]>
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 | ||
1da177e4 LT |
89 | #include <linux/slab.h> |
90 | #include <linux/mm.h> | |
c9cf5528 | 91 | #include <linux/poison.h> |
1da177e4 LT |
92 | #include <linux/swap.h> |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
101a5001 | 97 | #include <linux/cpuset.h> |
1da177e4 LT |
98 | #include <linux/seq_file.h> |
99 | #include <linux/notifier.h> | |
100 | #include <linux/kallsyms.h> | |
101 | #include <linux/cpu.h> | |
102 | #include <linux/sysctl.h> | |
103 | #include <linux/module.h> | |
104 | #include <linux/rcupdate.h> | |
543537bd | 105 | #include <linux/string.h> |
e498be7d | 106 | #include <linux/nodemask.h> |
dc85da15 | 107 | #include <linux/mempolicy.h> |
fc0abb14 | 108 | #include <linux/mutex.h> |
e7eebaf6 | 109 | #include <linux/rtmutex.h> |
1da177e4 LT |
110 | |
111 | #include <asm/uaccess.h> | |
112 | #include <asm/cacheflush.h> | |
113 | #include <asm/tlbflush.h> | |
114 | #include <asm/page.h> | |
115 | ||
116 | /* | |
117 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
118 | * SLAB_RED_ZONE & SLAB_POISON. | |
119 | * 0 for faster, smaller code (especially in the critical paths). | |
120 | * | |
121 | * STATS - 1 to collect stats for /proc/slabinfo. | |
122 | * 0 for faster, smaller code (especially in the critical paths). | |
123 | * | |
124 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
125 | */ | |
126 | ||
127 | #ifdef CONFIG_DEBUG_SLAB | |
128 | #define DEBUG 1 | |
129 | #define STATS 1 | |
130 | #define FORCED_DEBUG 1 | |
131 | #else | |
132 | #define DEBUG 0 | |
133 | #define STATS 0 | |
134 | #define FORCED_DEBUG 0 | |
135 | #endif | |
136 | ||
1da177e4 LT |
137 | /* Shouldn't this be in a header file somewhere? */ |
138 | #define BYTES_PER_WORD sizeof(void *) | |
139 | ||
140 | #ifndef cache_line_size | |
141 | #define cache_line_size() L1_CACHE_BYTES | |
142 | #endif | |
143 | ||
144 | #ifndef ARCH_KMALLOC_MINALIGN | |
145 | /* | |
146 | * Enforce a minimum alignment for the kmalloc caches. | |
147 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
148 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
149 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
150 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
151 | * Note that this flag disables some debug features. | |
152 | */ | |
153 | #define ARCH_KMALLOC_MINALIGN 0 | |
154 | #endif | |
155 | ||
156 | #ifndef ARCH_SLAB_MINALIGN | |
157 | /* | |
158 | * Enforce a minimum alignment for all caches. | |
159 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
160 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
161 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
162 | * some debug features. | |
163 | */ | |
164 | #define ARCH_SLAB_MINALIGN 0 | |
165 | #endif | |
166 | ||
167 | #ifndef ARCH_KMALLOC_FLAGS | |
168 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
169 | #endif | |
170 | ||
171 | /* Legal flag mask for kmem_cache_create(). */ | |
172 | #if DEBUG | |
173 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
174 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
ac2b898c | 175 | SLAB_CACHE_DMA | \ |
1da177e4 LT |
176 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ |
177 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 178 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 179 | #else |
ac2b898c | 180 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
1da177e4 LT |
181 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ |
182 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 183 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 LT |
184 | #endif |
185 | ||
186 | /* | |
187 | * kmem_bufctl_t: | |
188 | * | |
189 | * Bufctl's are used for linking objs within a slab | |
190 | * linked offsets. | |
191 | * | |
192 | * This implementation relies on "struct page" for locating the cache & | |
193 | * slab an object belongs to. | |
194 | * This allows the bufctl structure to be small (one int), but limits | |
195 | * the number of objects a slab (not a cache) can contain when off-slab | |
196 | * bufctls are used. The limit is the size of the largest general cache | |
197 | * that does not use off-slab slabs. | |
198 | * For 32bit archs with 4 kB pages, is this 56. | |
199 | * This is not serious, as it is only for large objects, when it is unwise | |
200 | * to have too many per slab. | |
201 | * Note: This limit can be raised by introducing a general cache whose size | |
202 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
203 | */ | |
204 | ||
fa5b08d5 | 205 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
206 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
207 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
208 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
209 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 210 | |
1da177e4 LT |
211 | /* |
212 | * struct slab | |
213 | * | |
214 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
215 | * for a slab, or allocated from an general cache. | |
216 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
217 | */ | |
218 | struct slab { | |
b28a02de PE |
219 | struct list_head list; |
220 | unsigned long colouroff; | |
221 | void *s_mem; /* including colour offset */ | |
222 | unsigned int inuse; /* num of objs active in slab */ | |
223 | kmem_bufctl_t free; | |
224 | unsigned short nodeid; | |
1da177e4 LT |
225 | }; |
226 | ||
227 | /* | |
228 | * struct slab_rcu | |
229 | * | |
230 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
231 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
232 | * we need to approach a kernel structure obliquely, from its address | |
233 | * obtained without the usual locking. We can lock the structure to | |
234 | * stabilize it and check it's still at the given address, only if we | |
235 | * can be sure that the memory has not been meanwhile reused for some | |
236 | * other kind of object (which our subsystem's lock might corrupt). | |
237 | * | |
238 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
239 | * taking the spinlock within the structure expected at that address. | |
240 | * | |
241 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
242 | */ | |
243 | struct slab_rcu { | |
b28a02de | 244 | struct rcu_head head; |
343e0d7a | 245 | struct kmem_cache *cachep; |
b28a02de | 246 | void *addr; |
1da177e4 LT |
247 | }; |
248 | ||
249 | /* | |
250 | * struct array_cache | |
251 | * | |
1da177e4 LT |
252 | * Purpose: |
253 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
254 | * - reduce the number of linked list operations | |
255 | * - reduce spinlock operations | |
256 | * | |
257 | * The limit is stored in the per-cpu structure to reduce the data cache | |
258 | * footprint. | |
259 | * | |
260 | */ | |
261 | struct array_cache { | |
262 | unsigned int avail; | |
263 | unsigned int limit; | |
264 | unsigned int batchcount; | |
265 | unsigned int touched; | |
e498be7d | 266 | spinlock_t lock; |
a737b3e2 AM |
267 | void *entry[0]; /* |
268 | * Must have this definition in here for the proper | |
269 | * alignment of array_cache. Also simplifies accessing | |
270 | * the entries. | |
271 | * [0] is for gcc 2.95. It should really be []. | |
272 | */ | |
1da177e4 LT |
273 | }; |
274 | ||
a737b3e2 AM |
275 | /* |
276 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
277 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
278 | */ |
279 | #define BOOT_CPUCACHE_ENTRIES 1 | |
280 | struct arraycache_init { | |
281 | struct array_cache cache; | |
b28a02de | 282 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
283 | }; |
284 | ||
285 | /* | |
e498be7d | 286 | * The slab lists for all objects. |
1da177e4 LT |
287 | */ |
288 | struct kmem_list3 { | |
b28a02de PE |
289 | struct list_head slabs_partial; /* partial list first, better asm code */ |
290 | struct list_head slabs_full; | |
291 | struct list_head slabs_free; | |
292 | unsigned long free_objects; | |
b28a02de | 293 | unsigned int free_limit; |
2e1217cf | 294 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
295 | spinlock_t list_lock; |
296 | struct array_cache *shared; /* shared per node */ | |
297 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
298 | unsigned long next_reap; /* updated without locking */ |
299 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
300 | }; |
301 | ||
e498be7d CL |
302 | /* |
303 | * Need this for bootstrapping a per node allocator. | |
304 | */ | |
305 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
306 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
307 | #define CACHE_CACHE 0 | |
308 | #define SIZE_AC 1 | |
309 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
310 | ||
ed11d9eb CL |
311 | static int drain_freelist(struct kmem_cache *cache, |
312 | struct kmem_list3 *l3, int tofree); | |
313 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
314 | int node); | |
2ed3a4ef | 315 | static int enable_cpucache(struct kmem_cache *cachep); |
ed11d9eb CL |
316 | static void cache_reap(void *unused); |
317 | ||
e498be7d | 318 | /* |
a737b3e2 AM |
319 | * This function must be completely optimized away if a constant is passed to |
320 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 321 | */ |
7243cc05 | 322 | static __always_inline int index_of(const size_t size) |
e498be7d | 323 | { |
5ec8a847 SR |
324 | extern void __bad_size(void); |
325 | ||
e498be7d CL |
326 | if (__builtin_constant_p(size)) { |
327 | int i = 0; | |
328 | ||
329 | #define CACHE(x) \ | |
330 | if (size <=x) \ | |
331 | return i; \ | |
332 | else \ | |
333 | i++; | |
334 | #include "linux/kmalloc_sizes.h" | |
335 | #undef CACHE | |
5ec8a847 | 336 | __bad_size(); |
7243cc05 | 337 | } else |
5ec8a847 | 338 | __bad_size(); |
e498be7d CL |
339 | return 0; |
340 | } | |
341 | ||
e0a42726 IM |
342 | static int slab_early_init = 1; |
343 | ||
e498be7d CL |
344 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
345 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 346 | |
5295a74c | 347 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
348 | { |
349 | INIT_LIST_HEAD(&parent->slabs_full); | |
350 | INIT_LIST_HEAD(&parent->slabs_partial); | |
351 | INIT_LIST_HEAD(&parent->slabs_free); | |
352 | parent->shared = NULL; | |
353 | parent->alien = NULL; | |
2e1217cf | 354 | parent->colour_next = 0; |
e498be7d CL |
355 | spin_lock_init(&parent->list_lock); |
356 | parent->free_objects = 0; | |
357 | parent->free_touched = 0; | |
358 | } | |
359 | ||
a737b3e2 AM |
360 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
361 | do { \ | |
362 | INIT_LIST_HEAD(listp); \ | |
363 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
364 | } while (0) |
365 | ||
a737b3e2 AM |
366 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
367 | do { \ | |
e498be7d CL |
368 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
369 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
370 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
371 | } while (0) | |
1da177e4 LT |
372 | |
373 | /* | |
343e0d7a | 374 | * struct kmem_cache |
1da177e4 LT |
375 | * |
376 | * manages a cache. | |
377 | */ | |
b28a02de | 378 | |
2109a2d1 | 379 | struct kmem_cache { |
1da177e4 | 380 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 381 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 382 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
383 | unsigned int batchcount; |
384 | unsigned int limit; | |
385 | unsigned int shared; | |
b5d8ca7c | 386 | |
3dafccf2 | 387 | unsigned int buffer_size; |
b5d8ca7c | 388 | /* 3) touched by every alloc & free from the backend */ |
b28a02de | 389 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
b5d8ca7c | 390 | |
a737b3e2 AM |
391 | unsigned int flags; /* constant flags */ |
392 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 393 | |
b5d8ca7c | 394 | /* 4) cache_grow/shrink */ |
1da177e4 | 395 | /* order of pgs per slab (2^n) */ |
b28a02de | 396 | unsigned int gfporder; |
1da177e4 LT |
397 | |
398 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 399 | gfp_t gfpflags; |
1da177e4 | 400 | |
a737b3e2 | 401 | size_t colour; /* cache colouring range */ |
b28a02de | 402 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 403 | struct kmem_cache *slabp_cache; |
b28a02de | 404 | unsigned int slab_size; |
a737b3e2 | 405 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
406 | |
407 | /* constructor func */ | |
343e0d7a | 408 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
409 | |
410 | /* de-constructor func */ | |
343e0d7a | 411 | void (*dtor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 | 412 | |
b5d8ca7c | 413 | /* 5) cache creation/removal */ |
b28a02de PE |
414 | const char *name; |
415 | struct list_head next; | |
1da177e4 | 416 | |
b5d8ca7c | 417 | /* 6) statistics */ |
1da177e4 | 418 | #if STATS |
b28a02de PE |
419 | unsigned long num_active; |
420 | unsigned long num_allocations; | |
421 | unsigned long high_mark; | |
422 | unsigned long grown; | |
423 | unsigned long reaped; | |
424 | unsigned long errors; | |
425 | unsigned long max_freeable; | |
426 | unsigned long node_allocs; | |
427 | unsigned long node_frees; | |
fb7faf33 | 428 | unsigned long node_overflow; |
b28a02de PE |
429 | atomic_t allochit; |
430 | atomic_t allocmiss; | |
431 | atomic_t freehit; | |
432 | atomic_t freemiss; | |
1da177e4 LT |
433 | #endif |
434 | #if DEBUG | |
3dafccf2 MS |
435 | /* |
436 | * If debugging is enabled, then the allocator can add additional | |
437 | * fields and/or padding to every object. buffer_size contains the total | |
438 | * object size including these internal fields, the following two | |
439 | * variables contain the offset to the user object and its size. | |
440 | */ | |
441 | int obj_offset; | |
442 | int obj_size; | |
1da177e4 LT |
443 | #endif |
444 | }; | |
445 | ||
446 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
447 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
448 | ||
449 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
450 | /* |
451 | * Optimization question: fewer reaps means less probability for unnessary | |
452 | * cpucache drain/refill cycles. | |
1da177e4 | 453 | * |
dc6f3f27 | 454 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
455 | * which could lock up otherwise freeable slabs. |
456 | */ | |
457 | #define REAPTIMEOUT_CPUC (2*HZ) | |
458 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
459 | ||
460 | #if STATS | |
461 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
462 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
463 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
464 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 465 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
466 | #define STATS_SET_HIGH(x) \ |
467 | do { \ | |
468 | if ((x)->num_active > (x)->high_mark) \ | |
469 | (x)->high_mark = (x)->num_active; \ | |
470 | } while (0) | |
1da177e4 LT |
471 | #define STATS_INC_ERR(x) ((x)->errors++) |
472 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 473 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 474 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
475 | #define STATS_SET_FREEABLE(x, i) \ |
476 | do { \ | |
477 | if ((x)->max_freeable < i) \ | |
478 | (x)->max_freeable = i; \ | |
479 | } while (0) | |
1da177e4 LT |
480 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
481 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
482 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
483 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
484 | #else | |
485 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
486 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
487 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
488 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 489 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
490 | #define STATS_SET_HIGH(x) do { } while (0) |
491 | #define STATS_INC_ERR(x) do { } while (0) | |
492 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 493 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 494 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 495 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
496 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
497 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
498 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
499 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
500 | #endif | |
501 | ||
502 | #if DEBUG | |
1da177e4 | 503 | |
a737b3e2 AM |
504 | /* |
505 | * memory layout of objects: | |
1da177e4 | 506 | * 0 : objp |
3dafccf2 | 507 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
508 | * the end of an object is aligned with the end of the real |
509 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 510 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 511 | * redzone word. |
3dafccf2 MS |
512 | * cachep->obj_offset: The real object. |
513 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
514 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
515 | * [BYTES_PER_WORD long] | |
1da177e4 | 516 | */ |
343e0d7a | 517 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 518 | { |
3dafccf2 | 519 | return cachep->obj_offset; |
1da177e4 LT |
520 | } |
521 | ||
343e0d7a | 522 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 523 | { |
3dafccf2 | 524 | return cachep->obj_size; |
1da177e4 LT |
525 | } |
526 | ||
343e0d7a | 527 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
528 | { |
529 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 530 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
531 | } |
532 | ||
343e0d7a | 533 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
534 | { |
535 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
536 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 537 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 538 | 2 * BYTES_PER_WORD); |
3dafccf2 | 539 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
540 | } |
541 | ||
343e0d7a | 542 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
543 | { |
544 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 545 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
546 | } |
547 | ||
548 | #else | |
549 | ||
3dafccf2 MS |
550 | #define obj_offset(x) 0 |
551 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
552 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
553 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
554 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
555 | ||
556 | #endif | |
557 | ||
558 | /* | |
a737b3e2 AM |
559 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
560 | * order. | |
1da177e4 LT |
561 | */ |
562 | #if defined(CONFIG_LARGE_ALLOCS) | |
563 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
564 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
565 | #elif defined(CONFIG_MMU) | |
566 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
567 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
568 | #else | |
569 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
570 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
571 | #endif | |
572 | ||
573 | /* | |
574 | * Do not go above this order unless 0 objects fit into the slab. | |
575 | */ | |
576 | #define BREAK_GFP_ORDER_HI 1 | |
577 | #define BREAK_GFP_ORDER_LO 0 | |
