpu's selected through mask bits | |
3437 | */ | |
3438 | void __percpu_depopulate_mask(void *__pdata, cpumask_t *mask) | |
3439 | { | |
3440 | int cpu; | |
3441 | for_each_cpu_mask(cpu, *mask) | |
3442 | percpu_depopulate(__pdata, cpu); | |
3443 | } | |
3444 | EXPORT_SYMBOL_GPL(__percpu_depopulate_mask); | |
1da177e4 | 3445 | |
7ff6f082 MP |
3446 | /** |
3447 | * percpu_populate - populate per-cpu data for given cpu | |
3448 | * @__pdata: per-cpu data to populate further | |
3449 | * @size: size of per-cpu object | |
3450 | * @gfp: may sleep or not etc. | |
3451 | * @cpu: populate per-data for this cpu | |
3452 | * | |
3453 | * Populating per-cpu data for a cpu coming online would be a typical | |
3454 | * use case. You need to register a cpu hotplug handler for that purpose. | |
3455 | * Per-cpu object is populated with zeroed buffer. | |
3456 | */ | |
3457 | void *percpu_populate(void *__pdata, size_t size, gfp_t gfp, int cpu) | |
3458 | { | |
3459 | struct percpu_data *pdata = __percpu_disguise(__pdata); | |
3460 | int node = cpu_to_node(cpu); | |
e498be7d | 3461 | |
7ff6f082 MP |
3462 | BUG_ON(pdata->ptrs[cpu]); |
3463 | if (node_online(node)) { | |
3464 | /* FIXME: kzalloc_node(size, gfp, node) */ | |
3465 | pdata->ptrs[cpu] = kmalloc_node(size, gfp, node); | |
3466 | if (pdata->ptrs[cpu]) | |
3467 | memset(pdata->ptrs[cpu], 0, size); | |
3468 | } else | |
3469 | pdata->ptrs[cpu] = kzalloc(size, gfp); | |
3470 | return pdata->ptrs[cpu]; | |
3471 | } | |
3472 | EXPORT_SYMBOL_GPL(percpu_populate); | |
1da177e4 | 3473 | |
7ff6f082 MP |
3474 | /** |
3475 | * percpu_populate_mask - populate per-cpu data for more cpu's | |
3476 | * @__pdata: per-cpu data to populate further | |
3477 | * @size: size of per-cpu object | |
3478 | * @gfp: may sleep or not etc. | |
3479 | * @mask: populate per-cpu data for cpu's selected through mask bits | |
3480 | * | |
3481 | * Per-cpu objects are populated with zeroed buffers. | |
3482 | */ | |
3483 | int __percpu_populate_mask(void *__pdata, size_t size, gfp_t gfp, | |
3484 | cpumask_t *mask) | |
3485 | { | |
3486 | cpumask_t populated = CPU_MASK_NONE; | |
3487 | int cpu; | |
1da177e4 | 3488 | |
7ff6f082 MP |
3489 | for_each_cpu_mask(cpu, *mask) |
3490 | if (unlikely(!percpu_populate(__pdata, size, gfp, cpu))) { | |
3491 | __percpu_depopulate_mask(__pdata, &populated); | |
3492 | return -ENOMEM; | |
3493 | } else | |
3494 | cpu_set(cpu, populated); | |
3495 | return 0; | |
3496 | } | |
3497 | EXPORT_SYMBOL_GPL(__percpu_populate_mask); | |
1da177e4 | 3498 | |
7ff6f082 MP |
3499 | /** |
3500 | * percpu_alloc_mask - initial setup of per-cpu data | |
3501 | * @size: size of per-cpu object | |
3502 | * @gfp: may sleep or not etc. | |
3503 | * @mask: populate per-data for cpu's selected through mask bits | |
3504 | * | |
3505 | * Populating per-cpu data for all online cpu's would be a typical use case, | |
3506 | * which is simplified by the percpu_alloc() wrapper. | |
3507 | * Per-cpu objects are populated with zeroed buffers. | |
3508 | */ | |
3509 | void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask) | |
3510 | { | |
3511 | void *pdata = kzalloc(sizeof(struct percpu_data), gfp); | |
3512 | void *__pdata = __percpu_disguise(pdata); | |
3513 | ||
3514 | if (unlikely(!pdata)) | |
3515 | return NULL; | |
3516 | if (likely(!__percpu_populate_mask(__pdata, size, gfp, mask))) | |
3517 | return __pdata; | |
1da177e4 LT |
3518 | kfree(pdata); |
3519 | return NULL; | |
3520 | } | |
7ff6f082 MP |
3521 | EXPORT_SYMBOL_GPL(__percpu_alloc_mask); |
3522 | ||
3523 | /** | |
3524 | * percpu_free - final cleanup of per-cpu data | |
3525 | * @__pdata: object to clean up | |
3526 | * | |
3527 | * We simply clean up any per-cpu object left. No need for the client to | |
3528 | * track and specify through a bis mask which per-cpu objects are to free. | |