578 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
579 | ||
a737b3e2 AM |
580 | /* |
581 | * Functions for storing/retrieving the cachep and or slab from the page | |
582 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
583 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 584 | */ |
065d41cb PE |
585 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
586 | { | |
587 | page->lru.next = (struct list_head *)cache; | |
588 | } | |
589 | ||
590 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
591 | { | |
84097518 NP |
592 | if (unlikely(PageCompound(page))) |
593 | page = (struct page *)page_private(page); | |
ddc2e812 | 594 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
595 | return (struct kmem_cache *)page->lru.next; |
596 | } | |
597 | ||
598 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
599 | { | |
600 | page->lru.prev = (struct list_head *)slab; | |
601 | } | |
602 | ||
603 | static inline struct slab *page_get_slab(struct page *page) | |
604 | { | |
84097518 NP |
605 | if (unlikely(PageCompound(page))) |
606 | page = (struct page *)page_private(page); | |
ddc2e812 | 607 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
608 | return (struct slab *)page->lru.prev; |
609 | } | |
1da177e4 | 610 | |
6ed5eb22 PE |
611 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
612 | { | |
613 | struct page *page = virt_to_page(obj); | |
614 | return page_get_cache(page); | |
615 | } | |
616 | ||
617 | static inline struct slab *virt_to_slab(const void *obj) | |
618 | { | |
619 | struct page *page = virt_to_page(obj); | |
620 | return page_get_slab(page); | |
621 | } | |
622 | ||
8fea4e96 PE |
623 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
624 | unsigned int idx) | |
625 | { | |
626 | return slab->s_mem + cache->buffer_size * idx; | |
627 | } | |
628 | ||
629 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
630 | struct slab *slab, void *obj) | |
631 | { | |
632 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
633 | } | |
634 | ||
a737b3e2 AM |
635 | /* |
636 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
637 | */ | |
1da177e4 LT |
638 | struct cache_sizes malloc_sizes[] = { |
639 | #define CACHE(x) { .cs_size = (x) }, | |
640 | #include <linux/kmalloc_sizes.h> | |
641 | CACHE(ULONG_MAX) | |
642 | #undef CACHE | |
643 | }; | |
644 | EXPORT_SYMBOL(malloc_sizes); | |
645 | ||
646 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
647 | struct cache_names { | |
648 | char *name; | |
649 | char *name_dma; | |
650 | }; | |
651 | ||
652 | static struct cache_names __initdata cache_names[] = { | |
653 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
654 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 655 | {NULL,} |
1da177e4 LT |
656 | #undef CACHE |
657 | }; | |
658 | ||
659 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 660 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 661 | static struct arraycache_init initarray_generic = |
b28a02de | 662 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
663 | |
664 | /* internal cache of cache description objs */ | |
343e0d7a | 665 | static struct kmem_cache cache_cache = { |
b28a02de PE |
666 | .batchcount = 1, |
667 | .limit = BOOT_CPUCACHE_ENTRIES, | |
668 | .shared = 1, | |
343e0d7a | 669 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 670 | .name = "kmem_cache", |
1da177e4 | 671 | #if DEBUG |
343e0d7a | 672 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
673 | #endif |
674 | }; | |
675 | ||
056c6241 RT |
676 | #define BAD_ALIEN_MAGIC 0x01020304ul |
677 | ||
f1aaee53 AV |
678 | #ifdef CONFIG_LOCKDEP |
679 | ||
680 | /* | |
681 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
682 | * for other slabs "off slab". | |
683 | * The locking for this is tricky in that it nests within the locks | |
684 | * of all other slabs in a few places; to deal with this special | |
685 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
686 | * |
687 | * We set lock class for alien array caches which are up during init. | |
688 | * The lock annotation will be lost if all cpus of a node goes down and | |
689 | * then comes back up during hotplug | |
f1aaee53 | 690 | */ |
056c6241 RT |
691 | static struct lock_class_key on_slab_l3_key; |
692 | static struct lock_class_key on_slab_alc_key; | |
693 | ||
694 | static inline void init_lock_keys(void) | |
f1aaee53 | 695 | |
f1aaee53 AV |
696 | { |
697 | int q; | |
056c6241 RT |
698 | struct cache_sizes *s = malloc_sizes; |
699 | ||
700 | while (s->cs_size != ULONG_MAX) { | |
701 | for_each_node(q) { | |
702 | struct array_cache **alc; | |
703 | int r; | |
704 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
705 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
706 | continue; | |
707 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
708 | alc = l3->alien; | |
709 | /* | |
710 | * FIXME: This check for BAD_ALIEN_MAGIC | |
711 | * should go away when common slab code is taught to | |
712 | * work even without alien caches. | |
713 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
714 | * for alloc_alien_cache, | |
715 | */ | |
716 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
717 | continue; | |
718 | for_each_node(r) { | |
719 | if (alc[r]) | |
720 | lockdep_set_class(&alc[r]->lock, | |
721 | &on_slab_alc_key); | |
722 | } | |
723 | } | |
724 | s++; | |
f1aaee53 AV |
725 | } |
726 | } | |
f1aaee53 | 727 | #else |
056c6241 | 728 | static inline void init_lock_keys(void) |
f1aaee53 AV |
729 | { |
730 | } | |
731 | #endif | |
732 | ||
1da177e4 | 733 | /* Guard access to the cache-chain. */ |
fc0abb14 | 734 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
735 | static struct list_head cache_chain; |
736 | ||
1da177e4 LT |
737 | /* |
738 | * chicken and egg problem: delay the per-cpu array allocation | |
739 | * until the general caches are up. | |
740 | */ | |
741 | static enum { | |
742 | NONE, | |
e498be7d CL |
743 | PARTIAL_AC, |
744 | PARTIAL_L3, | |
1da177e4 LT |
745 | FULL |
746 | } g_cpucache_up; | |
747 | ||
39d24e64 MK |
748 | /* |
749 | * used by boot code to determine if it can use slab based allocator | |
750 | */ | |
751 | int slab_is_available(void) | |
752 | { | |
753 | return g_cpucache_up == FULL; | |
754 | } | |
755 | ||
1da177e4 LT |
756 | static DEFINE_PER_CPU(struct work_struct, reap_work); |
757 | ||
343e0d7a | 758 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
759 | { |
760 | return cachep->array[smp_processor_id()]; | |
761 | } | |
762 | ||
a737b3e2 AM |
763 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
764 | gfp_t gfpflags) | |
1da177e4 LT |
765 | { |
766 | struct cache_sizes *csizep = malloc_sizes; | |
767 | ||
768 | #if DEBUG | |
769 | /* This happens if someone tries to call | |
b28a02de PE |
770 | * kmem_cache_create(), or __kmalloc(), before |
771 | * the generic caches are initialized. | |
772 | */ | |
c7e43c78 | 773 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
774 | #endif |
775 | while (size > csizep->cs_size) | |
776 | csizep++; | |
777 | ||
778 | /* | |
0abf40c1 | 779 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
780 | * has cs_{dma,}cachep==NULL. Thus no special case |
781 | * for large kmalloc calls required. | |
782 | */ | |
783 | if (unlikely(gfpflags & GFP_DMA)) | |
784 | return csizep->cs_dmacachep; | |
785 | return csizep->cs_cachep; | |
786 | } | |
787 | ||
b221385b | 788 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
789 | { |
790 | return __find_general_cachep(size, gfpflags); | |
791 | } | |
97e2bde4 | 792 | |
fbaccacf | 793 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 794 | { |
fbaccacf SR |
795 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
796 | } | |
1da177e4 | 797 | |
a737b3e2 AM |
798 | /* |
799 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
800 | */ | |
fbaccacf SR |
801 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
802 | size_t align, int flags, size_t *left_over, | |
803 | unsigned int *num) | |
804 | { | |
805 | int nr_objs; | |
806 | size_t mgmt_size; | |
807 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 808 | |
fbaccacf SR |
809 | /* |
810 | * The slab management structure can be either off the slab or | |
811 | * on it. For the latter case, the memory allocated for a | |
812 | * slab is used for: | |
813 | * | |
814 | * - The struct slab | |
815 | * - One kmem_bufctl_t for each object | |
816 | * - Padding to respect alignment of @align | |
817 | * - @buffer_size bytes for each object | |
818 | * | |
819 | * If the slab management structure is off the slab, then the | |
820 | * alignment will already be calculated into the size. Because | |
821 | * the slabs are all pages aligned, the objects will be at the | |
822 | * correct alignment when allocated. | |
823 | */ | |
824 | if (flags & CFLGS_OFF_SLAB) { | |
825 | mgmt_size = 0; | |
826 | nr_objs = slab_size / buffer_size; | |
827 | ||
828 | if (nr_objs > SLAB_LIMIT) | |
829 | nr_objs = SLAB_LIMIT; | |
830 | } else { | |
831 | /* | |
832 | * Ignore padding for the initial guess. The padding | |
833 | * is at most @align-1 bytes, and @buffer_size is at | |
834 | * least @align. In the worst case, this result will | |
835 | * be one greater than the number of objects that fit | |
836 | * into the memory allocation when taking the padding | |
837 | * into account. | |
838 | */ | |
839 | nr_objs = (slab_size - sizeof(struct slab)) / | |
840 | (buffer_size + sizeof(kmem_bufctl_t)); | |
841 | ||
842 | /* | |
843 | * This calculated number will be either the right | |
844 | * amount, or one greater than what we want. | |
845 | */ | |
846 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
847 | > slab_size) | |
848 | nr_objs--; | |
849 | ||
850 | if (nr_objs > SLAB_LIMIT) | |
851 | nr_objs = SLAB_LIMIT; | |
852 | ||
853 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
854 | } | |
855 | *num = nr_objs; | |
856 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
857 | } |
858 | ||
859 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
860 | ||
a737b3e2 AM |
861 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
862 | char *msg) | |
1da177e4 LT |
863 | { |
864 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 865 | function, cachep->name, msg); |
1da177e4 LT |
866 | dump_stack(); |
867 | } | |
868 | ||
8fce4d8e CL |
869 | #ifdef CONFIG_NUMA |
870 | /* | |
871 | * Special reaping functions for NUMA systems called from cache_reap(). | |
872 | * These take care of doing round robin flushing of alien caches (containing | |
873 | * objects freed on different nodes from which they were allocated) and the | |
874 | * flushing of remote pcps by calling drain_node_pages. | |
875 | */ | |
876 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
877 | ||
878 | static void init_reap_node(int cpu) | |
879 | { | |
880 | int node; | |
881 | ||
882 | node = next_node(cpu_to_node(cpu), node_online_map); | |
883 | if (node == MAX_NUMNODES) | |
442295c9 | 884 | node = first_node(node_online_map); |
8fce4d8e CL |
885 | |
886 | __get_cpu_var(reap_node) = node; | |
887 | } | |
888 | ||
889 | static void next_reap_node(void) | |
890 | { | |
891 | int node = __get_cpu_var(reap_node); | |
892 | ||
893 | /* | |
894 | * Also drain per cpu pages on remote zones | |
895 | */ | |
896 | if (node != numa_node_id()) | |
897 | drain_node_pages(node); | |
898 | ||
899 | node = next_node(node, node_online_map); | |
900 | if (unlikely(node >= MAX_NUMNODES)) | |
901 | node = first_node(node_online_map); | |
902 | __get_cpu_var(reap_node) = node; | |
903 | } | |
904 | ||
905 | #else | |
906 | #define init_reap_node(cpu) do { } while (0) | |
907 | #define next_reap_node(void) do { } while (0) | |
908 | #endif | |
909 | ||
1da177e4 LT |
910 | /* |
911 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
912 | * via the workqueue/eventd. | |
913 | * Add the CPU number into the expiration time to minimize the possibility of | |
914 | * the CPUs getting into lockstep and contending for the global cache chain | |
915 | * lock. | |
916 | */ | |
917 | static void __devinit start_cpu_timer(int cpu) | |
918 | { | |
919 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
920 | ||
921 | /* | |
922 | * When this gets called from do_initcalls via cpucache_init(), | |
923 | * init_workqueues() has already run, so keventd will be setup | |
924 | * at that time. | |
925 | */ | |
926 | if (keventd_up() && reap_work->func == NULL) { | |
8fce4d8e | 927 | init_reap_node(cpu); |
1da177e4 LT |
928 | INIT_WORK(reap_work, cache_reap, NULL); |
929 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
930 | } | |
931 | } | |
932 | ||
e498be7d | 933 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 934 | int batchcount) |
1da177e4 | 935 | { |
b28a02de | 936 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
937 | struct array_cache *nc = NULL; |
938 | ||
e498be7d | 939 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
940 | if (nc) { |
941 | nc->avail = 0; | |
942 | nc->limit = entries; | |
943 | nc->batchcount = batchcount; | |
944 | nc->touched = 0; | |
e498be7d | 945 | spin_lock_init(&nc->lock); |
1da177e4 LT |
946 | } |
947 | return nc; | |
948 | } | |
949 | ||
3ded175a CL |
950 | /* |
951 | * Transfer objects in one arraycache to another. | |
952 | * Locking must be handled by the caller. | |
953 | * | |
954 | * Return the number of entries transferred. | |
955 | */ | |
956 | static int transfer_objects(struct array_cache *to, | |
957 | struct array_cache *from, unsigned int max) | |
958 | { | |
959 | /* Figure out how many entries to transfer */ | |
960 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
961 | ||
962 | if (!nr) | |
963 | return 0; | |
964 | ||
965 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
966 | sizeof(void *) *nr); | |
967 | ||
968 | from->avail -= nr; | |
969 | to->avail += nr; | |
970 | to->touched = 1; | |
971 | return nr; | |
972 | } | |
973 | ||
765c4507 CL |
974 | #ifndef CONFIG_NUMA |
975 | ||
976 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
977 | #define reap_alien(cachep, l3) do { } while (0) | |
978 | ||
979 | static inline struct array_cache **alloc_alien_cache(int node, int limit) | |
980 | { | |
981 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
982 | } | |
983 | ||
984 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
985 | { | |
986 | } | |
987 | ||
988 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
989 | { | |
990 | return 0; | |
991 | } | |
992 | ||
993 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
994 | gfp_t flags) | |
995 | { | |
996 | return NULL; | |
997 | } | |
998 | ||
999 | static inline void *__cache_alloc_node(struct kmem_cache *cachep, | |
1000 | gfp_t flags, int nodeid) | |
1001 | { | |
1002 | return NULL; | |
1003 | } | |
1004 | ||
1005 | #else /* CONFIG_NUMA */ | |
1006 | ||
343e0d7a | 1007 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1008 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1009 | |
5295a74c | 1010 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
1011 | { |
1012 | struct array_cache **ac_ptr; | |
b28a02de | 1013 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
1014 | int i; |
1015 | ||
1016 | if (limit > 1) | |
1017 | limit = 12; | |
1018 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
1019 | if (ac_ptr) { | |
1020 | for_each_node(i) { | |
1021 | if (i == node || !node_online(i)) { | |
1022 | ac_ptr[i] = NULL; | |
1023 | continue; | |
1024 | } | |
1025 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
1026 | if (!ac_ptr[i]) { | |
b28a02de | 1027 | for (i--; i <= 0; i--) |
e498be7d CL |
1028 | kfree(ac_ptr[i]); |
1029 | kfree(ac_ptr); | |
1030 | return NULL; | |
1031 | } | |
1032 | } | |
1033 | } | |
1034 | return ac_ptr; | |
1035 | } | |
1036 | ||
5295a74c | 1037 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1038 | { |
1039 | int i; | |
1040 | ||
1041 | if (!ac_ptr) | |
1042 | return; | |
e498be7d | 1043 | for_each_node(i) |
b28a02de | 1044 | kfree(ac_ptr[i]); |
e498be7d CL |
1045 | kfree(ac_ptr); |
1046 | } | |
1047 | ||
343e0d7a | 1048 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1049 | struct array_cache *ac, int node) |
e498be7d CL |
1050 | { |
1051 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1052 | ||
1053 | if (ac->avail) { | |
1054 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1055 | /* |
1056 | * Stuff objects into the remote nodes shared array first. | |
1057 | * That way we could avoid the overhead of putting the objects | |
1058 | * into the free lists and getting them back later. | |
1059 | */ | |
693f7d36 JS |
1060 | if (rl3->shared) |
1061 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1062 | |
ff69416e | 1063 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1064 | ac->avail = 0; |
1065 | spin_unlock(&rl3->list_lock); | |
1066 | } | |
1067 | } | |
1068 | ||
8fce4d8e CL |
1069 | /* |
1070 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1071 | */ | |
1072 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1073 | { | |
1074 | int node = __get_cpu_var(reap_node); | |
1075 | ||
1076 | if (l3->alien) { | |
1077 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1078 | |
1079 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1080 | __drain_alien_cache(cachep, ac, node); |
1081 | spin_unlock_irq(&ac->lock); | |
1082 | } | |
1083 | } | |
1084 | } | |
1085 | ||
a737b3e2 AM |
1086 | static void drain_alien_cache(struct kmem_cache *cachep, |
1087 | struct array_cache **alien) | |
e498be7d | 1088 | { |
b28a02de | 1089 | int i = 0; |
e498be7d CL |
1090 | struct array_cache *ac; |
1091 | unsigned long flags; | |
1092 | ||
1093 | for_each_online_node(i) { | |
4484ebf1 | 1094 | ac = alien[i]; |
e498be7d CL |
1095 | if (ac) { |
1096 | spin_lock_irqsave(&ac->lock, flags); | |
1097 | __drain_alien_cache(cachep, ac, i); | |
1098 | spin_unlock_irqrestore(&ac->lock, flags); | |
1099 | } | |
1100 | } | |
1101 | } | |
729bd0b7 | 1102 | |
873623df | 1103 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1104 | { |
1105 | struct slab *slabp = virt_to_slab(objp); | |
1106 | int nodeid = slabp->nodeid; | |
1107 | struct kmem_list3 *l3; | |
1108 | struct array_cache *alien = NULL; | |
1109 | ||
1110 | /* | |
1111 | * Make sure we are not freeing a object from another node to the array | |
1112 | * cache on this cpu. | |
1113 | */ | |
1114 | if (likely(slabp->nodeid == numa_node_id())) | |
1115 | return 0; | |
1116 | ||
1117 | l3 = cachep->nodelists[numa_node_id()]; | |
1118 | STATS_INC_NODEFREES(cachep); | |
1119 | if (l3->alien && l3->alien[nodeid]) { | |
1120 | alien = l3->alien[nodeid]; | |
873623df | 1121 | spin_lock(&alien->lock); |
729bd0b7 PE |
1122 | if (unlikely(alien->avail == alien->limit)) { |
1123 | STATS_INC_ACOVERFLOW(cachep); | |
1124 | __drain_alien_cache(cachep, alien, nodeid); | |
1125 | } | |
1126 | alien->entry[alien->avail++] = objp; | |
1127 | spin_unlock(&alien->lock); | |
1128 | } else { | |
1129 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1130 | free_block(cachep, &objp, 1, nodeid); | |
1131 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1132 | } | |
1133 | return 1; | |
1134 | } | |
e498be7d CL |
1135 | #endif |
1136 | ||
8c78f307 | 1137 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1138 | unsigned long action, void *hcpu) |
1da177e4 LT |
1139 | { |
1140 | long cpu = (long)hcpu; | |
343e0d7a | 1141 | struct kmem_cache *cachep; |
e498be7d CL |
1142 | struct kmem_list3 *l3 = NULL; |
1143 | int node = cpu_to_node(cpu); | |
1144 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1145 | |
1146 | switch (action) { | |
1147 | case CPU_UP_PREPARE: | |
fc0abb14 | 1148 | mutex_lock(&cache_chain_mutex); |
a737b3e2 AM |
1149 | /* |
1150 | * We need to do this right in the beginning since | |
e498be7d CL |
1151 | * alloc_arraycache's are going to use this list. |
1152 | * kmalloc_node allows us to add the slab to the right | |
1153 | * kmem_list3 and not this cpu's kmem_list3 | |
1154 | */ | |
1155 | ||
1da177e4 | 1156 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1157 | /* |
1158 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1159 | * begin anything. Make sure some other cpu on this |
1160 | * node has not already allocated this | |
1161 | */ | |
1162 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1163 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1164 | if (!l3) | |
e498be7d CL |
1165 | goto bad; |
1166 | kmem_list3_init(l3); | |
1167 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1168 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1169 | |
4484ebf1 RT |
1170 | /* |
1171 | * The l3s don't come and go as CPUs come and | |
1172 | * go. cache_chain_mutex is sufficient | |
1173 | * protection here. | |
1174 | */ | |
e498be7d CL |
1175 | cachep->nodelists[node] = l3; |
1176 | } | |
1da177e4 | 1177 | |
e498be7d CL |
1178 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1179 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1180 | (1 + nr_cpus_node(node)) * |
1181 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1182 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1183 | } | |
1184 | ||
a737b3e2 AM |
1185 | /* |
1186 | * Now we can go ahead with allocating the shared arrays and | |
1187 | * array caches | |
1188 | */ | |
e498be7d | 1189 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1190 | struct array_cache *nc; |
4484ebf1 RT |
1191 | struct array_cache *shared; |
1192 | struct array_cache **alien; | |
cd105df4 | 1193 | |
e498be7d | 1194 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1195 | cachep->batchcount); |
1da177e4 LT |
1196 | if (!nc) |
1197 | goto bad; | |
4484ebf1 RT |
1198 | shared = alloc_arraycache(node, |
1199 | cachep->shared * cachep->batchcount, | |
1200 | 0xbaadf00d); | |
1201 | if (!shared) | |
1202 | goto bad; | |
7a21ef6f | 1203 | |
4484ebf1 RT |
1204 | alien = alloc_alien_cache(node, cachep->limit); |
1205 | if (!alien) | |
1206 | goto bad; | |
1da177e4 | 1207 | cachep->array[cpu] = nc; |
e498be7d CL |
1208 | l3 = cachep->nodelists[node]; |
1209 | BUG_ON(!l3); | |
e498be7d | 1210 | |
4484ebf1 RT |
1211 | spin_lock_irq(&l3->list_lock); |
1212 | if (!l3->shared) { | |
1213 | /* | |
1214 | * We are serialised from CPU_DEAD or | |
1215 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1216 | */ | |
1217 | l3->shared = shared; | |
1218 | shared = NULL; | |
e498be7d | 1219 | } |
4484ebf1 RT |
1220 | #ifdef CONFIG_NUMA |
1221 | if (!l3->alien) { | |
1222 | l3->alien = alien; | |
1223 | alien = NULL; | |
1224 | } | |
1225 | #endif | |
1226 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1227 | kfree(shared); |
1228 | free_alien_cache(alien); | |
1da177e4 | 1229 | } |
fc0abb14 | 1230 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1231 | break; |
1232 | case CPU_ONLINE: | |
1233 | start_cpu_timer(cpu); | |
1234 | break; | |
1235 | #ifdef CONFIG_HOTPLUG_CPU | |
1236 | case CPU_DEAD: | |
4484ebf1 RT |
1237 | /* |
1238 | * Even if all the cpus of a node are down, we don't free the | |
1239 | * kmem_list3 of any cache. This to avoid a race between | |
1240 | * cpu_down, and a kmalloc allocation from another cpu for | |
1241 | * memory from the node of the cpu going down. The list3 | |
1242 | * structure is usually allocated from kmem_cache_create() and | |
1243 | * gets destroyed at kmem_cache_destroy(). | |
1244 | */ | |
1da177e4 LT |
1245 | /* fall thru */ |
1246 | case CPU_UP_CANCELED: | |
fc0abb14 | 1247 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1248 | list_for_each_entry(cachep, &cache_chain, next) { |
1249 | struct array_cache *nc; | |
4484ebf1 RT |
1250 | struct array_cache *shared; |
1251 | struct array_cache **alien; | |
e498be7d | 1252 | cpumask_t mask; |
1da177e4 | 1253 | |
e498be7d | 1254 | mask = node_to_cpumask(node); |
1da177e4 LT |
1255 | /* cpu is dead; no one can alloc from it. */ |
1256 | nc = cachep->array[cpu]; | |
1257 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1258 | l3 = cachep->nodelists[node]; |
1259 | ||
1260 | if (!l3) | |
4484ebf1 | 1261 | goto free_array_cache; |
e498be7d | 1262 | |
ca3b9b91 | 1263 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1264 | |
1265 | /* Free limit for this kmem_list3 */ | |
1266 | l3->free_limit -= cachep->batchcount; | |
1267 | if (nc) | |
ff69416e | 1268 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1269 | |
1270 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1271 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1272 | goto free_array_cache; |
b28a02de | 1273 | } |
e498be7d | 1274 | |
4484ebf1 RT |
1275 | shared = l3->shared; |
1276 | if (shared) { | |
e498be7d | 1277 | free_block(cachep, l3->shared->entry, |
b28a02de | 1278 | l3->shared->avail, node); |
e498be7d CL |
1279 | l3->shared = NULL; |
1280 | } | |
e498be7d | 1281 | |
4484ebf1 RT |
1282 | alien = l3->alien; |
1283 | l3->alien = NULL; | |
1284 | ||
1285 | spin_unlock_irq(&l3->list_lock); | |
1286 | ||
1287 | kfree(shared); | |
1288 | if (alien) { | |
1289 | drain_alien_cache(cachep, alien); | |
1290 | free_alien_cache(alien); | |
e498be7d | 1291 | } |
4484ebf1 | 1292 | free_array_cache: |
1da177e4 LT |
1293 | kfree(nc); |
1294 | } | |
4484ebf1 RT |
1295 | /* |
1296 | * In the previous loop, all the objects were freed to | |
1297 | * the respective cache's slabs, now we can go ahead and | |
1298 | * shrink each nodelist to its limit. | |
1299 | */ | |
1300 | list_for_each_entry(cachep, &cache_chain, next) { | |
1301 | l3 = cachep->nodelists[node]; | |
1302 | if (!l3) | |
1303 | continue; | |
ed11d9eb | 1304 | drain_freelist(cachep, l3, l3->free_objects); |
4484ebf1 | 1305 | } |
fc0abb14 | 1306 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1307 | break; |
1308 | #endif | |
1309 | } | |
1310 | return NOTIFY_OK; | |
a737b3e2 | 1311 | bad: |
fc0abb14 | 1312 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1313 | return NOTIFY_BAD; |
1314 | } | |
1315 | ||
74b85f37 CS |
1316 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1317 | &cpuup_callback, NULL, 0 | |
1318 | }; | |
1da177e4 | 1319 | |
e498be7d CL |
1320 | /* |
1321 | * swap the static kmem_list3 with kmalloced memory | |
1322 | */ | |
a737b3e2 AM |
1323 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1324 | int nodeid) | |
e498be7d CL |
1325 | { |
1326 | struct kmem_list3 *ptr; | |
1327 | ||
1328 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1329 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1330 | BUG_ON(!ptr); | |
1331 | ||
1332 | local_irq_disable(); | |
1333 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1334 | /* |
1335 | * Do not assume that spinlocks can be initialized via memcpy: | |
1336 | */ | |
1337 | spin_lock_init(&ptr->list_lock); | |
1338 | ||
e498be7d CL |
1339 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1340 | cachep->nodelists[nodeid] = ptr; | |
1341 | local_irq_enable(); | |
1342 | } | |
1343 | ||
a737b3e2 AM |
1344 | /* |
1345 | * Initialisation. Called after the page allocator have been initialised and | |
1346 | * before smp_init(). | |
1da177e4 LT |
1347 | */ |
1348 | void __init kmem_cache_init(void) | |
1349 | { | |
1350 | size_t left_over; | |
1351 | struct cache_sizes *sizes; | |
1352 | struct cache_names *names; | |
e498be7d | 1353 | int i; |
07ed76b2 | 1354 | int order; |
e498be7d CL |
1355 | |
1356 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1357 | kmem_list3_init(&initkmem_list3[i]); | |
1358 | if (i < MAX_NUMNODES) | |
1359 | cache_cache.nodelists[i] = NULL; | |
1360 | } | |
1da177e4 LT |
1361 | |
1362 | /* | |
1363 | * Fragmentation resistance on low memory - only use bigger | |
1364 | * page orders on machines with more than 32MB of memory. | |
1365 | */ | |
1366 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1367 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1368 | ||
1da177e4 LT |
1369 | /* Bootstrap is tricky, because several objects are allocated |
1370 | * from caches that do not exist yet: | |
a737b3e2 AM |
1371 | * 1) initialize the cache_cache cache: it contains the struct |
1372 | * kmem_cache structures of all caches, except cache_cache itself: | |
1373 | * cache_cache is statically allocated. | |
e498be7d CL |
1374 | * Initially an __init data area is used for the head array and the |
1375 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1376 | * array at the end of the bootstrap. | |
1da177e4 | 1377 | * 2) Create the first kmalloc cache. |
343e0d7a | 1378 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1379 | * An __init data area is used for the head array. |
1380 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1381 | * head arrays. | |
1da177e4 LT |
1382 | * 4) Replace the __init data head arrays for cache_cache and the first |
1383 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1384 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1385 | * the other cache's with kmalloc allocated memory. | |
1386 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1387 | */ |
1388 | ||
1389 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1390 | INIT_LIST_HEAD(&cache_chain); |
1391 | list_add(&cache_cache.next, &cache_chain); | |
1392 | cache_cache.colour_off = cache_line_size(); | |
1393 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1394 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1395 | |
a737b3e2 AM |
1396 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1397 | cache_line_size()); | |
1da177e4 | 1398 | |
07ed76b2 JS |
1399 | for (order = 0; order < MAX_ORDER; order++) { |
1400 | cache_estimate(order, cache_cache.buffer_size, | |
1401 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1402 | if (cache_cache.num) | |
1403 | break; | |
1404 | } | |
40094fa6 | 1405 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1406 | cache_cache.gfporder = order; |
b28a02de | 1407 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1408 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1409 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1410 | |
1411 | /* 2+3) create the kmalloc caches */ | |
1412 | sizes = malloc_sizes; | |
1413 | names = cache_names; | |
1414 | ||
a737b3e2 AM |
1415 | /* |
1416 | * Initialize the caches that provide memory for the array cache and the | |
1417 | * kmem_list3 structures first. Without this, further allocations will | |
1418 | * bug. | |
e498be7d CL |
1419 | */ |
1420 | ||
1421 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1422 | sizes[INDEX_AC].cs_size, |
1423 | ARCH_KMALLOC_MINALIGN, | |
1424 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1425 | NULL, NULL); | |
e498be7d | 1426 | |
a737b3e2 | 1427 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1428 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1429 | kmem_cache_create(names[INDEX_L3].name, |
1430 | sizes[INDEX_L3].cs_size, | |
1431 | ARCH_KMALLOC_MINALIGN, | |
1432 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1433 | NULL, NULL); | |
1434 | } | |
e498be7d | 1435 | |
e0a42726 IM |
1436 | slab_early_init = 0; |
1437 | ||
1da177e4 | 1438 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1439 | /* |
1440 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1441 | * This should be particularly beneficial on SMP boxes, as it |
1442 | * eliminates "false sharing". | |
1443 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1444 | * allow tighter packing of the smaller caches. |
1445 | */ | |
a737b3e2 | 1446 | if (!sizes->cs_cachep) { |
e498be7d | 1447 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1448 | sizes->cs_size, |
1449 | ARCH_KMALLOC_MINALIGN, | |
1450 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1451 | NULL, NULL); | |
1452 | } | |
1da177e4 | 1453 | |
1da177e4 | 1454 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, |
a737b3e2 AM |
1455 | sizes->cs_size, |
1456 | ARCH_KMALLOC_MINALIGN, | |
1457 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1458 | SLAB_PANIC, | |
1459 | NULL, NULL); | |
1da177e4 LT |
1460 | sizes++; |
1461 | names++; | |
1462 | } | |
1463 | /* 4) Replace the bootstrap head arrays */ | |
1464 | { | |
2b2d5493 | 1465 | struct array_cache *ptr; |
e498be7d | 1466 | |
1da177e4 | 1467 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1468 | |
1da177e4 | 1469 | local_irq_disable(); |
9a2dba4b PE |
1470 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1471 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1472 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1473 | /* |
1474 | * Do not assume that spinlocks can be initialized via memcpy: | |
1475 | */ | |
1476 | spin_lock_init(&ptr->lock); | |
1477 | ||
1da177e4 LT |
1478 | cache_cache.array[smp_processor_id()] = ptr; |
1479 | local_irq_enable(); | |
e498be7d | 1480 | |
1da177e4 | 1481 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1482 | |
1da177e4 | 1483 | local_irq_disable(); |
9a2dba4b | 1484 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1485 | != &initarray_generic.cache); |
9a2dba4b | 1486 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1487 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1488 | /* |
1489 | * Do not assume that spinlocks can be initialized via memcpy: | |
1490 | */ | |
1491 | spin_lock_init(&ptr->lock); | |
1492 | ||
e498be7d | 1493 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1494 | ptr; |
1da177e4 LT |
1495 | local_irq_enable(); |
1496 | } | |
e498be7d CL |
1497 | /* 5) Replace the bootstrap kmem_list3's */ |
1498 | { | |
1499 | int node; | |
1500 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1501 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1502 | numa_node_id()); |
e498be7d CL |
1503 | |
1504 | for_each_online_node(node) { | |
1505 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1506 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1507 | |
1508 | if (INDEX_AC != INDEX_L3) { | |
1509 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1510 | &initkmem_list3[SIZE_L3 + node], |
1511 | node); | |
e498be7d CL |
1512 | } |
1513 | } | |
1514 | } | |
1da177e4 | 1515 | |
e498be7d | 1516 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1517 | { |
343e0d7a | 1518 | struct kmem_cache *cachep; |
fc0abb14 | 1519 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1520 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1521 | if (enable_cpucache(cachep)) |
1522 | BUG(); | |
fc0abb14 | 1523 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1524 | } |
1525 | ||
056c6241 RT |
1526 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1527 | init_lock_keys(); | |
1528 | ||
1529 | ||
1da177e4 LT |
1530 | /* Done! */ |
1531 | g_cpucache_up = FULL; | |
1532 | ||
a737b3e2 AM |
1533 | /* |
1534 | * Register a cpu startup notifier callback that initializes | |
1535 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1536 | */ |
1537 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1538 | |
a737b3e2 AM |
1539 | /* |
1540 | * The reap timers are started later, with a module init call: That part | |
1541 | * of the kernel is not yet operational. | |
1da177e4 LT |
1542 | */ |
1543 | } | |
1544 | ||
1545 | static int __init cpucache_init(void) | |
1546 | { | |
1547 | int cpu; | |
1548 | ||
a737b3e2 AM |
1549 | /* |
1550 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1551 | */ |
e498be7d | 1552 | for_each_online_cpu(cpu) |
a737b3e2 | 1553 | start_cpu_timer(cpu); |
1da177e4 LT |
1554 | return 0; |
1555 | } | |
1da177e4 LT |
1556 | __initcall(cpucache_init); |
1557 | ||
1558 | /* | |
1559 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1560 | * | |
1561 | * If we requested dmaable memory, we will get it. Even if we | |
1562 | * did not request dmaable memory, we might get it, but that | |
1563 | * would be relatively rare and ignorable. | |
1564 | */ | |
343e0d7a | 1565 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1566 | { |
1567 | struct page *page; | |
e1b6aa6f | 1568 | int nr_pages; |
1da177e4 LT |
1569 | int i; |
1570 | ||
d6fef9da | 1571 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1572 | /* |
1573 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1574 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1575 | */ |
e1b6aa6f | 1576 | flags |= __GFP_COMP; |
d6fef9da | 1577 | #endif |
765c4507 CL |
1578 | |
1579 | /* | |
1580 | * Under NUMA we want memory on the indicated node. We will handle | |
1581 | * the needed fallback ourselves since we want to serve from our | |
1582 | * per node object lists first for other nodes. | |
1583 | */ | |
1584 | flags |= cachep->gfpflags | GFP_THISNODE; | |
e1b6aa6f CH |
1585 | |
1586 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1587 | if (!page) |
1588 | return NULL; | |
1da177e4 | 1589 | |
e1b6aa6f | 1590 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1591 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1592 | add_zone_page_state(page_zone(page), |
1593 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1594 | else | |
1595 | add_zone_page_state(page_zone(page), | |
1596 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1597 | for (i = 0; i < nr_pages; i++) |
1598 | __SetPageSlab(page + i); | |
1599 | return page_address(page); | |
1da177e4 LT |
1600 | } |
1601 | ||
1602 | /* | |
1603 | * Interface to system's page release. | |
1604 | */ | |
343e0d7a | 1605 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1606 | { |
b28a02de | 1607 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1608 | struct page *page = virt_to_page(addr); |
1609 | const unsigned long nr_freed = i; | |
1610 | ||
972d1a7b CL |
1611 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1612 | sub_zone_page_state(page_zone(page), | |
1613 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1614 | else | |
1615 | sub_zone_page_state(page_zone(page), | |
1616 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1617 | while (i--) { |
f205b2fe NP |
1618 | BUG_ON(!PageSlab(page)); |
1619 | __ClearPageSlab(page); | |
1da177e4 LT |
1620 | page++; |
1621 | } | |
1da177e4 LT |
1622 | if (current->reclaim_state) |
1623 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1624 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1625 | } |
1626 | ||
1627 | static void kmem_rcu_free(struct rcu_head *head) | |
1628 | { | |
b28a02de | 1629 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1630 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1631 | |
1632 | kmem_freepages(cachep, slab_rcu->addr); | |
1633 | if (OFF_SLAB(cachep)) | |
1634 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1635 | } | |
1636 | ||
1637 | #if DEBUG | |
1638 | ||
1639 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1640 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1641 | unsigned long caller) |
1da177e4 | 1642 | { |
3dafccf2 | 1643 | int size = obj_size(cachep); |
1da177e4 | 1644 | |