3529 | */ | |
3530 | void percpu_free(void *__pdata) | |
3531 | { | |
3532 | __percpu_depopulate_mask(__pdata, &cpu_possible_map); | |
3533 | kfree(__percpu_disguise(__pdata)); | |
3534 | } | |
3535 | EXPORT_SYMBOL_GPL(percpu_free); | |
3536 | #endif /* CONFIG_SMP */ | |
1da177e4 LT |
3537 | |
3538 | /** | |
3539 | * kmem_cache_free - Deallocate an object | |
3540 | * @cachep: The cache the allocation was from. | |
3541 | * @objp: The previously allocated object. | |
3542 | * | |
3543 | * Free an object which was previously allocated from this | |
3544 | * cache. | |
3545 | */ | |
343e0d7a | 3546 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3547 | { |
3548 | unsigned long flags; | |
3549 | ||
ddc2e812 PE |
3550 | BUG_ON(virt_to_cache(objp) != cachep); |
3551 | ||
1da177e4 | 3552 | local_irq_save(flags); |
873623df | 3553 | __cache_free(cachep, objp); |
1da177e4 LT |
3554 | local_irq_restore(flags); |
3555 | } | |
3556 | EXPORT_SYMBOL(kmem_cache_free); | |
3557 | ||
1da177e4 LT |
3558 | /** |
3559 | * kfree - free previously allocated memory | |
3560 | * @objp: pointer returned by kmalloc. | |
3561 | * | |
80e93eff PE |
3562 | * If @objp is NULL, no operation is performed. |
3563 | * | |
1da177e4 LT |
3564 | * Don't free memory not originally allocated by kmalloc() |
3565 | * or you will run into trouble. | |
3566 | */ | |
3567 | void kfree(const void *objp) | |
3568 | { | |
343e0d7a | 3569 | struct kmem_cache *c; |
1da177e4 LT |
3570 | unsigned long flags; |
3571 | ||
3572 | if (unlikely(!objp)) | |
3573 | return; | |
3574 | local_irq_save(flags); | |
3575 | kfree_debugcheck(objp); | |
6ed5eb22 | 3576 | c = virt_to_cache(objp); |
f9b8404c | 3577 | debug_check_no_locks_freed(objp, obj_size(c)); |
873623df | 3578 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3579 | local_irq_restore(flags); |
3580 | } | |
3581 | EXPORT_SYMBOL(kfree); | |
3582 | ||
343e0d7a | 3583 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3584 | { |
3dafccf2 | 3585 | return obj_size(cachep); |
1da177e4 LT |
3586 | } |
3587 | EXPORT_SYMBOL(kmem_cache_size); | |
3588 | ||
343e0d7a | 3589 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3590 | { |
3591 | return cachep->name; | |
3592 | } | |
3593 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3594 | ||
e498be7d | 3595 | /* |
0718dc2a | 3596 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3597 | */ |
343e0d7a | 3598 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3599 | { |
3600 | int node; | |
3601 | struct kmem_list3 *l3; | |
cafeb02e CL |
3602 | struct array_cache *new_shared; |
3603 | struct array_cache **new_alien; | |
e498be7d CL |
3604 | |
3605 | for_each_online_node(node) { | |
cafeb02e | 3606 | |
a737b3e2 AM |
3607 | new_alien = alloc_alien_cache(node, cachep->limit); |
3608 | if (!new_alien) | |
e498be7d | 3609 | goto fail; |
cafeb02e | 3610 | |
0718dc2a CL |
3611 | new_shared = alloc_arraycache(node, |
3612 | cachep->shared*cachep->batchcount, | |
a737b3e2 | 3613 | 0xbaadf00d); |
0718dc2a CL |
3614 | if (!new_shared) { |
3615 | free_alien_cache(new_alien); | |
e498be7d | 3616 | goto fail; |
0718dc2a | 3617 | } |
cafeb02e | 3618 | |
a737b3e2 AM |
3619 | l3 = cachep->nodelists[node]; |
3620 | if (l3) { | |
cafeb02e CL |
3621 | struct array_cache *shared = l3->shared; |
3622 | ||
e498be7d CL |
3623 | spin_lock_irq(&l3->list_lock); |
3624 | ||
cafeb02e | 3625 | if (shared) |
0718dc2a CL |
3626 | free_block(cachep, shared->entry, |
3627 | shared->avail, node); | |
e498be7d | 3628 | |
cafeb02e CL |
3629 | l3->shared = new_shared; |
3630 | if (!l3->alien) { | |
e498be7d CL |
3631 | l3->alien = new_alien; |
3632 | new_alien = NULL; | |
3633 | } | |
b28a02de | 3634 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3635 | cachep->batchcount + cachep->num; |
e498be7d | 3636 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3637 | kfree(shared); |
e498be7d CL |
3638 | free_alien_cache(new_alien); |
3639 | continue; | |
3640 | } | |
a737b3e2 | 3641 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3642 | if (!l3) { |
3643 | free_alien_cache(new_alien); | |
3644 | kfree(new_shared); | |
e498be7d | 3645 | goto fail; |
0718dc2a | 3646 | } |
e498be7d CL |
3647 | |
3648 | kmem_list3_init(l3); | |
3649 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3650 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3651 | l3->shared = new_shared; |
e498be7d | 3652 | l3->alien = new_alien; |
b28a02de | 3653 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3654 | cachep->batchcount + cachep->num; |
e498be7d CL |
3655 | cachep->nodelists[node] = l3; |
3656 | } | |
cafeb02e | 3657 | return 0; |
0718dc2a | 3658 | |
a737b3e2 | 3659 | fail: |
0718dc2a CL |
3660 | if (!cachep->next.next) { |
3661 | /* Cache is not active yet. Roll back what we did */ | |
3662 | node--; | |
3663 | while (node >= 0) { | |
3664 | if (cachep->nodelists[node]) { | |
3665 | l3 = cachep->nodelists[node]; | |
3666 | ||
3667 | kfree(l3->shared); | |
3668 | free_alien_cache(l3->alien); | |
3669 | kfree(l3); | |
3670 | cachep->nodelists[node] = NULL; | |
3671 | } | |
3672 | node--; | |
3673 | } | |
3674 | } | |
cafeb02e | 3675 | return -ENOMEM; |
e498be7d CL |
3676 | } |
3677 | ||
1da177e4 | 3678 | struct ccupdate_struct { |
343e0d7a | 3679 | struct kmem_cache *cachep; |
1da177e4 LT |
3680 | struct array_cache *new[NR_CPUS]; |
3681 | }; | |
3682 | ||
3683 | static void do_ccupdate_local(void *info) | |
3684 | { | |
a737b3e2 | 3685 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3686 | struct array_cache *old; |
3687 | ||
3688 | check_irq_off(); | |
9a2dba4b | 3689 | old = cpu_cache_get(new->cachep); |
e498be7d | 3690 | |
1da177e4 LT |
3691 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3692 | new->new[smp_processor_id()] = old; | |
3693 | } | |
3694 | ||
b5d8ca7c | 3695 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3696 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3697 | int batchcount, int shared) | |
1da177e4 LT |
3698 | { |
3699 | struct ccupdate_struct new; | |
2ed3a4ef | 3700 | int i; |
1da177e4 | 3701 | |
b28a02de | 3702 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3703 | for_each_online_cpu(i) { |
a737b3e2 AM |
3704 | new.new[i] = alloc_arraycache(cpu_to_node(i), limit, |
3705 | batchcount); | |
e498be7d | 3706 | if (!new.new[i]) { |
b28a02de PE |
3707 | for (i--; i >= 0; i--) |
3708 | kfree(new.new[i]); | |
e498be7d | 3709 | return -ENOMEM; |
1da177e4 LT |
3710 | } |
3711 | } | |
3712 | new.cachep = cachep; | |
3713 | ||
a07fa394 | 3714 | on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1); |
e498be7d | 3715 | |
1da177e4 | 3716 | check_irq_on(); |
1da177e4 LT |
3717 | cachep->batchcount = batchcount; |
3718 | cachep->limit = limit; | |
e498be7d | 3719 | cachep->shared = shared; |
1da177e4 | 3720 | |
e498be7d | 3721 | for_each_online_cpu(i) { |
1da177e4 LT |
3722 | struct array_cache *ccold = new.new[i]; |
3723 | if (!ccold) | |
3724 | continue; | |
e498be7d | 3725 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3726 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3727 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3728 | kfree(ccold); |
3729 | } | |
1da177e4 | 3730 | |
2ed3a4ef | 3731 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3732 | } |
3733 | ||
b5d8ca7c | 3734 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3735 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3736 | { |
3737 | int err; | |
3738 | int limit, shared; | |
3739 | ||
a737b3e2 AM |
3740 | /* |
3741 | * The head array serves three purposes: | |
1da177e4 LT |