3dafccf2 | 1645 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1646 | |
b28a02de | 1647 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1648 | return; |
1649 | ||
b28a02de PE |
1650 | *addr++ = 0x12345678; |
1651 | *addr++ = caller; | |
1652 | *addr++ = smp_processor_id(); | |
1653 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1654 | { |
1655 | unsigned long *sptr = &caller; | |
1656 | unsigned long svalue; | |
1657 | ||
1658 | while (!kstack_end(sptr)) { | |
1659 | svalue = *sptr++; | |
1660 | if (kernel_text_address(svalue)) { | |
b28a02de | 1661 | *addr++ = svalue; |
1da177e4 LT |
1662 | size -= sizeof(unsigned long); |
1663 | if (size <= sizeof(unsigned long)) | |
1664 | break; | |
1665 | } | |
1666 | } | |
1667 | ||
1668 | } | |
b28a02de | 1669 | *addr++ = 0x87654321; |
1da177e4 LT |
1670 | } |
1671 | #endif | |
1672 | ||
343e0d7a | 1673 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1674 | { |
3dafccf2 MS |
1675 | int size = obj_size(cachep); |
1676 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1677 | |
1678 | memset(addr, val, size); | |
b28a02de | 1679 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1680 | } |
1681 | ||
1682 | static void dump_line(char *data, int offset, int limit) | |
1683 | { | |
1684 | int i; | |
aa83aa40 DJ |
1685 | unsigned char error = 0; |
1686 | int bad_count = 0; | |
1687 | ||
1da177e4 | 1688 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1689 | for (i = 0; i < limit; i++) { |
1690 | if (data[offset + i] != POISON_FREE) { | |
1691 | error = data[offset + i]; | |
1692 | bad_count++; | |
1693 | } | |
b28a02de | 1694 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1695 | } |
1da177e4 | 1696 | printk("\n"); |
aa83aa40 DJ |
1697 | |
1698 | if (bad_count == 1) { | |
1699 | error ^= POISON_FREE; | |
1700 | if (!(error & (error - 1))) { | |
1701 | printk(KERN_ERR "Single bit error detected. Probably " | |
1702 | "bad RAM.\n"); | |
1703 | #ifdef CONFIG_X86 | |
1704 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1705 | "test tool.\n"); | |
1706 | #else | |
1707 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1708 | #endif | |
1709 | } | |
1710 | } | |
1da177e4 LT |
1711 | } |
1712 | #endif | |
1713 | ||
1714 | #if DEBUG | |
1715 | ||
343e0d7a | 1716 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1717 | { |
1718 | int i, size; | |
1719 | char *realobj; | |
1720 | ||
1721 | if (cachep->flags & SLAB_RED_ZONE) { | |
1722 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
a737b3e2 AM |
1723 | *dbg_redzone1(cachep, objp), |
1724 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1725 | } |
1726 | ||
1727 | if (cachep->flags & SLAB_STORE_USER) { | |
1728 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1729 | *dbg_userword(cachep, objp)); |
1da177e4 | 1730 | print_symbol("(%s)", |
a737b3e2 | 1731 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1732 | printk("\n"); |
1733 | } | |
3dafccf2 MS |
1734 | realobj = (char *)objp + obj_offset(cachep); |
1735 | size = obj_size(cachep); | |
b28a02de | 1736 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1737 | int limit; |
1738 | limit = 16; | |
b28a02de PE |
1739 | if (i + limit > size) |
1740 | limit = size - i; | |
1da177e4 LT |
1741 | dump_line(realobj, i, limit); |
1742 | } | |
1743 | } | |
1744 | ||
343e0d7a | 1745 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1746 | { |
1747 | char *realobj; | |
1748 | int size, i; | |
1749 | int lines = 0; | |
1750 | ||
3dafccf2 MS |
1751 | realobj = (char *)objp + obj_offset(cachep); |
1752 | size = obj_size(cachep); | |
1da177e4 | 1753 | |
b28a02de | 1754 | for (i = 0; i < size; i++) { |
1da177e4 | 1755 | char exp = POISON_FREE; |
b28a02de | 1756 | if (i == size - 1) |
1da177e4 LT |
1757 | exp = POISON_END; |
1758 | if (realobj[i] != exp) { | |
1759 | int limit; | |
1760 | /* Mismatch ! */ | |
1761 | /* Print header */ | |
1762 | if (lines == 0) { | |
b28a02de | 1763 | printk(KERN_ERR |
a737b3e2 AM |
1764 | "Slab corruption: start=%p, len=%d\n", |
1765 | realobj, size); | |
1da177e4 LT |
1766 | print_objinfo(cachep, objp, 0); |
1767 | } | |
1768 | /* Hexdump the affected line */ | |
b28a02de | 1769 | i = (i / 16) * 16; |
1da177e4 | 1770 | limit = 16; |
b28a02de PE |
1771 | if (i + limit > size) |
1772 | limit = size - i; | |
1da177e4 LT |
1773 | dump_line(realobj, i, limit); |
1774 | i += 16; | |
1775 | lines++; | |
1776 | /* Limit to 5 lines */ | |
1777 | if (lines > 5) | |
1778 | break; | |
1779 | } | |
1780 | } | |
1781 | if (lines != 0) { | |
1782 | /* Print some data about the neighboring objects, if they | |
1783 | * exist: | |
1784 | */ | |
6ed5eb22 | 1785 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1786 | unsigned int objnr; |
1da177e4 | 1787 | |
8fea4e96 | 1788 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1789 | if (objnr) { |
8fea4e96 | 1790 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1791 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1792 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1793 | realobj, size); |
1da177e4 LT |
1794 | print_objinfo(cachep, objp, 2); |
1795 | } | |
b28a02de | 1796 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1797 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1798 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1799 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1800 | realobj, size); |
1da177e4 LT |
1801 | print_objinfo(cachep, objp, 2); |
1802 | } | |
1803 | } | |
1804 | } | |
1805 | #endif | |
1806 | ||
12dd36fa MD |
1807 | #if DEBUG |
1808 | /** | |
911851e6 RD |
1809 | * slab_destroy_objs - destroy a slab and its objects |
1810 | * @cachep: cache pointer being destroyed | |
1811 | * @slabp: slab pointer being destroyed | |
1812 | * | |
1813 | * Call the registered destructor for each object in a slab that is being | |
1814 | * destroyed. | |
1da177e4 | 1815 | */ |
343e0d7a | 1816 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1817 | { |
1da177e4 LT |
1818 | int i; |
1819 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1820 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1821 | |
1822 | if (cachep->flags & SLAB_POISON) { | |
1823 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1824 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1825 | OFF_SLAB(cachep)) | |
b28a02de | 1826 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1827 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1828 | else |
1829 | check_poison_obj(cachep, objp); | |
1830 | #else | |
1831 | check_poison_obj(cachep, objp); | |
1832 | #endif | |
1833 | } | |
1834 | if (cachep->flags & SLAB_RED_ZONE) { | |
1835 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1836 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1837 | "was overwritten"); |
1da177e4 LT |
1838 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1839 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1840 | "was overwritten"); |
1da177e4 LT |
1841 | } |
1842 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1843 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1844 | } |
12dd36fa | 1845 | } |
1da177e4 | 1846 | #else |
343e0d7a | 1847 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1848 | { |
1da177e4 LT |
1849 | if (cachep->dtor) { |
1850 | int i; | |
1851 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1852 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1853 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1854 | } |
1855 | } | |
12dd36fa | 1856 | } |
1da177e4 LT |
1857 | #endif |
1858 | ||
911851e6 RD |
1859 | /** |
1860 | * slab_destroy - destroy and release all objects in a slab | |
1861 | * @cachep: cache pointer being destroyed | |
1862 | * @slabp: slab pointer being destroyed | |
1863 | * | |
12dd36fa | 1864 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1865 | * Before calling the slab must have been unlinked from the cache. The |
1866 | * cache-lock is not held/needed. | |
12dd36fa | 1867 | */ |
343e0d7a | 1868 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1869 | { |
1870 | void *addr = slabp->s_mem - slabp->colouroff; | |
1871 | ||
1872 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1873 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1874 | struct slab_rcu *slab_rcu; | |
1875 | ||
b28a02de | 1876 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1877 | slab_rcu->cachep = cachep; |
1878 | slab_rcu->addr = addr; | |
1879 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1880 | } else { | |
1881 | kmem_freepages(cachep, addr); | |
873623df IM |
1882 | if (OFF_SLAB(cachep)) |
1883 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1884 | } |
1885 | } | |
1886 | ||
a737b3e2 AM |
1887 | /* |
1888 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1889 | * size of kmem_list3. | |
1890 | */ | |
343e0d7a | 1891 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1892 | { |
1893 | int node; | |
1894 | ||
1895 | for_each_online_node(node) { | |
b28a02de | 1896 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1897 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1898 | REAPTIMEOUT_LIST3 + |
1899 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1900 | } |
1901 | } | |
1902 | ||
117f6eb1 CL |
1903 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1904 | { | |
1905 | int i; | |
1906 | struct kmem_list3 *l3; | |
1907 | ||
1908 | for_each_online_cpu(i) | |
1909 | kfree(cachep->array[i]); | |
1910 | ||
1911 | /* NUMA: free the list3 structures */ | |
1912 | for_each_online_node(i) { | |
1913 | l3 = cachep->nodelists[i]; | |
1914 | if (l3) { | |
1915 | kfree(l3->shared); | |
1916 | free_alien_cache(l3->alien); | |
1917 | kfree(l3); | |
1918 | } | |
1919 | } | |
1920 | kmem_cache_free(&cache_cache, cachep); | |
1921 | } | |
1922 | ||
1923 | ||
4d268eba | 1924 | /** |
a70773dd RD |
1925 | * calculate_slab_order - calculate size (page order) of slabs |
1926 | * @cachep: pointer to the cache that is being created | |
1927 | * @size: size of objects to be created in this cache. | |
1928 | * @align: required alignment for the objects. | |
1929 | * @flags: slab allocation flags | |
1930 | * | |
1931 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1932 | * |
1933 | * This could be made much more intelligent. For now, try to avoid using | |
1934 | * high order pages for slabs. When the gfp() functions are more friendly | |
1935 | * towards high-order requests, this should be changed. | |
1936 | */ | |
a737b3e2 | 1937 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1938 | size_t size, size_t align, unsigned long flags) |
4d268eba | 1939 | { |
b1ab41c4 | 1940 | unsigned long offslab_limit; |
4d268eba | 1941 | size_t left_over = 0; |
9888e6fa | 1942 | int gfporder; |
4d268eba | 1943 | |
a737b3e2 | 1944 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1945 | unsigned int num; |
1946 | size_t remainder; | |
1947 | ||
9888e6fa | 1948 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1949 | if (!num) |
1950 | continue; | |
9888e6fa | 1951 | |
b1ab41c4 IM |
1952 | if (flags & CFLGS_OFF_SLAB) { |
1953 | /* | |
1954 | * Max number of objs-per-slab for caches which | |
1955 | * use off-slab slabs. Needed to avoid a possible | |
1956 | * looping condition in cache_grow(). | |
1957 | */ | |
1958 | offslab_limit = size - sizeof(struct slab); | |
1959 | offslab_limit /= sizeof(kmem_bufctl_t); | |
1960 | ||
1961 | if (num > offslab_limit) | |
1962 | break; | |
1963 | } | |
4d268eba | 1964 | |
9888e6fa | 1965 | /* Found something acceptable - save it away */ |
4d268eba | 1966 | cachep->num = num; |
9888e6fa | 1967 | cachep->gfporder = gfporder; |
4d268eba PE |
1968 | left_over = remainder; |
1969 | ||
f78bb8ad LT |
1970 | /* |
1971 | * A VFS-reclaimable slab tends to have most allocations | |
1972 | * as GFP_NOFS and we really don't want to have to be allocating | |
1973 | * higher-order pages when we are unable to shrink dcache. | |
1974 | */ | |
1975 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1976 | break; | |
1977 | ||
4d268eba PE |
1978 | /* |
1979 | * Large number of objects is good, but very large slabs are | |
1980 | * currently bad for the gfp()s. | |
1981 | */ | |
9888e6fa | 1982 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1983 | break; |
1984 | ||
9888e6fa LT |
1985 | /* |
1986 | * Acceptable internal fragmentation? | |
1987 | */ | |
a737b3e2 | 1988 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
1989 | break; |
1990 | } | |
1991 | return left_over; | |
1992 | } | |
1993 | ||
2ed3a4ef | 1994 | static int setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 1995 | { |
2ed3a4ef CL |
1996 | if (g_cpucache_up == FULL) |
1997 | return enable_cpucache(cachep); | |
1998 | ||
f30cf7d1 PE |
1999 | if (g_cpucache_up == NONE) { |
2000 | /* | |
2001 | * Note: the first kmem_cache_create must create the cache | |
2002 | * that's used by kmalloc(24), otherwise the creation of | |
2003 | * further caches will BUG(). | |
2004 | */ | |
2005 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2006 | ||
2007 | /* | |
2008 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2009 | * the first cache, then we need to set up all its list3s, | |
2010 | * otherwise the creation of further caches will BUG(). | |
2011 | */ | |
2012 | set_up_list3s(cachep, SIZE_AC); | |
2013 | if (INDEX_AC == INDEX_L3) | |
2014 | g_cpucache_up = PARTIAL_L3; | |
2015 | else | |
2016 | g_cpucache_up = PARTIAL_AC; | |
2017 | } else { | |
2018 | cachep->array[smp_processor_id()] = | |
2019 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
2020 | ||
2021 | if (g_cpucache_up == PARTIAL_AC) { | |
2022 | set_up_list3s(cachep, SIZE_L3); | |
2023 | g_cpucache_up = PARTIAL_L3; | |
2024 | } else { | |
2025 | int node; | |
2026 | for_each_online_node(node) { | |
2027 | cachep->nodelists[node] = | |
2028 | kmalloc_node(sizeof(struct kmem_list3), | |
2029 | GFP_KERNEL, node); | |
2030 | BUG_ON(!cachep->nodelists[node]); | |
2031 | kmem_list3_init(cachep->nodelists[node]); | |
2032 | } | |
2033 | } | |
2034 | } | |
2035 | cachep->nodelists[numa_node_id()]->next_reap = | |
2036 | jiffies + REAPTIMEOUT_LIST3 + | |
2037 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2038 | ||
2039 | cpu_cache_get(cachep)->avail = 0; | |
2040 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2041 | cpu_cache_get(cachep)->batchcount = 1; | |
2042 | cpu_cache_get(cachep)->touched = 0; | |
2043 | cachep->batchcount = 1; | |
2044 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2045 | return 0; |
f30cf7d1 PE |
2046 | } |
2047 | ||
1da177e4 LT |
2048 | /** |
2049 | * kmem_cache_create - Create a cache. | |
2050 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2051 | * @size: The size of objects to be created in this cache. | |
2052 | * @align: The required alignment for the objects. | |
2053 | * @flags: SLAB flags | |
2054 | * @ctor: A constructor for the objects. | |
2055 | * @dtor: A destructor for the objects. | |
2056 | * | |
2057 | * Returns a ptr to the cache on success, NULL on failure. | |
2058 | * Cannot be called within a int, but can be interrupted. | |
2059 | * The @ctor is run when new pages are allocated by the cache | |
2060 | * and the @dtor is run before the pages are handed back. | |
2061 | * | |
2062 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2063 | * the module calling this has to destroy the cache before getting unloaded. |
2064 | * | |
1da177e4 LT |
2065 | * The flags are |
2066 | * | |
2067 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2068 | * to catch references to uninitialised memory. | |
2069 | * | |
2070 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2071 | * for buffer overruns. | |
2072 | * | |
1da177e4 LT |
2073 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2074 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2075 | * as davem. | |
2076 | */ | |
343e0d7a | 2077 | struct kmem_cache * |
1da177e4 | 2078 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
2079 | unsigned long flags, |
2080 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 2081 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
2082 | { |
2083 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2084 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2085 | |
2086 | /* | |
2087 | * Sanity checks... these are all serious usage bugs. | |
2088 | */ | |
a737b3e2 | 2089 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
b28a02de | 2090 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { |
a737b3e2 AM |
2091 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
2092 | name); | |
b28a02de PE |
2093 | BUG(); |
2094 | } | |
1da177e4 | 2095 | |
f0188f47 RT |
2096 | /* |
2097 | * Prevent CPUs from coming and going. | |
2098 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
2099 | */ | |
2100 | lock_cpu_hotplug(); | |
2101 | ||
fc0abb14 | 2102 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2103 | |
7a7c381d | 2104 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2105 | mm_segment_t old_fs = get_fs(); |
2106 | char tmp; | |
2107 | int res; | |
2108 | ||
2109 | /* | |
2110 | * This happens when the module gets unloaded and doesn't | |
2111 | * destroy its slab cache and no-one else reuses the vmalloc | |
2112 | * area of the module. Print a warning. | |
2113 | */ | |
2114 | set_fs(KERNEL_DS); | |
2115 | res = __get_user(tmp, pc->name); | |
2116 | set_fs(old_fs); | |
2117 | if (res) { | |
2118 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2119 | pc->buffer_size); |
4f12bb4f AM |
2120 | continue; |
2121 | } | |
2122 | ||
b28a02de | 2123 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
2124 | printk("kmem_cache_create: duplicate cache %s\n", name); |
2125 | dump_stack(); | |
2126 | goto oops; | |
2127 | } | |
2128 | } | |
2129 | ||
1da177e4 LT |
2130 | #if DEBUG |
2131 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
2132 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
2133 | /* No constructor, but inital state check requested */ | |
2134 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 2135 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
2136 | flags &= ~SLAB_DEBUG_INITIAL; |
2137 | } | |
1da177e4 LT |
2138 | #if FORCED_DEBUG |
2139 | /* | |
2140 | * Enable redzoning and last user accounting, except for caches with | |
2141 | * large objects, if the increased size would increase the object size | |
2142 | * above the next power of two: caches with object sizes just above a | |
2143 | * power of two have a significant amount of internal fragmentation. | |
2144 | */ | |
a737b3e2 | 2145 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 2146 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2147 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2148 | flags |= SLAB_POISON; | |
2149 | #endif | |
2150 | if (flags & SLAB_DESTROY_BY_RCU) | |
2151 | BUG_ON(flags & SLAB_POISON); | |
2152 | #endif | |
2153 | if (flags & SLAB_DESTROY_BY_RCU) | |
2154 | BUG_ON(dtor); | |
2155 | ||
2156 | /* | |
a737b3e2 AM |
2157 | * Always checks flags, a caller might be expecting debug support which |
2158 | * isn't available. | |
1da177e4 | 2159 | */ |
40094fa6 | 2160 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2161 | |
a737b3e2 AM |
2162 | /* |
2163 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2164 | * unaligned accesses for some archs when redzoning is used, and makes |