3742 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3743 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3744 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3745 | * bufctl chains: array operations are cheaper. |
3746 | * The numbers are guessed, we should auto-tune as described by | |
3747 | * Bonwick. | |
3748 | */ | |
3dafccf2 | 3749 | if (cachep->buffer_size > 131072) |
1da177e4 | 3750 | limit = 1; |
3dafccf2 | 3751 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3752 | limit = 8; |
3dafccf2 | 3753 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3754 | limit = 24; |
3dafccf2 | 3755 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3756 | limit = 54; |
3757 | else | |
3758 | limit = 120; | |
3759 | ||
a737b3e2 AM |
3760 | /* |
3761 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3762 | * allocation behaviour: Most allocs on one cpu, most free operations |
3763 | * on another cpu. For these cases, an efficient object passing between | |
3764 | * cpus is necessary. This is provided by a shared array. The array | |
3765 | * replaces Bonwick's magazine layer. | |
3766 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3767 | * to a larger limit. Thus disabled by default. | |
3768 | */ | |
3769 | shared = 0; | |
3770 | #ifdef CONFIG_SMP | |
3dafccf2 | 3771 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3772 | shared = 8; |
3773 | #endif | |
3774 | ||
3775 | #if DEBUG | |
a737b3e2 AM |
3776 | /* |
3777 | * With debugging enabled, large batchcount lead to excessively long | |
3778 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3779 | */ |
3780 | if (limit > 32) | |
3781 | limit = 32; | |
3782 | #endif | |
b28a02de | 3783 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3784 | if (err) |
3785 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3786 | cachep->name, -err); |
2ed3a4ef | 3787 | return err; |
1da177e4 LT |
3788 | } |
3789 | ||
1b55253a CL |
3790 | /* |
3791 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3792 | * necessary. Note that the l3 listlock also protects the array_cache |
3793 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3794 | */ |
3795 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
3796 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
3797 | { |
3798 | int tofree; | |
3799 | ||
1b55253a CL |
3800 | if (!ac || !ac->avail) |
3801 | return; | |
1da177e4 LT |
3802 | if (ac->touched && !force) { |
3803 | ac->touched = 0; | |
b18e7e65 | 3804 | } else { |
1b55253a | 3805 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
3806 | if (ac->avail) { |
3807 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3808 | if (tofree > ac->avail) | |
3809 | tofree = (ac->avail + 1) / 2; | |
3810 | free_block(cachep, ac->entry, tofree, node); | |
3811 | ac->avail -= tofree; | |
3812 | memmove(ac->entry, &(ac->entry[tofree]), | |
3813 | sizeof(void *) * ac->avail); | |
3814 | } | |
1b55253a | 3815 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
3816 | } |
3817 | } | |
3818 | ||
3819 | /** | |
3820 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3821 | * @unused: unused parameter |
1da177e4 LT |
3822 | * |
3823 | * Called from workqueue/eventd every few seconds. | |
3824 | * Purpose: | |
3825 | * - clear the per-cpu caches for this CPU. | |
3826 | * - return freeable pages to the main free memory pool. | |
3827 | * | |
a737b3e2 AM |
3828 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3829 | * again on the next iteration. | |
1da177e4 LT |
3830 | */ |
3831 | static void cache_reap(void *unused) | |
3832 | { | |
7a7c381d | 3833 | struct kmem_cache *searchp; |
e498be7d | 3834 | struct kmem_list3 *l3; |
aab2207c | 3835 | int node = numa_node_id(); |
1da177e4 | 3836 | |
fc0abb14 | 3837 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3838 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3839 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3840 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3841 | return; |
3842 | } | |
3843 | ||
7a7c381d | 3844 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
3845 | check_irq_on(); |
3846 | ||
35386e3b CL |
3847 | /* |
3848 | * We only take the l3 lock if absolutely necessary and we | |
3849 | * have established with reasonable certainty that | |
3850 | * we can do some work if the lock was obtained. | |
3851 | */ | |
aab2207c | 3852 | l3 = searchp->nodelists[node]; |
35386e3b | 3853 | |
8fce4d8e | 3854 | reap_alien(searchp, l3); |
1da177e4 | 3855 | |
aab2207c | 3856 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 3857 | |
35386e3b CL |
3858 | /* |
3859 | * These are racy checks but it does not matter | |
3860 | * if we skip one check or scan twice. | |
3861 | */ | |
e498be7d | 3862 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 3863 | goto next; |
1da177e4 | 3864 | |
e498be7d | 3865 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3866 | |
aab2207c | 3867 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 3868 | |
ed11d9eb | 3869 | if (l3->free_touched) |
e498be7d | 3870 | l3->free_touched = 0; |
ed11d9eb CL |
3871 | else { |
3872 | int freed; | |
1da177e4 | 3873 | |
ed11d9eb CL |
3874 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
3875 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
3876 | STATS_ADD_REAPED(searchp, freed); | |
3877 | } | |
35386e3b | 3878 | next: |
1da177e4 LT |
3879 | cond_resched(); |
3880 | } | |
3881 | check_irq_on(); | |
fc0abb14 | 3882 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 3883 | next_reap_node(); |
2244b95a | 3884 | refresh_cpu_vm_stats(smp_processor_id()); |
a737b3e2 | 3885 | /* Set up the next iteration */ |
cd61ef62 | 3886 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3887 | } |
3888 | ||
3889 | #ifdef CONFIG_PROC_FS | |
3890 | ||
85289f98 | 3891 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3892 | { |
85289f98 PE |
3893 | /* |
3894 | * Output format version, so at least we can change it | |
3895 | * without _too_ many complaints. | |
3896 | */ | |
1da177e4 | 3897 | #if STATS |
85289f98 | 3898 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3899 | #else |
85289f98 | 3900 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3901 | #endif |
85289f98 PE |
3902 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3903 | "<objperslab> <pagesperslab>"); | |
3904 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3905 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3906 | #if STATS |
85289f98 | 3907 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 3908 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 3909 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 3910 | #endif |
85289f98 PE |
3911 | seq_putc(m, '\n'); |
3912 | } | |
3913 | ||
3914 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3915 | { | |
3916 | loff_t n = *pos; | |
3917 | struct list_head *p; | |
3918 | ||
fc0abb14 | 3919 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3920 | if (!n) |
3921 | print_slabinfo_header(m); | |
1da177e4 LT |
3922 | p = cache_chain.next; |
3923 | while (n--) { | |
3924 | p = p->next; | |
3925 | if (p == &cache_chain) | |
3926 | return NULL; | |
3927 | } | |
343e0d7a | 3928 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3929 | } |
3930 | ||
3931 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3932 | { | |
343e0d7a | 3933 | struct kmem_cache *cachep = p; |
1da177e4 | 3934 | ++*pos; |
a737b3e2 AM |
3935 | return cachep->next.next == &cache_chain ? |
3936 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
3937 | } |
3938 | ||
3939 | static void s_stop(struct seq_file *m, void *p) | |
3940 | { | |
fc0abb14 | 3941 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3942 | } |
3943 | ||
3944 | static int s_show(struct seq_file *m, void *p) | |
3945 | { | |