2165 | * sure any on-slab bufctl's are also correctly aligned. | |
2166 | */ | |
b28a02de PE |
2167 | if (size & (BYTES_PER_WORD - 1)) { |
2168 | size += (BYTES_PER_WORD - 1); | |
2169 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2170 | } |
2171 | ||
a737b3e2 AM |
2172 | /* calculate the final buffer alignment: */ |
2173 | ||
1da177e4 LT |
2174 | /* 1) arch recommendation: can be overridden for debug */ |
2175 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2176 | /* |
2177 | * Default alignment: as specified by the arch code. Except if | |
2178 | * an object is really small, then squeeze multiple objects into | |
2179 | * one cacheline. | |
1da177e4 LT |
2180 | */ |
2181 | ralign = cache_line_size(); | |
b28a02de | 2182 | while (size <= ralign / 2) |
1da177e4 LT |
2183 | ralign /= 2; |
2184 | } else { | |
2185 | ralign = BYTES_PER_WORD; | |
2186 | } | |
ca5f9703 PE |
2187 | |
2188 | /* | |
2189 | * Redzoning and user store require word alignment. Note this will be | |
2190 | * overridden by architecture or caller mandated alignment if either | |
2191 | * is greater than BYTES_PER_WORD. | |
2192 | */ | |
2193 | if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) | |
2194 | ralign = BYTES_PER_WORD; | |
2195 | ||
1da177e4 LT |
2196 | /* 2) arch mandated alignment: disables debug if necessary */ |
2197 | if (ralign < ARCH_SLAB_MINALIGN) { | |
2198 | ralign = ARCH_SLAB_MINALIGN; | |
2199 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2200 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
2201 | } |
2202 | /* 3) caller mandated alignment: disables debug if necessary */ | |
2203 | if (ralign < align) { | |
2204 | ralign = align; | |
2205 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2206 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 | 2207 | } |
a737b3e2 | 2208 | /* |
ca5f9703 | 2209 | * 4) Store it. |
1da177e4 LT |
2210 | */ |
2211 | align = ralign; | |
2212 | ||
2213 | /* Get cache's description obj. */ | |
c5e3b83e | 2214 | cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 2215 | if (!cachep) |
4f12bb4f | 2216 | goto oops; |
1da177e4 LT |
2217 | |
2218 | #if DEBUG | |
3dafccf2 | 2219 | cachep->obj_size = size; |
1da177e4 | 2220 | |
ca5f9703 PE |
2221 | /* |
2222 | * Both debugging options require word-alignment which is calculated | |
2223 | * into align above. | |
2224 | */ | |
1da177e4 | 2225 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2226 | /* add space for red zone words */ |
3dafccf2 | 2227 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 2228 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
2229 | } |
2230 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 PE |
2231 | /* user store requires one word storage behind the end of |
2232 | * the real object. | |
1da177e4 | 2233 | */ |
1da177e4 LT |
2234 | size += BYTES_PER_WORD; |
2235 | } | |
2236 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2237 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2238 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2239 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2240 | size = PAGE_SIZE; |
2241 | } | |
2242 | #endif | |
2243 | #endif | |
2244 | ||
e0a42726 IM |
2245 | /* |
2246 | * Determine if the slab management is 'on' or 'off' slab. | |
2247 | * (bootstrapping cannot cope with offslab caches so don't do | |
2248 | * it too early on.) | |
2249 | */ | |
2250 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2251 | /* |
2252 | * Size is large, assume best to place the slab management obj | |
2253 | * off-slab (should allow better packing of objs). | |
2254 | */ | |
2255 | flags |= CFLGS_OFF_SLAB; | |
2256 | ||
2257 | size = ALIGN(size, align); | |
2258 | ||
f78bb8ad | 2259 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2260 | |
2261 | if (!cachep->num) { | |
2262 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
2263 | kmem_cache_free(&cache_cache, cachep); | |
2264 | cachep = NULL; | |
4f12bb4f | 2265 | goto oops; |
1da177e4 | 2266 | } |
b28a02de PE |
2267 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2268 | + sizeof(struct slab), align); | |
1da177e4 LT |
2269 | |
2270 | /* | |
2271 | * If the slab has been placed off-slab, and we have enough space then | |
2272 | * move it on-slab. This is at the expense of any extra colouring. | |
2273 | */ | |
2274 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2275 | flags &= ~CFLGS_OFF_SLAB; | |
2276 | left_over -= slab_size; | |
2277 | } | |
2278 | ||
2279 | if (flags & CFLGS_OFF_SLAB) { | |
2280 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2281 | slab_size = |
2282 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2283 | } |
2284 | ||
2285 | cachep->colour_off = cache_line_size(); | |
2286 | /* Offset must be a multiple of the alignment. */ | |
2287 | if (cachep->colour_off < align) | |
2288 | cachep->colour_off = align; | |
b28a02de | 2289 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2290 | cachep->slab_size = slab_size; |
2291 | cachep->flags = flags; | |
2292 | cachep->gfpflags = 0; | |
2293 | if (flags & SLAB_CACHE_DMA) | |
2294 | cachep->gfpflags |= GFP_DMA; | |
3dafccf2 | 2295 | cachep->buffer_size = size; |
1da177e4 | 2296 | |
e5ac9c5a | 2297 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2298 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2299 | /* |
2300 | * This is a possibility for one of the malloc_sizes caches. | |
2301 | * But since we go off slab only for object size greater than | |
2302 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2303 | * this should not happen at all. | |
2304 | * But leave a BUG_ON for some lucky dude. | |
2305 | */ | |
2306 | BUG_ON(!cachep->slabp_cache); | |
2307 | } | |
1da177e4 LT |
2308 | cachep->ctor = ctor; |
2309 | cachep->dtor = dtor; | |
2310 | cachep->name = name; | |
2311 | ||
2ed3a4ef CL |
2312 | if (setup_cpu_cache(cachep)) { |
2313 | __kmem_cache_destroy(cachep); | |
2314 | cachep = NULL; | |
2315 | goto oops; | |
2316 | } | |
1da177e4 | 2317 | |
1da177e4 LT |
2318 | /* cache setup completed, link it into the list */ |
2319 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2320 | oops: |
1da177e4 LT |
2321 | if (!cachep && (flags & SLAB_PANIC)) |
2322 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2323 | name); |
fc0abb14 | 2324 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 2325 | unlock_cpu_hotplug(); |
1da177e4 LT |
2326 | return cachep; |
2327 | } | |
2328 | EXPORT_SYMBOL(kmem_cache_create); | |
2329 | ||
2330 | #if DEBUG | |
2331 | static void check_irq_off(void) | |
2332 | { | |
2333 | BUG_ON(!irqs_disabled()); | |
2334 | } | |
2335 | ||
2336 | static void check_irq_on(void) | |
2337 | { | |
2338 | BUG_ON(irqs_disabled()); | |
2339 | } | |
2340 | ||
343e0d7a | 2341 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2342 | { |
2343 | #ifdef CONFIG_SMP | |
2344 | check_irq_off(); | |
e498be7d | 2345 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2346 | #endif |
2347 | } | |
e498be7d | 2348 | |
343e0d7a | 2349 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2350 | { |
2351 | #ifdef CONFIG_SMP | |
2352 | check_irq_off(); | |
2353 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2354 | #endif | |
2355 | } | |
2356 | ||
1da177e4 LT |
2357 | #else |
2358 | #define check_irq_off() do { } while(0) | |
2359 | #define check_irq_on() do { } while(0) | |
2360 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2361 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2362 | #endif |
2363 | ||
aab2207c CL |
2364 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2365 | struct array_cache *ac, | |
2366 | int force, int node); | |
2367 | ||
1da177e4 LT |
2368 | static void do_drain(void *arg) |
2369 | { | |
a737b3e2 | 2370 | struct kmem_cache *cachep = arg; |
1da177e4 | 2371 | struct array_cache *ac; |
ff69416e | 2372 | int node = numa_node_id(); |
1da177e4 LT |
2373 | |
2374 | check_irq_off(); | |
9a2dba4b | 2375 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2376 | spin_lock(&cachep->nodelists[node]->list_lock); |
2377 | free_block(cachep, ac->entry, ac->avail, node); | |
2378 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2379 | ac->avail = 0; |
2380 | } | |
2381 | ||
343e0d7a | 2382 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2383 | { |
e498be7d CL |
2384 | struct kmem_list3 *l3; |
2385 | int node; | |
2386 | ||
a07fa394 | 2387 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2388 | check_irq_on(); |
b28a02de | 2389 | for_each_online_node(node) { |
e498be7d | 2390 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2391 | if (l3 && l3->alien) |
2392 | drain_alien_cache(cachep, l3->alien); | |
2393 | } | |
2394 | ||
2395 | for_each_online_node(node) { | |
2396 | l3 = cachep->nodelists[node]; | |
2397 | if (l3) | |
aab2207c | 2398 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2399 | } |
1da177e4 LT |
2400 | } |
2401 | ||
ed11d9eb CL |
2402 | /* |
2403 | * Remove slabs from the list of free slabs. | |
2404 | * Specify the number of slabs to drain in tofree. | |
2405 | * | |
2406 | * Returns the actual number of slabs released. | |
2407 | */ | |
2408 | static int drain_freelist(struct kmem_cache *cache, | |
2409 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2410 | { |
ed11d9eb CL |
2411 | struct list_head *p; |
2412 | int nr_freed; | |
1da177e4 | 2413 | struct slab *slabp; |
1da177e4 | 2414 | |
ed11d9eb CL |
2415 | nr_freed = 0; |
2416 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2417 | |
ed11d9eb | 2418 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2419 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2420 | if (p == &l3->slabs_free) { |
2421 | spin_unlock_irq(&l3->list_lock); | |
2422 | goto out; | |
2423 | } | |
1da177e4 | 2424 | |
ed11d9eb | 2425 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2426 | #if DEBUG |
40094fa6 | 2427 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2428 | #endif |
2429 | list_del(&slabp->list); | |
ed11d9eb CL |
2430 | /* |
2431 | * Safe to drop the lock. The slab is no longer linked | |
2432 | * to the cache. | |
2433 | */ | |
2434 | l3->free_objects -= cache->num; | |
e498be7d | 2435 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2436 | slab_destroy(cache, slabp); |
2437 | nr_freed++; | |
1da177e4 | 2438 | } |
ed11d9eb CL |
2439 | out: |
2440 | return nr_freed; | |
1da177e4 LT |
2441 | } |
2442 | ||
343e0d7a | 2443 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2444 | { |
2445 | int ret = 0, i = 0; | |
2446 | struct kmem_list3 *l3; | |
2447 | ||
2448 | drain_cpu_caches(cachep); | |
2449 | ||
2450 | check_irq_on(); | |
2451 | for_each_online_node(i) { | |
2452 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2453 | if (!l3) |
2454 | continue; | |
2455 | ||
2456 | drain_freelist(cachep, l3, l3->free_objects); | |
2457 | ||
2458 | ret += !list_empty(&l3->slabs_full) || | |
2459 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2460 | } |
2461 | return (ret ? 1 : 0); | |
2462 | } | |
2463 | ||
1da177e4 LT |
2464 | /** |
2465 | * kmem_cache_shrink - Shrink a cache. | |
2466 | * @cachep: The cache to shrink. | |
2467 | * | |
2468 | * Releases as many slabs as possible for a cache. | |
2469 | * To help debugging, a zero exit status indicates all slabs were released. | |
2470 | */ | |
343e0d7a | 2471 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2472 | { |
40094fa6 | 2473 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2474 | |
2475 | return __cache_shrink(cachep); | |
2476 | } | |
2477 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2478 | ||
2479 | /** | |
2480 | * kmem_cache_destroy - delete a cache | |
2481 | * @cachep: the cache to destroy | |
2482 | * | |
343e0d7a | 2483 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2484 | * |
2485 | * It is expected this function will be called by a module when it is | |
2486 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2487 | * cache being allocated each time a module is loaded and unloaded, if the | |
2488 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2489 | * | |
2490 | * The cache must be empty before calling this function. | |
2491 | * | |
2492 | * The caller must guarantee that noone will allocate memory from the cache | |
2493 | * during the kmem_cache_destroy(). | |
2494 | */ | |
133d205a | 2495 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2496 | { |
40094fa6 | 2497 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2498 | |
2499 | /* Don't let CPUs to come and go */ | |
2500 | lock_cpu_hotplug(); | |
2501 | ||
2502 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2503 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2504 | /* |
2505 | * the chain is never empty, cache_cache is never destroyed | |
2506 | */ | |
2507 | list_del(&cachep->next); | |
fc0abb14 | 2508 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2509 | |
2510 | if (__cache_shrink(cachep)) { | |
2511 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2512 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2513 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2514 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 2515 | unlock_cpu_hotplug(); |
133d205a | 2516 | return; |
1da177e4 LT |
2517 | } |
2518 | ||
2519 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2520 | synchronize_rcu(); |
1da177e4 | 2521 | |
117f6eb1 | 2522 | __kmem_cache_destroy(cachep); |
1da177e4 | 2523 | unlock_cpu_hotplug(); |
1da177e4 LT |
2524 | } |
2525 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2526 | ||
e5ac9c5a RT |
2527 | /* |
2528 | * Get the memory for a slab management obj. | |
2529 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2530 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2531 | * come from the same cache which is getting created because, | |
2532 | * when we are searching for an appropriate cache for these | |
2533 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2534 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2535 | * kmem_find_general_cachep till the initialization is complete. | |
2536 | * Hence we cannot have slabp_cache same as the original cache. | |
2537 | */ | |
343e0d7a | 2538 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2539 | int colour_off, gfp_t local_flags, |
2540 | int nodeid) | |
1da177e4 LT |
2541 | { |
2542 | struct slab *slabp; | |
b28a02de | 2543 | |
1da177e4 LT |
2544 | if (OFF_SLAB(cachep)) { |
2545 | /* Slab management obj is off-slab. */ | |
5b74ada7 RT |
2546 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
2547 | local_flags, nodeid); | |
1da177e4 LT |
2548 | if (!slabp) |
2549 | return NULL; | |
2550 | } else { | |
b28a02de | 2551 | slabp = objp + colour_off; |
1da177e4 LT |
2552 | colour_off += cachep->slab_size; |
2553 | } | |
2554 | slabp->inuse = 0; | |
2555 | slabp->colouroff = colour_off; | |
b28a02de | 2556 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2557 | slabp->nodeid = nodeid; |
1da177e4 LT |
2558 | return slabp; |
2559 | } | |
2560 | ||
2561 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2562 | { | |
b28a02de | 2563 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2564 | } |
2565 | ||
343e0d7a | 2566 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2567 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2568 | { |
2569 | int i; | |
2570 | ||
2571 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2572 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2573 | #if DEBUG |
2574 | /* need to poison the objs? */ | |
2575 | if (cachep->flags & SLAB_POISON) | |
2576 | poison_obj(cachep, objp, POISON_FREE); | |
2577 | if (cachep->flags & SLAB_STORE_USER) | |
2578 | *dbg_userword(cachep, objp) = NULL; | |
2579 | ||
2580 | if (cachep->flags & SLAB_RED_ZONE) { | |
2581 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2582 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2583 | } | |
2584 | /* | |
a737b3e2 AM |
2585 | * Constructors are not allowed to allocate memory from the same |
2586 | * cache which they are a constructor for. Otherwise, deadlock. | |
2587 | * They must also be threaded. | |
1da177e4 LT |
2588 | */ |
2589 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2590 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2591 | ctor_flags); |
1da177e4 LT |
2592 | |
2593 | if (cachep->flags & SLAB_RED_ZONE) { | |
2594 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2595 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2596 | " end of an object"); |
1da177e4 LT |
2597 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2598 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2599 | " start of an object"); |
1da177e4 | 2600 | } |
a737b3e2 AM |
2601 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2602 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2603 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2604 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2605 | #else |
2606 | if (cachep->ctor) | |
2607 | cachep->ctor(objp, cachep, ctor_flags); | |
2608 | #endif | |
b28a02de | 2609 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2610 | } |
b28a02de | 2611 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2612 | slabp->free = 0; |
2613 | } | |
2614 | ||
343e0d7a | 2615 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2616 | { |
a737b3e2 AM |
2617 | if (flags & SLAB_DMA) |
2618 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2619 | else | |
2620 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
1da177e4 LT |
2621 | } |
2622 | ||
a737b3e2 AM |
2623 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2624 | int nodeid) | |
78d382d7 | 2625 | { |
8fea4e96 | 2626 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2627 | kmem_bufctl_t next; |
2628 | ||
2629 | slabp->inuse++; | |
2630 | next = slab_bufctl(slabp)[slabp->free]; | |
2631 | #if DEBUG | |
2632 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2633 | WARN_ON(slabp->nodeid != nodeid); | |
2634 | #endif | |
2635 | slabp->free = next; | |
2636 | ||
2637 | return objp; | |
2638 | } | |
2639 | ||
a737b3e2 AM |
2640 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2641 | void *objp, int nodeid) | |
78d382d7 | 2642 | { |
8fea4e96 | 2643 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2644 | |
2645 | #if DEBUG | |
2646 | /* Verify that the slab belongs to the intended node */ | |
2647 | WARN_ON(slabp->nodeid != nodeid); | |
2648 | ||
871751e2 | 2649 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2650 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2651 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2652 | BUG(); |
2653 | } | |
2654 | #endif | |
2655 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2656 | slabp->free = objnr; | |
2657 | slabp->inuse--; | |
2658 | } | |
2659 | ||
4776874f PE |
2660 | /* |
2661 | * Map pages beginning at addr to the given cache and slab. This is required | |
2662 | * for the slab allocator to be able to lookup the cache and slab of a | |
2663 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2664 | */ | |
2665 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2666 | void *addr) | |
1da177e4 | 2667 | { |
4776874f | 2668 | int nr_pages; |
1da177e4 LT |
2669 | struct page *page; |
2670 | ||
4776874f | 2671 | page = virt_to_page(addr); |
84097518 | 2672 | |
4776874f | 2673 | nr_pages = 1; |
84097518 | 2674 | if (likely(!PageCompound(page))) |