343e0d7a | 3946 | struct kmem_cache *cachep = p; |
b28a02de PE |
3947 | struct slab *slabp; |
3948 | unsigned long active_objs; | |
3949 | unsigned long num_objs; | |
3950 | unsigned long active_slabs = 0; | |
3951 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3952 | const char *name; |
1da177e4 | 3953 | char *error = NULL; |
e498be7d CL |
3954 | int node; |
3955 | struct kmem_list3 *l3; | |
1da177e4 | 3956 | |
1da177e4 LT |
3957 | active_objs = 0; |
3958 | num_slabs = 0; | |
e498be7d CL |
3959 | for_each_online_node(node) { |
3960 | l3 = cachep->nodelists[node]; | |
3961 | if (!l3) | |
3962 | continue; | |
3963 | ||
ca3b9b91 RT |
3964 | check_irq_on(); |
3965 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3966 | |
7a7c381d | 3967 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
3968 | if (slabp->inuse != cachep->num && !error) |
3969 | error = "slabs_full accounting error"; | |
3970 | active_objs += cachep->num; | |
3971 | active_slabs++; | |
3972 | } | |
7a7c381d | 3973 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
3974 | if (slabp->inuse == cachep->num && !error) |
3975 | error = "slabs_partial inuse accounting error"; | |
3976 | if (!slabp->inuse && !error) | |
3977 | error = "slabs_partial/inuse accounting error"; | |
3978 | active_objs += slabp->inuse; | |
3979 | active_slabs++; | |
3980 | } | |
7a7c381d | 3981 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
3982 | if (slabp->inuse && !error) |
3983 | error = "slabs_free/inuse accounting error"; | |
3984 | num_slabs++; | |
3985 | } | |
3986 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3987 | if (l3->shared) |
3988 | shared_avail += l3->shared->avail; | |
e498be7d | 3989 | |
ca3b9b91 | 3990 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3991 | } |
b28a02de PE |
3992 | num_slabs += active_slabs; |
3993 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3994 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3995 | error = "free_objects accounting error"; |
3996 | ||
b28a02de | 3997 | name = cachep->name; |
1da177e4 LT |
3998 | if (error) |
3999 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4000 | ||
4001 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4002 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4003 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4004 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4005 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4006 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4007 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4008 | #if STATS |
b28a02de | 4009 | { /* list3 stats */ |
1da177e4 LT |
4010 | unsigned long high = cachep->high_mark; |
4011 | unsigned long allocs = cachep->num_allocations; | |
4012 | unsigned long grown = cachep->grown; | |
4013 | unsigned long reaped = cachep->reaped; | |
4014 | unsigned long errors = cachep->errors; | |
4015 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4016 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4017 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4018 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4019 | |
e498be7d | 4020 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4021 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4022 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4023 | node_frees, overflows); |
1da177e4 LT |
4024 | } |
4025 | /* cpu stats */ | |
4026 | { | |
4027 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4028 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4029 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4030 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4031 | ||
4032 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4033 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4034 | } |
4035 | #endif | |
4036 | seq_putc(m, '\n'); | |
1da177e4 LT |
4037 | return 0; |
4038 | } | |
4039 | ||
4040 | /* | |
4041 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4042 | * | |
4043 | * Output layout: | |
4044 | * cache-name | |
4045 | * num-active-objs | |
4046 | * total-objs | |
4047 | * object size | |
4048 | * num-active-slabs | |
4049 | * total-slabs | |
4050 | * num-pages-per-slab | |
4051 | * + further values on SMP and with statistics enabled | |
4052 | */ | |
4053 | ||
4054 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
4055 | .start = s_start, |
4056 | .next = s_next, | |
4057 | .stop = s_stop, | |
4058 | .show = s_show, | |
1da177e4 LT |
4059 | }; |
4060 | ||
4061 | #define MAX_SLABINFO_WRITE 128 | |
4062 | /** | |
4063 | * slabinfo_write - Tuning for the slab allocator | |
4064 | * @file: unused | |
4065 | * @buffer: user buffer | |
4066 | * @count: data length | |
4067 | * @ppos: unused | |
4068 | */ | |
b28a02de PE |
4069 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4070 | size_t count, loff_t *ppos) | |
1da177e4 | 4071 | { |
b28a02de | 4072 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4073 | int limit, batchcount, shared, res; |
7a7c381d | 4074 | struct kmem_cache *cachep; |
b28a02de | 4075 | |
1da177e4 LT |
4076 | if (count > MAX_SLABINFO_WRITE) |
4077 | return -EINVAL; | |
4078 | if (copy_from_user(&kbuf, buffer, count)) | |
4079 | return -EFAULT; | |
b28a02de | 4080 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4081 | |
4082 | tmp = strchr(kbuf, ' '); | |
4083 | if (!tmp) | |
4084 | return -EINVAL; | |
4085 | *tmp = '\0'; | |
4086 | tmp++; | |
4087 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4088 | return -EINVAL; | |
4089 | ||
4090 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4091 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4092 | res = -EINVAL; |
7a7c381d | 4093 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4094 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4095 | if (limit < 1 || batchcount < 1 || |
4096 | batchcount > limit || shared < 0) { | |
e498be7d | 4097 | res = 0; |
1da177e4 | 4098 | } else { |
e498be7d | 4099 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4100 | batchcount, shared); |
1da177e4 LT |
4101 | } |
4102 | break; | |
4103 | } | |
4104 | } | |
fc0abb14 | 4105 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4106 | if (res >= 0) |
4107 | res = count; | |
4108 | return res; | |
4109 | } | |
871751e2 AV |
4110 | |
4111 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4112 | ||
4113 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4114 | { | |
4115 | loff_t n = *pos; | |
4116 | struct list_head *p; | |
4117 | ||
4118 | mutex_lock(&cache_chain_mutex); | |
4119 | p = cache_chain.next; | |
4120 | while (n--) { | |
4121 | p = p->next; | |
4122 | if (p == &cache_chain) | |
4123 | return NULL; | |
4124 | } | |
4125 | return list_entry(p, struct kmem_cache, next); | |
4126 | } | |
4127 | ||
4128 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4129 | { | |
4130 | unsigned long *p; | |
4131 | int l; | |
4132 | if (!v) | |
4133 | return 1; | |
4134 | l = n[1]; | |
4135 | p = n + 2; | |
4136 | while (l) { | |
4137 | int i = l/2; | |
4138 | unsigned long *q = p + 2 * i; | |
4139 | if (*q == v) { | |
4140 | q[1]++; | |
4141 | return 1; | |
4142 | } | |
4143 | if (*q > v) { | |
4144 | l = i; | |
4145 | } else { | |
4146 | p = q + 2; | |
4147 | l -= i + 1; | |
4148 | } | |
4149 | } | |
4150 | if (++n[1] == n[0]) | |
4151 | return 0; | |
4152 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4153 | p[0] = v; | |
4154 | p[1] = 1; | |
4155 | return 1; | |
4156 | } | |
4157 | ||
4158 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4159 | { | |
4160 | void *p; | |
4161 | int i; | |
4162 | if (n[0] == n[1]) | |
4163 | return; | |
4164 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4165 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4166 | continue; | |