4776874f PE |
2675 | nr_pages <<= cache->gfporder; |
2676 | ||
1da177e4 | 2677 | do { |
4776874f PE |
2678 | page_set_cache(page, cache); |
2679 | page_set_slab(page, slab); | |
1da177e4 | 2680 | page++; |
4776874f | 2681 | } while (--nr_pages); |
1da177e4 LT |
2682 | } |
2683 | ||
2684 | /* | |
2685 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2686 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2687 | */ | |
343e0d7a | 2688 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2689 | { |
b28a02de PE |
2690 | struct slab *slabp; |
2691 | void *objp; | |
2692 | size_t offset; | |
2693 | gfp_t local_flags; | |
2694 | unsigned long ctor_flags; | |
e498be7d | 2695 | struct kmem_list3 *l3; |
1da177e4 | 2696 | |
a737b3e2 AM |
2697 | /* |
2698 | * Be lazy and only check for valid flags here, keeping it out of the | |
2699 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2700 | */ |
40094fa6 | 2701 | BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); |
1da177e4 LT |
2702 | if (flags & SLAB_NO_GROW) |
2703 | return 0; | |
2704 | ||
2705 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2706 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2707 | if (!(local_flags & __GFP_WAIT)) | |
2708 | /* | |
2709 | * Not allowed to sleep. Need to tell a constructor about | |
2710 | * this - it might need to know... | |
2711 | */ | |
2712 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2713 | ||
2e1217cf | 2714 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2715 | check_irq_off(); |
2e1217cf RT |
2716 | l3 = cachep->nodelists[nodeid]; |
2717 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2718 | |
2719 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2720 | offset = l3->colour_next; |
2721 | l3->colour_next++; | |
2722 | if (l3->colour_next >= cachep->colour) | |
2723 | l3->colour_next = 0; | |
2724 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2725 | |
2e1217cf | 2726 | offset *= cachep->colour_off; |
1da177e4 LT |
2727 | |
2728 | if (local_flags & __GFP_WAIT) | |
2729 | local_irq_enable(); | |
2730 | ||
2731 | /* | |
2732 | * The test for missing atomic flag is performed here, rather than | |
2733 | * the more obvious place, simply to reduce the critical path length | |
2734 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2735 | * will eventually be caught here (where it matters). | |
2736 | */ | |
2737 | kmem_flagcheck(cachep, flags); | |
2738 | ||
a737b3e2 AM |
2739 | /* |
2740 | * Get mem for the objs. Attempt to allocate a physical page from | |
2741 | * 'nodeid'. | |
e498be7d | 2742 | */ |
a737b3e2 AM |
2743 | objp = kmem_getpages(cachep, flags, nodeid); |
2744 | if (!objp) | |
1da177e4 LT |
2745 | goto failed; |
2746 | ||
2747 | /* Get slab management. */ | |
5b74ada7 | 2748 | slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); |
a737b3e2 | 2749 | if (!slabp) |
1da177e4 LT |
2750 | goto opps1; |
2751 | ||
e498be7d | 2752 | slabp->nodeid = nodeid; |
4776874f | 2753 | slab_map_pages(cachep, slabp, objp); |
1da177e4 LT |
2754 | |
2755 | cache_init_objs(cachep, slabp, ctor_flags); | |
2756 | ||
2757 | if (local_flags & __GFP_WAIT) | |
2758 | local_irq_disable(); | |
2759 | check_irq_off(); | |
e498be7d | 2760 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2761 | |
2762 | /* Make slab active. */ | |
e498be7d | 2763 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2764 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2765 | l3->free_objects += cachep->num; |
2766 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2767 | return 1; |
a737b3e2 | 2768 | opps1: |
1da177e4 | 2769 | kmem_freepages(cachep, objp); |
a737b3e2 | 2770 | failed: |
1da177e4 LT |
2771 | if (local_flags & __GFP_WAIT) |
2772 | local_irq_disable(); | |
2773 | return 0; | |
2774 | } | |
2775 | ||
2776 | #if DEBUG | |
2777 | ||
2778 | /* | |
2779 | * Perform extra freeing checks: | |
2780 | * - detect bad pointers. | |
2781 | * - POISON/RED_ZONE checking | |
2782 | * - destructor calls, for caches with POISON+dtor | |
2783 | */ | |
2784 | static void kfree_debugcheck(const void *objp) | |
2785 | { | |
2786 | struct page *page; | |
2787 | ||
2788 | if (!virt_addr_valid(objp)) { | |
2789 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2790 | (unsigned long)objp); |
2791 | BUG(); | |
1da177e4 LT |
2792 | } |
2793 | page = virt_to_page(objp); | |
2794 | if (!PageSlab(page)) { | |
b28a02de PE |
2795 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2796 | (unsigned long)objp); | |
1da177e4 LT |
2797 | BUG(); |
2798 | } | |
2799 | } | |
2800 | ||
58ce1fd5 PE |
2801 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2802 | { | |
2803 | unsigned long redzone1, redzone2; | |
2804 | ||
2805 | redzone1 = *dbg_redzone1(cache, obj); | |
2806 | redzone2 = *dbg_redzone2(cache, obj); | |
2807 | ||
2808 | /* | |
2809 | * Redzone is ok. | |
2810 | */ | |
2811 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2812 | return; | |
2813 | ||
2814 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2815 | slab_error(cache, "double free detected"); | |
2816 | else | |
2817 | slab_error(cache, "memory outside object was overwritten"); | |
2818 | ||
2819 | printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n", | |
2820 | obj, redzone1, redzone2); | |
2821 | } | |
2822 | ||
343e0d7a | 2823 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2824 | void *caller) |
1da177e4 LT |
2825 | { |
2826 | struct page *page; | |
2827 | unsigned int objnr; | |
2828 | struct slab *slabp; | |
2829 | ||
3dafccf2 | 2830 | objp -= obj_offset(cachep); |
1da177e4 LT |
2831 | kfree_debugcheck(objp); |
2832 | page = virt_to_page(objp); | |
2833 | ||
065d41cb | 2834 | slabp = page_get_slab(page); |
1da177e4 LT |
2835 | |
2836 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2837 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2838 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2839 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2840 | } | |
2841 | if (cachep->flags & SLAB_STORE_USER) | |
2842 | *dbg_userword(cachep, objp) = caller; | |
2843 | ||
8fea4e96 | 2844 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2845 | |
2846 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2847 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2848 | |
2849 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
a737b3e2 AM |
2850 | /* |
2851 | * Need to call the slab's constructor so the caller can | |
2852 | * perform a verify of its state (debugging). Called without | |
2853 | * the cache-lock held. | |
1da177e4 | 2854 | */ |
3dafccf2 | 2855 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2856 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2857 | } |
2858 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2859 | /* we want to cache poison the object, | |
2860 | * call the destruction callback | |
2861 | */ | |
3dafccf2 | 2862 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 2863 | } |
871751e2 AV |
2864 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2865 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2866 | #endif | |
1da177e4 LT |
2867 | if (cachep->flags & SLAB_POISON) { |
2868 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2869 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2870 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2871 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2872 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2873 | } else { |
2874 | poison_obj(cachep, objp, POISON_FREE); | |
2875 | } | |
2876 | #else | |
2877 | poison_obj(cachep, objp, POISON_FREE); | |
2878 | #endif | |
2879 | } | |
2880 | return objp; | |
2881 | } | |
2882 | ||
343e0d7a | 2883 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2884 | { |
2885 | kmem_bufctl_t i; | |
2886 | int entries = 0; | |
b28a02de | 2887 | |
1da177e4 LT |
2888 | /* Check slab's freelist to see if this obj is there. */ |
2889 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2890 | entries++; | |
2891 | if (entries > cachep->num || i >= cachep->num) | |
2892 | goto bad; | |
2893 | } | |
2894 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2895 | bad: |
2896 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2897 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2898 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2899 | for (i = 0; |
264132bc | 2900 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2901 | i++) { |
a737b3e2 | 2902 | if (i % 16 == 0) |
1da177e4 | 2903 | printk("\n%03x:", i); |
b28a02de | 2904 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2905 | } |
2906 | printk("\n"); | |
2907 | BUG(); | |
2908 | } | |
2909 | } | |
2910 | #else | |
2911 | #define kfree_debugcheck(x) do { } while(0) | |
2912 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2913 | #define check_slabp(x,y) do { } while(0) | |
2914 | #endif | |
2915 | ||
343e0d7a | 2916 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2917 | { |
2918 | int batchcount; | |
2919 | struct kmem_list3 *l3; | |
2920 | struct array_cache *ac; | |
2921 | ||
2922 | check_irq_off(); | |
9a2dba4b | 2923 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2924 | retry: |
1da177e4 LT |
2925 | batchcount = ac->batchcount; |
2926 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2927 | /* |
2928 | * If there was little recent activity on this cache, then | |
2929 | * perform only a partial refill. Otherwise we could generate | |
2930 | * refill bouncing. | |
1da177e4 LT |
2931 | */ |
2932 | batchcount = BATCHREFILL_LIMIT; | |
2933 | } | |
e498be7d CL |
2934 | l3 = cachep->nodelists[numa_node_id()]; |
2935 | ||
2936 | BUG_ON(ac->avail > 0 || !l3); | |
2937 | spin_lock(&l3->list_lock); | |
1da177e4 | 2938 | |
3ded175a CL |
2939 | /* See if we can refill from the shared array */ |
2940 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2941 | goto alloc_done; | |
2942 | ||
1da177e4 LT |
2943 | while (batchcount > 0) { |
2944 | struct list_head *entry; | |
2945 | struct slab *slabp; | |
2946 | /* Get slab alloc is to come from. */ | |
2947 | entry = l3->slabs_partial.next; | |
2948 | if (entry == &l3->slabs_partial) { | |
2949 | l3->free_touched = 1; | |
2950 | entry = l3->slabs_free.next; | |
2951 | if (entry == &l3->slabs_free) | |
2952 | goto must_grow; | |
2953 | } | |
2954 | ||
2955 | slabp = list_entry(entry, struct slab, list); | |
2956 | check_slabp(cachep, slabp); | |
2957 | check_spinlock_acquired(cachep); | |
2958 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2959 | STATS_INC_ALLOCED(cachep); |
2960 | STATS_INC_ACTIVE(cachep); | |
2961 | STATS_SET_HIGH(cachep); | |
2962 | ||
78d382d7 MD |
2963 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2964 | numa_node_id()); | |
1da177e4 LT |
2965 | } |
2966 | check_slabp(cachep, slabp); | |
2967 | ||
2968 | /* move slabp to correct slabp list: */ | |
2969 | list_del(&slabp->list); | |
2970 | if (slabp->free == BUFCTL_END) | |
2971 | list_add(&slabp->list, &l3->slabs_full); | |
2972 | else | |
2973 | list_add(&slabp->list, &l3->slabs_partial); | |
2974 | } | |
2975 | ||
a737b3e2 | 2976 | must_grow: |
1da177e4 | 2977 | l3->free_objects -= ac->avail; |
a737b3e2 | 2978 | alloc_done: |
e498be7d | 2979 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2980 | |
2981 | if (unlikely(!ac->avail)) { | |
2982 | int x; | |
e498be7d CL |
2983 | x = cache_grow(cachep, flags, numa_node_id()); |
2984 | ||
a737b3e2 | 2985 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 2986 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2987 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
2988 | return NULL; |
2989 | ||
a737b3e2 | 2990 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
2991 | goto retry; |
2992 | } | |
2993 | ac->touched = 1; | |
e498be7d | 2994 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2995 | } |
2996 | ||
a737b3e2 AM |
2997 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
2998 | gfp_t flags) | |
1da177e4 LT |
2999 | { |
3000 | might_sleep_if(flags & __GFP_WAIT); | |
3001 | #if DEBUG | |
3002 | kmem_flagcheck(cachep, flags); | |
3003 | #endif | |
3004 | } | |
3005 | ||
3006 | #if DEBUG | |
a737b3e2 AM |
3007 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3008 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3009 | { |
b28a02de | 3010 | if (!objp) |
1da177e4 | 3011 | return objp; |
b28a02de | 3012 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3013 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3014 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3015 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3016 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3017 | else |
3018 | check_poison_obj(cachep, objp); | |
3019 | #else | |
3020 | check_poison_obj(cachep, objp); | |
3021 | #endif | |
3022 | poison_obj(cachep, objp, POISON_INUSE); | |
3023 | } | |
3024 | if (cachep->flags & SLAB_STORE_USER) | |
3025 | *dbg_userword(cachep, objp) = caller; | |
3026 | ||
3027 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3028 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3029 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3030 | slab_error(cachep, "double free, or memory outside" | |
3031 | " object was overwritten"); | |
b28a02de | 3032 | printk(KERN_ERR |
a737b3e2 AM |
3033 | "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", |
3034 | objp, *dbg_redzone1(cachep, objp), | |
3035 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3036 | } |
3037 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3038 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3039 | } | |
871751e2 AV |
3040 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3041 | { | |
3042 | struct slab *slabp; | |
3043 | unsigned objnr; | |
3044 | ||
3045 | slabp = page_get_slab(virt_to_page(objp)); | |
3046 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; | |
3047 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3048 | } | |
3049 | #endif | |
3dafccf2 | 3050 | objp += obj_offset(cachep); |
1da177e4 | 3051 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 3052 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
3053 | |
3054 | if (!(flags & __GFP_WAIT)) | |
3055 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
3056 | ||
3057 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 3058 | } |
1da177e4 LT |
3059 | return objp; |
3060 | } | |
3061 | #else | |
3062 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3063 | #endif | |
3064 | ||
343e0d7a | 3065 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3066 | { |
b28a02de | 3067 | void *objp; |
1da177e4 LT |
3068 | struct array_cache *ac; |
3069 | ||
5c382300 | 3070 | check_irq_off(); |
9a2dba4b | 3071 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3072 | if (likely(ac->avail)) { |
3073 | STATS_INC_ALLOCHIT(cachep); | |
3074 | ac->touched = 1; | |
e498be7d | 3075 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3076 | } else { |
3077 | STATS_INC_ALLOCMISS(cachep); | |
3078 | objp = cache_alloc_refill(cachep, flags); | |
3079 | } | |
5c382300 AK |
3080 | return objp; |
3081 | } | |
3082 | ||
a737b3e2 AM |
3083 | static __always_inline void *__cache_alloc(struct kmem_cache *cachep, |
3084 | gfp_t flags, void *caller) | |
5c382300 AK |
3085 | { |
3086 | unsigned long save_flags; | |
de3083ec | 3087 | void *objp = NULL; |
5c382300 AK |
3088 | |
3089 | cache_alloc_debugcheck_before(cachep, flags); | |
3090 | ||
3091 | local_irq_save(save_flags); | |
de3083ec | 3092 | |
765c4507 CL |
3093 | if (unlikely(NUMA_BUILD && |
3094 | current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) | |
de3083ec | 3095 | objp = alternate_node_alloc(cachep, flags); |
de3083ec CL |
3096 | |
3097 | if (!objp) | |
3098 | objp = ____cache_alloc(cachep, flags); | |
765c4507 CL |
3099 | /* |
3100 | * We may just have run out of memory on the local node. | |
3101 | * __cache_alloc_node() knows how to locate memory on other nodes | |
3102 | */ | |
3103 | if (NUMA_BUILD && !objp) | |
3104 | objp = __cache_alloc_node(cachep, flags, numa_node_id()); | |
1da177e4 | 3105 | local_irq_restore(save_flags); |
34342e86 | 3106 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 3107 | caller); |
34342e86 | 3108 | prefetchw(objp); |
1da177e4 LT |
3109 | return objp; |
3110 | } | |
3111 | ||
e498be7d | 3112 | #ifdef CONFIG_NUMA |
c61afb18 | 3113 | /* |
b2455396 | 3114 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3115 | * |
3116 | * If we are in_interrupt, then process context, including cpusets and | |
3117 | * mempolicy, may not apply and should not be used for allocation policy. | |
3118 | */ | |
3119 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3120 | { | |
3121 | int nid_alloc, nid_here; | |
3122 | ||
765c4507 | 3123 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3124 | return NULL; |
3125 | nid_alloc = nid_here = numa_node_id(); | |
3126 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3127 | nid_alloc = cpuset_mem_spread_node(); | |
3128 | else if (current->mempolicy) | |
3129 | nid_alloc = slab_node(current->mempolicy); | |
3130 | if (nid_alloc != nid_here) | |
3131 | return __cache_alloc_node(cachep, flags, nid_alloc); | |
3132 | return NULL; | |
3133 | } | |
3134 | ||
765c4507 CL |
3135 | /* |
3136 | * Fallback function if there was no memory available and no objects on a | |
3137 | * certain node and we are allowed to fall back. We mimick the behavior of | |
3138 | * the page allocator. We fall back according to a zonelist determined by | |
3139 | * the policy layer while obeying cpuset constraints. | |
3140 | */ | |
3141 | void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) | |
3142 | { | |
3143 | struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
3144 | ->node_zonelists[gfp_zone(flags)]; | |
3145 | struct zone **z; | |
3146 | void *obj = NULL; | |
3147 | ||
3148 | for (z = zonelist->zones; *z && !obj; z++) | |
3149 | if (zone_idx(*z) <= ZONE_NORMAL && | |
3150 | cpuset_zone_allowed(*z, flags)) | |
3151 | obj = __cache_alloc_node(cache, | |
3152 | flags | __GFP_THISNODE, | |
3153 | zone_to_nid(*z)); | |
3154 | return obj; | |
3155 | } | |
3156 | ||
e498be7d CL |
3157 | /* |
3158 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3159 | */ |
a737b3e2 AM |
3160 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
3161 | int nodeid) | |
e498be7d CL |
3162 | { |
3163 | struct list_head *entry; | |
b28a02de PE |
3164 | struct slab *slabp; |
3165 | struct kmem_list3 *l3; | |
3166 | void *obj; | |
b28a02de PE |
3167 | int x; |
3168 | ||
3169 | l3 = cachep->nodelists[nodeid]; | |
3170 | BUG_ON(!l3); | |
3171 | ||
a737b3e2 | 3172 | retry: |
ca3b9b91 | 3173 | check_irq_off(); |
b28a02de PE |
3174 | spin_lock(&l3->list_lock); |
3175 | entry = l3->slabs_partial.next; | |
3176 | if (entry == &l3->slabs_partial) { | |
3177 | l3->free_touched = 1; | |
3178 | entry = l3->slabs_free.next; | |
3179 | if (entry == &l3->slabs_free) | |
3180 | goto must_grow; | |
3181 | } | |
3182 | ||
3183 | slabp = list_entry(entry, struct slab, list); | |
3184 | check_spinlock_acquired_node(cachep, nodeid); | |
3185 | check_slabp(cachep, slabp); | |
3186 | ||
3187 | STATS_INC_NODEALLOCS(cachep); | |
3188 | STATS_INC_ACTIVE(cachep); | |
3189 | STATS_SET_HIGH(cachep); | |
3190 | ||
3191 | BUG_ON(slabp->inuse == cachep->num); | |
3192 | ||