4167 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4168 | return; | |
4169 | } | |
4170 | } | |
4171 | ||
4172 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4173 | { | |
4174 | #ifdef CONFIG_KALLSYMS | |
4175 | char *modname; | |
4176 | const char *name; | |
4177 | unsigned long offset, size; | |
4178 | char namebuf[KSYM_NAME_LEN+1]; | |
4179 | ||
4180 | name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); | |
4181 | ||
4182 | if (name) { | |
4183 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4184 | if (modname) | |
4185 | seq_printf(m, " [%s]", modname); | |
4186 | return; | |
4187 | } | |
4188 | #endif | |
4189 | seq_printf(m, "%p", (void *)address); | |
4190 | } | |
4191 | ||
4192 | static int leaks_show(struct seq_file *m, void *p) | |
4193 | { | |
4194 | struct kmem_cache *cachep = p; | |
871751e2 AV |
4195 | struct slab *slabp; |
4196 | struct kmem_list3 *l3; | |
4197 | const char *name; | |
4198 | unsigned long *n = m->private; | |
4199 | int node; | |
4200 | int i; | |
4201 | ||
4202 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4203 | return 0; | |
4204 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4205 | return 0; | |
4206 | ||
4207 | /* OK, we can do it */ | |
4208 | ||
4209 | n[1] = 0; | |
4210 | ||
4211 | for_each_online_node(node) { | |
4212 | l3 = cachep->nodelists[node]; | |
4213 | if (!l3) | |
4214 | continue; | |
4215 | ||
4216 | check_irq_on(); | |
4217 | spin_lock_irq(&l3->list_lock); | |
4218 | ||
7a7c381d | 4219 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4220 | handle_slab(n, cachep, slabp); |
7a7c381d | 4221 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4222 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4223 | spin_unlock_irq(&l3->list_lock); |
4224 | } | |
4225 | name = cachep->name; | |
4226 | if (n[0] == n[1]) { | |
4227 | /* Increase the buffer size */ | |
4228 | mutex_unlock(&cache_chain_mutex); | |
4229 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4230 | if (!m->private) { | |
4231 | /* Too bad, we are really out */ | |
4232 | m->private = n; | |
4233 | mutex_lock(&cache_chain_mutex); | |
4234 | return -ENOMEM; | |
4235 | } | |
4236 | *(unsigned long *)m->private = n[0] * 2; | |
4237 | kfree(n); | |
4238 | mutex_lock(&cache_chain_mutex); | |
4239 | /* Now make sure this entry will be retried */ | |
4240 | m->count = m->size; | |
4241 | return 0; | |
4242 | } | |
4243 | for (i = 0; i < n[1]; i++) { | |
4244 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4245 | show_symbol(m, n[2*i+2]); | |
4246 | seq_putc(m, '\n'); | |
4247 | } | |
4248 | return 0; | |
4249 | } | |
4250 | ||
4251 | struct seq_operations slabstats_op = { | |
4252 | .start = leaks_start, | |
4253 | .next = s_next, | |
4254 | .stop = s_stop, | |
4255 | .show = leaks_show, | |
4256 | }; | |
4257 | #endif | |
1da177e4 LT |
4258 | #endif |
4259 | ||
00e145b6 MS |
4260 | /** |
4261 | * ksize - get the actual amount of memory allocated for a given object | |
4262 | * @objp: Pointer to the object | |
4263 | * | |
4264 | * kmalloc may internally round up allocations and return more memory | |
4265 | * than requested. ksize() can be used to determine the actual amount of | |
4266 | * memory allocated. The caller may use this additional memory, even though | |
4267 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4268 | * The caller must guarantee that objp points to a valid object previously | |
4269 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4270 | * must not be freed during the duration of the call. | |
4271 | */ | |
1da177e4 LT |
4272 | unsigned int ksize(const void *objp) |
4273 | { | |
00e145b6 MS |
4274 | if (unlikely(objp == NULL)) |
4275 | return 0; | |
1da177e4 | 4276 | |
6ed5eb22 | 4277 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4278 | } |