78d382d7 | 3193 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3194 | check_slabp(cachep, slabp); |
3195 | l3->free_objects--; | |
3196 | /* move slabp to correct slabp list: */ | |
3197 | list_del(&slabp->list); | |
3198 | ||
a737b3e2 | 3199 | if (slabp->free == BUFCTL_END) |
b28a02de | 3200 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3201 | else |
b28a02de | 3202 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3203 | |
b28a02de PE |
3204 | spin_unlock(&l3->list_lock); |
3205 | goto done; | |
e498be7d | 3206 | |
a737b3e2 | 3207 | must_grow: |
b28a02de PE |
3208 | spin_unlock(&l3->list_lock); |
3209 | x = cache_grow(cachep, flags, nodeid); | |
765c4507 CL |
3210 | if (x) |
3211 | goto retry; | |
1da177e4 | 3212 | |
765c4507 CL |
3213 | if (!(flags & __GFP_THISNODE)) |
3214 | /* Unable to grow the cache. Fall back to other nodes. */ | |
3215 | return fallback_alloc(cachep, flags); | |
3216 | ||
3217 | return NULL; | |
e498be7d | 3218 | |
a737b3e2 | 3219 | done: |
b28a02de | 3220 | return obj; |
e498be7d CL |
3221 | } |
3222 | #endif | |
3223 | ||
3224 | /* | |
3225 | * Caller needs to acquire correct kmem_list's list_lock | |
3226 | */ | |
343e0d7a | 3227 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3228 | int node) |
1da177e4 LT |
3229 | { |
3230 | int i; | |
e498be7d | 3231 | struct kmem_list3 *l3; |
1da177e4 LT |
3232 | |
3233 | for (i = 0; i < nr_objects; i++) { | |
3234 | void *objp = objpp[i]; | |
3235 | struct slab *slabp; | |
1da177e4 | 3236 | |
6ed5eb22 | 3237 | slabp = virt_to_slab(objp); |
ff69416e | 3238 | l3 = cachep->nodelists[node]; |
1da177e4 | 3239 | list_del(&slabp->list); |
ff69416e | 3240 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3241 | check_slabp(cachep, slabp); |
78d382d7 | 3242 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3243 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3244 | l3->free_objects++; |
1da177e4 LT |
3245 | check_slabp(cachep, slabp); |
3246 | ||
3247 | /* fixup slab chains */ | |
3248 | if (slabp->inuse == 0) { | |
e498be7d CL |
3249 | if (l3->free_objects > l3->free_limit) { |
3250 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3251 | /* No need to drop any previously held |
3252 | * lock here, even if we have a off-slab slab | |
3253 | * descriptor it is guaranteed to come from | |
3254 | * a different cache, refer to comments before | |
3255 | * alloc_slabmgmt. | |
3256 | */ | |
1da177e4 LT |
3257 | slab_destroy(cachep, slabp); |
3258 | } else { | |
e498be7d | 3259 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3260 | } |
3261 | } else { | |
3262 | /* Unconditionally move a slab to the end of the | |
3263 | * partial list on free - maximum time for the | |
3264 | * other objects to be freed, too. | |
3265 | */ | |
e498be7d | 3266 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3267 | } |
3268 | } | |
3269 | } | |
3270 | ||
343e0d7a | 3271 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3272 | { |
3273 | int batchcount; | |
e498be7d | 3274 | struct kmem_list3 *l3; |
ff69416e | 3275 | int node = numa_node_id(); |
1da177e4 LT |
3276 | |
3277 | batchcount = ac->batchcount; | |
3278 | #if DEBUG | |
3279 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3280 | #endif | |
3281 | check_irq_off(); | |
ff69416e | 3282 | l3 = cachep->nodelists[node]; |
873623df | 3283 | spin_lock(&l3->list_lock); |
e498be7d CL |
3284 | if (l3->shared) { |
3285 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3286 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3287 | if (max) { |
3288 | if (batchcount > max) | |
3289 | batchcount = max; | |
e498be7d | 3290 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3291 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3292 | shared_array->avail += batchcount; |
3293 | goto free_done; | |
3294 | } | |
3295 | } | |
3296 | ||
ff69416e | 3297 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3298 | free_done: |
1da177e4 LT |
3299 | #if STATS |
3300 | { | |
3301 | int i = 0; | |
3302 | struct list_head *p; | |
3303 | ||
e498be7d CL |
3304 | p = l3->slabs_free.next; |
3305 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3306 | struct slab *slabp; |
3307 | ||
3308 | slabp = list_entry(p, struct slab, list); | |
3309 | BUG_ON(slabp->inuse); | |
3310 | ||
3311 | i++; | |
3312 | p = p->next; | |
3313 | } | |
3314 | STATS_SET_FREEABLE(cachep, i); | |
3315 | } | |
3316 | #endif | |
e498be7d | 3317 | spin_unlock(&l3->list_lock); |
1da177e4 | 3318 | ac->avail -= batchcount; |
a737b3e2 | 3319 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3320 | } |
3321 | ||
3322 | /* | |
a737b3e2 AM |
3323 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3324 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3325 | */ |
873623df | 3326 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3327 | { |
9a2dba4b | 3328 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3329 | |
3330 | check_irq_off(); | |
3331 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3332 | ||
873623df | 3333 | if (cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3334 | return; |
3335 | ||
1da177e4 LT |
3336 | if (likely(ac->avail < ac->limit)) { |
3337 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3338 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3339 | return; |
3340 | } else { | |
3341 | STATS_INC_FREEMISS(cachep); | |
3342 | cache_flusharray(cachep, ac); | |
e498be7d | 3343 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3344 | } |
3345 | } | |
3346 | ||
3347 | /** | |
3348 | * kmem_cache_alloc - Allocate an object | |
3349 | * @cachep: The cache to allocate from. | |
3350 | * @flags: See kmalloc(). | |
3351 | * | |
3352 | * Allocate an object from this cache. The flags are only relevant | |
3353 | * if the cache has no available objects. | |
3354 | */ | |
343e0d7a | 3355 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3356 | { |
7fd6b141 | 3357 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3358 | } |
3359 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3360 | ||
a8c0f9a4 | 3361 | /** |
b8008b2b | 3362 | * kmem_cache_zalloc - Allocate an object. The memory is set to zero. |
a8c0f9a4 PE |
3363 | * @cache: The cache to allocate from. |
3364 | * @flags: See kmalloc(). | |
3365 | * | |
3366 | * Allocate an object from this cache and set the allocated memory to zero. | |
3367 | * The flags are only relevant if the cache has no available objects. | |
3368 | */ | |
3369 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3370 | { | |
3371 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3372 | if (ret) | |
3373 | memset(ret, 0, obj_size(cache)); | |
3374 | return ret; | |
3375 | } | |
3376 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3377 | ||
1da177e4 LT |
3378 | /** |
3379 | * kmem_ptr_validate - check if an untrusted pointer might | |
3380 | * be a slab entry. | |
3381 | * @cachep: the cache we're checking against | |
3382 | * @ptr: pointer to validate | |
3383 | * | |
3384 | * This verifies that the untrusted pointer looks sane: | |
3385 | * it is _not_ a guarantee that the pointer is actually | |
3386 | * part of the slab cache in question, but it at least | |
3387 | * validates that the pointer can be dereferenced and | |
3388 | * looks half-way sane. | |
3389 | * | |
3390 | * Currently only used for dentry validation. | |
3391 | */ | |
343e0d7a | 3392 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3393 | { |
b28a02de | 3394 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3395 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3396 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3397 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3398 | struct page *page; |
3399 | ||
3400 | if (unlikely(addr < min_addr)) | |
3401 | goto out; | |
3402 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3403 | goto out; | |
3404 | if (unlikely(addr & align_mask)) | |
3405 | goto out; | |
3406 | if (unlikely(!kern_addr_valid(addr))) | |
3407 | goto out; | |
3408 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3409 | goto out; | |
3410 | page = virt_to_page(ptr); | |
3411 | if (unlikely(!PageSlab(page))) | |
3412 | goto out; | |
065d41cb | 3413 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3414 | goto out; |
3415 | return 1; | |
a737b3e2 | 3416 | out: |
1da177e4 LT |
3417 | return 0; |
3418 | } | |
3419 | ||
3420 | #ifdef CONFIG_NUMA | |
3421 | /** | |
3422 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3423 | * @cachep: The cache to allocate from. | |
3424 | * @flags: See kmalloc(). | |
3425 | * @nodeid: node number of the target node. | |
3426 | * | |
3427 | * Identical to kmem_cache_alloc, except that this function is slow | |
3428 | * and can sleep. And it will allocate memory on the given node, which | |
3429 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3430 | * New and improved: it will now make sure that the object gets |
3431 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3432 | */ |
343e0d7a | 3433 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3434 | { |
e498be7d CL |
3435 | unsigned long save_flags; |
3436 | void *ptr; | |
1da177e4 | 3437 | |
e498be7d CL |
3438 | cache_alloc_debugcheck_before(cachep, flags); |
3439 | local_irq_save(save_flags); | |
18f820f6 CL |
3440 | |
3441 | if (nodeid == -1 || nodeid == numa_node_id() || | |
a737b3e2 | 3442 | !cachep->nodelists[nodeid]) |
5c382300 AK |
3443 | ptr = ____cache_alloc(cachep, flags); |
3444 | else | |
3445 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3446 | local_irq_restore(save_flags); |
18f820f6 CL |
3447 | |
3448 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3449 | __builtin_return_address(0)); | |
1da177e4 | 3450 | |
e498be7d | 3451 | return ptr; |
1da177e4 LT |
3452 | } |
3453 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3454 | ||
dbe5e69d | 3455 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3456 | { |
343e0d7a | 3457 | struct kmem_cache *cachep; |
97e2bde4 MS |
3458 | |
3459 | cachep = kmem_find_general_cachep(size, flags); | |
3460 | if (unlikely(cachep == NULL)) | |
3461 | return NULL; | |
3462 | return kmem_cache_alloc_node(cachep, flags, node); | |
3463 | } | |
dbe5e69d | 3464 | EXPORT_SYMBOL(__kmalloc_node); |
1da177e4 LT |
3465 | #endif |
3466 | ||
3467 | /** | |
800590f5 | 3468 | * __do_kmalloc - allocate memory |
1da177e4 | 3469 | * @size: how many bytes of memory are required. |
800590f5 | 3470 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3471 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3472 | */ |
7fd6b141 PE |
3473 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3474 | void *caller) | |
1da177e4 | 3475 | { |
343e0d7a | 3476 | struct kmem_cache *cachep; |
1da177e4 | 3477 | |
97e2bde4 MS |
3478 | /* If you want to save a few bytes .text space: replace |
3479 | * __ with kmem_. | |
3480 | * Then kmalloc uses the uninlined functions instead of the inline | |
3481 | * functions. | |
3482 | */ | |
3483 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3484 | if (unlikely(cachep == NULL)) |
3485 | return NULL; | |
7fd6b141 PE |
3486 | return __cache_alloc(cachep, flags, caller); |
3487 | } | |
3488 | ||
7fd6b141 | 3489 | |
1d2c8eea | 3490 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3491 | void *__kmalloc(size_t size, gfp_t flags) |
3492 | { | |
871751e2 | 3493 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3494 | } |
3495 | EXPORT_SYMBOL(__kmalloc); | |
3496 | ||
7fd6b141 PE |
3497 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3498 | { | |
3499 | return __do_kmalloc(size, flags, caller); | |
3500 | } | |
3501 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3502 | |
3503 | #else | |
3504 | void *__kmalloc(size_t size, gfp_t flags) | |
3505 | { | |
3506 | return __do_kmalloc(size, flags, NULL); | |
3507 | } | |
3508 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3509 | #endif |
3510 | ||
1da177e4 LT |
3511 | /** |
3512 | * kmem_cache_free - Deallocate an object | |
3513 | * @cachep: The cache the allocation was from. | |
3514 | * @objp: The previously allocated object. | |
3515 | * | |
3516 | * Free an object which was previously allocated from this | |
3517 | * cache. | |
3518 | */ | |
343e0d7a | 3519 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3520 | { |
3521 | unsigned long flags; | |
3522 | ||
ddc2e812 PE |
3523 | BUG_ON(virt_to_cache(objp) != cachep); |
3524 | ||
1da177e4 | 3525 | local_irq_save(flags); |
873623df | 3526 | __cache_free(cachep, objp); |
1da177e4 LT |
3527 | local_irq_restore(flags); |
3528 | } | |
3529 | EXPORT_SYMBOL(kmem_cache_free); | |
3530 | ||
1da177e4 LT |
3531 | /** |
3532 | * kfree - free previously allocated memory | |
3533 | * @objp: pointer returned by kmalloc. | |
3534 | * | |
80e93eff PE |
3535 | * If @objp is NULL, no operation is performed. |
3536 | * | |
1da177e4 LT |
3537 | * Don't free memory not originally allocated by kmalloc() |
3538 | * or you will run into trouble. | |
3539 | */ | |
3540 | void kfree(const void *objp) | |
3541 | { | |
343e0d7a | 3542 | struct kmem_cache *c; |
1da177e4 LT |
3543 | unsigned long flags; |
3544 | ||
3545 | if (unlikely(!objp)) | |
3546 | return; | |
3547 | local_irq_save(flags); | |
3548 | kfree_debugcheck(objp); | |
6ed5eb22 | 3549 | c = virt_to_cache(objp); |
f9b8404c | 3550 | debug_check_no_locks_freed(objp, obj_size(c)); |
873623df | 3551 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3552 | local_irq_restore(flags); |
3553 | } | |
3554 | EXPORT_SYMBOL(kfree); | |
3555 | ||
343e0d7a | 3556 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3557 | { |
3dafccf2 | 3558 | return obj_size(cachep); |
1da177e4 LT |
3559 | } |
3560 | EXPORT_SYMBOL(kmem_cache_size); | |
3561 | ||
343e0d7a | 3562 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3563 | { |
3564 | return cachep->name; | |
3565 | } | |
3566 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3567 | ||
e498be7d | 3568 | /* |
0718dc2a | 3569 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3570 | */ |
343e0d7a | 3571 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3572 | { |
3573 | int node; | |
3574 | struct kmem_list3 *l3; | |
cafeb02e CL |
3575 | struct array_cache *new_shared; |
3576 | struct array_cache **new_alien; | |
e498be7d CL |
3577 | |
3578 | for_each_online_node(node) { | |
cafeb02e | 3579 | |
a737b3e2 AM |
3580 | new_alien = alloc_alien_cache(node, cachep->limit); |
3581 | if (!new_alien) | |
e498be7d | 3582 | goto fail; |
cafeb02e | 3583 | |
0718dc2a CL |
3584 | new_shared = alloc_arraycache(node, |
3585 | cachep->shared*cachep->batchcount, | |
a737b3e2 | 3586 | 0xbaadf00d); |
0718dc2a CL |
3587 | if (!new_shared) { |
3588 | free_alien_cache(new_alien); | |
e498be7d | 3589 | goto fail; |
0718dc2a | 3590 | } |
cafeb02e | 3591 | |
a737b3e2 AM |
3592 | l3 = cachep->nodelists[node]; |
3593 | if (l3) { | |
cafeb02e CL |
3594 | struct array_cache *shared = l3->shared; |
3595 | ||
e498be7d CL |
3596 | spin_lock_irq(&l3->list_lock); |
3597 | ||
cafeb02e | 3598 | if (shared) |
0718dc2a CL |
3599 | free_block(cachep, shared->entry, |
3600 | shared->avail, node); | |
e498be7d | 3601 | |
cafeb02e CL |
3602 | l3->shared = new_shared; |
3603 | if (!l3->alien) { | |
e498be7d CL |
3604 | l3->alien = new_alien; |
3605 | new_alien = NULL; | |
3606 | } | |
b28a02de | 3607 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3608 | cachep->batchcount + cachep->num; |
e498be7d | 3609 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3610 | kfree(shared); |
e498be7d CL |
3611 | free_alien_cache(new_alien); |
3612 | continue; | |
3613 | } | |
a737b3e2 | 3614 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3615 | if (!l3) { |
3616 | free_alien_cache(new_alien); | |
3617 | kfree(new_shared); | |
e498be7d | 3618 | goto fail; |
0718dc2a | 3619 | } |
e498be7d CL |
3620 | |
3621 | kmem_list3_init(l3); | |
3622 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3623 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3624 | l3->shared = new_shared; |
e498be7d | 3625 | l3->alien = new_alien; |
b28a02de | 3626 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3627 | cachep->batchcount + cachep->num; |
e498be7d CL |
3628 | cachep->nodelists[node] = l3; |
3629 | } | |
cafeb02e | 3630 | return 0; |
0718dc2a | 3631 | |
a737b3e2 | 3632 | fail: |
0718dc2a CL |
3633 | if (!cachep->next.next) { |
3634 | /* Cache is not active yet. Roll back what we did */ | |
3635 | node--; | |
3636 | while (node >= 0) { | |
3637 | if (cachep->nodelists[node]) { | |
3638 | l3 = cachep->nodelists[node]; | |
3639 | ||
3640 | kfree(l3->shared); | |
3641 | free_alien_cache(l3->alien); | |
3642 | kfree(l3); | |
3643 | cachep->nodelists[node] = NULL; | |
3644 | } | |
3645 | node--; | |
3646 | } | |
3647 | } | |
cafeb02e | 3648 | return -ENOMEM; |
e498be7d CL |
3649 | } |
3650 | ||
1da177e4 | 3651 | struct ccupdate_struct { |
343e0d7a | 3652 | struct kmem_cache *cachep; |
1da177e4 LT |
3653 | struct array_cache *new[NR_CPUS]; |
3654 | }; | |
3655 | ||
3656 | static void do_ccupdate_local(void *info) | |
3657 | { | |
a737b3e2 | 3658 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3659 | struct array_cache *old; |
3660 | ||
3661 | check_irq_off(); | |
9a2dba4b | 3662 | old = cpu_cache_get(new->cachep); |
e498be7d | 3663 | |
1da177e4 LT |
3664 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3665 | new->new[smp_processor_id()] = old; | |
3666 | } | |
3667 | ||
b5d8ca7c | 3668 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3669 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3670 | int batchcount, int shared) | |
1da177e4 | 3671 | { |
d2e7b7d0 | 3672 | struct ccupdate_struct *new; |
2ed3a4ef | 3673 | int i; |
1da177e4 | 3674 | |
d2e7b7d0 SS |
3675 | new = kzalloc(sizeof(*new), GFP_KERNEL); |
3676 | if (!new) | |
3677 | return -ENOMEM; | |
3678 | ||
e498be7d | 3679 | for_each_online_cpu(i) { |
d2e7b7d0 | 3680 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
a737b3e2 | 3681 | batchcount); |
d2e7b7d0 | 3682 | if (!new->new[i]) { |
b28a02de | 3683 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3684 | kfree(new->new[i]); |
3685 | kfree(new); | |
e498be7d | 3686 | return -ENOMEM; |
1da177e4 LT |
3687 | } |
3688 | } | |
d2e7b7d0 | 3689 | new->cachep = cachep; |
1da177e4 | 3690 | |
d2e7b7d0 | 3691 | on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); |
e498be7d | 3692 | |
1da177e4 | 3693 | check_irq_on(); |
1da177e4 LT |
3694 | cachep->batchcount = batchcount; |
3695 | cachep->limit = limit; | |
e498be7d | 3696 | cachep->shared = shared; |
1da177e4 | 3697 | |
e498be7d | 3698 | for_each_online_cpu(i) { |
d2e7b7d0 | 3699 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3700 | if (!ccold) |
3701 | continue; | |
e498be7d | 3702 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3703 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3704 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3705 | kfree(ccold); |
3706 | } | |
d2e7b7d0 | 3707 | kfree(new); |
2ed3a4ef | 3708 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3709 | } |
3710 | ||
b5d8ca7c | 3711 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3712 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3713 | { |
3714 | int err; | |
3715 | int limit, shared; | |
3716 | ||
a737b3e2 AM |
3717 | /* |
3718 | * The head array serves three purposes: | |
1da177e4 LT |
3719 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3720 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3721 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3722 | * bufctl chains: array operations are cheaper. |
3723 | * The numbers are guessed, we should auto-tune as described by | |
3724 | * Bonwick. | |
3725 | */ | |
3dafccf2 | 3726 | if (cachep->buffer_size > 131072) |
1da177e4 | 3727 | limit = 1; |
3dafccf2 | 3728 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3729 | limit = 8; |
3dafccf2 | 3730 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3731 | limit = 24; |
3dafccf2 | 3732 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3733 | limit = 54; |
3734 | else | |
3735 | limit = 120; | |
3736 | ||
a737b3e2 AM |
3737 | /* |
3738 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3739 | * allocation behaviour: Most allocs on one cpu, most free operations |
3740 | * on another cpu. For these cases, an efficient object passing between | |
3741 | * cpus is necessary. This is provided by a shared array. The array | |
3742 | * replaces Bonwick's magazine layer. | |
3743 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3744 | * to a larger limit. Thus disabled by default. | |
3745 | */ | |
3746 | shared = 0; | |
3747 | #ifdef CONFIG_SMP | |
3dafccf2 | 3748 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3749 | shared = 8; |
3750 | #endif | |
3751 | ||
3752 | #if DEBUG | |
a737b3e2 AM |
3753 | /* |
3754 | * With debugging enabled, large batchcount lead to excessively long | |
3755 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3756 | */ |
3757 | if (limit > 32) | |
3758 | limit = 32; | |
3759 | #endif | |
b28a02de | 3760 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3761 | if (err) |
3762 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3763 | cachep->name, -err); |
2ed3a4ef | 3764 | return err; |
1da177e4 LT |
3765 | } |
3766 | ||
1b55253a CL |
3767 | /* |
3768 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3769 | * necessary. Note that the l3 listlock also protects the array_cache |
3770 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3771 | */ |
3772 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
3773 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
3774 | { |
3775 | int tofree; | |
3776 | ||
1b55253a CL |
3777 | if (!ac || !ac->avail) |
3778 | return; | |
1da177e4 LT |
3779 | if (ac->touched && !force) { |
3780 | ac->touched = 0; | |
b18e7e65 | 3781 | } else { |
1b55253a | 3782 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
3783 | if (ac->avail) { |
3784 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3785 | if (tofree > ac->avail) | |
3786 | tofree = (ac->avail + 1) / 2; | |
3787 | free_block(cachep, ac->entry, tofree, node); | |
3788 | ac->avail -= tofree; | |
3789 | memmove(ac->entry, &(ac->entry[tofree]), | |
3790 | sizeof(void *) * ac->avail); | |
3791 | } | |
1b55253a | 3792 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
3793 | } |
3794 | } | |
3795 | ||
3796 | /** | |
3797 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3798 | * @unused: unused parameter |
1da177e4 LT |
3799 | * |
3800 | * Called from workqueue/eventd every few seconds. | |
3801 | * Purpose: | |
3802 | * - clear the per-cpu caches for this CPU. | |
3803 | * - return freeable pages to the main free memory pool. | |
3804 | * | |
a737b3e2 AM |
3805 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3806 | * again on the next iteration. | |
1da177e4 LT |
3807 | */ |
3808 | static void cache_reap(void *unused) | |
3809 | { | |
7a7c381d | 3810 | struct kmem_cache *searchp; |
e498be7d | 3811 | struct kmem_list3 *l3; |
aab2207c | 3812 | int node = numa_node_id(); |
1da177e4 | 3813 | |
fc0abb14 | 3814 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3815 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3816 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3817 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3818 | return; |
3819 | } | |
3820 | ||
7a7c381d | 3821 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
3822 | check_irq_on(); |
3823 | ||
35386e3b CL |
3824 | /* |
3825 | * We only take the l3 lock if absolutely necessary and we | |
3826 | * have established with reasonable certainty that | |
3827 | * we can do some work if the lock was obtained. | |
3828 | */ | |
aab2207c | 3829 | l3 = searchp->nodelists[node]; |
35386e3b | 3830 | |
8fce4d8e | 3831 | reap_alien(searchp, l3); |
1da177e4 | 3832 | |
aab2207c | 3833 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 3834 | |
35386e3b CL |
3835 | /* |
3836 | * These are racy checks but it does not matter | |
3837 | * if we skip one check or scan twice. | |
3838 | */ | |
e498be7d | 3839 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 3840 | goto next; |
1da177e4 | 3841 | |
e498be7d | 3842 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3843 | |
aab2207c | 3844 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 3845 | |
ed11d9eb | 3846 | if (l3->free_touched) |
e498be7d | 3847 | l3->free_touched = 0; |
ed11d9eb CL |
3848 | else { |
3849 | int freed; | |
1da177e4 | 3850 | |
ed11d9eb CL |
3851 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
3852 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
3853 | STATS_ADD_REAPED(searchp, freed); | |
3854 | } | |
35386e3b | 3855 | next: |
1da177e4 LT |
3856 | cond_resched(); |
3857 | } | |
3858 | check_irq_on(); | |
fc0abb14 | 3859 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 3860 | next_reap_node(); |
2244b95a | 3861 | refresh_cpu_vm_stats(smp_processor_id()); |
a737b3e2 | 3862 | /* Set up the next iteration */ |
cd61ef62 | 3863 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3864 | } |
3865 | ||
3866 | #ifdef CONFIG_PROC_FS | |
3867 | ||
85289f98 | 3868 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3869 | { |
85289f98 PE |
3870 | /* |
3871 | * Output format version, so at least we can change it | |
3872 | * without _too_ many complaints. | |
3873 | */ | |
1da177e4 | 3874 | #if STATS |
85289f98 | 3875 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3876 | #else |
85289f98 | 3877 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3878 | #endif |
85289f98 PE |
3879 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3880 | "<objperslab> <pagesperslab>"); | |
3881 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3882 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3883 | #if STATS |
85289f98 | 3884 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 3885 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 3886 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 3887 | #endif |
85289f98 PE |
3888 | seq_putc(m, '\n'); |
3889 | } | |
3890 | ||
3891 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3892 | { | |
3893 | loff_t n = *pos; | |
3894 | struct list_head *p; | |
3895 | ||
fc0abb14 | 3896 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3897 | if (!n) |
3898 | print_slabinfo_header(m); | |
1da177e4 LT |
3899 | p = cache_chain.next; |
3900 | while (n--) { | |
3901 | p = p->next; | |
3902 | if (p == &cache_chain) | |
3903 | return NULL; | |
3904 | } | |
343e0d7a | 3905 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3906 | } |
3907 | ||
3908 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3909 | { | |
343e0d7a | 3910 | struct kmem_cache *cachep = p; |
1da177e4 | 3911 | ++*pos; |
a737b3e2 AM |
3912 | return cachep->next.next == &cache_chain ? |
3913 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
3914 | } |
3915 | ||
3916 | static void s_stop(struct seq_file *m, void *p) | |
3917 | { | |
fc0abb14 | 3918 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3919 | } |
3920 | ||
3921 | static int s_show(struct seq_file *m, void *p) | |
3922 | { | |
343e0d7a | 3923 | struct kmem_cache *cachep = p; |
b28a02de PE |
3924 | struct slab *slabp; |
3925 | unsigned long active_objs; | |
3926 | unsigned long num_objs; | |
3927 | unsigned long active_slabs = 0; | |
3928 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3929 | const char *name; |
1da177e4 | 3930 | char *error = NULL; |
e498be7d CL |
3931 | int node; |
3932 | struct kmem_list3 *l3; | |
1da177e4 | 3933 | |
1da177e4 LT |
3934 | active_objs = 0; |
3935 | num_slabs = 0; | |
e498be7d CL |
3936 | for_each_online_node(node) { |
3937 | l3 = cachep->nodelists[node]; | |
3938 | if (!l3) | |
3939 | continue; | |
3940 | ||
ca3b9b91 RT |
3941 | check_irq_on(); |
3942 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3943 | |
7a7c381d | 3944 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
3945 | if (slabp->inuse != cachep->num && !error) |
3946 | error = "slabs_full accounting error"; | |
3947 | active_objs += cachep->num; | |
3948 | active_slabs++; | |
3949 | } | |
7a7c381d | 3950 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
3951 | if (slabp->inuse == cachep->num && !error) |
3952 | error = "slabs_partial inuse accounting error"; | |
3953 | if (!slabp->inuse && !error) | |
3954 | error = "slabs_partial/inuse accounting error"; | |
3955 | active_objs += slabp->inuse; | |
3956 | active_slabs++; | |
3957 | } | |
7a7c381d | 3958 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
3959 | if (slabp->inuse && !error) |
3960 | error = "slabs_free/inuse accounting error"; | |
3961 | num_slabs++; | |
3962 | } | |
3963 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3964 | if (l3->shared) |
3965 | shared_avail += l3->shared->avail; | |
e498be7d | 3966 | |
ca3b9b91 | 3967 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3968 | } |
b28a02de PE |
3969 | num_slabs += active_slabs; |
3970 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3971 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3972 | error = "free_objects accounting error"; |
3973 | ||
b28a02de | 3974 | name = cachep->name; |
1da177e4 LT |
3975 | if (error) |
3976 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3977 | ||
3978 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3979 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3980 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3981 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3982 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3983 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3984 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3985 | #if STATS |
b28a02de | 3986 | { /* list3 stats */ |
1da177e4 LT |
3987 | unsigned long high = cachep->high_mark; |
3988 | unsigned long allocs = cachep->num_allocations; | |
3989 | unsigned long grown = cachep->grown; | |
3990 | unsigned long reaped = cachep->reaped; | |
3991 | unsigned long errors = cachep->errors; | |
3992 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3993 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3994 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 3995 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 3996 | |
e498be7d | 3997 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 3998 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 3999 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4000 | node_frees, overflows); |
1da177e4 LT |
4001 | } |
4002 | /* cpu stats */ | |
4003 | { | |
4004 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4005 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4006 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4007 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4008 | ||
4009 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4010 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4011 | } |
4012 | #endif | |
4013 | seq_putc(m, '\n'); | |
1da177e4 LT |
4014 | return 0; |
4015 | } | |
4016 | ||
4017 | /* | |
4018 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4019 | * | |
4020 | * Output layout: | |
4021 | * cache-name | |
4022 | * num-active-objs | |
4023 | * total-objs | |
4024 | * object size | |
4025 | * num-active-slabs | |
4026 | * total-slabs | |
4027 | * num-pages-per-slab | |
4028 | * + further values on SMP and with statistics enabled | |
4029 | */ | |
4030 | ||
4031 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
4032 | .start = s_start, |
4033 | .next = s_next, | |
4034 | .stop = s_stop, | |
4035 | .show = s_show, | |
1da177e4 LT |
4036 | }; |
4037 | ||
4038 | #define MAX_SLABINFO_WRITE 128 | |
4039 | /** | |
4040 | * slabinfo_write - Tuning for the slab allocator | |
4041 | * @file: unused | |
4042 | * @buffer: user buffer | |
4043 | * @count: data length | |
4044 | * @ppos: unused | |
4045 | */ | |
b28a02de PE |
4046 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4047 | size_t count, loff_t *ppos) | |
1da177e4 | 4048 | { |
b28a02de | 4049 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4050 | int limit, batchcount, shared, res; |
7a7c381d | 4051 | struct kmem_cache *cachep; |
b28a02de | 4052 | |
1da177e4 LT |
4053 | if (count > MAX_SLABINFO_WRITE) |
4054 | return -EINVAL; | |
4055 | if (copy_from_user(&kbuf, buffer, count)) | |
4056 | return -EFAULT; | |
b28a02de | 4057 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4058 | |
4059 | tmp = strchr(kbuf, ' '); | |
4060 | if (!tmp) | |
4061 | return -EINVAL; | |
4062 | *tmp = '\0'; | |
4063 | tmp++; | |
4064 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4065 | return -EINVAL; | |
4066 | ||
4067 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4068 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4069 | res = -EINVAL; |
7a7c381d | 4070 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4071 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4072 | if (limit < 1 || batchcount < 1 || |
4073 | batchcount > limit || shared < 0) { | |
e498be7d | 4074 | res = 0; |
1da177e4 | 4075 | } else { |
e498be7d | 4076 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4077 | batchcount, shared); |
1da177e4 LT |
4078 | } |
4079 | break; | |
4080 | } | |
4081 | } | |
fc0abb14 | 4082 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4083 | if (res >= 0) |
4084 | res = count; | |
4085 | return res; | |
4086 | } | |
871751e2 AV |
4087 | |
4088 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4089 | ||
4090 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4091 | { | |
4092 | loff_t n = *pos; | |
4093 | struct list_head *p; | |
4094 | ||
4095 | mutex_lock(&cache_chain_mutex); | |
4096 | p = cache_chain.next; | |
4097 | while (n--) { | |
4098 | p = p->next; | |
4099 | if (p == &cache_chain) | |
4100 | return NULL; | |
4101 | } | |
4102 | return list_entry(p, struct kmem_cache, next); | |
4103 | } | |
4104 | ||
4105 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4106 | { | |
4107 | unsigned long *p; | |
4108 | int l; | |
4109 | if (!v) | |
4110 | return 1; | |
4111 | l = n[1]; | |
4112 | p = n + 2; | |
4113 | while (l) { | |
4114 | int i = l/2; | |
4115 | unsigned long *q = p + 2 * i; | |
4116 | if (*q == v) { | |
4117 | q[1]++; | |
4118 | return 1; | |
4119 | } | |
4120 | if (*q > v) { | |
4121 | l = i; | |
4122 | } else { | |
4123 | p = q + 2; | |
4124 | l -= i + 1; | |
4125 | } | |
4126 | } | |
4127 | if (++n[1] == n[0]) | |
4128 | return 0; | |
4129 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4130 | p[0] = v; | |
4131 | p[1] = 1; | |
4132 | return 1; | |
4133 | } | |
4134 | ||
4135 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4136 | { | |
4137 | void *p; | |
4138 | int i; | |
4139 | if (n[0] == n[1]) | |
4140 | return; | |
4141 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4142 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4143 | continue; | |
4144 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4145 | return; | |
4146 | } | |
4147 | } | |
4148 | ||
4149 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4150 | { | |
4151 | #ifdef CONFIG_KALLSYMS | |
4152 | char *modname; | |
4153 | const char *name; | |
4154 | unsigned long offset, size; | |
4155 | char namebuf[KSYM_NAME_LEN+1]; | |
4156 | ||
4157 | name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); | |
4158 | ||
4159 | if (name) { | |
4160 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4161 | if (modname) | |
4162 | seq_printf(m, " [%s]", modname); | |
4163 | return; | |
4164 | } | |
4165 | #endif | |
4166 | seq_printf(m, "%p", (void *)address); | |
4167 | } | |
4168 | ||
4169 | static int leaks_show(struct seq_file *m, void *p) | |
4170 | { | |
4171 | struct kmem_cache *cachep = p; | |
871751e2 AV |
4172 | struct slab *slabp; |
4173 | struct kmem_list3 *l3; | |
4174 | const char *name; | |
4175 | unsigned long *n = m->private; | |
4176 | int node; | |
4177 | int i; | |
4178 | ||
4179 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4180 | return 0; | |
4181 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4182 | return 0; | |
4183 | ||
4184 | /* OK, we can do it */ | |
4185 | ||
4186 | n[1] = 0; | |
4187 | ||
4188 | for_each_online_node(node) { | |
4189 | l3 = cachep->nodelists[node]; | |
4190 | if (!l3) | |
4191 | continue; | |
4192 | ||
4193 | check_irq_on(); | |
4194 | spin_lock_irq(&l3->list_lock); | |
4195 | ||
7a7c381d | 4196 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4197 | handle_slab(n, cachep, slabp); |
7a7c381d | 4198 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4199 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4200 | spin_unlock_irq(&l3->list_lock); |
4201 | } | |
4202 | name = cachep->name; | |
4203 | if (n[0] == n[1]) { | |
4204 | /* Increase the buffer size */ | |
4205 | mutex_unlock(&cache_chain_mutex); | |
4206 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4207 | if (!m->private) { | |
4208 | /* Too bad, we are really out */ | |
4209 | m->private = n; | |
4210 | mutex_lock(&cache_chain_mutex); | |
4211 | return -ENOMEM; | |
4212 | } | |
4213 | *(unsigned long *)m->private = n[0] * 2; | |
4214 | kfree(n); | |
4215 | mutex_lock(&cache_chain_mutex); | |
4216 | /* Now make sure this entry will be retried */ | |
4217 | m->count = m->size; | |
4218 | return 0; | |
4219 | } | |
4220 | for (i = 0; i < n[1]; i++) { | |
4221 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4222 | show_symbol(m, n[2*i+2]); | |
4223 | seq_putc(m, '\n'); | |
4224 | } | |
d2e7b7d0 | 4225 | |
871751e2 AV |
4226 | return 0; |
4227 | } | |
4228 | ||
4229 | struct seq_operations slabstats_op = { | |
4230 | .start = leaks_start, | |
4231 | .next = s_next, | |
4232 | .stop = s_stop, | |
4233 | .show = leaks_show, | |
4234 | }; | |
4235 | #endif | |
1da177e4 LT |
4236 | #endif |
4237 | ||
00e145b6 MS |
4238 | /** |
4239 | * ksize - get the actual amount of memory allocated for a given object | |
4240 | * @objp: Pointer to the object | |
4241 | * | |
4242 | * kmalloc may internally round up allocations and return more memory | |
4243 | * than requested. ksize() can be used to determine the actual amount of | |
4244 | * memory allocated. The caller may use this additional memory, even though | |
4245 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4246 | * The caller must guarantee that objp points to a valid object previously | |
4247 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4248 | * must not be freed during the duration of the call. | |
4249 | */ | |
1da177e4 LT |
4250 | unsigned int ksize(const void *objp) |
4251 | { | |
00e145b6 MS |
4252 | if (unlikely(objp == NULL)) |
4253 | return 0; | |
1da177e4 | 4254 | |
6ed5eb22 | 4255 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4256 | } |